Developments in Mathematics

Volume 73

Series Editors
Krishnaswami Alladi, Department of Mathematics, University of Florida,
Gainesville, FL, USA
Pham Huu Tiep, Department of Mathematics, Rutgers University, Piscataway, NJ,
USA
Loring W. Tu, Department of Mathematics, Tufts University, Medford, MA, USA

Aims and Scope

The Developments in Mathematics (DEVM) book series is devoted to publishing
well-written monographs within the broad spectrum of pure and applied mathe-
matics. Ideally, each book should be self-contained and fairly comprehensive in
treating a particular subject. Topics in the forefront of mathematical research that
present new results and/or a unique and engaging approach with a potential rela-
tionship to other fields are most welcome. High-quality edited volumes conveying
current state-of-the-art research will occasionally also be considered for publication.
The DEVM series appeals to a variety of audiences including researchers, postdocs,
and advanced graduate students.
Dorina Mitrea Irina Mitrea
• •

Marius Mitrea

Geometric Harmonic
Analysis II
Function Spaces Measuring Size
and Smoothness on Rough Sets

123
Dorina Mitrea Irina Mitrea
Department of Mathematics Department of Mathematics
Baylor University Temple University
Waco, TX, USA Philadelphia, PA, USA

Marius Mitrea
Department of Mathematics
Baylor University
Waco, TX, USA

ISSN 1389-2177 ISSN 2197-795X (electronic)

Developments in Mathematics
ISBN 978-3-031-13717-4 ISBN 978-3-031-13718-1 (eBook)
https://doi.org/10.1007/978-3-031-13718-1

Mathematics Subject Classification: 32A, 26B20, 31B, 35J, 42B

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Dedicated with love to our parents
Prefacing the Full Series

The current work is part of a series, comprised of five volumes, [133], [134], [135],
[136], [137]. In broad terms, the principal aim is to develop tools in Real and
Harmonic Analysis, of geometric measure theoretic flavor, capable of treating a
broad spectrum of boundary value problems formulated in rather general geometric
and analytic settings.
In Volume I ([133]) we establish a sharp version of Divergence Theorem (aka
Fundamental Theorem of Calculus) which allows for an inclusive class of vector fields
whose boundary trace is only assumed to exist in a nontangential pointwise sense.
Volume II ([134]) is concerned with function spaces measuring size and/or smooth-
ness, such as Hardy spaces, Besov spaces, Triebel–Lizorkin spaces, Sobolev spaces,
Morrey spaces, Morrey-Campanato spaces, spaces of functions of Bounded Mean
Oscillations, etc., in general geometric settings. Work here also highlights the close
interplay between differentiability properties of functions and singular integral operators.
The topic of singular integral operators is properly considered in Volume III
([135]), where we develop a versatile Calderón-Zygmund theory for singular integral
operators of convolution type (and with variable coefficient kernels) on uniformly
rectifiable sets in the Euclidean ambient, and the setting of Riemannian manifolds.
Applications to scattering by rough obstacles are also discussed in this volume.
In Volume IV ([136]) we focus on singular integral operators of boundary layer
type which enjoy more specialized properties (compared with generic, garden
variety singular integral operators treated earlier in Volume III). Applications to
Complex Analysis in several variables are subsequently presented, starting from the
realizations that many natural integral operators in this setting, such as the
Bochner-Martinelli operator, are actual particular cases of double layer potential
operators associated with the complex Laplacian.
In Volume V ([137]), where everything comes together, finer estimates for a
certain class of singular integral operators (of chord-dot-normal type) are produced
in a manner which indicates how their size is affected by the (infinitesimal and
global) flatness of the “surfaces” on which they are defined. Among the library of
double layer potential operators associated with a given second-order system, we
then identify those double layers which fall under this category of singular integral

vii
viii Prefacing the Full Series

operators. It is precisely for this subclass of double layer potentials that Fredholm
theory may then be implemented assuming the underlying domain has a compact
boundary, which is sufficiently flat at infinitesimal scales. For domains with
unbounded boundaries, this very category of double layer potentials may be out-
right inverted, using a Neumann series argument, assuming the “surface” in
question is sufficiently flat globally. In turn, this opens the door for solving a large
variety of boundary value problems for second-order systems (involving boundary
data from Muckenhoupt weighted Lebesgue spaces, Lorentz spaces, Hardy spaces,
Sobolev spaces, BMO, VMO, Morrey spaces, Hölder spaces, etc.) in a large class
of domains which, for example, are allowed to have spiral singularities (hence more
general than domains locally described as upper-graphs of functions). In the
opposite direction, we show that the boundary value problems formulated for
systems lacking such special layer potentials may fail to be Fredholm solvable even
for really tame domains, like the upper half-space, or the unit disk. Save for the
announcement [132], all principal results appear here in print for the first time.
We close with a short epilogue, attempting to place the work undertaken in this
series into a broader picture. The main goal is to develop machinery of geometric
harmonic analysis flavor capable of ultimately dealing with boundary value prob-
lems of a very general nature. One of the principal tools (indeed, the piecè de
résistance) in this regard is a new and powerful version of the Divergence Theorem,
devised in Volume I, whose very formulation has been motivated and shaped from
the outset by its eventual applications to Harmonic Analysis, Partial Differential
Equations, Potential Theory, and Complex Analysis. The fact that its footprints may
be clearly recognized in the makeup of such a diverse body of results, as presented
in Volumes II-V, serves as testament to the versatility and potency of our brand of
Divergence Theorem. Alas, our enterprise is multifaceted, so its ssuccess is cru-
cially dependent on many other factors. For one thing, it is necessary to develop a
robust Calderón-Zygmund theory for singular integrals of boundary layer type (as
we do in Volumes III-IV), associated with generic weakly elliptic systems, capable
of accommodating a large variety of function spaces of interest considered in rather
inclusive geometric settings (of the sort discussed in Volume II). This renders these
(boundary-to-domain) layer potentials useful mechanisms for generating lots of
null-solutions for the given system of partial differential operators, whose format is
compatible with the demands in the very formulation of the boundary value
problem we seek to solve. Next, in order to be able to solve the boundary integral
equation to which matters are reduced in this fashion, the success of employing
Fredholm theory hinges on the ability to suitably estimate the essential norms of the
(boundary-to-boundary) layer potentials. In this vein, we succeed in relating the
distance from such layer potentials to the space of compact operators to the flatness
of the boundary of the domain in question (measured in terms of infinitesimal mean
oscillations of the unit normal) in a desirable manner which shows that, in a precise
quantitative fashion, the flatter the domain the smaller the proximity to compact
operators. This subtle and powerful result, bridging between analysis and geom-
etry, may be regarded as a far-reaching extension of the pioneering work of Radon
and Carleman in the early 1900’s.
Prefacing the Full Series ix

Ultimately, our work aligns itself with the program stemming from
A.P. Calderón’s 1978 ICM plenary address in which he advocates the use of layer
potentials “for much more general elliptic systems [than the Laplacian]” – see [24,
p.90], and may be regarded as an optimal extension of the pioneering work of E.B.
Fabes, M. Jodeit, and N.M. Rivière in [50] (where layer potential methods have
been first used to solve boundary value problems for the Laplacian in bounded C1
domains). In this endeavor, we have been also motivated by the problem1 posed by
A.P. Calderón on [24, p. 95], asking to identify the function spaces on which
singular integral operators (of boundary layer type) are well defined and continuous.
This is relevant since, as Calderón mentions, “A clarification of this question would
be very important in the study of boundary value problems for elliptic equations [in
rough domains]. The methods employed so far seem to be insufficient for the
treatment of these problems.” We also wish to mention that our work is also in line
with the issue raised as an open problem by C. Kenig in [113, Problem 3.2.2, pp.
116–117], where he asked whether operators of layer potential type may be inverted
on appropriate Lebesgue and Sobolev spaces in suitable subclasses on NTA
domains with compact Ahlfors regular boundaries.
The task of making geometry and analysis work in unison is fraught with diffi-
culties, and only seldom can a two-way street be built on which to move between
these two worlds without loss of information. Given this, it is actually surprising that
in many instances we come very close to having optimal hypotheses, almost an
accurate embodiment of the slogan if it makes sense to write it, then it’s true.
Acknowledgments: The authors gratefully acknowledge partial support from the
Simons Foundation (through grants # 426669, # 958374, # 318658, # 616050,
# 637481), as well as NSF (grant # 1900938). Portions of this work have been
completed at Baylor University in Waco, Temple University in Philadelphia, the
Institute for Advanced Study in Princeton, MSRI in Berkeley, and the American
Institute of Mathematics in San Jose. We wish to thank these institutions for their
generous hospitality. Last, but not least, we are grateful to Michael E. Taylor for
gently yet persistently encouraging us over the years to complete this project.

Waco, USA Dorina Mitrea

Philadelphia, USA Irina Mitrea
Waco, USA Marius Mitrea

1
In the last section of [24], simply titled “Problems,” Calderón singles two directions for further
study. The first one is the famous question whether the smallness condition on ka0 kL1 (the
Lipschitz constant of the curve fðx; aðxÞÞ : x 2 Rg on which he proved the L2 -boundedness
of the Cauchy operator) may be removed (as is well known, this has been solved in the
affirmative by Coifman, McIntosh, and Meyer in [32]). We are referring here to the second (and
final) problem formulated by Calderón on [24, p. 95].
Description of Volume II

An enthralling aspect of modern mathematics is the study of how analysis and

geometry affect one another. Webster’s dictionary defines the word ‘analysis’ as a
breaking up of a whole into its parts as to find out their nature. This point of view is
emblematic of a fundamental principle in Harmonic Analysis, in which one
decomposes a mathematical entity (such as a function/distribution, or an operator)
into simpler building blocks (enjoying certain specialized properties, typically
having to do with localization, cancellation, and size), analyzing these smaller
pieces individually, and then organizing this compartmentalized information in a
global, coherent manner, in order to derive conclusions about the original object of
study. Such a thesis goes at least as far back as the ground-breaking work of
J. Fourier in the early 1800’s, who had the vision of superimposing sine and cosine
curves with various amplitudes in an attempt to obtain the shape of the graph of a
fairly arbitrary given function. In this endeavor, the goal is then to understand the
quantitative correlation between the nature of said amplitudes (aka Fourier coeffi-
cients) on the one hand, and the functional-analytic properties of the original
function (such as membership to L2 or Lp ), on the other hand. This technology, as a
means of measuring size and smoothness for a given function, has received further
impetus through the advent of the Littlewood–Paley theory (particularly in relation
to the Lp -setting with p 6¼ 2), eventually culminating in the development of the
modern theory of function spaces of Triebel–Lizorkin and Besov type.
Yet another manifestation of Fourier’s pioneering ideas, that has fundamentally
altered the landscape of contemporary Harmonic Analysis, is the theory of Hardy
spaces viewed through the lenses of atomic and molecular techniques. In this sce-
nario, the so-called atoms and molecules play the role of the sine and cosine building
blocks (though, this times, they form an “overdetermined basis” as opposed to a
genuine linear basis). Originally considered in the work of R.R. Coifman in [31]
(on the real line), and R.H. Latter in [118] (in the higher-dimensional setting), then
benefiting from advances due to many authors (see, e.g., [9], [18], [25], [36], [40],
[56], [62], [71], [100], [88], [121], [138], [195], [176], [181], [182], [205], as well as

xi
xii Description of Volume II

the references therein) this body of work has matured into a complex, multifaceted
theory, with deep and far-reaching implications in many areas of mathematics.
To put matters into a broader perspective it is worth recalling that the theory of
H p spaces has been originally developed as a bridge between complex function
theory and Fourier Analysis, two branches of mathematics which tightly interfaced
with one-another in the lower dimensional setting (where methods of complex
function theory such as Blaschke products and conformal mappings played a
decisive role).
It was natural that higher-dimensional extensions of this theory would be sought
involving systems of harmonic functions in the upper half-space satisfying natural
generalizations of the Cauchy-Riemann equations (cf. [178]) and having their
nontangential maximal functions in Lp . This is the route taken in [52], building on
the earlier work in [177], [175], [179]. In this theory of harmonic H p spaces the
focus shifts from the harmonic functions themselves to their boundary values,
themselves tempered distributions in Rn , from which the harmonic functions can
then be recovered via Poisson’s integral formula.
In their pioneering work in [52], C. Fefferman and E. Stein have shown that the
n-dimensional Hardy spaces in the Euclidean setting, developed in [177], have
purely real-variable characterizations as the space of tempered distributions in Rn
having some sort of maximal function belonging to Lp ðRn Þ. Such Ra maximal
operator may be fashioned out of a fixed Schwartz function U with U 6¼ 0 (or
even the classical Poisson kernel, as historically has been the case). Alternatively,
one can take into account “all” such possible approximate identities, thus creating
the so-called grand maximal function. More specifically, for each tempered dis-
tribution f 2 S0 ðRn Þ, its grand maximal function is defined at each x 2 Rn as

f ] ðxÞ :¼ sup sup jðf Ut ÞðxÞj; ð0:0:1Þ

U2A t [ 0

where A is a family of suitably normalized Schwartz functions defined in Rn , and

Ut :¼ tn Uð=tÞ, for each U 2 A and t 2 ð0; 1Þ.
Concerning the role of geometry, one fundamental development (from the per-
spective of the work undertaken here) has been the consideration of environments
much more general than the Euclidean ambient, through the introduction of the
so-called spaces of homogeneous type.2 Indeed, by the late 1970’s it has become
increasingly evident that a large portion of basic real analysis (such as covering
lemmas, the Hardy-Littlewood maximal operator, functions of bounded mean
oscillation, Lebesgue’s differentiation type theorem, Whitney decompositions, and
even singular integral operators of Calderón-Zygmund type, among other things)
requires relatively little structure from the geometric ambient. The emergence of the
theory of Hardy spaces in spaces of homogeneous type perhaps best typifies these
developments. For example, one may introduce Hardy spaces in such a setting by

2
Basic references in this regard are [35] and [36], which have retained their significance many
decades after appearing in print.
Description of Volume II xiii

naturally adapting the definition of the Fefferman–Stein grand maximal function

reviewed in (0.0.1), now considering for any “distribution” f the function defined
pointwise as
 
f ] ðxÞ :¼ sup hU; f i for each x: ð0:0:2Þ
U2T ðxÞ

Above, h; i is interpreted as the duality paring between the space of Lipschitz
functions with bounded support and its dual, while T ðxÞ (playing the role of the
family A in the earlier discussion) is a collection of suitably normalized “test
functions” concentrated near the point x (i.e., Hölder continuous functions, with a
bounded support containing x, relative to which their size is correlated).
This is the point of view adopted in Chapter 4 where, in keeping with the
program initiated by R. Coifman and G. Weiss in the 1970’s, we develop a theory
of (Lebesgue-based and Lorentz-based) Hardy spaces on Ahlfors regular subsets of
Rn , by considering basic topics such as: the Fefferman–Stein grand maximal
function, real interpolation of Lebesgue-based and Lorentz-based Hardy spaces,
atoms and molecules, duality, weak- convergence, and the compatibility of various
duality pairings, among other things.
The starting point in Chapter 5 is the consideration of a more inclusive brand of
Banach function space (than traditionally used in the literature; cf. e.g., [14]), which
we dub Generalized Banach Function Space. This is done in x5.1. The relevance of
this extension is that a variety of scales of spaces of interest, such as Morrey spaces,
block spaces, as well as Beurling algebras and their pre-duals, now fit naturally into
this more accommodating label. Most significantly, in x5.2 we develop powerful
and versatile extrapolation results serving as portal, allowing us to pass from
estimates on Muckenhoupt weighted Lebesgue spaces (for a fixed integrability
exponent and arbitrary weights) to estimates on the brand of Generalized Banach
Function Spaces introduced earlier, on which the Hardy-Littlewood maximal
operator happens to be bounded. Finally, in x5.3 we focus on Orlicz spaces which,
in particular, are natural examples of classical Banach function spaces for which the
machinery developed so far applies.
Next, in Chapter 6, we introduce a natural brand of Morrey-Campanato and Morrey
spaces on Ahlfors regular subsets of Rn . The function theory developed in relation to
these spaces (in x6.1 and x6.2) features: embeddings, density results, atomic
decompositions, duality, extrapolation, as well as the boundedness of the Hardy-
Littlewood maximal operator and of fractional integration operators on these spaces.
Chapter 7 is reserved for a discussion pertaining to the scales of Besov and
Triebel–Lizorkin spaces on Ahlfors regular sets in Rn . Modeled upon their
Euclidean counterparts, these scales of spaces are considered through the per-
spective of Littlewood–Paley theory (suitably adapted to the context of spaces of
homogeneous type). Topics covered here include: atomic and molecular theory,
Calderón reproducing formula and frame theory, real and complex interpolation,
duality and embeddings, quasi-Banach envelopes, and various intrinsic
characterizations.
xiv Description of Volume II

In Chapter 8 the main focus is on boundary traces from weighted Sobolev spaces
in open subsets of the Euclidean ambient (in which the weight is a suitable power
of the distance to the boundary) into Besov spaces (discussed in Chapter 7) that are
defined on the boundary of the domain in question. In this vein, the main issues of
interest are: the well-definiteness and boundedness of such a boundary trace
operator, the fact that said trace operator is surjective (which we establish via
constructive methods, by showing the existence of a linear and bounded extension
operator), and the explicit description of its null-space. Of basic importance for
future applications is that this program is carried out in a very general geometric
setting. A program of somewhat similar aims is executed in Chapter 9, this time
starting from Besov and Triebel–Lizorkin spaces in open subsets of the Euclidean
ambient (in lieu of the weighted Sobolev spaces employed earlier).
Going further, in Chapter 10 we study the strong (i.e., pointwise) and weak (i.e.,
distributional, in the sense of our “bullet product”, introduced in [133, §4.2])
normal traces of vector fields, seeking conditions guaranteeing membership to
(Lebesgue-based and Lorentz-based) Hardy spaces on the boundary of a given
rough domain. This task is first accomplished working in the Euclidean ambient (in
x10.1 and x10.2), then in the setting of Riemannian manifolds (in x10.3). Our brand
of Divergence Theorem developed in Volume I ([133]) is a basic ingredient in
proving such results. In turn, they play a crucial role in much of the subsequent
work, from Volume III ([135]) where we discuss Fatou-type results in geometric
settings of a desirable degree of generality, all the way to Volume V ([137]) where
normal traces (more specifically, directional derivatives along the unit normal on
the boundary of a given domain) feature prominently in the very formulation of the
Neumann and Transmission boundary value problems.
In Chapter 11, the final chapter in this volume, we expand on the work in [97],
[139], [141] by considering a scale of “boundary” Sobolev spaces, which may be
meaningfully defined on the geometric measure theoretic boundary of sets of
locally finite perimeter (both in the Euclidean setting and that of manifolds). It is of
interest to compare our brand of Sobolev spaces with other types of Sobolev spaces
on generic measure metric spaces which have been introduced and studied else-
where in the literature (see, e.g., [79], [81] and the references there). The notion of
“weak gradient” employed in defining the latter is, by necessity (given the envi-
ronment considered), of a rather algebraic nature (no differential calculus is
involved). In sharp contrast, our Sobolev spaces make essential use of “weak
derivatives” and integration by parts along the boundary (itself a consequence of the
Divergence Theorem, suitably applied). This renders them particularly useful and
natural in connection with problems in Partial Differential Equations. Moreover, we
do show that the two scales are compatible, in the sense that they agree with one
another when considered on the boundary of an Ahlfors regular domain satisfying a
two-sided local John condition.
In these endeavors we shall make crucial use of the Calderón–Zygmund theory
for singular integrals of layer potential type, developed in [136] (independently
of the present considerations). This is not entirely surprising: indeed, it has long
been known that there is a close interplay between singular integrals and
Description of Volume II xv

differentiability properties of functions (the 1970 monograph [175] of E.M. Stein, is

famously titled as such). While such considerations have, so far, been largely
restricted to the entire Euclidean space, here we take the first steps in the direction
of employing a similar philosophy for more general geometries, i.e., when the
ambient is a set of more general nature.
The discussion in this volume is facilitated by the preparatory material presented
in Chapter 1. Here we collect and develop basic results concerning vector spaces
and operator theory (in x1.1), quasi-normed spaces (in x1.2), real interpolation of
quasilinear operators (in x1.3), complex interpolation of quasi-Banach spaces
(in x1.4), as well as Köthe function spaces and other categories of topological
vector spaces (in x1.5). Chapter 2 is concerned with Fredholm theory, not only in
the standard context of Banach spaces, but also of topological vector spaces of a
much more general nature (which may lack local convexity). The latter aspect is
relevant in view of the fact that many mainstream scales of function spaces (such as
Hardy, Besov, and Triebel–Lizorkin) contain spaces which are not necessarily
Banach.
In this volume we also study functions whose oscillations vanish at infinitesimal
scales (in Chapter 3). When said oscillations are measured in a mean (i.e., integral)
sense, this gives rise to the Sarason space VMO, discussed in x3.1. When the
oscillations in question are taken in a pointwise sense (via a local Hölder
semi-norm), we come across a new brand of Hölder spaces, defined in x3.2.
Lastly, we wish to note that while in this volume we study function spaces as an
independent topic, something of interest in its own right, our work here has also
been motivated by the problem posed by A.P. Calderón on [24, p. 95], asking to
identify the function spaces on which singular integral operators (of boundary layer
type) are well defined and continuous. We shall look into this latter aspect in
subsequent volumes, bearing in mind that, as Calderón mentions, “A clarification of
this question would be very important in the study of boundary value problems for
elliptic equations [in rough domains]. The methods employed so far seem to be
insufficient for the treatment of these problems.”
Contents

Prefacing the Full Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Description of Volume II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
1 Preliminary Functional Analytic Matters . . . . . . . . . . . . . . . . . . . . 1
1.1 Algebraic and Operator Theoretic Rudiments . . . . . . . . . . . . . 1
1.2 Quasi-Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Real Interpolation of Quasilinear Operators . . . . . . . . . . . . . . . 25
1.4 Complex Interpolation on Quasi-Banach Spaces . . . . . . . . . . . 36
1.5 Köthe Function Spaces and Other Categories
of Topological Vector Spaces . . . . . . . . . . . . . ........... 55
2 Abstract Fredholm Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.1 Fredholm Theory in Banach Spaces . . . . . . . . . . . . . . . . . . . . 63
2.2 Fredholm Theory in Topological Vector Spaces . . . . . . . . . . . 73
3 Functions of Vanishing Mean Oscillations and Vanishing
Hölder Moduli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.1 Functions of Vanishing Mean Oscillations on Measure
Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3.2 A New Brand of Hölder Spaces . . . . . . . . . . . . . . . . . . . . . . . 98
4 Hardy Spaces on Ahlfors Regular Sets . . . . . . . . . . . . . . . . . . . . . . 107
4.1 The Fefferman-Stein Grand Maximal Function . . . . . . . . . . . . 107
4.2 Lebesgue-Based and Lorentz-Based Hardy Spaces
on Ahlfors Regular Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.3 Real Interpolation of Hardy Spaces . . . . . . . . . . . . . . . . . . . . 119
4.4 Atomic Decompositions for Hardy Spaces . . . . . . . . . . . . . . . 143
4.5 Molecules in Hardy Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 174
4.6 Duality in Hardy Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
4.7 Richness of the Duals of Lorentz-Based Hardy Spaces . . . . . . 191

xvii
xviii Contents

4.8 Weak- Convergence and More on the Compatibility

of Pairings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
4.9 More on H p Versus Lp : The Filtering Operator . . . . . . . . . . . . 212
5 Banach Function Spaces, Extrapolation, and Orlicz Spaces . . . . . . 217
5.1 Generalized Banach Function Spaces . . . . . . . . . . . . . . . . . . . 217
5.2 Extrapolation Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
5.3 Young Functions and Orlicz Spaces . . . . . . . . . . . . . . . . . . . . 252
6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals
on Ahlfors Regular Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
6.1 Morrey-Campanato Spaces and Their Pre-Duals
on Ahlfors Regular Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular
Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets . . . . . . 363
7.1 Definitions with Sharp Ranges of Indices and Basic
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
7.2 Atomic and Molecular Theory . . . . . . . . . . . . . . . . . . . . . . . . 373
7.3 Calderón’s Reproducing Formula and Frame Theory . . . . . . . . 382
7.4 Interpolation of Besov and Triebel-Lizorkin Spaces
via the Real Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391
7.5 Complex Interpolation of Besov and Triebel-Lizorkin
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
7.6 Duality Results for Besov and Triebel-Lizorkin Spaces . . . . . . 405
7.7 Loose and Tight Embeddings . . . . . . . . . . . . . . . . . . . . . . . . . 408
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces . . . . . . . . . . 417
7.9 Intrinsic Characterizations of Besov Spaces . . . . . . . . . . . . . . 446
8 Boundary Traces from Weighted Sobolev Spaces into Besov
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 455
8.1 Traces from Weighted Sobolev Spaces Defined in Rn . . . . . . . 456
8.2 Technical Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 458
8.3 Traces from Weighted Sobolev Spaces Defined
in Open Subsets of Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467
8.4 Extension Operators from Boundary Besov Spaces
into Weighted Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . 481
8.5 Weighted Sobolev Spaces of Negative Order, Duality,
and the Conormal Derivative Operator . . . . . . . . . . . . . . . . . . 485
8.6 Traces from Weighted Maximal Sobolev Spaces
in Open Subsets of Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 498
9 Besov and Triebel-Lizorkin Spaces in Open Sets . . . . . . . . . . . . . . 517
9.1 Besov and Triebel-Lizorkin Spaces in Rn . . . . . . . . . . . . . . . . 517
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of Rn . . . 527
Contents xix

9.3 Envelopes of Besov and Triebel-Lizorkin Spaces

in Open Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
9.4 Traces of Functions from Besov and Triebel-Lizorkin
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572
9.5 The Conormal Derivative Operator on Besov
and Triebel-Lizorkin Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 592
9.6 Extension from Boundary Besov Spaces into Smoothness
Spaces in Open Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 602
10 Strong and Weak Normal Boundary Traces of Vector Fields
in Hardy and Morrey Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
10.1 Strong Normal Boundary Traces of Vector Fields
in Hardy Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 607
10.2 Weak Normal Boundary Traces of Vector Fields
in Hardy and Morrey Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 634
10.3 The Manifold Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 679
11 Sobolev Spaces on the Geometric Measure Theoretic Boundary
of Sets of Locally Finite Perimeter . . . . . . . . . . . . . . . . . . . . . . . . . 681
11.1 Weak Tangential Derivatives and Boundary Sobolev
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 682
11.2 Distributional Tangential Derivatives . . . . . . . . . . . . . . . . . . . 702
11.3 Distributional Derivatives Versus Weak Tangential
Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 705
11.4 The Tangential Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
11.5 Sobolev Spaces on Boundaries of Ahlfors Regular
Domains Satisfying a Two-Sided Local John Condition . . . . . 736
11.6 Boundary Sobolev Spaces on Manifolds . . . . . . . . . . . . . . . . . 784
11.7 A General Recipe for Sobolev-Like Spaces . . . . . . . . . . . . . . 791
11.8 Boundary Sobolev Spaces of Negative Smoothness . . . . . . . . . 796
11.9 Principal-Value Distributions on Countably Rectifiable
Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 804
11.10 Hardy-Based Boundary Sobolev Spaces . . . . . . . . . . . . . . . . . 811
11.11 Embedding Results Involving Boundary Sobolev Spaces . . . . . 824
11.12 Real Interpolation Involving Boundary Sobolev Spaces . . . . . . 831
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev
Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 848

Appendix A: Terms and Notation used in Volume II. . . . . . . . . . . . . . . . 869

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 905
Subject Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915
Symbol Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 921
Chapter 1
Preliminary Functional Analytic Matters

To set the stage for much of the subsequent discussion, in this chapter we elaborate
on basic terminology and results pertaining to vector spaces and operator theory
(in §1.1), quasi-normed spaces (in §1.2), real interpolation of quasilinear operators
(in §1.3), complex interpolation of quasi-Banach spaces (in §1.4), and some useful
categories of topological vector spaces (in §1.5).

1.1 Algebraic and Operator Theoretic Rudiments

For any given vector space X we denote by dim X its dimension. Let X be a vector
space, an suppose Y is a linear subspace of X. Denote by X/Y the quotient space,
consisting of all classes of equivalences [x]X/Y = x + Y with x ∈ X, naturally
regarded as a vector space.

Lemma 1.1.1 Let X be a vector space and suppose V, W are subspaces of X such
that V ⊆ W and      
dim X V = dim X W < +∞. (1.1.1)

Then necessarily V = W.
 
Proof Consider the mapping ι : X V → X W given by ι(x + V) := x + W for
every x ∈ X. Then the fact that V ⊆ W implies that ι is a well-defined, linear
mapping. Clearly, ι is surjective which, in concert with (1.1.1), actually forces ι to be
a bijection. Writing out what it means for ι to be injective then yields the inclusion
W ⊆ V. Thus, ultimately, V = W. 

Remark 1.1.2 Assume X, Y, Z are three given vector spaces with the property that
Z ⊆ Y ⊆ X. Then
dim(X/Z) = dim(X/Y ) + dim(Y /Z). (1.1.2)
As a consequence,
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 1
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_1
2 1 Preliminary Functional Analytic Matters

if dim(X/Z) < +∞ then Z = Y

(1.1.3)
if and only if dim(X/Z) = dim(X/Y ).

A proof of the following result may be found in [142].

Theorem 1.1.3 Let X be a Hausdorff linear topological space.

(1) If Y is a closed linear subspace of X and F is a finite-dimensional linear subspace

of X, then Y + F is closed in X.
(2) If Y is a closed linear subspace of X of finite codimension and F is any algebraic
complement of Y , then X = Y ⊕ F (direct topological sum).
(3) If Z is a closed subspace of finite codimension in X, then any linear subspace Y
of X with the property that Z ⊆ Y is necessarily closed and of finite codimension
in X.

We next discuss a few generalities of functional analytic nature. For further

reference, for each linear map T : X → Y where X, Y are vector spaces we shall let

Im(T : X → Y ) := {T x : x ∈ X } and
(1.1.4)
Ker(T : X → Y ) := {x ∈ X : T x = 0}

denote, respectively, the image (or range) and kernel (or null-space) of the operator
T. Moreover, in order to lighten notation, we shall occasionally simply write Im T,
Ker T when no confusion is likely to occur. Call T a finite-rank operator1 (or an
operator of finite rank, or an operator having finite rank) if Im(T : X → Y ) is a finite
dimensional subspace of Y .
We shall often work in the category of topological vector spaces (or linear topo-
logical spaces). As far as this setting is concerned, we first wish to note (see, e.g.,
the discussion in [142]) that
if X is a Hausdorff linear topological space with n := dim X < ∞, then
(1.1.5)
X is linearly homeomorphic to Cn (with the standard topology).
Hence,
on a finite dimensional vector space X over the field C there could
only be one Hausdorff vector topology; specifically, if {x1, . . . , xn } is
(1.1.6)
a basis in X, a set O ⊆ X is open if and only if there exist O1, . . . , On
open sets in C such that O = O1 x1 + · · · + On xn .
Recall that, given a topological vector space X,
a set A ⊆ X is said to be (topologically) bounded if for each zero-
neighborhood U in X there exists εU > 0 such that ε A ⊆ U for (1.1.7)
each ε ∈ (0, εU ].

1 we emphasize that, even when the spaces involved are Banach, an operator of finite rank is not
necessarily continuous
1.1 Algebraic and Operator Theoretic Rudiments 3

Hence, a set in a topological vector space is (topologically) bounded provided it is

absorbed by any zero-neighborhood of said space. Clearly, any subset of a bounded
set is itself bounded. Let us also recall that
a linear topological space X is said to be locally
(1.1.8)
bounded if it has a bounded zero-neighborhood.
If X, Y are two linear topological spaces, a linear operator T : X → Y is said to
be bounded if there exists a zero-neighborhood U of X such that T(U) is a bounded
subset of Y . Denote the set of bounded operators from X into Y by B(X → Y ), i.e.,
 
B(X → Y ) := T : X −→ Y : T is linear and bounded . (1.1.9)

From (1.1.9) and (1.1.7) we see that any operator T ∈ B(X → Y ) maps bounded
sets into bounded sets. Under the additional assumption that X is locally bounded,
the converse implication is also true. If we also consider
 
L(X → Y ) := T : X −→ Y : T is linear and continuous , (1.1.10)

then it is clear from definitions that every bounded operator is continuous, i.e., we
have2
B(X → Y ) ⊆ L(X → Y ). (1.1.11)
In general, continuous operators need not be bounded, though one may check without
difficulty that
if X, Y are two topological vector spaces, with X locally bounded, then
any linear continuous operator T : X → Y is bounded, i.e., we actually (1.1.12)
have B(X → Y ) = L(X → Y ) in this case.
In addition to (1.1.9) and (1.1.10), for any two topological vector spaces X, Y let
us also define3
 
Cp(X → Y ) := T : X → Y : T is a linear compact mapping (1.1.13)

and abbreviate

Cp(X) := Cp(X → X). (1.1.14)

It is known (cf. [142]) that

if X is a Hausdorff topological vector space, then every
(1.1.15)
continuous finite-rank linear operator on X is compact.

2 A companion result states that if X is a metrizable locally convex topological vector space and
Y is a locally convex topological vector space, and if T : X → Y is a linear operator mapping
(topologically) bounded sets into (topologically) bounded sets, then T ∈ L(X → Y).
3 recall that a mapping is said to be compact if it maps (topologically) bounded sets into relatively
compact sets
4 1 Preliminary Functional Analytic Matters

Straight from definitions it may be checked that every compact operator is bounded,
i.e., for any two topological vector spaces X, Y we have

Cp(X → Y ) ⊆ B(X → Y ). (1.1.16)

Also, it is clear from definitions that

if X is a topological vector space, T ∈ Cp(X), and Y is a closed
subspace of X which is invariant under T, then T Y ∈ Cp(Y ).
(1.1.17)

Given a topological vector space X, define its (topological) dual as

 
X ∗ := L(X → C) = Λ : X −→ C : Λ linear and continuous . (1.1.18)

Clearly, X ∗ is a vector space, and, as may be seen from (1.1.12),

if X is a locally bounded topological vector space
(1.1.19)
then one has X ∗ = L(X → C) = B(X → C).

Suppose X, Y are two topological vector spaces. For each T ∈ L(X → Y ) define

T ∗ : Y ∗ −→ X ∗, T ∗ (Λ) := Λ ◦ T for each Λ ∈ Y ∗ . (1.1.20)

Hence, T ∗ is a well-defined linear operator. In relation to this, we have the following

useful result.
Lemma 1.1.4 If X, Y are two topological vector spaces, with Y Hausdorff, and
T ∈ L(X → Y ) is a finite-rank operator, then T ∗ : Y ∗ → X ∗ is also a finite-rank
operator.
Proof Let {y1, . . . , yn } be a basis in Im(T : X → Y ). Hence, for each x ∈ X there
exist unique numbers Λi (x) ∈ C, for 1 ≤ i ≤ n, such that

n
Tx = Λi (x)yi . (1.1.21)
i=1

From the linearity and continuity of T : X → Im(T : X → Y ), the fact that the
latter is a finite dimensional Hausdorff topological vector space, and (1.1.6) we then
conclude that each assignment X x → Λi (x) ∈ C is linear and continuous. Hence,
Λi ∈ X ∗ for each {1, . . . , n}. Next, for any given Λ ∈ Y ∗ we rely on (1.1.20) and
(1.1.21) to compute

n
(T ∗ Λ)(x) = (Λ ◦ T)(x) = Λ(T x) = Λi (x)Λ(yi )
i=1

n 
= Λ(yi )Λi (x) for each x ∈ X. (1.1.22)
i=1

Hence,
1.2 Quasi-Normed Spaces 5


n
T ∗Λ = Λ(yi )Λi for each Λ ∈ Y ∗, (1.1.23)
i=1

which goes to show that Im(T ∗ : Y ∗ → X ∗ ) has dimension at most n. 

1.2 Quasi-Normed Spaces

For a quasi-normed space (X, · X ), we denote by ρ = ρ(X) its modulus of

concavity, i.e., the smallest positive constant for which

x+y X ≤ ρ(X)( x X + y X ), ∀x, y ∈ X. (1.2.1)

Note that always ρ(X) ≥ 1. Also, for any linear subspace Y of X, regarded as a
quasi-normed space equipped with the quasi-norm · X , we have ρ(Y ) ≤ ρ(X). Call
(X, · X ) a quasi-Banach space provided X is complete wit respect to · X . It
follows that

any closed linear subspace of a quasi-Banach space is itself a quasi-Banach.

(1.2.2)
Given p ∈ (0, 1], a p-norm on a vector space X is a quasi-norm · on X which
satisfies
x + y p ≤ x p + y p for all x, y ∈ X. (1.2.3)
It follows that
if · is a p-norm, for some p ∈ (0, 1], on a vector space X
(1.2.4)
then x + y ≤ 2(1/p)−1 x + y for each x, y ∈ X.

Also, a quasi-normed space X is said to be a p-Banach space for some p ∈ (0, 1]

provided there exists a p-norm which is equivalent to the original quasi-norm on
X, and with respect to which X is complete. We recall the Aoki-Rolewicz theorem,
which asserts (cf., e.g., [112], [138]) that
for each quasi-normed space X there exists an equivalent
p-norm where p ∈ (0, 1] is such that 2(1/p)−1 = ρ(X), i.e., (1.2.5)
when p := (1 + log2 ρ(X))−1 .

Note that by raising both sides of the inequality in (1.2.3) to an arbitrary power
θ ∈ (0, 1) and then using the general fact that (a + b)θ < aθ + bθ for each a, b ∈ [0, ∞)
leads to the conclusion that for any quasi-normed space X, · we have

x+y r
≤ x r
+ y r
for all r ∈ 0, (1 + log2 ρ(X))−1 and all x, y ∈ X. (1.2.6)

In general a quasi-norm need not be continuous but any p-norm is. Indeed, using the
elementary observation that
6 1 Preliminary Functional Analytic Matters
γ−1
|aγ − bγ | ≤ γ|a − b| · max{a, b} if a, b ∈ (0, ∞) and γ ≥ 1, (1.2.7)

for each x, y ∈ X we may estimate, using (1.2.7) with γ := 1/p ≥ 1 and (1.2.3),
   
 x − y  = ( x p )1/p − ( y p )1/p 
  (1/p)−1
≤ p−1  x p
− y p
· max{ x p
, y p
}

≤ p−1 x − y
1−p
p
· max{ x , y } , (1.2.8)

from which we conclude that

any p-norm is locally Hölder continuous,
(1.2.9)
with Hölder exponent p ∈ (0, 1].
By replacing the original quasi-norm by some equivalent p-norm (cf. (1.2.5)),
there is no loss of generality in assuming that
(1.2.10)
all quasi-norms considered are continuous.

Moving on, given two quasi-normed vector spaces X, · X and Y, · Y ,

consider a positively homogeneous mapping T : X → Y , i.e., a function T sending
X into Y and which satisfies T(λu) = λT(u) for each u ∈ X and each λ ∈ (0, ∞) (note
that taking u := 0 ∈ X and λ := 2 implies T(0) = 0 ∈ Y ). We shall denote by
 
T X→Y := sup Tu Y : u ∈ X, u X = 1 ∈ [0, +∞] (1.2.11)

the operator “norm” of such a mapping T. In particular,

Tu Y ≤ T X→Y u X for each u ∈ X. (1.2.12)

It is also useful to observe that for any positively homogeneous mapping T : X → Y

we have
Tx Y
T X→Y = sup = sup Tx Y. (1.2.13)
x ∈X\{0} x X x ∈X, x X =1

It is straightforward to check that

a positively homogeneous mapping T : X → Y is continuous at 0 ∈ X
if and only if T is bounded (i.e., it maps bounded subsets of X into (1.2.14)
bounded subsets of Y ) if and only if T X→Y < +∞.

Consider next the special case when (X, · X ) is an arbitrary quasi-normed

space, the quasi-normed space (Y, · Y ) is a sub-lattice of the space of measurable
functions associated with a generic measure space, and assume that T is some sub-
linear mapping of X into Y , i.e., T : X → Y satisfies T(λu) = |λ|T(u) for each u ∈ X
and each scalar λ, as well as T(u + w) ≤ Tu +T w at a.e. point in X, for each u, w ∈ X.
Then, for each u, w ∈ X we have |Tu − T w| ≤ T(u − w) at a.e. point in X, hence
1.2 Quasi-Normed Spaces 7

Tu − T w Y ≤ T(u − w) Y ≤ T X→Y u−w X thanks to using the lattice property

in Y . Consequently,
if X, Y are as above then a sub-linear map T : X → Y
(1.2.15)
is continuous if and only if one has T X→Y < +∞.

Next, assuming that X, · X as well as Y, · Y are quasi-normed vector

spaces, we set
 
Bd(X → Y ) := T : X → Y : T linear mapping with T X→Y < +∞ , (1.2.16)

and occasionally write T Bd(X→Y) in place of T X→Y . This is a quasi-normed

space which is compatible with (1.1.9), in the sense that if X, Y are quasi-normed
spaces then B(X → Y ) and Bd(X → Y ) are identical. Hence, in light of (1.1.12),
we have
Bd(X → Y ) = B(X → Y ) = L(X → Y )
(1.2.17)
whenever X, Y are quasi-normed spaces.
We also agree to abbreviate

Bd(X) := Bd(X → X). (1.2.18)

In particular, given a linear and bounded operator T mapping a quasi-normed space

(X, · X ) into itself,

Tx X
T Bd(X) = sup = sup Tx X. (1.2.19)
x ∈X\{0} x X x ∈X, x X =1

Note that if (X, · X ) is a quasi-normedspace, T ∈ Bd(X), and Y is a linear subspace

of X which is invariant under T, then T Y belongs to Bd(Y ) where Y is regarded as a
quasi-normed space equipped with the quasi-norm · X , and

T Y Bd(Y) ≤ T Bd(X) . (1.2.20)

As seen from definitions,

∗
if X is a quasi-normed
  → C) = X and
space then Bd(X
(1.2.21)
Λ X ∗ = sup |Λ(x) : x ∈ X, x X = 1 for each Λ ∈ X ∗ .

A companion result, itself a well-known consequence of the Hahn-Banach Theorem,

is that
if X is a normed
 space then, for each x ∈ X,
(1.2.22)
x X = sup |Λ(x) : Λ ∈ X ∗, Λ X ∗ = 1 .
The following elementary result is going to be useful later on.

Lemma 1.2.1 Assume that X, Y are quasi-normed spaces, such that Y ⊆ X and the
inclusion j : Y → X is continuous with dense range. Then the adjoint operator
8 1 Preliminary Functional Analytic Matters

ι := j ∗ : X ∗ −→ Y ∗ acting according to
(1.2.23)
ι(Λ) = j ∗ (Λ) = Λ ◦ j for every Λ ∈ X ∗ (cf. (1.1.20))

is well defined, linear, continuous, and injective. In particular, this induces a con-
tinuous embedding
X ∗ → Y ∗ . (1.2.24)
In addition, the embedding in (1.2.24) is in such a way that the duality pairings
X ∗ ·, · X and Y ∗ ·, ·Y are compatible, in the sense that

X ∗ Λ, yX = Y ∗ Λ, yY for every

(1.2.25)
Λ∈ X∗ → Y ∗ and y ∈ Y → X.

Finally,
if actually X, Y are Banach spaces and Y is reflexive,
(1.2.26)
then the embedding (1.2.24) also has dense range.
Proof That ι is well defined, linear, and continuous is clear from definitions. To
show that ι is injective, assume Λ ∈ X ∗ is such that ι(Λ) = 0 in Y ∗ . Then for each
y ∈ Y we have

0 = Y ∗ ι(Λ), yY = Y ∗  j ∗ (Λ), yY = X ∗ Λ, j(y)X (1.2.27)

hence the functional Λ ∈ X ∗ vanishes on the image of j : Y → X. Since the latter is

a dense subset of X, we conclude that Λ = 0. This proves that ι is indeed injective. As
regards (1.2.25), suppose Λ ∈ X ∗ and y ∈ Y are arbitrary. Regard Λ as a functional
in Y ∗ by identifying it with j ∗ (Λ), and regard y as a vector in X by identifying it with
j(y) ∈ X. Then

X ∗ Λ, yX = X ∗ Λ, j(y)X = Y ∗  j ∗ (Λ), yY = Y ∗ Λ, yY , (1.2.28)

proving (1.2.25). To justify the very last claim in the statement, work under the
stronger assumptions that X is a Banach space and Y is a reflexive Banach space. As
a consequence of the Hahn-Banach Theorem, it suffices to show that if ∈ (Y ∗ )∗ is
a functional which vanishes identically on j ∗ (X ∗ ) then necessarily = 0. Since Y is
reflexive, this comes down to checking that a given vector y ∈ Y is necessarily zero
if
∗ ∗
Y y, j (Λ)Y ∗ = 0 for each Λ ∈ X . (1.2.29)
To this end, note that (1.2.29) implies that for each Λ ∈ X ∗ we have

0 = Y y, j ∗ (Λ)Y ∗ = X  j(y), ΛX ∗ (1.2.30)

hence j(y) = 0 by the Hahn-Banach Theorem. Since j is injective, this ultimately

forces y = 0, as wanted. 
Given a vector space X along with a linear subspace Y , of X, the quotient map
1.2 Quasi-Normed Spaces 9

π : X → X/Y, π(x) := [x]X/Y = x + Y ∈ X/Y for each x ∈ X, (1.2.31)

is a well-defined linear operator. Assume next that (X, τX ) is a Hausdorff topological

vector space and that Y is a closed linear subspace of X. Then
 
τX/Y := A ⊆ X/Y : π −1 (A) ∈ τX (1.2.32)

is a topology on X/Y , called the quotient topology, which renders (X/Y, τX/Y ) a
topological vector space. In addition, τX/Y enjoys the properties listed in the lemma
below (for a proof, see [165, Theorem 1.41, p. 31]).

Lemma 1.2.2 Let Y be a closed linear subspace of a Hausdorff topological vector

space (X, τX ). Then, in relation to the quotient topology on X/Y , defined as in
(1.2.32), the following statements are true.

(1) The quotient map π : (X, τX ) → (X/Y, τX/Y ) is linear, continuous, surjective,
and open.
(2) If BX is a local base for τX , then {π(U)}U ∈B X is a local base for τX/Y .
(3) Each of the following properties of (X, τX ) is inherited by (X/Y, τX/Y ): local
convexity, local boundedness, metrizability, normability.
(4) If X is an F-space4, or a Fréchet space5, or a Banach space, then so is X/Y .

Next, consider a quasi-normed space X, · X and, given a linear subspace Y of

X, define

[x]X/Y X/Y
= x +Y X/Y := inf x + y X for each x ∈ X. (1.2.33)
y ∈Y

If Y is closed, then (1.2.33) becomes a quasi-norm on the quotient space X/Y , and
the associated modulus of concavity (cf. (1.2.1)) satisfies

ρ(X/Y ) = ρ(X). (1.2.34)

Moreover, the canonical projection

π : X, · X −→ X/Y, · X/Y defined by

π(x) := [x]X/Y = x + Y ∈ X/Y for each x ∈ X, (1.2.35)
is a well-defined, linear, and bounded operator.

Finally, we remark that if · X is a p-norm for some p ∈ (0, 1] then · X/Y is also
a p-norm, and that replacing · X in (1.2.33) by an equivalent quasi-norm on X
produces an equivalent quasi-norm on the quotient space X/Y .

4 recall that an F-space is a Hausdorff topological vector space (X, τX ) whose topology τX is
induced by a complete invariant metric
5 recall that a Hausdorff topological vector space (X, τX ) is said to be a Fréchet space if X is a
locally convex F-space
10 1 Preliminary Functional Analytic Matters

Lemma 1.2.3 Let X, · X be a p-Banach space for some p ∈ (0, 1] and let Y be a
closed linear subspace of X. Then X/Y, · X/Y is also a p-Banach space. Also, if
Xo is a closed linear subspace of X which contains Y , then Xo /Y is a closed linear
subspace of X/Y .
As a corollary, if X, · X is a quasi-Banach space and Y is a closed linear
subspace of X, then X/Y, · X/Y is also a quasi-Banach space, and for each
closed linear subspace Xo of X which contains Y it follows that Xo /Y is a closed
linear subspace of X/Y .

Proof From our earlier discussion, we already know that · X/Y is a p-norm on
X/Y . Suppose {ξn }n∈N ⊆ X/Y is Cauchy in the quotient quasi-norm. We wish to
prove this sequence converges, and for this it is enough to produce a convergent
subsequence. Begin by inductively choosing a subsequence {ξnk }k ∈N such that


∞
p
ξnk − ξnk+1 X/Y
< ∞. (1.2.36)
k=1

This may be accomplished as follows. Since {ξn }n∈N ⊆ X/Y is Cauchy, there exists
n1 ∈ N with the property that ξn − ξm X/Y < 2−1 whenever n, m ≥ n1 . Also, there
p

exists n2 > n1 with the property that ξn − ξm X/Y < 2−2 whenever n, m ≥ n2 .
p

Repeat this process, to produce a strictly increasing sequence {nk }k ∈N such that
ξnk − ξnk+1 X/Y < 2−k for all k ∈ N. From this, (1.2.36) follows. Going further,
p

pick x1 ∈ ξn1 and use the definition in (1.2.33) to conclude that for each k ∈ N there
exists a vector

xk+1 ∈ ξnk+1 − ξnk with xk+1 X ≤ ξnk+1 − ξnk X/Y + 2−k . (1.2.37)
∞ p
Together, (1.2.36) and (1.2.37) ensure that k=1 xk X < ∞. From this, the fact that
p
p
(1.2.3) implies N
k=M xk ≤ N
k=M xk X for all M, N ∈ N with N > M, and the
X
completeness of X we then conclude that the series ∞ k=1 xk converges in X, · X .
By continuity of the quotient map (1.2.35), the series

∞ 
∞
ξn1 + (ξnk+1 − ξnk ) = π(xk+1 ) (1.2.38)
k=1 k=0
 
∞
converges in X/Y to π k=0 xk+1 . Given that


M
ξn M +1 = ξn1 + (ξnk+1 − ξnk ) for each M ∈ N, (1.2.39)
k=1

this implies that the subsequence {ξn M +1 } M ∈N is convergent in the quotient p-norm,
as desired. Ultimately, this establishes that X/Y, · X/Y is indeed a p-Banach
space.
1.2 Quasi-Normed Spaces 11

Finally, consider a closed linear subspace Xo of X which contains Y . We wish

to show that Xo /Y is closed in X/Y, · X/Y . With this in mind, pick a sequence
{ξn }n∈N ⊆ Xo /Y which converges in X/Y, · X/Y to some ξ ∈ X/Y , with the
goal of proving that ξ ∈ Xo /Y . Since {ξn }n∈N is Cauchy in X/Y, · X/Y , we may
proceed as in the first part of the proof and produce vectors x1, x2, . . . , xk , · · · ∈ Xo
satisfying (1.2.37). As before, this ensures that the series ∞ k=1 xk converges in
X, · X , and since Xo is closed, it follows that the sum x := ∞ k=0 xk+1 belongs to
Xo . Ultimately,

∞ 
∞ 
ξ = ξn1 + (ξnk+1 − ξnk ) = π xk+1 = π(x) ∈ Xo /Y, (1.2.40)
k=1 k=0

as wanted. 

The point of our next lemma is that, for a linear continuous operator on quasi-
Banach spaces, modding out its kernel renders said operator continuous and injective.

Lemma 1.2.4 Let X, · X and Y, · Y be two quasi-Banach spaces and consider

a linear continuous operator T : X → Y . Then the operator

 : X/Ker T, ·
T X/Ker T −→ Y, · Y defined by
(1.2.41)
 [x]X/Ker T := T x ∈ Y for each x ∈ X,
T

is well defined, linear, continuous, and injective.

Proof We begin by noting that Ker T = Ker(T : X → Y ) is a closed subspace

 in (1.2.41) is
of X (as the null-space of a continuous linear operator), and that T

well defined, linear, and injective. There remains to show that T in (1.2.41) is also
continuous. Since Lemma 1.2.3 guarantees that X/Ker T, · X/Ker T is a quasi-
Banach space, thanks to the version of the Closed Graph
 Theorem recorded in [138,

Corollary 6.78, p. 442] it suffices to show that G := [x]X/Ker T , T x : x ∈ X is
closed in the space X/Ker T × Y equipped with the product quasi-norm

X/Ker T × Y [x]X/Ker T , y → [x]X/Ker T X/Ker T

+ y Y. (1.2.42)

To this end, suppose {ξn }n∈N ⊆ X/Ker T is a sequence convergent in X/Ker T, ·


X/Ker T to some ξ ∈ X/Ker T, with the property that {T(ξn )}n∈N converges in
Y, · Y to some y ∈ Y . The goal is to show that (ξ, y) belongs to G, i.e., that
 = y.
T(ξ)
In view of the Aoki-Rolewicz theorem (cf. (1.2.5)), there is no loss of generality in
assuming that · X is actually a p-norm, for some p ∈ (0, 1]. Since {ξn }n∈N ⊆ X/Y is
Cauchy, there exists n1 ∈ N with the property that ξn − ξm X/Ker T < 2−1 whenever
p

n, m ≥ n1 . Next, there exists n2 > n1 with the property that ξn − ξm X/Ker T < 2−2
p

whenever n, m ≥ n2 . Inductively, this process produces a strictly increasing sequence

{nk }k ∈N such that ξnk+1 − ξnk X/Ker T < 2−k for all k ∈ N. In particular,
p
12 1 Preliminary Functional Analytic Matters


∞
p
ξnk+1 − ξnk X/Ker T
< ∞. (1.2.43)
k=1

Next, pick an arbitrary vector x1 ∈ ξn1 and, using the definition in (1.2.33), for each
k ∈ N select some vector xk+1 ∈ ξnk+1 − ξnk such that

xk+1 X < ξnk+1 − ξnk X/Ker T + 2−k . (1.2.44)

∞ p
From (1.2.43) and (1.2.44) we deduce that k=1 xk X < ∞. Also, iterations of
p
p
(1.2.3) imply that N
k=M xk ≤ N
k=M M, N ∈ N with N > M.
xk X for all
X
Based on these observations and the completeness of X we conclude that the series
∞
k=1 xk converges in X, · X to some x ∈ X. In addition, from the definition of
xk and (1.2.41) we see that
 n1 ) = T x1 and T(ξ
T(ξ  nk+1 − ξnk ) = T xk+1 for each k ∈ N. (1.2.45)

In particular,


M 
M 
T ξn1 +
ξn M +1 = T ξnk+1 − T
(T ξnk ) = T x1 + T xk+1
k=1 k=1

 M+1
 
=T xk for each M ∈ N. (1.2.46)
k=1

∞
Since M+1
k=1 xk converges to x = in X,  ·
k=1 xk  → ∞ and
X as M the mapping
T : X → Y is continuous, we conclude that T M+1
k=1 xk converges to T x in the
space Y, · Y as M → ∞. Passing to limit as M → ∞ in (1.2.46) then yields
ξn M +1 = T x, as wanted.
y = lim M→∞ T 

In turn, Lemma 1.2.4 is one of the ingredients in the proof of the following
isomorphism result.

Lemma 1.2.5 Let X, · X and Y, · Y be two quasi-Banach spaces and consider

a linear continuous operator T : X → Y whose range Im(T : X → Y ) is closed in
Y, · Y . Then the operator

 : X/Ker T, ·
T X/Ker T −→ Im(T : X → Y ), · Y given by
(1.2.47)
 [x]X/Ker T := T x ∈ Y for each x ∈ X,
T

is well defined, linear, continuous, bijective, with a continuous inverse.

 in (1.2.47) is well defined, linear, continuous, and bijective
Proof That the operator T
is a consequence of definitions and Lemma 1.2.4. The continuity of its inverse comes
down to T  being open. That this is indeed the case follows from the version of the
1.2 Quasi-Normed Spaces 13

Open Mapping Theorem recorded in [138, Corollary 6.62, p. 423], bearing in mind
that both X/Ker T, · X/Ker T as well as Im(T : X → Y ), · Y are quasi-Banach
spaces, thanks to Lemma 1.2.3 and (1.2.2). 
Here is a result which further builds on the previous lemma.

Lemma 1.2.6 Let X, · X and Y, · Y be two quasi-Banach spaces and consider

a linear continuous operator T : X → Y whose range Im(T : X → Y ) is closed
in Y, · Y . Then for each closed linear subspace Z of X, · X which contains
Ker(T : X → Y ) it follows that T(Z) is closed in Y, · Y .

Proof Fix a linear subspace Z of X, · X containing Ker T = Ker(T : X → Y ).

Lemma 1.2.3 implies that Z/Ker T is a closed subspace of X/Ker T, · X/Ker T .
 defined as in (1.2.47)
Also, by virtue of Lemma 1.2.5, the corresponding operator T

is a topological isomorphism. As a result, T(Z) = T(Z/Ker T) is closed in the space
Im(T : X → Y ), · Y , hence in Y, · Y (given that T has closed range). 
We are now in a position to show that a linear continuous operator on quasi-
Banach spaces, the quality of having a finite dimensional cokernel entails closed
range.

Lemma 1.2.7 Let X, · X and Y, · Y be two quasi-Banach spaces and consider

a linear continuous operator T : X → Y with dim(Y /Im T) < ∞.
Then T has a closed range, i.e., Im(T : X → Y ) is a closed linear subspace of Y .

Proof First we observe that there exists a finite-dimensional linear subspace M of

Y , whose dimension is n := dim(Y /Im T), for which

M + Im(T : X → Y ) = Y and M ∩ Im(T : X → Y ) = {0}. (1.2.48)

Indeed, if {ξ1, . . . , ξn } is a basis Y /Im T, then each ξ j is of the form y j + Im T for

some y j ∈ Y , and one may check without difficulty that {y1, . . . , yn } are linearly
independent and span a finite-dimensional space M satisfying all desired properties.
Going further, since dim M < ∞ it follows that M is a closed linear subspace of
Y . Consequently, X1 := X × M equipped with the product quasi-norm

X×M (x, y) −→ (x, y) X1 := x X + y Y (1.2.49)

is a closed subspace of X × Y , hence itself a quasi-Banach space. Also, the operator

T1 : X1, · X1 −→ Y, · Y defined by
(1.2.50)
T1 (x, y) := T x + y ∈ Y for each (x, y) ∈ X1 = X × M,

is well defined, linear, continuous, and surjective (cf. (1.2.48)); in particular, T1

has closed range. Moreover, (1.2.48) implies that Ker (T1 : X1 → Y ) is equal to
Ker (T : X → Y ) × {0}. Because Z := X × {0} is a closed subspace of X1, · X1
which contains Ker (T1 : X1 → Y ), we then conclude from Lemma 1.2.6 that
T1 (Z) = T(X) is a closed subspace of Y, · Y . 
14 1 Preliminary Functional Analytic Matters

We shall also need the following self-improvement of Lemma 1.2.7.

Lemma 1.2.8 Let X, · X and Y, · Y be two quasi-Banach spaces and consider

a linear continuous operator T : X → Y with dim (Y /Im T) < ∞. Also, let M be a
finite dimensional linear subspace of Y . Then M + Im (T : X → Y ) is a closed linear
subspace of Y .

Proof This is a consequence of Lemma 1.2.7 and item (1) in Theorem 1.1.3. Another
proof goes as follows. With M having the current significance, consider the space
X1 := X × M equipped with the product quasi-norm (1.2.49), and define the operator
T1 : X1 → Y as in (1.2.50). As before, X1 is a quasi-Banach space and T1 is a
linear and continuous operator. In addition, Im (T : X → Y ) ⊆ Im (T1 : X1 → Y )
which implies dim (Y /Im T1 ) ≤ dim (Y /Im T) < ∞. Lemma 1.2.7 then implies that
Im (T1 : X1 → Y ) is a closed linear subspace of Y . Since, as seen from (1.2.50), we
have
Im (T1 : X1 → Y ) = M + Im (T : X → Y ), (1.2.51)
the desired conclusion follows. 
Recall (1.1.13). If actually X, Y are arbitrary quasi-normed spaces, then from
definitions and (1.2.14) it follows that

Cp(X → Y ) is a linear subspace of Bd(X → Y ). (1.2.52)

In addition, (again, see [142])

if X, Y are quasi-Banach spaces then Cp (X → Y ) is

(1.2.53)
a closed subspace of the space Bd (X → Y ).

Also, given T ∈ Bd(X → Y ) where X, Y are quasi-normed spaces, the quantity

(originally introduced by J. Radon in [156])
ess
T X→Y := dist T, Cp(X → Y )
 
= inf T − K X→Y : K ∈ Cp(X → Y ) (1.2.54)

is referred to as the essential norm of the operator T. It is then clear from definitions
and (1.2.53) that

if X, Y are quasi-Banach spaces then for any T ∈ Bd (X → Y )

ess (1.2.55)
we have T X→Y = 0 if and only if T ∈ Cp (X → Y ).

It is also apparent from definitions that for any quasi-normed spaces X, Y and any
T ∈ Bd(X → Y ) we have
ess
T X→Y ≤ T X→Y . (1.2.56)
ess
Let us also point out that, in the particular case when Y := X, it follows that T X→X
is the norm of the equivalence class of T in the Calkin algebra Bd(X)/Cp(X). As is
well known,
1.2 Quasi-Normed Spaces 15

a bounded linear operator between Banach spaces is compact if and

only if its adjoint is (Schauder’s theorem), and the set of all compact
(1.2.57)
operators from a Banach space to itself form a closed two-sided ideal
in the algebra of all bounded linear operators on this space.
Here is a lower estimate for the essential norm, complementing (1.2.56). This
shows that, given an invertible operator on an infinite dimensional Banach space, its
norm cannot be decreased arbitrarily by subtracting compact operators.
Lemma 1.2.9 Assume X is an infinite dimensional Banach space and suppose the
mapping T : X → X is an isomorphism. Then
−1
T −1
ess
T X→X ≥ X→Y . (1.2.58)

In particular, with I denoting the identity on X, for any compact operator K on X

one has
I − K X→X ≥ 1. (1.2.59)
Proof Pick an arbitrary compact operator K on X. Then (1.2.58) follows as soon as
we show that
−1
T − K X→X ≥ T −1 X→Y . (1.2.60)
Our strategy is to reason by contradiction. Concretely, define R := T − K and note
−1
that R ∈ Bd (X) satisfies R X→X < T −1 X→Y . Then

T −1 R X→X ≤ T −1 X→X R X→X < 1. (1.2.61)

Consequently, I −T −1 R is invertible (via a Neumann series). Since T is also invertible,

it follows that K = T −R = T(I −T −1 R) is invertible. Thus, X has a compact invertible
operator which, according to Riesz’ theorem, makes it a finite dimensional space.
This contradiction proves (1.2.58). Finally, (1.2.59) follows from (1.2.58) (used with
T := I) and (1.2.54). 
The essential norm of the composition may be estimated in the manned indicated
below.
Lemma 1.2.10 Let X j , Yj be quasi-normed spaces for j ∈ {0, 1}, and suppose that
S ∈ Bd(X1 → X0 ), T ∈ Bd(X0 → Y0 ), and R ∈ Bd(Y0 → Y1 ). Then
ess ess
R◦T ◦S X1 →Y1 ≤ R Y0 →Y1 T X0 →Y0 S X1 →X0 . (1.2.62)

Proof Let K ∈ Cp(X0 → Y0 ) be arbitrary. Then R ◦ K ◦ S ∈ Cp(X1 → Y1 ), hence

(1.2.54) permits us to estimate
ess ess ess
R◦T ◦S X1 →Y1 ≤ R◦T ◦S − R◦K ◦S X1 →Y1 = R ◦ (T − K) ◦ S X1 →Y1

≤ R Y0 →Y1 T −K X0 →Y0 S X1 →X0 . (1.2.63)

Taking the infimum over K ∈ Cp(X0 → Y0 ) of the most extreme sides in (1.2.63)
then yields (1.2.62), on account of (1.2.54). 
16 1 Preliminary Functional Analytic Matters

Let us now take a closer look at the space of bounded operators acting between
two given quasi-normed spaces.

Lemma 1.2.11 Whenever X, · X is a quasi-normed space and Y, · Y is a

quasi-Banach space, it follows that Bd(X → Y ), · X→Y is a quasi-Banach space.
Moreover, if Y is actually a p-Banach space for some p ∈ (0, 1], then Bd(X → Y )
is also a p-Banach space. In fact, with some self-explanatory notation, if · ∗ is an
equivalent p-norm on Y for some p ∈ (0, 1], then · (X, · X )→(Y, · ∗ ) is an equivalent
p-norm on Bd(X → Y ).
As a corollary, if X, · X is a quasi-normed space then its dual (cf. (1.1.18)),
i.e.,

X ∗, · X∗ with Λ X∗ := sup |Λ(x)| for each Λ ∈ X ∗ (1.2.64)

x ∈X, x X =1

is a Banach space.

Proof That Bd(X → Y ), · X→Y is a quasi-normed space is clear from definitions.

Assume {Tj } j ∈N is a Cauchy sequence in Bd(X → Y ), · X→Y . Since for each
x ∈ X and j, k ∈ N we have Tj x − Tk x Y ≤ Tj − Tk X→Y x X it follows that
{Tj x} j ∈N is a Cauchy sequence in the quasi-Banach space Y . Hence, T x := lim Tj x
j→∞
exists in Y , for each x ∈ X. Clearly, T : X → Y is linear. Also, since for each
x ∈ X and j ∈ N we may estimate T x Y ≤ C T x − Tj x Y + C Tj x Y , on the one
hand we have T x Y ≤ C lim inf Tj x Y . On the other hand, the fact that {Tj } j ∈N
j→∞
 
is Cauchy in Bd(X → Y ), · X→Y forces Tj X→Y j ∈N to be bounded. Hence,
M := sup Tj X→Y is a finite number, and T x Y ≤ C lim inf Tj x Y ≤ CM x for
j ∈N j→∞
each x ∈ X. This proves that T ∈ Bd(X → Y ). Next, pick a threshold ε > 0. Given
that {Tj } j ∈N is Cauchy in Bd(X → Y ), · X→Y , there exists some N ∈ N with
the property that Tk − Tj X→Y ≤ ε whenever j, k ∈ N satisfy j, k ≥ N. Fix now
an arbitrary x ∈ X along with an arbitrary j ∈ N satisfying j ≥ N. Since for each
k ∈ N we may write T x − Tj x Y ≤ C T x − Tk x Y + C Tk x − Tj x Y it follows that

T x − Tj x Y ≤ C lim inf Tk x − Tj x Y
k→∞

≤ C lim inf Tk − Tj X→Y x X ≤ Cε x X. (1.2.65)

k→∞

In view of the arbitrariness of x ∈ X, this entails T − Tj X→Y ≤ Cε for each j ≥ N.

Hence, {Tj } j ∈N converges to T in Bd(X → Y ), · X→Y , so the latter space is
complete. This takes care of the first claim in the statement of the lemma.
As regards the second claim in the statement of the lemma, assume · ∗ is an
equivalent p-norm on Y for some p ∈ (0, 1]. Then · (X, · X )→(Y, · ∗ ) is an equivalent
quasi-norm with · X→Y and for each T, R ∈ Bd(X → Y ) we have
1.2 Quasi-Normed Spaces 17
p p
T+R (X, · X )→(Y, · ∗)
= sup T x + Rx ∗
x ∈X, x X =1

 
p p
≤ sup Tx ∗ + Rx ∗
x ∈X, x X =1

p p
≤ sup Tx ∗ + sup Rx ∗
x ∈X, x X =1 x ∈X, x X =1

p p
= T (X, · X )→(Y, · ∗)
+ R (X, · X )→(Y, · ∗)
, (1.2.66)

so · (X, · X )→(Y, · ∗ ) is indeed an equivalent p-norm on Bd(X → Y ). Moreover,

since from the first part of the proof we know that Bd(X → Y ), · X→Y is
complete
 and since · (X, · X )→(Y,  · ∗ ) is equivalent with · X→Y , it follows that
Bd(X → Y ), · (X, · X )→(Y, · ∗) is complete as well. 

Recall (1.1.20). If X, Y are quasi-normed spaces it follows from Lemma 1.2.11

that X ∗, Y ∗ are Banach spaces and

for each T ∈ Bd(X → Y ) we have T ∗ ∈ Bd(Y ∗ → X ∗ )

(1.2.67)
and T ∗ Y ∗ →X ∗ ≤ T X→Y .

Moreover,

Bd(X → Y ) T −→ T ∗ ∈ Bd(Y ∗ → X ∗ )
(1.2.68)
is an (antilinear) isometry if Y is actually a normed space.

We next discuss Neumann series of operators on quasi-Banach spaces.

Lemma 1.2.12 Suppose X is a p-Banach space for some p ∈ (0, 1] and let I denote
the identity on X. Also, fix some T ∈ Bd(X) and pick

z ∈ C with |z| > T X→X . (1.2.69)

Then the operator zI −T is invertible on X and its inverse is given by the Neumann
series
∞
(zI − T)−1 = z −j−1T j , (1.2.70)
j=0

which converges in Bd(X), · X→X .

As a consequence of this and the classical Aoki-Rolewicz Theorem (cf. (1.2.5)),
whenever X is a quasi-Banach space and T ∈ Bd(X) it follows that the operator
zI − T is invertible on X for any z ∈ C with |z| > T X→X . Moreover, its inverse is
given by the Neumann series (1.2.70) which converges in Bd(X), · X→X provided
|z| > T X→X .
Proof Since by assumption X, · X is a p-Banach space, for some p ∈ (0, 1], the
last part in Lemma 1.2.11 ensures that Bd(X), · X→X is also a p-Banach space.
18 1 Preliminary Functional Analytic Matters

Granted this, if z is as in (1.2.69), for each N, M ∈ N with N > M we may write


N p 
N
z −j−1T j |z| −j p−p T j
p
≤ X→X
X→X
j=M j=M


N
≤ |z| −p |z| −1 T
jp
X→X (1.2.71)
j=M

which is the tail of a convergent geometric series. Thus, the Neumann series in
(1.2.70) converges in Bd(X), · X→X since the associated sequence of partial
sums is Cauchy. If R ∈ Bd(X) denotes the limit operator then, thanks to (1.2.69),


N 
(zI − T)R = lim (zI − T) z −j−1T j
N →∞
j=0


N 
= lim (zI − T) z −j−1T j
N →∞
j=0
 
= lim I − (T/z) N +1 = I. (1.2.72)
N →∞

In a similar fashion, R(zI − T) = I so zI − T is invertible and (zI − T)−1 = R. 

As a consequence of the Open Mapping Theorem, surjectivity for a linear operator

on quasi-Banach spaces self-improves to a quantitative version of itself, something
that may be summarized by saying that plain surjectivity implies “ontoness with
control.”

Proposition 1.2.13 Assume X, · X and Y, · Y are two quasi-Banach spaces

and suppose T : X → Y is a linear continuous operator with closed range. Then
there exists C ∈ (0, ∞) with the property that for each y ∈ Im (T → Y ) one may find
x ∈ X such that T x = y and x X ≤ C y Y .

Proof Without loss of generality we may assume that T is onto (otherwise replace Y
by Im (T → Y ), endowed with the quasi-norm inherited from Y ). In the latter setting,
the version of the Open Mapping Theorem for quasi-Banach spaces (cf., e.g., [138,
Corollary 6.62, p. 423]) tells us that T BX (0, 1) is an open neighborhood of 0Y in
Y . As such, there exists r ∈ (0, ∞) with the property that

BY (0, r) ⊆ T BX (0, 1) . (1.2.73)

Pick now y ∈ Y arbitrary. If y  0, then (r/2)(y/ y Y ) ∈ BY (0, r). As such, (1.2.73)

guarantees that there exists z ∈ BX (0, 1) with T z = (r/2)(y/ y Y ). Consequently,
x := (2 y Y /r)z ∈ X satisfies T x = y and x X ≤ (2 y Y /r) z X ≤ C y Y if we
define C := 2/r ∈ (0, ∞). Finally, if y = 0Y then x := 0X will do. 
1.2 Quasi-Normed Spaces 19

It is useful to note that passing to adjoints does not increase the distance to
compact operators.

Lemma 1.2.14 Let X, Y be two arbitrary Banach spaces. Then for each operator
T ∈ Bd(X → Y ) one has T ∗ ∈ Bd(Y ∗ → X ∗ ) and

dist T ∗, Cp(Y ∗ → X ∗ ) ≤ dist T, Cp(X → Y ) , (1.2.74)

with the distance in the left-hand side considered in Bd(Y ∗ → X ∗ ), and the distance
in the right-hand side considered in Bd(X → Y )
As a corollary, if X, Y are reflexive Banach spaces, for each T ∈ Bd(X → Y ) one
has
dist T ∗, Cp(Y ∗ → X ∗ ) = dist T, Cp(X → Y ) . (1.2.75)

With the piece of notation introduced in (1.2.54), we may rephrase the estimate
in (1.2.74) as

T∗
ess ess
Y ∗ →X ∗ ≤ T X→Y for each T ∈ Bd(X → Y ). (1.2.76)

In general, the inequality in (1.2.76) can be strict (see the discussion in [12]). Here
is the proof of Lemma 1.2.14.
Proof of Lemma 1.2.14 For each compact operator K ∈ Cp(X → Y ), Schauder’s
theorem (cf. (1.2.57)) implies that K ∗ ∈ Cp(Y ∗ → X ∗ ), thus for any T ∈ Bd(X → Y )
we may write

dist T ∗, Cp(Y ∗ → X ∗ ) ≤ T ∗ − K ∗ Y ∗ →X ∗

= (T − K)∗ Y ∗ →X ∗ = T −K X→Y . (1.2.77)

Taking the infimum over all K ∈ Cp(X → Y ) then yields (1.2.74). 

As a preamble to the proof of Lemma 1.2.17, the essential norm of the restriction
of a given linear continuous operator to the range of projection is looked at in the
next lemma.

Lemma 1.2.15 Let X be a quasi-normed space and suppose P is a projection6 on X.

Set Y := PX (viewed itself as a quasi-normed space, equipped with the quasi-norm
inherited from X) and assume T : X → X is a linear continuous operator with the
property that T(Y ) ⊆ Y . Finally, denote by T |Y : Y → Y the restriction of T to Y .
Then
ess
T |Y Y→Y
= dist T |Y , Cp(Y ) ≤ P Y→Y · dist T, Cp(X) ,
ess
= P Y→Y · T X→X (1.2.78)

where the distances above are considered in Bd(Y ) and Bd(X), respectively.

6 i.e., linear, continuous, idempotent operator

20 1 Preliminary Functional Analytic Matters

Proof Given any K ∈ Cp(X) it follows that K  : Y → Y , given by K  y := P(K y) for

each y ∈ Y , is a compact operator on Y . Since

dist T |Y , Cp(Y ) ≤ T |Y − K  Y→Y

 
= sup T y − K  y Y : y ∈ Y, y Y ≤1
 
= sup P(T y) − P(K y) Y : y ∈ Y, y Y ≤1
 
≤ P Y→Y · sup T y − K y X : y ∈ Y, y Y ≤1
 
≤ P Y→Y · sup T x − K x X : x ∈ X, x X ≤1

= P Y→Y · T −K X→X , (1.2.79)

the estimate claimed in (1.2.78) follows by taking the infimum of the most extreme
side in (1.2.79) over K ∈ Cp(X). 

The proof of Lemma 1.2.17 also requires the algebraic result described below.

Lemma 1.2.16 Let X be a quasi-normed space, and fix Λ1, . . . , Λn ∈ X ∗ which are

linearly independent. Then there exist linearly independent vectors x1, . . . , xn ∈ X
such that Λ j (xk ) = δ jk for each j, k ∈ {1, . . . , n}.

Proof We proceed by induction on n. When n = 1 the claim is trivial. Suppose

n + 1 linearly independent functionals Λ1, . . . , Λn, Λn+1 ∈ X ∗ have been given, and
assume y1, . . . , yn ∈ X are such that Λ j (yk ) = δ jk for each j, k ∈ {1, . . . , n}. Since
Λ1, . . . , Λn, Λn+1 ∈ X ∗ are linearly independent it follows that

Λ := Λn+1 − Λn+1 (y1 )Λ1 − Λn+1 (y2 )Λ2 − · · · − Λn+1 (yn )Λn (1.2.80)

belongs to X ∗ and is not identically zero. Consequently, we may select a vector z ∈ X

such that
 
Λn+1 − Λn+1 (y1 )Λ1 − Λn+1 (y2 )Λ2 − · · · − Λn+1 (yn )Λn (z) = 1. (1.2.81)

We then proceed to define

yn+1 := z − Λ1 (z)y1 − Λ2 (z)y2 − · · · − Λn (z)yn . (1.2.82)

This entails 
1 if k = n + 1,
Λk (yn+1 ) = (1.2.83)
0 if k ∈ {1, . . . , n}.
For each fixed i ∈ {1, . . . , n + 1} consider now the (n + 1) × (n + 1) system


n+1
ai j Λk (y j ) = δik , k ∈ {1, . . . , n + 1}. (1.2.84)
j=1
1.2 Quasi-Normed Spaces 21

Since this has determinant 1, it follows that for each i ∈ {1, . . . , n + 1} the system
(1.2.84) has a (unique) solution (ai j )1≤ j ≤n+1 . Then the vectors


n+1
xi := ai j y j for each i ∈ {1, . . . , n + 1} (1.2.85)
j=1

satisfy Λ j (xk ) = δ jk for each j, k ∈ {1, . . . , n}. Since the latter property enures that
the vectors x1, . . . , xn are linearly independent, the proof is complete. 
The stage is set for us to prove the following useful result.

Lemma 1.2.17 Let X be a quasi-normed space and T : X → X is a linear continuous

operator. Fix a finite dimensional linear subspace Z ⊆ Ker T ∗ : X ∗ → X ∗ and
define  
Y := x ∈ X : Λ(x) = 0 for each Λ ∈ Z (1.2.86)
regarded as a quasi-normed space of its own, equipped with the quasi-norm inherited
from X.
Then T(Y ) ⊆ Y and there exists a constant CX, Z ∈ (0, ∞) independent of T with
the property that, with T |Y : Y → Y denoting the restriction of T to Y , one has
ess
T |Y Y→Y
= dist T |Y , Cp(Y ) ≤ CX, Z · dist T, Cp(X)
ess
= CX, Z · T X→X (1.2.87)

with the distances above considered in Bd(Y ) and Bd(X), respectively.

Proof That T(Y ) ⊆ Y is clear from definitions. To proceed, set n := dim Z ∈ N0 .

If n = 0 then Z = {0}, hence Y = X. In this case, (1.2.87) is trivially true as
long as CX, Z ≥ 1. Consider the case when n ∈ N. In this scenario, pick a basis
Λ1, . . . , Λn for Z. Thus, Λ1, . . . , Λn ∈ X ∗ are linearly independent, so we may invoke
Lemma 1.2.16 according to which there exist x1, . . . , xn ∈ X such that Λ j (xk ) = δ jk
for each j, k ∈ {1, . . . , n}. We use these to define an operator P : X → X by setting

n
Px := x − Λ j (x)x j for each x ∈ X. (1.2.88)
j=1

Then one may check without difficulty that P is linear, continuous, idempotent (hence
a projection on X), and that Y = P(X). Granted these properties, Lemma 1.2.15
applies and gives (1.2.87) with CX, Z := P Y→Y . 
It is known (see, e.g., the discussion in [142]) that
given an arbitrary Hausdorff linear topological space X, it follows
that every closed subspace Y of X which has finite codimension in (1.2.89)
X is topologically complemented in X.
Here is a more general version of Lemma 1.2.17.
22 1 Preliminary Functional Analytic Matters

Lemma 1.2.18 Let X be a quasi-normed space and consider a closed subspace Y of

X of finite codimension in X. Let T : X → X be a linear continuous operator with
the property that T(Y ) ⊆ Y , and denote by T |Y : Y → Y the restriction of T to Y
(regarded as a quasi-normed space in its own right, equipped with the quasi-norm
inherited from X). Then there exists a constant CX,Y ∈ (0, ∞), independent of T with,
such that
ess
T |Y Y→Y
= dist T |Y , Cp(Y ) ≤ CX,Y · dist T, Cp(X)
ess
= CX,Y · T X→X (1.2.90)

with the distances above considered in Bd(Y ) and Bd(X), respectively.

The concrete manner in which Lemma 1.2.18 contains Lemma 1.2.17 is as follows.
Using notation from the proof of Lemma 1.2.17, the fact that PX = Y implies that if
x1, . . . , xn  is the linear span of the vectors x1, . . . , xn then Y + x1, . . . , xn  = X (cf.
 Thus, Y that has finite codimension in X (in fact, dim(X/Y ) ≤ n). Since
(1.2.88)).
Y= Ker Λ, we see that Y is also closed. As such, Y from Lemma 1.2.17 satisfies
Λ∈Z
all the hypotheses required of it in Lemma 1.2.18, and (1.2.87) gives (1.2.90).
Proof of Lemma 1.2.18 In light of (1.2.89), it follows that there exists a projection
P on X such that PX = Y . Granted this, Lemma 1.2.15 applies and yields the desired
conclusion. 

We continue by establishing the lower bound for the distance to compact operators
on Lebesgue spaces of the sort contained in the next lemma.

Lemma 1.2.19 Let (X, M, μ) be a measure space, fix some p ∈ (0, ∞), and define

1 if p ≥ 1,
Cp := (1.2.91)
21/p−1 if p < 1.

Then for each operator T ∈ Bd L p (X, μ) one has

 
lim+ sup 1 A · T L p (X,μ)→L p (X,μ) ≤ Cp · dist T, Cp L p (X, μ) , (1.2.92)
ε→0 A∈M
μ(A)<ε

where the distance in the right-hand side is measured in Bd L p (X, μ) .

Proof We begin with the claim that if T ∈ Cp L p (X, μ) then the left-hand side of
(1.2.92) vanishes. Reasoning by contradiction, suppose T ∈ Cp L p (X, μ) yet there
exists δ > 0 along with a sequence { A j } j ∈N ⊆ M with lim μ(A j ) = 0 and such that
j→∞
1 A j · T L p (X,μ)→L p (X,μ) > δ for each j ∈ N. In turn, the last condition ensures that
there exist functions f j ∈ L p (X, μ) with f j L p (X,μ) = 1 and

1 A j · T fj L p (X,μ)
> δ for each j ∈ N. (1.2.93)
1.2 Quasi-Normed Spaces 23

Since we are presently working under the hypothesis that T is compact on L p (X, μ),
there is no loss of generality in assuming that

{T f j } j ∈N converges in L p (X, μ) to some function g ∈ L p (X, μ). (1.2.94)

To proceed, define

N := x ∈ X : there exist natural numbers j1 < j2 < · · · < jk < jk+1 < · · ·

such that x ∈ A jk for each k ∈ N
(1.2.95)

and observe that, by design,

lim 1 A j (x) = 0 for each x ∈ X \ N. (1.2.96)

j→∞

Also, if J is the set of all increasing sequences of natural numbers, then

  
J is countable and N = Aj . (1.2.97)
J ∈J j ∈J
 
In particular, since μ j ∈J A j ≤ lim sup μ(A j ) = 0, we conclude from (1.2.97)
J j→∞
that μ(N) = 0. In concert with (1.2.96), this proves that

lim 1 A j (x) = 0 for μ-a.e. x ∈ X. (1.2.98)

j→∞

(Parenthetically, we wish to note that (1.2.98) also follows upon observing that
|g| p dμ is a measure absolutely continuous with respect to μ.) Having established
(1.2.98), Lebesgue’s Dominated Convergence Theorem then ensures that

lim 1 A j · g L p (X,μ)
= 0. (1.2.99)
j→∞

For each j ∈ N we may now write, based on (1.2.93),

δ < 1 A j · T fj L p (X,μ)
≤ Cp 1 A j · (T f j − g) L p (X,μ)
+ Cp 1 A j · g L p (X,μ)

≤ Cp T f j − g L p (X,μ) + Cp 1 A j · g L p (X,μ)
(1.2.100)

which, after passing to limit j → ∞, leads to a contradiction in view of (1.2.94) and

(1.2.99). This establishes the fact that
 
lim+ sup 1 A · T L p (X,μ)→L p (X,μ) = 0 for all T ∈ Cp L p (X, μ) . (1.2.101)
ε→0 A∈M
μ(A)<ε
24 1 Preliminary Functional Analytic Matters

Going further, fix an arbitrary operator T ∈ Bd L p (X, μ) . Then for each operator

K ∈ Cp L p (X, μ) we may estimate (with Cp as in (1.2.91))
 
lim+ sup 1 A · T L p (X,μ)→L p (X,μ)
ε→0 A∈M
μ(A)<ε
 
≤ Cp · lim+ sup 1 A · (T − K) L p (X,μ)→L p (X,μ)
ε→0 A∈M
μ(A)<ε
 
+ Cp · lim+ sup 1A · K L p (X,μ)→L p (X,μ)
ε→0 A∈M
μ(A)<ε

≤ Cp · T − K L p (X,μ)→L p (X,μ), (1.2.102)

where the last inequality uses (1.2.101). Taking the infimum over all operators
K ∈ Cp L p (X, μ) then yields (1.2.92). 

We conclude this discussion by recording a useful observation, to the effect that

extension by density of operators on quasi-Banach spaces, which are linear and
bounded, preserves those qualities. In addition, if the original operator is actually
invertible, so is its extension by density.

Lemma 1.2.20 Let X0, X1 be a pair of quasi-Banach spaces, and let T : X0 → X1

be a bounded, linear, operator. Consider Yj ⊆ X j linear subspaces, for j ∈ {0, 1},
Xj
and denote by Yj the closure of Yj in X j , for j ∈ {0, 1}.
X0 X1
(1) If T(Y0 ) ⊆ Y1 , then T induces a bounded linear mapping from Y0 into Y1 .
(2) If T is invertible and has the property that T(Y0 ) = Y1 , then T induces an
X0 X1
isomorphism from Y0 onto Y1 .

Proof Since T : X0 → X1 is continuous and T(Y0 ) ⊆ Y1 , standard topology implies

X0 X1
that T maps Y0 continuously into Y1 . In view of (1.2.17), this takes care of the
claim in item (1). To deal with the claim in item (1), assume T : X0 → X1 is
X0
invertible and T(Y0 ) = Y1 . Since T : X0 → X1 is injective, it follows that T maps Y0
X1
injectively into Y1 . To show that this mapping is also surjective, fix an arbitrary
X1
g ∈ Y1 . Then g ∈ X1 and there exists a sequence {g j } j ∈N ⊆ Y1 which converges
to g in X1 . Given that T(Y0 ) = Y1 , for each j ∈ N we may find f j ∈ Y0 ⊆ X0
such that T f j = g j . If T −1 : X1 → X0 denotes the inverse of T : X0 → X1 ,
then T −1 g j = T −1 (T f j ) = f j ∈ Y0 . Hence, {T −1 g j } j ∈N ⊆ Y0 . Also, since the Open
Mapping Theorem for quasi-Banach spaces (cf., e.g., [138, Corollary 6.62, p. 423])
guarantees that T −1 : X1 → X0 is continuous, it follows that the sequence {T −1 g j } j ∈N
X0
converges to T −1 g in X0 . Consequently, T −1 g ∈ Y0 and since T(T −1 g) = g we
X0 X1
conclude that T maps Y0 onto Y1 . 
1.3 Real Interpolation of Quasilinear Operators 25

1.3 Real Interpolation of Quasilinear Operators

Call two quasi-normed spaces, X0 = X0, · X0 and X1 = X1, · X1 , compatible

if there exists a Hausdorff topological vector space X such that

Xi → X continuously, i ∈ {0, 1}. (1.3.1)

Note that, in this scenario, we can talk about the algebraic sum X0 + X1 (⊆ X). This
becomes a quasi-normed space when equipped with

x X0 +X1 := inf x0 X0 + x1 X1 , ∀x ∈ X0 + X1, (1.3.2)

x=x0 +x1
x0 ∈X0, x1 ∈X1

and X0 + X1 → X continuously. Furthermore, Xi → X0 + X1 continuously, for

i ∈ {0, 1}. One may check (based on Lemma 1.4.13 and the Aoki-Rolewicz Theorem)
that if X0 , X1 are complete then so is X0 + X1 equipped with · X0 +X1 . Hence, X0 + X1
turns out to be a quasi-Banach space if X0 , X1 are quasi-Banach spaces to begin with.
We also equip the intersection X0 ∩ X1 with the quasi-norm
 
x X0 ∩X1 := max x X0 , x X1 , ∀x ∈ X0 ∩ X1 . (1.3.3)

Again, one may show (using Lemma 1.4.13 and the Aoki-Rolewicz Theorem) that
if X0 , X1 are complete then so is X0 ∩ X1 equipped with · X0 ∩X1 . Thus, X0 ∩ X1 is
a quasi-Banach space whenever X0 , X1 are so.
Consider next a Hausdorff topological vector space V and suppose

X0, · X0 , X1, · X1 are quasi-normed spaces

(1.3.4)
satisfying X0 → V and X1 → V continuously.

Define the sum of these spaces as

 
X0 + X1 := x ∈ V : x = x0 + x1, for some x0 ∈ X0, x1 ∈ X1 (1.3.5)

and set

x X0 +X1 := inf x0 X0 + x1 X1 for each x ∈ X0 + X1 . (1.3.6)

x=x0 +x1
x0 ∈X0, x1 ∈X1

Furthermore, define
 
x X0 ∩X1 := max x X0 , x X1 for each x ∈ X0 ∩ X1 . (1.3.7)

Lemma 1.3.1 If X0, · X0 , X1, · X1 are as in (1.3.4), then

X0 + X1, · X0 +X1 and X0 ∩ X1, · X0 ∩X1 are quasi-normed spaces. (1.3.8)

26 1 Preliminary Functional Analytic Matters

Proof Start by picking x ∈ X0 + X1 such that x X0 +X1 = 0. This and (1.3.6) imply
(j) (j)
that there exist sequences {x0 } j ∈N ⊆ X0 and {x1 } j ∈N ⊆ X1 with the property that
(j) (j) (j) (j)
x = x0 + x1 for each j ∈ N and such that lim x0 = 0 in X0 and lim x1 = 0 in X1 .
j→∞ j→∞
(j) (j) (j) (j)
In particular, lim x = 0 and lim x = 0 in V, forcing x = lim x0 + x1 = 0.
j→∞ 0 j→∞ 1 j→∞
This proves that · X0 +X1 satisfies the nondegeneracy property.
Next, to show homogeneity, let x ∈ X0 + X1 and λ ∈ C be arbitrary. Then, for
every x0 ∈ X0 and x1 ∈ X1 with x = x0 + x1 we have λx = λx0 + λx1 , thus

λx X0 +X1 ≤ λx0 X0 + λx1 X1 = |λ| x0 X0 + x1 X1 . (1.3.9)

Taking the infimum in (1.3.9) over all decompositions of x as x = x0 + x1 for x0 ∈ X0

and x1 ∈ X1 it follows that

|λ| x X0 +X1 ≥ λx X0 +X1 . (1.3.10)

For the reversed inequality, assuming λ  0 and using (1.3.10) with λ replaced by
λ−1 we obtain

x X0 +X1 = λ−1 (λx) X0 +X1

≤ |λ−1 | λx X0 +X1 (1.3.11)

hence |λ| x X0 +X1 ≤ λx X0 +X1 . The homogeneity in the case λ = 0 is immediate.

To see that · X0 +X1 satisfies a quasi-triangle inequality, pick x, y ∈ X0 + X1 and
for each

x0 , y0 ∈ X0 , x1 , y1 ∈ X1 , with x = x0 + x1 , y = y0 + y1 , (1.3.12)

write x + y = (x0 + y0 ) + (x1 + y1 ) and estimate

x+y X0 +X1 ≤ x0 + y0 X0
+ x1 + y1 X1

≤ C0 x0 X0 + y0 X0 + C1 x1 X1 + y1 X1

≤ C x0 X0 + x1 X1 + C y0 X0 + y1 X1 , (1.3.13)

where C0 and C1 , respectively, are the constants in the quasi-triangle inequality the
quasi-norms · X0 and · X0 satisfy, and C := max{C0, C1 }. Taking now the infimum
over all decompositions as in (1.3.12) in the resulting inequality from (1.3.13) we
obtain

x+y X0 +X1 ≤C x X0 +X1 + y X0 +X1 , (1.3.14)

as wanted. In summary, we have proved that X0 + X1, · X0 +X1 is a quasi-normed

space. That · X0 ∩X1 is a quasi-norm on X0 ∩ X1 is immediate from (1.3.7) and the
assumption that · X0 and · X1 are quasi-norms on X0 and X1 , respectively. 
Sums and intersections of compatible pairs of quasi-Banach spaces are themselves
quasi-Banach spaces; this is made precise in the next lemma.
1.3 Real Interpolation of Quasilinear Operators 27

Lemma 1.3.2 Suppose X0, · X0 , X1, · X1 are two quasi-Banach spaces with the
property that X0 → V and X1 → V continuously for some Hausdorff topological
vector space V. Then

X0 + X1, · X0 +X1 and X0 ∩ X1, · X0 ∩X1 are quasi-Banach spaces. (1.3.15)

Proof In light of Lemma 1.3.1 there remains to prove completeness. To this end, let
C0 , C1 be constants in [1, ∞) with the property that

x+y X0 ≤ C0 x X0 + y X0 for every x, y ∈ X0, (1.3.16)

x+y X1 ≤ C1 x X1 + y X1 for every x, y ∈ X1, (1.3.17)

and pick   
r ∈ 0, min (1 + log2 C0 )−1, (1 + log2 C1 )−1 . (1.3.18)

Then [138, Theorem 3.27, pp. 108-109] guarantees that there exist a quasi-norm · 0
on X0 and a quasi-norm · 1 on X1 such that

· 0 is equivalent with · X0 and · 1 is equivalent with · X1 , (1.3.19)

d0 (x, y) := x − y r
0 for x, y ∈ X0 is a distance on X0 , (1.3.20)
d1 (x, y) := x − y r
1 for x, y ∈ X1 is a distance on X1 . (1.3.21)

Thanks to (1.3.19), there is no loss of generality in assuming that

X0 + X1 is endowed with the quasi-norm · X0 +X1 defined as

(1.3.22)
in (1.3.5) but with · Xi replaced by · i for i ∈ {0, 1}.

With this convention, we will show that X0 + X1, · X0 +X1 is complete. Pick an
arbitrary sequence {x (n) }n∈N ⊆ X0 + X1 which is Cauchy in the quasi-norm · X0 +X1 .
Consequently, this sequence has a subsequence {y (n) }n∈N with the property that

y (k+1) − y (k) X0 +X1 < 2−k for all k ∈ N. (1.3.23)

In particular, there exist two sequences {y0(k) }n∈N ⊆ X0 and {y1(k) }n∈N ⊆ X1 such
that

y (k+1) − y (k) = y0(k) + y1(k), y0(k) 0 < 2−k , y1(k) 1 < 2−k , ∀k ∈ N. (1.3.24)
 N (k) 
This implies that the sequence k=1 y0 N ∈N
is Cauchy with respect to · 0r , thus
in light of (1.3.19), Cauchy with respect to · X0 . Given that X0, · X0 is quasi-
(k)
Banach, the latter sequence converges to some y0 ∈ X0 , so that y0 = ∞ k=1 y0 .
(k)
Similarly, there exists some y1 ∈ X1 such that y1 = ∞ k=1 y1 . Hence, recalling
(1.3.24) we obtain
28 1 Preliminary Functional Analytic Matters


n 
n 
n
y (n+1) − y (1) = (y (k+1) − y (k) ) = y0(k) + y1(k) (1.3.25)
k=1 k=1 k=1

which further implies


∞ 
∞
y (n+1) − y0 − y1 − y (1) X0 +X1
= − y0(k) − y1(k)
X0 +X1
k=n+1 k=n+1


∞ 
∞
≤ y0(k) + y1(k) . (1.3.26)
0 1
k=n+1 k=n+1

Observe that (1.3.20) implies

x+y r
0 = d0 (x, −y) ≤ d0 (x, 0) + d0 (0, −y) = x r
0 + y r
0 (1.3.27)

for every x, y ∈ X0 . An iteration of (1.3.27) implies that for each N, M ∈ N we have


N r 
N 
N
y0(k) y0(k) (2−r )k .
r
≤ 0
< (1.3.28)
0
k=M k=M k=M

Similarly,

N r 
N 
N
y1(k) y1(k) (2−r )k .
r
≤ 1
< (1.3.29)
1
k=M k=M k=M

Since the series ∞ −r k

k=1 (2 ) is convergent, (1.3.28)-(1.3.29) imply that the last ex-
pression in (1.3.26) may be made arbitrarily small by choosing n ∈ N sufficiently
large. Hence, {y (n) }n∈N converges to y0 + y1 + y (1) in X0 + X1 . This proves that
the Cauchy sequence {x (n) }n∈N contains a convergent subsequence which ultimately
forces {x (n) }n∈N to be convergent to the same limit. This completes the proof of the
fact that X0 + X1, · X0 +X1 is complete.
Next, pick {x (n) }n∈N ⊆ X0 ∩ X1 which is Cauchy in the quasi-norm · X0 ∩X1 and
let ε ∈ (0, ∞). Then there exists Nε ∈ N such that

x (n) − x (m) 0
< ε and x (n) − x (m) 1
< ε for all n, m ≥ Nε . (1.3.30)

In particular, {x (n) }n∈N is Cauchy in the quasi-Banach spaces X0, · X0 and

X1, · X1 , thus convergent. Let x0 ∈ X0 and x1 ∈ X1 be such that

lim x (n) − x0 0
= 0 and lim x (n) − x1 1
= 0. (1.3.31)
n→∞ n→∞

This, the continuous embeddings X0 → V, X1 → V, and the fact that V is

Hausdorff imply that x0 = x1 . Hence, x := x0 = x1 ∈ X0 ∩ X1 . From (1.3.30) we
know that

d0 x (n), x (m) < εr and d1 x (n), x (m) < εr for all n, m ≥ Nε . (1.3.32)
1.3 Real Interpolation of Quasilinear Operators 29

Keeping in mind (1.3.20)-(1.3.21) and that, in general, distances are Hölder, hence
continuous, in each argument, we may then pass to the limit with m → ∞ in the
inequalities in (1.3.32) to arrive at

x (n) − x 0
< ε and x (n) − x 1
< ε for all n ≥ Nε . (1.3.33)

Since ε > 0 is arbitrary, from (1.3.33) and (1.3.7) we obtain that {x (n) }n∈N converges
to x in · X0 ∩X1 . Ergo, the quasi-normed space X0 ∩ X1, · X0 ∩X1 is complete.

Going further, fix a parameter θ ∈ (0, 1) along with 0 < q ≤ ∞. For each t > 0
and x ∈ X0 + X1 define

K(t, x, X0, X1 ) := inf x0 X0 + t x1 X1 . (1.3.34)

x=x0 +x1
x0 ∈X0, x1 ∈X1

Note that for each fixed x ∈ X0 + X1 we have

K(t, x, X0, X1 ) = tK(t −1, x, X1, X0 ) for each t > 0, (1.3.35)

[0, ∞) t → K(t, x, X0, X1 ) is non-decreasing, and (1.3.36)

1
2 K(2t, x, X0, X1 ) ≤ K(t, x, X0, X1 ) ≤ K(2t, x, X0, X1 ) for each t ≥ 0. (1.3.37)

Then for each x ∈ X0 + X1 introduce the quasi-norm

⎧ ∫ ∞  1
⎪
⎪ −θ q dt q
⎪
⎨
⎪ t K(t, x, X ,
0 1X ) if 0 < q < ∞,
0 t
x (X0,X1 )θ, q :=  −θ (1.3.38)
⎪
⎪
⎪
⎪ sup t K(t, x, X0, X1 ) if q = ∞,
⎩ t>0
and define the intermediate space (for the real method of interpolation) between X0
and X1 as  
(X0, X1 )θ,q := x ∈ X0 + X1 : x (X0,X1 )θ, q < +∞ . (1.3.39)
From (1.3.39), (1.3.38), and (1.3.35) it is then clear that

(X0, X1 )θ,q = (X1, X0 )1−θ,q . (1.3.40)

It also turns out that the ambient space X plays only a minor role in the definition of
(X0, X1 )θ,q . For example, this may be replaced by X0 + X1 . In this vein, let us note
that, as is apparent from (1.3.39),

(X0, X1 )θ,q → X0 + X1 continuously,

(1.3.41)
for each θ ∈ (0, 1) and q ∈ (0, ∞].

Remark 1.3.3 The following are true (see, e.g., [187, pp. 63–64]).
30 1 Preliminary Functional Analytic Matters

(i) If A0 and A1 are two compatible quasi-Banach spaces and θ ∈ (0, 1) and
q ∈ (0, ∞] are arbitrary, the real interpolation space (A0, A1 )θ,q is itself a
quasi-Banach space.
(ii) If X is a Hausdorff topological vector space and A0, A1, B0, B1 are quasi-
Banach spaces such that A0 → B0 → X and A1 → B1 → X continuously,
then for each θ ∈ (0, 1) and q ∈ (0, ∞] it follows that

(A0, A1 )θ,q and (B0, B1 )θ,q are quasi-Banach spaces and

(1.3.42)
the embedding (A0, A1 )θ,q → (B0, B1 )θ,q is continuous.

Given any x ∈ X0 ∩ X1 along with any t > 0, writing x = x + 0 proves, in

the context of (1.3.34), that K(t, x, X0, X1 ) ≤ x X0 , while decomposing x = 0 + x
shows, once again the context of (1.3.34), that K(t, x, X0, X1 ) ≤ t x X1 . Together,
these observations establish the estimate
 
K(t, x, X0, X1 ) ≤ min x X0 , t x X1 for each x ∈ X0 ∩ X1 and t > 0. (1.3.43)

In turn, (1.3.43) and (1.3.38)-(1.3.39) permit us to conclude that

X0 ∩ X1 → (X0, X1 )θ,q continuously,

(1.3.44)
for each θ ∈ (0, 1) and q ∈ (0, ∞].

Other basic, general properties satisfied by the intermediate spaces obtained via
the real method of interpolation may be found in [15, §3.4]. For example, [15,
Theorem 3.4.2(b), p. 47] gives that

the set X0 ∩ X1 is dense in the space (X0, X1 )θ,q

(1.3.45)
for each parameters θ ∈ (0, 1) and q ∈ (0, ∞).
The next lemma shows that, given two quasi-normed spaces with the property that
one embeds continuously into the other, we may define an equivalent quasi-norm on
the corresponding intermediate spaces by restricting the parameter t in (1.3.38) to a
smaller sub-interval of (0, ∞).

Lemma 1.3.4 Assume X0 , X1 are two quasi-normed spaces, and fix θ ∈ (0, 1) along
with q ∈ (0, ∞]. Also, pick an arbitrary number t∗ ∈ (0, ∞). Then, if

X0 → X1 continuously, (1.3.46)

an equivalent quasi-norm on the intermediate space (X0, X1 )θ,q is given by

⎧ ∫ ∞  dt  q1
⎪
⎪ t −θ K(t, x, X0, X1 )
q
⎪
⎨
⎪ if 0 < q < ∞,
∗ t∗ t
x (X0,X1 ) θ, q :=  −θ (1.3.47)
⎪
⎪
⎪ sup t K(t, x, X0, X1 )
⎪ if q = ∞,
⎩ t>t∗
for each x ∈ (X0, X1 )θ,q . If, on the other hand,
1.3 Real Interpolation of Quasilinear Operators 31

X1 → X0 continuously, (1.3.48)

then an equivalent quasi-norm on (X0, X1 )θ,q is given by

 ∫
⎧
⎪
⎪
t∗  dt  q1
t −θ K(t, x, X0, X1 )
q
⎪
⎨
⎪ if 0 < q < ∞,
∗∗ 0 t
x (X0,X1 ) θ, q :=  −θ (1.3.49)
⎪
⎪
⎪ sup t K(t, x, X0, X1 )
⎪ if q = ∞,
⎩ t ∈(0,t∗ )
for each x ∈ (X0, X1 )θ,q .

Proof Granted (1.3.46), any x ∈ X0 + X1 = X1 admits the trivial decomposition

x = 0 + x with 0 ∈ X0 and x ∈ X1 , it follows from (1.3.34) that

K(t, x, X0, X1 ) ≤ t x X1 for each t ∈ (0, ∞)

(1.3.50)
and each x ∈ X0 + X1 = X1 .

On the other hand, if x ∈ X0 +X1 = X1 is written as x = x0 +x1 for some x0 ∈ X0 ⊆ X1

and x1 ∈ X1 , then for every t ∈ (t∗, ∞) we may estimate, using the quasi-triangle
inequality and the continuity of the embedding X0 → X1 ,

x X1 ≤ C x0 X1 + x1 X1 ≤ C x0 X0 + x1 X1

≤C· max{1, t∗−1 } · x0 X0 + t x1 X1 . (1.3.51)

Then, in view of (1.3.34), taking the infimum over all such decompositions of x
(with t fixed) yields

K(t, x, X0, X1 ) ≥ c x X1 for each t ∈ (t∗, ∞)

(1.3.52)
and each x ∈ X0 + X1 = X1,

where c ∈ (0, ∞) depends only on X0, X1 and t∗ . Consequently, assuming q ∈ (0, ∞),
for each vector x ∈ X0 + X1 = X1 we may estimate
 ∫ t∗  1  ∫ t∗  1
q dt q q dt q
t −θ K(t, x, X0, X1 ) ≤ x X1 · t 1−θ 
= Cq,θ,t ∗
x X1
0 t 0 t
∫ ∞  1
 q dt q
= Cq,θ,t∗ x X1 · t −θ
t∗ t
∫ ∞  1
−1  q dt q
≤ c · Cq,θ,t∗ t −θ K(t, x, X0, X1 )
t∗ t
(1.3.53)

with c as in (1.3.52) and Cq,θ,t 
, Cq,θ,t > 0 some finite constants depending only
∗ ∗
on q and θ, where the first inequality uses (1.3.50) and the last inequality is based
on (1.3.52). Similarly, corresponding to q = ∞, for each x ∈ X0 + X1 = X1 we may
32 1 Preliminary Functional Analytic Matters

estimate
 
sup t −θ K(t, x, X0, X1 ) ≤ x X1 · sup t 1−θ = t∗1−θ x X1
t ∈(0,t∗ ) t ∈(0,t∗ )

= t∗ x X1 · sup t −θ
t ∈(t∗,∞)

≤ c−1 t∗ · sup t −θ K(t, x, X0, X1 ) . (1.3.54)
t ∈(t∗,∞)

A combination of (1.3.38) with (1.3.53)-(1.3.54) then proves that (1.3.47) is indeed

an equivalent quasi-norm on the intermediate space (X0, X1 )θ,q if (1.3.46) holds.
In the second part of the proof assume (1.3.48) holds. Since any x ∈ X0 + X1 = X0
admits the trivial decomposition x = x + 0 with x ∈ X0 and 0 ∈ X1 , the definition of
the K-functional implies

K(t, x, X0, X1 ) ≤ x X0 for each t ∈ (0, ∞)

(1.3.55)
and each x ∈ X0 + X1 = X0 .

Suppose next that some x ∈ X0 +X1 = X0 has been given. Then for any decomposition
x = x0 + x1 with x0 ∈ X0 and x1 ∈ X1 ⊆ X0 and any t ∈ (0, t∗ ) we may estimate

t x X0 ≤ Ct x0 X0 + x1 X0 ≤ Ct x0 X0 + x1 X1

≤ C · max{1, t∗ } · x0 X0 + t x1 X1 , (1.3.56)

based on the quasi-triangle inequality and the continuity of the embedding X1 → X0 .

After taking the infimum over all such decompositions of x (with t ∈ (0, t∗ ) fixed)
we arrive at the conclusion that
K(t, x, X0, X1 ) ≥ c t x X0 for each t ∈ (0, t∗ )
(1.3.57)
and each x ∈ X0 + X1 = X0,

for some c ∈ (0, ∞) depending only on X0, X1 and t∗ . At this stage, if q ∈ (0, ∞) then
for each x ∈ X0 + X1 = X0 we may write, with Cq,θ,t  
, Cq,θ,t ∈ (0, ∞) depending
∗ ∗
only on q, θ, and t∗ ,
∫ ∞  1 ∫ ∞  1
q dt q q dt q
t −θ K(t, x, X0, X1 ) ≤ x X0 · t −θ 
= Cq,θ,t ∗
x X0
t∗ t t∗ t
 ∫ t∗  1
 1−θ q dt q
= Cq,θ,t∗
x X0 · t
0 t
∫
 t∗  1
q dt q
≤ c−1 · Cq,θ,t

∗
t −θ
K(t, x, X 0, X1 )
0 t
(1.3.58)
1.3 Real Interpolation of Quasilinear Operators 33

with the first inequality supplied by (1.3.55) and the last inequality originating in
(1.3.57). In an analogous fashion, if q = ∞ then for each x ∈ X0 + X1 = X0 we have
 
sup t −θ K(t, x, X0, X1 ) ≤ x X0 · sup t −θ ] = t∗−θ x X0
t ∈(t∗,∞) t ∈(t∗,∞)

= t∗−1 x X0 · sup t 1−θ
t ∈(0,t∗ )

≤ c−1 t∗−1 · sup t −θ K(t, x, X0, X1 ) . (1.3.59)
t ∈(0,t∗ )

From (1.3.38) and (1.3.58)-(1.3.59) we may now conclude that if (1.3.48) holds then
(1.3.49) is an equivalent quasi-norm on (X0, X1 )θ,q . 
We proceed by introducing quasi-normed lattices of functions on a given measure
space (Σ, μ). The reader is reminded that L 0 (Σ, μ) denotes the linear space consisting
of (equivalence classes of) μ-measurable functions which are finite μ-a.e. on Σ (see
the discussion in [133, S 3.2]).

Definition 1.3.5 Let (Σ, μ) be a measure space and assume that Y is a quasi-normed
space. Call Y a quasi-normed lattice of functions if

Y → L 0 (Σ, μ) continuously, (1.3.60)

and if there exists some constant C0 ∈ (0, ∞) with the property that whenever
f ∈ L 0 (Σ, μ) and g ∈ Y satisfy | f (x)| ≤ |g(x)| at μ-a.e. point x ∈ Σ then necessarily
f ∈ Y and f Y ≤ C0 g Y .

Parenthetically we note that if Y is as above then f ∈ Y is zero in Y if and only if

f (x) = 0 at μ-a.e. point x ∈ Σ.

Definition 1.3.6 Given a vector space X and assuming that Y is a linear subspace
of L 0 (Σ, μ), call the operator T : X → Y quasi-subadditive provided there exists
a finite constant C1 > 0 such that for all f , g ∈ X,

|T( f + g)| ≤ C1 |T f | + |T g| at μ-a.e. point in Σ. (1.3.61)

We are now in position to state and prove the following real interpolation result
for quasi-subadditive operators.

Proposition 1.3.7 Assume that X0 , X1 are two compatible quasi-normed spaces and
that Y0 , Y1 are two quasi-normed lattices of functions relative to a common measure
space (Σ, μ). Suppose
T : X0 + X1 −→ Y0 + Y1 (1.3.62)
is a quasi-subadditive operator satisfying T Xi ⊆ Yi for i ∈ {0, 1}, and such that there
exist M0 , M1 positive finite constants with the property that

Tx Yi ≤ Mi x Xi for all x ∈ Xi, with i ∈ {0, 1}. (1.3.63)

34 1 Preliminary Functional Analytic Matters

Then for each θ ∈ (0, 1) and q ∈ (0, ∞], the operator T maps (X0, X1 )θ,q into
(Y0, Y1 )θ,q and

Tx (Y0,Y1 ) θ, q ≤ C0 C1 M01−θ M1θ x (X0,X1 ) θ, q for every x ∈ (X0, X1 )θ,q (1.3.64)

where C0 ∈ (0, ∞) is a constant which works both for Y0 and Y1 as in Definition 1.3.5,
while the constant C1 ∈ (0, ∞) is associated with T as in Definition 1.3.6.

Before presenting the proof of this proposition we wish to make four comments.
First, if X is a Hausdorff topological vector space with the property that Xi → X
continuously for i ∈ {0, 1} (whose existence is guaranteed by the fact that X0 , X1 are
compatible quasi-normed spaces), then in place of (1.3.62) we may initially assume
that
T : X0 + X1 −→ X (1.3.65)
and the same proof will work. Our second comment is that, in a natural sense,
having an additive mapping T : X0 + X1 → Y0 + Y1 which satisfies
T Xi ⊆ Yi for i ∈ {0, 1} is equivalent to having two additive maps
Ti : Xi → Yi for i ∈ {0, 1} that are compatible with one another, in
the sense that T0 X0 ∩X1 = T1 X0 ∩X1 ; in the latter scenario, the operators (1.3.66)
induced by the mappings T0, T1 from (X0, X1 )θ,q into (Y0, Y1 )θ,q for
various values of θ ∈ (0, 1) and q ∈ (0, ∞] are additive and compatible
with one another (as they all agree with T : (X0, X1 )θ,q → (Y0, Y1 )θ,q ).

The third comment is that we may recast (1.3.64) simply as

1−θ θ
T (X0,X1 ) θ, q →(Y0,Y1 ) θ, q ≤ C0 C1 · T X0 →Y0 · T X1 →Y1 . (1.3.67)

Our fourth (and final) comment is that a version of Proposition 1.3.7 for linear
operators taking values in quasi-normed spaces is discussed in Remark 1.3.8.
Here is the proof of Proposition 1.3.7.
Proof of Proposition 1.3.7 Let x ∈ (X0, X1 )θ,q ⊆ X0 + X1 and consider x0 ∈ X0 ,
x1 ∈ X1 such that x = x0 + x1 . Thus, by quasi-additivity of T,

|T x| ≤ C1 |T x0 | + |T x1 | at μ-a.e. point on Σ. (1.3.68)

Introduce  
E := ξ ∈ Σ : |T x0 |(ξ) + |T x1 |(ξ) = 0 , (1.3.69)
so that E is μ-measurable and since (1.3.68) holds, it follows that

T x = 0 at μ-a.e. point in E. (1.3.70)

Next, for i ∈ {0, 1} set

1.3 Real Interpolation of Quasilinear Operators 35

⎧  |T xi | 
⎪
⎨
⎪ T x on Σ \ E,
yi := |T x0 | + |T x1 | (1.3.71)
⎪
⎪
⎩ 0 on E.

Then each yi is μ-measurable and, thanks to the quasi-subadditivity of T, for each

i ∈ {0, 1} we have the pointwise inequality
 |T xi | 
|yi | ≤ · C1 |T x0 | + |T x1 | = C1 |T xi | (1.3.72)
|T x0 | + |T x1 |
at μ-a.e. point in Σ. Consequently, since each Yi is a quasi-normed lattice of functions
on (Σ, μ), we may conclude (using the boundedness of T) that

yi ∈ Yi and yi Yi ≤ C0 C1 T xi Yi ≤ C0 C1 Mi xi Xi , i ∈ {0, 1}. (1.3.73)

Also, analyzing what happens on Σ \ E and E separately and using (1.3.70), we see
that
y0 + y1 = T x at μ-a.e. point in Σ, (1.3.74)
which shows that T x ∈ Y0 + Y1 . As a result,

K(t, T x, Y0, Y1 ) ≤ y0 + t y1
Y0 Y1
 tM  
1
≤ C0 C1 M0 x0 X0 + x1 X1 . (1.3.75)
M0

Taking the infimum over all representations of x = x0 + x1 yields

tM 
1
K(t, T x, Y0, Y1 ) ≤ C0 C1 M0 K , x, X0, X1 . (1.3.76)
M0
Now, if 0 < q < ∞ and θ ∈ (0, 1),
∫ ∞  tM   q dt  q1
t −θ K
1
Tx (Y0,Y1 ) θ, q ≤ C0 C1 M0 , x, X0, X1
0 M0 t
 M  −θ ∫ ∞   q dτ  q1
τ −θ K τ, x, X0, X1
0
= C0 C1 M0
M1 0 τ
= C0 C1 M01−θ M1θ x (X0,X1 ) θ, q < ∞, (1.3.77)

i.e., (1.3.64) holds. In concert with the membership T x ∈ Y0 + Y1 , this implies

T x ∈ (Y0, Y1 )θ,q , which completes the proof. 
Remark 1.3.8 A similar result to that established in Proposition 1.3.7 holds in the
case when T is linear and Y0, Y1 are now two compatible quasi-normed spaces.
Indeed, a cursory inspection shows that the previous proof works virtually verbatim
with the added benefit that we may now take C0 = C1 = 1, i.e., in place of (1.3.64)
we now have
36 1 Preliminary Functional Analytic Matters

Tx (Y0,Y1 ) θ, q ≤ M01−θ M1θ x (X0,X1 ) θ, q for every x ∈ (X0, X1 )θ,q . (1.3.78)

See also [15, p. 41] in this regard.

We close by recording a versatile version of the Reiteration Theorem, for quasi-
normed Abelian couples.
Theorem 1.3.9 Let A0, A1 be two quasi-normed Abelian groups which are subgroups
of some larger ambient Abelian group A, and which satisfy the following separability
axiom (or compatibility of convergence condition):

if the sequence {x j } j ∈N ⊆ A0 ∩ A1 and a0 ∈ A0 and a1 ∈ A1 are

such that lim x j = a0 in A0 and lim x j = a1 in A1 , then necessarily (1.3.79)
j→∞ j→∞
a0 = a1 .
Then for each θ 0, θ 1, θ ∈ (0, 1) with θ 0  θ 1 and each q0, q1, q ∈ (0, ∞] one has

(A0, A1 )θ0,q0 , (A0, A1 )θ1,q1 θ,q = (A0, A1 )(1−θ)θ0 +θθ1,q . (1.3.80)

In addition, the following two end-point versions of (1.3.80) (corresponding,

roughly speaking, to the cases θ 0 = 0 and θ 1 = 1, respectively) also hold:

A0, (A0, A1 )θ1,q1 θ,q = (A0, A1 )θθ1,q and (1.3.81)

(A0, A1 )θ0,q0 , A1 θ,q = (A0, A1 )(1−θ)θ0 +θ,q . (1.3.82)

Moreover, formulas (1.3.80)-(1.3.82) are quantitative in the sense that, in each

case, the quasi-norm of the space on the left is equivalent to the quasi-norm of the
corresponding space on the right, with constants of equivalence depending only on
θ 0, θ 1, θ, q0, q1, q as well as the constants in the quasi-triangle inequalities in A0 and
A1 .
Theorem 1.3.9 is proved for quasi-normed spaces in [98] but essentially the same
proof works for quasi-normed Abelian groups. Other variants and generalizations of
Theorem 1.3.9 may be found in [20] and [154].

1.4 Complex Interpolation on Quasi-Banach Spaces

The presentation in this section follows closely [111]. Throughout, (1.2.10) is tacitly
employed. To set the stage for adapting Calderón’s original complex method of
interpolation to the setting of quasi-Banach spaces, we first review some basic
results from the theory of analytic functions with values in quasi-Banach spaces as
developed in [193], [108], [109].
Recall that if X is a topological vector space and U is an open subset of the
complex plane then a function f : U → X is called analytic if given z0 ∈ U there
exists η > 0 so that there is a power series expansion
1.4 Complex Interpolation on Quasi-Banach Spaces 37


∞
f (z) = (z − z0 )n xn, xn ∈ X, uniformly convergent for |z − z0 | < η. (1.4.1)
j=0

As explained in [108], in the context of quasi-Banach spaces, this is the most

natural definition. Indeed, there are simple examples which show that complex
differentiability leads to an unreasonably weaker concept of analyticity (see also
[193] and [8] in this regard).

Proposition 1.4.1 Suppose 0 < p ≤ 1 and m ∈ N is such that m > p1 . Then there is
a constant C = C(m, p) with the following significance. If X is a p-normed quasi-
Banach space and f : D̄ → X is a continuous function which is analytic on the unit
f (n) (0) n
disk D := {z : |z| < 1} then for z ∈ D we have f (z) = ∞
n=0 n! z , and

f (n) (0) X ≤ C(m + n)! sup f (z) X. (1.4.2)

z ∈D

This is [108, Theorem 6.1].

Proposition 1.4.2 Let X be a quasi-Banach space and let U be an open subset of

the complex plane. Let fn : U → X, n ∈ N, be a sequence of analytic functions. If
lim fn = f uniformly on compacta then f is also analytic.
n→∞

This follows from [108, Theorem 6.3].

Proposition 1.4.3 Suppose X is a quasi-Banach space and that U is an open subset

of the complex plane. Let f : U → X be a locally bounded function. Suppose there
is a weaker Hausdorff vector topology τ0 on X which is locally p-convex for some
0 < p < 1 and such that f : U → (X, τ0 ) is analytic. Then f : U → X is analytic.

This is [111, Theorem 3.3]. It shows that, many times in practice, the ambient
space (within which the interpolation process is carried out) plays only a minor role
in the setup. More specifically, assume that Y is a space of distributions in which a
quasi-Banach space X is continuously embedded. Then, having an X-valued function
analytic for the quasi-norm topology is basically the same as requiring analyticity
for the weak topology (induced on X from Y ).
We are now prepared to elaborate on the complex method of interpolation for pairs
of quasi-Banach spaces. Consider a compatible couple (or pair) of quasi-Banach
spaces X0, X1 , i.e., X j with j ∈ {0, 1} are continuously embedded into a larger
topological vector space Y . Also, let U stand for the strip {z ∈ C : 0 < Re z < 1}.

Definition 1.4.4 A family F of functions which map U into X0 + X1 is called

admissible provided the following axioms are satisfied:
(i) F is a (complex) vector space endowed with a quasi-norm · F with respect to
which it is complete (i.e., F is a quasi-Banach space);
(ii) the point-evaluation mappings evw : F → X0 + X1 , w ∈ U, defined at each
f ∈ F by evw ( f ) := f (w), are continuous;
38 1 Preliminary Functional Analytic Matters

(iii) for any given compact subset K of U there exists some finite positive constant C
with the property that, for any fixed point w ∈ K and any fixed function f ∈ F
with f (w) = 0, the mapping U \ {w} z → f (z)/(z − w) ∈ X0 + X1 extends to
an element in F and
f (z)
≤ C f F. (1.4.3)
z−w F

Conditions (i)-(iii) in Definition 1.4.4 are the minimal requirements needed in

order develop a reasonable interpolation theory at an abstract level. In practice,
mimicking the Banach space theory, a common choice for F , the class of admissible
functions, is the space of bounded, analytic functions f : U → X0 + X1 , which is
extended continuously to the closure of the strip such that the traces t → f ( j + it) are
bounded continuous functions into X j , j ∈ {0, 1}. We endow F with the quasi-norm
!
f F := max sup f (it) X0 , sup f (1 + it) X1 . (1.4.4)
t t

However, there is an immediate problem that in general the evaluation maps evw are
not necessarily bounded on F and so this class is not always admissible. In fact in
the special case when X0 = X1 boundedness of the evaluation maps is equivalent to
the validity of a form of the Maximum Modulus Principle. A quasi-Banach space
X is analytically convex if there is a constant C such that for every polynomial
P : C → X we have P(0) X ≤ C max |z |=1 P(z) X . It is shown in [109] that if X is
analytically convex it has an equivalent quasi-norm which is plurisubharmonic (i.e.,
we can insist that the constant C above can be taken to be 1). Let us also point out
that being analytically convex is equivalent to the condition that

max f (z) X ≤C max f (z) X, (1.4.5)

0<Re z<1 Re z ∈ {0,1}

for any analytic function f : {z ∈ C : 0 < Re z < 1} → X which is continuous on

the closed strip {z ∈ C : 0 ≤ Re z ≤ 1}.
The relevance of the concept of analytic convexity in the current context suggests
we take a more systematic look at it. Clearly, any Banach space is analytically convex.
Other useful criteria for analytic convexity are summarized in the following result,
proved in [109], [44].

Theorem 1.4.5 For a quasi-Banach space (X, · X) the following conditions are
equivalent:
(i) X is analytically convex;
(ii) X has an equivalent quasi-norm · which is plurisubharmonic, i.e.
∫ 2π
1
x + eiθ y dθ ≥ x for each x, y ∈ X; (1.4.6)
2π 0

(iii) X has an equivalent quasi-norm · so that log · is plurisubharmonic;

1.4 Complex Interpolation on Quasi-Banach Spaces 39

(iv) X has an equivalent quasi-norm · so that · p is plurisubharmonic for some

0 < p < ∞;
(v) X has an equivalent quasi-norm · so that · p is plurisubharmonic for each
0 < p < ∞;
(vi) there exists a finite constant C > 0 with the property that
 
max{ f (z) X : 0 < Re z < 1} ≤ C · max f (z) X : Re z ∈ {0, 1} (1.4.7)

for any analytic function f : U → X, where U := {z ∈ C : 0 < Re z < 1},

which extends in a continuous and bounded fashion on the closed strip U.
We also have the following result.
Proposition 1.4.6 If X is an analytically convex quasi-Banach space and Y is a
closed subspace of X then Y is also analytically convex.
Proof This is seen directly from definitions, or as a consequence of the equivalence
(i) ⇔ (vi) in Theorem 1.4.5. 
Other examples of analytically convex quasi-Banach spaces can be manufactured
by means of the observation made in Proposition 1.4.15 stated later. This parallels
[44, Proposition 2.3] and involves Lebesgue spaces of quasi-Banach valued func-
tions. In contrast to the case of genuine Banach spaces, generally speaking there is
no satisfactory integration theory for functions which take values in quasi-Banach
spaces7. This being said, the notion of L p quasi-norm for functions taking values
in a quasi-normed space is meaningful. To discuss this in detail, we first make a
definition.
Definition 1.4.7 Let (Σ, M, μ) be a measure space, assume V is a topological vector
space, and consider a function f : Σ → V . Recall that f is said to be μ-measurable
if f −1 (O) ∈ M for each O open subset of V .
In this setting, call f countably simple if there exist a sequence of vectors
{a j } j ∈N ⊆ V and family of pairwise disjoint sets { A j } j ∈N ⊆ M such that there
holds f = aj 1Aj .
j ∈N
Finally, call f strongly μ-measurable if there exist a sequence { f j } j ∈N of
countably simple functions along with a μ-nullset A ∈ M such that for each ω ∈ Σ\ A
one has f j (ω) → f (ω) in the topology of V as j → ∞.
Remark 1.4.8 Let (Σ, M, μ) be a measure space, and let V be a topological vector
space.
(i) Straight from definitions it is clear that any countably simple function is
μ-measurable.
(ii) If the measure μ is complete, then any function which is the μ-a.e. limit of a
sequence of μ-measurable V -valued functions defined on Σ is itself μ-measurable.
7 Indeed, the existence of a reasonable notion of integral in the class functions taking values in a
given quasi-Banach space essentially forces the space in question to be locally convex
40 1 Preliminary Functional Analytic Matters

To see that this the case, assume A ∈ M is a μ-nullset and suppose { f j } j ∈N is a

sequence of μ-measurable functions such that the limit f (ω) := lim f j (ω) exists in
j→∞
V at each ω ∈ Σ \ A. To show that the function f itself is μ-measurable, for each
open subset O of V we write
" ∞ ∞ #

−1 −1 −1
f (O) = A ∩ f (O) ∪ f j (O) \ A . (1.4.8)
k=1 j=k

Bearing in mind that A ∩ f −1 (O is μ-measurable (since it is a subset of the μ-nullset

A, and the measure μ is complete), it follows that f −1 (O) ∈ M, so the desired
conclusion follows.
(iii) Combining parts (i)-(ii) above, we see that if the measure μ is complete then
any strongly μ-measurable function is μ-measurable.

Lemmas 1.4.9-1.4.12 build up to the conclusion that, in a suitable setting, the

quality of being a strongly measurable function is preserved under pointwise con-
vergence.

Lemma 1.4.9 Let Σ be an arbitrary set, and let V be a quasi-metric space. Consider
a sequence { f j } j ∈N of V -valued functions defined on Σ along with a family { A j } j ∈N
of subsets of Σ such that f j (Σ \ A j ) is a separable subset of V for each j ∈ N, and
the limit

f (ω) := lim f j (ω) exists at each ω ∈ Σ \ A, where A := Aj . (1.4.9)
j→∞
j ∈N

Then f (Σ \ A) is a separable subset of V .

Proof Since each f j (Σ \ A j ) is separable, it follows that (with ‘bar’ denoting closure
in the topology of V )

W := f j (Σ \ A j ) is a separable subset of V . (1.4.10)
j ∈N

Observe that f (Σ \ A) ⊆ W. Indeed, if w ∈ f (Σ \ A) then there exists ω ∈ Σ \ A

with the property that w = f (ω), hence w = lim f j (ω) ∈ W, as claimed. Granted
j→∞
this, the desired conclusion follows upon recalling that separability in a quasi-metric
space is hereditary (cf. [133, Lemma 7.1.1]). 

Lemma 1.4.10 Suppose (Σ, M, μ) is a measure space, and let V be a topological

vector space whose topology is induced by a quasi-metric. Then for each strongly
μ-measurable function f : Σ → V there exists some μ-nullset A ∈ M such that
f (Σ \ A) is a separable subset of V .

Proof Let { f j } j ∈N be a sequence of countably simple functions with the property

that there exists some μ-nullset A ∈ M such that f (ω) = lim f j (ω) in V for each
j→∞
1.4 Complex Interpolation on Quasi-Banach Spaces 41

ω ∈ Σ \ A. Since each f j (Σ \ A) is at most countable, hence a separable subset of V ,

the conclusion follows from Lemma 1.4.9 (used with A j := A). 
Lemma 1.4.11 Suppose (Σ, M, μ) is a measure space, and let V be a topological
vector space whose topology is induced by a quasi-metric. Assume f : Σ → V is
a μ-measurable function with the property that there exists some μ-nullset A ∈ M
such that f (Σ \ A) is a separable subset of V .
Then there exists a sequence { f j } j ∈N of countably simple functions which con-
verges to f pointwise on Σ \ A, in a uniform fashion. In particular, f is a strongly
μ-measurable function.
Proof Let ρ be the quasi-metric inducing the topology on V and denote by ρ# its
regularization constructed in [133, Theorem 7.1.2]. Also, let {a j } j ∈N be a countable
dense subset of f (Σ \ A). To proceed, fix ε > 0 arbitrary and, for each j ∈ N, define
" #

−1 −1
A j := f Bρ# (a j , ε) \ f Bρ# (ak , ε) . (1.4.11)
k< j

The fact that f is μ-measurable and that ρ# -balls are open implies that each A j
belongs to M. In addition, by design, the sets { A j } j ∈N are mutually disjoint. Granted
these properties, it follows that g := j ∈N a j 1 A j is a countably simple function. Next,
having {a j } j ∈N dense in f (Σ\ A) guarantees that for each ω ∈ Σ\ A there exists j ∈ N
such that ρ# (a j , f (ω)) < ε. As such, f (ω) ∈ Bρ# (a j , ε), hence ω ∈ f −1 Bρ# (a j , ε)
$
which further
$ entails Σ\ A ⊆ j ∈N f −1 Bρ# (a j , ε) . Ultimately, this goes to show that
Σ \ A = j ∈N A j . From this and the definition of the function g we then conclude that
ρ# g(ω), f (ω) < ε for each ω ∈ Σ \ A. With this in hand, the desired conclusions
follow on account of the arbitrariness of ε > 0, bearing in mind that ρ# yields the
same topology on V as ρ (cf. [133, Theorem 7.1.2]). 
Lemma 1.4.12 Let (Σ, M, μ) be a measure space with the property that μ is complete,
and let V be a topological vector space whose topology is induced by a quasi-
metric. Suppose { f j } j ∈N is a sequence of strongly μ-measurable V -valued functions
defined on Σ with the property that there exists some μ-nullset A ∈ M such that
f (ω) = lim f j (ω) in V for each ω ∈ Σ \ A.
j→∞
Then there exists a sequence {g j } j ∈N of countably simple functions which con-
verges to f pointwise on Σ \ A, in a uniform fashion. In particular, f is a strongly
μ-measurable function.
Proof From assumptions and parts (iii), (ii) in Remark 1.4.8 we see that f is μ-
measurable. Also, from Lemma 1.4.10 we know that for each j ∈ N there exists
some μ-nullset A j ∈ M such that f j (Σ \ A j ) is a separable subset of V . If for each
j ∈ N we now define A j := A ∪ A j ∈ M, it follows that A j is a μ-nullset. Also,
since separability in a quasi-metric space is hereditary (cf. [133, Lemma 7.1.1]),
 it
j ) is a separable subset of V . Then A $
 := j ∈N Aj ∈ M is
follows that each f j (Σ \ A
a μ-nullset and Lemma 1.4.9 implies that f (Σ \ A)  is a separable subset of V . With
this in hand, Lemma 1.4.11 applies and yields all desired conclusions. 
42 1 Preliminary Functional Analytic Matters

Given a complete measure space (Σ, M, μ) along with a quasi-Banach space

(X, · X ), for each p ∈ (0, ∞] we shall denote by L p (Σ, μ) ⊗ X the space of X-valued
strongly μ-measurable functions on Σ which are p-th power integrable with respect
to μ. More specifically,

L p (Σ, μ) ⊗ X is the collection of (equivalence classes of)

strongly μ-measurable functions f : Σ −→ X such that the (1.4.12)
mapping Σ ω −→ f (ω) X ∈ [0, ∞) belongs to L p (Σ, μ).

A few comments are in order here. First, observe that if f = a j 1 A j is a countably

j ∈N
simple function then f (·) X = aj X 1Aj becomes an ordinary scalar-valued
j ∈N
μ-measurable function on Σ. In particular, from this observation, Definition 1.4.7,
the convention made in (1.2.10), and [133, item (iii) in Remark 3.1.2] we conclude
that
given any strongly μ-measurable function f : Σ −→ X it
follows that the mapping Σ ω −→ f (ω) X ∈ [0, ∞) is a (1.4.13)
(ordinary, real-valued) μ-measurable function on Σ.

Second, in light of (1.4.13) we may rephrase the definition of L p (Σ, μ) ⊗ X given in

(1.4.12) as follows
L p (Σ, μ) ⊗ X is the collection of all (equivalence classes of) strongly
μ-measurable X-valued functions f defined on Σ with the property that
∫  1/p (1.4.14)
p
f (ω) X dμ(ω) < ∞ (with the usual alteration when p = ∞).
Σ

Third, it is clear that L p (Σ, μ) ⊗ X = L p (Σ, μ) ⊗ (X, · X) is a vector space, and

L p (Σ, μ) ⊗ X f −→ f L p (Σ, μ)⊗X := f (·) X (1.4.15)

L p (Σ, μ)

is a quasi-norm on the space L p (Σ, μ) ⊗ X. Also, it is clear from definitions that for
each r > 0 we have

f r
L p (Σ, μ)⊗(X, · X)
= f (·) r
X L p/r (Σ, μ)
(1.4.16)
for all f ∈ L p (Σ, μ) ⊗ (X, · X ).

Fourth,
if · X is an r-norm on X for some finite r ∈ (0, p], then
(1.4.17)
· L p (Σ, μ)⊗(X, ·X ) is an r-norm on L (Σ, μ) ⊗ (X,
p · X ).

Indeed, for each f , g ∈ L p (Σ, μ) ⊗ (X, · X) we may estimate

1.4 Complex Interpolation on Quasi-Banach Spaces 43

r
f +g r
L p (Σ, μ)⊗(X, · X)
= ( f + g)(·) X L p/r (Σ, μ)

≤ f (·) r
X + g(·) r
X
L p/r (Σ, μ)

≤ f (·) r
X + g(·) r
X
L p/r (Σ, μ) L p/r (Σ, μ)

= f r
L p (Σ, μ)⊗(X, · X)
+ g r
L p (Σ, μ)⊗(X, · X)
, (1.4.18)

from which (1.4.17) follows.

The space L p (Σ, μ) ⊗ X also turns out to be complete. Proving this takes some
work, and we address this issue separately, in Proposition 1.4.14 stated a little later
below. As a preamble, we first prove the following lemma.

Lemma 1.4.13 Let (V, · ) be a quasi-normed vector space with the property that
· is an r-norm for some r > 0. Also, assume that
∞
whenever {v j } j ∈N ⊆ V satisfies vj r < ∞ it follows
j=1
∞ (1.4.19)
that the series v j converges in the space (V, · ).
j=1

Then (V, · ) is complete.

Proof Let {w j } j ∈N ⊆ V be a Cauchy sequence in (V, · ). Then there exists a family

of integers 1 ≤ n1 < nn < · · · such that for each j ∈ N we have wn − wm < 2−j/r
whenever n, m ≥ n j . Define v1 := wn1 and v j := wn j − wn j−1 for each j ≥ 2. Then
k ∞ ∞
for each k ∈ N we have v j = wnk and vj r ≤ wn1 r + 2−(j−1) < +∞.
j=1 j=1 j=2
Granted these properties, (1.4.19) guarantees that {wnk }k ∈N converges in (V, · ).
Hence the Cauchy sequence {w j } j ∈N has a convergent sub-sequence in (V, · ).
This readily implies that actually the whole sequence {w j } j ∈N converges in (V, · ),
proving that (V, · ) is indeed complete. 

Here is the completeness result referred to above.

Proposition 1.4.14 Let (Σ, M, μ) be a complete measure

 space and consider a quasi-
Banach space (X, · X ). Also, fix p ∈ (0, ∞]. Then L p (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X
is
 a quasi-Banach space. Moreover,
 if X is Banach and p ∈ [1, ∞] then the space
L (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X is in fact Banach.
p

 
Proof From earlier comments we already know that L p (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X
is a quasi-normed vector space. To prove that this is also complete, consider first
the case when p < ∞. From (1.2.6) and (1.4.17) we know that there exists some
r ∈ 0, min{p, (1 + log2 ρ(X))−1 } with the property
44 1 Preliminary Functional Analytic Matters

· X is an r-norm on the space X, and also

(1.4.20)
· is an r-norm on L p (Σ, μ) ⊗ X.
L p (Σ, μ)⊗X

The idea is to use the criterion for completeness established earlier in Lemma 1.4.13
for the space V := L p (Σ, μ) ⊗ X and the r-norm · L p (Σ, μ)⊗X . With this strategy in
mind, consider a sequence { f j } j ∈N ⊆ L p (Σ, μ) ⊗ X satisfying


∞
M := fj r
L p (Σ, μ)⊗X < +∞. (1.4.21)
j=1

The goal is to prove that


∞  
the series f j converges in L p (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X . (1.4.22)
j=1

To this end, define


n
G n := f j (·) r
X : Σ → [0, ∞) for each n ∈ N, (1.4.23)
j=1

and

∞
G := f j (·) r
X : Σ → [0, ∞]. (1.4.24)
j=1

From (1.4.13) we know that each G n is μ-measurable, and G is μ-measurable as

well. Then for each n ∈ N we may estimate (bearing in mind that p/r > 1)

n 
∞
Gn L p/r (Σ, μ) ≤ f j (·) r
X ≤ fj r
L p (Σ, μ)⊗X = M < +∞,
L p/r (Σ, μ)
j=1 j=1
(1.4.25)
thanks to (1.4.16) and (1.4.21). Since we also have G n  G pointwise on Σ as
n → ∞, Lebesgue’s Monotone Convergence Theorem implies that

G L p/r (Σ, μ) = lim G n L p/r (Σ, μ) ≤ M < +∞. (1.4.26)

n→∞

Consequently,
G ∈ L p/r (Σ, μ). (1.4.27)
In particular, G(ω) < +∞ for μ-a.e. ω ∈ Σ which, in light of (1.4.24), goes to show
that
∞
f j (ω) Xr < +∞ for μ-a.e. ω ∈ Σ. (1.4.28)
j=1

If for each n ∈ N we now define

1.4 Complex Interpolation on Quasi-Banach Spaces 45


n
Fn (ω) := f j (ω) for each ω ∈ Σ, (1.4.29)
j=1

then whenever m, n ∈ N are such that m ≥ n it follows from the first assertion in
(1.4.20) that

m
Fm (ω) − Fn (ω) r
X ≤ f j (ω) r
X for each ω ∈ Σ. (1.4.30)
j=n

Together, (1.4.28) and (1.4.30) imply that for μ-a.e. ω ∈ Σ we have that {Fn (ω)}n∈N
is a Cauchy sequence in (X, · X ). Given that the latter space is complete, it is then
meaningful to define

∞
F(ω) := lim Fn (ω) = f j (ω) in (X, · ), for μ-a.e. ω ∈ Σ. (1.4.31)
n→∞
j=1

Since each f j is strongly μ-measurable, Lemma 1.4.12 guarantees that F is strongly

μ-measurable, while the first assertion in (1.4.20) ensure that

∞
F(ω) r
X ≤ f j (ω) r
X = G(ω) for μ-a.e. ω ∈ Σ. (1.4.32)
j=1

Based on these properties and (1.4.27) we deduce that

F ∈ L p (Σ, μ) ⊗ X. (1.4.33)

Relying on the first assertion in (1.4.20), (1.4.24), and (1.4.32) we also see that for
each n ∈ N we have

n r 
n
F(ω) − f j (ω) ≤ F(ω) r
X + f j (ω) r
X
X
j=1 j=1


∞
≤ F(ω) r
X + f j (ω) r
X
j=1

≤ 2G(ω) at μ-a.e. point ω ∈ Σ. (1.4.34)

Upon recalling from (1.4.27) that G belongs to L p/r (Σ, μ), Lebesgue’s Dominated
Convergence Theorem applies and, in concert with (1.4.16), gives that


n r 
n r
lim F − fj = lim F(·) − f j (·) = 0. (1.4.35)
n→∞ L p (Σ, μ)⊗X n→∞ X
j=1 j=1 L p/r (Σ, μ)
46 1 Preliminary Functional Analytic Matters

Ultimately, from (1.4.33) and (1.4.35) we conclude that (1.4.22) holds, and this
finishes the proof in the case when p < ∞. The case when p = ∞ is similar (and
simpler).  
Finally, the claim that L p (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X is in fact a Banach space
whenever X is Banach and p ∈ [1, ∞], is clear from what we have proved so far. 

At long last, we are now in a position to state and prove the following result which
provides a multitude of examples of analytically convex quasi-Banach spaces.

Proposition 1.4.15 Let (Σ, M, μ) be a complete measure space and assume that
(X, · X ) is an analytically convex quasi-Banach space. Then for each p ∈ (0, ∞]
it follows that L p (Σ, μ) ⊗ X (see (1.4.12)) is an analytically convex quasi-Banach
space.
 
Proof From Proposition 1.4.14 we already know that L p (Σ, μ) ⊗ X, · L p (Σ, μ)⊗X
is a quasi-Banach space, so there remains to show that this is analytically convex.
To this end, assume first that 0 < p < ∞. Then, by Theorem 1.4.5, there exists
an equivalent quasi-norm · on X so that · p is plurisubharmonic. It follows
that for any functions f , g ∈ L p (Σ, μ) ⊗ X and each fixed ω ∈ Σ, the function
uω (z) := f (ω) + zg(ω) p is subharmonic. Hence, so is
∫ ∫
U(z) : = uω (z) dμ(ω) = f (ω) + zg(ω) p dμ(ω)
Σ Σ
p
= f + zg L p (Σ, μ)⊗(X, · )
. (1.4.36)

Now, the desired conclusion follows from Theorem 1.4.5. When p = ∞, Theo-
rem 1.4.5 ensures that there exists an equivalent quasi-norm · on X so that ·
is plurisubharmonic, and we simply write
∫ 2π  1 ∫ 2π 
1
sup f (ω) + eiθ g(ω) dθ ≥ sup f (ω) + eiθ g(ω) dθ
2π 0 ω ∈Σ ω ∈Σ 2π 0

≥ sup f (ω) . (1.4.37)

ω ∈Σ

The proof of the lemma is finished. 

We now discuss several criteria for analytic convexity in the context of quasi-
Banach lattices of functions. To set the stage, assume that (Σ, M, μ) is a sigma-finite
measure space and denote by L0 the space of all complex-valued, μ-measurable
functions on Ω. Then a quasi-Banach function space X on (Σ, M, μ), equipped with
a quasi-norm · X so that (X, · X ) is complete, is an order-ideal in the space L0
if it contains a strictly positive function and if f ∈ X and g ∈ L0 with |g| ≤ | f | at
μ-a.e. point implies g ∈ X with g X ≤ f X . Going further, a quasi-Banach lattice
of functions (X, · X ) is called lattice r-convex if
1.4 Complex Interpolation on Quasi-Banach Spaces 47


m  1/r 
m  1/r
| f j |r ≤ fj r
X (1.4.38)
X
j=1 j=1

for any finite family { f j }1≤ j ≤m of functions from X (see, e.g., [107]; cf. also [119,
Vol. II]). This implies that the space
 
[X]r := f measurable : | f | 1/r ∈ X ,
(1.4.39)
normed by f [X]r := | f | 1/r Xr ,

is a Banach function space, called the r-convexification of X (cf. also [119,

Vol. II, pp. 53–54], at least if r > 1).

Remark 1.4.16 Let (Σ, M, μ) be a sigma-finite measure space and assume that X, Y
are two quasi-Banach lattices of functions on (Σ, M, μ), which are r-convex for some
r > 0. It is then clear from the above definition that [X]r = [Y ]r if and only if X = Y .

The theorem below was first proved in [107], [109], though in this particular form
it is stated in [127].

Theorem 1.4.17 Let X be a (complex) quasi-Banach lattice of functions and denote

by ρ(X) its modulus of concavity. Then the following assertions are equivalent:
(i) X is analytically convex;
(ii) X is lattice r-convex for some r > 0;
−1
(iii) X is lattice r-convex for each 0 < r < 1 + log2 ρ(X) .

Proof This follows directly from [109, Theorem 4.4] and [107, Theorem 2.2] pro-
−1
vided X satisfies an upper p-estimate with p := 1 + log2 ρ(X) . That is, for some
equivalent quasi-norm · and some constant C > 0,

n
|x1 | ∨ · · · ∨ |xn | p
≤C xj p
(1.4.40)
j=1

for any finite collection x1, . . . , xn ∈ X. However, this is a simple consequence of

the fact that in our case |x1 | ∨ · · · ∨ |xn | ≤ |x1 | + · · · + |xn | and the Aoki-Rolewicz
theorem (recalled earlier). 

Returning to the task of discussing the complex method of interpolation for a

compatible couple of quasi-Banach spaces X0, X1 , recall the class of admissible
functions F from Definition 1.4.4, and now make the additional assumption that
X0 + X1 is analytically convex. This entails
 
sup f (z) X0 +X1 : 0 < Re z < 1 ≤ C f F, (1.4.41)

for each f ∈ F . With this in hand, all the aforementioned deficiencies of the complex
method in the context of quasi-Banach spaces (such as the continuity of evaluation
functions and the completeness of space F ) are easily corrected. In this setting, we
48 1 Preliminary Functional Analytic Matters

define the (outer) complex interpolation spaces Xθ := [X0, X1 ]θ for θ ∈ (0, 1)

as

Xθ := [X0, X1 ]θ
 
:= x ∈ X0 + X1 : there exists f ∈ F such that x = f (θ) (1.4.42)

and introduce

x [X0,X1 ] θ := inf { f F : f ∈ F so that f (θ) = x}, ∀x ∈ Xθ . (1.4.43)

It then follows that Xθ is a quasi-Banach space for 0 < θ < 1.

Let us note at this point that there is an alternative choice for the class of admissible
functions. Specifically, define F0 to be a subspace of F consisting of the closure of
the space of functions f ∈ F with the property that f (w) ∈ X0 ∩ X1 for all w ∈ U.
We will use this class to define the inner complex interpolation spaces,
Xθi := [X0, X1 ]iθ for θ ∈ (0, 1), by asking that x ∈ Xθi if and only if there exists f ∈ F0
such that x = f (θ), and set

x [X0,X1 ]iθ := inf { f F : f ∈ F0 so that f (θ) = x}, ∀x ∈ Xθi . (1.4.44)

The inner spaces have the advantage that it is an immediate consequence of the
definition that X0 ∩ X1 is dense in each Xθi . Indeed, given any x ∈ Xθi , we have
x = f (θ) for some f ∈ F0 . The latter membership implies the existence of a
sequence { f j } j ∈N ⊆ F with f j (U) ⊆ X0 ∩ X1 for each j ∈ N with f j − f F → 0 as
j → ∞. Since, obviously, for each j ∈ N both f j and f j − f belong to the subspace
F0 , it follows that the vector x j := f j (θ) ∈ X0 ∩ X1 satisfies the norm estimate
x j − x0 [X0,X1 ]i = ( f j − f )(θ) [X0,X1 ]i ≤ f j − f F for each j ∈ N. The latter,
θ θ
when combined with lim j→∞ f j − f F = 0, allows us to conclude that X0 ∩ X1 is
dense in Xθi .
If X0 and X1 are Banach spaces the inner and outer complex methods yield exactly
the same spaces (isometrically). But the argument for this depends essentially on the
fact that X0 and X1 are Banach spaces (see [110] for more on this topic). Thus, it is far
from clear whether the inner and outer complex methods will always yield the same
result in our setting. However, in special cases the inner and outer methods do yield
the same result, as we will see below. Let us here note that complex interpolation
of quasi-Banach spaces contained in an analytically convex ambient space (not
necessarily X0 + X1 ) was first studied in [16].
To state the next result, recall that, given two quasi-Banach lattices of functions
(X j , · X j ), j ∈ {0, 1}, the Calderón product X01−θ X1θ , with 0 < θ < 1, is
 
X01−θ X1θ := h ∈ L0 : there exist f ∈ X0, g ∈ X1 such that |h| ≤ | f | 1−θ |g| θ ,
 
f X 1−θ X θ := inf f0 X1−θ f θ : | f | ≤ | f | 1−θ | f | θ , f ∈ X , j ∈ {0, 1} .
0
1 X1
0 1 j j
0 1
(1.4.45)
1.4 Complex Interpolation on Quasi-Banach Spaces 49

A simple yet important feature for us here is that the Calderón product “commutes"
with the process of convexification. More concretely, if X0 , X1 are as above and, in
addition, X0 , X1 are also lattice r-convex for some r > 0, it is straightforward to
check that
[X01−θ X1θ ]r = ([X0 ]r )1−θ ([X1 ]r )θ , ∀θ ∈ (0, 1), (1.4.46)
in the sense of equivalence of quasi-norms.
It has been pointed out in [111] that the complex method described above gives
the result predicted by the Calderón formula for nice pairs of function spaces. Let us
record a specific result, building on earlier work in [69] and which has been proved in
[111] for what we now call the outer method. To state it, recall that a Polish space is
a topological space that is homeomorphic to some complete separable metric space.

Theorem 1.4.18 Let Ω be a Polish space and let μ be a sigma-finite Borel measure
on Ω. Let X0, X1 be a pair of quasi-Banach function spaces on (Ω, μ). Suppose that
both X0 and X1 are analytically convex and separable. Then X0 + X1 is analytically
convex and, for each θ ∈ (0, 1),

[X0, X1 ]θ = [X0, X1 ]iθ = X01−θ X1θ (1.4.47)

in the sense of equivalence of quasi-norms.

Remark 1.4.19 As pointed out in [111], the hypothesis of separability in this case
is equivalent to sigma-order continuity. For a general quasi-normed space X, this
property asserts that a non-negative, non-increasing sequence of functions in X
which converges a.e. to zero also converges to zero in the quasi-norm topology of X
(cf., e.g., [119, Vol. II]). An equivalent reformulation is that if g ∈ X and | fn | ≤ |g|
for each n ∈ N and fn → f a.e. as n → ∞ then fn − f X → 0 as n → ∞. For us,
it is of interest to also note a result, proved in [39, Theorem 1.29], to the effect that

one of the lattices X0, X1 X 1−θ X1θ is sigma-order continuous

=⇒ 0 (1.4.48)
is sigma-order continuous for each index θ belonging to (0, 1).

Let us briefly indicate why the inner and outer methods agree here. It fact the
argument in [111] shows that for any given function f ∈ [X0, X1 ]θ a nearly optimal
choice for F ∈ F with F(θ) = f is of the form F(z) = u| f0 | 1−z | f1 | z where f0 ∈ X0 ,
f1 ∈ X1 and |u| = 1 a.e. But we can select a sequence of Borel sets En  Ω with
the property that 1En f0, 1En f1 ∈ X0 ∩ X1 , and consider Fn := 1En F. Thus Fn ∈ F0
and using order-continuity one sees that Fn − F F → 0 as n → ∞. Thus F ∈ F0 .

Remark 1.4.20 For sequence spaces, Theorem 1.4.18 continues to hold in the case
when just one of the two quasi-Banach lattices X0 , X1 is separable. Indeed, in
[111], the separability hypotheses on X j , j ∈ {0, 1}, was used to ensure that if
f0 ∈ X0 and f1 ∈ X1 then the function z → | f0 | 1−z | f1 | z is admissible (i.e. belongs
to F ). In fact, the one property which is not immediate is its continuity on the
closure of the strip 0 < Re z < 1. Nonetheless, this issue can be handled as follows.
If we now set g := | f0 | 1−θ | f1 | θ ∈ X01−θ X1θ for f j ∈ X j and 0 < θ < 1, then
50 1 Preliminary Functional Analytic Matters

g1E = | f0 1E | 1−θ | f1 1E | θ for any E ⊂ Ω. In particular, if E is finite then, clearly,

z → | f0 1E | 1−z | f1 1E | z is admissible. Thus, as in [111], we have that g1E ∈ [X0, X1 ]θ
and g1E [X0,X1 ]θ ≤ C g1E X 1−θ X θ . Consider now En  Ω, a nested family of
0 1
finite sets exhausting Ω (which can be arranged if the X j ’s are sequence spaces).
Replacing E by E j \ Ek and using the fact that, by (1.4.48), g1En → g in X01−θ X1θ as
n → ∞, ultimately gives that {g1En }n is Cauchy in [X0, X1 ]θ . The same argument
further yields that {g1En }n converges to g in X0 + X1 . Hence, g ∈ [X0, X1 ]θ and
g [X0,X1 ]θ ≤ C g X 1−θ X θ . From this point on, one proceeds as in the proof of [111,
0 1
Theorem 3.4].

The theorem below originates in [111]; see also [110] where other related results
can be found.

Theorem 1.4.21 Let X0, X1 and Y0, Y1 be two compatible couples of quasi-Banach
spaces and assume that X0 + X1 and Y0 + Y1 are analytically convex. Also, consider
a bounded, linear operator T : X j → Yj , j ∈ {0, 1}. Then the following results are
true.

(a) In the case when Xθ := [X0, X1 ]θ and Yθ := [Y0, Y1 ]θ , or if Xθ := [X0, X1 ]iθ and
Yθ := [Y0, Y1 ]iθ , it follows that T induces a bounded linear operator

Tθ : Xθ −→ Yθ for each θ ∈ (0, 1), (1.4.49)

in a natural fashion. Moreover,

θ
Tθ Xθ →Yθ ≤ T 1−θ
X0 →X0 T X1 →X1 for each θ ∈ (0, 1). (1.4.50)

(b) Assume that there exists θ o ∈ (0, 1) such that Tθo is an isomorphism. Then there
exists ε > 0 such that Tθ continues to be isomorphism whenever |θ − θ o | < ε.
(c) If I is any open sub-interval of (0, 1) with the property that the inverse Tθ−1 exists
for each θ ∈ I, then Tθ−1 agrees with Tθ−1 
 on Yθ ∩ Yθ  for every θ, θ ∈ I.

We continue by presenting a general interpolation result with constraints.

Theorem 1.4.22 Let Xi, · Xi , Yi, · Yi , Zi, · Zi , where i ∈ {0, 1}, be quasi-
Banach spaces such that X0 ∩ X1 is dense in both X0 and X1 , and Z0 ∩ Z1 is dense
in both Z0 and Z1 . Suppose that Yi → Zi continuously for i ∈ {0, 1} and that there
exists a linear operator D such that D : Xi → Zi boundedly for i ∈ {0, 1}. Define
the spaces  
Xi (D) := u ∈ Xi : Du ∈ Yi , i ∈ {0, 1}, (1.4.51)
equipped with the graph norm, i.e. u Xi (D) := u Xi + Du Yi for each u ∈ Xi (D).
Finally, suppose that there exist two continuous linear mappings, G : Zi → Xi and
K : Zi → Yi , with the property D ◦ G = I + K on Zi for i ∈ {0, 1}.
Then, whenever 0 < θ < 1 and 0 < q ≤ ∞, one has
 
X0 (D), X1 (D) θ,q = u ∈ (X0, X1 )θ,q : Du ∈ (Y0, Y1 )θ,q . (1.4.52)
1.4 Complex Interpolation on Quasi-Banach Spaces 51

Furthermore, if X0 + X1 and Y0 + Y1 are analytically convex, then X0 (D) + X1 (D)

is also analytically convex and
  
X0 (D), X1 (D) θ = u ∈ [X0, X1 ]θ : Du ∈ [Y0, Y1 ]θ , θ ∈ (0, 1). (1.4.53)

Proof For the real interpolation method, the crux of the matter is establishing the
following estimate involving the K-functionals of the pairs (X0 (D), X1 (D)), (X0, X1 ),
(Y0, Y1 ):

K(t, a; X0 (D), X1 (D)) ≈ K(t, a; X0, X1 ) + K(t, Da; Y0, Y1 ), (1.4.54)

uniformly in t and a. One direction is, of course, trivial. For the other one, given
a ∈ X0 + X1 along with t > 0, let

a = x0 + x1, xi ∈ Xi, and Da = y0 + y1, yi ∈ Yi, i ∈ {0, 1}, (1.4.55)

be nearly optimal splittings so that

x0 X0 + t x1 X1 ≈ K(t, a, X0, X1 ) (1.4.56)

and
y0 Y0 + t y1 Y1 ≈ K(t, Da, Y0, Y1 ). (1.4.57)
We then define a new splitting a = x0 + x1 , where

xi := xi − GDxi + Gyi, i ∈ {0, 1}. (1.4.58)

Then
xi Xi ≤ C( xi Xi + yi Yi ), i ∈ {0, 1}. (1.4.59)
Also
Dxi = −K Dxi + yi + K yi, i ∈ {0, 1}, (1.4.60)
so that
Dxi Yi ≤ C( xi Xi + yi Yi ), i ∈ {0, 1}, (1.4.61)
since K D maps Xi boundedly into Yi for i ∈ {0, 1}, and K maps Yi boundedly to
itself, i ∈ {0, 1}. Then (1.4.59)-(1.4.60) justify the equivalence in (1.4.54).
Formula (1.4.53), regarding complex interpolation, is due to J.-L. Lions and
E. Magenes (cf. [120]) when all spaces involved are Banach. However, their argument
goes through with minor modifications for quasi-Banach spaces given the analytic
convexity assumptions made in this portion of our theorem. The only thing we need
to check is that is that the space X0 (D) + X1 (D) is analytically convex, so that the
complex interpolation method outlined in the first part of this section applies to the
couple X0 (D), X1 (D).
In order to justify this we first note that
 
X0 (D) + X1 (D) = (X0 + X1 )(D) := u ∈ X0 + X1 : Du ∈ Y0 + Y1 , (1.4.62)
52 1 Preliminary Functional Analytic Matters

where the rightmost space is equipped with the natural graph norm. Indeed, (1.4.62)
follows readily from the decompositions (1.4.55), (1.4.58). Thus, it suffices to prove
that (X0 + X1 )(D) is analytically convex. To this end, let f : U → (X0 + X1 )(D)
be an analytic function which extends by continuity to U (where U is the open
unit strip {z : 0 < Re z < 1} in the complex plane). Since the inclusion mapping
ι : (X0 + X1 )(D) → X0 + X1 is linear and bounded, f may also be regarded as a
(X0 + X1 )-valued analytic function in U, extendible by continuity to U. Similarly,
the operator D : (X0 + X1 )(D) → Y0 + Y1 is linear and bounded, thus the function
D f is a (Y0 + Y1 )-valued analytic function in U, which extends by continuity to U.
Consequently, given that X0 + X1 and Y0 + Y1 are analytically convex, we may write

max f (z) (X0 +X1 )(D) ≈ max f (z) X0 +X1 + max D f (z) Y0 +Y1
0<Re z<1 0<Re z<1 0<Re z<1

≤ C max f (z) X0 +X1 + C max D f (z) Y0 +Y1

Re z=0,1 Re z=0,1

≈ C max f (z) (X0 +X1 )(D), (1.4.63)

Re z=0,1

as desired. 
We conclude this section with a simple, yet useful result, which is essentially
folklore. First, we make a definition. Let X0, X1 and Y0, Y1 be two compatible pairs of
quasi-Banach spaces. Call {Y0, Y1 } a retract of {X0, X1 } if there exist two bounded,
linear operators E : Yi → Xi , R : Xi → Yi , i ∈ {0, 1}, such that R ◦ E = I, the
identity map, on each Yi , i ∈ {0, 1}.

Lemma 1.4.23 Assume that X0, X1 and Y0, Y1 are two compatible pairs of quasi-
Banach spaces such that {Y0, Y1 } is a retract of {X0, X1 } (as before, the “extension-
restriction” operators are denoted by E and R, respectively). Then for each θ ∈ (0, 1)
and 0 < q ≤ ∞,
   
[Y0, Y1 ]θ = R [X0, X1 ]θ and (Y0, Y1 )θ,q = R (X0, X1 )θ,q . (1.4.64)

In the case of the complex method, it is assumed that X0 + X1 is analytically convex.

As a corollary, the following result holds. Assume that (X0, X1 ) is a compatible
pair of quasi-Banach spaces and that P is a common projection (i.e., a linear,
bounded operator on Xi , i ∈ {0, 1}, such that P2 = P). Then the real and complex
interpolation brackets commute with the action of P, i.e.,
   
[PX0, PX1 ]θ = P [X0, X1 ]θ and (PX0, PX1 )θ,q = P (X0, X1 )θ,q , (1.4.65)

for each θ ∈ (0, 1) and 0 < q ≤ ∞. In the case of the complex method, it is assumed
that X0 + X1 is analytically convex.

A couple of comments are in order. First, generally speaking, given two quasi-
normed spaces X, Y and a linear, bounded operator T : X → Y , by T X we shall
denote its image equipped with the quasi-norm
1.4 Complex Interpolation on Quasi-Banach Spaces 53

y TX := inf{ x X : x ∈ X such that y = T x}, y ∈ T X. (1.4.66)

In particular, this is the sense in which (1.4.64) and (1.4.65) should be understood.
Second, the portion of Lemma 1.4.23 referring to real interpolation remains valid
when the spaces in question are quasi-normed Abelian groups (in which case, the
operators involved are assumed to be group morphisms).
Proof of Lemma 1.4.23 The first order of business is to show that Y0 + Y1 is analyti-
cally convex (hence, justifying the use of the complex method of interpolation for the
pair Y0, Y1 ). One way to see this is by observing that E maps Y0 + Y1 isomorphically
onto E(Y0 +Y1 ) which, given that this operator has a left inverse, is a closed subspace
of the analytically convex space X0 + X1 . Hence, Y0 + Y1 is also analytically convex.
The remainder of the proof follows a well-known path. We, nonetheless, include
the details for the convenience of the reader. Fix θ ∈ (0, 1) and set Xθ = [X0, X1 ]θ ,
Yθ = [Y0, Y1 ]θ . By the interpolation property for bounded linear operators, R maps
Xθ to Yθ , i.e., R(Xθ ) ⊆ Yθ . To justify the opposite inclusion, note that E takes Yθ
into Xθ which is further mapped by R into R(Xθ ). Since the composition of these
two applications acts as the identity operator, we may conclude that Yθ ⊆ R(Xθ ), as
desired. The proof in the case of the real interpolation method is virtually the same
and this completes the proof of the first part of the lemma.
Turning to the second part of our lemma, we note that {PX0, PX1 } is a retract of
{X0, X1 } (taking E to be the inclusion and R the given projection). Thus, (1.4.65) is
a corollary of what we have proved so far. 
Given a quasi-normed space X, recall that Bd(X) stands for the space of linear
and bounded operators from X into X, and that Cp(X) denotes the space of compact
operators from X into X. Below we discuss a version of (1.3.78) and (1.4.50), for
the essential norm (cf. (1.2.54)) in place of the ordinary operator norm (cf (1.2.13));
this result should be also compared with Proposition 2.2.7.
Proposition 1.4.24 Let X0, · X0 , X1, · X1 be a compatible couple of quasi-
Banach spaces, and denote by C0, C1 ∈ [1, ∞) the moduli of concavity in X0, X1 (cf.
(1.2.1)). Suppose there exists a sequence of linear finite-rank operators {PN } N ∈N
from X0 + X1 into itself such that

PN Xi ⊆ Xi for each N ∈ N and each i ∈ {0, 1}, (1.4.67)

C := max sup PN Bd(Xi ) < +∞, (1.4.68)

i ∈ {0,1} N ∈N

and such that

for each index i ∈ {0, 1}, each relatively compact subset Oi of Xi,
and each ε > 0 there is some N(ε, Oi ) ∈ N with the property that (1.4.69)
sup f − PN f Xi < ε whenever N ∈ N satisfies N ≥ N(ε, Oi ).
f ∈ Oi

Also, let T be a linear mapping from X0 + X1 into itself such that

54 1 Preliminary Functional Analytic Matters

T(Xi ) ⊆ Xi and T Xi ∈ Bd(Xi ) for i ∈ {0, 1}. (1.4.70)

Then for each θ ∈ (0, 1) and each q ∈ (0, ∞] one has

 · dist T, Cp(X0 ) 1−θ θ

dist T, Cp((X0, X1 )θ,q ) ≤ C · dist T, Cp(X1 ) , (1.4.71)

where all distances are measured in the operator norm (in the corresponding space
of linear bounded operators), and
 
 := max C0 + C 2 C, C1 + C 2 C .
C (1.4.72)
0 1

Moreover, under the additional assumption that the quasi-Banach spaces X0, X1
are analytically convex it follows that for each θ ∈ (0, 1) one also has

 · dist T, Cp(X0 ) 1−θ θ

dist T, Cp([X0, X1 ]θ ) ≤ C · dist T, Cp(X1 ) . (1.4.73)

In terms of the piece of notation introduced in (1.2.54) we may recast (1.4.71) as

ess
· ess 1−θ ess θ
T (X0,X1 ) θ, q →(X0,X1 ) θ, q ≤C T X0 →X0 · T X1 →X1 , (1.4.74)

and recast (1.4.73) as

ess
· ess 1−θ ess θ
T [X0,X1 ] θ →[X0,X1 ] θ ≤C T X0 →X0 · T X1 →X1 . (1.4.75)

We next present the proof of Proposition 1.4.24.

Proof of Proposition 1.4.24 For i ∈ {0, 1}, pick δi > dist T, Cp(Xi ) . Then there
exists Ki ∈ Cp(Xi ) such that

T − Ki Bd(Xi ) < δi . (1.4.76)

Fix f ∈ BXi (0, 1). Since for each N ∈ N we may write

T f − PN T f = (T − Ki ) f + (Ki f − PN Ki f ) − PN (T − Ki ) f , (1.4.77)

the quasi-triangle inequality implies that

T f − PN T f Xi ≤ Ci T − Ki Bd(Xi ) f Xi + Ci2 Ki f − PN Ki f Xi

+ Ci2 PN Bd(Xi ) T − Ki Bd(Xi ) f Xi . (1.4.78)

Pick an arbitrary ε > 0. Since the set

Oi := Ki (BXi (0, 1)) (1.4.79)

is relatively compact in Xi , we may invoke (1.4.69) to conclude that

1.5 Köthe Function Spaces and Other Categories of Topological Vector Spaces 55

there exists N(ε, Oi ) ∈ N with the property that

(1.4.80)
Ki f − PN Ki f Xi < ε whenever N ≥ N(ε, Oi ).

By combining (1.4.78), (1.4.76), (1.4.80), (1.4.68) and keeping in mind the inequality
f Xi < 1, we obtain

T f − PN T f Xi ≤ Ci δi + Ci2 ε + Ci2 Cδi

(1.4.81)
for all f ∈ BXi (0, 1) and all N ≥ N(ε, Oi ).

This, in turn, gives that

T − PN T Bd(Xi ) ≤ (Ci + Ci2 C)δi + Ci2 ε for all N ≥ N(ε, Oi ). (1.4.82)

Pick θ ∈ (0, 1), q ∈ (0, ∞], and abbreviate Xθ,q := (X0, X1 )θ,q . Using real inter-
polation (see Proposition 1.3.7), from (1.4.82) we see that given any ε > 0, for each
integer N ∈ N with the property that N ≥ max{N(ε, O0 ), N(ε, O1 ) we have
1−θ θ
T − PN T Bd(Xθ, q ) ≤ (C0 + C02 C)δ0 + C02 ε (C1 + C12 C)δ1 + C12 ε . (1.4.83)

Since each PN T is a continuous linear operator, from (1.1.15) we conclude that

PN T is a finite-rank operator on Xθ,q , hence a
(1.4.84)
compact operator on Xθ,q , for each N ∈ N.

Granted this, from (1.4.83) we obtain that for each ε > 0

1−θ θ
dist T, Cp(Xθ,q ) ≤ (C0 + C02 C)δ0 + C02 ε (C1 + C12 C)δ1 + C12 ε . (1.4.85)

Sending ε  0 and δi  dist T, Cp(Xi ) for i ∈ {0, 1}, then yields (1.4.71).
Finally, (1.4.73) is proved in a similar fashion, using complex interpolation (cf.
Theorem 1.4.21 ). 

1.5 Köthe Function Spaces and Other Categories of Topological

Vector Spaces

Classically, Köthe function spaces (as defined in [115], [208], as well as in many
other works based on these references) associated with a given background measure
space are spaces of the form
 
L ρ := f measurable : f := ρ(| f |) < +∞ , (1.5.1)

where ρ is a mapping defined on M+ , the collection of all non-negative measurable

functions in the given environment, satisfying:
56 1 Preliminary Functional Analytic Matters

for each f ∈ M+, ρ( f ) ∈ [0, +∞], and ρ( f ) = 0 if and only if f = 0, (1.5.2)

ρ(λ f ) = λ ρ( f ) for each f ∈ M+ and each λ ≥ 0, (1.5.3)
ρ( f + g) ≤ ρ( f ) + ρ(g) for each f , g ∈ M+, (1.5.4)
ρ( f ) ≤ ρ(g) whenever f , g ∈ M+ satisfy f ≤ g a.e. (1.5.5)

The recipe described in (1.5.1)-(1.5.2) mimics in abstract the manner in which

many familiar normed vector spaces are constructed. This being said, the subadditiv-
ity property (1.5.4) precludes one from considering arbitrary quasi-normed spaces.
Following [140], [138], here we adopt a more general point of view which permits
a unified treatment of large classes of topological vector spaces, with particular em-
phasis on the issues of completeness and convergence. Our first result of this nature
is Theorem 1.5.3 below. Before stating it, we make a couple of definitions. First
we recall from [138, pp. 296–297] (see also [140]) a general recipe for constructing
topologies on a given group, and clarify the notion of completeness with respect to
such a topology.
Definition 1.5.1 Given a group (X, +), denote by 0 its neutral element, and by − f the
inverse of an element f ∈ X. In this context, for a given function ψ : X → [0, +∞]
with the property that ψ(0) = 0, define the topology τψ induced by ψ on X by
stipulating that a set O ⊆ X is open in τψ if and only if for each element f ∈ O there
exists some number r > 0 with the property that Bψ ( f , r) ⊆ O, where

Bψ ( f , r) := {g ∈ X : ψ( f − g) < r }. (1.5.6)

In such a setting, call a sequence { fn }n∈N ⊆ X Cauchy provided for every ε > 0
there exists Nε ∈ N such that ψ( fn − fm ) < ε whenever n, m ∈ N are such that
n, m ≥ Nε . Also, call (X, τψ ) complete if any Cauchy sequence in X is convergent
in τψ to some element in X.
Our second definition introduces a drastically weakened notion of measure.
Definition 1.5.2 Given a measurable space (Σ, M), call μ : M → [0, +∞] a feeble
measure provided the collection of its null-sets, i.e., Nμ := { A ∈ M : μ(A) = 0}
contains , is closed under countable union, and satisfies A ∈ Nμ whenever A ∈ M
and there exists B ∈ Nμ such that A ⊆ B.
Let (Σ, M) be a measurable space and let μ be a feeble measure on M. As in the case
of genuine measures, we shall say that a property is valid μ-a.e. provided the property
in question is valid with the possible exception of a set in Nμ . Identifying functions
coinciding μ-a.e. on Σ then becomes an equivalence relation, and we shall denote by
M (Σ, M, μ) the collection of all equivalence classes of scalar-valued, M-measurable
functions on Σ. Lastly, we define
 
M+ (Σ, M, μ) := f ∈ M (Σ, M, μ) : f ≥ 0 at μ-a.e. point on Σ . (1.5.7)

Here is the abstract completeness criterion alluded to earlier. More general results
of this nature are established in [140], [138].
1.5 Köthe Function Spaces and Other Categories of Topological Vector Spaces 57

Theorem 1.5.3 Assume that (Σ, M) is a measurable space and that μ is a feeble
measure on M. Suppose that the function

· : M+ (Σ, M, μ) −→ [0, +∞], (1.5.8)

satisfies the following properties:

(1) [Quasi-subadditivity] There exists a constant C0 ∈ [1, +∞) with the property
that

f + g ≤ C0 max{ f , g }, ∀ f , g ∈ M+ (Σ, M, μ); (1.5.9)

(2) [Pseudo-homogeneity] There exists a function ϕ : (0, +∞) → (0, +∞) satisfying

λ f ≤ ϕ(λ) f , ∀ f ∈ M+ (Σ, M, μ), ∀λ ∈ (0, +∞), (1.5.10)

and such that


sup ϕ(λ)ϕ(λ−1 ) < +∞ and lim+ ϕ(λ) = 0; (1.5.11)
λ>0 λ→0

(3) [Non-degeneracy] There holds

f = 0 ⇐⇒ f = 0, ∀ f ∈ M+ (Σ, M, μ); (1.5.12)

(4) [Quasi-monotonicity] There exists some constant C1 ∈ [1, +∞) such that for any
f , g ∈ M+ (Σ, M, μ) satisfying f ≤ g μ-a.e. on Σ there holds f ≤ C1 g ;
(5) [Weak Fatou property] If { fi }i ∈N ⊆ M+ (Σ, M, μ) is a sequence of functions
satisfying fi ≤ fi+1 μ-a.e. on Σ for each i ∈ N and such that sup fi < +∞,
i ∈N
then sup fi < +∞.
i ∈N

Finally, define
 
L := f ∈ M(Σ, M, μ) : f L := | f | < +∞ . (1.5.13)

Then functions in L are finite μ-a.e. on Σ and, with the topology τ · L considered
in the sense of Definition 1.5.1 (relative to the additive group structure on L),

L, τ · L
is a Hausdorff, complete, metrizable, topological vector space.
(1.5.14)

Formula (1.5.13) may be regarded as a general recipe according to which large

classes of function spaces are defined in practice. Of course, any genuine quasi-
norm · on the vector space L 0 (Σ, M, μ), consisting of (classes of) functions from
M(Σ, M) which are finite μ-a.e., satisfies the axioms (1)-(3). Any function of the
form ϕ(λ) := λ θ , with θ ∈ (0, ∞) fixed, satisfies
∫ (1.5.11). Such an example arises
naturally if, e.g., μ is a measure and f := Σ f θ dμ for each f ∈ M+ (Σ, M, μ)
(note that · satisfies all hypotheses of Theorem 1.5.3).
58 1 Preliminary Functional Analytic Matters

We also wish to note that the weak Fatou property mimics (in abstract, and with a
weaker conclusion) the familiar Fatou’s Lemma in the standard setting of Lebesgue
spaces. Indeed, in [140], [138] it has been proved that, given a feeble measure μ on
a measurable space (Σ, M) and a function · : M+ (Σ, M, μ) → [0, +∞] which is
quasi-subadditive and quasi-monotone, the weak Fatou property stated in item (5)
of Theorem 1.5.3 is equivalent to the demand that

∀{ fi }i ∈N ⊆ M+ (Σ, M, μ) such that lim inf i→∞ fi < +∞

(1.5.15)
we have lim inf i→∞ fi < +∞

and, in fact, a quantitative version of this implication holds.

In addition to completeness, we are also interested in abstract results pertaining to
the pointwise behavior of sequences of functions which are convergent in topolog-
ical vector spaces created according to formula (1.5.13). Specifically, the following
theorem appears in [140], [138].

Theorem 1.5.4 Retain the same hypotheses as in Theorem 1.5.3 and recall the vector
space L from (1.5.13). Then any sequence { f j } j ∈N in L which is convergent to some
f ∈ L in the topology τ · L has a sub-sequence which converges to f pointwise
μ-a.e. on Σ.
In particular, the positive cone in L equipped with the partial order induced by the
pointwise μ-a.e. inequality of functions, i.e., L + := { f ∈ L : f ≥ 0 μ-a.e. on Σ},
is closed in L, τ · L .

The work in [140], [138] also addresses the issue of having a continuous embed-
ding of the vector space L from (1.5.13) equipped with the topology τ · L into the
space of measurable, a.e. finite functions equipped with the topology of convergence
in measure.

Theorem 1.5.5 Suppose that (Σ, M, μ) is a measure space and consider a mapping
· : M+ (Σ, M, μ) → [0, +∞] satisfying properties (1)-(4) from Theorem 1.5.3. In
addition, assume that
$
∞
there exists some sequence {K j } j ∈N ⊆ M satisfying Kj = Σ
j=1 (1.5.16)
as well as K j ⊆ K j+1, μ(K j ) < +∞, 1K j < +∞ for each j ∈ N.

In this context, let L 0 (Σ, M, μ) := { f ∈ M(Σ, M, μ) : | f | < +∞ μ-a.e. on Σ} and

denote by τμ the topology on this space induced by convergence in measure on sets
of finite measure. Then, if L is as in (1.5.13), it follows that

L, τ · L
→ L 0 (Σ, M, μ), τμ continuously. (1.5.17)

Regarding the separability of the topological vector space from (1.5.14), we

recall the following result which appears in [138, Theorem 5.5, p. 300] (the reader
is referred to [133, Definition 3.6.1] for the notion of separable measure).
1.5 Köthe Function Spaces and Other Categories of Topological Vector Spaces 59

Theorem 1.5.6 Retain the hypotheses of Theorem 1.5.5 and, in addition, assume
that · is absolutely continuous, in the sense that for any given f ∈ L there
holds
&
∀(A j ) j ∈N ⊆ M such that 1 A j → 0
=⇒ lim | f | · 1 A j = 0. (1.5.18)
pointwise μ-a.e. on Σ as j → ∞ j→∞

Then L, τ · L
is a separable topological space whenever the measure μ is separa-
ble.
In the case when L, · L is a Banach function space, it follows from [14,
Theorem 5.5, p. 27] that both the absolute continuity property stated in (1.5.18) and
the separability of the measure μ are actually necessary conditions for the separability
of the topological space L, τ · L .
The remaining portion of this section contains material which is going to be
useful in justifying the boundedness result presented later, in Theorem 4.4.7, for
linear operators taking values in certain categories of topological vector spaces8. We
begin by reviewing a number of definitions and preliminary results which will play
a significant role in the formulation and proof of Theorem 4.4.7.
Definition 1.5.7 Suppose X is a vector space over C.
[1] Call a function · : X → [0, ∞) a θ-pseudo-quasi-norm (or simply pseudo-
quasi-norm) on X provided the following three conditions hold:
(i) (nondegeneracy) for each x ∈ X one has x = 0 if and only if x = 0;
(ii) (quasi-subadditivity) there exists a constant C0 ∈ [1, ∞) for which

x + y ≤ C0 max{ x , y }, ∀x, y ∈ X. (1.5.19)

(iii) (pseudo-homogeneity) there exist C1 ∈ (0, ∞) and θ ∈ R such that

λ x ≤ C1 |λ| θ x , ∀x ∈ X, ∀λ ∈ C \ {0}. (1.5.20)

[2] The pair (X, · ) (which shall be referred to as a pseudo-quasi-normed space)

is said to be a pseudo-quasi-Banach space provided (X, τ · ) is complete in
the sense of Definition 1.5.1, where τ · is the topology induced by · on X.
[3] Given s ∈ (0, ∞), call a pseudo-quasi-norm · on X s-norm provided it is
genuinely homogeneous and satisfies

x+y s
≤ x s
+ y s
for all x, y ∈ X. (1.5.21)

There are many classes of topological vector spaces which are of a basic im-
portance in analysis that are not Banach but merely quasi-Banach. Indeed, take for
8 Recall that we understand by a topological vector space, a pair (X , τ), where X is a vector
space over C and τ is a topology on X such that the vector space operations of addition and
scalar multiplication are continuous with respect to τ. We stress that under these assumptions, the
topological space (X , τ) may not be Hausdorff.
60 1 Preliminary Functional Analytic Matters

example the following familiar scales of spaces: sequence spaces, Lebesgue spaces,
weak-Lebesgue spaces, Lorentz spaces, Hardy spaces, weak-Hardy spaces, Besov
spaces, Triebel-Lizorkin spaces, as well as their weighted versions (just to name a
few). The class of pseudo-quasi-Banach spaces, given as in Definition 1.5.7, further
generalizes the notion of a quasi-Banach space (hence, the notion of genuine Banach
space) by allowing for the relaxation of the homogeneity condition in the manner
described in (1.5.20).
The following metrization result in the class of pseudo-quasi-norms, borrowed
from [138, Theorem 3.39, p. 130], may be regarded as a generalization of the Aoki-
Rolewicz Theorem (see [11], [162], [112]).

Theorem 1.5.8 Let X be a vector space over C and assume that · : X → [0, ∞)
is a function satisfying the following properties:
(1) there exists a constant C0 ∈ [1, ∞) for which

x + y ≤ C0 max{ x , y }, ∀x, y ∈ X; (1.5.22)

(2) there exist C1 ∈ (0, ∞) and θ ∈ R such that

λ x ≤ C1 |λ| θ x , ∀x ∈ X, ∀λ ∈ C \ {0}; (1.5.23)

Set  −1
α := log2 C0 ∈ (0, ∞], (1.5.24)
and, for each x ∈ X, define

N  α1
x := sup inf |λ| −θ λ xi α
: N ∈ N, and x1, . . . , x N ∈ X are
λ∈C\{0} i=1


N !
such that xi = x ,
i=1
(1.5.25)

if α < ∞ and, corresponding to the case when α = ∞,

x := sup inf |λ| −θ max λ xi : N ∈ N, and x1, . . . , x N ∈ X are

λ∈C\{0} 1≤i ≤ N


N !
such that xi = x , (1.5.26)
i=1

Then · : X → [0, ∞) satisfies:

1.5 Köthe Function Spaces and Other Categories of Topological Vector Spaces 61

C0−2 x ≤ x ≤ C1 x for all x ∈ X, (1.5.27)

λx = |λ| θ x for all x ∈ X and all λ ∈ C \ {0}, (1.5.28)

β β β
x+y ≤ x + y for all x, y ∈ X and each finite β ∈ (0, α], (1.5.29)
x+y ≤ C0 max{ x , y } for all x, y ∈ X. (1.5.30)

Moreover, if in addition to (1.5.22)-(1.5.23), the function · has the property

that for every x ∈ X one has x = 0 ⇔ x = 0 (i.e., if · is actually a pseudo-
quasi-norm on X), then the function
β
d : X × X → [0, ∞) given by d(x, y) := x − y for all x, y ∈ X (1.5.31)

is a genuine distance on X such that τd = τ ·  = τ · . In particular, the function

· is continuous on (X, τ · ). Hence, the balls with respect to · (see (1.5.6))
are open in τ · .

We continue by making a few remarks of general nature. For k ∈ {1, 2}, suppose
(Xk , τk ) is a topological vector space. Recall that a linear operator T : X1 → X2
is said to be bounded provided T maps topologically bounded subsets of X1 into
topologically bounded subsets of X2 . In particular, if for k ∈ {1, 2}, the function
· k : Xk → [0, ∞) is a θ k -pseudo-quasi-norm on Xk (for some θ k ∈ (0, ∞))
such that τ · k = τk , then by the homogeneity conditions for · k as well as the
coincidence between notions of topologically and geometrically bounded sets one
has
the linear operator T : X1 → X2 is bounded if and only if there exists
θ2 /θ1 (1.5.32)
C ∈ (0, ∞) such that T f X2
≤C f X1
for every f ∈ X1 .

Another property of general pseudo-quasi-normed spaces which will be of rele-

vance shortly is as follows:

if (X, · ) is a pseudo-quasi-normed vector space then there exists C ∈ [1, ∞)

such that if lim x j = x∗ in X, in the topology induced on X by · , then

j→∞

C −1 x∗ ≤ lim inf x j ≤ lim sup x j ≤ C x∗ .

j→∞ j→∞
(1.5.33)
The justification of (1.5.33) makes use of the continuity of the function · , given
as in Theorem 1.5.8, as well as (1.5.27).
We conclude with the following definition which is relevant later, in the context
of Theorem 4.4.7.

Definition 1.5.9 Two given topological vector spaces, (Xk , τk ) with k ∈ {1, 2} are
said to be weakly compatible provided there exists a topological vector space
(X , τX ) satisfying the following two conditions:
62 1 Preliminary Functional Analytic Matters

(i) every convergent sequence of points in X has a unique limit;

(ii) for each k ∈ {1, 2} there exists an injective linear mapping ιk : Xk → X
satisfying

for all {x j } j ∈N ⊆ Xk and x ∈ Xk ⎫

⎪
⎪
⎬
=⇒ lim ιk (x j ) = ιk (x) in (X , τX ).
such that lim x j = x in (Xk , τk ) ⎪
⎪ j→∞
j→∞ ⎭
(1.5.34)

Since not all topological vector spaces considered in this work are necessarily
Hausdorff, the additional demand on (X , τX ) in part (i) of Definition 1.5.9 is not
redundant. Also, the mapping ιk : (Xk , τk ) → (X , τX ) in part (ii) may not be
continuous given the minimal assumptions on the topological spaces (Xk , τk ). This
being said, if ιk is continuous then it necessarily satisfies (1.5.34). Finally, in light
of the injectivity of the mapping ιk in part (ii), there is no ambiguity in identifying
x ≡ ιk (x) ∈ X whenever x ∈ Xk .
Chapter 2
Abstract Fredholm Theory

With an eye on eventually employing Fredholm theory in the treatment of boundary

value problems for elliptic systems in rough domains (later on, in [137]), here we
review basic material on the topic of Fredholm theory, from a purely functional
analytic point of view. First, in §2.1, we recall useful results of this nature in the
classical setting of Banach spaces. This is where Fredholm theory is at its zenith
of beauty but, elegant as this may be, Fredholm theory is vastly restricted to such
a setting everywhere in the literature. This is because virtually all treatments make
essential use of the fact that Banach spaces have a rich duality theory, and rely on
the availability of the Hahn-Banach Theorem together with its usual plethora of very
useful consequences.
Since a variety of function spaces naturally employed in the formulation of
the boundary value problems we have in mind, such as certain Hardy, Besov, and
Triebel-Lizorkin spaces, fail to be Banach (as their “norms” only satisfy a quasi-
triangle inequality), we also require a refinement of the “standard” Fredholm theory
on Banach spaces which is applicable to more general topological vector spaces
(which are not necessarily locally convex). In §2.2 we recall some basic results to
this effect, following work in [142] (where a “dual-less” approach to Fredholm theory
has been developed) which shows that the core principle of the theory, namely that
a Fredholm operator is a linear homeomorphism modulo finite-dimensional spaces
is pervasive. Along the way, a number of new useful results are deduced.

2.1 Fredholm Theory in Banach Spaces

To get started, recall the following basic definition.

Definition 2.1.1 Let X, Y be two Banach spaces. Define the space of Fredholm
operators from X to Y as

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 63

D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_2
64 2 Abstract Fredholm Theory

Φ(X → Y ) := T ∈ Bd(X → Y ) : T has a finite-dimensional cokernel

and a finite-dimensional kernel . (2.1.1)

Also, consider the index function

ind : Φ(X → Y ) −→ Z defined by

(2.1.2)
ind T := dim (Ker T) − codim (Im T) for each T ∈ Φ(X → Y ).

Next, set

Φ+ (X → Y ) := T ∈ Bd(X → Y ) : T has closed range

and a finite-dimensional kernel , (2.1.3)

and
 
Φ− (X → Y ) := T ∈ Bd(X → Y ) : T has a finite-dimensional cokernel . (2.1.4)

The set of semi-Fredholm operators from X to Y is then defined as the union

Φ− (X → Y ) ∪ Φ+ (X → Y ). The index function (2.1.2) may be extended to the set of
all semi-Fredholm operators by setting

index : Φ− (X → Y ) ∪ Φ+ (X → Y ) −→ Z ∪ {±∞}
defined as index T := dim (Ker T) − dim (coker T) (2.1.5)
for each operator T ∈ Φ− (X → Y ) ∪ Φ+ (X → Y ).

Occasionally, if we desire to emphasize the spaces on which the operator T is defined,

we shall write index (T : X → Y ) in place of the less informative symbol ind T.
The following theorem summarizes various standard properties of Fredholm and
semi-Fredholm operators in Banach spaces which we will find useful later on. In this
functional analytic setting, the literature is enormous; see, e.g., [17, Chs. 1, 3], [65,
Chs. XI, XVII], [66, Vol. I, Sects. IV.6, IV.10], [128, Sect. I.3], [151, Ch. 2], [153,
Ch. III], [171, Chs. 5, 7].

Theorem 2.1.2 The following assertions hold for any Banach spaces X, Y, Z:
(1) Any semi-Fredholm operator T ∈ Φ− (X → Y ) ∪ Φ+ (X → Y ), hence any
Fredholm operator T ∈ Φ(X → Y ), has closed range.
(2) If T ∈ Φ± (X → Y ) and S ∈ Φ± (Y → Z) then ST ∈ Φ± (X → Z) and

index (ST) = index (S) + index (T). (2.1.6)

In particular, for any T ∈ Φ(X → Y ) and S ∈ Φ(Y → Z) it follows that

ST ∈ Φ(X → Z) and that (2.1.6) holds.
(3) If T ∈ Φ± (X → Y ) then T ∗ ∈ Φ∓ (Y ∗ → X ∗ ) and index (T) = −index (T ∗ ). In
particular, if T ∈ Φ(X → Y ) then T ∗ ∈ Φ(Y ∗ → X ∗ ) and their indexes satisfy
index (T) = −index (T ∗ ).
2.1 Fredholm Theory in Banach Spaces 65

(4) Given T ∈ Bd(X → Y ), one has T ∈ Φ+ (X → Y ) if and only if T is bounded

from below modulo compact operators, i.e., there exist a Banach space Z, a
compact operator K : X → Z, and a constant C ∈ (0, ∞) such that

 xX ≤ CT xY + K x Z for any x ∈ X. (2.1.7)

In particular, Φ+ (X → Y ) is both open in Bd(X → Y ) and stable under addition

of compact operators. Moreover, the index is unaffected by addition of compact
operators. Another consequence of (2.1.7) is the fact that

if X0 is a closed subspace of X and T ∈ Φ+ (X → X) with

(2.1.8)
T X0 ⊆ X0 , then also T |X0 ∈ Φ+ (X0 → X0 ).

(5) The set Φ− (X → Y ) is open in Bd(X → Y ) and Φ− (X → Y ) is stable under

addition of compact operators. Furthermore, the index is unaffected by addition
of compact operators.
(6) Under addition of compact operators, the set Φ(X → Y ) is stable and the index
is unaffected. Also,

Φ(X → Y ) = Φ− (X → Y ) ∩ Φ+ (X → Y ). (2.1.9)

(7) Given T ∈ Bd(X → Y ), one has T ∈ Φ(X → Y ) if and only if there exist
S1, S2 ∈ Bd(Y → X) and K1 ∈ Cp(Y → Y ) and K2 ∈ Cp(X → X) such that

T S1 = IY + K1 and S2T = IX + K2 (2.1.10)

where IX , IY are the identity operators on X, Y . In fact, one may actually take
S1 = S2 ∈ Φ(X → Y ), i.e., T is Fredholm if and only if it is invertible modulo
compact operators.
(8) The index function (2.1.5) is continuous, hence locally constant.

Several abstract results will play an important role. The first such result is an
index estimate, resembling a monotonicity property of the index on a nested scale
of spaces, is recalled below. For a proof see [139] (cf. also [144, Lemma 11.9.21,
p. 206]).

Lemma 2.1.3 Let X j , Yj , j = 0, 1, be two pairs of Banach spaces such that both
inclusions X0 → X1 and Y0 → Y1 are continuous with dense range. If T is a linear
operator which is Fredholm both from X0 into Y0 and from X1 into Y1 , then

index (T : X0 → Y0 ) ≤ index (T : X1 → Y1 ). (2.1.11)

Furthermore, (2.1.11) holds with equality if and only if

Ker (T : X0 → Y0 ) = Ker (T : X1 → Y1 )
and (2.1.12)
Ker (T ∗ : Y0∗ → X0∗ ) = Ker (T ∗ : Y1∗ → X1∗ ).
66 2 Abstract Fredholm Theory

We shall also need the following abstract regularity result from [139, Lemma 3.10,
pp. 136-137].

Lemma 2.1.4 Let X0 , X1 , be two Banach spaces such that X0 → X1 continuously

and densely. Assume that T is a linear operator which is Fredholm both from X0 into
X0 and from X1 into X1 , with the property that

index (T : X1 → X1 ) = index (T : X0 → X0 ). (2.1.13)

Then
x1 ∈ X1 and T x1 ∈ X0 =⇒ x1 ∈ X0 . (2.1.14)
As a corollary,

X0 ∩ Im (T : X1 → X1 ) = Im (T : X0 → X0 ), (2.1.15)
Ker (T : X1 → X1 ) = Ker (T : X0 → X0 ). (2.1.16)

Given a Banach space X, the annihilator V ⊥ of a subspace V of X is defined as

 
V ⊥ := Λ ∈ X ∗ : Λ(x) = 0 for all x ∈ V , (2.1.17)

and the annihilator ⊥W of a subspace W of X ∗ is defined as

⊥
 
W := x ∈ X : Λ(x) = 0 for all Λ ∈ W . (2.1.18)

For example, one of the standard corollaries of the Hahn-Banach Extension Theorem
gives
V ⊥ = {0} ⇐⇒ V = X, and V ⊥ = X ∗ ⇐⇒ V = {0}. (2.1.19)
It is also well known (cf., e.g., [165, Theorem 4.7, p. 96]) that, in the above context,
we have
⊥ ⊥
(V ) is the norm-closure of V in X, (2.1.20)
and
(⊥W)⊥ is the weak∗ -closure of W in X ∗ . (2.1.21)
Also, since X∗  Λ → Λ|V ∈ V∗ is a linear mapping which is surjective (thanks
to the Hahn-Banach Extension Theorem; cf. [165, Theorem 3.3, p. 58]) and whose
kernel is precisely V ⊥ , the First Group Isomorphism Theorem ensures that we may
identify
X∗
= V ∗. (2.1.22)
V⊥
Hence,  
dim X ∗ /V ⊥ = dim V ∗, (2.1.23)
and, in particular,
 
dim X ∗ /V ⊥ = dim V if V is finite dimensional. (2.1.24)
2.1 Fredholm Theory in Banach Spaces 67

A related (albeit weaker) result holds for the brand of annihilator introduced in
(2.1.18). Specifically, since X  x → Λx ∈ W ∗ where Λx () := (x) for each  ∈ W
is a linear mapping whose kernel is precisely ⊥W, the First Group Isomorphism
Theorem gives  
dim X/⊥W ≤ dim W ∗ . (2.1.25)
In relation to this we claim that, in fact,
 
dim X/⊥W = dim W if W is finite dimensional. (2.1.26)

To prove (2.1.26), suppose W is finite dimensional and let Λ1, . . . , Λn ∈ X ∗

be a basis for the space W. Then Lemma 1.2.16 ensures that there exist some
linearly independent vectors x1, . . . , xn ∈ X such that Λ j (xk ) = δ jk for each
j, k ∈ {1, . . . , n}. If x1, . . . , xn  denotes the linear span of the vectors x1, . . . , xn ,
then dimx1, . . . , xn  = n and we have
⊥
W ∩ x1, . . . , xn  = {0}. (2.1.27)

Indeed, if x0 ∈ ⊥W ∩ x1, . . . , xn  then x0 = a1 x1 +· · ·+an xn for some a1, . . . , an ∈ C

and for each j ∈ {1, . . . , n} we have 0 = Λ j (x0 ) = a1 Λ j (x1 ) + · · · + an Λ j (xn ) = a j
which
 ultimately
 forces x0 = 0. Having proved (2.1.27), we then conclude that
dim X/⊥W ≥ n. Since the opposite inequality is implied by (2.1.25), the claim
made in (2.1.26) follows.
Lemma 2.1.5 Let X be a Banach space, and let V ⊆ X be a closed linear subspace
of X. Denote by π the canonical projection

π : X −→ X
V given by
(2.1.28)
π(x) := x + V, for all x ∈ X,

and denote by ι the canonical inclusion

ι : V ⊥ → X ∗ . (2.1.29)

Then  ∗
Φ : V ⊥ −→ X
V given by
(2.1.30)
Φ(Λ), x + V := Λx, for all Λ ∈ V ⊥ and x ∈ X,
is an isomorphism, whose inverse is given by
 ∗
Φ−1 : VX −→ V ⊥
 ∗ (2.1.31)
Φ−1 () :=  ◦ π, for all  ∈ X
V .

As a consequence, Φ allows the identification

 X ∗
= V ⊥. (2.1.32)
V
68 2 Abstract Fredholm Theory

In turn, under the identification made in (2.1.32), it follows that

π ∗ : (X/V)∗ → X ∗ , the adjoint of π in (2.1.28), is in

(2.1.33)
fact the canonical inclusion operator ι : V ⊥ → X ∗ .
Finally, if X is reflexive then
 ∗
Ψ: −→ V ⊥ given by
X
V
 ∗
X
V  x + V → Ψ(x) := λx ∈ V ⊥ (2.1.34)
where λx (Λ) := Λx for all Λ ∈ V ⊥

is an isomorphism which allows the identification

 ⊥∗ X
V = . (2.1.35)
V
Proof The claims made in relation to (2.1.30)-(2.1.31) are seen from definitions. In
particular, we have (2.1.32). Next, the claim in (2.1.33) comes down to checking that
π ∗ ◦ Φ = ι on V ⊥ , which is then seen from definitions.
There remains to prove that, under the additional hypothesis that X is reflexive,
(2.1.34) is an isomorphism. Having Ψ injective reduces to verifying that if x ∈ X
is such that Λx = 0 for all Λ ∈ V ⊥ then necessarily x ∈ V. This, however, is
a consequence of one of the corollaries of the Hahn-Banach Extension Theorem
(as may be seen reasoning by contradiction, keeping in mind  the fact that V is
∗
closed). To check that Ψ is surjective, fix an arbitrary λ ∈ V ⊥ . Use the Hahn-
Banach Extension Theorem to produce some λ  ∈ (X ∗ )∗ with λ
 ⊥ = λ. Since X is
V
presently assumed to be reflexive, the membership just mentioned implies that there

exists some xo ∈ X with the property that λ(Λ) = Λ(xo ) for each Λ ∈ X ∗ . Then
⊥ ∗
λxo = Ψ(xo ) ∈ V satisfies λxo (Λ) = Λ(xo ) = λ(Λ) for all Λ ∈ V ⊥ , i.e., λxo = λ.
Thus, Ψ(xo ) = λ, proving that Ψ is indeed surjective. 

A useful generalization of Lemma 2.1.5 is presented below.

Lemma 2.1.6 Let X be a Banach space, and let V, W ⊆ X be closed linear subspaces
of X with the property that W ⊆ V. Then one has the identification
 V ∗ W⊥
= . (2.1.36)
W V⊥
The special case W = 0 of (2.1.36) reads

X∗
V∗ = (2.1.37)
V⊥
while the special case V = X of (2.1.36) corresponds precisely to (2.1.32) (after
re-denoting W as V).
Proof of Lemma 2.1.6 Denote by π the canonical projection
2.1 Fredholm Theory in Banach Spaces 69

π : V −→ V
W given by
(2.1.38)
π(x) := x + W, for all x ∈ V .
 ∗
Given any functional f ∈ V
W , it follows that f ◦ π : V → R is linear and
continuous. Pick any f ◦ π ∈ X ∗ such that f ◦ π V = f ◦ π (the Hahn-Banach
Theorem ensures that such an extension always exists). It follows that f ◦ π W ≡ 0
which ultimately shows that f ◦ π ∈ W ⊥ . Consider now the mapping
 ∗ ⊥
Φ: W V
−→ W V ⊥ given by
 ∗ (2.1.39)
Φ( f ) := f ◦ π + V ⊥, for each f ∈ WV
.

Note that this is meaningfully and unambiguously defined (since the difference of
any two extensions of f ◦ π ∈ V ∗ to functionals in X ∗ always belongs to V ⊥ ). In
particular, this makes Φ linear and bounded.
In the opposite direction, define
⊥
 ∗
Ψ:W V ⊥ −→ V
W by setting
(2.1.40)
Ψ(Λ + V ⊥ )(x + W) := Λ(x), for each Λ ∈ W ⊥ and x ∈ V .
 ∗
This is a well defined, linear, and bounded mapping. Then for each f ∈ V
W we
may write
 
Ψ Φ( f ) (x + W) = ( f ◦ π)(x) = ( f ◦ π)(x) = f (x + W) for each x ∈ V, (2.1.41)

hence  V ∗
 
Ψ Φ( f ) = f for each f ∈ . (2.1.42)
W
 ∗
Finally, pick an arbitrary Λ ∈ W ⊥ ⊆ X ∗ and set f := Ψ(Λ + V ⊥ ) ∈ V
W . Then
f ◦ π agrees with Λ on V, since

( f ◦ π)(x) = f (x + W) = Ψ(Λ + V ⊥ )(x + W) = Λ(x) for each x ∈ V . (2.1.43)

As such, Λ ∈ X ∗ is an extension of f ◦ π, hence Φ( f ) = Λ + V ⊥ . Ultimately, this

goes to show that
 
Φ Ψ(Λ + V ⊥ ) = Λ + V ⊥ for each Λ ∈ W ⊥ . (2.1.44)

Thus, Φ and Ψ are inverse to one another, and this finishes the proof of (2.1.36). 

Next, if X, Y are two Banach spaces and T : X → Y is a linear and bounded

operator, then (cf., e.g., [165, Theorem 4.12, p. 99])
70 2 Abstract Fredholm Theory
⊥
Ker (T ∗ : Y ∗ → X ∗ ) = Im (T : X → Y ) , (2.1.45)

Ker (T : X → Y ) = ⊥ Im (T ∗ : Y ∗ → X ∗ ) , (2.1.46)

and (cf., e.g., [165, Theorem 4.14, p. 101])

Im (T : X → Y ) is closed in Y ⇐⇒ Im (T ∗ : Y ∗ → X ∗ ) is weak∗ -closed in X ∗

⇐⇒ Im (T ∗ : Y ∗ → X ∗ ) is normed-closed in X ∗ .
(2.1.47)

As a consequence of (2.1.46) and (2.1.21) we have

⊥
Ker (T : X → Y ) is the weak∗ -closure of Im (T ∗ : Y ∗ → X ∗ ) in X ∗, (2.1.48)

and as a consequence of (2.1.45) and (2.1.20) we have

⊥
Ker (T ∗ : Y ∗ → X ∗ ) is the norm-closure of Im (T : X → Y ) in Y . (2.1.49)

In relation to (2.1.47)-(2.1.49) we also wish to note a couple of related results.

First, it is clear from definitions that
if X is a reflexive Banach space then the weak∗ -topology on X ∗
coincides with the weak topology on X ∗ (regarded as a Banach (2.1.50)
space in its own right).
Second,
given a locally convex Hausdorff topological vector space (X, τ) along
with a linear subspace V of X, the closure of V in (X, τ) coincides with
(2.1.51)
the closure of V in the weak topology on X; in particular, V is τ-closed
if and only if V is closed in the weak topology on X.

To justify this, denote by V the closure of V in (X, τ). In one direction, if xo ∈ X

but xo  V then the Hahn-Banach Theorem ensures the existence of some functional
Λ ∈ X ∗ with Λ(xo ) = 1 and such that Λ|V ≡ 0. Define O := {x ∈ X : |Λ(x)| < 1}.
Then O + xo is a neighborhood of xo in the weak topology on X which is disjoint
from V. Hence, xo does not belong to the closure of V in the weak topology on X.
Ultimately, this shows that the closure of V in the weak topology on X is contained
in V. In the opposite direction, if x ∈ V and O is any open set in the weak topology
on X, then O ∈ τ, hence O ∩ V  . This proves that x belongs to the closure of
V in the weak topology on X. Thus, V is contained in the closure of V in the weak
topology on X.
Third, combining (2.1.50) and (2.1.51) yields the following result:

if X is a reflexive Banach space and V is a linear subspace of X ∗ ,

then the weak∗ -closure of V coincides with the (norm) closure of (2.1.52)
V in the Banach space X ∗ .
2.1 Fredholm Theory in Banach Spaces 71

Hence, as a consequence of (2.1.48) and (2.1.52), we have

if X, Y are two Banach spaces, X being reflexive, and T : X → Y is a
⊥
linear and bounded operator, then Ker (T : X → Y ) is the (norm) (2.1.53)
closure of Im (T ∗ : Y ∗ → X ∗ ) in the Banach space X ∗ .
In particular, from (2.1.53) and (2.1.19) we see that
if X, Y are two Banach spaces, X being reflexive, and T : X → Y is a
linear and bounded operator, then T is injective if and only if T ∗ has (2.1.54)
dense range.
Finally, we wish to note that the reflexivity assumption is essential in the above
context. Indeed, take X :=  1 , Y :=  2 , and take T to be the canonical inclusion
ι :  1 →  2 . Then T is injective yet T ∗ , which becomes the canonical inclusion of  2
into  ∞ (cf. Lemma 1.2.1), fails to have dense range. Thus, the conclusion in (2.1.54)
may fail when the Banach space X is not reflexive.

Lemma 2.1.7 Let X be a Banach space, T : X → X a linear bounded operator and

V ⊆ X a closed subspace, invariant under T. Then

the annihilator V ⊥ is a closed subspace of X ∗ which

(2.1.55)
is invariant under the adjoint T ∗ : X ∗ → X ∗
and the quotient map
X X
[T] : −→ , [T](x + V) := (T x) + V, ∀x ∈ X, (2.1.56)
V V
is well defined, linear and bounded. Moreover, under the identification made in
(2.1.32), it follows that

[T]∗ : (X/V)∗ → (X/V)∗ , the adjoint of (2.1.56),

(2.1.57)
is the operator T ∗ V ⊥ : V ⊥ → V ⊥ .

Finally, if the Banach space X is actually reflexive then, under the identifications
made in (2.1.32) and (2.1.35), it follows that

the adjoint of T ∗ V ⊥ : V ⊥ → V ⊥
(2.1.58)
is the operator [T] : X/V → X/V,

and
T ∗ maps V ⊥ isomorphically onto itself if and only if the
(2.1.59)
quotient map [T] is an isomorphism in (2.1.56).
Proof That V ⊥ is a closed subspace of X ∗ , invariant under T ∗ , and the quotient
map in (2.1.56) is well defined, linear and bounded, are all routine consequences
of definitions. Next, observe that π ◦ T = [T] ◦ π where π is as in (2.1.28). Taking
72 2 Abstract Fredholm Theory

adjoints and using (2.1.33) then yields T ∗ ◦ ι = ι◦[T]∗ on (X/V)∗ = V ⊥ (cf. (2.1.32)),
where ι is the canonical inclusion (2.1.29). From this, the claim in (2.1.57) readily
follows.
Moving on, work under the assumption that X is reflexive. Then the identifications
made in (2.1.32) and (2.1.35) give
 X  ∗∗  ∗ X
= V⊥ = , (2.1.60)
V V
hence X/V is reflexive. Granted this, the claim made in (2.1.58) follows from (2.1.57).
There remains to deal with the equivalence in (2.1.59). In one direction, assume

T ∗ : V ⊥ −→ V ⊥ isomorphically. (2.1.61)

Let x + V ∈ Ker [T]. Then x ∈ X is such that T x ∈ V. We wish to show that x ∈ V.

Since V = V, according to one of the standard corollaries of the Hahn-Banach
Theorem the latter membership is equivalent to proving that

Λ(x) = 0 for each Λ ∈ V ⊥ . (2.1.62)

To this end, fix Λ ∈ V ⊥ arbitrary. Thanks to (2.1.61), there exists Ψ ∈ V ⊥ such that

T ∗ Ψ = Λ. Hence,
Λ(x) = (T ∗ Ψ)(x) = Ψ(T x) = 0, (2.1.63)
where the last equality is based on the fact that T x ∈ V and Ψ ∈ V ⊥ . This establishes
(2.1.62) which, in turn, finishes the proof of the fact that x ∈ V. In summary,
x + V ∈ Ker [T] forces x ∈ V, which goes to show that [T] is injective.
To prove that [T] is surjective, consider an arbitrary y ∈ X. Define
  −1 
Θ : V ⊥ −→ C, Θ(Λ) := T ∗ V ⊥ Λ (y) for each Λ ∈ V ⊥ . (2.1.64)

Given that Θ is a linear and continuous functional, the Hahn-Banach Theorem ensure
 from the subspace V ⊥ to the entire X ∗ . Hence,
that it has an extension, call it Θ,
 ∗ ∗
Θ ∈ (X ) and since we are presently assuming that X is reflexive, there exists x ∈ X

with the property that Θ(Λ) = Λ(x) for every Λ ∈ X ∗ . In particular,
  −1 
Λ(x) = T ∗ V ⊥ Λ (y) for each Λ ∈ V ⊥, (2.1.65)

hence
T ∗ Λ(x) = Λ(y) for each Λ ∈ V ⊥, (2.1.66)
which further implies

Λ(T x − y) = 0 for each Λ ∈ V ⊥ . (2.1.67)

From this and the same standard corollary of the Hahn-Banach Theorem invoked
earlier we deduce that T x−y ∈ V = V. Thus, ultimately, [T](x+V) = (T x)+V = y+V
2.2 Fredholm Theory in Topological Vector Spaces 73

which goes to show that [T] is indeed surjective. At this stage, we may therefore
conclude that if T ∗ is an isomorphism in the context of (2.1.61) then [T] is an
isomorphism in (2.1.56).
Conversely, assume [T] is an isomorphism in (2.1.56). Consider a functional
Λ ∈ V ⊥ such that T ∗ Λ = 0. Pick an arbitrary y ∈ X. Since [T] is surjective, there
exists x ∈ X with the property that T x − y ∈ V. As such, we may write

0 = Λ(T x − y) =⇒ Λ(y) = Λ(T x) = T ∗ Λ(x) = 0. (2.1.68)

Thus, Λ = 0 which proves that T ∗ V ⊥ is injective.

There remains to show that T ∗ : V ⊥ → V ⊥ is surjective. As a preamble, we
introduce some notation. For every y ∈ X, define Wy := {x ∈ X : (T x) − y ∈ V }.
The fact that [T] is surjective implies Wy  . Moreover, since [T] is injective if
x1, x2 ∈ Wy then necessarily x1 − x2 ∈ V. Also, since V is invariant under T, we have
V + Wy = Wy . The fact that [T] is an isomorphism also implies Wy1 +y2 = Wy1 + Wy2
for every y1, y2 ∈ X, and Wy = V whenever y ∈ V. Finally, WT x = x + V for every
x ∈ X.
Returning to the main topic of conversation, pick an arbitrary functional Λ ∈ V ⊥
and define the mapping Φ : X → C by setting

Φ(y) := Λ(x), for each y ∈ X and x ∈ Wy . (2.1.69)

The earlier digression ensures that Φ is a well-defined linear mapping whose restric-
tion to V vanishes identically. In concert, the continuity of [T] and the continuity of
Λ imply that Φ is continuous, hence Φ ∈ V ⊥ . In addition, for each x ∈ X and each
z ∈ WT x = x + V we have

(T ∗ Φ)(x) = Φ(T x) = Λ(z) = Λ(x). (2.1.70)

Hence, T ∗ Φ = Λ proving that the operator T ∗ : V ⊥ → V ⊥ is indeed surjective (thus,

ultimately, an isomorphism). 

2.2 Fredholm Theory in Topological Vector Spaces

In this section we refine some of the results in §2.1 through the consideration of
topological vector spaces more general than Banach spaces. The following definition
(which should be contrasted with Definition 2.1.1) plays a fundamental role.

Definition 2.2.1 Given two linear topological spaces X, Y , a linear continuous map-
ping T : X → Y is said to be Fredholm provided it verifies the following three
axioms:
(F1) Im T := Im (T : X → Y ) is a closed subspace of Y , of finite codimension in Y .
(F2) Ker T := Ker (T : X → Y ) is finite dimensional and topologically complemented
in X.
74 2 Abstract Fredholm Theory

(F3) T is a relatively open map (meaning that T(O) is a relatively open subset of
Im (T : X → Y ) whenever O is open in X).
Set  
Φ(X → Y ) := T : X → Y : T Fredholm (2.2.1)
and consider the index function

index : Φ(X → Y ) −→ Z defined by

(2.2.2)
index T := dim (Ker T) − codim (Im T) for each T ∈ Φ(X → Y ).

As in the case of Banach spaces, if we wish to stress the spaces on which the
operator T is considered, we may write index (T : X → Y ) in place of the less
revealing symbol index T.
In relation to the very general definition given above, it reassuring to note that in
a more restrictive setting we can get away with assuming less, namely,

if X and Y are actually Banach spaces, then T ∈ Bd(X → Y ) is

Fredholm in the sense of Definition 2.2.1 if and only if T has a finite
(2.2.3)
codimensional range, and a finite-dimensional kernel (hence, T is
Fredholm in the sense of Definition 2.1.1).
Specifically, we make the following remark:

Remark 2.2.2 (i) Regarding condition (F1) in Definition 2.2.1, the reader is alerted
to the fact that if X, Y are quasi-Banach spaces and T ∈ Bd(X → Y ) has a range of
finite codimension in Y , then necessarily Im (T : X → Y ) is a closed subspace of Y
(see Lemma 1.2.7).
(ii) Regarding condition (F2) in Definition 2.2.1, recall that a finite-dimensional
subspace of a Hausdorff linear topological space is always closed, but it need
not necessarily be topologically complemented. On the other hand, if X has enough
continuous linear functionals to separate points (e.g., if X is a normed space, or more
generally is locally convex) then every finite-dimensional subspace is topologically
complemented.
(iii) Regarding condition (F3) in Definition 2.2.1, we wish to note that for quasi-
Banach spaces the Open Mapping Theorem (for quasi-Banach spaces; cf., e.g., [138,
Corollary 6.62, p. 423]) provides this relative openness property once we know that
the range of T is closed.

In stark contrast with the case of Banach spaces just considered above, the topic
of Fredholm theory in non-locally convex spaces has remained largely unexplored.
Yet recent work in [142] shows that the core principle of the theory, namely that a
Fredholm operator is a linear homeomorphism modulo finite-dimensional spaces is
pervasive. This is particularly evident in the next result, a refinement of Atkinson’s
classical result in Banach spaces (cf. item (7) in Theorem 2.1.2), which should be
viewed as the “Fundamental Theorem of Fredholm Theory” in linear topological
spaces.
2.2 Fredholm Theory in Topological Vector Spaces 75

Theorem 2.2.3 Given a pair of Hausdorff linear topological spaces X, Y , along with
an operator T ∈ L (X → Y ), the following statements are equivalent:
(a) T is a Fredholm operator (cf. Definition 2.2.1).
(b) There exist S1, S2 ∈ L (Y → X) such that both IX − S1T and IY − T S2 are
compact operators.
(c) The operator T is invertible modulo compacts, in the sense that there exists
some operator S ∈ L (Y → X) such that both IX − ST and IY − T S are compact
operators.
(d) The operator T is invertible modulo operators with finite dimensional ranges, in
the sense that there exists S ∈ L(Y → X) such that both IX − ST and IY − T S
have finite dimensional ranges.
Moreover, if the operator T ∈ L(X → Y ) is Fredholm then matters may be
arranged so that the operators S ∈ L(Y → X) appearing in items (c) and (d) above
are also Fredholm.

As a consequence of the above theorem we see that

if X is a Hausdorff linear topological space and T ∈ L (X), then T

is a Fredholm operator on X if and only if its equivalence class is an (2.2.4)
invertible element of the Calkin algebra L (X)/Cp (X).
We wish to augment Theorem 2.2.3 with the following list of additional properties
of Fredholm operators in linear topological spaces (again, see [142] for a proof).

Theorem 2.2.4 Let X, Y , Z be Hausdorff linear topological spaces. Then the fol-
lowing properties hold.
(a) If T ∈ L (X → Y ) is a Fredholm operator (in the sense of Definition 2.2.1)
then for any K ∈ Cp (X → Y ) the operator T + K is Fredholm and one has
index (T + K) = index (T). Succinctly put, the class of Fredholm operators is
stable under compact perturbations and so is the index. In particular, if T is a
compact perturbation of the identity on X, then the operator T is Fredholm with
index zero, hence the Fredholm Alternative holds: T is one-to-one if and only if
T is onto.
More broadly, the following Generalized Fredholm Alternative holds: if an
operator T ∈ L (X → Y ) is invertible and K ∈ Cp (X → Y ), then T + K is a
Fredholm operator and index (T + K) = 0, hence T + K is one-to-one if and only
if it maps X onto Y .
(b) If T ∈ L (X → Y ) and S ∈ L (Y → Z) are Fredholm operators, then so is ST
and
index (T S) = index (T) + index (S). (2.2.5)
(c) Suppose Y ⊆ X a closed subspace of finite codimension in X. Assume that
T ∈ L (X) is a Fredholm operator on X such that T(Y ) ⊆ Y and define
 
T Y : Y → Y, T Y (x) := T x, ∀x ∈ Y . (2.2.6)
76 2 Abstract Fredholm Theory
 
Then the operator T Y is a Fredholm operator on Y and index T Y = index T.
(d) Assume T ∈ L (X) is a Fredholm operator and V ⊆ X is a finite dimensional
subspace of X such that T(V) ⊆ V. Then the quotient map
X X
[T] : −→ , [T](x + V) := (T x) + V, ∀x ∈ X, (2.2.7)
V V
is a Fredholm operator on X/V and index [T] = index T.
(e) If Y ⊆ X is a closed subspace of finite codimension in X, it follows that the canon-
ical inclusion ι : Y → X is a Fredholm operator with index ι = −dim (X/Y ).
Also, if V ⊆ X is a finite dimensional subspace of X then the canonical quotient
projection π : X → X/V is a Fredholm operator with index π = dim V.
(f) Let Y be a Hausdorff topological vector spaces such Y is a dense subset of X.
Consider a Fredholm operator T : X → X for which Y is an invariant subspace
and such that its restriction to Y , i.e., T Y defined as in (2.2.6), is a Fredholm
operator on Y . Then
   
index T Y = index (T) =⇒ ker T = ker T Y . (2.2.8)

(g) If T ∈ L (X → Y ) is a Fredholm operator and F is a finite dimensional linear

subspace of Y , then T −1 F is a finite dimensional linear subspace of X.
(h) If actually X, Y are quasi-Banach spaces, then the set Φ(X → Y ) of all Fredholm
operators from X into Y is open in Bd (X → Y ), and the index function

index : Φ(X → Y ) −→ Z (2.2.9)

is continuous, where Z is equipped with the discrete topology. Consequently, for

each fixed n ∈ Z, the set of all Fredholm operators of index n from X into Y is
both (relatively) open and closed in Φ (X → Y ).

For further use let us note that, as may be seen from Theorem 2.2.3 and Theo-
rem 2.2.4,
if X, Y are Hausdorff linear topological spaces and if T ∈ L (X → Y )
and S ∈ L (Y → Z) are such that ST is a Fredholm operator, then T (2.2.10)
is Fredholm if and only if S is Fredholm.
The notion of Fredholm radius, recalled below, has been originally introduced by
J. Radon in [156].

Definition 2.2.5 Let X be a quasi-Banach space and denote by I the identity operator
on X. Given an operator T ∈ Bd (X), define its spectral radius as

ρinv (T; X) := inf r > 0 : zI − T :X → X is a homeomorphism

for each z ∈ C \ B(0, r) . (2.2.11)
2.2 Fredholm Theory in Topological Vector Spaces 77

Also, define the Fredholm radius1 of T as

 
ρFred (T; X) := inf r > 0 : zI − T ∈ Φ(X → X) for each z ∈ C \ B(0, r) .
(2.2.12)

Lemma 1.2.12 implies that the infima in (2.2.11)-(2.2.12) are meaningful, hence
both ρinv (T; X) and ρFred (T; X) are well-defined numbers in [0, ∞). In fact,

0 ≤ ρFred (T; X) ≤ ρinv (T; X) ≤ T X→X . (2.2.13)

It is also clear from Definition 2.2.5 that for any T ∈ Bd (X) we have

ρFred (T; X) = 0 whenever T ∈ Cp (X). (2.2.14)

It is also useful to note that item (3) of Theorem 2.1.2 implies that

if X is a Banach space, then for each operator T ∈ Bd (X) we

(2.2.15)
have ρFred (T ∗ ; X ∗ ) ≤ ρFred (T; X) (with equality if X is reflexive).

Lemma 2.2.6 Let X be a p-Banach space with p ∈ (0, 1] and consider an operator
T ∈ Bd (X). Then
ess
ρFred (T; X) ≤ T X→X ≤ T X→X (2.2.16)

and
zI − T is a Fredholm operator with index zero
whenever z ∈ C satisfies |z| > ρFred (T; X). (2.2.17)

Moreover,
if X is actually a Banach space and X0 is a closed sub-
space of X which is invariant under T (i.e., T X0 ⊆ X0 ), then (2.2.18)
ρFred (T |X0 ; X0 ) ≤ ρFred (T; X).
ess
Proof Fix an arbitrary real number λ > T X→X . Then (1.2.54) guarantees the
existence of some operator K ∈ Cp(X → X) with the property that T −K X→X < λ.
Then, for each z ∈ C with |z| ≥ λ, the operator zI − (T − K) is invertible (via a
Neumann series; cf. Lemma 1.2.12). Consequently, by item (a) in Theorem 2.2.4,
we have zI −T = zI − (T − K) − K ∈ Φ(X → X). In view of (2.2.12), this proves that
ρFred (T; X) ≤ λ. With this in hand, the first inequality in (2.2.16) follows by letting
ess
λ  T X→X . The second inequality in (2.2.16) is a particular case of (1.2.56).
Finally, the mapping
 
z ∈ C : |z| > ρFred (T; X)  z −→ index (zI − T) ∈ Z (2.2.19)

1 sometimes referred to as the essential spectral radius

78 2 Abstract Fredholm Theory

is continuous, hence constant (cf. item (h) in Theorem 2.2.4), and takes the value
zero whenever |z| is large enough (cf. Lemma 1.2.12). As such, said mapping is
identically zero. This finishes the proof of (2.2.17).
As regards (2.2.18), assume X is actually a Banach space and suppose X0 is
a closed subspace of X which is invariant under T. From (2.2.17) we see that
zI − T ∈ Φ+ (X → X) for each z ∈ C with |z| > ρFred (T; X). Since X0 is a closed
subspace of X which is invariant under zI − T, from (2.1.8) we conclude that z ∈ C
with |z| > ρFred (T; X) we have zIX0 − T |X0 ∈ Φ+ (X0 → X0 ), where IX0 is the identity
operator on X0 . Hence, the mapping
 
z ∈ C : |z| > ρFred (T; X)  z −→ index (zIX0 − T |X0 ) ∈ Z (2.2.20)

is continuous, hence constant (cf. item (8) in Theorem 2.1.2), and takes the value zero
whenever |z| is large enough (cf. Lemma 1.2.12). As a consequence, the mapping
(2.2.20) is identically zero. This implies that for each z ∈ C with |z| > ρFred (T; X)
we have
   
dim coker (zIX0 − T |X0 ) = dim Ker (zIX0 − T |X0 ) < ∞, (2.2.21)

thus

zIX0 − T |X0 ∈ Φ(X0 → X0 ) for each z ∈ C with |z| > ρFred (T; X). (2.2.22)

Ultimately, this proves that ρFred (T |X0 ; X0 ) ≤ ρFred (T; X). 

It has been known for a while that the Fredholm radius (aka essential spectral
radius; cf. Definition 2.2.5) behaves naturally under real interpolation of Banach
spaces. Specifically, [30, Corollary 1.3, p. 37] reads as follows:
Proposition 2.2.7 Suppose X0, X1 are two compatible Banach spaces and consider
a linear operator T : X0 + X1 → X0 + X1 such that its restrictions T : X j → X j are
bounded for j = 0, 1. Then for each θ ∈ (0, 1) and q ∈ [1, ∞] one has
  
ρFred T; (X0, X1 )θ,q ≤ ρFred (T; X0 )1−θ · ρFred T; X1 )θ . (2.2.23)

This extends earlier work in [7] where the case 1 ≤ q < ∞ (a scenario which ensures
that X0 ∩ X1 is dense in (X0, X1 )θ,q ) has been treated via a different approach.
All these results deal with Banach spaces and the real method of interpolation. In
this regard, Proposition 2.2.7 should be compared with Proposition 1.4.24 where a
related interpolation result for the essential norm is established, both for the real and
complex interpolation method, for linear and bounded operators on quasi-Banach
spaces.
The notion of spectral radius makes sense in a general unital Banach algebra
(A,  · ). Specifically, if 1 denotes the unit in A then for each x ∈ A we define
 
ρinv (x) := inf r > 0 : z1 − x invertible in A for each z ∈ C \ B(0, r) . (2.2.24)

In relation to this, we have the following classical result:

2.2 Fredholm Theory in Topological Vector Spaces 79

Theorem 2.2.8 (Spectral Radius Formula) Let (A,  · ) be a unital Banach algebra.
Then for each x ∈ A the limit lim  x n  1/n exists and satisfies
n→∞

lim  x n  1/n = ρinv (x) = inf  x n  1/n . (2.2.25)

n→∞ n∈N

Let (X,  · X ) be a Banach space. We may then define

ρinv (T; X) to be the spectral radius of T ∈ Bd (X), in the sense of

(2.2.26)
(2.2.24), regarding Bd (X) a unital Banach algebra.
Also, the Calkin algebra

A := Bd (X)/Cp (X) equipped with the norm

ess
(2.2.27)
Bd (X)/Cp (X)  [T] → [T] A := T X→X

is a unital Banach algebra, and invertible elements in A are precisely equivalence

classes [T] of Fredholm operators T on X. In view of this and (2.2.12), we then
conclude from Theorem 2.2.8 that for each T ∈ Φ(X → X) we have
 n ess  1/n
ρFred (T; X) = lim [T]n  1/n
A = lim T  X→X (2.2.28)
n→∞ n→∞

and
 n ess  1/n
ρFred (T; X) = inf [T]n  1/n
A = inf T  X→X . (2.2.29)
n∈N n∈N
Chapter 3
Functions of Vanishing Mean Oscillations and
Vanishing Hölder Moduli

In this chapter we study functions whose oscillations vanish at infinitesimal scales.

When said oscillations are measured in a mean (i.e., integral) sense, this gives rise
to the Sarason space VMO, discussed in §3.1. When the oscillations in question are
taken in a pointwise sense (via a local Hölder semi-norm), we come across a new
brand of Hölder spaces, defined in §3.2.
Originally, Sarason’s space VMO has been defined as the closure of uniformly
continuous functions in BMO. Via a PDE approach (relying on these well-posedness
of the Dirichlet Problem in the upper half-space with data in BMO) in [124] the
authors succeed in producing characterizations of the space VMO as the closure
in BMO of classes of smooth functions contained in BMO within which uniform
continuity may be suitably quantified, such as the class of smooth functions satisfying
a Hölder or Lipschitz condition (thus improving Sarason’s classical result describing
VMO as the closure in BMO of the space of uniformly continuous functions with
bounded mean oscillations). In turn, this stronger density result was used in [124] to
show that any Calderón-Zygmund operator T satisfying T(1) = 0 extends as a linear
and bounded mapping from VMO (modulo constants) into itself. One of the main
goals in §3.1 is to give a new proof of the aforementioned density result of a purely
real analysis flavor, which works in the general context of spaces of homogeneous
type.
Subsequently, in §3.2 we make further use of this technology to study a related
scale, namely a new brand of Hölder spaces, obtained by imposing the condition
that the local Hölder semi-norm vanishes as the scale (at which this is considered)
approaches zero.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 81

D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_3
82 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

3.1 Functions of Vanishing Mean Oscillations on Measure Metric

Spaces

In this section we shall concern ourselves with the space VMO which, heuristically,
should be thought of as an integral version1 of uniform continuity. Specifically, given
a measure metric space (X, ρ, μ), define the Sarason space VMO(X, μ) of functions
of vanishing mean oscillations on X (cf. [170]) as

VMO(X, μ) := the closure of UC(X, ρ) ∩ BMO(X, μ) in BMO(X, μ), (3.1.1)

where UC(X, ρ) stands for the space of uniformly continuous functions on the metric
space (X, ρ). That is, VMO(X, μ) is the linear subspace of BMO(X, μ) consisting of
f ∈ BMO(X, μ) for which one may find { f j } j ∈N ⊆ UC(X, ρ) ∩ BMO(X, μ) with the
property that lim  f − f j BMO(X,μ) = 0. Note that
j→∞

if X is bounded then VMO(X, μ) is a closed subspace of BMO(X, μ), a

scenario in which VMO(X, μ) is a Banach space when equipped with
(3.1.2)
the norm  · BMO(X,μ) ; hence, VMO(X, μ) → BMO(X, μ) isometri-
cally in this case.
In the case when X is unbounded, the quotient space
  

VMO(X, μ) := VMO(X, μ) ∼ = [ f ] : f ∈ VMO(X, μ) (3.1.3)


equipped with the norm inherited from BMO(X, μ) (cf. [133, (7.4.95)]) is complete,
hence Banach. It is known that the BMO space remains unchanged if the underlying
measure is multiplied by an A∞ weight. Here is a version of this result for functions
of vanishing mean oscillations.

Lemma 3.1.1 Let (X, ρ, μ) be a measure metric space. Then for each w ∈ A∞ (X, μ)
one has
VMO(X, μ) = VMO(X, w). (3.1.4)

Proof This is a direct consequence of [133, Lemma 7.7.5] and (3.1.1). 

Even though almost everything in this section is valid in the context of arbitrary
spaces of homogeneous type, we find it convenient to restrict ourselves to working
on a closed Ahlfors regular subset Σ of Rn , which we equip with the “surface
measure” σ := H n−1 Σ. To set the stage for what follows, for each f ∈ Lloc 1 (Σ, σ)

and p ∈ [1, ∞) abbreviate

1 as opposed to a pointwise version

3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 83
⨏  ⨏  p  1/p
 
Mp ( f ; R) := sup sup f − f dσ  dσ for all R ∈ (0, ∞),
x ∈Σ r ∈(0,R) Δ(x,r) Δ(x,r)

where Δ(x, r) := Σ ∩ B(x, r) for each x ∈ Σ and r ∈ (0, ∞).

(3.1.5)
Obviously, Mp ( f ; R) is non-decreasing in R. Also, for each fixed p ∈ [1, ∞), we see
from [133, (7.4.112)] that
⎧
⎨ supR>0 Mp ( f ; R)
⎪
⎪ if Σ is unbounded,
 f BMO(Σ,σ) ≈  ∫ 
 (3.1.6)
⎪
⎪  Σ f dσ  + supR>0 Mp ( f ; R) if Σ is bounded,
⎩
uniformly in f ∈ Lloc
1 (Σ, σ). Since for each p ∈ [1, ∞) we have

lim Mp ( f ; R) = 0 whenever f is uniformly continuous on Σ, (3.1.7)

R→0+

it follows from (3.1.1), (3.1.7), and (3.1.6) that

lim Mp ( f ; R) = 0 for each f ∈ VMO(Σ, σ) and p ∈ [1, ∞). (3.1.8)

R→0+

Next, the goal is to prove a quantitative characterization of Sarason’s space VMO,

in which uniformly continuous functions are replaced by Hölder functions. Such
a version is contained in Theorem 3.1.3, formulated a little further below. As a
preamble, we establish the result in the following lemma.

Lemma 3.1.2 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, fix some β ∈ (0, ∞). Then the collection of all functions
f ∈ BMO(Σ, σ) with the property that there exists some r0 = r0 ( f ) ∈ (0, ∞) such
that
| f (x) − f (y)|
sup < +∞ (3.1.9)
x,y ∈Σ |x − y| β
0< |x−y |<r0
.
is contained in the homogeneous Hölder space C α (Σ) for each α ∈ (0, β).

Proof The claim is trivially true if Σ is compact, so we shall focus on the case when
the set Σ is unbounded. Fix an arbitrary point x ∈ Σ and radius r ∈ (0, ∞). Abbreviate
Δ(x, r) := ⨏Σ ∩ B(x, r) and, having picked a function f as in the statement, introduce
fΔ(x,r) := Δ(x,r) f dσ. Then, on the one hand, (A.0.8) and (A.0.9) imply
⨏
 
 f − fΔ(x,r)  dσ ≤  f BMO(Σ,σ) . (3.1.10)
Δ(x,r)

On the other hand, if C ∈ [0, ∞) denotes the number defined as the supremum in
(3.1.9), then for each r ∈ (0, r0 /2) we may estimate
84 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli
⨏ ⨏ ⨏
 
 f − fΔ(x,r)  dσ ≤ | f (y) − f (z)| dσ(y) dσ(z)
Δ(x,r) Δ(x,r) Δ(x,r)
⨏ ⨏
≤C |y − z| β dσ(y) dσ(z) ≤ C(2r)β .
Δ(x,r) Δ(x,r)
(3.1.11)

Pick now some α ∈ (0, β) arbitrary. Raise (3.1.10) to the power 1 − (α/β), raise
(3.1.11) to the power α/β, then multiply the resulting inequalities. This leads to the
conclusion that
⨏
 
 f − fΔ(x,r)  dσ ≤ C α/β (2r)α  f  1−(α/β) for each r ∈ (0, r0 /2). (3.1.12)
BMO(Σ,σ)
Δ(x,r)

Straight from (3.1.10), we also have

⨏
 
 f − fΔ(x,r)  dσ ≤ (r0 /2)α  f BMO(Σ,σ) · r α for each r ∈ [r0 /2, ∞). (3.1.13)
Δ(x,r)

Together, (3.1.12) and (3.1.13) prove that there exists some C0 ∈ (0, ∞) with the
property that
⨏
 
 f − fΔ(x,r)  dσ ≤ C0 · r α for each r ∈ (0, ∞). (3.1.14)
Δ(x,r)

Having shown this and bearing in mind that Σ is Ahlfors regular, [133, Proposi-
tion 7.4.9] applies with p := 1 (thanks to [133, Lemma 3.6.4] presently used with
s := n − 1) and gives that f may be redefined on a σ-nullset as to become a function
.
in C α (Σ). 

Our next theorem (alluded to earlier) extends a similar density result first estab-
lished in [124] in the important special case when Σ = Rn−1 × {0}. The original
argument in [124] is PDE-based, making essential use of specific properties of the
Euclidean space, and here we develop a new approach, of a purely real-variable
nature, which actually works in arbitrary spaces of homogeneous type.

Theorem 3.1.3 Let Σ be a closed Ahlfors regular subset of Rn and set σ := H n−1 Σ.
Then for each α ∈ (0, 1) it follows that
.
BMO(Σ, σ) ∩ C α (Σ) is a dense subspace of VMO(Σ, σ). (3.1.15)

Proof Fix an exponent α ∈ (0, 1). For starters, it is relevant to observe that the
definition in (3.1.1) shows that, indeed,
.
BMO(Σ, σ) ∩ C α (Σ) is a subspace of VMO(Σ, σ). (3.1.16)

The idea now is to bring in a sufficiently regular approximation to the identity.

Concretely, [9, Theorem 3.22, pp. 102-103] guarantees the existence of a family of
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 85
 
(σ ⊗ σ)-measurable functions St : Σ × Σ → R indexed by t ∈ 0, 2 diam Σ and
satisfying,
 for some
 positive finite constant C0 , the following properties for every
t ∈ 0, 2 diam Σ :

0 ≤ St (x, y) ≤ C0 t 1−n for all x, y ∈ Σ,

St (x, y) = 0 for all x, y ∈ Σ with |x − y| ≥ C0 t,
(3.1.17)
|St (x, y) − St (x , y)| ≤ C0 t −n |x − x  | for all x, x , y ∈ Σ,
∫
Σ t
S (x, z) dσ(z) = 1 for all x ∈ Σ.

Without loss of generality, we mayassume that C0 ≥ 2. Next, fix an arbitrary function

f ∈ BMO(Σ, σ) and for each t ∈ 0, 2 diam Σ define
∫
gt (x) := St (x, y) f (y) dσ(y) for each x ∈ Σ. (3.1.18)
Σ

In view of (3.1.17), the fact that f is locally absolutely integrable, and the measure

σ is locally finite, this definition is meaningful. Also, for each t ∈ 0, 2 diam Σ and
x, x  ∈ Σ with |x − x  | < C0 t, the properties listed in (3.1.17) permit us to write
∫ ∫ 
 
   
|gt (x) − gt (x )| =  St (x, y) f (y) dσ(y) − St (x , y) f (y) dσ(y)
 Σ Σ 
∫ ⨏
  

=  St (x, y) f (y) − f dσ dσ(y)
 Σ Δ(x,2C0 t)

∫ ⨏ 
  
 
− St (x , y) f (y) − f dσ dσ(y)
Σ Δ(x,2C0 t) 
∫  ⨏ 
 
≤ |St (x, y) − St (x , y)|  f (y) − f dσ  dσ(y)
Σ Δ(x,2C0 t)
∫  ⨏ 
 
≤ C0 t −n |x − x  |  f (y) − f dσ  dσ(y).
Δ(x,2C0 t) Δ(x,2C0 t)
(3.1.19)

Bearing in mind the fact that Σ is Ahlfors regular and the definition in (3.1.5), this
implies that there exist C1, C2 ∈ (1, ∞) such that

|gt (x) − gt (x )| ≤ C1 t −1 |x − x  |M1 ( f ; C2 t) whenever

  (3.1.20)
t ∈ 0, 2 diam Σ and x, x  ∈ Σ satisfy |x − x  | < C0 t.
 
As a consequence, for each t ∈ 0, 2 diam Σ we have
86 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

|gt (x) − gt (x )|

sup ≤ C1 t −1  f BMO(Σ,σ) < +∞, (3.1.21)
x,x  ∈Σ |x − x  |
0< |x−x  |<C0 t

which goes to show that each  gt is a continuous

 function; in particular, we have
gt ∈ Lloc
1 (Σ, σ) for each t ∈ 0, 2 diam Σ .

We now claim that there exist two constants C3, C4 ∈ (0, ∞), independent of f ,
such that
 
sup M1 ( f − gt ; R) ≤ C3 M1 ( f ; C4 t) for each t ∈ 0, 2 diam Σ . (3.1.22)
R>0
 
To justify this inequality, fix some number t ∈ 0, 2 diam Σ along with an arbitrary
point x ∈ Σ and a radius R ∈ (0, ∞). In the regime 0 < R < t, we estimate
⨏  ⨏ 
 
( f − gt ) − ( f − gt ) dσ  dσ
Δ(x,R) Δ(x,R)
⨏  ⨏  ⨏  ⨏ 
   
≤ f − f dσ  dσ + gt − gt dσ  dσ
Δ(x,R) Δ(x,R) Δ(x,R) Δ(x,R)

=: I + II. (3.1.23)

On the one hand, from (3.1.5) and the fact that R ∈ (0, t) is arbitrary it follows that
I ≤ M1 ( f ; t). Note that

|y − z| < 2R < C0 t for each y, z ∈ Δ(x, R). (3.1.24)

thanks to the fact that C0 ≥ 2 and that we are currently assuming R < t. So, on the
other hand,
⨏ ⨏
II ≤ |gt (y) − gt (z)| dσ(y) dσ(z)
Δ(x,R) Δ(x,R)
⨏ ⨏
−1
≤ C1 t M1 ( f ; C2 t) |y − z| dσ(y) dσ(z)
Δ(x,R) Δ(x,R)

≤ C1 t −1 M1 ( f ; C2 t) 2R < 2C1 M1 ( f ; C2 t), (3.1.25)

by (3.1.24), (3.1.20), and the fact that we are presently working in the regime
0 < R < t (hence t −1 R < 1). Bearing in mind that M1 ( f ; ·) is nondecreasing, the
estimates on I, II above yield the following conclusion:
⨏  ⨏ 
 
0 < R < t =⇒ ( f − gt ) − ( f − gt ) dσ  dσ ≤ 3C1 M1 ( f ; C2 t).
Δ(x,R) Δ(x,R)
(3.1.26)
In view of the goal set forth in (3.1.22), this is a step in the right direction.
Hence, as far as (3.1.22) is concerned, there remains to consider the case when
R ≥ t. In such a scenario, pick k ∈ Z such that
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 87

2−k ≤ t < 2−k+1 (3.1.27)

and bring in the dyadic grid structure from [133, Proposition 7.5.4], used with X := Σ
(in view of [133, (7.5.6)] and (3.1.27) such a choice guarantees that we have k ≥ κΣ ).
In particular, from [133, (7.5.17)] we know that there exists a σ-measurable subset
Nk of Σ such that (with notation introduced in [133, Proposition 7.5.4])
  
Σ= Q αk Nk with σ(Nk ) = 0. (3.1.28)
α∈Ik

From [133, (7.5.9)] and (3.1.27) we see that there exist two finite constants, c > 0
and λ > 0, such that

Δ(xαk , c t) ⊆ Q αk ⊆ Δ(xαk , λt) for each α ∈ Ik . (3.1.29)

In turn, from (3.1.29) and the fact that we are currently assuming R ≥ t we conclude
that
having Q αk ∩ Δ(x, R)   for some α ∈ Ik
  (3.1.30)
forces Q αk ⊆ Δ x, (2λ + 1)R .
In particular,
    
Q αk ⊆ Δ x, (2λ + 1)R if J(k, x, R) := α ∈ Ik : Q αk ∩ Δ(x, R)   .
α∈J(k,x,R)
(3.1.31)
Granted these properties, we may now estimate2

2 alternatively, we could have used Vitali’s Covering Lemma (see (3.2.32) in this regard)
88 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli
⨏  ⨏ 
 
( f − gt ) − ( f − gt ) dσ  dσ
Δ(x,R) Δ(x,R)
⨏
≤2 | f (y) − gt (y)| dσ(y)
Δ(x,R)
∫  ⨏
  
2 
≤  f (y) − f dσ
σ(Δ(x, R)) k 
Qα Δ(xαk ,(λ+C0 )t)
α∈J(k,x,R)
∫ ⨏ 
  

− St (y, z) f (z) − f dσ dσ(z) dσ(y)
Σ Δ(xαk ,(λ+C0 )t) 
  
σ Δ(xαk , (λ + C0 )t)
≤C ×
σ(Δ(x, R))
α∈J(k,x,R)
⨏  ⨏ 
 
× f − f dσ  dσ
Δ(xαk ,(λ+C0 )t) Δ(xαk ,(λ+C0 )t)

   σ(Q αk ) 
≤ C M1 f ; (λ + C0 )t
σ(Δ(x, R))
α∈J(k,x,R)
 
σ
k
 α∈J(k,x,R) Q α
= C M1 f ; (λ + C0 )t
σ(Δ(x, R))
 
  σ Δ(x, (2λ + 1)R)
≤ C M1 f ; (λ + C0 )t
σ(Δ(x, R))
 
≤ C M1 f ; (λ + C0 )t , (3.1.32)

for some geometric constant C ∈ (0, ∞), independent of f , x, and R. Specifically, the
first inequality in (3.1.32) is obvious. The second inequality in (3.1.32) is obtained
by unpacking the integral average on Δ(x, R), using the definition of gt from (3.1.18),
the last property in (3.1.19), the fact that the “dyadic cubes” {Q αk }α∈Ik cover Σ up
to a σ-nullset (cf. (3.1.28)), and the definition of the index set J(k, x, R) given in
(3.1.31). The third inequality in (3.1.32) is based on the first property in (3.1.19) and
the Ahlfors regularity of Σ, the fact that thanks to the second inclusion in (3.1.29)
we have  
Q αk ⊆ Δ(xαk , λt) ⊆ Δ xαk , (λ + C0 )t , (3.1.33)
plus the observation that if y ∈ Q αk and z ∈ Σ are such that |y − z| < C0 t then

|z − xαk | ≤ |xαk − y| + |y − z| < λt + C0 t = (λ + C0 )t (3.1.34)

 k 
so we necessarily have z ∈ Δ xα, (λ + C0 )t . The fourth inequality in (3.1.32) is seen
from the definition made in (3.1.5), the fact that Σ is Ahlfors regular, and (3.1.29).
The subsequent equality in (3.1.32) is a result of the fact that the “dyadic cubes”
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 89

{Q αk }α∈Ik are mutually disjoint. The penultimate inequality in (3.1.32) comes from
(3.1.31). The final inequality in (3.1.32) uses the fact that the measure σ is doubling.
In summary, the implication in (3.1.26) also holds (with possibly different con-
stants) when R ≥ t. We may therefore allow R ∈ (0, ∞) arbitrary, and taking the
supremum in R yields (3.1.22), on account of (3.1.5). With (3.1.22) in hand, we then
conclude from (A.0.8) and (3.1.5) that
 
.
 f − gt BMO(Σ,σ) ≤ C3 M1 ( f ; C4 t) for each t ∈ 0, 2 diam Σ . (3.1.35)
 
As a consequence, for each t ∈ 0, 2 diam Σ we have

.
gt BMO(Σ,σ) .
≤ gt − f BMO(Σ,σ) .
+  f BMO(Σ,σ)

.
≤ C3 M1 ( f ; C4 t) +  f BMO(Σ,σ) < +∞, (3.1.36)

from which we conclude that

 
gt ∈ BMO(Σ, σ) for each t ∈ 0, 2 diam Σ . (3.1.37)

Collectively, (3.1.37), (3.1.20), and Lemma 3.1.2 (with β := 1) then show that
.
gt ∈ C α (Σ) ∩ BMO(Σ, σ) for each
  (3.1.38)
t ∈ 0, 2 diam Σ and α ∈ (0, 1).

Passing to limit in (3.1.35) also proves that for each function h ∈ VMO(Σ, σ) we
have

.
lim sup  f − gt BMO(Σ,σ) ≤ C3 lim+ M1 ( f ; C4 t)
t→0+ t→0

≤ C3 lim+ M1 ( f − h; C4 t) + C3 lim+ M1 (h; C4 t)

t→0 t→0

.
≤ C3  f − hBMO(Σ,σ), (3.1.39)

where the last step takes into account (3.1.8) and (A.0.8) written for f − h in place
of f . After taking the infimum over all h ∈ VMO(Σ, σ) in (3.1.39) we ultimately
arrive at the conclusion that
 
.
lim sup  f − gt BMO(Σ,σ) ≤ C3 · dist f , VMO(Σ, σ) , (3.1.40)
t→0+

.
where the distance is measured with respect to  · BMO(Σ,σ) . Having established this,
the desired conclusion readily follows in the case when Σ is unbounded, since in this
.
case  · BMO(Σ,σ) coincides with  · BMO(Σ,σ) .
To finish the proof, there remains to treat the case when Σ is compact. This time,
.
in addition to the semi-norm  · BMO(Σ,σ) , the norm  · BMO(Σ,σ) contains an extra
90 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

term (see (A.0.9)), but the same type of argument applies as soon as we show that
∫ 
 
lim+  ( f − gt ) dσ  = 0. (3.1.41)
t→0 Σ

To justify this, by Hölder’s inequality it suffices to prove that for each function
ψ ∈ L 2 (Σ, σ) we have

G(ψ; t) −→ ψ in L 2 (Σ, σ) as t → 0+, (3.1.42)

where ∫
G(ψ; t)(x) := St (x, y)ψ(y) dσ(y) for each x ∈ Σ. (3.1.43)
Σ
Since the integral kernel of the operator G(·; t) is an approximation to the identity of
the sort described in (3.1.17), the following properties hold:

G(ψ; t) → ψ pointwise on Σ, as t → 0+ if ψ ∈ Lip(Σ), and

  (3.1.44)
|G(ψ; t)| ≤ CM Σ ψ on Σ, for each ψ ∈ L 2 (Σ, σ) and t ∈ 0, 2 diam Σ ,

where M Σ denotes the Hardy-Littlewood maximal operator on Σ. Hence, the claim

in (3.1.42) is implied by (3.1.44), Lebesgue’s Dominated Convergence Theorem, the
boundedness of M Σ on L 2 (Σ, σ) (cf. [133, Corollary 7.6.2]), and the density result
contained in [133, (3.7.22)]. This finishes the proof of Theorem 3.1.3. 

Remark 3.1.4 Our proof of the density result stated in.Theorem 3.1.3 is constructive,
and the sequence {gt }0<t<2 diam Σ ⊆ BMO(Σ, σ) ∩ C α (Σ) defined as in (3.1.18) in
relation to any given function f ∈ BMO(Σ, σ) enjoys additional useful properties.
In particular, it is of interest to point out that, as seen from (3.1.18) and (3.1.17),
 constantsC1, C0 ∈ (0, ∞) depending only on Σ with the property that
there exist two
for each t ∈ 0, 2 diam Σ and each x ∈ Σ we have
 ⨏  ∫  ⨏  
   
gt (x) − f dσ  =  St (x, y) f (y) − f dσ dσ(y)
Δ(x,C0 t) Σ Δ(x,C0 t)
⨏  ⨏ 
 
≤ C1  f (y) − f dσ  dσ(y)
Δ(x,C0 t) Δ(x,C0 t)

≤ C1 M1 ( f ; C0 t), (3.1.45)

where the latter piece of notation has been introduced in (3.1.5). In addition,
∫ 
 
|gt (x) − f (x)| =  St (x, y)( f (y) − f (x)) dσ(y)
Σ
⨏
≤ C1 | f (y) − f (x)| dσ(y), (3.1.46)
Δ(x,C0 t)
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 91

from which we deduce that

lim gt (x) = f (x) at each Lebesgue point x ∈ Σ of f . (3.1.47)

t→0+

Let us also note that, as is apparent from (3.1.18) and (3.1.17),

 
|gt (x)| ≤  f  L ∞ (Σ,σ) at each point x ∈ Σ, for each t ∈ 0, 2 diam Σ , (3.1.48)

and there exists a constant C = C(Σ) ∈ (0, ∞) with the property that
 
|gt | ≤ CM Σ f on Σ, for each t ∈ 0, 2 diam Σ , (3.1.49)

where M Σ denotes the Hardy-Littlewood maximal operator on Σ.

When specialized to bounded sets (in which context the homogeneous and inho-
mogeneous Hölder spaces coincide), Theorem 3.1.3 tells us that:

if Σ ⊆ Rn is a compact Ahlfors regular set and σ := H n−1 Σ, then

for each α ∈ (0, 1) the Sarason space VMO(Σ, σ) is the closure of
(3.1.50)
the inhomogeneous Hölder space C α (Σ) in the John-Nirenberg space
BMO(Σ, σ).
In fact, (3.1.50) may be further improved to the following stronger statement:

if Σ ⊆ Rn is a compact Ahlfors regular set andσ := H n−1 Σ, then

the Sarason space VMO(Σ, σ) is the closure of φ|Σ : φ ∈ Cc∞ (Rn )
in the John-Nirenberg space BMO(Σ, σ); in particular, VMO(Σ, σ) is (3.1.51)
also the closure of Lip(Σ), the space of Lipschitz functions on Σ, in
BMO(Σ, σ).
To justify (3.1.51), thanks to (3.1.50) it suffices to approximate Hölder functions on
Σ arbitrarily well by Lipschitz functions on Σ, in the space BMO(Σ, σ). To fix ideas,
pick a function f ∈ C α (Σ) with α ∈ (0, 1). Extending this to the entire Euclidean
space with preservation of class, then truncate this extension by a smooth function
which is identically one near the compact Σ. Finally, mollify the resulting function
to produce a sequence {Fj } j ∈N ⊆ Cc∞ (Rn ) which converges uniformly on Σ to f .
Then φ j := Fj |Σ for each j ∈ N has the property that {φ j } j ∈N converges to f in
L ∞ (Σ, σ), hence also in BMO(Σ, σ). This establishes (3.1.51).
Theorem 3.1.3 has other consequences of interest, and we discuss some of these
in the next five corollaries.

Corollary 3.1.5 Let Σ be a closed Ahlfors regular subset of Rn and define the
measure σ := H n−1 Σ. Also, recall the approximation to the identity {St }0<t<2 diam Σ
consisting of real-valued
 (σ ⊗ σ)-measurable functions on Σ × Σ satisfying (3.1.17).
For each t ∈ 0, 2 diam Σ use the symbol St to also denote the integral operator on
Σ with integral kernel St (·, ·). Then for each fixed function f ∈ BMO(Σ, σ) one has
the equivalence
92 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

f ∈ VMO(Σ, σ) ⇐⇒ lim+ St f = f in BMO(Σ, σ). (3.1.52)

t→0

Proof This is implicit in the proof of Theorem 3.1.3. 

Here is the second corollary mentioned above, providing alternative proofs of

[97, Proposition 2.22, p. 2611] and [97, Corollary 2.24, p. 2615].

Corollary 3.1.6 Let Σ be a closed Ahlfors regular subset of Rn and set σ := H n−1 Σ.
Also, fix some p ∈ [1, ∞). Then for each function f ∈ BMO(Σ, σ) one has (with the
distance measured in BMO(Σ, σ))
 
dist f , VMO(Σ, σ)
 ⨏ ⨏  1/p
≈ lim sup sup | f (y) − f (z)| dσ(y) dσ(z)
p
r→0+ x ∈Σ Δ(x,r) Δ(x,r)
 ⨏  ⨏  p  1/p
 
≈ lim sup sup  f − f dσ  dσ
r→0+ x ∈Σ Δ(x,r) Δ(x,r)
 
⨏  ⨏  p  1/p
 
≈ lim+ sup f − f dσ  dσ , (3.1.53)
R→0 x ∈Σ, r ∈(0,R) Δ(x,r) Δ(x,r)

where the implicit proportionality constants depend only on p and the environment.
In particular, for each given function f ∈ BMO(Σ, σ) one has the equivalence
⨏ ⨏  p1
 p
 
f ∈ VMO(Σ, σ) ⇐⇒ lim+ sup  f (y) − f dσ  dσ(y) =0
R→0 x ∈Σ and Δ(x,r) Δ(x,r)
r ∈(0,R)
(3.1.54)
for some (or all) p ∈ [1, ∞).
Finally, with the piece of notation introduced in (A.0.7), one has (again, with the
distance is measured in BMO(Σ, σ))
   
lim+ sup  f ∗ (Δ(x, R)) ≈ dist f , VMO(Σ, σ) ,
R→0 x ∈Σ (3.1.55)
uniformly for f ∈ BMO(Σ, σ),

where the implicit proportionality constants depend only on the environment.

Proof Fix an arbitrary function f ∈ BMO(Σ, σ) . Using notation and results from
the proof of Theorem 3.1.3, for each function h ∈ VMO(Σ, σ) we may write
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 93
 
dist f , VMO(Σ, σ) ≤ lim sup  f − gt BMO(Σ,σ)
t→0+

≤ C lim+ M1 ( f ; t) ≤ C lim+ Mp ( f ; t)
t→0 t→0

≤ C lim+ Mp ( f − h; t) + C lim+ Mp (h; t)

t→0 t→0

≤ C f − hBMO(Σ,σ) . (3.1.56)

The first inequality above is a consequence of (3.1.16) and (3.1.38). The second
inequality above comes directly from (3.1.39) when Σ is unbounded, and from
(3.1.39) together with (3.1.41) when Σ is bounded. The next step in (3.1.56) is
implied by Hölder’s inequality. The penultimate step in (3.1.56) is simply the triangle
inequality. The final inequality takes in (3.1.56) follows on account of (3.1.8) and
(A.0.9) written for f − h in place of f .
Taking the infimum over all h ∈ VMO(Σ, σ) in (3.1.56) then leads to the conclu-
sion that
   
dist f , VMO(Σ, σ) ≤ C lim+ Mp ( f ; t) ≤ C · dist f , VMO(Σ, σ) . (3.1.57)
t→0

With this in hand, routine estimates yield all equivalences in (3.1.53). In turn, (3.1.54)
is a direct consequence of (3.1.53). Finally, from definitions we see that

sup  f ∗ (Δ(x, R)) ≈ M1 ( f ; R) (3.1.58)

x ∈Σ

so (3.1.55) follows from this by invoking (3.1.57) (with p = 1). 

Here is the third corollary mentioned earlier, providing useful characterization of

membership to the Sarason class VMO.

Corollary 3.1.7 Let Σ be a closed Ahlfors regular subset of Rn and set σ := H n−1 Σ.
Then for each given function f ∈ BMO(Σ, σ) the following conditions are equivalent:

(1) The function f belongs to the Sarason space VMO(Σ, σ), of functions of van-
ishing mean oscillations on Σ.
(2) For some (or every) smoothness exponent α ∈ (0, 1) there exists a sequence
.
{ f j } j ∈N ⊆ BMO(Σ, σ) ∩ C α (Σ) (3.1.59)

with the property that

lim  f − f j BMO(Σ,σ) = 0. (3.1.60)

j→∞

(3) For some (or every) integrability exponent p ∈ [1, ∞) one has
94 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli
⨏ ⨏  p1
 p
 
lim sup  f (y) − f dσ  dσ(y) = 0. (3.1.61)
R→0+ x ∈Σ and Σ∩B(x,r) Σ∩B(x,r)
r ∈(0,R)

It is important to note that, as a cursory inspection of the proof of Corollary 3.1.7

reveals, a result similar in spirit is valid in arbitrary spaces of homogeneous type.
Proof of Corollary 3.1.7 The fact that (1) implies the strongest form of (2) has been
established in Theorem 3.1.3. To see that the weakest version of (2) implies the
strongest form of (3), fix an arbitrary integrability exponent p ∈ [1, ∞) and assume
the function f ∈ BMO(Σ, σ) has the property that there exists some α ∈ (0, 1) along
.
with a sequence { f j } j ∈N ⊆ BMO(Σ, σ) ∩ C α (Σ) satisfying (3.1.60). Fix an arbitrary
ε > 0. Then there exists jε ∈ N for which
 
 f − fj 
ε BMO(Σ,σ) < ε. (3.1.62)
 
Also, for each x ∈ Σ and r ∈ 0, 2 diam Σ we have
⨏ ⨏  p1
 p
 
 f jε (y) − f jε dσ  dσ(y)
Σ∩B(x,r) Σ∩B(x,r)

⨏ ⨏  p1
 
≤  f j (y) − f j (z) p dσ(y) dσ(z)
ε ε
Σ∩B(x,r) Σ∩B(x,r)

⨏ ⨏  p1
 
≤  f jε C.α (Σ) |y − z| αp
dσ(y) dσ(z)
Σ∩B(x,r) Σ∩B(x,r)
 
≤  f jε C.α (Σ) (2r)α . (3.1.63)

With the piece of notation introduced in (3.1.5), this implies

   
Mp ( f jε ; R) ≤  f jε C.α (Σ) (2R)α for each R ∈ 0, 2 diam Σ . (3.1.64)

In particular,
lim Mp ( f jε ; R) = 0. (3.1.65)
R→0+

while from (3.1.5) and (3.1.62) we deduce that

   
Mp ( f − f jε ; R) ≤  f − f jε BMO(Σ,σ) < ε for each R ∈ 0, 2 diam Σ . (3.1.66)

Consequently,

lim sup Mp ( f ; R) ≤ lim sup Mp ( f jε ; R) + lim sup Mp ( f jε ; R) < ε. (3.1.67)

R→0+ R→0+ R→0+
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 95

In view of the arbitrariness of ε > 0, we conclude from (3.1.67) that (3.1.61) holds.
This concludes the proof of the implication (2) ⇒ (3). Finally, that the weakest
version of (3) implies (1) is seen from (3.1.54). 

Continuing out series of corollaries to Theorem 3.1.3, we next show that the
collection of essentially bounded functions with vanishing mean oscillations is a
stable environment.

Corollary 3.1.8 Let Σ be a closed Ahlfors regular subset of Rn and set σ := H n−1 Σ.
Then there exists C ∈ (0, ∞) which depends only on the Ahlfors regularity constants
of Σ with the property that for each two functions f , g ∈ L ∞ (Σ, σ) one has
   
dist f g, VMO(Σ, σ) ≤ C f  L ∞ (Σ,σ) · dist g, VMO(Σ, σ)
 
+ Cg L ∞ (Σ,σ) · dist f , VMO(Σ, σ) , (3.1.68)

with all distances measured in BMO(Σ, σ).

As a consequence, L ∞ (Σ, σ)∩VMO(Σ, σ) is a closed linear subspace of L ∞ (Σ, σ),
closed under products, hence a closed ∗-subalgebra of the C ∗ -algebra L ∞ (Σ, σ).

Proof Pick an arbitrary surface ball Δ ⊆ Σ. Abbreviate

⨏ ⨏ ⨏
fΔ := f dσ, gΔ := g dσ, ( f g)Δ := f g dσ, (3.1.69)
Δ Δ Δ

then estimate
⨏ ⨏
   
 f g − ( f g)Δ  dσ ≤ 2  f g − fΔ gΔ  dσ (3.1.70)
Δ Δ
⨏ ⨏
   
≤ 2 f  L ∞ (Σ,σ)
g − gΔ  dσ + 2g L ∞ (Σ,σ)  f − fΔ  dσ.
Δ Δ

Then the estimate claimed in (3.1.68) follows in view of this and (3.1.53). In turn,
(3.1.68) readily implies that L ∞ (Σ, σ) ∩ VMO(Σ, σ) is closed under pointwise prod-
uct.
Assume next that { f j } j ∈N is a sequence in L ∞ (Σ, σ) ∩ VMO(Σ, σ) that converges
in L ∞ (Σ, σ) to some f ∈ L ∞ (Σ, σ). Since  f − f j BMO(Σ,σ) ≤ 2 f − f j  L ∞ (Σ,σ) for
each j ∈ N, it follows that { f j } j ∈N ⊆ VMO(Σ, σ) converges to f in BMO(Σ, σ).
This places the limit function f in VMO(Σ, σ) since the latter is a closed sub-
space of BMO(Σ, σ). Hence, f ∈ L ∞ (Σ, σ) ∩ VMO(Σ, σ) ultimately proving
that L ∞ (Σ, σ) ∩ VMO(Σ, σ) is a closed linear subspace of L ∞ (Σ, σ). Given that
L ∞ (Σ, σ) ∩ VMO(Σ, σ) is also stable under complex conjugation, it follows that this
is indeed a closed ∗-subalgebra of the C ∗ -algebra L ∞ (Σ, σ). 

Below we single out a useful consequence of Corollary 3.1.8 which, in particular,

applies to matrix-valued functions with entries in L ∞ ∩ VMO.
96 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

Corollary 3.1.9 Let Σ be a closed Ahlfors regular subset

 of R and set σ := H  Σ.
n n−1

∗ ∞
Suppose A is a finite-dimensional C -algebra. Then L (Σ, σ) ∩ VMO(Σ, σ) ⊗ A
 
is a C ∗ -algebra, and if φ ∈ L ∞ (Σ, σ) ∩ VMO(Σ, σ) ⊗ A , regarded as an element
in the C ∗ -algebra ∞ −1 ∞
 L (Σ, σ) ⊗ A , has an inverse φ in L (Σ, σ) ⊗ A , then actually
φ−1 belongs to L ∞ (Σ, σ) ∩ VMO(Σ, σ) ⊗ A .
Proof This is a consequence of Corollary 3.1.8 and an abstract general result to the
effect that if A is a C ∗ -algebra with unit 1 and B a C ∗ -subalgebra, containing 1, then
an element φ ∈ B is invertible in B if and only if it is invertible in A. 
We conclude this section by circling back to Theorem 3.1.3 and proving a more
nuanced version of the density result stated there for unimodular functions (i.e.,
vector-valued functions of modulus one).
Theorem 3.1.10 Let Σ be a closed Ahlfors regular subset of Rn and abbreviate
σ := H n−1 Σ. Also, fix an exponent α ∈ (0, 1) along with an integer N ∈ N. Then
 N
for each function f ∈ VMO(Σ, σ) which is unimodular (i.e., satisfying | f | = 1
at σ-a.e. point on Σ) there exists
 N
a sequence {φ} j ∈N ⊆ C α (Σ) of unimodular functions
(3.1.71)
(i.e., satisfying |φ j | = 1 at σ-a.e. point on Σ, for each j ∈ N)
 N
which converges to f in BMO(Σ, σ) .
Proof We shall use notation and results from the proof of Theorem 3.1.3 with
 N  
impunity. Fix an arbitrary function f ∈ VMO(Σ, σ) and for each t ∈ 0, 2 diam Σ
 . N
define gt as in (3.1.18). Then each gt belongs to C α (Σ) and since | f | = 1 it
follows from (3.1.48) that
 
|gt (x)| ≤ 1 for each x ∈ Σ and t ∈ 0, 2 diam Σ . (3.1.72)
 α N
As a consequence,
  gt belongs to C (Σ) . In addition, for each x ∈ Σ and
each
each t ∈ 0, 2 diam Σ we may write
     
 |gt (x)| − 1 ≤  |gt (x)| − | fΔ(x,C t) |  +  | fΔ(x,C t) | − 1
0 0
⨏
   
≤ gt (x) − fΔ(x,C0 t)  + 1 − | fΔ(x,C t) |  dσ
0
Δ(x,C0 t)
⨏
 
≤ C1 M1 ( f ; C0 t) +  | f (x)| − | fΔ(x,C t) |  dσ(x)
0
Δ(x,C0 t)
⨏
 
≤ C1 M1 ( f ; C0 t) +  f (x) − fΔ(x,C t)  dσ(x)
0
Δ(x,C0 t)

≤ C2 M1 ( f ; C0 t), (3.1.73)
3.1 Functions of Vanishing Mean Oscillations on Measure Metric Spaces 97

thanks to the reverse triangle inequality, Remark 3.1.4, the fact that f is unimodular,
and the definition made in (3.1.5). From (3.1.73) we then conclude that
   
sup  |gt (x)| − 1 ≤ C2 M1 ( f ; C0 t) for each t ∈ 0, 2 diam Σ . (3.1.74)
x ∈Σ
 N
Since f ∈ VMO(Σ, σ) , it follows from Corollary 3.1.6 that

lim M1 ( f ; t) = 0, (3.1.75)
t→0+
 
hence there exists t∗ ∈ 0, 2 diam Σ with the property that
 
sup  |gt (x)| − 1 ≤ 1 whenever t ∈ (0, t∗ ].
2 (3.1.76)
x ∈Σ

In turn, this forces

|gt (x)| ≥ 1
2 for each t ∈ (0, t∗ ] and x ∈ Σ. (3.1.77)

In particular, for each t ∈ (0, t∗ ] the function gt /|gt | is well defined, unimodular, and
 N
belongs to the Hölder space C α (Σ) . To proceed, for each t ∈ (0, t∗ ] write
gt gt  
f− = ( f − gt ) + |gt | − 1 on Σ, (3.1.78)
|gt | |gt |
and use this to estimate
 gt   g   
 .  t
f −  . ≤  f − gt  [BMO(Σ,σ)] N
+  |g t | − 1  .
|gt | [BMO(Σ,σ)] N |gt | [BMO(Σ,σ)] N
 g   
 t
≤ C3 M1 ( f ; C4 t) + 2 |gt | − 1  ∞
|gt | [L (Σ,σ)] N
 
≤ C3 M1 ( f ; C4 t) + 2 |gt | − 1 L ∞ (Σ,σ)

≤ C5 M1 ( f ; C6 t), (3.1.79)

where we have relied on (3.1.35) and (3.1.74). In concert with (3.1.75), this proves
that the sequence {φ j } j ∈N , with φ j := gt∗ /j /|gt∗ /j | for each j ∈ N, consists of
 N  N
unimodular functions in C α (Σ) , and converges to f in BMO(Σ, σ) if Σ is
unbounded (cf. (A.0.9)).
.
Finally, in the case when Σ is bounded, in addition to the semi-norm  · BMO(Σ,σ) ,
the norm  · BMO(Σ,σ) contains an extra term (see (A.0.9)), so we also need to check
that ∫  gt  

lim+  f− dσ  = 0 (3.1.80)
t→0 Σ |gt |
98 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

holds in such a scenario. This, however, is a consequence of Lebesgue’s Dominated

Convergence Theorem (bearing in mind (3.1.77), (3.1.72), (3.1.47), and the fact that
f is unimodular). 

3.2 A New Brand of Hölder Spaces

We have discussed the scale of (homogeneous and inhomogeneous) Hölder spaces

in the general setting of quasi-metric spaces in [133, § 7.3]. Here the goal is to
introduce and study a sub-scale obtained by imposing the condition that the local
Hölder seminorm vanishes as the scale (at which this is considered) approaches zero.
To set the stage, let Σ be an arbitrary (nonempty) subset of Rn , and fix some
γ ∈ (0, ∞). We are interested in the class of real-valued functions f defined on Σ
satisfying the following property3:

for each ε > 0 there exists δ > 0 such that

| f (x) − f (x )| ≤ ε|x − x  |γ whenever (3.2.1)
x, x ∈ Σ satisfy |x − x| < δ.

Specifically, introduce4
 
V γ (Σ) := f ∈ C 0 (Σ) : f satisfies (3.2.1) . (3.2.2)

It may be check without difficulty that V γ1 (Σ) ⊆ V γ2 (Σ) whenever γ1 > γ2 > 0, and
that .
C β (Σ) ⊆ V γ (Σ) for each β > γ. (3.2.3)
It is also apparent from definitions that
   
γ
V γ (Σ) = f ∈ Cloc (Σ) : lim+ sup  f C.γ (B(x,r)∩Σ) = 0 . (3.2.4)
r→0 x ∈Σ

We shall use (3.2.2) to define what we call the homogeneous “vanishing” Hölder
.γ
space Cvan (Σ), of order γ on the set Σ, as follows:
.
γ
.
Cvan (Σ) := C γ (Σ) ∩ V γ (Σ)
 .   
= f ∈ C γ (Σ) : lim+ sup  f C.γ (B(x,r)∩Σ) = 0 . (3.2.5)
r→0 x ∈Σ

It is then clear from this definition and (3.2.3) that

. . .
C γ (Σ) ∩ C β (Σ) ⊆ Cvan
γ
(Σ) for each β > γ. (3.2.6)

3 think of this as a Hölder-styled version of uniform continuity

4 at least formally, V 0 (Σ) corresponds to UC (Σ), the space of uniformly continuous functions on Σ
3.2 A New Brand of Hölder Spaces 99

Lemma 3.2.1 Let Σ be an arbitrary nonempty

 . subset ofRn , and fix some γ ∈ (0, ∞).
.γ
Then Cvan (Σ) is a closed subspace of C (Σ),  · C.γ (Σ) .
γ

.
γ (Σ) equipped with the semi-norm inher-
Henceforth we shall always consider Cvan
.γ
ited from the space C (Σ), i.e.,  · C.γ (Σ) .
.
Proof of Lemma 3.2.1 Assume f ∈ C γ (Σ) is such that there exists a sequence
.γ
{ f j } j ∈N ⊆ Cvan (Σ) with the property that  f − f j C.γ (Σ) → 0 as j → ∞. Also,
fix an arbitrary threshold ε > 0. Then we may find an integer jε ∈ N for which
 f − f jε C.γ (Σ) < ε/2. In addition, since f jε ∈ V γ (Σ), there exists δ > 0 such that
| f jε (x) − f jε (x )| ≤ (ε/2)|x − x  |γ whenever x, x  ∈ Σ satisfy |x − x  | < δ.
For each pair of points x, x  ∈ Σ satisfying |x − x  | < δ we may then estimate
 
| f (x) − f (x )| ≤ ( f − f jε )(x) − ( f − f jε )(x ) + | f jε (x) − f jε (x )|

≤  f − f jε C.γ (Σ) · |x − x  |γ + (ε/2)|x − x  |γ

≤ ε|x − x  |γ . (3.2.7)
. .
This shows that f ∈ V γ (Σ). Hence, ultimately, f ∈ C γ (Σ) ∩ V γ (Σ) = Cvan
γ (Σ), as

wanted. 
Retaining the assumptions that Σ is an arbitrary nonempty subset of Rn and
that γ ∈ (0, ∞) is an arbitrary number, let us also introduce the inhomogeneous
“vanishing” Hölder space
γ
Cvan (Σ) := C γ (Σ) ∩ V γ (Σ)
   
= f ∈ C γ (Σ) : lim+ sup  f C.γ (B(x,r)∩Σ) = 0 . (3.2.8)
r→0 x ∈Σ

The same argument as in the proof of Lemma 3.2.1 then shows that
γ (Σ) is a closed subspace of C γ (Σ), so it becomes a
Cvan
(3.2.9)
Banach space when equipped with the norm  · C γ (Σ) .

It is also clear from the definition in (3.2.8) and (3.2.3) that

C γ (Σ) ∩ C β (Σ) ⊆ Cvan

γ
(Σ) for each β > γ. (3.2.10)

Lastly, we wish to note that

γ
 .γ 
Cvan (Σ) = f ∈ Cvan (Σ) : f bounded . (3.2.11)

Much like the space of essentially bounded functions (in the context of a mea-
surable space), the ordinary Hölder spaces are notoriously difficult when it comes
100 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

to issues pertaining to density. This being said, for the brands of Hölder spaces
introduced above we have the following remarkable density result.
Theorem 3.2.2 Let Σ be a closed Ahlfors regular subset of Rn and set σ := H n−1 Σ.
Also, fix an exponent γ ∈ (0, 1). Then
. . .
C γ (Σ) ∩ C β (Σ) ⊆ Cvan
γ
(Σ) densely for each β ∈ (γ, 1). (3.2.12)

Equivalently, . . .
γ (Σ) is the closure of C γ (Σ) ∩ C β (Σ) in
Cvan
. (3.2.13)
the space C γ (Σ), for each β ∈ (γ, 1).
Furthermore,
γ (Σ) is the closure of C γ (Σ) ∩ C β (Σ) in
Cvan
(3.2.14)
the space C γ (Σ), for each β ∈ (γ, 1).
. . .
As a by-product, we see that the closure of C γ (Σ) ∩ C β (Σ) in the space C γ (Σ)
is independent of the choice of the exponent β ∈ (γ, 1), and so is the closure of
C γ (Σ) ∩ C β (Σ) in C γ (Σ).
Let us also observe that, as is clear from (3.2.14), in the context of the above
theorem the following density result holds:

if the set Σ is bounded, then for each β ∈ (γ, 1) the

γ (Σ) is the closure of C β (Σ) in C γ (Σ). (3.2.15)
space Cvan
γ (Σ) is the answer to the question of providing an intrinsic description of
Hence, Cvan
the closure of C β (Σ) in the space C γ (Σ), for each given exponent β ∈ (γ, 1), when
Σ is a bounded set.
We now turn to the task of proving Theorem 3.2.2.
Proof of Theorem 3.2.2 We shall adapt the argument used in the proof of Theo-
rem 3.1.3. To set the stage, for each function f ∈ Lloc 1 (Σ, σ) define

⨏  ⨏ 
1  
Hγ ( f ; R) := sup sup γ f − f dσ  dσ for all R ∈ (0, ∞),
x ∈Σ r ∈(0,R) r Δ(x,r) Δ(x,r)

where Δ(x, r) := Σ ∩ B(x, r) for each x ∈ Σ and r ∈ (0, ∞).

(3.2.16)
As is apparent from definitions,
.
sup Hγ ( f ; R) ≤ 2γ  f C.γ (Σ) for each f ∈ C γ (Σ), (3.2.17)
R>0

and .
γ
lim+ Hγ ( f ; R) = 0 for each f ∈ Cvan (Σ). (3.2.18)
R→0

To proceed, bring back the approximation to the identity used in the proof of
Theorem 3.1.3, i.e., a family of (σ ⊗ σ)-measurable functions St : Σ × Σ → R
3.2 A New Brand of Hölder Spaces 101
 
indexed by a parameter t ∈ 0, 2 diam Σ satisfying the properties listed in (3.1.17)
.
for some finite constant C0 ≥ 2. Next, let us fix an arbitrary function f ∈ C γ (Σ),
and for each t ∈ (0, ∞) define
∫
gt (x) := St (x, y) f (y) dσ(y) for each x ∈ Σ. (3.2.19)
Σ

Bearing in mind the fact that  Σ is an Ahlfors regular set, much as in (3.1.19) we
conclude that for each t ∈ 0, 2 diam Σ and x, x  ∈ Σ with |x − x  | < C0 t we have
⨏  ⨏ 
 
|gt (x) − gt (x )| ≤ C t −1 |x − x  | f − f dσ  dσ. (3.2.20)
Δ(x,2C0 t) Δ(x,2C0 t)

In view of the definition made in (3.2.16), this implies that there exist C1, C2 ∈ (1, ∞)
such that
|gt (x) − gt (x )| ≤ C1 t −1+γ |x − x  | · Hγ ( f ; C2 t) whenever
(3.2.21)
t ∈ (0, ∞) and x, x  ∈ Σ satisfy |x − x  | < C0 t.
 
As a consequence, for each t ∈ 0, 2 diam Σ we have

|gt (x) − gt (x )|

sup ≤ C1 t −1+γ  f C.γ (Σ) < +∞. (3.2.22)
x,x  ∈Σ |x − x  |
0< |x−x  |<C0 t

We now claim that there exist two constants C3, C4 ∈ (0, ∞), independent of f ,
such that
 
sup Hγ ( f − gt ; R) ≤ C3 Hγ ( f ; C4 t) for each t ∈ 0, 2 diam Σ . (3.2.23)
R>0
 
To justify this inequality, fix some number t ∈ 0, 2 diam Σ along with an arbitrary
point x ∈ Σ and a radius R ∈ (0, ∞). In the regime 0 < R < t, we estimate
⨏  ⨏ 
1  
( f − gt ) − ( f − gt ) dσ  dσ ≤ I + II, (3.2.24)
Rγ Δ(x,R) Δ(x,R)

where ⨏  ⨏ 
1  
I := f − f dσ  dσ, (3.2.25)
Rγ Δ(x,R) Δ(x,R)

and ⨏  ⨏ 
1  
II := gt − gt dσ  dσ. (3.2.26)
Rγ Δ(x,R) Δ(x,R)

From (3.2.16) and the fact that R ∈ (0, t) is arbitrary it follows that

I ≤ Hγ ( f ; t). (3.2.27)
102 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

Since |y − z| < 2R < C0 t for each y, z ∈ Δ(x, R), thanks to the fact that C0 ≥ 2 and
that we are currently assuming R < t, we also have
⨏ ⨏
1
II ≤ γ |gt (y) − gt (z)| dσ(y) dσ(z)
R Δ(x,R) Δ(x,R)
⨏ ⨏
t −1+γ
≤ C1 Hγ ( f ; C2 t) |y − z| dσ(y) dσ(z)
Rγ Δ(x,R) Δ(x,R)

t −1+γ
≤ C1 Hγ ( f ; C2 t) 2R ≤ C1 Hγ ( f ; C2 t). (3.2.28)
Rγ
The last step above uses the observation that

t −1+γ
R = (t −1 R)1−γ < 1, (3.2.29)
Rγ
given that we are presently working in the regime 0 < R < t (hence t −1 R < 1), and
1 − γ > 0. Bearing in mind that Hγ ( f ; ·) is nondecreasing, the estimates on I, II
above lead to the following conclusion:
⨏  ⨏ 
1  
0 < R < t =⇒ γ ( f − gt ) − ( f − gt ) dσ  dσ ≤ C1 Hγ ( f ; C2 t).
R Δ(x,R) Δ(x,R)
(3.2.30)
Thus, as far as (3.2.23) is concerned, there remains to consider the case when R ≥ t.
In such a scenario, start from the realization that Δ(z, t) : z ∈ Δ(x, R) is a family
of surface balls of same (finite) radius covering the bounded set Δ(x, R). Vitali’s
Covering Lemma (cf. [133, Lemma 7.5.7], or [133, Lemma 7.5.8]) then guarantees
that there exist some dilation parameter λ ∈ (1, ∞) and an at most countable family
of points {z j } j ∈J ⊆ Δ(x, R) such that

{Δ(z j , t)} j ∈J are mutually disjoint, and Δ(x, R) ⊆ Δ(z j , λt). (3.2.31)
j ∈J

Availing ourselves of these properties, as well as (3.1.17), the definition in (3.2.16),

and keeping in mind that (t/R)γ ≤ 1 (given that γ > 0 and we are presently assuming
R ≥ t), we may now estimate
3.2 A New Brand of Hölder Spaces 103
⨏  ⨏ 
1  
( f − gt ) − ( f − gt ) dσ  dσ
Rγ Δ(x,R) Δ(x,R)
⨏
2
≤ | f (y) − gt (y)| dσ(y)
Rγ Δ(x,R)
∫  ⨏
  
2 
≤ γ  f (y) − f dσ
R · σ(Δ(x, R)) j ∈J Δ(z j ,λt)  Δ(z j ,(λ+C0 )t)

∫ ⨏ 
  

− St (y, z) f (z) − f dσ dσ(z) dσ(y)
Σ Δ(z j ,(λ+C0 )t) 
 ⨏  ⨏ 
C  σ Δ(z j , (λ + C0 )t)  
≤ γ f − f dσ  dσ
R j ∈J σ(Δ(x, R)) Δ(z j ,(λ+C0 )t) Δ(z j ,(λ+C0 )t)
 
γ     σ Δ(z j , t) 
≤ C(t/R) Hγ f ; (λ + C0 )t  
j ∈J
σ Δ(x, R)
 
  σ j ∈J Δ(z j , t)
≤ C Hγ f ; (λ + C0 )t  
σ Δ(x, R)
 
≤ C Hγ f ; (λ + C0 )t , (3.2.32)

for some geometric constant C ∈ (0, ∞), independent of f , x, and R.

In summary, the implication in (3.2.30) also holds (with possibly different con-
stants) when R ≥ t. We may therefore allow R ∈ (0, ∞) arbitrary, and taking the
supremum in R yields (3.2.23), on account of (3.2.16). With (3.2.23) in hand, we
then conclude from [133, Proposition 7.4.9] (bearing in mind [133, Lemma 3.6.4]
with s := n − 1) that

 f − gt C.γ (Σ) ≤ C supR>0 Hγ ( f − gt ; R) ≤ C Hγ ( f ; C4 t)

  (3.2.33)
for each t ∈ 0, 2 diam Σ .
 
As a consequence, for each t ∈ 0, 2 diam Σ we have

gt C.γ (Σ) ≤ gt − f C.γ (Σ) +  f C.γ (Σ)

≤ C3 Hγ ( f ; C4 t) +  f C.γ (Σ) < +∞, (3.2.34)

from which we conclude that

.  
gt ∈ C γ (Σ) for each t ∈ 0, 2 diam Σ . (3.2.35)

Focusing now directly on (3.2.12), fix β ∈ (γ, 1) and assume the function f

.γ
actually belongs to the homogeneous vanishing Hölder space Cvan (Σ). Also, fix
104 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

t ∈ (0, ∞) and pick two arbitrary points x, x  ∈ Σ. On the one hand, if |x − x  | < C0 t
then from (3.2.21) we see that

|gt (x) − gt (x )| ≤ C1 t −1+γ Hγ ( f ; C2 t) · |x − x  |

= C1 t −1+γ Hγ ( f ; C2 t) · |x − x  | 1−β |x − x  | β

≤ C1 t −1+γ Hγ ( f ; C2 t) · (C0 t)1−β |x − x  | β . (3.2.36)

On the other hand, if |x − x  | ≥ C0 t then (3.2.35) implies

|gt (x) − gt (x )| ≤ gt C.γ (Σ) |x − x  |γ ≤ gt C.γ (Σ) (C0 t)γ−β |x − x  | β (3.2.37)

Together, (3.2.36)-(3.2.37) and (3.2.35) prove that

. .  
gt ∈ C γ (Σ) ∩ C β (Σ) for each t ∈ 0, 2 diam Σ . (3.2.38)

Passing to limit in (3.2.33) also proves that

lim sup  f − gt C.γ (Σ) ≤ C3 lim+ Hγ ( f ; C4 t) = 0, (3.2.39)

t→0+ t→0

where the last inequality takes into account (3.2.18). Having established this, (3.2.12)
follows from (3.2.6), (3.2.38), and (3.2.39).
In fact, a stronger conclusion holds. This hinges on the observation that if the
function f ∈ UC(Σ) (i.e., f is uniformly continuous on Σ) then gt defined in (3.2.19)
also satisfies
lim+ sup | f (x) − gt (x)| = 0. (3.2.40)
t→0 x ∈Σ

Indeed, for each t ∈ (0, ∞) and x ∈ Σ we have

∫   
 
| f (x) − gt (x)| =  St (x, y) f (x) − f (y) dσ(y)
Σ
⨏
 
≤C  f (x) − f (y) dσ(y), (3.2.41)
Δ(x,C0 t)

so (3.2.40) follows in view of the fact that f is uniformly continuous. Consequently,

since any Hölder function is uniformly continuous, (3.2.39) may be improved to

lim  f − gt C γ (Σ) = 0. (3.2.42)

t→0+

In particular, the conclusion in (3.2.14) readily follows from this. The proof of
Theorem 3.2.2 is therefore complete. 

Much like the John-Nirenberg space BMO(Σ, σ) may be regarded as the end-
.
point space of the Hölder scale C γ (Σ) with γ ∈ (0, 1), we may view the Sarason
.γ
space VMO(Σ, σ) as the end-point space of our vanishing Hölder scale Cvan (Σ)
3.2 A New Brand of Hölder Spaces 105

with γ ∈ (0, 1). This analogy is strengthened by the following result, which mirrors
Corollary 3.1.6.

Corollary 3.2.3 Let Σ be a closed Ahlfors regular subset of Rn and . abbreviate

σ := H n−1 Σ. Also, fix some γ ∈ (0, 1). Then for each function f ∈ C γ (Σ) one has
 ⨏ ⨏
.γ  
1  
f ∈ Cvan (Σ) ⇐⇒ lim+ sup sup γ f − f dσ  dσ = 0,
R→0 x ∈Σ r ∈(0,R) r Δ(x,r) Δ(x,r)
(3.2.43)
while for each function f ∈ C γ (Σ) one has
 ⨏ ⨏
 
γ 1  
f ∈ Cvan (Σ) ⇐⇒ lim+ sup sup γ f − f dσ  dσ = 0.
R→0 x ∈Σ r ∈(0,R) r Δ(x,r) Δ(x,r)

.γ (3.2.44)
.
In addition, for each given function f ∈ C (Σ) one has (with the distance mea-
sured in C γ (Σ))
 ⨏ ⨏ 
 .γ  1
dist f , Cvan (Σ) ≈ lim sup sup γ | f (y) − f (z)| dσ(y) dσ(z)
r→0+ x ∈Σ r Δ(x,r) Δ(x,r)
 ⨏  ⨏  
1  
≈ lim sup γ sup f − f dσ  dσ
r→0+ r x ∈Σ Δ(x,r) Δ(x,r)
 ⨏ ⨏ 
1    
 
≈ lim+ sup f − f dσ  dσ , (3.2.45)
R→0 x ∈Σ, r ∈(0,R) r γ Δ(x,r) Δ(x,r)

where the implicit proportionality constants depend only on the environment. Finally,
a similar result to the one described in (3.2.45) is true for functions f ∈ C γ (Σ) (now
measuring the distance in C γ (Σ)).
.
Proof Fix an arbitrary function f ∈ C γ (Σ). The left-to-right implication in (3.2.43)
is implied by (3.2.16) and (3.2.18), while the right-to-left implication in (3.2.43) is
seen from (3.2.39), (3.2.38), (3.2.6), and Lemma 3.2.1. The equivalence in (3.2.44)
is justified in a similar manner, now also keeping in mind (3.2.40).
.
As regards (3.2.45), pick a function f ∈ C γ (Σ). Using notation employed in the
.γ
proof of Theorem 3.2.2, for each h ∈ Cvan (Σ) we may write
 .γ 
dist f , Cvan (Σ) ≤ lim sup  f − gt C.γ (Σ) ≤ C lim+ Hγ ( f ; t)
t→0+ t→0

≤ C lim+ Hγ ( f − h; t) + C lim+ Hγ (h; t)

t→0 t→0

≤ C f − hC.γ (Σ) . (3.2.46)

The first step above is a consequence of (3.2.38) and (3.2.6). The second step above
is directly implied by the inequality (3.2.39). The penultimate step in (3.2.46) is
106 3 Functions of Vanishing Mean Oscillations and Vanishing Hölder Moduli

simply the triangle inequality. The final step takes in (3.2.46) follows on account of
(3.2.17) (written for f − h in place of f ), and (3.2.18) (written for h in place of f ).
.γ
Taking the infimum over all h ∈ Cvan (Σ) in (3.2.46) then leads to the conclusion
that
 .γ   .γ 
dist f , Cvan (Σ) ≤ C lim+ Hγ ( f ; t) ≤ C · dist f , Cvan (Σ) . (3.2.47)
t→0

With this in hand, routine estimates yield all equivalences in (3.2.45). Finally, the
task of establishing equivalences analogous to those recorded in (3.2.45) in the case
when f ∈ C γ (Σ) is dealt with similarly. 
Chapter 4
Hardy Spaces on Ahlfors Regular Sets

The pioneering work of R. Coifman and G. Weiss ([35], [36]) has fundamentally
reconfigured the classical theory of Hardy spaces by fully recognizing the usefulness
and versatility of a brand of Hardy spaces which places minimal regularity and
structural demands on the underlying space. Here we are concerned with Hardy
spaces on Ahlfors regular subsets of the Euclidean ambient and, by further building
on the work in [9], consider topics such as the Fefferman-Stein grand maximal
function, Lebesgue-based and Lorentz-based Hardy spaces, real interpolation, atoms
and molecules, duality, weak-∗ convergence, and the compatibility of various duality
pairings.

4.1 The Fefferman-Stein Grand Maximal Function

Given a closed set Σ ⊆ Rn which is Ahlfors regular, abbreviate σ := H n−1 Σ, and
fix γ ∈ (0, 1). In this context, for each x ∈ Σ define Tγ (x) to be
 the collection
 of all
functions ψ ∈ Lipc (Σ) with the property that there exists r ∈ 0, 2 diam Σ such that

1
supp ψ ⊆ Σ ∩ B(x, r), sup |ψ(z)| ≤  ,
z ∈Σ σ Σ ∩ B(x, r)
(4.1.1)
|ψ(z1 ) − ψ(z2 )| 1
and ψC.γ (Σ) := sup γ
≤  .
z 1 , z 2 ∈Σ |z1 − z2 | r γ · σ Σ ∩ B(x, r)
z1 z2

That is, for each x ∈ Σ, define

  
Tγ (x) := ψ ∈ Lipc (Σ) : (4.1.1) holds for some r ∈ 0, 2 diam Σ . (4.1.2)

Regarding this set we note the following useful fact:

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 107
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_4
108 4 Hardy Spaces on Ahlfors Regular Sets

given φ ∈ Lipc (Σ), not identically

 zero, and some point x ∈ Σ,
then for each R ∈ 0, 2 diam Σ  such that supp φ ⊆ B(x,
 R) ∩ Σ
it follows that the function φ Cφ,R · σ(B(x, R) ∩ Σ) belongs (4.1.3)
to Tγ (x), provided
 the normalization constant is defined
 as
Cφ,R := max (RφLip(Σ) )γ (2 · supΣ |φ|)1−γ, supΣ |φ| .

Indeed, if φ ∈ Lipc (Σ) then for each x, y ∈ Σ and each R > 0 we have
 
|φ(x) − φ(y)| ≤ R−1 RφLip(Σ) |x − y| and |φ(x) − φ(y)| ≤ 2 · sup |φ|. (4.1.4)
Σ

Raising the first inequality in (4.1.4) to the power γ, the second one to the power
1 − γ, then multiplying the resulting inequalities eventually yields

 γ 1−γ Cφ,R · σ(B(x, R) ∩ Σ)

φC.γ (Σ) ≤ R−γ RφLip(Σ) 2 · sup |φ| ≤   . (4.1.5)
Σ Rγ · σ Σ ∩ B(x, R)

Granted this estimate, it is easy to check that if φ, x, R are as in (4.1.3) then the
function defined as ψ := φ/[Cφ,R · σ(B(x, R) ∩ Σ)] satisfies the conditions in (4.1.1)
with r := R. This proves (4.1.3).  
Pressing on, for any distribution f ∈ Lipc (Σ) define the Fefferman-Stein
grand maximal function fγ of f as the function pointwise defined on Σ according
to
fγ (x) := sup f , ψ , ∀x ∈ Σ, (4.1.6)
ψ ∈ Tγ (x)
 
where ·, · currently stands for the duality paring between Lipc (Σ) and Lipc (Σ)
(recall that this is compatible with the ordinary integral pairing on Σ, with respect
to the measure σ, when such a pairing is meaningful; see [133, Proposition 4.1.4]
and the subsequent comment). From (4.1.3) and (4.1.6) we see that with C ∈ (0, ∞)
denoting the upper Ahlfors regularity constant of the set Σ,
 
given any f ∈ Lipc (Σ) and γ ∈ (0, 1), for each φ ∈ Lipc (Σ) we have
 
f , φ ≤ C Rn−1 · max (RφLip(Σ) )γ (2 · supΣ |φ|)1−γ, supΣ |φ| · fγ (x) (4.1.7)
whenever x ∈ Σ and R > 0 are selected such that supp φ ⊆ B(x, R) ∩ Σ.

It is useful to observe that other related variants of the inequality in (4.1.7) are valid
for radii which are small (relative to the diameter of Σ). Specifically, if
  1/(n−1)
λΣ := 2 CΣ /cΣ , (4.1.8)

where CΣ , cΣ are, respectively, the upper and lower Ahlfors regularity constants of Σ
(cf. [133, Definition 5.9.1]), then we have the following companion result for (4.1.7):
4.1 The Fefferman-Stein Grand Maximal Function 109
 
given any f ∈ Lipc (Σ) and γ ∈ (0, 1), for each φ ∈ Lipc (Σ) we have
 γ   1−γ 
f , φ ≤ C Rn−1+γ · φLip(Σ) · sup |φ| · fγ (x)
Σ
(4.1.9)
≤ C R · φLip(Σ) ·
n
fγ (x)
 
if x ∈ Σ and R ∈ 0, λΣ−1 · 2 diam Σ are such that supp φ ⊆ B(x, R) ∩ Σ,

where now C ∈ (0, ∞) is allowed to depend on n, γ, and the Ahlfors regularity

 of Σ. Indeed,
constants  in the context of (4.1.9),
 the choice of λΣ in (4.1.8) ensures
that σ B(x, λΣ R) ∩ Σ is strictly bigger that σ B(x, R) ∩ Σ . In particular, there exists
a point    
x∗ ∈ B(x, λΣ R) ∩ Σ \ B(x, R) ∩ Σ . (4.1.10)
As such, for each y ∈ supp φ ⊆ B(x, R) ∩ Σ we may estimate

|φ(y)| = |φ(y) − φ(x∗ )| ≤ |y − x∗ |φLip(Σ) ≤ 2λΣ RφLip(Σ), (4.1.11)

hence
sup |φ| ≤ 2λΣ RφLip(Σ), (4.1.12)
Σ

and (4.1.9) readily follows by combining (4.1.7) with (4.1.12).

In this vein, we also wish to note that whenever ψ ∈ Lipc (Σ) satisfies
 
supp ψ ⊆ Σ ∩ B(x, r) for some x ∈ Σ and r ∈ 0, λΣ−1 · 2 diam Σ ,
(4.1.13)
and ψC.γ (Σ) ≤  1 ,
r γ ·σ Σ∩B(x,r)

then γ
λΣ
sup |ψ(z)| ≤  . (4.1.14)
z ∈Σ σ Σ ∩ B(x, r)
   
Indeed, much as in (4.1.10), there exists some x∗ ∈ B(x, λΣ r) ∩ Σ \ B(x, r) ∩ Σ ,
so for each y ∈ supp ψ ⊆ Σ ∩ B(x, r) we may estimate

|ψ(y)| = |ψ(y) − ψ(x∗ )| ≤ |y − x∗ |γ ψC.γ (Σ)

γ
γ 1 λΣ
≤ (λΣ r) γ   =  , (4.1.15)
r · σ Σ ∩ B(x, r) σ Σ ∩ B(x, r)

as wanted.
Going further, we remark that (4.1.7) entails that, for each γ ∈ (0, 1),
 
if f ∈ Lipc (Σ) is such that fγ vanishes at one point (4.1.16)
on Σ then necessarily f is the zero distribution on Σ.
Let us also point out that, as visible from (4.1.1)-(4.1.2) and (4.1.6), we have
110 4 Hardy Spaces on Ahlfors Regular Sets

1γ ≤ 1 on Σ. (4.1.17)

Another relevant example is discussed below.

Example 4.1.1 For each fixed point  xo ∈ Σ, recall the Dirac distribution with mass
at xo , i.e., the functional δxo ∈ Lipc (Σ) , defined in [133, (4.1.49)]. Then for each
γ ∈ (0, 1) we have

(δxo )γ (x) ≈ |x − xo | 1−n, uniformly for x ∈ Σ. (4.1.18)

Indeed,
 fix a point
 x ∈ Σ and pick some ψ ∈ Tγ (x). Then there exists some number
r ∈ 0, 2 diam Σ for which the properties of ψ listed in (4.1.1) are true. In particular,
if xo  Σ ∩ B(x, r) then δxo ψ = ψ(xo ) = 0, while if xo ∈ Σ ∩ B(x, r) then

1
| δxo ψ | = |ψ(xo )| ≤ sup |ψ(z)| ≤  
z ∈Σ σ Σ ∩ B(x, r)
C C
≤ ≤ . (4.1.19)
r n−1 |x − xo | n−1

In light of (4.1.6), the estimate in (4.1.19) shows that (δxo )γ (x) ≤ C|x − xo | 1−n for
some constant C ∈ (0, ∞) independent of x and xo . In the opposite direction, fix a
 
point x ∈ Σ \ {xo } and pick a function θ ∈ Cc∞ B(0, 2) such that θ ≡ 1 on B(0, 1).
With ρ := |x − xo | ∈ (0, ∞), define the Lipschitz function ψ(z) := ρ1−n θ (z − x)/ρ
for each z ∈ Σ. Then there exists some (typically large) geometric constant C ∈ (0, ∞)
such that the properties in (4.1.1) hold for ψ/C with r := 2ρ. Consequently, ψ/C
belongs to Tγ (x) which, upon noting that |(xo − x)/ρ| = 1, allows us to write
 
C · (δxo )γ (x) ≥ | δxo ψ | = |ψ(xo )| = ρ1−n θ (xo − x)/ρ

= ρ1−n = |x − xo | 1−n . (4.1.20)

To finish the proof of (4.1.18), there remains to show that (δxo )γ (xo ) = +∞, which
we prove next. Specifically,
  with θ as before, define the family of Lipschitz functions
ψt (z) := t 1−n θ (z − xo )/t for each z ∈ Σ and t ∈ (0, ∞). Once again, there exists
C ∈ (0, ∞) independent of t such that the properties in (4.1.1) are valid for ψt /C
with x := xo and r := 2t. Hence, ψt /C ∈ Tγ (xo ) for each t > 0. As such, for each
t > 0 we may write

C · (δxo )γ (xo ) ≥ | δxo ψt | = |ψt (xo )| = t 1−n |θ(0)| = t 1−n . (4.1.21)

Sending t → 0+ then yields (δxo )γ (xo ) = +∞, completing the proof of (4.1.18).
 
Moving on, from [9, Lemma 4.7, p. 131] we know that for each f ∈ Lipc (Σ)
the function
fγ : Σ −→ [0, +∞] is lower-semicontinuous, (4.1.22)
4.1 The Fefferman-Stein Grand Maximal Function 111

when the set Σ is equipped with the topology

 inherited from the Euclidean ambient.
In the case when f ∈ Lloc
1 (Σ, σ) ⊂ Lip (Σ) the duality pairing in (4.1.6) becomes
c
the integral pairing on Σ which, on account of (4.1.1), then readily implies

fγ ≤ M Σ f at σ-a.e. point on Σ, for every f ∈ Lloc

1
(Σ, σ), (4.1.23)

where M Σ is the Hardy-Littlewood maximal operator on (Σ, | · − · |, σ) (recall

(A.0.71)). In the opposite direction we have the following result.

Lemma 4.1.2 Let Σ ⊆ Rn be a closed set which is Ahlfors regular. Abbreviate

σ := H n−1 Σ and fix γ ∈ (0, 1). Then there exists a geometric constant C ∈ (0, ∞)
1 (Σ, σ) one has
with the property that for each function f ∈ Lloc

| f | ≤ C fγ at σ-a.e. point on Σ. (4.1.24)

Proof To get started, pick some f ∈ Lloc 1 (Σ, σ) and consider a bump-function
∞
θ ∈ Cc (R ) satisfying θ ≡ 1 on B(0, 1), supp θ ⊆ B(0, 2), and 0 ≤ θ ≤ 1. Also, fix
n

some x ∈ Σ that  is a Lebesgue

 point for f and select some small r > 0. Then since
the function θ (· − x)/r is non-negative and is identically one on B(x, r) ∩ Σ, we
may estimate
∫
1 y−x
| f (x)| ≤ f (x) ·   θ dσ(y)
σ B(x, r) ∩ Σ Σ r
∫
1 y−x
≤   f (y)θ dσ(y)
σ B(x, r) ∩ Σ Σ r
∫
1   y−x
+   f (y) − f (x) θ dσ(y)
σ B(x, r) ∩ Σ Σ r

=: I1 + I2 . (4.1.25)
   
Since supp θ (· − x)/r ⊆ B(x, 2r) and 0 ≤ θ (· − x)/r ≤ 1, the set Σ is Ahlfors
regular, and x is a Lebesgue point for f , it follows that
∫
C
I2 ≤   | f (y) − f (x)| dσ(y) −→ 0 as r → 0+ . (4.1.26)
σ B(x, 2r) ∩ Σ B(x,2r)∩Σ

Also, (4.1.7) used for φ := θ((·− x)/r) ∈ Lipc (Σ), the point x, and the radius R := 2r,
guarantees (keeping in mind that RφLip(Σ) ≤ supz ∈Rn |(∇θ)(z)| and supΣ |φ| ≤ 1)
that

I1 ≤ C · fγ (x), (4.1.27)

for some purely geometric constant C ∈ (0, ∞), which is independent of x and r. At
this stage, combining (4.1.25)-(4.1.27) yields | f (x)| ≤ C fγ (x) from which (4.1.23)
112 4 Hardy Spaces on Ahlfors Regular Sets

follows given that, according to [133, Proposition 7.4.4] and [133, Lemma 3.6.4],
σ-a.e. x ∈ Σ is a Lebesgue point for the function f . 

Continue to assume that Σ ⊆ Rn is a closed set which is Ahlfors regular. As

before, abbreviate σ := H n−1 Σ and fix γ ∈ (0, 1). Another useful fact about the
grand maximal function is the existence of a geometric constant C ∈ (0, ∞) with the
property that
∫
 
f, φ ≤ C fγ |φ| dσ for every f ∈ Lipc (Σ) and φ ∈ Lipc (Σ). (4.1.28)
Σ

This follows from [9, Proposition 4.15, p. 148] where a more general geometric
setting was considered.

Lemma 4.1.3 Let Σ ⊆ Rn be a closed set which is Ahlfors regular. Abbreviate

σ := H n−1 Σ then fix a parameter γ ∈ (0, 1) along with two integrability exponents

p ∈ [1, ∞] and q ∈ (0, ∞]. Finally, pick an arbitrary distribution f ∈ Lipc (Σ) .
Then

fγ ∈ Lloc (Σ, σ) implies f ∈ Lloc (Σ, σ),

p p
(4.1.29)

fγ ∈ L p (Σ, σ) implies f ∈ L p (Σ, σ), (4.1.30)

fγ ∈ L p,q (Σ, σ) implies f ∈ L p,q (Σ, σ). (4.1.31)

Proof Assume first that fγ ∈ Lloc (Σ, σ) with p ∈ (1, ∞]. In this situation, consider
p

the exponent p ∈ [1, ∞) such that 1/p+1/p = 1, with the understanding that p := 1
when p = ∞. Fix xo ∈ Σ and r ∈ (0, ∞) arbitrary, then abbreviate Δr := B(xo, r) ∩ Σ.
Next, define
L f : Lipc (Δr ) −→ C by setting
(4.1.32)
L f φ := f , φ for every φ ∈ Lipc (Δr ).
Then by design, L f is a well-defined linear functional on Lipc (Δr ), which is a
subspace of the Banach space L p (Δr , σ). Also, from (4.1.28) we know that there
exists a purely geometric constant C ∈ (0, ∞) with the property that for every
φ ∈ Lipc (Δr ) we have
∫
|L f φ| = f , φ ≤ C fγ |φ| dσ ≤ C fγ  L p (Δr ,σ) φ| L p (Δr ,σ) . (4.1.33)
Δr

By the Hahn-Banach Theorem there exists a linear and bounded functional

L f : L p (Δr , σ) −→ C (4.1.34)

which extends L f in (4.1.32). Hence, L f belongs to the dual of L p (Δr , σ) which,

by the Riesz Representation Theorem, can be identified with L p (Δr , σ), given that
p ∈ [1, ∞). This proves that
4.2 Lebesgue-Based and Lorentz-Based Hardy Spaces on Ahlfors Regular Sets 113

there exists a function gr ∈ L p (Δr , σ) such that

∫ (4.1.35)
Lf h = hgr dσ for each h ∈ L p (Δr , σ).
Δr

In particular,
∫
f , φ = Lf φ = φgr dσ for all φ ∈ Lipc (Δr ). (4.1.36)
Δr

In concert with [133, Corollary 3.7.3], this shows that whenever 0 < r1 < r2 < ∞
we necessarily have gr2 Δr = gr1 at σ-a.e. point in Δr1 . In turn, this pointwise
1
compatibility property guarantees the existence of a well-defined function g : Σ → C
such that for each r > 0 we have g Δr = gr at σ-a.e. point in Δr . Consequently,
∫
p
g∈ Lloc (Σ, σ) and f, φ = φg dσ for all φ ∈ Lipc (Σ), (4.1.37)
Σ
p
which goes to show that, indeed, f ∈ Lloc (Σ, σ).
To address the case p = 1 in (4.1.29), define σf := fγ · σ, which is a positive,
locally finite, Borel-regular measure on Σ. By reasoning as in the case p ∈ (1, ∞]
using
 1 the measure
∗ σf in place of σ, and keeping in mind that for each r > 0 we have
L (Δr , σf ) = L ∞ (Δr , σf ), we deduce that
∞ (Σ, σ ) such that
there exists g ∈ Lloc f
∫ (4.1.38)
f, φ = φg dσf for all φ ∈ Lipc (Σ).
Σ

Since we are currently assuming that fγ ∈ Lloc 1 (Σ, σ), from (4.1.38) we conclude

∫
that the function g := g · fγ belongs to Lloc (Σ, σ) and satisfies f , φ = Σ φg dσ
1

for all φ ∈ Lipc (Σ). Ultimately, this shows that f ∈ Lloc

1 (Σ, σ), finishing the proof of

(4.1.29) for all p ∈ [1, ∞].

Next, (4.1.30) is a direct consequence of (4.1.29) and Lemma 4.1.2. In a similar
manner, (4.1.31) is implied by [133, (6.2.36)], (4.1.29), and Lemma 4.1.2. 

4.2 Lebesgue-Based and Lorentz-Based Hardy Spaces on Ahlfors

Regular Sets

We begin by recalling the classical Hardy space H p (Rn, L n )∫(cf., e.g., [176, Theo-
rem 1, p. 91]). Fix a background function Φ ∈ Cc∞ (Rn ) with Rn Φ dL n  0 and set
Φt (x) := t −n Φ(x/t) for all t > 0 and x ∈ Rn . Then for each p ∈ (0, ∞), the classical
Hardy space H p (Rn, L n ) consists of all tempered distributions f in Rn with the
property that their radial maximal function, defined as
114 4 Hardy Spaces on Ahlfors Regular Sets

( frad )(x) := sup |(Φt ∗ f )(x)| = sup Φt (x − ·), f for each x ∈ Rn, (4.2.1)
t>0 t>0

satisfies
frad ∈ L p (Rn, L n ). (4.2.2)
To be able to define Hardy spaces in more general geometric settings, we should
find a suitable substitute for the radial maximal function used above. In this vein,
note that all Φt (x − ·)’s behave like normalized “bump functions” (at scale r ≈ t,
centered at x), of the sort considered in (4.1.1). This observation suggests that in
place of the radial maximal function frad we should use the more geometrically
friendly Fefferman-Stein grand maximal function fγ introduced in (4.1.6).
With this motivation in mind, we proceed to define the scale of (Lebesgue-based)
Hardy spaces as follows. The reader is reminded that, generally speaking,

(a)+ := max{a, 0} for each a ∈ R. (4.2.3)

Definition 4.2.1 Pick a closed set Σ ⊆ Rn which is Ahlfors regular and abbreviate
σ := H n−1 Σ. Having selected p, γ such that
1 
n < p < ∞ and (n − 1) p − 1 + < γ < 1,
n−1
(4.2.4)

define the (Lebesgue-based) Hardy space

   
H p (Σ, σ) := f ∈ Lipc (Σ) : fγ ∈ L p (Σ, σ) (4.2.5)

and equip it with the quasi-norm

 
 f  H p (Σ,σ) :=  fγ  L p (Σ,σ), ∀ f ∈ H p (Σ, σ). (4.2.6)

Note that (4.2.6) is a p-norm (see Definition 1.5.7 for a broader point of view), in
the sense that
p p p
 f + g H p (Σ,σ) ≤  f  H p (Σ,σ) + g H p (Σ,σ), ∀ f , g ∈ H p (Σ, σ). (4.2.7)

We refer the reader to [9] where this brand of Hardy spaces has been studied
at length. Among other things, the Hardy space (4.2.5) is complete (hence, quasi-
Banach) space, and is independent on the parameter γ (chosen as  in (4.2.4)).
 In
general, the Hardy space H p (Σ, σ) is not locally convex when p ∈ n−1n , 1 (see [62,
Theorem 6.2, p. 71] for a proof in the Euclidean setting).
Bearing in mind item (3) of [133, Proposition 4.1.2], from (4.1.7) we see that
   
H p (Σ, σ) → Lipc (Σ) continuously, for each p ∈ n−1 n ,∞ . (4.2.8)

Also, Lemma 4.1.3, (4.1.23), and [133, (7.6.18)] imply that

H p (Σ, σ) coincides, in a quantitative sense, with the

(4.2.9)
Lebesgue space L p (Σ, σ) whenever p belongs to (1, ∞)
4.2 Lebesgue-Based and Lorentz-Based Hardy Spaces on Ahlfors Regular Sets 115

and, corresponding to the end-point case p = 1,

H 1 (Σ, σ) → L 1 (Σ, σ) continuous, proper inclusion. (4.2.10)

Lastly, from (4.2.5), (4.1.17), and Lemma 4.1.2 we conclude that, in the context of
Definition 4.2.1,
1 ∈ H p (Σ, σ) ⇐⇒ Σ is bounded. (4.2.11)
In view of this, in the same setting as that considered in Definition 4.2.1, we may
find it useful to define the homogeneous Hardy space
 
.p f ∈ H p (Σ, σ) : f , 1 = 0 if Σ is bounded,
H (Σ, σ) := (4.2.12)
H p (Σ, σ) if Σ is unbounded.

Finally, we wish to note that, as is apparent from (4.1.22)-(4.1.23), [133, (7.6.18)],

and (4.2.5)-(4.2.6),

L q (Σ, σ) → H p (Σ, σ) continuously

(4.2.13)
if Σ is bounded and n−1
n < p ≤ 1 < q ≤ ∞.

There is also a useful characterization of Hardy spaces on a given closed Ahlfors

regular set Σ ⊆ Rn , with respect to the associated measure σ := H n−1 Σ, in terms of
a sufficiently regular approximation to the identity (aka ATTI). Specifically,
in such a setting [9, Theorem 3.22, pp. 102-103] guarantees the existence of a
 of (σ ⊗
family  σ)-measurable functions St : Σ × Σ → R indexed by a parameter
t ∈ 0, diam Σ and satisfying,for some constant
 C ∈ (0, ∞), the following properties
for all x, x , y, y ∈ Σ and t ∈ 0, diam Σ :

0 ≤ St (x, y) ≤ Ct 1−n and St (x, y) = 0 if |x − y| ≥ Ct,

|St (x, y) − St (x , y)| ≤ Ct −n |x − x |, (4.2.14)

|x − x ||y − y |
[St (x, y) − St (x , y)] − [St (x, y ) − St (x , y )] ≤ C ,
t n+1
∫ ∫
St (x, y) = St (y, x) and St (x, z) dσ(z) = 1 = St (z, y) dσ(z).
Σ Σ

Then, according to [9, Theorem 4.11, p. 140], the Hardy scale may be characterized1
as follows:
   
given f ∈ Lipc (Σ) and p ∈ n−1 n , ∞ , we have f ∈ H (Σ, σ)
p
 
if and only if Σ  x −→ sup Lip c (Σ) St (x, ·), f (Lip c (Σ)) (4.2.15)
0<t<diam(Σ)
belongs to L p (Σ, σ).

1 this characterization turns out to be quantitative as well

116 4 Hardy Spaces on Ahlfors Regular Sets
 
As an example, consider the case when f := δx0 − δx1 ∈ Lipc (Σ) for some fixed
points x0, x1 ∈ Σ. Then from (4.2.14) we see that there exists a constant C ∈ (0, ∞)
with the property that for each x ∈ Σ we have
⎧
⎪ C(1 + |x|)−n if x is away from x0, x1,
⎪
⎪
  ⎨
⎪
Lip c (Σ) St (x, ·), f (Lip c (Σ)) ≤ C|x − x0 |
1−n if x is near x ,
sup 0
⎪
⎪
0<t<diam(Σ) ⎪
⎪ C|x − x1 | 1−n if x is near x1 .
⎩
(4.2.16)
In turn, from (4.2.15), (4.2.16), and [133, (7.2.5)] we conclude that
 
for each given x0, x1 ∈ Σ, the distribution δx0 − δx 1 ∈ Lip  c (Σ) (4.2.17)
belongs to the Hardy space H p (Σ, σ) for every p ∈ n−1 n ,1 .

In particular, this points to the fact that,

 in sharp contrast to (4.2.9)-(4.2.10), the
Hardy space H p (Σ, σ) with p ∈ n−1 n , 1 contains genuine2 distributions on Σ.
Moving in, recall the log-convexity law for the Lebesgue scale, to the effect that on
an arbitrary measure space (X, μ) if 0 < p0 < p1 ≤ ∞ and f ∈ L p0 (X, μ) ∩ L p1 (X, μ)
then f ∈ L p (X, μ) for each p ∈ [p0, p1 ] and we have

 f  L p θ (X,μ) ≤  f  L1−θ θ
p0 (X,μ)  f  L p1 (X,μ) for each θ ∈ [0, 1],

  −1 (4.2.18)
where pθ := (1 − θ)/p0 + θ/p1 .

Moreover, if X has finite measure, then Hölder’s inequality ensures that the Lebesgue
scale is nested, i.e.,

L p1 (X, μ) → L p0 (X, μ) continuously,

(4.2.19)
if μ(X) < ∞ and 0 < p0 ≤ p1 ≤ ∞.

Below we record results for Hardy spaces of a similar nature.

Proposition 4.2.2 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and introduce
σ := H n−1 Σ. Also, fix n−1 n < p0 < p1 < ∞ and consider an arbitrary function
f ∈ H p0 (Σ, σ) ∩ H p1 (Σ, σ). Then f ∈ H p (Σ, σ) for each p ∈ [p0, p1 ], and there
exists a constant C ∈ (0, ∞) independent of f such that

 f  H p θ (Σ,σ) ≤ C f  H θ
p0 (Σ,σ)  f  H p1 (Σ,σ) for each θ ∈ [0, 1],
1−θ

  −1 (4.2.20)
where pθ := (1 − θ)/p0 + θ/p1 .

Furthermore,
H p1 (Σ, σ) → H p0 (Σ, σ) continuously
(4.2.21)
if Σ is bounded and n−1
n < p0 ≤ p1 < ∞.
Proof The claims in the first part of the statement are direct consequences of the
log-convexity law for the Lebesgue scale and (4.2.5)-(4.2.6) (alternatively, we may
2 i.e., distributions which are not induced by integration against a locally integrable function
4.2 Lebesgue-Based and Lorentz-Based Hardy Spaces on Ahlfors Regular Sets 117

use the area-function characterization of Hardy spaces recorded later, in (4.4.37)).

Likewise, (4.2.21) follows from (4.2.19) in a similar manner. 
The scale of Lorentz-based Hardy spaces, defined below, is an environment in
which a number of basic operators in Harmonic Analysis, including certain types of
singular integral operators and square-functions, act in a natural fashion.
Definition 4.2.3 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and abbreviate
σ := H n−1 Σ. Having selected p, q, γ such that
 
n < p < ∞, 0 < q ≤ ∞, and (n − 1) p1 − 1 + < γ < 1,
n−1
(4.2.22)

define the Lorentz-based Hardy space

   
H p,q (Σ, σ) := f ∈ Lipc (Σ) : fγ ∈ L p,q (Σ, σ) (4.2.23)

and equip it with the quasi-norm

 
 f  H p, q (Σ,σ) :=  fγ  L p, q (Σ,σ), ∀ f ∈ H p,q (Σ, σ). (4.2.24)

In the case when Σ is the hyperplane Rn−1 ≡ Rn−1 × {0} ⊆ Rn , these spaces
have been introduced by C. Fefferman, N. Rivière, and Y. Sagher in [51]. The case
q = ∞, when the terminology weak Hardy space is sometimes employed, has been
studied in the same Euclidean setting by R. Fefferman and F. Soria in [53]. Weak
Hardy spaces on spaces of homogeneous type have been considered in [197].
The Lorentz-based Hardy spaces considered in Definition 4.2.3 make up a more
inclusive scale than the family of (Lebesgue-based) Hardy spaces introduced earlier,
in Definition 4.2.1 since, as is apparent from definitions and [133, (6.2.25)],
 
H p, p (Σ, σ) = H p (Σ, σ) for each p ∈ n−1
n ,∞ . (4.2.25)

Let us also observe that we have the following continuous embeddings:

 
H p (Σ, σ) → H p,∞ (Σ, σ) for each p ∈ n−1 n ,∞ , (4.2.26)
 
H p,q1 (Σ, σ) → H p,q2 (Σ, σ) if p ∈ n−1
n , ∞ and 0 < q1 ≤ q2 ≤ ∞, (4.2.27)

H 1 (Σ, σ) → L 1 (Σ, σ) → H 1,∞ (Σ, σ). (4.2.28)

Indeed, (4.2.26)-(4.2.27) are clear from definitions and similar properties enjoyed
by Lorentz spaces, while (4.2.28) uses (4.2.10), (4.1.23), and [133, (7.6.19)]. From
(4.1.7) and [133, (6.2.16)] we also conclude (also bearing in mind item (3) of [133,
Proposition 4.1.2]) that
 
H p,q (Σ, σ) → Lipc (Σ) continuously,
  (4.2.29)
for p ∈ n−1 n , ∞ and q ∈ (0, ∞].

In addition, from Lemma 4.1.3, (4.1.23), and [133, (6.2.16), (6.2.20)] we see that
118 4 Hardy Spaces on Ahlfors Regular Sets

H p,q (Σ, σ) coincides, in a quantitative sense, with the

(4.2.30)
Lorentz space L p,q (Σ, σ) if p ∈ (1, ∞) and q ∈ (0, ∞],

and there exists a geometric constant C ∈ (0, ∞) for which

H 1,∞ (Σ, σ) ∩ Lloc 1 (Σ, σ) ⊂ L 1,∞ (Σ, σ), and we have the estimate
(4.2.31)
 f  L 1,∞ (Σ,σ) ≤ C f  H 1,∞ (Σ,σ) for all f ∈ H 1,∞ (Σ, σ) ∩ Lloc
1 (Σ, σ).

In particular, from (4.2.30) and [133, (6.2.36)] we conclude that

p∗
H p,q (Σ, σ) → Lloc (Σ, σ) provided
(4.2.32)
p ∈ (1, ∞), q ∈ (0, ∞], and p∗ ∈ (0, p).

There is also a log-convexity law for the scale of Lorentz-based Hardy spaces,
and [133, Lemma 6.2.6] implies that
if Σ is bounded, then whenever 0 < p0 < p1 < ∞ and
0 < q0, q1 ≤ ∞, we have the continuous embedding (4.2.33)
H p1,q1 (Σ, σ) → H p0,q0 (Σ, σ).
From Definition 4.2.3, Example 4.1.1, and [133, Example 6.2.2] it follows that
Dirac distributions with mass at points on Σ belong to the weak Hardy space
H 1,∞ (Σ, σ) in a uniform fashion with respect to the location of the point where the
mass is concentrated. This, which in particular shows that H 1,∞ (Σ, σ)  Lloc
1 (Σ, σ),

is made explicit in the example below (which should be compared with (4.2.17)).

Example 4.2.4 Let Σ ⊆ Rn be an Ahlfors regular set and abbreviate σ := H n−1 Σ.
Then there exists a purely geometric constant C ∈ (0, ∞) with the property that for
each point xo ∈ Σ the distribution
 
δxo ∈ Lipc (Σ) belongs to H 1,∞ (Σ, σ)
(4.2.34)
and satisfies δxo  H 1, ∞ (Σ,σ) ≤ C.

Recall that the j-th Riesz transform R j in Rn is (up to normalization) the operator
xj
of convolution with the principal-value distribution P.V. |x | n+1 (cf., e.g., [130, (4.9.1),
p. 181]). With δ0 denoting the Dirac distribution with mass at the origin in Rn , from
(4.2.34) we have δ0 ∈ H 1,∞ (Rn, L n ), and the definition of R j implies R j δ0 is a
xj
(nonzero dimensional) multiple of P.V. |x | n+1 . Hence

xj
P.V. ∈ H 1,∞ (Rn, L n ) for each j ∈ {1, . . . , n}, (4.2.35)
|x| n+1

since it turns out that the Riesz transforms map the weak Hardy space H 1,∞ (Rn, L n )
into itself. The latter fact may be established based on the observation that the 
Riesz transforms map ordinary Hardy spaces, H p (Rn, L n ) with p ∈ n+1 n
,∞ ,
boundedly into themselves (cf. [52, Remark 2, p. 191]), via real interpolation. A
general interpolation result of this nature makes the object of our next theorem.
4.3 Real Interpolation of Hardy Spaces 119

4.3 Real Interpolation of Hardy Spaces

Our main result concerning the real interpolation of Hardy spaces reads as follows.

Theorem 4.3.1 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and introduce
σ := H n−1 Σ. Then

H p (Σ, σ), L ∞ (Σ, σ) = H pθ ,q (Σ, σ) whenever

θ,q (4.3.1)
 
n ,1 ,
p ∈ n−1 q ∈ (0, ∞), θ ∈ (0, 1), and p1θ = 1−θ p .

Moreover,

H p0,q0 (Σ, σ), H p1,q1 (Σ, σ) = H p,q (Σ, σ) provided

θ,q (4.3.2)
  θ
p0, p1, p ∈ n−1n , ∞ , q0, q1, q ∈ (0, ∞], θ ∈ (0, 1), and p = p0 +
1 1−θ
p1 .

In particular,

H p0 (Σ, σ), H p1 (Σ, σ) = H p,q (Σ, σ) provided

θ,q
  θ
(4.3.3)
p0, p1, p ∈ n−1
n ,∞ , q ∈ (0, ∞], θ ∈ (0, 1), and p1 = 1−θ p0 + p1 .

As a consequence of this and (4.2.25),

H p0 (Σ, σ), H p1 (Σ, σ) = H p (Σ, σ) whenever

θ, p
 n−1  θ
(4.3.4)
p0, p1, p ∈ n , ∞ , θ ∈ (0, 1), and p1 = 1−θ p0 + p1 .

A byproduct of Theorem 4.3.1 is the independence of the Lorentz-based Hardy

space defined in (4.2.23) on the particular value of the parameter γ (chosen as
in (4.2.22)). Theorem 4.3.1 also shows that Lorentz-based Hardy spaces may be
realized as intermediate spaces obtained by interpolating using the real method
between standard (Lebesgue-based) Hardy spaces. Our primary interest in this result
stems from its applications to the theory of singular integral operators in rough
settings, as discussed later on, in [136, Chapter 2].
The proof of Theorem 4.3.1 is presented later. The strategy involves the following
steps:
Step I: Prove (4.3.1). This is done using a Calderón-Zygmund decomposition
lemma for locally integrable functions, in the spirit of [51, Lemma A, p. 76] which
deals with the Euclidean setting. Given the nature of the functions decomposable as in
said lemma, we also need a result asserting the density of locally integrable functions
in Lorentz-based Hardy spaces. It is precisely here that the limitation q < ∞ occurs
in (4.3.1). Unfortunately, this aspect is apparently overlooked in [51] where q = ∞
is permitted in the formulation of the analogue of (4.3.1) in the Euclidean setting,
even though the proof in [51] certainly fails in this case (as Schwartz functions are
not dense in weak Hardy spaces).
120 4 Hardy Spaces on Ahlfors Regular Sets

Step II: Use formula (4.3.1) and the reiteration theorem for the real method of
interpolation to conclude that (4.3.2) holds when q0, q1, q < ∞.
Step III: Prove (4.3.3) when q = ∞. Since the density result from Step I does not
hold, we are forced to employ a new Calderón-Zygmund decomposition, valid for
distributions which are not necessarily associated with locally integrable functions.
Such a result has been produced in [9, Theorem 5.16, p. 207], and this is the main
tool in dealing with the present task.
Step IV: Combine the results proved in Steps II-III to justify (4.3.3) in full
generality.
Step V: Rely on the full force of (4.3.3) and the reiteration theorem to conclude
that (4.3.2) holds in full generality. In the Euclidean setting, such a formula first
appeared in [51], though there is a gap in the proof when q = ∞ (as indicated in
Step I above). Our result here both generalizes and corrects the work in [51].
As a preamble to the proof of Theorem 4.3.1, we prove several preliminary results.
First, in Lemma 4.3.2 below we discuss a Calderón-Zygmund type decomposition
for locally integrable functions, in the spirit of [9, Theorem 5.16, p. 207] and [51,
Lemma A, p. 76].
Lemma 4.3.2 Let Σ ⊆ Rn be a closed set which
 is Ahlfors
1  regular
 and define
σ := H n−1 Σ. Suppose p ∈ n−1
n , 1 and fix γ ∈ (n − 1) p − 1 , 1 . Also, pick some
1 (Σ, σ) and recall the Fefferman-Stein grand maximal function f  associated
f ∈ Lloc γ
with f as in (4.1.6). Then there exists some constant C = C(Σ, p, γ) ∈ (0, ∞),
independent of f , with following significance.
Suppose α ∈ [0, ∞) and consider the set
 
Oα := x ∈ Σ : fγ (x) > α . (4.3.5)

In the case when α > 0 make the additional assumption that Oα is a proper subset
of Σ. Then one can find g ∈ L ∞ (Σ, σ) and b ∈ Lloc
1 (Σ, σ) satisfying

f = g + b at σ-a.e. point on Σ,
|g| ≤ C · α at σ-a.e. point on Σ,
∫ ∫ (4.3.6)
 p   p
bγ dσ ≤ C fγ dσ.
Σ Oα

Proof Let us first dispense with the case α = 0. Then g := 0 and b := f will do. For
the remainder of the proof assume α ∈ (0, ∞) and that Oα from (4.3.5) is a proper
subset of Σ. Note that (4.1.24) implies

| f | ≤ C · α at σ-a.e point on Σ \ Oα, (4.3.7)

for some C ∈ (0, ∞) depending only on Σ. Since fγ is lower-semicontinuous it

follows that Oα is a relatively open subset of Σ. If Oα = , take g := f and b := 0
(which, thanks to (4.3.7), will do). Henceforth consider the case when Oα  , and
recall that we are assuming Oα to be a proper subset of Σ. These observations pave
4.3 Real Interpolation of Hardy Spaces 121

the way for performing a Whitney decomposition of Oα . Concretely, fix some λ > 1

and apply [133, Proposition 7.5.3] (with X := Σ and ρ the Euclidean distance in Rn ).
This guarantees the existence of two constants, Λ ∈ (λ, ∞) and M ∈ N, a sequence
of points {x j } j ∈N ⊆ Oα , and a sequence of radii {r j } j ∈N ⊆ (0, ∞), enjoying the
following properties:
 
Oα = B(x j , r j ) ∩ Σ, 1B(x j ,λr j ) ≤ M on Oα, (4.3.8)
j ∈N j ∈N
 
B(x j , λr j ) ∩ Σ ⊆ Oα and B(x j , Λr j ) ∩ Σ \ Oα   for every j ∈ N, (4.3.9)
ri ≈ r j uniformly for i, j ∈ N such that B(xi, λri ) ∩ B(x j , λr j )  . (4.3.10)

In fact, as noted in [133, (7.5.5)], a stronger condition that the finite overlap property
mentioned in (4.3.8) holds, namely there exists a small dilation parameter ε ∈ (0, 1)
such that
   
# j ∈ N : B x, ε ·dist(x, Σ \ Oα ) ∩ B(x j , λr j )∩Σ   ≤ M, ∀x ∈ Oα . (4.3.11)

Next, apply [138, Comment 4.22, p. 189] to construct a partition of unity sub-
ordinate to this cover of Oα of the following sort. Impose the additional restriction
λ > 4 and pick an arbitrary λ ∈ (2, λ/2). Then, for each j ∈ N introduce the sets
j := B(x j , λr j ) ∩ Σ, and observe that
E j := B(x j , r j ) ∩ Σ, E j := B(x j , λ r j ) ∩ Σ, E
they satisfy hypotheses (a)-(d) of [138, Theorem 4.18, pp. 178-179] (with {r j } j ∈N as
above). Hence, the latter theorem applies and yields a constant C ∈ [1, ∞) together
with a family of real-valued functions {φ j } j ∈N defined on Σ with the property that
for each j ∈ N one has

φ j ∈ Lip(Σ), φ j Lip(Σ) ≤ Cr j−1 (4.3.12)

0 ≤ φ j ≤ 1 on Σ, φ j ≡ 0 on Σ \ B(x j , λ r j ), (4.3.13)
and φ j ≥ 1/C on B(x j , r j ) ∩ Σ, (4.3.14)

and also 
φ j = 1 j ∈N B(x j ,r j )∩Σ = 1 Oα . (4.3.15)
j ∈N

This partition of unity enjoys a number of additional properties. First, we claim

that there exists a geometric constant C ∈ (0, ∞) such that
−γ
φ j C.γ (Σ) ≤ Cr j for every j ∈ N. (4.3.16)

Indeed, if x, y ∈ Σ, then (4.3.12)-(4.3.13) imply

|φ j (x) − φ j (y)| ≤ Cr j−1 |x − y| and |φ j (x) − φ j (y)| ≤ 2. (4.3.17)

122 4 Hardy Spaces on Ahlfors Regular Sets

Now (4.3.16) follows by raising the first inequality in (4.3.17) to the power γ,
raising the second inequality to the power 1 − γ, and then multiplying the resulting
inequalities.
A second property we wish to single out is as follows. To set the stage, observe
that for each fixed j ∈ N we may, thanks to (4.3.9), select a point
 
y j ∈ B(x j , Λr j ) ∩ Σ \ Oα . (4.3.18)

Since λ < λ/2 < Λ/2, there exists some constant c ∈ (0, ∞), independent of j, such
that
if R j := cr j , then B(x j , λ r j ) ⊆ B(y j , R j ). (4.3.19)
Finally, recall the collection of sets Tγ (x) for x ∈ Σ defined in (4.1.2). Then, in
relation to the partition of unity {φ j } j ∈N introduced earlier, the second claim we
make is that there exists some large constant C0 ∈ (0, ∞) such that
  
φ j C0 · σ B(x j , r j ) ∩ Σ belongs to Tγ (y j ), for every j ∈ N. (4.3.20)

This is seen by applying (4.1.3) for φ := φ j , R := R j , and x := y j . Specifically, since

from (4.3.12)-(4.3.13) and (4.3.19) we have that, for each j ∈ N,

the Lipschitz function φ j is supported in B(y j , R j ) ∩ Σ, (4.3.21)

   
and since sup j ∈N supΣ |φ j | ≤ 1 and sup j ∈N r j · φ j Lip(Σ) < +∞, it follows that
the quantity Cφ,R (defined in (4.1.3)) may be bounded independently of j. By also
taking into account the fact that, thanks to the Ahlfors regularity of Σ, we have
σ(B(x j , r j ) ∩ Σ) ≈ σ(B(y j , R j ) ∩ Σ) uniformly in j, we conclude that C0 ∈ (0, ∞)
doing the job in (4.3.20) does exist.
The third claim we make is that
∫
 
f φ j dσ ≤ C · α · σ B(x j , r j ) ∩ Σ , for every j ∈ N. (4.3.22)
Σ

To prove (4.3.22), for every j ∈ N, we use the fact that f ∈ Lloc

1 (Σ, σ), together with

(4.3.20), (4.1.6), (4.3.5), and (4.3.18), to write

∫
   φj 
f φ j dσ = f , φ j = C0 · σ B(x j , r j ) ∩ Σ f ,  
Σ C0 · σ B(x j , r j ) ∩ Σ
   
≤ C0 · σ B(x j , r j ) ∩ Σ fγ (y j ) ≤ C0 · α · σ B(x j , r j ) ∩ Σ . (4.3.23)

At this stage we define the sequence of numbers

∫
f φ j dσ
λ j := ∫Σ , for every j ∈ N. (4.3.24)
φ dσ
Σ j
4.3 Real Interpolation of Hardy Spaces 123

Based on the fact that f ∈ Lloc

1 (Σ, σ) and the properties of each φ , the λ ’s are
j j
meaningfully defined. Also, it is immediate from (4.3.24) that
∫
( f − λ j )φ j dσ = 0, for every j ∈ N. (4.3.25)
Σ

Moreover, (4.3.22) implies

|λ j | ≤ C · α, for every j ∈ N. (4.3.26)

This newly introduced sequence allows us to define the function


g := f · 1Σ\Oα + λ j φ j on Σ. (4.3.27)
j ∈N

As is apparent from (4.3.11) and (4.3.13) (bearing in mind that λ < λ),  any point
 in
Oα has a neighborhood which intersects at most M sets from the family supp φ j j ∈N .
This property implies that the series in (4.3.27) converges uniformly on compact
subsets of Oα . As such, the function g is both well-defined and σ-measurable on Σ.
In addition, a combination of (4.3.27), (4.3.7), (4.3.26), and (4.3.15), gives

|g| ≤ C · α at σ-a.e. point on Σ. (4.3.28)

In particular, g ∈ L ∞ (Σ, σ). If we now set b := f − g, then b ∈ Lloc1 (Σ, σ) and (4.3.6)

is satisfied. To finish the proof of Lemma 4.3.2, we are left with showing that the
last estimate in (4.3.6) holds.
With this goal in mind, we first remark that (4.3.27) and (4.3.15) imply
 
b = f − g = f · 1 Oα − λj φj = ( f − λ j )φ j at σ-a.e. point on Σ. (4.3.29)
j ∈N j ∈N

Consequently, if we define b j := ( f − λ j )φ j on Σ, for each j ∈ N, then


b j ∈ Lloc
1
(Σ, σ) and b = b j at σ-a.e. point on Σ. (4.3.30)
j ∈N

Our next goal is to prove that there exists a constant C ∈ (0, ∞) such that, for each
j ∈ N and at each x ∈ Σ, the following estimate holds:
⎧
⎪ 
⎨ C · fγ (x)
⎪
⎪ if x ∈ B(x j , λr j ) ∩ Σ,
(b j )γ (x) ≤ rj n−1+γ (4.3.31)
⎪
⎪ C · α · |x − x j |
⎪ if x ∈ Σ \ B(x j , λr j ).
⎩
Proving this takes some work. To set the stage, fix j ∈ N, x ∈ Σ, and ψ ∈ Tγ (x).
Hence, ψ is Lipschitz, compactly supported on Σ, and satisfies (4.1.1) for some
r ∈ (0, ∞), i.e.,
124 4 Hardy Spaces on Ahlfors Regular Sets

1
supp ψ ⊆ Σ ∩ B(x, r), sup |ψ(z)| ≤  ,
z ∈Σ σ Σ ∩ B(x, r)
(4.3.32)
1
and ψC.γ (Σ) ≤  .
r γ · σ Σ ∩ B(x, r)

It is then immediate that

∫
|ψ(y)| dσ(y) ≤ 1. (4.3.33)
Σ

We proceed with the proof of (4.3.31) by analyzing three separate cases, as follows.
Case 1. Assume x ∈ B(x j , λr j ) ∩ Σ and r ≤ r j . Since b j ∈ Lloc
1 (Σ, σ) we may write

∫
b j, ψ = b j (y)ψ(y) dσ(y)
Σ
∫ ∫
= f (y)φ j (y)ψ(y) dσ(y) − λ j φ j (y)ψ(y) dσ(y) =: I + I I. (4.3.34)
Σ Σ

Next, using (4.3.26), (4.3.13), and (4.3.33), we estimate

∫
|I I | ≤ |λ j φ j  L ∞ (Σ,σ) |ψ(y)| dσ(y) ≤ Cα < C fγ (x), (4.3.35)
Σ

where the last inequality in (4.3.35) is a consequence of the fact that in the current case
x ∈ Oα . To obtain a similar upper bound for |I |, the idea is to show that there exists
some constant C ∈ (0, ∞), independent of j, x, and ψ, for which φ j ψ/C ∈ Tγ (x). To
see why the later membership holds, first note that (4.3.13) and (4.3.32) imply
 
supp (φ j ψ) ⊆ Σ ∩ B(x, r) and |φ j ψ| ≤ 1/σ Σ ∩ B(x, r) on Σ. (4.3.36)

Second, relying on (4.3.13), (4.3.16), (4.3.32), and the fact that we are currently
assuming r ≤ r j , we further estimate

φ j ψC.γ (Σ) ≤ φ j  L ∞ (Σ,σ) ψC.γ (Σ) + φ j C.γ (Σ) ψ L ∞ (Σ,σ)

1 C 1
≤   + γ ·  
rγ · σ Σ ∩ B(x, r) r j σ Σ ∩ B(x, r)

C
≤  . (4.3.37)
rγ · σ Σ ∩ B(x, r)

This proves that there exists some constant C ∈ (0, ∞) independent of j, x, and ψ,
with the property that φ j ψ/C ∈ Tγ (x). Consequently,

|I | = f , φ j ψ ≤ C · fγ (x). (4.3.38)

4.3 Real Interpolation of Hardy Spaces 125

In concert, (4.3.34), (4.3.35), and (4.3.38) prove that | b j , ψ | ≤ C · fγ (x). Since ψ
was arbitrarily selected from Tγ (x), this proves (4.3.31) in the current case.
Case 2. Assume x ∈ B(x j , λr j ) ∩ Σ and r > r j . The argument in this case is similar
to the one in the proof of Case 1. Specifically, start by writing (4.3.34). The estimate
in (4.3.35) continues to be valid for the same reasons. To estimate I, again we claim
that there exists some constant C ∈ (0, ∞) independent of j, x, and ψ, for which
φ j ψ/C ∈ Tγ (x). In the current case, B(x j , r j ) ⊆ B(x, 2λr j ), hence we may view the
function φ j ψ as being supported in B(x, 2λr j ) ∩ Σ. Using the Ahlfors regularity of
Σ and the fact that r > r j we obtain
   
|φ j ψ| ≤ 1/σ Σ ∩ B(x, r) ≤ C/σ Σ ∩ B(x, 2λr j ) on Σ. (4.3.39)

To estimate φ j ψC.γ (Σ) , we may again proceed as in the first two lines in (4.3.37),
and then rely on the the Ahlfors regularity of Σ and the fact that r > r j to write

1 C 1
φ j ψC.γ (Σ) ≤   + γ ·  
rγ · σ Σ ∩ B(x, r) r j σ Σ ∩ B(x, r)

C
≤  . (4.3.40)
(2λr j )γ · σ Σ ∩ B(x, 2λr j )

Hence, φ j ψ/C ∈ Tγ (x) for some constant C ∈ (0, ∞) independent of j, x, ψ. The

proof of estimate (4.3.31) in the current case is now finished exactly as in the
end-game of Case 1.
Case 3. Assume x ∈ Σ \ B(x j , λr j ). The first observation we make is that the
function b j ψ vanishes
∫ identically on Σ if B(x, r) ∩ B(x j , r j ) ∩ Σ = . Since we
seek to estimate Σ b j ψ dσ, we may therefore assume that there exists some point
y∗ ∈ B(x, r) ∩ B(x j , r j ) ∩ Σ. Hence,

(λ − 1)r j = λr j − r j < |x − x j | − |y∗ − x j | ≤ |x − y∗ | < r (4.3.41)

which, given that λ > 1, further permits us to estimate

|x − x j | ≤ |x − y∗ | + |y∗ − x j | ≤ r + r j ≤ Cr, (4.3.42)

∫
for some geometric constant C ∈ (0, ∞). In order to estimate Σ b j ψ dσ, introduce
the function ψ j := ψ − ψ(x j ) and invoke the cancellation property (4.3.25) to write
∫
b j, ψ = b j (y)ψ j (y) dσ(y)
Σ
∫ ∫
= f (y)φ j (y)ψ j (y) dσ(y) − λ j φ j (y)ψ j (y) dσ(y)
Σ Σ

=: I I I + IV . (4.3.43)
126 4 Hardy Spaces on Ahlfors Regular Sets

To treat IV, rely on (4.3.26) and (4.3.13) to first obtain

∫
|IV | ≤ |λ j |φ j  L ∞ (Σ,σ) |ψ j (y)| dσ(y)
B(x j ,r j )∩Σ
∫
≤ Cα |ψ j (y)| dσ(y). (4.3.44)
B(x j ,r j )∩Σ

Making use of (4.3.32) we write

|ψ j (y)| = |ψ(y) − ψ(x j )| ≤ |y − x j |γ ψC.γ (Σ)

γ
rj
≤   for each y ∈ B(x j , r j ) ∩ Σ. (4.3.45)
r γ · σ B(x, r) ∩ Σ

Returning with this in (4.3.44) and also employing the Ahlfors regularity of Σ and
(4.3.42) we may further estimate
γ
rj   rj n−1+γ
|IV | ≤ C · α ·   σ B(x j , r j ) ∩ Σ ≤ Cα ·
r γ · σ B(x, r) ∩ Σ r
rj n−1+γ
≤ Cα · . (4.3.46)
|x − x j |

We desire a similar upper bound for |I I I |. To obtain this desired bound recall y j
and R j from (4.3.18) and (4.3.19) and observe that (4.3.21) ensures that

φ j ψ j ∈ Lipc (Σ) and supp (φ j ψ j ) ⊂ B(y j , R j ) ∩ Σ. (4.3.47)

Also, the properties of φ j and ψ (from (4.3.12)-(4.3.13) and (4.3.32)) combined with
(4.3.42), (4.3.19), and the Ahlfors regularity of Σ give

φ j ψ j  L ∞ (Σ,σ) ≤ sup |ψ(x) − ψ(x j )| ≤ sup |x − x j |γ ψC.γ (Σ)

x ∈B(x j ,r j )∩Σ x ∈B(x j ,r j )∩Σ

γ 1
≤ rj ·  
r γ · σ B(x, r) ∩ Σ
rj n−1+γ 1
≤C ·  . (4.3.48)
|x − x j | σ B(y j , R j ) ∩ Σ

Moreover, we claim that

rj n−1+γ 1
φ j ψ j C.γ (Σ) ≤ C · γ  . (4.3.49)
|x − x j | Rj · σ B(y j , R j ) ∩ Σ

Assume (4.3.49) for now. Then (4.3.47), (4.3.48), and (4.3.49) imply that there exists
C ∈ (0, ∞) independent of j, x, ψ, such that
4.3 Real Interpolation of Hardy Spaces 127

φ j ψj
n−1+γ
∈ Tγ (y j ). (4.3.50)
rj
C |x−x j |

In turn, we may employ this membership to estimate |I I I | (from (4.3.43)) as


rj n−1+γ φ j ψj
|I I I | = f , φ j ψ j = C f,
|x − x j | rj n−1+γ
C |x−x j |

rj n−1+γ rj n−1+γ
≤C fγ (x) ≤ Cα · . (4.3.51)
|x − x j | |x − x j |

This is the same upper bound as the one obtained for IV. Collectively, (4.3.43),
(4.3.46), and (4.3.51) give

rj n−1+γ
b j , ψ ≤ Cα · |x−x j | (4.3.52)
for every x ∈ Σ \ B(x j , λr j ) and ψ ∈ Tγ (x).

The estimate in (4.3.31) corresponding to x ∈ Σ \ B(x j , λr j ) now follows from

(4.3.52).
To finish the proof in Case 3 we are left with establishing (4.3.49). As an
intermediate step, we rely on the fact that r j ≈ R j , the Ahlfors regularity of Σ, and
(4.3.42) to write
1 C C
  ≤ n−1+γ ≤
rγ · σ B(x, r) ∩ Σ r |x − x j|
n−1+γ

rj n−1+γ 1
≤C · γ  . (4.3.53)
|x − x j | Rj · σ B(y j , R j ) ∩ Σ

Note that, in light of estimate (4.3.53), in order to conclude (4.3.49) it suffices to

prove that, for every y, z ∈ Σ,

(φ j ψ j )(y) − (φ j ψ j )(z) C
≤ γ  . (4.3.54)
|y − z|γ r · σ B(x, r) ∩ Σ

The proof of (4.3.54) is handled by analyzing separately the following three cases:
Case (a): y ∈ Σ \ B(x j , r j ) and z ∈ Σ \ B(x j , r j );
Case (b): z ∈ B(x j , r j ) ∩ Σ;
Case (c): y ∈ B(x j , r j ) ∩ Σ.
Since supp (φ j ψ j ) ⊂ B(x j , r j ) ∩ Σ, the estimate in (4.3.54) is immediate in
Case (a). Also, we observe that the estimate in (4.3.54) is symmetric in y and
z, thus its proof in Case (b) is the same as its proof in Case (c). As such, we are
only left with treating Case (c). Having fixed y ∈ B(x j , r j ) ∩ Σ and an arbitrary
z ∈ Σ, we then write
128 4 Hardy Spaces on Ahlfors Regular Sets

(φ j ψ j )(y) − (φ j ψ j )(z)

≤ φ j (z) ψ j (y) − ψ j (z) + ψ j (y) φ j (y) − φ j (z)

≤ |y − z|γ ψC.γ (Σ) + |y − x j |γ ψC.γ (Σ) |y − z|γ φ j C.γ (Σ)

1 1 C|y − z|γ
≤ |y − z|γ ·   + r γj · γ   · γ
rγ · σ B(x, r) ∩ Σ r · σ B(x, r) ∩ Σ rj

C|y − z|γ
≤  , (4.3.55)
r γ · σ B(x, r) ∩ Σ

where for the third inequality we used (4.3.32) and (4.3.16). Hence, (4.3.54) is proved
to be true, and this completes the proof of the estimate in (4.3.31) corresponding to
Case 3.
Having finished the proof of (4.3.31), the next order of business is showing that
b (recall (4.3.29)) satisfies the estimate in the last line of (4.3.6). In this regard, we
first claim that
∫ N ∫

bψ dσ = lim b j ψ dσ, ∀ψ ∈ Lipc (Σ). (4.3.56)
Σ N →∞ Σ
j=1

To justify this claim, fix ψ ∈ Lipc (Σ) and note that (4.3.30) guarantees that
!
j=1 b j ψ → bψ at σ-a.e. point in Σ, while (4.3.26), (4.3.13), and (4.3.8) collectively
N

imply


N 
N
 
bj ψ ≤ | f − λ j ||ψ|φ j ≤ M | f | + C · α |ψ|1 Oα ∈ L 1 (Σ, σ). (4.3.57)
j=1 j=1

Based on these two observations and Lebesgue’s Dominated Convergence Theorem

we conclude that the claim made in (4.3.56) is true. Going further, from (4.3.56) and
(4.1.6) we see that

bγ ≤ (b j )γ at each point on Σ. (4.3.58)
j ∈N

Given that 0 < p < 1, this further implies

 p   p
bγ ≤ (b j )γ on Σ. (4.3.59)
j ∈N

To estimate the L p -(quasi)norm of each (b j )γ , fix j ∈ N and use (4.3.31) to write

4.3 Real Interpolation of Hardy Spaces 129
∫ ∫ ∫
 p  p  p
(b j )γ dσ = (b j )γ dσ + (b j )γ dσ
Σ B(x j ,λr j )∩Σ Σ\B(x j ,λr j )
∫
 p
≤C fγ dσ
B(x j ,λr j )∩Σ
∫ (n−1+γ)p
rj
+ Cα p
dσ(x). (4.3.60)
Σ\B(x j ,λr j ) |x − x j |
1 
Since γ > (n − 1) p − 1 , [133, Lemma 7.2.1] applies and [133, (7.2.5)] yields
∫ (n−1+γ)p
rj n−1−(n−1+γ)p
dσ ≤ Cr j . (4.3.61)
Σ\B(x j ,λr j ) |x − x j |

In turn, (4.3.61), the Ahlfors regularity of Σ, the fact that B(x j , λr j ) ∩ Σ ⊂ Oα , and
the definition of Oα may be employed to estimate
∫
rj (n−1+γ)p  
α p
dσ(x) ≤ Cα p · r jn−1 ≤ Cα p · σ B(x j , r j ) ∩ Σ
Σ\B(x j ,λr j ) |x − x j |
∫
≤ C α p · 1B(x j ,λr j )∩Σ dσ
Σ
∫
 p
≤C fγ dσ. (4.3.62)
B(x j ,λr j )∩Σ

In concert, (4.3.60) and (4.3.62) imply

∫ ∫
 p   p
(b j )γ dσ ≤ C fγ dσ for each j ∈ N. (4.3.63)
Σ B(x j ,λr j )∩Σ

Consequently, (4.3.59), (4.3.63), and (4.3.8) may be combined to compute

∫  ∫  p ∫
 p   p
bγ dσ ≤ fγ 1B(x j ,λr j )∩Σ dσ ≤ C fγ dσ, (4.3.64)
Σ j ∈N Σ Oα

hence the estimate in the last line of (4.3.6) is satisfied. The proof of Lemma 4.3.2
is therefore complete. 

Along the way, we find it useful to have a density result of the sort described in
the next lemma.

Lemma 4.3.3 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and define
σ := H n−1 Σ. Then for each p ∈ n−1
n , 1 and q ∈ (0, ∞),
"
H p,q (Σ, σ) ∩ H d (Σ, σ) is dense in H p,q (Σ, σ). (4.3.65)
p<d<∞
130 4 Hardy Spaces on Ahlfors Regular Sets

Proof Pick a distribution f ∈ H p,q (Σ, σ). The goal is to show that f may be
approximated arbitrarily accurately in the space H p,q (Σ, σ) by elements
 belonging
 
to the space in the left side of (4.3.65). To this end, fix some γ ∈ (n − 1) p1 − 1 , 1
and for each threshold α ∈ (0, ∞) define
 
Oα := x ∈ Σ : fγ (x) > α . (4.3.66)

Then (4.1.22) ensures that each Oα is a relatively open subset of Σ. Also, by design,
Oα  A as α  ∞, where A := {x ∈ Σ : fγ (x) = +∞}. Since from item (a) of
[133, Lemma 6.2.5] we know that σ(Oα ) < +∞ and σ(A) = 0, it follows that there
exists α∗ ∈ (0, ∞) with the property that σ(Oα ) < σ(Σ) whenever α ∈ (α∗, ∞). In
particular,
Oα is a relatively open proper subset of Σ
if α ∈ (α∗, ∞), and lim σ(Oα ) = 0. (4.3.67)
α→∞

Observe that if there exists α ∈ (0, ∞) such that Oα =  then 0 ≤ fγ (x) ≤ α for
each x ∈ Σ. In light of (4.1.22) and (4.2.23) this entails fγ ∈ L ∞ (Σ, σ) ∩ L p,q (Σ, σ).
In concert with the natural log-convex estimates accompanying [133, (6.2.47)], this
further implies fγ ∈ L d (Σ, σ) for each d ∈ (p, ∞), hence f ∈ H d (Σ, σ) for each
d ∈ (p, ∞). Given the present aims (outlined at the beginning of the proof), the
desired conclusion trivially holds in such a scenario.
We shall therefore focus on the remaining case, when Oα   for every α ∈ (0, ∞).
In this situation, it follows from (4.3.67) that Oα is a nonempty, relatively open, proper
subset of Σ whenever α ∈ (α∗, ∞). Let us now fix α ∈ (α∗, ∞). The qualities of Oα
just mentioned allow us to perform a Whitney decomposition of Oα as in (4.3.8)-
(4.3.11). Granted this and reasoning as in the proof of [9, Theorem 5.16, p. 207-209]
we may perform a Calderón-Zygmund type decomposition of the distribution f of
the following sort,
 
f = g + b in Lipc (Σ) , (4.3.68)
 
for some g, b ∈ Lipc (Σ) satisfying two additional properties. Specifically, with
notation introduced in relation to (4.3.8)-(4.3.11), if we define
 rj n−1+γ
h(x) := for all x ∈ Σ, (4.3.69)
j ∈N
|x − x j | + r j

then there exists a geometric constant C ∈ (0, ∞) for which (cf. [9, (5.2.18), p. 208])

bγ (x) ≤ Cα · h(x) + C · fγ (x) · 1 Oα (x) for all x ∈ Σ, (4.3.70)

and (cf. [9, (5.2.20), p. 208])

gγ (x) ≤ Cα · h(x) + C · fγ (x) · 1Σ\Oα (x) for all x ∈ Σ. (4.3.71)
4.3 Real Interpolation of Hardy Spaces 131

In addition,
 from [9, Lemma 5.14, p. 202] and (4.3.8) we have that for each number
d ∈ n−1n , ∞ there exists a constant Cd ∈ (0, ∞) such that
∫   
h(x)d dσ(x) ≤ Cd · σ B(x j , r j ) ∩ Σ = Cd · σ Oα . (4.3.72)
Σ j ∈N

 n−1 
In particular, for each d ∈ n , ∞ , based on (4.3.72), (4.3.66), and [133, (6.2.21)],
we may estimate
   1/d
α · h L d (Σ,σ) ≤ Cd · α · σ Oα
   1/d
= Cd · α · σ {x ∈ Σ : (1 Oα · fγ )(x) > α}
 
≤ Cd 1 Oα · fγ  L d,∞ (Σ,σ) . (4.3.73)

As a byproduct of (4.3.73) and [133, (6.2.16), (6.2.26)] we see that

 
h ∈ L d (Σ, σ) for each d ∈ n−1 n ,∞ . (4.3.74)

Also, whenever d ∈ (p, ∞), the log-convex estimate accompanying [133, (6.2.47)]
guarantees (also keeping [133, (6.2.25)] in mind) the existence of some constant
C = C(d, p) ∈ (0, ∞) with the property that
   1−θ  θ
1Σ\O · fγ  d ≤ C 1Σ\Oα · fγ  L p, q (Σ,σ) 1Σ\Oα · fγ  L ∞ (Σ,σ)
α L (Σ,σ)
 1−θ
≤ Cα θ ·  fγ  L p, q (Σ,σ) < +∞, (4.3.75)

where θ ∈ (0, 1) is such that 1/d = (1 − θ)/p. As a result,

1Σ\Oα · fγ ∈ L d (Σ, σ) for each d ∈ (p, ∞). (4.3.76)

Combining (4.3.71), (4.3.74), and (4.3.76) yields gγ ∈ L d (Σ, σ) for each d ∈ (p, ∞)
which, in light of (4.2.5), further entails
"
g∈ H d (Σ, σ). (4.3.77)
p<d<∞

 n−1

Going further, fix d0, d1 ∈ n , ∞ such that d0 < p < d1 and pick some number
θ ∈ (0, 1) with the property that 1/p = (1 − θ)/d0 + θ/d1 . Invoking [133, (6.2.49)],
then estimating as in (4.3.73), and finally relying on [133, (6.2.26)], we may write
132 4 Hardy Spaces on Ahlfors Regular Sets

θ
α · h L p, q (Σ,σ) ≤ Cα · h L1−θ
d0 (Σ,σ) h L d1 (Σ,σ)

    (1−θ)/d0  θ     θ/d1 
≤ C α(1−θ) · σ Oα α · σ Oα
   1/p  
= Cα · σ Oα ≤ C 1 Oα · fγ  L p,∞ (Σ,σ)
 
≤ C 1 Oα · fγ  L p, q (Σ,σ), (4.3.78)

for some constant C = C(d1, d2, p) ∈ (0, ∞). In concert, (4.3.70) and (4.3.78) imply
(also keeping in mind (4.1.22)) that
   
bγ ∈ L p,q (Σ, σ) and bγ  L p, q (Σ,σ) ≤ C 1 Oα · fγ  L p, q (Σ,σ) . (4.3.79)

In particular,
 
b ∈ H p,q (Σ, σ) and b H p, q (Σ,σ) ≤ C 1 Oα · fγ  L p, q (Σ,σ) . (4.3.80)

This has two consequences. First, (4.3.80) implies that g = f − b also belongs to
H p,q (Σ, σ) which, together with (4.3.77), proves that g belongs to the space in the
left side of (4.3.65). Second,
 
 f − g H p, q (Σ,σ) = b H p, q (Σ,σ) ≤ C 1 Oα · fγ  L p, q (Σ,σ) (4.3.81)

and (4.3.80) together with [133, Lemma 6.2.8] and the last property in (4.3.67) imply
that the last expression above can be made arbitrarily small by taking α sufficiently
large. From this, the density conclusion we seek follows. 

Lemma 4.3.3 may be further augmented as follows.

Lemma 4.3.4 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and define
σ := H n−1 Σ. Then for each p ∈ n−1
n , ∞ and q ∈ (0, ∞),

H p,q (Σ, σ) ∩ Lloc

1
(Σ, σ) is dense in H p,q (Σ, σ). (4.3.82)
 
Proof If p ∈ n−1n , 1 this follows from Lemma 4.3.3, whereas if p ∈ (1, ∞) this is a
trivial consequence of (4.2.32). 

Finally, we record a useful variant of the classical Hardy inequalities.

Lemma 4.3.5 Fix M ∈ (0, ∞] and assume f : (0, M) → [0, ∞] is a Lebesgue

measurable function. Then for each p, r ∈ (0, ∞) and each q ∈ R there exists a
constant C(p, q, r) ∈ (0, ∞), independent of f and M, with the property that the
estimates
4.3 Real Interpolation of Hardy Spaces 133
#∫ M #∫ x $p $ 1/p
y q f (y) dy x −r−1 dx
0 0
#∫ $ 1/p
M  p
≤ C(p, q, r) y q+1 f (y) y −r−1 dy (4.3.83)
0

and
#∫ M #∫ M $p $ 1/p
y q f (y) dy x r−1 dx
0 x
#∫ $ 1/p
M  p
≤ C(p, q, r) y q+1 f (y) y r−1 dy (4.3.84)
0

hold provided either 1 ≤ p < ∞, or 0 < p < 1 and f is non-increasing.

Proof When M = ∞ and p ≥ 1 these are just the classical Hardy inequalities
(cf. [175, Appendix A.4]). The case M = ∞, 0 < p < 1, and f non-increasing is
proved in [180, §5] (by dyadically discretizing integrals using the monotonicity of the
integrands, then distributing powers to the terms thus created, bearing in mind that
p ∈ (0, 1)). The case when M < ∞ then follows from these results after extending
the given function f : (0, M) → [0, ∞] by zero to the entire interval (0, ∞). 

The stage is now set for us to present the proof of Theorem 4.3.1.
Proof of Theorem 4.3.1 For ease of exposition we divide the argument into several
steps.
 
Step I: The proof of (4.3.1). Fix p ∈ n−1 n , 1 along with θ ∈ (0, 1) and introduce
   
pθ := p/(1−θ) ∈ (p, ∞). Also, choose some γ ∈ (n−1) p1 −1 , 1 and fix q ∈ (0, ∞).
We shall first treat the case when Σ is unbounded. Assuming this is the case, pick
1 (Σ, σ). With f  denoting the Fefferman-Stein
an arbitrary f ∈ H pθ ,q (Σ, σ) ∩ Lloc γ
grand maximal function of f , let F be the non-increasing re-arrangement of fγ
(recall (A.0.42)), that is,
 ∗    
F(t) := fγ Σ (t) = inf λ > 0 : σ {x ∈ Σ : fγ (x) > λ} ≤ t , ∀t > 0. (4.3.85)

From [133, (6.2.41), (6.2.26)] and (4.2.24) it follows that, for some C ∈ (0, ∞)
independent of f ,

F(t) ≤ Ct −1/pθ  f  H p θ , q (Σ,σ) for each t ∈ (0, ∞). (4.3.86)

For each fixed t ∈ (0, ∞), the idea is now to apply Lemma 4.3.2 for the threshold
α := F(t p ). In this regard, note that (4.3.86) implies α ∈ [0, ∞). Whenever applicable,
Lemma 4.3.2 allows us to decompose f as

f = gt + bt with gt ∈ L ∞ (Σ, σ) and bt ∈ Lloc

1
(Σ, σ) (4.3.87)
134 4 Hardy Spaces on Ahlfors Regular Sets

satisfying, for some constant C ∈ (0, ∞) independent of t,

|gt | ≤ C · F(t p ) for σ-a.e. point on Σ, and

∫ ∫
 p   p (4.3.88)
(bt )γ dσ ≤ C fγ dσ,
Σ Ut

where  
Ut := Oα = x ∈ Σ : fγ (x) > F(t p ) . (4.3.89)
Note that Ut is a relatively open subset of Σ, hence σ-measurable. Also, whenever
α > 0, Chebytcheff’s inequality gives that
∫
 
σ(Ut ) = σ Oα ≤ α−p ( fγ ) p dσ < +∞, (4.3.90)
Σ

where the last inequality uses the fact that f ∈ H p (Σ, σ). In particular, having
σ(Ut ) < +∞ forces Ut to be a proper subspace of the unbounded Ahlfors regular
set Σ. This property guarantees the applicability of Lemma 4.3.2. In relation to
(4.3.87), it is also useful to observe that thanks to (4.3.88), [133, (6.2.5)] (whose
validity is ensured by the fact that F(t p ) < +∞, as seen from (4.3.86)), and (4.3.86)
we have
∫ ∫ ∫ tp
 p   p
(bt )γ dσ ≤ C fγ dσ ≤ C F(s) p ds
Σ Ut 0
∫ tp
s−p/pθ ds
p
≤ C f  H p θ , q (Σ,σ)
0
∫ tp
s θ−1 ds
p
= C f  H p θ , q (Σ,σ)
0

= C f  H p θ , q (Σ,σ) t θ ·p < +∞.

p
(4.3.91)

In view of (4.2.5), this implies that

∫ tp 1/p
bt ∈ H (Σ, σ) and bt 
p
H p (Σ,σ) ≤C F(s) p ds . (4.3.92)
0

Moving on, recall from (1.3.34) that the K-functional is presently given by
 
K(t, f ) := K t, f , H p (Σ, σ), L ∞ (Σ, σ) (4.3.93)
 
= inf b H p (Σ,σ) + t g L ∞ (Σ,σ) : f = b + g, b ∈ H p (Σ, σ), g ∈ L ∞ (Σ, σ) .

Then, on account of (4.3.87)-(4.3.88) and (4.3.92), we may estimate

4.3 Real Interpolation of Hardy Spaces 135
∫ tp 1/p
K(t, f ) ≤ C F(s) p ds + Ct · F(t p ). (4.3.94)
0

Next, we use (1.3.38) to write

∫ ∞   q dt 1/q
 f (H p (Σ,σ), L ∞ (Σ,σ))θ, q = t −θ K(t, f )
0 t
%∫ ∫ & 1/q
∞ tp q/p
−θq dt
≤C t p
F(s) ds
0 0 t
∫ ∞   q dt 1/q
+C t (1−θ)q F(t p ) =: I + I I. (4.3.95)
0 t
To treat I, we make the change of variables ρ := t p and then apply Hardy’s inequality,
as stated in Lemma 4.3.5, to write
∫ ∞
dt 1/q
I≤C t (1−θ)q/p F(t)q
0 t
   
= C t 1/pθ F(t) L q (0,∞), dt  = C  fγ  L p θ , q (Σ,σ) . (4.3.96)
t

Note that after making the change of variables ρ := t p in I I we also obtain, much as
above,
∫ ∞
dρ 1/q
II ≤ C ρ(1−θ)q/p F(ρ)q
0 ρ
   
= C  ρ1/pθ F(ρ) L q (0,∞), dt  = C  fγ  L p θ , q (Σ,σ) . (4.3.97)
t

Let us summarize our progress. Combining (4.3.95)-(4.3.97) we arrive at the

conclusion that there exists some constant C ∈ (0, ∞) with the property that
 
 f (H p (Σ,σ), L ∞ (Σ,σ))θ, q ≤ C  fγ  L p θ , q (Σ,σ) = C f  H p θ , q (Σ,σ),
(4.3.98)
for each f ∈ H pθ ,q (Σ, σ) ∩ Lloc
1 (Σ, σ).

Consider now f ∈ H pθ ,q (Σ, σ) arbitrary. The density result from Lemma 4.3.4
ensures the existence of a sequence { f j } j ∈N ⊆ H pθ ,q (Σ, σ) ∩ Lloc
1 (Σ, σ) such that

f j −→ f in H pθ ,q (Σ, σ) as j → ∞. (4.3.99)

Thus, { f j } j ∈N is Cauchy in H pθ ,q (Σ, σ) which, in light of (4.3.98), forces it to be

a Cauchy sequence in the intermediate space H p (Σ, σ), L ∞ (Σ, σ) θ,q . Given that
the latter
 is complete (cf. [15,  Theorem 3.4.2(b), p. 47]), it follows that there exists 
h ∈ H p (Σ, σ), L ∞ (Σ, σ) θ,q such that lim f j = h in H p (Σ, σ), L ∞ (Σ, σ) θ,q .
j→∞
Recall that the latter space embeds continuously in the ambient topological space
136 4 Hardy Spaces on Ahlfors Regular Sets

within which the interpolation process takes place (cf. (1.3.1)), which we presently
take to be the space of distributions on Σ, i.e.,
 p   
H (Σ, σ), L ∞ (Σ, σ) θ,q → Lipc (Σ) continuously. (4.3.100)
 
This further implies that f j → h in Lip
 c (Σ)  as j → ∞. Since, thanks to (4.3.99)
and (4.2.29), we also have f j → f in Lipc (Σ) as j → ∞, we finally conclude that
h = f as distributions on Σ. From this, we ultimately conclude that if Σ is unbounded
we have
 
H pθ ,q (Σ, σ) → H p (Σ, σ), L ∞ (Σ, σ) θ,q continuously. (4.3.101)

We claim that the inclusion in (4.3.101) continues to be valid in the case when
Σ is bounded. Observe that in such a scenario we have σ(Σ) < +∞ which further
implies
L ∞ (Σ, σ) → H p (Σ, σ) continuously, (4.3.102)
thanks to (4.2.5) and (4.1.23). Granted this embedding, Lemma 1.3.4 applied with
  1/p
t∗ := σ(Σ) ∈ (0, ∞) (4.3.103)

gives that
∫ t∗   q dt 1
q
 f (H p (Σ,σ), L ∞ (Σ,σ))θ, q ≈ t −θ K(t, f ) , (4.3.104)
0 t

uniformly for f ∈ H p (Σ, σ). Since (4.3.89) and [133, (6.2.3)] (used here with f
replaced by fγ and t replaced by t p ) presently imply
 
σ(Ut ) = σ x ∈ Σ : fγ (x) > F(t p ) ≤ t p for each t ∈ (0, ∞), (4.3.105)

it follows from (4.3.103) and (4.3.105) that σ(Ut ) < σ(Σ) for every t ∈ (0, t∗ ).
Consequently,

Ut is a proper subset of Σ, for every t ∈ (0, t∗ ). (4.3.106)

Having established (4.3.106), we may invoke Lemma 4.3.2 with α := F(t p ) ∈ [0, ∞)
for each t in the range (0, t∗ ). As a result, we are still able to decompose f as in
(4.3.87)-(4.3.88), whenever t ∈ (0, t∗ ). Then, if q ∈ (0, ∞), on account of (4.3.93)
and (4.3.87)-(4.3.88), we may continue to estimate K(t, f ) as in (4.3.94) for each
t ∈ (0, t∗ ). With this estimate in hand, we may then run the same argument which
has produced (4.3.96)-(4.3.97), by once gain relying on Hardy’s inequality (as in
Lemma 4.3.5, now used with M := t∗ ), to conclude that (4.3.98) continues to be
valid in the case when Σ is bounded. Ultimately, the inclusion in (4.3.101) remains
true when Σ is bounded.
Let us record our progress. The argument so far proves that the inclusion in
(4.3.101) is valid for Σ as in the statement of the theorem. To establish the opposite
4.3 Real Interpolation of Hardy Spaces 137

inclusion in (4.3.101), observe that the grand maximal function induces well-defined,
sub-linear, bounded operators

H p (Σ, σ)  f −→ fγ ∈ L p (Σ, σ), (4.3.107)

thanks to (4.2.1)-(4.2.6), as well as

L ∞ (Σ, σ)  f −→ fγ ∈ L ∞ (Σ, σ), (4.3.108)

thanks to (4.1.23). On account of (4.3.107)-(4.3.108) and the real interpolation result

from Proposition 1.3.7, we then conclude that
 
H p (Σ, σ), L ∞ (Σ, σ) θ,q  f −→ fγ ∈ L pθ (Σ, σ) (4.3.109)

is a well-defined bounded sub-linear mapping. In light of the definition of Lorentz-

based Hardy spaces, this implies
 p 
H (Σ, σ), L ∞ (Σ, σ) θ,q → H pθ ,q (Σ, σ) continuously. (4.3.110)

At this stage, (4.3.1) follows from (4.3.98) and (4.3.110).

Step II: The proof of (4.3.2) in the case when q0, q1, q ∈ (0, ∞). In such a sce-
nario, (4.3.2) is a consequence of (4.3.1) and the reiteration theorem for the real
method of interpolation recalled in Theorem 1.3.9.
Step III: The proof of (4.3.3) in the case when q = ∞. Fix θ ∈ (0, 1) and suppose

n−1
n < p0 < p < p1 < ∞ satisfy 1
p = 1−θ
p0 + θp . (4.3.111)
   
Also, choose some γ ∈ (n − 1) p10 − 1 , 1 , and define
  p
η := θ 1
p0 − 1
p1 ∈ (0, ∞), β := 1 − p1 ∈ (0, 1). (4.3.112)

The goal is to prove that, with equivalence of quasi-norms,

 p0 
H (Σ, σ), H p1 (Σ, σ) θ,∞ = H p,∞ (Σ, σ). (4.3.113)

We shall first deal with the case when Σ is unbounded. In such a scenario, pick
an arbitrary f ∈ H p,∞ (Σ, σ) and denote by F the non-increasing re-arrangement of
fγ (cf. (4.3.85)). As was the case with (4.3.86), we presently have

F(t) ≤ Ct −1/p  f  H p,∞ (Σ,σ) for each t ∈ (0, ∞). (4.3.114)

The strategy is to invoke, for each fixed t ∈ (0, ∞), the Calderón-Zygmund type
decomposition of the distribution f as in [9, Theorem 5.16, p. 207-209] at level
α := F(t η ) ∈ [0, ∞). Specifically, for each t ∈ (0, ∞) this allows us to split
 
f = gt + bt in Lipc (Σ) , (4.3.115)
138 4 Hardy Spaces on Ahlfors Regular Sets
 
for some gt , bt ∈ Lipc (Σ) satisfying (cf. [9, (5.2.18) and (5.2.20), p. 208])

(bt )γ ≤ Cα · ht + C · fγ · 1 Oα on Σ, (4.3.116)

(gt )γ ≤ Cα · ht + C · fγ · 1Σ\Oα on Σ, (4.3.117)

where C ∈ (0, ∞) is a geometric constant independent of f and t, the function ht is

as in (4.3.69), and the level set Oα is given by
 
Oα := x ∈ Σ : fγ (x) > α = F(t η ) . (4.3.118)

The applicability of [9, Theorem 5.16, p. 207-209] (which ensures the existence
of such a decomposition) requires that in the case when α > 0 the set Oα does
not coincide with Σ. Given that σ(Σ) = +∞ (since we are presently assuming Σ
to be unbounded), the latter condition is satisfied since, as seen from Chebytcheff’s
inequality, [133, (6.2.5)], and (4.3.114), we have
∫ ∫ tη
 
σ Oα ≤ α −p0
( fγ ) p0 dσ ≤ Cα −p0
F(s) p0 ds
Oα 0
∫ tη
≤ Cα−p0  f  H0p,∞ (Σ,σ) s−p0 /p ds
p
0

= Cα−p0  f  H0p,∞ (Σ,σ) t θ ·p0 < +∞,

p
(4.3.119)

where the last inequality uses the fact that f ∈ H p,∞ (Σ, σ).
Having justified the decomposition of the  distribution
 f as in (4.3.115)-(4.3.118),
we recall from (4.3.72) that for each d ∈ n−1 n , ∞ there exists Cd ∈ (0, ∞) indepen-
dent of t such that
    1/d
ht  L d (Σ,σ) ≤ Cd · σ Oα . (4.3.120)

Two consequences of (4.3.120) are of a particular interest to us. First, based on

(4.3.120) with d := p0 , (4.3.118), and [133, (6.2.21), (6.2.27)], we may estimate
   1/p0
α · ht  L p0 (Σ,σ) ≤ Cα · σ Oα
   1/p0
= Cα · σ {x ∈ Σ : ( fγ · 1 Oα )(x) > α}
   
≤ C  fγ · 1 Oα  L p0,∞ (Σ,σ) ≤ C  fγ · 1 Oα  L p0 (Σ,σ), (4.3.121)

for some constant C ∈ (0, ∞) independent of t. Second, based on (4.3.120) with

d := p1 , (4.3.118), [133, (6.2.21)], and (4.2.24), we conclude that, once again for
some C ∈ (0, ∞) independent of t,
4.3 Real Interpolation of Hardy Spaces 139
   1/p1     1/p1
α · ht  L p1 (Σ,σ) ≤ Cα · σ Oα = Cα β · α1−β · σ Oα
'    ( 1−β  1−β
= Cα β · α · σ Oα ≤ Cα β  fγ  L p,∞ (Σ,σ)
1/p

= Cα β  f  H p,∞ (Σ,σ) .
1−β
(4.3.122)

We augment (4.3.121)-(4.3.122) with another useful estimate. Specifically, keeping

this in mind that 1/p1 = (1 − β)/p and invoking [133, (6.2.49)] implies (by also
taking into account (4.3.118))
    1−β  β
 fγ · 1Σ\O  ≤ C  fγ · 1Σ\Oα  L p,∞ (Σ,σ) ·  fγ · 1Σ\Oα  L ∞ (Σ,σ)
α L p1 (Σ,σ)
 1−β
≤ C  fγ  L p,∞ (Σ,σ) · α β = Cα β  f  H p,∞ (Σ,σ)
1−β
(4.3.123)

for some constant C ∈ (0, ∞) independent of t.

The nest step is to make use of (4.3.121)-(4.3.123) to estimate bt  H p0 (Σ,σ) and
gt  H p1 (Σ,σ) . First, from (4.3.116), (4.3.121), and [133, (6.2.5)], we obtain
   
(bt )γ  ≤ Cαht  L p0 (Σ,σ) + C  fγ · 1 Oα  L p0 (Σ,σ)
L p0 (Σ,σ)
∫
    p0 1/p0
≤ C  fγ · 1 Oα  L p0 (Σ,σ) = C fγ dσ
Oα
∫ tη 1/p0
≤C F(s) p0 ds . (4.3.124)
0

In view of (4.3.119) and (4.2.5), this implies that for some constant C ∈ (0, ∞)
independent of t we have
∫ tη 1/p0
bt ∈ H p0 (Σ, σ) and bt  H p0 (Σ,σ) ≤ C F(s) p0 ds . (4.3.125)
0

Second, from (4.3.117), (4.3.122), and (4.3.123) we see that

 
(gt )γ  p ≤ Cα β  f  H p,∞ (Σ,σ) .
1−β
(4.3.126)
L 1 (Σ,σ)

Collectively, (4.2.5), (4.3.126), and the earlier definitions of α and β imply that for
some constant C ∈ (0, ∞) independent of t we have
  1−(p/p1 )
gt ∈ H p1 (Σ, σ) and gt  H p1 (Σ,σ) ≤ C F(t η )
p/p
·  f  H p,∞1 (Σ,σ) . (4.3.127)

Pressing on, recall from (1.3.34) that for each t ∈ (0, ∞) the K-functional is
presently given by
140 4 Hardy Spaces on Ahlfors Regular Sets
 
K(t, f ) := K t, f , H p0 (Σ, σ), H p1 ∞(Σ, σ) (4.3.128)
 
= inf b H p0 (Σ,σ) + t g H p1 (Σ,σ) : f = b + g, b ∈ H p0 (Σ, σ), g ∈ H p1 (Σ, σ) .

On account of (4.3.128), (4.3.115), (4.3.125), and (4.3.127), for each t ∈ (0, ∞) we

may therefore estimate (with η as in (4.3.112))
∫ tη 1/p0   1−(p/p1 )
+ Ct · F(t η )
p/p
K(t, f ) ≤ C F(s) p0 ds ·  f  H p,∞1 (Σ,σ) (4.3.129)
0

for some constant C ∈ (0, ∞) independent of t. Based on (1.3.38) and (4.3.129) we

may now estimate
 
 f (H p0 (Σ,σ), H p1 (Σ,σ))θ,∞ = sup t −θ K(t, f )
t>0
) ∫ *
tη 1/p0
−θ
≤ C · sup t F(s) ds p0
t>0 0
'   1−(p/p1 ) (
+ C f  H p,∞1 (Σ,σ) · sup t (1−θ) · F(t η )
p/p
t>0

=: I I I + IV . (4.3.130)

Note that, thanks to (A.0.60) and (4.2.5), we have

) ∫ tη *
−θ
 1/p p 1/p0
I I I = C · sup t s F(s) 0 s−p0 /p ds
t>0 0
) ∫ *
  tη 1/p0
≤ C · sup s1/p F(s) · sup t −θ s−p0 /p ds
s>0 t>0 0

   
= C · sup s1/p F(s) = C  fγ  L p,∞ (Σ,σ)
s>0

= C f  H p,∞ (Σ,σ), (4.3.131)

where the first equality uses

 the fact that, as seen from (4.3.111)-(4.3.112), we have
p0 /p ∈ (0, 1) and −θ + 1 − (p0 /p) (η/p0 ) = 0. In addition, after changing s := t η
   −1
and observing that (4.3.111)-(4.3.112) imply (1 − θ)/η 1 − (p/p1 ) = 1/p, we
conclude that
4.3 Real Interpolation of Hardy Spaces 141
'   1−(p/p1 ) (
IV = C f  H p,∞1 (Σ,σ) · sup s(1−θ)/η · F(s)
p/p
s>0

p/p   1−(p/p1 )
= C f  H p,∞1 (Σ,σ) · sup s1/p F(s)
s>0
 1−(p/p )
·  fγ  p,∞ 1
p/p
= C f  H p,∞1 (Σ,σ) L (Σ,σ)

p/p 1−(p/p )
= C f  H p,∞1 (Σ,σ) ·  f  H p,∞ (Σ,σ)
1

= C f  H p,∞ (Σ,σ) . (4.3.132)

From (4.3.130)-(4.3.132) we arrive at the conclusion that there exists some constant
C ∈ (0, ∞) with the property that

 f (H p0 (Σ,σ), H p1 (Σ,σ))θ,∞ ≤ C f  H p,∞ (Σ,σ) for all f ∈ H p,∞ (Σ, σ). (4.3.133)

Since, by definition,
 p0 
H (Σ, σ),H p1 (Σ, σ) θ,∞
  
= f ∈ H p0 (Σ, σ) + H p1 (Σ, σ) ⊆ Lipc (Σ) :

 f (H p0 (Σ,σ), H p1 (Σ,σ))θ,∞ < +∞ , (4.3.134)

this ultimately proves that if Σ is unbounded then

 
H p,∞ (Σ, σ) → H p0 (Σ, σ), H p1 (Σ, σ) θ,∞ continuously. (4.3.135)

We claim that the inclusion in (4.3.135) remains well-defined and continuous

when Σ is bounded. Since in such a scenario (4.2.21) implies

H p1 (Σ, σ) → H p0 (Σ, σ) continuously, (4.3.136)

Lemma 1.3.4 applied with

  1/η
t∗ := σ(Σ) ∈ (0, ∞) (4.3.137)

gives that
 
 f (H p0 (Σ,σ), H p1 (Σ,σ))θ,∞ ≈ sup t −θ K(t, f ) , (4.3.138)
t ∈(0,t∗ )

with proportionality constants independent of f . Since (4.3.89) and [133, (6.2.3)]

(presently used with f replaced by fγ and t replaced by t p ) presently imply
 
σ(Oα ) = σ x ∈ Σ : fγ (x) > F(t η ) ≤ t η for each t ∈ (0, ∞), (4.3.139)
142 4 Hardy Spaces on Ahlfors Regular Sets

it follows from (4.3.137) and (4.3.139) that σ(Oα ) < σ(Σ) for every t ∈ (0, t∗ ).
Hence,
Oα is a proper subset of Σ, for every t ∈ (0, t∗ ). (4.3.140)
In turn, (4.3.140) guarantees that we are still able to decompose the distribution
f as in (4.3.115)-(4.3.118) for each t ∈ (0, t∗ ). Consequently, the estimate for the
K-functional in (4.3.129) continues to be valid for t ∈ (0, t∗ ). Granted this estimate,
we may then run the same argument which has led to (4.3.131)-(4.3.132). Ultimately,
this proves that the inclusion in (4.3.135) is also valid when Σ is bounded.
As regards the opposite inclusion in (4.3.135), observe that thanks to (4.2.5) the
grand maximal function induces well-defined, sub-linear, bounded operators

H p0 (Σ, σ)  f −→ fγ ∈ L p0 (Σ, σ), (4.3.141)

H p1 (Σ, σ)  f −→ fγ ∈ L p1 (Σ, σ). (4.3.142)

Based on (4.3.141)-(4.3.142) and the real interpolation result from Proposition 1.3.7
and [133, (6.2.48)] (both used with q = ∞) we then conclude that
 
H p0 (Σ, σ), H p1 (Σ, σ) θ,∞  f −→ fγ ∈ L p,∞ (Σ, σ) (4.3.143)

is a well-defined bounded sub-linear mapping. Upon recalling the definition of

Lorentz-based Hardy spaces, this implies
 p0 
H (Σ, σ), H p1 (Σ, σ) θ,∞ → H p,∞ (Σ, σ) continuously. (4.3.144)

Finally, (4.3.135) and (4.3.144) justify (4.3.113).

Step IV: The proof of (4.3.3) in full generality. When q ∈ (0, ∞) then the interpo-
lation formula in (4.3.3) follows from (4.3.2) with q0 := p0 and q1 := p1 (a scenario
which has been already dealt with in Step II), bearing in mind (4.2.25). Finally,
formula (4.3.3) with q = ∞ has been proved in Step III.
Step V: The proof of (4.3.2) in full generality. The fact that formula (4.3.2) holds
as stated follows from (4.3.3) and the reiteration theorem for the real method of
interpolation recorded in Theorem 1.3.9.
The proof of Theorem 4.3.1 is therefore complete. 

Once Theorem 4.3.1 has been established, from (4.3.3), (1.3.44), and (1.3.45) we
conclude that
if Σ ⊆ Rn is a closed set which is Ahlfors regular and σ := H n−1 Σ,
then H p0 (Σ, σ) ∩ H p1 (Σ, σ) embeds continuously (cf. (1.3.3)) and
(4.3.145)
densely into the space H p,q (Σ, σ) granted n−1
n < p0 < p < p1 < ∞
and q ∈ (0, ∞).
Also, from Theorem 4.3.1 and (1.3.41) we see that
4.4 Atomic Decompositions for Hardy Spaces 143

if Σ ⊆ Rn is a closed set which is Ahlfors regular and σ := H n−1 Σ,

then H p,q (Σ, σ) embeds continuously into H p0 (Σ, σ) + H p1 (Σ, σ) (4.3.146)
whenever n−1n < p0 < p < p1 < ∞ and q ∈ (0, ∞].

4.4 Atomic Decompositions for Hardy Spaces

We begin by specializing results from [9] pertaining to atomic decompositions of

Hardy spaces to the setting of Ahlfors regular subsets of the Euclidean ambient. To
set the stage, given any closed subset Σ of Rn along with some β ∈ (0, ∞) define
 β
C (Σ) if Σ is bounded,
L β (Σ) := .β (4.4.1)
C (Σ)/∼ if Σ is unbounded,

where, as before, the equivalence identifies functions which differ by a constant on

Σ. Next, assume that Σ ⊆ Rn is a closed set which is Ahlfors regular and abbreviate
σ := H n−1 Σ. In this context, given two exponents p ∈ (0, 1] and q ∈ [1, ∞] with
q > p, call a σ-measurable function a : Σ → C a (p, q)-atom provided there exist a
point x ∈ Σ and a number r ∈ (0, 2 diam Σ) with the property that
∫
  1/q−1/p
supp a ⊆ B(x, r) ∩ Σ, a L q (Σ,σ) ≤ σ B(x, r) ∩ Σ , a dσ = 0. (4.4.2)
Σ

It is also agreed that

in the case when Σ is bounded the constant function
a(x) := [σ(Σ)]−1/p for every x ∈ Σ is also considered (4.4.3)
to be a (p, q)-atom.
p,q
We then define the atomic Hardy space Hat (Σ, σ) as
  ∗
Hat (Σ, σ) := f ∈ L (n−1)(1/p−1) (Σ) : there exist {λ j } j ∈N ∈  p (N)
p,q

along with a sequence {a j } j ∈N of (p, q)-atoms


∞
 ∗
such that f = λ j a j in L (n−1)(1/p−1) (Σ) (4.4.4)
j=1

when p ∈ (0, 1) and, corresponding to the case when p = 1,


1,q
Hat (Σ, σ) := f ∈ L 1 (Σ, σ) : there exist {λ j } j ∈N ∈  1 (N)and (1, q)-atoms


∞ 
{a j } j ∈N such that f = λ j a j in L 1 (Σ, σ) . (4.4.5)
j=1
144 4 Hardy Spaces on Ahlfors Regular Sets
p,q
In all cases, we equip Hat (Σ, σ) with the quasi-norm


∞ 1/p
 f  Hatp, q (Σ,σ) := inf |λ j | p (4.4.6)
j=1

!∞ p,q
where the infimum is taken over all writings f = j=1 λ j a j of f ∈ Hat (Σ, σ) as in
(4.4.4)-(4.4.5). The above definitions imply that

p,q !
∞  
if f ∈ Hat (Σ, σ) is expanded as f = λ j a j in Lipc (Σ) for some
j=1
sequence {λ j } j ∈N ∈  1 (N) and with each a j a (p, q)-atom on Σ, then (4.4.7)
!
∞
p,q
the series λ j a j also converges to f in Hat (Σ, σ).
j=1

Theorem 4.4.1 Given a closed set Σ ⊆ Rn which is Ahlfors regular, abbreviate

σ := H n−1 Σ, and select p, q, γ such that
 
n < p ≤ 1, 1 ≤ q ≤ ∞, p < q, and (n − 1) p1 − 1 < γ < 1.
n−1
(4.4.8)

Associated with these parameters, consider the Hardy space H p (Σ, σ) defined using
the grand-maximal function as in Definition 4.2.1, as well as the atomic Hardy space
p,q
Hat (Σ, σ) defined as in (4.4.4)-(4.4.5).
p,q
Then H p (Σ, σ) may be naturally identified with Hat (Σ, σ) both algebraically
and quantitatively (in the sense of equivalence of quasi-norms). In particular, these
spaces do not depend on the choice of the exponent q and the index γ as in (4.4.8).

Proof This follows by combining [9, Theorem 5.27, p. 263] with [9, Theorem 7.5,
p. 302]. 

As a consequence of Theorem 4.4.1 and (4.4.5) we see that

if Σ ⊆ Rn is a closed Ahlfors regular set and σ := H n−1 Σ, then

H 1 (Σ, σ) embeds continuously into
 L (Σ,1σ); moreover,
1
∫ if in addition
 (4.4.9)
Σ is unbounded, then H (Σ, σ) ⊂ f ∈ L (Σ, σ) : Σ f dσ = 0 .
1

In particular, from (4.4.9) and (4.2.12) we deduce that

 ∫ 
.
H 1 (Σ, σ) = f ∈ H 1 (Σ, σ) : f dσ = 0 . (4.4.10)
Σ

Theorem 4.4.1 also implies, in concert with (4.2.21), that

if Σ ⊆ Rn is a closed set which is Ahlfors regular and σ := H n−1 Σ

then H p1 (Σ, σ) → H p0 (Σ, σ) continuously and densely (4.4.11)
provided Σ is bounded and n−1
n < p0 ≤ p1 < ∞.
4.4 Atomic Decompositions for Hardy Spaces 145

For later purposes it is useful to know that Hardy spaces are separable quasi-
Banach spaces, an issue addressed in our next proposition.

Proposition 4.4.2 Let Σ ⊆ Rn be a closed set

 which is Ahlfors regular and define
σ := H n−1 Σ. Then, for each p ∈ n−1n , ∞ , the quasi-Banach space H p (Σ, σ) is

separable.

Proof When p ∈ (1, ∞), the fact that H p (Σ, σ) is separable is a consequence
 of
(4.2.9) and [133, (3.6.27)]. Consider next the case when p ∈ n−1 n , 1 . For now,
suppose Σ is unbounded. Fix a point x0 ∈ Σ along with an exponent r ∈ (1, ∞) and,
for each j ∈ N, abbreviate Δ j := B(x0, j) ∩ Σ. From [133, Lemma 3.6.4] we know
that L r (Δ j , σ) is separable for each j ∈ N. If we now introduce
 ∫ 
L0 (Δ j , σ) := f ∈ L (Δ j , σ) :
r r
f dσ = 0 , ∀ j ∈ N, (4.4.12)
Δj

the fact that any subset of a separable metric space is separable implies that

for each j ∈ N the space L0r (Δ j , σ) is separable. (4.4.13)

Granted this, for each j ∈ N we may pick a set

Fj which is a countable and dense subset of L0r (Δ j , σ). (4.4.14)

Next, extend each function in Fj by zero to the entire Σ and denote the collection of
all these extensions by Fj . Then every f ∈ Fj is a multiple of a (p, r)-atom on Σ,
hence Fj ⊂ H p (Σ, σ) for each j ∈ N. Let us record our progress:

if F denotes the collection of all finite sums of functions in j ∈N Fj , (4.4.15)
then F is a countable subset of H p (Σ, σ) which is stable under addition.
The claim we make at this stage is that

F is dense in H p (Σ, σ). (4.4.16)

Keeping in mind the atomic characterization of H p (Σ, σ) from Theorem 4.4.1, this
claim will follow as soon as we prove that

for every (p, r)-atom a on Σ, every λ ∈ R, and every ε > 0,

there exists some f ∈ F such that λ a − f  H p (Σ,σ) < ε. (4.4.17)

To this end, fix a, λ, and ε as in (4.4.17). Then there exists some natural number
j0 ∈ N such that supp a ⊆ Δ j0 and (λa) Δ j ∈ L0r Δ j0 , σ . Relying on (4.4.14)
0
(presently used with j = j0 ) consider f ∈ Fj0 such that

λ a − f  L r (Δ j0 ,σ) < ε · σ(Δ j0 )1/r−1/p . (4.4.18)

146 4 Hardy Spaces on Ahlfors Regular Sets

If f denotes the extension of f by zero to Σ, it follows that f ∈ Fj0 ⊆ F . Moreover,

the function g := λ a − f satisfies
∫
supp g ⊆ Δ j0 , g dσ = 0, and g L r (Σ,σ) < ε · σ(Δ j0 )1/r−1/p . (4.4.19)
Δ j0

The properties in (4.4.19) imply that ε −1 g is a (p, r)-atom on Σ. Consequently,

ε −1 g H p (Σ,σ) ≤ C for some constant C ∈ (0, ∞) which depends only on Σ, p, and
r. In turn, this inequality translates into λ a − f  H p (Σ,σ) < Cε. After re-adjusting
ε, this completes the proof of (4.4.17) which, in turn, finishes the proof of the fact
that F is dense in H p (Σ, σ). The desired conclusion now follows in the case when
Σ is unbounded.
Finally, when Σ is bounded, the argument proceeds as before with one minor
adjustment. Specifically, since we now are also allowing constant functions to be

multiples of (p, r)-atoms, we enrich the union j ∈N Fj by also including the constant
functions with rational values. Reasoning as before then shows that taking F to the
collection of all finite sums of functions in this family continues yields a countable
and dense subset of H p (Σ, σ). This finishes the proof of Proposition 4.4.2. 

It turns out that if a given distribution f in some Hardy space additionally

belongs to a Lebesgue space, or another Hardy space, then one may perform an
atomic decomposition which converges to f simultaneously in all said spaces. This
is made precise in the theorem below.

Theorem 4.4.3 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and define
σ := H n−1 Σ. Suppose n−1 n < p ≤ 1, n < q < ∞, 1 < r < ∞, and pick
n−1

some arbitrary f ∈ H (Σ, σ) ∩ H (Σ, σ). Then there exist a numerical sequence
p q

{λ j } j ∈N ∈  p (N) along with a sequence of (p, r)-atoms {a j } j ∈N on Σ such that


∞ 1/p
|λ j | p ≤ C f  H p (Σ,σ) (4.4.20)
j=1

for some finite constant C > 0 which depends only on the ambient, and

∞
f = λ j a j with convergence both in H p (Σ, σ) and in H q (Σ, σ). (4.4.21)
j=1

As a consequence,
!
∞
f = λ j a j with convergence in H p (Σ, σ) ∩ H q (Σ, σ) e-
j=1 (4.4.22)
quipped with the intersection quasi-norm (cf. (1.3.3)).
It is also possible to allow r = ∞ in the above theorem. Indeed, as observed
in [159], the argument from [176, pp. 107-109] yields just that. Originally carried
4.4 Atomic Decompositions for Hardy Spaces 147

out in the whole Euclidean space as ambient, said argument is robust enough to be
adapted to the setting considered here.
Proof of Theorem 4.4.3 First we consider the case when Σ is unbounded. We begin
by reviewing the technology of decomposing a distributions belonging to a Hardy
space as an infinite linear span of atoms. The approach we take hinges on the existence
of a family {St }t>0 of integral operators on Σ which constitute an approximation to
the identity of order ε1 ∈ (0, 1]. This means that, for each t ∈ (0, ∞), St is an integral
operator of the form
∫
St f (x) := St (x, y) f (y) dσ(y), x ∈ Σ, (4.4.23)
Σ

with an integral kernel St : Σ × Σ → R satisfying, for some C ∈ (0, ∞) independent

of t, the following properties:
(i) 0 ≤ St (x, y) ≤ Ct 1−n for all x, y ∈ Σ, and St (x, y) = 0 if |x − y| ≥ Ct;
(ii) |St (x, y) − St (x , y)| ≤ Ct −(n−1+ε1 ) |x − x |∫ε1 for every x, x , y ∈ Σ;
(iii) St (x, y) = St (y, x) for every x, y ∈ Σ, and Σ St (x, y) dσ(y) = 1 for every x ∈ Σ.
The construction of such an approximation to the identity for any given exponent
ε1 ∈ (0, 1] is carried out in detail in [9, Theorem 3.22, pp. 102-103], following
R. Coifman’s template, outlined in [43]. Specifically, fix a function h ∈ C 1 (R) with
the property that 0 ≤ h ≤ 1 pointwise on R, h ≡ 1 on [−1/2, 1/2], and h ≡ 0
on R \ (−2, 2). Next, for each t ∈ (0, ∞), let Tt be the integral operator acting on
functions f defined on Σ according to
∫
|x − y|
(Tt f )(x) := t −(n−1) h f (y) dσ(y), x ∈ Σ. (4.4.24)
Σ t
Based on properties of the function h and the Ahlfors regularity of Σ, it is straight-
forward to check that there exists a finite constant Co ≥ 1 such that

Co−1 ≤ (Tt 1)(x) ≤ Co for each x ∈ X and each t ∈ (0, ∞). (4.4.25)

Keeping this in mind, for each t ∈ (0, ∞) it is then meaningful to define for each
x, y ∈ Σ
∫    
t −2(n−1) h |x − z|/t) h |z − y|/t
St (x, y) := dσ(z). (4.4.26)
(Tt 1)(x)(Tt 1)(y) T 1 (z)
Σ t Tt 1

Based on this formula, one may check that properties (i)-(iii) above hold. More-
over, if for each t ∈ (0, ∞) we now define the integral kernel

∂  
Dt (x, y) := t St (x, y) , x, y ∈ Σ, (4.4.27)
∂t
a direct computation shows that Dt (x, y) behaves much like St (x, y), both in terms
of size and regularity. Ultimately, this permits us to conclude that the family
148 4 Hardy Spaces on Ahlfors Regular Sets
 
Dt (x, y) t>0 satisfies properties similar to those recorded in (i)-(iii)
∫ above, the only
exception being that we now have the cancellation condition Σ Dt (x, y) dσ(y) = 0
for every x ∈ Σ. Hence, if for each t > 0 we define the integral operator
∫
Dt f (x) := Dt (x, y) f (y) dσ(y), x ∈ Σ, (4.4.28)
Σ

it follows that for every ε2 > 0 the family {Dt }t>0 is, in the language of [88, Main
Assumption, p. 1510], a Calderón Reproducing Formula family of order (ε1, ε2 ) on
Σ (or a (ε1, ε2 )-CRF, for short). This terminology is inspired by the fact that each Dt
is a bounded operator on L po (Σ, σ) for every po ∈ (1, ∞) and the following identity,
in the spirit of the classical Calderón reproducing formula, holds (see [83]-[84], and
also [88, Proposition 2.13, p. 1517])
∫ ∞
dt
f = Dt2 f , ∀ f ∈ L po (Σ, σ) with po ∈ (1, ∞). (4.4.29)
0 t

Henceforth fix ε1 ∈ (0, 1]. Having also fixed an arbitrary background parameter

N ∈ (0, ∞), to be specified later, it is convenient to further decompose each kernel
Dt (x, y) as a weighted superposition of bump functions of the following sort:

∞
Dt (x, y) = 2−N ϕ2 t (x, y), x, y ∈ Σ, t > 0, (4.4.30)
=0

where each ϕ2 t (x, y) is an adjusted bump function in the variable x associated with
the ball B(y, 2 t), which means that there exists some constant C ∈ (0, ∞) with the
property that for every  ∈ N0 and t > 0 we have:
(1) supp ϕ2 t (·, y) ⊆ B(y, 2 t) for each y ∈ Σ;
|ϕ2 t (x, y)| ≤ C(2 t)
(2) 1−n for each x, y ∈ Σ;

(3) ϕ2 t (·, y) .η ≤ C(2 t)1−n−η for each η ∈ (0, ε1 ] and each y ∈ Σ;
∫ C (Σ)
(4) Σ ϕ2 t (x, y) dσ(x) = 0 for each y ∈ Σ.
Together, (4.4.29) and (4.4.30) permit us to write, for each f ∈ L po (Σ, σ) with
po ∈ (1, ∞),
∫ ∞ ∫ ∞ ∫
dt dt
f (x) = Dt (Dt f )(x) = Dt (x, y)(Dt f )(y) dσ(y)
0 t 0 Σ t
∞ ∫
dt
= 2−N ϕ2 t (x, y)(Dt f )(y) dσ(y) , ∀x ∈ Σ. (4.4.31)
=0 Σ×(0,∞) t

In turn, formula (4.4.31) is the staring point in the quest of decomposing a given
distribution f belonging to the Hardy space H p (Σ, σ) into atoms. To describe this in
more details, we shall freely borrow notation and results from [88, pp. 1524-1525]
and [185, pp. 347-350].
4.4 Atomic Decompositions for Hardy Spaces 149

As a preamble, we briefly review the characterization of Hardy spaces in terms

of the Lusin area-function. Concretely, having fixed some background parameter
κ > 0, define the Lusin area-function A κ F of a given (σ ⊗ L 1 )-measurable function
F : Σ × (0, ∞) → R as
∫
dσ(y) dt 12
(A κ F)(x) := |F(y, t)| 2 , ∀x ∈ Σ, (4.4.32)
Υκ (x) tn

where
 
Υκ (x) := (y, t) ∈ Σ × (0, ∞) : |x − y| < κ t , ∀x ∈ Σ. (4.4.33)

In relation to this, first observe that

A κ F : Σ −→ [0, ∞] is lower-semicontinuous. (4.4.34)

To justify this, fix xo ∈ Σ arbitrary and consider a sequence {x j } j ∈N ⊆ Σ convergent

to xo . Then,
 as is easily seen by analyzing each of the cases (y, t) ∈ Υκ (xo ) and
(y, t) ∈ Σ × (0, ∞) \ Υκ (xo ), we have

lim inf 1Υκ (x j ) (y, t) ≥ 1Υκ (xo ) (y, t), ∀(y, t) ∈ Σ × (0, ∞). (4.4.35)
j→∞

Granted this, Fatou’s lemma permits us to estimate

∫
dσ(y) dt 12
(A κ F)(xo ) = |F(y, t)| 2
Υκ (x o ) tn
∫
dσ(y) dt 12
= 1Υκ (xo ) (y, t)|F(y, t)| 2
Σ×(0,∞) tn
∫
dσ(y) dt 12
≤ lim inf 1Υκ (x j ) (y, t)|F(y, t)| 2
Σ×(0,∞) j→∞ tn
∫
dσ(y) dt 12
≤ lim inf ≤ 1Υκ (x j ) (y, t)|F(y, t)| 2
j→∞ Σ×(0,∞) tn
∫
dσ(y) dt 12
= lim inf |F(y, t)| 2 = lim inf (A κ F)(x j ), (4.4.36)
j→∞ Υκ (x j ) tn j→∞

finishing the proof of (4.4.34). Now, with the family {Dt }t>0 as before, here is the
promised characterization of the scale of Hardy spaces on Σ (cf., e.g., [88]3):

3 Under the current assumptions, the Hardy spaces used in [88] coincide with those from Defini-
tion 4.2.1; see [88, Remarks 2.27,2.30] and Theorem 4.4.1
150 4 Hardy Spaces on Ahlfors Regular Sets

n < s < ∞ then a distribution g on Σ belongs to H (Σ, σ) if and

if n−1 s

only if A κ G ∈ L (Σ, σ) where the function G : Σ × (0, ∞) → R is

r
(4.4.37)
defined as G(x, t) := (Dt g)(x) for each x ∈ Σ and t ∈ (0, ∞); moreover,
g H s (Σ,σ) ≈ A κ G L s (Σ,σ) .

Continuing the discussion about the Lusin area-function, the goal to show that there
exists some constant C ∈ (0, ∞), depending only on Σ, κ, and n, such that for any
two (σ ⊗ L 1 )-measurable functions H1, H2 : Σ × (0, ∞) → R we have
∫ ∫
dt
|H1 (y, t) H2 (y, t)| dσ(y) ≤ C (A κ H1 )(x)(A κ H2 )(x) dσ(x). (4.4.38)
Σ×(0,∞) t Σ

First note that, thanks to (4.4.34), the integral in the right-hand side of (4.4.38) is
meaningful. Next, the idea is to estimate the expression
∫ ∫
dt
I := |H1 (y, t) H2 (y, t)| dσ(y) n dσ(x) (4.4.39)
Σ Υκ (x) t

in two ways. On the one hand, using Fubini-Tonelli’s Theorem we may write
∫ ∫
dt
I= |H1 (y, t)||H2 (y, t)| 1Υκ (x) (y, t) dσ(x) dσ(y) n
Σ×(0,∞) Σ t
∫
  dt
= |H1 (y, t)||H2 (y, t)|σ B(y, κ t) ∩ Σ dσ(y) n
Σ×(0,∞) t
∫
dt
≈ |H1 (y, t)||H2 (y, t)| dσ(y) , (4.4.40)
Σ×(0,∞) t
 
since σ B(y, κ t) ∩ Σ ≈ t n−1 uniformly in y ∈ Σ and t > 0. On the other hand, based
on Cauchy-Schwarz’ inequality we may estimate
∫ ∫ ∫
dt 1/2 dt 1/2
I≤ |H1 (y, t)| 2 dσ(y) n |H2 (y, t)| 2 dσ(y) n dσ(x)
Σ Υκ (x) t Υκ (x) t
∫
= (A κ H1 )(x)(A κ H2 )(x) dσ(x). (4.4.41)
Σ

Now, (4.4.38) follows from (4.4.40) and (4.4.41).

We need to make yet another quick detour in order to address a relevant technical
issue. Specifically, let M Σ be the Hardy-Littlewood maximal operator associated
with the space of homogeneous type (Σ, | · − · |, σ) as in [133, Corollary 7.6.3], and
1 (Σ, σ). While we know from [133, (7.6.17)] that
fix some arbitrary function h ∈ Lloc
M Σ h : Σ → [0, +∞] is a σ-measurable function, as opposed to the Euclidean case,
M Σh is not necessarily semicontinuous.
 In particular, given some λ > 0, the level
set x ∈ Σ : (M Σ h)(x) > λ , while σ-measurable, is not necessarily open. To
rectify this, we propose to work with the following version of M Σ , defined for each
σ-measurable function h : Σ → C as
4.4 Atomic Decompositions for Hardy Spaces 151
∫
+Σ h(x) := sup 1
M n−1
|h| dσ, ∀x ∈ Σ. (4.4.42)
r ∈(0,∞) r B(x,r)∩Σ

In relation to this, we claim that

+Σ h : Σ −→ [0, +∞] is a lower-semicontinuous function
M
(4.4.43)
for every h ∈ Lloc
1 (Σ, σ).

To prove (4.4.43), fix h ∈ Lloc

1 (Σ, σ) along with some λ > 0, the goal being to show

that  
the set x ∈ Σ : (M +Σ h)(x) > λ is relatively open in Σ. (4.4.44)
With this finality in mind, pick a point xo belonging to the above set and note that,
in light of (4.4.42), this entails the existence of some r > 0 with the property that
∫
A := |h| dσ > λ r n−1 . (4.4.45)
B(x o ,r)∩Σ
 
Granted this, it becomes possible to choose a number s ∈ r, (A/λ)1/(n−1) . Such
a choice then forces both s > r and A > λ s n−1 . Consider now an arbitrary point
y ∈ B(xo, s − r) ∩ Σ. Then simple geometry gives B(xo, r) ∩ Σ ⊆ B(y, s) ∩ Σ, which
then permits us to estimate
∫ ∫
|h| dσ ≥ |h| dσ = A > λ s n−1 . (4.4.46)
B(y,s)∩Σ B(x o ,r)∩Σ

Having established this, we conclude from (4.4.42) +Σ h(y) > λ for every
that M
 
y ∈ B(xo, s − r) ∩ Σ. Thus, B(xo, s − r) ∩ Σ ⊆ x ∈ Σ : (M Σ h)(x) > λ which
ultimately proves (4.4.44). Thus, the claim made in (4.4.43) is indeed true.
In addition, thanks to the fact that we are currently assuming that Σ is Ahlfors
regular, it follows that for each σ-measurable function h : Σ → C the maximal
function M +Σ h from (4.4.42) is pointwise comparable with the standard Hardy-
Littlewood maximal function M Σ h defined as in (A.0.71). More specifically, if
0 < co ≤ Co < +∞ are the lower and upper Ahlfors regularity constants for Σ, then
for each σ-measurable function h : Σ → C we have
+Σ h ≤ Co · M Σ h everywhere on Σ.
co · M Σ h ≤ M (4.4.47)

As a consequence of this and [133, (7.6.18)-(7.6.19)], we have that

+Σ : L p (Σ, σ) −→ L p (Σ, σ) is well defined,

M
(4.4.48)
sub-linear and bounded for each p ∈ (1, ∞],

and, corresponding to the case p = 1, the modified maximal operator

152 4 Hardy Spaces on Ahlfors Regular Sets

+Σ : L 1 (Σ, σ) −→ L 1,∞ (Σ, σ) is well defined, sub-linear and bounded. (4.4.49)

M

Let us also note that, thanks to (4.4.47) and [133, Corollary 7.6.5] (whose applica-
bility is ensured by [133, (3.6.26)] in [133, Lemma 3.6.4]),

for each σ-measurable function h : Σ → C we have co · h ≤ M +Σ h at

σ-a.e. point on Σ, and this inequality actually holds everywhere if h is (4.4.50)
the characteristic function of a relatively open subset of Σ.

Returning to the main topic of discussion, assume next that f ∈ H p (Σ, σ) with
n−1
n < p ≤ 1 and consider

the function F : Σ×(0, ∞) −→ R defined as F(x, t) :=

(4.4.51)
(Dt f )(x) for each x ∈ Σ and t ∈ (0, ∞).
Then (4.4.37) implies that for each κ > 0 we have

A κ F ∈ L p (Σ, σ). (4.4.52)

To proceed, for each k ∈ Z define

 
Ek := x ∈ Σ : (A1 F)(x) > 2k , (4.4.53)

then, with co ∈ (0, ∞) denoting the lower Ahlfors regularity constant of Σ, introduce
 
Ok := x ∈ Σ : (M +Σ 1E )(x) > co /2n , (4.4.54)
k

where M +Σ stands for the modified Hardy-Littlewood maximal operator defined in

(4.4.42). Note that (4.4.34) ensures that for each k ∈ Z the set Ek is relatively open
(in Σ), while Chebytcheff’s inequality and (4.4.52) allow us to estimate

σ(Ek ) ≤ 2−k p A1 F  L p (Σ,σ) < +∞.

p
(4.4.55)

In particular, 1Ek ∈ L 1 (Σ, σ) hence, thanks to (4.4.44) and (4.4.49),

Ok is relatively open in Σ, and σ(Ok ) ≤ C · σ(Ek ) < +∞, (4.4.56)

for some C ∈ (0, ∞) independent of k. In particular, σ(Ok ) < +∞ implies, since we

are presently assuming that Σ is unbounded (hence σ(Σ) = +∞),

Ok is a proper subset of Σ, for each k ∈ Z. (4.4.57)

Parenthetically, we note that since (4.4.50) implies that for each k ∈ Z we have
+Σ 1E )(x) ≥ co for each point x ∈ Ek , we also have Ek ⊆ Ok for each k ∈ Z.
(M k
Going forward, let us now define the “stopping time” function φ : Σ × (0, ∞) → Z
by setting for each (y, t) ∈ Σ × (0, ∞)
4.4 Atomic Decompositions for Hardy Spaces 153

φ(y, t) to be the largest  k ∈ Z with

 integer  the property (4.4.58)
that σ B(y, t) ∩ Ek > σ B(y, t) ∩ Σ /2.

Since Ek decreases from Σ to a σ-nullset as k increases from −∞ to +∞, the function

φ is well defined. Moreover, φ enjoys the following properties:
  
(a) the sets φ−1 ({k}) k ∈Z are pair-wise disjoint and k ∈Z φ−1 ({k}) = Σ × (0, ∞);
(b) if (y, t) ∈ φ−1 ({k}) then B(y,
 t) ∩ Σ ⊆ Ok ;   
(c) if (y, t) ∈ φ−1 ({k}) then σ B(y, t) ∩ (Σ \ Ek+1 ) ≥ σ B(y, t) ∩ Σ /2.
Indeed, (a) is obvious. To prove (b), pick some (y, t) ∈ φ−1 ({k}) and for each
z ∈ B(y, t) ∩ Σ write
∫  
+Σ 1E )(z) ≥ 1 σ B(z, 2t) ∩ E k
(M k 1E dσ =
(2t)n−1 B(z,2t)∩Σ k (2t)n−1
 
co σ B(y, t) ∩ Ek co
≥ n−1 ·   > n. (4.4.59)
2 σ B(y, t) ∩ Σ 2

The first inequality in (4.4.59) is clear from (4.4.42) while the subsequent equality
is self-evident. The second inequality in (4.4.59) follows by combining the inclusion
B(y, t) ⊆ B(z, 2t) for each z ∈ B(y, t) with the fact that co is the lower Ahlfors
regularity constant of Σ. The last inequality in (4.4.59) is implied by (4.4.58). In
view of (4.4.54), the estimate in (4.4.59) ultimately proves that B(y, t) ∩ Σ ⊆ Ok .
Hence, the claim made in (b) is justified. Finally,
 as far as the
 estimate
 in (c) is
concerned, observe that this is equivalent to σ B(y, t) ∩ Ek+1 ) ≤ σ B(y, t) ∩ Σ /2
which is valid thanks to the manner in which the number k = φ(y, t) has been defined
in (4.4.58).
Moving on, fix a sufficiently large geometric constant λ ∈ (0, ∞). Given the
properties of Ok noted in (4.4.56)-(4.4.57) (and keeping in mind that Σ is Ahlfors
regular), for each fixed k ∈ Z for which Ok is not empty we may invoke [133,
Proposition 7.5.6] to guarantee the existence of some number Λ ∈ (0, ∞) independent
of k along with a sequence of dyadic cubes {Q j,k } j ∈N (relative to a dyadic grit
D(Σ) associated with the space of homogeneous type (Σ, | · − · |, σ) as in [133,
Proposition 7.5.4]) with the property that

the cubes {Q j,k } j ∈N are mutually disjoint, contained


in Ok , and satisfy σ Ok \ Q j,k = 0, (4.4.60)
j ∈N

and
   
B x j,k , λ (Q j,k ) ∩ Σ ⊆ Ok and dist Q j,k , Σ \ Ok ≤ Λ · (Q j,k ) for
each j ∈ N, where x j,k and (Q j,k ) denote, respectively, the center (4.4.61)
and side-length of the dyadic cube Q j,k .

For each k ∈ Z and j ∈ N we then define

154 4 Hardy Spaces on Ahlfors Regular Sets
 
Tj,k := (y, t) ∈ φ−1 ({k}) : y ∈ Q j,k ⊆ Σ × (0, ∞), (4.4.62)

and observe that

 
Tj,k j ∈N, k ∈Z is a family of pairwise disjoint sets, satisfying
, ,
Tj,k = φ−1 ({k}) for each k ∈ Z, and Tj,k = Σ × (0, ∞), (4.4.63)
j ∈N j ∈N, k ∈Z

up to a null-set for the product measure σ ⊗ L 1 . Indeed, that the sets Tj,k ’s are
mutually disjoint is clear from (4.4.62) upon recalling that the cubes {Q j,k } j ∈N are
mutually disjoint. The first equality in (4.4.63) is a consequence of (4.4.62), property
(b) of the stopping time function φ, and (4.4.60). Finally, the last equality in (4.4.63)
is implied by the first equality in (4.4.63) together with property (a) of the stopping
time function φ.
In the context of (4.4.31), the fact that
% &
,,
(σ ⊗ L 1 ) Σ × (0, ∞) \ Tj,k = 0 (4.4.64)
j ∈N k ∈Z

suggests that we attempt to express the given distribution f as the series


∞  ∫
dt
2−N ϕ2 t (·, y)(Dt f )(y) dσ(y) . (4.4.65)
=0 j ∈N k ∈Z Tj, k t

In turn, this bring into focus the family of building blocks

∫
dt
a j,k (x) := ϕ2 t (x, y)(Dt f )(y) dσ(y) , ∀x ∈ Σ, (4.4.66)
Tj, k t

indexed by  ∈ N0 , k ∈ Z, and j ∈ N. Note that each function a j,k satisfies

∫ ∫ ∫
dt
a j,k (x) dσ(x) = ϕ2 t (x, y) dσ(x) (Dt f )(y) dσ(y) = 0, (4.4.67)
Σ Tj, k Σ t

thanks to property (4) for the adjusted bump function ϕ2 t . To study the supports of
these building blocks, fix  ∈ N0 , k ∈ Z, and j ∈ N and pick some x ∈ supp a j,k .
Property (1) of ϕ2 t then forces that on the domain of integration in (4.4.66) we have

|x − y| ≤ 2 t. (4.4.68)

In addition, from (4.4.62) we see that

if (y, t) ∈ Tj,k then y ∈ Q j,k and (y, t) ∈ φ−1 ({k}). (4.4.69)

In view of property (b) of the stopping time function φ, the latter membership further
guarantees that B(y, t) ∩ Σ ⊆ Ok hence, on the one hand,
4.4 Atomic Decompositions for Hardy Spaces 155
 
dist y, Σ \ Ok ≥ t. (4.4.70)

On the other hand, based on triangle inequality and (4.4.61) we may bound, bearing
in mind that y ∈ Q j,k ,
   
dist y, Σ \ Ok ≤ dist Q j,k , Σ \ Ok + diam(Q j,k ) ≤ C · (Q j,k ), (4.4.71)

for some purely geometric constant C ∈ (0, ∞), independent of j, k. Collectively,

(4.4.70) and (4.4.71) then give

t ≤ C · (Q j,k ) (4.4.72)

which then allows us to estimate (with x, y, t as above; recall that x j,k stands for the
center of Q j,k and that  ∈ N0 )

|x − x j,k | ≤ |x − y| + |y − x j,k | ≤ 2 t + diam(Q j,k ) ≤ C · 2 · diam(Q j,k ).

(4.4.73)

All in all, this establishes that, for some purely geometric constant C > 0, independent
of  and j, k,
 
supp a j,k ⊆ Δ j,k := B x j,k , C · 2 · diam Q j,k ∩ Σ. (4.4.74)

Thus, when appropriately normalized, the function a j,k becomes a genuine atom
on Σ. Seeking a normalization in L r (Σ, σ), we proceed by duality, following the
approach in [185, p. 349]. Specifically, let r ∈ (1, ∞) be the Hölder conjugate
exponent of r and pick some g ∈ L r (Σ, σ). Define
∫
G (y, t) := ϕ2 t (x, y)g(x) dσ(x) for y ∈ Σ and t > 0, (4.4.75)
Σ

and pick some κ > 0. We claim that for each fixed parameter υ > 0 there exists a
finite constant C = Cυ > 0, independent of g, such that

A κ G  L r (Σ,σ) ≤ C · 2 υ g L r (Σ,σ) for every  ∈ N0 . (4.4.76)

The main ingredient in the proof of (4.4.76) is the L p -square function estimate from
[95]. To establish a dictionary between the current setting and that used in [95,
Theorem 1.1, pp. 6-7], we first observe that if for each  ∈ N0 we introduce
 
θ (y, t), x := t −υ ϕ2 t (x, y), x, y ∈ Σ, t > 0, (4.4.77)

and
∫
 
(Θ g)(y, t) := θ (y, t), x g(x) dσ(x), y ∈ Σ, t > 0, (4.4.78)
Σ
156 4 Hardy Spaces on Ahlfors Regular Sets

then for each  ∈ N0 and x ∈ Σ we have (on account of (4.4.77)-(4.4.78), (4.4.102),

and (4.4.32)-(4.4.33))
% ∫ & 12
2 dσ(y) dt
(A κ G )(x) = (Θ g)(y, t)   n−2υ
dist (y, t), Σ × {0}
(y,t)∈Σ×(0,∞)
|(x,0)−(y,t) |<(1+κ)dist((y,t),Σ×{0})
(4.4.79)

where the distance function is relative to the ambient metric space Rn × R, equipped
with the product metric d((a, r), (b, s)) := |a − b| + |r − s| for each a, b ∈ Rn and
r, s ∈ R, in which the set Σ × {0} is contained.
 Incidentally, in the context of (4.4.79)
is easy to check that dist (y, t), Σ × {0} is actually equal to t but we prefer the former,
more elaborate, piece of notation as this matches the set-up in [95, Theorem 1.1,
pp. 6-7]. Next, observe from (4.4.77) and property (4) of the adjusted bump function
ϕ2 t (x, y) that we have the cancellation condition
∫
 
(Θ 1)(y, t) := θ (y, t), x dσ(x)
Σ
∫
= t −υ ϕ2 t (x, y) dσ(x) = 0, ∀y ∈ Σ, ∀t > 0, (4.4.80)
Σ

and note that (4.4.77) and properties (1)-(2) of the adjusted bump function ϕ2 t (x, y)
readily imply the estimate
 
θ (y, t), x = t −υ ϕ2 t (x, y)

C·2 υ
≤ , ∀x, y ∈ Σ, ∀t > 0, (4.4.81)
|(x, 0) − (y, t)| n−1+υ

for some constant C ∈ (0, ∞) independent of . Finally, from (4.4.77) and property
(3) of the adjusted bump function ϕ2 t (x, y) we may easily conclude that once some
α ∈ (0, ε] has been fixed then there exists a constant C ∈ (0, ∞) independent of 
such that
    C · 2 υ |x − x | α
θ (y, t), x − θ (y, t), x ≤
|(x, 0) − (y, t)| n−1+α+υ (4.4.82)
for all x, x, y ∈ Σ and t > 0 satisfying |x − x | ≤ 12 |(x, 0) − (y, t)|.

Granted the identification in (4.4.79) and the cancellation, size, and regularity prop-
erties established in (4.4.80)-(4.4.82), we may invoke the equivalence (2) ⇔ (12)
from [95, Theorem 1.1, pp. 6-7] (with X := Σ × [0, ∞), E := Σ × {0}, d := n − 1,
μ := σ ⊗ L 1 , m := n, and r in place of p) in order to conclude that (4.4.76) holds.
Pushing forth in our quest of establishing suitable bounds for the L r -norms of
the building blocks introduced in (4.4.66), with the function F as in (4.4.51) and
the function G as in (4.4.75), for each , j, k fixed we may estimate (with constants
4.4 Atomic Decompositions for Hardy Spaces 157

independent of , j, k and the function g)

∫
a j,k (x)g(x) dσ(x) (4.4.83)
Σ
∫
  dt
= 1Tj, k · F (y, t)G (y, t) dσ(y)
Σ×(0,∞) t
∫
  dt
≤ 1Tj, k (y, t)|F(y, t)||G (y, t)| dσ(y)
Σ×(0,∞) t
∫  
  σ B(y, t) ∩ (Σ \ Ek+1 )
≤2 1Tj, k (y, t)   ×
Σ×(0,∞) σ B(y, t) ∩ Σ
dt
× |F(y, t)||G (y, t)| dσ(y)
t
∫ %∫ &
   
≤C 1Tj, k (y, t) 1B(y,t)∩Σ (x) dσ(x) ×
Σ×(0,∞) Σ\Ek+1

dt
× |F(y, t)||G (y, t)| dσ(y)
tn
∫ %∫
   
≤C 1Tj, k (y, t) 1B(y,t)∩Σ (x)|F(y, t)|×
Σ\Ek+1 Σ×(0,∞)
&
dt
× |G (y, t)| dσ(y) n dσ(x).
t

The first equality in (4.4.83) is based on Fubini’s Theorem, and the subsequent
inequality is clear. The second inequality relies on property (b) for the stopping time
function φ and the fact that we have (y, t) ∈ φ−1 ({k}) whenever  (y, t) ∈ Tj,k (cf.
(4.4.62)).
 The third equality is justified by observing that σ B(y, t) ∩ Σ ≈ t n−1 and
that σ B(y, t) ∩ (Σ \ Ek+1 ) may be recast as the inner integral appearing in the
penultimate line of (4.4.83). Finally, the last inequality above is Fubini’s Theorem.
Pressing on, observe that there exists a purely geometric constant C ∈ (0, ∞) (in
particular, independent of j, k, ) such that for each x, y ∈ Σ and t > 0 we have
       
1Tj, k (y, t) 1B(y,t)∩Σ (x) ≤ 1C ·Q j, k (x) 1B(x,t)∩Σ (y). (4.4.84)

To justify this inequality it suffices to prove that whenever the left-hand side is
nonzero then the right-hand side is nonzero as well. To this end, arguing as in
(4.4.68)-(4.4.73) with  = 0 we obtain that whenever x, y ∈ Σ and t ∈ (0, ∞) are
such that the left-hand side of (4.4.84) is nonzero then necessarily |x − y| < t and
|x − x j,k | ≤ C · diam(Q j,k ). These place x in C · Q j,k and y in B(x, t) ∩ Σ which,
in turn, make the right-hand side of (4.4.84) equal to one. Thus, (4.4.84) holds and,
158 4 Hardy Spaces on Ahlfors Regular Sets

when used back in (4.4.83), it permits us to further estimate

∫
a j,k (x)g(x) dσ(x) (4.4.85)
Σ
∫ %∫
   
≤C 1C ·Q j, k (x) 1B(x,t)∩Σ (y)×
Σ\Ek+1 Σ×(0,∞)
&
dt
× |F(y, t)||G (y, t)| dσ(y) n dσ(x)
t
∫ %∫ &
dt
≤C |F(y, t)||G (y, t)| dσ(y) n dσ(x)
C ·Q j, k ∩(Σ\Ek+1 ) Υ1 (x) t

∫ %∫ & 1/2
dt
≤C |F(y, t)| 2 dσ(y) n ×
C ·Q j, k ∩(Σ\Ek+1 ) Υ1 (x) t
%∫ & 1/2
dt
× |G (y, t)| dσ(y) n
2
dσ(x)
Υ1 (x) t
∫
  
=C A1 F (x) A1 G )(x) dσ(x)
C ·Q j, k ∩(Σ\Ek+1 )
∫
 
≤C max A1 F (x) A1 G dσ
x ∈Σ\Ek+1 C ·Q j, k
    1/r
≤ C 2k+1 A1 G  L r (Σ,σ)
σ C · Q j,k
  1/r
≤ C · 2k 2 υ σ 2− · Δ j,k g L r (Σ,σ)

υ   1/r
≤ C · 2k 2 2− (n−1)/r
σ Δ j,k g L r (Σ,σ) .

Above, the first inequality is obtained by combining  (4.4.83)

 with (4.4.84). The
second inequality is based on the observation that 1B(x,t)∩Σ (y)  0 if and only if
(y, t) ∈ Υ1 (x) (cf. (4.4.33)). The third inequality use the Cauchy-Schwarz inequality,
while the subsequent equality is seen from (4.4.32). The fourth inequality is self-
evident, while the next one uses (4.4.53) and Hölder’s inequality. Finally, the first
inequality in the last line of (4.4.85) is implied by (4.4.76) and the definition of Δ j,k
in (4.4.74), while the very last inequality is a consequence of the Ahlfors regularity
of Σ.
Having proved (4.4.85) for arbitrary functions g ∈ L r (Σ, σ), from Riesz’ Duality
Theorem we may now conclude that
4.4 Atomic Decompositions for Hardy Spaces 159
    1/r
a  ≤ C 2k 2 υ
2− (n−1)/r
σ Δ j,k . (4.4.86)
j,k L r (Σ,σ)

Consequently, for each  ∈ N0 , j ∈ N, and k ∈ Z, the function

a j,k
a j,k := is a (p, r)-atom on Σ (4.4.87)
λ j,k

if we take
υ   1/p
λ j,k := C · 2k 2 2− (n−1)/r
σ Δ j,k . (4.4.88)

At this stage, assuming that the parameter N has been selected so that

N > υ + (n − 1)(1 − p/r) (4.4.89)

to begin with, we may estimate

 - 1/p

∞ 
−N p p
2 λ j,k
=0 j ∈N k ∈Z
 - 1/p

∞ 
 
−N p k p υp − (n−1)p/r
=C 2 2 2 2 σ Δ j,k
=0 j ∈N k ∈Z
 - 1/p

∞ 
 
−N p k p υp (n−1)(1−p/r)
≤C 2 2 2 2 σ Q j,k
=0 j ∈N k ∈Z
 - 1/p  - 1/p
     
≤C 2 kp
σ Q j,k ≤C 2 kp
σ Ok
j ∈N k ∈Z k ∈Z
 - 1/p

≤C 2 kp
σ(Ek ) ≤ CA1 F  L p (Σ,σ)
k ∈Z

≤ C f  H p (Σ,σ), (4.4.90)

where, in the first inequality we have used the definition of Δ j,k from (4.4.74) and the
Ahlfors regularity of Σ, in the second inequality we have employed (4.4.89), in the
third inequality we have invoked (4.4.60), in the fourth inequality we have recalled
the estimate in (4.4.56), in the fifth inequality we have recalled a well-known formula
in real variable analysis estimating weighted sums of level sets of a given function
in terms of its Lebesgue quasi-norm, and in the last inequality we have appealed to
(4.4.37).  
In particular (4.4.90) shows that the sequence 2−N λ j,k belongs to the
j,k,
space  p . If we now re-index said sequence simply as {λ j } j ∈N , and also re-index the
160 4 Hardy Spaces on Ahlfors Regular Sets

family of (p, r)-atoms from (4.4.87) simply as {a j } j ∈N , then from [88, Lemma 2.23,
p. 1523], (4.4.64), (4.4.66), (4.4.87), and (4.4.90), we conclude that

∞ 
∞ 1/p
f = λ j a j in H p (Σ, σ) and |λ j | p ≤ C f  H p (Σ,σ) . (4.4.91)
j=1 j=1

In relation to (4.4.91), let us remark that, thanks to our earlier identifications and
(4.4.87),
!
the partial sums of the series ∞ λ j a j are the same as
! j=1! ! (4.4.92)
the partial sums of the series ∞=0 j ∈N k ∈Z 2−N a j,k

and that, as seen from (4.4.66),

all
!∞ the ! partial
! sums associated with the series
2−N a are independent of the ex-
=0 j ∈N k ∈Z j,k (4.4.93)
ponent p (labeling the Hardy space in which the dis-
tribution f originates).

Keeping (4.4.92)-(4.4.93) in mind, it follows that if f ∈ H p (Σ, σ) ∩ H q (Σ, σ) with

n < q ≤ 1, then running the same argument that has led to (4.4.91) we also
n−1

obtain that, with the same {λ j } j ∈N and {a j } j ∈N constructed as before (in relation to
f ∈ H p (Σ, σ)),
∞
f = λ j a j in H q (Σ, σ). (4.4.94)
j=1
 
The claim in (4.4.21) is therefore established in the case when q ∈ n−1 n ,1 .
Henceforth, consider the case 1 < q < ∞, i.e., assume f ∈ H p (Σ, σ) ∩ L q (Σ, σ)
(see [133, (3.6.27)] in this regard). The goal is to show that, with {λ j } j ∈N and {a j } j ∈N
constructed as before we also have
the
!∞ sequence of partial sums of the series
λ a converges to f in L q (Σ, σ). (4.4.95)
j=1 j j

In light of (4.4.92), this is further equivalent to proving that

the sequence
! !of partial
! sums associated with the
series ∞=0 j ∈N k ∈Z 2−N a j,k is convergent (4.4.96)
to the function f in L q (Σ, σ).

Justifying the claim in (4.4.96) requires some preparation. For starters, let { fm }m∈N
be the sequence of partial sums referred to in (4.4.96). Concretely, in view of (4.4.66)
for each m ∈ N we may express

m ∫
dt
fm (x) = 2−N ϕ2 t (x, y)(Dt f )(y) dσ(y) , ∀x ∈ Σ, (4.4.97)
=0 Om t
4.4 Atomic Decompositions for Hardy Spaces 161

where
,
m ,
m
Om := Tj,k . (4.4.98)
j=1 k=−m

Observe that {Om }m∈N is a nested increasing sequence of sets which, up to a (σ⊗L 1 )-
nullset, exhaust Σ × (0, ∞) (cf. (4.4.64)). In this notation, the claim in (4.4.96) may
be simply restated as
lim fm = f in L q (Σ, σ). (4.4.99)
m→∞

To this end, fix some κ > 0 and recall the function F from (4.4.51). Granted the
assumptions on f , Proposition 4.2.2 gives (bearing in mind that presently we are
assuming p ≤ 1 < q) that f also belongs to the Hardy space H 1 (Σ, σ). With this
membership in hand, from (4.4.92)-(4.4.93) and (4.4.91) (used now with p = 1)
we conclude that lim fm = f in H 1 (Σ, σ). This further implies lim fm = f in
m→∞ m→∞
L 1 (Σ, σ), hence also (by eventually restricting the index m to a sub-sequence of N)

lim fm (x) = f (x) for σ-a.e. x ∈ Σ. (4.4.100)

m→∞

Having noted (4.4.100), the claim in (4.4.99) follows as soon as we show that
{ fm }m∈N is a Cauchy sequence in L q (Σ, σ). With this goal in mind, for each m, k ∈ N
and each x ∈ Σ write

m ∫
dt
fm+k (x) − fm (x) = 2−N ϕ2 t (x, y)(Dt f )(y) dσ(y)
=0 Om+k \Om t


m+k ∫
−N dt
+ 2 ϕ2 t (x, y)(Dt f )(y) dσ(y) . (4.4.101)
=m+1 Om+k t

Also, fix an arbitrary g ∈ L q (Σ, σ) where q is such that 1/q + 1/q = 1, and for
each  ∈ N0 and m ∈ N define
∫
G (y, t) := ϕ2 t (x, y)g(x) dσ(x) for every y ∈ Σ and t > 0,
Σ (4.4.102)
and Fm := 1 Om · F (with the function F as in (4.4.51)).

In this notation, for each m, k ∈ N we may write

∫
( fm+k − fm )(x)g(x) dσ(x) (4.4.103)
Σ

m ∫
  dt
= 2−N G (y, t) Fm+k (y, t) − Fm (y, t) dσ(y)
=0 Σ×(0,∞) t


m+k ∫
−N dt
+ 2 G (y, t)Fm+k (y, t) dσ(y) .
=m+1 Σ×(0,∞) t
162 4 Hardy Spaces on Ahlfors Regular Sets

Returning to the mainstream discussion, starting with (4.4.103) and then employ-
ing (4.4.38) and Hölder’s inequality, for each m, k ∈ N we may estimate
∫
( fm+k − fm )(x)g(x) dσ(x) (4.4.104)
Σ

m
≤C 2−N A κ (Fm+k − Fm ) L q (Σ,σ) A κ G  L q (Σ,σ)
=0


m+k
+C 2−N A κ Fm+k  L q (Σ,σ) A κ G  L q (Σ,σ) .
=m+1

To proceed, we recall from (4.4.76) (presently used with p in place of r ) that for
each fixed parameter υ > 0 there exists a finite constant C = Cυ > 0, independent
of g, such that

A κ G  L p (Σ,σ) ≤ C · 2 υ g L p (Σ,σ) for every  ∈ N0 . (4.4.105)

In addition to this basic estimate, we shall also need that

0 ≤ A κ Fm ≤ A κ F on Σ for each m ∈ N, (4.4.106)

A κ (F − Fm ) → 0 in L q (Σ, σ) as m → ∞. (4.4.107)

Accepting the properties recorded in (4.4.106)-(4.4.107) for the time being, we may
readily conclude the proof of (4.4.99). Indeed, from (4.4.104) and (4.4.105) we see
that for each m, k ∈ N we have
4.4 Atomic Decompositions for Hardy Spaces 163

 fm+k − fm  L q (Σ,σ)
∫
= sup ( fm+k − fm )(x)g(x) dσ(x)
g ∈L q (Σ,σ), g  L q ≤1 Σ
(Σ, σ)


m
≤C 2−(N −υ) A κ (Fm+k − Fm ) L q (Σ,σ)
=0


m+k
+C 2−(N −υ) A κ Fm+k  L q (Σ,σ)
=m+1


m
≤C 2−(N −υ) A κ (Fm+k − F) L q (Σ,σ)
=0


m
+C 2−(N −υ) A κ (Fm − F) L q (Σ,σ)
=0


m+k
+C 2−(N −υ) A κ Fm+k  L q (Σ,σ) . (4.4.108)
=m+1

Assume that N > υ to begin with. Then from (4.4.108), (4.4.106)-(4.4.107), and
(4.4.52) we may conclude that  fm+k − fm  L q (Σ,σ) can be made as small as desired,
uniformly in k ∈ N, by taking m sufficiently large. Thus, the sequence { fm }m∈N is
indeed Cauchy in L q (Σ, σ). As remarked earlier, this finishes the proof of (4.4.99),
modulo the justification of (4.4.76)-(4.4.107).
Moving on, (4.4.106) is clear from (4.4.32) and (4.4.102). Finally, to prove
(4.4.107), make use of (4.4.52) to conclude that there exists a σ-measurable set
E ⊆ Σ satisfying

σ(E) = 0 and (A κ F)(x) < +∞ for each x ∈ Σ \ E. (4.4.109)

For each fixed x ∈ Σ \ E it follows that for every m ∈ N we have

∫
  dt 1/2
A κ (F − Fm ) (x) = 1(Σ×(0,∞))\Om (y, t)|F(y, t)| 2 dσ(y) n , (4.4.110)
Υκ (x) t

and the fact that (A κ F)(x) < +∞ implies that

  dt
1(Σ×(0,∞))\Om |F | Υκ (x)
≤ |F | Υκ (x) ∈ L 2 Υκ (x), σ ⊗ , ∀m ∈ N. (4.4.111)
tn
Since Om  Σ × (0, ∞) as m → ∞ (up to a (σ ⊗ L 1 )-nullset; cf. (4.4.98) and
(4.4.64)), it follows that 1(Σ×(0,∞))\Om |F | converges pointwise to zero as m → ∞. As
such, Lebesgue’s Dominated Convergence Theorem applies (in the ambient Υκ (x)
equipped with the measure σ ⊗ tdtn ) and, in light of (4.4.110), gives that
164 4 Hardy Spaces on Ahlfors Regular Sets
 
A κ (F − Fm ) (x) → 0 as m → ∞ for each x ∈ Σ \ E. (4.4.112)

With this in hand, one more application of Lebesgue’s Dominated Convergence

Theorem (bearing in mind (4.4.52), (4.4.106), and the fact that σ(E) = 0) proves
(4.4.107). This finishes the proof of the theorem in the case when Σ is unbounded.
Finally, we note that the main ingredients in used in the proof so far may be adapted
to the case when Σ is bounded. For example, a discrete Calderón-type reproducing
formula valid in arbitrary spaces of homogeneous type (bounded and unbounded)
may be found in [95, Proposition 2.14, p. 25]. As such, the same type of argument
applies with minor natural adjustments to yield similar conclusions in such a scenario.
This completes the proof of Theorem 4.4.3. 

As illustrated by the next proposition, the space of finite linear combinations of

atoms has excellent approximation qualities vis-a-vis to the entire Hardy space.

Proposition 4.4.4 Fix a closed set Σ ⊆ Rn which is Ahlfors regular and abbreviate
 n−1 p,q
σ := H Σ. Consider p ∈ n , 1 and q ∈ [1, ∞] with q > p, and let Hfin (Σ, σ)
n−1

stand for the vector space consisting of all finite linear combinations of (p, q)-atoms
p,q p,q
on Σ. Also, define a quasi-norm on Hfin (Σ, σ) by setting, for each f ∈ Hfin (Σ, σ),


N 1/p 
N
 f  H p, q (Σ,σ) := inf |λ j | p : N ∈ N and f = λ j a j for some
fin
j=1 j=1
-
numbers {λ j }1≤ j ≤ N ⊆ C and some (p, q)-atoms {a j }1≤ j ≤ N .

(4.4.113)

Then
 ∫ 
p,q q
Hfin (Σ, σ) = f ∈ Lcomp (Σ, σ) : Σ f dσ = 0 if Σ is unbounded,
p,q
Hfin (Σ, σ) = L q (Σ, σ) if Σ is bounded and, in all cases, (4.4.114)
p,q
Hfin (Σ, σ) is a dense linear subspace of H p (Σ, σ),

and
under the additional assumption that q < ∞, it follows that
 ·  H p, q (Σ,σ) and  ·  H p (Σ,σ) are equivalent quasi-norms on (4.4.115)
fin
p,q
the space Hfin (Σ, σ).

Proof The two equalities in (4.4.114) are clear from definitions, while the claim in
the last line is a consequence of Theorem 4.4.1. Finally, (4.4.115) is implied by [71,
Theorem 5.6, p. 2276]. 

The issue of the separability of the Lorentz-based Hardy spaces is discussed

below.
4.4 Atomic Decompositions for Hardy Spaces 165

Proposition 4.4.5 Let Σ ⊆ Rn be a closed  set which is Ahlfors regular and define
σ := H n−1 Σ. Then, for each p ∈ n−1 n , ∞ and q ∈ (0, ∞), the Lorentz-based Hardy
space H p,q (Σ, σ) is a separable quasi-Banach space.
 
Proof Consider first the case when we have p ∈ n−1 n , 1 . In such a scenario, pick
 n−1 
r ∈ (1, ∞) then select p0 ∈ n , p and p1 ∈ (1, r). In relation to the exponent r,
bring back the set F from the proof of Proposition 4.4.2 (cf. (4.4.15)) and observe
that, according to what has been established there,
 
F ⊆ H s (Σ, σ) for each s ∈ n−1 n ,r . (4.4.116)

In particular, F ⊆ H p0 (Σ, σ) ∩ H p1 (Σ, σ) hence also F ⊆ H p,q (Σ, σ) (cf. (4.3.145)).

We now claim that if the quasi-norm  ·  H p0 (Σ,σ)∩H p1 (Σ,σ) is defined as in (1.3.3),
then
for every (p0, r)-atom a on Σ, every λ ∈ R, and every ε > 0, there
exists some f ∈ F such that λ a − f  H p0 (Σ,σ)∩H p1 (Σ,σ) < ε. (4.4.117)

Recycling notation originally introduced in the proof of Proposition 4.4.2, this may
be justified along the lines of (4.4.17), by choosing f ∈ Fj0 such that in place of
(4.4.18) we now have
 
λ a − f  L r (Δ j0 ,σ) < ε · min σ(Δ j0 )1/r−1/p0 , σ(Δ j0 )1/r−1/p1 . (4.4.118)

Then, with f denoting the extension of f by zero to Σ, we have f ∈ Fj0 ⊆ F . Also,

the function g := λ a − f satisfies
∫
supp g ⊆ Δ j0 , g dσ = 0, g L r (Σ,σ) < ε · σ(Δ j0 )1/r−1/p0 . (4.4.119)
Δ j0

Since these properties imply that ε −1 g is a (p0, r)-atom on Σ, it follows that

ε −1 g H p0 (Σ,σ) ≤ C for some constant C ∈ (0, ∞) which depends only on Σ, p0 ,
and r. Thus, λ a − f  H p0 (Σ,σ) < Cε. Also, based on [133, (3.6.27)], (4.4.118), and
Hölder’s inequality we may write

λ a − f  H p1 (Σ,σ) ≈ λ a − f  L p1 (Σ,σ) = λ a − f  L p1 (Δ j0 ,σ)

≤ λ a − f  L r (Δ j0 ,σ) · σ(Δ j0 )1/p1 −1/r

< ε · σ(Δ j0 )1/r−1/p1 · σ(Δ j0 )1/p1 −1/r = ε. (4.4.120)

Hence, ultimately, λ a − f  H p0 (Σ,σ)∩H p1 (Σ,σ) < Cε which, after re-adjusting ε,

completes the proof of the claim made in (4.4.117).
In turn, from (4.4.117), (4.4.22), and the fact that F is stable under addition (cf.
(4.4.15)) we deduce that
166 4 Hardy Spaces on Ahlfors Regular Sets

F is dense in H p0 (Σ, σ) ∩ H p1 (Σ, σ),  ·  H p0 (Σ,σ)∩H p1 (Σ,σ) . (4.4.121)

Since, in turn, H p0 (Σ, σ) ∩ H p1 (Σ, σ) embeds continuously and densely into the
space H p,q (Σ, σ) (as noted in (4.3.145)), we ultimately conclude that

F is dense in H p,q (Σ, σ). (4.4.122)

Given that F is also known to be a countable set (cf. (4.4.15)),

 we conclude
 that the
quasi-Banach space H p,q (Σ, σ) is indeed separable if p ∈ n−1 n , 1 . Finally, when
p ∈ (1, ∞) the same conclusion follows from (4.2.30) and [133, Proposition 6.2.7].
This completes the proof of Proposition 4.4.5. 

Moving on, assume Σ ⊆ Rn is a closed Ahlfors regular set, and abbreviate

σ := H n−1 Σ. The goal now is to characterize the Hardy space H p (Σ, σ) for
n , 1 in terms of ions. Following [143, § A], these are defined as follows.
p ∈ n−1
Pick α ∈ (0, 1). We say that a function
 f ∈ L∞ (Σ, σ) is an ion, or an (α, p)-ion,
provided there exist x0 ∈ Σ and r ∈ 0, 2 diam Σ such that
∫
−(n−1)/p
supp f ⊆ B(x0, r) ∩ Σ,  f  L ∞ (Σ,σ) ≤ r , f dσ ≤ r α . (4.4.123)
Σ

Also, call a function h ∈ L ∞ (Σ, σ) a charge

 (or an α-charge) provided there exist a
point x0 ∈ Σ along with a radius r ∈ 0, 2 diam Σ such that

supp h ⊆ B(x0, r) ∩ Σ and h L ∞ (Σ,σ) ≤ r α−(n−1) . (4.4.124)

Of course, any (p, ∞)-atom f on Σ is an (α, p)-ion for any α ∈ (0, 1). It is also clear
from definitions that
if the set Σ is compact then any α-charge is a
(4.4.125)
fixed multiple of an (α, p)-ion.

In addition, any α-charge h satisfies, with q := (n − 1)/(n − 1 − α) > 1,

  1/q
h L q (Σ,σ) ≤ r α−(n−1) σ B(x0, r) ∩ Σ

≤ C 1/q r α−(n−1)+(n−1)/q = C 1/q, (4.4.126)

where C ∈ (0, ∞) is the upper Ahlfors regularity constant of Σ.

We have the following ionic characterization of Hardy spaces, in the spirit of
Theorem 4.4.1.

Proposition 4.4.6 Let Σ ⊆ Rn be a compact

 Ahlfors regular set, and abbreviate
σ := H n−1 Σ. Then for each p ∈ n−1
n , 1 one has
4.4 Atomic Decompositions for Hardy Spaces 167

H p (Σ, σ) (4.4.127)
    
= λ j f j in Lip (Σ) : each f j is an (α, p)-ion, and |λ j | p < ∞
j j

as sets, and
 1/p
 f  H p (Σ,σ) ≈ inf |λ j | p uniformly for f ∈ H p (Σ, σ), (4.4.128)
j

where the infimum in the right side is taken over all ionic decompositions of f .
p
Proof Temporarily denote the right side of (4.4.127) by Hion (Σ, σ). Since, as ob-
served earlier, any (p, ∞)-atom is an ion, from (4.4.127) and Theorem 4.4.1 we
p
conclude that H p (Σ, σ) ⊆ Hion (Σ, σ) and there exists C ∈ (0, ∞) such that for each
H (Σ, σ) we have
p
 1/p
inf |λ j | p ≤ C f  H p (Σ,σ), (4.4.129)
j

where the infimum in the left side is taken over all ionic decompositions of f .
To establish the reverse inclusion together with the opposite inequality in
(4.4.129), fix some α ∈ (0, 1). We claim that there exists a constant C ∈ (0, ∞)
such that

f (α, p)-ion =⇒ f ∈ H p (Σ, σ) and  f  H p (Σ,σ) ≤ C. (4.4.130)

To justify this claim, start with a function f as in (4.4.123) and introduce

∫
1
h := λ · 1B(x0,r)∩Σ where λ :=   f dσ. (4.4.131)
σ B(x0, r) ∩ Σ Σ

Note that |λ| ≤ r −(n−1)/p by the first two conditions in (4.4.123). As such, if we
define
g := f − h on Σ, (4.4.132)
then
∫
−(n−1)/p
supp g ⊆ B(x0, r) ∩ Σ, g L ∞ (Σ,σ) ≤ 2r , g dσ = 0. (4.4.133)
Σ

Hence, g/2 is a (p, ∞)-atom. In particular, g H p (Σ,σ) ≤ 2. As seen from its definition
in (4.4.131), the remainder h satisfies

supp h ⊆ B(x0, r) ∩ Σ and h L ∞ (Σ,σ) ≤ r −(n−1) . (4.4.134)

Since Σ is compact, it follows that h is a β-charge for any β ∈ (0, 1). Consequently,
if we define q := (n − 1)/(n − 1 − β) > 1 then (4.4.126) gives
168 4 Hardy Spaces on Ahlfors Regular Sets

h L q (Σ,σ) ≤ C, (4.4.135)

where C ∈ (0, ∞) depends only on β and the upper Ahlfors regularity constant of Σ.
By design, we have f = g + h, so the claim in (4.4.130) follows on account of the
fact that we presently have L q (Σ, σ) ⊆ H p (Σ, σ) continuously for each q > 1 (cf.
(4.2.13)).
p
In turn, (4.4.130) is the main ingredient in the proof of Hion (Σ, σ) ⊆ H p (Σ, σ).
p
Specifically, start with some f ∈ Hion (Σ, σ) and consider an arbitrary decomposition
!   !
f = j λ j f j in Lip (Σ) , in which each f j is an (α, p)-ion and j |λ j | p < ∞. For
any two integers M, N ∈ B with M > N we may then estimate
M p 
M
  p
 λ j fj  p ≤ λ j f j  H p (Σ,σ)
H (Σ,σ)
j=N j=N


M 
M
p
= |λ j | p  f j  H p (Σ,σ) ≤ C |λ j | p, (4.4.136)
j=N j=N

thanks to (4.2.7) and (4.4.130). This proves that the sequence of partial sums in
!
the series j λ j f j is Cauchy in H p (Σ, σ). Since the latter space is quasi-Banach
 
and, according to (4.2.8), embeds continuously into Lip (Σ) (whose topology is
Hausdorff; cf. [133, (4.1.39)]), we ultimately conclude that f ∈ H p (Σ, σ) and
 1/p
 f  H p (Σ,σ) ≤ C |λ j | p . (4.4.137)
j

p
This goes to show that Hion (Σ, σ) ⊆ H p (Σ, σ), finishing the proof of (4.4.127).
Taking the infimum in (4.4.137) over all ionic decompositions of f also yields
 1/p
 f  H p (Σ,σ) ≤ C inf |λ j | p . (4.4.138)
j

Together with (4.4.129), this concludes the proof of (4.4.128). 

Our next theorem contains a useful boundedness criterion for linear operators
defined on Hardy spaces and taking values in certain categories of topological
vector spaces.
Theorem 4.4.7 Suppose Σ ⊆ Rn is a closed  set which is Ahlfors
 regular and define
σ := H n−1 Σ. Pick three exponents p ∈ n−1 n ,1 ,q ∈ n , ∞ , and r ∈ (1, ∞). Next,
n−1

fix a topological vector space (X, τ) along with a pseudo-quasi-Banach space (Y, ·)
such that (X, τ) and (Y, τ ·  ) are weakly compatible (in the sense of Definition 1.5.9).
Denote by θ ∈ R the parameter quantifying the homogeneity of  ·  (cf. (1.5.20))
and assume % &
 f + g
θ ≥ p · log2 sup . (4.4.139)
f , g ∈Y max{ f , g}
not both zero
4.4 Atomic Decompositions for Hardy Spaces 169

Finally, consider a continuous linear operator

T : H q (Σ, σ) −→ (X, τ) (4.4.140)

with the property that

there exists some C ∈ (0, ∞) such that T a ∈ Y and

(4.4.141)
T a ≤ C, for every (p, r)-atom a on Σ.
Then there exists a unique linear and bounded (hence continuous) operator

T : H p (Σ, σ) −→ (Y,  · ), (4.4.142)

whose operator norm is bounded by a multiple of the constant C appearing in

(4.4.141), and which extends T, in the sense that if (X , τX ) is the ambient topological
vector space in which (X, τ) and (Y, τ ·  ) embed as in Definition 1.5.9 then

T f = T f in X , for each f ∈ H q (Σ, σ) ∩ H p (Σ, σ). (4.4.143)

In fact,

the operator T : H p (Σ, σ) −→ Y acts on any given f ∈ H p (Σ, σ)

according to T f := lim j→∞ T f j in (Y, τ ·  ) whenever the sequence (4.4.144)
p,r
{ f j } j ∈N ⊆ Hfin (Σ, σ) is such that f = lim j→∞ f j in H p (Σ, σ).

Before presenting the proof of Theorem 4.4.7 a couple of comments are in order.
First, in [18] one may find an example of a linear functional which is uniformly
bounded on all (1, ∞)-atoms (in the Euclidean setting), yet cannot be extended to a
bounded linear functional defined on all of H 1 .
Second, if  ·  is actually an s-norm (in the sense of item [3] of Definition 1.5.7)
for some s ∈ (0, ∞) then
' (s
 f + g s ≤  f  s + g s ≤ 21/s max{ f , g} , ∀ f , g ∈ Y . (4.4.145)

This implies that the supremum in (4.4.139) is ≤ 21/s , which means that

condition (4.4.139) is automatically satisfied if  ·  is

(4.4.146)
actually an s-norm for some s ∈ (0, ∞) and θ ≥ p/s.

Proof of Theorem 4.4.7 The property that T map (p, r)-atoms on Σ into the space Y
p,r
implies that T f ∈ Y for each f ∈ Hfin (Σ, σ). In relation to this, we will first establish
that there exists some C ∈ (0, ∞) such that
θ p,r
T f  ≤ C f  H p (Σ,σ) for each function f ∈ Hfin (Σ, σ). (4.4.147)
!
To this end, given f as above, consider an arbitrary way of expressing f = N j=1 λ j a j
on Σ where N ∈ N, {λ j } j=1 ⊆ C, and {a j } j=1 is a sequence of (p, r)-atoms on Σ.
N N
170 4 Hardy Spaces on Ahlfors Regular Sets

We claim that there exists a finite constant C > 0 (independent of f and its atomic
decomposition) with the property that
#
N $ θ/p
T f  ≤ C |λ j | p . (4.4.148)
j=1

In order to justify (4.4.148) we proceed by considering two cases. Suppose first that
the supremum displayed in (4.4.139) is one, i.e., suppose  f + g ≤ max{ f , g}
for every f , g ∈ Y . In this scenario write
 
 N   

T f  =  λ j T a j  ≤ max λ j T a j  ≤ C max |λ j | θ T a j 
1≤ j ≤ N 1≤ j ≤ N
j=1
θ
≤ C max |λ j | θ = C max |λ j |
1≤ j ≤ N 1≤ j ≤ N

#
N $θ #
N $ θ/p
≤C |λ j | ≤C |λ j | p , (4.4.149)
j=1 j=1

where the second inequality is a consequence of the pseudo-homogeneity of  · , the

third inequality follows from the uniform bound in (4.4.141), and the last inequality
makes use of the fact p ≤ 1.
Next, assume that the supremum displayed in (4.4.139) is strictly greater than one
and let
) % &* −1
 f + g
β := log2 sup ∈ (0, ∞). (4.4.150)
f , g ∈Y max{ f , g}
not both zero

Then for some C ∈ (0, ∞) which depends only on β and the proportionality constants
in (1.5.27) we may estimate
 β  β 
 N   N  N
T f  =  λ j T a j  ≤ C  λ j T a j  ≤ C
β
|λ j | θβ · T a j 
j=1 j=1  j=1


N 
N
≤C |λ j | θβ · T a j  β ≤ C |λ j | θβ, (4.4.151)
j=1 j=1

where the first and third inequalities follow from (1.5.27) in Theorem 1.5.8, the
second inequality comes from (1.5.28)-(1.5.29) in Theorem 1.5.8, and the last in-
equality is a consequence of (4.4.141). Note that the usage of (1.5.29) is valid given
the definition of β ∈ (0, ∞). Combining the estimate in (4.4.151) with the fact that
θ β ≥ p (as a result of (4.4.139) and the definition of β) ultimately permits us to
write
4.4 Atomic Decompositions for Hardy Spaces 171
  # $ 1/β # $ θ/p
 N  N N
T f  =  λ j T a j  ≤ C |λ j | θβ ≤C |λ j | p , (4.4.152)
j=1 j=1 j=1

finishing the proof of (4.4.148). Taking the infimum in (4.4.148) over all finite atomic
decompositions of f then yields T f  ≤ C f  H θ , from which (4.4.147)
p, r
(Σ,σ)
fin
follows on account of (4.4.115).
Given (1.5.32), the pseudo-homogeneity of  · , and the homogeneity of  ·
 H p (Σ,σ) , the estimate in (4.4.147) implies that
 p,r 
T p, r
Hfin (Σ,σ)
: Hfin (Σ, σ),  ·  H p (Σ,σ) −→ (Y,  · )
(4.4.153)
is a well-defined, linear, and bounded mapping.
p,r
In particular, T maps Cauchy sequences in the space Hfin (Σ, σ) (equipped with
the quasi-norm  ·  H p (Σ,σ) ) into Cauchy sequences in (Y, τ ·  ) (in the sense of
Definition 1.5.1). Based on this, the density result in (4.4.114), and the completeness
of (Y, τ ·  ) (again, in the sense of Definition 1.5.1), it follows that (4.4.153) may
be further extended to an operator T : H p (Σ, σ) → Y as in (4.4.144). That T is
unambiguously defined may be seen by interlacing sequences. In turn, this further
p,r
implies that T is linear and that T agrees with T on Hfin (Σ, σ). In addition, from
the definition of T and the property displayed in (1.5.33) it follows that there exists
C ∈ (0, ∞) such that
θ
T f  ≤ C f  H p (Σ,σ) for each f ∈ H (Σ, σ).
p
(4.4.154)

In concert with (1.5.32) this shows that T is indeed bounded in the context of
(4.4.142).
There remains to justify (4.4.143). Fix a function f ∈ H q (Σ, σ) ∩ H p (Σ, σ).
p,r
Theorem 4.4.3 implies that there exists a sequence { f j } j ∈N ⊆ Hfin (Σ, σ) such that
lim f j = f both in H (Σ, σ) and in H (Σ, σ). Relying on the convergence in
q p
j→∞
H p (Σ, σ), from the continuity of T in (4.4.142) and the fact that T agrees with T on
p,r
Hfin (Σ, σ) we may conclude that

T f = lim T f j = lim T f j in (Y, τ ·  ). (4.4.155)

j→∞ j→∞

On the other hand, from the H q -convergence and the boundedness of T in (4.4.140)
we have
T f = lim T f j in (X, τ). (4.4.156)
j→∞

Recall that (X, τ) and (Y, τ ·  ) are weakly compatible. Then, if (X , τX ) is as in

Definition 1.5.9, from (4.4.155), (4.4.156), and the fact that convergent sequences in
(X , τX ) have unique limits, we conclude that

T f = lim T f j = lim T f j = T f in (X , τX ). (4.4.157)

j→∞ j→∞
172 4 Hardy Spaces on Ahlfors Regular Sets

This justifies (4.4.143), hence the proof of Theorem 4.4.7 is complete. 

It is of interest to show that Hardy spaces are local whenever the underlying set
is compact, in the following precise sense.

Proposition 4.4.8 Fix n ∈ N with n ≥ 2 and suppose Σ ⊆ Rn be a compact Ahlfors

regular set. Abbreviate σ := H n−1 Σ and pick some p ∈ n−1 n , 1 . Also, suppose
(n − 1)(p−1 − 1) < γ < 1. Then the Hardy space H p (Σ, σ) is a module over the
Hölder space C γ (Σ), in the sense that the operator Mϕ of multiplication by any
given function ϕ ∈ C γ (Σ) induces a well defined linear and bounded mapping

Mϕ : H p (Σ, σ) −→ H p (Σ, σ). (4.4.158)

Moreover, there exists a constant C = C(Σ, n, p, γ) ∈ (0, ∞) with the property that

Mϕ  H p (Σ,σ)→H p (Σ,σ) ≤ CϕC γ (Σ) for each ϕ ∈ C γ (Σ). (4.4.159)

 
Proof Since γ − (n − 1) p1 − 1 ∈ (0, 1), there exists q ∈ (1, ∞) large enough so that
1 
α := γ − (n − 1) p −1− 1
q ∈ (0, 1). (4.4.160)

This choice of α also ensures that

α
q∗ := q satisfies 1 < q∗ < q, (4.4.161)
n−1
which is going to be relevant shortly.
For now, pick a function ϕ ∈ C γ (Σ), then select A, B ∈ (0, ∞) such that

sup |ϕ| ≤ A and |ϕ(x) − ϕ(y)| ≤ B|x − y|γ for all x, y ∈ Σ. (4.4.162)
Σ

With q ∈ (1, ∞) as above, let f be a (p, q)-atom on Σ, Hence, there exist a point
x0 ∈ Σ and a number r ∈ 0, diam Σ with the property that
∫
  1/q−1/p
supp f ⊆ B(x0, r) ∩ Σ,  f  L q (Σ,σ) ≤ σ B(x0, r) ∩ Σ , f dσ = 0.
Σ
(4.4.163)
If we set λ := ϕ(x0 ) then

|λ| ≤ A and (ϕ − λ) f  L q (Σ,σ) ≤ CB r α−(n−1), (4.4.164)

where C ∈ (0, ∞) depends only on Σ, p, q. Based on the last property above, the
support condition in (4.4.163), (4.4.161), Hölder’s inequality, and the fact that Σ is
upper Ahlfors regular, we may estimate
  1−(q∗ /q)
(ϕ − λ) f  L q∗ (Σ,σ) ≤ (ϕ − λ) f  L q∗ (Σ,σ) · σ B(x0, r) ∩ Σ

≤ C r α−(n−1)+(n−1)(1−(q∗ /q)) = C, (4.4.165)

4.4 Atomic Decompositions for Hardy Spaces 173

where C ∈ (0, ∞) depends only on the upper Ahlfors regularity constant of Σ as well
as B, p, q, and γ. Consequently, the decomposition

ϕ f = λ f + g, where g := (ϕ − λ) f , (4.4.166)

expresses f as a sum of a multiple of an (p, q)-atom, for some coefficient bounded

independently of f , x0, r, and a function whose norm in L qα (Σ, σ) is also bounded
independently of f , x0, r. In view of this, (4.4.160), (4.4.126), and the fact that
L q∗ (Σ, σ) embeds continuously into H p (Σ, σ) (cf. (4.2.13)), we then conclude that
ϕ f belongs to the space H p (Σ, σ) and ϕ f  H p (Σ,σ) ≤ C, for some finite constant
C > 0 depending only on A, B, Σ, n, p, and γ.
Granted this, the fact that Mϕ induces a well-defined, linear, and bounded operator
in the context of (4.4.158) is then a consequence of Theorem 4.4.7, bearing in mind
that Mϕ is a linear and bounded operator on any L s (Σ, σ) with s > 1. 

We conclude this section with a remark concerning Hardy spaces of vector distri-
butions. To set the stage, suppose Σ ⊆ Rn is a closed set which is Ahlfors regular and
define σ := H n−1 Σ. Also, pick some p ∈ n−1 n , 1 along with q ∈ [1, ∞] satisfying
q > p, and fix some integer M ∈ N. The goal is to discuss atomic decompositions
 M
for distributions belonging to the vector Hardy space H p (Σ, σ) . In this regard,
we find it convenient to work with C M -valued (p, q)-atoms on Σ, i.e., functions
 M
a ∈ L q (Σ, σ) such that for some x ∈ Σ and r ∈ (0, 2 diam Σ) one has
  1/q−1/p
supp a ⊆ Σ ∩ B(x, r), a[L q (Σ,σ)] M ≤ σ Σ ∩ B(x, r) ,
∫
as well as a dσ = 0 ∈ C M .
Σ
(4.4.167)
Furthermore,
in the case when Σ is bounded we also agree that constant vectors
(4.4.168)
a ∈ C M with |a| ≤ [σ(Σ)]−1/p are C M -valued (p, q)-atoms on Σ.
 M
Suppose now that some f = ( fβ )1≤β ≤M ∈ H p (Σ, σ) has been given. Then
!
according to Theorem 4.4.1 each fβ has an atomic decomposition fβ = ∞ j=1 λβ j aβ j
(no summation on β here) where each aβ j is a (p, q)-atom on Σ (cf. (4.4.2)) and
{λβ j } j ∈N ∈  p . Moreover, it may be assumed that said atomic decomposition is
quasi-optimal, in the sense that (for some absolute constants)


∞
 1/p
 fβ  H p (Σ,σ) ≈ |λβ j | p for each β ∈ {1, . . . , M }. (4.4.169)
j=1

For each index β ∈ {1, . . . , M } introduce eβ := (δγβ )1≤γ ≤M ∈ C M (using the

Kronecker symbol formalism) then write
174 4 Hardy Spaces on Ahlfors Regular Sets


M 
M 
∞ 
M 
∞
f = fβ eβ = λβ j aβ j eβ = λβ j Aβ j (4.4.170)
β=1 β=1 j=1 β=1 j=1

 M
with convergence in H p (Σ, σ) , where

Aβ j := aβ j eβ for each β ∈ {1, . . . , M } and j ∈ N. (4.4.171)

These are C M -valued functions as in (4.4.167)-(4.4.168),

   C  -valued (p, q)-
hence M

atoms on Σ. If we then relabel the sequences Aβ j 1≤β ≤M and λβ j 1≤β ≤M simply

    j ∈N j ∈N
as a j j ∈N and λ j j ∈N , respectively, we may re-cast (4.4.169)-(4.4.171) in the form
of the following conclusion:
 M
each distribution f ∈ H p (Σ, σ) may be decomposed as
!∞  M
f = λ j a j with convergence in H p (Σ, σ) , where each
j=1
(4.4.172)
a j is a C M -valued (p, q)-atom on Σ (cf. (4.4.167)-(4.4.168)),
!
∞  p1
and  f [H p (Σ,σ)] M ≈ |λ j | p (with absolute constants).
j=1

Moreover, as is apparent from the above discussion,

the analogues of Theorem 4.4.3 and Proposition 4.4.4

(4.4.173)
remain valid for vector Hardy space.

4.5 Molecules in Hardy Spaces

A molecule may be regarded as weakened version of an atom, in which the compact

support condition has been replaced by some power-type decay condition over dyadic
annuli (while the vanishing moment condition is retained). This is made precise in
the definition below (see [9, Definition 6.1, p. 266] for a more general setting).

Definition 4.5.1 Suppose Σ ⊆ Rn is a closed

 set
 which is Ahlfors regular and define
σ := H n−1 Σ. Pick two exponents p ∈ n−1 n , 1 and q ∈ [1, ∞] such that q > p, and
fix some ε ∈ (0, ∞). Having chosen a point xo ∈ Σ along with some finite number
r ∈ (0, diam Σ], call a σ-measurable function m defined on Σ a (p, q, ε)-molecule
centered near the ball B(xo, r) on Σ (or, simply, a H p -molecule on Σ) provided for
each k ∈ N0 one has
∫ 1/q   1/q−1/p
|m| q dσ ≤ 2k(n−1)[1/q−1−ε] σ B(xo, r) ∩ Σ (4.5.1)
A k (x o ,r)

where
4.5 Molecules in Hardy Spaces 175

⎧
⎨ B(xo, r) ∩ Σ
⎪
⎪ if k = 0,
Ak (xo, r) :=   (4.5.2)
⎪
⎪
⎩ B(xo, 2 r) \ B(xo, 2 r) ∩ Σ if k ≥ 1,
k k−1

and
∫
m dσ = 0. (4.5.3)
Σ

In addition, in the case when Σ is bounded it is also agreed that the constant
function given by m(x) := [σ(Σ)]−1/p for every x ∈ Σ is a (p, q, ε)-molecule on Σ.

In the context of Definition 4.5.1, from (4.5.1)-(4.5.2) we conclude that

if m is a (p, q, ε)-molecule on Σ then

(4.5.4)
m ∈ L qo (Σ, σ) for each qo ∈ 1+ε 1
,q .

In particular, the case qo := 1 ensures that the cancellation condition in (4.5.3) is

meaningfully formulated. Of course,

every (p, q)-atom is a (p, q, ε)-molecule for each ε > 0. (4.5.5)

It turns out (see, e.g., [9, Theorem 6.4, pp. 273-274]) that
any H p -molecule m on Σ (cf. Definition 4.5.1) belongs to the
Hardy space H p (Σ, σ) and satisfies m H p (Σ,σ) ≤ C for some (4.5.6)
constant C ∈ (0, ∞) independent of the molecule m.

In fact, H p -molecules on Σ are building blocks for the Hardy space H p (Σ, σ),
much as in Theorem 4.4.1. See [9, Chapter 6] for more information (including
pertinent references) in this regard. In the lemma below we provide a more intrinsic
characterization of molecules.

Lemma 4.5.2 Let Σ ⊆ Rn be a closed set which is Ahlfors regular. Abbreviate

σ := H n−1 Σ and fix some point xo ∈ Σ.
Then for each exponents p, q, θ, d such that
 
p ∈ n−1 n ,1 , q ∈ [1, ∞) with q > p,
 (4.5.7)
θ ∈ (0, 1), and d := θ −1 (n − 1) q/p − 1),

there exists a constant C = C(Σ, n, p, q, θ) ∈ (0, ∞) with the property that every
function
   
m ∈ L q (Σ, σ) ∩ L q Σ, | · −xo | d σ = L q Σ, (1 + | · −xo | d )σ (4.5.8)

satisfying ∫
m dσ = 0 (4.5.9)
Σ
176 4 Hardy Spaces on Ahlfors Regular Sets

necessarily is a multiple of a H p -molecule on Σ (hence, in particular m belongs to

the Hardy space H p (Σ, σ)) and
θ
m H p (Σ,σ) ≤ Cm L1−θ
q (Σ,σ) · m q
L (Σ, | ·−x d σ)
o|

≤ Cm L q (Σ,(1+ | ·−xo | d )σ) . (4.5.10)

Conversely, for each exponents p, q, ε, θ, d such that

 
p ∈ n−1 n ,1 , q ∈ [1, ∞) with q > p, ε > 1/p − 1,
1/p−1/q −1
 (4.5.11)
1+ε−1/q < θ < 1, and d := θ (n − 1) q/p − 1),

there exists a constant C = C(Σ, n, p, q, ε, θ) ∈ (0, ∞) with the property that every
(p, q, ε)-molecule m centered near the ball B(xo, r) on Σ, for some finite number
r ∈ (0, diam Σ], satisfies (4.5.8) and
θ
m L1−θ
q (Σ,σ) · m q
L (Σ, | ·−x d σ) ≤ C. (4.5.12)
o|

Proof As a preamble, we note that, as may be seen using Hölder’s inequality and
[133, Lemma 7.2.1],

if qo ∈ (0, q] and a > (n − 1) q/qo − 1) then
  (4.5.13)
L q Σ, (1 + | · −xo | a )σ → L qo (Σ, σ).
 
In turn, since d > (n − 1) q/qo − 1 whenever qo ≥ p, this implies
  "
L q Σ, (1 + | · −xo | d )σ ⊆ L qo (Σ, σ) ⊆ L 1 (Σ, σ). (4.5.14)
p ≤qo ≤q

Hence, any function m as in (4.5.8) is absolutely integrable on Σ, which goes to show

that the property formulated in (4.5.9) is indeed meaningful.
Fix now a function m as in (4.5.8)-(4.5.9) and, without loss of generality, assume
m does not vanish σ-a.e. on Σ. To show that m is a multiple of a H p -molecule on Σ,
in the precise quantitative fashion described in (4.5.10), we distinguish two cases.
 
Case I: Assume p ∈ n−1 n , 1 . In this scenario, choose ε > 0 such that 1/p = 1 + ε.
For some r > 0 to be specified later, recall (4.5.2) and, for each k ∈ N, estimate
4.5 Molecules in Hardy Spaces 177
∫ 1/q
|m| q dσ (4.5.15)
A k (x o ,r)
∫ 1/q
= |m| q | · −xo | d | · −xo | −d dσ
A k (x o ,r)

≤ C(2k r)−d/q m L q (Σ, | ·−xo | d σ)

−1 k(n−1)[1/q−1−ε]   θ −1 (1/q−1/p)
≤ C2θ σ B(xo, r) ∩ Σ m L q (Σ, | ·−xo | d σ),

based on the definition of d, ε, and the Ahlfors regularity of Σ. Consequently,

∫ 1/q
∫ θ/q
∫ (1−θ)/q
|m| q dσ = |m| q dσ |m| q dσ
A k (x o ,r) A k (x o ,r) A k (x o ,r)
  1/q−1/p
≤ C2k(n−1)[1/q−1−ε] σ B(xo, r) ∩ Σ ×
θ
q (Σ,σ) m q
× m L1−θ L (Σ, | ·−x d σ) (4.5.16)
o|

for each k ∈ N. To also be able to estimate the left-most side of (4.5.16) when k = 0,
choose
# $
m L q (Σ, | ·−xo | d σ) q/d
r := ∈ (0, ∞). (4.5.17)
m L q (Σ,σ)

Together with the Ahlfors regularity of Σ, this choice of r permits us to write

∫ 1/q
|m| q dσ ≤ m L q (Σ,σ) (4.5.18)
A0 (x o ,r)

= r (n−1)(1/q−1/p) m L1−θ θ

q (Σ,σ) m q
L (Σ, | ·−x d σ)
o|

  1/q−1/p θ
≤ Cσ B(xo, r) ∩ Σ m L1−θ
q (Σ,σ) m q
L (Σ, | ·−x d σ) .
o|

Assume m is does not vanish σ-a.e. on Σ. Then from (4.5.15), (4.5.18), and (4.5.9),
we deduce that the function
m
m := θ
(4.5.19)
q (Σ,σ) m q
Cm L1−θ L (Σ, | ·−x o | d σ)

is a (p, q, ε)-molecule centered near the ball B(xo, r) on Σ, in the sense of Defini-
tion 4.5.1. Hence, m is a multiple of a H p -molecule on Σ. In light of (4.5.6), we
conclude that m belongs to the Hardy space H p (Σ, σ) and the first inequality in
(4.5.10) follows with the help of (4.5.19). Since the second inequality in (4.5.10) is
clear, and since the case when m = 0 at σ-a.e. point in Σ is obvious, the treatment
of Case I is complete.
178 4 Hardy Spaces on Ahlfors Regular Sets

Case II: Assume p = 1. Recall that this entails q ∈ (1, ∞) and d = (n−1)θ −1 (q−1).
In particular, θ −1 (1 − 1/q) + 1/q > 1 which means that it is possible to choose
  −1 " n−1
po ∈ θ −1 (1 − 1/q) + 1/q ,1 ,1 . (4.5.20)
n
In turn, such a choice implies that

θ(1/po − 1/q)
θ o := belongs to (0, 1) and d = (n − 1)θ o−1 (q/po − 1). (4.5.21)
1 − 1/q
Starting with the given function m as in (4.5.8)-(4.5.9) we may, granted (4.5.20)-
(4.5.21), run the same argument as in Case I with p replaced
 by po and θ replaced
by θ o (while retaining the same q and d). Since po ∈ n−1 n , 1), this yields

θo
m ∈ H po (Σ, σ) and m H p o (Σ,σ) ≤ Cm L1−θ
q (Σ,σ) · m q
o
L (Σ, | ·−x
,
d σ) (4.5.22)
o|

for some constant C = C(Σ, n, po, q, θ o ) ∈ (0, ∞). Moreover, m is a multiple of a

H po -molecule on Σ, hence m is also a multiple of a H 1 -molecule on Σ. On the
other hand, since m ∈ L q (Σ, σ) and we are currently assuming that q ∈ (1, ∞), we
conclude from [133, (3.6.27)] that

m ∈ H q (Σ, σ) and m H q (Σ,σ) ≤ Cm L q (Σ,σ) . (4.5.23)

Given that 1 ∈ (po, q) we may select θ ∈ (0, 1) with the property that

1 = (1 − θ)/po + θ/q (4.5.24)

and apply Proposition 4.2.2 (with p0 := po , q0 := q, and θ := θ) to obtain that

m ∈ H 1 (Σ, σ) and
1−θ θ
m H 1 (Σ,σ) ≤ Cm H p o (Σ,σ) · m H q (Σ,σ)

≤ Cm L(1−θ o )(1−θ)

q (Σ,σ) · m Lθoq(1− θ)
(Σ, | ·−x d σ) · m Lθ q (Σ,σ)
o|

= Cm L1−θ o +θo θ

q (Σ,σ) · m Lθoq(1− θ)
(Σ, | ·−x d σ) , (4.5.25)
o|

for some constant C = C(Σ, n, po, q, θ o ) ∈ (0, ∞). The second inequality in (4.5.25)
is based on (4.5.22)-(4.5.23). Since (4.5.21) and (4.5.24) imply

θ o (1 − θ) = θ (4.5.26)

which further entails 1 − θ o + θ o θ = 1 − θ o (1 − θ) = 1 − θ, we may refashion (4.5.25)

as
θ
m H 1 (Σ,σ) ≤ Cm L1−θ
q (Σ,σ) · m q
L (Σ, | ·−x
,
d σ) (4.5.27)
o|
4.5 Molecules in Hardy Spaces 179

which goes to prove that estimate (4.5.10) also holds when p = 1. This finishes the
treatment of Case II, and completes the proof of the first part in the statement of the
lemma.
In the opposite direction, assume the exponents p, q, ε, θ, d are as in (4.5.11), and
suppose m is a (p, q, ε)-molecule centered near the ball B(xo, r) on Σ, for some finite
r ∈ (0, diam Σ] which is not a constant function. Then, using notation and estimates
from Definition 4.5.1, for some C ∈ (0, ∞) depending only on n, p, q, ε, and the
Ahlfors regularity constants of Σ we may write
∞ ∫

q
m L q (Σ,σ) = |m| q dσ
k=0 A k (x o ,r)


∞
≤ Cr (n−1)(1−q/p) 2k(n−1)[1−(1+ε)q] = Cr (n−1)(1−q/p), (4.5.28)
k=0

and
∞ ∫

q
m L q (Σ, | ·−x d σ) = |m| q | · −xo | d dσ
o|
k=0 A k (x o ,r)


∞  
≤ Cr d+(n−1)(1−q/p) 2k d+(n−1)[1−(1+ε)q]

k=0

= Cr d+(n−1)(1−q/p)
, (4.5.29)

where in the last step above we have used the fact that our choice of parameters in
(4.5.11) guarantees that d + (n − 1)[1 − (1 + ε)q] < 0. From (4.5.28)-(4.5.29) and
the definition of d in (4.5.11) we then readily see that (4.5.12) holds in this case.
Finally, there remains to observe that in the case when Σ is bounded the constant
function given by m(x) := [σ(Σ)]−1/p for every x ∈ Σ also satisfies (4.5.12). 

Here is a useful consequence of the above lemma.

Corollary 4.5.3 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and abbreviate
σ := H n−1 Σ. Also, fix some point xo ∈ Σ and consider exponents p, q, d such that
  
p ∈ n−1
n ,1 , q ∈ [1, ∞) with q > p, and d > (n − 1) q/p − 1). (4.5.30)

Then  ∫ 
q 
m ∈ L Σ, (1 + | · −xo | )σ : d
m dσ = 0 (4.5.31)
Σ
embeds continuously into the Hardy space H p (Σ, σ). Moreover, each function be-
longing to the space defined in (4.5.31) is a multiple of a H p -molecule on Σ (in the
sense of Definition 4.5.1), and there exists a finite constant C > 0, independent of
xo and m, with the property that
180 4 Hardy Spaces on Ahlfors Regular Sets

m H p (Σ,σ) ≤ Cm L1−θ θ

q (Σ,σ) · m q d σ)
L (Σ, | ·−x o|
(4.5.32)
where θ := (n − 1)(q/p − 1)/d ∈ (0, 1).

Proof All claims in the statement are direct consequences of Lemma 4.5.2. 

In relation to the space defined in (4.5.31), it is useful to point out that having
fixed p, q as in (4.5.30) then a given (p, q, ε)-molecule m on Σ, in the sense of
Definition 4.5.1, belongs to (4.5.31) for some d as in (4.5.30) provided p > 1/(1 + ε)
(a condition automatically satisfied if ε ≥ 1/(n − 1)).
The integral pairing between a molecule and a function of bounded mean os-
cillations is always absolutely convergent. More specifically, we have the following
result.

Lemma 4.5.4 Let Σ ⊆ Rn be a closed set which  is Ahlfors

 regular and abbreviate
σ := H n−1 Σ. Also, pick two exponents p ∈ n−1 n , 1 and q ∈ (1, ∞], and fix some
ε ∈ (0, ∞). Then
∫
| f ||m| dσ < +∞ whenever f ∈ BMO(Σ, σ) and m
Σ (4.5.33)
is a (p, q, ε)-molecule on Σ (cf. Definition 4.5.1).

Proof Fix some function f ∈ BMO(Σ, σ) and suppose m is a (p, q, ε)-molecule

centered near the ball B(xo, r) on Σ (where the point xo ∈ Σ and r > 0). Also, let
the exponent q ∈ [1, ∞) be such that 1/q + 1/q = 1. Observe that [133, (7.4.122)]
(considered with (X, ρ, μ) := (Σ, | · − · |, σ), p := q and (n − 1)ε in place of ε/p)
presently gives

∞ ⨏ 1/q
2−k(n−1)ε
q
f − fB(xo,r)∩Σ dσ
k=1 B(x o ,2k r)∩Σ


∞ ⨏ 1/q
−k(n−1)ε q
≤C 2 f − fB(xo,2k r)∩Σ dσ
k=1 B(x o ,2k r)

≤ C f BMO(Σ,σ) . (4.5.34)

To proceed, recall the set Ak (xo, r) defined for each k ∈ N0 as in (4.5.2) and estimate
4.5 Molecules in Hardy Spaces 181
∫
f − fB(xo,r)∩Σ |m| dσ
Σ\B(x o ,r)


∞ ∫ 1/q
∫ 1/q
q
≤ f − fB(xo,r)∩Σ dσ |m| q dσ
k=1 A k (x o ,r) A k (x o ,r)


∞ ⨏
  1/q q 1/q
≤ σ B(xo, 2k r) ∩ Σ f − fB(xo,r)∩Σ dσ ×
k=1 B(x o ,2k r)∩Σ

  1/q−1/p
× 2k(n−1)[1/q−1−ε] σ B(xo, r) ∩ Σ

  1−1/p 
∞
≤ Cσ B(xo, r) ∩ Σ 2−k(n−1)ε ×
k=1
⨏ 1/q
q
× f − fB(xo,r)∩Σ dσ
B(x o ,2k r)∩Σ
  1−1/p
≤ Cσ B(xo, r) ∩ Σ  f BMO(Σ,σ) < +∞. (4.5.35)

Above, the first inequality is obtained by breaking up the domain of integration into
mutually disjoint dyadic annuli and then using Hölder’s inequality. Next, the second
inequality is a consequence of (4.5.1), the third inequality follows from the Ahlfors
regularity of Σ, while the last inequality is simply (4.5.34).∫Since, as seen from (4.5.4),
we also have m ∈ L 1 (Σ, σ), the reasoning so far gives that Σ\B(x ,r) | f ||m| dσ < +∞.
∫ o
As such, there remains to show that Σ∩B(x ,r) | f ||m| dσ < +∞. This, however, is a
o
consequence of [133, (7.4.105)] and (4.5.4). 

Here is a companion to Lemma 4.5.4, showing that the integral pairing between
a molecule and a Hölder function of appropriate order is also absolutely convergent.

Lemma 4.5.5 Let Σ ⊆ Rn be a closed set which  is Ahlfors

 regular and abbreviate
σ := H n−1 Σ. Also, pick two exponents p ∈ n−1 n , 1 and q ∈ (1, ∞], and fix some
ε ∈ (0, ∞). Then for each exponent α ∈ (0, 1) such that α < (n − 1)ε one has
∫
.
| f ||m| dσ < +∞ whenever f ∈ C α (Σ) and m is a
Σ (4.5.36)
(p, q, ε)-molecule on Σ (cf. Definition 4.5.1).
.
Proof Fix some function f ∈ C α (Σ) and consider a (p, q, ε)-molecule m centered
near the ball B(xo, r) on Σ. If q ∈ [1, ∞) is such that 1/q + 1/q = 1 we may reason
as in (4.5.34) then invoke [133, (7.4.127)] to estimate
182 4 Hardy Spaces on Ahlfors Regular Sets


∞ ⨏ 1/q
2−k(n−1)ε
q
f − fB(xo,r)∩Σ dσ
k=1 B(x o ,2k r)∩Σ


∞ ⨏ 1/q
2−k(n−1)ε
q
≤C f − fB(xo,2k r)∩Σ dσ
k=1 B(x o ,2k r)


∞
 α
≤C 2−k(n−1)ε 2k r  f C.α (Σ)
k=1

≤ Cr α  f C.α (Σ), (4.5.37)

where the convergence of the geometric series uses the fact that α < (n − 1)ε. Next,
if we proceed as in (4.5.35) and, this time, use (4.5.37) in the last inequality, we
obtain
∫
f − fB(xo,r)∩Σ |m| dσ
Σ\B(x o ,r)
  1−1/p
≤ Cσ B(xo, r) ∩ Σ ×

∞ ⨏ 1/q
2−k(n−1)ε
q
× f − fB(xo,r)∩Σ dσ
k=1 B(x o ,2k r)∩Σ

  1−1/p
≤ Cr α σ B(xo, r) ∩ Σ  f C.α (Σ) < +∞. (4.5.38)
∫
Given that m ∈ L 1 (Σ, σ) (cf. (4.5.4)), we conclude that Σ\B(x ,r) | f ||m| dσ < +∞.
o
Upon observing that we also have
∫
| f ||m| dσ ≤ sup | f (x)| m L 1 (Σ,σ) < +∞, (4.5.39)
Σ∩B(x o ,r) x ∈Σ∩B(x o ,r)

the proof is complete. 

4.6 Duality in Hardy Spaces

We have the following result (cf., e.g., [36, Theorem B, p. 593], [9, Theorem 7.20,
p. 323]) describing the dual of the Hardy spaces from Definition 4.2.1 in the range
p ≤ 1.

Theorem 4.6.1 Given a closed set Σ ⊆ Rn which is Ahlfors regular, abbreviate

σ := H n−1 Σ and fix an exponent p ∈ n−1
n , 1 . Then, in a quantitative fashion,
4.6 Duality in Hardy Spaces 183

⎧
.   
 ∗ ⎨ C (n−1)(1/p−1) (Σ) ∼ if p ∈ n−1
⎪ n ,1 ,
Σ unbounded =⇒ H p (Σ, σ) =
⎪ BMO(Σ, σ)  ∼ if p = 1,
⎩
(4.6.1)
(where the equivalence identifies functions which differ by a constant on Σ) and
 (n−1)(1/p−1)  
 p ∗ C (Σ) if p ∈ n−1n ,1 ,
Σ bounded =⇒ H (Σ, σ) = (4.6.2)
BMO(Σ, σ) if p = 1.

The actual nature of the identification of the dual of the Hardy spaces in (4.6.1) is
as follows. Having fixed some q ∈ [1, ∞] satisfying q > p, consider the assignment
taking any equivalence class [ f ] belonging to

⎧
.   
⎨ C (n−1)(1/p−1) (Σ) ∼ if p ∈ n−1
⎪ n ,1 ,
(4.6.3)
⎪ BMO(Σ, σ)  ∼ if p = 1,
⎩
 ∗
into the functional Λ[ f ] ∈ H p (Σ, σ) defined as

 ∫
  N
Λ[ f ], g := lim λj f a j dσ, (4.6.4)
N →∞ Σ
j=1

whenever
g ∈ H p (Σ, σ), {λ j } j ∈N ∈  p (N), and {a j } j ∈N is a
sequence of (p, q)-atoms on Σ, with the property that (4.6.5)
!
g= ∞ j=1 λ j a j in H (Σ, σ).
p

Then Λ[ f ] is unambiguously defined, and the assignment  [ f ] → Λ[ f ] is a bounded

∗
linear isomorphism from the space in (4.6.3) onto H p (Σ, σ) , with bounded inverse.
In particular, identifying Λ[ f ] with [ f ] yields (bearing in mind the format of the norm
.
on BMO(Σ, σ); cf. [133, (7.4.95)])
   
 f BMO(Σ,σ) =  [ f ] BMO(Σ,σ)/∼ ≈ Λ[ f ] (H 1 (Σ,σ))∗
  
= sup Λ[ f ], g : g ∈ H 1 (Σ, σ) with g H 1 (Σ,σ) = 1
  
= sup [ f ], g : g ∈ H 1 (Σ, σ) with g H 1 (Σ,σ) = 1 , (4.6.6)

uniformly for f ∈ BMO(Σ, σ) (compare with [133, Proposition 7.4.12]), and

184 4 Hardy Spaces on Ahlfors Regular Sets
   
 f C.(n−1)(1/p−1) (Σ) =  [ f ]  .(n−1)(1/p−1)
C (Σ)/∼
≈ Λ[ f ] (H p (Σ,σ))∗
  
= sup Λ[ f ], g : g ∈ H p (Σ, σ) with g H p (Σ,σ) = 1
  
= sup [ f ], g : g ∈ H p (Σ, σ) with g H p (Σ,σ) = 1 , (4.6.7)
.  
uniformly for f ∈ C (n−1)(1/p−1) (Σ) with p ∈ n−1 n , 1 (compare with [133, Proposi-
tion 7.4.8]). Also, (4.6.4) yields the duality formula
  ! ∫
[ f ], g = ∞ j=1 λ j Σ f a j dσ whenever [ f ] belongs to the
space in (4.6.3), g ∈ H p (Σ, σ), {λ j } j ∈N ∈  p (N), and {a j } j ∈N
! (4.6.8)
is a sequence of (p, q)-atoms on Σ, such that g = ∞ j=1 λ j a j in
H p (Σ, σ).
Another consequence is that there exists a finite constant C > 0, which depends only
on the environment, having the property that
  
   f C.(n−1)(1/p−1) (Σ) · g H p (Σ,σ) if p ∈ n−1n ,1 ,
[ f ], g ≤ C (4.6.9)
 f BMO(Σ,σ) · g H 1 (Σ,σ) if p = 1.

Finally, there is an analogous interpretation of the identification of the dual of the

Hardy space in (4.6.2) (plus naturally accompanying results as in (4.6.6), (4.6.7),
(4.6.8), (4.6.9)), this time involving inhomogeneous Hölder spaces and without hav-
ing to consider equivalence classes of functions (modulo constants).
 
Hence, if Σ ⊆ Rn is a closed Ahlfors regular set, σ := H n−1 Σ, and p ∈ n−1 n ,∞ ,
we have:
.   
⎧ C (n−1)(1/p−1) (Σ) ∼ if p ∈ n−1
⎪
⎪
⎪ n , 1 and Σ unbounded,
⎪
⎪  
⎪
⎪ C (n−1)(1/p−1) (Σ) if p ∈ n−1
⎪
⎪ n , 1 and Σ bounded,
 p ∗ ⎪ ⎨
⎪ 
H (Σ, σ) = BMO(Σ, σ) ∼ if p = 1 and Σ unbounded,
⎪
⎪
⎪
⎪
⎪
⎪ BMO(Σ, σ) if p = 1 and Σ bounded,
⎪
⎪
⎪
⎪
⎪ L p/(p−1) (Σ, σ) if p ∈ (1, ∞).
⎩
(4.6.10)
Continue to assume that Σ ⊆ Rn is a closed Ahlfors regular  set,
 abbreviate
σ := H n−1 Σ. In general, the Hardy space H p (Σ, σ) with p ∈ n−1 n , 1 is not a dual
space, but the case p = 1 is different. To elaborate, following [36] we shall4

denote by CMO(Σ, σ) the closure in BMO(Σ, σ) of C00 (Σ), the space

(4.6.11)
of all continuous functions on Σ which vanish at infinity.

4 In [36] the authors employ the notation VMO(Σ, σ), in place of CMO(Σ, σ) as we do here (since
the former already has a precise, typically distinct, meaning; see (3.1.1))
4.6 Duality in Hardy Spaces 185

Hence, by design,

CMO(Σ, σ) is a closed linear subspace of BMO(Σ, σ). (4.6.12)

Since Lipc (Σ) is a linear subspace of the space of continuous functions on Σ van-
ishing at infinity which is dense with respect to  ·  L ∞ (Σ,σ) , and since convergence
in L ∞ (Σ, σ) implies convergence in BMO(Σ, σ), we conclude that (4.6.11) self-
improves to

CMO(Σ, σ) is the closure of Lipc (Σ) in BMO(Σ, σ). (4.6.13)

Let us also observe that

f + c belongs to CMO(Σ, σ) whenever f ∈ CMO(Σ, σ) and c ∈ C,
(4.6.14)
and CMO(Σ, σ) is the closure of Lipc (Σ) + C in BMO(Σ, σ).

Indeed, if Σ is bounded then constants belong to Lipc (Σ), whereas if Σ is unbounded

this follows from the fact that  · BMO(Σ,σ) is insensitive to additive constants. Hence,
constants are contained in CMO(Σ, σ) and

.
CMO(Σ, σ) := CMO(Σ, σ)/∼ is a closed linear
(4.6.15)
.
subspace of BMO(Σ, σ) := BMO(Σ, σ)/∼.

In addition, from (4.6.13) and (3.1.50) we see that

if Σ ⊆ Rn is a compact Ahlfors regular set and
(4.6.16)
σ := H n−1 Σ, then CMO(Σ, σ) = VMO(Σ, σ).
Also, according to [36, Theorem 4.1, p. 638],

H 1 (Σ, σ) is the dual space of CMO(Σ, σ); more precisely, for each
continuous linear functional Λ ∫on CMO(Σ, σ) there exists a unique
f ∈ H 1 (Σ, σ) such that Λ(φ) = Σ f φ dσ for all continuous functions (4.6.17)
φ on Σ which vanish at infinity, and  f  H 1 (Σ,σ) is equivalent to the
linear functional norm of Λ.
From (4.6.11), (4.6.17),
 Lemma4.6.5 (with p := 1), and Theorem 4.6.1 we see that
∗
if a functional Λ ∈ CMO(Σ, σ) is identified with some f ∈ H 1 (Σ, σ) in the sense
of (4.6.17), it follows that

H 1 (Σ,σ) f , [·] BMO(Σ,σ)/∼ if Σ unbounded,
Λ= on CMO(Σ, σ). (4.6.18)
H 1 (Σ,σ) f , · BMO(Σ,σ) if Σ bounded,

Let us take a separate look at the case when the set Σ is unbounded. In this
scenario, denote by π : CMO(Σ, σ) → CMO(Σ, σ)/∼ the canonical projection
∗
onto equivalence classes modulo constants. Then for each Λ ∈ CMO(Σ, σ)/∼ it
 ∗
follows that the composition Λ := Λ ◦ π belongs to CMO(Σ, σ) . Based on this
186 4 Hardy Spaces on Ahlfors Regular Sets

and (4.6.18) we then conclude that there exists a unique function f ∈ H 1 (Σ, σ) such
that

Λ= H 1 (Σ,σ) f, · BMO(Σ,σ)/∼ on CMO(Σ, σ)/∼ and

(4.6.19)
 f  H 1 (Σ,σ) is equivalent to the linear functional norm of Λ.

This proves that the assignment

 ∗
H 1 (Σ, σ)  f → Λ f := H 1 (Σ,σ) f, · BMO(Σ,σ)/∼ ∈ CMO(Σ, σ)/∼ (4.6.20)

is well defined, linear, continuous, and surjective. From (4.6.11), Lemma 4.6.5 (with
p := 1), and [133, Corollary 3.7.3] we also see that the assignment (4.6.20) is
injective. Ultimately, this discussion shows that

 Σ is unbounded
 ∗ then H (Σ, σ) may be identified with
if 1
(4.6.21)
CMO(Σ, σ)/∼ via the isomorphism (4.6.20).

Moving on, we note that, collectively, (4.6.16) and (4.6.17) give

 ∗
H 1 (Σ, σ) = VMO(Σ, σ) if Σ is bounded. (4.6.22)

More precisely, whenever the set Σ is assumed

 to be bounded, from (4.6.16)-(4.6.18)
∗
we see that for any functional Λ ∈ VMO(Σ, σ) there exists a unique function
f ∈ H (Σ, σ) with  f  H 1 (Σ,σ) equivalent to the linear functional norm of Λ and
1

such that
Λ = H 1 (Σ,σ) f , · BMO(Σ,σ) on VMO(Σ, σ). (4.6.23)
Another point of view on (4.6.17), stemming from the discussion in [36, p. 638]
(and (4.6.13)), is as follows. If μ is a Borel measure on Σ with the property that there
exists a constant C ∈ (0, ∞) such that
∫
φ dμ ≤ CφBMO(Σ,σ) for all φ ∈ Lipc (Σ), (4.6.24)
Σ

then μ << σ and the Radon-Nikodym derivative dμ/dσ belongs to H 1 (Σ, σ) and
satisfies  dμ 
 
  1 ≤ C. (4.6.25)
dσ H (Σ,σ)
In particular, this implies that

if f ∈ L 1 (Σ, σ) has the property that there exists a constant C ∈ (0, ∞)

∫
such that Σ φ f dσ ≤ CφBMO(Σ,σ) for each φ ∈ Lipc (Σ), then the (4.6.26)
function f belongs to H 1 (Σ, σ) and satisfies  f  H 1 (Σ,σ) ≤ C.

Moving on, we next discuss the following result in relation to the space (4.6.11).
4.6 Duality in Hardy Spaces 187

Proposition 4.6.2 Suppose Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, let { f j } j ∈N be a bounded sequence in H 1 (Σ, σ). Then there
exist f ∈ H 1 (Σ, σ) and a subsequence { f jk }k ∈N of { f j } j ∈N such that

lim f jk , φ = f , φ for every φ ∈ CMO(Σ, σ), (4.6.27)

k→∞

where ·, · is the duality pairing described in (4.6.17). In addition, there exists some
constant C = C(Σ) ∈ (0, ∞) with the property that

 f  H 1 (Σ,σ) ≤ C · sup  f j  H 1 (Σ,σ) . (4.6.28)

j ∈N

Proof For the claims pertaining to (4.6.27) see [36, Lemma 4.2, p. 638]. With this
in hand, the estimate in (4.6.28) is a consequence of (4.6.17). 
Proposition 4.6.2 should be compared with the weak-∗ convergence result record-
ed next, which was first proved Jones and Journé in [104] in the case when Σ is a
hyperplane.
Proposition 4.6.3 Let Σ be a closed Ahlfors regular subset of Rn and abreviate
σ := H n−1 Σ. Suppose { f j } j ∈N is a bounded sequence in H 1 (Σ, σ) with the property
that the limit f (x) := lim f j (x) exists for σ-a.e. point x ∈ Σ. Then f ∈ H 1 (Σ, σ)
j→∞
and { f j } j ∈N is weak-∗ convergent to f , that is,

lim f j , φ = f , φ for every φ ∈ CMO(Σ, σ), (4.6.29)

j→∞

where ·, · is the duality pairing described in (4.6.17). In addition, there exists some
constant C = C(Σ) ∈ (0, ∞) such that

 f  H 1 (Σ,σ) ≤ C · sup  f j  H 1 (Σ,σ) . (4.6.30)

j ∈N

Proof All claims pertaining to (4.6.29) follow from [99, Theorem 1.1]. The estimate
in (4.6.30) then becomes a consequence of this and (4.6.17). 
We next proceed to describe a number of useful compatibility results involving
duality pairings. The first such result establishes the compatibility between the duality
pairing from Theorem 4.6.1 and the distributional pairing from [133, (4.1.35)].
Lemma 4.6.4 Let Σ be a closed set in Rn which is Ahlfors regular and abbreviate
σ := H n−1 Σ.
   
(a) Suppose p ∈ n−1 n , 1 and define α := (n − 1) p1 − 1 ∈ (0, 1). Then for each
 
distribution f ∈ H p (Σ, σ) → Lipc (Σ) and each g ∈ Lipc (Σ) → C α (Σ) one
has
  
  H p (Σ,σ) f , [g] C.α (Σ)/∼ if Σ is unbounded,
(Lip c (Σ)) f , g Lip c (Σ) =  
H p (Σ,σ) f , g C α (Σ) if Σ is bounded.
(4.6.31)
188 4 Hardy Spaces on Ahlfors Regular Sets
 
(b) If p = 1, for each distribution f ∈ H 1 (Σ, σ) → L 1 (Σ, σ) → Lipc (Σ) and
each function g ∈ Lipc (Σ) → L ∞ (Σ, σ) → BMO(Σ, σ) one has
  
  H 1 (Σ,σ) f , [g] BMO(Σ,σ)/∼ if Σ is unbounded,
(Lip c (Σ)) f , g Lip c (Σ) =  
H 1 (Σ,σ) f , g BMO(Σ,σ) if Σ is bounded.
(4.6.32)
 
Proof Fix q ∈ (1, ∞), and consider f ∈ H p (Σ, σ) with p ∈ n−1 n , 1 . Theorem 4.4.1
guarantees the existence of a numerical sequence {λ! j } j ∈N ∈  (N) and a family of
p

(p, q)-atoms a j , indexed by j ∈ N, such that if f N := j=1 λ j a j for each N ∈ N then

N

f = lim f N in H p (Σ, σ). (4.6.33)

N →∞

Assume next that we are in the setting of part (a) and that Σ is unbounded. Based
on (4.6.33) and (4.2.8) we may write

  
N
 
(Lip c (Σ)) f, g Lip c (Σ) = lim λ j (Lipc (Σ)) a j , g Lipc (Σ)
N →∞
j=1

 ∫
N
 
= lim λj a j g dσ = H p (Σ,σ) f , [g] C.α (Σ)/∼ (4.6.34)
N →∞ Σ
j=1

where the second equality in (4.6.34) relies on [133, Proposition 4.1.4], and the
last equality is provided by (4.6.8). This establishes (4.6.31) in the case when Σ
is unbounded. When Σ is bounded a similar argument proves the corresponding
equality in (4.6.31). This finishes the proof of the claim in part (a). Finally, the claim
stated in part (b) may be justified in a similar fashion, this time relying on the duality
results in (4.6.1)-(4.6.2) corresponding to p = 1. 
The second compatibility result involves the duality brackets between the Hardy
space H 1 and the John-Nirenberg space, on the one hand, and dual Lebesgue spaces,
on the other hand.

Lemma 4.6.5 Suppose Σ ⊆ Rn is a closed unbounded set which is also Ahlfors

regular, and abbreviate σ := H n−1 Σ. Consider f ∈ BMO(Σ, σ) ∩ L p (Σ, σ) and
g ∈ H 1 (Σ, σ) ∩ L p (Σ, σ) where the exponents p ∈ [1, ∞), p ∈ (1, ∞] are such that
1/p + 1/p = 1. Then, with ·, · denoting the duality bracket between the John-
Nirenberg space of functions of bounded mean oscillations on Σ, modulo constants,
and the Hardy space H 1 on Σ (cf. Theorem 4.6.1), one has (using the piece of notation
introduced in [133, (7.4.94)]
∫
[ f ], g = f g dσ. (4.6.35)
Σ

In particular, identity (4.6.35) holds for each f ∈ L ∞ (Σ, σ) ⊆ BMO(Σ, σ) and

g ∈ H 1 (Σ, σ) ⊆ L 1 (Σ, σ).
4.6 Duality in Hardy Spaces 189

Moreover, a similar formula is valid in the case when Σ is bounded, replacing [ f ]

by f in the left-hand side of (4.6.35).
Proof Theorem 4.4.3 guarantees the existence of a numerical sequence {λ j } j ∈N ⊆ C
along with a sequence {a j } j ∈N of (1, 2)-atoms on Σ with the property that
!
g N := N j=1 λ j a j converges to g as N → ∞, both in (4.6.36)
the space H 1 (Σ, σ) and the space L p (Σ, σ).

Then, thanks to Theorem 4.6.1 (cf. (4.6.4)) and L p -L p duality, in the case when Σ
is unbounded we may write
∫ ∫
[ f ], g = lim f g N dσ = f g dσ, (4.6.37)
N →∞ Σ Σ

which establishes (4.6.35). Finally, the case when Σ is bounded is treated in a

completely analogous manner. 
In a parallel fashion to Lemma 4.6.5 we have the following compatibility result.
Lemma 4.6.6 Let Σ ⊆ Rn be a closed set which is Ahlfors regular and unbounded.
 n−1 σ := H Σ and pick p, . p ∈ (1, ∞) such that 1/p+1/p = 1, along with
Abbreviate n−1

q ∈ n , 1 . Also, suppose f ∈ C (n−1)(1/q−1) (Σ) ∩ L p (Σ, σ) and g ∈ H q (Σ, σ) ∩

L p (Σ, σ). Then, with ·, · denoting the duality bracket between homogeneous Hölder
spaces, modulo constants, and Hardy spaces (cf. Theorem 4.6.1), there holds
∫
[ f ], g = f g dσ. (4.6.38)
Σ

In addition, corresponding to the case when p = 1, formula (4.6.38) holds for each
function f ∈ C (n−1)(1/q−1) (Σ) and each g ∈ H q (Σ, σ) ∩ H 1 (Σ, σ).
Moreover, similar formulas are valid in the case when Σ is bounded, this time
using inhomogeneous Hölder spaces to begin with and replacing [ f ] by f in the
left-hand side of (4.6.38).
Proof The same argument as in the proof of Lemma 4.6.5, based on Theorem 4.4.3
and Theorem 4.6.1, works here just as well. 
It turns out that the duality brackets from Lemma 4.6.5 and Lemma 4.6.6 are also
compatible, in the manner indicated below.
Lemma 4.6.7 Let Σ ⊆ Rn be a closed set which is Ahlfors  regular and un-
bounded. Abbreviate σ := H n−1 Σ and pick some p ∈ n−1 , 1 . Also, fix some
.(n−1)(1/p−1) n
f ∈C (Σ) ∩ BMO(Σ, σ) along with g ∈ H (Σ, σ) ∩ H (Σ, σ). Then
p 1

   
. [ f ], g H p (Σ,σ) = BMO(Σ,σ)/∼ [ f ], g H 1 (Σ,σ) . (4.6.39)
C (n−1)(1/p−1) (Σ)/∼

Furthermore, a similar formula is valid in the case when Σ is bounded, this time
using inhomogeneous Hölder spaces to begin with, no longer modding out constants,
and replacing [ f ] by f .
190 4 Hardy Spaces on Ahlfors Regular Sets

Proof Once again, the argument in the proof of Lemma 4.6.5, relying on Theo-
rem 4.4.3 and Theorem 4.6.1, applies here as well. 
Finally, we record a compatibility result for the Hölder-Hardy duality bracket
considered for two choices of the parameters involved in the definitions of these
spaces.

Lemma 4.6.8 Suppose Σ is a closed unbounded Ahlfors regular  n−1 subset

 on Rn and,
as before, abbreviate σ := H Σ. Pick some p1, p2 ∈ n , 1 . Consider some
n−1
. .
f ∈ C (n−1)(1/p1 −1) (Σ) ∩ C (n−1)(1/p2 −1) (Σ) and g ∈ H p1 (Σ, σ) ∩ H p2 (Σ, σ). Then
   
. [ f ], g H p1 (Σ,σ) = C.(n−1)(1/p2 −1) (Σ)/∼ [ f ], g H p2 (Σ,σ) . (4.6.40)
C (n−1)(1/p1 −1) (Σ)/∼

In addition, a similar formula is valid in the case when Σ is bounded, this time using
inhomogeneous Hölder spaces to begin with, no longer modding out constants, and
replacing [ f ] by f .

Proof Once more, the same argument as in the proof of Lemma 4.6.5, based on
Theorem 4.4.3 and Theorem 4.6.1, works in the present setting. 
For further use we note in the lemma below that the integral pairings with arbitrary
atoms determine a function uniquely (eventually, up to constants).

Lemma 4.6.9 Let Σ ⊆ Rn be a closed

 set
 which is Ahlfors regular and abbreviate
σ := H n−1 Σ. Also, assume p ∈ n−1n , 1 and q, q ∈ [1, ∞] are such that p < q and
q
1/q + 1/q = 1. Then a function φ ∈ Lloc (Σ, σ) satisfying
∫
φ a dσ = 0 for each (p, q)-atom a on Σ (4.6.41)
Σ

is necessarily constant on Σ. Moreover, if Σ is bounded then any φ ∈ L q (Σ, σ)

satisfying (4.6.41) necessarily vanishes identically on Σ.

Proof Pick a point x0 ∈ Σ along with some number r0 ∈ (0, diam Σ). Then for each
x ∈ Σ and r ∈ (0, diam Σ) arbitrary, the function

a = 1B(x,r)∩Σ /σ(B(x, r) ∩ Σ) − 1B(x0,r0 )∩Σ /σ(B(x0, r0 ) ∩ Σ) (4.6.42)

∫
is a scalar multiple of a (p, q)-atom on Σ. Consequently, Σ φ a dσ = 0, which then
implies ⨏ ⨏
φ dσ = c, where c := φ dσ. (4.6.43)
B(x,r)∩Σ B(x0,r0 )∩Σ

Upon letting r → 0+ in (4.6.43) ultimately gives φ(x) = c for σ-a.e. x ∈ Σ, by

virtue of Lebesgue’s Differentiation Theorem (see [133, Proposition 7.4.4], which
presently applies thanks to [133, Lemma 3.6.4]). In the case when Σ is bounded, the
fact that a ≡ 1 is a multiple of a (p, q)-atom forces, in the context of (4.6.41), the
constant function φ to actually vanish identically on Σ. 
4.7 Richness of the Duals of Lorentz-Based Hardy Spaces 191

4.7 Richness of the Duals of Lorentz-Based Hardy Spaces

The dual of the Hardy spaces from Definition 4.2.1 has been precisely identified in
Theorem 4.6.1 in the range p ≤ 1. Our present goal is to show that Lorentz-based
Hardy spaces continue to have fairly rich dual spaces. A first result of this flavor is
described next.

Lemma 4.7.1 Consider a set Σ ⊆ Rn which is closed and Ahlfors regular and
abbreviate σ := H n−1 Σ. Fix p ∈ n−1
n , ∞ and q ∈ (0, ∞], then pick α0 ∈ (0, 1) and
p1 ∈ (1, ∞) such that
 α0  −1  
1 −1
p0 := 1 + n−1 < p and p1 := 1 − p1 > p. (4.7.1)

Then .  ∗
C α0 (Σ) ∩ L p1 (Σ, σ) → H p,q (Σ, σ)
(4.7.2)
in a continuous and injective fashion.
Moreover,
∫
 
(H p, q (Σ,σ))∗ f, g H p, q (Σ,σ) = f g1 dσ (4.7.3)
Σ
 .  
C α 0 (Σ)/∼
[ f ], g0 H p0 (Σ,σ) if Σ is unbounded,
+  
C α0 (Σ) f , g0 H p0 (Σ,σ) if Σ is bounded,

whenever
.  ∗
f ∈ C α0 (Σ) ∩ L p1 (Σ, σ) → H p,q (Σ, σ) and
g ∈ H p,q (Σ, σ) ⊆ H p0 (Σ, σ) + L p1 (Σ, σ) is written as5 (4.7.4)
g = g0 + g1 with g0 ∈ H p0 (Σ, σ) and g1 ∈ L p1 (Σ, σ).
.
Proof Pick an arbitrary function f ∈ C α0 (Σ)∩ L p1 (Σ, σ). In relation to this, consider
the linear and continuous functionals

Λ(0)
f : H (Σ, σ) −→ C
p0
(4.7.5)

acting on each g ∈ H p0 (Σ, σ) according to (cf. Theorem 4.6.1)

  
. [ f ], g H p0 (Σ,σ) if Σ is unbounded,
(0) C α0 (Σ)/∼
Λ f (g) :=   (4.7.6)
C α0 (Σ) f , g H p0 (Σ,σ) if Σ is bounded,

and
∫
Λ(1)
f : L (Σ, σ) −→ C,
p1
Λ(1)
f (g) := f g dσ, ∀g ∈ L p1 (Σ, σ). (4.7.7)
Σ
192 4 Hardy Spaces on Ahlfors Regular Sets

Since Lemma 4.6.6 then ensures that Λ(0) (1)

f and Λ f are compatible with one another,
we may unambiguously define the linear and bounded mapping

Λ f : H p0 (Σ, σ) + L p1 (Σ, σ) → C, Λ f (g) := Λ(0) (1)

f (g0 ) + Λ f (g1 ),

whenever g ∈ H p0 (Σ, σ) + L p1 (Σ, σ) is expressed as (4.7.8)

g = g0 + g1 with g0 ∈ H p0 (Σ, σ) and g1 ∈ L p1 (Σ, σ).

By design, this satisfies

Λf = Λ(0)
f and Λ f p
= Λ(1)
f . (4.7.9)
H p0 (Σ,σ) L 1 (Σ,σ)

Thanks to Proposition 1.3.7, Theorem 4.3.1, and [133, (3.6.27)] to conclude that
Λ f maps H p,q (Σ, σ) (itself, a subspace of H p0 (Σ, σ) + L p1 (Σ, σ)), equipped with
the quasi-norm  ·  H p, q (Σ,σ) , boundedly into C. Thus, regarding this as a linear and
bounded mapping Λ f : H p,q (Σ, σ) → C, the reasoning so far yields a well-defined
and linear assignment
.  ∗
C α0 (Σ) ∩ L p1 (Σ, σ)  f −→ Λ f ∈ H p,q (Σ, σ) . (4.7.10)

From (4.7.5)-(4.7.7), (4.6.9), Hölder’s inequality, and (1.3.64) it follows that this
assignment is also continuous. As such, the claim in (4.7.2) follows (by identifying
each f with Λ f ) as soon as we show that the assignment (4.7.10) is also injective.
.
To this end, suppose
 f ∈ C α0 (Σ) ∩ L p1 (Σ, σ) is such that Λ f = 0 as a functional
∗
in H p,q (Σ, σ) . In such a scenario, (1.3.44) and (4.3.3) allow us to conclude that
actually Λ f H p0 (Σ,σ)∩L p1 (Σ,σ) ≡ 0. In particular, Λ f (a) = 0 for each atom a on Σ
∫which, on account of the very last property recorded in (1.3.66) and (4.7.7), implies
Σ
f a dσ = 0 for each atom a on Σ. With this in hand, Lemma 4.6.9 applies and
ultimately proves that f = 0, as wanted. This finishes the proof of (4.7.2).
Finally, whenever f , g are as in (4.7.4), by virtue of (4.7.10), and (4.7.9) we may
write
 
(H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) = Λ f (g) = Λ f (g0 + g1 )

= Λ f (g0 ) + Λ f (g1 ) = Λ(0) (0)

f (g0 ) + Λ f (g1 ), (4.7.11)

from which (4.7.3) follows on account of (4.7.5)-(4.7.6), and (4.7.7). 

Here is a companion result to Lemma 4.7.1, once again pointing to the fact that
Lorentz-based Hardy spaces have reasonably rich dual spaces.

Lemma 4.7.2 Suppose Σ ⊆ Rn is an unbounded closed  set which is also Ahlfors

regular, and abbreviate σ := H n−1 Σ. Fix p ∈ n−1
n , 1 and q ∈ (0, ∞], then pick
α0, α1 ∈ (0, 1) such that
 α0  −1  α1  −1
p0 := 1 + n−1 < p < p1 := 1 + n−1 . (4.7.12)
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 193

Then
 .α .   ∗
C 0 (Σ) ∩ C α1 (Σ) ∼ embeds into H p,q (Σ, σ)
(4.7.13)
in a continuous and injective fashion,

and
   
(H p, q (Σ,σ))∗ [ f ], g H p, q (Σ,σ) = C.α0 (Σ)/∼ [ f ], g0 H p0 (Σ,σ)
 
+ C.α1 (Σ)/∼ [ f ], g1 H p1 (Σ,σ) (4.7.14)

whenever
. .
f belongs to the intersection C α0 (Σ) ∩ C α1 (Σ) and
g ∈ H p,q (Σ, σ) ⊆ H p0 (Σ, σ) + H p1 (Σ, σ) is written as (4.7.15)
g = g0 + g1 with g0 ∈ H p0 (Σ, σ) and g1 ∈ H p1 (Σ, σ).

Furthermore, similar claims are valid in the case when Σ is bounded, this time
using inhomogeneous Hölder spaces and no longer modding out constants.
Proof The same type of argument as in the proof of Lemma 4.7.1 works in the
present setting, with Lemma 4.6.8 now employed in place of Lemma 4.6.6. 

4.8 Weak-∗ Convergence and More on the Compatibility of

Pairings

In concert, the Sequential Banach-Alaoglu Theorem recorded in [133, (3.6.22)]

(whose present applicability is ensured by [133, (3.6.3)]) and Lebesgue’s Dominated
Convergence Theorem give that

if (X, μ) is a sigma-finite measure space, the measure μ is separable,

and if { f j } j ∈N is a bounded sequence in L ∞ (X, μ) with the property
that f (x) := lim f j (x) exists for μ-a.e. point x ∈ X, then f ∈ L ∞ (X, μ) (4.8.1)
j→∞
 ∗
and lim f j = f weak-∗ in L ∞ (X, μ) = L 1 (X, μ) .
j→∞

Here is a version of this result in which the space of essentially bounded functions
is replaced by the John-Nirenberg space of functions of bounded mean oscillations.
Lemma 4.8.1 Consider a closed, unbounded set Σ ⊆ Rn , which is also Ahlfors
regular, and abbreviate σ := H n−1 Σ. Pick f ∈ Lloc
1 (Σ, σ) and suppose { f }
j j ∈N is
a sequence of functions in BMO(Σ, σ) satisfying

sup  f j BMO(Σ,σ) < +∞ and f j → f in Lloc

1
(Σ, σ) as j → ∞. (4.8.2)
j ∈N
194 4 Hardy Spaces on Ahlfors Regular Sets

Then f belongs to BMO(Σ, σ) and

.
lim [ f j ] = [ f ] weak-∗ in BMO(Σ, σ). (4.8.3)
j→∞

That is, with ·, · denoting the duality bracket between the John-Nirenberg space
of functions of bounded mean oscillations on Σ, modulo constants, and the Hardy
space H 1 on Σ (cf. Theorem 4.6.1),
   
for each g ∈ H 1 (Σ, σ) one has [ f j ], g → [ f ], g as j → ∞. (4.8.4)

In addition, a similar weak-∗ convergence result holds in the case when Σ is

bounded, namely
lim f j = f weak-∗ in BMO(Σ, σ). (4.8.5)
j→∞

Proof Suppose (4.8.4) fails. Then there exists some function g ∈ H 1 (Σ, σ) along
with some strictly increasing sequence { jk }k ∈N ⊆ N such that
   
any sub-sequence of [ f jk ], g does not converge to [ f ], g . (4.8.6)
k ∈N
 
On the other hand, the first condition in (4.8.2) implies that the sequence [ f jk ] k ∈N
 ∗
.
is bounded in BMO(Σ, .
σ). Since BMO(Σ, σ) = H 1 (Σ, σ) (cf. (4.6.1) in The-
orem 4.6.1), the Sequential Banach-Alaoglu Theorem recalled in [133, (3.6.22)]
(whose applicability in the present case is ensured by the separability
 result from
Proposition 4.4.2) implies the existence of a sub-sequence, say [ f jk i ] i ∈N , along
with some function f ∈ BMO(Σ, σ), such that
.
[ f jk i ] −→ [ f ] weak-∗ in BMO(Σ, σ), as i → ∞. (4.8.7)

Then for each (1, ∞)-atom a on Σ we may write

∫ ∫
 
f a dσ = lim f jk i a dσ = lim [ f jk i ], a
Σ i→∞ Σ i→∞
∫
 
= [ f ], a = f a dσ. (4.8.8)
Σ

Above, the first equality is a consequence of the last property in (4.8.2), the second
and fourth equalities are implied by (4.6.4), and the third equality follows from
(4.8.7). Granted (4.8.8), we may invoke Lemma 4.6.9 to conclude that

f − f is constant on Σ. (4.8.9)
.
In turn, from (4.8.7) and (4.8.9) we see that [ f jk i ] → [ f ] weak-∗ in BMO(Σ, σ) as
i → ∞. Hence, in particular,
   
[ f jk i ], g −→ [ f ], g as i → ∞, (4.8.10)
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 195

which contradicts (4.8.6). This finishes the proof of (4.8.3). Finally, the weak-
convergence result recorded in (4.8.5) in the case when Σ is bounded is proved
similarly. 
The formal convolution paring between the logarithm, viewed as a function in
BMO, and a fixed function in the Hardy space H 1 , turns out to be continuous and
bounded. This is made precise in the lemma below.
Lemma 4.8.2 Suppose Σ is a closed Ahlfors regular set in Rn . Let σ := H n−1 Σ,
fix some f ∈ H 1 (Σ, σ), and consider the function
 
ln |x − ·|, f if Σ is bounded,
F(x) :=   ∀x ∈ Rn, (4.8.11)
[ln |x − ·|], f if Σ is unbounded,

where ·, · stands for the duality bracket between the John-Nirenberg space of
functions of bounded mean oscillations on Σ (modulo constants, if Σ is unbounded)
and the Hardy space H 1 on Σ, described in Theorem 4.6.1.
Then F is a well-defined continuous function in Rn . Moreover, the restriction of
F to Rn \ Σ is of class C ∞ , and its partial derivatives in Rn \ Σ may be computed by
differentiating directly under the integral pairing in (4.8.11).
Finally, if Σ is unbounded then there exists a constant C = C(Σ) ∈ (0, ∞) with
the property that
sup |F(x)| ≤ C f  H 1 (Σ,σ) . (4.8.12)
x ∈R n

Proof In a first stage, work under the assumption that Σ is unbounded. From [133,
Lemma 7.4.13] and Theorem 4.6.1 we see that the function F is well defined,
bounded, and that (4.8.12) holds. To establish the continuity of F, let {x j } j ∈N ⊆ Rn
be a sequence convergent to a point x ∈ Rn . Define
   
g := ln |x − ·| and g j := ln |x j − ·| for each j ∈ N. (4.8.13)
Σ Σ

Then [133, Lemma 7.4.13] ensures that {g j } j ∈N is a bounded sequence in BMO(Σ, σ).
Also, Vitali’s Convergence Theorem (cf., e.g., [130, Theorem 14.29, p. 559]) guar-
antees that g j → g in Lloc
1 (Σ, σ) as j → ∞. Granted these properties, Lemma 4.8.1
   
applies and gives that [g j ], f → [g], f as j → ∞. In turn, this readily implies
that F is continuous in Rn .
To prove that F is smooth on the complement of Σ, recall from Theorem 4.4.1
that there exist a numerical sequence {λ j } j ∈N ∈  p (N) and a sequence {a j } j ∈N of
(1, ∞)-atoms on Σ such that

f = lim fm in H 1 (Σ, σ) where, for each m ∈ N,

m→∞

m
(4.8.14)
∞
fm := λ j a j ∈ Lcomp (Σ, σ).
j=1

If for each m ∈ N we now set

196 4 Hardy Spaces on Ahlfors Regular Sets

Fm (x) := [ln |x − ·|], fm , ∀x ∈ Rn \ Σ, (4.8.15)

then (4.6.8) permits us to express


m ∫
Fm (x) = λj ln |x − y|a j (y) dσ(y)
j=1 Σ

∫
= ln |x − y| fm (y) dσ(y), ∀x ∈ Rn \ Σ. (4.8.16)
Σ

It is then clear from this representation that for each m ∈ N we have

Fm ∈ C ∞ (Rn \ Σ) and, for each α ∈ N0n and x ∈ Rn \ Σ,

∫   
α    (4.8.17)
(∂ Fm )(x) = ∂xα ln |x − y| fm (y) dσ(y) = ∂xα ln |x − ·| , fm ,
Σ
 
where the last equality relies on (4.6.8) plus the fact that ∂xα ln |x − ·| belongs to
L ∞ (Σ, σ) if |α| > 0 and to BMO(Σ, σ) if |α| = 0. If for each given α ∈ N0n and each
given x ∈ Rn \ Σ we now introduce
  α  
∂x ln |x − ·|  L ∞ (Σ,σ) if |α| > 0,
Cα (x) := (4.8.18)
ln |x − ·| BMO(Σ,σ) if |α| = 0,

then based on (4.8.17), (4.6.9), (4.8.18), and (4.2.10) we may estimate

   
∂xα ln |x − ·| , f − (∂ α Fm )(x)
       
= ∂xα ln |x − ·| , f − ∂xα ln |x − ·| , fm
   
= ∂xα ln |x − ·| , f − fm

≤ Cα (x) f − fm  H 1 (Σ,σ), (4.8.19)

for each α ∈ N0n , m ∈ N, and x ∈ Rn \ Σ. For each xo ∈ Rn \ Σ and α ∈ N0n fixed,

[133, Lemma 7.4.13] and (4.8.18) imply that

sup Cα (x) < +∞ if 0 < r < dist(xo, Σ). (4.8.20)

x ∈B(x o ,r)

Together with (4.8.14) and (4.8.19), this proves that

 
for each multi-index α ∈ N0n the sequence (∂ α Fm )(x) m∈N converges
    (4.8.21)
to ∂xα ln |x − ·| , f uniformly for x in compact subsets of Rn \ Σ.
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 197

Based this and the differentiation theorem for sequences of functions we then readily
conclude that the restriction of F to Rn \ Σ is of class C ∞ , and its partial derivatives
may be computed by differentiating directly under the integral pairing.
Finally, the case when Σ is bounded is treated very similarly. 

Going back to the result described in Lemma 4.8.1, it turns out that truncating (in
the range) any given BMO function, in the most basic fashion, yields a sequence of
bounded functions which is weak-∗ convergent in BMO to the original function.

Lemma 4.8.3 Let Σ ⊆ Rn be a closed, unbounded set, which is Ahlfors regular, and
abbreviate σ := H n−1 Σ. Given a real-valued function f ∈ BMO(Σ, σ), for each
M ∈ N define
⎧
⎪ M if f (x) > M
⎪
⎨
⎪
fM (x) := f (x) if − M ≤ f (x) ≤ M, ∀x ∈ Σ. (4.8.22)
⎪
⎪
⎪ −M if f (x) < −M,
⎩
Then
.
lim [ fM ] = [ f ] weak-∗ in BMO(Σ, σ). (4.8.23)
M→∞

That is, with ·, · denoting the duality bracket between the John-Nirenberg space
of functions of bounded mean oscillations on Σ, modulo constants, and the Hardy
space H 1 on Σ (cf. Theorem 4.6.1),
   
for each g ∈ H 1 (Σ, σ) one has [ fM ], g → [ f ], g as M → ∞. (4.8.24)

In addition, a similar weak-∗ convergence result holds in the case when Σ is

bounded, namely
lim fM = f weak-∗ in BMO(Σ, σ). (4.8.25)
M→∞

Proof In view of [133, (7.4.104), (7.4.102)], this is a consequence of Lemma 4.8.1.

We aim to replicate the results in Lemma 4.8.1 and Lemma 4.8.3 working with
Hölder functions. In this regard, the analogue of Lemma 4.8.1 reads as follows.

Lemma 4.8.4 Let Σ ⊆ Rn be a closed unbounded Ahlfors regular set. Abbreviate

 
σ := H n−1 Σ, pick α ∈ (0, 1), and introduce p := (n − 1)/(n − 1 + α) ∈ n−1
n ,1 .
.α
Consider a sequence { f j } j ∈N ⊆ C (Σ) of functions satisfying

sup  f j C.α (Σ) < +∞ and f (x) := lim f j (x) exists for each x ∈ Σ. (4.8.26)
j ∈N j→∞

.
Then the function f belongs to C α (Σ) and
. 
lim [ f j ] = [ f ] weak-∗ in C α (Σ) ∼ . (4.8.27)
j→∞
198 4 Hardy Spaces on Ahlfors Regular Sets

Specifically, with ·, · denoting the duality bracket between functions satisfying a

homogeneous Hölder condition of order α on Σ, modulo constants, and the Hardy
space H p on Σ (cf. Theorem 4.6.1),
   
for each g ∈ H p (Σ, σ) one has [ f j ], g → [ f ], g as j → ∞. (4.8.28)

Moreover, a similar weak-∗ convergence result holds in the case when Σ is

bounded, namely starting with a bounded sequence { f j } j ∈N in C α (Σ) which con-
verges pointwise to a function f and concluding that

lim f j = f weak-∗ in C α (Σ). (4.8.29)

j→∞

Proof If M ∈ [0, ∞) denotes the supremum in (4.8.26), it follows that for each j ∈ N
we have | f j (x) − f j (y)| ≤ M |x − y| α for every x, y ∈ Σ. Passing to limit j → ∞ then
.
yields | f (x) − f (y)| ≤ M |x − y| α for each x, y ∈ Σ, which shows that f ∈ C α (Σ).
Next, seeking a contradiction, assume the claim in (4.8.28) fails. Then there exists
a distribution g ∈ H p (Σ, σ) together with a strictly increasing sequence { jk }k ∈N of
natural numbers such that
   
any sub-sequence of [ f jk ], g does not converge to [ f ], g . (4.8.30)
k ∈N

To proceed, observe that the first condition in (4.8.26) implies that the sequence
  .  .   ∗
[ f jk ] k ∈N is bounded in C α (Σ) ∼. Given that C α (Σ) ∼ = H p (Σ, σ) (cf.
(4.6.1) in Theorem 4.6.1), the separability result from Proposition 4.4.2 together
with the Sequential Banach-Alaoglu Theorem from [133, (3.6.22)] (which covers
the class of separable quasi-Banach spaces) imply the existence of a sub-sequence,
  .
say [ f jk i ] i ∈N , along with some function f ∈ C α (Σ), such that
. 
[ f jk i ] −→ [ f ] weak-∗ in C α (Σ) ∼, as i → ∞. (4.8.31)

Having fixed an arbitrary point xo ∈ Σ, for each (p, ∞)-atom a on Σ we then write
∫ ∫
 
f a dσ = f (x) − f (xo ) a(x) dσ(x)
Σ Σ
∫
 
= lim f jk i (x) − f jk i (xo ) a(x) dσ(x)
i→∞ Σ
∫
   
= lim [ f jk i ], a = [ f ], a = f a dσ. (4.8.32)
i→∞ Σ

Above, the first equality is a consequence of the cancellation property of the atom, the
second equality is implied by Lebesgue’s Dominated Convergence Theorem (with
the uniform domination provided by M |x − xo | α · |a(x)| ∈ Lcomp∞ (Σ, σ)), the third

and final equalities comes from (4.6.4), while the fourth equality is guaranteed by
(4.8.31). With (4.8.32) in hand, we may rely on Lemma 4.6.9 to conclude that
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 199

f − f is constant on Σ. (4.8.33)

In turn, from (4.8.31) and (4.8.33) we conclude that [ f jk i ] → [ f ] weak-∗ in

. 
C α (Σ) ∼ as i → ∞. Consequently,
   
[ f jk i ], g −→ [ f ], g as i → ∞, (4.8.34)

which contradicts (4.8.30). This finishes the proof of (4.8.27). Finally, the weak-
convergence result recorded in (4.8.29) in the case when Σ is bounded is proved
similarly. 

Here is a companion result for Lemma 4.8.3, valid for Hölder functions.

Lemma 4.8.5 Let Σ ⊆ Rn be a closed, unbounded set,  which is Ahlfors reg-

ular, and abbreviate σ := H n−1 Σ. Also, pick p ∈ n−1 n , 1 and introduce
1  .α
α := (n − 1) p − 1 ∈ (0, 1). Given a real-valued function f ∈ C (Σ), define
   
fM := min max{ f , −M }, M = max min{ f , M }, −M for each M ∈ N. Then
. 
lim [ fM ] = [ f ] weak-∗ in C α (Σ) ∼ . (4.8.35)
M→∞

Specifically, with ·, · denoting the duality bracket between functions satisfying a

homogeneous Hölder condition of order α on Σ, modulo constants, and the Hardy
space H p on Σ (cf. Theorem 4.6.1),
   
for each g ∈ H p (Σ, σ) one has [ fM ], g → [ f ], g as M → ∞. (4.8.36)

Moreover, a similar weak-∗ convergence result holds in the case when Σ is

bounded, namely
lim fM = f weak-∗ in C α (Σ). (4.8.37)
M→∞

Proof This is a consequence of Lemma 4.8.4, bearing in mind [133, (7.3.13)-

(7.3.14)]. 

The next proposition further refines the compatibility result from Lemma 4.6.5.

Proposition 4.8.6 Let Σ ⊆ Rn be a closed, unbounded set, which is Ahlfors regular,

and abbreviate σ := H n−1 Σ. In this setting, suppose
∫
f ∈ BMO(Σ, σ) and g ∈ H 1 (Σ, σ) are such that | f ||g| dσ < +∞. (4.8.38)
Σ

Then, with ·, · denoting the duality bracket between the John-Nirenberg space
of functions of bounded mean oscillations on Σ, modulo constants, and the Hardy
space H 1 on Σ (cf. Theorem 4.6.1), one has
∫
 
[ f ], g = f g dσ. (4.8.39)
Σ
200 4 Hardy Spaces on Ahlfors Regular Sets

In particular,
∫
 
[ f ], g = f g dσ for each f ∈ L ∞ (Σ, σ) and g ∈ H 1 (Σ, σ). (4.8.40)
Σ

Moreover, similar formulas are valid in the case when Σ is bounded, this time
without having to consider equivalence classes of functions (modulo constants).

Proof If for each integer M ∈ N we define the function fM as in (4.8.22), then from
[133, (7.4.102)], Lemma 4.8.3, and Lebesgue’s Dominated Convergence Theorem
we have
fM ∈ L ∞ (Σ, σ) ⊂ BMO(Σ, σ),
.
lim [ fM ] = [ f ] weak-∗ in BMO(Σ, σ), (4.8.41)
M→∞
lim fM g = f g in L 1 (Σ, σ).
M→∞

Having fixed some q ∈ (1, ∞), Theorem 4.4.1 guarantees the existence of a numerical
sequence {λ j } j ∈N ∈  1 (N) along with a sequence {a j } j ∈N of (p, q)-atoms on Σ with
the property that


N
g N := λ j a j converges to g in H 1 (Σ, σ) as N → ∞. (4.8.42)
j=1

We may then write

     
[ f ], g = lim [ fM ], g = lim lim [ fM ], g N
M→∞ M→∞ N →∞
∫ ∫
= lim lim fM g N dσ = lim fM g dσ
M→∞ N →∞ Σ M→∞ Σ
∫
= f g dσ. (4.8.43)
Σ

The first equality in (4.8.43) is a consequence of the second property in (4.8.41),

while the second equality in (4.8.43) is implied by (4.8.42), and the third equality in
(4.8.43) follows from (4.6.8). The fourth equality in (4.8.43) may be justified based
on the first property in (4.8.41) and the fact that, since H 1 (Σ, σ) embeds continuously
into L 1 (Σ, σ), from (4.8.42) we also have that lim g N = g in L 1 (Σ, σ). The last
N →∞
equality in (4.8.43) is simply the last formula in (4.8.41). This establishes (4.8.39).
Finally, the version of the formula (4.8.39) in the case when Σ is bounded is proved
in a similar fashion. 
In the next proposition we refine the compatibility result from Lemma 4.6.6.

Proposition 4.8.7 Let Σ ⊆ Rn be a closed, unbounded set, which

 is Ahlfors regular,
and abbreviate σ := H n−1 Σ. Also, pick an arbitrary p ∈ n−1 ,
n 1 and introduce
 
α := (n − 1) p1 − 1 ∈ (0, 1). In this setting, suppose
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 201
∫
.
f ∈ C α (Σ) and g ∈ H p (Σ, σ) ∩ H 1 (Σ, σ) are such that | f ||g| dσ < +∞.
Σ
(4.8.44)
Then, with ·, · denoting the duality bracket between functions satisfying a homo-
geneous Hölder condition of order α on Σ, modulo constants, and the Hardy space
H p on Σ (cf. Theorem 4.6.1), one has
∫
 
[ f ], g = f g dσ. (4.8.45)
Σ

In addition, a similar formula is valid in the case when Σ is bounded, this

time working with inhomogeneous Hölder spaces and without having to consider
equivalence classes of functions (modulo constants). Moreover, in the case when Σ is
bounded there exists a finite constant C > 0, which depends only on the environment,
with the property that
 
| f , g | ≤ C f C.α (Σ) · g H p (Σ,σ) for each f ∈ C α (Σ)
∫ (4.8.46)
and g ∈ H (Σ, σ) ∩ H (Σ, σ) such that
p 1
g dσ = 0.
Σ

The reader is alerted to the fact that the inequality in (4.8.46) is finer than
the generic manner in which the paring between a given function f ∈ C α (Σ)
with an arbitrary g ∈ H p (Σ, σ) may be estimated according to (the last part of)
Theorem 4.6.1 when Σ is bounded, since the estimate in (4.8.46) only involves the
.
semi-norm associated with the homogeneous Hölder space C α (Σ) (as opposed to
the larger norm in C α (Σ)).
Proof of Proposition 4.8.7 The argument is similar to that in the
 proof of Proposi-

tion 4.8.6. To get started, for each M ∈ N consider fM := min max{ f , −M }, M .
Then from [133, (7.3.13)], Lemma 4.8.5, and Lebesgue’s Dominated Convergence
Theorem we have
.
fM ∈ L ∞ (Σ, σ) ∩ C α (Σ) = C α (Σ),
. 
lim [ fM ] = [ f ] weak-∗ in C α (Σ) ∼, (4.8.47)
M→∞
lim fM g = f g in L 1 (Σ, σ).
M→∞

Next, Theorem 4.4.3 ensures the existence of a numerical sequence {λ j } j ∈N ∈  1 (N)

along with a sequence {a j } j ∈N of (p, 2)-atoms on Σ with the property that


N
g N := λ j a j converges to g as N → ∞ both in H p (Σ, σ) and in H 1 (Σ, σ).
j=1
(4.8.48)
In particular, since H 1 (Σ, σ) embeds continuously into L 1 (Σ, σ), we also have that

g N converges to g as N → ∞ in L 1 (Σ, σ). (4.8.49)

202 4 Hardy Spaces on Ahlfors Regular Sets

We may now write

     
[ f ], g = lim [ fM ], g = lim lim [ fM ], g N
M→∞ M→∞ N →∞
∫ ∫ ∫
= lim lim fM g N dσ = lim fM g dσ = f g dσ. (4.8.50)
M→∞ N →∞ Σ M→∞ Σ Σ

The first equality in (4.8.50) is a consequence of the second property in (4.8.47),

while the second equality in (4.8.50) is implied by (4.8.48), and the third equality in
(4.8.50) follows from (4.6.8). The fourth equality in (4.8.50) follows from (4.8.49)
and the first property in (4.8.47). The last equality in (4.8.50) is simply the last
formula in (4.8.47). This establishes (4.8.45).
∫ when Σ is bounded. If g ∈ H (Σ, σ) ∩ H (Σ, σ) and
Consider next the case p 1
α
f ∈ C (Σ) are such that Σ | f ||g| dσ < ∞, then the fact that
∫
 
f, g = f g dσ (4.8.51)
Σ

may be justified in a very similar fashion to the proof of formula (4.8.45). As regards
the claim∫ in (4.8.46), suppose now that f ∈ C α (Σ) and g ∈ H p (Σ, σ) ∩ H 1 (Σ, σ)
satisfies Σ g dσ = 0. As before, invoke Theorem 4.4.3 to produce a numerical
sequence {λ j } j ∈N ∈  1 (N) satisfying


∞ 1/p
|λ j | p ≤ Cg H p (Σ,σ) (4.8.52)
j=1

for some finite constant C > 0 which depends only on the ambient, along with a
∫sequence {a j }∫j ∈N of (p, 2)-atoms on Σ such that (4.8.48)-(4.8.49) hold. In particular,
g dσ → Σ g dσ = 0 as N → ∞. Even though the constant function a ≡ σ(Σ)−1
Σ N
is now considered an atom, this convergence ensures that none of the atoms a j is this
special atom. Consequently, each a j has vanishing moment which, in turn, permits
us to estimate ∫
f a j dσ ≤ C f C.α (Σ) for each j ∈ N, (4.8.53)
Σ
where the constant C ∈ (0, ∞) depends only on the ambient. At this stage, we may
write
∞ ∫ 
∞ ∫
 
f, g = λj f a j dσ ≤ |λ j | f a j dσ
j=1 Σ j=1 Σ


∞ 
∞ 1/p
≤C |λ j |  f C.α (Σ) ≤ C |λ j | p  f C.α (Σ)
j=1 j=1

≤ C f C.α (Σ) g H p (Σ,σ), (4.8.54)

4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 203

using the last part of Theorem 4.6.1, (4.8.53), the fact that p < 1, and (4.8.52). This
establishes (4.8.46), and finishes the proof of Proposition 4.8.7. 
A useful version of Proposition 4.8.7, in which the membership of g to H 1 (Σ, σ)
1 (Σ, σ) (at the
(cf. (4.8.44)) is replaced by the less restrictive membership to Lloc
expense of strengthening the assumptions on f ) is discussed next.
Proposition 4.8.8 Let Σ ⊆ Rn be a closed, unbounded set, which
 is Ahlfors regular,
and abbreviate σ := H n−1 Σ. Also, pick an arbitrary p ∈ n−1 ,
n 1 and introduce
1 
α := (n − 1) p − 1 ∈ (0, 1). In this setting, suppose
.
f ∈ C α (Σ) ∩ Lip(Σ) and g ∈ H p (Σ, σ) ∩ Lloc
1 (Σ, σ)

∫ (4.8.55)
are such that | f ||g| dσ < +∞.
Σ

Then, with ·, · denoting the duality bracket between functions satisfying a homo-
geneous Hölder condition of order α on Σ, modulo constants, and the Hardy space
H p on Σ (cf. Theorem 4.6.1), one has
∫
 
[ f ], g = f g dσ. (4.8.56)
Σ

In addition, a similar formula is valid when Σ is bounded (in which case, the
last property in (4.8.55) becomes redundant), this time working with inhomogeneous
Hölder spaces and without having to consider equivalence classes of functions
(modulo constants).
In relation to the hypotheses made on f in (4.8.55) it is relevant to observe that
(compare with [133, (7.3.15)])
"
if f ∈ Lip(Σ) and sup | f | < +∞ then f ∈ C α (Σ). (4.8.57)
Σ 0<α ≤1

Proof of Proposition 4.8.8 To fix ideas, assume that Σ is unbounded, and first con-
sider the special case when f is as in (4.8.57), i.e., f ∈ Lip(Σ) and supΣ | f | < +∞.
In this scenario, pick a function φ ∈ Cc∞ (Rn ) with φ ≡ 0 near the origin in Rn , then
define φ j (x) := φ(x/ j) for each j ∈ N and x ∈ Rn . Clearly, φ j f ∈ Lipc (Σ) for each
j ∈ N, and Lemma 4.8.4 applies to give that
.   ∗
lim [φ j f ] = [ f ] weak-∗ in C α (Σ) ∼ = H p (Σ, σ) . (4.8.58)
j→∞

For each g ∈ H p (Σ, σ) we may then write

     
[ f ], g = lim [φ j f ], g = lim (Lip c (Σ)) g, φ j f Lip c (Σ)
j→∞ j→∞
∫ ∫
= lim φ j f g dσ = f g dσ. (4.8.59)
j→∞ Σ Σ
204 4 Hardy Spaces on Ahlfors Regular Sets

Above, the first equality is a consequence of (4.8.58), the second equality follows
from part (a) in Lemma 4.6.4, the third equality is implied by [133, Proposition 4.1.4],
and the last equality may be justified using Lebesgue’s Dominated Convergence
Theorem (bearing in mind the last property in (4.8.55)). This establishes (4.8.56) in
the case when Σ is unbounded and f is as in (4.8.57).
Consider next the case when Σ  is unbounded, without
 making
 any additional

assumptions on f . Set fM := min max{ f , −M }, M = max min{ f , M }, −M for
each M ∈ N. Then, as seen from [133, (7.3.13)-(7.3.14)], we have fM ∈ Lip(Σ) and
supΣ | fM | < +∞ for every M ∈ N. Consequently, for each g ∈ H p (Σ, σ) we may
write
∫ ∫
   
[ f ], g = lim [ fM ], g = lim fM g dσ = f g dσ, (4.8.60)
M→∞ M→∞ Σ Σ

thanks to Lemma 4.8.5, the result established in the first part of the proof (for fM
in place of f ), and Lebesgue’s Dominated Convergence Theorem (keeping in mind
[133, (7.3.13)] and the last property in (4.8.55)). This proves (4.8.56) in the case
when Σ is unbounded. Finally, the case when Σ is bounded is dealt with similarly. 

It is also desirable to have conditions guaranteeing that the duality pairing between
Lorentz-based Hardy spaces and their duals agrees with the ordinary integral pairing.

 n−1 Σ ⊆ R is a closed Ahlfors regular set and abbreviate

Proposition 4.8.9 Suppose n

σ := H Σ. Fix p ∈ n , ∞ and q ∈ (0, ∞], then pick some α0 ∈ (0, 1) and some
n−1

p1 ∈ (1, ∞) such that

 α0  −1  
1 −1
p0 := 1 + n−1 < p and p1 := 1 − p1 > p. (4.8.61)

Then
∫
 
(H p, q (Σ,σ))∗ f, g H p, q (Σ,σ) = f g dσ (4.8.62)
Σ

provided (with the inclusion considered in the sense of (4.7.2))

.  ∗
f ∈ Lip(Σ) ∩ C α0 (Σ) ∩ L p1 (Σ, σ) → H p,q (Σ, σ)
∫ (4.8.63)
and g ∈ H (Σ, σ) ∩ Lloc (Σ, σ) with
p,q 1
| f ||g| dσ < +∞.
Σ

Proof Let f , g be as in (4.8.63) and decompose g = g0 +g1 , with g0 ∈ H p0 (Σ, σ) and

g1 ∈ L p1 (Σ, σ) (recall that H p,q (Σ, σ) embeds into H p0 (Σ, σ) + L p1 (Σ, σ)). Given
that by assumption f ∈ L p1 (Σ, σ) and g ∈ Lloc
1 (Σ, σ), it follows that actually

∫
g0 ∈ H (Σ, σ) ∩ Lloc (Σ, σ) and
p0 1
| f ||g0 | dσ < +∞ (4.8.64)
Σ

since
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 205
∫ ∫
 
| f ||g0 | dσ ≤ | f | |g| + |g1 | dσ
Σ Σ
∫
≤ | f ||g| dσ +  f  L p1 (Σ,σ) g1  L p1 (Σ,σ) < +∞. (4.8.65)
Σ

Based on (4.8.64), (4.7.3), and Proposition 4.8.8 applied   with α, g, p replaced,

here
respectively, by α0 , g0 , p0 (note that we have p0 ∈ n−1 n , 1 by design; cf. (4.8.61))
we may then write
∫ ∫ ∫
 
(H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) = f g1 dσ + f g0 dσ = f g dσ, (4.8.66)
Σ Σ Σ

as wanted. 
Let us formally record the fact that the duality pairing between any BMO function
and any H 1 -molecule agrees with the integral pairing of said functions.
Corollary 4.8.10 Let Σ ⊆ Rn be a closed, unbounded set, which is Ahlfors regular,
and abbreviate σ := H n−1 Σ. Then for any function f ∈ BMO(Σ, σ) and any
H 1 -molecule m on Σ (cf. Definition 4.5.1) one has
∫
 
[ f ], m = f m dσ, (4.8.67)
Σ

where ·, · denotes the duality bracket between the John-Nirenberg space of functions
of bounded mean oscillations on Σ, modulo constants, and the Hardy space H 1 on
Σ (cf. Theorem 4.6.1). As a consequence,
  ! ∫
one has [ f ], g = lim N j=1 λ j Σ f m j dσ whenever
N →∞
f ∈ BMO(Σ, σ), g ∈ H 1 (Σ, σ), {λ j } j ∈N ∈  1 (N), and (4.8.68)
{m j }!j ∈N is a sequence of H 1 -molecules on Σ, such that
g= ∞ j=1 λ j m j in H (Σ, σ).
1

Moreover, similar results are valid in the case when Σ is bounded, this time
without having to consider equivalence classes of functions (modulo constants).
Proof This is an immediate consequence of Proposition 4.8.6 and Lemma 4.5.4. 
There is also a version of Corollary 4.8.10 for Hölder functions and H p Hardy
spaces with p < 1, of the sort described below.
Corollary 4.8.11 Let Σ ⊆ Rn be a closed, unbounded set, which  is Ahlfors regular,
and abbreviate σ := H n−1 Σ. Also, pick an arbitrary p ∈ n−1 n , 1 and introduce
 
α := (n − 1) p1 − 1 ∈ (0, 1). Finally, choose q ∈ (1, ∞) along with ε > α/(n − 1).
.
Then for any function f ∈ C α (Σ) and any (p, q, ε)-molecule m on Σ (cf. Defini-
tion 4.5.1) one has ∫
 
[ f ], m = f m dσ, (4.8.69)
Σ
206 4 Hardy Spaces on Ahlfors Regular Sets

where ·, · stands for the duality bracket between functions satisfying a homogeneous
Hölder condition of order α on Σ, modulo constants, and the Hardy space H p on Σ
(cf. Theorem 4.6.1). As a consequence,
  ! ∫
one has [ f ], g = lim N j=1 λ j Σ f m j dσ whenev-
. N →∞
er f ∈ C α (Σ), g ∈ H p (Σ, σ), {λ j } j ∈N ∈  1 (N), and (4.8.70)
{m j } j ∈N is a sequence of (p, q, ε)-molecules on Σ, with
!
the property that g = ∞ j=1 λ j m j in H (Σ, σ).
p

In addition, similar results are valid in the case when Σ is bounded, this time
working with the inhomogeneous Hölder space and without having to consider
equivalence classes of functions (modulo constants).

Proof All claims are readily justified with the help of Proposition 4.8.7 and Lem-
ma 4.5.5, bearing in mind that any (p, q, ε)-molecule on Σ is a scalar multiple of a
(1, q, ε)-molecule on Σ and, hence, belongs to the Hardy space H 1 (Σ, σ). 

Lemma 4.8.12, discussed below, is a weak-∗ density result which is relevant when
studying Hardy-based Sobolev spaces, introduced later.

Lemma 4.8.12 Suppose Σ is a closed set in Rn which is Ahlfors regular and let
σ := H n−1 Σ. Fix α ∈ (0, 1) and define p = n−1+α n−1
∈ n−1n , 1 . Then for each
.α ∞
g ∈ C (Σ) there exists a sequence {φ j } j ∈N ⊆ Cc (R ) with the following property.
n
  . 
If Σ is unbounded, then φ j Σ , viewed as an element in C α (Σ) ∼, converges to [g]
.α 
weak-∗ in C (Σ) ∼ as j → ∞, i.e.,
  
φ j Σ , h −→ [g], h as j → ∞ for every h ∈ H p (Σ, σ), (4.8.71)

where the pairings are understood in the sense of Theorem 4.6.1. Also, if Σ is bounded
then, with a similar interpretation,
 
φ j Σ, h −→ g, h as j → ∞ for every h ∈ H p (Σ, σ). (4.8.72)
.
Proof Fix some G ∈ C α (Rn ) such that G Σ = g (such an extension always exists;
see the discussion in [138]). Also consider a cut-off function ξ ∈ Cc∞ (Rn ) with
supp ξ ⊆ B(0, 2), 0 ≤ ξ ≤ 1, ξ ≡ 1 on B(0, 1). Then, having picked some x0 ∈ Σ, for
each ε ∈ (0, ∞) define ξε (x) := ξ ε(x − x0 ) for x ∈ Rn . Fix ε ∈ (0, 1). We claim that
there exists some constant C ∈ (0, ∞), independent of ε (actually depending only on
inf Σ |g| and GC.α (Rn ) ) such that

(ξε G)(x) − (ξε G)(x) ≤ C|x − y| α for every x, y ∈ Rn . (4.8.73)

We split the proof of (4.8.73) into four cases.

Case I: Assume x ∈ B(x0, 2/ε) and y ∈ B(x0, 4/ε). Then
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 207

(ξε G)(x) − (ξε G)(x) ≤ ξε (x) G(x) − G(y) + |G(y) − G(x0 )| ξε (x) − ξε (y)

+ |g(x0 )| ξε (x) − ξε (y) . (4.8.74)

Observe that, the properties of ξ, the definition of ξε , and the current assumptions
on x and y, imply

ξε (x) − ξε (y) ≤ |x − y| α |x − y| 1−α sup ∇ξε ≤ C|x − y| α ε α−1 · Cε

Rn

= C|x − y| α ε α . (4.8.75)
.
Making use of (4.8.75), the fact that G ∈ C α (Σ), that 0 ≤ ξε ≤ 1, that |y − x0 | < 4/ε,
and that |ε| ≤ 1, we may now return to (4.8.74) and further estimate

(ξε G)(x) − (ξε G)(x)

≤ C|x − y| α + C|y − x0 | α |x − y| α ε α + |g(x0 )| · C|x − y| α ε α
≤ C|x − y| α, (4.8.76)

for some constant C ∈ (0, ∞) independent of ε, x, and y. This proves (4.8.73) in the
current case.
Case II: Assume x ∈ B(x0, 2/ε) and y ∈ Rn \ B(x0, 4/ε). Then |x − y| ≥ 2/ε and
ξε (y) = 0. As such,

(ξε G)(x) − (ξε G)(x) = (ξε G)(x) ≤ |G(x)| ≤ |G(x) − G(x0 )| + |g(x0 )|
1 C
≤ C|x − x0 | α + |g(x0 )|ε α · ≤ α ≤ C|x − y| α, (4.8.77)
εα ε
for some constant C ∈ (0, ∞) independent of ε, x, and y. Hence, (4.8.73) holds in
the current case as well.
Case III: Assume y ∈ B(x0, 2/ε) and x ∈ Rn arbitrary. Since the estimate in
(4.8.73) is symmetric in x and y, the desired conclusion follows from what we
proved in Case I and Case II.
Case IV: Assume x, y ∈ Rn \ B(x0, 2/ε). In this scenario ξε (x) = 0 = ξε (y), so
(4.8.73) is trivial.
Having completed the proof of (4.8.73), bring in ∫a mollifier of the following
sort. Pick a non-negative function θ ∈ Cc∞ (Rn ) with Rn θ dL n = 1 and, for each
ε ∈ (0, 1), set θ ε (x) := ε −n θ(x/ε) for every x ∈ Rn . Finally, define
 
φε := ξε G ∗ θ ε for each ε ∈ (0, 1). (4.8.78)

Then each φε belongs to Cc∞ (Rn ). Moreover, for every x, y ∈ Rn we may estimate,
thanks to (4.8.73) and the qualities of θ,
208 4 Hardy Spaces on Ahlfors Regular Sets
∫
φε (x) − φε (y) ≤ (ξε G)(x − z) − (ξε G)(y − z) θ ε (z) dz
Rn

≤ C|x − y| α, for each ε ∈ (0, 1), (4.8.79)

where the constant C ∈ (0, ∞) is as in (4.8.73) (hence, independent of ε). Thus,

.
φε ∈ C α (Rn ) and (4.8.79) shows that, for some constant C ∈ (0, ∞) independent of
ε, we have
   
φε  .α ≤ φε  .α n ≤ C, for each ε ∈ (0, 1). (4.8.80)
Σ C (Σ) C (R )

Also, given any compact set K ⊂ Σ, observe that ξε becomes identically 1 near K
when ε ∈ (0, 1) is sufficiently small. Together with (4.8.73), this also permits us to
conclude that there exists some small εK ∈ (0, 1) such that
∫
φε (x) − g(x) ≤ (ξε G)(x − z) − (ξε G)(x) θ ε (z) dz
Rn

≤ Cε α whenever ε ∈ (0, εK ). (4.8.81)

As a result,
φε Σ converges uniformly on arbitrary bounded
(4.8.82)
subsets of Σ to the function g, as ε −→ 0+ .
Suppose next that Σ is unbounded. Then (4.8.80) implies that the sequence
  .   ∗
φε Σ ε ∈(0,1) is bounded in C α (Σ) ∼= H p (Σ, σ) . Granted this and (4.8.82),
Lemma 4.8.4 applies and yields (4.8.71).
Finally, when Σ is bounded, the end-game in the proof is similar. This time, in
addition to (4.8.80), we also have

sup φε Σ
≤ sup |G| < +∞ for each ε ∈ (0, 1), (4.8.83)
Σ U

where
  U := {x + t y : x ∈ Σ, y ∈ supp θ, ∗|t| ≤ 1} is a compact set. Thus,
φε Σ ε ∈(0,1) is bounded in C α (Σ) = H p (Σ, σ) . From this point on, the same type
of argument as before, based on Lemma 4.8.4, then justifies (4.8.72). 

It is also useful to have a weak-∗ convergence result involving duals of Lorentz-

based Hardy spaces, of the sort described in our next lemma.

Lemma 4.8.13 Let Σ ⊆ Rn be a closed set which is Ahlfors regular, and  abbreviate

σ := H n−1 Σ. Pick α0 ∈ (0, 1), introduce p0 := (n − 1)/(n − 1 + α0 ) ∈ n−1 n , 1 , and
select p1 ∈ (1, ∞). Also, fix some p ∈ (p0, p1 ) along. with some q ∈ (0, ∞). In this
setting, consider a bounded sequence { f j } j ∈N ⊆ C α0 (Σ) ∩ L p1 (Σ, σ), i.e., satisfying

sup  f j C.α0 (Σ) < +∞ and sup  f j  L p1 (Σ,σ) < +∞, (4.8.84)
j ∈N j ∈N

and make the assumption that

4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 209

f (x) := lim f j (x) exists for each x ∈ Σ. (4.8.85)

j→∞

.
 Then the function f belongs to C α0 (Σ) ∩ L p1 (Σ, σ) and, when viewed in
∗
H (Σ, σ) (cf. Lemma 4.7.1),
p,q
∗ is the weak-∗ limit of { f j } j ∈N , itself regarded
as a sequence in H p,q (Σ, σ) (again, in the sense of Lemma 4.7.1). Specifically,

for each g ∈ H p,q (Σ, σ) one has

    (4.8.86)
(H p, q (Σ,σ))∗ f j , g H p, q (Σ,σ) → (H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) as j → ∞.

Proof Let M ∈ [0, ∞) denote the first supremum in (4.8.84). Consequently, for each
j ∈ N we have | f j (x) − f j (y)| ≤ M |x − y| α0 for all x, y ∈ Σ. Passing to limit j → ∞
.
then yields | f (x)− f (y)| ≤ M |x − y| α0 for all x, y ∈ Σ, which proves that f ∈ C α0 (Σ).
Next, we make the claim that
f belongs to L p1 (Σ, σ) and is the weak-∗ limit
(4.8.87)
of the sequence { f j } j ∈N in L p1 (Σ, σ).

To justify this, fix g ∈ L p1 (Σ, σ) where p1 ∈ (1, ∞) is such that 1/p1 + 1/p1 = 1 and
consider an arbitrary sub-sequence { f jk }k ∈N
 of { f j } j ∈N . Since from (4.8.84) we know
∗
that { f jk }k ∈N is bounded in L p1 (Σ, σ) = L p1 (Σ, σ) , and since [133, (3.6.27)] en-
sures that L p1 (Σ, σ) is a separable Banach space, the Sequential Banach-Alaoglu
Theorem (recalled in [133, (3.6.22)]) permits us to extract a sub-sub-sequence
{ f jk i }i ∈N which converges weak-∗ in L p1 (Σ, σ) to some function h ∈ L p1 (Σ, σ).
As such, for each φ ∈ Cc∞ (Rn ) we may write
∫ ∫ ∫
h φ dσ = lim f jk i φ dσ = f φ dσ, (4.8.88)
Σ i→∞ Σ Σ

where the last equality is a consequence of Lebesgue’s Dominated Convergence The-

orem, whose applicability is presently guaranteed by (4.8.85) and the first property
in (4.8.84), as they together imply that

| f j (x)| ≤ M · sup |y − x0 | α0 + sup | fk (x0 )| ≤ C < +∞, (4.8.89)

y ∈suppφ∩Σ k ∈N

for each x ∈ suppφ ∩ Σ and each j ∈ N, where x0 ∈ Σ is some fixed reference point.
In turn, from (4.8.88) and [133, (3.7.23)] we conclude that h = f on Σ. This has two
consequences of interest. First, it implies that f ∈ L p1 (Σ, σ). Second, it allows us to
write (for the function g selected earlier)
∫ ∫
lim f jk i g dσ = f φ dσ. (4.8.90)
i→∞ Σ Σ

Let us summarize our progress

 ∫ at this stage.
 The
 ∫argumentpresented so far shows
that for any sub-sequence Σ f jk g dσ of Σ f j g dσ there exists a sub-
k ∈N j ∈N
210 4 Hardy Spaces on Ahlfors Regular Sets
∫ 
sub-sequence Σ f jk i g dσ with the property that (4.8.90) holds. Ultimately,
i ∈N
from this we conclude that actually
∫ ∫
lim f j g dσ = f g dσ, (4.8.91)
j→∞ Σ Σ

and since g ∈ L p1 (Σ, σ) has been chosen arbitrarily it follows that the sequence
{ f j } j ∈N is weak-∗ convergent in L p1 (Σ, σ) to the function f . The proof of the claims
made in (4.8.87) is therefore complete.
Going further, consider again an arbitrary sub-sequence { f jk }k ∈N of { f j } j ∈N and
pick some arbitrary g ∈ H p,q (Σ, σ) (which is going to come into play a little later; see
.
(4.8.94)). Since { f j } j ∈N is assumed to be bounded  in the space C α0 (Σ)  ∩ L (Σ, σ)
p1
.
(equipped with the natural quasi-norm max  · C α0 (Σ),  ·  L 1 (Σ,σ) ; cf. (1.3.3)),
p
 p,q ∗
and since the latter space is continuously embedded  p,q  ∗ H (Σ, σ) (cf. (4.7.2)),
into
we conclude that { f j } j ∈N is bounded in H (Σ, σ) . Given that we know from
Proposition 4.4.5 that the Lorentz-based Hardy space H p,q (Σ, σ) is a separable quasi-
Banach space, we may once more rely on the Sequential Banach-Alaoglu Theorem
(cf. [133, (3.6.22)]) to extract a sub-sub-sequence { f jk i }i ∈N which converges weak-∗
 ∗  ∗
in H p,q (Σ, σ) to some functional Λ ∈ H p,q (Σ, σ) . For any (p0, p1 )-atom a on
Σ we may then write
   
(H p, q (Σ,σ))∗ Λ, a H p, q (Σ,σ) = lim (H p, q (Σ,σ))∗ f jk i , a H p, q (Σ,σ)
i→∞
 
= lim (H p0 (Σ,σ))∗ f jk i , a H p0 (Σ,σ)
i→∞
∫
= lim f jk i a dσ
i→∞ Σ
∫
 
= f a dσ = (H p, q (Σ,σ))∗ f , a H p, q (Σ,σ) . (4.8.92)
Σ

The first equality above is implied by the fact that { f jk i }i ∈N converges weak-∗
 ∗  ∗
in H p,q (Σ, σ) to Λ ∈ H p,q (Σ, σ) , and the membership of a to the space
H p0 (Σ, σ) ∩ L p1 (Σ, σ) ⊆ H p,q (Σ, σ) (cf. (4.3.145)). The second equality in (4.8.92)
is a consequence of (4.7.3)-(4.7.4). The third equality in (4.8.92) is seen from (4.6.4)-
(4.6.5). The fourth equality in (4.8.92) may be justified using Lebesgue’s Dominated
Convergence Theorem, whose applicability is currently ensured by (4.8.85) and
(4.8.89). Finally, the fifth equality in (4.8.92) may be justified much as the second
equality in (4.8.92).
Having proved (4.8.92), bring on the family F used in the proof of Proposi-
tion 4.4.2, constructed as in (4.4.15) for an exponent r ∈ (p1, ∞). From (4.8.92)
and the fact that any function in F may be written as a finite linear combination of
(p0, p1 )-atoms on Σ we conclude that
4.8 Weak-∗ Convergence and More on the Compatibility of Pairings 211
   
(H p, q (Σ,σ))∗ Λ, h H p, q (Σ,σ) = (H p, q (Σ,σ))∗ f , h H p, q (Σ,σ) for each h ∈ F .
(4.8.93)

In concert with (4.4.122), this further gives (with g as above)

   
(H p, q (Σ,σ))∗ Λ, g H p, q (Σ,σ) = (H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) . (4.8.94)
 ∗
Thus, necessarily, Λ = f in Hp,q (Σ, σ) . In summary, the proof so far shows
 
that for any given sub-sequence (H p, q (Σ,σ))∗ f jk , g H p, q (Σ,σ) of the sequence
    k ∈N
(H p, q (Σ,σ))∗ f j , g H p, q (Σ,σ) , wre are able to find a sub-sub-sequence
 j∈N
 
(H p, q (Σ,σ))∗ f jk i , g H p, q (Σ,σ) with the property that
i ∈N
   
lim (H p, q (Σ,σ))∗ f jk i , g H p, q (Σ,σ) = (H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) . (4.8.95)
i→∞

From this we eventually conclude that, in fact,

   
lim (H p, q (Σ,σ))∗ f j , g H p, q (Σ,σ) = (H p, q (Σ,σ))∗ f , g H p, q (Σ,σ) (4.8.96)
j→∞

and since g ∈ H p,q (Σ, σ) has been  chosen arbitrarily

∗ we finally see that the sequence
{ f j } j ∈N is weak-∗ convergent in H p,q (Σ, σ) to the function f . 
It is also useful to have a result pertaining to the compatibility of the paring of
Lorentz-based Hardy spaces and their duals, on the one hand, and the distributional
pairing, on the other hand, of the sort described in the lemma below.

Lemma 4.8.14 Let Σ be a closed set in Rn which is Ahlfors  regular

 and abbreviate
σ := H n−1 Σ. Fix some integrability exponent p ∈ n−1 n , ∞ and some q ∈ (0, ∞).

Then for each given distribution
 f ∈
∗ H p,q (Σ, σ) → Lip (Σ) and each given
c
function g ∈ Lipc (Σ) → H p,q (Σ, σ) (with the inclusion considered in the sense
of Lemma 4.7.1; cf. (4.7.2)) one has
   
(Lip c (Σ)) f , g Lip c (Σ) = H p, q (Σ,σ) f , g (H p, q (Σ,σ))∗ . (4.8.97)
 
Proof Consider first the case when we have p ∈ n−1 n , 1 . In such a scenario,
r ∈ (1, ∞) then bring back the set F first used in the proof of Proposition 4.4.2 (cf.
(4.4.15)). Thanks to (4.4.122) and (4.2.29), it suffices to show that (4.8.97) holds if
f ∈ F . However, whenever this is the case we may write
∫
   
(Lip c (Σ)) f , g Lip c (Σ) = f g dσ = H p, q (Σ,σ) f , g (H p, q (Σ,σ))∗ (4.8.98)
Σ

thanks to [133, (4.1.47)]and (4.7.3)-(4.7.4).

Suppose now that p ∈ (1, ∞), and pick p0 ∈ (1, p) along with p1 ∈ (p, ∞). We
run the same argument as before, with one adjustment. Specifically, in place of the
family F this time we consider L p0 (Σ, σ) ∩ L p1 (Σ, σ) which, according to [133,
(6.2.51)] and (4.2.30), continues to be a dense subset of H p,q (Σ, σ). 
212 4 Hardy Spaces on Ahlfors Regular Sets

4.9 More on H p Versus L p : The Filtering Operator

On a given closed Ahlfors regular set in Rn , it turns out that the identity map
between the Hardy scale H p and the Lebesgue scale L p when p ∈ (1, ∞) (cf. [133,
(3.6.27)]) may
 be further extended uniquely to a linear and bounded mapping in the
range p ∈ n−1n , 1 . We shall refer to the identity map thus extended as an operator

from H into L with p ∈ n , ∞ as the L p -filtering operator. Details are
p p n−1

contained in the theorem below.

Theorem 4.9.1 Suppose Σ ⊆ Rn is a closed Ahlfors regular set. Let σ := H n−1

 Σ
and consider a family of kernels St : Σ × Σ → R indexed by t ∈ 0, diam Σ and
satisfying (4.2.14).
Then the limit6
 
(H f )(x) := lim+ (H p (Σ,σ))∗ St (x, ·), f H p (Σ,σ)
t→0 (4.9.1)
for each f ∈ H p (Σ, σ) with n−1 n < p < ∞

exists for σ-a.e. point x ∈ Σ and is unambiguously defined7. Moreover, the assign-
ment f → H f induces a well-defined linear and bounded operator
 
H : H p (Σ, σ) −→ L p (Σ, σ) for each p ∈ n−1
n ,∞ , (4.9.2)
 
and the operators H associated with various values of p in n−1 n , ∞ are compatible
with one another. Also,
 
H f = f whenever f ∈ H p (Σ, σ) ∩ Lloc
1 (Σ, σ) with p ∈ n−1 , ∞ , hence
n (4.9.3)
in particular for each function f ∈ H p (Σ, σ) with 1 ≤ p < ∞,
properties (4.9.2)-(4.9.3) determine uniquely the operator H, and
 
if p ∈ n−1 n , 1 then the operator H : H (Σ, σ) −→ L (Σ, σ) acts
p p

on any given f ∈ H (Σ, σ) according to H f = lim f j in L p (Σ, σ)

p
j→∞
p,q (4.9.4)
provided f = lim f j in H p (Σ, σ) for some { f j } j ∈N ⊆ Hfin (Σ, σ) with
j→∞
q ∈ (1, ∞).
Finally, the operator (4.9.2) further induces well-defined linear and bounded
mappings
 
H : H p,q (Σ, σ) −→ L p,q (Σ, σ) for p ∈ n−1
n , ∞ and q ∈ (0, ∞], (4.9.5)

(henceforth referred to as L p,q -filtering operators) according to

6 with the convention that St (x, ·) in (4.9.1)

 is replaced
 by [St (x, ·)], its class modulo constants, in
the case when Σ is unbounded and p ∈ n−1 n , 1
 
7 in the sense that it is not affected by the particular index p ∈ n−1 n , ∞ labeling the Hardy space
to which f happen to belong to
4.9 More on H p Versus L p : The Filtering Operator 213
 
(H f )(x) = lim+ Lipc (Σ) St (x, ·), f (Lip c (Σ)) at σ-a.e. x ∈ Σ,
t→0 (4.9.6)
for each f ∈ H p,q (Σ, σ) with n−1
n < p < ∞ and q ∈ (0, ∞]

which are compatible with one another as well as with the operator in (4.9.2), and

H f = f if f ∈ H p,q (Σ, σ) ∩ Lloc

1 (Σ, σ)
 n−1  (4.9.7)
with p ∈ n , ∞ and q ∈ (0, ∞].

We wish to note that while H in (4.9.5) becomes the identification of H p,q (Σ, σ)
with L p,q (Σ, σ) when 1 < p < ∞ and 0 < q ≤ ∞ (cf. (4.2.30)), this operator is
not injective in the range p ≤ 1. To substantiate this  claim, recall from (4.2.34) that
for each point y ∈ Σ the Dirac distribution δy ∈ Lipc (Σ) belongs to H 1,∞ (Σ, σ).
Then, on account of (4.9.6) and (4.2.14), we may write
 
(Hδy )(x) = lim+ Lipc (Σ) St (x, ·), δy (Lipc (Σ)) = lim+ St (x, y) = 0 (4.9.8)
t→0 t→0

first for σ-a.e. x ∈ Σ \ {y},

 hence  ultimately for σ-a.e. x ∈ Σ if n ≥ 2.
Likewise, when p ∈ n−1 n , 1 , the L p -filtering operator (4.9.1)-(4.9.2) fails to be

injective. Indeed, according to (4.2.17), for each two  distinct points x0, x1 ∈ Σ we
have δx0 − δx1 ∈ H p (Σ, σ) for every p ∈ n−1 n , 1 and (4.9.8) implies (assuming
n ≥ 2) that
H(δx0 − δx1 ) = 0 at σ-a.e. point on Σ. (4.9.9)
After this digression, we now turn to the proof of Theorem 4.9.1.
 n−1 
Proof of Theorem 4.9.1 Suppose f ∈ H p (Σ, σ) ∩ Lloc
1 (Σ, σ) with p ∈
n ,∞ .
Then for σ-a.e. x ∈ Σ we have
∫
lim+ St (x, y) f (y) dσ(y) − f (x)
t→0 Σ
∫
≤ lim sup St (x, y)| f (y) − f (x)| dσ(y)
t→0+ Σ
⨏
≤ C lim sup | f (y) − f (x)| dσ(y) = 0, (4.9.10)
t→0+ B(x,Ct)∩Σ

thanks to (4.2.14), the current assumptions on Σ, and [133, Proposition 7.4.4] (whose
applicability in the present setting is ensured by [133, (3.6.26)]). We may then rely
on (4.6.10), (4.2.14), part (a) in Lemma 4.6.4, and [133, Proposition 4.1.4] to write
   
lim+ (H p (Σ,σ))∗ St (x, ·), f H p (Σ,σ) = lim+ Lipc (Σ) St (x, ·), f (Lipc (Σ))
t→0 t→0
∫
= lim+ St (x, y) f (y) dσ(y)
t→0 Σ

= f (x) at σ-a.e. x ∈ Σ, (4.9.11)

214 4 Hardy Spaces on Ahlfors Regular Sets

where the last equality follows from (4.9.10). From (4.9.11) (and also keeping in
mind [133, (3.6.27)], (4.2.10)) the claims in (4.9.3) follow.
Having shown that H acts as the identity on H p (Σ, σ) when p ∈ [1, ∞), consider
  1 
next the case when p ∈ n−1 n , 1 . Having fixed some γ ∈ (n − 1) p − 1 , 1 , we
may rely on Lemma 4.6.4, (4.1.7), and (4.2.14) to conclude that there is a constant
C ∈ (0, ∞) with the property that for every f ∈ H p (Σ, σ) we have
 
sup (H p (Σ,σ))∗ St (x, ·), f H p (Σ,σ) ≤ C fγ (x) at each x ∈ Σ. (4.9.12)
t>0

In concert with (4.2.6), this implies that there exists some C ∈ (0, ∞) such that8
   
 
 sup (H p (Σ,σ))∗ St (x, ·), f H p (Σ,σ)  p ≤ C f  H p (Σ,σ) (4.9.13)
t>0 L x (Σ,σ)

for all f ∈ H p (Σ, σ). At this stage in the proof, the idea is to invoke [133, Propo-
sition 6.2.11] in the following concrete context: X := Σ × 0, diam Σ viewed as a
topological space with the topology inherited from Rn ×R, the set X := Σ×{0} ⊆ X ,
the measure μ :=  σ on  X = Σ × {0} ≡ Σ which,  according to [133, Lemma 3.6.4],
is complete, Γ (x, 0) := {x} × (0, diam Σ ⊆ X \ X for each (x, 0) ∈ X (so
that the condition in [133, (6.2.71)] is actually satisfied at every point), the space
Y := H p (Σ, σ) equipped with the quasi-norm  · Y :=  ·  H p (Σ,σ) , the linear opera-
tor T mapping mapping vectors  from Y into functions defined on X \ X according
to (T f )(x, t) := (H p (Σ,σ))
 ∗ St (x, ·), f H p (Σ,σ) for each f ∈ Y = H (Σ, σ) and each

p
p,q
(x, t) ∈ X \ X = Σ× 0, diam Σ , and with V := Hfin (Σ, σ) for some fixed q ∈ (1, ∞)
which, according to (4.4.114), is a dense linear subspace of Y .
For these choices, (4.9.13) implies that the maximal operator T associated with
T as in [133, (6.2.73)] satisfies [133, (6.2.74)]. Also, (4.9.11) used with p := q
guarantees that the condition in [133, (6.2.75)] holds for every f ∈ V. Granted
these, we may rely on [133, Proposition 6.2.11] to ultimately conclude that the limit
in (4.9.1) exists at σ-a.e. point x ∈ Σ. In turn, the existence of this limit together with
(4.9.13) proves that the operator (4.9.1)-(4.9.2) is well-defined, linear, and bounded.
 
Furthermore, that the operators H associated with various values of p ∈ n−1 n ,∞
are compatible with one another is a consequence of (4.9.1) and the compatibility
results from Lemmas 4.6.4-4.6.8.
p,q   Next, recall from (4.9.3) that H f = f for each
f ∈ Hfin (Σ, σ) with p ∈ n−1 n , 1 and q ∈ (1, ∞). In concert with the continuity
of (4.9.2), this then yields (4.9.4). Also, that properties (4.9.2)-(4.9.3) determine H
uniquely is a consequence of the density result in (4.4.114). At this stage, the claims
about (4.9.5) follow from what we have proved so far and real-interpolation (cf.
(4.3.3), [133, (6.2.48)], as well as Proposition 1.3.7). That the operator (4.9.5) acts
according to (4.9.6) is a consequence of (4.9.1), Lemma 4.6.4, and (1.3.39). Finally,
(4.9.7) may be justified much as (4.9.3). 
 
Given a closed Ahlfors regular set Σ ⊆ Rn along with some exponent p ∈ n , 1 , n−1

in general there is no set-theoretic relationship between the Hardy scale H p (Σ, σ)

8 with  ·  L xp (Σ, σ) indicating that the L p quasi-norm on Σ is taken in the variable x
4.9 More on H p Versus L p : The Filtering Operator 215

and the Lebesgue space L p (Σ, σ) (where, as usual, σ := H n−1 Σ). This being
said, Theorem 4.9.1 permits us to canonically associate to any linear and bounded
operator T : H p (Σ, σ) → H p (Σ, σ) a linear and bounded operator T from H p (Σ, σ)
into L p (Σ, σ) via T := H ◦ T.
Chapter 5
Banach Function Spaces, Extrapolation, and
Orlicz Spaces

In this chapter we begin by providing a definition which is more inclusive than that
of a “standard” Banach function space (as traditionally used in the literature; cf.
e.g., [14]), and indicate that a significant portion of the classical theory goes through
for this more general brand, which we dub Generalized Banach Function Spaces.
This is done in §5.1. The relevance of this extension is that a variety of scales of
spaces of interest, such as Morrey spaces, block spaces, as well as Beurling algebras
and their pre-duals, now fit naturally into this more accommodating label. Most
significantly, in §5.2 we develop powerful and versatile extrapolation results serving
as portal, allowing us to pass from estimates on Muckenhoupt weighted Lebesgue
spaces (for a fixed integrability exponent and arbitrary weights) to estimates on
the brand of Generalized Banach Function Spaces introduced earlier, on which the
Hardy-Littlewood maximal operator happens to be bounded. Finally, in §5.3 we
focus on Orlicz spaces which, in particular, are natural examples of classical Banach
function spaces for which the machinery developed so far applies.

5.1 Generalized Banach Function Spaces

The notion of Banach function space is a well-established concept in Functional

and Harmonic Analysis, and there is a multitude of accounts where the theory of
“standard” Banach function space is discussed at length (see, e.g., [14] and the
references therein). Alas, this classical construct is not inclusive enough, as certain
scales of spaces we are interested in (such as Morrey and block spaces, as well as
Beurling algebras) fail to be Banach function spaces in a traditional sense. The goal
in this section is to develop a more inclusive brand of Banach function space, which
we dub Generalized Banach Function Spaces, in which the aforementioned scales of
spaces now fit naturally. We emphasize that, throughout this section, we follow [14]
very closely (with necessary alterations to accommodate our more general setting).
To get started, we make the following definition:

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 217
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_5
218 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Definition 5.1.1 Let (X, M, μ) be an arbitrary measure space1 and denote

 
M (X, μ) : = f : X → R : f is μ-measurable , (5.1.1)
 
M+ (X, μ) : = f ∈ M (X, μ) : f ≥ 0 . (5.1.2)

A mapping ρ : M+ (X, μ) → [0, ∞] is called a function norm provided the follow-

ing properties are satisfied for all f , g ∈ M+ (X, μ):
(P1) ρ( f ) = 0 if and only if f = 0 at μ-a.e. point in X, for each λ ∈ [0, ∞) one has
ρ(λ f ) = λρ( f ), and ρ( f + g) ≤ ρ( f ) + ρ(g);
(P2) if f ≤ g at μ-a.e. point in X then ρ( f ) ≤ ρ(g);
(P3) if { fk }k ∈N ⊂ M+ (X, μ) is such that fk increases to f pointwise μ-a.e. in X as
k → ∞, then ρ( fk ) increases to ρ( f ) as k → ∞;
∞
(P4) there exists a collection of sets {Yj } j ∈N ⊆ M with the property that X = Yj
  j=1
and ρ 1Yj < ∞ for all j ∈ N;

∞
(P5) there exists a collection of sets {Z j } j ∈N ⊆ M with the property that X = Zj
j=1
∫ such that for every j ∈ N there exists some constant c j ∈ (0, ∞) satisfying
and
Z
f dμ ≤ c j ρ( f ).
j

Note that (P2) implies that, for any two functions f , g ∈ M+ (X, μ),

ρ( f ) = ρ(g) whenever f = g at μ-a.e. point in X. (5.1.3)

Remark 5.1.2 Without loss of generality, we may assume that the sets in (P4)-(P5)
are such that Yj = Z j for all j ∈ N. Indeed, if (P4)-(P5) are true as stated, we may
consider the family {Wi j }(i, j)∈N2 defined by Wi j := Yi ∩ Z j , for (i, j) ∈ N2 . Then after
re-denoting this collection as {Wk }k ∈N it is immediate that the latter satisfies the
properties listed in (P4) and (P5).
Furthermore, we may assume that {Wk }k ∈N just defined is an increasing nested
sequence of sets. This may be achieved by replacing it with the sequence {W k }k ∈N
k :=  W j , for each k ∈ N. In summary, it is possible to combine properties
k
where W
j=1
(P4) and (P5) in Definition 5.1.1 as the following “locality” property:
in the context of Definition 5.1.1, properties (P4)-(P5) are equivalent
with the existence of a collection of sets {W N } N ∈N ⊆ M with the
property
  that W N X as N → ∞, and for each N ∈ N one has (5.1.4)
ρ
∫ 1W N < ∞ and there exists some constant c N ∈ (0, ∞) such that
W
f dμ ≤ c N ρ( f ) for every f ∈ M+ (X, μ).
N

Let us now compare our notion of Generalized Banach Function Space introduced
in Definition 5.1.1 with the classical concept of Banach function space.

1 we emphasize that we do not require that the measure μ is complete, or sigma-finite

 5.1 Generalized Banach Function Spaces 219

Remark 5.1.3 The class of function norms defined in Definition 5.1.1 is a more
general than the class of Banach function norms defined as in [14, Definition 1.1,
p. 2]. The difference is that [14, Definition 1.1, p. 2] also makes the assumption that
μ is sigma-finite and properties (P4)-(P5) above are replaced by the following more
restrictive axioms:
 
(P4’) for every μ-measurable set E ⊆ X with μ(E) < ∞ one has ρ 1E < ∞;
(P5’) for each μ-measurable
∫ set E ⊆ X with μ(E) < ∞ there exists CE ∈ (0, ∞) with
the property E f dμ ≤ CE ρ( f ) for every f ∈ M+ (X, μ).
Observe that if the measure μ is sigma-finite then there exists some countable
collection {X j } j ∈N of μ-measurable sets with the property that X = ∞
j=1 X j and
μ(X j ) < ∞ for all j ∈ N, and properties (P4’)-(P5’) being satisfied imply that
(P4)-(P5) hold with Yj := X j and Z j := X j for each j ∈ N.
Even in a standard metric-measure theoretic setting, such as the Euclidean space
Rn equipped with the Lebesgue measure, the classical scales of Morrey spaces and
block spaces (see §6.2) fail to satisfy properties (P4’)-(P5’) above, but they do satisfy
(P4)-(P5) in Definition 5.1.1 (see Proposition 6.2.17).
Moving on, we make the following definition, formally introducing the concept
of Generalized Banach Function Space.
Definition 5.1.4 Let (X, M, μ) be an arbitrary measure space and let ρ be a function
norm as in Definition 5.1.1. Then the set
 
X := f ∈ M (X, μ) : ρ(| f |) < ∞ (5.1.5)

is called a Generalized Banach Function Space (GBFS) on X. In such a scenario,

for each f ∈ M (X, μ) define

ρ(| f |) if f ∈ X,
f X := ρ(| f |) = (5.1.6)
+∞ if f  X.

Remark 5.1.5 Whenever (X, M, μ) is a sigma-finite measure space, and ρ is a func-

tion norm in a “classical” sense, as defined in [14, Definition 1.1, p. 2] (i.e., with
properties (P4)-(P5) in our Definition 5.1.1 replaced by properties (P4’)-(P5’) from
Remark 5.1.3), we shall refer to the X defined in (5.1.5) as being a (classical)
Banach function space.
Going back to Definition 5.1.4, in view of (5.1.6) one may then re-cast (5.1.5)
simply as  
X = f ∈ M (X, μ) : f X < ∞ . (5.1.7)
Also, from (5.1.6) and (P1) we see that

f X= 0 if and only if f = 0 at μ-a.e. point in X. (5.1.8)

 
Ultimately, we conclude that X, · X is a normed vector space. In addition, (P5)
simply says that, for each j ∈ N,
220 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

the restriction operator X f → f Zj

∈ L 1 (Z j , μ)
(5.1.9)
is well defined and bounded, with norm ≤ c j .

For future reference, we observe that this implies

if f ∈ X then | f | < ∞ at μ-a.e. point in X. (5.1.10)

Indeed, given any f ∈ X, from (5.1.9) we see that f | Z j ∈ L 1 (Z j , μ) for each j ∈ N,

hence | f | < ∞ at μ-a.e. point on Z j for each j ∈ N. Now (5.1.10) follows since

∞
X= Zj.
j=1
It is also useful to remark that, as seen from (5.1.3) and (5.1.6), for any two given
functions f , g ∈ M (X, μ) we have

f X = g X whenever f = g at μ-a.e. point in X. (5.1.11)

Finally, for later use we observe here that

if f , g : X → R are two μ-measurable functions such that
|g| ≤ | f | at μ-a.e. point on X and f ∈ X, then g also (5.1.12)
belongs to the space X and g X ≤ f X ,
and
the operator of pointwise multiplication by some given function b in
L ∞ (X, μ) is a bounded mapping from the space X into itself, with (5.1.13)
operator norm ≤ b L ∞ (X,μ) .

Proposition 5.1.6 Let (X, M, μ) be an arbitrary measure space, let ρ be a function

norm
 as in Definition 5.1.1, and let X and · X be as in Definition 5.1.4. Then
X, · X is a normed vector space. In addition, if f ∈ X and { fn }n∈N ⊂ X is
a sequence with the property that fn − f X → 0 as n → ∞, then there exists
a sub-sequence { fnk }k ∈N of { fn }n∈N that converges to f pointwise μ-a.e. in X as
k → ∞.
Proof As seen in (5.1.10), every f ∈ X is finite μ-a.e. on X. This and the properties
in (P1)-(P2) further imply that X, · X is a normed vector space.
Next, suppose { fn }n∈N ⊂ X and f ∈ X are such that fn − f X → 0 as n → ∞.
Recall the sets {W N } N ∈N from (5.1.4). In particular, for each N ∈ N there exists
some CN ∈ (0, ∞) with the property that for each g ∈ X we have

g|WN L 1 (WN ,μ)

≤ CN g X. (5.1.14)

 g := fn − f , n ∈ N, and then use the

Hence, if we write (5.1.14) for N := 1 and with
current assumptions, we obtain that fn |W1 n∈N converges to f |W1 in L 1 (W1, μ) as
n → ∞. As a consequence, there exists a sub-sequence { fnk }k ∈N that converges to f
at μ-a.e. point in W1 as k → ∞. Next, iterate this process. That is, we write (5.1.14)
for N := 2 and g := fnk − f for each k ∈ N, to obtain a sub-sub-sequence { fnk j } j ∈N
5.1 Generalized Banach Function Spaces 221

that converges to f at μ-a.e. point in W2 as j → ∞. Inductively, we do this for each

N ∈ N. The desired final sub-sequence is then selected via a Cantor diagonalization
argument. That this final sub-sequence converges to f pointwise μ-a.e. on X is then
seen from its construction and the fact that W N X as N ∞ (cf. (5.1.4)). 

Any Generalized Banach Function Space enjoys suitable versions of Lebesgue’s

Monotone Convergence Theorem and Fatou’s Lemma, as indicated in the next lem-
ma.
 
Lemma 5.1.7 Let X, · X be a Generalized Banach Function Space associated
as in Definition 5.1.4 with a measure space (X, M, μ) and a function norm ρ. Then
for any given sequence { fk }k ∈N in X the following properties are true.
(i) Suppose that for each k ∈ N one has 0 ≤ fk (x) ≤ fk+1 (x) for μ-a.e. point x ∈ X,
and define f (x) := lim fk (x) at μ-a.e. point x ∈ X. Redefine2 f on a μ-nullset
k→∞
as to make it a non-negative μ-measurable (see [133, (3.1.30)]). Then

either f  X and fk X ∞ as k → ∞,
(5.1.15)
or f ∈ X and fk X f X as k → ∞.

Simply put, with the convention made in (5.1.6), in all cases one has

fk X f X as k → ∞. (5.1.16)

(ii) Assume lim inf fk X < ∞ and suppose f (x) := lim fk (x) exists at μ-a.e. point
k→∞ k→∞
x ∈ X. Redefine3 the function f on a μ-nullset as to make it a non-negative
μ-measurable (cf. [133, (3.1.30)]). Then

f ∈ X and f X ≤ lim inf fk X. (5.1.17)

k→∞

Proof The statement in (i) is a consequence of (P3), bearing in mind that f redefined
as indicated in the statement belongs to M+ (X, μ). To deal with the statement in (ii),
note that f redefined belongs to M+ (X, μ) and there exists some set A ∈ M of
μ-measure zero such that sequence { fk (x)}k ∈N converges pointwise to f (x) at each
point x ∈ X \ A. To proceed, consider hk (x) := inf | fm (x)| for each x ∈ X \ A and
m≥k
every k ∈ N. This is a non-negative, increasing sequence of functions, and
 
lim hk (x) = sup hk (x) = sup inf | fm (x)| = lim inf | fk (x)|
k→∞ k ∈N k ∈N m≥k k→∞

= lim | fn (x)| = | f (x)| at each x ∈ X \ A. (5.1.18)

k→∞

Thus, 0 ≤ hk | f | pointwise on X \ A as k → ∞. Since as a consequence of (P2),

for each k ∈ N we have ρ(hk ) ≤ ρ(| fm |) for all m ≥ k, we may apply (P3) to write

2 this step is not necessary if the measure space (X, M, μ) is complete

3 again, this is no longer necessary if the measure μ is complete
222 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
 
f X = ρ(| f |) = lim ρ(hk ) ≤ lim inf ρ(| fm |)
k→∞ k→∞ m≥k

= lim inf ρ(| fk |) = lim inf fk X < +∞. (5.1.19)

k→∞ k→∞

With this in hand, (5.1.17) follows, so the proof of (ii) is complete. 

The point of the next proposition is that any Generalized Function Space also
enjoys the Riesz-Fischer property (which renders said space Banach).
 
Proposition 5.1.8 Let X, · X be a Generalized Banach Function Space associated
as in Definition 5.1.4 with a measure space (X, M, μ). Suppose fk ∈ X, with k ∈ N,
is a family of functions satisfying
∞
fk X < ∞. (5.1.20)
k=1

∞
Then fk converges in X to a function f ∈ X and
k=1

∞
f X ≤ fk X. (5.1.21)
k=1
 
As a consequence, the space X, · X is complete, hence Banach.

Proof For each n ∈ N, consider the function tn := n

k=1 | fk | ∈ M+ (X, μ). Invoking
(P1) and (5.1.20) we may write
n ∞
tn X ≤ fk X ≤ fk X < ∞ for all n ∈ N. (5.1.22)
k=1 k=1

∞
Thus, tn ∈ X for all n ∈ N, and if we set t := | fk |, then 0 ≤ tn t as n → ∞.
k=1
Since this convergence happens at every point in X, it follows that t is a μ-measurable
function. In light of (5.1.22), by part (i) in Lemma 5.1.7 we have t ∈ X. The latter
∞
and (5.1.10) ensure that | fk (x)| < ∞ for μ-a.e. x ∈ X which, in turn, implies that
k=1
∞
fk (x) is absolutely convergent for μ-a.e. x ∈ X. From [133, (3.1.30)] we know
k=1
∞
that there exists a μ-measurable function f on X with the property that f = fk
k=1
n
at μ-a.e. point in X. In particular, if we define sn := fk for each n ∈ N, then
k=1
sn ∈ X for each n ∈ N and sn → f at μ-a.e. point in X as n → ∞. Moreover, for
each m ∈ N, we have sn − sm → f − sm at μ-a.e. point in X as n → ∞ and
5.1 Generalized Banach Function Spaces 223
∞ ∞
lim inf sn − sm X ≤ fk X ≤ fk X < ∞. (5.1.23)
n→∞
k=m+1 k=1

We may therefore invoke part (ii) in Lemma 5.1.7 to conclude that f − sm ∈ X for
each m ∈ N and
∞
f − sm X ≤ lim inf sn − sm X ≤ fk X. (5.1.24)
n→∞
k=m+1

Consequently, f = ( f − sm ) + sm ∈ X and if we pass to the limit as m → ∞ in

(5.1.24) while relying on assumption (5.1.20), we obtain lim f − sm X = 0. This
m→∞
proves that sm → f in X as m → ∞. Furthermore, for each m ∈ N, we may write
m
f X ≤ f − sm X + sm X ≤ f − sm X+ fk X. (5.1.25)
k=1

The estimate in (5.1.21) now follows from (5.1.25) by passing to limit m → ∞.

To show that X is complete, pick an arbitrary Cauchy sequence {gn }n∈N in X.
Then for each k ∈ N there exists nk ∈ N such that gnk+1 −gnk X < 2−k . In particular,
the sequence fk := gnk+1 − gnk , indexed by k ∈ N, is contained in X and satisfies
(5.1.20). Invoking what we have proved so far, we obtain that there exists some f ∈ X
such that
N
f = lim fk = lim gn N +1 − gn1 in X. (5.1.26)
N →∞ N →∞
k=1

This shows that the sub-sequence {gnk }k ∈N converges in X to f + gn1 , which further
implies that the
 entire sequence
 {gn }n∈N converges in X to f + gn1 . Ultimately, we
conclude that X, · X is complete. 

To each given function norm one can associate a new function norm of the sort
described below.

Definition 5.1.9 Let (X, M, μ) be a measure space and let ρ be a function norm as
in Definition 5.1.1. Its associated norm is the function ρ : M+ (X, μ) → [0, ∞]
defined by
∫ 
ρ(g) := sup f g dμ : f ∈ M+ (X, μ), ρ( f ) ≤ 1 for all g ∈ M+ (X, μ).
X
(5.1.27)

Our next proposition clarifies the fact that ρ defined above is indeed a function
norm.

Proposition 5.1.10 If (X, M, μ) is a measure space and ρ is a function norm as in

Definition 5.1.1, then its associated norm ρ defined in Definition 5.1.9 is a function
norm.
224 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Proof Let g ∈ M+ (X, μ) be such that ρ(g) = 0. Then

∫
f g dμ = 0 for all f ∈ M+ (X, μ) with ρ( f ) ≤ 1. (5.1.28)
X

Since the goal is to show that g = 0 at μ-a.e. point in X, without loss of generality
we may assume that μ(X) > 0. By (P4) in Definition 5.1.1, there exists a collection

∞  
of sets {Yj } j ∈N ⊆ M with the property that X = Yj and ρ 1Yj < ∞ for all j ∈ N.
j=1
Let
J := { j ∈ N : μ(Yj ) = 0}. (5.1.29)
Then  
ρ 1Yj > 0 for each j ∈ N \ J. (5.1.30)
 
Indeed, otherwise ρ 1Yj = 0 would imply 1Yj = 0 at μ-a.e. point in X, thus
μ(Yj ) = 0, contradicting the fact that j  J.
In light of (5.1.30), for each integer j ∈ N \ J, we may set f j := μ(11Y ) 1Yj which
j
is μ-measurable and ρ( f j ) = 1 (the latter a consequence of property (P1) for ρ with
λ := 1/ρ(1Yj )). Applying (5.1.28), we have
∫ ∫
0= f j g dμ = g dμ, for all j ∈ N \ J. (5.1.31)
X Yj

Hence, g = 0 at μ-a.e. point in Yj . Since by the definition of J we have
j ∈N\J
 
μ Yj = 0, we may conclude that g = 0 at μ-a.e. point in X.
j ∈J
Conversely, assume g = 0 at μ-a.e. point in X and pick an arbitrary f ∈ M+ (X, μ)
with ρ( f ) ≤ 1. Then f ∈ X and | f | < ∞ ∫at μ-a.e. point in X (recall (5.1.10)) which
implies f g = 0 at μ-a.e. point in X, thus X f g dμ = 0. Hence, ρ(g) = 0 as wanted.
That ρ also satisfies the rest of the properties listed in (P1)-(P2) in Definition 5.1.1
follows from (5.1.27) and standard properties of integrals.
We also observe that, with the sets {Yj } j ∈N ⊆ M and J as above, if f ∈ M+ (X, μ)
then
∫ ∫ ∫
  1Y
f dμ = f 1Yk dμ = ρ 1Yk f ·  k  dμ
Yk X X ρ 1Yk
  
≤ ρ 1Yk ρ ( f ) (5.1.32)
∫
for each k ∈ N \ J, and Y f dμ = 0 if k ∈ J. Hence, ρ satisfies (P5) in Defini-
k  
tion 5.1.1 with Z j := Yj for each j ∈ N, and with ck := ρ 1Yk if k ∈ N \ J and
ck := 0 if k ∈ J.
To show that ρ satisfies (P3), let g ∈ M+ (X, μ) and {gk }k ∈N ⊂ M+ (X, μ) be
such that gk increases to g pointwise μ-a.e. in X as k → ∞. Since ρ satisfies (P2),
it follows that the sequence {ρ(gk )}k ∈N is increasing and that ρ(gk ) ≤ ρ(g) for all
k ∈ N. To finish proving that ρ satisfies (P3), we have to show that ρ(gk ) ρ(g)
5.1 Generalized Banach Function Spaces 225

as k → ∞. Without loss of generality we can assume ρ(gk ) < ∞ for all k ∈ N.

If c ∈ R satisfies c < ρ(g), then the definition 
∫ of ρ implies the existence of some
f ∈ M+ (X, μ) with ρ( f ) ≤ 1 and such that X f g dμ > c. Since gk f increases to
g f pointwise μ-a.e.
∫ in X as k ∫→ ∞, by the Monotone Convergence Theorem we
necessarily have X g fk dμ f g dμ as k → ∞. Hence, there exists some N ∈ N
∫ X
with the property that X g fk dμ > c for all k ≥ N. The arbitrariness of c satisfying
the condition c < ρ(g) now allows us to conclude that, as wanted, ρ(gk ) increases
to ρ(g) as k → ∞.
Finally, the fact that ρ satisfies (P5) in Definition 5.1.1, implies the existence of
∞
a collection {Z j } j ∈N ⊆ M with the property that X = Z j and such that for every
∫
j=1
k ∈ N there exists some constant ck ∈ (0, ∞) satisfying Z f dμ ≤ ck ρ( f ) for each
k
f ∈ M+ (X, μ). Then, for each k ∈ N,
  ∫ 
ρ 1 Zk = sup f dμ : f ∈ M+ (X, μ), ρ( f ) ≤ 1 ≤ ck < ∞. (5.1.33)
Zk

This shows that ρ satisfies (P4) in Definition 5.1.1 with Yj := Z j , for each j ∈ N,
and completes the proof of the proposition. 

We are now in a position to define X, the associated space of a given Generalized

Banach Function Space X.

Definition 5.1.11 Let (X, M, μ) be a measure space, ρ be a function norm as in Def-

inition 5.1.1, and ρ its associated norm as in Definition 5.1.9. Then the Generalized
Banach Function Space on (X, μ) defined in relation to ρ:
 
X := g ∈ M (X, μ) : ρ(|g|) < ∞ (5.1.34)

is called the associated space of X (defined in relation to ρ as in Definition 5.1.4).

According to the recipe in (5.1.6) implemented for ρ in place of ρ, the formula

given in (5.1.27), and (5.1.5), for every g ∈ M (X, μ) we have
∫ 
g X = ρ(|g|) = sup f |g| dμ : f ∈ M+ (X, μ), ρ( f ) ≤ 1
X
∫ 
= sup | f g| dμ : f ∈ X, f X ≤1 , (5.1.35)
X

and the Generalized Banach Function Space X may be characterized as

 
X = g ∈ M (X, μ) : g X < ∞ . (5.1.36)

There is a very useful (generalized) Hölder inequality accompanying the pair X,

X, where X is a given Generalized Banach Function Space and X is its associated
space.
226 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Proposition 5.1.12 Let X be a Generalized Banach Function Space on a measure

space (X, M, μ) with associated space X. Then for every f ∈ X and every g ∈ X
the function f g is absolutely integrable on X and
∫
| f g| dμ ≤ f X g X . (5.1.37)
X

Proof Let f ∈ X and g ∈ X be arbitrary. According to (5.1.10) (used for X in

place of X), we have |g| < ∞ at μ-a.e. point in X. If f X = 0, then f = 0 at μ-a.e.
point in X, so f g = 0 at μ-a.e. point in X. As such, both sides in the inequality in
(5.1.37) are zero. If f X > 0 then f / f X ∈ X has norm equal to 1 and, due to
(5.1.35), we have ∫
f
g dμ ≤ g X (5.1.38)
X f X
which in turn yields (5.1.37). 
The lemma below plays a role in the proof of Proposition 5.1.14, given a little
later.
Lemma 5.1.13 Let X be a Generalized Banach Function Space on a measure space
(X, M, μ) with associated space X. Then a μ-measurable function f on X belongs
to X if and only if f g is absolutely integrable on X for every g ∈ X.
Proof The left-to-right implication is an immediate consequence of Proposi-
tion 5.1.12. For the opposite implication we reason by contradiction. Suppose
f ∈ M (X, μ) is such that f g ∈ L 1 (X, μ) for every g ∈ X and ρ(| f |) = ∞. Re-
calling (5.1.35), the latter implies the existence of a sequence {gn }n∈N ⊂ M (X, μ)
satisfying
∫
gn X ≤ 1 and |gn f | dμ > n3 for every n ∈ N. (5.1.39)
X

∞
Invoking Proposition 5.1.8, it follows that n−2 gn converges in X to some function
n=1
g. However, by the second inequality in (5.1.39) we have
∫ ∫
|g f | dμ ≥ n−2 |gn f | dμ > n for every n ∈ N, (5.1.40)
X X

which contradicts the assumption that f g is absolutely integrable on X. 

Here is a natural version of the classical Lorentz-Luxemburg theorem correspond-
ing to Generalized Banach Function Spaces, which generalizes (and corrects4) [14,
Theorem 2.7, p. 10].
Proposition 5.1.14 Let X be a Generalized Banach Function Space on a measure
space (X, M, μ). Assume the measure μ is sigma-finite. Then X = X isometrically,
4 note that [14, (2.12), p. 11] forces γ to be non-negative, which is not necessarily the case
5.1 Generalized Banach Function Spaces 227

i.e., X coincides with its second associated space X := (X) and f X = f X
for each f ∈ X. In fact, one has

f X = f X for each f ∈ M (X, μ). (5.1.41)

Proof Let f ∈ X be arbitrary. Applying Proposition 5.1.12, it follows that f g is in

L 1 (X, μ) for every g ∈ X. Hence, Lemma 5.1.13 applies (with X in place of X)
and gives that f belongs to the associated space X. This establishes the inclusion
X ⊆ X. In addition, (5.1.35) and (5.1.37) imply
∫ 
f X = sup | f g| dμ : g ∈ X, g X ≤ 1 ≤ f X . (5.1.42)
X

To finish the proof of the fact that X coincides with its second associated space
X = (X) in an isometric fashion, it remains to show that X ⊆ X and that

f X ≤ f X for all f ∈ X . (5.1.43)

Pick f ∈ X arbitrary, let {W N } N ∈N be as in (5.1.4), and fix N ∈ N. Define the

μ-measurable function  
f N := min | f |, N 1WN , (5.1.44)
and observe that since 1WN ∈ X and 0 ≤ f N ≤ N1WN , by (P2) we have f N ∈ X.
Hence,
0 ≤ fN | f | at every point in X as N → ∞,
(5.1.45)
and f N ∈ X ⊆ X for each N ∈ N.
We may now invoke (ii) in Lemma 5.1.7 (with X in place of X) to conclude

f ∈ X and fN X f X as N → ∞. (5.1.46)

Moreover, a second application of (ii) in Lemma 5.1.7 (as stated for X) implies

either f  X and fN X ∞, (5.1.47)

or f ∈ X and fN X f X. (5.1.48)

At this stage we make the claim that

fN X ≤ fN X for all N ∈ N. (5.1.49)

Assuming (5.1.49) to be true, we may then send N → ∞ in the inequality in (5.1.49)

and (keeping in mind (5.1.46)) obtain

lim fN X ≤ lim fN X = f X < ∞. (5.1.50)

N →∞ N →∞

In light of (5.1.47)-(5.1.48), this further implies f ∈ X and f X ≤ f X , as

wanted.
228 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Let us now return to the task of justifying (5.1.49). Henceforth fix N ∈ N. To set

the stage, we introduce a couple of symbols and make some preliminary observations.
First, we define
 
L 1N := f ∈ L 1 (X, μ) : f = 0 μ-a.e. on X \ W N (5.1.51)

which is a closed subspace of L 1 (X, μ). Indeed, if {gn }n∈N ⊂ L 1N converges in

L 1 (X, μ) to some g ∈ L 1 (X, μ), then there exists a sub-sequence {gnk }k ∈N that
converges pointwise at μ-a.e. point in X to g (recall that μ is complete). In turn, this
forces g = 0 at μ-a.e. point in X \ W N (given that gnk = 0 at μ-a.e. point in X \ W N ),
so g ∈ L 1N , as wanted. As a consequence,
 
L 1N , · L 1 (X,μ) is a Banach space. (5.1.52)

Second, we consider a subset of L 1N defined by

 
U N := L 1N ∩ S where S := g ∈ X : g X ≤1 . (5.1.53)

It is clear that the set U N is convex and we claim that U N is also closed in L 1N . To see
why the latter is true, let {hn }n∈N ⊂ U N be a sequence that converges in L 1 (X, μ)
to some h ∈ L 1 (X, μ). Then there exists a sub-sequence {hnk }k ∈N that converges
pointwise at μ-a.e. point in X to h. Since also hnk X ≤ 1 for all k ∈ N, we may
apply item (ii) in Lemma 5.1.7 to obtain h X ≤ 1. Thus, h ∈ L 1N .
Third, we have
 1 ∗  
L N = φ ∈ L ∞ (X, μ) : φ = 0 μ-a.e. on X \ W N . (5.1.54)

To establish (5.1.54) we introduce the mapping

   ∗
Φ : φ ∈ L ∞ (X, μ) : φ = 0 at μ-a.e. point in X \ W N −→ L 1N
Φ(φ) := Λφ for each φ ∈ L ∞ (X, μ) vanishing at μ-a.e. point in X \ W N ,
∫
where Λφ ( f ) := φ f dμ for each function f ∈ L 1N .
X
(5.1.55)
This is a well-defined map and (5.1.54) will follow as soon as we prove that Φ is a
bijection.
∫ Suppose φ ∈ L ∞ (X, μ) is such that φ = 0 at μ-a.e. point in X \ W N and
φ f dμ = 0 for all f ∈ L 1N . Corresponding to f := (sgn φ)1WN ∈ L 1N , the latter
X ∫
gives 0 = W |φ| dμ, which further implies φ = 0 at μ-a.e. point in W N , hence
N
φ = 0 at μ-a.e. point in X. This proves that Φ is injective.
∗
To show that Φ is also surjective, let Λ ∈ L 1N be arbitrary. The Hahn-Banach
theorem guarantees the existence of an extension Λ  ∈ (L 1 (X, μ))∗ of Λ. Given that
we have the isometric identification (L (X, μ)) = L ∞ (X, μ) since the measure μ is
1 ∗

sigma-finite (cf., e.g., [55, Theorem 6.15, p. 190]), it follows that there exists some
function η ∈ L ∞ (X, μ) with the property that
5.1 Generalized Banach Function Spaces 229
∫
     
 f · 1WN =
Λ f · 1WN = Λ f η · 1WN dμ for all f ∈ L 1 (X, μ). (5.1.56)
X

If we now define φ0 := η · 1WN , then φ0 = 0 on X \ W N and (5.1.56)  implies

∗
Λ f = Λφ0 ( f ) for all f ∈ L 1N . Thus, Λ = Λφ0 as functionals in L 1N . This
establishes that Φ is also surjective, so the proof of (5.1.54) is finished.
Returning to the claim in (5.1.49), we remark that without loss of generality we
may assume f N X > 0. With this a standing assumption, we may then consider
the function f N / f N X which belongs to U N . On account of this and (P1), for any
dilation factor λ ∈ (1, ∞) we have
λ
gλ := f N ∈ L 1N and gλ  U N . (5.1.57)
fN X

Fix λ > 1 and invoke the Hahn-Banach Separation Theorem (cf., e.g., [46, Corol-
lary 12, p. 418], [165, Theorem 3.4, p. 59]) in the context of the Banach space L 1N ,
the closed and convex subset U N of L 1N , and the function gλ  U N . This guarantees
 ∗
the existence of a number γ ∈ R and a functional Λ ∈ L 1N such that

Re Λ(h) < γ < Re Λ(gλ ) for all h ∈ U N . (5.1.58)

Upon recalling (5.1.54), we know that there

∫ exists φ ∈ L ∞ (X, μ) satisfying φ = 0
at μ-a.e. point in X \ W N and Λ(h) = X φh dμ for all h ∈ L 1N . Then (5.1.58) is
equivalent with having
∫  ∫ 
Re φh dμ < γ < Re φgλ dμ for all h ∈ U N . (5.1.59)
WN WN

In turn, (5.1.59) implies

∫  ∫
Re φh dμ ≤ |φgλ | dμ for all h ∈ U N . (5.1.60)
WN WN

Let us now consider the function ψ : X → R defined by

⎧
⎪ φ(x)
⎪
⎨
⎪ if φ(x)  0,
ψ(x) := |φ(x)| at μ-a.e. x ∈ X. (5.1.61)
⎪
⎪
⎪1 if φ(x) = 0,
⎩
Then ψ is μ-measurable,

|ψ| = 1 and φ = |φ|ψ at μ-a.e. point in X, (5.1.62)

and for every h ∈ L 1N ,

h ∈ U N ⇐⇒ ψh ∈ U N ⇐⇒ ψ −1 |h| ∈ U N ⇐⇒ |h| ∈ U N . (5.1.63)

230 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

(consequences of the fact that h X = ψ −1 h X = ψh X for all h ∈ L 1N ). Together,

(5.1.63) and (5.1.60) imply
∫ ∫ ∫
|φh| dμ = Re |φh| dμ ≤ |φgλ | dμ for all h ∈ U N , (5.1.64)
WN WN WN

thus ∫ ∫
sup |φh| dμ ≤ |φgλ | dμ. (5.1.65)
h ∈U N WN WN

Next, we will show that (5.1.65) remains valid if the supremum is taken over
h ∈ S. To see why this is true, for each h ∈ S introduce the sequence of functions

hn := min{|h|, n} · 1WN for all n ∈ N. (5.1.66)

Then 0 ≤ hn ≤ |h| for each n ∈ N. From this, (P2), and the fact that h X ≤ 1
we conclude that hn X ≤ 1 for each n ∈ N. Moreover, in light of (5.1.4), for each
n ∈ N we have hn ∈ L 1N , thus hn ∈ U N for each n ∈ N. In addition,

hn |h|1WN as n → ∞ at every point in X, (5.1.67)

so an application of Lebesgue’s Monotone Convergence Theorem gives

∫ ∫
|φhn | dμ |φh| dμ as n → ∞, for each h ∈ S. (5.1.68)
WN WN

This ultimately implies

∫ ∫
sup |φh| dμ = sup |φh| dμ. (5.1.69)
h ∈U N WN h ∈S WN

Recalling that φ = 0 at μ-a.e. point in X \ W N , then using (5.1.69), then (5.1.65)

and finally, recalling the definition of gλ , we arrive at
∫ ∫ ∫
φ X = sup |φh| dμ = sup |φh| dμ = sup |φh| dμ
h ∈S X h ∈S WN h ∈U N WN
∫ ∫
λ
≤ |φgλ | dμ = |φ f N | dμ. (5.1.70)
WN fN X WN

After multiplying the resulting inequality in (5.1.70) by f N X / φ X and invoking

Proposition 5.1.12 we obtain
∫
φ
fN X ≤ λ f N dμ ≤ λ f N X . (5.1.71)
X φ X

Now (5.1.49) follows by passing to the limit with λ  1 in (5.1.71). This finishes
the proof of the fact that X coincides with its second associated space X = (X) in
an isometric fashion.
5.1 Generalized Banach Function Spaces 231

In turn, this shows that (5.1.41) holds whenever f ∈ X. If f  X then f  (X)

so f X = ∞ by (5.1.6) and f (X ) = ∞ by (5.1.36) written for X in place of X.
Hence, (5.1.41) holds in all cases. 
A simple yet useful consequence of Proposition 5.1.14 is the following charac-
terization of the norm in given Generalized Banach Function Space, in terms of its
associated space.
Corollary 5.1.15 Let X be a Generalized Banach Function Space on a measure space
(X, M, μ). Assume that the measure μ is sigma-finite. Then for each f ∈ M (X, μ)
one has ∫ 
f X = sup | f g| dμ : g ∈ X, g X ≤ 1 . (5.1.72)
X
Proof This is a direct consequence of (5.1.35) and Proposition 5.1.14. 
Other equivalent descriptions of the norms in a given Generalized Banach Func-
tion Space and its associated space are provided below.
 
Proposition 5.1.16 Let X, · X be a Generalized Banach Function Space associ-
ated as in Definition 5.1.4 with a measure space (X, M, μ). Consider its associated
space X from Definition 5.1.11 with norm defined in (5.1.35). Then
 ∫ 
g X = sup f g dμ : f ∈ X, f X ≤ 1 for all g ∈ X (5.1.73)
X

and if the measure μ is sigma-finite one also has

 ∫ 
f X = sup f g dμ : g ∈ X, g X ≤ 1 for all f ∈ X. (5.1.74)
X

Proof Fix an arbitrary g ∈ X. That the supremum in the right-hand side of (5.1.73)
∫ ∫
is ≤ g X follows from (5.1.35) and the fact that X f g dμ ≤ X | f g| dμ for every
f ∈ X. It remains to prove the opposite inequality, that is,
∫ ∫
sup | f g| dμ ≤ sup f g dμ (5.1.75)
f ∈S X f ∈S X
 
where S := f ∈ X : f X ≤ 1 . To proceed, set E := {x ∈ X : g(x)  0} and
define φ : X → R by φ(x) := g(x)/|g(x)| if x ∈ E and φ(x) := 0 if x ∈ X \ E. In
particular, φ is μ-measurable. If f ∈ S is arbitrary, then the function h := | f |φ is
μ-measurable and satisfies |h| ≤ | f |. Hence h X ≤ f X ≤ 1, which also implies
h ∈ X. Ultimately, h ∈ S. This together with the fact that |g| = φg then allow us to
write
∫ ∫ ∫ ∫ ∫
| f g| dμ = | f |φg dμ = hg dμ ≤ hg dμ ≤ sup ψg dμ . (5.1.76)
X X X X ψ ∈S X

Now (5.1.75) follows by taking the supremum over f ∈ S in (5.1.76). This completes
the proof of (5.1.73).
232 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Finally, if the measure μ is sigma-finite, Proposition 5.1.14 is applicable. Bearing

this in mind, that the equality claimed in (5.1.74) holds for each f ∈ X = (X) is
seen from (5.1.41) and (5.1.73) written for X in place of X. 

We conclude with result to the effect that set theoretic inclusions for Generalized
Banach Function Spaces are always continuous injections.
 
Proposition 5.1.17 Let X, · X be a Generalized Banach Function Space associat-
ed as in Definition 5.1.4 with a measure space (X, M, μ). Suppose V ⊆ M (X, μ) is a
linear space such that X ⊆ V and for which there exists a mapping · V : V → [0, ∞)
satisfying the following properties:
(i) | f | ∈ V and | f | V = f V for every f ∈ V;
(ii) λ f V = λ f V for all f ∈ V and all λ ∈ [0, ∞);
(iii) if f , g ∈ V are such that 0 ≤ f ≤ g at μ-a.e. point in X, then f V ≤ g V.

Then there exists C ∈ (0, ∞) such that

f V ≤C f X for all f ∈ X. (5.1.77)

Proof First we observe that if f ∈ V is such that f = 0 at μ-a.e. point in X then

f V = 0. Indeed, if f ∈ V satisfies f = 0 at μ-a.e. point in X, then | f | ∈ V by
(i), and 2| f | = | f | at μ-a.e. point in X. This and (iii) imply f V = 2 f V , hence
f V = 2 f V by (ii), which in the end gives f V = 0 as wanted.
Suppose (5.1.77) is false. Then for each n ∈ N there exists gn ∈ X such that

gn V > n3 g n X. (5.1.78)

If some gn X = 0, then gn = 0 at μ-a.e. point in X, which by the earlier observation

gives gn V = 0 contradicting (5.1.78). This shows that we necessarily have that
∞ > gn X  0 for every n ∈ N and we can define fn := g|gnn |X for each n ∈ N. This
sequence of μ-measurable functions satisfies

fn ≥ 0, fn X = 1, and fn V > n3 for all n ∈ N. (5.1.79)

∞ ∞
Consequently, n−2 fn X < ∞ and Proposition 5.1.8 gives that n−2 fn con-
n=1 n=1
∞
verges in X to some f ∈ X. Proposition 5.1.6 then ensures that n−2 fn also con-
n=1
verges pointwise μ-a.e. to f . In particular, for every n ∈ N, we have f ≥ n−2 fn ≥ 0
at μ-a.e. point in X. Invoking (iii), (ii), and the last inequality in (5.1.79) it follows
that
f V ≥ n−2 fn V = n−2 fn V > n for each n ∈ N. (5.1.80)
which contradicts the fact that f V is finite. This proves that (5.1.77) holds for
some constant C ∈ (0, ∞). 
5.2 Extrapolation Theory 233

5.2 Extrapolation Theory

The proof of Rubio de Francia’s extrapolation theorem, formulated in [133, Propo-

sition 7.7.6], lends itself to a considerably more versatile and powerful result to
the effect that estimates on Muckenhoupt weighted Lebesgue spaces (for a fixed
integrability exponent and arbitrary weights) yield estimates on Generalized Banach
Function Spaces (on which the Hardy-Littlewood maximal operator turns out to be
bounded). Our main result to this effect (refining work in [37], [123]) is the theorem
below, itself a generalization of [133, Proposition 7.7.6].

Theorem 5.2.1 Let (X, d, μ) be a space of homogeneous type with the property that
the quasi-distance d : X × X → [0, ∞) is continuous5 in the product topology τd ×τd .
Also, denote by M X the Hardy-Littlewood maximal operator on (X, d, μ) and, given
a Generalized Banach Function Space X on (X, μ), denote by X its associated space
(cf. Definition 5.1.4 and Definition 5.1.11).
Then there exists some CX ∈ (0, ∞), depending only on the quasi-distance d (via
the constants Cd, Cd defined as in (A.0.19)-(A.0.20)) and the doubling charter of μ,
such that the following statements are true:

(1) Fix some integrability exponent p0 ∈ (1, ∞) and, assuming

M X : X → X and M X : X → X
(5.2.1)
are well-defined bounded mappings,

introduce
p p0 −1
WX, p0 := CX0 · M X X→X · MX X →X
(5.2.2)
p p0 −1
and WX, p0 := CX0 · M X X →X · MX X→X .

Also, suppose f , g are two μ-measurable real-valued functions defined on X

with the property that for each Muckenhoupt weight w ∈ Ap0 (X, d, μ) with
[w] A p0 ≤ WX, p0 one has

f L p0 (X,w) ≤ Cw g L p0 (X,w) (5.2.3)

for some constant Cw ∈ (0, ∞) which depends only on the ambient (via d, μ, p0 ),
and w. Then  
f X ≤ 22−1/p0 · sup Cw g X. (5.2.4)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

Moreover, if (5.2.3) holds for each Muckenhoupt weight w ∈ Ap0 (X, d, μ) with
[w] A p0 ≤ WX, p0 then

5 The result proved in [133, Theorem 7.1.2] guarantees that any quasi-metric space has an equivalent
quasi-distance satisfying this property
234 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
 
f X ≤2 2−1/p0
· sup Cw g X . (5.2.5)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

(2) Assume
M X : X → X is a well-defined bounded mapping, (5.2.6)
and define
WX,1 := CX · M X X →X . (5.2.7)
In addition, suppose f , g are two μ-measurable real-valued functions defined
on X with the property that for each Muckenhoupt weight w ∈ A1 (X, d, μ) with
[w] A1 ≤ WX,1 one has

f L 1 (X,w) ≤ Cw g L 1 (X,w) (5.2.8)

for some constant Cw ∈ (0, ∞) which depends only on the ambient (via d, μ),
and w. Then  
f X ≤ 2· sup Cw g X. (5.2.9)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

(3) Assume

M X : X → X is a well-defined bounded mapping, (5.2.10)

and define
WX,1 := CX · M X X→X . (5.2.11)
In addition, suppose f , g are two μ-measurable real-valued functions defined
on X with the property that for each Muckenhoupt weight w ∈ A1 (X, d, μ) with
[w] A1 ≤ WX,1 one has

f L 1 (X,w) ≤ Cw g L 1 (X,w) (5.2.12)

for some constant Cw ∈ (0, ∞) which depends only on the ambient (via d, μ),
and w. Then  
f X ≤ 2· sup Cw g X . (5.2.13)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

Proof To prove the claims made in item (1), assume p0 ∈ (1, ∞) and work under
the assumptions made in (5.2.1). Then M X X→X < ∞ and M X X →X < ∞. We
claim that we also have M X X→X > 0. To justify this, bring in the family {W j } j ∈N
of μ-measurable subsets of X considered in (5.1.4). Then 1Wj ∈ X for each j ∈ N
and μ(W j ) μ(X) > 0 as j → ∞. In particular,  there
 exists some jo ∈ N such
that μ(W jo ) > 0. If M X X→X = 0 then M X 1Wj o = 0 in X which, by (P1) in
Definition 5.1.1 forces M X 1Wj o = 0 at μ-a.e. point in X hence, further, 1Wj o = 0 at
μ-a.e. point in X, in contradiction with the choice of jo . Likewise, M X X →X > 0.
5.2 Extrapolation Theory 235

These considerations allow us to implement a version of Rubio de Francia’s iterative

algorithm in the present setting, i.e., define

T : X −→ X by setting
∞ j
MX φ (5.2.14)
T φ := j
for each function φ ∈ X,
j=0 2j M X X→X

j
where M X0 φ := |φ| for each φ ∈ X, and M X := M X ◦ · · · ◦ M X is the j-fold
composition of M X : X → X with itself, for each j ∈ N, as well as

T  : X −→ X by setting
∞ j

MX ψ (5.2.15)
T ψ := j
for each function ψ ∈ X,
j=0 2j MX X →X

where M X0 ψ := |ψ| for each ψ ∈ X, and M X := M X ◦ · · · ◦ M X is the j-fold

j

composition of the operator M X : X → X with itself, for each j ∈ N. Then,

thanks to item (1) in [133, Lemma 7.7.1], T, T  are well-defined bounded sublinear
operators with
 
T X→X := sup T φ X : φ X = 1 ≤ 2,
  (5.2.16)
T  X →X := sup T  ψ X : ψ X = 1 ≤ 2.

Bearing in mind the nature of the first term in (5.2.14) it follows that

for each function φ ∈ X one has

(5.2.17)
|φ| ≤ T φ and M X (T φ) ≤ 2 M X X→X T φ.

The above properties further imply

0 < T φ < ∞ at μ-a.e. point on X whenever 0  φ ∈ X. (5.2.18)

In a similar fashion,

for each function ψ ∈ X one has

|ψ| ≤ T ψ and M X (T ψ) ≤ 2 M X X →X T
 ψ,
(5.2.19)
plus 0 < T ψ < ∞ at μ-a.e. point on X if ψ  0.

Based on the properties recorded in (5.2.17)-(5.2.18) and item (4) in [133,

Lemma 7.7.1] we then conclude that

if φ ∈ X is not identically zero (μ-a.e.), then

(5.2.20)
T φ ∈ A1 (X, μ) and [T φ] A1 ≤ C M X X→X
236 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

d
where C ∈ (0, ∞) depends only on the quasi-distance d (via the constants Cd, C
defined as in (A.0.19)-(A.0.20)), and the doubling charter of μ. Likewise, from
(5.2.19) and item (4) in [133, Lemma 7.7.1] we see that

if ψ ∈ X is not identically zero (μ-a.e.), then

(5.2.21)
T ψ ∈ A1 (X, μ) and [T ψ] A1 ≤ C M X X →X .

Let us focus on (5.2.4). Note that this is trivially true if f = 0 at μ-a.e. point in
X, or if g X is either ∞ or 0 (since in the latter case (5.2.3) with, say, w ≡ 1, forces
f = 0 at μ-a.e. point in X). As far as (5.2.4) is concerned, we are therefore left with
considering the situation when

f is not identically zero (μ-a.e.), and 0 < g X < ∞. (5.2.22)

Hence, f X > 0 and 0  g ∈ X. In particular, it is meaningful to define

g

g := , (5.2.23)
g X
which is a function satisfying


g ∈ X and 
g X = 1. (5.2.24)

The remainder of the proof of (5.2.4) is divided into several steps, starting with:
Step I. Proof of (5.2.4) under the additional assumption f ∈ X. In particular,
0 < f X < ∞, and Proposition 5.1.14 together with Definition 5.1.9 ensure that
∫
f X = f X = sup | f ||h| dμ. (5.2.25)
h ∈X X
h X =1

Fix an arbitrary number θ ∈ (0, 1). Then the above considerations guarantee the
existence of a function
∫
hθ ∈ X with hθ X = 1 and θ f X ≤ | f ||hθ | dμ. (5.2.26)
X

Granted (5.2.24) and the first two properties in (5.2.26), we conclude from (5.2.20)
and item (2) in [133, Lemma 7.7.1] that there exists CX ∈ (0, ∞) such that

g ∈ A1 (X, μ) and T  hθ ∈ A1 (X, μ) with

T
(5.2.27)
[T
g ] A1 ≤ C X M X X→X and [T  hθ ] A1 ≤ CX M X X →X .

In turn, based on these properties and item (3) in [133, Lemma 7.7.1] we see that if

g )1−p0 (T  hθ )
wθ := (T (5.2.28)

then
5.2 Extrapolation Theory 237

wθ ∈ Ap0 (X, μ) and [wθ ] A p0 ≤ WX, p0 (5.2.29)

with the inequality in (5.2.29) implied by (5.2.2). With q0 ∈ (1, ∞) denoting the
Hölder conjugate exponent of p0 , may now write
∫ ∫ ∫

θ f X≤ | f ||hθ | dμ ≤ | f |(T hθ ) dμ = g ) p0 −1 wθ dμ
| f |(T
X X X
∫ ∫  1/p0  ∫  1/q0
= g ) p0 −1 dwθ ≤
| f |(T | f | p0 dwθ g )(p0 −1)q0 dwθ
(T
X X X
∫  1/q0
= f L p0 (X,wθ ) g )(p0 −1)q0 (T
(T g )1−p0 (T  hθ ) dμ
X
∫  1/q0
= f L p0 (X,w θ) g )(T  hθ ) dμ
(T
X
  1/q0
≤ f L p0 (X,wθ ) g
T X T  hθ X

≤ 41/q0 f L p0 (X,wθ ) . (5.2.30)

Above, the first two inequalities come from (5.2.26) and (5.2.19), respectively. In
the next two equalities we have used the definition of the function wθ given in
(5.2.28), and the fact that dwθ = wθ dμ. Going further, we have employed Hölder’s
inequality and once again (5.2.28), then relied on the identity (p0 −1)(q0 −1) = 1 and
invoked Proposition 5.1.12. The final inequality in (5.2.30) is implied by (5.2.16).
In summary, (5.2.30) proves that

θ f X ≤ 41/q0 f L p0 (X,wθ ) . (5.2.31)

We next estimate g. To get started, recall from (5.2.18) and (5.2.24) that

0 < T
g < ∞ at μ-a.e. point in X. (5.2.32)

Also, use (5.2.23) and the first property in (5.2.17) to write

|g| = g X
g≤ g g.
XT (5.2.33)

Bearing (5.2.32) in mind, this implies

g )−1 ≤ g
|g|(T X at μ-a.e. point in X. (5.2.34)

After rising both sides of (5.2.34) to the power p0 − 1 > 0 and multiplying by |g|,
we arrive (again, bearing (5.2.32) in mind) at the conclusion that
  1−p0 p0 −1
g
|g| p0 T ≤ g X |g| at μ-a.e. point in X. (5.2.35)

Based on the definition of the weight wθ given in (5.2.28), (5.2.35), Proposi-

tion 5.1.12, (5.2.16), and (5.2.26) we therefore obtain
238 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
∫  1/p0
  1−p0 
g L p0 (X,w θ) = |g| p0 T
g (T hθ ) dμ
X
p0 −1 ∫  1/p0
p0 
≤ g X |g| (T hθ ) dμ
X
p0 −1   1/p0
T  hθ T  hθ
p0 1/p0
≤ g X g X X = g X X

≤ 21/p0 g X, (5.2.36)

hence
g L p0 (X,wθ ) ≤ 21/p0 g X. (5.2.37)
Combining (5.2.31), (5.2.3), (5.2.37) (in this order) and recalling (5.2.29) yields
 
θ f X ≤ 22−1/p0 · sup Cw g X. (5.2.38)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

Upon sending θ 1 then justifies (5.2.4) in the scenario in which f ∈ X.

Step II. The end-game in the proof of (5.2.4). From Step I we know that (5.2.4)
is true if f ∈ X. To eliminate the latter additional assumption, bring in the family
{Wk }k ∈N of μ-measurable subsets of X considered in (5.1.4). Then 1Wk ∈ X for each
k ∈ N and Wk X as k → ∞. If we now define

fk := min{| f |, k} · 1Wk for each k ∈ N, (5.2.39)

then each fk is a μ-measurable function on X satisfying 0 ≤ fk ≤ k · 1Wk . As such,

(P2) in Definition 5.1.1 together with Definition 5.1.4 imply that

each fk belongs to X. (5.2.40)

Also, 0 ≤ fk | f | as k → ∞. In particular, (5.2.3) implies that for every Mucken-

houpt weight w ∈ Ap0 (X, d, μ) with [w] A p0 ≤ WX, p0 we have

fk L p0 (X,w) ≤ f L p0 (X,w) ≤ Cw g L p0 (X,w) . (5.2.41)

On account of this and (5.2.40), we conclude from (5.2.41) and Step I that
 
fk X ≤ 22−1/p0 · sup Cw g X for each k ∈ N. (5.2.42)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

Passing to limit k → ∞ and relying on item (ii) in Lemma 5.1.7 (while bearing in
mind (5.2.22)) we then arrive at the conclusion that (5.2.4) holds as stated.
5.2 Extrapolation Theory 239

Step III. The proof of (5.2.5). In view of Proposition 5.1.14, the estimate claimed
in (5.2.5) is a consequence of (5.2.4) written for the Generalized Banach Function
Space X.
At this stage, all claims in item (1) have been justified, and we turn our attention
to item (2). The goal is to establish (5.2.9), now working under the assumption made
in (5.2.6) in place of (5.2.1), and with (5.2.8) in place of (5.2.3). To this end, we
shall largely follow the same approach as in the treatment of item (1) in which we
now formally take p0 := 1, so we will only focus on the novel aspects. As before,
suppose first that f ∈ X. In this scenario, in place of (5.2.28) we now simply define

wθ := T  hθ (5.2.43)

which in view of (5.2.27) implies

wθ ∈ A1 (X, μ) and [wθ ] A1 ≤ CX M X X →X = WX,1 (5.2.44)

where the final equality in (5.2.44) is a consequence of (5.2.7). Much as in (5.2.30),

we then have
∫ ∫
θ f X≤ | f ||hθ | dμ ≤ | f |(T  hθ ) dμ
X X
∫
= | f | dwθ = f L 1 (X,wθ ), (5.2.45)
X

hence
θ f X ≤ f L 1 (X,wθ ) . (5.2.46)
As far as the function g is concerned, (5.2.43), Proposition 5.1.12, (5.2.16), and
(5.2.26) yield
∫
g L 1 (X,wθ ) = |g| (T  hθ ) dμ ≤ g X T  hθ X ≤ 2 g X . (5.2.47)
X

Collectively, (5.2.46), (5.2.8), (5.2.47), and (5.2.44) permit us to write

 
θ f X ≤ 2· sup Cw g X. (5.2.48)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

The estimate claimed in (5.2.9) is then obtained from this, in the case when f ∈ X,
after sending θ 1. Lastly, the case when f is arbitrary is dealt using the same type
of reasoning as the one employed in Step II above, with p0 := 1. This completes the
proof of item (2).
Finally, item (3) is a consequence of the result in item (2) written for the General-
ized Banach Function Space X in place of X, bearing in mind the fact that X = X
(cf. Proposition 5.1.14). 
240 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

A byproduct of the proof of the extrapolation result in Theorem 5.2.1 is the (ab-
stract) embedding result of Generalized Banach Function Spaces into Muckenhoupt
weighted Lebesgue spaces, described in the corollary below.

Corollary 5.2.2 Assume (X, d, μ) is a space of homogeneous type with the property
that the quasi-distance d : X × X → [0, ∞) is continuous6 in the product topology
τd × τd . Let M X be the Hardy-Littlewood maximal operator on (X, d, μ). Finally,
suppose X is a Generalized Banach Function Space on (X, μ) and denote by X its
associated space (cf. Definition 5.1.4 and Definition 5.1.11).
Then there exists some CX ∈ (0, ∞), depending only on the quasi-distance d (via
the constants Cd, Cd defined as in (A.0.19)-(A.0.20)) and the doubling charter of μ,
such that

M X : X → X boundedly =⇒ X ⊆ L 1 (X, w μ), (5.2.49)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

where
WX,1 := CX · M X X →X (5.2.50)
and

M X : X → X boundedly =⇒ X ⊆ L 1 (X, w μ) (5.2.51)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

where
WX,1 := CX · M X X→X . (5.2.52)
Finally, if
M X : X → X and M X : X → X
(5.2.53)
are well-defined bounded mappings,
then for each p0 ∈ (1, ∞) one has
 
X⊆ L p0 (X, w μ) and X ⊆ L p0 (X, w μ) (5.2.54)
w ∈ A p0 (X,d,μ) w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0 [w] A p ≤WX, p0
0 0

where
p p0 −1
WX, p0 := CX0 · M X X→X · MX X →X
(5.2.55)
p p0 −1
and WX, p0 := CX0 · M X X →X · MX X→X .

Proof All desired conclusions are implicit in the proof of Theorem 5.2.1. To be
specific, assume first that M X is bounded on X and fix an arbitrary function h ∈ X
with h X = 1. For this choice of h used in place of hθ , the estimates in (5.2.47)
6 The result proved in [133, Theorem 7.1.2] guarantees that any quasi-metric space has an equivalent
quasi-distance satisfying this property
5.2 Extrapolation Theory 241

prove that any function g ∈ X belongs to the space L 1 (X, w μ) for some choice of
a weight w ∈ A1 (X, μ) with [w] A1 ≤ WX,1 . This justifies the claim in (5.2.49). The
implication in (5.2.51) is then a consequence of (5.2.49) and Proposition 5.1.14.
To deal with (5.2.54), assume (5.2.53) and pick an exponent p0 ∈ (1, ∞). Given
any nontrivial function g ∈ X, run the argument which has produced (5.2.37) (with
h as above). This shows that there exists some weight w ∈ Ap0 (X, μ) satisfying
[w] A p0 ≤ WX, p0 and with the property that g ∈ L p0 (X, w), thus establishing the first
inclusion in (5.2.54). Finally, the second inclusion in (5.2.54) becomes a consequence
of the first, keeping in mind the fact that X = X (cf. Proposition 5.1.14). 

For certain applications, it is useful to formulate an extrapolation result in the

spirit of Theorem 5.2.1 for operators (which are not necessarily linear), of the sort
described in the next corollary.

Corollary 5.2.3 Assume (X, d, μ) is a space of homogeneous type with the property
that the quasi-distance d : X × X → [0, ∞) is continuous7 in the product topology
τd × τd . Let M X be the Hardy-Littlewood maximal operator on (X, d, μ). Finally,
suppose X is a Generalized Banach Function Space on (X, μ) and denote by X its
associated space (cf. Definition 5.1.4 and Definition 5.1.11). Fix some integrability
exponent p0 ∈ [1, ∞) and assume
 

L (X, w)
p0
f −→ Θ( f ) ∈ M (X, μ) (5.2.56)
w ∈ A p0 (X,μ)

constitutes an assignment8 with the property that for any given Muckenhoupt weight
w ∈ Ap0 (X, d, μ) one has

Θ( f ) L p0 (X,w) ≤ Cw f L p0 (X,w) for each f ∈ L p0 (X, w), (5.2.57)

where the constant9 Cw ∈ (0, ∞] depends only on Θ, d, μ, p0 , and w. Finally,

denote by M X the Hardy-Littlewood maximal operator on (X, d, μ) and, given a
Generalized Banach Function Space X on (X, μ), denote by X its associated space
(cf. Definition 5.1.4 and Definition 5.1.11).
Then there exists some CX ∈ (0, ∞), depending only on the quasi-distance d (via
the constants Cd, Cd defined as in (A.0.19)-(A.0.20)) and the doubling charter of μ,
such that the following statements are true:

(1) If p0 ∈ (1, ∞) and

M X : X → X and M X : X → X
(5.2.58)
are well-defined bounded mappings,

7 from [133, Theorem 7.1.2] it is known that any quasi-metric space has an equivalent quasi-distance
satisfying this property
8 not necessarily linear
9 note that Cw = ∞ is allowed, and we make the convention that ∞ · 0 = ∞
242 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

then, with
p p0 −1
WX, p0 := CX0 · M X X→X · MX X →X
(5.2.59)
p p0 −1
and WX, p0 := CX0 · M X X →X · MX X→X,

it follows that for each f ∈ X the function Θ( f ) is well defined and μ-measurable
on X (see the first inclusion in (5.2.54)) and
 
Θ( f ) X ≤ 22−1/p0 · sup Cw f X, (5.2.60)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

while for each f ∈ X the function Θ( f ) is well defined and μ-measurable on X

(see the second inclusion in (5.2.54)) and
 
Θ( f ) X ≤ 22−1/p0 · sup Cw f X . (5.2.61)
w ∈ A p0 (X,d,μ)
[w] A p ≤WX, p0
0

(2) If p0 = 1 and

M X : X → X is a well-defined bounded mapping, (5.2.62)

then, with
WX,1 := CX · M X X →X , (5.2.63)
it follows that for each f ∈ X the function Θ( f ) is well defined and μ-measurable
on X (see (5.2.49)) and
 
Θ( f ) X ≤ 2· sup Cw f X. (5.2.64)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1

(3) If p0 = 1 and

M X : X → X is a well-defined bounded mapping, (5.2.65)

then, with
WX,1 := CX · M X X→X, (5.2.66)
it follows that for each f ∈ X the function Θ( f ) is well defined and μ-measurable
on X (see (5.2.51)) and
 
Θ( f ) X ≤ 2· sup Cw f X . (5.2.67)
w ∈ A1 (X,d,μ)
[w] A1 ≤WX,1
5.2 Extrapolation Theory 243

Before presenting the proof of this corollary, we make two comments. First, the
fact that the constant Cw appearing in (5.2.57) is allowed to be ∞ permits us to
obtain a version of Corollary 5.2.3 in which the demand in (5.2.57) is imposed for
a restrictive class of weights, namely only for those w ∈ Ap0 (X, d, μ) satisfying
[w] A p0 ≤ W where W ∈ (0, ∞] is some (a priori) fixed threshold. Specifically, in
the scenario in which (5.2.57) is known to hold only for this smaller category of
weights, we define

Cw for each w ∈ Ap0 (X, μ) with [w] A p0 ≤ W,

w :=
C (5.2.68)
∞ for each w ∈ Ap0 (X, μ) with [w] A p0 > W,

and note that (5.2.57) is satisfied for any weight w ∈ Ap0 (X, d, μ) provided Cw is
replaced by Cw (given that the latter is larger than the former). Wit this adjustment,
Corollary 5.2.3 applies and yields bounds like (5.2.61), (5.2.64), (5.2.67) written
with Cw in place of Cw . However, since the characteristic of the weight is restricted
in all aforementioned estimates, we may return to the original constant Cw in said
bounds provided the threshold W is sufficiently large, to begin with.
The second comment we wish to make in relation to Corollary 5.2.3  is that in
the case when the space of homogeneous type (X, d, μ) is actually Σ, | · − · |, σ
where Σ is a closed Ahlfors regular subset of Rn (for some n ∈ N with n ≥ 2) and
σ := H n−1 Σ, then (5.2.56) naturally takes effect if we start from the premise that
the operator Θ acts at the level
 σ(x) 
L 1 Σ, f −→ Θ( f ) ∈ M (Σ, σ). (5.2.69)
1 + |x| n−1
See [133, (7.7.104)] in this regard.
Here is the proof of Corollary 5.2.3.
Proof of Corollary 5.2.3 This is a direct consequence of the extrapolation results
from Theorem 5.2.1 and the embeddings established in Corollary 5.2.2. 
The results so far in this section highlight the importance of the boundedness of
the Hardy-Littlewood maximal operator on a given Generalized Banach Function
Space X and/or its associate space X. Our next proposition sheds light on the delicate
balance between these properties.
 
Proposition 5.2.4 Let X, d, μ be a space of homogeneous type with the property
that the quasi-distance d : X × X → [0, ∞) is continuous10 in the product topology
τd × τd . Fix an arbitrary exponent s ∈ (0, 1) and denote by M X and M X,s , respec-
tively, the Hardy-Littlewood maximal operator on X (cf. (A.0.71)) and its L s -based
version (defined as in (A.0.69)). Finally, suppose X is a Generalized Banach Func-
tion Space on (X, μ) and denote by X its associated space (cf. Definition 5.1.4 and
Definition 5.1.11). In this setting,
10 The result proved in [133, Theorem 7.1.2] guarantees that any quasi-metric space has an equivalent
quasi-distance satisfying this property
244 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

if M X : X → X is a well-defined bounded mapping

(5.2.70)
then M X,s : X → X is a well-defined bounded mapping,

and
if M X : X → X is a well-defined bounded mapping
(5.2.71)
then M X,s : X → X is a well-defined bounded mapping.

It is straightforward to check that for any sigma-finite measure space (X, M, μ)

both L 1 (X, μ) and L ∞ (X, μ) are Generalized Banach Function spaces for which
 1   
L (X, μ) = L ∞ (X, μ) and L ∞ (X, μ) = L 1 (X, μ) (5.2.72)
 
(see, e.g., [14, Theorem 2.5, p. 10]). Whenever X, d, μ happens to be a space of
homogeneous type with the property that the quasi-distance d : X × X → [0, ∞) is
continuous, we know that M X is a well-defined bounded mapping on L ∞ (X, μ) but
(in general) M X,s maps L 1 (X, μ) boundedly into itself only if s < 1. This explains
the restriction of the exponent s to the interval (0, 1) and highlights the optimality of
Proposition 5.2.4.
Proof of Proposition 5.2.4 Assume first that M X is bounded on X, and pick an
arbitrary function f ∈ X. In addition, suppose the functions g ∈ X has g X ≤ 1.
Then M X,s f : X → [0, ∞] is a well-defined μ-measurable function, and the version
of Fefferman-Stein’s maximal inequality from [133, Proposition 7.6.7] ensures the
existence of a constant C = C(μ, d, s) ∈ (0, ∞) such that
∫ ∫
   
M X,s f |g| dμ ≤ C | f | M X g dμ ≤ C f X M X g X
X X

≤ C MX X →X f X g X ≤C f X, (5.2.73)

where we have also employed the Hölder type inequality established in Propo-
sition 5.1.12. From this, (5.1.72), and Definition 5.1.4 we then conclude that
M X,s f ∈ X and M X,s f X ≤ C f X . In view of the arbitrariness of f , this
shows that M X,s induces a well-defined sublinear bounded mapping from X into
itself. The claim in (5.2.70) is therefore established. Finally, (5.2.71) follows from
(5.2.70) written for X in place of X, keeping in mind the fact that X = X (see
Proposition 5.1.14). 

It is also of interest to note that if a Generalized Banach Function Space, con-

sidered over a space of homogeneous type, is invariant under the action of the
Hardy-Littlewood maximal operator then the Generalized Banach Function Space
enjoys some useful embedding properties, of the sort discussed in the lemma below.
5.2 Extrapolation Theory 245
 
Lemma 5.2.5 Assume X, d, μ is a space of homogeneous type with the property
that the quasi-distance d : X × X → [0, ∞) is continuous11 in the product topology
τd × τd . Denote by Lloc
1 (X, μ) the space of μ-measurable functions on X which are
∞ (X, μ) the space of μ-
absolutely integrable on any d-ball in X, and denote by Lcomp
measurable functions on X which are essentially bounded and vanish μ-a.e. outside
of some d-ball in X. Also, bring in the Hardy-Littlewood maximal operator M X on X
(cf. (A.0.71)). Finally, suppose X is a Generalized Banach Function Space on (X, μ)
and denote by X its associated space (cf. Definition 5.1.4 and Definition 5.1.11). In
relation to these, make the assumption that

either M X (X) ⊆ X, or M X (X) ⊆ X . (5.2.74)

Then
∞ ∞
Lcomp (X, μ) ⊆ X ⊆ Lloc
1
(X, μ) and Lcomp (X, μ) ⊆ X ⊆ Lloc
1
(X, μ). (5.2.75)

In addition12,

whenever μ(X) < ∞ one has the continuous embeddings

(5.2.76)
L ∞ (X, μ) → X → L 1 (X, μ) and L ∞ (X, μ) → X → L 1 (X, μ).

Proof To fix ideas, assume first that M X (X) ⊆ X. Pick an arbitrary f ∈ X. Then

M X f ∈ X, so (5.1.10) implies that there
⨏ exists xo ∈ X such that (M X f )(xo ) < ∞. In
view of (A.0.71), this further entails B (x ,r) | f | dμ < ∞ for each r ∈ (0, ∞), which
d o
ultimately shows that f ∈ Lloc 1 (X, μ). The inclusion X ⊆ L 1 (X, μ) has therefore
loc
been established.
Moving on, recall from Definition 5.1.1 and Definition 5.1.4 that there exists a
collection {Yj } j ∈N of μ-measurable subsets sets of X with the property that we have

∞
X = Yj and 1Yj ∈ X for each j ∈ N. Pick j∗ ∈ N such that μ(Yj∗ ) > 0. Fix an
j=1
arbitrary point x∗ ∈ X. Then there exists r∗ ∈ (0, ∞) such that
 
0 < μ Yj∗ ∩ Bd (x∗, r∗ ) < ∞. (5.2.77)

Fix an arbitrary real number R ≥ r∗ . Then there exists a constant C = C(d) ∈ (0, ∞)
for which we have

Bd (x, CR) ⊆ Bd (x∗, R) for each x ∈ Bd (x∗, R). (5.2.78)

Consequently, with Cμ ∈ (0, ∞) and Dμ ∈ [0, ∞) denoting the doubling constant and
doubling order of μ (see [133, (7.7.4)]), for each x ∈ Bd (x∗, R) we may estimate

11 from [133, Theorem 7.1.2] it is known that any quasi-metric space has an equivalent quasi-
distance satisfying this property
 
12 recall that for a space of homogeneous type X, d, μ one has μ(X) < ∞ if and only if X is
d-bounded; see, e.g., [9, Proposition 2.12(7), pp. 49-50]
246 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
 ∫
  1
M X 1Yj∗ (x) ≥   1Yj∗ dμ
μ Bd (x, CR) B d (x,C R)
∫
1
≥   1Y dμ
μ Bd (x, CR) B d (x∗,R) j∗
 
μ Bd (x∗, R)  
=   · μ Yj∗ ∩ Bd (x∗, R)
μ Bd (x, CR)
1  R  Dμ  
≥ · μ Yj∗ ∩ Bd (x∗, r∗ )
Cμ CR
1  1  Dμ  
≥ c := · μ Yj∗ ∩ Bd (x∗, r∗ ) > 0. (5.2.79)
Cμ C
 
This shows that M X 1Yj∗ ≥ c · 1B d (x∗,R) with c ∈ (0, ∞), hence 1B d (x∗,R) ∈ X for
each real number R ≥ r∗ , by the lattice property enjoyed by X (cf. (5.1.12)). Since
for any R ∈ (0, r∗ ) we have 1B d (x∗,r∗ ) ≥ 1B d (x∗,R) , the lattice property just mentioned
also implies that 1B d (x∗,R) ∈ X for each R ∈ (0, r∗ ). All together, 1B d (x∗,R) ∈ X for
each x∗ ∈ X and each R ∈ (0, ∞), from which we then conclude that Lcomp ∞ (X, μ) ⊆ X

by one last application of the lattice property for the vector space X.
Moving on, given any g ∈ X, Proposition 5.1.12 guarantees that f g is absolutely
integrable on X for each f ∈ X. In particular, in view of what we have just proved,
1B d (x,R) · g is an absolutely integrable function on X for each x ∈ X and each
R ∈ (0, ∞). Thus, g ∈ Lloc 1 (X, μ), which goes to show that X  ⊆ L 1 (X, μ). To
loc
finish the proof of (5.2.75) in the current case, there remains to notice that (5.1.35)-
(5.1.36) and the fact that X ⊆ Lloc 1 (X, μ) guarantee that we also have the inclusion
∞
Lcomp (X, μ) ⊆ X . 

Going further, if in (5.2.74) we now assume that M X (X) ⊆ X, then all desired
conclusions in (5.2.75) follow from what we have proved so far applied to the
Generalized Banach Function Space X (in place of X) and Proposition 5.1.14.
Finally, to deal with (5.2.76), assume μ(X) < ∞ and recall (see., e.g., [9]) that
this is equivalent to having X bounded (relative to the quasi-distance d). Bearing
this in mind, (5.2.75) presently implies X ⊆ L 1 (X, μ) and X ⊆ L 1 (X, μ). In view of
Proposition 5.1.17, these set-theoretic inclusions self-improve to continuous embed-
dings X → L 1 (X, μ) and X → L 1 (X, μ). Having established these, we may now
rely on (5.1.7) together with Corollary 5.1.15 to conclude that L ∞ (X, μ) → X, and
invoke (5.1.35)-(5.1.36) to also obtain that L ∞ (X, μ) → X. 

Our next definition brings forth a distinguished linear subspace of a Generalized

Banach Function Space, which is going to play a significant role in future endeavors.
Indeed, this offers a natural functional analytic environment in which the Hardy-
Littlewood maximal operator as well as singular integral operators are well behaved.
5.2 Extrapolation Theory 247
 
Definition 5.2.6 Suppose X, d, μ is a space of homogeneous type with the property
that the quasi-distance d : X × X → [0, ∞) is continuous13 in the product topology
τd × τd . Denote by M X the Hardy-Littlewood maximal operator on X (cf. (A.0.71)).
Also, assume X is a Generalized Banach Function Space on (X, μ) with the property
that
either M X (X) ⊆ X, or M X (X) ⊆ X, (5.2.80)
where X is its associated space (cf. Definition 5.1.4 and Definition 5.1.11). In this
setting, recall that (5.2.75) holds, soit is meaningful
 to define X̊ as being the closure
∞ (X, μ) in the Banach space X, ·
of Lcomp X , i.e.,

∞ (X, μ) · X
X̊ := Lcomp . (5.2.81)

Thus, by design,
 
X̊ is a closed linear subspace of X, hence X̊, · X becomes a Banach
∞ (X, μ) is a dense subspace. (5.2.82)
space in which Lcomp

In addition, from (5.2.75) we see that

∞
Lcomp (X, μ) ⊆ X̊ ⊆ X ⊆ Lloc
1
(X, μ), (5.2.83)

while from (5.2.82) and (5.2.76) we conclude that

· X
if μ(X) < ∞ ∞
 then X̊ = L ∞(X, μ) , hence L ∞ (X, μ) is a dense
subspace of X̊, · X and L (X, μ) → X̊ → X → L 1 (X, μ) contin- (5.2.84)
uously in this case.
Finally, from (5.1.13), (5.2.81), and Lemma 1.2.20 we see that
the operator of pointwise multiplication by some given function b in
L ∞ (X, μ) is a bounded mapping from the space X̊ into itself, with (5.2.85)
operator norm ≤ b L ∞ (X,μ) ,

which, in turn, readily implies the lattice property for X̊, to the effect that
if f , g : X → R are two μ-measurable functions such that
|g| ≤ | f | at μ-a.e. point on X and f ∈ X̊, then g also (5.2.86)
belongs to the space X̊.
In the case when the underlying space of homogeneous type is an Ahlfors regu-
lar set equipped with its “surface” measure (and the standard Euclidean distance),
we may combine the abstract embedding results from Corollary 5.2.2 with the em-
bedding results involving weighted Lebesgue spaces from [133, Lemma 7.7.13] to

13 from [133, Theorem 7.1.2] it is known that any quasi-metric space has an equivalent quasi-
distance satisfying this property
248 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

establish useful embedding properties, like the ones formulated in the proposition
below.

Proposition 5.2.7 Let Σ ⊆ Rn (n ∈ N with n ≥ 2) be a closed set which is Ahlfors

regular and abbreviate σ := H n−1 Σ. Also, denote by M Σ the Hardy-Littlewood
 
maximal operator associated with the space of homogeneous type Σ, | · − · |, σ .
Finally, consider a Generalized Banach Function Space X on (Σ, σ) and denote
by X its associated space (cf. Definition 5.1.4 and Definition 5.1.11). Then, with
M (Σ, σ) standing for the space of σ-measurable functions on Σ,

whenever M Σ (X) ⊆ X one has the continuous embeddings

 
1+ |x | n−1
f ∈ M (Σ, σ) : ess-supx ∈Σ 1+log |x | | f (x)| < ∞ → X (5.2.87)
+
   
(1+log+ |x |)σ(x)

as well as X → L Σ, 1+ |x | n−1
1 → L 1 Σ, 1+σ(x)
|x | n−1
,

and
whenever M Σ (X) ⊆ X one has the continuous embeddings
 
1+ |x | n−1
g ∈ M (Σ, σ) : ess-supx ∈Σ 1+log |g(x)| < ∞ → X (5.2.88)
+ |x |
   
(1+log |x |)σ(x)
as well as X → L 1 Σ, 1+ +|x | n−1 → L 1 Σ, 1+σ(x)
|x | n−1 .

Furthermore, if
M Σ : X → X and M Σ : X → X
(5.2.89)
are well-defined bounded mappings
then one can find an exponent p∗ = p∗ (Σ, X) ∈ (1, ∞) with the property that

for each q ∈ (0, p∗ ) there exists ε = ε(Σ, X, q) ∈ (0, 1) such that

 σ(x)   σ(x)  (5.2.90)

X → L q Σ, and X → L q
Σ, .
1 + |x| n−1−ε 1 + |x| n−1−ε

Moreover, under the assumptions made in (5.2.89) and with p∗ = p∗ (Σ, X) ∈ (1, ∞)
as above, whenever q ∈ (1, p∗ ) and ε = ε(Σ, X, q) ∈ (0, 1) is defined as before, if
q  := (1 − 1/q)−1 ∈ (1, ∞) is the Hölder conjugate exponent of q, it follows that one
also has

 

L q Σ, (1 + |x| n−1−ε )q −1 σ(x) → X and

 
 (5.2.91)
L q Σ, (1 + |x| n−1−ε )q −1 σ(x) → X .

As a corollary,
whenever the Ahlfors regular set Σ is compact and (5.2.89) holds,

there exists q ∈ (1, 2) such that L q (Σ, σ) → X → L q (Σ, σ) and (5.2.92)

L q (Σ, σ) → X → L q (Σ, σ) continuously, with q  := (1 − 1/q)−1 .
5.2 Extrapolation Theory 249

Proof To deal with (5.2.87), assume M Σ (X) ⊆ X. Having fixed some x0 ∈ Σ,

Lemma 5.2.5 then guarantees that 1Δ(x0,1) ∈ X, hence also M Σ (1Δ(x0,1) ) ∈ X and,
further, (M Σ ◦ M Σ )(1Δ(x0,1) ) ∈ X. Then, on account of [133, (7.6.69)] and Proposi-
tion 5.1.12, for each function g ∈ X we may estimate
∫ ∫
1 + log+ |x| 1 + log+ |x − x0 |
|g(x)| dσ(x) ≈ |g(x)| dσ(x)
Σ 1 + |x| n−1
Σ 1 + |x − x0 | n−1
∫
 
≈ |g(x)| (M Σ ◦ M Σ ) 1Δ(x0,1) (x) dσ(x)
Σ

≤ CΣ,x0 (M Σ ◦ M Σ )(1Δ(x0,1) ) X
g X

=C g X (5.2.93)

where
C := CΣ,x0 · (M Σ ◦ M Σ )(1Δ(x0,1) ) X
∈ (0, ∞). (5.2.94)
This establishes the second embedding in (5.2.87). With this in hand, the first
embedding in (5.2.87) then follows on account of Corollary 5.1.15. At this stage,
(5.2.87) has been fully justified and (5.2.88) follows from it and Proposition 5.1.14.
Moving on, work under the assumptions made in (5.2.89). From these, the first
inclusion in (5.2.54) (written for some arbitrary fixed exponent p0 ∈ (1, ∞)), and
[133, (7.7.103)] we then conclude that, with q and ε as in the statement, we have the
set theoretic inclusion  
σ(x)
X ⊆ L q Σ, . (5.2.95)
1 + |x| n−1−ε
Granted this, we may invoke  Proposition  5.1.17 with (V, · ) taken to be the
weighted Lebesgue space L q Σ, 1+ |xσ(x)
| n−1−ε
equipped with its canonical norm and, on
account of (5.1.77), conclude that we have the first continuous embedding specified
in (5.2.90).
Parenthetically, we note that while (5.1.77) only ensures that the operator norm
of the inclusion (5.2.90) is finite, we can actually estimate said operator norm more
concretely. To be specific, fix some point x0 ∈ Σ and recall from Lemma 5.2.5 that
we presently have 1Δ(x0,1) ∈ X. Pick an arbitrary f ∈ X and, with q and ε as before,
consider the functions
∫ | f (x)| q  1/q
F := dσ(x) · 1Δ(x0,1) and G := f . (5.2.96)
Σ 1 + |x|
n−1−ε

Then, thanks to [133, (7.7.113)] (written with ε in place of θ), the extrapolation
result established in Theorem 5.2.1 applies to the pair (F, G) and ultimately provides
a constant C ∈ (0, ∞) which depends only on X and Σ for which
∫ | f (x)| q  1/q
dσ(x) · 1Δ(x0,1) X ≤ C f X . (5.2.97)
Σ 1 + |x|
n−1−ε
250 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
−1
Hence the operator norm of the first inclusion in (5.2.90) is ≤ C 1Δ(x0,1) X .
Finally, the second embedding claimed in (5.2.90) is a consequence of the first
embedding specified in (5.2.90) written for the Generalized Banach Function Space
X (in place of X) and Proposition 5.1.14. 

Other basic properties of the space X̊, introduced in relation to a given Generalized
Banach Function Space X as in Definition 5.2.6, are described in the next lemma.

Lemma 5.2.8 Let Σ ⊆ Rn (n ∈ N with n ≥ 2) be a closed set which is Ahlfors

regular and abbreviate σ := H n−1 Σ. Also, denote by M Σ the Hardy-Littlewood
 
maximal operator associated with the space of homogeneous type Σ, | · − · |, σ .
Consider a Generalized Banach Function Space X on (Σ, σ) with the property that

M Σ : X → X and M Σ : X → X
(5.2.98)
are well-defined bounded mappings,

where X is its associated space (cf. Definition 5.1.4 and Definition 5.1.11). Finally,
bring in the exponent p∗ = p∗ (Σ, X) ∈ (1, ∞) from Proposition 5.2.7.
Then for each q ∈ (1, p∗ ), if q := (1 − 1/q )−1 ∈ (1, ∞) is its Hölder conjugate
exponent it follows that X̊ originally introduced in Definition 5.2.6 (with X := Σ and
q
μ := σ) may bealternatively described as the closure of Lcomp (Σ, σ) in the Banach
space X, · X , i.e.,
· X
q
X̊ = Lcomp (Σ, σ) , (5.2.99)
 
as well as the closure of Lipc (Σ) in the Banach space X, · X , i.e.,
· X
X̊ = Lipc (Σ) . (5.2.100)

In particular, (5.2.99), (5.2.100), and (5.2.92) imply that

whenever the Ahlfors regular set Σ is compact and (5.2.89) holds, it

follows that X̊ is the closure of L q (Σ, σ) in the Banach space X, · X ,
 (5.2.101)
L q (Σ, σ) → X̊ → X → L q (Σ, σ) continuously, and Lipc (Σ) ⊆ X̊
densely.
Furthermore,
M Σ : X̊ −→ X̊
(5.2.102)
is a well-defined bounded mapping.

∞ (Σ, σ) ⊆ L q
Proof Staring with Lcomp comp (Σ, σ) and taking closures in X yields, on
account of (5.2.81),
· X
q
X̊ ⊆ Lcomp (Σ, σ) . (5.2.103)
We next claim that
q
Lcomp (Σ, σ) ⊆ X̊. (5.2.104)
5.2 Extrapolation Theory 251

To prove this, let the number ε = ε(Σ, X, q) ∈ (0, 1) be as in Proposition 5.2.7 and
q
pick an arbitrary function f ∈ Lcomp (Σ, σ). Then there exists a sequence {s j } j ∈N of

simple functions which converges to f in L q (Σ, σ) (see [133, (3.1.11)]), and matters
may be arranged so that each s j vanishes outside of the support of f . Since f is
compactly supported,  we conclude that the sequence
 {s j } j ∈N actually converges to
 
f in the space L q Σ, (1 + |x| n−1−ε )q −1 σ(x) . Together with (5.2.91), this proves
∞ (Σ, σ), we conclude
that {s j } j ∈N converges to f in X. Since each s j belongs to Lcomp
from this and (5.2.81) that f ∈ X̊. In view of the arbitrariness of f , this establishes
(5.2.104).
In turn, from (5.2.104) and the fact that X̊ is a closed subspace of X (cf. (5.2.82)),
we deduce that
· X
q
Lcomp (Σ, σ) ⊆ X̊. (5.2.105)
At this stage, (5.2.99) follows from (5.2.103) and (5.2.105). Also, (5.2.100) is a
consequence of (5.2.99), [133, (3.7.22)], and the continuity of the first embedding
in (5.2.91).
Let us now turn our attention to (5.2.102). Since M Σ maps L ∞ (Σ, σ) into itself,
it follows that  ∞ 
M Σ Lcomp (Σ, σ) ⊆ L ∞ (Σ, σ). (5.2.106)
∞ (Σ, σ). Then there exists
Fix a point x0 ∈ Σ, and pick an arbitrary function f ∈ Lcomp
some R ∈ (0, ∞) such that | f | ≤ R · 1Δ(x0,R) at σ-a.e. point in Σ. In concert with
[133, (7.6.66)], this permits us to write
     Rn
0 ≤ M Σ f (x) ≤ R · M Σ 1Δ(x0,R) (x) ≤ CΣ R ·
(R + |x − x0 |)n−1
CΣ,x0,R
≤ at σ-a.e. point x ∈ Σ. (5.2.107)
1 + |x − x0 | n−1
Also, (5.2.106) implies that
∞
g j := 1Δ(x0, j) · M Σ f ∈ Lcomp (Σ, σ) for each j ∈ N. (5.2.108)

For each large j ∈ N now estimate

252 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

MΣ f − g j X
= 1Σ\Δ(x0, j) · M Σ f X

1
≤ C 1Σ\Δ(x0, j) ·
1 + | · −x0 | n−1 X

1Σ\Δ(x0, j) 1 + log+ | · −x0 |

=C ·
1 + log+ | · −x0 | 1 + | · −x0 | n−1 X

C 1 + log+ | · −x0 |
≤
ln j 1 + | · −x0 | n−1 X

C
≤ (M Σ ◦ M Σ )(1Δ(x0,1) )
ln j X

C 2
≤ MΣ X→X
1Δ(x0,1) X
(5.2.109)
ln j
where the first equality comes from (5.2.108), the subsequent inequality is implied
by (5.2.107) (bearing in mind property (P2) in Definition 5.1.1), the second equality
is obvious, the penultimate inequality uses [133, (7.6.69)] and once again the mono-
tonicity of the norm in X, and the final inequality is a consequence of (5.2.98). In
view of the fact that M Σ X→X < ∞ thanks to (5.2.98), and that 1Δ(x0,1) X < ∞
thanks to (5.2.75), this proves that

g j → M Σ f in X as j → ∞. (5.2.110)

In concert with (5.2.108) and (5.2.81), this ultimately shows that M Σ f ∈ X̊. Hence,
 ∞ 
M Σ Lcomp (Σ, σ) ⊆ X̊. (5.2.111)

From (5.2.111), (5.2.98), and item (1) in Lemma 1.2.20 we then conclude that the
claim made in (5.2.102) is true. 

5.3 Young Functions and Orlicz Spaces

Orlicz spaces are natural examples of classical Banach function spaces (cf. Re-
mark 5.1.5). Moreover, it is possible to discern when the Hardy-Littlewood operator
is bounded on an Orlicz space constructed in relation to a given space of homogenous
type, based on the nature of the dilation indices associated with the corresponding
Young function. Here we elaborate on these aspects. This ties up with the material
from §5.1 and §5.2. See Proposition 5.3.15 for a concrete result of such flavor. For
the benefit of the reader, we begin by reviewing the class of Young functions.
5.3 Young Functions and Orlicz Spaces 253

Call Φ a Young function14 provided

Φ : [0, ∞) → [0, ∞) is convex, does not vanish on (0, ∞), and

satisfies lim+ Φ(t)/t = 0 as well as lim Φ(t)/t = ∞. (5.3.1)
t→0 t→∞

There are other qualities implicit in (5.3.1). For one thing,

Φ(λt) ≤ λΦ(t) for each λ ∈ [0, 1] and t ∈ [0, ∞), (5.3.2)

which may be justified by writing

 
Φ(λt) = Φ λt + (1 − λ)0 ≤ λΦ(t) + (1 − λ)Φ(0) = λΦ(t)
(5.3.3)
for each λ ∈ [0, 1] and t ∈ [0, ∞).

In turn, (5.3.2) has several notable consequences. First, we may use (5.3.3) and the
fact that Φ is non-negative on [0, ∞) to obtain 0 ≤ Φ(λ) ≤ λΦ(1) for each λ ∈ [0, 1].
This implies that Φ vanishes continuously at 0. Upon recalling that any convex
function on an open interval is continuous (cf., e.g., [164, Theorem 3.2, p. 61]), we
then conclude that Φ is continuous on [0, ∞). We may also use (5.3.3) and the fact
that Φ is strictly positive on (0, ∞) to conclude that Φ(λt) < Φ(t) for each λ ∈ (0, 1)
and t ∈ (0, ∞). In particular, Φ is strictly increasing on [0, ∞). Another consequence
of (5.3.3) is that Φ(λt)/(λt) ≤ Φ(t)/t for each λ ∈ (0, 1] and t ∈ (0, ∞), which readily
shows that
Φ(t)
the assignment (0, ∞) t→ ∈ (0, ∞) is non-decreasing. (5.3.4)
t
As such, the limits lim+ Φ(t)/t and lim Φ(t)/t are guaranteed to exist (in the interval
t→0 t→∞
[0, ∞]). Having lim+ Φ(t)/t = 0 implies, since Φ(0) = 0, that Φ is differentiable from
t→0
the right at 0, and Φ(0) = 0. Being a convex function makes Φ differentiable15 at
L 1 -a.e. point in (0, ∞) (cf., e.g., [160, Theorem 25.5, p. 246]). Bearing this in mind,
we conclude from (5.3.4) that
tΦ(t)
≥ 1 at any differentiability point t ∈ (0, ∞) for Φ,
Φ(t) (5.3.5)
hence at L 1 -a.e. point t in the interval (0, ∞).

It also turns out that Φ is super-additive, i.e., Φ(t1 ) + Φ(t2 ) ≤ Φ(t1 + t2 ) for each
t1, t2 ∈ [0, ∞). Indeed, this is clear if either t1 = 0 or t2 = 0, and in the case when
t1, t2 ∈ (0, ∞) we employ (5.3.2) to write

14 in the literature, the demands in the second line of (5.3.1) are not always imposed at full strength,
but even when they are replaced by weaker hypotheses the corresponding Orlicz spaces (cf. (5.3.27))
continue to enjoy a significant number of properties deduced in this section
15 In fact, any convex function is locally Lipschitz (cf., e.g., [158]), hence a.e. differentiable by the
classical Rademacher theorem. Aleksandov’s theorem also guarantees that Φ has a second-order
derivative at L 1 -a.e. point in (0, ∞) (see, e.g., [49, Theorem 1, p. 242]).
254 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

t1    t  
2
Φ(t1 ) + Φ(t2 ) = Φ (t1 + t2 ) + Φ (t1 + t2 )
t1 + t2 t1 + t2
 t   t 
1 2
≤ Φ(t1 + t2 ) + Φ(t1 + t2 ) = Φ(t1 + t2 ), (5.3.6)
t1 + t2 t1 + t2
as wanted. Finally, having the property lim Φ(t)/t = ∞ forces16 lim Φ(t) = ∞,
t→∞ t→∞
and since the function Φ has already been shown to be strictly increasing, it follows
that the function Φ : [0, ∞) → [0, ∞) is a bijection. Granted this, it makes sense
to consider its inverse Φ−1 : [0, ∞) → [0, ∞), itself a continuous strictly increasing
function, which vanishes at 0, and has lim Φ−1 (t) = ∞.
t→∞

Call a Young function Φ doubling, and indicate this by simply writing Φ ∈ Δ2 ,

if there exists a constant CΔ ∈ (0, ∞) with the property that

Φ(2t) ≤ CΔ Φ(t) for each t ∈ [0, ∞). (5.3.7)

For future use we note that Φ being strictly increasing forces CΔ > 1, hence

# := log2 CΔ ∈ (0, ∞). (5.3.8)

For the Young function Φ, the doubling property is intimately related with the
behavior of the function
Φ(st)
hΦ (t) := sup ∈ (0, ∞] for each t ∈ (0, ∞) (5.3.9)
s>0 Φ(s)

at infinity. Let us elaborate on this. First, it is immediate from its definition that
the function hΦ is non-decreasing (given that Φ is increasing), sub-
multiplicative (i.e., hΦ (t1 t2 ) ≤ hΦ (t1 )hΦ (t2 ) for all t1, t2 > 0), and (5.3.10)
hΦ (1) = 1; also, (5.3.2) implies hΦ (t) ≤ t for each t ∈ (0, 1].
 m
As a consequence, hΦ (t) ≤ hΦ (t 1/m ) for each t ∈ (0, ∞) and m ∈ N. Bearing in
mind that hΦ is non-decreasing, this further implies that

if hΦ (t∗ ) = ∞ for some t∗ ∈ (1, ∞) then

(5.3.11)
hΦ is identically equal to +∞ on the interval (1, ∞).

In relation to (5.3.11) we claim that, for any Young function Φ,

hΦ is not identically equal to +∞ on the interval (1, ∞) if and only if

Φ ∈ Δ2 if and only if there exists CΔ ∈ (1, ∞) such that hΦ (t) ≤ CΔ t # (5.3.12)
for each t ∈ (1, ∞) where # := log2 CΔ ∈ (0, ∞).

To justify this, assume first that Φ ∈ Δ2 . Pick some arbitrary t ∈ (1, ∞). Then there
exists a unique N ∈ N with 2 N −1 < t ≤ 2 N , which also entails N − 1 < log2 t ≤ N.

16 henceforth, we shall tacitly assume that Φ is extended to [0, ∞] with Φ(∞) = ∞

5.3 Young Functions and Orlicz Spaces 255

As such, given any s ∈ (0, ∞) we can iterate (5.3.7) and make use of (5.3.8) to write

Φ(ts) ≤ Φ(2 N s) ≤ (CΔ ) N Φ(s) = CΔ (CΔ ) N −1 Φ(s) < CΔ (CΔ )log2 t Φ(s)

= CΔ (2# )log2 t Φ(s) = CΔ t # Φ(s). (5.3.13)

Hence, hΦ (t) = sups>0 Φ(ts)

Φ(s) ≤ CΔ t for each t ∈ (1, ∞), as wanted. In the opposite
#

direction, if hΦ is not identically equal to +∞ on the interval (1, ∞) it follows

that hΦ is finite at any point in (1, ∞) (cf. (5.3.11)). As such, (5.3.7) holds with
CΔ := hΦ (2) < ∞, so Φ is doubling. This finishes the proof of (5.3.12).
The growth of a Young function Φ is encoded in its lower and upper dilation
indices defined, respectively, as

ln hΦ (t) ln hΦ (t)
i(Φ) := sup = lim+ , (5.3.14)
0<t<1 ln t t→0 ln t

ln hΦ (t) ln hΦ (t)
I(Φ) := inf = lim . (5.3.15)
1<t<∞ ln t t→∞ ln t
The second equalities in (5.3.14) and (5.3.15) are consequences of (5.3.10)-
(5.3.12) and [94,  Theorem 7.6.2, p. 244] applied to the sub-additive function
f (t) := ln hΦ (et ) for all t ∈ R (in this regard, see also [14, pp. 146-147]). We
also have
1 ≤ i(Φ) ≤ I(Φ) ≤ ∞. (5.3.16)
Indeed, the first inequality in (5.3.16) is a consequence of the last property in (5.3.10).
To justify the second inequality in (5.3.16), invoke (5.3.10) to write

1 = hΦ (1) ≤ hΦ (t)hΦ (1/t) for each t ∈ (0, ∞), (5.3.17)

then use this to estimate (bearing in mind that hΦ is strictly positive and finite on
(0, 1))
  
ln hΦ (1/t) ln 1 hΦ (1/t) ln hΦ (t)
= ≤ for each t ∈ (1, ∞). (5.3.18)
ln(1/t) ln t ln t
The second inequality in (5.3.16) now follows from (5.3.18) and (5.3.14)-(5.3.15)
by sending t → ∞.
Let us also record here that, as is apparent from (5.3.12) and (5.3.15), for any
Young function Φ we have

Φ ∈ Δ2 if and only if I(Φ) < ∞. (5.3.19)

Unraveling definitions in (5.3.14)-(5.3.15) readily shows that, given any Young

function Φ,
256 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

if 0 < α < i(Φ) ≤ I(Φ) < β < ∞ there exists Cα,β ∈ [1, ∞) so that
(5.3.20)
Φ(λt) ≤ Cα,β · max{λα, λ β } · Φ(t) for all λ ∈ (0, ∞) and t ∈ [0, ∞).

After replacing t by λ−1 t and subsequently re-denoting λ−1 by λ, from this we further
deduce that
if 0 < α < i(Φ) ≤ I(Φ) < β < ∞ there exists cα,β ∈ (0, 1] so that
(5.3.21)
cα,β · min{λα, λ β } · Φ(t) ≤ Φ(λt) for all λ ∈ (0, ∞) and t ∈ [0, ∞).

Forging ahead, suppose (X, M, μ) is a sigma-finite measure space. Also, fix a

Young function Φ. In this context, for each f ∈ M+ (X, μ) define
 ∫ 
Φ
ρ ( f ) := inf λ > 0 : Φ( f (x)/λ) dμ(x) ≤ 1 ∈ [0, +∞], (5.3.22)
X

with the convention that inf  := +∞. For each f ∈ M (X, μ) the quantity
 ∫ 
f L Φ (X,μ) := ρΦ (| f |) = inf λ > 0 : Φ(| f (x)|/λ) dμ(x) ≤ 1 ∈ [0, +∞]
X
(5.3.23)
is referred to as the Luxemburg norm of f . Note that we may equivalently express
the Luxemburg norm of any given function f ∈ M (X, μ) as
 ∫ ∞ 
  
f L Φ (X,μ) = inf λ > 0 : μ x ∈ X : | f (x)| > λΦ−1 (t) dt ≤ 1 . (5.3.24)
0

One can also see directly from definitions that for each f ∈ M (X, μ) we have
∫
 
Φ | f (x)|/ f L Φ (X,μ) dμ(x) ≤ 1 whenever 0 < f L Φ (X,μ) < ∞ (5.3.25)
X

and ∫
 
f L Φ (X,μ) ≤ 1 if and only if Φ | f (x)| dμ(x) ≤ 1. (5.3.26)
X

The Luxemburg norm is a genuine norm on the Orlicz space L Φ (X, μ), defined as
 
L Φ (X, μ) := f ∈ M (X, μ) : f L Φ (X,μ) <∞ . (5.3.27)
It turns out that the assignment ρΦ : M+ (X, μ) → [0, +∞] defined in (5.3.22) is a
function norm, in the sense of Definition 5.1.1. In fact, it may be easily checked from
(5.3.22) that for each E ∈ M with μ(E) < ∞ we have
  
Φ
1 Φ−1 1/μ(E) if μ(E) > 0,
ρ (1E ) = (5.3.28)
0 if μ(E) = 0.

In particular, ρΦ (1E ) < ∞ whenever E ∈ M has μ(E) < ∞, hence

5.3 Young Functions and Orlicz Spaces 257

1E ∈ L Φ (X, μ) for each E ∈ M with μ(E) < ∞. (5.3.29)

Also,∫ for each set E ∈ M with μ(E) < ∞ there exists a constant CE ∈ (0, ∞) such
that E f dμ ≤ CE ρΦ ( f ) for each function f ∈ M+ (X, μ). As a consequence,

for any given set E ∈ M with μ(E) < ∞, the restriction operator
(5.3.30)
L Φ (X, μ) f −→ f E ∈ L 1 (E, μ) is well defined and bounded.

For all these and related matters see, e.g., [14, pp. 268–271]. In particular,

the Orlicz space L Φ (X, μ) associated with any sigma-finite measure

space (X, M, μ) and any Young function Φ is a classical Banach func-
tion space (as introduced in [14, Definition 1.3, p. 3]; see Remark 5.1.5), (5.3.31)
hence also a Generalized Banach Function Space, in the sense of our
Definition 5.1.4.
If the sigma-finite measure space (X, M, μ) is non-atomic, recall that a (classic)
Banach function space X on (X, μ) is called rearrangement invariant if

for any two function f , g ∈ M (X, μ) one has f X = g X provided

    (5.3.32)
μ {x ∈ X : | f (x)| > λ} = μ {x ∈ X : |g(x)| > λ} for each λ ≥ 0.

See, e.g., [14, Definition 4.1, p. 59]. It is then apparent from this definition, (5.3.31),
and (5.3.24) that
if (X, M, μ) is a non-atomic sigma-finite measure space then the Orlicz
space L Φ (X, μ) associated with any Young function Φ is a rearrange- (5.3.33)
ment invariant Banach function space.
Let us also record here the well known fact that
given a sigma-finite measure space (X, M, μ) along with a Young
function Φ, the Orlicz space L Φ (X, μ) is reflexive if one has (5.3.34)
1 < i(Φ) ≤ I(Φ) < ∞.
Our next lemma contributes to the understanding of the Luxemburg norm of an
Orlicz space.

Lemma 5.3.1 Assume (X, M, μ) is a sigma-finite measure space, and pick some
Young function17 Φ ∈ Δ2 . In addition, suppose α, β ∈ R are two numbers with
the property that 0 < α < i(Φ) ≤ I(Φ) < β < ∞. Then there exist constants
cα,β ∈ (0, ∞) and Cα,β ∈ (0, ∞) such that for each function f ∈ M (X, μ) one has

17 the reader is reminded of the convention that Φ(∞) := ∞

258 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
  ∫
α β  
cα,β · min f L Φ (X,μ)
, f L Φ (X,μ)
≤ Φ | f (x)| dμ(x)
X
 
α β
≤ Cα,β · max f L Φ (X,μ)
, f L Φ (X,μ)
.
(5.3.35)

As a corollary, for each function f ∈ M (X, μ) one has

∫
 
f ∈ L Φ (X, μ) ⇐⇒ Φ | f (x)| dμ(x) < +∞, (5.3.36)
X

hence
 ∫ 
Φ  
L (X, μ) = f ∈ M (X, μ) : Φ | f (x)| dμ(x) < +∞ . (5.3.37)
X

In view of this result, given a sigma-finite measure space (X, M, μ) and some
Young function Φ, for each f ∈ M (X, μ) we henceforth agree to abbreviate
∫
 
N Φ ( f ) := Φ | f (x)| dμ(x) ∈ [0, +∞], (5.3.38)
X

and refer to it as the modular size of the function f . In terms of this symbol, we
may then recast (5.3.37) simply as
 
L Φ (X, μ) = f ∈ M (X, μ) : N Φ ( f ) < +∞ . (5.3.39)

So, while the assignment M (X, μ) f → N Φ ( f ) ∈ [0, +∞] does not give rise to a
conventional norm (as it lacks homogeneity and the triangle inequality), if Φ ∈ Δ2
it does satisfy weaker substitutes. For example, it is clear from (5.3.20) and (5.3.21)
that, given any Young function Φ ∈ Δ2 ,

whenever α, β ∈ R satisfy 0 < α < i(Φ) ≤ I(Φ) < β < ∞ it follows that
there exist two constants, cα,β ∈ (0, 1] and Cα,β ∈ [1, ∞), such that
(5.3.40)
cα,β · min{λα, λ β } · N Φ ( f ) ≤ N Φ (λ f ) ≤ Cα,β · max{λα, λ β } · N Φ ( f )
for each number λ ∈ (0, ∞) and each function f ∈ M (X, μ).

Let us now present the proof of Lemma 5.3.1.

Proof of Lemma 5.3.1 Fix an arbitrary function f ∈ M (X, μ). Let us justify the
second inequality in (5.3.35). This is clear if f L Φ (X,μ) = 0, and if f L Φ (X,μ) = ∞
there is nothing to prove. Therefore assume 0 < f L Φ (X,μ) < ∞, in which case
(5.3.25) and (5.3.21) permit us to estimate
5.3 Young Functions and Orlicz Spaces 259
  ∫  
−β
cα,β · min f L−αΦ (X,μ), f L Φ (X,μ) · Φ | f (x)| dμ(x)
X
∫
 
≤ Φ | f (x)|/ f L Φ (X,μ) dμ(x) ≤ 1.
X
(5.3.41)

The second inequality in (5.3.35) now readily follows from this.

Turning our attention
∫ to
 the first inequality in (5.3.35), we may assume that
0 < f L Φ (X,μ) and X Φ | f (x)| dμ(x) < ∞, since otherwise there is nothing to
prove. Since in view of (5.3.2) we may bound
∫ ∫
   
Φ λ| f (x)| dμ(x) ≤ λ Φ | f (x)| dμ(x) for each λ ∈ (0, 1), (5.3.42)
X X

the assumed finiteness ∫ property

  ensures that there exists some small number
λ ∈ (0, 1) such that X Φ λ| f (x)| dμ(x) ≤ 1. Together with (5.3.26), this implies
λ f L Φ (X,μ) ≤ 1 from which −1
 we further conclude that f L Φ (X,μ) ≤ λ < ∞. Then
for each ε ∈ 0, f L Φ (X,μ) we may rely on (5.3.26) to write
∫  
f | f (x)|
> 1 =⇒ Φ dμ(x) > 1, (5.3.43)
f L Φ (X,μ) − ε L Φ (X,μ) X f L Φ (X,μ) − ε
  −1
and then use (5.3.20) with λ := f L Φ (X,μ) −ε to deduce from the latter inequality
that
 ∫
 −α   −β   
Cα,β · max f L Φ (X,μ) − ε , f L Φ (X,μ) − ε · Φ | f (x)| dμ(x) > 1.
X
(5.3.44)
After sending ε  0 we eventually arrive at the first inequality in (5.3.35) after some
simple algebra. 
In turn, the above lemma is an ingredient in the proof of the following useful
density result.

Lemma 5.3.2 Suppose (X, M, μ) is a sigma-finite measure space, and select a Young
function Φ ∈ Δ2 . Then the space of simple functions Sfin (X, μ) defined in (A.0.107)
is dense in the Orlicz space L Φ (X, μ).

To offer an example, consider

Φ : [0, ∞) → [0, ∞), Φ(t) := t ln(1 + t) for each t ∈ [0, ∞). (5.3.45)

It may be verified without difficulty that this is a Young function satisfying

Φ(2t) Φ(2t)
lim = 2 and lim = 1, (5.3.46)
t→0+ Φ(t) t→∞ Φ(t)
260 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

hence Φ ∈ Δ2 (in fact i(Φ) = 1 and I(Φ) = 2; see (5.3.103)). Consequently, given
any sigma-finite measure space (X, M, μ), the corresponding Orlicz space L Φ (X, μ),
aka the Zygmund space
 ∫ 
L log L(X, μ) := f ∈ M (X, μ) : | f (x)|ln(1 + | f (x)|) dμ(x) < ∞ , (5.3.47)
X

has Sfin (X, μ) as a dense subspace.

Proof of Lemma 5.3.2 From (5.3.29) and (A.0.107) we see that Sfin (X, μ) is a sub-
space of L Φ (X, μ). To show that this subspace is actually dense in the space L Φ (X, μ),
pick two exponents α, β such that 0 < α < i(Φ) ≤ I(Φ) < β < ∞. Given an arbi-
trary function f ∈ L Φ (X, μ), by decomposing f = f+ − f− where f± := 12 (| f | ± f )
matters are reduced to the case when f is non-negative to begin with. Assuming
the latter, bring in the sequence {s j } j ∈N ⊂ S(X, μ) associated with the non-negative,
μ-measurable, function f on X as in [133, (3.1.12)]. In particular, each simple func-
tion s j is of the form i=1 N
ai 1 Ai , where N ∈ N, the ai ’s are strictly positive real
numbers, and the Ai ’s are mutually disjoint sets in M. Then, for each i, the fact that
0 ≤ ai 1 Ai ≤
 f at every point in X implies, in view of the monotonicity of Φ, that
Φ ai 1 Ai (x) ≤ Φ f (x) for each x ∈ X. As a consequence of this and the second
inequality in (5.3.35), for each i we may write
∫ ∫
   
Φ(ai ) · μ(Ai ) = Φ ai 1 Ai (x) dμ(x) ≤ Φ f (x) dμ(x)
X X
 
α β
≤ Cα,β · max f L Φ (X,μ)
, f L Φ (X,μ)
< ∞, (5.3.48)

with the last inequality above a consequence of the fact that f ∈ L Φ (X, μ). Since
Φ(ai ) ∈ (0, ∞), this forces μ(Ai ) < ∞. Ultimately, this shows that each function in
the sequence {s j } j ∈N actually belongs the space Sfin (X, μ) defined in (A.0.107).
Going further, [133, (3.1.12)] implies 0 ≤ f (x) − s j (x) ≤ f (x) for each x ∈ X
and each j ∈ N which,
 in viewof the monotonicity
 of the function Φ, leads to the
conclusion that Φ f (x) − s j (x) ≤ Φ f (x) for each point x ∈ X and each integer
j ∈ N. The function X x → Φ f (x) is, by the second inequality in (5.3.35), in
the space L 1 (X, μ).Also, thanks to [133, (3.1.12)] and the continuity of the function
Φ, we have lim Φ f (x) − s j (x) = Φ(0) = 0 at every point x ∈ X. Granted these
j→∞
properties, Lebesgue’s Dominated Convergence Theorem applies and gives
∫
 
lim Φ f (x) − s j (x) dμ(x) = 0. (5.3.49)
j→∞ X

In concert with the first inequality in (5.3.35), this proves

 
β
lim min f − s j LαΦ (X,μ), f − s j L Φ (X,μ) = 0, (5.3.50)
j→∞
5.3 Young Functions and Orlicz Spaces 261

which ultimately forces lim f − sj L Φ (X,μ) = 0. Now the desired conclusion fol-
j→∞
lows. 
Here is another noteworthy consequence of Lemma 5.3.1.
Lemma 5.3.3 Let (X, M, μ) be a sigma-finite measure space, and fix a Young function
Φ ∈ Δ2 . Suppose T is a homogeneous mapping18 from the Orlicz space L Φ (X, μ)
into M (X, μ). Then T is bounded on L Φ (X, μ), i.e., there exists a constant C ∈ (0, ∞)
such that

Tf L Φ (X,μ) ≤C f L Φ (X,μ) for each f ∈ L Φ (X, μ), (5.3.51)

if and only if whenever α, β are such that 0 < α < i(Φ) ≤ I(Φ) < β < ∞ there exists
a constant Cα,β ∈ (0, ∞) with the property that for each function f ∈ L Φ (X, μ) one
has
∫
 
Φ |T f | dμ (5.3.52)
X
 ∫   α/β ∫   β/α 
≤ Cα,β · max Φ | f | dμ , Φ | f | dμ .
X X

As a corollary, T is bounded on L Φ (X, μ) whenever the following modular in-

equality ∫ ∫
   
Φ |(T f )(x)| dμ(x) ≤ C Φ | f (x)| dμ(x) (5.3.53)
X X

holds for each function f ∈ L Φ (X, μ), where C ∈ (0, ∞) is a constant independent
of f .
Proof If (5.3.52) holds, then (5.3.35) shows that T maps the unit ball of L Φ (X, μ)
into a bounded set in L Φ (X, μ). In view of the fact that T is homogeneous, this proves
(5.3.51). Conversely, if (5.3.51) holds, then (5.3.35) readily yields (5.3.52). 
Pressing on, fix a Young function Φ. The complementary function of Φ is the
function
 : [0, ∞) −→ [0, ∞) given by
Φ
   (5.3.54)
 := sup{st − Φ(s)} = sup s t − Φ(s) for each t ≥ 0.
Φ(t) s
s>0 s>0

The last equality (5.3.54) together with the second line in (5.3.1) show (also bearing
in mind that Φ is continuous) that the supremum is attained and is a positive number

for each t > 0. Also, it is clear from the first equality (5.3.54) that Φ(0) = 0, In
addition, as the supremum of a family of linear (hence convex) functions, Φ  is itself
a convex function. It actually turns out that the complementary function Φ  is itself a
Young function, and
18 not necessarily linear
262 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces


 = Φ for each Young function Φ.
Φ (5.3.55)

See [160, Theorem 12.2, p. 104] for a more general result of this flavor, involving
generic convex functions in Rn . An immediate consequence of the definition given
in (5.3.54) is Young’s inequality, to the effect that
 for all s, t ∈ [0, ∞).
st ≤ Φ(s) + Φ(t) (5.3.56)

It is also known that

 −1 (t) ≤ 2t for each t ∈ [0, ∞),
t ≤ Φ−1 (t)Φ (5.3.57)

which further implies, after replacing t by Φ(t) and applying the strictly increasing

function Φ,  
Φ Φ(t)/t ≤ Φ(t) for each t ∈ (0, ∞). (5.3.58)
With “prime” denoting Hölder conjugation, the lower and upper dilation indices
satisfy (cf. [38, p. 71], [54], and the references therein)
 and I(Φ) = i(Φ).
i(Φ) = I(Φ)  (5.3.59)

In view of (5.3.19), this further implies

 ∈ Δ2 if and only if i(Φ) > 1.
Φ (5.3.60)

Thus, as a consequence of (5.3.19) and (5.3.60),

 ∈ Δ2 if and only if 1 < i(Φ) ≤ I(Φ) < ∞.
Φ, Φ (5.3.61)

Suppose (X, M, μ) is a sigma-finite measure space and fix a Young function Φ.

It has already been noted in (5.3.31) that the Orlicz space L Φ (X, μ) a classical
Banach function space (cf. Remark 5.1.5) hence, in particular, a Generalized Banach
Function Space. Then, it turns out that the associated space X of the Generalized
Banach Function Space X := L Φ (X, μ) (in the sense of Definition 5.1.11) is the Orlicz
 
space L Φ (X, μ) defined as in (5.3.27) in relation to the complementary function Φ
of the original Young function Φ. See [14, Corollary 8.15, p. 275]. In view of this,
Proposition 5.1.14, and (5.3.55), we therefore have

if X := L Φ (X, μ) then X = L Φ (X, μ),
(5.3.62)

and if X := L Φ (X, μ) then X = L Φ (X, μ).

Moreover, as a classical Banach function space (cf. Remark 5.1.5), X contains the
family of all essentially bounded functions vanishing outside sets of finite measure,
and we agree to denote by X̊ the closure of the latter family in X (compare with Defi-
nition 5.2.6 in the case of a Generalized Banach Function Space). Then Lemma 5.3.2
tells us that
5.3 Young Functions and Orlicz Spaces 263

if X := L Φ (X, μ) then X̊ = L Φ (X, μ) provided Φ ∈ Δ2 . (5.3.63)

Time to shift perspectives, and make the following definition.

Definition 5.3.4 (i) Call a function φ : (0, ∞) → R quasi-increasing if there

exists C ∈ (0, ∞) such that φ(t1 ) ≤ Cφ(t2 ) whenever 0 < t1 < t2 . Also, call a given
function φ : (0, ∞) → R quasi-decreasing if there exists C ∈ (0, ∞) such that
φ(t2 ) ≤ Cφ(t1 ) whenever 0 < t1 < t2 .
(ii) Call a function φ : (0, ∞) → R pseudo-increasing if there exist two
constants C1, C2 ∈ (0, ∞) such that φ(t1 ) ≤ C1 φ(C2 t2 ) whenever 0 < t1 < t2 . Finally,
call a function φ : (0, ∞) → R pseudo-decreasing if there exist two constants
C1, C2 ∈ (0, ∞) such that φ(t2 ) ≤ C1 φ(C2 t1 ) whenever 0 < t1 < t2 .

The next lemma sheds some light on the interplay between quasi/pseudo mono-
tonicity properties of power-modulated Young functions and their complementary
functions.

Lemma 5.3.5 Let p, p ∈ (1, ∞) be such that 1/p + 1/p = 1, and consider a Young
function Φ, along with its complementary function Φ.  Then t −p Φ(t) is pseudo-


increasing if and only if t Φ(t) is pseudo-decreasing, and t −p Φ(t) is pseudo-
−p

 is pseudo-increasing.
decreasing if and only if t −p Φ(t)

Proof Suppose t −p Φ(t) is pseudo-increasing, i.e., there exist C1, C2 ∈ (0, ∞) such
that
−p
t1 Φ(t1 ) ≤ C1 (C2 t2 )−p Φ(C2 t2 ) whenever 0 < t1 < t2 . (5.3.64)
  p −1
Fix 0 < t1 < t2 . Since 0 < s < t2 /t1 · s, we see from (5.3.64) that
  t  p −1    p
p t2
· s > C1−1 C2
2
Φ C2 · Φ(s) > 0. (5.3.65)
t1 t1
Based on this and (5.3.54) we may then write
!
 t  p −1   t  p −1 
 2 ) = sup{st2 − Φ(s)} = sup C2
Φ(t
2
· st2 − Φ C2
2
·s
s>0 s>0 t1 t1
!
 t  p  t  p
st1 − C1−1 C2
2 p 2
≤ sup C2 · · Φ(s)
s>0 t1 t1
!
 t  p
C1−1 C2
p 2 1−p
= · · sup C1 C2 · st1 − Φ(s)
t1 s>0

 t  p  
= C1−1 C2 ·
p 2  C1 C 1−p · t1 .
·Φ (5.3.66)
t1 2


In view of Definition 5.3.4, this tells us that the function t −p Φ(t) is pseudo-


decreasing.
264 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Assume next that t −p Φ(t) is pseudo-decreasing, i.e., there exist C1, C2 ∈ (0, ∞)
such that
−p
t2 Φ(t2 ) ≤ C1 (C2 t1 )−p Φ(C2 t1 ) whenever 0 < t1 < t2 . (5.3.67)
  p −1
Consider 0 < t1 < t2 . Given that 0 < t1 /t2 · s < 2, we conclude from (5.3.67)
that
  t  p −1  −p    p −1 
0 < s−p Φ(s) ≤ C1−1 C2
1
· s Φ C2 t1 /t2 ·s . (5.3.68)
t2
Use this and (5.3.54) to write
!
 t  p −1   t  p −1 
 1 ) = sup{st1 − Φ(s)} = sup C2
Φ(t
1
· st1 − Φ C2
1
·s
s>0 s>0 t2 t2
!
 t  p  t  p
st2 − C1−1 C2
1 p 1
≤ sup C2 · · Φ(s)
s>0 t2 t2
!
 t  p
C1−1 C2
p 1 1−p
= · · sup C1 C2 · st2 − Φ(s)
t2 s>0

 t  p  
= C1−1 C2 ·
p 1  C1 C 1−p · t2
·Φ (5.3.69)
t2 2


which, in light of Definition 5.3.4, implies that function t −p Φ(t)  is pseudo-
increasing.
At this stage, the equivalences claimed in the statement become consequences of
what we have proved and (5.3.55). 

The relevance of Definition 5.3.4 is most apparent in the context of the lemma
below, containing useful characterizations of the dilation indices of a Young function.

Lemma 5.3.6 For each Young function Φ one has

 
i(Φ) = sup p ∈ (0, ∞) : t −p Φ(t) is quasi-increasing
 
= sup p ∈ (0, ∞) : t −p Φ(t) is pseudo-increasing , (5.3.70)

and
 
I(Φ) = inf p ∈ (1, ∞) : t −p Φ(t) is pseudo-decreasing
 
= inf p ∈ (1, ∞) : t −p Φ(t) is quasi-decreasing , (5.3.71)

with the usual convention inf  = ∞ in place.

5.3 Young Functions and Orlicz Spaces 265

Proof Recall from the first inequality in (5.3.16) that 1 ≤ i(Φ), and pick a strictly
positive number p < i(Φ). In light of (5.3.14) this enures the existence of a number
t∗ ∈ (0, 1) with the property that

ln hΦ (t)
> p for each t ∈ (0, t∗ ). (5.3.72)
ln t
Since ln < 0 on (0, 1), this implies ln hΦ (t) < p ln t = ln(t p ). Hence, on the one
hand, hΦ (t) < t p for each t ∈ (0, t∗ ). On the other hand, for each t ∈ (t∗, 1) we may
−p
write hΦ (t) ≤ 1 < Ct p with C := t∗ ∈ (1, ∞). On account of (5.3.9), this analysis
shows that

Φ(st) < Ct p Φ(s) for each t ∈ (0, 1) and s ∈ (0, ∞). (5.3.73)

Suppose next that 0 < t1 < t2 . Since any Young function is strictly increasing,
(5.3.73) used with t := t1 /t2 ∈ (0, 1) and s := t2 ∈ (0, ∞) gives
 
Φ(t1 ) = Φ (t1 /t2 )t2 ≤ C(t1 /t2 ) p Φ(t2 ), (5.3.74)
−p −p
from which we deduce that t1 Φ(t1 ) ≤ Ct2 Φ(t2 ). According to Definition 5.3.4,
the function t −p Φ(t) is therefore quasi-increasing, hence p is less than, or equal to,
the first supremum in (5.3.70). Upon letting p i(Φ) we conclude that
 
i(Φ) ≤ sup p ∈ (0, ∞) : t −p Φ(t) is quasi-increasing
 
≤ sup p ∈ (0, ∞) : t −p Φ(t) is pseudo-increasing , (5.3.75)

where the last inequality is a simple consequence of the fact that any quasi-increasing
is also pseudo-increasing (cf. Definition 5.3.4).
Going further, from what we have proved
 earlier we know that any strictly positive

number p < i(Φ) belongs to the set p ∈ (0, ∞) : t −p Φ(t) is pseudo-increasing .
As such, the set in question is nonempty. Let us now pick an element in said set,
i.e., suppose p ∈ (0, ∞) is such that t −p Φ(t) is pseudo-increasing. Definition 5.3.4
guarantees the existence of two constants C1, C2 ∈ (0, ∞) with the property that
−p
t1 Φ(t1 ) ≤ C1 (C2 t2 )−p Φ(C2 t2 ) whenever 0 < t1 < t2 . To proceed, fix an arbitrary
t ∈ (0, 1). Then for each s ∈ (0, ∞) we have 0 < st < s so

(st)−p Φ(st) < C1 (C2 s)−p Φ(C2 s) for each s ∈ (0, ∞). (5.3.76)

In turn, this implies

Φ(st)
< C1 (C2 )−p t p for each s ∈ (0, ∞) (5.3.77)
Φ(C2 s)
which, after re-naming C2 s simply as s, become equivalent to
 
Φ s(t/C2 )
< C1 (t/C2 ) p for each s ∈ (0, ∞). (5.3.78)
Φ(s)
266 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

If we also re-name t/C2 ∈ (0, 1/C2 ) as t, we arrive at the conclusion that

Φ(st)
< C1 t p for each t ∈ (0, 1/C2 ) and each s ∈ (0, ∞). (5.3.79)
Φ(s)
From this and (5.3.9) we then obtain
Φ(st)
hΦ (t) = sup ≤ C1 t p for each t ∈ (0, 1/C2 ). (5.3.80)
s>0 Φ(s)

As such, ln hΦ (t) ≤ ln C1 + p ln t for all t > 0 small which, in concert with (5.3.9),
permits us to write

ln hΦ (t) C1
i(Φ) = lim+ ≥ p + lim+ = p. (5.3.81)
t→0 ln t t→0 ln t

Taking the supremum over all p’s satisfying the current working assumptions then
proves that
 
i(Φ) ≥ sup p ∈ (0, ∞) : t −p Φ(t) is pseudo-increasing . (5.3.82)

At this stage, (5.3.70) follows from (5.3.75) and (5.3.82).

Turing to the task of proving (5.3.71), first consider the question whether
 
I(Φ) ≤ inf p ∈ (1, ∞) : t −p Φ(t) is pseudo-decreasing . (5.3.83)

Since inf  = ∞, this is trivially true when there is no p ∈ (1, ∞) such that t −p Φ(t)
is quasi-decreasing. To treat the remaining case, suppose that p ∈ (1, ∞) is such

 is quasi-
that t −p Φ(t) is quasi-decreasing. Then Lemma (5.3.5) implies that t −p Φ(t)

increasing, where p ∈ (1, ∞) is the Hölder conjugate exponent of p and Φ  is the
complementary function of Φ. As such, from (5.3.70) we conclude that i(Φ)  ≥ p,
hence p ≥ i(Φ)   = I(Φ), with the last equality a consequence of (5.3.59) and (5.3.55).
With this in hand, (5.3.83) follows. Given that any quasi-decreasing function is also
pseudo-decreasing, we then have
 
I(Φ) ≤ inf p ∈ (1, ∞) : t −p Φ(t) is pseudo-decreasing
 
≤ inf p ∈ (1, ∞) : t −p Φ(t) is quasi-decreasing . (5.3.84)

Consider the issue whether

 
I(Φ) ≥ inf p ∈ (1, ∞) : t −p Φ(t) is quasi-decreasing . (5.3.85)

If I(Φ) = ∞ there is nothing to prove, so assume I(Φ) < ∞. Recall that we always
have I(Φ) ≥ 1 (cf. (5.3.16)). Pick a number p > I(Φ). In light of (5.3.15), this enures
the existence of some t∗ ∈ (1, ∞) such that

ln hΦ (t)
< p for each t ∈ (t∗, ∞), (5.3.86)
ln t
5.3 Young Functions and Orlicz Spaces 267

which implies that for each t ∈ (t∗, ∞) we have ln hΦ (t) < p ln t = ln(t p ), hence
hΦ (t) < t p for each t ∈ (t∗, ∞). Since hΦ is non-decreasing, therefore there exists
some C ∈ (0, ∞) such that hΦ (t) ≤ Ct p for each t ∈ (1, ∞). On account of (5.3.9),
this forces

Φ(st) ≤ Ct p Φ(s) for each t ∈ (1, ∞) and each s ∈ (0, ∞). (5.3.87)

Suppose next that 0 < t1 < t2 . Keeping in mind that any Young function is strictly
increasing, (5.3.87) used with t := t2 /t1 ∈ (1, ∞) and s := t1 ∈ (0, ∞) yields
 
Φ(t2 ) = Φ (t2 /t1 )t1 ≤ C(t2 /t1 ) p Φ(t1 ), (5.3.88)
−p −p
from which we conclude that t2 Φ(t2 ) ≤ Ct1 Φ(t1 ). The function t −p Φ(t) is there-
fore quasi-decreasing. Since we also know that p ∈ (1, ∞), we conclude that p is
greater than, or equal to, the infimum in (5.3.85). Sending p  I(Φ) then proves
(5.3.85). 

The characterizations of the dilation indices of a Young function given in

Lemma 5.3.6 are the key ingredient in the estimates contained in our next lemma,
itself a basic tool in the proof of the version of the Marcinkiewicz interpolation result
established a little later, in Theorem 5.3.16.

Lemma 5.3.7 Consider a Young function Φ. Then whenever 0 < p < i(Φ) one has

Φ(t)
lim+ =0 (5.3.89)
t→0 tp
and there exists a constant Cp ∈ (0, ∞) such that
∫ t
Φ(s) Φ(t)
p+1
ds ≤ Cp p for each t ∈ (0, ∞). (5.3.90)
0 s t

Moreover, whenever I(Φ) < q < ∞ one has

Φ(t)
lim = 0, (5.3.91)
t→∞ tq
and there exists a constant Cq ∈ (0, ∞) such that
∫ ∞
Φ(s) Φ(t)
q+1
ds ≤ Cq q for each t ∈ (0, ∞). (5.3.92)
t s t

Proof If 0 < p < i(Φ) then (5.3.70) guarantees the existence of some p∗ > p
with the property that the function t −p∗ Φ(t) is quasi-increasing. Using this, for each
t ∈ (0, ∞) we may estimate
268 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
∫ ∫
t
Φ(s) t
1 Φ(s)
ds = ds
0 s p+1 0 s1−(p∗ −p) s p∗
∫
Φ(t) t
1 Φ(t)
≤C ds = C , (5.3.93)
t p∗ 0 s1−(p∗ −p) tp
proving (5.3.90). In addition,

Φ(t)  Φ(t)
lim+ p
= lim + t p∗ −p p = 0, (5.3.94)
t→0 t (0,1) t→0 t ∗

since p∗ − p > 0 and 0 < Φ(t) t p∗ ≤ CΦ(1) < ∞ for each t ∈ (0, 1).
Next, if I(Φ) < q < ∞ then (5.3.71) ensures that there exists q∗ ∈ (1, q) such that
t −q Φ(t) is quasi-decreasing. Keeping this in mind, we may now write
∫ ∞ ∫ ∞
Φ(s) 1 Φ(s)
ds = )
ds
t s q+1
t s 1+(q−q∗ s q∗
∫
Φ(t) ∞ 1 Φ(t)
≤C q 1+(q−q )
ds = C q , (5.3.95)
t ∗ s ∗ t
t

which establishes (5.3.92). Finally,

Φ(t)  1 Φ(t)
lim q
= lim = 0, (5.3.96)
t→∞ t (1,∞) t→∞ t q−q∗ t q∗

Φ(t)
since q − q∗ > 0 and 0 < t q∗ ≤ CΦ(1) < ∞ for each t ∈ (1, ∞). 

Moving on, in regard to Young functions which are continuously differentiable

on (0, ∞), we know from [116, Theorem 5.1] that

for any Young function Φ of class C 1 on (0, ∞) one has

 ∈ Δ2 if and only if 1 < inf tΦ (t) 
(t) (5.3.97)
Φ, Φ ≤ sup tΦ
Φ(t) < ∞.
t>0 Φ(t) t>0

Moreover, given any Young function Φ of class C 1 on (0, ∞), it has been noted in
[54, Theorem 1.3, p. 435] that
 tΦ(t) tΦ(t) 
i(Φ) ≥ min lim inf , lim inf , (5.3.98)
t→0+ Φ(t) t→∞ Φ(t)
 tΦ(t) tΦ(t) 
i(Φ) ≤ min lim sup , lim sup , (5.3.99)
t→0+ Φ(t) t→∞ Φ(t)

and
5.3 Young Functions and Orlicz Spaces 269
 tΦ(t) tΦ(t) 
I(Φ) ≤ max lim sup , lim sup , (5.3.100)
t→0+ Φ(t) t→∞ Φ(t)
 tΦ(t) tΦ(t) 
I(Φ) ≥ max lim inf , lim inf . (5.3.101)
t→0+ Φ(t) t→∞ Φ(t)
In particular,

tΦ(t) tΦ(t)
if the limits r0 := lim+ and r∞ := lim exist (5.3.102)
t→0 Φ(t) t→∞ Φ(t)

then
i(Φ) = min(r0, r∞ ) and I(Φ) = max(r0, r∞ ). (5.3.103)
The differentiability condition imposed above on a Young function is not essential.
To elaborate on this aspect, let us introduce an equivalence relation in the class of
non-decreasing non-negative functions defined on [0, ∞) and which vanish at the
origin, by setting

Φ ∼ Ψ if and only if there exists c ∈ (0, 1]

(5.3.104)
such that Ψ(c t) ≤ Φ(t) ≤ Ψ(t/c) for each t ≥ 0.

Then, for one thing,

for two Young functions Φ, Ψ, having Φ ∼ Ψ implies that i(Φ) = i(Ψ),

I(Φ) = I(Ψ), and Φ ∼ Ψ; in addition, Φ ∈ Δ2 if and only if Ψ ∈ Δ2 ,
 ∈ Δ2 if and only if Ψ
while Φ  ∈ Δ2 ; also, CΦ ∼ Φ for any constant (5.3.105)
C ∈ (0, ∞) if Φ ∈ Δ2 (cf. (5.3.20)-(5.3.21)).
For another thing, from the discussion in (5.3.1)-(5.3.4) we see that if Φ is a Young
function then ∫ t
Φ(s)
Θ(t) := ds for each t ∈ [0, ∞) (5.3.106)
0 s
is a Young function which is continuously differentiable on the interval (0, ∞), and
is equivalent to Φ (i.e., Θ ∼ Φ), specifically

Φ(t/2) ≤ Θ(t) ≤ Φ(t) for each t ≥ 0. (5.3.107)

In addition, Θ is doubling if Φ is doubling, and has the property that its comple-
mentary function Θ is continuously differentiable on (0, ∞) and equivalent to Φ (i.e.,
 
Θ ∼ Φ). As a consequence of (5.3.105), we also have i(Θ) = i(Φ) plus I(Θ) = I(Φ),
 ∈ Δ2 if Φ
and Θ  ∈ Δ2 , while Φ
 ∈ Δ2 if and only if Φ
 ∈ Δ2 . Of course, this smoothing
procedure can be iterated, yielding successively more regular (equivalent) Young
functions.
In applications, given a function φ which may fail to be globally convex, it is of
interest to associate a genuine Young function Φ having roughly the same size at φ.
This is made precise in our next lemma.
270 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

Lemma 5.3.8 Suppose φ : [0, ∞) → [0, ∞) is a function satisfying the following

properties:
 
φ(0) = 0, φ > 0 on (0, ∞), φ ∈ C 0 (0, ∞) ,
near zero and near infinity (i.e., outside of a compact
(5.3.108)
subinterval of (0, ∞)) the function φ is of class C 2 and φ  > 0,
tφ (t)
and one has lim+ φ (t) = 0, lim φ (t) = ∞, lim inf φ(t) > 1.
t→0 t→∞ t→∞

Then there exists a Young function Φ which is continuously differentiable on

(0, ∞) and a constant C ∈ (0, ∞) such that

Φ(t) ≤ φ(t) ≤ Φ(Ct) for each t ∈ (0, ∞), and

Φ coincides with φ near zero and near infinity (5.3.109)
(i.e., outside of a compact subinterval of (0, ∞)).

Moreover, if one also assumes that

tφ (t) tφ (t)
lim sup < ∞ and lim sup < ∞, (5.3.110)
t→0+ φ(t) t→∞ φ(t)

then Φ ∈ Δ2 and, for any given sigma-finite measure space (X, M, μ), the Orlicz
space originally defined in relation to Φ as in (5.3.27) may alternatively described
(compare with (5.3.37)) as
 ∫ 
 
L Φ (X, μ) = f ∈ M (X, μ) : φ | f (x)| dμ(x) < +∞ . (5.3.111)
X

 ∈ Δ2 , if one
Finally, the complementary function of Φ is also doubling, i.e., Φ
also assumes that
tφ (t)
lim inf > 1. (5.3.112)
t→0+ φ(t)
In relation to the situation described in the above lemma, we shall henceforth
adopt the following:
Convention 5.3.9 Whenever φ : [0, ∞) → [0, ∞) is a function as described in
(5.3.108), we shall write Φ(t) ≈ φ(t) (or, simply, Φ ≈ φ) to indicate that Φ is a Young
function continuously differentiable on (0, ∞) which is related to φ as in (5.3.109).

Proof of Lemma 5.3.8 We begin by noting that for each number t > 0 sufficiently
small there exists some ξt ∈ (0, t) such that

φ(t) φ(t) − φ(0)

0< = = φ (ξt ) < φ (t), (5.3.113)
t t
thanks to the Mean Value Theorem and the fact that the function φ  is increasing
near the origin (given that φ  > 0 there). In particular,
5.3 Young Functions and Orlicz Spaces 271

tφ (t)
> 1 for each t > 0 sufficiently small. (5.3.114)
φ(t)
We claim that the properties listed in (5.3.108) also imply

φ(t) φ(t)
lim = 0 and lim = ∞. (5.3.115)
t→0+ t t→∞ t

Indeed, the first formula in (5.3.115) follows from (5.3.113), the Squeeze Theorem,
and the fact that lim+ φ (t) = 0. Consider next the second formula in (5.3.115). The
t→0
fact that φ is convex near infinity implies that in that regime the graph of φ lies
above any of its tangent lines, i.e., there exists some T ∈ (0, ∞) with the property
that for each t∗ ≥ T we have φ(t) ≥ φ (t∗ )(t − t∗ ) + φ(t∗ ) at every t ≥ T. This forces
lim inf φ(t)  
t ≥ φ (t∗ ), and since lim φ (t∗ ) = ∞ the second formula in (5.3.115) is
t→∞ t∗ →∞
proved. In particular, (5.3.115) shows that

lim φ(t) = 0 and lim φ(t) = ∞. (5.3.116)

t→0+ t→∞

Going further, let ε ∈ (0, 1) be small enough such that the function φ is of class
C 2 and φ  > 0 on (0, ε) ∪ (ε −1, ∞). We may also assume that ε ∈ (0, 1) is sufficiently
small so that φ (t) > 0 for each t ∈ [ε −1, ∞). Given
 that φ is continuous on (0, ∞),
it follows that m := min φ(t) : ε ≤ t ≤ ε −1 is a well-defined strictly positive
number, Moreover, since lim+ φ (t) = 0 = lim+ φ(t) there exists some small number
t→0 t→0
t0 ∈ (0, ε) such that

m > φ (t0 )(ε −1 − t0 ) + φ(t0 ) and φ (ε −1 ) > φ (t0 ). (5.3.117)

For this choice, define Φ0 : [0, ∞) → [0, ∞) by setting, for each t ≥ 0,

φ(t) if t ∈ (0, t0 ],
Φ0 (t) := (5.3.118)
φ (t0 )(t − t0 ) + φ(t0 ) if t ∈ [t0, ∞).

Then
Φ0 convex and strictly increasing on [0, ∞), Φ0 > 0 on (0, ∞),
 
Φ0 ∈ C 1 (0, ∞) , Φ0 (t) ≤ φ(t) for each t ∈ [0, ∞), and Φ0 = φ on [0, t0 ].
(5.3.119)
Next, since the very last property in (5.3.108) entails
 φ(t)    φ(t) 
lim inf t −  = lim inf t 1 −  = ∞, (5.3.120)
t→∞ φ (t) t→∞ tφ (t)

we may find some large number t1 ∈ (ε −1, ∞) such that

φ(t1 )
t1 − > ε −1 . (5.3.121)
φ (t1 )
272 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

For this choice, define Φ1 : [0, ∞) → [0, ∞) by setting, for each t ≥ 0,

⎧
⎪ 0 if t ≤ t1 − φφ(t 1)
 (t ) ,
⎪
⎪
⎨
⎪ 1

 φ(t 1 )
Φ1 (t) := φ (t1 )(t − t1 ) + φ(t1 ) if t1 − φ (t ) ≤ t ≤ t1, (5.3.122)
⎪
⎪ 1
⎪
⎪ φ(t)
⎩ if t ≥ t1 .

Then
 
φ(t1 )
Φ1 is convex on [0, ∞), C 1 and strictly increasing on t1 − φ (t1 ) , ∞ ,
(5.3.123)
Φ1 (t) ≤ φ(t) for each t ∈ [0, ∞), and Φ1 = φ on [t1, ∞).

If we now define Ψ : [0, ∞) → [0, ∞) by setting

Ψ(t) := max{Φ0 (t), Φ1 (t)} for each t ≥ 0, (5.3.124)

then the function Ψ is convex and strictly increasing on [0, ∞), such that Ψ(t) ≤ φ(t)
for each t ∈ [0, ∞) and which actually coincides with φ outside the compact interval
[t0, t1 ]. In addition, the function Ψ is continuously differentiable everywhere on
(0, ∞) except at the point

t1 φ (t1 ) − t0 φ (t0 ) − φ(t1 ) + φ(t0 )  −1 

t# := ∈ ε ,∞ (5.3.125)
φ (t1 ) − φ (t0 )
where the two slanted lines from (5.3.118) and (5.3.122)
  actually meet. Finally,
slightly round off the graph of Ψ near the corner t#, Ψ(t# ) (using a small convex
arc tangent to both lines) to yield Φ : [0, ∞) → [0, ∞) with Ψ(t) ≤ Φ(t) ≤ φ(t)
for each t ∈ [0, ∞), such that Φ is strictly increasing, convex, and of class C 1 on
the entire interval (0, ∞), while continues to coincide with φ outside the compact
interval [t0, t1 ]. In particular, Φ(0) = φ(0) = 0 and lim Φ(t) = ∞ (cf. (5.3.116)). By
t→∞
design, Φ(t) ≥ Φ0 (t) > 0 for each t ∈ (0, ∞). We also claim that there exists some
C ∈ [1, ∞) such that

Φ(Ct) ≥ φ(t) for each t ∈ [0, ∞). (5.3.126)

Indeed, having C ≥ 1 implies, since Φ is increasing everywhere and agrees with

φ on [0, t0 ] ∪ [t1, ∞), that Φ(Ct) ≥ Φ(t) = φ(t) for each t ∈ [0, t0 ] ∪ [t1, ∞). To
deal with the interval [t0, t1 ], define M := max{φ(t) : t0 ≤ t ≤ t1 }, and notice that
M ∈ (0, ∞). To finish the proof of (5.3.126), it suffices to show that Φ(Ct) ≥ M for
each t ∈ [t0, t1 ] which, in view of the monotonicity of Φ, is implied by Φ(Ct0 ) ≥ M.
However, since lim Φ(t) = ∞, this may always be arranged by taking C ∈ [1, ∞)
t→∞
sufficiently large. With (5.3.126) in hand, all properties claimed in (5.3.109) have
been justified. In light of these and (5.3.115) we then readily conclude that

Φ(t) Φ(t)
lim = 0 and lim = ∞. (5.3.127)
t→0+ t t→∞ t
5.3 Young Functions and Orlicz Spaces 273

Hence, Φ is actually a Young function (cf. (5.3.1)).

For the remainder of the proof work under the additional assumption made in
(5.3.110). Based on this, (5.3.100), and the fact that Φ coincides with φ near zero
and near infinity we then estimate
 tΦ(t) tΦ(t) 
I(Φ) ≤ max lim sup , lim sup ,
t→0+ Φ(t) t→∞ Φ(t)
 tφ (t) tφ (t) 
= max lim sup , lim sup < ∞. (5.3.128)
t→0+ φ(t) t→∞ φ(t)

On account of (5.3.19), this permits us to conclude that Φ ∈ Δ2 . The description of

the Orlicz space given in (5.3.111) is a consequence of the double inequality in the
first line of (5.3.109), and (5.3.37) (bearing in mind that Φ is doubling).
Finally, the additional assumption in (5.3.112) ensures that
 tΦ(t) tΦ(t) 
i(Φ) ≥ min lim inf , lim inf ,
t→0+ Φ(t) t→∞ Φ(t)
 tφ (t) tφ (t) 
= min lim inf , lim inf > 1, (5.3.129)
t→0 + φ(t) t→∞ φ(t)
thanks to (5.3.98), the fact that Φ coincides with φ near zero and near infinity, and
the last property in (5.3.108). Once this has been established, (5.3.60) tells us that
the complementary function of Φ is doubling, i.e., Φ ∈ Δ2 . 

The usefulness of Lemma 5.3.8 becomes apparent in the discussion below, where
a certain variety of Orlicz spaces, called Zygmund spaces, are introduced.

Example 5.3.10 (The Zygmund Spaces L p (log L)α ) Fix an exponent p ∈ (1, ∞)
along with a power α ∈ R and define the function φ : [0, ∞) → [0, ∞) by setting

φ(t) := t p [ln(e + t)]α for each t ∈ [0, ∞). (5.3.130)

Then φ(0) = 0, the function φ is strictly positive and of class C ∞ on (0, ∞), and for
each t ∈ (0, ∞) we have
  t  1 
φ (t) = t p−1 [ln(e + t)]α p + α , (5.3.131)
e + t ln(e + t)
hence  t 
tφ (t) 1
= p+α . (5.3.132)
φ(t) e + t ln(e + t)
From these we conclude that

lim φ (t) = 0 and lim φ (t) = ∞, (5.3.133)

t→0+ t→∞

and
274 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

tφ (t) tφ (t)
lim+ = lim = p ∈ (1, ∞). (5.3.134)
t→0 φ(t) t→∞ φ(t)

In addition, for each t ∈ (0, ∞) we have

 t  1
φ (t) = t p−2 [ln(e + t)]α p(p − 1) + 2αp (5.3.135)
e + t ln(e + t)
!
 t 2 1  t 2 1
+ α(α − 1) −α .
e + t [ln(e + t)]2 e + t ln(e + t)

Hence
 
φ (t) = t p−2 [ln(e + t)]α p(p − 1) + o(1) as either t → 0+ or t → ∞, (5.3.136)

from which we see that φ  > 0 near zero and near infinity19. As such, all properties
demanded in (5.3.108) and (5.3.110) are satisfied. Granted these, Lemma 5.3.8
then guarantees the existence of a Young function Φ ∈ Δ2 as in (5.3.109), which is
continuously differentiable on (0, ∞), and whose complementary function is doubling
 ∈ Δ2 . In fact, i(Φ) = I(Φ) = p ∈ (1, ∞) (see (5.3.61)). Also, in
as well, i.e., Φ
accordance with Convention 5.3.9, we may write

Φ(t) ≈ t p [ln(e + t)]α, (5.3.137)

and it turns out that Φ,  the complementary function of Φ, has the property that
 
Φ(t) ≈ t [ln(e + t)]
p α(1−p )
where p ∈ (1, ∞) is the Hölder conjugate exponent of p
(cf. [38, p. 72]).
Given a sigma-finite measure space (X, M, μ), we agree to denote the Orlicz
space L Φ (X, μ), defined in relation to Φ as in (5.3.27), by L p (log L)α (X, μ). This is
typically referred to as Zygmund’s space (cf., e.g., [14, Example 8.3(e), p. 266]).
According to (5.3.111), we then have the following description of the Zygmund space
just introduced:
 ∫ 
 α
L p (log L)α (X, μ) = f ∈ M (X, μ) : | f (x)| p ln(e + | f (x)|) dμ(x) < +∞ .
X
(5.3.138)
In fact, with Convention 5.3.9 in place, similar considerations are valid in relation
to    β
Φ(t) ≈ t p [ln(e + t)]α ln ln(ee + t)
(5.3.139)
with p ∈ (1, ∞) and α, β ∈ R,
or, more generally, given any p ∈ (1, ∞), m ∈ N, and α1, . . . , αm ∈ R, in relation to

19 if α ≥ 2 − 2p then actually φ > 0 everywhere on (0, ∞), hence in this case φ is globally convex
and, ultimately, a Young function itself
5.3 Young Functions and Orlicz Spaces 275
   α
Φ(t) ≈ t p [ln(e + t)]α1 ln ln(ee + t) 2 × · · · ×
"   # αm
   ..
e

e.
× ln ln · · · ln ln(e + t) · · · , (5.3.140)

where the last expression involves m logarithms and e is m-fold exponentiated.

A related construction is described in the next example.

Example 5.3.11 (The Space L p exp (logθ L)) Pick an exponent p ∈ (1, ∞) together
with a number θ ∈ [0, 1) and define the function φ : [0, ∞) → [0, ∞) by setting
 
φ(t) := t p exp [ln(e + t)]θ for each t ∈ [0, ∞). (5.3.141)

Then φ(0) = 0, the function φ is strictly positive and of class C ∞ on (0, ∞), and for
each t ∈ (0, ∞) we have
   t  1 
φ (t) = t p−1 exp [ln(e + t)]θ p + θ , (5.3.142)
e + t [ln(e + t)]1−θ

which further implies

tφ (t)  t  1
= p+θ . (5.3.143)
φ(t) e + t [ln(e + t)]1−θ

It is then apparent from these formulas that

lim φ (t) = 0 and lim φ (t) = ∞, (5.3.144)

t→0+ t→∞

and
tφ (t) tφ (t)
lim+ = lim = p ∈ (1, ∞). (5.3.145)
t→0 φ(t) t→∞ φ(t)

Moreover, for each t ∈ (0, ∞) we have

   t  1
φ (t) = t p−2 exp [ln(e + t)]θ p(p − 1) + 2pθ
e + t [ln(e + t)]1−θ
 t 2 1
−θ ·
e+t [ln(e + t)]1−θ
 t 2 1
+ θ(θ − 1) ·
e+t [ln(e + t)]2−θ
!
 t 2 1
+ θ2 · . (5.3.146)
e+t [ln(e + t)]2(1−θ)

Thus,
276 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
  
φ (t) = t p−2 exp [ln(e + t)]θ p(p − 1) + o(1) as either t → 0+ or t → ∞,
(5.3.147)
so φ  > 0 near zero and near infinity. Consequently, all properties specified in
(5.3.108) and (5.3.110) are presently satisfied. Lemma 5.3.8 then guarantees the
existence of a Young function
 
Φ(t) ≈ t p exp [ln(e + t)]θ (5.3.148)

such that Φ, Φ ∈ Δ2 . In addition, i(Φ) = I(Φ) = p ∈ (1, ∞) (see (5.3.61)).

Given a sigma-finite measure space (X, M, μ), it is natural to denote the Orlicz
space L Φ (X, μ), defined in relation to Φ as in (5.3.27), by L p exp (logθ L)(X, μ). In
light of (5.3.111), we then have the following description of the space just introduced:

L p exp (logθ L)(X, μ) (5.3.149)

 ∫ 
 θ
= f ∈ M (X, μ) : | f (x)| p exp ln(e + | f (x)|) dμ(x) < ∞ .
X

In our last example we note that intersections and sums of Lebesgue spaces may
be construed as Orlicz spaces.

Example 5.3.12 Given any two Young functions Φ1, Φ2 , we see from definitions that
max{Φ1, Φ2 } and Φ1 + Φ2 are also Young functions, satisfying

max{Φ1, Φ2 } ≤ Φ1 + Φ2 ≤ 2 max{Φ1, Φ2 }. (5.3.150)

Moreover, the Young functions max{Φ1, Φ2 } and Φ1 + Φ2 are doubling and actually
equivalent (in the sense of (5.3.104)) if Φ1, Φ2 ∈ Δ2 . In such a scenario, for any
sigma-finite measure space (X, M, μ) it is apparent from (5.3.37) and (5.3.150) that

L max{Φ1,Φ2 } (X, μ) = L Φ1 +Φ2 (X, μ) = L Φ1 (X, μ) ∩ L Φ1 (X, μ). (5.3.151)

For example, having fixed two exponents p, q ∈ (1, ∞), then for the Young function
defined as
Φ(t) := max{t p, t q } for each t ∈ [0, ∞) (5.3.152)
we have
i(Φ) = min{p, q}, I(Φ) = max{p, q},
(5.3.153)
and L Φ (X, μ) = L p (X, μ) ∩ L q (X, μ),
with the first two formulas implied by (5.3.103). In this vein, let us also observe that
we may use (5.3.54) to identify the complementary function Φ  of Φ in (5.3.152) as a
Young function satisfying

 ≈ min{t p, t q } for each t ∈ [0, ∞),

Φ(t)
(5.3.154)
where p, q  ∈ (1, ∞) are the Hölder conjugates of p, q.
5.3 Young Functions and Orlicz Spaces 277

In general, for a Young function Ψ(t) ≈ min{t p, t q } with p, q ∈ (1, ∞) we have

i(Ψ) = min{p, q}, I(Ψ) = max{p, q},

(5.3.155)
and L Ψ (X, μ) = L p (X, μ) + L q (X, μ).

To justify the last equality above, suppose 1 < p∫≤ q < ∞. Then for  f in M (X, μ)
arbitrary we have f ∈ L Ψ (X, μ) if and only if X min | f | p, | f | q dμ < ∞ if and
∫ ∫ ∫
only if | f |<1 | f | q dμ + | f | ≥1 | f | p dμ < ∞ if and only if | f |<1 | f | q dμ < ∞ and
∫
| f | ≥1
| f | p dμ < ∞. Hence, any given f ∈ L Ψ (X, μ) may be decomposed as the sum
f = f 1 { | f |<1} + f 1 { | f | ≥1} with f 1 { | f |<1} ∈ L q (X, μ) and f 1 { | f | ≥1} ∈ L p (X, μ). This
proves∫ that L Ψ (X, μ) ∫⊆ L p (X, μ) + L q (X, μ).∫Conversely, given f ∈ L p (X, μ) we
have | f |<1 | f | q dμ ≤ | f |<1 | f | p dμ < ∞ and | f | ≥1 | f | p dμ < ∞, so f ∈ L Ψ (X, μ).
∫ ∫
Likewise, any given function f ∈ L q (X, μ) has | f | ≥1 | f | p dμ ≤ | f | ≥1 | f | q dμ < ∞
∫
and | f |<1 | f | q dμ < ∞, so f ∈ L Ψ (X, μ). Ultimately, this shows that in fact we have
L p (X, μ) + L q (X, μ) ⊆ L Ψ (X, μ) as wanted.
Lastly, from (5.3.153), (5.3.154), and (5.3.155) we see that for any integrability
exponents p, q, p, q  ∈ (1, ∞) with 1/p + 1/p = 1 and 1/q + 1/q  = 1 we have
 p   
L (X, μ) ∩ L q (X, μ) = L p (X, μ) + L q (X, μ)
   
(5.3.156)
and L p (X, μ) + L q (X, μ) = L p (X, μ) ∩ L q (X, μ).

Continuing on, recall that the lower and upper Boyd indices, 1 ≤ pX ≤ qX ≤ ∞,
of a rearrangement invariant Banach function space X on a non-atomic sigma-finite
measure space (X, M, μ) are computed in terms of the norms of certain dilation
operators (see, e.g., [14, Definition 5.12, p. 149], where actually what we consider
the reciprocals of said indices are defined, and also the discussion in [123, §8]). In
relation to these, we record an interpolation result proved in [123, §8], along the lines
of the classical Boyd’s interpolation theorem (cf., e.g., [14, Theorem 5.16, p. 153]),
allowing one to pass from Lorentz spaces a given measure space to estimates on
certain rearrangement invariant Banach function spaces (over said measure space)
whose Boyd indices are flanked by first integrability exponents involved in the
Lorentz spaces

Proposition 5.3.13 Assume (X, M) is a measurable space, and μ is a positive, non-

atomic, sigma-finite measure on the sigma-algebra M, with μ(X) > 0. Let X be a
rearrangement invariant Banach function space over (X, μ). Denote by pX , qX its
lower and upper Boyd indices and pick two integrability exponents p, q ∈ (0, ∞].
Make the assumption that

either 0 < p < pX ≤ qX < q < ∞, or

(5.3.157)
0 < p < pX and q = ∞.

Suppose
T : L p,1 (X, μ) + L q,1 (X, μ) −→ M (X, μ) (5.3.158)
278 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

is a quasi-subadditive operator20 such that

Tf L p,∞ (X,μ) ≤ Mp f L p,1 (X,μ) for every f ∈ L p,1 (X, μ), (5.3.159)

and21

Tf L q,∞ (X,μ) ≤ Mq f L q,1 (X,μ) for every f ∈ L q,1 (X, μ), (5.3.160)

for some Mp, Mq ∈ (0, ∞). Then

X ⊆ L p,1 (X, μ) + L q,1 (X, μ) (5.3.161)

and there exists some C ∈ (0, ∞), which depends only on p, q, and the quasi-triangle
inequality constant for T, such that

Tf X ≤ C (Mp + Mq ) f X for every f ∈ X. (5.3.162)

The proposition below singles out a distinguished subcategory of rearrangement

invariant Banach function space X, considered in relation to the underlying measure
space of a given space of homogeneous type, for which the corresponding Hardy-
Littlewood maximal operator is bounded both on X and on its associated space
X.
 
Proposition 5.3.14 Assume X, d, μ is a space of homogeneous type with the
property that the measure μ does not charge singletons22 and the quasi-distance
d : X × X → [0, ∞) is continuous23 in the product topology τd × τd . Also, fix a rear-
rangement invariant Banach function space X on (X, μ). Then the Hardy-Littlewood
maximal operator M X on X (cf. (A.0.71)) induces a well-defined sublinear bounded
mapping on X whenever its lower Boyd index is super unital, i.e.,

M X : X −→ X provided pX > 1. (5.3.163)

Moreover,

M X : X −→ X and M X : X −→ X
(5.3.164)
are well-defined and bounded if 1 < pX ≤ qX < ∞.

Proof The assumption that the measure μ does not charge singletons renders the
measure space (X, μ) non-atomic (see [133, Lemma 7.4.2]). Bearing this in mind,
in the case when pX > 1 we may invoke Proposition 5.3.13 to conclude that the
sublinear operator M X maps X boundedly into itself. That (5.3.164) then holds

20 i.e., there exists a constant C ∈ (0, ∞) with the property that for any two functions f , g in the
domain of T one has |T ( f + g)| ≤ C(|T f | + |T g |) at μ-a.e. point
21 with the convention that L ∞,1 (X, μ) := L ∞ (X, μ) and L ∞,∞ (X, μ) := L ∞ (X, μ)
22 i.e., μ({x }) = 0 for each x ∈ X
23 The result proved in [133, Theorem 7.1.2] tells us that any quasi-metric space has an equivalent
quasi-distance satisfying this property
5.3 Young Functions and Orlicz Spaces 279

follows from (5.3.163), in view of the fact that pX = qX (and qX = pX ); see [14],
[123, §8]. 

Below we shall specialize Proposition 5.3.14 to the class of Orlicz space. In

relation to this we wish to remark that, given a non-atomic sigma-finite measure
space (X, M, μ) along with a Young function Φ, then it is know (cf., e.g., [14,
Theorem 8.18, p. 277], [122], [38, Remark 4.5, p. 71]) that
the dilation indices of the Young function Φ are the same as the
Boyd indices of its corresponding Orlicz space (regarded as a
(5.3.165)
rearrangement-invariant Banach function space), i.e., in the case when
Φ
X := L (X, μ) one has i(Φ) = pX and I(Φ) = qX .
 
Proposition 5.3.15 Suppose X, d, μ is a space of homogeneous type such that the
measure μ does not charge singletons24 and the quasi-distance d : X × X → [0, ∞)
is continuous25 in the product topology τd × τd . Also, fix a Young function Φ and
denote by i(Φ), I(Φ) ∈ [1, ∞] its upper and lower dilation indices.
Then the Hardy-Littlewood maximal operator M X on X (cf. (A.0.71)) induces a
well-defined sublinear bounded mapping on the corresponding Orlicz space when-
ever the lower dilation index of Φ is super unital, i.e.,

M X : L Φ (X, μ) −→ L Φ (X, μ) provided i(Φ) > 1. (5.3.166)

 denoting the complementary function of Φ (cf. (5.3.54)),
As a consequence, with Φ
 
M X : L Φ (X, μ) → L Φ (X, μ) and M X : L Φ (X, μ) → L Φ (X, μ)
(5.3.167)
are well-defined bounded operators if 1 < i(Φ) ≤ I(Φ) < ∞.

Proof This is a direct consequence of Proposition 5.3.14, bearing in mind (5.3.31),

 is itself a Young function.
(5.3.33), and (5.3.165), (5.3.59), and the fact that Φ 

In general, the boundedness result for the Hardy-Littlewood maximal operator in

the context of (5.3.166) fails for the (doubling) Young function satisfying defined in
(5.3.45) (recall that i(Φ) = 1 in this case). In relation to this, we wish to bring in a
positive result stating that
 
if X, d, μ is a space of homogeneous type such that the quasi-distance
d : X × X → [0, ∞) is continuous in the product topology τd × τd ,
then there exists some constant CX ∈ (0, ∞) with the property that for (5.3.168)
each μ-measurable set  E ⊆ X∫and each function f ∈ M (X, μ) one  has
∫
E
M X f dμ ≤ CX · μ(E) + X | f (x)| · max{0, ln| f (x)|} dμ(x) .

See [62, Theorem 2.7, p. 146] for a version of this result in the entire Euclidean
space, whose proof readily adapts to the present setting. In fact, it is known that

24 i.e., μ({x }) = 0 for each x ∈ X

25 from [133, Theorem 7.1.2] we know that any quasi-metric space has an equivalent quasi-distance
satisfying this property
280 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

if (X, μ) is a finite measure space, then any quasilinear operator T which

is of weak type (1, 1) and of weak type (p, p) on (X, μ), for some p > 1, (5.3.169)
maps the Zygmund space L log L(X, μ) boundedly into L 1 (X, μ).
A slightly more general result is found in [14, Theorem 6.6, p. 248]. This is also
closely related to Yano’s Extrapolation Theorem (see [207]; cf. also the discussion
and additional references in [13], [26], [47]). The latter asserts that
if (X, μ) is a finite measure space and T is a sublinear operator satisfying
T L p (X,μ)→L p (X,μ) ≤ C(p − 1)−α for some fixed α > 0, C ∈ (0, ∞),
and all p’s near 1 with p > 1 then T maps L (log L)α (X, μ) bound- (5.3.170)
edly into L 1 (X, μ); in fact, if μ is merely sigma-finite then T maps
L (log L)α (X, μ) boundedly into L 1 (X, μ) + L ∞ (X, μ).
In particular, from Lemma
 5.3.1 and either (5.3.168), or (5.3.170) and [133, (7.6.18)],
it may be seen that if X, d, μ is a space of homogeneous type such that the quasi-
distance d : X × X → [0, ∞) is continuous in the product topology τd × τd , then the
Hardy-Littlewood maximal operator

M X : L log L(X, μ) −→ L 1 (X, μ) is well defined and bounded if μ(X) < ∞.

(5.3.171)
The proof of the classical Marcinkiewicz interpolation theorem naturally extends
to produce modular estimates in the context described in the theorem below.
 
Theorem 5.3.16 (Modular Interpolation Theorem) Let X, μ be a non-atomic
sigma-finite measure space. Consider a doubling Young function Φ (i.e., a Young
function whose upper and lower dilation indices satisfy 1 ≤ i(Φ) ≤ I(Φ) < ∞; cf.
(5.3.19)), and pick two integrability exponents p, q ∈ (0, ∞] such that

0 < p < i(Φ) ≤ I(Φ) < q ≤ ∞. (5.3.172)

Then
L Φ (X, μ) ⊆ L p (X, μ) + L q (X, μ). (5.3.173)
Also, for each quasi-subadditive operator26

T : L p (X, μ) + L q (X, μ) −→ M (X, μ) (5.3.174)

which is both of weak-type (p, p) and of weak-type (q, q), i.e., there exist two constants
Mp, Mq ∈ (0, ∞) such that

Tf L p,∞ (X,μ) ≤ Mp f L p (X,μ) for every f ∈ L p (X, μ), (5.3.175)

26 i.e., there exists some

 constant C ∈ (0, ∞) such that for all functions f , g in the domain of T one
has |T ( f + g)| ≤ C |T f | + |T g | at μ-a.e. point
5.3 Young Functions and Orlicz Spaces 281

and27

Tf L q,∞ (X,μ) ≤ Mq f L q (X,μ) for every f ∈ L q (X, μ), (5.3.176)

it follows that there exists C ∈ (0, ∞), depending only on p, q, i(Φ), I(Φ), and the
quasi-subadditivity constant of T, such that for each function f ∈ L Φ (X, μ) one has
∫ ∫
   
Φ |(T f )(x)| dμ(x) ≤ C · CT Φ | f (x)| dμ(x), (5.3.177)
X X

for
p q
Mp + Mq if q < ∞,
CT := p  α, β
(5.3.178)
Cα,β · Mp · max M∞ M∞ if q = ∞,
where α, β ∈ R are any numbers satisfying 0 < α < i(Φ) ≤ I(Φ) < β < ∞ and
Cα,β ∈ [1, ∞) depends only on α, β, and the doubling character of Φ.
As a corollary of this modular estimate and the last property in the statement of
Lemma 5.3.3, T is bounded on the Orlicz space L Φ (X, μ).
Proof Choose another exponent p0 ∈ (p, ∞) such that 1 ≤ p0 < i(Φ). Then

L Φ (X, μ) ⊆ L p0 (X, μ) + L q (X, μ) ⊆ L p (X, μ) + L q (X, μ) (5.3.179)

with the first inclusion implied by (5.3.161) (keeping in mind (5.3.165)), and the
second inclusion implied by the fact that L p0 (X, μ) ⊆ L p (X, μ) + L q (X, μ) since
p < p0 < q. This establishes (5.3.173).
Next, assume T is a quasi-sublinear operator as in (5.3.174) which is both of
weak-type (p, p) and of weak-type (q, q). The goal is to establish (5.3.177). To this
end, we first note that, as seen from the discussion surrounding (5.3.106), there is
no loss of generality in assuming that the Young function Φ is also continuously
differentiable on (0, ∞). Given any number t ∈ (0, ∞), we may then integrate by parts
∫ t ∫
Φ(s) 1 t −p 
p+1
ds = − (s ) Φ(s) ds (5.3.180)
0 s p 0
∫ t  ! ∫
1 Φ(s) t Φ (s) 1 Φ(t) 1 t Φ(s)
=− − ds = − + ds,
p sp 0 0 sp p tp p 0 sp

where the last equality uses (5.3.89) and the fact that 0 < p < i(Φ). From (5.3.180)
and (5.3.90) we then conclude that there exists a constant C ∈ (0, ∞) such that
∫ t 
Φ (s) Φ(t)
p
ds ≤ C p for each t ∈ (0, ∞). (5.3.181)
0 s t

To proceed, consider the case when q < ∞. Then for any t ∈ (0, ∞) we may
integrate by parts
27 with the convention that L ∞,∞ (X, μ) := L ∞ (X, μ)
282 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
∫ ∞ ∫ ∞
Φ(s) 1
ds = − (s−q )Φ(s) ds (5.3.182)
t s q+1 q t
∫ ! ∫
∞
1 Φ(s) ∞ Φ(s) 1 Φ(t) 1 ∞ Φ(s)
=− − ds = + ds,
q sp t t sq q tq q t sq

where the last equality uses (5.3.91) and the fact that I(Φ) < q < ∞. In concert with
(5.3.92), this implies that there exists a constant C ∈ (0, ∞) such that
∫ ∞ 
Φ (s) Φ(t)
q
ds ≤ C q for each t ∈ (0, ∞). (5.3.183)
t s t

Going further, pick α, β ∈ R satisfying 0 < α < i(Φ) ≤ I(Φ) < β < ∞. Also,
select an arbitrary function f ∈ L p (X, μ) + L q (X, μ). In particular, | f (x)| < ∞ for
μ-a.e. point x ∈ X. Finally, fix some constant γ ∈ (0, ∞), to be determined later. For
each t ∈ (0, ∞) decompose f = gt + ht where

gt := f 1 { | f |>t/γ } and ht := f 1 {0< | f | ≤t/γ } . (5.3.184)

Then for each t ∈ (0, ∞) we have |gt | ≤ | f | and |ht | ≤ | f |, which places gt , ht in
L p (X, μ)+ L q (X, μ) (since this is a lattice). In view of this and the quasi-subadditivity
property of T, for each t ∈ (0, ∞) we may estimate
 
|(T f )(x)| ≤ C |(T gt )(x)| + |(T ht )(x)| at μ-a.e. point x ∈ X. (5.3.185)

Based on this and upon recalling that (X, μ) is sigma-finite, we may estimate (cf.
[133, (8.4.102)])
∫ ∫ ∞ 
   
Φ |(T f )(x)| dμ(x) = μ x ∈ X : |(T f )(x)| > t Φ(t) dt
X 0
∫ 
∞ 
≤ μ x ∈ X : |(T gt )(x)| > t/(2C) Φ(t) dt
0
∫ 
∞ 
+ μ x ∈ X : |(T ht )(x)| > t/(2C) Φ(t) dt
0

=: I + II. (5.3.186)

The fact that T is of weak-type (p, p) with p ∈ (0, ∞) allows us to estimate

5.3 Young Functions and Orlicz Spaces 283
∫ t  −p
∞
gt L p (X,μ) Φ(t) dt
p p
I ≤ Mp
0 2C
∫ ∞ ∫ 
−p
p p
= (2C) Mp t | f (x)| p dμ(x) Φ(t) dt
0 {x ∈X: | f (x) |>t/γ }
∫ ∫
p Φ(t) 
| f (x) |γ
= (2C) p
Mp | f (x)| p
dt dμ(x)
X 0 tp
∫  ∫ | f (x) |γ Φ(t) 
p
= (2C) p Mp | f (x)| p dt dμ(x)
{x ∈X: ∞> | f (x) |>0} 0 tp
∫
p Φ(| f (x)|γ)
≤ (2C) p Mp | f (x)| p dμ(x)
{x ∈X: ∞> | f (x) |>0} | f (x)| p
∫
p
= (2C) p Mp Φ(| f (x)|γ) dμ(x)
{x ∈X: ∞> | f (x) |>0}
∫
p
= (2C) p Mp Φ(| f (x)|γ) dμ(x)
X
∫
≤ CMp · max{γ α, γ β } ·
p
Φ(| f (x)|) dμ(x), (5.3.187)
X

also using Fubini’s theorem (recall that (X, μ) is sigma-finite), (5.3.184), (5.3.181),
and (5.3.40).
In addition, the fact that T is of weak-type (q, q) with q ∈ (1, ∞) permits us to
estimate
284 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
∫
 t  −q
∞
Φ(t) dt
q q
II ≤ Mq ht L p (X,μ)
0 2C
∫ ∞ ∫ 
−q
q q
= (2C) Mq t | f (x)| q dμ(x) Φ(t) dt
0 {x ∈X: 0< | f (x) | ≤t/γ }
∫ ∫
q
∞
Φ(t) 
= (2C)q Mq | f (x)| q dt dμ(x)
{x ∈X: ∞> | f (x) |>0} | f (x) |γ tp
∫
q Φ(| f (x)|γ)
≤ (2C)q Mq | f (x)| q dμ(x)
{x ∈X: ∞> | f (x) |>0} | f (x)| q
∫
q
= (2C)q Mq Φ(| f (x)|γ) dμ(x)
{x ∈X: ∞> | f (x) |>0}
∫
q
= (2C) q
Mq Φ(| f (x)|γ) dμ(x)
X
∫
q α β
≤ CMq · max{γ , γ } · Φ(| f (x)|) dμ(x), (5.3.188)
X

after also availing ourselves of Fubini’s theorem, (5.3.184), (5.3.183), and (5.3.40).
From (5.3.186), (5.3.187), (5.3.188), and (5.3.173) we see that the modular estimate
claimed in (5.3.177) holds if q < ∞ (in this scenario, the actual value of the parameter
γ ∈ (0, ∞) is inconsequential; e.g., we make take γ := 1).
In the case when q = ∞, for each t ∈ (0, ∞) we see from the fact that T is bounded
on L ∞ (X, μ) and (5.3.184) that

|(T ht )(x)| ≤ T ht L ∞ (X,μ) ≤ M∞ ht L ∞ (X,μ)

≤ M∞ · (t/γ) < t/(2C) at μ-a.e. point x ∈ X, (5.3.189)

with the very last inequality valid provided we choose, e.g.,

γ := 4CM∞ (5.3.190)

to begin with28. For such a choice,

 
μ x ∈ X : |(T ht )(x)| > t/(2C) = 0 for each t ∈ (0, ∞). (5.3.191)

hence II = 0. Based on this, (5.3.186), and (5.3.187) we then once again conclude
that the modular estimate claimed in (5.3.177) is true in the case when q = ∞ as
well. 

In turn, the modular interpolation result established Theorem 5.3.16 readily yields
weighted modular estimates for the Hardy-Littlewood maximal operator, of the

28 assuming T is not identically zero, in which case everything is trivially true

5.3 Young Functions and Orlicz Spaces 285

sort described in the corollary below. In particular, they allow us to recapture the
boundedness properties of the Hardy-Littlewood maximal operator on Orlicz spaces
already established in (5.3.167).
 
Corollary 5.3.17 Let X, d, μ be a space of homogeneous type such that the measure
μ does not charge singletons29 and the quasi-distance d : X × X → [0, ∞) is
continuous30 in the product topology τd ×τd . Bring in the Hardy-Littlewood maximal
operator M X on X (cf. (A.0.71)). Finally, consider a Young function Φ whose upper
and lower dilation indices satisfy

1 < i(Φ) ≤ I(Φ) < ∞. (5.3.192)

Then for each Muckenhoupt weight w ∈ Ai(Φ) (X, μ) there exists some constant
C ∈ (0, ∞), which depends only on the ambient (X, d, μ), the dilation indices of Φ,
and [w] Ai(Φ) , with the property that
∫ ∫
   
Φ |(M X f )(x)| w(x) dμ(x) ≤ C Φ | f (x)| w(x) dμ(x)
X X (5.3.193)
for each function f belonging to the weighted Orlicz space L Φ (X, w μ).

In particular, corresponding to the case when w ≡ 1, one has the “plain” modular
estimate
∫ ∫
   
Φ |(M X f )(x)| dμ(x) ≤ C Φ | f (x)| dμ(x) for each f ∈ L Φ (X, μ).
X X
(5.3.194)
Finally, as a corollary of (5.3.194), (5.3.59), and the last property in the statement
of Lemma 5.3.3, the Hardy-Littlewood maximal operator M X induces well-defined
sublinear bounded mappings both from the Orlicz space L Φ (X, μ) into itself, and
  is the complementary function
from the Orlicz space L Φ (X, μ) into itself, where Φ
of Φ.
 
Proof Given any weight w ∈ Ai(Φ) (X, μ) there exists ε ∈ 0, i(Φ) − 1 with the
property that
w ∈ Ai(Φ)−ε (X, μ) ⊆ AI(Φ)+ε (X, μ). (5.3.195)
See items (9) and (5) in [133, Lemma 7.7.1] which also show that ε as well as
[w] Ai(Φ)−ε and [w] A I (Φ)+ε are controlled in terms the ambient (X, d, μ), the lower
dilation index i(Φ), and [w] Ai(Φ) . Then the claim made in (5.3.193) is seen from
 
Theorem 5.3.16,presently  used for the choices T := M X , p := i(Φ) − ε ∈ 1, i(Φ) ,
q := I(Φ) + ε ∈ I(Φ), ∞ , and with μ replaced by the measure w μ, bearing in mind
item (1) in [133, Lemma 7.7.1], [133, Lemma 7.4.2], and [133, Lemma 7.4.3]. 
We conclude by presenting an extrapolation result from Muckenhoupt weighted
Lebesgue space estimates to modular estimates on Orlicz spaces. This refines and
29 i.e., μ({x }) = 0 for each x ∈ X
30 it is known from [133, Theorem 7.1.2] that any quasi-metric space has an equivalent quasi-
distance satisfying this property
286 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

extends [38, Theorem 4.15, p. 78]. For example, in contrast to the argument given
on [38, pp. 80-83] in the entire Euclidean space, our proof is carried out in spaces of
homogeneous type and does not rely on the classical Rubio de Francia extrapolation
theorem (see [133, Proposition 7.7.6]). Also, we provide a more transparent formula
for the constant involved in the main estimate.
Theorem 5.3.18 (Modular Extrapolation Theorem) Let (X, d, μ) be a space of
homogeneous type such that the measure μ does not charge singletons31 and the
quasi-distance d : X × X → [0, ∞) is continuous32 in the product topology τd × τd .
Let M X stand for the Hardy-Littlewood maximal operator on (X, d, μ). Also, pick a
Young function Φ whose upper and lower dilation indices satisfy

1 < i(Φ) ≤ I(Φ) < ∞, (5.3.196)

 Then the following statements are true.
and denote its complementary function by Φ.
(1) The functions Φ, Φ  are doubling, and the operator M X is bounded both on

L Φ (X, μ) and on L Φ (X, μ). In particular, with the piece of notation introduced
in (5.3.38),
N Φ (M X f )
N Φ [M X ] := sup (5.3.197)
f ∈L Φ (X,μ) N Φ( f )
f 0

and

 N Φ (M X f )
N Φ [M X ] := sup 
(5.3.198)

f ∈L Φ (X,μ) NΦ (f)
f 0

are well-defined real numbers, belonging to the interval (0, ∞).

(2) There exists some CX,Φ ∈ (0, ∞), depending only on the quasi-distance d (via
d defined as in (A.0.19)-(A.0.20)) the doubling charter of μ,
the constants Cd, C
as well as the dilation indices of Φ, with the following significance. Fix some
integrability exponent p0 ∈ (1, ∞) and define
 p 
WΦ, p0 := CX,Φ 0 · N Φ [M X ] p0 −1 · N Φ [M X ]. (5.3.199)

Also, suppose the number CW ∈ (0, ∞) and the μ-measurable real-valued

functions f , g defined on X are such that for any given Muckenhoupt weight
w ∈ Ap0 (X, d, μ) with [w] A p0 ≤ WΦ, p0 one has

f L p0 (X,w) ≤ CW g L p0 (X,w) . (5.3.200)

Then whenever 0 < α < i(Φ) ≤ I(Φ) < β < ∞ there exists some constant
CΦ,α,β ∈ (0, ∞), which depends only on p0 , the dilation indices of Φ, as well as

31 i.e., μ({x }) = 0 for each x ∈ X

32 The result proved in [133, Theorem 7.1.2] guarantees that any quasi-metric space has an equivalent
quasi-distance satisfying this property
5.3 Young Functions and Orlicz Spaces 287

α and β, with the property that

∫ ∫
     
Φ | f (x)| dμ(x) ≤ CΦ,α,β · max (CW )α, (CW )β · Φ |g(x)| dμ(x).
X X
(5.3.201)

Proof Based on (5.3.61) and (5.3.196) we conclude that Φ, Φ  ∈ Δ2 . Also, from

(5.3.196) and (5.3.167) (or the very last property in the statement of Corollary 5.3.17)

we see that M X is bounded both on L Φ (X, μ) and on L Φ (X, μ). To proceed, from
Φ 
Φ
(5.3.193)-(5.3.194) we know that N [M X ] and N [M X ] defined in (5.3.197)-
(5.3.198) are are well-defined numbers in the interval [0, ∞), with the property
that
N Φ (M X f ) ≤ N Φ [M X ] · N Φ ( f ) for each f ∈ L Φ (X, μ), (5.3.202)
and
   
N Φ (M X f ) ≤ N Φ [M X ] · N Φ ( f ) for each f ∈ L Φ (X, μ). (5.3.203)

To get strictly positive lower bounds for N Φ [M X ], N Φ [M
X ], fix a point xo ∈ X
along with a radius r ∈ (0, ∞), and consider the d-ball B := Bd (xo, r) in X. Note that
∫
   
N Φ 1B = Φ 1B (x) dμ(x) = Φ(1)μ(B). (5.3.204)
X

Also, it may be seen from (A.0.71) that there exists some CX ∈ (1, ∞), which depends
only on the doubling constant of μ and the quasi-distance d, with the property that33
 
CX · M X 1B ≥ 1B pointwise on X. (5.3.205)

In concert with (5.3.204), (5.3.202), and (5.3.20) (with, say α := 1 and β := 2I(Φ)),
this permits us to estimate (also bearing in mind that Φ is increasing, as well as
doubling)
    
Φ(1)μ(B) = N Φ 1B ≤ N Φ CX · M X 1B
  
≤ CΦ · (CX )2I(Φ) · N Φ M X 1B
 
≤ CΦ · (CX )2I(Φ) · N Φ [M X ]N Φ 1B

= CΦ · (CX )2I(Φ) · N Φ [M X ]Φ(1)μ(B), (5.3.206)

for some CΦ ∈ (1, ∞) depending only on I(Φ). Ultimately, this proves that

33 if the measure μ is Borel-semiregular on (X, τρ ), then Lebesgue’s Differentiation Theorem (see

(1) ⇔ (3) in [133, Proposition 7.4.4]) ensures that (5.3.205) holds with C X := 1
288 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

N Φ [M X ] ≥ CΦ · (CX )−2I(Φ) . (5.3.207)

If we now define   −1
ΛΦ := 2CΦ · (CX )2I(Φ) · N Φ [M X ] (5.3.208)

then
∞
j
ΛΦ ∈ (0, 1/2] and (1 − ΛΦ )ΛΦ = 1. (5.3.209)
j=0

These considerations allow us to implement a version of Rubio de Francia’s

iterative algorithm in the present setting. Specifically, define

T : L Φ (X, μ) −→ M+ (X, μ) by setting

∞
(5.3.210)
ΛΦ M X φ for each function φ ∈ L Φ (X, μ),
j j
T φ :=
j=0

where M X0 φ := |φ| for each φ ∈ L Φ (X, μ), and M X := M X ◦ · · · ◦ M X is the j-fold

j

composition of M X : L Φ (X, μ) → L Φ (X, μ) with itself. In view of the nature of the

first term in (5.3.210), we have

|φ| ≤ T φ for each function φ ∈ L Φ (X, μ). (5.3.211)

By also taking into account to the sublinearity of M X on M+ (X, μ) and (5.3.208),

it follows that for each function φ ∈ L Φ (X, μ) we may write
∞ ∞
ΛΦ M X φ = Λ−1 ΛΦ M X φ ≤ Λ−1
j j+1 j+1 j+1
M X (T φ) ≤ Φ Φ Tφ
j=0 j=0

= 2CΦ · (CX )2I(Φ) · N Φ [M X ] · T φ. (5.3.212)

Next, given any function φ ∈ L Φ (X, μ), we make use of (5.3.38), (5.3.210), the
fact that Φ is convex, (5.3.209), iterations of (5.3.202), and the monotonicity of N Φ
to write
5.3 Young Functions and Orlicz Spaces 289
∫  ∞ 
j  j 
N Φ (T φ) = Φ ΛΦ M X φ (x) dμ(x)
X j=0

∫  ∞
j  j  
= Φ (1 − ΛΦ )ΛΦ M X (φ/(1 − ΛΦ )) (x) dμ(x)
X j=0

∫ 
∞
j j  
≤ (1 − ΛΦ )ΛΦ Φ M X (φ/(1 − ΛΦ )) (x) dμ(x)
j=0 X

∞ 
j  
ΛΦ N Φ M X φ/(1 − ΛΦ )
j
≤
j=0

∞
 
ΛΦ N Φ [M X ] j N Φ φ/(1 − ΛΦ )
j
≤
j=0

  ∞  j
= N Φ φ/(1 − ΛΦ ) ΛΦ · N Φ [M X ]
j=0

Φ
 
N φ/(1 − ΛΦ )
= ≤ 2N Φ (2φ)
1 − ΛΦ · N Φ [M X ]

≤ CΦ · 22I(Φ)+1 · N Φ (φ), (5.3.213)

where in the penultimate line we have also used that

  −1
ΛΦ · N Φ [M X ] = 2CΦ · (CX )2I(Φ) ∈ (0, 1/2), (5.3.214)

itself a consequence of (5.3.208) and the fact that CΦ, CX ∈ (1, ∞), while the very
last line in (5.3.213) employs (5.3.40) (with λ := 2 and, say, α := 1, β := 2I(Φ)) and
(5.3.8). Hence,

N Φ (T φ) ≤ CΦ · 22I(Φ)+1 · N Φ (φ) for each φ ∈ L Φ (X, μ). (5.3.215)

From (5.3.215) and Lemma 5.3.3 we deduce that T actually maps the Orlicz space
L Φ (X, μ) boundedly into itself, thus

the operator T : L Φ (X, μ) −→ L Φ (X, μ)

(5.3.216)
is well-defined, sublinear, and bounded.

Together with (5.3.211) and (5.3.212), this further implies

0 < T φ < ∞ at μ-a.e. point on X whenever 0  φ ∈ L Φ (X, μ). (5.3.217)

290 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces

From (5.3.217), (5.3.212), and item (4) in [133, Lemma 7.7.1] we then conclude that

if φ ∈ L Φ (X, μ) is not identically zero (μ-a.e.), then

(5.3.218)
T φ ∈ A1 (X, μ) and [T φ] A1 ≤ CX,Φ · N Φ [M X ]

where CX,Φ ∈ (0, ∞) depends only on the quasi-distance d (via the constants Cd, C d
defined as in (A.0.19)-(A.0.20)), the doubling charter of μ, and upper dilation index
of Φ.
In a similar fashion, we define

 : LΦ
T (X, μ) −→ M+ (X, μ) by setting
∞ (5.3.219)
 := 
Λ M X ψ for each function φ ∈ L Φ (X, μ),
j j
Tψ
Φ
j=0

where ΛΦ 
 ∈ (0, 1/2] is defined in relation to Φ the same way ΛΦ has been introduced
in reference to Φ in (5.3.209), i.e., for some CΦ 
 ∈ (1, ∞) depending only on I(Φ),

  −1   −1

2I(Φ)  
ΛΦ  · (CX )
 := 2CΦ · N Φ [M X ] = 2CΦ
 · (CX )
2i(Φ)
· N Φ [M X ] . (5.3.220)

Then, much as above,


 : LΦ 
the operator T (X, μ) −→ L Φ (X, μ)
is well-defined, sublinear, bounded, (5.3.221)
 for each ψ ∈ 
and |ψ| ≤ Tψ L Φ (X, μ).

Also,
    
N Φ (Tψ) ≤ CΦ
 ·2
2I(Φ)+1 · N Φ (ψ) = CΦ
 ·2
2i(Φ)+1 · N Φ (ψ)
(5.3.222)

for each function ψ ∈ L Φ (X, μ),

and

if ψ ∈ L Φ (X, μ) is not identically zero (μ-a.e.), then
(5.3.223)
 ∈ A1 (X, μ) and [Tψ]
 A1 ≤ C 
X,Φ · N Φ
Tψ [M X ]
where C X,Φ ∈ (0, ∞) depends only on the quasi-distance d (via the constants Cd, C d
defined as in (A.0.19)-(A.0.20)), the doubling charter of μ, and lower dilation index
of Φ.
Let us turn our attention to (5.3.201). Note that this is trivially true if f = 0 at
μ-a.e. point in X, or if N Φ (g) is either ∞ or 0 (since in the latter case (5.3.200) with,
say, w ≡ 1, forces f = 0 at μ-a.e. point in X). As far as (5.3.201) is concerned, we
are therefore left with considering the situation when
5.3 Young Functions and Orlicz Spaces 291

f is not identically zero (μ-a.e.), and 0 < N Φ (g) < ∞. (5.3.224)

Define ψ : X → [0, ∞) by setting, for each x ∈ X,
 
Φ | f (x)| | f (x)| if f (x)  0,
ψ(x) := (5.3.225)
0 if f (x)  0.

In particular, since Φ(0) = 0 and Φ > 0 on (0, ∞),

ψ is not identically zero (μ-a.e.) on X, (5.3.226)

and  
| f (x)|ψ(x) = Φ | f (x)| for each x ∈ X. (5.3.227)
 = 0 to conclude that
We may also use (5.3.225), (5.3.58), and the fact that Φ(0)
 
 ψ(x) ≤ Φ(| f (x)|) for each x ∈ X.
Φ (5.3.228)

As a consequence of this, (5.3.226), and (5.3.38), we have


0 < N Φ (ψ) ≤ N Φ ( f ). (5.3.229)

The remainder of the proof of (5.3.201) is divided into two steps.

Step I. Proof of (5.3.201) under the additional assumption f ∈ L Φ (X, μ). In
particular, thanks to (5.3.39) and (5.3.229),

0 < N Φ (ψ) ≤ N Φ ( f ) < ∞, (5.3.230)

which further entails (cf. (5.3.39))


0  ψ ∈ L Φ (X, μ). (5.3.231)

For some δ ∈ (0, 1/2) to be chosen later, define

φ := δ| f | + (1 − δ)|g| ∈ M+ (X, μ). (5.3.232)

Upon recalling that Φ is convex, we have

N Φ (φ) ≤ δN Φ ( f ) + (1 − δ)N Φ (g) < ∞, (5.3.233)

with the last inequality a consequence of (5.3.224) and the current working assump-
tion. In view of (5.3.39) and (5.3.224), this implies that

0  φ ∈ L Φ (X, μ). (5.3.234)

From the properties listed in (5.3.231), (5.3.218), (5.3.234), (5.3.223), and item (3)
in [133, Lemma 7.7.1] we see that if
292 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces


w := (T φ)1−p0 (Tψ) (5.3.235)

then
w is a weight in the Muckenhoupt class Ap0 (X, μ), and
  p −1    (5.3.236)
[w] A p0 ≤ CX,Φ · N Φ [M X ] 0 · C X,Φ · N Φ [M X ] ≤ WΦ, p0

with the final inequality in (5.3.236) coming from (5.3.199). With q0 ∈ (1, ∞)
denoting the Hölder conjugate exponent of p0 , we are now in a position to write
(with justifications provided below)
∫ ∫ ∫
 
N Φ( f ) = Φ | f (x)| dμ(x) = | f |ψ dμ ≤  dμ
| f |(Tψ)
X X X
∫
=  dμ
| f |(T φ)−1/q0 (T φ)1/q0 (Tψ)
X
∫  1/p0  ∫  1/q0
≤  dμ
| f | p0 (T φ)1−p0 (Tψ)  dμ
(T φ)(Tψ)
X X
∫  1/p0 ∫  1/q0
= | f | p0 dw  dμ
(T φ)(Tψ)
X X
∫  1/p0  ∫  1/q0
≤ CW |g| p0 dw  dμ
(T φ)(Tψ)
X X
∫  1/p0  ∫  1/q0
= CW |g| (T φ)
p0 1−p0  dμ
(Tψ)  dμ
(T φ)(Tψ)
X X
∫  1/p0  ∫  1/q0
= 2 p0 CW  dμ
(T φ)(Tψ)  dμ
(T φ)(Tψ)
X X
∫
= 2 p0 CW ·  dμ.
(T φ)(Tψ) (5.3.237)
X

Above, the first two equalities come from (5.3.38) and (5.3.227), the subsequent
inequality is a consequence of (5.3.221), and the third equality is clear from (5.3.217)
and (5.3.234). We next use Hölder’s inequality for the measure (Tψ)  dμ, then use
the definition of the weight w (cf. (5.3.235)). The third inequality is implied by the
hypothesis made in (5.3.200) plus (5.3.236), after which we unpack the definition
the weight w (cf. (5.3.235)). Next, bearing in mind that δ ∈ (0, 1/2), from (5.3.232)
together with (5.3.211) and (5.3.234) we see that

2−1 |g| ≤ (1 − δ)|g| ≤ φ ≤ T φ pointwise on X, (5.3.238)

so |g| p0 ≤ 2 p0 (T φ) p0 . Based on this we conclude that the equality in the penultimate

line of (5.3.237) is true. Finally, we use the fact that 1/p0 + 1/q0 = 1 in the last
equality of (5.3.237).
5.3 Young Functions and Orlicz Spaces 293

Going further, for each ε ∈ (0, 1) we may estimate

∫ ∫
    
 dμ = (ε −1T φ)(εTψ)
(T φ)(Tψ)  dμ ≤ N Φ ε −1T φ + N Φ εTψ
X X

 
≤ CΦ · ε −2I(Φ) · N Φ (T φ) + ε · N Φ (Tψ)

≤ (CΦ )2 · 22I(Φ)+1 ε −2I(Φ) · N Φ (φ) + CΦ ·2
2i(Φ)+1
ε · N Φ (ψ)
 
≤ (CΦ )2 · 22I(Φ)+1 ε −2I(Φ) · δN Φ ( f ) + (1 − δ)N Φ (g)

 ·2
+ CΦ 2i(Φ)+1
ε · N Φ( f )
 
≤ 2 (CΦ )2 · 22I(Φ) ε −2I(Φ) · δ + CΦ
 ·2
2i(Φ)
ε · N Φ( f )

+ (CΦ )2 · 22I(Φ) ε −2I(Φ) · N Φ (g), (5.3.239)

where we have used Young’s inequality (recorded in (5.3.56)) and (5.3.38) in the
first inequality, (5.3.40) (with, say, α := 1 and β := 2I(Φ)) as well as (5.3.2) (written
 in place of Φ and for λ := ε ∈ (0, 1)) in the second inequality, (5.3.215) and
for Φ
(5.3.222) in the third inequality, (5.3.230) and (5.3.233) in the fourth inequality, and
the fact that δ ∈ (0, 1/2) in the final inequality. If we further restrict ε to (0, 1/2) and
take
δ := ε 1+2I(Φ) ∈ (0, 1/2), (5.3.240)
we see that (5.3.239) becomes
∫  
 dμ ≤ 2ε · (CΦ )2 · 22I(Φ) + C · 22i(Φ) · N Φ ( f )
(T φ)(Tψ) Φ
X

+ (CΦ )2 · 22I(Φ) ε −2I(Φ) · N Φ (g). (5.3.241)

At this stage, use (5.3.241) back into (5.3.237) for the choice
   −1 
ε := min 2−1, 2−(p0 +2) (CΦ )2 · 22I(Φ) + CΦ
 ·2
2i(Φ)
(CW )−1 ∈ (0, 1/2]. (5.3.242)

After absorbing the term N Φ ( f ) from the right side, whose coefficient is ≤ 1/2, into
the left side we obtain
 
N Φ ( f ) ≤ CΦ · CW · max 1, CΦ · (CW )2I(Φ) · N Φ (g), (5.3.243)

for some constant CΦ ∈ (0, ∞) which depends only on p0 and the dilation indices of
Φ.
Having proved (5.3.243), running the same argument as above for the function
CW · g in place of g produces the version of (5.3.243) with CW = 1 and g replaced
by CW · g, i.e.,
294 5 Banach Function Spaces, Extrapolation, and Orlicz Spaces
 
N Φ ( f ) ≤ CΦ · N Φ CW · g , (5.3.244)
for some CΦ ∈ (0, ∞) depending only on p0 and the dilation indices of Φ. At this
stage, whenever 0 < α < i(Φ) ≤ I(Φ) < β < ∞ we conclude from (5.3.244) and
(5.3.40) that (5.3.201) holds.
Step II. The end-game in the proof of (5.3.201). From Step I we know that
(5.3.201) is true if f ∈ L Φ (X, μ). The goal here is to eliminate the latter additional
assumption. To this end, fix a reference point xo ∈ X and define

f j := min{| f |, j} · 1B d (xo, j) for each j ∈ N. (5.3.245)

Then each f j is a μ-measurable function on X satisfying 0 ≤ f j ≤ j · 1B d (xo, j) . From

the monotonicity of Φ, (5.3.29) and (5.3.37) we see that

each f j belongs to L Φ (X, μ). (5.3.246)

Also, 0 ≤ f j | f | as j → ∞ in a pointwise fashion on X. In particular, (5.3.200)

implies that for every Muckenhoupt weight w ∈ Ap0 (X, d, μ) with [w] A p0 ≤ WΦ, p0
we have
f j L p0 (X,w) ≤ f L p0 (X,w) ≤ CW g L p0 (X,w) . (5.3.247)
On account of this and (5.3.246), we conclude from (5.3.247) and Step I that whenever
we have 0 < α < i(Φ) ≤ I(Φ) < β < ∞ there exists a constant CΦ,α,β ∈ (0, ∞),
depending only on p0 , the dilation indices of Φ, as well as α and β, with the property
that for each j ∈ N we have
∫ ∫
     
Φ | f j (x)| dμ(x) ≤ CΦ,α,β · max (CW )α, (CW )β · Φ |g(x)| dμ(x).
X X
(5.3.248)
Passing to limit j → ∞ and relying on Lebesgue’s Monotone Convergence Theorem
(while bearing in mind the monotonicity of Φ) we then arrive at the conclusion that
(5.3.201) holds as stated. 
Chapter 6
Morrey-Campanato Spaces, Morrey Spaces, and
Their Pre-Duals on Ahlfors Regular Sets

General information pertaining to the classical Morrey and Morrey-Campanato

spaces in the entire Euclidean ambient may be found in a multitude of sources
(cf., e.g., [2], [6], [63], [117], [155], [157], [184] and the references therein). The
goal here is to develop a theory for these scales of spaces, which is comparable
in scope and power to its Euclidean counterpart, in more general geometric settings.

6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors

Regular Sets

To set the stage, throughout we let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a

closed Ahlfors regular set and abbreviate σ := H n−1 Σ. For each p ∈ (1, ∞) and
.
λ ∈ (0, n − 1) define the homogeneous Morrey-Campanato space L p,λ (Σ, σ) as
.  
L p,λ (Σ, σ) := f ∈ Lloc
1
(Σ, σ) :  f L. p, λ (Σ,σ) < +∞ (6.1.1)
⨏
where, for each f ∈ Lloc1 (Σ, σ) and with f
Δ(x,R) := Σ∩B(x,R)
f dσ for each R > 0
and x ∈ Σ, we have set
 
n−1−λ  ⨏  p  p1
 f L. p, λ (Σ,σ) := sup R p  f (y) − fΔ(x,R)  dσ(y) .
x ∈Σ and Σ∩B(x,R)
0<R<2 diam(Σ)
(6.1.2)

Note that, as is apparent from (6.1.1)-(6.1.2),

. p
L p,λ (Σ, σ) ⊆ Lloc (Σ, σ), (6.1.3)

and [133, (7.4.59)] gives, with Δ(x, R) := Σ ∩ B(x, R) for each R > 0 and x ∈ Σ,

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 295
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_6
296 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

 f L. p, λ (Σ,σ) (6.1.4)

 
n−1−λ ⨏ ⨏  p1
≈ sup R p | f (y) − f (z)| dσ(y) dσ(z)
p
,
x ∈Σ and Δ(x,R) Δ(x,R)
0<R<2 diam(Σ)

uniformly for f ∈ Lloc

1 (Σ, σ). In particular, this goes to show that there is a constant

C ∈ (0, ∞) such that

 | f | L. p, λ (Σ,σ) ≤ C f L. p, λ (Σ,σ) for each f ∈ Lloc

1
(Σ, σ). (6.1.5)

1 (Σ, σ), each R ∈ 0, 2 diam(Σ) , and each x ∈ Σ we

Also, given that for each f ∈ Lloc
have
n−1−λ  ⨏   1
R p  f − fΔ(x,R)  p dσ p
Σ∩B(x,R)

n−1−λ ⨏  p1
≤ 2R p | f | p dσ
Σ∩B(x,R)

n−1−λ  ⨏  s1
≤ 2R p | f | s dσ
Σ∩B(x,R)

n−1−λ
− s1
≤ 2R p σ Σ ∩ B(x, R)  f  L s (Σ,σ)

p(n − 1)
≤ C f  L r (Σ,σ) if s := , (6.1.6)
n−1−λ
it follows that there exists a constant C ∈ (0, ∞) with the property that for each
1 (Σ, σ) we have
function f ∈ Lloc

p(n − 1)
 f L. p, λ (Σ,σ) ≤ C f  L s (Σ,σ) provided s := . (6.1.7)
n−1−λ
We also wish to remark that while  · L. p, λ (Σ,σ) is merely a semi-norm, since

 f − cL. p, λ (Σ,σ) =  f L. p, λ (Σ,σ) for every f ∈ Lloc

1
(Σ, σ) and c ∈ C, (6.1.8)

if we introduce1

[ f ] L. p, λ (Σ,σ)/∼ :=  f L. p, λ (Σ,σ) for each f ∈ Lloc

1
(Σ, σ), (6.1.9)

then [·] L. p, λ (Σ,σ)/∼ becomes a genuine norm on the quotient space

1 as in the past, [ f ] denotes the equivalence class, modulo constants, of a given function f ; cf. [133,
(7.3.4)]
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 297
. p,λ . p,λ
L (Σ, σ)/∼ := [ f ] : f ∈ L (Σ, σ) . (6.1.10)

Let us next define the inhomogeneous Morrey-Campanato space L p,λ (Σ, σ) as

.
L p,λ (Σ, σ) := L p (Σ, σ) ∩ L p,λ (Σ, σ)
 
= f ∈ L p (Σ, σ) :  f L. p, λ (Σ,σ) < +∞ (6.1.11)

where
 f L p, λ (Σ,σ) :=  f  L p (Σ,σ) +  f L. p, λ (Σ,σ) . (6.1.12)
In particular, it follows from (6.1.11)-(6.1.12) and (6.1.3) that
.
L p,λ (Σ, σ) = L p,λ (Σ, σ) if Σ is bounded. (6.1.13)

From [133, Lemma 7.4.14] we also have

.  
L p,λ (Σ, σ) ⊂ L p,λ (Σ, σ) ⊂ L 1 Σ, 1+σ(x)
|x | n
(6.1.14)
for all p ∈ (1, ∞), λ ∈ (0, n − 1).

Going forward, given an integrability exponent q ∈ (1, ∞) and λ ∈ (0, n − 1), a

function a ∈ L q (Σ, σ) is said to be a H q,λ -atom on Σ provided there exist some
point xo ∈ Σ and some radius R ∈ 0, 2 diam(Σ) such that
1 ∫
λ q −1
supp a ⊆ B(xo, R) ∩ Σ, a L q (Σ,σ) ≤ R , a dσ = 0. (6.1.15)
Σ

In the case when the set Σ is bounded we agree that the constant function a ≡ 1 is
also considered to be a H q,λ -atom on Σ. We then consider the following space


H q,λ (Σ, σ) := f ∈ Lipc (Σ) : there exist a sequence {λ j } j ∈N ∈  1 (N) and
a family {a j } j ∈N of H q,λ -atoms on Σ such
∞ 

that f = λ j a j in Lipc (Σ) , (6.1.16)
j=1

and for each f ∈ H q,λ (Σ, σ) define


∞ 
∞

 f H q, λ (Σ,σ) := inf |λ j | : f = λ j a j in Lipc (Σ) with (6.1.17)
j=1 j=1

{λ j } j ∈N ∈  1 (N) and each a j a H q,λ
-atom on Σ .

In particular, as is visible from (6.1.17),

298 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

for each H q,λ

-atom a on Σ we have aH q, λ (Σ,σ) ≤ 1, (6.1.18)

while Hölder’s inequality together with the support condition and the normalization
of the atoms in (6.1.17) imply that

for each H q,λ -atom a on Σ we have a L r (Σ,σ) ≤ C with

q(n−1) (6.1.19)
r := n−1+λ(q−1) and with C ∈ (0, ∞) independent of the atom.

The above definitions also guarantee that


∞ 
if f ∈ H q,λ (Σ, σ) is expanded as f = λ j a j in Lipc (Σ) for
j=1
some sequence {λ j } j ∈N ∈  1 (N) and with each a j a H q,λ -atom on (6.1.20)

∞
Σ, then the series λ j a j also converges to f in H q,λ (Σ, σ).
j=1

By reasoning as in the proof of [209, Proposition 4, p. 589] (for a completeness

result in the entire Euclidean space which may be readily adapted to the current
setting) and as in the proof of Proposition 4.4.2 (for separability) we see that
 
H q,λ (Σ, σ),  · H q, λ (Σ,σ) is a separable Banach space. (6.1.21)

Moreover, from (6.1.17), (6.1.19), Riesz’s Duality Theorem for Lebesgue spaces,

and [133, Corollary 3.7.3] (ensuring the density of Lipc (Σ) in Lebesgue spaces
on Σ, it follows that we have a continuous embedding

q(n − 1)
H q,λ
(Σ, σ) → L r (Σ, σ) with r := ∈ (1, q). (6.1.22)
n − 1 + λ(q − 1)
In turn, from (6.1.20) and (6.1.22) we see that

∞ 
if f ∈ H q,λ (Σ, σ) is expressed as f = λ j a j in Lipc (Σ) for a
j=1
sequence {λ j } j ∈N ∈  1 (N) and with each a j a H q,λ -atom on Σ, then (6.1.23)
∞
q(n−1)
λ j a j converges to f in L r (Σ, σ) if we take r := n−1+λ(q−1) .
j=1

It is also useful to note that

if Σ is bounded then L q (Σ, σ) ⊆ H q,λ

(Σ, σ). (6.1.24)
⨏
Indeed, any function f ∈ L q (Σ, σ) may be written as the sum between f − Σ f dσ
which
⨏ is a multiple of an H q,λ -atom on Σ (with “large” support) and the constant
Σ
f dσ which, by definition, is considered to be a multiple of an H q,λ -atom in
this case.
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 299

The chief significance of H q,λ (Σ, σ) is that this is the pre-dual of Morrey-
Campanato spaces, defined earlier. Concretely, if 1/p + 1/q = 1 we have (see [209,
Proposition 5, p. 590])
. 
 ∗ L p,λ (Σ, σ) ∼ if Σ unbounded,
H (Σ, σ) =
q,λ
(6.1.25)
L p,λ (Σ, σ) if Σ bounded.

The precise sense in which (6.1.25) should be understood is as follows. In the case
when the set Σ is unbounded, consider the assignment taking any equivalence class
.  ∗
[ f ] ∈ L p,λ (Σ, σ) ∼ into the functional Λ[ f ] ∈ H q,λ (Σ, σ) defined as

 ∫
  N
Λ[ f ], g := lim λj f a j dσ, (6.1.26)
N →∞ Σ
j=1

whenever
g ∈ H q,λ (Σ, σ), {λ j } j ∈N ∈  1 (N), and {a j } j ∈N is
∞H -atoms on Σ, having the proper-
a sequence of q,λ
(6.1.27)
ty that g = j=1 λ j a j with convergence in the space
H q,λ (Σ, σ).

Then Λ[ f ] is unambiguously defined, and the assignment [ f ] → Λ[ f ] is a bounded

.  ∗
linear isomorphism from L p,λ (Σ, σ) ∼ onto H q,λ (Σ, σ) , with bounded inverse.
In particular, identifying Λ[ f ] with [ f ] yields the duality formula
  ∞ ∫ . 
[ f ], g = λ j Σ f a j dσ when [ f ] ∈ L p,λ (Σ, σ) ∼, the function g
j=1
(6.1.28)
belongs to H q,λ (Σ, σ), {λ j } j ∈N ∈  1 (N), and {a j } j ∈N is a sequence

of H q,λ -atoms on Σ, such that g = ∞ j=1 λ j a j in H
q,λ (Σ, σ).

As a consequence, there exists a finite constant C > 0, which depends only on the
environment, having the property that
 
 [ f ], g  ≤ C [ f ] . · gH q, λ (Σ,σ) .
L p, λ (Σ,σ)/∼
. p,λ (6.1.29)
for all f ∈ L (Σ, σ) and g ∈ H q,λ (Σ, σ).

Finally, there is an analogous interpretation of the identification of the dual in (6.1.25)

(plus naturally accompanying results as in (6.1.28)-(6.1.29)) in the case when Σ is
bounded, this time without having to consider equivalence classes of functions
modulo constants.

Lemma 6.1.1 Suppose Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a closed, unbounded

set, which is Ahlfors regular, and abbreviate σ := H n−1 Σ. Also, pick p, q ∈ (1, ∞)
with .1/p + 1/q = 1 and fix some λ ∈ (0, n − 1). Given a real-valued function
f ∈ L p,λ (Σ, σ), for each number M ∈ N define fM := min max{ f , −M }, M . Then
300 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .
.
fM ∈ L p,λ (Σ, σ) for each M ∈ N and
.  (6.1.30)
lim [ fM ] = [ f ] weak-∗ in L p,λ (Σ, σ) ∼ .
M→∞

Specifically, with ·, · denoting the duality bracket between the homogeneous
Morrey-Campanato space, modulo constants, and its pre-dual (cf. (6.1.25)), one
has
   
[ fM ], g → [ f ], g as M → ∞, for each g ∈ H q,λ (Σ, σ). (6.1.31)

Moreover, a similar weak-∗ convergence result holds in the case when Σ is

bounded, namely
fM ∈ L p,λ (Σ, σ) for each M ∈ N and
(6.1.32)
lim fM = f weak-∗ in L p,λ (Σ, σ).
M→∞

Proof The first claim in (6.1.30) is a consequence of (6.1.5). The second claim in
claim in (6.1.30) may be established via a similar argument to the one used in the
proof of Lemma 4.8.3, making use of (6.1.21). 

We continue to assume that Σ ⊆ Rn is a closed Ahlfors regular set, and abbreviate

σ := H n−1 Σ. In this setting, in addition to atoms (cf. (6.1.15)) there is a natural
concept of molecule. To describe it, fix three parameters, q ∈ (1, ∞), λ ∈ (0, n − 1),
and θ > (n − 1)(q − 1). Then a given function m ∈ L q (Σ, σ) is said to be a
H q,λ,θ -dome on Σ provided there exist some point xo ∈ Σ along with some radius
R ∈ 0, 2 diam(Σ) such that
∫
|m(x)| q dσ(x) ≤ Rλ(1−q) and
Σ
∫ (6.1.33)
θ λ(1−q)+θ
|m(x)| |x − xo | dσ(x) ≤ R
q
.
Σ

Observe that if p ∈ (1, ∞) is such that 1/p + 1/q = 1 then the conditions above imply
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 301
∫ ∫ ∫
|m| dσ = |m| dσ + |m| dσ
Σ Σ∩B(x o ,R) Σ\B(x o ,R)
∫  1/q
1/p
≤ σ Σ ∩ B(xo, R) |m| q dσ
Σ∩B(x o ,R)


∞ ∫  1/q
1/p
+ σ Σ ∩ B(xo, 2k R) |m| q dσ
k=1 Σ∩[B(x o ,2k R)\B(x o ,2k−1 R)]

∫  1/q 
∞
(n−1)/p 1/p
≤ CR |m| q dσ + σ Σ ∩ B(xo, 2k R) ×
Σ k=1
∫  1/q
× |m(x)| q |x − xo | θ |x − xo | −θ dσ(x)
Σ∩[B(x o ,2k R)\B(x o ,2k−1 R)]


∞ ∫  1/q
≤ CR(n−1)/p+λ(1/q−1) + (2k R)(n−1)/p−θ/q |m(x)| q |x − xo | θ dσ(x)
k=1 Σ


∞
≤ CR(n−1)/p+λ(1/q−1) + (2k R)(n−1)/p−θ/q Rλ(1/q−1)+θ/q
k=1


∞ 
≤ CR(n−1)/p+λ(1/q−1) 2−k[θ/q−(n−1)/p]
k=1

= CR(n−1)/p+λ(1/q−1), (6.1.34)

for some constant C ∈ (0, ∞) independent of m, since the series in the curly brackets
converges given that θ/q −(n−1)/p = [θ −(n−1)q/p]/q = [θ −(n−1)(q −1)]/q > 0.
Hence,

m ∈ L 1 (Σ, σ) and m L 1 (Σ,σ) ≤ CR(n−1)/p+λ(1/q−1) . (6.1.35)

Together with the first property in (6.1.33) this also proves that

any H q,λ,θ -dome on Σ belongs to L r (Σ, σ). (6.1.36)
1≤r ≤q

A function m ∈ L q (Σ, σ) is said to be a H q,λ,θ -molecule on Σ provided

∫
m is a H q,λ,θ
-dome on Σ and m dσ = 0. (6.1.37)
Σ

In particular, (6.1.34) ensures that the last condition in (6.1.37) is meaningful. Our
next lemma extends an Euclidean result found in [10, Proposition 17.4.2, p. 312].
302 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Lemma 6.1.2 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular set,

and abbreviate σ := H n−1 Σ. Also, fix q ∈ (1, ∞) along with λ ∈ (0, n − 1), then
pick θ > (n − 1)(q − 1). Then

there exists C = C(Σ, q, λ, θ) ∈ (0, ∞) such that any H q,λ,θ -molecule

(6.1.38)
m on Σ belongs to the space H q,λ (Σ, σ) and has mH q, λ (Σ,σ) ≤ C.

Proof If cΣ, CΣ denote the lower and upper Ahlfors regularity constants of Σ, then
0 < cΣ ≤ CΣ < ∞. Throughout, fix a number
1/(n−1)
2∗ > CΣ /cΣ . (6.1.39)

Let us first treat the case when Σ is unbounded. In such a scanario, the choice
made in (6.1.39) ensures that we have

σ Δ(x, ρ) \ Δ(x, ρ/2∗ ) ≈ ρn−1,

(6.1.40)
uniformly for x ∈ Σ and ρ ∈ (0, ∞).

Given a H q,λ,θ -molecule m on Σ, let xo ∈ Σ and R ∈ (0, ∞) be as in (6.1.33). From

(6.1.35) we know that m ∈ L 1 (Σ, σ).
j
For each j ∈ N0 abbreviate Δ j := Σ ∩ B(xo, 2∗ R) and set Δ−1 := . For each
j ∈ N0 let us also introduce A j := Δ j \ Δ j−1 and observe that thanks to (6.1.40) we
have
j
σ(A j ) ≈ (2∗ R)n−1 uniformly for j ∈ N0 . (6.1.41)
⨏
In particular, it is meaningful to define m A j := A m dσ for each j ∈ N0 . Next, fix
j
some N ∈ N and write

N 
N ∫
1Aj
mAj 1Aj = m dσ (6.1.42)
j=0 j=0
σ(A j ) Δ j \Δ j−1

∫ ∫

N
1Aj  
= m dσ − m dσ
j=0
σ(A j ) Σ\Δ j−1 Σ\Δ j

 1A ∫ ∫

N −1
j+1 1Aj  1AN
= − m dσ − m dσ.
j=0
σ(A j+1 ) σ(A j ) Σ\Δ j σ(A N ) Σ\Δ N

The last equality above uses the summation by parts formula, to the effect that

for any a0, a1, . . . , a N ∈ C and b−1, b0, . . . , b N ∈ C we have


N 
N −1
(6.1.43)
a j (b j−1 − b j ) = (a j+1 − a j )b j − a N b N + a0 b−1 .
j=0 j=0

In the context of (6.1.42), at any point x ∈ Σ this is used with

6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 303

a j := 1 A j (x)/σ(A j ) for 0 ≤ j ≤ N and

∫ (6.1.44)
b j := m dσ for −1 ≤ j ≤ N.
Σ\Δ j

The vanishing moment property of the molecule m translates into b−1 = 0, so the
very last term in (6.1.43) presently drops out. This explains the structure of the
expression in the final line of (6.1.42). Going further, express


N 
N 
N 
N
m − mAj 1Aj = m1 A j − m A j 1 A j = m1Δ N − mAj 1Aj . (6.1.45)
j=0 j=0 j=0 j=0

Hence, if for each j ∈ N0 we introduce

 1A ∫
j+1 1Aj 
α j := (m − m A j 1 A j and β j := − m dσ, (6.1.46)
σ(A j+1 ) σ(A j ) Σ\Δ j

then combining (6.1.42) and (6.1.45)-(6.1.46) we arrive at the formula


N 
N −1 ∫
1AN
m1Δ N = αj + βj − m dσ. (6.1.47)
j=0 j=0
σ(A N ) Σ\Δ N

Note that for each j ∈ N0 we have supp α j ⊆ Δ j ,

∫ ∫ ∫ ∫
α j dσ = m − m A j dσ = m dσ − m dσ = 0. (6.1.48)
Σ Aj Aj Aj

Also, if j ≥ 1
∫  1/q ∫  1/q
α j  L q (Σ,σ) ≤ 2 |m| q dσ =2 |m(x)| q |x − xo | θ |x − xo | −θ dσ(x)
Aj Aj

∫  1/q
≤ C(2∗ R)−θ/q |m(x)| q |x − xo | θ dσ(x)
j
Σ

−jθ/q
≤ C(2∗ R)−θ/q Rλ(1/q−1)+θ/q = C2∗ Rλ(1/q−1)
j

−j[θ/q+λ(1/q−1)]
(2∗ R)λ(1/q−1),
j
= C2∗ (6.1.49)

while if j = 0
∫  1/q ∫  1/q
α0  L q (Σ,σ) ≤ 2 |m| q dσ ≤2 |m| q dσ ≤ 2Rλ(1/q−1) . (6.1.50)
A0 Σ

Thus, for each j ∈ N0 ,

304 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

−j[θ/q+λ(1/q−1)]
if λ j := C2∗ then 
α j := α j /λ j is a H q,λ
-atom on Σ. (6.1.51)

Also, for each j ∈ N0 we have supp β j ⊆ Δ j+1 ,

∫  
∫  ∫ 1A ∫
1Aj
j+1
β j dσ = m dσ dσ − dσ = 0, (6.1.52)
Σ Σ\Δ j Σ σ(A j+1 ) Σ σ(A j )

and
 
∫  1 A j+1 1Aj
 β j  L q (Σ,σ) ≤ |m| dσ +
Σ\Δ j σ(A j+1 ) L q (Σ,σ) σ(A j ) L q (Σ,σ)

∫ 
= |m(x)||x − xo | θ/q |x − xo | −θ/q dσ(x) ×
Σ\Δ j
 
× σ(A j+1 )1/q−1 + σ(A j )1/q−1
∫  1/q
≤ C(2∗ R)(n−1)(1/q−1) |m(x)| q |x − xo | θ dσ(x)
j
×
Σ
∫ dσ(x)  1/p
×
Σ\Δ j |x − xo | pθ/q

≤ C(2∗ R)(n−1)(1/q−1) Rλ(1/q−1)+θ/q (2∗ R)(n−1)/p−θ/q

j j

−jθ/q −j[θ/q+λ(1/q−1)]
Rλ(1/q−1) = C2∗ (2∗ R)λ(1/q−1) .
j
= C2∗ (6.1.53)

Hence, for each j ∈ N0 ,

−j[θ/q+λ(1/q−1)]
if η j := C2∗ then βj := β j /η j is a H q,λ
-atom on Σ. (6.1.54)

We may then re-write (6.1.47) as


N 
N −1 ∫
1AN
m1Δ N = λj
αj + η j βj − m dσ, (6.1.55)
j=0 j=0
σ(A N ) Σ\Δ N

where

α j and βj are H

 q,λ
-atoms on Σ for each j ∈ N0 (6.1.56)

and

∞ 
∞
−j[θ/q+λ(1/q−1)]
|λ j | + |η j | ≤ C 2∗ < +∞, (6.1.57)
j=0 j=0

given that θ/q + λ(1/q − 1) > (n − 1)(q − 1)/q + λ(1/q − 1) = (n − 1 − λ)(1 − 1/q) > 0
since θ > (n − 1)(q − 1), q ∈ (1, ∞), and λ ∈ (0, n − 1). Since m ∈ L 1 (Σ, σ) (cf.
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 305

1A ∫
(6.1.35)), it follows that both m1Δ N → m and σ(ANN ) Σ\Δ m dσ → 0 in L 1 (Σ, σ)
N
as N → ∞. Consequently, from (6.1.55) we deduce that

∞ 
∞

m= λj
αj + η j βj in Lipc (Σ) , (6.1.58)
j=0 j=0

while (6.1.57) ensures that


∞
C := |λ j | + |η j | < +∞ is independent of m. (6.1.59)
j=0

In view of (6.1.56) and (6.1.17), the claims recorded in (6.1.38) now follow from
(6.1.58) and (6.1.59). This finishes the proof of the lemma in the case when Σ is
unbounded.
The argument in the case when Σ is bounded follows along similar lines, with
some important alterations, indicated below. For starters, when Σ is bounded the
choice made in (6.1.39) only guarantees that

σ Δ(x, ρ) \ Δ(x, ρ/2∗ ) ≈ ρn−1 uniformly for

(6.1.60)
x ∈ Σ and ρ ∈ 0, 2 diam(Σ) .

To proceed, given a H q,λ,θ -molecule m on Σ, let xo ∈ Σ and R ∈ 0, 2 diam(Σ) be

as in (6.1.33). Once again, set Δ−1 :=  and, for each j ∈ N0 , define the surface balls
j
Δ j := Σ ∩ B(xo, 2∗ R) and the annuli A j := Δ j \ Δ j−1 as before. If we now introduce
j
NR := min j ∈ N : 2∗ R ≥ 2 diam(Σ) , (6.1.61)

then thanks to (6.1.40) and the definition of NR we have (compare with (6.1.41))
j
σ(A j ) ≈ (2∗ R)n−1 uniformly for j ≤ NR − 1, (6.1.62)

and
A j =  whenever j ≥ NR + 1. (6.1.63)
Moreover, corresponding to j = NR ,

Δ N R = Σ and A N R = Σ \ Δ N R −1 (6.1.64)

which, in particular, implies that

∫
m dσ = 0 if σ(A N R ) = 0. (6.1.65)
Σ\Δ N R −1

Glancing back at the definition of NR given in (6.1.61) we see that

306 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

2∗N R R ≥ 2 diam(Σ) and 2∗N R −1 R < 2 diam(Σ)

(6.1.66)
hence 2∗N R R ≈ diam(Σ) uniformly with respect to R.

We now fix an arbitrary point x ∈ Σ and write the summation by parts formula

(6.1.43) for N := NR and the following choices: for each j ∈ {0, 1, . . . , NR } take
⎧
⎪ 1 A j (x)/σ(A j ) if j ≤ NR − 1,
⎪
⎪
⎨
⎪
a j := 1 A N R (x)/σ(A N R ) if j = NR and σ(A N R ) > 0, (6.1.67)
⎪
⎪
⎪
⎪0
⎩ if j = NR and σ(A N R ) = 0,

while for each j ∈ {−1, 0, . . . , NR } take

∫
b j := m dσ. (6.1.68)
Σ\Δ j

As a consequence of (6.1.68), (6.1.64), and the vanishing moment condition for m,

b N R = 0 and b−1 = 0. (6.1.69)

For these choices, (6.1.43) gives (bearing in mind (6.1.64)-(6.1.65))


NR N
R −1

m= αj + βj (6.1.70)
j=0 j=0

where, if for each j ∈ {0, 1, . . . , NR },

⨏
⎧
⎪ m dσ if j ≤ NR − 1,
⎪
⎪ A
⎨⨏ j
⎪
α j := (m − m A j 1 A j with m A j := m dσ if j = NR and σ(A N R ) > 0,
⎪
⎪ ANR
⎪
⎪0
⎩ if j = NR and σ(A N R ) = 0,
(6.1.71)
and, for each j ∈ {0, 1, . . . , NR − 1},
 1A ∫
⎧
⎪ j+1 1Aj
⎪
⎪ σ(A ) − σ(A ) m dσ if j ≤ NR − 2,
⎪
⎨
⎪ j+1 j Σ\Δ j
1 A N −1  ∫
β j := 1AN
⎪ σ(A N R ) − σ(A N R −1 ) Σ\Δ N R −1 m dσ if j = NR − 1 and σ(A N R ) > 0,
R R
⎪
⎪
⎪
⎪0
⎩ if j = NR − 1 and σ(A N R ) = 0.
(6.1.72)
Much as in (6.1.48), (6.1.52), these formulas imply that
∫
α j dσ = 0 for each j ∈ {0, 1, . . . , NR }, (6.1.73)
Σ

and
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 307
∫
β j dσ = 0 for each j ∈ {0, 1, . . . , NR − 1}. (6.1.74)
Σ
Also, for each j ∈ {0, 1, . . . , NR } we have supp α j ⊆ Δ j and the same type of
estimates as in (6.1.49)-(6.1.50) give that
−j[θ/q+λ(1/q−1)]
(2∗ R)λ(1/q−1)
j
α j  L q (Σ,σ) ≤ C2∗
(6.1.75)
for j ∈ {0, 1, . . . , NR }.

Thus,
−j[θ/q+λ(1/q−1)]
if λ j := C2∗ then 
α j := α j /λ j
(6.1.76)
is a H q,λ -atom on Σ, for each {0, 1, . . . , NR }.
Going further, for each j ∈ {0, 1, . . . , NR − 1} we have supp β j ⊆ Δ j+1 and the
same type of estimate as in (6.1.53) shows that
−j[θ/q+λ(1/q−1)]
(2∗ R)λ(1/q−1)
j
 β j  L q (Σ,σ) ≤ C2∗
(6.1.77)
for j ∈ {0, 1, . . . , NR − 2}.

As far as the case j = NR − 1 is concerned, observe first that

1 A N R −1 ∫ −N R [θ/q+λ(1/q−1)]
m dσ ≤ C2∗ (2∗N R R)λ(1/q−1)
σ(A N R −1 ) Σ\Δ N R −1
L q (Σ,σ)

≤ C(2∗N R R)λ(1/q−1) . (6.1.78)

Indeed, the first estimate above is justified by the same type of argument used in
(6.1.53) (thanks to (6.1.62) for j = NR − 1), while the second estimate is a simple
consequence of the fact that 2∗ > 1 and θ/q + λ(1/q − 1) > 0.
We complement the estimate established in (6.1.78) by showing that there exists
a constant C = C(Σ, q, λ, θ) ∈ (0, ∞) (in particular, independent of x0, R, m) with the
property that in the case when σ(A N R ) > 0 we also have

1 ANR ∫
m dσ ≤ C(2∗N R R)λ(1/q−1) . (6.1.79)
σ(A N R ) Σ\Δ N R −1
L q (Σ,σ)

(The main difference between (6.1.78) and (6.1.79) is the absence of a concrete
lower bound for σ(A N R ) of the sort appearing in (6.1.62).) To justify (6.1.79),
assume σ(A N R ) > 0 and first estimate
308 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .
 ∫ 
 
 1 
 m dσ 
 σ(A N R ) Σ\Δ N −1

R
∫
1
≤ |m(x)||x − xo | θ/q |x − xo | −θ/q dσ(x)
σ(A N R ) Σ\Δ N R −1
∫
1
≤ C(2∗N R R)−θ/q |m(x)||x − xo | θ/q dσ(x)
σ(A N R ) Σ\Δ N R −1
⨏
−θ/q
NR
= C(2∗ R) |m(x)||x − xo | θ/q dσ(x)
ANR

⨏  1/q
≤ C(2∗N R R)−θ/q |m(x)| q |x − xo | θ dσ(x)
ANR

1 ∫  1/q
≤ C(2∗N R R)−θ/q |m(x)| q |x − xo | θ dσ(x)
σ(A N R )1/q Σ

1
≤ C(2∗N R R)−θ/q Rλ(1/q−1)+θ/q
σ(A N R )1/q
1 λ(1/q−1)+θ/q
≤ C(2∗N R R)−θ/q 2 diam(Σ)
σ(A N R )1/q
1
≤ C(2∗N R R)−θ/q (2∗N R R)λ(1/q−1)+θ/q
σ(A N R )1/q
C
= (2∗N R R)λ(1/q−1), (6.1.80)
σ(A N R )1/q

where C = C(Σ, q, λ, θ) ∈ (0, ∞) is a constant independent of xo, R, m. Above, the

first two inequalities are clear, the first equality is based on (6.1.64) and, subsequently,
we have used Hölder’s inequality. Next, we unpack the integral average and extend
the domain of integration, then we rely on the second inequality in (6.1.33). The fact
that R ∈ 0, 2 diam(Σ) and λ(1/q − 1) + θ/q > 0 permit us to justify the penultimate
inequality. The last inequality is implied by (6.1.66). The final equality in (6.1.80) is
obvious.
Having established (6.1.80), the claim made in (6.1.79) follows by writing
 
1 ANR ∫ 
 1
∫ 

m dσ = m dσ  1 A N R q
σ(A N R ) Σ\Δ N −1 q
 σ(A NR ) Σ\Δ N −1  L (Σ,σ)
R L (Σ,σ) R

C
≤ (2∗N R R)λ(1/q−1) σ(A N R )1/q
σ(A N R )1/q

= C(2∗N R R)λ(1/q−1) . (6.1.81)

6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 309

Collectively, (6.1.72) (with j := NR − 1), (6.1.78), (6.1.79), (6.1.66), and (6.1.64)

prove that
λ(1/q−1)
 β N R −1  L q (Σ,σ) ≤ C diam(Σ) and supp β N R −1 ⊆ Δ N R = Σ. (6.1.82)

As such, from (6.1.77), (6.1.82), (6.1.74), the fact that

supp β j ⊆ Δ j+1 for each j ∈ {0, 1, . . . , NR − 1}, (6.1.83)

and (6.1.64) we conclude that if

 −j[θ/q+λ(1/q−1)]
C2∗ if j ∈ {0, 1, . . . , NR − 2}
η j := (6.1.84)
C if j = NR − 1,

then

βj := β j /η j is a H q,λ
-atom on Σ for each j ∈ {0, 1, . . . , NR − 1}. (6.1.85)

Additionally, it is apparent from definitions that


NR N
R −1 
∞ 
−j[θ/q+λ(1/q−1)]
|λ j | + |η j | ≤ C 2∗ + C =: C(Σ, q, λ, θ) < +∞, (6.1.86)
j=0 j=0 j=0

since θ/q + λ(1/q − 1) > 0. From (6.1.70), (6.1.76), (6.1.84), (6.1.86), and (6.1.16)
we see that the function

NR N
R −1

m= λj
αj + η j βj (6.1.87)
j=0 j=0

belongs to the space H q,λ (Σ, σ) and satisfies mH q, λ (Σ,σ) ≤ C(Σ, q, λ, θ). This
establishes the desired conclusions in the case when Σ is bounded, so the proof of
Lemma 6.1.2 is complete. 

It is also useful to observe (from our earlier definitions) that, given q ∈ (1, ∞) and
λ ∈ (0, n − 1),

any H q,λ -atom on Σ is also a H q,λ,θ -molecule on Σ for any choice

(6.1.88)
of the parameter θ in the interval (n − 1)(q − 1), ∞ .

In the lemma below we address the fundamental issue of correlating the integral
pairing of a function belonging to the homogeneous Morrey-Campanato space (de-
fined on a given Ahlfors regular set) with a molecule belonging to the pre-dual of this
space. This compatibility result is essential in establishing boundedness properties
for singular integral operators (of double layer type) acting on Morrey-Campanato
spaces, a topic considered a little later in this section.
310 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Lemma 6.1.3 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed set which is Ahlfors

regular and abbreviate σ := H n−1 Σ. Also, pick two exponents p, q ∈ (1, ∞) such
that 1/p + 1/q = 1, along with
. two parameters λ ∈ (0, n − 1) and θ > (n − 1)(q − 1).
Then for any function f ∈ L p,λ (Σ, σ) and any H q,λ,θ -molecule m on Σ (cf. (6.1.37))
one has
∫ ∫  
[ f ], m if Σ is unbounded,
| f ||m| dσ < +∞ and f m dσ =   (6.1.89)
Σ Σ f, m if Σ is bounded,

where ·, · denotes the duality bracket between the homogeneous Morrey-Campanato
space and its pre-dual (cf. (6.1.25)). .
Furthermore, for each function f ∈ L p,λ (Σ, σ) ⊆ L p,λ (Σ, σ) and each function
g ∈ H q,λ (Σ, σ) with the property that
∫
| f ||g| dσ < +∞ (6.1.90)
Σ

one has
∫  
[ f ], g if Σ is unbounded,
f g dσ =   (6.1.91)
Σ f, g if Σ is bounded,
with ·, · denoting, as before, the duality bracket between the inhomogeneous
Morrey-Campanato space and its pre-dual.

Proof To fix ideas, suppose first that the set Σ is unbounded. ⨏ For each x ∈ Σ and
R > 0 abbreviate Δ(x, R) := Σ ∩ B(x, R) and fΔ(x,R) := Δ(x,R) f dσ. Fix a point
x ∈ Σ along with a scale R ∈ (0, ∞), then use [133, (7.4.121)] with X := Σ, μ := σ,
ρ := | · − · |, q := p, x0 := x, and r := R to write for each j ∈ N0
⨏    1/p
 f − fΔ(x,R)  p dσ
Δ(x,2 j+1 R)
 

j λ−n+1 n−1−λ  ⨏    1/p
≤C +1
(2 R) p (2+1 R) p  f − fΔ(x,2+1 R)  p dσ
=0 Δ(x,2+1 R)


j λ−n+1 λ−n+1
≤ C f L p, λ (Σ,σ) (2+1 R) p ≤ C f L p, λ (Σ,σ) R p , (6.1.92)
=0


j λ−n+1
since (λ − n + 1)/p < 0 ensures that (2+1 ) p < +∞. To proceed, let xo ∈ Σ
=0
and R ∈ (0, ∞) be as in (6.1.33) for the given H q,λ,θ -molecule m on Σ. We may
then estimate:
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 311
∫
 
 f − fΔ(x ,R)  |m| dσ
o
Σ\Δ(xo ,R)

∞ ∫
  1/p
 
≤  f − fΔ(x ,R)  p dσ ×
o
k=1 Δ(x o ,2k R)\Δ(x o ,2k−1 R)

∫  1/q
× |m| q dσ
Δ(x o ,2k R)\Δ(x o ,2k−1 R)


∞ ⨏    1/p
= σ Δ(xo, 2 R) k 1/p  f − fΔ(x ,R)  p dσ ×
o
k=1 Δ(x o ,2k R)

∫  1/q
× |m(x)| q |x − xo | θ |x − xo | −θ dσ(x)
Δ(x o ,2k R)\Δ(x o ,2k−1 R)

λ−n+1 
∞ 
≤ C f L p, λ (Σ,σ) R p (2k R)(n−1)/p−θ/q ×
k=1
∫  1/q
× |m(x)| q |x − xo | θ dσ(x)
Σ

λ−n+1 
∞ 
≤ C f L p, λ (Σ,σ) R p (2k R)(n−1)/p−θ/q Rλ(1/q−1)+θ/q
k=1

λ−n+1 n−1 θ 1 θ 
∞ 
= C f L p, λ (Σ,σ) · R p + p − q +λ q −1 + q 2−k[θ/q−(n−1)/p]
k=1

= C f L p, λ (Σ,σ) . (6.1.93)

The first inequality above is obtained by breaking up the domain of integration into
mutually disjoint dyadic annuli and then using Hölder’s inequality (keeping in mind
that 1/p + 1/q = 1). The subsequent equality is clear, and the second inequality
is a consequence of (6.1.92) and the upper Ahlfors regularity of Σ. Next, the third
inequality follows from (6.1.33), while the ensuing equality is just algebra. Finally,
the last equality uses the fact that the series in the curly brackets converges, since
θ/q − (n − 1)/p = [θ − (n − 1)q/p]/q = [θ − (n − 1)(q − 1)]/q
∫ > 0. Since from (6.1.36)
we know that m ∈ L 1 (Σ, σ), the argument so far gives Σ\B(x ,R) | f ||m| dσ < +∞.
∫ o
As such, there remains to show that Σ∩B(x ,R) | f ||m| dσ < +∞. This, however, is
o
a consequence of (6.1.3) and (6.1.36). The first property in (6.1.89) has now been
established if Σ is unbounded. In the case when Σ is bounded, the first property
in (6.1.89) follows by observing that we now have f ∈ L p (Σ, σ) (cf. (6.1.13)) and
recalling from (6.1.36) that m ∈ L q (Σ, σ).
To prove the second property in (6.1.89), suppose first that Σ is unbounded. Note
that, in light of the nature of the conclusion we seek, there is no loss of generality
312 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

in assuming that the given function f is real-valued. For each M ∈ N we may then
define fM := min max{ f , −M }, M . Lemma 6.1.1 ensures that
.
{ fM } M ∈N ⊆ L ∞ (Σ, σ) ∩ L p,λ (Σ, σ) and
.  (6.1.94)
lim [ fM ] = [ f ] weak-∗ in L p,λ (Σ, σ) ∼ .
M→∞

As such, since m belongs to H q,λ (Σ, σ) (cf. (6.1.38)), we may write

   
[ f ], m = lim [ fM ], m . (6.1.95)
M→∞
∞
To proceed, express m ∈ H q,λ (Σ, σ) as a series m = j=1 λ j a j convergent in
H q,λ (Σ, σ), where {λ j } j ∈N ∈  1 (N) and each a j is a H q,λ -atom on Σ. Then for

each M ∈ N fixed, (6.1.26)-(6.1.27) permit us to write

∫  
  N
[ fM ], m = lim fM λ j a j dσ. (6.1.96)
N →∞ Σ j=1

Let xo ∈ Σ and R ∈ 0, 2 diam(Σ) be such that m satisfies the properties listed in

(6.1.33). Also, recall the number 2∗ from (6.1.39). For each j ∈ N0 let us now define
j
Δ j := Σ ∩ B(xo, 2∗ R), set Δ−1 := , and introduce A j := Δ j \ Δ j−1 . Then from
(6.1.55)-(6.1.57) it follows that matters may be arranged so that for each N ∈ N we
have
 N ∫
1AN
λ j a j = m1Δ N + m dσ. (6.1.97)
j=1
σ(A N ) Σ\Δ N

The availability of such an explicit formula for the partial sums of an atomic decom-
position series for m is a crucial element in the present proof. For this allows us to
compute, on the one hand,
∫ 
N  ∫  ∫ 
1AN
lim fM λ j a j dσ = lim fM m1Δ N + m dσ dσ
N →∞ Σ j=1
N →∞ Σ σ(A N ) Σ\Δ N
∫
= lim fM m dσ
N →∞ ΔN
∫  1
∫ 
+ lim m dσ fM dσ .
N →∞ Σ\Δ N σ(A N ) AN
(6.1.98)

On the other hand, the fact that m ∈ L 1 (Σ, σ) and fM ∈ L ∞ (Σ, σ) implies
6.1 Morrey-Campanato Spaces and Their Pre-Duals on Ahlfors Regular Sets 313
∫ ∫ ∫
lim fM m dσ = fM m dσ, lim m dσ = 0,
N →∞ ΔN Σ N →∞ Σ\Δ N
 ∫  (6.1.99)
 1 
and lim sup  fM dσ  ≤  fM  L ∞ (Σ,σ) .
N →∞ σ(A N ) AN

All together, from (6.1.96), (6.1.98), and (6.1.99) we deduce that

∫
 
[ fM ], m = fM m dσ, ∀M ∈ N. (6.1.100)
Σ

Given that fM → f pointwise∫ on Σ as M → ∞, that | fM | ≤ | f | for each M ∈ N

and that, as already noted, Σ | f ||m| dσ < +∞, Lebesgue’s Dominated Convergence
Theorem applies and gives that
∫ ∫
lim fM m dσ = f m dσ. (6.1.101)
M→∞ Σ Σ

At this stage, the second property claimed in (6.1.89) follows in the case when Σ is
unbounded by combining (6.1.95), (6.1.100), and (6.1.101).
To justify the second property in (6.1.89) in the case when Σ is bounded, recall
.
from (6.1.13) and (6.1.11) that f ∈ L p,λ (Σ, σ) = L p,λ (Σ, σ) ⊆ L p (Σ, σ). Also, in
the current setting, ⨏
  ⨏
m= m− m dσ + m dσ (6.1.102)
Σ Σ

constitutes an atomic decomposition of the function m ∈ L q (Σ, σ) ⊆ H q,λ (Σ, σ)

(in the sense of (6.1.16)), so the discussion regarding the sense in which (6.1.25) is
to be understood implies that
∫  ⨏  ∫ ⨏  ∫
 f , m = m− m dσ f dσ + m dσ f dσ = f m dσ, (6.1.103)
Σ Σ Σ Σ Σ

as wanted.
Let us now turn to the task of proving (6.1.91) for any given f ∈ L p,λ (Σ, σ) and
g ∈ H q,λ (Σ, σ) satisfying (6.1.90). As above, the crux of the matter is establishing
that for each fixed M ∈ N we have
∫ 
N  ∫
lim fM λ j a j dσ = fM g dσ, (6.1.104)
N →∞ Σ Σ
j=1


if g = ∞j=1 λ j a j in H
q,λ (Σ, σ), where {λ }
j j ∈N ∈  (N) and each a j is a H
1 q,λ -atom

on Σ. To this end, we first recall from (6.1.22) that


N
q(n − 1)
lim λ j a j = g in L r (Σ, σ) where r := . (6.1.105)
N →∞
j=1
n − 1 + λ(q − 1)
314 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Note that 1/p + 1/r = 1/p + 1/q + [λ(q − 1)]/[q(n − 1)] > 1. Hence, if r  ∈ (1, ∞)
is such that 1/r + 1/r  = 1, then necessarily r  ∈ (p, ∞). In particular, for each

M ∈ N the function fM belongs to L ∞ (Σ, σ) ∩ L p (Σ, σ) ⊆ L r (Σ, σ). Granted this
membership, we may now conclude from (6.1.105) that (6.1.104) is indeed valid.
The desired conclusion now follows by reasoning as before. 

6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets

We begin by discussing the scale of Morrey spaces on Ahlfors regular sets. As in the
past, we assume that Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a closed Ahlfors regular
set, and that σ := H n−1 Σ. Given p ∈ (0, ∞) and λ ∈ (0, n − 1), we then define the
Morrey space M p,λ (Σ, σ) as
 
M p,λ (Σ, σ) := f : Σ → C : f is σ-measurable and  f  M p, λ (Σ,σ) < +∞
(6.2.1)
where, for each σ-measurable function f on Σ, we have set
 
n−1−λ  ⨏  p1
 f  M p, λ (Σ,σ) := sup R p | f | p dσ . (6.2.2)
x ∈Σ and Σ∩B(x,R)
0<R<2 diam(Σ)

It is then apparent from the above definitions that

if f , g : Σ −→ C are two σ-measurable functions such that

|g| ≤ | f | at σ-a.e. point on Σ and f ∈ M p,λ (Σ, σ), then g (6.2.3)
belongs to M p,λ (Σ, σ) and g M p, λ (Σ,σ) ≤  f  M p, λ (Σ,σ) ,

that
for each σ-measurable function f : Σ −→ C and each given
power θ ∈ (0, ∞) we have | f | θ M p, λ (Σ,σ) =  f  M
θ (6.2.4)
p θ, λ (Σ,σ) ,

and that
the operator of pointwise multiplication by a given function
b ∈ L ∞ (Σ, σ) is a bounded mapping from the Morrey space (6.2.5)
M p,λ (Σ, σ) into itself, with operator norm ≤ b L ∞ (Σ,σ) .

Another observation worth making here is that (6.2.1)-(6.2.2) together with (6.2.3)
and Lebesgue’s Monotone Convergence Theorem imply the following monotonicity
property of the Morrey norm:

if { f j } j ∈N is a sequence of non-negative σ-measurable functions on

Σ which converges to some function f in a non-decreasing fashion at (6.2.6)
σ-a.e. point on Σ, then  f j  M p, λ (Σ,σ)   f  M p, λ (Σ,σ) as j → ∞.
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 315

Retaining the above setting, as may be seen from (6.1.6) and (6.2.1)-(6.2.2) we
have
p
L s (Σ, σ) → M p,λ (Σ, σ) → Lloc (Σ, σ)
p(n−1)
(6.2.7)
continuously, with s := n−1−λ ∈ (p, ∞).

In particular, the continuity of the second inclusion in (6.2.7) may be used to prove
that2
M p,λ (Σ, σ) is a Banach space if p ≥ 1, and a p-Banach space if
0 < p < 1; also, for any f , g ∈ M p,λ (Σ, σ), in all cases (cf. (1.2.4)) (6.2.8)
one has  f + g M p, λ (Σ,σ) ≤ 2(1/p−1)+  f  M p, λ (Σ,σ) + g M p, λ (Σ,σ) .

Also, the continuity of the first inclusion in (6.2.7) yields useful estimates in the
Morrey norm for characteristic functions of measurable sets. Specifically, there exists
some C ∈ (0, ∞) which depends only on n, p, λ, and the upper Ahlfors regularity
constant of Σ, with the property that for each σ-measurable set E ⊆ Σ we have

1E M p, λ (Σ,σ)
≤ C 1E L s (Σ,σ)
= C · σ(E)(n−1−λ)/[p(n−1)] . (6.2.9)

In particular, from this and (A.0.107) we see that

Sfin (Σ, σ) ⊂ M p,λ (Σ, σ) (6.2.10)

(though the Morrey space M p,λ (Σ, σ) is not separable, generally speaking). Other
examples of functions belonging to Morrey spaces are presented below.

Example 6.2.1 As before, let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors

regular set, and abbreviate σ := H n−1 Σ. Also, select p ∈ (0, ∞) and λ ∈ (0, n − 1).
For each fixed point xo ∈ Σ consider the function fxo : Σ → R defined for each
x ∈ Σ \ {xo } as fxo (x) := |x − xo | −(n−1−λ)/p . Then each fxo belongs to the Morrey
space M p,λ (Σ, σ) and, in fact,

sup  fxo  M p, λ (Σ,σ) < +∞. (6.2.11)

x o ∈Σ

p(n−1)
This being said, each fxo fails to be in L s (Σ, σ) with s := n−1−λ , so the inclusions
in (6.2.7) are strict.

The claim made in (6.2.11) is justified by studying two cases. First, whenever x ∈ Σ
and 0 < R < |x − xo |/2 we have |x − xo |/2 < |y − xo | < 3|x − xo |/2 for each
y ∈ B(x, R), so | fxo (y)| ≈ |x − xo | −(n−1−λ)/p , uniformly for y ∈ Σ ∩ B(x, R). As such,

2 recall that (a)+ := max{a, 0} for each a ∈ R

316 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .
⨏    1/p
R(n−1−λ)/p  fx (y) p dσ(y)
o
Σ∩B(x,R)

≈ R(n−1−λ)/p |x − xo | −(n−1−λ)/p
(n−1−λ)/p
= R/|x − xo | < (1/2)(n−1−λ)/p, (6.2.12)

in this case. Finally, if x ∈ Σ and R ≥ |x − xo |/2 it follows that B(x, R) ⊆ B(xo, 3R)
which, together with the first estimate in [133, (7.2.5)], further entails
⨏    1/p
R(n−1−λ)/p  fx (y) p dσ(y)
o
Σ∩B(x,R)
∫ dσ(y)  1/p
−λ/p
≈R
Σ∩B(x,R) |y − xo | n−1−λ
∫ dσ(y)  1/p
≤ R−λ/p
Σ∩B(x o ,3R) |y − xo | n−1−λ

≤ C(Σ, n, p, λ) < +∞. (6.2.13)

At this stage, the claim in (6.2.11) is seen from (6.2.12)-(6.2.13) and (6.2.2).
In view of the above considerations it is of interest to define the space
p(n−1)
M̊ p,λ (Σ, σ) := the closure of L s (Σ, σ) with s := n−1−λ in M p,λ (Σ, σ). (6.2.14)

Hence, by design,

M̊ p,λ (Σ, σ) is a closed linear subspace of M p,λ (Σ, σ) with the property
(6.2.15)
that L s (Σ, σ) → M̊ p,λ (Σ, σ) continuously and densely.

As a consequence of this and the fact that L s (Σ, σ) is a separable Banach space (see
[133, (3.6.27)]),
M̊ p,λ (Σ, σ) is a separable Banach space. (6.2.16)
From (6.2.15) and (6.2.5) we also conclude that
the operator of pointwise multiplication by some given function
b ∈ L ∞ (Σ, σ) is a bounded mapping from the space M̊ p,λ (Σ, σ) (6.2.17)
into itself, with operator norm ≤ b L ∞ (Σ,σ) .

The latter property further implies that

if f , g : Σ → C are two σ-measurable functions such

p,λ
that |g| ≤ | f | at σ-a.e. point on Σ and f ∈ M̊ (Σ, σ), (6.2.18)
p,λ
then g also belongs to the space M̊ (Σ, σ),
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 317

since

g/ f if f  0,
g = b f with b := satisfying b L ∞ (Σ,σ) ≤ 1. (6.2.19)
0 if f = 0,

In relation to the space introduced in (6.2.14) we wish to remark that since [133,
Corollary 3.7.3] guarantees that Lipc (Σ) is dense in L s (Σ, σ) and since, according
to (6.2.7), the latter space embeds continuously into M p,λ (Σ, σ), we have

M̊ p,λ (Σ, σ) = the closure of Lipc (Σ) in M p,λ (Σ, σ). (6.2.20)

An immediate corollary of the latter description of the space M̊ p,λ (Σ, σ) worth men-
tioning is that functions f belonging to M̊ p,λ (Σ, σ) enjoy the “vanishing” property
 
n−1−λ  ⨏  p1
lim+ sup R p | f | dσ
p
= 0. (6.2.21)
ρ→0 x ∈Σ and Σ∩B(x,R)
R ∈(0,ρ)

Our final comment pertaining to (6.2.14) at the moment is that from (6.2.15) and
duality it follows that
 ∗ p(n − 1)
M̊ p,λ (Σ, σ) → L r (Σ, σ) with r = . (6.2.22)
(n − 1 − λ)(p − 1)
Later on, in Proposition 6.2.16, we shall identify the above dual space more precisely.
Moving on, it is also useful to observe that we have the continuous embedding
(compare with (6.1.14))
 σ(x) 
M p,λ (Σ, σ) → L τ Σ, for each
1 + |x| a (6.2.23)
(n−1−λ)τ
τ ∈ (0, p] and a > n − 1 − p .

Indeed, if Σ is bounded then this is clear from (6.2.1)-(6.2.2). If Σ is unbounded,

fix some reference point x0 ∈ Σ and for each function f ∈ M p,λ (Σ, σ), integrability
exponent τ ∈ (0, p], and power a > n − 1 − (n−1−λ)τ
p , estimate
318 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .
∫ ∫
| f (x)|τ | f (x)|τ
dσ(x) ≤ C dσ(x)
Σ 1 + |x| a Σ 1 + |x − x0 | a
∫
| f (x)|τ
≤C dσ(x)
B(x0,1)∩Σ 1 + |x − x0 | a
∞ ∫
 | f (x)|τ
+C dσ(x)
j=1 [B(x0,2 j+1 )\B(x0,2 j )]∩Σ 1 + |x − x0 | a

∫ 
∞ ∫
τ −j a
≤C | f | dσ + C 2 | f |τ dσ
B(x0,1)∩Σ j=1 B(x0,2 j+1 )∩Σ


∞ ⨏  τ/p
≤C 2−j(a−n+1) | f | p dσ
j=0 B(x0,2 j+1 )∩Σ


∞
a−n+1+
(n−1−λ)τ 
≤C (2−j ) p  f  M p, λ (Σ,σ)
j=0

= C f  M p, λ (Σ,σ), (6.2.24)

from which the desired conclusion follows. In particular, from (6.2.15) and (6.2.23)
(used with τ := p and τ := 1) we see that we have continuous embeddings
 σ(x)   σ(x) 
M̊ p,λ (Σ, σ) → M p,λ (Σ, σ) → L p Σ, ∩ L 1 Σ,
1 + |x| a 1 + |x| n−1
 σ(x) 
p
→ Lloc (Σ, σ) ∩ L 1 Σ,
1 + |x| n−1

if p ∈ (1, ∞), λ ∈ (0, n − 1), and a > λ. (6.2.25)

As it turns out, there is a natural description of Morrey spaces in terms of

the fractional Hardy-Littlewood maximal operator (cf. (A.0.68) with X := Σ), as
indicated in the lemma below.

Lemma 6.2.2 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular set,

and abbreviate σ := H n−1 Σ. Also, fix p ∈ (0, ∞) along with λ ∈ (0, n − 1), and
recall from (A.0.68) p
 1 that L -based fractional Hardy-Littlewood maximal operator
of order α ∈ 0, p on Σ acts on each σ-measurable function f on Σ as
 ⨏  p1 
α
M Σ, p,α f (x) := sup σ B(x, r) ∩ Σ | f | p dσ , ∀x ∈ Σ. (6.2.26)
r >0 B(x,r)∩Σ

Then, if
n−1−λ
α := , (6.2.27)
(n − 1)p
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 319

one has
M Σ, p,α f L ∞ (Σ,σ)
≈  f  M p, λ (Σ,σ),
(6.2.28)
uniformly for σ-measurable functions f on Σ,

and
 
M p,λ (Σ, σ) = f : Σ → C : f is σ-measurable and M Σ, p,α f ∈ L ∞ (Σ, σ) .
(6.2.29)

The description of the Morrey space in (6.2.29) offers a natural interpretation

of the embedding L s (Σ, σ) → M p,λ (Σ, σ) with s := p(n−1)
n−1−λ (cf. (6.2.7)) in view of
the fact that the fractional Hardy-Littlewood maximal operator M Σ, p,α with α as in
(6.2.27), i.e., α = 1/s, maps L s (Σ, σ) into L ∞ (Σ, σ) (see [133, (7.6.6)]).
Proof of Lemma 6.2.2 As in (4.4.42), we find it convenient to work with a version
of M Σ, p,α which exhibits better measurability properties than the original fractional
Hardy-Littlewood maximal operator. Specifically, if for each σ-measurable function
f : Σ → C we define
 ∫  1/p
Σ, p,α f (x) := sup 1
M (n−1)(1−αp)
| f | p
dσ
R ∈(0,∞) R B(x,R)∩Σ

 1 ∫  1/p
= sup λ
| f | p
dσ, ∀x ∈ Σ, (6.2.30)
R ∈(0,∞) R B(x,R)∩Σ

then, in a similar fashion to (4.4.43), we have

Σ, p,α f : Σ −→ [0, +∞] is a lower-semicontinuous function

M
(6.2.31)
for each given each σ-measurable function f : Σ → C.

The fact that Σ is an Ahlfors regular set ensures that there exist 0 < co ≤ Co < +∞
with the property that for each σ-measurable function f : Σ → C we have

Σ, p,α f ≤ Co · M Σ, p,α f everywhere on Σ.

co · M Σ, p,α f ≤ M (6.2.32)

As such, for each σ-measurable function f on Σ we may write

M Σ, p,α f Σ, p,α f

≈ M Σ, p,α f (x)
= sup M
L ∞ (Σ,σ) L ∞ (Σ,σ)
x ∈Σ

≈ sup M Σ, p,α f (x) ≈  f  M p, λ (Σ,σ), (6.2.33)

x ∈Σ

with proportionality constants independent of f . The first two equivalences above

are implied by (6.2.32), the equality is a consequence of (6.2.31), and the last
equivalence is clear from (6.2.26), (6.2.27), and (6.2.2). This proves (6.2.28) which,
in turn, justifies (6.2.29). 
320 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

The Morrey spaces from (6.2.1)-(6.2.2) interface tightly with the Morrey-
Campanato spaces discussed earlier. For starters let us observe that, as is apparent
from [133, (7.4.58)] and definitions, if p ∈ (1, ∞) and λ ∈ (0, n − 1) we have
.
M p,λ (Σ, σ) ⊆ L p,λ (Σ, σ) and there exists a constant C ∈ (0, ∞) such
(6.2.34)
that  f L. p, λ (Σ,σ) ≤ C f  M p, λ (Σ,σ) for all functions f ∈ M p,λ (Σ, σ).

To be able to further elaborate on this relationship, we first establish the following

result.
Lemma 6.2.3 Suppose Σ ⊆ Rn (where n ∈ N with n ≥ 2) is an unbounded closed
. p,λσ := H Σ. Also, fix p ∈ (1, ∞) and
set which is Ahlfors regular, and abbreviate n−1

λ ∈ (0, n − 1). Then for each function f ∈ L (Σ, σ) and each point x0 ∈ Σ, the limit
⨏
fΣ := lim f dσ (6.2.35)
R→∞ B(x0,R)∩Σ

exists and is actually independent of the point x0 . In addition, there exists a constant
C ∈ (0, ∞) with the property that
 
n−1−λ  ⨏   p  p1
 f − fΣ  M p, λ (Σ,σ) = sup R p  f − fΣ  dσ
x ∈Σ and Σ∩B(x,R)
R ∈(0,∞)

≤ C f L. p, λ (Σ,σ) . (6.2.36)

It is worth pointing out that, as seen with the help of Hölder’s inequality,

whenever f ∈ L q (Σ, σ) for some q ∈ [1, ∞) the

(6.2.37)
limit in (6.2.35) is zero for each point x0 ∈ Σ.
Proof of Lemma 6.2.3⨏ As in the past, for each point x ∈ Σ and each radius R > 0,
abbreviate fΔ(x,R) := B(x,R)∩Σ f dσ. Let us now fix some x ∈ Σ. Then, for each
r, R > 0 such that R > r, based on [133, Lemma 7.4.15], (A.0.83), and (6.1.2) we
may estimate
∫
  2R
dt
 fΔ(x,r) − fΔ(x,R)  ≤ C osc p ( f ; t)
r t
∫ ∞ n−1−λ
−1−
≤ C f L. p, λ (Σ,σ) t p dt
r
n−1−λ
−
≤ C f L. p, λ (Σ,σ) r p = o(1) as r → ∞. (6.2.38)

This proves that the sequence fΔ(x,r) r >0 is Cauchy, hence convergent to some
complex number which we shall call fΣ,x .
Let us now fix two arbitrary points x, y ∈ Σ. Then for each r > |x − y| we
have B(y, r) ⊆ B(x, 2r) so we may invoke [133, (7.4.56)] (with p := 1), Hölder’s
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 321

inequality, the Ahlfors regularity of Σ, (6.1.2), and what we have just proved in order
to write
     
 fΔ(y,r) − fΣ,x  ≤  fΔ(y,r) − fΔ(x,2r)  +  fΣ,x − fΔ(x,2r) 
⨏
 
≤ | f − fΔ(y,r) | dσ +  fΣ,x − fΔ(x,2r) 
Δ(x,2r)
 σ Σ ∩ B(x, 2r) ⨏  p1
≤ 1+ | f − fΔ(x,2r) | p dσ
σ Σ ∩ B(y, r) Δ(x,2r)
 
+  fΣ,x − fΔ(x,2r) 

−
n−1−λ  
≤ C f L. p, λ (Σ,σ) r p +  fΣ,x − fΔ(x,2r) 

= o(1) as r → ∞. (6.2.39)

Ultimately, this shows that fΣ,y = fΣ,x which establishes the independence of fΣ,x
on the point x.
As regards (6.2.36), recall from (6.2.38) that for each for each x ∈ Σ and each
r, R > 0 such that R > r we have
  n−1−λ
− p
 fΔ(x,r) − fΔ(x,R)  ≤ C f  . r . (6.2.40)
L p, λ (Σ,σ)

Upon passing to limit as R → ∞ we then obtain, on account of (6.2.35), that

  n−1−λ
− p
 fΔ(x,r) − fΣ  ≤ C f  . r . (6.2.41)
p, λ
L (Σ,σ)

Hence,
n−1−λ ⨏   1 n−1−λ  ⨏
  1
r p  f − fΣ  p dσ p ≤ r p  f − fΔ(x,r)  p dσ p
Σ∩B(x,r) Σ∩B(x,r)

n−1−λ
+r p |f
Σ − fΔ(x,r) |

≤ C f L. p, λ (Σ,σ) . (6.2.42)

Taking the supremum over all x ∈ Σ and r ∈ (0, ∞) then yields (6.2.36). This finishes
the proof of the lemma. 

We are now prepared to fully clarify the relationship between the Morrey spaces
and the Morrey-Campanato spaces in the case when the underlying Ahlfors regular
is unbounded.

Proposition 6.2.4 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be an unbounded closed

set which is Ahlfors regular, and abbreviate σ := H n−1 Σ. Also, fix p ∈ (1, ∞) and
322 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

λ ∈ (0, n − 1). Then, with the piece of notation introduced in (6.2.35), the assignment
.  
L p,λ (Σ, σ) ∼ ,  · L. p, λ (Σ,σ)  ∼  [ f ] −→ f − fΣ ∈ M p,λ (Σ, σ),  ·  M p, λ (Σ,σ)
(6.2.43)
is a well-defined linear and bounded isomorphism, whose inverse is the mapping
.  
M p,λ (Σ, σ),  ·  M p, λ (Σ,σ)  f −→ [ f ] ∈ L p,λ (Σ, σ) ∼ ,  · L. p, λ (Σ,σ)  ∼ .
(6.2.44)
Moreover, .
for each f ∈ L p,λ (Σ, σ) one has
(6.2.45)
f ∈ M p,λ (Σ, σ) ⇐⇒ fΣ = 0.

We wish to note that, as a consequence of (6.2.45), (6.2.37), and (6.2.34), in the

context of the above proposition we have
.
L q (Σ, σ) ∩ L p,λ (Σ, σ) = L q (Σ, σ) ∩ M p,λ (Σ, σ) for each q ∈ [1, ∞). (6.2.46)
.
Proof of Proposition 6.2.4 Observe that for each f ∈ M p,λ (Σ, σ) ⊆ L p,λ (Σ, σ) we
have
⨏ ⨏  1/p
| fΣ | ≤ lim sup | f | dσ ≤ lim sup | f | p dσ
R→∞ B(x0,R)∩Σ R→∞ B(x0,R)∩Σ

n−1−λ
−
≤ lim sup R p  f  M p, λ (Σ,σ) = 0, (6.2.47)
R→∞

hence
fΣ = 0 for each f ∈ M p,λ (Σ, σ). (6.2.48)
. p,λ
In the opposite direction, if f ∈ L (Σ, σ) has fΣ = 0 then (6.2.36) implies that
f ∈ M p,λ (Σ, σ). This finishes the proof of (6.2.45).
Next, the fact the assignment considered in (6.2.43) is well defined, linear, and
bounded is clear from Lemma 6.2.3 and (6.2.1)-(6.2.2). Its injectivity is obvious,
since f − fΣ = 0 in M p,λ (Σ, σ) forces f to be a constant, hence [ f ] = 0. To prove
that the assignment in (6.2.43) is also surjective, pick f ∈ M p,λ (Σ, σ). Then (6.2.34)
.
implies f ∈ L p,λ (Σ, σ), and [ f ] is mapped into f − fΣ = f , by (6.2.48). This
reasoning also shows that the inverse of (6.2.43) is the mapping in (6.2.44). 

Here is a companion result to Proposition 6.2.4, establishing the coincidence

of Morrey and Morrey-Campanato spaces in the case when the underlying Ahlfors
regular is bounded.

Proposition 6.2.5 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a bounded closed set

which is Ahlfors regular, and abbreviate σ := H n−1 Σ. Also, fix p ∈ (1, ∞) and
λ ∈ (0, n − 1). Then
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 323

.
L p,λ (Σ, σ) = L p,λ (Σ, σ) = M p,λ (Σ, σ) (6.2.49)
and ∫ 
 
 f L p, λ (Σ,σ) ≈  f  . p, λ
L (Σ,σ)
+  f dσ  ≈  f  M p, λ (Σ,σ),
Σ (6.2.50)
uniformly for f ∈ L 1 (Σ, σ).

Proof The first equality in (6.2.49) has been already noted in (6.1.13), and from
.
(6.2.34) we know that M p,λ (Σ, σ) ⊆ L p,λ (Σ, σ). To prove the opposite inclusion, fix
.
a function f ∈ L p,λ (Σ, σ), some point x ∈ Σ, and take R := 4 diam(Σ). Then for
each r ∈ 0, 2 diam(Σ) we may rely on [133, Lemma 7.4.15], (A.0.83), and (6.1.2)
to estimate
 ∫   ∫ 8diam(Σ)
    dt
 fΔ(x,r) − σ(Σ) −1
f dσ  = fΔ(x,r) − fΔ(x,R) ≤ C osc p ( f ; t)
Σ r t
∫ 8diam(Σ) n−1−λ
−1− p
≤ C f L. p, λ (Σ,σ) t dt
r
n−1−λ
−
≤ C f L. p, λ (Σ,σ) r p . (6.2.51)

Consequently,
⨏  p1 ⨏   1  
| f | dσ
p
≤  f − fΔ(x,r)  p dσ p +  fΔ(x,r) 
Σ∩B(x,r) Σ∩B(x,r)

n−1−λ  ∫ 
−  
≤ C f L. p, λ (Σ,σ) r p +  fΔ(x,r) − σ(Σ)−1 f dσ 
Σ
∫ 
 
+ σ(Σ)−1  f dσ 
Σ

n−1−λ ∫ 
−  
≤ C f L. p, λ (Σ,σ) r p + σ(Σ)−1  f dσ . (6.2.52)
Σ

n−1−λ
Multiplying the most extreme sides by r p , then taking the supremum over all
x ∈ Σ and all r ∈ 0, 2 diam(Σ) then yields
∫ 
 
 f  M p, λ (Σ,σ) ≤ C f L. p, λ (Σ,σ) + C  f dσ . (6.2.53)
Σ

This shows that f ∈ M p,λ (Σ, σ), which finishes the proof of the inclusion
.
L p,λ (Σ, σ) ⊆ M p,λ (Σ, σ). At this stage, (6.2.49) is fully justified and the equiva-
lences in (6.2.50) are implicit in the above argument. The proof of the proposition is
therefore complete. 
324 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

In Lemma 6.2.7 we shall take a closer look at the nature of H q,λ,θ -domes on
compact Ahlfors regular sets Σ ⊆ Rn . One useful technical result in this regard is
isolated in the lemma below.

Lemma 6.2.6 Suppose Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a bounded closed set

which is Ahlfors regular, and abbreviate σ := H n−1 Σ. Also, fix p, q ∈ (1, ∞) such
that 1/p + 1/q = 1, along with some λ ∈ (0, n − 1). Then

L q (Σ, σ) → H q,λ
(Σ, σ) continuously, (6.2.54)

and
there exists C = C(Σ, q, λ) ∈ (0, ∞) such that for any x0 ∈ Σ and any
R ∈ 0, 2 diam(Σ) the function f := R(n−1)(1/p−1)+λ(1/q−1) 1B(x0,R)∩Σ (6.2.55)
belongs to the space H q,λ (Σ, σ) and satisfies  f H q, λ (Σ,σ) ≤ C.

For further reference let us observe that as a consequence of (6.2.54) and (6.1.90)-
(6.1.91), in the context of Lemma 6.2.6, we have that
∫
 
f, g = f g dσ for each f ∈ L p,λ (Σ, σ) and g ∈ L q (Σ, σ), (6.2.56)
Σ

where ·, · denotes the duality bracket between the inhomogeneous Morrey-
Campanato space and its pre-dual (regarding g in H q,λ (Σ, σ)). In particular, given
any θ > (n − 1)(q − 1), we have
∫
 
f, m = f m dσ for each f ∈ L p,λ (Σ, σ) and each H q,λ,θ -dome m on Σ.
Σ
(6.2.57)
Proof of Lemma 6.2.6 Any given function∫ f ∈ L q (Σ, σ) may∫ be decomposed as
f = g + c, where g := f − σ(Σ) Σ f dσ and c := σ(Σ)−1 Σ f dσ. Since g may
−1

be regarded as a scalar multiple of a H q,λ -atom on Σ with “large” support (i.e.,

a function satisfying (6.1.15) with R := diam(Σ)) and g is a constant hence, by
definition, also a multiple of a H q,λ -atom on Σ (given that Σ is bounded), the
inclusion in (6.2.54) follows. Its continuity is implicit in the above reasoning.
Next, fix x0 ∈ Σ along with R ∈ 0, 2 diam(Σ) , and define f as in (6.2.55).
Thanks to (6.2.54) we have f ∈ H q,λ (Σ, σ). To estimate the norm of f in the space
H q,λ (Σ, σ) we write
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 325
 
 f H q, λ (Σ,σ) = sup |Λ( f )| ≤ sup  g, f  
Λ∈(H q, λ (Σ,σ))∗ g ∈L p, λ (Σ,σ)
Λ (H q, λ (Σ, σ))∗ ≤1 g L p, λ (Σ, σ) ≤C

∫ 
 
≤ sup  f g dσ 
g ∈M p, λ (Σ,σ) Σ
g  M p, λ (Σ, σ) ≤C

⨏
(n−1)/p+λ(1/q−1)
≤ sup R |g| dσ
g ∈M p, λ (Σ,σ) B(x0,R)∩Σ
g  M p, λ (Σ, σ) ≤C

 
⨏  1/p
(n−1−λ)/p
≤C sup R |g| dσ
p
g ∈M p, λ (Σ,σ) B(x0,R)∩Σ
g  M p, λ (Σ, σ) ≤C

=C sup g M p, λ (Σ,σ) ≤ C. (6.2.58)

g ∈M p, λ (Σ,σ)
g  M p, λ (Σ, σ) ≤C

Above, the first step is one of the basic consequences of the Hahn-Banach Theo-
rem, the second step is implied by (6.1.25), the third step originates in (6.1.91) in
Lemma 6.1.3 and (6.2.49) in Proposition 6.2.5, the fourth step takes into account the
specific format of the function f , the fifth step is based on Hölder’s inequality and
the fact that 1/p + 1/q = 1, the sixth step is seen from (6.2.2), and the final step is
tautological. 

Remarkably, there is a companion result for Lemma 6.1.2 valid for domes (in lieu
of molecules), provided the Ahlfors regular set is now assumed to be bounded.

Lemma 6.2.7 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a bounded set which is

closed and Ahlfors regular, and abbreviate σ := H n−1 Σ. Also, fix q ∈ (1, ∞),
λ ∈ (0, n − 1), and θ > (n − 1)(q − 1). Then

there exists C = C(Σ, q, λ, θ) ∈ (0, ∞) such that any H q,λ,θ -dome m

(6.2.59)
on Σ belongs to the space H q,λ (Σ, σ) and satisfies mH q, λ (Σ,σ) ≤ C.

As a consequence,
whenever Σ is compact, the operator of pointwise multiplica-
tion by some given function belonging to L ∞ (Σ, σ) is a bounded (6.2.60)
mapping from the space H q,λ (Σ, σ) into itself.

Proof Fix a H q,λ,θ -dome m on Σ and pick x0 ∈ Σ along with R ∈ 0, 2 diam(Σ)

such that the estimates in (6.1.33) are true. We perform the same decomposition
procedure as in the proof of Lemma 6.1.2 with one basic difference. Specifically,
in the process of writing (6.1.42) (presently with N := NR ) by once again making
use of (6.1.43) for the same choices of the a j ’s and b j ’s as in (6.1.67)-(6.1.67), we
326 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

no longer have b−1 = 0 as the∫ dome m may be lacking any cancellation property.
Instead, we now have b−1 = Σ m dσ so we currently have an additional term, namely
−1 ∫
a0 b−1 = σ B(x0, R) ∩ Σ 1B(x0,R)∩Σ Σ m dσ, in the last line of (6.1.42), hence also
in the right-hand side of (6.1.70) and the right-hand side of (6.1.87). Thus, in place
of (6.1.87) we now arrive at


NR N
R −1

m= λj
αj + η j βj + m0 pointwise on Σ,
j=0 j=0 (6.2.61)
∫
1B(x0,R)∩Σ
with m0 := m dσ,
σ B(x0, R) ∩ Σ Σ

α j ’s and βj ’s
where the numerical coefficients λ j , η j satisfy (6.1.86) and where the 
are H q,λ -atoms on Σ. Denote by p ∈ (1, ∞) the Hölder conjugate exponent of q.
Since from (6.1.34) and the Ahlfors regularity of Σ we obtain
∫
1B(x0,R)∩Σ
|m0 | ≤ |m| dσ ≤ CR(n−1)/p+λ(1/q−1) · R−(n−1) 1B(x0,R)∩Σ
σ B(x0, R) ∩ Σ Σ

= CR(n−1)(1/p−1)+λ(1/q−1) 1B(x0,R)∩Σ pointwise on Σ, (6.2.62)

it follows that (6.2.55) applies to the function m0 defined in (6.2.61). Hence, m0

belongs to the space H q,λ (Σ, σ) and satisfies m0 H q, λ (Σ,σ) ≤ C for some constant
C ∈ (0, ∞) independent of m, x0, R. In concert with (6.1.17), this proves the claim
made in (6.2.59).
Finally, the claim in (6.2.60) is a consequence of (6.1.17), (6.1.22), (6.2.59), and
the fact that the product between a bounded function and a H q,λ -dome on Σ is a
multiple of a H q,λ -dome on Σ. 

The topic addressed next pertains to the pre-duals of Morrey spaces. To set the
stage, given an integrability exponent q ∈ (1, ∞) and a parameter λ ∈ (0, n − 1), a
function b ∈ L q (Σ, σ) is said to be a B q,λ -block on Σ (or, simply, a block) provided
there exist some point xo ∈ Σ and some radius R ∈ 0, 2 diam(Σ) such that
1
λ q −1
supp b ⊆ B(xo, R) ∩ Σ and b L q (Σ,σ) ≤ R . (6.2.63)

We then define the block space



B q,λ (Σ, σ) := f ∈ Lipc (Σ) : there exist a sequence {λ j } j ∈N ∈  1 (N) and a
family {b j } j ∈N of B q,λ -blocks on Σ such that

∞ 

f = λ j b j with convergence in Lipc (Σ) ,
j=1
(6.2.64)
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 327

and for each f ∈ B q,λ (Σ, σ) define


∞ 
∞

 f  B q, λ (Σ,σ) := inf |λ j | : f = λ j b j in Lipc (Σ) with (6.2.65)
j=1 j=1

{λ j } j ∈N ∈  1 (N) and each b j a B q,λ -block on Σ .

In particular, it is clear from (6.2.65) that

each B q,λ -block b on Σ belongs to

(6.2.66)
B q,λ (Σ, σ) and b B q, λ (Σ,σ) ≤ 1,

and that

∞ 
if f ∈ B q,λ (Σ, σ) is expanded as f = λ j b j in Lipc (Σ) for some
j=1
sequence {λ j } j ∈N ∈  1 (N) and with each b j a B q,λ -block on Σ, then (6.2.67)

∞
the series λ j b j also converges to f in B q,λ (Σ, σ).
j=1

Additionally, it is of significance to note that (6.2.66) implies that

if f ∈ L q (Σ, σ) has supp f ⊆ Δ(x0, R) for some point x0 ∈ Σ

and some radius R ∈ 0, 2 diam(Σ) , then f ∈ B q,λ (Σ, σ) and (6.2.68)
1
λ q −1
the estimate  f  B q, λ (Σ,σ) ≤ R  f  L q (Σ,σ) holds.

As a consequence of (6.2.67) and (6.2.68) we have that

q
Lcomp (Σ, σ) → B q,λ (Σ, σ) continuously and densely (6.2.69)

which, thanks to [133, (3.7.22)] and (6.2.66), further self-improves to3:

Lipc (Σ) is a dense linear subspace of B q,λ (Σ, σ). (6.2.70)

Also, as in the case of the space introduced in (6.1.16), we see that

 
B q,λ (Σ, σ),  ·  B q, λ (Σ,σ) is a separable Banach space,

and B q,λ (Σ, σ) → L r (Σ, σ) with r := q(n−1)

∈ (1, q) (6.2.71)
n−1+λ(q−1)
is a well-defined, continuous embedding, with dense range.

In particular, (6.2.71) and [133, (7.7.106)] imply that

 σ(x) 
B q,λ (Σ, σ) → L 1 Σ, continuously. (6.2.72)
1 + |x| n−1
3 another proof of (6.2.70) may be given based on the Hahn-Banach theorem and Proposition 6.2.8
328 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

In addition,

∞ 
if f ∈ B q,λ (Σ, σ) is expressed as f = λ j b j in Lipc (Σ) for a
j=1
sequence {λ j } j ∈N ∈  1 (N) and with each b j a B q,λ -block on Σ, then (6.2.73)
∞
q(n−1)
λ j b j converges to f L r (Σ, σ) if we take r := n−1+λ(q−1) ,
j=1

and
the operator of pointwise multiplication by a given function
h ∈ L ∞ (Σ, σ) is a bounded mapping from B q,λ (Σ, σ) into (6.2.74)
itself, with operator norm ≤ h L ∞ (Σ,σ) ,

since such a mapping does not distort dull atoms by more than a fixed multiplicative
constant, which is ≤ h L ∞ (Σ,σ) . Note that the latter property further implies that

if f , g : Σ −→ C are two σ-measurable functions such that

|g| ≤ | f | at σ-a.e. point on Σ and f ∈ B q,λ (Σ, σ), then g is in (6.2.75)
B q,λ (Σ, σ) and one has g B q, λ (Σ,σ) ≤  f  B q, λ (Σ,σ) ,

since

g/ f if f  0,
g = h f with h := satisfying h L ∞ (Σ,σ) ≤ 1. (6.2.76)
0 if f = 0,

It is instructive to compare the space B q,λ (Σ, σ) introduced in (6.2.64) with

H q,λ (Σ, σ) from (6.1.16) (the latter being the pre-dual of a Morrey-Campanato
space; cf. (6.1.25)). In this regard, we wish to remark that

H q,λ (Σ, σ) ⊆ B q,λ (Σ, σ), and actually

(6.2.77)
H q,λ (Σ, σ) = B q,λ (Σ, σ) whenever Σ is bounded.

Indeed, the inclusion in the first line of (6.2.77) is clear from definitions. Also, if Σ is
bounded then any B q,λ -block on Σ is a H q,λ,θ -dome on Σ (in the sense of (6.1.33))
for any θ > (n − 1)(q − 1), so Lemma 6.2.7 now provides the opposite inclusion
B q,λ (Σ, σ) ⊆ H q,λ (Σ, σ) from which the equality in the second line of (6.2.77)
follows.
This being said, our primary interest in the space (6.2.64) stems from the fact
that this turns out to be the pre-dual of a Morrey space. Specifically, we have the
following result.

Proposition 6.2.8 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Fix two integrability exponents p, q ∈ (1, ∞)
satisfying 1/p + 1/q = 1, along with a parameter λ ∈ (0, n − 1). Then there exists a
constant C ∈ (0, ∞) with the property that
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 329
∫
| f ||g| dσ ≤ C f  M p, λ (Σ,σ) g B q, λ (Σ,σ)
Σ (6.2.78)
for all f ∈ M p,λ (Σ, σ) and g ∈ B q,λ (Σ, σ).

Moreover, the mapping

∗
M p,λ (Σ, σ)  f −→ Λ f ∈ B q,λ (Σ, σ) given by
∫ (6.2.79)
Λ f (g) := f g dσ for each g ∈ B q,λ (Σ, σ)
Σ

is a well-defined, linear, bounded isomorphism, with bounded inverse. Simply put,

the integral paring yields the quantitative identification
 ∗
B q,λ (Σ, σ) = M p,λ (Σ, σ). (6.2.80)


As a corollary, whenever Λ ∈ Lipc (Σ) is a distribution with the property that
there exists some constant C ∈ (0, ∞) such that
 
 Λ, φ  ≤ Cφ B q, λ (Σ,σ) for every φ ∈ Lip (Σ), (6.2.81)
c

there exists a unique function f ∈ M p,λ (Σ, σ) such that

∫
Λ, φ = f φ dσ for every φ ∈ Lipc (Σ). (6.2.82)
Σ

In addition, with C denoting the constant appearing in (6.2.81), the function f

satisfies  f  M p, λ (Σ,σ) ≤ c C, where c ∈ (0, ∞) is a fixed number which depends only
the ambient.

It is useful to note that, in the context of Proposition 6.2.8, for each function
f ∈ M p,λ (Σ, σ) we may write

 f  M p, λ (Σ,σ) ≈ Λ f (B q, λ (Σ,σ))∗ = sup |Λ f (g)|

g ∈B q, λ (Σ,σ)
g B q, λ (Σ, σ) ≤1

∫ 
 
= sup  f g dσ , (6.2.83)
g ∈B q, λ (Σ,σ) Σ
g B q, λ (Σ, σ) ≤1

thanks to the fact that (6.2.79) is an isomorphism. A closely related result is the fact
that for each σ-measurable function f : Σ → R we have
∫
f ∈ M p,λ (Σ, σ) if and only if sup | f ||g| dσ < +∞, (6.2.84)
g ∈B q, λ (Σ,σ) Σ
g B q, λ (Σ, σ) ≤1
330 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

and ∫
 f  M p, λ (Σ,σ) ≈ sup | f ||g| dσ. (6.2.85)
g ∈B q, λ (Σ,σ) Σ
g B q, λ (Σ, σ) ≤1

Indeed, the right-pointing implication in (6.2.84) is seen from (6.2.78), while the
left-pointing implication can be proved using formula (6.2.80). To elaborate, assume
f : Σ → R is a σ-measurable function for which the supremum in (6.2.84) is finite.
∫ that f ∈ Lloc (Σ, σ). Next, observe that the assignment
Thanks to (6.2.69), this ensures 1

B q,λ (Σ, σ)  g → Λ(g) := Σ f g dσ is a continuous linear functional on B q,λ (Σ, σ),

 ∗
hence Λ ∈ B q,λ (Σ, σ) = M p,λ (Σ, σ) by (6.2.80). The latter identification implies
∫
that there exists f ∈ M p,λ (Σ, σ) such that Λ(g) = Σ  f g dσ for each g ∈ B q,λ (Σ, σ).
∫ ∫
Hence, Σ f g dσ = Σ  f g dσ for each g ∈ B q,λ (Σ, σ). In light of (6.2.70) and [133,
Proposition 3.7.2], this ultimately permits us to conclude that f =  f ∈ M p,λ (Σ, σ),
as wanted.
Parenthetically, we wish to remark that the left-pointing implication in (6.2.84)
may also be justified by pairing the given function f with individual blocks. Specif-
ically, if C f ∈ [0, ∞) denotes the supremum appearing in (6.2.84) then, having fixed
some point xo ∈ ∫Σ and some radius R ∈ 0, 2 diam(Σ) , for each function b as in
(6.2.63) we have Σ | f ||b| dσ ≤ C f . By Riesz’s Representation Theorem, this places

f Δ(xo,R) in L p Δ(xo, R), σ with

∫  p1 1
λ 1− q
λ
| f | p dσ ≤ Cf · R = Cf · R p . (6.2.86)
Δ(x o ,R)

In turn, from this and (6.2.2) we conclude that

 
n−1−λ  ⨏  p1
 f  M p, λ (Σ,σ) = sup R p | f | dσ
p
≤ C(Σ) · C f < +∞,
x o ∈Σ and Δ(x o ,R)
0<R<2 diam(Σ)
(6.2.87)
hence f ∈ M p,λ (Σ, σ). This finishes the proof of (6.2.84). Finally, (6.2.85) is a
consequence of (6.2.83) and (6.2.87).
Let us also note here that for each g ∈ B q,λ (Σ, σ) we have

g B q, λ (Σ,σ) = sup |Λ(g)| ≈ sup |Λ f (g)|

Λ∈(B q, λ (Σ,σ))∗ f ∈M p, λ (Σ,σ)
Λ (B q, λ (Σ, σ))∗ ≤1  f  M p, λ (Σ, σ) ≤1

∫ 
 
= sup  f g dσ , (6.2.88)
f ∈M p, λ (Σ,σ) Σ
 f  M p, λ (Σ, σ) ≤1

with the first equality a well-known corollary of the Hahn-Banach theorem, the
subsequent equivalence provided by the fact that (6.2.79) is an isomorphism, and the
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 331

final equality simply the definition in (6.2.79). Related formulas are discussed later,
in (6.2.158) and (6.2.160).
For now, we present the proof of Proposition 6.2.8.
Proof of Proposition 6.2.8 To justify (6.2.78), fix two functions, f ∈ M p,λ (Σ, σ)
and g ∈ B q,λ (Σ, σ), arbitrary. Then (6.2.73) guarantees that there exist some nu-
merical sequence {λ j } j ∈N ∈  1 (N) along with some family {b j } j ∈N of B q,λ -blocks

∞
on Σ with the property that g = λ j b j with convergence in the space L r (Σ, σ) for
j=1
q(n−1)
r := n−1+λ(q−1) , and which is quasi-optimal, in the sense that we have the estimate

∞ 
∞
|λ j | ≤ 2g B q, λ (Σ,σ) . In particular, |g| ≤ |λ j ||b j | at σ-a.e. point on Σ. If for
j=1 j=1
each j ∈ N we pick a point x j ∈ Σ and a radius R j ∈ 0, 2 diam(Σ) such that (cf.
(6.2.63))
1
λ( q −1)
supp b j ⊆ B(x j , R j ) ∩ Σ and b j  L q (Σ,σ) ≤ Rj , (6.2.89)

we may then estimate, using Hölder’s inequality, (6.2.89), and assumptions,

∫ 
∞ ∫
| f ||g| dσ ≤ |λ j | | f ||b j | dσ
Σ j=1 Σ∩B(x j ,R j )


∞ ⨏
= |λ j |σ Σ ∩ B(x j , R j ) | f ||b j | dσ
j=1 Σ∩B(x j ,R j )


∞
≤ |λ j |σ Σ ∩ B(x j , R j ) ×
j=1

⨏  1/p  ⨏  1/q
× | f | p dσ |b j | q dσ
Σ∩B(x j ,R j ) Σ∩B(x j ,R j )


∞
1−1/q −
n−1−λ
p
≤ |λ j |σ Σ ∩ B(x j , R j ) Rj  f  M p, λ (Σ,σ) b j  L q (Σ,σ)
j=1


∞ 1
(n−1) 1− q −
n−1−λ 1
λ q −1
p
≤ C f  M p, λ (Σ,σ) |λ j |R j Rj Rj
j=1


∞ 
= C f  M p, λ (Σ,σ) |λ j |
j=1

≤ C f  M p, λ (Σ,σ) g B q, λ (Σ,σ), (6.2.90)

332 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

which establishes (6.2.78). In turn, this proves that the assignment (6.2.79) is indeed
well defined, linear, and bounded. Moreover, (3.7.23) in [133, Corollary 3.7.3]
ensures that said assignment is injective.
Let us now turn our attention to establishing that the assignment (6.2.79) is
∗
also surjective. To this end, fix an arbitrary functional Λ ∈ B q,λ (Σ, σ) with
Λ q, λ ∗ ≤ 1. For each fixed point x ∈ Σ and radius R ∈ 0, 2 diam(Σ)
B (Σ,σ)
abbreviate Δ := Δ(x, R) := Σ ∩ B(x, R) and consider the extension by zero mapping

g on Δ(x, R),
ιΔ : L Δ(x, R), σ −→ B (Σ, σ),
q q,λ
ιΔ (g) := (6.2.91)
0 on Σ \ Δ(x, R),

for each g ∈ L q Δ(x, R), σ . Then for each function g ∈ L q Δ(x, R), σ which is not
identically zero we have

supp ιΔ (g) ⊆ Δ(x, R) and ιΔ (g) L q (Σ,σ)

= g L q (Δ(x,R),σ), (6.2.92)

hence
1
λ( −1)
R q
ιΔ (g) is a B q,λ -block on Σ. (6.2.93)
g L q (Δ(x,R),σ)

From this and (6.2.66) we then conclude that

1
−λ q −1
ιΔ (g) B q, λ (Σ,σ)
≤R g L q (Δ(x,R),σ), ∀g ∈ L q Δ(x, R), σ . (6.2.94)

This proves that ιΔ in (6.2.91) is well defined, linear and bounded, with operator
norm
1
−λ q −1
ιΔ L q (Δ(x,R),σ)→B q, λ (Σ,σ)
≤R . (6.2.95)

As such, the composition Λ ◦ ιΔ is a well-defined, linear and bounded functional on

the Lebesgue space L q Δ(x, R), σ satisfying
1
−λ q −1
Λ ◦ ιΔ (L q (Δ(x,R),σ))∗
≤R . (6.2.96)

Granted this, Riesz’s Duality Theorem then guarantees the existence of some function
1
−λ −1
fΔ(x,R) ∈ L p Δ(x, R), σ with fΔ(x,R) L p (Δ(x,R),σ) ≤ R q and
∫ (6.2.97)
Λ ιΔ (g) = fΔ(x,R) g dσ for each given g ∈ L q Δ(x, R), σ .
Δ(x,R)

Consider now two points x1, x2 ∈ Σ along with two radii R1, R2 ∈ 0, 2 diam(Σ)
such that Δ1 := Δ(x1, R1 ) is contained in Δ2 := Δ(x2, R2 ). Also, fix an arbitrary
function g ∈ L q (Δ1, σ) and define
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 333

g on Δ1,
h ∈ L q (Δ2, σ), h := (6.2.98)
0 on Δ2 \ Δ1 .

Then (6.2.91) implies that ιΔ1 g = ιΔ2 h ∈ B q,λ (Σ, σ). Granted this, (6.2.97) permits
us to write
∫
fΔ(x1,R1 ) g dσ = Λ ιΔ1 (g) = Λ ιΔ2 (h) (6.2.99)
Δ(x1,R1 )
∫ ∫
= fΔ(x2,R2 ) h dσ = fΔ(x2,R2 ) g dσ.
Δ(x2,R2 ) Δ(x1,R1 )

On account of (6.2.99) and [133, (3.7.23)] we then conclude that



fΔ(x2,R2 )  = fΔ(x1,R1 ) at σ-a.e. point in Δ1 . (6.2.100)
Δ(x1,R1 )

This
 compatibility
 property allows us to glue together the family of functions
fΔ(x,R) thus giving rise to a function f which is globally defined on Σ and
x,R
which satisfies, for each x ∈ Σ and R ∈ 0, 2 diam(Σ) ,

f Δ(x,R) = fΔ(x,R) at σ-a.e. point in Δ(x, R). (6.2.101)

In particular, f is σ-measurable, and by combing (6.2.101) with the first line in

(6.2.97) we see that for each x ∈ Σ and R ∈ 0, 2 diam(Σ) we obtain
n−1−λ ⨏  p1 λ
−p
R p | f | p dσ ≤ CR fΔ(x,R) L p (Δ(x,R),σ)
Σ∩B(x,R)
λ 1
−p −λ q −1
≤ CR ·R = C, (6.2.102)

since 1/p + 1/q = 1, for some constant C = C(Σ, p, λ) ∈ (0, ∞) independent of

f , x, R. In view of this and (6.2.1)-(6.2.1) we may now conclude that

f ∈ M p,λ (Σ, σ) and  f  M p, λ (Σ,σ) ≤ C = C(Σ, p, λ) < +∞. (6.2.103)

Also, from the second line in (6.2.97), (6.2.91), (6.2.101), (6.2.79) it follows that
∫
Λ(b) = f b dσ = Λ f (b) for each B q,λ -block b on Σ. (6.2.104)
Σ
∗
Thus, the linear continuous functionals Λ, Λ f ∈ B q,λ (Σ, σ) agree when acting
on arbitrary B q,λ -blocks on Σ. On account of this, (6.2.64), and (6.2.67) we finally
conclude that Λ = Λ f on the entire space B q,λ (Σ, σ). This argument shows that the
assignment (6.2.79) is also surjective thus, ultimately, a bounded isomorphism, with
bounded inverse.
334 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

As regards the very last claim in the statement of the proposition, since Lipc (Σ)
is dense in the space B q,λ (Σ, σ) (something readily seen from (6.2.70)) and since
the mapping Lipc (Σ)  φ → Λ, φ ∈ C is linear and bounded with respect to the
norm in B q,λ (Σ, σ) it follows that said mapping extends (by density) to a unique
∗
functional Θ ∈ B q,λ (Σ, σ) whose norm is ≤ c C, where the constant C is as in
(6.2.81) and where c ∈ (0, ∞) is a fixed number which depends only the ambient.
In light of (6.2.80), this proves ∫that there exists some f ∈ M p,λ (Σ, σ) such that
 f  M p, λ (Σ,σ) ≤ c C, and Θ(g) = Σ f g dσ for each g ∈ B q,λ (Σ, σ). In concert with
the fact that Θ(φ) = Λ, φ for each φ ∈ Lipc (Σ), this establishes (6.2.82). The
uniqueness of f is then a consequence of [133, Corollary 3.7.3]. 

In the setting of Proposition 6.2.8, from (6.2.80), (6.2.71), and the Sequential
Banach-Alaoglu Theorem (recalled in [133, (3.6.22)]) we conclude that any bounded
sequence in M p,λ (Σ, σ) has a subsequence which is weak-∗ convergent. Here is a
result in this spirit in which a stronger conclusion is reached, provided one assumes
more than mere boundedness for said sequence.

Proposition 6.2.9 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Fix two integrability exponents p, q ∈ (1, ∞)
satisfying 1/p + 1/q = 1, along with a parameter λ ∈ (0, n − 1). In this setting,
suppose { f j } j ∈N ⊆ M p,λ (Σ, σ) is a sequence of functions with the property that

f (x) := lim f j (x) exists for σ-a.e. x ∈ Σ, and

j→∞
there exists g ∈ M p,λ (Σ, σ) such that for each (6.2.105)
j ∈ N one has | f j (x)| ≤ |g(x)| for σ-a.e. x ∈ Σ.

Then f ∈ M p,λ (Σ, σ) and f j → f as j → ∞ weak-∗ in M p,λ (Σ, σ), i.e.,

∫ ∫
lim f j h dσ = f h dσ for each h ∈ B q,λ (Σ, σ). (6.2.106)
j→∞ Σ Σ

Proof From (6.2.105), [133, Remark 3.1.2], and [133, (3.6.26)] it follows that the
function f is σ-measurable and satisfies | f (x)| ≤ |g(x)| for σ-a.e. point x ∈ Σ.
In concert with (6.2.3), this further implies that f ∈ M p,λ (Σ, σ). With an eye on
(6.2.106), pick an arbitrary function h ∈ B q,λ (Σ, σ). Then (6.2.78) guarantees that
gh ∈ L 1 (Σ, σ), and

lim ( f j (x) − f (x))h(x) = 0 for σ-a.e. x ∈ Σ, and for each j ∈ N

j→∞ (6.2.107)
we have |( f j (x) − f (x))h(x)| ≤ 2|g(x)h(x)| at σ-a.e. point x ∈ Σ.

On account of these properties,

∫ Lebesgue’s Dominated Convergence Theorem ap-
plies and gives that lim Σ |( f j − f )h| dσ = 0. This ultimately establishes (6.2.106),
j→∞
finishing the proof of the proposition. 
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 335

Our next lemma contains a rather useful and versatile membership criterion (in-
volving local integrability and decay at infinity) to the space B q,λ (Σ, σ), introduced
in (6.2.64).

Lemma 6.2.10 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be an Ahlfors regular set and

abbreviate σ := H n−1 Σ. Also, fix q ∈ (1, ∞) along with λ ∈ (0, n − 1), and assume
that
 1
θ < (n − 1 − λ) 1 − . (6.2.108)
q
In this setting, suppose f is a σ-measurable function on Σ with the property that
there exist a point xo ∈ Σ and a radius R ∈ 0, 2 diam(Σ) such that
∫  1/q 1
λ q −1
| f (x)| q dσ(x) ≤R and
Σ∩B(x o ,R)
1 (6.2.109)
−θ−(n−1−λ) q −1
R
| f (x)| ≤ for σ-a.e. x ∈ Σ \ B(xo, R).
|x − xo | n−1−θ

Then there exists a constant C = C(Σ, n, q, λ, θ) ∈ (0, ∞), independent of f , x0, R,

such that
f ∈ B q,λ (Σ, σ) and  f  B q, λ (Σ,σ) ≤ C. (6.2.110)

Proof To set the stage, consider the function b0 := f 1Σ∩B(xo,R) and, for each integer
j ∈ N, define b j := f 1Σ∩[B(xo,2 j R)\B(xo,2 j−1 R)] . Then the property in the first line of
(6.2.109) implies that b0 is a B q,λ -block on Σ. As such, (6.2.66) gives

b0 ∈ B q,λ (Σ, σ) and b0  B q, λ (Σ,σ) ≤ 1. (6.2.111)

Next, for each j ∈ N, the property in the second line of (6.2.109) allows us to estimate
∫  1/q  ∫  1/q
|b j (x)| q dσ(x) = | f (x)| q dσ(x)
Σ Σ∩[B(x o ,2 j R)\B(x o ,2 j−1 R)]

1
−θ−(n−1−λ) q −1 n−1
CR
≤ (2 j R) q
(2 j R)n−1−θ
1 1
θ+(n−1−λ) q −1 λ q −1
= C(2 j ) (2 j R) , (6.2.112)

for some constant C ∈ (0, ∞) independent of j ∈ N, xo ∈ Σ, and R > 0. Since, by

design, we also have supp b j ⊆ Σ ∩ B(xo, 2 j R), it follows that if we define
1
θ+(n−1−λ) q −1
λ j := C(2 j ) (6.2.113)

then each b j /λ j becomes a B q,λ -block on Σ, and we have

336 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .


∞


∞
f = b0 + λ j (b j /λ j ) in Lipc (Σ) with |λ j | = C(Σ, n, q, λ, θ) < +∞,
j=1 j=1
(6.2.114)

where the last property uses (6.2.108). By combining (6.2.111), (6.2.114), (6.2.64),
and (6.2.65), we then arrive at the conclusion recorded in (6.2.110). 

Among other applications, Lemma 6.2.10 allows us to deal with the issue of the
boundedness of the Hardy-Littlewood maximal operator acting on the pre-dual of a
Morrey space, in the manner described below.

Corollary 6.2.11 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Also, fix q ∈ (1, ∞) along with λ ∈ (0, n − 1). Then
the Hardy-Littlewood maximal operator on Σ induces a well-defined, sub-linear, and
bounded mapping
M Σ : B q,λ (Σ, σ) −→ B q,λ (Σ, σ). (6.2.115)

Proof Let b be an arbitrary B q,λ -block on Σ and set f := M Σ b. Then, with the
point xo ∈ Σ and the radius R ∈ 0, 2 diam(Σ) chosen so that the properties listed
in (6.2.63) hold, we may estimate
∫  1/q 1
λ −1
| f (x)| q dσ(x) ≤ Cb L q (Σ,σ) ≤ CR q , (6.2.116)
Σ

thanks to [133, (7.6.18)]. Also, given that supp b ⊆ B(xo, R) ∩ Σ, from the very
definition of Hardy-Littlewood maximal operator (cf. (A.0.71)) we conclude that for
each x ∈ Σ \ B(xo, 2R) we have
∫ 
| f (x)| = M Σ b (x) ≤ C |b| dσ |x − xo | −(n−1) (6.2.117)
Σ
⨏ 
≤ CRn−1 |b| dσ |x − xo | −(n−1)
Σ∩B(x o ,R)
⨏  1/q
≤ CRn−1 |b| q dσ |x − xo | −(n−1)
Σ∩B(x o ,R)

1
n−1 −(n−1−λ) q −1
(n−1)− q −(n−1) CR
≤ CR b L q (Σ,σ) |x − xo | ≤ .
|x − xo | n−1

Thus, the function f satisfies the conditions listed (6.2.109) (with θ := 0, which
happens to be a permissible value in (6.2.108)), so from (6.2.110) we ultimately
conclude that there exists some constant C = C(Σ, n, q, λ) ∈ (0, ∞) with the property
that
M Σ b ∈ B q,λ (Σ, σ) and M Σ b B q, λ (Σ,σ) ≤ C
(6.2.118)
for each given B q,λ -block b on Σ.
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 337

With this in hand, the end-game in the proof of (6.2.115) is as follows. Having
fixed a function f ∈ B q,λ (Σ, σ), there exist a numerical sequence {λ j } j ∈N ∈  1 (N)
∞
along with a sequence {b j } j ∈N of B q,λ -blocks on Σ such that
∞f = j=1 λ j b j in the
sense of distributions on Σ. Thanks to (6.2.73), the series j=1 λ j b j actually con-
q(n−1)
verges to f in L r (Σ, σ) with r := .
Collectively, [133, (7.6.28)], (6.2.3),

n−1+λ(q−1)
 
and (6.2.118) also imply that the sequence M Σ Jj=1 λ j b j is Cauchy, hence
J ∈N
convergent, in the Banach space B q,λ (Σ, σ). In light of (6.2.71), this latter conver-
gence takes place in Lr (Σ, σ) as well. However, thanks to [133, (7.6.28)], it follows

that M Σ Jj=1 λ j a j actually converges to M Σ f in L r (Σ, σ). By also taking
J ∈N
(6.2.118) into account, this argument ultimately gives that

M Σ f ∈ B q,λ (Σ, σ) and M Σ f B q, λ (Σ,σ)

≤ C f  B q, λ (Σ,σ), (6.2.119)

for some constant C = C(Σ, n, q, λ) ∈ (0, ∞). This finishes the proof of (6.2.115). 
Remarkably, certain types of estimates on Muckenhoupt weighted Lebesgue
spaces imply estimates on Morrey spaces. Here is an extrapolation result of this
flavor, refining work in [27].

Proposition 6.2.12 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set, and abbreviate σ := H n−1 Σ. Also, fix an integrability exponent p ∈ (1, ∞)
along with a parameter λ ∈ (0, n − 1). Finally, let F be a family of pairs ( f , g) of
σ-measurable functions defined on Σ with the property that

for each Muckenhoupt weight w ∈ A1 (Σ, σ) there exists a constant

Cw = C([w] A1 ) ∈ (0, ∞), which stays bounded as [w] A1 stays bounded, (6.2.120)
and such that  f  L p (Σ,wσ) ≤ Cw g L p (Σ,wσ) for each pair ( f , g) ∈ F .

Then there exist two constants CΣ, p ∈ (0, ∞) (depending only on the Ahlfors
regularity character of Σ and p) and Wn,λ ∈ (0, ∞) (depending only on n and λ) such
that, with ! "
CΣ, p,λ := CΣ, p · sup Cw ∈ (0, ∞), (6.2.121)
w ∈ A1 (Σ,σ)
[w] A1 ≤Wn, λ

one has

 f  M p, λ (Σ,σ) ≤ CΣ, p,λ g M p, λ (Σ,σ) for each pair ( f , g) ∈ F . (6.2.122)

Comment 1. The family F is quite general and, in principle, it might not even
be specifically related to the Morrey space under consideration. For example,
given some (not necessarily linear) assignment f → Θ( f ), mapping functions
f ∈ L 1 Σ, 1+σ(x)
|x | n−1
into σ-measurable functions on Σ, one may take
 
F := (Θ f , f ) : f ∈ L 1 Σ, 1+σ(x)
|x | n−1 . (6.2.123)
338 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

In this case, from [133, (7.7.104)] and item (5) in [133, Lemma 7.7.1] it follows that
for each f ∈ L p (Σ, wσ) with w ∈ A1 (Σ, σ) we have that Θ( f ) is a meaningfully
defined σ-measurable function on Σ. In addition, (6.2.25) ensures that Θ( f ) is also
a meaningfully defined σ-measurable function on Σ for each f ∈ M p,λ (Σ, σ).
Comment 2. Continue to assume that f → Θ( f ) is an assignment (not necessarily
linear), mapping functions f ∈ L 1 Σ, 1+σ(x)
|x | n−1
into σ-measurable functions on Σ.
Combining Rubio de Francia’s extrapolation theorem (for Muckenhoupt weighted
Lebesgue spaces, recalled in [133, Proposition 7.7.6]) with Proposition 6.2.12 (ap-
plied to F as in (6.2.123)) then yields the following extrapolation result: whenever
(6.2.120) is strengthened to

for each Muckenhoupt weight w ∈ Ap (Σ, σ) there exists some constant

Cw = C([w] A p ) ∈ (0, ∞), which stays bounded as [w] A p stays bounded,
(6.2.124)
and such that Θ( f ) L p (Σ,wσ) ≤ Cw  f  L p (Σ,wσ) for each f in the
Muckenhoupt weighted Lebesgue space L p (Σ, wσ),

then for each given integrability exponent q ∈ (1, ∞) we may find two constants,
CΣ,q ∈ (0, ∞) and Wn,λ ∈ (0, ∞), such that

Θ( f ) M q, λ (Σ,σ) ≤ Cq  f  M q, λ (Σ,σ) for each f

(6.2.125)
belonging to the Morrey space M q,λ (Σ, σ),
where ! "
Cq := CΣ,q · sup Cw ∈ (0, ∞). (6.2.126)
w ∈ A p (Σ,σ)
[w] A p ≤Wn, λ

In addition,
if Θ is actually a linear operator, then Θ turns out to be bounded on
M̊ q,λ (Σ, σ) for any q ∈ (1, ∞), so its transpose Θ is bounded on
(6.2.127)
B q,λ (Σ, σ) for any q ∈ (1, ∞) (with similar operator norm estimates to
those in (6.2.125)-(6.2.126)).
We now turn to the proof of Proposition 6.2.12.
Proof of Proposition 6.2.12 Fix an arbitrary point x ∈ Σ along with an arbitrary
λ
scale R ∈ 0, 2 diam(Σ) , and select a power θ ∈ n−1 , 1 . Then, if M Σ denotes the
Hardy-Littlewood maximal operator on Σ, from [133, Corollary 7.6.5] (whose appli-
cability the present setting is guaranteed by [133, (3.6.26)] in [133, Lemma 3.6.4]),
we have

1B(x,R)∩Σ ≤ M Σ 1B(x,R)∩Σ at σ-a.e. point on Σ. (6.2.128)

Hence, from (6.2.128) and item (8) in [133, Lemma 7.7.1] we have
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 339
 θ
θ
1B(x,R)∩Σ = 1B(x,R)∩Σ ≤ w := M Σ 1B(x,R)∩Σ at σ-a.e. point on Σ,
and w ∈ A1 (Σ, σ) with [w] A1 ≤ Cθ ∈ (0, ∞), independent of x, R.
(6.2.129)
Let us also observe that since there exists a constant C ∈ (0, ∞) with the property
that at each point on Σ we have

∞
σ 1B(x,R)∩Σ
M Σ 1B(x,R)∩Σ ≤ C1B(x,R)∩Σ + C 1[B(x,2 j R)\B(x,2 j−1 R)]∩Σ,
j=1
σ 1B(x,2 j R)∩Σ
(6.2.130)
it follows that
 θ 
∞
w = M Σ 1B(x,R)∩Σ ≤ C1B(x,R)∩Σ + C 2−j(n−1)θ 1[B(x,2 j R)\B(x,2 j−1 R)]∩Σ
j=1


∞
≤C 2−j(n−1)θ 1B(x,2 j R)∩Σ at every point on Σ. (6.2.131)
j=0

Then there exists C = C(Σ, p) ∈ (0, ∞) with the property that for any σ-measurable
function g on Σ we may write
∫ 
∞ ∫
|g| p w dσ ≤ 2−j(n−1)θ |g| p dσ
Σ j=0 B(x,2 j R)∩Σ
 p

∞
n−1−λ
⨏  1/p
−j(n−1)θ λ
≤C 2 (2 R)
j
(2 R)
j p |g| dσ
p

j=0 B(x,2 j R)∩Σ


∞ 
≤ CRλ 2−j[(n−1)θ−λ] g M p, λ (Σ,σ)
p

j=0

λ p
= CR g M p, λ (Σ,σ), (6.2.132)

based on (6.2.131), (6.2.2), and the fact that, by design, (n − 1)θ − λ > 0.
Take Wn,λ := Cθ and CΣ, p := C 1/p with C = C(Σ, p) ∈ (0, ∞) as in (6.2.132). For
( f , g) ∈ F arbitrary we may then estimate
∫ ∫
| f | p dσ = | f | p 1B(x,R)∩Σ dσ
B(x,R)∩Σ Σ
∫ ∫
≤ | f | p w dσ ≤ (Cw ) p |g| p w dσ
Σ Σ

≤ (CΣ, p,λ ) p · Rλ g M p, λ (Σ,σ) .

p
(6.2.133)
340 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Above, the first inequality uses the top line in (6.2.129), the second inequality is
based on (6.2.120) and the bottom line in (6.2.129), while the final inequality is
provided by (6.2.132) and (6.2.121). Dividing by σ B(x, R) ∩ Σ , raising to the
n−1−λ
power 1/p, then multiplying by R p and, finally, taking the supremum over all
x ∈ Σ and R ∈ 0, 2 diam(Σ) then yields (6.2.122). 

It turns out that the Hardy-Littlewood maximal operator is bounded on Morrey

spaces defined on Ahlfors regular sets. The full-space Euclidean version of the result
below first appeared in [27].

Corollary 6.2.13 Suppose Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a closed Ahlfors

regular set and abbreviate σ := H n−1 Σ. Then, given any exponent p ∈ (1, ∞) and
any number λ ∈ (0, n − 1), the Hardy-Littlewood maximal operator on Σ induces a
well-defined, sub-linear and bounded mapping

M Σ : M p,λ (Σ, σ) −→ M p,λ (Σ, σ). (6.2.134)

As a consequence of this, (6.2.14), and [133, (7.6.18)], one also has that

M Σ : M̊ p,λ (Σ, σ) −→ M̊ p,λ (Σ, σ) (6.2.135)

is a well-defined, sub-linear and bounded mapping.

Proof This is a direct consequence of Proposition 6.2.12 and items (1), (5) in [133,
Lemma 7.7.1].
It is also possible to provide an alternative proof of (6.2.134) based on the
boundedness result from Corollary 6.2.11, the duality result from Proposition 6.2.8,
and Fefferman-Stein’s maximal inequality in the version recorded in [133, Proposi-
tion 7.6.7], which presently gives that

for each s ∈ (1, ∞) ∫ C = C(Σ, s) ∈ (0, ∞) with

∫ there exists a constant
s
the property that M Σ f g dσ ≤ C f s M Σ g dσ for any two (6.2.136)
Σ Σ
non-negative σ-measurable functions f , g defined on Σ.

Specifically, pick some number θ ∈ (1, p) and denote by (p/θ) ∈ (1, ∞) the Hölder
conjugate exponent of p/θ ∈ (1, ∞). Then, given any function f ∈ M p,λ (Σ, σ), we
may write:
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 341

MΣ f M p, λ (Σ,σ)

 θ 1/θ  θ 1/θ
= M Σ f  = M Σ f  
M p/θ, λ (Σ,σ) (H (p/θ ) , λ (Σ,σ))∗
∫  θ 1/θ
  
≤C sup  M Σ f  g dσ 
 Σ
g ∈H (p/θ ) , λ (Σ,σ)
g H (p/θ ), λ (Σ, σ) ≤1

∫  1/θ
θ
≤C· sup M Σ | f | |g| dσ
 Σ
g ∈H (p/θ ) , λ (Σ,σ)
g H (p/θ ), λ (Σ, σ) ≤1

∫  1/θ
≤C· sup | f | θ M Σ g dσ
 Σ
g ∈H (p/θ ) , λ (Σ,σ)
g H (p/θ ), λ (Σ, σ) ≤1

  1/θ
≤C· sup | f |θ M p/θ, λ (Σ,σ)
MΣ g H (p/θ ), λ (Σ,σ)

g ∈H (p/θ ) , λ (Σ,σ)
g H (p/θ ), λ (Σ, σ) ≤1

  1/θ
≤C· sup | f |θ M p/θ, λ (Σ,σ)
gH (p/θ ), λ (Σ,σ)

g ∈H (p/θ ) , λ (Σ,σ)
g H (p/θ ), λ (Σ, σ) ≤1

≤ C | f |θ
1/θ
M p/θ, λ (Σ,σ)
= C f  M p, λ (Σ,σ), (6.2.137)

thanks to (6.2.4), Proposition 6.2.8, (6.2.136), and Corollary 6.2.11. This establish-
es the boundedness of the Hardy-Littlewood maximal operator in the context of
(6.2.134). 

Corollary 6.2.13 is a key ingredient the proof of the following version of the
Fractional Integration Theorem in the context of Morrey spaces defined on Ahlfors
regular sets, generalizing the classical result in [1].

Proposition 6.2.14 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors

regular set and abbreviate σ := H n−1 Σ. Fix p ∈ (1, ∞) along with λ ∈ (0, n − 1)
and suppose

n−1−λ 1 α  −1
0<α< , p∗ := − . (6.2.138)
p p n−1−λ
Finally, recall the fractional integral operator IΣ,α of order α on Σ (cf. [133, (7.8.5)];
see also (A.0.51))
342 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

IΣ,α : L 1 Σ, 1+ |xσ(x)
| n−1−α
−→ Lloc
1
(Σ, σ),
∫ (6.2.139)
f (y)
IΣ,α f (x) := dσ(y) for σ-a.e. x ∈ Σ.
Σ |x − y|
n−1−α

Then  
σ(x)
M p,λ (Σ, σ) → L 1 Σ, (6.2.140)
1 + |x| n−1−α

and there exists a constant C = C(Σ, n, p, α, λ) ∈ (0, ∞) with the property that

IΣ,α f M p∗ , λ (Σ,σ)
≤ C f  M p, λ (Σ,σ) for each f ∈ M p,λ (Σ, σ). (6.2.141)

Proof The inclusion in (6.2.140) is a consequence of (6.2.23) and the current as-
sumptions on p, λ, α. To proceed, denote by M Σ the Hardy-Littlewood maximal
operator on Σ, and by p ∈ (1, ∞) the Hölder conjugate exponent of p. The assump-
tions on α ensure that we may decompose
λ n−1
n − 1 − α = a + b with a > and b > . (6.2.142)
p p

Next, fix a function f ∈ M p,λ (Σ, σ) along with a point x ∈ Σ. For some r > 0 to be
specified later, use [133, (7.8.20)] to estimate
∫
| f (y)|
dσ(y) ≤ Cr α (M Σ f )(x), (6.2.143)
B(x,r)∩Σ |x − y| n−1−α

for some constant C ∈ (0, ∞) which is independent of x, r, f . Also, based on

(6.2.142), Hölder’s inequality, [133, (7.2.5)], and (6.2.2) we may write
∫
| f (y)|
dσ(y)
Σ\B(x,r) |x − y| n−1−α
∫ | f (y)| p  1/p  ∫ dσ(y)  1/p
≤ dσ(y) bp
Σ\B(x,r) |x − y| Σ\B(x,r) |x − y|
ap

∞ ∫
  1/p
  | f (y)| p
≤ Cr (n−1−bp )/p dσ(y)
j=0 Σ∩[B(x,2 j+1 r)\B(x,2 j r)] |x − y| ap


∞ ∫  1/p
 
≤ Cr (n−1−bp )/p (2 j r)−ap | f (y)| p dσ(y)
j=0 Σ∩B(x,2 j+1 r)

 

∞  1/p
≤ Cr (n−1−bp )/p (2 j r)−ap (2 j r)λ  f  M p, λ (Σ,σ)
p

j=0

= Cr (λ−n−1)/p+α  f  M p, λ (Σ,σ) . (6.2.144)

6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 343

Combining (6.2.143) with (6.2.144) then yields, after solving an elementary opti-
mization problem,
   
 IΣ,α f (x) ≤ C · inf r α (M Σ f )(x) + r (λ−n−1)/p+α  f  M p, λ (Σ,σ)
r >0

αp   1− n−1−λ
αp

= C f  M p, λ (Σ,σ) (M Σ f )(x)
n−1−λ

1− pp∗
  pp
∗
= C f  M p, λ (Σ,σ) (M Σ f )(x) , ∀x ∈ Σ, (6.2.145)

where the final equality taken into account the formula for p∗ given in (6.2.138). For
each point x ∈ Σ and each scale R ∈ 0, 2 diam(Σ) we may then use (6.2.145) and
Corollary 6.2.13 to estimate
⨏    1 p ⨏  p1
 IΣ,α f  p∗ dσ p∗ ≤ C f  1−p,p∗λ M Σ f
p
dσ
∗
M (Σ,σ)
Σ∩B(x,R) Σ∩B(x,R)

1− p
 λ−n−1  pp
∗
≤ C f  M p,p∗λ (Σ,σ) R p  f  M p, λ (Σ,σ)

λ−n−1
≤ C f  M p, λ (Σ,σ) R p∗ . (6.2.146)
λ−n−1
Dividing by R p∗ and taking the supremum over all x ∈ Σ and R ∈ 0, 2 diam(Σ)
then yields (6.2.141), on account of (6.2.2). 
Corollary 6.2.13 may now be thought as the limiting case (corresponding to
α = 0) of the result pertaining to the action of the fractional Hardy-Littlewood
maximal operator of order α on Morrey spaces, presented below.
Corollary 6.2.15 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular
set and abbreviate σ := H n−1 Σ. Suppose
n−1−λ n−1−λ
s ∈ (0, ∞), 0 < λ < n − 1, 0<α< , s<p< ,
s(n − 1) (n − 1)α
1 (n − 1)α  −1
and consider the integrability exponent p := − .
p n−1−λ
(6.2.147)
Also, recall from (A.0.68) that the L s -based fractional Hardy-Littlewood maximal
operator of order α on Σ acts on σ-measurable functions f defined on Σ according
to
 ⨏  s1 
α
M Σ,s,α f (x) := sup σ B(x, r) ∩ Σ | f | dσ
s
, ∀x ∈ Σ. (6.2.148)
r >0 B(x,r)∩Σ

Then there exists a constant C = C(Σ, n, p, λ, α, s) ∈ (0, ∞) with the property that

M Σ,s,α f M p , λ (Σ,σ)
≤ C f  M p, λ (Σ,σ) for each f ∈ M p,λ (Σ, σ). (6.2.149)
344 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

As a consequence of this, (6.2.14), and [133, (7.6.4)], one also has that

M Σ,s,α : M̊ p,λ (Σ, σ) −→ M̊ p,λ (Σ, σ) (6.2.150)

is a well-defined, sub-linear and bounded mapping.

Proof Define
p  n − 1 − λ
p := ∈ (1, ∞) and 
α := αs(n − 1) ∈ 0, . (6.2.151)
s p
Given any σ-measurable functions f defined on Σ, we have
 1
s
 1
s
M Σ,s,α f (x) = M Σ,1,αs | f | s (x) ≤ C IΣ, α | f | s (x) , ∀x ∈ Σ, (6.2.152)

thanks to (6.2.148), (6.2.139), and [133, (7.8.19)]. Observe that the assignments

M p,λ (Σ, σ)  f → | f | s ∈ M p,λ (Σ, σ),

M p,λ (Σ, σ)  g → IΣ, α g ∈ M p /s,λ (Σ, σ), (6.2.153)

M p /s,λ (Σ, σ)  h → |h| s ∈ M p,λ (Σ, σ),

1

map bounded sets to bounded sets. Indeed, in the case of the first and third as-
signments this is clear from (6.2.4), while for the middle assignment the desired
conclusion is furnished by Proposition 6.2.14. At this stage, (6.2.149) follows on
account of the observation made in relation to (6.2.153), (6.2.152), (6.2.3), and [133,
(7.6.2)]. 
Here is the duality result advertised earlier (in conjunction with (6.2.22)), i-
dentifying precisely the dual of the space M̊ p,λ (Σ, σ), defined in (6.2.14) (cf. also
(6.2.20)).

Proposition 6.2.16 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Fix two integrability exponents p, q ∈ (1, ∞)
satisfying 1/p + 1/q = 1, along with a parameter λ ∈ (0, n − 1). Throughout, treat
M̊ p,λ (Σ, σ) as a Banach space equipped with the norm inherited from M p,λ (Σ, σ).
Then the mapping
∗
B q,λ (Σ, σ)  g −
 → Λg ∈ M̊ p,λ (Σ, σ) given by
∫ (6.2.154)
Λg ( f ) := f g dσ for each f ∈ M̊ p,λ (Σ, σ)
Σ

is a well-defined, linear, bounded isomorphism, with bounded inverse. As such, the

integral paring yields the identification
 ∗
M̊ p,λ (Σ, σ) = B q,λ (Σ, σ). (6.2.155)
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 345

As a corollary, whenever Λ ∈ Lipc (Σ) is a distribution with the property that
there exists some constant C ∈ (0, ∞) such that
 
 Λ, φ  ≤ Cφ M p, λ (Σ,σ) for every φ ∈ Lip (Σ), (6.2.156)
c

there exists a unique function g ∈ B q,λ (Σ, σ) such that

∫
Λ, φ = g φ dσ for every φ ∈ Lipc (Σ). (6.2.157)
Σ

In addition, with C denoting the constant appearing in (6.2.156), the function g

satisfies g B q, λ (Σ,σ) ≤ c C, where c ∈ (0, ∞) is a fixed number which depends only
the ambient.

Retaining the setting of Proposition 6.2.16, for each function g ∈ B q,λ (Σ, σ) we
have

g B q, λ (Σ,σ) ≈ Λg ( M̊ p, λ (Σ,σ))∗ = sup |Λg ( f )|

f ∈ M̊ p, λ (Σ,σ)
 f  M p, λ (Σ, σ) ≤1

∫ 
 
= sup  f g dσ , (6.2.158)
f ∈ M̊ p, λ (Σ,σ) Σ
 f  M p, λ (Σ, σ) ≤1

by virtue of the fact that (6.2.154) is an isomorphism (compare with (6.2.88)).

In relation to this, we also wish to remark that for each σ-measurable function
g : Σ → R we have
∫
g ∈ B (Σ, σ) if and only if
q,λ
sup | f ||g| dσ < +∞, (6.2.159)
f ∈M p, λ (Σ,σ) Σ
 f  M p, λ (Σ, σ) ≤1

and for each function g ∈ B q,λ (Σ, σ) we have

∫
g B q, λ (Σ,σ) ≈ sup | f ||g| dσ =: |||g||| B q, λ (Σ,σ) . (6.2.160)
f ∈M p, λ (Σ,σ) Σ
 f  M p, λ (Σ, σ) ≤1

Indeed, the right-pointing implication in (6.2.159) is a consequence of (6.2.78). To

prove the left-pointing implication in (6.2.159), denote by Cg ∈ [0, ∞) the supremum
appearing there. The fact that Cg is a finite number forces g ∈ L r (Σ, σ) with
integrability index r := s  = n−1+λ(q−1)
q(n−1)
∈ (1, ∞), by virtue of (6.2.7) and Riesz’s
Representation Theorem. In particular, g ∈ Lloc 1 (Σ, σ). Define next

∫
Λg : M̊ p,λ (Σ, σ) → C, Λg ( f ) := f g dσ for each f ∈ M̊ p,λ (Σ, σ). (6.2.161)
Σ
346 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Thanks to (6.2.15) and the fact that Cg is finite, it follows that Λg is a well-defined,
linear, and bounded functional on M̊ p,λ (Σ, σ), with norm ≤ C(Σ, q, λ) · Cg . Granted
this, we may now invoke Proposition 6.2.16 to conclude that there exists a (unique)
function g ∈ B q,λ (Σ, σ) satisfying 
g  B q, λ (Σ,σ) ≈ Cg and such that
∫ ∫
f g dσ = f
g dσ for each f ∈ M̊ p,λ (Σ, σ). (6.2.162)
Σ Σ

Bearing in mind that g ∈ Lloc 1 (Σ, σ), from (6.2.162), [133, Corollary 3.7.3], and

(6.2.15) we then conclude that g =  g at σ-a.e. point on Σ. As such, we have

that g ∈ B q,λ (Σ, σ) and g B q, λ (Σ,σ) ≤ C(Σ, q, λ) · Cg . This finishes the proof of
(6.2.159), and also establishes the left-pointing inequality in (6.2.160). Finally, the
right-pointing inequality in (6.2.160) is a consequence of (6.2.78).
We next turn to the task of presenting the proof of Proposition 6.2.16.
Proof of Proposition 6.2.16 Given that M̊ p,λ (Σ, σ) is a closed linear subspace of
M p,λ (Σ, σ) (cf. (6.2.15)), we conclude from this and (6.2.78) that the mapping
defined in (6.2.154) is well defined, linear, and bounded. In addition, from (6.2.71),
(6.2.15), and [133, (3.7.23)] we see that the map in question is also injective. The
bulk of the proof deals with the issue of the surjectivity (with norm-control) of the
mapping introduced in (6.2.154). Following [163] where a similar result has been
proved in the full Euclidean ambient, we shall establish this working with a discrete
version of the Morrey space. Defining it requires some preparations, and we divide
the argument devoted to this issue into four steps.
Step 1: Functional Analytic Preliminaries. Let A = { A j } j ∈N be a sequence of
Banach spaces. Define the Banach spaces
 
∞ 
1 (A) := a = {a j } j ∈N ∈ Π j ∈N A j : a1 (A) := a j  A j < ∞ , (6.2.163)
j=1

∞ (A) := a = {a j } j ∈N ∈ Π j ∈N A j : a∞ (A) := sup a j  A j < ∞}, (6.2.164)
j ∈N

and consider
 
c0 (A) := a = {a j } j ∈N ∈ ∞ (A) : lim a j  A j = 0 . (6.2.165)
j→∞

Then c0 (A) is a closed subspace of ∞ (A) which means that c0 (A) becomes itself
a Banach space when equipped with the norm  · ∞ (A) .
Next, define A ∗ := { A∗j } j ∈N where each A∗j is the dual of A j , and introduce the
pairing

(a, a ) := ∞ 
j=1 (a j , a j ) for each
(6.2.166)
a = {a j } j ∈N ∈ Π j ∈N A j and a  = {a j } j ∈N ∈ Π j ∈N A∗j .
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 347

Then, according to [48, pp. 73–74] (cf. also [186, Lemma 1.11.1, p. 68]),

under the pairing (6.2.165), the dual of the Banach space c0 (A)
(6.2.167)
may be canonically identified with the space 1 (A ∗ ).
Finally, given any two Banach X, Y , we agree to call a linear map Φ : X → Y a
quasi-isometry, if there exists θ ∈ (0, 1] such that θ  xX ≤ ΦxY ≤ θ −1  xX
for each x ∈ X. In particular, in such a scenario Φ maps any closed subspace of X
into a closed subspace of Y in an isomorphic fashion.
Step 2: Labeling the Dyadic Grid on Σ. The triplet Σ, | · − · |, σ is a space of
homogeneous type, whose underlying measure is complete, locally finite, and Borel-
regular (cf. [133, (3.6.26)]). As such, [133, Proposition 7.5.4] ensures that there
exists a dyadic cube structure (or dyadic grid) on Σ, denoted by D(Σ) (cf. (A.0.36)).
As mentioned in [133, Remark 7.5.5], for each dyadic cube Q ∈ D(Σ) we shall
denote by (Q) and xQ , respectively, the side-length and center of Q. Given that
D(Σ) is a countable set, it may be bijectively identified with N, though not in a
canonical fashion. To single out a specific identification which suits our purposes,
fix a reference point x0 ∈ Σ and define
 
rank (Q) := max (Q), (Q)−1, dist(x0, Q) , ∀Q ∈ D(Σ). (6.2.168)

If for each j ∈ N we introduce

 
D j (Σ) := Q ∈ D(Σ) : j − 1 < rank (Q) ≤ j (6.2.169)

then each set D j (Σ) will have finite cardinality, i.e., #D j (Σ) < +∞ for each j,
#
and there holds D(Σ) = j ∈N D j (Σ), disjoint union. Let us now label D1 (Σ) using
indices from the set 1, 2, . . . , #D1 (Σ) , then next label D2 (Σ) using indices from
the set #D1 (Σ) + 1, #D1 (Σ) + 2, . . . , #D1 (Σ) + #D2 (Σ) , and so on. Ultimately, this
yields

an enumeration Q1, Q2, . . . , Q j , Q j+1, . . . of D(Σ) with

(6.2.170)
the property that rank (Q j ) → ∞ as j → ∞.

Whenever convenient, henceforth we agree to identify D(Σ) with N in the manner

just indicated.
Step 3: Discretizing Morrey Spaces. Define the family of Banach spaces

A Σ := AQ Q ∈D(Σ)
with AQ := L p Q, (Q)−λ σ for each Q ∈ D(Σ). (6.2.171)

Then, according to the discussion in Steps 1-2 (in particular, bearing in mind the
identification of D(Σ) with N from (6.2.170)), we have

A Σ∗ = A∗Q Q ∈D(Σ)
with A∗Q = L q Q, (Q)−λ σ for each Q ∈ D(Σ), (6.2.172)
348 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

and for each F = { fQ }Q ∈D(Σ) ∈ ∞ A Σ and G = {gQ }Q ∈D(Σ) ∈ 1 A Σ∗ we may

write
∫  1/p
F ∞ (AΣ ) = sup | fQ | p (Q)−λ dσ , (6.2.173)
Q ∈D(Σ) Q

 ∫  1/q
G1 (AΣ∗ ) = |gQ | q (Q)−λ dσ , (6.2.174)
Q ∈D(Σ) Q

 ∫
(F, G) = fQ gQ (Q)−λ dσ. (6.2.175)
Q ∈D(Σ) Q

Consider next the linear assignment

Φ 
M p,λ (Σ, σ)  f −→ F := f Q Q ∈D(Σ)
∈ ∞ A Σ , (6.2.176)

in relation to which we claim that

Φ is a quasi-isometry which maps Lipc (Σ) into c0 A Σ . (6.2.177)

To justify this, fix an arbitrary function f ∈ M p,λ (Σ, σ). Parts (6) and (9) of [133,
Proposition 7.5.4] ensure that there exists a number N ∈ N with the property that
for each x ∈ Σ and R ∈ 0, 2 diam(Σ) there exists at most N dyadic cubes, say
{Qi }i ∈I ⊆ D(Σ) with # I ≤ N, such that for each i ∈ I we have

(Qi ) ≈ R with proportionality constants independent of x, R,

  (6.2.178)
and σ B(x, R) ∩ Σ \ ∪i ∈I Qi = 0.

In turn, these properties permit us to estimate

n−1−λ  ⨏  p1  ∫  p1
−λ
R p | f | dσ
p
≈ R | f | p dσ
Σ∩B(x,R) Σ∩B(x,R)

 ∫  p1
−λ
≤C (Qi ) | f | p dσ
i ∈I Qi

∫  1/p
≤ C · sup | f | p (Q)−λ dσ
Q ∈D(Σ) Q

= CΦ( f )∞ (AΣ ) . (6.2.179)

Taking the supremum over x and R then proves that

 f  M p, λ (Σ,σ) ≤ CΦ( f )∞ (AΣ ) . (6.2.180)

6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 349

Since part (5) of [133, Proposition 7.5.4] gives the inequality in the opposite direction,
we conclude that Φ in (6.2.176) is indeed a quasi-isometry.
To see that Φ maps compactly supported Lipschitz functions on Σ into c0 A Σ ,
fix an arbitrary f ∈ Lipc (Σ). Then the fact that λ > 0 implies

f |Q L p (Q,(Q)−λ σ)
≤ (Q)−λ/p  f  L p (Σ,σ) −→ 0 as (Q) → ∞, (6.2.181)

while having λ < n − 1 forces

⨏  1/p
f |Q L p (Q,(Q)−λ σ)
≈ (Q)(n−1−λ)/p | f | p dσ
Q

≤ (Q)(n−1−λ)/p  f  L ∞ (Σ,σ) −→ 0 as (Q) → 0. (6.2.182)

In addition, the fact that f has compact support ensures that

for ε ∈ (0, 1) fixed, f |Q L p (Q,(Q)−λ σ) → 0 as dist(x0, Q) → ∞ within

(6.2.183)
the class of cubes Q ∈ D(Σ) satisfying ε < (Q) < ε−1.
In turn, from (6.2.181)-(6.2.183) and (6.2.170) (also bearing in mind (6.2.168)) we
conclude that, as wanted, Φ( f ) ∈ c0 A Σ . This finishes the proof of (6.2.177).
Having established (6.2.177), the fact that Φ is continuous implies, on account of
(6.2.20), that actually
the quasi-isometry Φ, defined in (6.2.176),
(6.2.184)
maps the space M̊ p,λ (Σ, σ) into c0 A Σ .

As a consequence,

Φ M̊ p,λ (Σ, σ) is a closed subspace of c0 A Σ and

(6.2.185)
Φ : M̊ p,λ (Σ, σ) → Φ M̊ p,λ (Σ, σ) is an isomorphism.

Henceforth we agree to denote by Φ−1 the inverse of the above isomorphism. Hence,

Φ−1 : Φ M̊ p,λ (Σ, σ) −→ M̊ p,λ (Σ, σ) (6.2.186)

is itself an isomorphism with the property that

Φ−1 Φ( f ) = f for each f ∈ M̊ p,λ (Σ, σ). (6.2.187)

∗
Step 4: Proof of the Duality Formula. Fix an arbitrary Λ ∈ M̊ p,λ (Σ, σ) . Then
Λ ◦ Φ−1 is a linear continuous functional on Φ M̊ p,λ (Σ, σ) , which is a closed
subspace of c0 A Σ . The Hahn-Banach Theorem then guarantees the existence of
some linear continuous functional Λ  on c0 A Σ satisfying

 
Λ = Λ ◦ Φ−1 on Φ M̊ p,λ (Σ, σ) and
Φ( M̊ p, λ (Σ,σ))
(6.2.188)
 −1
 Λc0 (AΣ )∗ = Λ ◦ Φ (Φ( M̊ p, λ (Σ,σ)))∗ ≈ Λ( M̊ p, λ (Σ,σ))∗ .
350 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

Hence, Λ  belongs to the dual of c0 A Σ , which has been identified as 1 A ∗

Σ
in (6.2.167). Specifically, there exists some G = {gQ }Q ∈D(Σ) ∈ 1 A Σ∗ with the
property that
 ∫

Λ(F) = (F, G) = fQ gQ (Q)−λ dσ
Q ∈D(Σ) Q
(6.2.189)
for each F = { fQ }Q ∈D(Σ) ∈ c0 A Σ ,
and the second line in (6.2.188) further gives

G1 (AΣ∗ ) ≈ Λ( M̊ p, λ (Σ,σ))∗ . (6.2.190)


In particular, for each f in M̊ p,λ (Σ, σ), the fact that F := f Q Q ∈D(Σ)
= Φ( f )
belongs to Φ M̊ p,λ (Σ, σ) ⊆ c0 A Σ permits us to write


Λ( f ) = Λ Φ−1 (Φ( f )) = (Λ ◦ Φ−1 ) Φ( f ) = Λ(F)
 ∫   ∫
= f Q gQ (Q)−λ dσ = f gQ (Q)−λ dσ, (6.2.191)
Q ∈D(Σ) Q Q ∈D(Σ) Q

thanks to (6.2.187), (6.2.188), and (6.2.189).

To proceed, for each dyadic cube Q ∈ D(Σ) denote by 
gQ the extension by zero of
the function gQ ∈ A∗Q = L q Q, (Q)−λ σ to the entire set Σ. Then each 
gQ belongs
to L q (Σ, σ) and


gQ L q (Σ,σ)
= gQ L q (Q,σ)
= (Q)λ/q gQ L q (Q,(Q)−λ σ)
. (6.2.192)

In concert with (6.2.190), this implies

 
(Q)−λ/q gQ L q (Σ,σ) = gQ L q (Q,(Q)−λ σ)
Q ∈D(Σ) Q ∈D(Σ)

= G1 (AΣ∗ ) ≈ Λ( M̊ p, λ (Σ,σ))∗ . (6.2.193)

q(n−1)
Consequently, if r := n−1+λ(q−1) ∈ (1, q) then
 
gQ (Q)−λ
 L r (Σ,σ)
≤C· (Q)−λ+(n−1)(1/r−1/q) 
gQ L q (Σ,σ)
Q ∈D(Σ) Q ∈D(Σ)

=C· (Q)−λ/q 
gQ L q (Σ,σ)
Q ∈D(Σ)

≤ CΛ( M̊ p, λ (Σ,σ))∗ < +∞, (6.2.194)

thanks to Hölder’s inequality (together with the nature of the support of 

gQ , property
[133, (7.5.9)] satisfied by dyadic cubes, and the Ahlfors regularity of Σ), the fact that
(n − 1)(1/r − 1/q) = λ(1 − 1/q), and (6.2.193). This proves that
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 351

the series g := gQ (Q)−λ converges absolutely in L r (Σ, σ),
 (6.2.195)
Q ∈D(Σ)


hence also pointwise σ-a.e. on Σ, as well as in Lipc (Σ) . If for each Q ∈ D(Σ) we
now define

λQ := (Q)−λ/q 
gQ L q (Σ,σ)
∈ [0, ∞) (6.2.196)

together with
 
gQ (Q)−λ λQ if λQ  0,

bQ := (6.2.197)
0 if λQ = 0,

then we may express g, originally defined in (6.2.195), as

 
g= λQ bQ in L r (Σ, σ), hence also in Lipc (Σ) , (6.2.198)
Q ∈D(Σ)

while (6.2.193) tells that


|λQ | ≤ CΛ( M̊ p, λ (Σ,σ))∗ < +∞. (6.2.199)
Q ∈D(Σ)

In addition, the bQ ’s are designed such that for each Q ∈ D(Σ) we have
1
λ q −1
supp bQ ⊆ Q and bQ  L q (Σ,σ) ≤ (Q) . (6.2.200)

From (6.2.198)-(6.2.200) and (6.2.63)-(6.2.65) we conclude (by also bearing in mind

[133, (7.5.9)]) that

g ∈ B q,λ (Σ, σ) and g B q, λ (Σ,σ) ≤ CΛ( M̊ p, λ (Σ,σ))∗ . (6.2.201)

Pressing on, given any f ∈ L s (Σ, σ) with s := p(n−1)

n−1−λ , we may write
 ∫ ∫   
Λ( f ) = f gQ (Q)−λ dσ = f gQ (Q)−λ dσ

Q ∈D(Σ) Q Q ∈D(Σ)
Σ
∫
= f g dσ = Λg ( f ), (6.2.202)
Σ

where the first equality is implied by (6.2.191) and (6.2.14), the second equality
is a consequence of (6.2.195) and the observation that 1/r + 1/s = 1, the third
equality uses (6.2.195), while the last equality comes from (6.2.154) (again, bearing
in mind (6.2.14)). Since L s (Σ, σ) is a dense subspace of M̊ p,λ (Σ, σ) (cf. (6.2.15)),
and since both Λ and Λg are continuous functions on M̊ p,λ (Σ, σ), we deduce from
∗
(6.2.202) that actually Λg = Λ as functionals in M̊ p,λ (Σ, σ) . Given that Λ has
352 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .
∗
been arbitrarily selected in M̊ p,λ (Σ, σ) to begin with, this ultimately proves that
the mapping defined in (6.2.154) is surjective.
At this point, we may conclude that the mapping defined in (6.2.154) is actually
well-defined, linear, bounded, bijective, and the estimate in (6.2.201) also shows that
its inverse is also bounded.
To complete the proof of Proposition 6.2.16 there remains to address the very last
claim in the statement. In this regard, since Lipc (Σ) is dense in the space M̊ p,λ (Σ, σ)
(cf. (6.2.20)) and since the mapping Lipc (Σ)  φ → Λ, φ ∈ C is linear and bounded
with respect to the norm in M̊ p,λ (Σ, σ) (inherited from M p,λ (Σ, σ)) it follows that
∗
said mapping extends (by density) to a unique functional Θ ∈ M̊ p,λ (Σ, σ) whose
norm is ≤ c C, where the constant C is as in (6.2.156) and where c ∈ (0, ∞) is a
fixed number which depends only the ambient. In view of (6.2.155), this proves that
∫ some function g in the space B (Σ, σ) such that g B q, λ (Σ,σ) ≤ c C and
there exists q,λ

Θ( f ) = Σ f g dσ for each f ∈ M p,λ (Σ, σ). In concert with the fact that Θ(φ) = Λ, φ
for each φ ∈ Lipc (Σ), this establishes (6.2.157). Finally, the uniqueness of g is a
consequence of [133, Corollary 3.7.3]. 

One of the main reasons we have altered the standard definition of a Banach
function space (as given in, e.g., [14, pp. 2-3]) is to be able to allow Morrey and
block spaces to be part of this theory. We deliver on this promise in the proposition
below.

Proposition 6.2.17 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Pick two integrability exponents p, q ∈ (1, ∞) with
1/p+1/q = 1, along with an arbitrary parameter λ ∈ (0, n−1). Then both M p,λ (Σ, σ)
and B q,λ (Σ, σ) are Generalized Banach Function Spaces (cf. Definition 5.1.4) and
theirs associated spaces (in the sense of Definition 5.1.11) are, respectively, given
by
   
M p,λ (Σ, σ) = B q,λ (Σ, σ) and B q,λ (Σ, σ) = M p,λ (Σ, σ). (6.2.203)

Simply put, M p,λ (Σ, σ) and B q,λ (Σ, σ) are mutually associated with one another in
the sense of Definition 5.1.11.

Once the Morrey and block spaces space have been identified as Generalized Ba-
nach Function Spaces, which are actually mutually associated with one another, the
abstract embedding results from Proposition 5.2.7 apply (thanks to Corollary 6.2.11
and Corollary 6.2.13) and yield embeddings like the ones in Proposition 5.2.7 for
Morrey and block spaces. We shall revisit this point of view later, in Proposi-
tion 6.2.22.
Proof of Proposition 6.2.17 First, the fact that the Morrey space M p,λ (Σ, σ) is a
Generalized Banach Function Space may be checked from Definition 5.1.4, presently
used with X := Σ, μ := σ, and with the function norm ρ : M+ (Σ, σ) → [0, ∞] given
by
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 353
 
n−1−λ ⨏  p1
ρ( f ) := sup R p f dσp
(6.2.204)
x ∈Σ and Δ(x,R)
0<R<2 diam(Σ)

for each non-negative σ-measurable function f on Σ. Specifically, properties (P1)-

(P3) in Definition 5.1.4 are seen from (6.2.204) plus (6.2.8), (6.2.3), and (6.2.6). We
now claim that properties (P4)-(P5) in Definition 5.1.4 are true if we fix a reference
point x0 ∈ Σ and take

Yj := Z j := Δ(x0, j) for each j ∈ N with j < 2 diam(Σ). (6.2.205)

Indeed, for this choice, (P4) becomes a consequence of (6.2.204) and (6.2.9), while
(P5) is seen directly from (6.2.204) and Hölder’s inequality. Concretely, bearing in
mind that Σ is upper Ahlfors regular, for each non-negative σ-measurable function
f on Σ and each j ∈ N with j < 2 diam(Σ) we may write
∫ ⨏ ⨏  p1
f dσ = σ Δ(x0, j) f dσ ≤ CΣ · j n−1
f p dσ
Δ(x0, j) Δ(x0, j) Δ(x0, j)
 
n−1 λ
+
n−1−λ ⨏  p1
= CΣ · j q p j p p
f dσ ≤ c j ρ( f ), (6.2.206)
Δ(x0, j)

n−1 λ
+
with c j := CΣ · j q p ∈ (0, ∞). This shows that M p,λ (Σ, σ) is indeed a Generalized
Banach Function Space. Having established that, from (5.1.35)-(5.1.36) (used with
X := M p,λ (Σ, σ)) and (6.2.159)-(6.2.160) we then conclude that the associated space
of M p,λ (Σ, σ) (in the sense of Definition 5.1.11) coincides, as a set, with B q,λ (Σ, σ),
and

g  = |||g||| B q, λ (Σ,σ) ≈ g B q, λ (Σ,σ), uniformly for g ∈ B q,λ (Σ, σ).
M p, λ (Σ,σ)
(6.2.207)
Granted this, from Definition 5.1.11 we see that B q,λ (Σ, σ) equipped with the
(equivalent) norm ||| · ||| B q, λ (Σ,σ) is itself a Generalized Banach Function Space,
and Proposition 5.1.14 further implies that its associated space is M p,λ (Σ, σ). The
claims in (6.2.203) are therefore true (in this interpretation). 
In particular, Proposition 6.2.17 implies the following corollary concerning point-
wise convergence of functions in block spaces, strongly reminiscent of Lebesgue’s
Monotone Convergence Theorem and Fatou’s Lemma.
Corollary 6.2.18 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular
set and abbreviate σ := H n−1 Σ. Select an integrability exponent q ∈ (1, ∞) and a
parameter λ ∈ (0, n − 1). Then there exists a constant C = C(Σ, q, λ) ∈ (0, ∞) with
the property that the following statements are true for any given sequence {g j } j ∈N
of σ-measurable functions on Σ.
(i) If for each j ∈ N one has 0 ≤ g j (x) ≤ g j+1 (x) at σ-a.e. point x ∈ Σ, and if one
defines g(x) := lim g j (x) at σ-a.e. point x ∈ Σ, it follows that
j→∞
354 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

either g  B q,λ (Σ, σ) and g j  B q, λ (Σ,σ)  ∞ as j → ∞,

(6.2.208)
or g ∈ B q,λ (Σ, σ) and lim g j  B q, λ (Σ,σ) ≤ Cg B q, λ (Σ,σ) .
j→∞

(ii) If lim inf g j  B q, λ (Σ,σ) < ∞ and g(x) := lim g j (x) exists at σ-a.e. point x ∈ Σ,
j→∞ j→∞
then

g ∈ B q,λ (Σ, σ) and g B q, λ (Σ,σ) ≤ C lim inf g j  B q, λ (Σ,σ) . (6.2.209)

j→∞

Proof All claims are direct consequences of Proposition 6.2.17 plus (6.2.207), and
Lemma 5.1.7 (bearing in mind the first property in [133, (3.6.26)] and [133, item
(iii) in Remark 3.1.2]). 
By allowing us to invoke the Banach-Alaoglu Theorem, Proposition 6.2.16 makes
it possible to prove the following weak-∗ convergence result involving the pre-duals
of Morrey spaces, which is somewhat akin to the weak-∗ convergence theorem of
Jones-Journé for the Hardy space H 1 (Rn ) found in [104] (see Proposition 4.6.3).

Proposition 6.2.19 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors

regular set and abbreviate σ := H n−1 Σ. Pick p, q ∈ (1, ∞) with 1/p + 1/q = 1 and
fix some λ ∈ (0, n − 1). In this setting, suppose

{g j } j ∈N is a bounded sequence in B q,λ (Σ, σ)

(6.2.210)
and g(x) := lim g j (x) exists for σ-a.e. x ∈ Σ.
j→∞

Then
∫ ∫
g∈B q,λ
(Σ, σ) and lim f g j dσ = f g dσ for each f ∈ M̊ p,λ (Σ, σ).
j→∞ Σ Σ
(6.2.211)

Proof Granted (6.2.210), item (ii) in Corollary 6.2.18 applies and gives the mem-
bership g ∈ B q,λ (Σ, σ). A direct proof of this fact, which actually gives more, is as
follows. For starters, we claim that
q(n − 1)
g ∈ L r (Σ, σ) with r := ∈ (1, q). (6.2.212)
n − 1 + λ(q − 1)

Indeed, since thanks to (6.2.71) and assumptions {g j } j ∈N is a bounded sequence in

L r (Σ, σ), the membership in (6.2.212) becomes a consequence of Fatou’s Lemma
which gives
∫ ∫ ∫
|g| dσ =
r
lim inf |g j | dσ ≤ lim inf |g j | r dσ
r
Σ Σ j→∞ j→∞ Σ
 r
≤ sup g j  L r (Σ,σ) < +∞. (6.2.213)
j ∈N
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 355

To proceed, fix f ∈ Lipc (Σ) along with some arbitrary threshold ε > 0. Since thanks
to (6.2.212) we have f g ∈ L 1 (Σ, σ) (hence the measure f g σ is finite and absolutely
continuous with respect to σ) it follows that there exists some δ > 0 with the property
that
∫
| f g| dσ < ε for each σ-measurable set A ⊆ Σ with σ(A) < δ. (6.2.214)
A

Also, Egoroff’s Theorem (cf., e.g., [55, Theorem 2.33, p. 62]) guarantees the exis-
tence of some σ-measurable set E ⊆ Σ with the property that

σ(E) < min{ε, δ} and lim g j − g L ∞ (supp f \E,σ) = 0. (6.2.215)

j→∞

For each j ∈ N write

∫ ∫ ∫ ∫
(g − g j ) f dσ = (g − g j ) f dσ − g j f dσ + g f dσ (6.2.216)
Σ supp f \E E E

and note that, thanks to the second property in (6.2.215),

∫ 
 
lim  (g−g j ) f dσ  ≤  f  L 1 (Σ,σ) lim g j −g L ∞ (supp f \E,σ) = 0, (6.2.217)
j→∞ supp f \E j→∞

whereas (6.2.214) and the first property in (6.2.215) ensure that

∫ 
 
 g f dσ  < ε. (6.2.218)
E

Finally, consider the middle term in the right-hand side of (6.2.216). With the
exponent p ∈ (1, ∞) satisfying 1/p + 1/q = 1, we rely on (6.2.78) to estimate this
term as follows:
∫  ∫  
 
 g j f dσ  ≤ |g j || f |1E dσ ≤ C sup g j  B q, λ (Σ,σ) | f |1E M p, λ (Σ,σ) .
E Σ j ∈N
(6.2.219)
On the other hand, combining (6.2.5) with (6.2.9) and (6.2.215) yields

| f |1E M p, λ (Σ,σ)
≤  f  L ∞ (Σ,σ) · ε (n−1−λ)/[p(n−1)] . (6.2.220)

Collectively, (6.2.216), (6.2.217), (6.2.218), (6.2.219), and (6.2.220) then prove, on

account of the arbitrariness of ε > 0 (and the fact that λ ∈ (0, n − 1)), that
∫ ∫
lim f g j dσ = f g dσ for each f ∈ Lipc (Σ). (6.2.221)
j→∞ Σ Σ

Next, the hypothesis that the sequence of functions {g j } j ∈N is bounded in the

∗
space B q,λ (Σ, σ) = M̊ p,λ (Σ, σ) (cf. (6.2.155) for the latter equality) together
with the Sequential Banach-Alaoglu Theorem (as recalled in [133, (3.6.22)], whose
356 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

applicability in the present context is ensured by (6.2.16)), imply the existence of

a sub-sequence {g jk }k ∈N of {g j } j ∈N which is weak-∗ convergent to some function
h ∈ B q,λ (Σ, σ), i.e.,
∫ ∫
lim f g jk dσ = f h dσ for each f ∈ M̊ p,λ (Σ, σ). (6.2.222)
k→∞ Σ Σ

In turn, from (6.2.221), (6.2.222), and (6.2.20), we see that

∫ ∫
f g dσ = f h dσ for each f ∈ Lipc (Σ). (6.2.223)
Σ Σ

With this in hand, [133, Corollary 3.7.3] applies and gives that g = h at σ-a.e.
point on Σ. In particular, g ∈ B q,λ (Σ, σ), as claimed in (6.2.211). Granted this
membership, (6.2.221) self-improves to
∫ ∫
lim f g j dσ = f g dσ for each f ∈ M̊ p,λ (Σ, σ). (6.2.224)
j→∞ Σ Σ

Indeed, given any f ∈ M̊ p,λ (Σ, σ), with φ ∈ Lipc (Σ) arbitrary we may write
∫ ∫ ∫
f g j dσ = ( f − φ)g j dσ + φg j dσ for each j ∈ N. (6.2.225)
Σ Σ Σ

Then (6.2.221) implies that

∫ ∫ ∫ ∫
lim φg j dσ = φg dσ = f g dσ − ( f − φ)g dσ, (6.2.226)
j→∞ Σ Σ Σ Σ

and (6.2.78) permits us to estimate

∫ 
 
 ( f − φ)g dσ  ≤ C f − φ M p, λ (Σ,σ) g B q, λ (Σ,σ), (6.2.227)
Σ

as well as
∫   
 
 ( f − φ)g j dσ  ≤ C f − φ M p, λ (Σ,σ) sup g j  B q, λ (Σ,σ) . (6.2.228)
Σ j ∈N

The fact that φ may be selected in Lipc (Σ) so that  f − φ M p, λ (Σ,σ) becomes arbitrary
small (cf. (6.2.20)) then ultimately leads to the conclusion that (6.2.224) holds. This
finishes the proof of (6.2.211). 
We continue by presenting a companion result to Proposition 6.2.14, detailing on
the mapping properties of the fractional integral operator acting on the pre-duals of
Morrey spaces.
Proposition 6.2.20 Suppose Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a closed Ahlfors
regular set and abbreviate σ := H n−1 Σ. Select q ∈ (1, ∞) along with λ ∈ (0, n − 1)
and assume
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 357

n−1−λ 1 α  −1
0<α< , q∗ := − . (6.2.229)
q q n−1−λ
Then  
σ(x)
B q,λ (Σ, σ) → L 1 Σ, (6.2.230)
1 + |x| n−1−α
and the fractional integral operator IΣ,α of order α on Σ, originally considered as
in (6.2.139), induces a well-defined, linear and bounded mapping

IΣ,α : B q,λ (Σ, σ) −→ B q∗,λ (Σ, σ). (6.2.231)

q(n−1)
Proof If r := n−1+λ(q−1) then
 σ(x) 
B q,λ (Σ, σ) → L r (Σ, σ) → L 1 Σ, . (6.2.232)
1 + |x| n−1−α

Indeed, the first embedding comes from (6.2.71), while the second embedding is a
consequence of [133, Lemma 7.2.1], Hölder’s inequality, and the observation that
if r  is the conjugate exponent of r then r (n − 1 − α) > n − 1 (which ultimately
is implied by the first condition in (6.2.229)). This proves (6.2.230). As regards the
claims pertaining to the fractional integral operator in (6.2.231), define p := (q∗ ),
the Hölder conjugate exponent of q∗ , and observe that
−1 −1 −1 −1
1 α
p − n−1−λ = 1
(q∗ )
α
− n−1−λ α
= 1− q1∗ − n−1−λ = 1− q1 = q , (6.2.233)

the Hölder conjugate exponent of q. The next step is to fix an arbitrary function

g ∈ B q,λ (Σ, σ). Then (6.2.139) and [133, Proposition 4.1.4] ensure that

IΣ,α g ∈ Lloc
1
(Σ, σ) ⊆ Lipc (Σ) (6.2.234)

and for each φ ∈ Lipc (Σ) we may write

∫  ∫
   
 IΣ,α g, φ  =  (IΣ,α g)φ dσ  =  g(IΣ,α φ) dσ 
Σ Σ

≤ Cg B q, λ (Σ,σ) IΣ,α φ M q, λ (Σ,σ)

≤ Cg B q, λ (Σ,σ) φ M p, λ (Σ,σ) . (6.2.235)

Above, the first equality is a consequence of [133, Proposition 4.1.4], the second
equality is implied by (6.2.230) and [133, (7.8.6)], the first inequality comes from
(6.2.78), and the final inequality follows from (6.2.141), bearing (6.2.233) in mind.
Having established (6.2.235), the very last claim in Proposition 6.2.16 (cf. (6.2.156)-

(6.2.157)) then guarantees the existence of a function h ∈ B p ,λ (Σ, σ) satisfying

h B q∗, λ (Σ,σ) = h B p, λ (Σ,σ) ≤ Cg B q, λ (Σ,σ), (6.2.236)

358 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

(where the first equality uses the fact that p = q∗ since p = (q∗ ) to begin with), and
with the property that
∫
IΣ,α g, φ = h φ dσ for every φ ∈ Lipc (Σ). (6.2.237)
Σ

In turn, from (6.2.237), (6.2.234), and [133, Proposition 4.1.4] we deduce that
∫ ∫
(IΣ,α g)φ dσ = h φ dσ for every φ ∈ Lipc (Σ) (6.2.238)
Σ Σ

which, in light of [133, Corollary 3.7.3], ultimately forces IΣ,α g = h at σ-a.e. point
on Σ. From this and (6.2.236) we then conclude that

IΣ,α g ∈ B q∗,λ (Σ, σ) and IΣ,α g B q∗, λ (Σ,σ) ≤ Cg B q, λ (Σ,σ) . (6.2.239)

With this in hand, all desired conclusions follow. 

Proposition 6.2.16 also permits us to establish useful embeddings of weighted

Lebesgue spaces into the pre-duals of Morrey spaces, of the sort discussed below.

Proposition 6.2.21 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Also, fix q ∈ (1, ∞) along with λ ∈ (0, n − 1) and
a > λ. Then one has the continuous and dense embedding
 
L q Σ, (1 + |x|)a(q−1) σ(x) → B q,λ (Σ, σ). (6.2.240)

Proof Pick an arbitrary function g ∈ L q Σ, (1 + |x|)a(q−1) σ(x) → Lloc

1 (Σ, σ) and

consider the distribution Λg ∈ Lipc (Σ) given by
∫
 
Λg, φ = gφ dσ for every φ ∈ Lipc (Σ). (6.2.241)
Σ

Note that if p ∈ (1, ∞) is such that 1/p + 1/q = 1, then for each φ ∈ Lipc (Σ) we may
use Hölder’s inequality and (6.2.25) to estimate
∫
  
 Λg, φ  ≤ g(x)|(1 + |x|)a(1−1/q) |φ(x)| dσ(x)
Σ (1 + |x|)a/p

≤ Cg L q (Σ,(1+ |x |) a(q−1) σ(x)) φ L p σ(x)

Σ, 1+| x|a

≤ Cg L q (Σ,(1+ |x |) a(q−1) σ(x)) φ M p, λ (Σ,σ), (6.2.242)

for some constant C ∈ (0, ∞) independent of g and φ. With this in hand, the last claim
in Proposition 6.2.16 guarantees that there exists a (unique) function h ∈ B q,λ (Σ, σ)
such that ∫
 
Λg, φ = h φ dσ for every φ ∈ Lipc (Σ) (6.2.243)
Σ
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 359

and
h B q, λ (Σ,σ) ≤ Cg L q (Σ,(1+ |x |) a(q−1) σ(x)) . (6.2.244)
There remains to observe that, collectively, (6.2.241), (6.2.243), and [133, (3.7.23)]
force g = h at σ-a.e. point in in Σ. With this in hand, the fact that the embedding in
(6.2.240) is well-defined and continuous follows from (6.2.244). The fact that said
embedding has also dense range is then seen from (6.2.70). 
For example, (6.2.240) implies that, in the context of Proposition 6.2.21,

if N > λ(q−1)+n−1
q and f N (x) := (1 + |x|)−N for all x ∈ Σ,
(6.2.245)
then the function f N belongs to the space B q,λ (Σ, σ).
Returning to Proposition 6.2.17, we shall use it to prove the following useful
embedding result of Morrey spaces into Muckenhoupt weighted Lebesgue spaces.

Proposition 6.2.22 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors

regular set, and abbreviate σ := H n−1 Σ. Then for each p0 ∈ (1, ∞) one has
$ $
M p,λ (Σ, σ) ⊆ L p0 (Σ, w). (6.2.246)
1<p<∞ w ∈ A p0 (Σ,σ)
0<λ<n−1

In fact, this inclusion is quantitative, in the following sense: having fixed

p0 ∈ (1, ∞), for each p ∈ (1, ∞) and λ ∈ (0, n − 1) there exists a constant
C = C(Σ, n, p0, p, λ) ∈ (0, ∞) with the property that for each function f ∈ M p,λ (Σ, σ)
one can find a weight w = w f ∈ Ap0 (Σ, σ) with [w] A p0 ≤ C and such that
f ∈ L p0 (Σ, w) with  f  L p0 (Σ,w) ≤ C f  M p, λ (Σ,σ) .

Proof This is a consequence of Proposition 6.2.17 together with the abstract em-
bedding results proved in Corollary 5.2.2 (whose applicability in the present setting
is ensuresd by Corollary 6.2.11 and Corollary 6.2.13). 
In the last portion of this section we introduce and briefly discuss a new function
space which is relevant for our work. To set the stage, we first make the following
definition.

Definition 6.2.23 Assume Σ ⊆ Rn (where n ∈ N with n ≥ 2) is a closed Ahlfors

regular set and abbreviate σ := H n−1 Σ. Also, select an integrability exponent
q ∈ (1, ∞), a parameter λ ∈ (0, n − 1), along with an arbitrary point xo ∈ Σ and an
arbitrary radius R ∈ 0, 2 diam Σ . In this context, call a function m ∈ L q (Σ, σ) a
(q, λ)-dull molecule concentrated near Δ(xo, R) := Σ ∩ B(xo, R) provided

m L q (Σ,σ) ≤ Rλ q −1
1
(6.2.247)

and
1
1+(n−1−λ) 1− q
R
|m(x)| ≤ for σ-a.e. x ∈ Σ \ B(xo, R). (6.2.248)
|x − xo | n−1
360 6 Morrey-Campanato Spaces, Morrey Spaces, and Their Pre-Duals . . .

In the same context as in Definition 6.2.23 we then introduce the (q, λ)-midway
space as


M q,λ (Σ, σ) := f ∈ Lipc (Σ) : there exist {λ j } j ∈N ∈  1 (N) and a family
{m j } j ∈N of (q, λ)-dull molecules on Σ
∞ 

such that f = λ j m j in Lipc (Σ) ,
j=1
(6.2.249)

and for each f ∈ M q,λ (Σ, σ) define


∞ 
∞

 f M q, λ (Σ,σ) := inf |λ j | : f = λ j m j in Lipc (Σ) with (6.2.250)
j=1 j=1

{λ j } j ∈N ∈  1 (N) and each m j a (q, λ)-dull molecule on Σ .

In particular, it is clear from (6.2.250) that

each (q, λ)-dull molecule m on Σ belongs to

(6.2.251)
M q,λ (Σ, σ) and mM q, λ (Σ,σ) ≤ 1,

and that

∞ 
if f ∈ M q,λ (Σ, σ) is expanded as f = λ j m j in Lipc (Σ) for some
j=1
sequence {λ j } j ∈N ∈  1 (N) and with each m j a (q, λ)-dull molecule on (6.2.252)

∞
Σ, then λ j m j also converges to f in the space M q,λ (Σ, σ).
j=1

The reason we have called M q,λ (Σ, σ) the “midway” space is that this is some-
where in between H q,λ (Σ, σ) (which is the pre-dual of a Morrey-Campanato space)
and the block space B q,λ (Σ, σ). Specifically, we have the following result.

Lemma 6.2.24 Let Σ ⊆ Rn (where n ∈ N with n ≥ 2) be a closed Ahlfors regular

set and abbreviate σ := H n−1 Σ. Fix an integrability exponent q ∈ (1, ∞) and a
parameter λ ∈ (0, n − 1). Then one has the continuous inclusions

H q,λ
(Σ, σ) → M q,λ (Σ, σ) → B q,λ (Σ, σ). (6.2.253)

In particular,
6.2 Morrey Spaces and Their Pre-Duals on Ahlfors Regular Sets 361
 
M q,λ (Σ, σ),  · M q, λ (Σ,σ) is a separable Banach space,

and M q,λ (Σ, σ) → L r (Σ, σ) with r := q(n−1)

∈ (1, q) (6.2.254)
n−1+λ(q−1)
is a well-defined continuous embedding with dense range.

Proof The first inclusion in (6.2.253) is a consequence of (6.1.15)-(6.1.17) and the

observation that any H q,λ -atom on Σ is also a (q, λ)-dull molecule on Σ (concen-
trated near the same surface ball in which said atom is supported). To justify the
second inclusion in (6.2.253), suppose m ∈ L q (Σ, σ) is a (q, λ)-dull molecule on Σ
concentrated near Δ(xo, R) := Σ ∩ B(xo, R), for some xo ∈ Σ and R ∈ 0, 2 diam Σ .
Then for σ-a.e. point x ∈ Σ \ B(xo, R) we may invoke (6.2.248) to estimate
1 1
1+(n−1−λ) 1− q −θ−(n−1−λ) q −1
R R
|m(x)| ≤ ≤ . (6.2.255)
|x − xo | n−1 |x − xo | n−1−θ
provided θ ≥ −1. Since it is possible to choose a number θ such that
 1
−1 ≤ θ < (n − 1 − λ) 1 − , (6.2.256)
q
Lemma 6.2.10 becomes applicable (keeping in mind (6.2.247) as well) and guaran-
tees the existence of a constant C = C(Σ, n, q, λ, θ) ∈ (0, ∞), independent of m, such
that
m ∈ B q,λ (Σ, σ) and m B q, λ (Σ,σ) ≤ C. (6.2.257)
Together with (6.2.249)-(6.2.250) and (6.2.71) this ultimately proves that the second
inclusion in (6.2.253) is well defined and continuous. Finally, (6.2.254) may be
justified with the help of (6.2.71). 
Chapter 7
Besov and Triebel-Lizorkin Spaces on Ahlfors
Regular Sets

We refer the reader to [190], [188], [187] for a thorough exposition regarding the
history and the nature of Besov and Triebel-Lizorkin spaces in the Euclidean setting.
Here we are concerned with adaptations of these scales of spaces to more general
ambients, which only enjoy but a small fraction of the structural richness of the
Euclidean space. This is in line with efforts made in the direction of extending the
standard theory of Besov and Triebel-Lizorkin spaces to the geometric measure
theoretic context of spaces of homogeneous type; see, e.g., [90], [86], [91], [92],
[203], [89], [152], and [206].

7.1 Definitions with Sharp Ranges of Indices and Basic Results

In this section we introduce the scales of homogeneous and inhomogeneous Besov

and Triebel-Lizorkin spaces in the context of Ahlfors regular sets, and review some
of their properties. We begin by adopting a number of basic definitions from [89],
[92].

Definition 7.1.1 Assume that Σ ⊆ Rn is a closed Ahlfors

 regular set and abbreviate
σ := H n−1 Σ. Fix two numbers γ ∈ (0, ∞) and β ∈ 0, 1]. A function f : Σ → R
is said to be a test function of type (x0, r, β, γ) with x0 ∈ Σ and r ∈ (0, ∞)
provided the following two conditions are satisfied for some constant C ∈ (0, ∞):
rγ
| f (x)| ≤ C   n−1+γ , ∀x ∈ Σ, (7.1.1)
r + |x − x0 |

and, for every x, y ∈ Σ,

r γ |x − y| β r + |x − x0 |
| f (x) − f (y)| ≤ C   n−1+γ+β if |x − y| < . (7.1.2)
r + |x − x0 | 4

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 363
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_7
364 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

In what follows, the collection of all test functions of type (x0, r, β, γ) on Σ will be
denoted by GΣ (x0, r, β, γ) and for each f ∈ GΣ (x0, r, β, γ) one defines

 f  GΣ (x0,r,β,γ) := inf {C > 0 : (7.1.1)-(7.1.2) hold}. (7.1.3)

As noted in [92], GΣ (x0, r, β, γ) is a Banach space, and a different choice of the

base point x0 and the scale r > 0 yields the same topological vector space, with an
equivalent norm. This justifies dropping the dependence on x0 and r in the definition
of the space of test functions of a certain type. Concretely, for a fixed x0 ∈ Σ, we
abbreviate
G β,γ (Σ) := GΣ (x0, 1, β, γ). (7.1.4)
To circumvent the inconvenience created by the fact that G β1,γ (Σ) is not densely
embedded into G β2,γ (Σ) if β1 > β2 , introduce
β,γ
G0 (Σ) := the closure of G 1,1 (Σ) in G β,γ (Σ) whenever 0 < β, γ < 1. (7.1.5)

In particular,
β,γ
Lipc (Σ) ⊆ G 1,1 (Σ) ⊆ G0 (Σ) whenever 0 < β, γ < 1. (7.1.6)

In analogy with the classical setting in Rn , in the environment provided by Σ as

the ambient space one may informally think of Lipc (Σ) as being test functions on
β,γ
Σ, and G0 (Σ) as being Schwartz functions on Σ. Let also note that any functional
 β,γ  ∗
Λ ∈ G0 (Σ) induces, by considering its restriction to Lipc (Σ), a distribution in
 
Lipc (Σ) (cf. (A.0.58)). We also wish to remark that

given x0 ∈ Σ, 0 < βo < β < 1, and 0 < γo < γ < 1, it follows

any f ∈ GΣ (x0, 1, β, γ) may be approximated with functions (7.1.7)
from the space Lipc (Σ) in the norm  ·  GΣ (x0,1,βo,γo ) .
∞ n
To justify this claim, bring in a function θ ∈ Cc (R ) satisfying θ ≡ 1 near the
origin, and set θ R (x) := θ (x − x0 )/R for each x ∈ Σ and each R > 0. One
may then check without difficulty that fR := f θ R ∈ Lipc (Σ) for each R > 0 and
 f − fR  GΣ (x0,1,βo,γo ) → 0 as R → ∞. This justifies (7.1.7). Finally, observe that
⎧
⎪ β,γ β
⎨ G (Σ) = Lip (Σ) and G (Σ) = C (Σ);
1,1
⎪
Σ compact =⇒ ·
(7.1.8)
⎪
⎪ hence G β,γ (Σ) = Lip (Σ) C β (Σ) for 0 < β, γ < 1.
⎩ 0

In the same setting as above, the space of test functions with mean zero is defined
as  ∫
G̊ β,γ (Σ) := f ∈ GΣ (x0, r, β, γ) : f dσ = 0 (7.1.9)
Σ
which is considered equipped with the norm inherited from GΣ (x0, r, β, γ) (cf.
(7.1.3)). Also, much as before, introduce
7.1 Definitions with Sharp Ranges of Indices and Basic Results 365

β,γ
G̊0 (Σ) := the closure of G̊ 1,1 (Σ) in G̊ β,γ (Σ) whenever 0 < β, γ < 1. (7.1.10)

We now proceed to introduce the scale of homogeneous Besov and Triebel-Lizorkin

spaces on Ahlfors regular sets (compare with [89, Definition 5.8, p.120]). The reader
is reminded that (a)+ := max{a, 0} for each a ∈ R.

Definition 7.1.2 Suppose Σ ⊆ Rn is an unbounded closed Ahlfors regular set and

abbreviate σ := H n−1 Σ. Since (Σ, | · − · |, σ) is a space of homogeneous type, [133,
Proposition 7.5.4] ensures the existence of a dyadic grid on Σ,

D(Σ) = Q αk k ∈Z,α∈Ik
. (7.1.11)

Denote by {St }t>0 the family of operators with integral kernels {St (·, ·)}t>0 as in
(4.2.14) and define the conditional expectation operators {Ek }k ∈Z by setting

Ek := S2−k − S2−k+1 for each k ∈ Z. (7.1.12)

Then, if

s ∈ (−1, 1), max n , n+s

n−1 n−1
< p ≤ ∞, 0 < q ≤ ∞,
 
max (s)+, −s + (n − 1) 1
p −1 < β < 1, (7.1.13)
+
  
max s − p , (n
n−1
− 1) 1
p − 1 , −s + (n − 1) 1
p −1 < γ < 1,
+
. p,q
the homogeneous Besov  ∗ Bs (Σ, σ) on the set Σ is defined as the collection
space
 β,γ
of all functionals f ∈ G̊0 (Σ) for which
  1/q
 q
 f B. p, q (Σ,σ) := 2 Ek f  L p (Σ,σ)
ks
< ∞, (7.1.14)
s
k ∈Z

with natural alterations when p = ∞ or q = ∞. Also, if

s ∈ (−1, 1), max n , n+s

n−1 n−1
< p ≤ ∞, max n , n+s
n−1 n−1
< q ≤ ∞,
 
max (s)+, −s + (n − 1) 1
p −1 < β < 1, (7.1.15)
+
  
max s − p , (n
n−1
− 1) 1
p − 1 , −s + (n − 1) 1
p −1 < γ < 1,
+
. p,q
then the homogeneous Triebel-Lizorkin  ∗ Fs (Σ, σ) on the set Σ is defined
 β,γspace
as the collection of all functionals f ∈ G̊0 (Σ) with the property that
   
  ks  q 1/q 
 f F. p, q (Σ,σ) :=  2 |Ek f | 
 <∞ (7.1.16)
s
k ∈Z L p (Σ,σ)
366 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

whenever p < ∞ (with natural alterations when q = ∞) and, corresponding to the

case when p = ∞,
⨏ 
∞  1/q
 q
 f F. ∞, q (Σ,σ) := sup sup 2 |Ek f |
ks
dσ < ∞, (7.1.17)
s
 ∈Z τ ∈I Qτ k=

again, with natural alterations when q = ∞.

Remark 7.1.3 In the context of Definition 7.1.2, the following are true.
. p,q . p,q
(1) The definitions of Bs (Σ, σ) and Fs (Σ, σ) are independent of the approxima-
. p,q 5.6, p. 115]. p,q
tion of identity used (see [89, Proposition and [89, Proposition 6.5,
p. 180]). Moreover, the definitions of Bs (Σ, σ) and Fs (Σ, σ) are independent
of the indices β, γ (see [89, Proposition 5.7, p. 116] and [89, Proposition 6.6,
p. 180]).
(2) Under the assumptions made in Definition 7.1.2 we have
. p, p . p, p
Bs (Σ, σ) = Fs (Σ, σ) (7.1.18)

(see [89, Proposition 5.10 (ii), p. 120] when p < ∞ and [89, Proposition 6.9
(ii), p. 182] for p = ∞).
. p,q 
(3) Modulo constants, the homogeneous Besov space Bs (Σ, σ) ∼ along with
. p,q 
the homogeneous Triebel-Lizorkin space Fs (Σ, σ) ∼ are quasi-Banach for
the range of indices specified in Definition 7.1.2, and become genuine Banach
spaces when 1 ≤ p, q ≤ ∞. See [89, Proposition 5.10(vi), p. 121] and [89,
Proposition 6.9(v), p. 182].

The following proposition describes how many important spaces we have dealt
with in this work relate to the homogeneous Besov and Triebel-Lizorkin spaces just
introduced.

Proposition 7.1.4 Suppose Σ ⊆ Rn is an unbounded closed Ahlfors regular set and

abbreviate σ := H n−1 Σ. Then
. p,2
F0 (Σ, σ) = L p (Σ, σ) whenever p ∈ (1, ∞), (7.1.19)
. p,2  
F0 (Σ, σ) = H p (Σ, σ) whenever p ∈ n−1n ,1 , (7.1.20)
. ∞,2
F0 (Σ, σ) = BMO(Σ, σ), (7.1.21)
. .
Fs∞,∞ (Σ, σ) = C s (Σ) whenever s ∈ (0, 1). (7.1.22)

Proof The identification in (7.1.19) is contained in [89, Proposition 5.10(v), p. 140],

(7.1.20) follows from [89, Definition 5.14 and Theorem 5.16, p. 124] (see also [89,
Remark 5.17, p. 124] and [88, Remark 2.30, p. 1527] in this regard), while (7.1.22)
and (7.1.21) may be deduced from [89, Theorem 6.11, p. 184]. 
7.1 Definitions with Sharp Ranges of Indices and Basic Results 367

We shall next give the definition of the inhomogeneous Besov and Triebel-
Lizorkin spaces on Ahlfors regular sets. This requires some preparation and we
begin by fixing a closed Ahlfors regular set Σ ⊆ Rn . Let κΣ ∈ Z ∪ {−∞} be such that

2−κΣ −1 < diam(Σ) ≤ 2−κΣ , (7.1.23)

and define 
κΣ if Σ is bounded,

κΣ := (7.1.24)
0 if Σ is unbounded.
Next, abbreviate σ := H n−1 Σ, and since (Σ, | · − · |, σ) is a space of homogeneous
type [133, Proposition 7.5.4] ensures the existence of a dyadic grid on Σ,

D(Σ) = Q αk k ∈Z, k ≥κΣ . (7.1.25)
α∈Ik

In this vein, recall the constant a1 appearing in [133, (7.5.9)] and pick some

jΣ ∈ N large enough so that 2−jΣ a1 < 13 . (7.1.26)

Then, having fixed such a background parameter jΣ ,

for each k ∈ Z with k ≥ κΣ and τ ∈ Ik we agree to organize the family
k+j k+j
Qτ Σ : Qτ Σ ⊂ Qτk as the collection Qτk,ν ν ∈ {1,..., N (k,τ)} . (7.1.27)

Finally, denote by yτk,ν the center of the cube Qτk,ν and, for any dyadic cube Qτk,ν and
1 (Σ, σ), define
any function f ∈ Lloc
∫
1
mQ k,ν ( f ) := f dσ. (7.1.28)
τ
σ(Qτk,ν ) Qτk,ν
We are now ready to present the definition of the inhomogeneous Besov and Triebel-
Lizorkin spaces on arbitrary closed subsets of Rn which are Ahlfors regular.

Definition 7.1.5 Given a closed Ahlfors regular set Σ ⊆ Rn , let σ := H n−1 Σ and
consider the collection of dyadic cubes

Qτk,ν k ∈Z, k ≥κΣ,τ ∈Ik , (7.1.29)
ν ∈ {1,..., N (κ,τ)}

produced according to the procedure described in (7.1.27). Also, bring in the family
of operators {St }0<t<diam(Σ) with integral kernels {St (·, ·)}0<t<diam(Σ) as in (4.2.14),
and define the conditional expectation operators {Ek }k ∈Z,k ≥κΣ by setting (with  κΣ as
in (7.1.24))

EκΣ := S2−κΣ and Ek := S2−k − S2−k+1 for k ∈ Z with k ≥ 

κΣ + 1. (7.1.30)

Then, if
368 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

s ∈ (−1, 1), max n , n+s

n−1 n−1
< p ≤ ∞, 0 < q ≤ ∞,
   (7.1.31)
max (s)+, −s + (n − 1) 1
p −1 < β < 1, (n − 1) p1 − 1 < γ < 1,
+ +
p,q
the inhomogeneous Besov  ∗ Bs (Σ, σ) on the set Σ is defined as the collection
 β,γ space
of all functionals f ∈ G0 (Σ) for which
   
  p 1/p
N
(
κΣ,τ)

 f Bsp, q (Σ,σ) := σ(QτκΣ,ν ) mQ κΣ ,ν |EκΣ f |
τ
τ ∈IκΣ ν=1

  1/q
  q
+ 2 Ek f  L p (Σ,σ)
ks
< ∞, (7.1.32)
k ∈Z
κΣ +1
k ≥

with natural alterations when p = ∞, or q = ∞.

In a similar fashion, whenever the parameters s, p, q, β, γ satisfy

s ∈ (−1, 1), max n−1 n , n+s < p ≤ ∞,

n−1
max n−1n , n+s < q ≤ ∞,
n−1

   (7.1.33)
max (s)+, −s + (n − 1) p1 − 1 < β < 1, (n − 1) p1 − 1 < γ < 1,
+ +
p,q
 β,γ  ∗ space Fs (Σ, σ) on the set
one defines the inhomogeneous Triebel-Lizorkin
Σ as the collection of all functionals f ∈ G0 (Σ) with the property that
   
  p 1/p
N
(
κΣ,τ)

 f Fsp, q (Σ,σ) := σ(QτκΣ,ν ) mQ κΣ ,ν |EκΣ f |
τ
τ ∈IκΣ ν=1
  
  ks q 1/q 
+  2 |Ek f | 
 <∞ (7.1.34)
k ∈Z L p (Σ,σ)
κΣ +1
k ≥

whenever p < ∞ (with natural alterations when q = ∞) and, corresponding to the

case when p = ∞,

 f Fs∞, q (Σ,σ) := max sup mQτκΣ ,ν (|EκΣ f |) : τ ∈ IκΣ , 1 ≤ ν ≤ N(
κΣ, τ) ,

⨏ 
∞  1/q 
 q
sup sup 2ks |Ek f | dσ < ∞, (7.1.35)
 ∈Z τ ∈I Qτ k=
κΣ +1
 ≥

again, with natural alterations when q = ∞.

Remark 7.1.6 In the context of Definition 7.1.5, the following properties hold.
7.1 Definitions with Sharp Ranges of Indices and Basic Results 369
p,q p,q
(1) The definitions of Bs (Σ, σ) and Fs (Σ, σ) are independent of the approxima-
tion of identity used (see [89, Proposition 5.27, p. 136] and [89, Proposition 6.17,
p,q p,q
p. 193]). Moreover, the definitions of Bs (Σ, σ) and Fs (Σ, σ) are independent
of the indices β, γ (see [89, Proposition 5.28, p. 137] and [89, Proposition 6.18,
p. 193]). Also, according to [89, Proposition 5.31(iv), p. 140] and [89, Proposi-
tion 6.21(iv), p. 195],
p,q  β,γ  ∗
Bs (Σ, σ) → G0 (Σ) continuously when the parameter-
p,q  β,γ  ∗
s s, p, q, β, γ are as in (7.1.31), and Fs (Σ, σ) → G0 (Σ) (7.1.36)
continuously when the parameters s, p, q, β, γ are as in (7.1.33).
Finally, with s, p, q as in Definition 7.1.5,

G β,γ (Σ) → Bs (Σ, σ) and G β,γ (Σ) → Fs (Σ, σ)

p,q p,q

  (7.1.37)
provided (s)+ < β < 1 and (n − 1) p1 − 1 + < γ < 1.

See [89, Proposition 5.31(v), p. 140].

(2) Under the assumptions made in Definition 7.1.5 we have
p, p p, p
Bs (Σ, σ) = Fs (Σ, σ) (7.1.38)

(see [89, Proposition 5.31(iii), p. 140] when p < ∞ and [89, Proposition 6.21
(iii), p. 195] for p = ∞).
p,q
(3) The inhomogeneous Besov space Bs (Σ, σ) along with the inhomogeneous
p,q
Triebel-Lizorkin space Fs (Σ, σ) are quasi-Banach for the range of indices
specified in Definition 7.1.5, and become genuine Banach spaces whenever
1 ≤ p, q ≤ ∞ (see [89, Proposition 5.31(vii), p. 140] for the scale of Besov spaces
and, for the scale of Triebel-Lizorkin spaces, see [89, Proposition 5.31(vii),
p. 140] when p < ∞ and [89, Proposition 6.21(iii), p. 195] for p = ∞).
p,q
n , n+s < p ≤ 1, and 0 < q ≤ p, the space Bs (Σ, σ) is
(4) If s ∈ (−1, 1), max n−1 n−1

a q-Banach space. Indeed, it may be checked without difficulty that

  q/p
   N
(
κΣ,τ)
 p
 f  p, q := σ(QτκΣ,ν ) mQ κΣ ,ν (|EκΣ f |)
B s (Σ,σ) τ
τ ∈IκΣ ν=1

   1/q
q
+ 2ks Ek f  L p (Σ,σ) (7.1.39)
k ∈Z
κΣ +1
k ≥

p,q p,q
for each f ∈ Bs (Σ, σ) is a quasi-norm on Bs (Σ, σ) which is equivalent to
 · Bsp, q (Σ,σ) given in (7.1.32), and with the property that
 q
Bs (Σ, σ)  f , g →  f − g B p, q (Σ, σ) ∈ [0, ∞)
p,q
(7.1.40)
s
370 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
p,q
is a metric (with respect to which the Besov space Bs (Σ, σ) is complete). In view
of this and (7.1.38) we conclude that the diagonal Besov and Triebel-Lizorkin
p, p p, p
spaces, Fs (Σ, σ) and Bs (Σ, σ), are p-Banach spaces whenever s ∈ (−1, 1)
and max n , n+s < p ≤ 1.
n−1 n−1

(5) In the case when Σ is unbounded, and if 0 < s < 1, 1 ≤ p ≤ ∞, 0 < q ≤ ∞,

then
p,q . p,q
Bs (Σ, σ) = L p (Σ, σ) ∩ Bs (Σ, σ) and
(7.1.41)
 f Bsp, q (Σ,σ) ≈  f  L p (Σ,σ) +  f B. p, q (Σ,σ) for f ∈ Bs (Σ, σ)
p,q
s

(see [89, Proposition 5.39(i), p. 149]), while if 0 < s < 1, 1 ≤ p < ∞, and
n < q ≤ ∞, then
n−1

p,q . p,q
Fs (Σ, σ) = L p (Σ, σ) ∩ Fs (Σ, σ) and
(7.1.42)
 f Fsp, q (Σ,σ) ≈  f  L p (Σ,σ) +  f F. p, q (Σ,σ) for f ∈ Fs (Σ, σ)
p,q
s

(see [89, Proposition 5.39(ii), p. 149]).

It is also significant to observe that the inhomogeneous Besov and Triebel-

Lizorkin spaces introduced in Definition 7.1.5 embed continuously in the space
of distributions. Specifically, we have the following result.

Lemma 7.1.7 Retain the context of Definition 7.1.5. Then the

 mapping
 ∗ which associ-
p,q β,γ
ated to each f ∈ Bs (Σ, σ), regarded as functional in f ∈ G0 (Σ) , its restriction
to Lipc (Σ) induces a well-defined continuous embedding
p,q  
Bs (Σ, σ) → Lipc (Σ) . (7.1.43)
p,q
Likewise, then the mapping
 β,γ which associated to each f ∈ Fs (Σ, σ), regarded
∗
as functional in f ∈ G0 (Σ) , its restriction to Lipc (Σ) induces a well-defined
continuous embedding  
p,q
Fs (Σ, σ) → Lipc (Σ) (7.1.44)
p,q
Proof Suppose f ∈ Bs (Σ, σ) is such that
!
β,γ
(G (Σ))∗
f , ψ G β,γ (Σ) = 0 for each ψ ∈ Lipc (Σ). (7.1.45)
0 0

If we choose βo ∈ (0, β) and γo ∈ (0, γ) so that the last two lines in (7.1.15) are still
 ∗ by βo, γo , item
valid with β, γ replaced (1) in Remark 7.1.6 guarantees that we also
β ,γ β,γ
have f ∈ G0 o o (Σ) . Pick φ ∈ G0 (Σ) arbitrary and recall from (7.1.7) that there
β ,γ
exists a sequence {ψ j } j ∈N ⊆ Lipc (Σ) which converges to φ in G0 o o (Σ). We may
then write
7.1 Definitions with Sharp Ranges of Indices and Basic Results 371
! !
β,γ
(G0 (Σ))∗
f, φ β,γ
G0 (Σ)
= (G β o ,γo (Σ))∗ f , φ β ,γ o
G0 o (Σ)
0
!
= lim β o ,γ o
(Σ))∗
f , ψj β ,γ o
G0 o
= 0, (7.1.46)
j→∞ (G0 (Σ)

 β,γ  ∗
thanks to (7.1.45). This proves that f is zero as a functional in G0 (Σ) . This
proves that the mapping described in the statement of the lemma is injective. Since
its continuity follows by unraveling definitions, we ultimately conclude that we have
the continuous embedding in (7.1.43). The claims pertaining to (7.1.44) are dealt
with in a similar manner. 
Once again, many important spaces dealt with in this work fit within the hierarchy
of inhomogeneous Besov and Triebel-Lizorkin scales. To elaborate on this topic, let
us first briefly discuss the scale of local Hardy spaces and the local version of the
space of functions of bounded mean oscillations. Throughout, assume that Σ ⊆ Rn
is a closed set which is Ahlfors regular and abbreviate σ := H n−1Σ. In addition, fix
a background threshold rΣ ∈ (0, diam Σ) and select p ∈ n−1 n , 1 . A σ-measurable
function a : Σ → C with the property that there exist a point x ∈ Σ and a number
r ∈ (0, 2 diam Σ) such that
  1/2−1/p
supp a ⊆ B(x, r) ∩ Σ, a L 2 (Σ,σ) ≤ σ B(x, r) ∩ Σ ,
∫ (7.1.47)
and Σ a dσ = 0

is called a p-atom. When the last property is not necessarily enforced but one now
insists that r ≥ rΣ , the function a is said to be a p-block. In particular,

the collection of all p-blocks is a bounded subset of L 2 (Σ, σ). (7.1.48)

 
Definition 7.1.8 Retain the above context, and suppose (n − 1) p1 − 1 <β, γ < 1.
 ∗ local Hardy space h (Σ, σ) as the collection of functionals f in
Define p
 β,γ the
G0 (Σ) with the property that there exist two sequences of numbers, {λ j } j ∈N
# #
and {μ j } j ∈N , satisfying j ∈N |λ j | p < +∞ and j ∈N | μ j | p < +∞, along with a
sequence {a j } j ∈N of p-atoms and a sequence {b j } j ∈N of p-blocks such that


∞ 
∞
 β,γ  ∗
f = λj aj + μ j b j in G0 (Σ) . (7.1.49)
j=1 j=1

Moreover, for each f ∈ h p (Σ, σ) define

   1/p   1/p
 f h p (Σ,σ) := inf |λ j | p + | μj |p (7.1.50)
j ∈N j ∈N

where the infimum is taken over all the decompositions of f as in (7.1.49).

In relation to the local Hardy space just introduced and the brand of Hardy space
introduced earlier in Definition 4.2.1 we wish to note that, as is apparent from
372 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Definition 7.1.8 and Theorem 4.4.1 we have

 n−1 
H p (Σ, σ) → h p (Σ, σ) for each p ∈ n ,1 , (7.1.51)

and, by also taking into account (7.1.48), and (4.4.114),

if Σ ⊆ Rn is a compact Ahlfors regular set and σ := H n−1  Σ, (7.1.52)

it follows that h p (Σ, σ) = H p (Σ, σ) for each p ∈ n−1
n , 1 .

In the same geometric setting as above, the space of functions of local bounded
mean oscillations bmo (Σ, σ) is defined to be the collection of all functions
1 (Σ, σ) with the property that
f ∈ Lloc
⨏  ⨏ 
 
 f bmo (Σ,σ) := sup  f (y) − f dσ  dσ(y)
x ∈Σ, 0<r <rΣ B(x,r)∩Σ B(x,r)∩Σ
⨏
+ sup | f | dσ < +∞. (7.1.53)
x ∈Σ, r ≥rΣ B(x,r)∩Σ

See [89, Definition 2.27, p. 51] for a more inclusive point of view. Here we only wish
to note that, in the sense described in [89, Theorem 5.44, p. 155],

bmo (Σ, σ) is the dual space of h1 (Σ, σ). (7.1.54)

After this preamble, here is the list of identifications mentioned a little earlier in
the narrative.
Proposition 7.1.9 Suppose Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then
p,2
F0 (Σ, σ) = L p (Σ, σ) whenever p ∈ (1, ∞), (7.1.55)
p,2  
F0 (Σ, σ) = h p (Σ, σ), whenever p ∈ n−1n ,1 , (7.1.56)

F0∞,2 (Σ, σ) = bmo(Σ, σ), (7.1.57)

Fs∞,∞ (Σ, σ) = C s (Σ) whenever s ∈ (0, 1). (7.1.58)

Hence, as a consequence of (7.1.58) and (7.1.38), one also has

Bs∞,∞ (Σ, σ) = C s (Σ) whenever s ∈ (0, 1), (7.1.59)

while (7.1.56) and (7.1.52) imply that

p,2  n−1 
F0 (Σ, σ) = H p (Σ, σ) if Σ is compact and p ∈ n ,1 . (7.1.60)

Proof The identification in (7.1.55) is contained in [87, Theorem 3, p. 578] and

[89, Proposition 5.31(vi), p. 140], (7.1.56) follows from Definition 7.1.8 in concert
with [89, Definition 5.40 and Theorem 5.42, p. 151], (7.1.57) may be deduced from
7.2 Atomic and Molecular Theory 373

[89, Theorem 6.28, p. 204], while (7.1.58) is immediate from [89, Corollary 6.24,
p. 200]. 
We conclude this section by discussing a density result. As a preamble, we make
a definition. Given a closed Ahlfors regular set Σ ⊆ Rn and any β, γ > 0, following
[89, (5.176)] we set
β,γ
Gb (Σ) := { f ∈ G β,γ (Σ) : f has bounded support}. (7.1.61)

Lemma 7.1.10 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix a smoothness index s ∈ (−1, 1). Then the space Gb1,1 (Σ) is
p,q p,q
dense in Bs (Σ, σ) if max n−1 n , n+s < p < ∞ and 0 < q < ∞, and in Fs (Σ, σ)
n−1

n , n+s < p, q < ∞. As a consequence, with Lipc (Σ) denoting the space
if max n−1 n−1

of Lipschitz functions on Σ with compact support, for the same ranges of indices as
above one has
p,q p,q
Lipc (Σ) → Bs (Σ, σ) densely, and Lipc (Σ) → Fs (Σ, σ) densely. (7.1.62)

Proof See [89, Proposition 5.46, p. 155] (cf. also [86, Proposition 3.3, p. 56] in the
case when 1 ≤ p, q < ∞). 

7.2 Atomic and Molecular Theory

The goal here is to discuss atomic and molecular characterization of the inhomoge-
neous Besov and Triebel-Lizorkin spaces introduced in Definition 7.1.5. To set the
stage, assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate σ := H n−1 Σ.
Also, recall the dyadic cubes defined in (7.1.27) and, with  κΣ as in (7.1.24) and
jΣ ∈ N as in (7.1.26), introduce

D∗ (Σ) := Qτk,ν : k ∈ Z, k ≥ 
κΣ, τ ∈ Ik , 1 ≤ ν ≤ N(k, τ) . (7.2.1)

Hence, in terms of the piece of notation introduced in (A.0.35),

$
D∗ (Σ) = Dk+jΣ (Σ). (7.2.2)
k ∈Z, k ≥
κΣ

The following definition of atoms and blocks agrees, up to a renormalization, with

the definition introduced in [86, Definition 2.1, p. 45] for spaces of homogeneous
type (see also [92, Definition 7, p. 74]).
Definition 7.2.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, pick s ∈ (−1, 1) along with p ∈ (0, ∞] and η > 0. Also, fix a
background constant C0 ∈ (0, ∞).
Given a cube Qτk,ν ∈ D∗ (Σ), call a function aQ k,ν : Σ → R an η-smooth atom of
τ
type (p, s) (for the dyadic cube Qτk,ν ) on Σ if the following four conditions hold:
374 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 
supp (aQ k,ν ) ⊆ B yτk,ν, C0 2−k ∩ Σ where yτk,ν is the center of Qτk,ν, (7.2.3)
τ
  n−1
sup aQ k,ν  ≤ (2−k )s− p , (7.2.4)
τ
Σ
 
.
n−1
a ≤ (2−k )s− p −η , (7.2.5)
Qτk,ν C η (Σ)
∫
aQ k,ν dσ = 0. (7.2.6)
τ
Σ

In the case when (7.2.3)-(7.2.5) hold but (7.2.6) is not necessarily satisfied, we say
that the function aQ k,ν is an η-smooth block of type (p, s) on Σ.
τ

Definition 7.2.2 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, pick s ∈ (−1, 1) along with p ∈ (0, ∞] and η, ε > 0. Given a
cube Qτk,ν ∈ D∗ (Σ), call a function uQ k,ν : Σ → R a (η, ε)-smooth molecule of
τ
type (p, s) (for the dyadic cube Qτk,ν ) on Σ if the following three conditions hold:
n−1
|uQ k,ν (x)| ≤ (2−k )s− p (1 + 2k |x − yτk,ν |)−(n−1+ε) for each x ∈ Σ, (7.2.7)
τ

n−1
|uQ k,ν (x) − uQ k,ν (y)| ≤ (2−k )s− p −η |x − y| η × (7.2.8)
τ τ

× (1 + 2k |x − yτk,ν |)−(n−1+ε) + (1 + 2k |y − yτk,ν |)−(n−1+ε) , ∀x, y ∈ Σ,
∫
uQ k,ν (x) dσ(x) = 0, (7.2.9)
τ
Σ

A function uQ k,ν : Σ → R is called a (η, ε)-smooth unit of type (p, s) (for the dyadic
τ
cube Qτk,ν ) on Σ if it satisfies (7.2.7) and (7.2.8).

A cursory inspection of definitions reveals that

given any ε > 0, it follows that any η-smooth atom of type (p, s) is a
fixed multiple of an (η, ε)-smooth molecule of type (p, s) (for the same
(7.2.10)
dyadic cube), and any η-smooth block of type (p, s) is a fixed multiple
of an (η, ε)-smooth unit of type (p, s) (again, for the same dyadic cube).
It is also clear from Definition 7.2.1 and Definition 7.2.2 that
for any η, ε > 0 fixed, the class of η-smooth atoms of type (p, s),
the class of η-smooth blocks of type (p, s), the class of (η, ε)-smooth
molecules of type (p, s), and the class of (η, ε)-smooth units of type (7.2.11)
(p, s) do not change if the quantity p1 − n−1
s
does not change.

There is also the useful notion of rough molecule, formally introduced next.

Definition 7.2.3 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Assume
7.2 Atomic and Molecular Theory 375
1 
n−1
n < p ≤ 1, (n − 1) p − 1 < s < 1. (7.2.12)

Also, pick qo ∈ [1, ∞] with qo > p and fix some ε ∈ (0, ∞). Finally, chose a point
xo ∈ Σ along with some finite number r ∈ (0, diam Σ].
In this setting, call a σ-measurable function m : Σ → C a (p, qo, ε, s)-rough
molecule centered near the ball B(xo, r) on Σ (or, simply, a rough molecule on Σ)
provided for each k ∈ N0 one has
∫  1/qo   1 − 1 + s−1
|m| qo dσ ≤ 2k(n−1)[1/qo −1−ε] σ B(xo, r) ∩ Σ qo p n−1 (7.2.13)
A k (x o ,r)

where
⎧
⎨ B(xo, r) ∩ Σ
⎪
⎪ if k = 0,
Ak (xo, r) :=   (7.2.14)
⎪
⎪ B(xo, 2k r) \ B(xo, 2k−1 r) ∩ Σ if k ≥ 1,
⎩
and
∫
m dσ = 0. (7.2.15)
Σ

Moreover, in the case when Σ is bounded it is also agreed that the constant function
given by m(x) := [σ(Σ)]−1/p for every x ∈ Σ is a (p, qo, ε)-rough molecule on Σ.
In the context of Definition 7.2.3, it follows from Definition 4.5.1 that
 
s−1 −1
if po := p1 − n−1 , then a function m : Σ → C is a (p, qo, ε, s)-rough
molecule in the sense of Definition 7.2.3 if and only if the function m (7.2.16)
is a (po, qo, ε)-molecule on Σ in the sense of Definition 4.5.1.
In particular, from this and (4.5.6) we conclude that

any given (p, qo, ε, s)-rough molecule m : Σ → C belongs to the

 
s−1 −1
Hardy space H po (Σ, σ) with po := p1 − n−1 , and the estimate (7.2.17)
m H p o (Σ,σ) ≤ C holds for a constant C ∈ (0, ∞) independent of m.

For further use, let us also observe here that, as may be seen from (7.2.3), (7.2.4),
(7.2.6), and (4.4.2),
if s, p are as in (7.2.12) and η > 0, then any η-smooth atom of type
(p, s − 1) on Σ (in the sense of Definition 7.2.1) is a fixed multiple of
some (po, qo )-atom on Σ (as described in (4.4.2)) provided one defines (7.2.18)
 
s−1 −1
 
po := p1 − n−1 ∈ n−1
n , 1 and takes qo ∈ [1, ∞] arbitrary.

It turns out that rough molecules belong to a suitable Besov space, have uniform
control of their quasi-norms in that setting, and linear combinations with  p coeffi-
cients of rough molecules stay in said Besov space. These properties are discussed
next.
376 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Theorem 7.2.4 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Assume
 
n < p ≤ 1, (n − 1) p1 − 1 < s < 1, 0 < q ≤ ∞.
n−1
(7.2.19)

Also, suppose qo ∈ [1, ∞] and ε ∈ (0, ∞).

Then any (p, qo, ε, s)-rough molecule m on Σ belongs to the Besov space
p,q
Bs−1 (Σ, σ) and there exists some constant C = C(Σ, n, p, q, s, qo, ε) ∈ (0, ∞) in-
dependent of m with the property that

mB p, q (Σ,σ) ≤ C. (7.2.20)

s−1

Moreover, in the case when 0 < q ≤ p there exists a constant C ∈ (0, ∞) with the
property that, for any family of (p, qo, ε, s)-rough molecules {m j } j ∈N on Σ and any
# p,q
numerical sequence {λ j } j ∈N ∈  q , the series j ∈N λ j m j converges in Bs−1 (Σ, σ)
and     1/q
 
 λ j m j  p, q ≤C |λ j | q . (7.2.21)
B s−1 (Σ,σ)
j ∈N j ∈N
1 
s−1 −1
Proof Define po := p − n−1 and observe that the current assumptions on
p, q, s, qo entail
 
n−1
n < po < p ≤ 1, qo > po, and − (n − 1) 1
po − 1
p = s − 1. (7.2.22)

Granted these, we conclude from (7.7.54) that

p,q
H po (Σ, σ) → Bs−1 (Σ, σ) continuously. (7.2.23)

The claims pertaining to (7.2.20) are then consequences of (7.2.16)-(7.2.17) and

(7.2.23). Finally, in the case when 0 < q ≤ p the claims in the last portion of the
statement follow (7.2.20) and item (4) in Remark 7.1.6. 
We next introduce discrete Besov and Triebel-Lizorkin spaces on Ahlfors regular
sets. Our definition is adjusted to the normalization of our atoms and yields results in
line with the situation when the underlying space is Rn (see [57], [59]). A different
normalization appears in [92, p. 74]. The choice we have made in the normalization
of atoms is designed so that the discrete Besov and Triebel-Lizorkin spaces have
definitions which are independent of the smoothness index (which we choose not to
include in the notation employed for these discrete spaces).
Definition 7.2.5 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall the family of cubes (7.2.1), fix p, q ∈ (0, ∞], and assume  κΣ
is as in (7.1.24). In this setting, denote by bp,q (Σ) the space of numerical sequences
λ = {λQ }Q ∈D∗ (Σ) such that
    N
(k,τ)  q/p  1/q
λb p, q (Σ) := |λQ k,ν | p < ∞, (7.2.24)
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ
7.2 Atomic and Molecular Theory 377

with natural modifications when p = ∞ or q = ∞.

Moreover, let f p,q (Σ) be the space of numerical sequences λ = {λQ }Q ∈D∗ (Σ) with
the property that
 
   N (k,τ) 
   q  1/q 
 
λ f p, q (Σ) :=  σ(Qτ ) λQ k,ν  1Q k,ν
k,ν − p1 
 < ∞,
 τ τ  p
k ∈Z τ ∈Ik ν=1 L (Σ,σ)
k ≥κΣ
(7.2.25)
when p < ∞ (with a natural adaptation when q = ∞). Finally, corresponding to the
case when p = ∞, and q ∈ (0, ∞], the space f ∞,q (Σ) is defined as the collection of
sequences λ = {λQ }Q ∈D∗ (Σ) having the property that the discrete Carleson measure
condition λ f ∞, q (Σ) < ∞, where

λ f ∞, q (Σ) := max sup |λτκΣ,ν | : τ ∈ IκΣ , 1 ≤ ν ≤ N(
κΣ, τ) ,

  q1 
1  
∞ N (k,τ)
sup sup 
σ(Qτ )|λτ | 1 {(τ,ν): Q k,ν ⊂Q  } (τ, ν)
k,ν k,ν q
 ∈Z α∈I σ(Q α ) k= τ ∈I ν=1 τ α
 ≥
κΣ +1 k

(7.2.26)

with D(Σ) = {Qα :  ∈ Z,  ≥ κΣ, α ∈ I } denoting the dyadic grid associated with
Σ as in [133, Proposition 7.5.4].

Based on the fact that  p0 →  p1 whenever 0 < p0 ≤ p1 ≤ ∞ (applied twice) we

conclude that
bp0,q0 (Σ) → bp1,q1 (Σ) whenever
(7.2.27)
0 < p0 ≤ p1 ≤ ∞ and 0 < q0 ≤ q1 ≤ ∞.
Later on, we shall nonetheless also use the standard definition of discrete Besov and
Triebel-Lizorkin spaces, so we record this below (compare with [92, p. 74]).

Definition 7.2.6 Let Σ ⊆ Rn be a closed Ahlfors regular set and let σ := H n−1 Σ.
Recall the family of cubes (7.2.1) and fix parameters s ∈ R and p, q ∈ (0, ∞].
p,q
Also, assume  κΣ is as in (7.1.24). Then bs (Σ) denotes the space of sequences
λ = {λQ }Q ∈D∗ (Σ) with the property that
    N  1/q
 p  q/p
(k,τ)

2ks σ(Qτk,ν ) p − 2 |λQ k,ν |
1 1
λbsp, q (Σ) := < ∞, (7.2.28)
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

with natural modifications when p = ∞, or q = ∞.

p,q
Furthermore, denote by fs (Σ) the space of sequences λ = {λQ }Q ∈D∗ (Σ) for
which
378 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 
   N
(k,τ)
1   q  1/q 
 
λ fsp, q (Σ) :=  2ks σ(Qτk,ν )− 2 λQ k,ν  1Q k,ν  < ∞,
 τ τ 
k ∈Z τ ∈Ik ν=1 L p (Σ,σ)
k ≥κΣ
(7.2.29)
when p < ∞ (with a natural adaptation when q = ∞). Finally, corresponding to the
∞,q
case when p = ∞, and q ∈ (0, ∞], the space fs (Σ) is defined as the collection
of sequences λ = {λQ }Q ∈D∗ (Σ) satisfying the discrete Carleson measure condition
λ fs∞, q (Σ) < ∞, where

λ fs∞, q (Σ) := max sup |λτκΣ,ν | : τ ∈ IκΣ , 1 ≤ ν ≤ N(
κΣ, τ) ,
  q1 
1    ksq
∞ N (k,τ)
sup sup 
2 σ(Q k,ν
τ )|λτ |
k,ν q
1 k,ν 
{(τ,ν): Qτ ⊂Qα } (τ, ν)
 ∈Z α∈I σ(Q α ) k= τ ∈I ν=1
κΣ +1
 ≥ k

(7.2.30)

with D(Σ) = {Qα :  ∈ Z,  ≥ κΣ, α ∈ I } denoting the dyadic grid associated with
Σ as in [133, Proposition 7.5.4].

The theorem below describes the decomposition of distributions from the “con-
tinuous” scale of Besov spaces into series of atoms and blocks with coefficients
belonging to discrete Besov spaces.

Theorem 7.2.7 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall the family of dyadic cubes defined in (7.2.1) and recall the
parameter 
κΣ from (7.1.24).
p,q
(i) Given any f ∈ Bs (Σ, σ) with s, p, q, β, γ as in (7.1.31), there exist a numerical
sequence λ = {λQ k,ν }Q k,ν ∈D∗ (Σ) , an exponent η ∈ (|s|, 1], along with η-smooth
τ τ
blocks aQ κΣ ,ν of type (p, s) for τ ∈ IκΣ and ν ∈ {1, . . . , N(
κΣ, τ)}, and η-smooth
τ
atoms aQ k,ν of type (p, s) for k ∈ Z, k ≥  κΣ + 1, τ ∈ Ik , ν ∈ {1, . . . , N(k, τ)},
τ
such that
  N (k,τ)
f = λQ k,ν aQ k,ν (7.2.31)
τ τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ
p,q  β,γ  ∗
with convergence taking place both in Bs (Σ, σ) and in G0 (Σ) when
 β,γ  ∗
max{p, q} < ∞, and only in G0 (Σ) when max{p, q} = ∞. In addition,
matters may be arranged so that

λb p, q (Σ) ≤ C f Bsp, q (Σ,σ), (7.2.32)

for some finite constant C > 0, which depends on Σ, p,q, and s.

p,q
(ii) Given any f ∈ Fs (Σ, σ) with s, p, q, β, γ as in (7.1.33), there exist a numerical
sequence λ = λQ k,ν Q k,ν ∈D∗ (Σ) , an exponent η ∈ (|s|, 1], along with η-smooth
τ τ
7.2 Atomic and Molecular Theory 379

blocks aQ κΣ ,ν of type (p, s) for τ ∈ IκΣ and ν ∈ {1, . . . , N(

κΣ, τ)}, and η-smooth
τ
atoms aQ k,ν of type (p, s) for k ∈ Z, k ≥  κΣ + 1, τ ∈ Ik , ν ∈ {1, . . . , N(k, τ)},
τ
p,q
such
 that (7.2.31)
∗ holds with convergence taking place
∗ both in Fs (Σ, σ) and
β,γ β,γ
in G0 (Σ) when q < ∞, and only in G0 (Σ) when q = ∞. In addition,
matters may be arranged so that

λ f p, q (Σ) ≤ C f Fsp, q (Σ,σ) (7.2.33)

for some finite constant C = C(Σ, p, q, s) > 0.

Proof This follows from [92, Theorem 4, p. 75] by taking into account the renor-
malization we presently consider for our atoms. 
In the converse direction to Theorem 7.2.7, the extent to which linear combinations
of units and molecules with coefficients in a discrete Besov space belong to the
corresponding continuous Besov space is studied next.

Theorem 7.2.8 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, recall the family of dyadic cubes defined in (7.2.1) and recall

κΣ from (7.1.24).
(i) Suppose

s ∈ (−1, 1), max n , n+s

n−1 n−1
< p ≤ ∞, 0 < q ≤ ∞,

(s)+ < η < 1, (7.2.34)

  
max (n − 1) 1
p − 1 , −s + (n − 1) 1
p −1 < ε < 1.
+ +

Let uQ κΣ ,ν be a (η, ε)-smooth unit of type (p, s) for each τ ∈ IκΣ and each
τ
ν ∈ {1, . . . , N(
κΣ, τ) , and let uQ k,ν be a (η, ε)-smooth molecule of type (p, s)
τ
for each k ∈ Z, k ≥  κΣ + 1, τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}. Then for every
numerical sequence λ = λQ k,ν Q k,ν ∈D∗ (Σ) ∈ bp,q (Σ) it follows that the series
τ τ

  N
(k,τ)
f := λQ k,ν uQ k,ν (7.2.35)
τ τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

β,γ
converges in the space Bs (Σ, σ) whenever max{p, q} < ∞, and in (G0 (Σ))∗
p,q

whenever
 
max (s)+, −s + (n − 1) p1 − 1 < β < 1 and 0 < γ < 1. (7.2.36)
+

Furthermore, when max{p, q} < ∞ there exists C ∈ (0, ∞) such that

 f Bsp, q (Σ,σ) ≤ Cλb p, q (Σ) . (7.2.37)

380 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Moreover,
when s ∈ (0, 1), the same conclusions as above continue to hold in the
situation when each uQ k,ν is actually a (η, ε)-smooth unit of type (p, s)
τ
for every k ∈ Z, k ≥ κΣ + 1, τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}.
(7.2.38)
(ii) Assume
 
s ∈ (−1, 1), max n−1 ,
n n+s
n−1
< p < ∞, max n , n+s < q ≤ ∞,
n−1 n−1

(s)+ < η < 1,

   
max (n − 1) 1
min(p,q) − 1 , −s + (n − 1) 1
min(p,q) −1 < ε < 1.
+ +
(7.2.39)

Let uQ κΣ ,ν be a (η, ε)-smooth unit of type (p, s) for every τ ∈ IκΣ and every
τ
ν ∈ {1, . . . , N(
κΣ, τ)}, and let uQ k,ν be a (η, ε)-smooth molecule of type (p, s)
τ
for each k ∈ Z, k ≥  κΣ + 1, τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}. Then for each
numerical sequence λ = λQ k,ν Q k,ν ∈D∗ (Σ) ∈ f p,q (Σ) it follows that the series
p,q
τ τ
 β,γ  ∗
(7.2.35) converges in Fs (Σ, σ) whenever q < ∞, and in G0 (Σ) whenever
β, γ verify (7.2.36). Furthermore, when q < ∞ there exists C ∈ (0, ∞) with the
property that
 f Fsp, q (Σ,σ) ≤ Cλ f p, q (Σ) . (7.2.40)
Moreover, when s ∈ (0, 1), the same conclusions as above continue to hold in
the situation when each uQ k,ν is actually a (η, ε)-smooth unit of type (p, s) for
τ
every k ∈ Z, k ≥ 
κΣ + 1, τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}.

Proof This follows from [92, Theorem 5, p. 76] (cf. also [86, Theorem 2.2, p. 51] for
the case when p, q ≥ 1), after readjusting notation. The last claim in the statement
of the theorem is seen from an inspection of the proof of [92, Theorem 5, p. 76]. In
this regard, see also the second remark in [92, § 3, p. 95]. 

We may further refine the decomposition (7.2.31) by retaining the atoms as they
are while bundling together the blocks into a single function belonging to Lebesgue
spaces, as indicated in the corollary below.

Corollary 7.2.9 Assume that Σ ⊆ Rn is a closed Ahlfors regular set, and abbreviate
σ := H n−1 Σ. Recall the family of dyadic cubes defined in (7.2.1) and recall the
parameter 
κΣ from (7.1.24). Also, suppose

s ∈ (−1, 1), n , n+s < p ≤ 1,

max n−1 n−1

   (7.2.41)
max (s)+, −s + (n − 1) 1
p − 1 < β < 1, (n − 1) p1 − 1 < γ < 1.
7.2 Atomic and Molecular Theory 381

Then there exist an exponent η ∈ (|s|, 1] along with some constant C ∈ (0, ∞)
p, p
such that any f ∈ Bs (Σ, σ) may be decomposed as

  N
(k,τ)
f =g+ λQ k,ν aQ k,ν (7.2.42)
τ τ
k ∈Z τ ∈Ik ν=1
k ≥
κΣ +1

p,q  β,γ  ∗
with convergence both in Bs (Σ, σ) and in G0 (Σ) , for some function
% ∗

p, p
g ∈ Bs (Σ, σ) ∩ L p (Σ, σ) (7.2.43)
1≤p ∗ ≤∞

satisfying

gBsp, p (Σ,σ) + sup g L p ∗ (Σ,σ) ≤ C f Bsp, p (Σ,σ) (7.2.44)

1≤p ∗ ≤∞

and for some family of η-smooth atoms aQ k,ν of type (p, s) along with some family
τ
of numbers λQ k,ν ∈ C, both indexed by k ∈ Z with k ≥  κΣ + 1, τ ∈ Ik , and
τ
ν ∈ {1, . . . , N(k, τ)}, satisfying

   N
(k,τ)
 p 1/p
λ  ≤ C f Bsp, p (Σ,σ) . (7.2.45)
Qτk,ν
k ∈Z τ ∈Ik ν=1
k ≥
κΣ +1

Proof From item (i) of Theorem 7.2.7, specialized to the case when s, p, β, γ are as
p, p
in (7.2.41) and when q := p, we know that any f ∈ Bs (Σ, σ) may be decomposed
as in (7.2.42) with
 N (
κΣ,τ)
g := λQ κΣ ,ν aQ κΣ ,ν (7.2.46)
τ τ
τ ∈IκΣ ν=1

for some numerical sequence λ = {λQ k,ν }Q k,ν ∈D∗ (Σ) satisfying (cf. (7.2.32))
τ τ

λ p = λb p, p (Σ) ≤ C f Bsp, p (Σ,σ) . (7.2.47)

To proceed, fix an arbitrary integrability exponent p∗ ∈ [1, ∞]. From (7.2.3)-

(7.2.3) with k :=  κΣ it follows that there exists a constant C ∈ (0, ∞) with the
property that
 
a κΣ ,ν  p ∗ ≤ C for all τ ∈ IκΣ and ν ∈ {1, . . . , N(
κΣ, τ)}. (7.2.48)
Q τ L (Σ,σ)

As such, based on (7.2.46)-(7.2.48) we may estimate

382 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

  N
(
κΣ,τ) 
 
g L p ∗ (Σ,σ) =  λQ κΣ ,ν aQ κΣ ,ν  ∗
τ τ L p (Σ,σ)
τ ∈IκΣ ν=1

 N
(
κΣ,τ)
  
≤ λ 
κ ,ν
a 
κ ,ν
 ∗
Qτ Σ Qτ Σ L p (Σ,σ)
τ ∈IκΣ ν=1

 N
(
κΣ,τ)
 
≤C λ 
κ ,ν
 ≤ Cλ p (Σ)
Qτ Σ
τ ∈IκΣ ν=1

= Cλb p, p (Σ) ≤ C f Bsp, p (Σ,σ), (7.2.49)

thanks to the fact that p ∈ (0, 1] (cf. (7.2.41)). In addition, from (7.2.42), (7.2.10),
and item (i) of Theorem 7.2.8 we conclude that g belongs to the Besov space
p, p
Bs (Σ, σ) and there exists a constant C ∈ (0, ∞) independent of f with the property
that gBsp, p (Σ,σ) ≤ C f Bsp, p (Σ,σ) . All together, this analysis shows that g is as in
(7.2.43)-(7.2.44) and that (7.2.45) holds. 

7.3 Calderón’s Reproducing Formula and Frame Theory

The result in Lemma 7.3.1 describes a general version of Calderón’s reproducing

formula proved in [87, Theorem 1, p. 575], although the present formulation follows
[92, Lemma 2, pp. 76-77], [199, Lemma 2.2, p. 573]. Related results may be found in
[91, Theorem 4.1, p. 69], [202, Lemma 2.4, p. 100], and [89, Theorem 4.16, p. 109].

Lemma 7.3.1 Let Σ ⊆ Rn be a closed Ahlfors regular set. Abbreviate σ := H n−1 Σ

and recall  κΣ from (7.1.24). Bring in the family of conditional expectation operators
{Ek }k ∈Z, k ≥κΣ introduced in Definition 7.1.5 and, for each k ∈ Z with k ≥  κΣ , denote
by Ek (·, ·) the integral kernel of Ek .
Then there exist a family of functions E  κΣ ,ν (·) defined on
Q τ ∈I , ν ∈ {1,..., N (
τ κΣ,τ)}
κΣ
k (·, ·)
Σ, along with a family of functions E defined on Σ × Σ, satisfying the
k ∈Z, k ≥
κΣ
following properties:

(a) Given any ε ∈ (0, 1) there exists C ∈ (0, ∞) with the following significance. For
each k ∈ Z with k ≥  κΣ one has

C2−kε
| Ek (x, y)| ≤ , ∀x, y ∈ Σ, (7.3.1)
(2−k + |x − y|)n−1+ε

for every x, x , y ∈ Σ,
7.3 Calderón’s Reproducing Formula and Frame Theory 383

C2−kε |x − x | ε 2−k + |x − y|
| Ek (x, y) − E
k (x , y)| ≤ if |x − x | < ,
(2−k + |x − y|)n−1+2ε 10
(7.3.2)
and ∫ ∫
k (x, y) dσ(y) =
E k (y, x) dσ(y) = 0, ∀x ∈ Σ,
E (7.3.3)
Σ Σ
whereas for each τ ∈ IκΣ and ν ∈ {1, . . . , N(
κΣ, τ)} one has
 
  C
EQ κΣ ,ν (x) ≤ , ∀x ∈ Σ, ∀y ∈ QτκΣ,ν, (7.3.4)
τ (1 + |x − y|)n−1+ε

for every x, z ∈ Σ and y ∈ QτκΣ,ν ,

  C|x − z| ε 1 + |x − y|
  κΣ ,ν (z) ≤
EQ κΣ ,ν (x) − E if |x − z| < , (7.3.5)
τ Q τ (1 + |x − y|)n−1+2ε 10

and ∫
 κΣ ,ν (x) dσ(x) = 1.
E (7.3.6)
Q τ
Σ
Moreover, the constant C appearing in (7.3.4) and (7.3.4) may be taken to
be independent of jΣ , the large integer fixed earlier1 with the property that
 
diam QτκΣ,ν ≈ 2−jΣ uniformly for τ ∈ IκΣ and ν ∈ {1, . . . , N(
κΣ, τ)}.
 β,γ  ∗
(b) For each functional f ∈ G0 (Σ) with 0 < β, γ < 1 one has (with yτk,ν
denoting the center of the cube Qτk,ν )

 N
(
κΣ,τ)
f =  κΣ ,ν (·)
σ(QτκΣ,ν ) mQ κΣ ,ν (EκΣ f )E (7.3.7)
τ Q τ
τ ∈IκΣ ν=1

  N
(k,τ)
+ k (·, yτk,ν )
σ(Qτk,ν )(Ek f )(yτk,ν )E
k ∈Z τ ∈Ik ν=1
k ≥
κΣ +1

 β ,γ ∗
where the series converges in G0 1 1 (Σ) for any β1 ∈ (β, 1) and γ1 ∈ (γ, 1).

The following two propositions provide a natural mechanism for moving back
p,q
and forth between discrete Besov spaces, bs (Σ), and continuous Besov spaces,
p,q p,q
Bs (Σ, σ), as well as between the discrete Triebel-Lizorkin spaces, fs (Σ), and
p,q
continuous Triebel-Lizorkin spaces, Fs (Σ, σ).

Proposition 7.3.2 Let Σ ⊆ Rn be some closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. With 
κΣ introduced as in (7.1.24), consider the family of functions

1 see the comment right before (7.1.27)

384 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

 κΣ ,ν (·)
E defined on Σ, along with the family of functions
Q τ τ ∈IκΣ , ν ∈ {1,..., N (
κΣ,τ)}
k (·, ·)
E defined on Σ × Σ, from Lemma 7.3.1.
k ∈Z, k ≥
κΣ
Next, fix s ∈ (−1, 1) along with p ∈ (0, ∞] satisfying max n−1
n , n+s < p ≤ ∞,
n−1

and suppose
    
max (s)+, −s + (n − 1) p1 − 1 + < β < 1, (n − 1) p1 − 1 + < γ < 1. (7.3.8)

Finally, for each sequence of complex numbers of the form

λ = λτk,ν ∈ C : k ∈ Z, k ≥ 
κΣ, τ ∈ Ik , ν ∈ {1, . . . , N(k, τ)} (7.3.9)

consider the (formal) series

 N
(
κΣ,τ)
Φ(λ) :=  κΣ ,ν (·)
λτκΣ,ν E Qτ
τ ∈IκΣ ν=1

  N
(k,τ)
+ k (·, yτk,ν ).
λτk,ν σ(Qτk,ν )E (7.3.10)
k ∈Z τ ∈Ik ν=1
κΣ +1
k ≥

Then the following properties hold.

(1) If q ∈ (0, ∞] and λbsp, q (Σ) < ∞ then the series (7.3.10) converges to some
p,q p,q  β,γ  ∗
distribution belonging to Bs (Σ, σ) both in Bs (Σ, σ) and G0 (Σ) when
 β,γ  ∗
when max{p, q} < ∞, and only in G0 (Σ) in the case when max{p, q} = ∞.
Moreover, in all cases there exists some constant C = C(Σ, s, p, q) ∈ (0, ∞) such
that
Φ(λ)Bsp, q (Σ,σ) ≤ Cλbsp, q (Σ), (7.3.11)
which, in particular, implies that the application
p,q p,q
Φ : bs (Σ) −→ Bs (Σ, σ) (7.3.12)

induced by (7.3.10) is well defined, linear, and bounded.

n , n+s
(2) If max n−1 < q ≤ ∞ and λ fsp, q (Σ) < ∞, then the series in (7.3.10)
n−1
p,q p,q
 β,γ  ∗to some distribution belonging to Fs (Σ, σ) both in
converges  F s (Σ,
β,γ  ∗σ) and
G0 (Σ) in the case when max{p, q} < ∞, and only in G0 (Σ) in the
case when max{p, q} = ∞. Furthermore, in all cases there exists some constant
C = C(Σ, s, p, q) ∈ (0, ∞) such that

Φ(λ)Fsp, q (Σ,σ) ≤ Cλ fsp, q (Σ) . (7.3.13)

Hence, the application

p,q p,q
Φ : fs (Σ) −→ Fs (Σ, σ) (7.3.14)
7.3 Calderón’s Reproducing Formula and Frame Theory 385

is also well-defined, linear, and bounded.

(3) For each k ∈ Z with k ≥ 
κΣ , τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}, define

⎧
⎪ κΣ,ν s − 1 + 1 

⎨ σ(Qτ ) n−1 p 2 E
⎪ κ ,ν (·) if k = 

Qτ Σ
κΣ ,
ψτk,ν := (7.3.15)
⎪
⎪ σ(Q k,ν ) n−1
s
− p1 + 32 
⎩ τ Ek (·, yτk,ν ) if k ≥ 
κΣ + 1,

(where, as before, yτk,ν is the center of Qτk,ν ). Then for each q ∈ (0, ∞] there
exists a constant C = C(Σ, p, q, s) ∈ (0, ∞) with the property that
 
each ψτk,ν belongs to Bs (Σ, σ) and ψτk,ν B p, q (Σ,σ) ≤ C.
p,q
(7.3.16)
s

In addition, whenever max n−1 n , n+s < q ≤ ∞ there exists some constant
n−1

C = C(Σ, p, q, s) ∈ (0, ∞) with the property that

 
each ψτk,ν belongs to Fs (Σ, σ) and ψτk,ν F p, q (Σ,σ) ≤ C.
p,q
(7.3.17)
s

Proof For items (1)-(2) see [89, Proposition 7.3, p. 214] and also [199, Theorem 2.1,
p. 575]. To deal with the claims in item (3), fix some k o ∈ Z with k o ≥  κΣ , along
with some τo ∈ Iko and some νo ∈ {1, . . . , N(k o, τo )}. For each k ∈ Z with k ≥  κΣ ,
τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)}, we then define
⎧
⎪ 
κ ,ν s
− 1 +1
⎨ σ(QτΣo o ) n−1 p 2 if k = k o , and τ = τo , and ν = νo,
⎪
λQ k,ν := (7.3.18)
τ ⎪
⎪0
⎩ if either k  k o , or τ  τo , or ν  νo .

Finally, with D∗ (Σ) as in (7.2.1), consider the numerical sequence

λ := λQ k,ν Qτk,ν ∈D∗ (Σ)

. (7.3.19)
τ

p,q
This has only one nonzero term, hence λ ∈ bs (Σ) for each q ∈ (0, ∞] and, as seen
from (7.3.18)-(7.3.19) and (7.2.28), there exists a constant C = C(Σ, s) ∈ (0, ∞) with
the property that
    N
(k,τ)  q/p  1/q
 p
σ(Qτk,ν ) p − 2 |λQ k,ν |
1 1
λ p, q
b s (Σ) = 2 ks
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

s
= 2ko s σ(Qτkoo,νo ) p − 2 σ(Qτkoo,νo ) n−1 − p + 2 ≤ C,
1 1 1 1
(7.3.20)

where inequality uses the fact that 2ko ≈ σ(Qτkoo,νo )− n−1 . Since, as seen from (7.3.10),
1

(7.3.18)-(7.3.19), and (7.3.15),

Ψ(λ) = ψτkoo,νo , (7.3.21)

386 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

we may invoke the current item (1) to conclude that

 
ψτkoo,νo ∈ Bs (Σ, σ) and ψτkoo,νo B p, q (Σ,σ) ≤ C,
p,q
(7.3.22)
s

for some constant C = C(Σ, p, q, s) ∈ (0, ∞). This establishes (7.3.16).

Finally, the proof of (7.3.17) is similar, this time invoke the current item (2) and
using
 
   N (k,τ)
1   q  1/q 
 
λ fsp, q (Σ) =  2ks σ(Qτk,ν )− 2 λQ k,ν  1Q k,ν 
 τ τ  p
k ∈Z τ ∈Ik ν=1 L (Σ,σ)
k ≥κΣ

s 1 1 
= 2ko s σ(Qτkoo,νo )− 2 σ(Qτkoo,νo ) n−1 − p + 2 1Q k o ,νo  L p (Σ,σ) ≤ C, (7.3.23)
1

τo

in place of (7.3.20). 

Here is the second proposition alluded to above.

Proposition 7.3.3 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. With  κΣ as in (7.1.24), consider the family of conditional expectation
operators {Ek }k ∈Z, k ≥κΣ from Definition 7.1.5. Also, bring in the family of functions
 κΣ ,ν (·)
E defined on Σ, along with the family of functions
Q τ τ ∈I , ν ∈ {1,..., N (

κΣ κΣ,τ)}
k (·, ·)
E defined on Σ × Σ, from Lemma 7.3.1. Finally, for every functional
 β,γk ∈Z, k ∗≥κΣ
f ∈ G0 (Σ) with 0 < β, γ < 1 let

λτκΣ,ν := σ(QτκΣ,ν ) mQ κΣ ,ν (EκΣ f ) for τ ∈ IκΣ and ν ∈ {1, . . . , N(

κΣ, τ)},
τ

λτk,ν := (Ek f )(yτk,ν ) for k ∈ Z, k ≥ 

κΣ + 1, τ ∈ Ik , and ν ∈ {1, . . . , N(k, τ)},
(7.3.24)
where yτk,ν is the center of Qτk,ν and, with D∗ (Σ) as in (7.2.1), define

Ψ( f ) := λτk,ν Qτk,ν ∈D∗ (Σ)

. (7.3.25)

Then for each s ∈ (−1, 1) and p ∈ (0, ∞] satisfying max n−1

n , n+s < p ≤ ∞ the
n−1

following conclusions are valid.

p,q  β,γ  ∗
(i) If q ∈ (0, ∞] then f ∈ Bs (Σ, σ) if and only if f ∈ G0 (Σ) for some β, γ
satisfying
  
max (s)+, −s +(n −1) p1 −1 < β < 1, (n −1) p1 −1 < γ < 1, (7.3.26)
+ +

the sequence λ = λτk,ν Q k,ν ∈D∗ (Σ) := Ψ( f ) defined as in (7.3.25) belongs to

τ
p,q
bs (Σ), and the discrete Calderón reproducing formula
7.3 Calderón’s Reproducing Formula and Frame Theory 387

 N
(
κΣ,τ)
f =  κΣ ,ν (·)
λτκΣ,ν E Q τ
τ ∈IκΣ ν=1

  N
(k,τ)
+ k (·, yτk,ν )
λτk,ν σ(Qτk,ν )E (7.3.27)
k ∈Z τ ∈Ik ν=1
κΣ +1
k ≥

 β,γ  ∗
holds in G0 (Σ) . Moreover, the coefficients associated as in (7.3.25) with
p,q
each f ∈ Bs (Σ, σ) satisfy the following frame property:
p,q
Ψ( f ) ∈ bs (Σ) and  f Bsp, q (Σ,σ) ≈ Ψ( f )bsp, q (Σ),
p,q
(7.3.28)
uniformly for f ∈ Bs (Σ, σ).
p,q  β,γ  ∗
(ii) If max n−1n , n+s < q ≤ ∞, then f ∈ Fs (Σ, σ) if and only if f ∈ G0 (Σ)
n−1

for some β, γ as in (7.3.26), the discrete Calderón reproducing formula (7.3.27)
β,γ ∗
holds in G0 (Σ) , and the sequence λ = λτk,ν Q k,ν ∈D∗ (Σ) := Ψ( f ) defined as
τ
p,q p,q
in (7.3.25) belongs to fs (Σ). In addition, for each f ∈ Fs (Σ, σ) one has
p,q
Ψ( f ) ∈ fs (Σ) and  f Fsp, q (Σ,σ) ≈ Ψ( f ) fsp, q (Σ),
p,q
(7.3.29)
uniformly for f ∈ Fs (Σ, σ).

Proof See [89, Theorem 7.4, p. 219] and also [199, Theorem 2.2, p. 585]. 
We also have the following discrete Calderón reproducing formula in Besov
spaces.
Proposition 7.3.4 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Pick a smoothness index s ∈ (−1, 1) along with an integrability
n , n+s < p, and consider β, γ satisfying
exponent p ∈ (0, ∞) for which max n−1 n−1

  
max (s)+, −s + (n − 1) 1
p −1 < β < 1, (n − 1) 1
p −1 < γ < 1. (7.3.30)
+ +
p,q
Then for each f belonging to the space Bs (Σ, σ) with q ∈ (0, ∞), if the sequence
λ = λτk,ν Q k,ν ∈D∗ (Σ) := Ψ( f ) is defined as in (7.3.24), one has the discrete Calderón
τ
reproducing formula

 N
(
κΣ,τ)
f =  κΣ ,ν (·)
λτκΣ,ν E Qτ
τ ∈IκΣ ν=1

  N
(k,τ)
+ k (·, yτk,ν )
λτk,ν σ(Qτk,ν )E (7.3.31)
k ∈Z τ ∈Ik ν=1
κΣ +1
k ≥

p,q
with convergence in Bs (Σ, σ).
388 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
p,q
Furthermore, if for some f ∈ Fs (Σ, σ) with max n−1 n , n+s < q < ∞, then
n−1

one considers the sequence λ = λτk,ν Q k,ν ∈D∗ (Σ) := Ψ( f ) is defined as in (7.3.24),
τ
then the discrete Calderón reproducing formula (7.3.31) holds with convergence in
p,q
Fs (Σ, σ).
p,q
Proof To begin with, since f ∈ Bs (Σ, σ), Proposition 7.3.3 gives that λ is in ∗
p,q β ,γ
bs (Σ) and the series in the right-hand side of (7.3.31) converges to f in G0 1 1 (Σ)
for any indices β1 ∈ (β, 1) and γ1 ∈ (γ, 1) satisfying similar properties as β, γ do in
(7.3.30). On the other hand, since we have max{p, q} < ∞, Proposition 7.3.2  β ,γapplies
∗
p,q
and gives that the series in (7.3.31) converges both in Bs (Σ, σ) and in G0 1 1 (Σ)
for any β1 ∈ (β, 1) and γ1 ∈ (γ, 1) satisfying similar properties as β, γ do in (7.3.30).
p,q
As such, the series in the right-hand side of (7.3.31) converges to f in Bs (Σ, σ).
p,q
Finally, the case when f ∈ Fs (Σ, σ) with max n , n+s < q < ∞ is handled
n−1 n−1

similarly. 

When considered together, Proposition 7.3.2 and Proposition 7.3.3 yield some
very useful consequences which we describe next.

Proposition 7.3.5 In the context of Propositions 7.3.2-7.3.3, the bounded linear

maps Φ, Ψ satisfy
Φ ◦ Ψ = I, the identity operator, (7.3.32)
both on the scales of Besov and Triebel-Lizorkin spaces. As a result, in the context
of Propositions 7.3.2-7.3.3,
Φ is onto, and Ψ is a quasi-isometric embedding (i.e., Ψ is injective
and distorts quasi-norms only up to a fixed multiplicative factor) of
(7.3.33)
the continuous scales of Besov and Triebel-Lizorkin spaces into their
respective discrete versions.
 β,γ  ∗
Finally, when suitably interpreted, formula (7.3.32) also holds in G0 (Σ) .

Proof This is a straightforward consequence of Proposition 7.3.2, Proposition 7.3.3,

as well as Calderón’s reproducing formula described in Lemma 7.3.1. 

The frame theory developed so far may, in turn, be used to prove the existence
of some very useful approximations to the identity on Besov and Triebel-Lizorkin
scales of the sort described in the proposition below.

Proposition 7.3.6 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then there exists a family of operators {PN } N ∈N satisfying the
following properties.

n , n+s < p ≤ ∞, and 0 < q ≤ ∞. Then each
(1) Assume that s ∈ (−1, 1), max n−1 n−1

p,q p,q
operator PN : Bs (Σ, σ) → Bs (Σ, σ) is linear, bounded, of finite rank, and

sup PN Bsp, q (Σ,σ)→Bsp, q (Σ,σ) < +∞. (7.3.34)

N ∈N
7.3 Calderón’s Reproducing Formula and Frame Theory 389

(2) Assume that s ∈ (−1, 1), max n , n+s
n−1 n−1
< p < ∞, and 0 < q < ∞. Also, fix
p,q
some relatively compact subset O of Bs (Σ, σ). Then for every ε > 0 there
exists N(ε) ∈ N such that

sup  f − PN f Bsp, q (Σ,σ) < ε whenever N ≥ N(ε). (7.3.35)

f ∈O

In particular, corresponding to the case when O is a singleton, one has (with I

p,q
denoting the identity operator on Bs (Σ, σ))
p,q
PN → I pointwise on Bs (Σ, σ) as N → ∞. (7.3.36)

(3) The same family of operators {PN } N ∈N enjoys similar properties as in items
p,q
(1)-(2) above, now formulated
 on the Triebel-Lizorkin scale Fs (Σ, σ) with
s ∈ (−1, 1) and max n−1 n , n+s < p, q ≤ ∞ for item (1) and s ∈ (−1, 1) and
n−1

max n−1 n , n+s < p, q < ∞ for item (2).
n−1

Proof Recall the family of dyadic cubes defined in (7.2.1) and recall the parameter

κΣ from (7.1.24). For each k ∈ Z with k ≥  κΣ , consider a nested family {IkN } N ∈N
of finite subsets of Ik such that Ik  Ik as N → ∞. Also, fix β, γ satisfying
N

(7.3.26).  ∗ any N ∈ N, define the operator f → PN f mapping each functional

 β,γGiven
f ∈ G0 (Σ) into

 N
(
κΣ,τ)
PN f :=  κΣ ,ν (·)
σ(QτκΣ,ν ) mQ κΣ ,ν (EκΣ f )E
τ Q τ
τ ∈IκN ν=1
Σ

  N
(k,τ)
+ k (·, yτk,ν ).
σ(Qτk,ν )(Ek f )(yτk,ν )E (7.3.37)
k ∈Z τ ∈IkN ν=1
N ≥k ≥
κΣ +1

Note that, by design, the sums in the right-hand side run over finite sets of indices.
p,q
Suppose now that some arbitrary f ∈ Bs (Σ, σ) has been fixed. Then Proposi-
k,ν
tion 7.3.3, used with λ := Ψ( f ) = λτ Q k,ν ∈D∗ (Σ) as in (7.3.24)-(7.3.25), ensures
τ
p,q
that λ ∈ bs (Σ), the equality

 N
(
κΣ,τ)
f =  κΣ ,ν (·)
λτκΣ,ν E Qτ
τ ∈IκΣ ν=1

  N
(k,τ)
+ k (·, yτk,ν )
λτk,ν σ(Qτk,ν )E (7.3.38)
k ∈Z τ ∈Ik ν=1
k ≥
κΣ +1

 β ,γ ∗
holds in G0 1 1 (Σ) for any β1 ∈ (β, 1) and γ1 ∈ (γ, 1) satisfying (7.3.26), and
390 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

 f Bsp, q (Σ,σ) ≈ λbsp, q (Σ) . (7.3.39)

k,ν
For each N ∈ N consider λ N := {λτ, N }k ∈Z, k ≥
κΣ, τ ∈Ik , 1≤ν ≤ N (k,τ) where

k,ν λτk,ν if 
κΣ ≤ k ≤ N, τ ∈ IkN , 1 ≤ ν ≤ N(k, τ),
λτ, N := (7.3.40)
0 otherwise.
p,q
Then clearly λ N ∈ bs (Σ) and λ N bsp, q (Σ) ≤ λbsp, q (Σ) < ∞ for each N ∈ N.
Moreover, by Proposition 7.3.2 we have PN f Bsp, q (Σ,σ) ≤ Cλ N bsp, q (Σ) for some
C > 0 independent of N. In combination with (7.3.39), these properties yield

PN f Bsp, q (Σ,σ) ≤ Cλ N bsp, q (Σ) ≤ Cλbsp, q (Σ)

≤ C f Bsp, q (Σ,σ) for each N ∈ N. (7.3.41)

p,q p,q
This proves (7.3.34). In particular, each operator PN : Bs (Σ, σ) → Bs (Σ, σ) is
p,q p,q
linear and bounded. The fact that each operator PN : Bs (Σ, σ) → Bs (Σ, σ) has
finite rank is apparent from the definition of PN in (7.3.37).
To proceed, observe that

 f − PN f Bsp, q (Σ,σ) (7.3.42)


  N (
κΣ,τ)
  κΣ ,ν (·)
= σ(QτκΣ,ν ) mQ κΣ ,ν (EκΣ f )E
 τ Qτ
N
τ ∈IκΣ \Iκ ν=1
Σ 
  N
(k,τ) 
+ k (·, yτk,ν )
σ(Qτk,ν )(Ek f )(yτk,ν )E .
 p, q
k>N τ ∈I N ν=1 Bs (Σ,σ)
k

p,q
Since f ∈ Bs (Σ, σ) and we are presently assuming max{p, q} < ∞, Proposi-
tion 7.3.4 ensures that the series in the right-hand side of (7.3.38) converges to f
p,q
in Bs (Σ, σ). This implies that, in this case, the norm in (7.3.42) converges to 0 as
N → ∞, establishing (7.3.36).
We are left with proving (7.3.35). To this end, fix ε > 0, a relatively compact
p,q
subset O of Bs (Σ, σ), and define

C := max 1, sup PN Bsp, q (Σ,σ)→Bsp, q (Σ,σ) ∈ [1, ∞). (7.3.43)
N ∈N

To facilitate the subsequent discussion, for a generic quasi-metric space (X,  · X ),

g ∈ X, and r > 0, we agree to use the notation

B(g, r; X) := { f ∈ X :  f − gX < r }. (7.3.44)

 ε p,q 
Let f0 ∈ O be arbitrary but fixed, and consider some f ∈ B f0, 3C ; Bs (Σ, σ) . By
(7.3.36), there exists some N( f0 ) ∈ N such that
7.4 Interpolation of Besov and Triebel-Lizorkin Spaces via the Real Method 391
ε
 f0 − PN f0 Bsp, q (Σ,σ) < if N ≥ N( f0 ). (7.3.45)
3C
Moreover, (7.3.43) entails
ε
PN f0 − PN f Bsp, q (Σ,σ) < for each N ∈ N. (7.3.46)
3
Combining (7.3.45) and (7.3.46), we conclude that

 f − PN f Bsp, q (Σ,σ) < ε for each N ≥ N( f0 ). (7.3.47)

Thus, so far we have proved that for each f0 ∈ O there exists N( f0 ) ∈ N such that

sup  f − PN f Bsp, q (Σ,σ) < ε whenever N ≥ N( f0 ). (7.3.48)

ε p, q
f ∈B( f0, 3C ; Bs (Σ,σ))

p,q
From the fact that O is relatively compact in Bs (Σ, σ), it follows that there exist
J ∈ N and f j ∈ O with j ∈ {1, . . . , J}, such that

$
J
 
ε p,q
O⊆ B f j , 3C ; Bs (Σ, σ) . (7.3.49)
j=1

If for each j ∈ {1, . . . , J} we denote by N( f j ) the natural number obtained by

applying the reasoning above to f j instead of f0 , then (7.3.48) holds with f0 replaced
by f j , for each j ∈ {1, . . . , J}. Hence, (7.3.49) shows that
 for any f ∈ O there
ε p,q
exists jo ∈ {1, . . . , J} such that f ∈ B f jo , 3C ; Bs (Σ, σ) . Granted this, (7.3.48)
applies and gives that  f − PN f Bsp, q (Σ,σ) < ε for all integers N ≥ N(ε), where
N(ε) := max1≤ j ≤J N( f j ) is independent of f . This completes the proof of (7.3.35).
Finally, that the same family of operators {PN } N ∈N satisfies analogous prop-
p,q
erties when acting on the Triebel-Lizorkin scale Fs (Σ, σ) with s ∈ (−1, 1) and
max n−1 n , n+s < p, q ≤ ∞ is established similarly. 
n−1

7.4 Interpolation of Besov and Triebel-Lizorkin Spaces via the

Real Method

This section deals with two theorems regarding the behavior of both the inhomoge-
neous and homogeneous Besov and Triebel-Lizorkin spaces under the real method
of interpolation method. Such results have been well-understood in the Euclidean
setting for a long time (see [187] and [15] for excellent references) and have subse-
quently been generalized to the context of d-Ahlfors-regular quasi-metric spaces in
[202] and to reverse-doubling spaces in [89]. Below, we present some results found
in [202] and [89], but recorded here for an optimal range of indices.
We begin by discussing several results regarding the real interpolation of the
inhomogeneous Besov and Triebel-Lizorkin spaces.
392 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Theorem 7.4.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix q ∈ (0, ∞] and θ ∈ (0, 1). Suppose s0, s1 ∈ (−1, 1) with
s0  s1 and set s := (1 − θ)s0 + θs1 .
Then for each q0, q1 ∈ (0, ∞] and p ∈ (0, ∞] satisfying

n , n+s0 , n+s1 < p,

n−1 n−1 n−1
max (7.4.1)

one has  
p,q p,q p,q
Bs0 0 (Σ, σ), Bs1 1 (Σ, σ) θ,q = Bs (Σ, σ). (7.4.2)
Moreover, if p ∈ (0, ∞) is as in (7.4.1) and q0, q1 ∈ (0, ∞] satisfy

n , n+s j < q j ≤ ∞ for j ∈ {0, 1},

n−1 n−1
max (7.4.3)

then  
p,q0 p,q1 p,q
Fs0 (Σ, σ), Fs1 (Σ, σ) θ,q = Bs (Σ, σ). (7.4.4)

Proof See [202, Theorem 2.3, p. 100] and [89, Theorem 8.9, p. 230]. 

We augment the above theorem with the following companion results in which
the smoothness index stays unchanged.

Theorem 7.4.2 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix θ ∈ (0, 1) along with s ∈ (−1, 1).
 
θ −1
(i) Assume 1 ≤ p ≤ ∞, 0 < q0, q1 ≤ ∞, and define q := 1−θ q0 + q1 . Then
 p,q0 p,q1  p,q
Bs (Σ, σ), Bs (Σ, σ) θ,q = Bs (Σ, σ). (7.4.5)

(ii) Assume max{ n−1 , n−1 } < p j < ∞ for j ∈ {0, 1}, 1 ≤ q ≤ ∞, and define
 1−θ n n+s
θ −1
p := p0 + p1 . Then
 p ,q p ,q  p,q
Fs 0 (Σ, σ), Fs 1 (Σ, σ) θ, p = Fs (Σ, σ). (7.4.6)

Proof For (i) see [199, Theorem 3.2, p. 587], while for (ii) see [199, Theorem 3.3,
p. 587]. 

The next result. describes the .behavior of the homogeneous Besov and Triebel-
p,q p,q
Lizorkin spaces Bs (Σ, σ) and Fs (Σ, σ) via the real method.

Theorem 7.4.3 Assume Σ ⊆ Rn is an unbounded, closed, Ahlfors regular set, and

abbreviate σ := H n−1 Σ. Also, fix q ∈ (0, ∞] and θ ∈ (0, 1). Finally, suppose
s0, s1 ∈ (−1, 1) with s0  s1 and set s := (1 − θ)s0 + θs1 .
Then for each q0, q1 ∈ (0, ∞] and p ∈ (0, ∞] satisfying

p > max n , n+s0 , n+s1

n−1 n−1 n−1
, (7.4.7)

one has
7.4 Interpolation of Besov and Triebel-Lizorkin Spaces via the Real Method 393
 . p,q0 . p,q  . p,q
Bs0 (Σ, σ), Bs1 1 (Σ, σ) θ,q = Bs (Σ, σ). (7.4.8)
Moreover, if p ∈ (0, ∞) is as in (7.4.1) and q0, q1 ∈ (0, ∞] satisfy

n , n+s j < q j ≤ ∞ for j ∈ {0, 1},

n−1 n−1
max (7.4.9)

then  . p,q0 . p,q  . p,q

Fs0 (Σ, σ), Fs1 1 (Σ, σ) θ,q = Bs (Σ, σ). (7.4.10)

Proof See [202, Theorem 3.1, p. 111] and [89, Theorem 8.8, p. 225]. 

In contrast to Theorem 7.4.1, Theorem 7.4.2, and Theorem 7.4.3, our next result
deals with the real interpolation of the homogeneous and inhomogeneous Besov
spaces where both integrability exponents are allowed to vary.

Theorem 7.4.4 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix an arbitrary parameter θ ∈ (0, 1).

(i) Suppose s0, s1 ∈ (−1, 1) and consider p0, p1 ∈ (0, ∞] with p0  p1 satisfying

p j > max n , n+s j

n−1 n−1
for j ∈ {0, 1}. (7.4.11)

In this context, set s := (1 − θ)s0 + θs1 and let the number p ∈ (0, ∞] be such
that 1/p = (1 − θ)/p0 + θ/p1 . Then one has
 p0, p0 p ,p  p, p
Bs0 (Σ, σ), Bs11 1 (Σ, σ) θ, p = Bs (Σ, σ), (7.4.12)
 p , p0 p , p1  p, p
Fs00 (Σ, σ), Fs11 (Σ, σ) θ, p = Fs (Σ, σ). (7.4.13)

Additionally, if Σ is unbounded, then

 . p0, p0 .p ,p  . p, p
Bs0 (Σ, σ), Bs11 1 (Σ, σ) θ, p = Bs (Σ, σ), (7.4.14)
 . p0, p0 . p ,p  . p, p
Fs0 (Σ, σ), Fs11 1 (Σ, σ) θ, p = Fs (Σ, σ). (7.4.15)

(ii) Suppose s0, s1 ∈ (−1, 1) with s0  s1 , max{ n−1

n , } < p j < ∞ for j ∈ {0, 1}, n−1
n+s j
 
θ −1
1 ≤ q0, q1 ≤ ∞. Set s := (1 − θ)s0 + θs1 and define p := 1−θ
p0 + p1 . Then
 p ,q0 p ,q1  p, p  p, p 
Fs00 (Σ, σ), Fs11 (Σ, σ) θ, p = Bs (Σ, σ) = Fs (Σ, σ) . (7.4.16)
 
θ −1
(iii) Assume that s ∈ (−1, 1), 1 ≤ p0, p1 ≤ ∞, and define p := 1−θ p0 + p1 ∈ [1, ∞].
Then
 p0, p0 p ,p  p, p
Fs (Σ, σ), Fs 1 1 (Σ, σ) θ, p = Fs (Σ, σ), (7.4.17)
 p , p0 p , p1  p, p
Bs 0 (Σ, σ), Bs 1 (Σ, σ) θ, p = Bs (Σ, σ). (7.4.18)
394 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Proof For formulas (7.4.12), (7.4.14) see [89, Theorem 8.7, p. 224]. Then formulas
(7.4.13), (7.4.15) follow on account of these and (7.1.18), (7.1.38). For (ii)-(iii) see
[199, Theorem 3.3, p. 587]. 

As a special case, it is useful to single out the following interpolation result.

Theorem 7.4.5 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, assume n−1n < p0 ≤ 1 < p1 < ∞. Then, if θ ∈ (0, 1) and
θ
1
p = 1−θ
p0 + p1 , it follows that

  H p (Σ, σ) if n−1
n < p ≤ 1,
H (Σ, σ), L (Σ, σ)
p0 p1
θ, p = (7.4.19)
L p (Σ, σ) if 1 < p < ∞.

p ,2
Proof Use the identifications (7.1.20), ensuring that H p0 (Σ, σ) = F0 0 (Σ, σ), and
p ,2
(7.1.19) according to which L p1 (Σ, σ) = F0 1 (Σ, σ), along with (7.4.6). 

7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces

The complex method of interpolation, denoted by [·, ·]θ , has been originally intro-
duced in [23] in the context of Banach spaces. Since the full scales of Besov and
Triebel-Lizorkin spaces contain spaces which are not locally convex, we shall work
here with the adaptation of the complex method of interpolation to analytically con-
vex quasi-Banach spaces as described in §1.4. We begin, nonetheless, by recording
the following theorem dealing with the classical complex interpolation method for
the portion of the Besov and Triebel-Lizorkin scales consisting of Banach spaces.

Theorem 7.5.1 Assume that Σ ⊆ Rn is an arbitrary closed Ahlfors regular set,

and abbreviate σ := H n−1 Σ. Fix some θ ∈ (0, 1), along with s0, s1 ∈ (−1, 1)
and p0, q0, p1, q1 ∈ (1, ∞). Finally, consider s := (1 − θ)s0 + θs1 and introduce
 
θ −1
 
θ −1
p0 + p1
p := 1−θ q0 + q1
together with q := 1−θ . Then
 p ,q0 p ,q1  p,q
Bs00 (Σ, σ), Bs11 (Σ, σ) θ
= Bs (Σ, σ), (7.5.1)
 p ,q0 p ,q1  p,q
Fs00 (Σ, σ), Fs11 (Σ, σ) θ
= Fs (Σ, σ). (7.5.2)

Moreover, when Σ is unbounded, similar results hold for the homogeneous scales of
Besov and Triebel-Lizorkin spaces on Σ.

A proof may be found in [86, Theorem 5.2, p. 64] and [90, Theorem 7.7, p. 121].
This should be compared with results in the standard Euclidean setting to the effect
that if s0, s1 ∈ R, 1 ≤ p0, p1, q0, q1 ≤ ∞, and if
θ θ
0 < θ < 1, s = (1 − θ)s0 + θs1, 1
p = 1−θ
p0 + p1 ,
1
q = 1−θ
q0 + q1 , (7.5.3)
7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces 395

then  p ,q0 p ,q1  p,q

Fs00 (Rn ), Fs11 (Rn ) θ
= Fs (Rn ) provided
(7.5.4)
either max {p0, q0 } < ∞, or max {p1, q1 } < ∞,
and also  p ,q0 p ,q1  p,q
Bs00 (Rn ), Bs11 (Rn ) θ
= Bs (Rn ),
(7.5.5)
granted that min {q0, q1 } < ∞.
In addition, analogous formulas are valid for the homogeneous versions of the Besov
and Triebel-Lizorkin spaces. Indeed, this particular collection of results is obtained
by a careful inspection of the proofs of several well-known results in [15], [58], [166],
[186]. Here we wish to mention that formulas (7.5.4), (7.5.5) have established in
[127] for the full range of indices for which the scales of Besov and Triebel-Lizorkin
spaces are defined (cf. also [110]). See Theorem 9.1.6 below.
The main result of this section is an extension of Theorem 7.5.1 to a larger range
of indices, for which the scales of Besov and Triebel-Lizorkin spaces include non
locally convex spaces. More specifically, we have:

Theorem 7.5.2 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, select some θ ∈ (0, 1) along with s0, s1 ∈ (−1, 1), and finally
define s := (1 − θ)s0 + θs1 . Then whenever

max n−1n , n+s j < p j ≤ ∞, 0 < q j ≤ ∞, for j ∈ {0, 1},
n−1

 
θ −1
 
θ −1
p0 + p1 , q := 1−θq0 + q1 , and
p := 1−θ (7.5.6)

either max{p0, q0 } < ∞, or max{p1, q1 } < ∞,

it follows that
p ,q0 p ,q1 p ,q0 p ,q1
Bs00 (Σ, σ), Bs11 (Σ, σ), and Bs00 (Σ, σ) + Bs11 (Σ, σ)
(7.5.7)
are all analytically convex quasi-Banach spaces

and  
p ,q0 p ,q1 p,q
Bs00 (Σ, σ), Bs11 (Σ, σ) θ
= Bs (Σ, σ). (7.5.8)
Furthermore, if
 
max n−1 , n−1
n n+s j < p j < ∞, max n , n+s < q j < ∞, for j ∈ {0, 1},
n−1 n−1

    (7.5.9)
θ −1 θ −1
and p := 1−θ p0 + p1 , q := 1−θq0 + q1 ,

then
p ,q0 p ,q1 p ,q0 p ,q1
Fs00 (Σ, σ), Fs11 (Σ, σ), and Fs00 (Σ, σ) + Fs11 (Σ, σ)
(7.5.10)
are all analytically convex quasi-Banach spaces

and
396 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 p ,q0 p ,q1  p,q
Fs00 (Σ, σ), Fs11 (Σ, σ) θ
= Fs (Σ, σ). (7.5.11)
Moreover, analogous results (for the same ranges of indices as above) are valid
for the discrete scales of Besov and Triebel-Lizorkin spaces on Σ.
Finally, when Σ is unbounded, similar results hold for the homogeneous scales of
Besov and Triebel-Lizorkin spaces on Σ.

As a preamble to the proof of Theorem 7.5.2, we discuss several useful aux-

iliary results. The following two results appear in [15, Theorem 5.5.3, p. 120 and
Theorem 5.6.3, p. 123].
First, it is possible to identify the intermediate interpolation spaces produced by
the complex method between weighted Lebesgue spaces with change of weights.

Proposition 7.5.3 Let (Σ, μ) be a measure space and let ω0, ω1 : Σ → [0, ∞) be two
μ-measurable functions. Then whenever 1 ≤ p0, p1 < ∞ and θ ∈ (0, 1) the following
formula holds  p 
L 0 (Σ, ω0 μ), L p1 (Σ, ω1 μ) θ = L p (Σ, ω μ), (7.5.12)
where
 1−θ 
θ −1
  (1−θ )p   θ p
p := p0 + p1 and ω := ω0 p0 ω1 p1 , (7.5.13)
with equality of norms, i.e., for each f ∈ L p (Σ, ω μ) one has

 f [L p0 (Σ, ω0 μ), L p1 (Σ, ω1 μ)]θ =  f  L p (Σ, ω μ) . (7.5.14)

Proof See [15, Theorem 5.5.3, p. 120]. 

Having fixed a background set Σ ⊆ Rn , consider 

κΣ as in (7.1.24) and abbreviate

ZΣ := k ∈ Z : k ≥ 
κΣ . (7.5.15)

Definition 7.5.4 Assume that X is a quasi-Banach space, q ∈ (0, ∞) and that s ∈ R.

q
In this context, define s (ZΣ, X) as the space of sequences { fk }k ∈ZΣ , with fk ∈ X for
each k ∈ ZΣ , such that
  q1
  
∞
 q
{ fk }k ∈Z  q := 2  fk X
ks
< ∞. (7.5.16)
Σ s (ZΣ,X)
κΣ
k=

Remark 7.5.5 In the context of Definition 7.5.4, if μs denotes the measure defined
#
on ZΣ as ∞ κΣ 2 δ {k } (where δ {x } denoting the Dirac mass at a point x), it is
k=
ks

natural to make the following identification (see (1.4.12)-(1.4.14))

q
s/q (ZΣ, X) = L q (ZΣ, μs ) ⊗ X. (7.5.17)

Proposition 7.5.6 Let X0, X1 be two compatible Banach spaces, and suppose that
1 ≤ q0, q1 ≤ ∞, s0, s1 ∈ R, and θ ∈ (0, 1). Then the following complex interpolation
result holds:
7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces 397
 
q q q 
s00 (ZΣ, X0 ), s11 (ZΣ, X1 ) = s ZΣ, [X0, X1 ]θ , (7.5.18)
θ
 1−θ 
θ −1
where s := (1 − θ)s0 + θs1 and q := q0 + q1 .

Proof See [15, Theorem 5.6.3, p. 123]. 

To continue, assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Retaining earlier notation used in such a setting, whenever 0 < p ≤ ∞
p
define the quasi-Banach space w (D∗ (Σ)) (where the purely decorative subscript “w"
is meant to suggest that this is a weighted space) as

p
w (D∗ (Σ)) := {λQ }Q ∈D∗ (Σ) : λQ ∈ R for every Q ∈ D∗ (Σ),
 
such that {λQ }Q ∈D∗ (Σ)  p (D∗ (Σ)) < ∞ , (7.5.19)
w

where
  p1
    p
{λQ }Q ∈D  p−2
1 1
∗ (Σ) p
w (D∗ (Σ))
:= σ(Q) |λQ | . (7.5.20)
Q ∈D∗ (Σ)

p
Remark 7.5.7 If 0 < p ≤ ∞, a moment’s reflection shows that the space w (D∗ (Σ))
defined
 above may
 be naturally identified with the weighted Lebesgue space
L p D∗ (Σ), ω p m where m denotes the standard counting measure on the count-
able set D∗ (Σ) and

ω p is the weight defined on D∗ (Σ) by

p (7.5.21)
ω p (Q) := σ(Q)1− 2 for every Q ∈ D∗ (Σ).

Proposition 7.5.8 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Suppose 1 ≤ p0, p1 ≤ ∞, 1 ≤ q0, q1 ≤ ∞, and s0, s1 ∈ R. Then the
following complex interpolation result holds:
   q  
q p p q p 
s00 ZΣ, w0 (D∗ (Σ)) , s11 ZΣ, w1 (D∗ (Σ)) = s ZΣ, w (D∗ (Σ)) , (7.5.22)
θ
 1−θ 
θ −1
 1−θ 
θ −1
whenever θ ∈ (0, 1), p := p0 + p1 , q := q0 + q1 , and s := (1 − θ)s0 + θs1 .
p p
Proof Since w0 (D∗ (Σ)) and w1 (D∗ (Σ)) are compatible Banach spaces, by Proposi-
tion 7.5.6 it suffices to show that
 p0 p  p
w (D∗ (Σ)), w1 (D∗ (Σ)) θ = w (D∗ (Σ)). (7.5.23)

This, however, is a consequence of Proposition 7.5.3 (presently employed with Σ

replaced by D∗ (Σ)
 equipped with the counting measure m) and the identification
p
w (D∗ (Σ)) = L p D∗ (Σ), ω p m valid for all p’s (cf. Remark 7.5.7 where the weight
ω p is also defined), together with the observation that
398 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
  (1−θ) pp   θ pp  p0  (1−θ) p  p1  θ p
ω p0 (Q) 0 ω p1 (Q) 1 = σ(Q)1− 2 p0
σ(Q)1− 2 p1
p
= σ(Q)1− 2 = ω p (Q), (7.5.24)

for each cube Q ∈ D∗ (Σ). 

Proposition 7.5.9 Let Σ ⊆ Rn be a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Given q ∈ (0, ∞) along with s ∈ R, consider the weighted space
L q (D∗ (Σ), ws,q m) where m is the counting measure on D∗ (Σ) and the weight ws,q is
defined by   s
ws,q (Q) := σ(Q)−q n−1 + 2
1
, ∀Q ∈ D∗ (Σ). (7.5.25)
For each integrability exponent p ∈ (0, ∞] consider the vector-valued Lebesgue
space L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m , consisting of all p-th power integrable func-
tions on Σ with respect to the measure σ, taking values in the quasi-Banach space
L q (D∗ (Σ), ws,q m) (see (1.4.12)-(1.4.14)).
Then, if θ ∈ (0, 1), 1 ≤ p0, p1 ≤ ∞, 1 ≤ q0, q1 < ∞, and s0, s1 ∈ R, the following
complex interpolation result (involving Banach spaces) holds:
    
L p0 (Σ, σ) ⊗ L q0 D∗ (Σ), ws0,q0 m , L p1 (Σ, σ) ⊗ L q1 D∗ (Σ), ws1,q1 m
θ
q 
= L (Σ, σ) ⊗ L D∗ (Σ), ws,q m ,
p
(7.5.26)
 1−θ  −1  1−θ  −1
where s := (1−θ)s0 +θs1 and p, q are defined by p := p0 + pθ1 , q := q0 + qθ1 .

Proof The justification of (7.5.26) is analogous to the argument used in the proof
of Proposition 7.5.8. 

Lemma 7.5.10 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Assuming 0 < p, q ≤ ∞ and s ∈ R, consider the mapping
q p  p,q
R b : s ZΣ, w (D∗ (Σ)) −→ bs (Σ) (7.5.27)

defined by
R b ({ fk }k ∈ZΣ ) := {λQ }Q ∈D∗ (Σ) for each
q 
p
(7.5.28)
{ fk }k ∈ZΣ ∈ s ZΣ, w (D∗ (Σ)) ,
where
 for each Q
λQ := fk (Q)  ∈ D∗ (Σ) where the integer k ∈ ZΣ is uniquely
(7.5.29)
 ∈ Dk+jΣ (Σ).
determined (thanks to (7.2.2)) by the requirement that Q

Furthermore, introduce a second mapping,

p,q q p 
E b : bs (Σ) −→ s ZΣ, w (D∗ (Σ)) which acts according to
p,q
(7.5.30)
E b ({λQ }Q ∈D∗ (Σ) ) := { fk }k ∈ZΣ for every {λQ }Q ∈D∗ (Σ) ∈ bs (Σ),
7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces 399
p,q
where, given any {λQ }Q ∈D∗ (Σ) ∈ bs (Σ) and k ∈ ZΣ , the function fk : D∗ (Σ) → R
is defined by

λQ if Q ∈ Dk+jΣ (Σ),
fk (Q) := ∀Q ∈ D∗ (Σ). (7.5.31)
0 if Q ∈ D∗ (Σ) \ Dk+jΣ (Σ),

Then the mappings (7.5.27), (7.5.30) are well-defined, linear, bounded and
p,q
R b ◦ E b = I, the identity operator on bs (Σ). (7.5.32)

In particular, the mapping R b in (7.5.27) is onto, i.e.,

  
q p p,q
R b s ZΣ, w (D∗ (Σ)) = bs (Σ). (7.5.33)

q p 
Proof Given { fk }k ∈ZΣ ∈ s ZΣ, w (D∗ (Σ)) arbitrary, define the numerical sequence
{λQ }Q ∈D∗ (Σ) as in (7.5.29). Based on definitions (cf. (7.2.28), (7.5.29), (7.2.2),
(7.5.16), and (7.5.20)), we may then write

⎧
⎪ ⎫ 1/q
⎨
⎪ ∞     p  q/p ⎪
⎬
⎪
p−2
1 1
R b ({ fk }k ∈ZΣ )bsp, q (Σ) = 2 σ(Q)
ks
|λQ |
⎪
⎪ k=κΣ ⎪
⎪
⎩ Q ∈Dk (Σ) ⎭
⎧
⎪ ⎫ 1/q
⎨
⎪ ∞     p  q/p ⎪
⎬
⎪
p−2
1 1
= 2ks σ(Q) | fk (Q)|
⎪
⎪ k=κΣ ⎪
⎪
⎩ Q ∈Dk+ jΣ (Σ) ⎭
⎧
⎪ ⎫ 1/q
⎨
⎪ ∞  
  p  q/p ⎪
⎬
⎪
p−2
1 1
≤ 2 σ(Q)
ks
| fk (Q)|
⎪
⎪ k=κΣ ⎪
⎪
⎩ Q ∈D∗ (Σ) ⎭
= { fk }k ∈ZΣ  q Z p
, (7.5.34)
s Σ,w (D∗ (Σ))

which proves that the mapping R b is well defined and bounded. Clearly, the map-
p,q
ping R b is also linear. Assume next that {λQ }Q ∈D∗ (Σ) ∈ bs (Σ) is arbitrary and
define { fk }k ∈ZΣ as in (7.5.30). Then, by unraveling definitions (cf. (7.5.16), (7.5.20),
(7.5.31), (7.2.2), and (7.2.28)) we may write
400 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

E b ({λQ }Q ∈D∗ (Σ) ) q Z p


s Σ,w (D∗ (Σ))

⎧
⎪ ⎫ 1/q
⎨
⎪ ∞     p  1/p  q ⎪
⎬
⎪
p−2
1 1
= 2 ks
σ(Q) | fk (Q)|
⎪
⎪ k=κΣ ⎪
⎪
⎩ Q ∈D∗ (Σ) ⎭
⎧
⎪ ⎫ 1/q
⎨
⎪ ∞     p  q/p ⎪
⎬
⎪
p−2
1 1
= 2ks σ(Q) |λQ |
⎪
⎪ k=κΣ ⎪
⎪
⎩ Q ∈Dk+ jΣ (Σ) ⎭
⎧
⎨
⎪ ∞   N (k,τ)
 ks  p  q/p ⎫
⎬
⎪
1/q
k,ν p1 − 12
= 2 σ(Qτ ) |λQ k,ν |
⎪ τ ⎪
⎩ k=κΣ τ ∈Ik ν=1 ⎭
= {λQ }Q ∈D∗ (Σ) bsp, q (Σ), (7.5.35)

which shows that the mapping E b is also well-defined and isometric (hence, in
particular, bounded).
To justify the formula claimed in (7.5.32), let {λQ }Q ∈D∗ (Σ) ∈ D∗ (Σ) be an arbitrary
fixed sequence and consider { fk }k ∈ZΣ := E b ({λQ }Q ∈D∗ (Σ) ) defined as in (7.5.30).
Then, if we further set { λ Q }Q ∈D∗ (Σ) := R b ({ fk }k ∈ZΣ ) as defined in (7.5.29), it
follows that for each arbitrary, fixed cube Q  ∈ D∗ (Σ), and for each k ∈ ZΣ such that
  
Q ∈ Dk+jΣ (Σ) we have λQ = fk (Q) = λQ . This establishes that R b ◦ E b = I on
p,q
bs (Σ). 

Lemma 7.5.11 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Assume 0 < p, q < ∞, s ∈ R, and recall the weighted counted
measure ws,q m on D∗ (Σ) introduced before (cf. (7.5.25)). In this context, define the
mapping
p,q  
E f : fs (Σ) −→ L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m which acts according to
p,q
E f ({λQ }Q ∈D∗ (Σ) ) := g for every {λQ }Q ∈D∗ (Σ) ∈ fs (Σ)
(7.5.36)
p,q
where, given any {λQ }Q ∈D∗ (Σ) ∈ fs (Σ), the function g is defined by

(g(x))(Q) := λQ 1Q for every x ∈ Σ and Q ∈ D∗ (Σ). (7.5.37)

Then for the full range of indices p, q, s as above, the mapping (7.5.36) is well
p,q
defined, linear, and isometric, i.e., for each {λQ }Q ∈D∗ (Σ) ∈ fs (Σ) one has
   
E f ({λQ }Q ∈D (Σ) ) p = {λQ }Q ∈D∗ (Σ)  f p, q (Σ), (7.5.38)
∗ L (Σ,σ)⊗L q (D∗ (Σ),ws, q m) s

Going further, assume 1 < p, q < ∞, s ∈ R and, in this context, introduce a

second mapping, namely
  p,q
R f : L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m −→ fs (Σ), (7.5.39)
7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces 401

by setting
 
R f (g) := {λQ }Q ∈D∗ (Σ) for each g ∈ L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m , (7.5.40)

where ⨏
λQ := (g(x))(Q) dσ(x) for each Q ∈ D∗ (Σ). (7.5.41)
Q

Then, for the specified range of indices, the mapping (7.5.39) is well defined,
linear, bounded and
p,q
R f ◦ E f = I, the identity operator on fs (Σ). (7.5.42)

In particular, whenever 1 < p, q < ∞, s ∈ R, the mapping R f in (7.5.39) is onto,

i.e.,    p,q
R f L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m = fs (Σ). (7.5.43)

Proof The claims made about the operator E f , including (7.5.38), are straightfor-
ward from definitions (bearing in mind (1.4.14) and Lemma 1.4.11). Moving on,
assume that 1 < p, q < ∞, s ∈ R. In this setting, the crux of the matter is showing
that R f is a well-defined and  bounded operator.
 With this goal in mind, pick an
arbitrary g ∈ L p (Σ, σ) ⊗ L q D∗ (Σ), ws,q m and introduce
s
gQ (x) := σ(Q)− n−1 − 2 (g(x))(Q),
1
∀Q ∈ D∗ (Σ), ∀x ∈ Σ. (7.5.44)

Also, recall that M Σ denotes the Hardy-Littlewood maximal operator on Σ (cf.

(A.0.71)). Then (with justifications to follow shortly) we may write
 
   ⨏ q 
     1/q 
R f (g) p, q ≈   g dσ  1 
fs (Σ)  Q Q 
 Q ∈D∗ (Σ) Q  p
L (Σ,σ)
 
  
   1/q 
≤ C  M Σ (gQ )q 

 Q ∈D∗ (Σ) 
L p (Σ,σ)
 
  
  q 1/q 
≤ C  gQ  

 Q ∈D∗ (Σ) 
L p (Σ,σ)
 
   s 1 
 1/q 
= C  σ(Q)−q n−1 + 2 |(g(·))(Q)| q 

 Q ∈D∗ (Σ) 
L p (Σ,σ)

= Cg L p (Σ,σ)⊗L q (D∗ (Σ),ws, q m)), (7.5.45)

for some constant C ∈ (0, ∞) independent of g. Above, the first step is a matter of
unraveling definitions and observing that
402 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
q
ws,q (Q) ≈ 2ksq σ(Q)− 2 , uniformly for Q ∈ Dk+jΣ (Σ), k ∈ ZΣ . (7.5.46)

To justify the second step, it suffices to note that, generally speaking, there exists some
constant C = C(Σ) ∈ (0, ∞) with the property that if x∗ ∈ Σ, r > 0, Δ := B(x∗, r) ∩ Σ,
and h ∈ Lloc1 (Σ, σ) then

⨏ 
|h| dσ 1Δ ≤ CM Σ h pointwise on Σ. (7.5.47)
Δ

The third step uses the version of a classical inequality due to Fefferman-Stein
regarding the boundedness of the vector-valued Hardy-Littlewood maximal operator
on spaces of homogeneous type (cf. [133, Theorem 7.6.6] for a precise formulation).
This is where we make crucial use of the fact that 1 < p, q < ∞. Steps four and five
are mere reinterpretations of definitions. This finishes the proof of the fact that R f
is a well-defined, linear and bounded operator.
Finally, formula claimed in (7.5.42) follows directly from (7.5.36)-(7.5.37) and
(7.5.39)-(7.5.41). 

The reader is reminded that the property being analytically convex has been
defined in §1.4.

Proposition 7.5.12 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ.
(i) If

s ∈ (−1, 1), max n , n+s
n−1 n−1
< p ≤ ∞, 0 < q ≤ ∞, (7.5.48)

p,q
then the Besov space Bs (Σ, σ) is analytically convex.
(ii) If 
s ∈ (−1, 1), max n , n+s
n−1 n−1
< p < ∞,
 (7.5.49)
n , n+s < q ≤ ∞,
n−1 n−1
and max
p,q
then the Triebel-Lizorkin space Fs (Σ, σ) is analytically convex.
(iii) Similar results are valid for the associated sequence spaces. More precisely,
p,q p,q
bs (Σ) is analytically convex whenever (7.5.48) holds, while fs (Σ) is ana-
lytically convex if (7.5.49) holds. Finally, when Σ is unbounded, similar results
hold for the homogeneous scales of Besov and Triebel-Lizorkin spaces on Σ.

Proof Thanks to Proposition 7.3.5 and Proposition 1.4.6, it suffices to work with
sequence spaces. In this setting, the conclusion pertaining to Besov spaces follows
from the fact that the operator (7.5.30) is a linear isomorphism onto its image (see
(7.5.35)), Proposition 1.4.6, and repeated applications of Proposition 1.4.15.
The argument for the scale of discrete Triebel-Lizorkin spaces largely follows
along these lines, making use of the first part of Lemma 7.5.11 (cf. (7.5.36)-(7.5.38)
7.5 Complex Interpolation of Besov and Triebel-Lizorkin Spaces 403

which imply that E f is a linear isomorphism onto its image), then once again invoking
Proposition 1.4.6 together with Proposition 1.4.15. 
Lemma 7.5.13 Let Σ ⊆ Rn be a closed Ahlfors regular set. Fix s ∈ R, 0 < p, q ≤ ∞,
p,q p,q
and assume that 0 < r < min{p, q}. Then the spaces fs (Σ) and bs (Σ) are lattice
r-convex. Moreover,
p,q p ,q p,q p ,q
[ fs (Σ)]r = fs (Σ), [bs (Σ)]r = bs (Σ), (7.5.50)

where the indices are related by

p := p/r, q := q/r, s := s + (n − 1)(r − 1)/(2r). (7.5.51)

Proof The first claim in the lemma is a consequence of Proposition 7.5.12 and
Theorem 1.4.17. Next, it follows straight from definitions that for each r > 0 we
have
p,q p ,q
{|λQ | 1/r }Q ∈D∗ (Σ) ∈ fs (Σ) ⇐⇒ {λQ }Q ∈D∗ (Σ) ∈ fs (Σ), (7.5.52)

where the primed indices are as in (7.5.51) (incidentally, this can also be used to
p,q
show that fs (Σ) is lattice r-convex if 0 < r < min{p, q} will do). The formula in
p,q
(7.5.50) dealing with the scale fs (Σ) is clear from this. The treatment of the scale
p,q
bs (Σ) is similar, finishing the proof of the lemma. 
After these preparations, we are finally ready to present the proof of Theo-
rem 7.5.2.
Proof of Theorem 7.5.2 We shall first focus on proving the claims made in Theo-
rem 7.5.2 for the scale of Besov spaces. Fix pi, qi, p, q, for i ∈ {0, 1}, as in (7.5.6)
and choose β, γ ∈ (0, 1) sufficiently close 1 (depending on the aforementioned pa-
β,γ
rameters). Then we may take (G0 (Σ))∗ as an ambient topological vector space
pi ,qi
in which the Besov spaces Bsi (Σ, σ), i ∈ {0, 1}, are continuously embedded. In
p ,q
addition, from Proposition 7.5.12 we know that the spaces Bsii i (Σ, σ), i ∈ {0, 1},
are analytically convex (cf. also (7.5.56) below), so the version of the method of
complex interpolation described in §1.4 applies.
The rest of the proof is organized in a series of steps, each of which deals with a
separate claim, building up to the desired conclusion.
Step 1. The sum
p ,q0 p ,q1
bs00 (Σ) + bs11 (Σ) is an analytically convex space. (7.5.53)

Proof. From item (iii) in Proposition 7.5.12 we know that

p ,qi
each space bsii (Σ), i ∈ {0, 1}, is analytically convex, (7.5.54)

whereas from definitions it is apparent that, generally speaking,

p,q
bs (Σ) is separable if 0 < p < ∞, 0 < q < ∞, and s ∈ R. (7.5.55)
404 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

In concert with Theorem 1.4.18 and the second remark following it, this proves
(7.5.53).
Step 2. The sum
p ,q0 p ,q1
Bs00 (Σ, σ) + Bs11 (Σ, σ) is an analytically convex space. (7.5.56)

Proof. Recall the mappings Φ and Ψ defined in Proposition 7.3.2 and Propo-
sition 7.3.3, respectively. One may then check that the operator Ψ maps the
p ,q p ,q
sum space Bs00 0 (Σ, σ) + Bs11 1 (Σ, σ) isomorphically onto a closed subspace of
p0,q0 p1,q1
bs0 (Σ) + bs1 (Σ). Indeed, this follows readily by taking into account that Ψ acts
β,γ
linearly on (G0 (Σ))∗ and that, as already pointed out, formula (7.3.32) continues
to hold in this setting. Then (7.5.56) follows from this observation, (7.5.53) and
Proposition 1.4.6.
p ,q
Step 3. The pair of continuous Besov spaces Bsii i (Σ), i ∈ {0, 1}, is a retract of
p ,q
the pair of discrete Besov spaces bsii i (Σ), i ∈ {0, 1}.
Proof. This follows from Proposition 7.3.5, with Φ, Ψ doing the job.
Step 4. For Φ defined in Proposition 7.3.2, there holds
 p ,q0 p ,q1   p ,q p ,q  
Bs00 (Σ, σ), Bs11 (Σ, σ) θ
= Φ bs00 0 (Σ), bs11 1 (Σ) θ . (7.5.57)

Proof. This is a consequence of the first formula in (1.4.64).

Step 5. Formula (7.5.8) holds provided
 p0,q0 p ,q  p,q
bs0 (Σ), bs11 1 (Σ) θ = bs (Σ). (7.5.58)

p,q p,q
Proof. This follows from (7.5.57) and the fact that Φ maps bs (Σ) onto Bs (Σ, σ)
(cf. the first part of (7.3.33)).
Step 6. With the mapping R b defined as in (7.5.27)-(7.5.29), there holds
 p0,q0 p ,q 
bs0 (Σ), bs11 1 (Σ) θ
   q   
q p p
= R b s00 ZΣ, w0 (D∗ (Σ)) , s11 ZΣ, w1 (D∗ (Σ)) . (7.5.59)
θ

Proof. This follows from Lemma 7.5.10 and (1.4.64) in Lemma 1.4.23.
Step 7. One has
 p ,q p ,q  p,q
p0, q0, p1, q1 ≥ 1 =⇒ bs00 0 (Σ), bs11 1 (Σ) θ = bs (Σ). (7.5.60)

Proof. This is a consequence of formula  (7.5.59), the interpolation result proved in

q p  p,q
Proposition 7.5.8, and the fact that R b s ZΣ, w (D∗ (Σ)) = bs (Σ) (cf. (7.5.33)).
7.6 Duality Results for Besov and Triebel-Lizorkin Spaces 405

Step 8. Under the conditions on the indices stipulated in the (first part of the)
statement of Theorem 7.5.2, formula (7.5.58) holds.
Proof. Suppose that 0 < r < min{p0, q0, p1, q1 }. We may then write the following
string of equalities (which are individually justified shortly):
   r   p ,q  1−θ  p1,q1  θ  r
p ,q p ,q
bs00 0 (Σ), bs11 1 (Σ) θ = bs00 0 (Σ) bs1 (Σ)

 p ,q0  r  1−θ  p ,q1 r θ

= bs00 (Σ) bs11 (Σ)
 p /r,q0 /r  1−θ  p1 /r,q1 /r θ
= bs 0+(n−1)(r−1)/(2r) (Σ) bs +(n−1)(r−1)/(2r) (Σ)
0 1
 p /r,q0 /r p /r,q1 /r 
= bs 0+(n−1)(r−1)/(2r) (Σ), bs 1+(n−1)(r−1)/(2r) (Σ) θ
0 1

p/r,q/r
= bs+(n−1)(r−1)/(2r) (Σ)
 p,q  r
= bs (Σ) . (7.5.61)

The first and fourth equalities in (7.5.61) are based on Theorem 1.4.18, the second
and sixth equalities follow from formula (1.4.46), the third equality is a consequence
of (7.5.50), while in the fifth equality we made use of the result proved in Step 7 and
the fact that pi /r > 1 and qi /r> 1 for i ∈ {0, 1}.
 p ,q p ,q  r  p,q  r
In summary, we have that bs00 0 (Σ), bs11 1 (Σ) θ = bs (Σ) , so the claim
made in Step 8 follows from this and the comment in Remark 1.4.16.
Step 9. Formula (7.5.8) holds under the conditions on the indices stipulated in the
(first part of the) statement of Theorem 7.5.2.
Proof. This follows from the results proved in Step 5 and Step 8.
This finishes the proof of the portion of Theorem 7.5.2 dealing with the scale
of Besov spaces. The corresponding statements made for the scale of Triebel-
Lizorkin spaces are proved analogously, making only natural changes in the above
scheme (most notably, we use Proposition 7.5.9 in place of Proposition 7.5.8, and
Lemma 7.5.11 in place of Lemma 7.5.10). The proof of Theorem 7.5.2 is therefore
complete. 

7.6 Duality Results for Besov and Triebel-Lizorkin Spaces

To discuss the duality theory of Besov and Triebel-Lizorkin spaces defined on Ahlfors
regular sets, for any given p ∈ (0, ∞) define its Hölder conjugate exponent p as
 p
if 1 < p < ∞,
p := p−1 (7.6.1)
∞ if 0 < p ≤ 1.
406 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Proposition 7.6.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix s ∈ (−1, 1).
p,q p ,q
(i) If 1 ≤ p < ∞, 0 < q < ∞, then the dual space of Bs (Σ, σ) is B−s (Σ, σ),
where p and q are the Hölder conjugate exponents of p and q, respectively
p ,q
 ∗ any given g ∈ B−s (Σ, σ) may be
(cf. (7.6.1)). More precisely, inone direction,
p,q
identified with a functional in Bs (Σ, σ) in the following fashion. First, if
 1 
(−s)+ < β < 1 and max − s − n−1 p , s + (n − 1) p − 1 < γ < 1, (7.6.2)

then
β,γ p,q !
G0 (Σ) ∩ Bs (Σ, σ)  f −→ Λg ( f ) := G0
β,γ
(Σ)
f, g β,γ
(G0 (Σ))∗
(7.6.3)

is an assignment which satisfies

|Λg ( f )| ≤ C f Bsp, q (Σ,σ) gB p , q (Σ,σ)

−s
β,γ p,q p ,q
(7.6.4)
for each f ∈ G0 (Σ) ∩ Bs (Σ, σ) and g ∈ B−s (Σ, σ),

for some constant C ∈ (0, ∞) which is independent of f , g. Granted this, the

mapping (7.6.3) may be then extended, on account of Lemma 7.1.10, to a linear
p,q ∗
functional Λ ∈ Bs (Σ, σ) satisfying Λ(Bsp, q (Σ,σ))∗ ≤ CgB p , q (Σ,σ) and
−s
one defines
!
B s (Σ,σ) f , g (B s (Σ,σ))∗ := Λ( f )
p, q p, q

p,q p ,q
(7.6.5)
for each f ∈ Bs (Σ, σ) and g ∈ B−s (Σ, σ).

In the opposite direction, there exists some constant C ∈ (0, ∞) with the property
p,q
that if Λ is any given linear functional on Bs (Σ, σ) then there exists a unique
p ,q
g ∈ B−s (Σ, σ) with gB p , q (Σ,σ) ≤ CΛ and such that Λ extends Λg (defined
−s
above).
(ii) If 1 < p, q < ∞, or p = 1 and max n−1 n , n+s < q < ∞, then the dual space of
n−1
p,q p ,q
Fs (Σ, σ) is F−s (Σ, σ), where p and q are the Hölder conjugate exponents
of p and q, respectively (cf. (7.6.1)). More specifically, in one direction, any
p ,q  p,q ∗
given g ∈ F−s (Σ, σ) may be identified with a functional in Fs (Σ, σ) in the
following manner. First, if β, γ are as in (7.6.2), it turns out that the assignment
β,γ p,q !
G0 (Σ) ∩ Fs (Σ, σ)  f −→ Λg ( f ) := G β,γ (Σ) f , g (G β,γ (Σ))∗ (7.6.6)
0 0

satisfies

|Λg ( f )| ≤ C f Fsp, q (Σ,σ) gF p , q (Σ,σ)

−s
β,γ p,q p ,q
(7.6.7)
for each f ∈ G0 (Σ) ∩ Fs (Σ, σ) and g ∈ F−s (Σ, σ),
7.6 Duality Results for Besov and Triebel-Lizorkin Spaces 407

for some C ∈ (0, ∞) independent of f , g. Hence, this mapping

 ∗ extends,
p,q
thanks to Lemma 7.1.10, to a linear functional Λ ∈ Fs (Σ, σ) satisfying
Λ(Fsp, q (Σ,σ))∗ ≤ CgF p , q (Σ,σ) and one defines
−s

!
p, q
Fs (Σ,σ) f, g p, q
(Fs (Σ,σ))∗ := Λ( f )
p,q p ,q
(7.6.8)
for each f ∈ Fs (Σ, σ) and g ∈ F−s (Σ, σ).

In the opposite direction, there exists C ∈ (0, ∞) with the property that if Λ is a
p,q p ,q
linear functional on Fs (Σ, σ) then there exists a unique g ∈ F−s (Σ, σ) with
gF p , q (Σ,σ) ≤ CΛ and such that Λ extends Λg .
−s

Proof This is immediate from [89, Theorem 8.18, p. 246]. See also [91, Lemma 1.8,
p. 18] or [86, Theorem 5.1, p. 64] for the case when p, q > 1. 

In the context of Proposition 7.6.1, the manner in which Besov and Triebel-
Lizorkin spaces embed into distributions (cf. Lemma 7.1.7) implies
! !
B s (Σ,σ) f , g B−s = Lip (Σ,σ) f , g (Lip (Σ,σ))
p, q p ,q
(Σ,σ)
p ,q
(7.6.9)
whenever f ∈ Lip (Σ, σ) and g ∈ B−s (Σ, σ),

and ! !
p, q
Fs (Σ,σ) f, g p ,q
F−s (Σ,σ)
= Lip (Σ,σ) f , g (Lip (Σ,σ))
p ,q
(7.6.10)
whenever f ∈ Lip (Σ, σ) and g ∈ F−s (Σ, σ).
It is significant to note that the duality pairings from Proposition 7.6.1 are actually
compatible with one another. More precisely, we have the following result.

Proposition 7.6.2 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ.
(i) Suppose 1 ≤ p0, p1 < ∞, 0 < q0, q1 < ∞, s0, s1 ∈ (−1, 1), and denote by
p0, p1, q0, q1 the Hölder conjugate exponents of p0, p1, q0, q1 (cf. (7.6.1)). Then
one has
! !
B s 0 0 (Σ,σ) f , g = Bsp1, q1 (Σ,σ) f , g p1, q1
p ,q p ,q
0 0 0
B−s0 (Σ,σ) 1 B−s1 (Σ,σ)
p ,q0 p ,q1
for any f ∈ Bs00 (Σ, σ) ∩ Bs11 (Σ, σ) (7.6.11)
p ,q p ,q
and g ∈ B−s00 0 (Σ, σ) ∩ B−s11 1 (Σ, σ).

(ii) Fix two smoothness exponents s0, s1 ∈ (−1, 1) and suppose that there holds
either 1 < p0, p1, q0, q1 < ∞, or p j = 1 and max n−1 n , n+s j
n−1
< q j < ∞ for
j ∈ {0, 1}. Once again, denote by p0, p1, q0, q1 the Hölder conjugate exponents
of p0, p1, q0, q1 (cf. (7.6.1)). Then one has
408 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
! !
p , q0
Fs00 (Σ,σ) f, g p ,q = Fsp1, q1 (Σ,σ) f , g p ,q
1 (Σ,σ)
F−s00 0 (Σ,σ) 1 F−s11
p0,q0 p ,q
for any f ∈ F s0 (Σ, σ) ∩ Fs11 1 (Σ, σ) (7.6.12)
p ,q p ,q
and g ∈ F−s00 0 (Σ, σ) ∩ F−s11 1 (Σ, σ).

Proof Fix β, γ ∈ (0, 1) sufficiently close to 1 so that the inequalities in (7.6.2) remain
valid both with p, s replaced by p0, s0 and with p, s replaced by p1, s1 . Also, recall
the family of operators PN , with N ∈ N, defined as in (7.3.37). Then for each
p ,q p ,q p ,q p ,q
f ∈ Bs00 0 (Σ, σ) ∩ Bs11 1 (Σ, σ) and each g ∈ B−s00 0 (Σ, σ) ∩ B−s11 1 (Σ, σ) we may
write
! !
B s 0 0 (Σ,σ) f , g = lim Bsp0, q0 (Σ,σ) PN f , g p0, q0
p ,q p ,q
0 0
B−s0 (Σ,σ) N →∞ B−s0 (Σ,σ)
!
= lim β,γ PN f , g β,γ
(Σ))∗
N →∞ G0 (Σ) (G0
!
= lim 1 p , q1
(Σ,σ) PN f , g p ,q
N →∞ B s1 B−s11 1 (Σ,σ)

!
= p , q1
Bs 1 (Σ,σ) f, g p ,q
1 (Σ,σ)
, (7.6.13)
B−s11

thanks to Proposition 7.6.1 and (7.3.36). This establishes (7.6.11). The claim in item
(ii) is dealt with analogously. 

7.7 Loose and Tight Embeddings

We begin by reviewing what we refer to as loose embeddings, i.e., continuous

inclusions within the Besov and Triebel-Lizorkin scales in which the smoothness
and integrability exponents involved are related via inequalities. Here is the first such
result.

Proposition 7.7.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, let s ∈ (−1, 1). Then the following are true.

n , n+s < p ≤ ∞ and 0 < q0 ≤ q1 ≤ ∞, one has

n−1 n−1
(i) If max
p,q0 p,q1
Bs (Σ, σ) → Bs (Σ, σ). (7.7.1)

n , n+s < p < ∞ and max n , n+s < q0 ≤ q1 ≤ ∞, then

n−1 n−1 n−1 n−1
(ii) If max
p,q0 p,q1
Fs (Σ, σ) → Fs (Σ, σ). (7.7.2)

(iii) If θ ∈ (0, 1 − s), n−1

n < p ≤ ∞ and 0 < q0, q1 ≤ ∞, then
p,q p,q1
Bs+θ0 (Σ, σ) → Bs (Σ, σ). (7.7.3)
7.7 Loose and Tight Embeddings 409

(iv) If θ ∈ (0, 1 − s), max n , n+s

n−1 n−1
< p < ∞ and max n , n+s
n−1 n−1
< q0, q1 ≤ ∞,
then
p,q p,q1
Fs+θ0 (Σ, σ) → Fs (Σ, σ). (7.7.4)

n , n+s < p < ∞ and max n , n+s < q ≤ ∞,

n−1 n−1 n−1 n−1
(v) If max
p,min{p,q } p,q p,max{p,q }
Bs (Σ, σ) → Fs (Σ, σ) → Bs (Σ, σ). (7.7.5)

Proof See [89, Proposition 5.31, p. 140] (cf. also [91, Proposition 5.1, p. 84] in the
case when 1 ≤ p ≤ ∞ and 1 ≤ q ≤ ∞). 
The next proposition describes loose embedding which happen to be compact,
assuming the underlying closed Ahlfors regular set is bounded.
Proposition 7.7.2 Assume that Σ ⊆ Rn is a compact Ahlfors regular set and abbre-
viate σ := H n−1 Σ. Then the following compact embeddings hold.
(i) Whenever −1 < s2 < s1 < 1, max n−1 , n−1i < pi ≤ ∞ and 0 < qi ≤ ∞ for
 1 n 1n+s

i ∈ {1, 2}, and also s1 − s2 − (n − 1) p1 − p2 + > 0, it follows that the embedding
p ,q1 p ,q2
Bs11 (Σ, σ) → Bs22 (Σ, σ) is compact. (7.7.6)

(ii) If −1 < s2 < s1 < 1, max n−1 n , n+si < pi < ∞, max
n−1
n , n+si < qi ≤ ∞ for
n−1 n−1

i ∈ {1, 2} and s1 − s2 − (n − 1) p1 − p2 + > 0, then the embedding

1 1

p ,q1 p ,q2
Fs11 (Σ) → Fs22 (Σ, σ) is compact. (7.7.7)

Proof See [199, Theorem 3.1, p. 586] (cf. also [91, Proposition 5.1, p.84] in the case
1 ≤ p ≤ ∞ and 1 ≤ q ≤ ∞). 
As a corollary of (7.7.6), (7.7.4), (7.1.55), and (7.1.38), we see that
∗ p,q
L p (Σ, σ) → B−s (Σ, σ) compactly (hence also continuously)
if Σ ⊆ Rn is a compact Ahlfors regular set, σ := H n−1 Σ, and (7.7.8)
 
1 < p∗ < ∞, n+s < p ≤ ∞, 0 < q ≤ ∞, (n − 1) p∗ − p + < s < 1.
n−1 1 1

Likewise, (7.7.6), (7.7.4), (7.1.55), and (7.1.38) imply that

p,q ∗
Bs (Σ, σ) → L p (Σ, σ) compactly (hence also continuously)
if Σ ⊆ Rn is a compact Ahlfors regular set, σ := H n−1 Σ, and (7.7.9)
 
1 < p∗ < ∞, n−1 n < p ≤ ∞, 0 < q ≤ ∞, (n − 1) p1 − p1∗ + < s < 1.

We next turn to “tight” embeddings, i.e., continuous inclusions within the Besov
and Triebel-Lizorkin scales in which the smoothness and integrability exponents
involved are related via precise algebraic equalities (among other things). First, we
record the following result of this flavor.
410 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Proposition 7.7.3 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then the following embeddings hold.
(i) Whenever −1 < s2 < s1 < 1, max n−1 n , n+si < pi ≤ ∞ and 0 < qi ≤ ∞ for
n−1

i ∈ {1, 2}, and also −1 < s1 − n−1

p1 = s2 − p2 < 1, it follows that the embedding
n−1

p ,q1 p ,q2
Bs11 (Σ, σ) → Bs22 (Σ, σ) is continuous. (7.7.10)

In addition, (7.7.10) holds whenever 1 ≤ pi = qi ≤ ∞ for i ∈ {1, 2} and

0 < s0 ≤ s1 < 1 are such that s1 − n−1
p1 = s2 − p2 .
n−1

(ii) If −1 < s2 < s1 < 1, max n−1n , n+si < pi < ∞, max
n−1
n , n+si < qi ≤ ∞ for
n−1 n−1

i ∈ {1, 2} and −1 < s1 − p1 = s2 − p2 < 1, then the embedding

n−1 n−1

p ,q1 p ,q2
Fs11 (Σ, σ) → Fs22 (Σ, σ) is continuous. (7.7.11)

Proof See [200, Theorem 2, p. 191]. The version of the embedding in (7.7.10) corre-
sponding to 1 ≤ pi = qi ≤ ∞ for i ∈ {1, 2} and 0 < s0 ≤ s1 < 1 may be seen to hold
thanks to [106, Proposition 5, p. 213] and Corollary 7.9.2. Parenthetically, we also
wish to mention [91, Theorem 5.2, p. 89] as well as [85, Theorem 2, p. 294] for relat-
ed results (valid for more restrictive ranges of indices) in the case of homogeneous
Besov and Triebel-Lizorkin spaces. 

The tight embedding of Besov spaces recorded in (7.7.10) may be established for
other ranges of indices, as discussed in the theorem below.

Theorem 7.7.4 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. If −1 < s0, s1 < 1 and max n−1n , n+si < pi ≤ ∞ for i ∈ {0, 1} are
n−1

such that
s1 s0
p1 − n−1 = p0 − n−1 ,
1 1
(7.7.12)
then the embedding
p ,q0 p ,q1
Bs00 (Σ, σ) → Bs11 (Σ, σ) is continuous (7.7.13)

provided either
0 < q0 ≤ q1 ≤ ∞ and s0 ≥ s1, (7.7.14)
or
0 < q0 < ∞, 0 < q1 ≤ ∞, and s0 > s1 . (7.7.15)

Proof For starters, observe that (7.2.10), (7.2.11), (7.2.27), Theorem 7.2.7, and
Theorem 7.2.8 ensures that (7.7.13) holds in the case when the conditions in (7.7.14)
are satisfied. To deal with the scenario in which the demands in (7.7.15) are assumed,
suppose
7.7 Loose and Tight Embeddings 411

− 1 < s∗ < s < 1, 0 < q∗ < q < ∞, (7.7.16)

n , n+s < p ≤ ∞,
n−1 n−1
max (7.7.17)

n , n+s∗ < p∗ ≤ ∞,
n−1 n−1
max (7.7.18)
s∗
1
p − s
n−1 = 1
p∗ − n−1 . (7.7.19)

Pick q0 ∈ (0, q) then define q1 := q0 . Taking q0 very close to q ensures that we may
select a number θ ∈ (0, 1) satisfying

1−s s∗ +1
1 − qq0 < θ < min 1−s ∗ np , 1 − (n+s)p .
, 2 , 1 − n−1 n−1
(7.7.20)

The hypotheses in (7.7.16)-(7.7.19) guarantee that this is a meaningful demand on

the parameter θ and, given that they additionally entail 0 < p < p∗ ≤ ∞ as well as
p(n + s) < p∗ (n + s∗ ), we also have

θ < min s+12 , 1 − np∗ , 1 − (n+s∗ )p∗ .
n−1 n−1
(7.7.21)

Next, since 1/q∗ > 1/q > (1 − θ)/q0 = (1 − θ)/q1 , we may select q0, q1 > 0 small
enough so that
1 1−θ θ 1 1−θ θ
= + and = + . (7.7.22)
q q0 q0 q∗ q1 q1

Since (7.7.21) forces 1/p∗ < (n + s∗ )(1 − θ)/(n − 1) while (7.7.20) guarantees that
∗ −θ
we also have −1 < s1−θ < s∗ < 1, we may choose some
 s∗ −θ 
s1 ∈ 1−θ , s∗ ⊆ (−1, 1) (7.7.23)

with the property that

1 (n + s1 )(1 − θ)
< . (7.7.24)
p∗ n−1

Pick p1 ∈ (0, ∞) satisfying

 (n − 1)θ (n − 1)θ
p1 > max θp∗, ,   (7.7.25)
θ − s∗ + (1 − θ)s1 2 s∗ − (1 − θ)s1
  −1
then define p1 := (1 − θ) 1/p∗ − θ/ p1 ∈ (0, ∞). Hence,

1 1−θ θ
= + (7.7.26)
p∗ p1 p1

and the choice of θ in (7.7.20) ensures that we have p1 > (n − 1)/n. Moreover, from
(7.7.24) we see that p1 > n+s
n−1
1
hence, ultimately,
412 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

n , n+s1 < p1 < ∞.

n−1 n−1
max (7.7.27)

To proceed, pick some p0 ∈ (0, ∞) which, for now, is assumed to satisfy
 1 1 θ  −1 (n − 1)θ
p1 > p0 > max θ − + , , θp
p p∗ p1 θ − s∗ + (1 − θ)s1 + (n − 1)θ/ p1
(7.7.28)

and remark that, in concert with (7.7.25), the aforementioned choice guarantees that

(n − 1)θ
p0 > . (7.7.29)
θ + s∗ − (1 − θ)s1

Consider next p0 := (1 − θ)(1/p − θ/ p0 )−1 ∈ (0, ∞). The latter definition entails
1 1−θ θ
= + . (7.7.30)
p p0 p0

Also, (7.7.20) forces 1/p < n(1 − θ)/(n − 1) hence

1 n(1 − θ) θ n−1
− <0< =⇒ p0 > . (7.7.31)
p n−1 p0 n
In addition, from (7.7.28) and definitions we see that

0 < p0 < p1 < ∞. (7.7.32)

Pressing on, define

 
s0 := s1 + (n − 1) 1
p0 − 1
p1 . (7.7.33)

s−s∗
As seen from (7.7.20), we have θ < 1−s
1−s∗ which implies s1 < s∗ < 1 − 1−θ hence,
further,
(1 − s1 )(1 − θ) s − s∗ 1 1 1 1 θ θ
> = − > − − +
n−1 n−1 p p∗ p p∗ p0 p1
1 1  (1 − θ)(s0 − s1 )
= (1 − θ) − = , (7.7.34)
p0 p1 n−1
from which we readily conclude that s0 < 1. From this, (7.7.33), (7.7.32), and
(7.7.23) we see that
s0 s1
−1 < s1 < s0 < 1 and 1
p0 − n−1 = 1
p1 − n−1 . (7.7.35)

Also, the fact that p1 > n−1

n+s1 (cf. (7.7.27)) permits us to write

1 n + s1 n s1 n s0 1 1
< = + = + + − (7.7.36)
p1 n−1 n − 1 n − 1 n − 1 n − 1 p1 p0
7.7 Loose and Tight Embeddings 413

s0
hence 1
p0 < n
n−1 + n−1 which gives p0 > n−1
n+s0 . Consequently,

n , n+s0 < p0 < ∞.

n−1 n−1
max (7.7.37)

Let us now define

   
s0 := θ −1 s − (1 − θ)s0 , s1 := θ −1 s∗ − (1 − θ)s1 , (7.7.38)

which entail

s = (1 − θ)s0 + θ s0 and s∗ = (1 − θ)s1 + θ s1 . (7.7.39)

Additionally, (7.7.38), (7.7.19), (7.7.33), and the definitions of p0, p1 give

 
s0 = θ −1 s∗ − (1 − θ)s1 + θ −1 s − s∗ − (1 − θ)(s0 − s1 )
 1 1 1 1
= s1 + θ −1 (n − 1) − − (1 − θ)(n − 1) −
p p∗ p0 p1
 1 1 1 θ 1 θ 
= s1 + θ −1 (n − 1) − − (n − 1) − − +
p p∗ p p0 p∗ p1
θ θ  n−1 n−1
= s1 + θ −1 (n − 1) − = s1 + − . (7.7.40)
p0 p1 p0 p1

Starting with −θs∗ < θ, we may estimate s∗ < (s∗ + θ)/(1 − θ). As such,

s∗ + θ s∗ − (1 − θ)s1
s1 < s ∗ < =⇒ s∗ > (1 − θ)s1 − θ =⇒ s1 = > −1. (7.7.41)
1−θ θ
Also, based on (7.7.28), (7.7.19), the definitions of p0, p1 , (7.7.33), and (7.7.38) we
may write
θ θ
s∗ − (1 − θ)s1 +(n − 1) − (n − 1) < θ
p0 p1
s − s∗  θ θ 
=⇒ s − (1 − θ)s1 − (n − 1) + (n − 1) − <θ
n−1 p0 p1
1 θ 1 θ 
=⇒ s − (1 − θ)s1 − (n − 1) − − + <θ
p p0 p∗ p1
 1 1 
=⇒ s − (1 − θ) s1 + (n − 1) − <θ
p0 p1

=⇒ s − (1 − θ)s0 < θ

s − (1 − θ)s0
=⇒ s0 = < 1. (7.7.42)
θ
414 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Collectively, (7.7.40), (7.7.41), and (7.7.42) prove that

1 s0 1 s1
−1 < s1 < s0 < 1 and − = − . (7.7.43)
p0 n − 1 p1 n − 1
Furthermore, from (7.7.29), (7.7.38), and (7.7.43) we deduce that
n−1 n−1
p0 > > . (7.7.44)
s1 + 1 s0 + 1
In particular,
n−1 n−1
−1 < s0 − = s1 − < 1. (7.7.45)
p0 p1
Finally, (7.7.44) readily gives
n − 1 n − 1
p0 > max , , (7.7.46)
n n + s0
while (7.7.44) and the first inequality in (7.7.28) give
n − 1 n − 1
p1 > max , . (7.7.47)
n n + s1
At this stage, from the version of (7.7.13) corresponding to (7.7.14) (which is
presently applicable, thanks to (7.7.27), (7.7.35), (7.7.37), and bearing in mind that
q0 = q1 ∈ (0, ∞)) we obtain
p ,q0 p ,q1
Bs00 (Σ, σ) → Bs11 (Σ, σ) continuously, (7.7.48)

while from part (i) in Proposition 7.7.3 (which may be invoked here with p1 := p0 ,
q1 := q0 , s1 := s0 , p2 := p1 , q2 := q1 , s2 := s1 , thanks to (7.7.43), (7.7.45), and
(7.7.46)) we see that
)
p ,)
q0 )
p ,)
q1
Bs0 (Σ, σ) → Bs1 (Σ, σ) continuously. (7.7.49)
0 1

In turn, from (7.7.48)-(7.7.49) and (7.5.8) in Theorem 7.5.2 (also bearing in mind
(7.7.39)), we conclude that
p,q  p ,q )
p ,)
q 
Bs (Σ, σ) = Bs00 0 (Σ, σ), Bs0 0 (Σ, σ) θ
0

 p ,q1 )
p ,)
q1 
→ Bs11 (Σ, σ), Bs1 (Σ, σ) θ
1

p ,q∗
= Bs∗∗ (Σ, σ) continuously. (7.7.50)

The conclusion so far is that

7.7 Loose and Tight Embeddings 415
p,q p ,q∗
Bs (Σ, σ) → Bs∗∗ (Σ, σ) continuously
(7.7.51)
whenever s, s∗, p, p∗, q, q∗ are as in (7.7.16)-(7.7.19).

Having established this, we may now eliminate the restriction that q∗ < q (appearing
in (7.7.16)) by invoking item (i) in Proposition 7.7.1. This shows that (7.7.13) holds
in the case when the conditions in (7.7.15) are satisfied. 

From Proposition 7.1.9 we know that local Hardy spaces are contained in the
Triebel-Lizorkin scale (cf. (7.1.56)). Using this, together with the relationship be-
tween the Triebel-Lizorkin scale and Besov scale (cf. (7.7.5)), and Theorem 7.7.4,
we are then able to show that local Hardy spaces embed naturally into Besov spaces,
in the precise manner described below.

Corollary 7.7.5 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then whenever

−1 < s < 1, max n , n+s

n−1 n−1
< p < p∗ ≤ 1,
1 
satisfy − 1 + (n − 1) p − p1∗ < s, and (7.7.52)
max n , n+s
n−1 n−1
< q < ∞, 0 < q∗ ≤ ∞,

one has a well-defined continuous embedding

p,q p ∗,q ∗
Fs (Σ, σ) → B (Σ, σ). (7.7.53)
s−(n−1)( p1 − p1∗ )

As a particular case, corresponding to s := 0 and q := 2, one has the well-defined

continuous embeddings
p ∗,q ∗
H p (Σ, σ) → h p (Σ, σ) → B (Σ, σ) whenever
−(n−1)( p1 − p1∗ )
(7.7.54)
n−1
n < p < p∗ ≤ 1 and 0 < q∗ ≤ ∞.

Proof Assuming (7.7.52), the idea is to write

p,q p,max{p,q } p ∗,q ∗
Fs (Σ, σ) → Bs (Σ, σ) → B (Σ, σ), (7.7.55)
s−(n−1)( p1 − p1∗ )

with the first embedding a consequence of (7.7.5), and the second embedding implied
by Theorem 7.7.4. This proves (7.7.53), then (7.7.54) follows by specializing it to
the case s := 0 and q := 2 (a valid choice in the context of (7.7.52)), bearing in mind
the identification made in (7.1.56) and (7.1.51). 

Having an adequate arsenal of embedding results at our disposal, we may now

revisit the duals of Besov and Triebel-Lizorkin spaces and prove that, most of the
time, these happen to be reasonably rich, even in the case when said spaces are
merely quasi-Banach.
416 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Proposition 7.7.6 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. In addition, assume −1 < s < 1, max n−1 n , n+s < p ≤ ∞, and
n−1

0 < q < ∞.
p,q p,q
Then the dual of Bs (Σ, σ) separates its points2. Moreover, the dual of Fs (Σ, σ)
separates its points whenever max n , n+s < p, q < ∞.
n−1 n−1

Proof We shall first treat the scale of Besov spaces in the scenario when
 1 we  have
max n−1n , n+s < p < 1. The lower bound on p implies that s − (n − 1) p − 1 > −1,
n−1

hence we may choose some exponent p∗ ∈ (1, ∞) with the property that
 
s∗ := s − (n − 1) p1 − p1∗ > −1. (7.7.56)

Granted this, and also picking q∗ ∈ (1, ∞), we may now rely on Theorem 7.7.4 to
conclude that the embedding
p,q p ,q∗
ι : Bs (Σ, σ) → Bs∗∗ (Σ, σ) is continuous. (7.7.57)
p,q
Consider now 0  f ∈ Bs (Σ, σ). The goal is to show that there exists some linear
p,q
and continuous functional  : Bs (Σ, σ) → C with the property that ( f )  0. To
p ,q
justify this, since (7.7.57) ensures that 0  f ∈ Bs∗∗ ∗ (Σ, σ) and since the latter
space is Banach (cf. item (3) in Remark 7.1.6), we may invoke  p ,qthe Hahn-Banach
∗
Theorem to guarantee the existence of some functional Λ ∈ Bs∗∗ ∗ (Σ, σ) with  ∗ the
property that Λ( f )  0. Then  := ΛB p, q (Σ,σ) = Λ ◦ ι belongs to Bs (Σ, σ) (cf.
p,q
s
(7.7.57)) and ( f ) = Λ( f )  0. This finishes the proof of the fact that the dual of
p,q
Bs (Σ, σ) separates points in the case when p is strictly sub-unital. In the case when
1 ≤ p ≤ ∞, take q∗ := max{1, q}, and recall from item (i) of Proposition 7.7.1 that
p,q p,q∗
Bs (Σ, σ) → Bs (Σ, σ). (7.7.58)
p,q
Since Bs ∗ (Σ, σ) is a Banach space, the same type of reasoning as above may be
p,q
employed to conclude that the dual of Bs (Σ, σ) separates points in this case as
well.
In order to deal with the case of Triebel-Lizorkin spaces, assume next that we
have max n−1 n , n+s < p, q < ∞. Then (7.7.5) gives that the embedding
n−1

p,q p,max{p,q }
Fs (Σ, σ) → Bs (Σ, σ) is continuous. (7.7.59)

Since we presently have max{p, q} < ∞, from what we have established in the first
p,max{p,q }
half of the proof we know that the dual of Bs (Σ, σ) separates points. With
this in hand, the same type of reasoning then shows that the same is true for the
p,q
Triebel-Lizorkin space Fs (Σ, σ). 

p, q  p, q ∗
2 in the sense that for any distinct f , g ∈ B s (Σ, σ) there exists a functional Λ ∈ B s (Σ, σ)
with the property that Λ f  Λg
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 417

7.8 Envelopes of Besov and Triebel-Lizorkin Spaces

Recall that a given topological vector space is said to be locally convex provided
there exists a fundamental system of neighborhoods for the zero vector consisting
of absorbing, balanced, and convex sets. We are interested in a similar yet more
nuanced notion of convexity, which involves a parameter p ∈ (0, 1]. Specifically,
assume X is a vector space and fix p ∈ (0, 1]. A set A ⊆ X is said to be p-convex
provided
p p
μ1 v1 + μ2 v2 : μ1, μ2 ≥ 0 and μ1 + μ2 = 1 ⊆ A for every v1, v2 ∈ A. (7.8.1)

An alternative characterization (proved inductively on the number of vectors in-

volved) is that A ⊆ X is p-convex if and only if A coincides with its p-convex hull,
defined as

M 
M
p
μ j v j : M ∈ N, {v j }1≤ j ≤M ⊆ A, {μ j }1≤ j ≤M ⊆ [0, ∞), μ j = 1 . (7.8.2)
j=1 j=1

Given p ∈ (0, 1], call a topological vector space locally p-convex if there exists
a fundamental system of neighborhoods for the zero vector consisting of absorbing,
balanced, and p-convex sets.
There is yet another related brand of convexity which we shall now describe.
Consider a (complex) vector space X and fix p ∈ (0, 1]. The absolutely p-convex
hull of a given set A ⊆ X is defined as


M 
M
A p := λ j v j : M ∈ N, {v j }1≤ j ≤M ⊆ A, {λ j }1≤ j ≤M ⊆ C, |λ j | p ≤ 1 .
j=1 j=1
(7.8.3)

Obviously, A ⊆ A p , and A is said to be absolutely p-convex if A coincides with

A p . It is clear from definitions that

zA p ⊆ A p for each complex number z with |z| ≤ 1, (7.8.4)

!
A p = A p p and A p ⊆ Aq if 0 < p ≤ q ≤ 1. (7.8.5)

Also, if A is such that 0 ∈ A and za ∈ A for all a ∈ A and z ∈ C with |z| = 1, then
A p is contained in the convex hull of A for any p ∈ (0, 1]. Whenever important to
stress the dependence of the absolutely p-convex hull of a given set A ⊆ X on the
space X, in place of A p we shall write AX, p .
Let X,  · X be a quasi-normed space and denote by BX (0, 1) its (“open”) unit
ball, that is, BX (0, 1) := x ∈ X :  xX < 1 (this is not an open set per se, as the
quasi-norm is not necessarily a continuous function, but is a neighborhood of the
origin, nonetheless). For each p ∈ (0, 1], define the Minkowski functional associated
with the absolutely p-convex hull of the unit ball in X, i.e.,
418 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
   !
 x  := inf λ > 0 : λ−1 x ∈ BX (0, 1) , ∀x ∈ X. (7.8.6)
p p

Whenever convenient
  to emphasize
  the dependence on X of the functional in (7.8.6),
we shall denote  ·  p by  · X, p .
! !
Remark 7.8.1 (i) Since BX (0, 1) ⊆ BX (0, 1) p , it follows that BX (0, 1) p is a
neighborhood of the origin in X. In particular, for each given  x ∈ X the infimum
in (7.8.6) is taken over a nonempty subset of (0, ∞). As such  x  p is a well-defined
number in [0, ∞). Also,
  !
 x  < 1 for each x ∈ BX (0, 1) . (7.8.7)
p p
! #
Indeed, according to (7.8.3), any x ∈ BX (0, 1) p may be written as x = M j=1 λ j v j

# Msome pfinite family {v j }1≤ j ≤M ⊆ BX (0, 1) and with {λ j }1≤ j ≤M ⊆ C satisfying

for
j=1 |λ j | ≤ 1. Hence, if we pick λ ∈ (0, 1) such that λ > max1≤ j ≤M v j  X , it
#
follows that λ−1 x = M λ (v /λ) with {v j /λ}1≤ j ≤M still in BX (0, 1). In view of
  j=1 j j
 
(7.8.6), this forces x p ≤ λ < 1, proving (7.8.7).
!
(ii) It turns out that any functional Λ ∈ X ∗ is actually bounded on BX (0, 1) p
!
by ΛX ∗ . To see that this is the case, recall that any vector x ∈ BX (0, 1) p may be
#
written as x = M j=1 λ j v j for some finite collection of vectors {v j }1≤ j ≤M ⊆ BX (0, 1)
#
and some finite family of complex numbers {λ j }1≤ j ≤M satisfying M j=1 |λ j | ≤ 1.
p
∗
Consequently, given any Λ ∈ X we may estimate


M 
M
|Λ(x)| ≤ |λ j ||Λ(v j )| ≤ ΛX ∗ |λ j |
j=1 j=1


M  1/p
≤ ΛX ∗ |λ j | p ≤ ΛX ∗ . (7.8.8)
j=1

Hence, as claimed,
!
sup |Λ(x)| : x ∈ BX (0, 1) p
≤ ΛX ∗ . (7.8.9)
!
In fact, since BX (0, 1) ⊆ BX (0, 1) p
the opposite inequality in (7.8.9) is also true,
so we ultimately conclude that
!
sup |Λ(x)| : x ∈ BX (0, 1) p
= ΛX ∗
(7.8.10)
for each functional Λ ∈ X ∗ .

As a consequence, for each functional!Λ ∈ X ∗ , each vector x ∈ X, and each positive

number λ such that λ−1 x ∈ BX (0, 1) p we may compute
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 419
!
|Λ(x)| = λ|Λ(λ−1 x)| ≤ λ sup |Λ(z)| : z ∈ BX (0, 1) p
= λΛX ∗ , (7.8.11)

thanks to (7.8.10). Hence, taking the infimum over all λ’s with the specified properties
and keeping (7.8.6) in mind, we arrive at
 
|Λ(x)| ≤  x  p ΛX ∗ . (7.8.12)

(iii) If  · X is a p-norm on X for some p ∈ (0, 1] then the absolutely p-convex

hull of unit ball in X actually coincides with the unit ball in that space, hence the
quasi-norm given by (7.8.6) coincides with the quasi-norm in the original space, i.e.,

if  · X is
! a p-norm on X for some
 p ∈ (0, 1] then
BX (0, 1) p = BX (0, 1) and  ·  p =  · X on X. (7.8.13)

 
Given a quasi-normed space X,  · X , we shall say that the dual of X separates
its points (or that X is dual rich, or that X has a separating dual) provided for each
x ∈ X \ {0} there exists a functional Λ ∈ X ∗ with the property that Λx  0.
 
Proposition 7.8.2 If X,  · X is a quasi-normed space and p ∈ (0, 1], then for each
x, y ∈ X and z ∈ C one has
         
 zx  = |z|  x  ,  x + y  p ≤  x  p +  y  p .
p p p p p
    (7.8.14)
and  x q ≤  x  p whenever q ∈ [p, 1].

Also,  

 x  ≤  xX , (7.8.15)
p

  equality in the case when  · X is actually a p-norm (cf. (7.8.13)). In addition,
with
 ·  satisfies the following “maximality” property:
p

if  ·  is some p-norm on X satisfying  · ≤

 C · X for
some constant C ∈ (0, ∞), then  ·  ≤ C  ·  p . (7.8.16)

 
Finally, if the dual of X separates its points, then  x  p = 0 if and only if x = 0.
 
Hence, in such a scenario,  ·  p is a p-norm (hence, in particular, a quasi-norm)
on X.

Proof If z = 0 then the first equality in (7.8.14) is clear from (7.8.6). When z  0,
expressing z as ρeiθ for some ρ ∈ (0, ∞) and θ ∈ R, permits us to write
420 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
   !
 zx  = inf λ > 0 : eiθ (ρ−1 λ)−1 x ∈ BX (0, 1)
p p
 !
= inf λ > 0 : (ρ−1 λ)−1 x ∈ BX (0, 1) p
 !
= ρ inf λ > 0 : λ−1 x ∈ BX (0, 1) p
   
= ρ x  p = |z|  x  p, (7.8.17)

where we have used (7.8.4) in the second equality. This establishes the first formula
in (7.8.14).
As regards the second formula in (7.8.14), observe that for each arbitrary ε > 0
there exist λx, λy > 0 satisfying
       
 x  < λx <  x  + ε,  y  < λy <  y  + ε, (7.8.18)
p p p p

and ! !
λ−1
x x ∈ BX (0, 1) p, λy−1 y ∈ BX (0, 1) p
. (7.8.19)
With this in hand, one can write
x+y λx x λy y
p 1/p = + p
p
(λx + λy )
p p 1/p
(λx + λy ) λx (λx + λy )
p 1/p
λy
 −1   −1 
= μ1 λx x + μ2 λy y , (7.8.20)

λx λy
where we have set μ1 := p p
(λ x +λ y )1/p
and μ2 := p p
(λ x +λ y )1/p
. Since these coefficients
satisfy | μ1 |p + | μ2 |p
= 1 we conclude from (7.8.3) and (7.8.19) that
x+y * ! + !
p p 1/p ∈ B X (0, 1) = BX (0, 1) p, (7.8.21)
(λx + λy ) p p

with the equality provided by the first formula in (7.8.5). As such,

          1/p
 x + y  ≤ (λxp + λyp )1/p <  x  + ε p +  y  + ε p . (7.8.22)
p p p

Sending ε  0 then yields the second formula in (7.8.14). Next, the third formula
in (7.8.14) is  seen
 directly from (7.8.6) and the second property recorded in (7.8.5).
Clearly, 0 p = 0, so the claim made in (7.8.15) is obviously true when x = 0. In
!
the case when x  0, for each θ ∈ (0, 1) we have θ x/ xX ∈ BX (0, 1) ⊆ BX (0, 1) p ,
   
hence θ  x  p = θ x  p ≤  xX by the first formula in (7.8.14) and (7.8.6). Sending
θ  1 then establishes (7.8.15).
Moving on, suppose  ·  is some p-norm on X satisfying  ·  ≤ C · X for
some constant C ∈! (0, ∞). In light of (7.8.3), these properties imply x  ≤ C for
each x ∈ BX (0, 1) p which, in concert with (7.8.6), forces  x ≤ C  x  p for each
x ∈ X. This proves (7.8.16).
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 421

Finally, assume that  dual of X separates its points and consider some x ∈ X
 the
with the property that  x  p = 0. If x  0 then there exists Λ ∈ X ∗ such that Λ(x)  0.
However, (7.8.12) written for this Λ and x then yields a contradiction. This shows
that we necessarily have x = 0. 
 
Remark 7.8.3 (i) In general, one does not expect that a finite multiple of  ·  p
dominates  · X on X, as otherwise   the inequality in (7.8.15) would make the quasi-
norm  · X equivalent with  ·  p on X, eventually rendering X a locally p-convex
space. This, of course, may not be the case for a generic quasi-normed space X
(although, as a consequence of the Aoki-Rolewicz Theorem, X is locally r-convex
for some r ∈ (0, 1]; cf. [138, Theorem 1.4, p. 5]).  
(ii) For each p ∈ (0, 1], the quasi-norm  ·  p generates a locally p-convex
topology, weaker (hence, having fewer open sets)  than
 the original topology on
X. Indeed, from Proposition 7.8.2 we know that  ·  p is a p-norm, while (7.8.15)
implies that
 
BX (x, r) ⊆ x ∈ X :  x  p < r for each x ∈ X and r > 0 (7.8.23)
   
which, in turn, readily implies that any open set in the space X,  ·  p is also open
 
in the space X,  · X .

We are now prepared to introduce the notion of p-envelope of a given quasi-

normed space X. Heuristically, this envelope may be thought of as the “smallest”
locally p-convex topological space containing the original space X.

Definition 7.8.4 Given a quasi-normed space X whose dual separates points and
some exponent p ∈ (0, 1], the p-envelope of X,henceforth
 denoted by E p (X), is
defined as the completion of X in the quasi-norm  ·  p .

A few comments are in order.  First,

 according to the manner in which the com-
pletion of X in the quasi-norm  ·  p is abstractly carried out, we may view E p (X)
as the space of equivalence classes (which we shall indicate using brackets   [·]) of
sequences of vectors in X which are Cauchy relative to the quasi-norm  ·  p . More
precisely,
 
E p (X) = {x j } j ∈N : {x j } j ∈N ⊆ X and for each ε > 0 there exists
 
N ∈ N such that  x j − xk  p < ε whenever j, k ≥ N ,
(7.8.24)

with the convention that two sequences {x j } j ∈N and {y j } j ∈N of vectors in X belong

to the same equivalence class if
for each ε > 0 there exists
 a rank N ∈ N with the
property that  x j − y j  p < ε whenever j ≥ N. (7.8.25)
422 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets


Second, given any {x j } j ∈N ∈ E p (X), Proposition 7.8.2 implies that
  
  p   p    p
 x j p − xk p  ≤ x j − xk  p, ∀ j, k ∈ N. (7.8.26)

Hence,
 
 x j  is a Cauchy sequence of numbers in [0, ∞). (7.8.27)
p j ∈N

In particular, this permits us to define

     
 {x j } j ∈N  := lim  x j  p for all {x j } j ∈N ∈ E p (X). (7.8.28)
E p (X) j→∞

We may once again rely on Proposition 7.8.2 to see that the above definition is
unambiguous, and that
      p   p
 {x j } j ∈N + {y j } j ∈N  p ≤  {x j } j ∈N  E p (X) +  {y j } j ∈N  E p (X)
E p (X)
    (7.8.29)
for all {x j } j ∈N , {y j } j ∈N ∈ E p (X).

Ultimately,

E p (X) is a vector space and  ·  E p (X) is a p-norm on it. (7.8.30)

Third, the mapping

   
ιX : X,  · X −→ E p (X),  ·  E p (X) acting according to

ιX (x) := {x, x, . . . }] for each vector x ∈ X, (7.8.31)
is linear, injective, continuous, with dense image,

and satisfies  

ιX (x) E p (X) =  x  p ≤  xX for each x ∈ X, (7.8.32)
hence, in particular,
 
 x 
p
ιX X→E p (X) = sup ≤ 1. (7.8.33)
0x ∈X  xX

This is the manner in which we shall think of X as being embedded into E p (X).
Remark 7.8.5 (i) When p = 1 in Definition 7.8.4, the space E p (X) corresponds to
the so-called Banach envelope of X. For additional information on this matter we
refer to [112], [127].
(ii) As seen from (7.8.13),

if X is a p-Banach space, for some p ∈ (0, 1], whose dual

separates points, then E p (X) = X and ιX is the identity (7.8.34)
operator in the context of (7.8.31).
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 423

Hence, in particular,

if X is a Banach space, then for each p ∈ (0, 1] we have

E p (X) = X and the mapping ιX from (7.8.31) becomes (7.8.35)
the identity operator.
A basic result is the fact that the p-envelope of a quasi-normed space with
separating dual is a p-Banach space.
 
Proposition 7.8.6 Let X,  · X be a quasi-normed space whose duals separate
points and fix some p ∈ (0, 1]. Then E p (X),  ·  E p (X) is a p-Banach space (hence,
in particular, a quasi-Banach space).

Proof
 In light of (7.8.30), there remains to show that the quasi-normed space
E p (X),  ·  E p (X) is complete. To this end, consider a Cauchy sequence {Xα }α∈N
 
in E p (X),  ·  E p (X) . This implies several things. First, for each α ∈ N it follows
that there exists a sequence {x αj } j ∈N ⊆ X which is Cauchy in  ·  p and such that
 
Xα = {x αj } j ∈N . Second, for each ε > 0 there exists Nε ∈ N with the property that
 Xα − Xβ  E p (X) < ε whenever α, β ≥ Nε . According to (7.8.28), the latter property
implies  β 
lim  x αj − x j  < ε if α, β ≥ Nε . (7.8.36)
j→∞ p
 
Since for each α ∈ N fixed the sequence {x αj } j ∈N is Cauchy in  ·  p , we may select
Jα ∈ N such that  α 
 x − x α  < α−1 if j, k ≥ Jα . (7.8.37)
j k p

If for each α ∈ N we now define yα := xJαα ∈ X, then (7.8.37) implies that for each
α ∈ N we have  
 yα − x α  < α−1 if j ≥ Jα . (7.8.38)
j p

The claim we make at this stage is that

 
{yα }α∈N is a Cauchy sequence in  ·  p . (7.8.39)

To see that this is the case, pick an arbitrary threshold ε > 0 and pick α, β ∈ N with
α, β ≥ Nε . Then, if j ∈ N is such that j ≥ max{Jα, Jβ }, we may invoke (7.8.37)
(twice) to estimate
       
 yα − yβ  p ≤  yα − x α  p +  x α − x β  p +  yβ − x β  p
p j p j j p j p
 β  p
≤ α−p +  x αj − x j  + β−p . (7.8.40)
p

Passing to the limit j → ∞ and relying on (7.8.36) we conclude that

 
 yα − yβ  p ≤ α−p + ε p + β−p . (7.8.41)
p
424 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

From this, the claim made in (7.8.39) follows. As a consequence of (7.8.39), it is

meaningful to consider
 
X ∗ := {yα }α∈N ∈ E p (X). (7.8.42)

To proceed, for each α ∈ N fixed consider the equivalence class of the stationary
sequence whose terms are all equal to yα , i.e.,
 
Xα∗ := {yα, yα, . . . } ∈ E p (X). (7.8.43)

Then for each α ∈ N we may compute

 
 Xα − Xα∗  E p (X) = lim  yα − x αj  p < α−1, (7.8.44)
j→∞

on account of (7.8.28), (7.8.38), and (7.8.43). Let us also observe that for each α ∈ N
fixed,  
 X ∗ − Xα∗  E p (X) = lim  yβ − yα  p, (7.8.45)
β→∞

thanks to (7.8.42), (7.8.43), and (7.8.28). On account of (7.8.45) and (7.8.39), we

conclude that
Xα∗ → X ∗ as α → ∞, in E p (X). (7.8.46)
Combining this with (7.8.44) (and keeping in mind that  ·  E p (X) is a quasi-norm;
cf. (7.8.30)), we finally conclude that

Xα → X ∗ as α → ∞, in E p (X). (7.8.47)
 
This shows that the quasi-normed space E p (X),  ·  E p (X) is complete, finishing
the proof of the proposition. 

Remark 7.8.7 From (7.8.34) and Proposition 7.8.6 we see that

if X is a quasi-normed space whose dual separates its points and,
for some p ∈ (0, 1], the dual of E p (X) also separates its points then (7.8.48)
E p E p (X) = E p (X).

In fact, the following more general result holds:

if X is a p-Banach space, for some p ∈ (0, 1], whose dual

(7.8.49)
separates its points, then Eq (X) = X whenever 0 < q ≤ p.

Indeed, (7.8.49) follows from (7.8.34) upon noticing that if 0 < q ≤ p ≤ 1 then any
p-Banach space is also a q-Banach space. As a consequence,
if X is a quasi-normed space whose dual separates its points and,
for some p ∈ (0, 1], the dual of E p (X) also separates its points (7.8.50)
then Eq E p (X) = E p (X) whenever 0 < q ≤ p.
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 425

Our next result identifies a dense subset of the unit ball in the envelope of a given
dual-rich quasi-normed space.
 
Proposition 7.8.8 Let X,  · X be a quasi-normed space whose dual separates
its points and fix some p ∈ (0, 1]. Also, recall the mapping ιX from (7.8.31). Then
!   
ιX BX (0, 1) p is a dense subset of B E p (X) (0, 1), the unit ball in E p (X),  ·  E p (X) .

Proof The fact that

! 
ιX BX (0, 1) p
⊆ B E p (X) (0, 1) (7.8.51)

is a consequence of (7.8.31), (7.8.28), and (7.8.7). To see that this inclusion  is dense

with respect to the topology induced by  ·  E p (X) , pick an arbitrary {x j } j ∈N in
B E p (X) (0, 1). This means that {x j } j ∈N is a sequence of vectors from
 X which is
Cauchy with respect to the quasi-norm  ·  p and satisfies lim  x j  p < 1. Fix an
j→∞  
arbitrary threshold ε > 0. Then there exist j∗ ∈ N and N ∈ N such that  x j∗  p < 1
 
and  x j − x j∗  p < ε whenever j ≥ N. In particular,
!  
x j∗ ∈ BX (0, 1) p
and ιX (x j∗ ) = {x j∗ , x j∗ , . . . } . (7.8.52)

Hence,
     
  
 {x j } j ∈N − ιX (x j∗ ) =  {x j − x j∗ } j ∈N 
E p (X) E p (X)
 
= lim  x j − x j∗  p < ε. (7.8.53)
j→∞

This ultimately shows that any element from B E p (X) (0, 1) may be approximated with-
! 
in ε by an element from ιX BX (0, 1) p . Since ε > 0 has been chosen arbitrarily,
it follows that the inclusion in (7.8.51) is indeed dense. 

We next discuss a fundamental extrapolation procedure extending the action of a

given bounded linear operator acting between two dual-rich quasi-normed spaces to
their respective envelopes, with preservation of linearity and boundedness.
   
Proposition 7.8.9 Let X,  · X and Y,  · Y be two quasi-normed spaces whose
duals separate theirs points and fix some p ∈ (0, 1]. Also, consider a linear and
bounded operator
   
T : X,  · X −→ Y,  · Y . (7.8.54)

Then T induces a linear and bounded mapping in the context

       
T : X,  · X, p −→ Y,  · Y, p . (7.8.55)
426 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Moreover, the operator T from (7.8.54) extends uniquely to a linear and bounded
operator
   
, : E p (X),  ·  E p (X) → E p (Y ),  ·  E p (Y) .
T (7.8.56)

Its operator norm satisfies

 
T (X,  | ·  | X, p )→(Y,  | ·  |Y, p ) ≤ ,
T  E p (X)→E p (Y) ≤ T (X,  ·  X )→(Y,  · Y ) (7.8.57)

and
, ◦ ιX = ιY ◦ T on X.
T (7.8.58)
   
Proof Abbreviate by T X→Y the operator norm of T : X,  · X → Y,  · Y .
If T X→Y = 0 the statement is obvious, so we shall assume in what follows that
T X→Y > 0. For starters, we claim that
   
T x  ≤ T X→Y  x X, p, ∀x ∈ X. (7.8.59)
Y, p
!
To justify this estimate, first recall that any vector x ∈ BX (0, 1) p may be written as
#
x= M j=1 λ j v j for some finite collection of vectors {v j }1≤ j ≤M ⊆ BX (0, 1) and some
#
finite family of complex numbers {λ j }1≤ j ≤M satisfying M j=1 |λ j | ≤ 1. Hence, since
p
#M
T x/T X→Y = j=1 λ j (T v j /T X→Y ) and each T v j /T X→Y belongs to BY (0, 1),
!
we conclude from (7.8.3) that T x/T X→Y ∈ BY (0, 1) p . Thus, we ultimately have
  !
T x  ≤ T X→Y , ∀x ∈ BX (0, 1) . (7.8.60)
Y, p p
!
Consider now an arbitrary x ∈ X and pick some λ > 0 such that λ−1 x ∈ BX (0, 1) p .
 
Then (7.8.60) permits us to estimate T(λ−1 x)Y, p ≤ T X→Y , which further implies
 
T x  ≤ λT X→Y for all x ∈ X and λ > 0
Y, p
! (7.8.61)
such that λ−1 x ∈ BX (0, 1) p .

With x ∈ X fixed, taking the infimum over all λ’s satisfying the aforementioned
properties then yields (7.8.59) on account of (7.8.6). This proves the claim pertaining
to (7.8.55).
To proceed, observe that (7.8.59) implies that
   
T maps Cauchy sequences in X,  · X, p
    (7.8.62)
into Cauchy sequences in Y,  · Y, p .

Next, we rely on (7.8.62) to (meaningfully) define

7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 427

, : E p (X) −→ E p (Y ) by setting
T
      (7.8.63)
, {x j } j ∈N
T := {T x j } j ∈N for each {x j } j ∈N ∈ E p (X).

To also show that the definition made in (7.8.63) is unambiguous start with two
sequences {x j } j ∈N and {y j } j ∈N from the same equivalence class. Then (7.8.25)
holds which, in concert with (7.8.62), implies that for every ε > 0 we have
   
T x j − T y j  ≤ T X→Y  x j − y j X, p < εT X→Y (7.8.64)
Y, p

whenever j ∈ N is large enough. Ultimately, this proves that {T x j } j ∈N and {T y j } j ∈N

belong to the same equivalence class, as desired. In turn, this readily implies ,
  that T
is linear in the context of (7.8.63). Going further, for any given {x j } j ∈N ∈ E p (X)
we may write
       
, 
T {x j } j ∈N  =  {T x j } j ∈N 
E p (Y) E p (Y)
   
= lim T x j Y, p ≤ T X→Y lim  x j X, p
j→∞ j→∞
 
= T X→Y  {x j } j ∈N  E p (X) , (7.8.65)

thanks to (7.8.63), (7.8.28), and (7.8.59).

To summarize, the argument so far shows the operator T , is linear and bounded
the context of (7.8.63), with operator norm ≤ T X→Y (which takes care of the
second estimate in (7.8.57)). Moreover, (7.8.58) is visible from (7.8.31) and (7.8.63).
This implies that T , : E p (X) → E p (Y ) is an extension of T : X → Y , under the
embeddings of X, Y into E p (X), E p (Y ) as in (7.8.31). The aforementioned extension
is also unique since the aforementioned embeddings have dense ranges. Finally, the
first estimate in (7.8.57) is justified by writing
    
,
  T {x j } j ∈N 
E p (Y)
,
T  E p (X)→E p (Y) = sup  
[{x j } j ∈N ]∈ E p (X)\{0}  {x }
j j ∈N

E p (X)
 
,   
T ([{x, x, . . . }])  T x 
E p (Y) Y, p
≥ sup   = sup  
x ∈X\{0}  {x, x, . . . }  x ∈X\{0}  x 
E p (X) X, p

= T (X,  | ·  | X, p )→(Y,  | ·  |Y, p ), (7.8.66)

thanks to definitions and (7.8.31). The proof of the proposition is therefore com-
plete. 
Remark 7.8.10 Let X be a quasi-normed space whose dual separates its points. Fix
p ∈ (0, 1] and assume that the dual of E p (X) also separates its points. In particular,
if Y := E p (X) then E p (Y ) = Y by (7.8.48). Recall the inclusion map ιX of X into
428 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Y from (7.8.31), and denote by , ιX : Y → Y the mapping associated with ιX as in

Proposition 7.8.9. Then, based on definitions, for each x ∈ X we may write
     
ιX ιX (x) = ,
, ιX [{x, x, . . . ] = {ιX (x), ιX (x), . . . } = ιX (x). (7.8.67)
       
ιX ιX (x) = ,
(Alternatively, , ιX ◦ ιX (x) = ιY ◦ ιX (x) = IY ιX (x) = ιX (x), thanks
to (7.8.58) and the fact that ιY = IY .) Thus, the bounded operator , ιX acts as the
identity on the image of ιX . Since the latter is dense in Y (cf. (7.8.31)), we ultimately
conclude that
ιX = I E p (X) , the identity operator on E p (X).
, (7.8.68)
A remarkable consequence of Proposition 7.8.9 is the fact that the envelope of a
dual-rich quasi-Banach space has the same dual as the original space.
Proposition 7.8.11 Let X be a quasi-normed space whose dual separates its points
and fix some p ∈ (0, 1]. Then one has the identification
 ∗
E p (X) = X ∗ isometrically via
  (7.8.69)
, ∈ E p (X) ∗ .
X ∗  Λ −→ Λ
 ∗
Proof Consider the mapping Φ : X ∗ → E p (X) defined by Φ(Λ) := Λ , for each
Λ ∈ X ∗ , with the “hat” understood in the sense of (7.8.63). Since E p (C) = C (cf.
(7.8.35)), Proposition 7.8.9 ensures that the mapping Φ is well defined, linear, and
bounded. We claim that, in fact, Φ is an isometry. To prove this is the case, for each
Λ ∈ X ∗ compute
       !   
, ,
Λ(E p (X))∗ = sup ,
Λ B E p (X) (0, 1)  = sup Λ ιX BX (0, 1) p 
  !   !
 ,
= sup  Λ ◦ ιX BX (0, 1) p  = sup |Λ(x)| : x ∈ BX (0, 1) p

= ΛX ∗ . (7.8.70)

Above, first equality is simply the definition of the quasi-norm of Λ , in the space
 the
∗
E p (X) (as in the past, B E p (X) (0, 1) denotes the unit ball in E p (X)). The second
equality is a consequence of Proposition 7.8.8, while the third equality is plain
algebra. Next, the fourth equality is justified by observing that for each x ∈ X we
have
 
, ◦ ιX (x) = Λ
Λ , ([{x, x, . . . }]) = [{Λx, Λx, . . . }] = Λx, (7.8.71)

thanks to (7.8.31), (7.8.63), and the manner in which E p (C) is identified with C.
Lastly, the final equality in (7.8.70) is provided by (7.8.10). This establishes (7.8.70)
which, in turn, proves that Φ is indeed an isometry.  ∗ ∗
Going further, let us also consider
  ∗ mapping Ψ : E p (X) → X defined by
the
Ψ(Θ) := Θ ◦ ιX for each Θ ∈ E p (X) , where ιX is the mapping introduced in
(7.8.31). Clearly, Ψ is also well-defined, linear, and bounded. Since from (7.8.71)
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 429

we know that
, ◦ ιX = Λ on X, for each Λ ∈ X ∗,
Λ (7.8.72)

we conclude
 that Ψ(Φ(Λ)) = Λ for each Λ ∈ X ∗ . In the opposite direction, given
∗
any Θ ∈ E p (X) , it follows that Θ ◦ ιX ∈ X ∗ satisfies
   
Θ-◦ ιX ιX (x) = (Θ ◦ ιX )(x) = Θ ιX (x) for each x ∈ X, (7.8.73)

where the first equality is a consequence  of (7.8.72). In turn, (7.8.73) implies that
- ∗
Θ ◦ ιX and Θ are two functionals in E p (X) which agree on ιX (X), the image of
the mapping ιX . Since, as noted in (7.8.31), the space ιX (X) is dense in E p (X), we
- -
 ∗ conclude that Θ ◦ ιX and Θ agree on E p (X). The fact that Θ ◦ ιX = Θ in
ultimately

E p (X) translates into Φ(Ψ(Θ)) = Θ. The above reasoning shows that the mappings
Φ and Ψ are inverse to one another. In particular, since Φ is an isometry it follows
that Ψ is also an isometry, thus finishing the proof of (7.8.69). 
 
On a given dual-rich quasi-Banach space, the action of the norm  · 1 may be
characterized via duality, in the manner described in the proposition below.
 
Proposition 7.8.12 Let X,  · X be a quasi-normed space whose duals separate
points. Recall that E1 (X),  ·  E1 (X) is the Banach
 envelope
  of X (cf. item (i) of
Remark 7.8.5), and recall the embedding ιX : X,  · X → E1 (X),  ·  E1 (X) from
(7.8.31) (with p := 1). Then for each x ∈ X one has
   
 x  = ιX (x) E (X) = sup Λ(x) : Λ ∈ X ∗, ΛX ∗ = 1 . (7.8.74)
1 1

Proof Indeed, for each x ∈ X we may write

 
 x  = ιX (x) E (X)
1 1

    ∗
= sup  ιX (x)  :  ∈ E1 (X) , (E1 (X))∗ = 1
  
= sup ,
Λ ιX (x)  : Λ ∈ X ∗, ΛX ∗ = 1
 
= sup Λ(x) : Λ ∈ X ∗, ΛX ∗ = 1 , (7.8.75)
 
thanks to (7.8.32), the fact that E1 (X),  ·  E1 (X) is a Banach space (so the Hahn-
Banach Theorem and its usual consequences are valid), the isometric identification
made in Proposition 7.8.11, and (7.8.71). 

For future purposes it is useful to note that taking “hats” preserves the identity.
 
Proposition 7.8.13 Let X,  · X be a quasi-normed space whose dual separates
its points and fix some p ∈ (0, 1].
430 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
   
Then the mapping ,IX : E p (X),  ·  E p (X) → E p (X),  ·  E p (X) associated as in
Proposition 7.8.9 with IX , the identity operator on X, is actually I E p (X) , the identity
 
operator on E p (X),  ·  E p (X) .

Proof With ιX denoting the mapping introduced in (7.8.31), repeated applications

of (7.8.58) allow us to write
,IX ◦ ιX = ιX ◦ IX = ιX = I E p (X) ◦ ιX . (7.8.76)

Hence, ,IX and I E p (X) coincide on ιX (X), the image of the mapping ιX . Since the
latter is a dense subspace of E p (X) (cf. (7.8.31)), this ultimately forces ,IX = I E p (X)
on E p (X). 

It turns out that taking “hats” commutes with composition of operators. This is
made precise in our next proposition.
     
Proposition 7.8.14 Let X,  · X , Y,  · Y , Z,  ·  Z , be three quasi-normed
spaces whose duals separate their points and fix some p ∈ (0, 1]. Also, let
       
T : X,  · X → Y,  · Y and S : Y,  · Y → Z,  ·  Z (7.8.77)

be two linear and bounded operators. Then, with “hats” considered in the sense of
Proposition 7.8.9, it follows that

S.◦ T = S, ◦ T , as operators
    (7.8.78)
from E p (X),  ·  E p (X) into E p (Z),  ·  E p (Z) .

Proof Proposition 7.8.9 ensures that the operators S. ◦ T and S, ◦ T

, are well-defined

linear and bounded mappings from E p (X),  ·  E p (X) into E p (Z),  ·  E p (Z) . In
addition, if ιX , ιY , ι Z are the mappings associated with X, Y, Z as in (7.8.31), based
on repeated applications of (7.8.58) we may write
       
S.
◦ T ◦ ιX = ι Z ◦ S ◦ T = ι Z ◦ S ◦ T = S, ◦ ιY ◦ T
     
= S, ◦ ιY ◦ T = S, ◦ T , ◦ ιX = S, ◦ T
, ◦ ιX . (7.8.79)

Consequently, S.◦ T and S, ◦ T

, agree on ιX (X), the image of the mapping ιX . Given
that, as remarked in (7.8.31), the space ιX (X) is dense in E p (X), we ultimately
conclude that S.
◦ T and S, ◦ T
, agree on E p (X). 

Other functorial properties of the “hat” operation are discussed in the proposition
below.
   
Proposition 7.8.15 Suppose X,  · X and Y,  · Y are two quasi-normed spaces
with separating duals and
 fix some
 exponent
 p ∈ (0, 1]. Given some linear and
bounded operator T : X,  · X → Y,  · Y , consider the linear and bounded
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 431
   
operator T, : E p (X),  ·  E p (X) → E p (Y ),  ·  E p (Y) associated with T as in
Proposition 7.8.9. Then the following statements are true.
  −1 /
(1) If T is an isomorphism, then T , is an isomorphism and, in fact, T , = T −1 .
,
(2) If X, Y are quasi-Banach spaces and T is surjective, then T is surjective.
Proof To justify the claim made in item (1), invoke Proposition 7.8.14 and Proposi-
tion 7.8.13 to write
, ◦ T/
T 0
−1 = T ◦ T −1 = ,IX = I E p (X),
(7.8.80)
T/ , = T0
−1 ◦ T −1 ◦ T = ,
IX = I E p (X) .
  −1 /
This proves that T , is an isomorphism and, in fact, T , = T −1 .
Let us now deal with the claim made in item (2). Consider an arbitrary vector
y ∈ E p (Y ) satisfying 
 y  E p (Y) ≤ 1. Then
 Proposition
 7.8.8 ensures that there exists
!
y∗ ∈ BY (0, 1) p with the property that  y −ιY (y∗ ) E p (Y) ≤ 1/2. The aforementioned
#
M
membership guarantees that y∗ may be represented in the form y∗ = λ j y j for some
j=1
finite family of vectors {y j }#
1≤ j ≤M ⊆ BY (0, 1) and some finite family of numbers
{λ j }1≤ j ≤M ⊆ C satisfying M j=1 |λ j | ≤ 1. Given that we are currently assuming
p

that X, Y are quasi-Banach spaces and the operator T is surjective, we may invoke
[138, Corollary 6.62, p. 423] to conclude that there exist a number C ∈ (0, ∞)
and a family of vectors {x j }1≤ j ≤M ⊆ X such that T x j = y j and  x j X ≤ C for
#
M
every j ∈ {1, . . . , M }. Define x∗ := λ j x j ∈ X and pick Co ∈ (C, ∞). Since
j=1
#
M
x∗ /Co may be represented as λ j (x j /Co ) with {x j /Co }1≤ j ≤M ⊆ BX (0, 1) and
j=1
#
M    
|λ j | p ≤ 1, it follows that  x∗ /Co X, p ≤ 1. Hence,  x∗ X, p ≤ Co which then
j=1    
implies ιX (x∗ ) E p (X) =  x∗ X, p ≤ Co . Let us also observe that by design T x∗ = y∗
which, in concert with (7.8.58), allows us to write
 
T, ιX (x∗ ) = ιY (T x∗ ) = ιY (y∗ ). (7.8.81)
    
Consequently,  , ιX (x∗ ) 
y −T E p (Y)
= y − ιY (y∗ ) E p (Y) ≤ 1/2. Let us summarize
our progress. Abbreviating x := ιX (x∗ ) ∈ E p (X), the argument so far shows that

there exists a constant Co ∈ (0, ∞) with the property that for each vector

y ∈ E p (Y ) with  y  E p (Y) ≤ 1 there exists some  x ∈ E p (X) satisfying (7.8.82)
 
x  E p (X) ≤ Co and such that  ,
y −T x  E p (Y) ≤ 1/2.

With this in hand, and recalling from Proposition 7.8.6 that E p (Y ) is a quasi-Banach
space, we may now rely on [138, Proposition 6.60, pp. 421–422] to conclude that
, is onto, as desired.
the operator T 
432 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 
Let X,  · X be a quasi-normed
  space. A projection of X is a linear, bounded,
idempotent operator P on X,  · X , where the last property amounts to P2 = P on
X. As a result, the image Y := P(X) of a projection P of X satisfies
 
Y = Ker IX − P : X → X and Y ⊕ Ker (P : X → X) = X, (7.8.83)

where IX is the identity operator on X. In particular, Y is a (topologically) comple-

mented closed subspace of X. Conversely, any (topologically) complemented closed
subspace of X is the range of a projection of X.
It turns out that the envelope of the a (topologically) complemented space or,
equivalently, the envelope of the range of a projection, behaves in a natural fashion,
as indicated in the proposition below.
 
Proposition 7.8.16 Let X,  · X be a quasi-normed space with separating dual and
fix some p ∈ (0, 1]. Also, consider a (topologically) complemented closed subspace
Y of X.  
Then Y := Y,  · X is a quasi-normed space with separating dual and
 
E p (Y ) := E p Y,  · X embeds continuously into E p (X). (7.8.84)
 
In fact, if X,  · X is actually a quasi-Banach space and P : X → X is a projection
of X onto Y , then P , : E p (X) → E p (X) considered in the sense of Proposition 7.8.9
is a projection of E p (X) onto E p (Y ); in particular, in such a scenario one has
   
, E p (X) , or E p (PX) = P
E p (Y ) = P , E p (X) ,
   
or E p Ker IX − P : X → X = P , E p (X) , (7.8.85)
  
or E p Ker IX − P : X → X = Ker I E p (X) − P , : E p (X) → E p (X) .

 
Proof Since X,  · X is a quasi-normed space with separating dual it follows that
Y := Y,  · X inherits these properties. To proceed, let ι be the inclusion map of Y
into X. Hence, ι : Y → X is linear and bounded. As such, Proposition 7.8.9 ensures
that ι extends to a linear and bounded map
 
,
ι : E p (Y ) := E p Y,  · X −→ E p (X). (7.8.86)

We claim that , ι is also injective. To prove this claim, according to (7.8.24), (7.8.63),
and (7.8.28) it suffices to show that
 
if {y j } j ∈N ⊆ Y is such that y j → 0 as j → ∞ in  · X, p
  (7.8.87)
then we necessarily have y j → 0 as j → ∞ in  · Y, p .

We shall actually prove that there exists a constant C ∈ (0, ∞) with the property that
   
 ·  ≤ C  · X, p on Y . (7.8.88)
Y, p
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 433

Of course, (7.8.87) is immediate from (7.8.88). To justify

  (7.8.88), pick an arbitrary
vector y ∈ Y and select a positive number a >  y X, p . In view of (7.8.6), this
!
guarantees the existence of some λ ∈ (0, a) such that λ−1 y ∈ BX (0, 1) X, p . Upon
recalling (7.8.6), we see that we may express


M 
M
λ−1 y = λi vi with |λi | p ≤ 1 and {vi }1≤i ≤M ⊆ BX (0, 1). (7.8.89)
i=1 i=1

Since Y is assumed to be a (topologically) complemented closed subspace of X,

there exist a closed subspace Z of X along with a constant C ∈ (0, ∞) such that
any x ∈ X may be uniquely decomposed as y + z with y ∈ Y , z ∈ Z
(7.8.90)
and, in addition, the estimate max  yX , zX ≤ C xX holds.

Use (7.8.90) to split each vi appearing in (7.8.89) as vi = wi + zi with wi ∈ Y and

zi ∈ Z satisfying max wi X , zi X ≤ Cvi X < C. Then


M 
M  
M 
λ−1 y = λi vi = λ i wi + λi zi . (7.8.91)
i=1 i=1 i=1
#M
Since Y ∩ Z = {0}, the fact that λ−1 y ∈ Y forces i=1 λi zi = 0 and


M
λ−1 y = λ i wi . (7.8.92)
i=1
#M
Given that wi /C ∈ BY (0, 1) and also ! i=1 |λi | ≤ 1, from (7.8.92) and (7.8.3) we
p
−1
conclude that λ y/C ∈ BY (0, 1) Y, p . On account of (7.8.6), this further implies
 −1     
λ y  ≤ C. Thus,  y Y, p ≤ Cλ < Ca. By letting a   y X, p we ultimately
 
Y, p  
arrive at  y Y, p ≤ C  y X, p . This establishes (7.8.88). The claim made in (7.8.87)
is therefore justified and, hence, the mapping , ι in (7.8.86) is indeed injective.
 
Going further, strengthen the original hypotheses by assuming X,  · X is
actually a quasi-Banach space, and observe that this further entails that Y,  · X
is a quasi-Banach space. Also, suppose P : X → X is a projection of X onto
Y and consider P , : E p (X) → E p (X) in the sense of Proposition 7.8.9, which is
known to be linear and bounded. Since Proposition 7.8.14 also guarantees that P ,
is idempotent, ,
that P is a projection of the space E p (X). Next, define
 it follows

P# : X,  · X → Y,  · X by setting P# x := Px for each x ∈ X. Given that
Y = P(X), this is a well-defined linear and bounded operator which, by design,
is surjective. Granted this, from part (2) of Proposition 7.8.15 we conclude that
1
P# : E p (X) → E p (Y ) is also surjective, hence
 
1
P# E p (X) = E p (Y ). (7.8.93)

In addition, P = ι ◦ P# so Proposition 7.8.14 implies

434 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

,=,
P ι◦1
P# . (7.8.94)

Consequently, on account of (7.8.94) and (7.8.93) we may write

     
P, E p (X) = ,ι 1
P# E p (X) = , ι E p (Y ) ≡ E p (Y ) (7.8.95)

where the last step takes into consideration the specific manner in which E p (Y ) is
identified with a subspace of E p (X) in (7.8.84) (via,
ι, that is). Since (7.8.95) may be
,
interpreted as saying that the projection P of E p (X) maps onto E p (Y ), the proof of
the proposition is complete. 

Next we present the following general criterion for identifying the p-envelope of
a given quasi-normed space with a rich dual.
 
Theorem 7.8.17 Let E,  · E be a quasi-normed
  space whose dual separates its
points. Consider a p-Banach space F,  · F , for some p ∈ (0, 1], whose dual
separates its points and with the property that E ⊆ F and the inclusion ι : E → F
is continuous with dense range. Then the following statements are equivalent.

(1) There exists an isomorphism3 from E p (E) onto F which fixes E. More specifi-
cally,

 exists Φ : E p (E) → F isomorphism with the property that

there
ΦE = I, the identity operator on E, where E is regarded as a subspace
of E p (E) (via the embedding ιE ; cf. (7.8.31)).
  (7.8.96)
(2) The quasi-norms  · E, p and  · F are equivalent on E.
(3) There exists a constant C ∈ (0, ∞) such that
 
 x  ≤ C xF for each x ∈ E. (7.8.97)
E, p

(4) The space E has a good approximation of the identity (relative to the ambient
F), in the sense that there exists a sequence of linear operators PN : E → E
indexed by N ∈ N with the property that one may find a constant C ∈ (0, ∞)
such that  
sup PN x E, p ≤ C xF for each x ∈ E, (7.8.98)
N ∈N

and
PN x − xE → 0 as N → ∞ for each fixed x ∈ E. (7.8.99)
(5) There exists a constant c ∈ (0, ∞) with the property that
!  · E
E ∩ BF (0, c) ⊆ BE (0, 1) E, p
(7.8.100)

3 linear, continuous, bijective map

7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 435

where BF (0, c) := x ∈ F :  xF < c and the vertical bar stands for closure
in the topology induced by the indicated quasi-norm.
(6) There exists a constant C ∈ (0, ∞) with the property that for each vector x ∈ E
one may find a sequence {x N } N ∈N ⊆ E such that
 
sup  x N E, p ≤ C xF (7.8.101)
N ∈N

and  
x = lim x N in E,  · E . (7.8.102)
N →∞

Remark 7.8.18 Whenever (7.8.96) holds we shall simply say that F is the p-envelope
of E, and (via a slight abuse of notation) write F = E p (E), in place of the identifi-
cation of F with E p (E) via the isomorphism Φ.

Before presenting the proof of Theorem 7.8.17 we wish to make two comments.
First, as evident from item (6), the operators PN : E → E with N ∈ N from item
(4) need not be linear. Second, Theorem 7.8.17 specialized to the case p = 1 yields
criteria for identifying the Banach envelope of a given quasi-normed space with
separating dual.
Proof of Theorem 7.8.17 To check the implication (1) ⇒ (2), for each x ∈ E write
  
 xF = Φ ιE (x) F ≈ ιE (x) E p (E)
   
=  {x, x, . . . }  E p (E) =  x E, p . (7.8.103)

Above, the first equality is a consequence of the fact that ΦE = I, the identity
operator on E, given the manner in which E is regarded as a subspace of E p (E) (via
the embedding ιE ; cf. (7.8.31)). The subsequent equivalence is a consequence of the
Open Mapping Theorem for quasi-Banach spaces (cf., e.g., [138, Corollary 6.62,
p. 423]), bearing in mind Proposition 7.8.6. For the remaining equalities in (7.8.103)
see (7.8.31), (7.8.32).
Next, the implication (2) ⇒ (3) is trivial. As regards the implication (3) ⇒ (4),
assume (7.8.97) holds and take PN := I, the identity operator on E, for each N ∈ N.
Obviously, these are linear operators on E satisfying (7.8.98)-(7.8.99).
Moving on, consider the implication (4) ⇒ (5). With C ∈ (0, ∞) as in (7.8.98),
define c := C −1 . Then for each x ∈ E ∩ BF (0, c) we may rely on (7.8.98) to estimate
 
PN x  ≤ C xF < 1 for each N ∈ N. (7.8.104)
E, p

In light of this, (7.8.6) implies that for each !N ∈ N there exists some number !
λ N ∈ (0, 1) such that λ−1 N · PN x ∈ BE (0, 1) E, p . Hence, PN x ∈ BE (0, 1) E, p
for each N ∈ N, thanks to (7.8.4). From this and (7.8.99) we then conclude that
!  · E
x ∈ BE (0, 1) E, p . In view of the arbitrariness of x, this ultimately establishes
(7.8.100).
436 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Turning to the implication (5) ⇒ (6), pick an arbitrary vector x ∈ E. If x = 0,

then (7.8.101)-(7.8.102) are trivially satisfied (for any C ∈ (0, ∞))by taking x N := 0
for each N ∈ N. When x  0, choose an arbitrary number λ ∈ 0, c/ xF where
c ∈ (0, ∞) is as in (7.8.100), and note that this entails λ x ∈ E ∩ BF (0, c). Upon
recalling that we are presently assuming
! (7.8.100), we then conclude that there
 exists
a sequence { x N } N ∈N ⊆ BE (0, 1) E, p such that λ x = lim  x N in E,  · E . If we
N →∞  
now define x N := λ−1  x N for each N ∈ N, it follows that x = lim x N in E,  · E
N →∞
and    
sup  x N E, p = λ−1 sup 
x N E, p ≤ λ−1 . (7.8.105)
N ∈N N ∈N

Upon letting λ  c/ xF in (7.8.105) we therefore arrive at

 
sup  x N E, p ≤ c−1  xF (7.8.106)
N ∈N

which proves (7.8.101) with C := c−1 .

Finally, consider the implication (6) ⇒ (1). Assume the existence of C ∈ (0, ∞)
with the property that for each x ∈ E one may find a sequence {x N } N ∈N ⊆ E
satisfying (7.8.101)-(7.8.102). The goal is to show that F is isomorphic to the p-
envelope of E, in a manner that fixes E. To get started, observe that (7.8.13) and
(7.8.59) presently ensure that
   
 xF =  x F, p ≤ ιE→F  x E, p for all x ∈ E. (7.8.107)

(Alternatively, the same conclusion may be justified using (7.8.16).) Also, thanks to
Proposition 7.8.9 and (7.8.34), the inclusion ι : E → F extends to a bounded, linear
operator
,ι : E p (E) −→ F (7.8.108)
which fixes E (regarded as a subspace of E p (E) in the manner described in (7.8.31))
and acts according to the following scheme:
 
Start with {x j } j ∈N ∈ E p (E).
 Then the sequence {x j } j ∈N ⊆ E
is Cauchy with respect to  · E, p hence, thanks to the estimate in
 
(7.8.107), {x j } j ∈N ⊆ E is Cauchy with respect to  · F . Since F,  · F (7.8.109)
is complete, the limit lim x j exists in the latter space, and we set
   j→∞
ι {x j } j ∈N E = lim x j ∈ F.
,
j→∞

The goal now becomes proving that the extension described in (7.8.108)-(7.8.109)
is, in fact, an isomorphism. From (7.8.108)-(7.8.109) it is clear that this operator is
one-to-one if we succeed in showing that
 
if {x j } j ⊆ E is Cauchy with respect to  · E, p and x j → 0 in F
  (7.8.110)
then necessarily x j → 0 in  · E, p as j → ∞.
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 437
 
Consider a sequence as in the first line of (7.8.110). Upon recalling that  ·  is a E, p
p-norm on E (cf. Proposition 7.8.2), for each j ∈ N and N ∈ N we may estimate
  p   p   p
 x j  ≤  x j − (x j ) N E, p + (x j ) N E, p (7.8.111)
E, p

where {(x j ) N } N ∈N is the sequence associated with each vector x j as in item (6). In
particular, the current hypotheses guarantee  that for each  given ε > 0 we may find
some jε ∈ N such that  x jε F ≤ ε and  x jε − (x jε ) N E ≤ ε whenever N is large
enough. Assuming this is the case, (7.8.15) used with X := E permits us to estimate
   
 x j − (x j ) N  ≤  x jε − (x jε ) N E ≤ ε. (7.8.112)
ε ε E, p
 
Also, (7.8.101) gives (x jε ) N E, p ≤ C x jε F ≤ Cε. In concert with (7.8.111),
 
these estimates allow us to conclude that  x jε E, p ≤ ε(1 + C p )1/p . This proves that
 
{x j } j ∈N contains a sub-sequence which is convergent to zero with respect to  · E, p .
 
Being Cauchy in the p-norm  · E, p , then the entire sequence {x j } j ∈N converges
 
to zero with respect to  · E, p . This finishes the proof of (7.8.110) which, in turn,
shows that , ι in (7.8.108)-(7.8.109) is indeed one-to-one.
Let us turn our attention to the ontoness of, ι in (7.8.108)-(7.8.109). Fix an arbitrary
x ∈ E and consider the sequence {x N } N ∈N associated with the vector x as in item
(6). Then, thanks to (7.8.15) and (7.8.101) (also mindful of the fact that  · E, p is a
p-norm on E; cf. Proposition 7.8.2), for each N ∈ N we may write
  p   p   p
 x  ≤  x N − x E, p +  x N E, p ≤  x N − xE + C p  xF .
p p
(7.8.113)
E, p

Based on this and (7.8.102), sending N → ∞ yields

 
 x  ≤ C xF for every x ∈ E, (7.8.114)
E, p

where C ∈ (0, ∞) is as in (7.8.114). Next, keeping in mind that E is viewed as a

subspace of E p (E) as indicated in (7.8.31), we claim that
 
E ∩ BF (0, C −1 ) ⊆ ,
ι E ∩ B E p (E) (0, 1) . (7.8.115)
 
To see this, pick an arbitrary vector x ∈ E ∩ BF (0, C −1 ). Then {x, x, . . . } belongs
to the identification of E with a subspace of E p (E) (cf. (7.8.31)) and satisfies
   
 {x, x, . . . }  =  x E, p ≤ C xF < 1, (7.8.116)
E p (E)
 
thanks to (7.8.114) and the fact that x ∈ BF (0, C −1 ). Thus, {x, x, . . . } belongs to
  
E ∩ B E p (E) (0, 1) and (7.8.109) gives ,
ι {x, x, . . . } = x. The proof of the claim
made in (7.8.115) is therefore finished.
−1
 Moving  on, we wish to remark that, in general, BF (0, C ) is not an open set in
F,  · F although this is a neighborhood of the origin (cf. [138, (3.543), p. 149]).
438 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 
In particular, there exists an open set O in the space F,  · F with 0 ∈ O and such
that O ⊆ BF (0, C −1 ). Since we assume that E is densely embedded into F, we may
then use (7.8.115) to write
 · F  · F  · F    · F
O⊆O =E∩O ⊆ E ∩ BF (0, C −1 ) ⊆,
ι B E p (E) (0, 1) (7.8.117)

where, in each case, the vertical bar stands for closure in the topology induced by
the indicated quasi-norm. Noting that the set on the left-most side of (7.8.117) is
a neighborhood of the zero vector in F, and recalling that E p (E),  ·  E p (E) is a
p-Banach space, hence complete (cf. Proposition 7.8.6), we may now invoke the
version of the Open Mapping Theorem established in [138, Theorem 6.49, p. 408]
to conclude that ,
ι in (7.8.108)-(7.8.109) is an open mapping. Consequently, , ι is
onto (hence, ultimately, bijective) and has a continuous inverse in the context of
(7.8.108). 

As an example, Theorem 7.8.17 yields an elegant proof of the fact that

  ∗
E p∗  p (N) =  p (N) whenever 0 < p ≤ p∗ ≤ 1. (7.8.118)
∗
Indeed, for starters, E :=  p (N) and F :=  p (N) satisfy the background hypotheses
made in the statement of Theorem 7.8.17 (with p replaced by p∗ ). Next, for each
N ∈ N define PN : E → E by setting
 
PN {λ j } j ∈N := λ1, λ2, . . . , λ N −1, λ N , 0, 0, 0, . . .
(7.8.119)
for each {λ j } j ∈N ∈  p (N).

Hence, for each fixed {λ j } j ∈N ∈ E, we have

    
  1/p
PN {λ j } j ∈N − {λ j } j ∈N  = |λ j | p → 0 as N → ∞. (7.8.120)
E
j>N

Also, if for each j ∈ N we consider v j := {δ jk }k ∈N ∈ BE (0, 1), then for each N ∈ N

we have
  N
PN {λ j } j ∈N = λ j v j for each {λ j } j ∈N ∈ E. (7.8.121)
j=1

In turn, from (7.8.121), (7.8.6), and (7.8.3) we conclude that

    # 1/p ∗  
PN {λ j } j ∈N  ≤ N
|λ j | p
∗
≤ {λ j } j ∈N F
E, p ∗ j=1 (7.8.122)
for each N ∈ N and {λ j } j ∈N ∈ E.

Together, (7.8.120) and (7.8.122) show that the demands in item (6) of Theo-
rem 7.8.17 are satisfied by the family of operators {PN } N ∈N . As such, Theo-
rem 7.8.17 applies and gives (7.8.118).
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 439

In turn, Theorem 7.8.17 is the basis for other procedures capable of identifying the
p-envelope of a given quasi-normed space with separating dual. In Proposition 7.8.19
below we discuss such a procedure which allows transferring the quality of being the
p-envelope from one pair of quasi-normed spaces to another granted the existence
of (abstract) restriction/extension operators between said pairs.
 
Proposition 7.8.19 Let Xi,  · Xi with i ∈ {1, 2} be two quasi-normed spaces with
separating duals, satisfying X1 ⊆ X2 , and such that

the inclusion ι1 : X1 → X2 is continuous. (7.8.123)

 
Consider also two quasi-normed spaces Yi,  · Yi with i ∈ {1, 2} satisfying Y1 ⊆ Y2 ,
and such that
the inclusion ι2 : Y1 → Y2 is continuous. (7.8.124)
Next, for i ∈ {1, 2} let
       
Ri : Xi,  · Xi → Yi,  · Yi and Ei : Yi,  · Yi → Xi,  · Xi (7.8.125)

be linear and bounded operators satisfying (again, for i ∈ {1, 2})

Ri ◦ Ei = IYi , the identity operator on Yi . (7.8.126)

In addition, assume that R2 extends R1 and E2 extends E1 , in the sense that

ι2 ◦ R1 = R2 ◦ ι1 as operators from X1 into Y2, (7.8.127)

ι1 ◦ E1 = E2 ◦ ι2 as operators from Y1 into X2 . (7.8.128)

Then, given any p ∈ (0, 1], the following implication holds:

E p (X1 ) = X2 =⇒ Y1, Y2 have separating duals and E p (Y1 ) = Y2 . (7.8.129)

Proof Fix p ∈ (0, 1] and work under the assumption that E p (X1 ) = X2 . In particular,
after eventually replacing  · X2 with an equivalent quasi-norm on X2 (something
which does not affect any of the hypotheses made in the statement) we may, in light
of Proposition 7.8.6 and (7.8.31), assume that
 
X2,  · X2 is a p-Banach space, and the inclusion
(7.8.130)
ι1 : X1 → X2 is continuous with dense range.
We next claim that
 
Y2,  · Y2 is a quasi-Banach space. (7.8.131)
 
By assumption, Y2,  · Y2 is a quasi-normed space so we only need  to check 
completeness. To this end, suppose {y j } j ∈N is a Cauchy
 sequence in Y2,  · Y2 .
Then {E2 y j } j ∈N is a Cauchy sequence in X2,  · X2 and, since the latter
 space has
been shown to be complete, it follows that {E2 y j } j ∈N converges in X2,  · X2 to
440 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

some x ∈ X2 . In concert with (7.8.126), this  allows us to conclude that the sequence
{y j } j ∈N = {R2 (E2 y j )} j ∈N converges in Y2,  · Y2 to R2 x ∈ Y2 . Hence, the claim
made in (7.8.131) is indeed true.
To proceed, define

 yY2,∗ := E2 yX1 for each y ∈ Y2 . (7.8.132)

In relation to this, we claim that

 · Y2,∗ is a p-norm on Y2 , which is equivalent
(7.8.133)
(as a quasi-norm) with  · Y2 on the space Y2 .

To justify this, first observe that for each y ∈ Y2 we have

 
 yY2 =  R2 (E2 y)Y2 ≤ R2 X2 →Y2 E2 yX2 = R2 X2 →Y2  yY2,∗ (7.8.134)

and
 yY2,∗ = E2 yX2 ≤ E2 Y2 →X2  yY2 . (7.8.135)
Also, for each y ∈ Y2 and λ ∈ C we may write

λyY2,∗ = E2 (λy)X2 = λE2 yX2 = |λ|E2 yX2 = |λ| yY2,∗ . (7.8.136)

From (7.8.134)-(7.8.136) we conclude that  · Y2,∗ is a quasi-norm on Y2 , which is

equivalent with  · Y2 on Y2 . Finally, for each y , y ∈ Y2 we may estimate (keeping
in mind (7.8.132) and the fact that  · X2 is a p-norm)
p p p
 y + y Y2,∗ = E2 (y + y )X2 = E2 y + E2 y X2
p p p p
≤ E2 y X2 + E2 y X2 =  y Y2,∗ +  y Y2,∗ (7.8.137)

which, ultimately, goes to show that  · Y2,∗ is a p-norm on Y2 . This finishes the proof
of (7.8.133). On account of this and (7.8.131), there is therefore no loss of generality
in assuming that actually
 
Y2,  · Y2 is a p-Banach space. (7.8.138)

Next, we shall show that

 
Yi,  · Yi have separating duals, for i ∈ {1, 2}. (7.8.139)

To prove this, assume first that y ∈ Y1 \ {0}. Then E1 y ∈ X1 \ {0} (since
 E1 y = 0
would force y = R1 (E1 y) = 0, an impossibility). Given that X1,  · X1 is assumed

to have a separating dual, there exists Λ1 ∈ X1∗ with the property that Λ1 E1 y  0.
define Λ := Λ1 ◦ E1 , it follows that ∗
If we now
  Λ ∈ Y1 and Λy = Λ1 E1 y  0.
Thus, Y1,  · Y1 has a separating dual. That Y2,  · Y2 also has a separating dual
is established similarly, finishing the proof of (7.8.139).
Pressing on, we claim that
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 441

the inclusion ι2 : Y1 → Y2 has dense range. (7.8.140)

To see this is the case, pick an arbitrary y ∈ Y2 . Then E2 y ∈ X2 , so according to

the last claim
 in (7.8.130) there exists a sequence {x j } j ∈N ⊆ X1 which converges
to E2 y in X2,  · X2 . Consequently,
 {R2 x j } j ∈N = {R1 x j } j ∈N ⊆ Y1 converges to
R2 (E2 y) = y in Y2,  · Y2 , proving (7.8.140).
Finally, for each y ∈ Y1 we may estimate (with C ∈ (0, ∞) independent of y)
     
 y  =  R1 (E1 y) ≤ E1 y 
Y1, p Y1, p X1, p

≤ CE1 yX2 = CE2 yX2 ≤ C yY2 (7.8.141)

thanks to (7.8.126) (with i := 1), the first claim in Proposition 7.8.9, the estimate
from (7.8.97) (with x := y, E := X1 , and F := Y2 ), thefact that E 2 extends
 E1 (cf.
(7.8.128)), and the boundedness of the operator E2 : Y2,  · Y2 → X2,  · X2
(cf. (7.8.125)). In turn, from (7.8.141) and the equivalence of items (3) and (1) in
Theorem 7.8.17 we ultimately conclude that E p (Y1 ) = Y2 .
Parenthetically, we wish to note that other equivalences from Theorem 7.8.17
work just as well as far as the present goal is concerned. For example, it may be
checked without difficulty that if {PN } N ∈N is a family of linear operators on Y1
which constitute a good approximation of the identity on the space X1 (relative to the
ambient X2 ) in the sense of item (4) in Theorem 7.8.17, then the family { P N } N ∈N
where
N := R1 ◦ PN ◦ E1 for each N ∈ N
P (7.8.142)
becomes a good approximation of the identity on the space Y1 (relative to the ambient
Y2 ). In view of this and the equivalence of items (4) and (1) in Theorem 7.8.17 we
may then once again arrive at the conclusion that E p (Y1 ) = Y2 . 

We conclude the abstract considerations in this section by establishing the follow-

ing criterion for identifying the Banach envelope of a given dual-rich quasi-normed
space.
 
Proposition 7.8.20 Assume X, · X is a quasi-normed space whose dual separates
its points and suppose Y,  · Y is a Banach space with the property that X ⊆ Y
and    
X,  · X → Y,  · Y continuously and densely. (7.8.143)
Then
E1 (X) = Y if and only if X ∗ = Y ∗ (7.8.144)

(with the last equality understood in the sense that Y∗  ΛX ∈ X ∗ is an
 Λ →
isomorphism, with bounded inverse).

Proof The fact that E1 (X) = Y implies X ∗ = Y ∗ is a consequence of Proposi-

tion 7.8.11 (used with p := 1). Conversely, assume X ∗ = Y ∗ . Then
 the identification

E1 (X) = Y follows from Definition 7.8.4 (bearing in mind that Y,  · Y is a Banach
space and that (7.8.143) holds) as soon as we show that
442 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
 
 x  ≈  xY uniformly for x ∈ X. (7.8.145)
X,1

To this end, observe first that since (7.8.143) implies the existence of a constant
C ∈ (0, ∞) with the property that  · Y ≤ C · X on X, we may invoke (7.8.16) to
conclude that  
 · Y ≤ C  · X,1 on X. (7.8.146)
As far as (7.8.145) is concerned, there remains to establish a similar inequality
in the opposite direction. With this goal in mind, pick an arbitrary x ∈ X. Then
ιX (x) belongs to E1 (X) which, according to Proposition 7.8.6, is a Banach space
when equipped with  ·  E1 (X). As such,
 ∗ the Hahn-Banach Theorem guarantees the
existence of a functional Λ ∈ E1 (X) with
      
Λ(E1 (X))∗ ≤ 1 and Λ ιX (x)  = ιX (x) E1 (X) =  x X,1 . (7.8.147)
 ∗
Recall from the proof of Proposition
 7.8.11 that the identification of E1 (X) with
∗
X ∗ is via the assignment E1 (X)  Θ → Θ ◦ ιX ∈ X ∗ . Also, since we are presently
∗ ∗
assuming that X = Y (in the manner described in the statement), it follows that
Λ ◦ ιX ∈ X ∗ extends to a functional Λ 2 ◦ ιX ∈ Y ∗ with
 
Λ2◦ ιX Y ∗ ≤ CΛ ◦ ιX X ∗ ≤ CΛ(E1 (X))∗ ≤ C. (7.8.148)

At this point, we may combine (7.8.147) and (7.8.148) to estimate

          
 x  = Λ ιX (x)  =  Λ ◦ ιX (x) =  Λ
2 ◦ ιX (x)
X,1
 
≤ Λ2◦ ιX Y ∗  xY ≤ C xY . (7.8.149)

Together with (7.8.146), this completes the proof of (7.8.145). 

The considerations up to the present point in this section have been of a purely
abstract nature. We shall now demonstrate the value of this body of results by
identifying the envelopes of certain concrete Besov and Triebel-Lizorkin spaces. We
do so in a sequence of three theorems, starting with the following result.

Theorem 7.8.21 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then whenever

−1 < s < 1, max n , n+s

n−1 n−1
< q < ∞, max n , n+s
n−1 n−1
< p < p∗ ≤ 1,
(7.8.150)
one has  p,q  p ∗, p ∗
E p∗ Fs (Σ, σ) = B (Σ, σ). (7.8.151)
s−(n−1)( p1 − p1∗ )

As a particular case, corresponding to s := 0 and q := 2, one has

  p ∗, p ∗
E p∗ h p (Σ, σ) = B (Σ, σ) whenever n−1
< p < p∗ ≤ 1, (7.8.152)
−(n−1)( p1 − p1∗ ) n
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 443

and   p ∗, p ∗
E p∗ H p (Σ, σ) = B (Σ, σ)
−(n−1)( p1 − p1∗ )
(7.8.153)
provided Σ is compact and ∗
n < p < p ≤ 1.
n−1

p,q
In particular, the Banach envelope of Fs (Σ, σ) is
 p,q 
E1 Fs (Σ, σ) = B1,1 (Σ, σ), (7.8.154)
s−(n−1)( p1 −1)

hence, corresponding to s := 0 and q := 2,

the Banach envelope
 of the local Hardy space
h p (Σ, σ) is E1 h p (Σ, σ) = B1,1 (Σ, σ), (7.8.155)
1
−(n−1)( p −1)

and
if the set Σ is compact
 then
 the Banach envelope of
H p (Σ, σ) is E1 H p (Σ, σ) = B1,1 (Σ, σ). (7.8.156)
1
−(n−1)( p −1)

Proof The strategy is to invoke Theorem 7.8.17 (cf. also Remark 7.8.18) with
p,q p ∗, p ∗
E := Fs (Σ, σ), F := B (Σ, σ), (7.8.157)
s−(n−1)( p1 − p1∗ )

and with p∗ ∈ (0, 1] presently playing the role of the old p. As noted in (3) in
Remark 7.1.6, E, F are quasi-Banach spaces. In fact, from item (4) in Remark 7.1.6
we know that F is actually a p∗ -Banach space. Also, Proposition 7.7.6 guarantees
that the duals of E, F separate the points in these spaces, while Corollary 7.7.5
together with Lemma 7.1.10 ensure that E ⊆ F and the inclusion ι : E → F is
continuous with dense range.
As regards the existence of a good approximation of the identity for the space
E, relative to the ambient F (in the sense described in item (4) of Theorem 7.8.17),
consider the family {PN } N ∈N with each PN as in (7.3.37) (for some β, γ ∈ (0, 1)
sufficiently close to 1). Then (7.8.99) holds for such a choice, thanks to (7.3.36).
As far as the applicability of Theorem 7.8.17 is concerned, there remains to verify
p,q
(7.8.98). To this end, fix an arbitrary distribution f ∈ E = Fs (Σ, σ) along with
some N ∈ N. We shall make use of (7.3.24) and (7.3.15) to recast (7.3.37) as

  N
(k,τ)
s
σ(Qτk,ν )− n−1 + p − 2 λτk,ν ψτk,ν
1 1
PN f =
k ∈Z τ ∈I N ν=1
N ≥k ≥
κΣ k

  N
(k,τ)
s ∗
− n−1 + p1∗ − 12
= σ(Qτk,ν ) λτk,ν ψτk,ν . (7.8.158)
k ∈Z τ ∈IkN ν=1
N ≥k ≥κΣ

where the last equality simply uses the abbreviation s∗ := s −(n−1)( p1 − p1∗ ). Observe
that if we now consider the numerical sequence
444 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

λ := λτk,ν Qτk,ν ∈D∗ (Σ)

, (7.8.159)
∗
then the  p -norm of the coefficients intervening in the last line of (7.8.158) may be
estimated as follows:
   N
(k,τ)  1/p∗
 s ∗  p∗
|λτk,ν 
− n−1 + p1∗ − 12
σ(Qτk,ν )
k ∈Z τ ∈I N ν=1
N ≥k ≥
κΣ k
   N
(k,τ)  1/p∗
 ks ∗ − 1 1 p ∗
≤C 2 σ(Qτk,ν ) p ∗ 2 |λQ k,ν |
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

= Cλb p ∗, p ∗ (Σ) = CΨ( f )b p ∗, p ∗ (Σ) ≤ C f B p ∗ , p ∗ (Σ,σ)

s∗ s∗ s∗

= C f B p ∗ , p ∗ (Σ,σ)
, (7.8.160)
1− 1 )
s−(n−1)( p
p∗

∗ s∗
thanks to the fact that 2s ≈ σ(Qτk,ν )− n−1 , the definitions made in (7.2.28), (7.3.25)
and, finally, (7.3.28). Above, C ∈ (0, ∞) is a constant independent of f and N. In
turn, from (7.8.158), (7.8.160), (7.3.17), and (7.8.3) we conclude that there exists
some large constant C ∈ (0, ∞), independent of f and N, such that if f  0 then
PN f !
∈ BE (0, 1) p∗
. (7.8.161)
C f B p ∗ , p ∗ (Σ,σ)
1− 1 )
s−(n−1)( p
p∗

In concert, (7.8.161) and (7.8.6) ultimately prove (now also allowing f = 0) that
 
PN f  ∗ ≤ C f  p ∗, p ∗ (7.8.162)
E, p B (Σ,σ) 1− 1 )
s−(n−1)( p
p∗

for some constant C ∈ (0, ∞) independent of f and N. This establishes (7.8.98) in

the current context (again, with p replaced by p∗ ). Hence, for E, F as in (7.8.157), we
may now invoke the equivalence of items (1) and (4) in Theorem 7.8.17 to conclude
that (7.8.151) holds. Finally, (7.8.152), (7.8.153) are particular cases of (7.8.151)
(cf. (7.1.56) and (7.1.60)). 

We pause to record the following intriguing observation.

Remark 7.8.22 Assume that Σ ⊆ Rn is a compact Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, pick p ∈ n−1 n , 1 and introduce α := (n − 1)( p − 1). Then, even
1
α
though the dual of H (Σ, σ) is C (Σ) (see Theorem 4.6.1), from Proposition 7.8.12
p

and (7.8.156) it follows that, for each f ∈ H p (Σ, σ), the quantity
 
sup   f , φ  : φ ∈ C α (Σ), φC α (Σ) = 1
(7.8.163)
is equal to  f B1,1 (Σ,σ) , and not to  f  H p (Σ,σ) .
−α
7.8 Envelopes of Besov and Triebel-Lizorkin Spaces 445

This is a striking

 manifestation of the failure of the Hardy space H p (Σ, σ) with
p ∈ n−1 n , 1 to be a locally convex space4 (hence, in particular, a genuine normed
vector space; compare with (1.2.21)-(1.2.22)).
Here is the second theorem alluded to earlier, pertaining to envelopes of Besov
spaces.
Theorem 7.8.23 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, assume

−1 < s < 1, 0 < q < ∞, max n , n+s

n−1 n−1
< p < p∗ ≤ 1,
1  (7.8.164)
satisfy − 1 + (n − 1) p − 1
p∗ < s.

Then  p,q  p ∗, p ∗
E p∗ Bs (Σ, σ) = B (Σ, σ). (7.8.165)
s−(n−1)( p1 − p1∗ )

Proof This is established by reasoning much as in the proof of Theorem 7.8.21,

p,q p ∗, p ∗
now choosing E := Bs (Σ, σ) and F := B 1 (Σ, σ), and relying on the
1
s−(n−1)( p − p ∗ )
embedding (7.7.13). 
We conclude with the following result in which further envelopes of Besov spaces
are concretely identified.
Theorem 7.8.24 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then, whenever

−1 < s < 1, max n , n+s

n−1 n−1
< p ≤ 1, 0 < q ≤ p∗ ≤ p, (7.8.166)

it follows that
 p,q  p, p ∗
E p∗ Bs (Σ, σ) = Bs (Σ, σ). (7.8.167)
Proof Again, this is justified by arguing along the lines of the proof of Theo-
p,q p, p ∗
rem 7.8.21, now taking E := Bs (Σ, σ) and F := Bs (Σ, σ), relying on the em-
bedding (7.7.1), and keeping in mind that, as mentioned in item (4) of Remark 7.1.6,
p, p ∗
the space Bs (Σ, σ) is p∗ -Banach granted the assumptions made in (7.8.166). 
Lastly, we wish to remark that a special case of Theorems 7.8.21-7.8.21 worth
considering is when Σ := Rn−1 × {0} (a scenario in which one may find it con-
venient to express the resulting envelope identifications with n replaced by n + 1
and Σ simply regarded as being Rn ). If this venue is pursued then certain redundant
limitations on the indices involved are inherited for the formulation of the results in
Theorems 7.8.21-7.8.21 which deal with much rougher settings. In order to avoid
such an issue, in the theorem below (which refines results from [127]) we shall di-
rectly tackle the task of computing envelopes of Besov and Triebel-Lizorkin spaces
in the entire Euclidean setting, albeit using the same blue-print as in the proofs of
Theorems 7.8.21-7.8.21.
4 cf. [62, Theorem 6.2, p. 71] for a proof in the Euclidean setting
446 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

Theorem 7.8.25 For Besov and Triebel-Lizorkin spaces in Rn , one has the following
envelope identifications:

if s ∈ R, 0 < q < ∞, 0 < p < p∗ ≤ 1,

 p,q  p ∗, p ∗ (7.8.168)
then E p∗ Fs (Rn ) = B 1 (R ),
n
s−n( p − p ∗ )
1

if s ∈ R, 0 < q < ∞, 0 < p < p∗ ≤ 1,

 p,q  p ∗, p ∗ (7.8.169)
then E p∗ Bs (Rn ) = B 1 (R ),
n
s−n( p − p ∗ )
1

and
if s ∈ R and 0 < q ≤ p∗ ≤ p ≤ 1,
 p,q  p, p ∗
(7.8.170)
then E p∗ Bs (Rn ) = Bs (Rn ).

Proof All formulas in the statement of the theorem may be established by reasoning
along the lines of the proofs of Theorems 7.8.21-7.8.21. Once again, the idea is to
implement the abstract criterion from Theorem 7.8.17, whose applicability in the
present setting hinges on two ingredients. First, one requires embeddings for Besov
and Triebel-Lizorkin spaces in Rn , and these are standard results (cf., e.g., [166]).
Second, one needs to ensure the existence of a good approximation of the identity
(in the sense described in item (4) of Theorem 7.8.17). In this regard, one may
consider operators PN , with N ∈ N, mapping a distribution into its N-th partial
sum associated with its series expansion with respect to a wavelet basis (which is
sufficiently smooth and has sufficiently many vanishing moments). The coefficients
in such an expansion are known to characterize the size of said distribution measured
on Besov and Triebel-Lizorkin spaces (cf., e.g., [57], [59]) and this makes the family
{PN } N ∈N an ideal replacement for the frame decomposition operators used earlier
in the proofs of Theorems 7.8.21-7.8.21. 

7.9 Intrinsic Characterizations of Besov Spaces

It is useful to provide intrinsic characterizations of the Besov spaces originally intro-

duced in Definition 7.1.5, without recourse to the conditional expectation operators
(7.1.30). A first result of this nature is contained in the theorem below.

Theorem 7.9.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, assume 0 < s < 1, 1 ≤ p ≤ ∞, 0 < q ≤ ∞. Then
7.9 Intrinsic Characterizations of Besov Spaces 447

 f Bsp, q (Σ,σ)

≈  f  L p (Σ,σ) (7.9.1)
 3 ∬ 5 qp  q1

∞
+ 2 jsq
2 j(n−1)
| f (x) − f (y)| dσ(x) dσ(y)
p

j=0 (x,y)∈Σ×Σ
|x−y |<2− j

≈  f  L p (Σ,σ) (7.9.2)
 3∫ ∫ 5 qp  q1

∞  p
+ 2 jsq
2 j(n−1)
| f (x) − f (y)| dσ(y) dσ(x)
j=0 Σ y ∈Σ
|x−y |<2− j

p,q
uniformly for f ∈ Bs (Σ, σ). In particular, corresponding to p = q = ∞,

 f B∞,∞
s (Σ,σ) ≈  f  L ∞ (Σ,σ) (7.9.3)

| f (x)− f (y) |
+ σ-ess sup |x−y | s : x, y ∈ Σ, 0 < |x − y| < 1 ,

uniformly for f ∈ Bs∞,∞ (Σ, σ).

Proof This is proved in [204, Theorem 3.8, p. 732]. 
Corollary 7.9.2 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Then, in the range 1 ≤ p, q ≤ ∞ and 0 < s < 1, the Besov scale
p,q
Bs (Σ, σ) considered by A. Jonsson and H. Wallin in [106] coincides with the scale
of Besov spaces introduced by Y. Han and D. Yang in [92].
Proof This follows from Theorem 7.9.1 and [106, Proposition 3, p. 125]. 
Here is a companion result to Theorem 7.9.1.
Theorem 7.9.3 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also suppose that
1 
n < p ≤ 1 and (n − 1) p − 1 < s < 1.
n−1
(7.9.4)

Then there exists a finite constant C = C(Σ, p, s) > 0 with the property that for
p, p
each function f ∈ Bs (Σ, σ) one has f ∈ L q (Σ, σ) and
∫ ∫  1/p
| f (x) − f (y)| p
 f  L q (Σ,σ) + dσ(x) dσ(y) ≤ C f Bsp, p (Σ,σ), (7.9.5)
Σ Σ |x − y|
n−1+sp

where 1 
s −1
q := p − n−1 ∈ (1, ∞). (7.9.6)
Moreover, whenever
448 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

n−1
n < p ≤ ∞, 0 < q ≤ ∞, 0 < s < 1, (7.9.7)
p,q
there exists C = C(Σ, p, q, s) ∈ (0, ∞) such that for each f ∈ Bs (Σ, σ) one has
(with natural alterations when p = ∞ or q = ∞)
∫ ∫ ⨏  1/q
∞  q/p dt
 f  L p (Σ,σ) + | f (x) − f (y)| dσ(y) dσ(x)
p
0 Σ Σ∩B(x,t) t 1+sq

≤ C f Bsp, q (Σ,σ) . (7.9.8)

As a consequence, corresponding to q := p, whenever n−1 n < p ≤ ∞ and 0 < s < 1

there exists some constant C = C(Σ, p, s) ∈ (0, ∞) with the property that for each
p, p
f ∈ Bs (Σ, σ) one has (with a natural alteration when p = ∞)
∫ ∫  1/p
| f (x) − f (y)| p
 f  L (Σ,σ) +
p dσ(x) dσ(y) ≤ C f Bsp, p (Σ,σ) . (7.9.9)
Σ Σ |x − y|
n−1+sp

Finally, in the case when 1 ≤ p ≤ ∞ and 0 < s < 1 one actually has (with a
natural alteration when p = ∞)
⎧
⎪ f ∈ L p (Σ, σ) and
⎪
⎪
⎪
⎨ ∫∫
⎪
p, p
f ∈ Bs (Σ, σ) ⇐⇒ | f (x) − f (y)| p  p1 (7.9.10)
⎪
⎪ dσ(x) dσ(y) <∞
⎪
⎪ |x − y| n−1+sp
⎪
⎩ Σ Σ
as well as
∫ ∫  p1
| f (x) − f (y)| p
 f Bsp, p (Σ,σ) ≈  f  L p (Σ,σ) + dσ(x) dσ(y) ,
Σ Σ |x − y| n−1+sp (7.9.11)
p, p
uniformly for f ∈ Bs (Σ, σ).

Several comments are in order. First, we wish to point out that for each fixed r > 0
an equivalent quasi-norm to the expression in the right-hand side of (7.9.11) is given
by
1/p
6 ∬ 9
7 | f (x) − f (y)| p :
 f  L p (Σ,σ) + 77 dσ(x) dσ(y) :
: . (7.9.12)
7 |x − y| n−1+sp
:
(x,y)∈Σ×Σ
8 |x−y |<r ;
This is a simple consequence of the last inequality in [133, (7.2.5)]. Let us also bring
in the L p -modulus of continuity at scale t ∈ (0, ∞) of a given σ-measurable function
f : Σ → R, defined as
7.9 Intrinsic Characterizations of Besov Spaces 449
∬  1/p
ω p ( f ; t) := | f (x) − f (y)| p dσ(x) dσ(y) . (7.9.13)
(x,y)∈Σ×Σ
|x−y |<t

Observe that (0, ∞)  t → ω p ( f ; t) ∈ [0, ∞] is nondecreasing, and

∬
ω p ( f ; t) p = | f (x) − f (y)| p dσ(x) dσ(y)
(x,y)∈Σ×Σ
|x−y |<t
∬
 
≤c | f (x)| p + | f (y)| p 1 |x−y |<t dσ(x) dσ(y)
Σ×Σ
∬
= 2c | f (x)| p 1 |x−y |<t dσ(x) dσ(y)
Σ×Σ
∫ ∫ 
= 2c | f (x)| p 1 |x−y |<t dσ(y) dσ(y)
Σ Σ
p
≤ Ct n−1  f  L p (Σ,σ), (7.9.14)

hence

ω p ( f ; t) ≤ Ct (n−1)/p  f  L p (Σ,σ) for each t ∈ (0, ∞). (7.9.15)

In addition, for each s, p > 0 we have

∫ ∞ ∫ ∞3 ∬ 5
ω p ( f ; t) p dt dt
= | f (x) − f (y)| dσ(x) dσ(y)
p
0 t n−1+sp t 0 t n+sp
(x,y)∈Σ×Σ
|x−y |<t

∫ 3∬ 5
∞
dt
= | f (x) − f (y)| 1 |x−y |<t dσ(x) dσ(y)
p
0 t n+sp
Σ×Σ
∬ ∫ ∞
dt 
= | f (x) − f (y)| p
dσ(x) dσ(y)
|x−y | t n+sp
Σ×Σ
∬
1 | f (x) − f (y)| p
= dσ(x) dσ(y). (7.9.16)
n − 1 + sp |x − y| n−1+sp
Σ×Σ

Ultimately, this proves that the expression in the right-hand side of (7.9.11) is further
equivalent to the quasi-norm
450 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets
∫ ∞ ω p ( f ; t) p dt  1/p
f L p (Σ,σ) + . (7.9.17)
0 t n−1+sp t
Let us also note that since for any s, p > 0 the monotonicity of the L p -modulus of
continuity function t → ω p ( f ; t) (cf. (7.9.13)) permits us to estimate
∬
| f (x) − f (y)| p
dσ(x) dσ(y)
|x − y| n−1+sp
(x,y)∈Σ×Σ
|x−y |<1/2 ∬

∞
| f (x) − f (y)| p
= dσ(x) dσ(y)
j=1
|x − y| n−1+sp
(x,y)∈Σ×Σ
2− j−1 ≤ |x−y |<2− j


∞ ∬
≤C 2 j(n−1+sp)
| f (x) − f (y)| p dσ(x) dσ(y)
j=1 (x,y)∈Σ×Σ
|x−y |<2− j


∞
=C 2 j(n−1+sp) ω p ( f ; 2−j ) p
j=1


∞ ∫ 2− j+1
dt
≤C 2 j(n−1+sp) ω p ( f ; t) p
j=1 2− j t

∞ ∫
 2− j+1 ω p ( f ; t) p dt
≤C
j=1 2− j t n−1+sp t
∫ 1 ω p ( f ; t) p dt
=C , (7.9.18)
0 t n−1+sp t
we conclude from the remarks made in relation to (7.9.12) and (7.9.17) that yet
another equivalent quasi-norm to the expression in the right-hand side of (7.9.11) is
given by
∫ 1 ω p ( f ; t) p dt  1/p
 f  L p (Σ,σ) + . (7.9.19)
0 t n−1+sp t
In the same vein, from the work in [106] and Corollary 7.9.2 we conclude that
p,q
in the range 1 ≤ p, q ≤ ∞ and 0 < s < 1 the Besov space Bs (Σ, σ) consists of
functions f ∈ L (Σ, σ) having the property that
p

 3 ∬ 5 q/p  1/q

∞
2 jsq
2 j(n−1)
| f (x) − f (y)| dσ(x) dσ(y)
p
< +∞, (7.9.20)
j=0 (x,y)∈Σ×Σ
|x−y |<2− j
7.9 Intrinsic Characterizations of Besov Spaces 451

and equivalent quasi-norms on this space are given by


∞ 1/q
 f  L p (Σ,σ) + 2 jq[(n−1)/p+s] ω p ( f ; 2−j )q
j=0
∫ ∞ ω p ( f ; t)q dt  1/q
≈ f L p (Σ,σ) +
0 t (n−1)q/p+sq t
∫ 1 ω p ( f ; t)q dt  1/q
≈  f  L p (Σ,σ) + . (7.9.21)
0 t (n−1)q/p+sq t
For more on this topic, see also [64].
Our last comment is that, in the case when Σ is the boundary of a Lipschitz domain
Ω ⊆ Rn , one may describe the class of Besov spaces on Σ = ∂Ω via localization and
pull-back to the Euclidean space Rn−1 . This procedure has the distinct advantage that
a number of useful properties valid in the latter setting automatically carry over to
the scale of Besov spaces defined on Lipschitz surfaces (see, e.g., the discussion in
[141, §2.7, p. 88]). To illustrate this point we wish to note here the following intrinsic
characterization (further refining (7.9.10)-(7.9.11) in the class of Lipschitz surfaces)
which is a direct consequence of (9.1.14) and the aforementioned principle:

Suppose Ω ⊂ Rn is a Lipschitz domain and abbreviate σ := H n−1 ∂Ω.

Then, if (n − 1)/n < p ≤ ∞ and (n − 1)(1/p − 1)+ < s < 1, one has
∫ ∫  1/p
| f (x) − f (y)| p
f p, p
B s (∂Ω,σ) ≈ dσ(x) dσ(y) (7.9.22)
∂Ω ∂Ω |x − y| n−1+sp

+  f  L p (∂Ω,σ) .

We now turn to the task of giving the proof of Theorem 7.9.3.

Proof of Theorem 7.9.3 In general, given any radius R > 0 and any point x ∈ Σ
we shall abbreviate Δ(x, R) := B(x, R) ∩ Σ. Fix some η ∈ (s, 1]. As far as (7.9.5)
is concerned, thanks to Theorem 7.2.7 it suffices to show that there exists a finite
constant C = C(Σ, p, s, η) > 0 such that
∫ ∫
|a(x) − a(y)| p
dσ(x) dσ(y) ≤ C, (7.9.23)
Σ Σ |x − y|
n−1+sp

and ∫
|a(x)| q dσ(x) ≤ C, (7.9.24)
Σ
whenever a : Σ → R is a η-smooth block of type (p, s). More specifically, assume
that (cf. (7.2.3)-(7.2.5))
452 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

supp a ⊆ Δ(x0, r) for some x0 ∈ Σ and r > 0, (7.9.25)

n−1
sup |a(x)| ≤ r s− p and (7.9.26)
x ∈Σ
n−1
|a(x) − a(y)| ≤ r s−η− p |x − y| η, ∀x, y ∈ Σ. (7.9.27)

Then (7.9.24) is an immediate consequence of (7.9.6), (7.9.25)-(7.9.26), and the fact

that Σ is upper Ahlfors regular.
To prove (7.9.23) we proceed by splitting the respective double integral into four
double integrals corresponding to the following regions of integration:

R1 := {(x, y) ∈ Σ × Σ : x ∈ Σ \ Δ(x0, 2r), y ∈ Δ(x0, r)}, (7.9.28)

R2 := {(x, y) ∈ Σ × Σ : x ∈ Δ(x0, r), y ∈ Σ \ Δ(x0, 2r)}, (7.9.29)
R3 := {(x, y) ∈ Σ × Σ : x ∈ Δ(x0, 2r), y ∈ Δ(x0, r)}, (7.9.30)
R4 := {(x, y) ∈ Σ × Σ : x ∈ Σ \ Δ(x0, r), y ∈ Σ \ Δ(x0, 2r)}. (7.9.31)

First, we estimate the part of (7.9.23) corresponding to R1 . Starting with (7.9.25)

and (7.9.26) we may write
∫∫
|a(x) − a(y)| p
dσ(x) dσ(y)
R1 |x − y|
n−1+sp

∫ ∫
r sp−(n−1)
≤ dσ(x) dσ(y)
Δ(x0,r) Σ\Δ(x0,2r) |x − y|
n−1+sp

∫
r sp−(n−1)
≤ CintΔ(x0,r) dσ(x) dσ(y)
Σ\Δ(x0,2r) |x − x0 |
n−1+sp
∫
dσ(x)
= Cr sp
Σ\Δ(x0,2r) |x − x0 | n−1+sp
≤ Cr sp r −sp = C, (7.9.32)

for some finite constant C = C(Σ, p, s) > 0. For the second inequality in (7.9.32)
we have used the fact that if (x, y) ∈ R1 then |x − y| ≈ |x − x0 |, while the second
to the last inequality in (7.9.32) is based on the second inequality in [133, (7.2.5)],
presently used with d := n − 1 and δ := sp > 0. A similar computation also gives
that ∫∫
|a(x) − a(y)| p
dσ(x) dσ(y) ≤ C. (7.9.33)
R2 |x − y|
n−1+sp

Next, we estimate the part of (7.9.23) corresponding to R3 . For this portion we start
by using (7.9.27) to write
7.9 Intrinsic Characterizations of Besov Spaces 453
∫∫
|a(x) − a(y)| p
dσ(x) dσ(y)
R3 |x − y| n−1+sp
∫ ∫
r sp−η p−(n−1)
≤C dσ(x) dσ(y)
Δ(x0,r) Δ(x0,2r) |x − y|
n−1+sp−η p
∫ ∫ 
dσ(y)
≤ Cr sp−η p−(n−1)
dσ(x)
Δ(x,3r) |x − y|
n−1+sp−η p
Δ(x0,2r)

≤ Cr sp−η p−(n−1) r n−1 r −sp+η p = C, (7.9.34)

for some finite constant C = C(Σ, p, s, η) > 0. For the second inequality in (7.9.34)
we have used the fact that if x ∈ Δ(x0, 2r) then for any y ∈ Δ(x0, r) we have that
y ∈ Δ(x, 3r), while the last inequality in (7.9.34) is based on the first inequality in
[133, (7.2.5)], currently employed with d := n − 1 and δ := ηp − sp (the fact that
s < η ensures that δ > 0). Finally, the support condition on a (cf. (7.9.25)) trivially
forces ∫∫
|a(x) − a(y)| p
dσ(x) dσ(y) = 0. (7.9.35)
R4 |x − y|
n−1+sp

Summing up the estimates (7.9.32), (7.9.33), (7.9.34), (7.9.35) ultimately yields

(7.9.23).
Next, the estimate claimed in (7.9.8) (under the conditions stipulated in (7.9.7))
is implied by [152, Proposition 4.2, p. 275]. In turn, (7.9.9) follows from (7.9.8) with
q := p via a simple application of Fubini’s Theorem (keeping in mind that Σ is upper
Ahlfors regular). Finally, (7.9.10)-(7.9.11) are consequences of Corollary 7.9.2,
(7.9.1), and [106, Proposition 1, p. 114]. 

We conclude by establishing an embedding result which is going to be relevant

in the context of singular integral operators acting on Besov spaces.

Proposition 7.9.4 Suppose Σ ⊆ Rn is a closed Ahlfors regular and abbreviate

σ := H n−1 Σ. Also, assume that
1 
n < p < ∞ and (n − 1) p − 1 + < s < 1.
n−1
(7.9.36)

Then
p, p σ(x) 
Bs (Σ, σ) → L 1 Σ, . (7.9.37)
1 + |x| n−1
Proof In the case when 1 ≤ p < ∞ (which entails 0 < s < 1) we may write

p, p σ(x) 
Bs (Σ, σ) → L p (Σ, σ) → L 1 Σ, , (7.9.38)
1 + |x| n−1

thanks to the results in the second part of Theorem 7.9.3. In the case n−1
n < p < 1,
define
454 7 Besov and Triebel-Lizorkin Spaces on Ahlfors Regular Sets

1 
s∗ := s − (n − 1) p − 1 ∈ (0, s), (7.9.39)
then invoke Theorem 7.7.4 and (7.9.10)-(7.9.11) (with p := 1) to conclude that

p, p σ(x) 
Bs (Σ, σ) → Bs1,1 (Σ, σ) → L 1 (Σ, σ) → L 1 Σ, , (7.9.40)
∗
1 + |x| n−1
as wanted. 
Chapter 8
Boundary Traces from Weighted Sobolev Spaces
into Besov Spaces

The starting point in this chapter is the consideration of boundary traces from a
weighted Sobolev space defined in the entire Euclidean space in §8.1. The ordinary
scale of Sobolev spaces W k, p (Ω) in a given open subset Ω of Rn , defined intrinsically,
via the demand that weak partial derivatives up to an order k ∈ N exist and are p-th
power integrable in Ω, for p ∈ (1, ∞), may be considerably enhanced by replacing
the ordinary Lebesguemeasure L n by the weighted measure μa := dist(·, ∂Ω)ap L n ,
where a ∈ − p1 , 1 − p1 is a fixed parameter. The resulting scale of weighted Sobolev
k, p
spaces is denoted by Wa (Ω).
1, p
Of particular interest is the space Wa (Ω). This weighted scale of Sobolev spaces
exhibits a number
 of salient
 features. First, whenever ∂Ω is Ahlfors regular, the
range a ∈ − p1 , 1 − p1 ensures that dist(·, ∂Ω)ap is a Muckenhoupt weight in
Rn , belonging to the class Ap (Rn, L n ). Second, a byproduct of the presence of the
k, p
weight is that functions in Wa (Ω) exhibit k − a “units” of smoothness measured
in a plain (unweighted) L p sense (which is more, respectively less, than precisely k
k, p
units, depending on whether a < 0, or a > 0). In particular, Wa (Ω) allows for the
consideration of fractional amounts of smoothness, without the use of the Fourier
transform (as has traditionally been the case). Third, there is a rather satisfactory
trace/extension theory associated with this brand of weighted Sobolev spaces in
Ahlfors regular domains. Among other things, the boundary trace operator maps
1, p p, p
Wa (Ω) boundedly and surjectively into the Besov space Bs (∂Ω, σ), where the
index s := 1 − a − 1/p ∈ (0, 1). Fourth, certain conormal derivatives may be
defined naturally in relation to functions belonging to this scale of weighted Sobolev
spaces. Fifth, boundary-to-domain singular integral operators (of boundary layer
type), associated with weakly elliptic systems in an (ε, δ)-domain Ω with compact
p, p
boundary, turn out to act naturally from the Besov space Bs (∂Ω, σ) into the
1, p
weighted Sobolev space Wa (Ω) (with a := 1 − s − 1/p). This is going to be
particularly relevant in the treatment of boundary value problems later on, in [137].
We shall establish most of these (and other related) results in the present chapter.
The starting point is the consideration of boundary traces from a weighted Sobolev

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 455
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_8
456 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

space defined in the entire Euclidean space in §8.1 (the bulk of the proofs are
contained in §8.2). Next, in §8.3, we consider traces from weighted Sobolev spaces
defined in a given (ε, δ)-domain Ω ⊆ Rn by relying on P. Jones’ extension theorem
(originally proved in [103, Theorem 1, p. 73], then further generalized as to be
applicable to Muckenhoupt weighted Sobolev spaces in [29]) to reduce matters to
the full Euclidean setting considered earlier. The next order of business is to construct
extension operators from boundary Besov spaces into our weighted Sobolev spaces,
something we accomplish in §8.4. Weighted Sobolev spaces of negative order,
duality, and the conormal derivative operators are discussed in §8.5. In the last
section of this chapter (§8.6), we study boundary traces from weighted “maximal”
Sobolev spaces, defined by requiring the membership of distributional derivatives to
solid maximal Lebesgue spaces introduced in [133, §6.6].

8.1 Traces from Weighted Sobolev Spaces Defined in R n

We begin by defining weighted Sobolev spaces in the entire Euclidean ambient.

Consider a weight w in Rn , i.e., a Lebesgue measurable function w : Rn → [0, ∞]
with the property that 0 < w(x) < ∞ for L n -a.e. x ∈ Rn . Given p ∈ (0, ∞), define
the weighted Sobolev spaces of order k ∈ N0 in Rn as
 
k,1 n
W k, p (Rn, wL n ) := f ∈ Wloc (R ) :  f W k, p (Rn,w L n ) < +∞ , (8.1.1)

where, with the derivatives taken in the sense of distributions, for any given function
f ∈ W k, p (Rn, wL n ) we have set
 ∫ 1/p
 f W k, p (Rn,w L n ) := |∂ β f | p w dL n . (8.1.2)
|β | ≤k Rn

A prominent role is played by Muckenhoupt weighted Sobolev spaces, defined

as above with p ∈ (1, ∞) and w ∈ Ap (Rn, L n ). In this regard, we wish to record a
useful density result to the effect that

Cc∞ (Rn ) → W k, p (Rn, wL n ) densely

(8.1.3)
if 1 < p < ∞ and w ∈ Ap (Rn, L n ).

This may be justified via smooth truncation and mollification, using the boundedness
of the Hardy-Littlewood maximal operator on L p (Rn, wL n ) when p ∈ (1, ∞) and
w ∈ Ap (Rn, L n ). As a consequence of (8.1.3) and [133, Proposition 8.7.4] we
therefore obtain
if Σ ⊆ Rn is closed and Ahlfors regular then
 
Cc∞ (Rn ) → W k, p Rn, δΣ L n has dense range
ap
(8.1.4)
whenever k ∈ N0, 1 < p < ∞, and − 1
p < a < 1− p.
1
8.1 Traces from Weighted Sobolev Spaces Defined in R n 457

Moving on, for each α > 0 recall that the Bessel kernel of order α is the function
Gα defined by (with F denoting the Fourier transform)
 
Gα (x) := Fξ→x (1 + |ξ | 2 )−α/2 (x) for x ∈ Rn . (8.1.5)

The Bessel kernel is a radial, positive, decreasing function (of |x|), belongs to
L 1 (Rn, L n ) as well as to C ∞ (Rn \ {0}), and has the property that for any multi-index
β ∈ N0n there exists a finite constant C = C(n, α, β) > 0 such that the following
estimate holds:

C|x| α− |β |−n for all x ∈ Rn \ {0} if | β| > α − n,

(∂ β Gα )(x) ≤ (8.1.6)
Ce−c |x | for all |x| ≥ 1, for some fixed c > 0.

Call a function u the Bessel potential of order α of f provided

u = Gα ∗ f in Rn . (8.1.7)

For each α ∈ (0, ∞), p ∈ (1, ∞), and w ∈ Ap (Rn, L n ), define the weighted Bessel
potential space of order α in Rn as
p
Lα (Rn, wL n ) := u : Rn → C Lebesgue-measurable :

u = Gα ∗ f for some f ∈ L p (Rn, wL n ) , (8.1.8)

endowed with the norm

p
u Lαp (Rn,w L n ) :=  f  L p (Rn,w L n ) for each u ∈ Lα (Rn, wL n )
(8.1.9)
expressed as u = Gα ∗ f with f ∈ L p (Rn, wL n ).

In this regard, recall from [129, Theorem 3.3, p. 104] that

for each given k ∈ N, p ∈ (1, ∞), and w ∈ Ap (Rn, L n ) we have

p
Lk (Rn, wL n ) = W k, p (Rn, wL n ) (as vector spaces) and the norms (8.1.10)
 ·  L p (Rn,w L n ) and  · W k, p (Rn,w L n ) are actually equivalent.
k

Moving on, assume next that E ⊆ Rn is a L n -measurable set which is n-thick.

Given a function f ∈ Lbdd
1 (E, L n ), we introduce the strictly defined version

of f at every point x ∈ E as
∫
⎧
⎪ 1
⎨ lim+ n
⎪ f dL n if the limit exists,
[ f ]E (x) := r→0 L (E ∩ B(x, r)) E∩B(x,r) (8.1.11)
⎪
⎪0
⎩ otherwise.

Of course, Lebesgue’s Differentiation Theorem guarantees that [ f ]E (x) = f (x) for

L n -a.e. x ∈ E, but the role of [ f ]E is to provide us with a canonical choice for a
representative in the class of functions which agree with f at L n -a.e. point in E.
458 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Our major result in this section is contained in the theorem below. This describes
the traces of functions from a weighted Sobolev space in the entire Euclidean ambient
on a given closed Ahlfors regular set as functions belonging to certain Besov space
on that set.
Theorem 8.1.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Also, fix an exponent p ∈ (1, ∞) along with a number s ∈(0, 1) and
ap
denote a := 1 − s − p1 ∈ − p1 , 1 − p1 . Then for each u ∈ W 1, p Rn, δΣ L n the limit
⨏
(Tr R n →Σ u)(x) := [u] (x) := lim+
Rn u(y) dy (8.1.12)
r→0 B(x,r)

exists at σ-a.e. point x ∈ Σ, and the induced trace operator

 ap  p, p
TrRn →Σ : W 1, p Rn, δΣ L n −→ Bs (Σ, σ) (8.1.13)

is a well-defined, linear, bounded mapping.

Of course, when a = 0, i.e., when s = 1 − p1 ∈ (0, 1), Theorem 8.1.1 becomes a
statement about traces on Σ of functions belonging to the ordinary (i.e., unweighted)
Sobolev space W 1, p (Rn ), in which case the result is well known; see (e.g., [106]).
The novelty here is the consideration of the weighted case. The very format of the
result in (8.1.13) is dictated by homogeneity considerations. The heuristic principle
is that when measuring smoothness on the L p scale, taking the trace of a function
which has a certain amount of smoothness in Rn on a “surface” of co-dimension one
decreases the smoothness by 1/p units; see item (ii) in Theorem 9.4.5 in this regard.
p, p
So having the trace of u on Σ be in Bs (Σ, σ) would require that u has s + 1/p units
of smoothness in R . Given that we insist on working with
n
 weighted
 Sobolev spaces
of order one in Rn , the disagreement quantity a := 1 − s + p1 = 1 − s − p1 needs to
be accounted for by the actual weight. By envisioning pointwise multiplication by
γ
δΣ as a fractional differentiation/integration operator of order γ, for each γ ∈ R, this
provides an explanation of the nature of the power of the weight in (8.1.13).
Theorem 8.1.1 is a culmination of a number of technical results, which we deal
with separately in the next section (the actual proof of Theorem 8.1.1 is presented at
the end of §8.2).

8.2 Technical Lemmas

Before proving the above theorem we deal with several preparatory lemmas, starting
with the following.
Lemma 8.2.1 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall that G1 is the Bessel kernel of order
 1 one in  R , defined as
n

in (8.1.5) with α := 1. Fix p ∈ (1, ∞) along with a ∈ − p , 1 − p , and consider

1
ap
f ∈ L p (Rn, δΣ L n ). Then
8.2 Technical Lemmas 459
∫
|G1 (x − y)|| f (y)| dy < +∞ for H n−1 -a.e. x ∈ Rn, (8.2.1)
Rn

and the function

∫
u(x) := (G1 ∗ f )(x) = G1 (x − y) f (y) dy at H n−1 -a.e. point x ∈ Rn (8.2.2)
Rn

satisfies the following two properties:

u may be strictly defined at H n−1 -a.e. point in Rn, (8.2.3)

[u]Rn = G1 ∗ f at H n−1 -a.e. point in Rn . (8.2.4)

Proof The unweighted case (when a := 0) has been treated in [106, Proposition 1,
p. 151] and our proof is an adaptation of this result. For starters observe that [133,
Proposition 8.7.4] and [133, (7.7.105)] (presently used with n replaced by n + 1 and
Σ := Rn ≡ Rn × {0} ⊂ Rn+1 ) ensures that

f ∈ r
Lloc (Rn, L n ). (8.2.5)
1<r <p

Consider the claim made in (8.2.1). For each number j ∈ N with j ≥ 2, decompose
f = f1, j + f2, j , where f1, j := f 1B(0, j) and f2, j := f 1Rn \B(0, j) . Then
∫ ∫
|G1 (x − y)|| f (y)| dy ≤ |G1 (x − y)|| f1, j (y)| dy
Rn Rn
∫
+ |G1 (x − y)|| f2, j (y)| dy
Rn

=: I j (x) + I I j (x). (8.2.6)

Since, by (8.2.5), we have f1, j ∈ L r (Rn, L n ) for some r > 1, the aforementioned
unweighted result from [106] applies and gives that there exists E j ⊆ Rn with
H n−1 (E j ) = 0 and such that I j (x) < +∞ for every x ∈ Rn \ E j . On the other hand,
if p := (1 − 1/p)−1 and x ∈ B(0, j − 1), we have
460 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
∫
I I j (x) ≤ C e−c |x−y | | f2, j (y)| dy
y ∈R n, |y |> j

1/p
 ∫  ∫ 1/p
 
≤ C  e−c |x−y | δΣ (y)−ap dy 

 | f2, j (y)| δΣ (y) dy 
p ap

 y ∈Rn, |x−y |>1  R n 

1/p
 ∫ 

δΣ (y)−ap
≤ CN  dy   f  L p (Rn,δ a p L n ) (8.2.7)
|x − y| N Σ

 y ∈R n, |x−y |>1 
for every N > 0. Choosing N sufficiently large then ensures (by virtue of [133,
(8.7.6)]) that the last integral in (8.2.7) is finite. Thus, I I j (x) < +∞ for every

∞
x ∈ B(0, j − 1). Set E := E j so that H n−1 (E) = 0. Then I j (x) + I I j (x) < +∞ for
j=2
every integer
∫ j ≥ 2 and every point x ∈ B(0, j − 1) \ E, which ultimately shows that
we have Rn |G1 (x − y)|| f (y)| dy ≤ I j (x) + I I j (x) < +∞ for H n−1 -a.e. x ∈ Rn . This
proves (8.2.1).
Turning our attention to the claim made in (8.2.3), let us now decompose (retaining
earlier notation) u(x) = u1, j (x) + u2, j (x) where, for each j ≥ 2, we have set
∫ ∫
u1, j (x) := G1 (x − y) f1, j (y) dy, u2, j (x) := G1 (x − y) f2, j (y) dy, (8.2.8)
Rn Rn

for H n−1 -a.e. x ∈ Rn . Since, as already observed, f1, j ∈ L r (Rn, L n ) for some r > 1,
the unweighted version of (8.2.3) applies and gives that, for each j ≥ 2, the function
u1, j can be strictly defined at each point in Rn , with the possible exception of a set
E j satisfying H n−1 (E j ) = 0. On the other hand, one can show (by recycling earlier
estimates and invoking Lebesgue’s Dominated Convergence Theorem) that for each
j ≥ 2 the function u2, j is continuous at each point in the ball B(0, j − 1). Altogether,
this shows that the function u can be strictly defined at each point in Rn not belonging
∞
to the set E := E j . Observing that H n−1 (E) = 0 proves (8.2.3). Finally, (8.2.4)
j=2
can be proved in a similar fashion. 

Lemma 8.2.2 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall that G1 is the Bessel kernel of order
 1 one in R , defined as in
n

(8.1.5) with α := 1. Then for each p ∈ (1, ∞) and θ ∈ 0, p there exists a constant
C = C(Σ, p, θ) > 0 such that
∫
|G1 (x − y)| θ p dσ(x) ≤ C for all y ∈ Rn . (8.2.9)
Σ

Proof Let p and θ be as in the statement of the lemma, and fix y ∈ Rn . Then

8.2 Technical Lemmas 461
∫
|G1 (x − y)| θ p dσ(x) (8.2.10)
Σ
∫ ∫
= |G1 (x − y)| θ p dσ(x) + |G1 (x − y)| θ p dσ(x) =: I + I I.
x ∈Σ x ∈Σ
|x−y |<1 |x−y | ≥1

To estimate I, we recall that (8.1.6) implies |G1 (z)| ≤ C|z| 1−n for all z ∈ Rn \ {0},
so that ∫
dσ(x)
I≤C ≤ C, (8.2.11)
|x − y| (n−1)θ p
x ∈Σ
|x−y |<1

where the last inequality in (8.2.11) is obtained by applying (the first formula in)
[133, (7.2.5)] (note that (n − 1)θp < n − 1 due to the choice of θ). To estimate I I, we
recall the second bound in (8.1.6), which gives the existence of constants C and c,
such that |G1 (z)| ≤ C e−c |z | if |z| ≥ 1. Therefore, we may choose N > (n − 1)/(θp)
and CN ∈ (0, ∞) so that |G1 (z)| ≤ CN |z| −N whenever |z| ≥ 1. Granted this, we may
write ∫
dσ(x)
II ≤ C ≤ C, (8.2.12)
|x − y| N θ p
x ∈Σ
|x−y | ≥1

where for the last inequality in (8.2.12) we have applied (the second formula in)
[133, (7.2.5)]. Now (8.2.9) follows by combining (8.2.10)-(8.2.12). 

Lemma 8.2.3 Assume that Σ ⊆ Rn be a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall that G1 is the Bessel kernel of order one in Rn , defined as in
(8.1.5) with α := 1. Let p, p ∈ (1, ∞) be two exponents satisfying p1 + p1 = 1, and
 
select a ∈ − p1 , 1 − p1 along with θ ∈ (0, 1). Then there exists some finite constant
C = C(Σ, p, a, θ) > 0 with the property that
∫
   
θ > n−1 a + p − 1 + =⇒ sup
1 n
|G1 (x − y)| (1−θ)p δΣ (y)−ap dy ≤ C. (8.2.13)
x ∈Σ Rn

Also, if  
1
n−1 a+ n
p −1 + <θ< a
n + 1
p (8.2.14)
then there exists a finite constant C = C(Σ, p, a, θ) > 0 such that
∫ p
p
(1−θ)p −ap
|G1 (x − t) − G1 (y − t)| δΣ (t) dt ≤ C |x − y| (n−1)(θ p−1)+p−ap−1
Rn
(8.2.15)
for all x, y ∈ Σ.

Proof We first prove (8.2.13). Fix x ∈ Σ arbitrary. Recall that by the second bound
in (8.1.6), we have |G1 (z)| ≤ C|z| −N for all |z| ≥ 1, provided N is chosen large
enough (N > np−n+1
p−1 ). Having selected such a number N, with the help of [133,
462 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

(8.7.6)] we may then write

∫
 
|G1 (x − y)| (1−θ)p δΣ (y)−ap dy
|y−x | ≥1
∫ 
δΣ (y)−ap
≤C dy ≤ C, (8.2.16)
|y−x | ≥1 |x − y| N (1−θ)p

since by assumption a < 1 − p1 . Also, making use of the first estimate in (8.1.6)
(to the effect that |G1 (z)| ≤ C|z| 1−n for all z ∈ Rn \ {0}) and of [133, (8.7.3)] with
N := (n − 1)(1 − θ)p, we obtain that
∫
 
|G1 (x − y)| (1−θ)p δΣ (y)−ap dy
|y−x | ≤1
∫ 
δΣ (y)−ap
≤C  dy ≤ C, (8.2.17)
|y−x | ≤1 |x − y| (n−1)(1−θ)p

provided (n − 1)(1 − θ)p < n − ap. The latter condition is equivalent with having
 
θ > n−1
1
a + np − 1 . (8.2.18)

Now granted (8.2.18), from (8.2.16) and (8.2.17) we obtain (8.2.13) as desired.
Next, we turn to the proof of (8.2.15). With p, a, θ and x as before, let y ∈ Σ,
y  x, be arbitrary. Then, for some constant finite C = C(θ, p) > 0 we have
∫
 
|G1 (x − t) − G1 (y − t)| (1−θ)p δΣ (t)−ap dt
Rn
∫
 
≤C |G1 (x − t)| (1−θ)p δΣ (t)−ap dt
|x−t |<2 |x−y |
∫
 
+C |G1 (y − t)| (1−θ)p δΣ (t)−ap dt
|x−t |<2 |x−y |
∫
 
+ |G1 (x − t) − G1 (y − t)| (1−θ)p δΣ (t)−ap dt
|x−t | ≥2 |x−y |

=: I + I I + I I I. (8.2.19)

To further estimate I, we apply (8.1.6) and [133, (8.7.3)] with α := ap < 1 and
N := (n − 1)(1 − θ)p and obtain that if (8.2.18) holds then
8.2 Technical Lemmas 463
∫
   
I≤C |x − t| (1−n)(1−θ)p δΣ (t)−ap dt ≤ C|x − y| n−ap −(n−1)(1−θ)p .
|x−t |<2 |x−y |
(8.2.20)

Noting that B(x, 2|x − y|) ⊆ B(y, 3|x − y|), the reasoning we applied to estimate I
also gives that if (8.2.18) holds, then
 
I I ≤ C|x − y| n−ap −(n−1)(1−θ)p . (8.2.21)

Returning to I I I, we first observe that if t is such that |x − t| ≥ 2|x − y|, then for
any ξ ∈ [0, 1], we have |ξ x + (1 − ξ)y − t| ≥ 2|x − t|. Hence, by the Mean Value
Theorem and (8.1.6), for each t with |x − t| ≥ 2|x − y| we have

|G1 (x − t) − G1 (y − t)| ≤ |x − y| sup |(∇G1 )(ξ x + (1 − ξ)y − t)|

ξ ∈[0,1]

≤ C|x − y| sup |ξ x + (1 − ξ)y − t| −n

ξ ∈[0,1]

≤ C|x − y||x − t| −n . (8.2.22)

As such, making use of (8.2.22) and then of [133, (8.7.5)] with α := ap < 1 and
N := n(1 − θ)p, we obtain
∫ 
 δΣ (t)−ap
I I I ≤ C|x − y| (1−θ)p  dt
|x−t | ≥2 |x−y | |x − t| n(1−θ)p
(1−θ)p  
≤ C|x − y| |x − y| n−ap −n(1−θ)p
 
= C|x − y| n−ap −(n−1)(1−θ)p , (8.2.23)

provided ap > n − n(1 − θ)p. The latter condition is equivalent to

a 1
θ<+ . (8.2.24)
n p
   
Since by assumption a < 1− p1 < n 1− p1 , we have n−1 1
a + np − 1 < an + p1 , thus the
conditions (8.2.18) and (8.2.24) are compatible. Moreover, for θ satisfying (8.2.18)
and (8.2.24), the estimates (8.2.20), (8.2.21), (8.2.23) hold, which when used back
in (8.2.19), ultimately yield (8.2.15). 

Lemma 8.2.4 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall that G1 is the Bessel kernel of order
 1 one in  R , defined as
n

in (8.1.5) with α := 1. Fix p ∈ (1, ∞) along with a ∈ − p , 1 − p , and for some

1
ap
g ∈ L p (Rn, δΣ L n ) consider u := G1 ∗ g in Rn (which, by Lemma 8.2.1 is well
defined at H n−1 -a.e. point in Rn ).
Then there exists a finite constant C = C(Σ, p, a) > 0 such that
464 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
∫
p
|u(x)| p dσ(x) ≤ Cg L p (Rn,δ a p L n ) (8.2.25)
Σ Σ

and, for all r > 0,

 ∫∫  1/p
1 1
|u(x) − u(y)| dσ(x) dσ(y)
p
≤ C r 1−a− p g L p (Rn,δ a p L n ) .
r n−1 Σ
(x,y)∈Σ×Σ
|x−y |<r
(8.2.26)
 
1 −1
Proof Let p := 1 − p . Then, with θ ∈ (0, 1) to be chosen later, we write
∫ ∫ ∫ p
|u(x)| p dσ(x) = G1 (x − y)g(y) dy dσ(x)
Σ Σ Rn
∫ ∫    p
≤ |G1 (x − y)| θ δΣ (y)a |g(y)| |G1 (x − y)| 1−θ δΣ (y)−a dy dσ(x)
Σ Rn
∫ ∫ 
≤ |G1 (x − y)| θ p δΣ (y)ap |g(y)| p dy ×
Σ Rn
∫ p
p
(1−θ)p −ap
× |G1 (x − y)| δΣ (y) dy dσ(x)
Rn
∫ ∫
≤C |G1 (x − y)| θ p dσ(x) δΣ (y)ap |g(y)| p dy
Rn Σ
∫
p
≤C δΣ (y)ap |g(y)| p dy = Cg L p (Rn,δ a p L n ) . (8.2.27)
Rn Σ

For the second inequality in (8.2.27) we have applied Hölder’s inequality,  for the
third inequality we have used (8.2.13), under the restriction θ > n−1 a + p − 1 ,
1 n
+
and Fubini-Tonelli’s Theorem. The fourth inequality in (8.2.27) is a consequence
of (8.2.9), under the additional assumption that θ < p1 . Note that we can select θ
 
satisfying the above restrictions since a < 1 − p1 is equivalent to n−1
1
a + np − 1 < p1 .
Next, we turn to the proof of (8.2.26). Again we start with θ ∈ (0, 1) to be specified
later and fix r > 0 arbitrary. Then, based on the definition of u and Hölder’s inequality,
we have
8.2 Technical Lemmas 465
∫∫
|u(x) − u(y)| p dσ(x) dσ(y) (8.2.28)
(x,y)∈Σ×Σ
|x−y |<r
∫∫ ∫ p
≤ |G1 (x − t) − G1 (y − t)||g(t)| dt dσ(x) dσ(y)
Rn
(x,y)∈Σ×Σ
|x−y |<r
∫∫ ∫
≤ |G1 (x − t) − G1 (y − t)| θ p |g(t)| p δΣ (t)ap dt ×
Rn
(x,y)∈Σ×Σ
|x−y |<r

∫ p
p
(1−θ)p −ap
× |G1 (x − t) − G1 (y − t)| δΣ (t) dt dσ(x) dσ(y)
Rn

≤ C r (n−1)(θ p−1)+p−ap−1 ×
∫  ∫∫ 
θp
× |g(t)| δΣ (t)
p ap
|G1 (x − t) − G1 (y − t)| dσ(x) dσ(y) dt,
Rn (x,y)∈Σ×Σ
|x−y |<r

provided θ satisfies (8.2.14), where for the last inequality in (8.2.28) we have applied
(8.2.15) and Fubini-Tonelli’s Theorem. Next, [106, Lemma A, p. 105] (used with
d := n − 1, α := 1, and s := θp), whose application requires that
n−1
np < θ < p1 , (8.2.29)

implies the existence of a finite constant C = C(p, θ, n) > 0 such that for each t ∈ Rn
we have
∫∫
|G1 (x − t) − G1 (y − t)| θ p dσ(x) dσ(y) ≤ C r 2(n−1)−(n−1)θ p . (8.2.30)
(x,y)∈Σ×Σ
|x−y |<r

When used back in (8.2.28), this finally yields (8.2.26). There remains to point
out that it is possible to select θ ∈ (0, 1) which simultaneously satisfies conditions
(8.2.14), (8.2.29) since
     
0 < max n−1 1
a + np − 1 +, n−1
np < min ,
1 a
p n + p < 1,
1
(8.2.31)

given the current assumptions on p and a. 

Lemma 8.2.5 Assume that Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. Recall that G1 is the Bessel kernel of order one in Rn , defined as in
(8.1.5) with α := 1. Also, assume 1 < p < ∞, 0 < s < 1, and set a := 1 − s − p1 .
466 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Then the operator

 ap  p, p  ap 
T : L p Rn, δΣ L n −→ Bs (Σ, σ) acting on each g ∈ L p Rn, δΣ L n
(8.2.32)
according to T g(x) := (G1 ∗ g)(x) for σ-a.e. point x ∈ Σ,

is well defined, linear, and bounded.

Proof From Lemma 8.2.1 and Lemma 8.2.4 we know that G1 ∗ g is well defined
σ-a.e. on Σ, and
 ap 
T : L p Rn, δΣ L n −→ L p (Σ, σ) boundedly. (8.2.33)
 
Abbreviate A p := L p Σ × Σ, |x − y| 1−n σ(x) ⊗ σ(y) and denote by q (A p ),
s

1 ≤ q ≤ ∞, the space
 ! ! 
∞ 1 
:= (a j ) j ≥0 ⊆ A p : !(a j ) j ≥0 ! s (A p ) =
q
q (A p ) (2 js a j  A p )q <∞
s
q
j=0
(8.2.34)
if q < ∞ and, corresponding to q = ∞,
 ! ! 
s
(A ) := (α ) ≥0 ⊆ A : !(α j ) j ≥0 ! s = sup 2 js
α  A < ∞ . (8.2.35)
∞ p j j p  (A p )
∞
j p
j ≥0

Let u := T g and define

α j (x, y) := |u(x) − u(y)|12− j−1 < |x−y |<2− j for all x, y ∈ Σ and j ≥ 0. (8.2.36)

By Lemma 8.2.4, (8.2.26), we have for all p ∈ (1, ∞) that (α j ) j ≥0 ∈ ∞ (A p )

s and
! !
!(α j ) j ≥0 ! s ≤ Cg L p (Rn,δ a p L n ) . (8.2.37)
 (A p )∞ Σ

Let now p ∈ (1, ∞) be fixed and pick s0, s1 ∈ (0, 1) and θ ∈ (0, 1). Denote
s := (1 − θ)s0 + θs1 , a j := 1 − s j − p1 , for j ∈ {0, 1}, and set a := 1 − s − p1 .
By [15, Theorem 5.4.1, p. 115], interpolating by the real method we obtain

a p a p ap
L p (Rn, δΣ 0 L n ), L p (Rn, δΣ 1 L n ) = L p (Rn, δΣ L n ). (8.2.38)
θ, p

Moreover, by [15, Theorem 5.6.1, p. 122],


s0 s1
∞ (A p ), ∞ (A p ) = p (A p ).
s
(8.2.39)
θ, p

Therefore,
ap
T : L p (Rn, δΣ L n ) −→ p (A p )
s
boundedly. (8.2.40)
The last step of this proof is to identify the norm of (a j ) j ≥0 in ps (A p ) in the special
case when α j (x, y) := |u(x) − u(y)|12− j−1 ≤ |x−y |<2− j for each j ≥ 0. Concretely,
8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 467

! ! 
∞ 1
!(α j ) j ≥0 ! =
p
2 jsp α j  A p
p
 ps (A p )
j=0

 ∫ ∫  p1

∞
|u(x) − u(y)| p
= 2 jsp
dσ(x) dσ(y)
j=0
|x − y| n−1
(x,y)∈Σ×Σ
2− j−1 ≤ |x−y |<2− j

 ∫ ∫  p1

∞
|u(x) − u(y)| p
≈ dσ(x) dσ(y)
j=0
|x − y| n−1+sp
(x,y)∈Σ×Σ
2− j−1 ≤ |x−y |<2− j

 ∫∫  p1
|u(x) − u(y)| p
= dσ(x) dσ(y) , (8.2.41)
|x − y| n−1+sp
(x,y)∈Σ×Σ
|x−y |<1

p, p
which is the second part
 of the norm
 defining Bs (Σ, σ). Therefore, we may now
ap p, p
conclude that T : L p Rn, δΣ L n → Bs (Σ, σ) boundedly. 

We are now in a position to present the

Proof of Theorem 8.1.1 This is an immediate consequence of Lemma 8.2.5 and
(8.1.10) (cf. also (8.2.4)). 

8.3 Traces from Weighted Sobolev Spaces Defined in Open

Subsets of R n

Consider an open set Ω ⊆ Rn along with a weight w in Ω, i.e., a Lebesgue measurable

function w : Ω → [0, ∞] satisfying 0 < w(x) < ∞ for L n -a.e. x ∈ Ω. Given
p ∈ (0, ∞), the weighted L p Lebesgue space over Ω is defined as

L p (Ω, wL n ) := u : Ω → C : u Lebesgue-measurable, and u L p (Ω,w L n ) < ∞ ,

(8.3.1)
where, for each Lebesgue-measurable function u : Ω → C, we have set
∫ 1/p
u L p (Ω,w L n ) := |u| p w dL n . (8.3.2)
Ω

Lemma 8.3.1 If Ω is an arbitrary open set in Rn , 1 < p < ∞, and w ∈ Ap (Rn, L n ),

then
468 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

L p (Ω, wL n ) is a module over Lip (Ω),

Cc∞ (Ω) → L p (Ω, wL n ) densely, and (8.3.3)
L p (Ω, wL n ) → 1 (Ω, L n )
Lbdd → 1 (Ω, L n ).
Lloc

Proof The first claim in (8.3.3) is a direct consequence of definitions. As for the
second claim, for the reader’s convenience we record the standard∫argument. Let
η ∈ C ∞ (Rn ) be non-negative, radial, non-increasing and such that Rn η dL n = 1,
and define the usual mollifier ηε := ε −n η(·/ε) for each ε > 0. The crux of the matter
is then the observation that
for each function u ∈ L p (Rn, wL n ) we have
(8.3.4)
ηε ∗ u −→ u in L p (Rn, wL n ) as ε → 0+ .

The proof of (8.3.4) is straightforward, the main ingredients being the pointwise
domination of the convolution product by the Hardy-Littlewood maximal function–
cf. [175, pp. 62-63] – plus the well-known (cf., e.g., [62]) boundedness of the latter
operator on L p (Rn, wL n ). With this in hand, the second claim in (8.3.3) readily
follows.
Finally, the last claim in (8.3.3) is seen from [133, (7.7.105)] (used with n replaced
by n + 1 and Σ := Rn ≡ Rn × {0} ⊂ Rn+1 ). 

Next, given an open set Ω ⊆ Rn and an arbitrary weight w in Ω, for each p ∈ (0, ∞)
and k ∈ N define the weighted L p -based Sobolev space of order k over Ω as
k,1
W k, p (Ω, wL n ) := u ∈ Wloc (Ω) : ∂ β u ∈ L p (Ω, wL n ) for all β ∈ N0n, | β| ≤ k ,
(8.3.5)
and equip this with the quasi-norm given for each u ∈ W k, p (Ω, wL n ) by
 ∫ 1/p
uW k, p (Ω,w L n ) := |∂ β u| p w dL n . (8.3.6)
|β | ≤k Ω

From [133, (7.7.105)] (presently used with the dimension n replaced by n + 1 and
for the choice Σ := Rn ≡ Rn × {0} ⊂ Rn+1 ) we see that, whenever p ∈ (1, ∞), for
each Muckenhoupt weight w ∈ Ap (Rn, L n ), each number k ∈ N0 , and each open set
Ω ⊆ Rn we have 
k,q
W k, p (Ω, wL n ) → Wbdd (Ω), (8.3.7)
1<q<p

where we have used the piece of notation introduced in (A.0.121). Moreover, for
each p ∈ (1, ∞), w ∈ Ap (Rn, L n ), k ∈ N0 , and each open set Ω ⊆ Rn ,

W k, p (Ω, wL n ),  · W k, p (Ω,w L n ) is a reflexive Banach
(8.3.8)
space, which is a module over the ring φ Ω : φ ∈ Cck (Rn ) .

Indeed, the map

8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 469

Φ : W k, p (Ω, wL n ) → L p (Ω, wL n ) ⊕ · · · ⊕ L p (Ω, wL n )

(8.3.9)
Φ u := (∂ β u) |β | ≤k for each u ∈ W k, p (Ω, wL n )

is a well-defined linear quasi-isometry, hence injective with closed range. This

establishes an isomorphism between the weighted Sobolev space W k, p (Ω, wL n ) and
a closed subspace of a reflexive Banach space. Hence, ultimately, W k, p (Ω, wL n ) is
a reflexive Banach space itself. The last claim in (8.3.8) is clear from definitions.
Let us also observe that for each p ∈ (1, ∞), w ∈ Ap (Rn, L n ), and k ∈ N0 we
have the continuous embeddings

W k,∞ (Ω) → W k, p (Ω, wL n ) → W k,1 (Ω),

(8.3.10)
if Ω ⊆ Rn is a bounded open set.

Finally, recall from [68, Theorem 1, p. 3831] that

if p ∈ (1, ∞), w ∈ Ap (Rn, L n ), and k ∈ N0 , then for each open set

Ω ⊆ Rn it follows that C ∞ (Ω)∩W k, p (Ω, wL n ) is a dense subspace (8.3.11)
of the weighted Sobolev space W k, p (Ω, wL n ).
According to a celebrated result of P. Jones (see [103, Theorem 1, p. 73]), if
Ω ⊆ Rn is an (ε, δ)-domain then there exists a bounded linear extension operator
from the Sobolev space W k, p (Ω) (defined intrinsically in the open set Ω) into the
Sobolev space W k, p (Rn ). In fact, [103, Theorem 3, p. 74] asserts that a finitely
connected planar open set is a Sobolev extension domain if and only if the set in
question is an (ε, δ)-domain for some ε, δ > 0.
To give a formal statement of Jones’ result, we first recall Whitney’s decompo-
sition lemma (cf. [175, Theorem 1, p. 167]) according to which one can associate
to any given open, nonempty, proper subset O of Rn a family W(O) = {Q j } j ∈N of
countably many closed dyadic cubes from Rn such that

O= Q j, (8.3.12)
j ∈N
√ √
n (Q j ) ≤ dist(Q j , ∂O) ≤ 4 n (Q j ) for all j ∈ N, (8.3.13)

Q̊ j ∩ Q̊ k =  for all j, k ∈ N with j  k, (8.3.14)

1
4 (Q j ) ≤ (Q k ) ≤ 4 (Q j ) for all j, k ∈ N with Q j ∩ Q k  , (8.3.15)

and there exists M ∈ N such that 1 5 Q j ≤ M. (8.3.16)
4
j ∈N

Above, (Q) denotes the side-length of a generic cube Q in Rn . Also, given a positive
number λ and a cube Q, we denote by λQ the cube with the same center as Q, and
side-length λ (Q). With this convention, it is then straightforward to check that
(8.3.13) implies
470 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

if λ ∈ (0, 3) then λQ j ⊆ O for all j ∈ N. (8.3.17)

In fact, for each λ ∈ (0, 3) there exists some cλ ∈ (0, 1) such that
 
dist λQ j , ∂O
cλ ≤ ≤ cλ−1 for all j ∈ N. (8.3.18)
(Q j )

Here is the result advertised earlier. It is a detailed statement of [103, Theorem 1,

p. 73] which also takes into account the comments made in the top paragraph on
p. 76 of [103] (where the role of (A.0.104) is emphasized).

Theorem 8.3.2 Let Ω be an (ε, δ)-domain in Rn with rad (Ω) > 0, and fix k ∈ N.
 pick aWhitney decomposition W(Ω) of Ω along with a Whitney
Also,  decomposition

W Rn \ Ω of (Rn \ Ω)◦ = Rn \ Ω (with the convention that W Rn \ Ω =  in the
case when Ω = Rn ), and consider the collection of all small cubes in the latter, i.e.,
define
   
Ws Rn \ Ω := Q ∈ W Rn \ Ω : (Q) ≤ εδ/(16n) . (8.3.19)
k−1,1
For any function u ∈ Wloc (Ω) and any dyadic cube Q ∈ W(Ω) let PQ (u) denote
the unique polynomial of degree k − 1 which best fits u on Q, in the sense that
∫
 
∂ α u − PQ (u) dL n = 0 for each α ∈ N0n with |α| ≤ k − 1; (8.3.20)
Q

specifically, set
 ⨏
xα
PQ (u)(x) := ∂ α u dL n, ∀x ∈ Rn . (8.3.21)
α! Q
|α | ≤k−1

 
Next, to each Q ∈ Ws Rn \ Ω assign a reflected cube, i.e., some Q∗ ∈ W(Ω)
satisfying

(Q) ≤ (Q∗ ) ≤ 4 (Q) and dist(Q, Q∗ ) ≤ Cn,ε (Q). (8.3.22)

k−1,1
Finally, to each u ∈ Wloc (Ω) associate the function Λk u defined L n -a.e. in Rn by

⎧
⎪ u in Ω,
⎪
⎨
⎪ 
Λk u := PQ∗ (u) ϕQ in Rn \ Ω, (8.3.23)
⎪
⎪
⎪
⎩ Q ∈Ws (Rn \Ω)

where the family ϕQ Q ∈Ws (R n \Ω)

consists of functions satisfying

ϕQ ∈ Cc∞ (Rn ), supp ϕQ ⊆ 17

16 Q, 0 ≤ ϕQ ≤ 1,
(8.3.24)
∂α ϕ Q ≤ Cα (Q)− |α |, ∀α ∈ N0n,
8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 471
 
for every Q ∈ Ws Rn \ Ω , as well as
 
ϕQ ≡ 1 in Q. (8.3.25)
Q ∈Ws (R n \Ω) Q ∈Ws (R n \Ω)

Then for every p ∈ [1, ∞] the operator Λk satisfies

Λk : W k, p (Ω) −→ W k, p (Rn ) linearly and boundedly, and

(8.3.26)
Λk u Ω = u at L n -a.e. point in Ω for every u ∈ W k, p (Ω).

In addition, the operator norm of Λk in the first line of (8.3.26) may be controlled
solely in terms of ε, δ, n, p, k, rad (Ω) in the following fashion:

given p ∈ [1, ∞], k ∈ N, M ∈ (0, ∞) there exists C(n, k, p, M) ∈ (0, ∞)

! !
with the property that !Λk !W k, p (Ω)→W k, p (Rn ) ≤ C(n, k, p, M) whenever

the set Ω is an (ε, δ)-domain in Rn with max ε −1, δ−1, rad (Ω)−1 ≤ M.
(8.3.27)

As seen from (8.3.23), P. Jones’ extension operator is degree-dependent, though

it does not depend on the integrability exponent. A degree-independent extension
operator for the scale of Sobolev spaces on locally uniform domains has been
constructed in [161]. A variant of P. Jones’ extension operator for function spaces
involving fractional orders of differentiability has been given in [28]. See also [167]
for a universal extension operator for the scales of Besov and Triebel-Lizorkin spaces
in Lipschitz domains.
Remarkably, P. Jones’ result just cited also extends to the class of weighted
Sobolev spaces, with arbitrary weights in the Muckenhoupt class. More specifi-
cally, when Ω is an (ε, δ)-domain in Rn , k ∈ N, p ∈ (1, ∞) and w ∈ Ap (Rn, L n ),
S.K. Chua has proved the existence of a bounded, linear extension operator from
W k, p (Ω, wL n ) into W k, p (Rn, wL n ). This result, which appears in [29, Theorem 1.1,
p. 1029], reads as follows.

Theorem 8.3.3 Let Ω be an (ε, δ)-domain in Rn with rad (Ω) > 0 (cf. (A.0.104)),
and let k be a fixed positive integer. Then there exists a linear operator

Λ : W k, p (Ω, wL n ) −→ W k, p (Rn, wL n ) (8.3.28)

(depending only on k and Ω) which is bounded whenever 1 ≤ p < ∞ and for any
weight w ∈ Ap (Rn, L n ). More specifically, in such a setting one has there exists a
linear operator as in (8.3.28) with

Λ f W k, p (Ω,w L n ) ≤ C f W k, p (Ω,w L n ) for all f ∈ W k, p (Ω, wL n ), (8.3.29)

where the constant C ∈ (0, ∞) depends only on ε, δ, k, w, p, n and rad (Ω), and
which has the property that
472 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

(Λ f ) Ω = f at L n -a.e. point on Ω,
(8.3.30)
for every f ∈ W k, p (Ω, wL n ).

Let us now formally introduce the specific class of weighted Sobolev spaces we
wish to consider.

Definition 8.3.4 Assume Ω is an open, nonempty, proper subset of Rn , and fix

k ∈ N0, p ∈ (0, ∞), and a ∈ R. (8.3.31)

k, p
Then Wa (Ω) is the weighted Sobolev space defined as in (8.3.5) for the weight
ap
w := δ∂Ω in Ω. In particular, the quasi-norm associated with this space is
 ∫ 1/p
uW k, p (Ω) := |(∂ α u)(x)| p δ∂Ω (x)ap dx (8.3.32)
a
|α | ≤k Ω

k, p
for each u ∈ Wa (Ω).

Obviously, when a = 0, i.e., when s = 1 − p1 ∈ (0, 1), we are talking ordinary (i.e.,
unweighted) Sobolev spaces in the open set Ω.
Whenever Ω is an open nonempty proper subset of Rn with an Ahlfors regular
ap
boundary, [133, Proposition 8.7.4] guarantees that the weight w := δ∂Ω belongs to
 
the Muckenhoupt class Ap (Rn, L n ) whenever p ∈ (1, ∞) and a ∈ − p1 , 1 − p1 . As
a corollary of this and (8.3.7), we conclude that
if Ω ⊆ Rn is a bounded
 open set
 with an Ahlfors regular boundary,
p ∈ (1, ∞) and a ∈ − p1 , 1 − p1 , then there exists q ∈ (1, ∞) with the (8.3.33)
k, p
property that the inclusion Wa (Ω) → W k,q (Ω) is continuous.
Other basic properties of the class of weighted Sobolev spaces introduced in
Definition 8.3.4 are described in the proposition below.

Proposition 8.3.5 Let Ω ⊆ Rn be an open set whose boundary is Ahlfors regular

  k, p
and fix p ∈ (1, ∞) along with a ∈ − p1 , 1 − p1 and k ∈ N. Then Wa (Ω) equipped
with the norm  · W k, p (Ω) (cf. (8.3.32)) becomes a reflexive Banach space, which is
a
a module over φ Ω : φ ∈ Cck (Rn ) . Furthermore,

under the additional assumption that Ω is an (ε, δ)-domain with

(8.3.34)
rad (Ω) > 0, one has φ Ω : φ ∈ Cc∞ (Rn ) → Wa (Ω) densely.
k, p

Finally,
k, p
if Ω is bounded then w k,∞ (Ω) → Wa (Ω) → W k,1 (Ω), (8.3.35)

in a continuous fashion.
8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 473

Proof The claims in the first part of the statement are direct consequences of (8.3.8),
while (8.3.34) follows from Theorem 8.3.3, (8.1.3), and [133, Proposition 8.7.4] (cf.
also [29, Theorem 6.1, p. 1048]), and (8.3.35) is seen from (8.3.10). 
For future use, let us also remark that since any Muckenhoupt weight is locally
integrable,
for any open set Ω ⊆ Rn with an Ahlfors regular boundary, any
p ∈ (1, ∞), any s ∈ (0, 1), and with a := 1 − s − p1 , it follows that (8.3.36)
1, p
φ Ω : φ ∈ Lipc (Rn ) is a linear subspace of Wa (Ω).

Our next major result in this section is Theorem 8.3.6, stated a little below, which
expresses the well-definiteness of the boundary trace operator acting from a weighted
Sobolev space, considered in an (ε, δ)-domain with an Ahlfors regular boundary, onto
a suitable boundary Besov space. In this vein we wish to remark that, in the standard
case of the upper half-space, a similar result has been proved in [194]. Given that the
spaces involved are bi-Lipschitz invariant, it is straightforward to extend the trace
result from [194] to the case of Lipschitz domains. A more elaborate trace result
in the class of Lipschitz domains, involving higher-order weighted Sobolev spaces
and higher-smoothness boundary Besov spaces, has been proved in [126]. Here we
continue this line of work and consider much larger classes of domains, whose
boundaries may lack a manifold structure. As a result, methods based on flattening
the boundary are no longer applicable and we are forced to deal with the roughness
directly.
We are ready to state our main result concerning traces of functions belonging to
power weighted L p -based Sobolev spaces, with p ∈ (1, ∞), " in (ε, δ)-domains.
 defined
A suitable variant, valid in the complement range p ∈ n−1 n , 1 and involving what
we call weighted maximal Sobolev spaces, is presented later in Theorem 8.6.8.

Theorem 8.3.6 Let Ω ⊆ Rn be an (ε, δ)-domain with rad (Ω) > 0 whose boundary
is an Ahlfors regular set, and abbreviate σ = H n−1 ∂Ω. In addition, select an
integrability exponent p ∈ (1, ∞) together with a smoothness index s ∈ (0, 1), and
  1, p
introduce a := 1 − s − p1 ∈ − p1 , 1 − p1 . Then for each function u ∈ Wa (Ω) the
trace ⨏
(TrΩ→∂Ω u)(x) := [u]Ω (x) := lim+ u(y) dy (8.3.37)
r→0 B(x,r)∩Ω

exists at σ-a.e. point x ∈ ∂Ω and the induced mapping

1, p p, p
TrΩ→∂Ω : Wa (Ω) −→ Bs (∂Ω, σ)
(8.3.38)
is well defined, linear, and bounded.

In particular,
1, p
TrΩ→∂Ω u = u ∂Ω for each u ∈ Wa (Ω) ∩ C 0 (Ω). (8.3.39)

Moreover,
474 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
⨏
(TrΩ→∂Ω u)(x) = lim+ w(y) dy at σ-a.e. point x ∈ ∂Ω,
r→0 B(x,r) (8.3.40)
1, p  ap 
for each u ∈ Wa (Ω) and w ∈ W 1, p Rn, δ∂Ω L n with w Ω = u.

We wish to include a couple of remarks pertaining to the nature and scope of

Theorem 8.3.6.
Remark 1. The choice a := 0, corresponding to having s := 1 − p1 ∈ (0, 1), is
permissible and leads to a version of Theorem 8.3.6 involving plain (i.e., unweighted)
Sobolev space in the set Ω. This is known (see [106]), so the novel aspect here is
the consideration of the weighted case. Once again, the very format of the result
in (8.3.38) is prefigured by homogeneity considerations. Indeed, recall the heuristic
principle according to which when measuring smoothness on the L p scale taking
the trace of a function defined in Ω on ∂Ω, regarded as a “surface” of co-dimension
one, decreases the smoothness by 1/p units; see item (ii) in Theorem 9.4.5 for a
rigorous manifestation of this principle. As such, for the trace of u on ∂Ω be in
p, p
Bs (∂Ω, σ) requires that u has s + 1/p units of smoothness in Ω. Since we insist
on working  with weighted Sobolev spaces of order one in Ω, the disagreement
a := 1 − s + p1 = 1 − s − p1 needs to be accounted for by the actual weight. By
γ
regarding the operator of pointwise multiplication by δ∂Ω as having the effect of
fractional differentiation/integration of order γ ∈ R in Ω, this explains the nature of
the power of the weight in (8.3.38).
Remark 2. Retain the setting of Theorem 8.3.6 and make the additional assumption
that ∂Ω is compact. In this case, we may naturally extend the action of the boundary
trace operator (8.3.38) to the larger space
 
(Ω, L n ) : ψu ∈ Wa (Ω) for all ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near ∂Ω
1, p
u ∈ Lloc
1

(8.3.41)
p, p
by defining TrΩ→∂Ω u in the “new” sense, to be TrΩ→∂Ω (ψu) ∈ Bs (∂Ω, σ) in the
“old” sense, whenever u and ψ as in (8.3.41). It turns out that this definition is
meaningful and unambiguous (in this regard, (8.3.39) is helpful).
These considerations suggest adopting the following convention:

Convention 8.3.7 Let Ω ⊆ Rn be an arbitrary open set, and suppose X (Ω) is a

space of distributions in Ω (i.e., X (Ω) is a linear subspace of D (Ω)). Then Xbdd (Ω),
or X (Ω)bdd (used interchangeably), shall be used to denote
   
u ∈ D (Ω) : ψ Ω u ∈ X (Ω) for each ψ ∈ Cc∞ (Rn ) . (8.3.42)

Moreover, said space will be equipped with the initial topology1 with respect to the
family of functions Xbdd (Ω)  u → ψu ∈ X (Ω) with ψ ∈ Cc∞ (Rn ), should the space
X (Ω) come equipped with a topology (i.e., Xbdd (Ω) is equipped with the coarsest
topology which makes each map Xbdd (Ω)  u → ψu ∈ X (Ω) continuous).

1 or weak topology, or limit topology, or projective topology

8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 475

For example, with this piece of notation, the space in (8.3.41) simply becomes
1, p
Wa (Ω)bdd (assuming ∂Ω is compact), and Remark 2 above may be concisely restated
as the statement that, in the setting of Theorem 8.3.6,
1, p p, p
TrΩ→∂Ω : Wa (Ω)bdd −→ Bs (∂Ω, σ)
(8.3.43)
is well defined, linear, and bounded, if ∂Ω is compact.

After this digression, we now turn to the proof of Theorem 8.3.6.

1, p
Proof of Theorem 8.3.6 Consider an arbitrary function u ∈ Wa (Ω). Then Theo-
rem 8.3.3 together with
 [133, Proposition 8.7.4] guarantee the existence of some
ap
w ∈ W 1, p Rn, δ∂Ω L n such that

w Ω = u and w W 1, p (Rn,δ a p L n ) ≤ CuW 1, p (Ω) (8.3.44)

∂Ω a

where the constant C ∈ (0, ∞) is independent of u. Also, Theorem 8.1.1 ensures that
p, p
the restriction of [w]Rn to ∂Ω is well defined, belongs to Bs (∂Ω, σ), and its norm
in this Besov space is dominated by a fixed multiple of w W 1, p (Rn,δ a p L n ) . As such,
∂Ω
as far as the claims pertaining to (8.3.37)-(8.3.38) are concerned it suffices to show
that
[w]Rn = [u]Ω at σ-a.e. point on ∂Ω. (8.3.45)
To justify this, pick an arbitrary number R > 0 along with a function ψ ∈ Cc∞ (Rn )
with the property that ψ ≡ 1 on BR := B(0, R). Next, from [133, (7.7.105)] (used
with n replaced by n + 1 and Σ := Rn ≡ Rn × {0} ⊂ Rn+1 ) we see that there exists
q ∈ (1, ∞) such that ψw ∈ W 1,q (Rn ). In concert with (8.3.44), which ensures that
(ψw) Ω = ψu in Ω, this also implies ψu ∈ W 1,q (Ω). Granted these properties, [106,
Proposition 2, p. 206] applies and gives that [ψw]Rn = [ψu]Ω at σ-a.e. point on ∂Ω.
In particular,
[w]Rn = [u]Ω at σ-a.e. point on BR ∩ ∂Ω. (8.3.46)
Since R > 0 was arbitrary, (8.3.45) follows. Finally, (8.3.39) is a consequence of
(8.3.37), while (8.3.40) is seen from (8.3.45). This finishes the proof of the theorem.

Moving on, recall from [133, Lemma 5.11.9)] that any (ε, δ)-domain Ω ⊆ Rn
with rad (Ω) > 0 satisfies an interior corkscrew condition.

Proposition 8.3.8 Let Ω ⊆ Rn be an (ε, δ)-domain with rad (Ω) > 0 and whose
boundary
 is Ahlfors
 regular. Abbreviate σ = H n−1 ∂Ω. For each z ∈ ∂Ω and each
r ∈ 0, diam Ω denote by Ar (z) an interior corkscrew point for z at scale r, and fix
a sufficiently
  constant C > 0 such that B(Ar (z), r/C) ⊆ Ω for all z ∈ ∂Ω and
large
r ∈ 0, diam Ω . Finally,
 fix an arbitrary integrability exponent p ∈ (1, ∞) along
with a power a ∈ − p1 , 1 − p1 .
1, p
Then for each function u ∈ Wa (Ω) it follows that
476 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
⨏
the limit lim+ u(y) dy exists, and is
r→0 B(Ar (z),r/C) (8.3.47)
equal to (TrΩ→∂Ω u)(z), for σ-a.e. z ∈ ∂Ω.
1, p
Proof Fix an arbitrary function u ∈ Wa (Ω). Since from (8.3.7) we know that
there exists some q ∈ (1, ∞) with the property that for each ψ ∈ Cc∞ (Rn ) we have
ψu ∈ W 1,q (Ω), we may invoke [19, Theorem 5.6] to conclude that (8.3.47) holds
with u replaced by ψu. In light of the arbitrariness of the cutoff function ψ, the
desired conclusion follows.
A more direct proof of (8.3.47), of intrinsic value, is as follows. By Theorem 8.3.3,
Theorem 8.3.6, and  Theorem  8.1.1 there is no loss of generality in assuming that
ap
actually u ∈ W 1, p Rn, δ∂Ω L n , with the goal of showing that
⨏
lim+ u(y) dy = (TrRn →Ω u)(z) for σ-a.e. z ∈ ∂Ω. (8.3.48)
r→0 B(Ar (z),r/C)

Let this be the case and denote by F ⊆ ∂Ω the collection of points z ∈ ∂Ω with
the property that (TrRn →Ω u)(z) exists; hence, σ(∂Ω \ F) = 0 by Theorem 8.3.6 and
Theorem 8.1.1. Also, for each λ > 0 fixed, introduce
# ⨏ ⨏ $
Eλ := z ∈ F : lim+ u(y) dy − lim inf
+
u(y) dy > λ .
r→0 B(z,r)∩Ω r→0 B(Ar (z),r/C)
(8.3.49)
Then, with M = MRn denoting the classical Hardy-Littlewood maximal operator
in Rn , for every z ∈ Eλ and each φ ∈ Cc∞ (Rn ) we have
⨏ ⨏
λ < lim+ u(y) dy − lim inf
+
u(y) dy
r→0 B(z,r)∩Ω r→0 B(Ar (z),r/C)
⨏ ⨏
= lim+ (u − φ)(y) dy − lim inf
+
(u − φ)(y) dy
r→0 B(z,r)∩Ω r→0 B(Ar (z),r/C)

≤ C M(u − φ)(z), (8.3.50)

by the interior corkscrew condition for Ω. To continue, we need a couple of estimates

of independent interest, to the effect that

 ap  (Mw)(z) ≤ [Mw]Rn (z)

w ∈ W 1, p Rn, δ∂Ω L n =⇒ (8.3.51)
for H n−1 -a.e. z ∈ Rn,

and
Mw W 1, p (Rn,δ a p L n ) ≤ C(n, p, a)w W 1, p (Rn,δ a p L n ) . (8.3.52)
∂Ω ∂Ω

Assuming these for the time being, we may then deduce from (8.3.50)-(8.3.51) that
8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 477
 
λ < [M(u − φ)]Rn (z) = TrRn →∂Ω M(u − φ) (z)
(8.3.53)
for σ-a.e. z ∈ Eλ , and every φ ∈ Cc∞ (Rn ).

Thus, on account of (8.3.52)-(8.3.53), Chebytcheff’s inequality, and the fact that

 ap 
TrRn →∂Ω : W 1, p Rn, δ∂Ω L n −→ L p (∂Ω, σ) (8.3.54)

is a well-defined, bounded operator (cf. Theorem 8.1.1 and (7.9.10)-(7.9.11)), we

may estimate
∫
 p p
λ p σ(Eλ ) ≤ TrRn →∂Ω M(u − φ) dσ ≤ Cu − φW 1, p (Rn,δ a p L n ), (8.3.55)
∂Ω ∂Ω

where C > 0 is a finite constant, independent of φ. Upon recalling that

 
Cc∞ (Rn ) → W 1, p Rn, δ∂Ω L n densely
ap
(8.3.56)

(cf. (8.1.4)), we may therefore conclude from (8.3.55) that σ(Eλ ) = 0 for every
λ > 0. In the context of (8.3.49), this entails
⨏ ⨏
lim inf
+
u(y) dy = lim+ u(y) dy for σ-a.e. z ∈ ∂Ω. (8.3.57)
r→0 B(Ar (z),r/C) r→0 B(z,r)∩Ω

A similar reasoning yields

⨏ ⨏
lim sup u(y) dy = lim+ u(y) dy for σ-a.e. z ∈ ∂Ω,
r→0+ B(Ar (z),r/C) r→0 B(z,r)∩Ω
(8.3.58)
so that (8.3.48) follows from (8.3.57)-(8.3.58). This completes the proof of the
proposition, modulo the justification of (8.3.51)-(8.3.52).  
ap
Consider first (8.3.51). To prove this implication, fix w ∈ W 1, p Rn, δ∂Ω L n and
denote by S ⊆ Rn the collection of points z ∈ Rn where the function  Mw may 
ap
be strictly defined in Rn . Since, by (8.3.52), Mw belongs to W 1, p Rn, δ∂Ω L n , it
follows that
H n−1 (Rn \ S) = 0. (8.3.59)
Then at every z ∈ S and for each r > 0 we may write
478 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
⨏
[Mw]Rn (z) = lim+ (Mw)(y) dy
ε→0 B(z,ε)
% ∫ ∫ &
1 1
≥ lim+     |w(x)| dx dy
ε→0 L n B(z, ε) B(z,ε) L n B(y, r) B(y,r)
% ∫ &
1 n 
= lim+   n  |w(x)| L B(z, ε) ∩ B(x, r) dx
ε→0 L B(0, ε) L B(0, r)
n
B(z,r+ε)

%⨏   &
L n B(z, ε) ∩ B(x, r)
≥ lim+ |w(x)|   dx
ε→0 B(z,r) L n B(z, ε)
⨏
≥ |w(x)| dx, (8.3.60)
B(z,r)

where the last inequality in (8.3.60) follows from the Lebesgue Dominated Conver-
gence Theorem (here it helps to observe that if x ∈ B(z, r) then B(z, ε) ⊂ B(x, r) for
ε sufficiently small). On account of (8.3.59), the estimate in (8.3.51) now follows by
taking the supremum over r > 0 of the most extreme sides in (8.3.60). This finishes
the proof of (8.3.51).
Finally, as far as (8.3.52) is concerned, we employ the point-wise inequality

|∂j (Mw)| ≤ M(∂j w) at L n -a.e. point in Rn,

 ap  (8.3.61)
for each w ∈ W 1, p Rn, δ∂Ω L n and j ∈ {1, . . . , n}.

It is implicit in the
 above considerations
 that ∂j (Mw) is a regular distribution for
ap
every w ∈ W 1, p Rn, δ∂Ω L n . This is proved in [114] in the unweighted case (i.e.,
when a = 0) and the proof carries over to the present case mutandis-mutatis. With
this in hand we may then estimate

n
Mw W 1, p (Rn,δ a p L n ) = Mw  L p (Rn,δ a p L n ) + ∂j (Mw) L p (Rn,δ a p L n )
∂Ω ∂Ω ∂Ω
j=1


n
≤ Mw  L p (Rn,δ a p L n ) + M(∂j w) L p (Rn,δ a p L n )
∂Ω ∂Ω
j=1


n
≤ Cw  L p (Rn,δ a p L n ) + C ∂j w  L p (Rn,δ a p L n )
∂Ω ∂Ω
j=1

≤ Cw W 1, p (Rn,δ a p L n ), (8.3.62)

∂Ω

 ap 
by the boundedness of the Hardy-Littlewood maximal function on L p Rn, δ∂Ω L n .
This completes the justification of (8.3.52) and finishes the proof of the proposition.
8.3 Traces from Weighted Sobolev Spaces Defined in Open Subsets of R n 479

The stage is now set for us to show that, whenever meaningful, the boundary trace
taken in the nontangential pointwise limit sense (cf. [133, §8.9]) is compatible with
the boundary trace considered in the sense of Theorem 8.3.6.
Corollary 8.3.9 Let Ω ⊆ Rn be an (ε, δ)-domain with rad (Ω) > 0 and whose
 Also, abbreviate σ = H
boundary is Ahlfors n−1 ∂Ω and fix p ∈ (1, ∞)
 1 regular.
along with a ∈ − p , 1 − p . Then there exists an aperture parameter κ > 0 with the
1

1, p
property that for each function u ∈ Wa (Ω) which has a κ-nontangential pointwise
limit at σ-a.e. point on ∂Ω, i.e., such that
 κ−n.t.
u ∂Ω (x) exists (in the sense of [133, Definition 8.9.1]; cf. (A.0.82))
(8.3.63)
for σ-a.e. point x ∈ ∂Ω,

it follows that the boundary trace of the function u in the sense of Theorem 8.3.6
coincides with the κ-nontangential pointwise limit of u at σ-a.e. point on ∂Ω, i.e.,
κ−n.t.
TrΩ→∂Ω u = u ∂Ω at σ-a.e. point on ∂Ω. (8.3.64)

Proof This is a consequence of Proposition 8.3.8 and [133, Definition 8.9.1] (cf.
(A.0.82)), bearing in mind that, as seen from [133, Corollary 8.8.9] and [133,
Lemma 5.11.9)] (in combination
 with
 [133, (5.11.35)]), the present geometric as-
sumptions ensure that σ ∂Ω \ ∂nta Ω = 0. 
Our next result elaborates on the nature of the kernel (or null-space) of the bound-
ary trace operator studied in this chapter. To facilitate the subsequent discussion,
we make the following definition. If Ω ⊆ Rn is an open set with an Ahlfors regular
boundary, then for every p ∈ (1, ∞) and a ∈ (−1/p, 1 − 1/p) we set

W̊a (Ω) := the closure of Cc∞ (Ω) in Wa (Ω).

1, p 1, p
(8.3.65)
1, p 1, p
Since, by design, W̊a (Ω) is a closed subspace of Wa (Ω), Proposition 8.3.5 guar-
antees that
1, p
W̊a (Ω) is a reflexive Banach space when equipped
1, p (8.3.66)
with the norm inherited from the space Wa (Ω).

Proposition 8.3.10 Let Ω ⊆ Rn be an (ε, δ)-domain with2 rad (Ω) > 0, whose
boundary is an Ahlfors regular set, and such that Ωc := Rn \ Ω is n-thick. Abbreviate
σ := H n−1 ∂Ω. Then whenever
 
1 < p < ∞, a ∈ − p1 , 1 − p1 , s := 1 − a − p1 ∈ (0, 1) (8.3.67)

the null-space of the trace operator (8.3.37) may be described as

 
1, p p, p 1, p
Ker TrΩ→∂Ω : Wa (Ω) → Bs (∂Ω, σ) = W̊a (Ω). (8.3.68)

2 according to [133, (5.11.35)] this is the case if, e.g., Ω has compact boundary
480 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

In particular,
 
Cc∞ (Ω) → u ∈ Wa (Ω) : TrΩ→∂Ω u = 0 on ∂Ω densely.
1, p
(8.3.69)

Proof We are working in a context in which the boundary trace operator (8.3.37)
is linear and continuous. In particular, its null-space is closed. From (8.3.39) it
is also clear that Cc∞ (Ω) is contained into the null-space of (8.3.37). Combining
these observations, the inclusion of the closure in the Wa (Ω) norm of Cc∞ (Ω) into
1, p

the null-space of (8.3.37) follows. To prove the opposite inclusion, fix an arbitrary

1, p
function u ∈ Wa (Ω) with the property that TrΩ→∂Ω u = 0. From (8.3.34) we know
that there exists a sequence of functions {uk }k ∈N ⊂ φ Ω : φ ∈ Cc∞ (Rn ) such that
1, p
uk → u as k → ∞ in Wa (Ω). To proceed, observe that Ω is a domain of locally finite
perimeter (since its boundary is Ahlfors regular) and denote by ν = (ν j )1≤ j ≤n its
geometric measure theoretic outward unit normal. Then, if ' u stands for the extension
by zero of u outside Ω, for each j ∈ {1, . . . , n} and ϕ ∈ Cc∞ (Rn ) we may rely
on the De Giorgi-Federer version of the Divergence Theorem (formulated in [133,
Theorem 1.1.1]) to write
∫ ∫ ∫
'
u(∂j ϕ) dL n = u(∂j ϕ) dL n = lim uk (∂j ϕ) dL n
Rn Ω k→∞ Ω
% ∫ ∫ &
= lim − (∂j uk )ϕ dL n + uk ϕ ν j dσ
k→∞ Ω ∂∗ Ω
∫ ∫
=− (∂j u)ϕ dL n = − (
(∂ j u)ϕ dL .
n
(8.3.70)
Ω Rn

For the convergence of the solid integrals intervening in the second and third equality
in (8.3.70) we have used (8.3.7) and [133, Proposition 8.7.4]. Moreover, (8.3.39)
guarantees that uk ∂Ω = TrΩ→∂Ω uk for each k ∈ N, and since TrΩ→∂Ω uk converges
∫as k → ∞ to TrΩ→∂Ω u = 0 in Lloc (∂Ω, σ) (by (8.3.38) and (7.9.40)), we have
1

∂∗ Ω k
u ϕ ν j dσ → 0 as k → ∞. Having established (8.3.70), we conclude that
 
∂j '
u=∂ ) ap
u ∈ W 1, p Rn, δ∂Ω L n . By design we have '
j u, hence ' u Ωc = 0, so on account
of the weighted version of the spectral synthesis theory (cf. Theorem 4.4.2 on p. 157
and the comment at the end of § 4.4.1 on p. 160 in [192]; it is here that the assumption
on the n-thickness of Ωc is used) it follows
 there exists a sequence {ui }i ∈N ⊆ Cc∞ (Ω)
ap
such that u'i → ' u in W 1, p Rn, δ∂Ω L n as i → ∞. The latter implies that ui → u in
1, p
Wa (Ω) as i → ∞, proving (8.3.69). This completes the proof of the proposition. 
8.4 Extension Operators from Boundary Besov Spaces into Weighted Sobolev Spaces 481

8.4 Extension Operators from Boundary Besov Spaces into

Weighted Sobolev Spaces

The goal here is to discuss the counterpart of Theorem 8.3.6, pertaining to the
existence and properties of an extension operator, mapping functions on the boundary
into functions inside of a given domain. Here is our main result in this regard.

Theorem 8.4.1 Let Ω ⊆ Rn (where n ∈ N, n ≥ 2) be an (ε, δ)-domain with a compact

Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. Then there exists a linear
operator
Ex∂Ω→Ω : L 1 (∂Ω, σ) −→ C ∞ (Ω) (8.4.1)
satisfying the following properties.

(1) Assume
 
p ∈ (1, ∞), s ∈ (0, 1), and a := 1 − s − 1
p ∈ − p1 , 1 − p1 . (8.4.2)

Then the operator (8.4.1) induces a well-defined, linear, and bounded mapping
in the context
p, p 1, p
Ex∂Ω→Ω : Bs (∂Ω, σ) −→ Wa (Ω). (8.4.3)
Moreover, if TrΩ→∂Ω is the boundary trace operator from (8.3.38) and I denotes
the identity operator then (8.4.3) satisfies
p, p
TrΩ→∂Ω ◦ Ex∂Ω→Ω = I on Bs (∂Ω, σ). (8.4.4)

In particular, the trace operator (8.3.38) is onto.

(2) For every γ ∈ (0, 1) the operator

Ex∂Ω→Ω : C γ (∂Ω) −→ C γ (Ω ) (8.4.5)

is well defined, linear, and bounded, and there exists C = C(Ω, γ) ∈ (0, ∞) such
that
! !
! 1−γ !
!δ∂Ω ∇Ex∂Ω→Ω f ! ∞ n
≤ C f C γ (∂Ω), ∀ f ∈ C γ (∂Ω). (8.4.6)
L (Ω, L )

(3) Given any p ∈ (1, ∞) and κ > 0 there exists some constant C ∈ (0, ∞) with the
property that for each function f ∈ L p (∂Ω, σ) one has
! !
!Nκ (Ex∂Ω→Ω f )! p ≤ C f  L p (∂Ω,σ) (8.4.7)
L (∂Ω,σ)

and κ−n.t.
(Ex∂Ω→Ω f ) ∂Ω = f at σ-a.e. point on ∂Ω. (8.4.8)
(4) Assume 1 
n−1
n < p ≤ ∞ and (n − 1) p −1 + < s < 1. (8.4.9)
482 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Then for each k ∈ N0 and θ ∈ (0, 1) there exists C ∈ (0, ∞) with the property
p, p
that for each function f ∈ Bs (∂Ω, σ) one has
! k+1− 1 −s !
! p !
!δ∂Ω ∇k+1 Ex∂Ω→Ω f ,θ !
≤ C f Bsp, p (∂Ω,σ), (8.4.10)
L p (Ω, L n )

with the understanding that the subscripts , θ are omitted when p > 1.
Proof To get started, observe that there is no loss of generality in assuming that Ω
is bounded. Indeed, in view of [133, Lemma 5.10.10], the alternative would be for
Ω to be an exterior domain. In such a scenario, we may rely on the result assumed to
hold in the bounded domain B(0, R) ∩ Ω (where R > 0 is a sufficiently large radius)
and multiply the extension operator in such a setting by a smooth cutoff function
which is identically 1 near ∂Ω.
Also, since Ω is an NTA domain with finitely many connected components
at a strictly positive distance from each other (cf. [133, (5.11.66)] and [133,
Lemma 5.11.3]), working in one connected component at a time there is no loss
of generality in assuming that Ω is also connected.
For the remainder of the proof we shall therefore assume that Ω is a bounded con-
nected (ε, δ)-domain with a compact Ahlfors regular boundary. Define the operator
Ex∂Ω→Ω to act on any function f ∈ L 1 (∂Ω, σ) according to
∫
(Ex∂Ω→Ω f )(x) := K(x, y) f (y) dσ(y), ∀ ∈ Ω, (8.4.11)
∂Ω

where the integral kernel K : Ω× ∂Ω → R is assumed to be measurable, C ∞ smooth

in the first variable, and satisfy
∫
K(x, y) dσ(y) = 1 for all points x ∈ Ω, (8.4.12)
∂Ω

|∂xα K(x, y)| ≤ cα δ∂Ω (x)1−n− |α |, ∀x ∈ Ω, ∀y ∈ ∂Ω, (8.4.13)

for each multi-index α ∈ N0n , and

K(x, y) = 0 for all x ∈ Ω and y ∈ ∂Ω with |x − y| ≥ λδ∂Ω (x), (8.4.14)

for some λ > 1 to be specified later. One can take, for instance, the kernel

∫  −1
 x−y x−z
 
K(x, y) := η reg  η reg dσ(z) for x ∈ Ω, y ∈ ∂Ω,
 · δ∂Ω (x)  · δ∂Ω (x)
 ∂Ω 
  (8.4.15)
where η ∈ Cc∞ B(0, λ) is a fixed function satisfying η ≡ 1 on B(0, 3λ/4), η ≥ 0, and
reg
 is a positive constant which will be specified momentarily. Above, δ∂Ω (x) stands
for the regularized distance from x ∈ R to the set ∂Ω (cf. [133, Proposition 6.1.1]).
n

Hence, this is a C ∞ function in Rn \ ∂Ω with the property that there exist two (purely
dimensional) constants C0, C1 > 0 such that
8.4 Extension Operators from Boundary Besov Spaces into Weighted Sobolev Spaces 483

∂ α δ∂Ω (x) ≤ Cn,α δ∂Ω (x)1− |α |,

reg reg
C0 δ∂Ω (x) ≤ δ∂Ω (x) ≤ C1 δ∂Ω (x), (8.4.16)

for each x ∈ Rn \ ∂Ω and each multi-index α. Hence, choosing

2C1 1
λ := and  := , (8.4.17)
C0 C1
ensures that
 
x−y
η reg = 0 if x ∈ Ω, y ∈ ∂Ω with |x − y| ≥ λδ∂Ω (x). (8.4.18)
 · δ∂Ω (x)

We next wish to show that K is well defined, and we shall do so by producing a

positive lower bound for the integral appearing in (8.4.15) for each x ∈ Ω. With this
goal in mind, fix x ∈ Ω arbitrary and consider x ∗ ∈ ∂Ω such that |x − x ∗ | = δ∂Ω (x).
Set r := δ∂Ω (x)/(2C0 ) and consider z ∈ B(x ∗, r) ∩ ∂Ω. Then
reg

|x − z| ≤ |x − x ∗ | + |x ∗ − z| ≤ δ∂Ω (x) + r

≤ (3/2)C0−1 δ∂Ω (x) =  · δ∂Ω (x) 3λ/4,

reg reg
(8.4.19)

which forces  
x−z
η reg = 1 for all z ∈ B(x ∗, r) ∩ ∂Ω. (8.4.20)
 · δ∂Ω (x)
 
Given that σ B(x ∗, r) ∩ ∂Ω ≈ r n−1 ≈ δ∂Ω (x)n−1 ≈ δ∂Ω (x)n−1 as ∂Ω is Ahlfors
reg

regular, and since η ≥ 0, it follows that

∫  −1
 x−z 
 η reg dσ(z) ≤ Cδ∂Ω (x)1−n, (8.4.21)
 · δ∂Ω (x)
∂Ω 
for all x ∈ Ω, where C > 0 is a finite constant independent of x. Of course, once
(8.4.21) has been established then it is clear that K is well defined and (8.4.13) holds
when α = 0. Inductively, a direct calculation based on (8.4.15), (8.4.16), (8.4.18)
and (8.4.21) allows us to estimate

∂xα K(x, y) ≤ Cα δ∂Ω (x)− |α |+1−n 1(x,y)∈Ω×∂Ω: |x−y | ≤λδ∂Ω (x), (8.4.22)

for all multi-indices α ∈ N0n . Furthermore, (8.4.12) holds by design. All together,
this analysis shows that K in (8.4.15) satisfies (8.4.12)-(8.4.14).
At this stage, let us make the observation that the operator (8.4.11) falls under
the scope of [133, Proposition 8.4.12]. More specifically, from (8.4.13) (used with
|α| = 0) and (8.4.14) we conclude that |K(x, y)| satisfies [133, (8.4.150)] (for any
α > 0). Consequently, [133, (8.4.153)] implies (8.4.7).
Since we are presently assuming that Ω is bounded, (8.4.22) implies for each
ε∗ < 1 there exists a constant C ∈ (0, ∞) such that |K(x, y)| ≤ C|x − y| −(n−ε∗ ) for all
484 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

x ∈ Ω and y ∈ ∂Ω. Hence, given any p ∈ (1, ∞) and s ∈ (0, 1), choosing ε∗ ∈ (s, 1)
permits us to invoke [133, Proposition 8.7.10] (for the operator T := Ex∂Ω→Ω ) with
β := 1 − s − 1/p and r := diam Ω to conclude that there exists a constant C ∈ (0, ∞)
such that for every f ∈ L p (∂Ω, σ) we have
! 1− 1 −s !
! p !
!δ∂Ω · Ex∂Ω→Ω f ! ≤ C f  L p (∂Ω,σ) . (8.4.23)
L p (Ω, L n )

Let us also observe from (8.4.22) that for each α ∈ N0n with |α| = k + 1 there
exists C ∈ (0, ∞) such that the function q(x, y) := ∂xα K(x, y) satisfies

C
|q(x, y)| ≤ for all x ∈ Ω and y ∈ ∂Ω. (8.4.24)
δ∂Ω (x)k |x − y| n
Also, (8.4.12) guarantees that the boundary-to-domain integral operator with kernel
q(x, y) annihilates constants provided |α| > 0. As such, whenever p, s are as in
(8.4.9) we may apply [136, Theorem 4.2.1] (with ε := 0) and conclude that for any
k ∈ N0 and any θ ∈ (0, 1) there exists a finite constant C = C(Ω, p, s, k, θ) > 0 such
p, p
that (8.4.10) holds for every f ∈ Bs (∂Ω, σ).
Next, choosing k := 0 and restricting p and s to (1, ∞) and (0, 1), respectively, we
may then conclude from (8.4.10), (8.4.23), (7.9.11), and (8.3.32) (with k := 1) that
the operator (8.4.3) is indeed bounded.
The case p = ∞ of (8.4.10) implies (8.4.6) (keeping in mind (7.1.59)). In turn,
from (8.4.6), [133, (5.11.78)], and the fact that in the class of connected sets with
compact boundaries being an (ε, δ)-domain is equivalent to being a uniform domain
(cf. [133, (5.11.66)] and items (2), (4) in [133, Proposition 5.11.14]) we conclude
that (8.4.5) is a bounded operator.
Consider next (8.4.4). For starters, based on the approximation of identity nature
of Ex∂Ω→Ω (cf. (8.4.12)-(8.4.14)) we may readily conclude that
* "
Ex∂Ω→Ω f = f for every f ∈ Lip (∂Ω). (8.4.25)
∂Ω
p, p
On the other hand, Lip (∂Ω) is a dense subspace of Bs (∂Ω, σ) (as seen from
Lemma 7.1.10, bearing in mind that ∂Ω is compact). Given that both the composition
p, p
TrΩ→∂Ω ◦ Ex∂Ω→Ω and the identity are bounded linear operators on Bs (∂Ω, σ), on
account of the aforementioned properties together with (8.4.5) and Proposition 8.3.8
we may now conclude that (8.4.4) is true.
There remains to show that (8.4.8) holds for each function f ∈ L p (∂Ω, σ) with
p ∈ (1, ∞). The idea is to invoke [133, Proposition 6.2.11] with the following choices:
X := Ω ∪ ∂nta Ω equipped with the topology inherited from the ambient Euclidean
space, X := ∂nta Ω, μ := H n−1 ∂nta Ω which is a locally finite complete Borel-regular
measure on X by virtue of [133, Lemma 3.6.4] used with parameter s := n − 1
(that the hypotheses in [133, (3.6.25)] are presently satisfied is a consequence of
item (i) in [133, Proposition 8.8.6] and the upper Ahlfors regularity of ∂Ω), the
sets Γ(x) := Γκ (x) for each x ∈ ∂nta Ω which satisfy [133, (6.2.71)] thanks to
(A.0.94), the Banach space Y := L p (∂Ω, σ), the linear operator Ex∂Ω→Ω originally
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 485

defined as in (8.4.1), and V := Lip(∂Ω) which is a dense subset of Y (by [133,

(3.7.22)]) with the property that for each f ∈ V the limit in [133, (6.2.75)] exists
(cf. (8.4.25)). In this scenario, [133, (6.2.74)] becomes a consequence of (8.4.7), so
[133, Proposition 6.2.11] applies and gives that for each function f ∈ L p (∂Ω, σ)
κ−n.t.
with p ∈ (1, ∞) we have (Ex∂Ω→Ω f ) ∂Ω = f at σ-a.e. point on ∂nta Ω. Granted this,
(8.4.8) follows on account of [133, Corollary 8.8.9]. 

8.5 Weighted Sobolev Spaces of Negative Order, Duality, and the

Conormal Derivative Operator

Suppose Ω ⊆ Rn is an open set with an Ahlfors regular boundary and fix an exponent

p ∈ (1, ∞) along with a ∈ (−1/p, 1 − 1/p). As in the past, D (Ω) denotes the space
of distributions in Ω. In this setting, we define the weighted Sobolev space of order
minus one as
 
n
  
−1, p
∂j f j ∈ D (Ω) : f j ∈ L p Ω, δ∂Ω L n , 0 ≤ j ≤ n .
ap
Wa (Ω) := f = f0 +
j=1
(8.5.1)
According to [133, Proposition 8.7.4] and the last claim in (8.3.3) it is meaningful
to take distributional
  derivatives of functions belonging to the weighted Lebesgue
ap
space L p Ω, δ∂Ω L n (which are themselves, generally speaking, distributions in Ω).
−1, p
It may be checked that, in the above setting, Wa (Ω) becomes a Banach space when
equipped with the norm  · W −1, p (Ω) given by
a


n+1 
n
−1, p
Wa (Ω)  f → inf  f j  L p (Ω,δ a p L n ) : f = f0 + ∂j f j ,
∂Ω
j=0 j=1
  
ap
f j ∈ L p Ω, δ∂Ω L n , 0 ≤ j ≤ n . (8.5.2)

Remark 8.5.1 Let Ω ⊆ Rn be an open set with an Ahlfors regular boundary and
fix two exponents, 1 < p < ∞ and −1/p < a < 1 − 1/p. Then there exists a finite
−1, p
constant C > 0 with the property that given any f ∈ Wa (Ω) it is possible to find
' −1, p n
f ∈ Wa (R ) such that

'
f = f and  '
f W −1, p (Rn ) ≤ C f W −1, p (Ω) . (8.5.3)
Ω a a

+
Indeed, start with a representation f = f0 + nj=1 ∂j f j which is quasi-optimal (in the
+
norm sense), then take ' f := ' f j where '
f0 + nj=1 ∂j ' f j is the extension to Rn of f j by
setting it equal to zero outside Ω, for 0 ≤ j ≤ n.
486 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

The weighted Sobolev space of negative smoothness arise naturally as duals, in

the manner described in the proposition below. The reader is reminded that the space
1, p
W̊a (Ω) has been defined in (8.3.65).

Proposition 8.5.2 Let Ω ⊆ Rn be an (ε, δ)-domain with rad (Ω) > 0, whose bound-
ary is an Ahlfors regular set, and such that Ωc := Rn \ Ω is n-thick. Also, as-
sume p, p ∈ (1, ∞) are two exponents such that 1/p + 1/p = 1 and fix some
a ∈ (−1/p, 1 − 1/p). Then
 1, p ∗ −1, p
W̊a (Ω) = W−a (Ω) (8.5.4)

and  −1, p ∗ −1, p

Wa (Ω) = W̊−a (Ω). (8.5.5)
 
Note that since 1 < p < ∞ and −a ∈ − 1/p, 1 − 1/p , it makes sense to
−1, p 
consider the weighted Sobolev space of negative order W−a (Ω) (cf. the discussion
pertaining to (8.5.1)).
Proof of Proposition 8.5.2 To justify (8.5.4), we begin by defining the mapping
−1, p  1, p ∗
Φ : W−a (Ω) −→ W̊a (Ω) (8.5.6)
−1, p
by setting Φ( f ) := Λ f with the agreement that, if f ∈ W−a (Ω) admits the repre-
+  −ap 
sentation f = f0 + nj=1 ∂j f j in D (Ω) with f j ∈ L p Ω, δ∂Ω L n , 0 ≤ j ≤ n, we
 1, p ∗
define Λ f ∈ W̊a (Ω) as the functional
∫ n ∫

1, p
Λ f (u) := u f0 dL n − (∂j u) f j dL n, ∀u ∈ W̊a (Ω). (8.5.7)
Ω j=1 Ω

Then (8.3.65) ensures that the above definition of Λ f is independent of the

particular representation of the distribution f . In particular, Φ in (8.5.6)-(8.5.7) is
well defined, linear, and bounded. In fact, another application of (8.3.65) gives that
Φ is one-to-one.
Hence, as far as (8.5.4) is concerned, there remains to show that Φ is onto (once
this is established, we conclude that Φ is an isomorphism, which finishes the proof of
 1, p ∗
(8.5.4)). With this goal in mind, fix an arbitrary functional Λ ∈ W̊a (Ω) . Our aim
−1, p 
is to find f ∈ W−a (Ω) such that Λ f = Λ. To get started, consider the application
    n+1
1, p ap
ι : W̊a (Ω) −→ L p Ω, δ∂Ω L n
(8.5.8)
1, p
defined by ι(u) := (u, ∇u) for each u ∈ W̊a (Ω).

Clearly, ι maps its domain in a one-to-one fashion onto its range, Range ι. Thanks
to Proposition 8.3.10, the latter happens to be a closed subspace of the Banach
    n+1
ap
space L p Ω, δ∂Ω L n . In particular, ι is an isomorphism onto the Banach space
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 487

Range ι. Note that if ι−1 denotes the inverse of ι : W̊a (Ω) → Range ι, the Open
1, p

Mapping Theorem ensures that ι−1 is continuous, hence

    n+1
Λ ◦ ι−1 : Range ι −→ L p Ω, δ∂Ω L n
ap
(8.5.9)

is a bounded linear map. By the  * Hahn-Banach Extension Theorem, this map may
 ap n " n+1
∗
be extended to a functional in L Ω, δ∂Ω L
p . We may then invoke Riesz’s
 −ap 
Representation Theorem to conclude that there exist f j ∈ L p Ω, δ∂Ω L n with
1, p
0 ≤ j ≤ n such that for all u ∈ W̊a (Ω) we have
∫ n ∫

  
Λ(u) = Λ ◦ ι−1 ι(u) = u f0 dL n − (∂j u) f j dL n . (8.5.10)
Ω j=1 Ω

+ −1, p
As a consequence, we have Λ = Λ f with f := f0 + nj=1 ∂j f j ∈ W−a (Ω), as
wanted.
Having proved (8.5.4), formula (8.5.5) follows from this and (8.3.66) (after re-
adjusting notation). 

Moving on, the goal is to associate a natural notion of conormal derivative operator
with any given second-order system. To set the stage, fix an open set Ω ⊆ Rn with
an Ahlfors regular boundary. Also, suppose 1 < p, p < ∞ satisfy 1/p + 1/p = 1
and select some a ∈ (−1/p, 1 − 1/p). Then, as seen from (8.5.1)-(8.5.2), any given
second-order homogeneous constant (complex) coefficient M ' × M system L in Rn
becomes a linear and bounded operator in the context
* 1, p "M * −1, p "M
,
L : Wa (Ω) −→ Wa (Ω) . (8.5.11)
* 1, p "M
We wish to associate a notion of conormal derivative for functions in Wa (Ω) .
* 1, p "M
Our definition of the conormal derivative of a function u ∈ Wa (Ω) strongly
depends on the choice of an extension f of Lu, which is a vector-distribution in
* −1, p "M
, * 1, p ∗ ,
Wa (Ω) , to a vector functional belonging to the space W−a (Ω) ] M . By
* 1, p ∗" M
,
this, it is meant that f ∈ W−a (Ω) satisfies
* "M
,
 f , ψ = Lu, ψ, ∀ψ ∈ Cc∞ (Ω) , (8.5.12)

where the left side is understood in the sense of the duality paring between the
*  1, p ∗" M
, * 1, p "M
,
functional f ∈ W−a (Ω) and ψ, viewed as an element in W−a (Ω) , while
the right side is considered in the sense of distributional pairing in Ω. To facilitate
the subsequent discussion we make the following convention.
488 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
*  1, p ∗" M
,
given an arbitrary vector functional f ∈ W−a (Ω) , we
henceforth agree to denote by f Ω the vector distribution in-
duced by f on the open set Ω, that is, the linear assignment (8.5.13)
* ∞ "M , - .
Cc (Ω)  ψ → [(W 1, p (Ω))∗ ] M
, f, ψ
[W
1, p  , ∈ C.
(Ω)] M
−a −a

In this notation, the condition demanded in (8.5.12) then simply reads Lu = f Ω in

*  "M ,
D (Ω) .

Proposition 8.5.3 Let Ω ⊆ Rn be an (ε, δ)-domain with a compact Ahlfors regular

boundary, and such that Rn \ Ω is n-thick. Abbreviate σ := H n−1 ∂Ω, and fix 
p ∈ (1, ∞) along with s ∈ (0, 1), then define a := 1 − s − p1 ∈ − p1 , 1 − p1 .
  −1
Also, let p := 1 − p1 be the Hölder conjugate exponent of p. Next, for some
 
' M ∈ N, select a coefficient tensor A = aαβ
M, , with constant (complex)
jk 1≤α ≤ M
1≤β ≤M
1≤ j,k ≤n
entries, and associate with this coefficient tensor the
 second-order
 homogeneous
' × M system L A := aαβ ∂j ∂k
1≤α ≤ M in R . In this
constant (complex) coefficient M n
jk ,
1≤β ≤M
setting, consider the conormal derivative map
 * 1, p " M *  1, p ∗" M
, * "M
,

∂νA : (u, f ) ∈ Wa (Ω) ⊕ W−a (Ω) : L Au = f Ω in D (Ω)

* p, p "M
,
−→ Bs−1 (∂Ω, σ) (8.5.14)
* 1, p "M
assigning to each pair (u, f ), where u = (uβ )1≤β ≤M ∈ Wa (Ω) and where
*  1, p ∗" M
, *  "M ,
f ∈ W−a (Ω) has the property that L Au = f Ω in D (Ω) , the functional
* "M
, ∗ * p, p "M
,
p, p
∂νA(u, f ) ∈ B1−s (∂Ω, σ) = Bs−1 (∂Ω, σ) (8.5.15)

acting according to (with the summation convention over repeated indices in effect)
- A . αβ - .
p , p  , ∗ ∂ν (u, f ), ϕ p , p  , := a jk ∂k uβ , ∂ j Φα +  f , Φ
([B1−s
(∂Ω,σ)] M ) [B 1−s
(∂Ω,σ)] M

* p, p "M
, * 1, p "M
,
for all ϕ ∈ B1−s (∂Ω, σ) and Φ = (Φα )α ∈ W−a (Ω)
with the property that TrΩ→∂Ω Φ = ϕ,
(8.5.16)
where the first set of brackets in the right side is understood as the integral pairing
 ap   −ap 
between functions from L p Ω, δ∂Ω L n and functions from L p Ω, δ∂Ω L n (Ω), and
the second set of brackets in the right side is the canonical duality pairing between
*  1, p ∗" M
, * 1, p "M
,
W−a (Ω) and W−a (Ω) .
Then the operator (8.5.14)-(8.5.16) is well defined, linear, and bounded in the
sense that there exists a constant C ∈ (0, ∞) with the property that
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 489
! A !  
!∂ (u, f )! p, p , ≤ C u 1, p +  f [(W 1, p (Ω))∗ ] M (8.5.17)
ν [B (∂Ω,σ)] M [W (Ω)] M a
,
−a
s−1

for each pair (u, f ) belonging to the domain of ∂νA (cf. (8.5.14)).
* 1, p "M
Furthermore, for each pair (u, f ) with u = (uβ )1≤β ≤M ∈ Wa (Ω) and
*  1, p ∗" M
, *  "M ,
f ∈ W−a (Ω) such that L Au = f Ω in D (Ω) one has the following
generalized “half” Green’s formula:
- A .
p , p  , ∗ ∂ν (u, f ), TrΩ→∂Ω w p , p  , (8.5.18)
([B 1−s
(∂Ω,σ)] M ) [B (∂Ω,σ)] M
1−s
∫
αβ - .
= a jk , f, w
∂k uβ ∂j wα dL n + [(W 1, p (Ω))∗ ] M [W
1, p  ,
(Ω)] M
−a −a
Ω

* 1, p "M
,
for each w = (wα )α ∈ W−a (Ω) .
* 1, p " M * 1, p ∗" M
,
Finally, for any two given pairs, (u, f ) ∈ Wa (Ω) ⊕ W−a (Ω) such
*  "M , * 1, p "M
, * 1, p ∗" M
that L Au = f Ω in D (Ω) and (w, g) ∈ W−a (Ω) ⊕ Wa (Ω) such that
*  "M 
L A w = g Ω in D (Ω)
 (where A is the transpose of A), one has the following
generalized “full” (or “symmetric”) Green’s formula:
- A .
p , p  , ∗ ∂ν (u, f ), TrΩ→∂Ω w p , p  , (8.5.19)
([B 1−s
(∂Ω,σ)] M ) [B (∂Ω,σ)] M
1−s

- A
.
− ([Bsp, p (∂Ω,σ)] M )∗ ∂ν (w, g), TrΩ→∂Ω u [Bsp, p (∂Ω,σ)] M
- . - .
, f, w
= [(W 1, p (Ω))∗ ] M [W
1, p  , −
(Ω)] M
1, p
[(W (Ω))∗ ] M
g, u [W 1, p (Ω)] M .
−a −a a a

Proof Let us check that the definition of the action of the functional (8.5.15) on any
* p, p "M,
given function ϕ ∈ B1−s (∂Ω, σ) (as described in (8.5.16)) does not depend on
* 1, p "M,
the choice of Φ ∈ W−a (Ω) as long as TrΩ→∂Ω Φ = ϕ. With this goal in mind,
* 1, p "M ,
assuming Φ j ∈ W−a (Ω) and TrΩ→∂Ω Φ j = ϕ for j ∈ {1, 2}, then Φ := Φ1 − Φ2
* 1, p "M,
belongs to W−a (Ω) and has TrΩ→∂Ω Φ = 0. Granted these properties, we
may invoke Proposition 8.3.10 to conclude that there exists a sequence of (vector-
* "M
,
valued) test functions, say Ψ(i) = (Ψα(i) )α ∈ Cc∞ (Ω) indexed by i ∈ N, for which
(i)
* 1, p "M
,
lim Ψ = Φ in W−a (Ω) . Then, if (Φα )α denote the scalar components of Φ,
i→∞
for each pair (u, f ) belonging to the domain of the conormal derivative operator
(8.5.14) we may write
490 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
αβ - - ..
a jk ∂k uβ, ∂j Φα + f , Φ

αβ - . - .
= a jk lim ∂k uβ, ∂j Ψα(i) + lim f , Ψ(i)
i→∞ i→∞

∫ /
αβ - .
= lim a jk ∂k uβ ∂j Ψα(i) dL n + [D  (Ω)] M L Au, Ψ (i)
[D(Ω)] M
i→∞ Ω

∫ /
αβ - .
= lim − a jk uβ ∂k ∂j Ψα(i) dL n + [D  (Ω)] M L Au, Ψ (i)
[D(Ω)] M
i→∞ Ω

= lim 0 = 0, (8.5.20)
i→∞

thanks to the convention made in (8.5.13). This proves that the definition made in
(8.5.16) is indeed coherent. Once this has been established, we may then conclude that
the conormal operator is well defined and linear in the context of (8.5.14). To prove
* 1, p "M *  1, p ∗" M
,
its boundedness, fix a pair (u, f ) ∈ Wa (Ω) ⊕ ∈ W−a (Ω) satisfying
*  "M , * p, p "M
,
L Au = f Ω in D (Ω) , and pick an arbitrary ϕ ∈ B1−s (∂Ω, σ) . Then, if
Ex∂Ω→Ω denotes the vector version of the extension operator from Theorem 8.4.1,
the function Φ := Ex∂Ω→Ω ϕ satisfies
* 1, p "M
,
Φ ∈ W−a (Ω) and Φ[W 1, p (Ω)] M
, ≤ Cϕ
[B
p , p  ,
(∂Ω,σ)] M
(8.5.21)
−a 1−s

for some constant C ∈ (0, ∞) independent of ϕ. Based on (8.5.21) and (8.5.16) we

may then estimate
- .
p , p  , ∗
([B1−s (∂Ω,σ)] M )
∂νA(u, f ), ϕ p , p  ,
[B1−s (∂Ω,σ)] M

αβ
≤ a jk ∂k uβ  L p (Ω,δ a p L n ) ∂j Φα  L p (Ω,δ −a p L n )
∂Ω ∂Ω

+  f [(W 1, p (Ω))∗ ] M
, Φ
[W
1, p  ,
(Ω)] M
−a −a
 
, ϕ
≤ C u[W 1, p (Ω)] M +  f [(W 1, p (Ω))∗ ] M [(B
p , p  ,.
(∂Ω,σ)] M
(8.5.22)
a −a 1−s

* p, p "M,
On account of this, the arbitrariness of ϕ ∈ B1−s (∂Ω, σ) , and Proposition 7.6.1
we conclude that estimate (8.5.17) holds. This finishes the proof of the fact that
the conormal operator is also bounded in the context of (8.5.14). Finally, the gen-
eralized “half” Green’s formula (8.5.18) is a consequence of (8.5.16) used with
ϕ := TrΩ→∂Ω w and Φ := w, while the “full” Green’s formula (8.5.19) is obtained
by subtracting two versions of (8.5.18) (one written for A, and one written for the
transpose coefficient tensor A ). 
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 491

Occasionally, we need the notion of conormal derivative in the more general

setting described below.

Remark 8.5.4 Retain the assumptions on Ω, A, p, s, a, from Proposition 8.5.3. In

addition, fix an arbitrary cutoff function ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near ∂Ω.
* 1, p "M
For each u = (uβ )1≤β ≤M ∈ Wa (Ω)bdd (cf. Convention 8.3.7) and each
*  1, p ∗" M
, * "M,
f ∈ W−a (Ω) with the property that L Au = f Ω in D (Ω) define
  * p, p "M
,
∂νA(u, f ) := ∂νA 'u, '
f ∈ Bs−1 (∂Ω, σ) (8.5.23)

with
* 1, p "M *  1, p ∗" M
,
u := ψu ∈ Wa (Ω)
' and '
f ∈ W−a (Ω) given by

' αβ αβ αβ
f := a jk (∂j ∂k ψ)uβ + a jk (∂j ψ)(∂k uβ ) + a jk (∂k ψ)(∂j uβ ) +ψf.
,
1≤α ≤ M
(8.5.24)
Then the above definition is meaningful, and does not depend on the particular
* 1, p "M
,
cutoff function ψ. Furthermore, for each function w = (wα )α ∈ W−a (Ω) which
vanishes outside a bounded subset of Ω the following generalized “half” Green’s
formula holds:
- A .
p , p  , ∗ ∂ν (u, f ), TrΩ→∂Ω w p , p  , (8.5.25)
([B 1−s
(∂Ω,σ)] M ) [B (∂Ω,σ)] M
1−s
∫
αβ - .
= a jk ∂k uβ ∂j wα dL n + [(W 1, p (Ω))∗ ] M
, f, w
[W
1, p  ,.
(Ω)] M
−a −a
Ω

* 1, p "M
, *  1, p ∗" M
Finally, for each pair (w, g) ∈ W−a (Ω) ⊕ Wa (Ω) with L A w = g Ω
*  "M
in D (Ω) (where A is the transpose of A) and with w vanishing outside a
bounded subset of Ω, the generalized “full” Green’s formula
- A .
p , p  , ∗ ∂ν (u, f ), TrΩ→∂Ω w p , p  , (8.5.26)
([B 1−s
(∂Ω,σ)] M ) [B 1−s
(∂Ω,σ)] M

- A
.
− ([Bsp, p (∂Ω,σ)] M )∗ ∂ν (w, g), TrΩ→∂Ω u [Bsp, p (∂Ω,σ)] M
- . - .
, f, w
= [(W 1, p (Ω))∗ ] M [W
1, p  , −
(Ω)] M
1, p
[(W (Ω))∗ ] M
g, ψu [W 1, p (Ω)] M
−a −a a a

holds for each cutoff function ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near both ∂Ω and the support
of w.

A couple of clarifications regarding Remark 8.5.4 are in order. First, the fact that
* 1, p "M * 1, p "M
u belongs to Wa (Ω)bdd puts '
u := ψu in Wa (Ω) (cf. Convention 8.3.7).
1, p  1, p ∗
Second, since W−a (Ω) is a module over Cc∞ (Rn ), so is W−a (Ω) , via duality. As
*  1, p ∗" M
,
such, ψ f ∈ W−a (Ω) . In addition,
492 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
αβ αβ αβ
a jk (∂j ∂k ψ)uβ + a jk (∂j ψ)(∂k uβ ) + a jk (∂k ψ)(∂j uβ ) belongs to
   1, p ∗ (8.5.27)
ap
L p Ω, δ∂Ω L n → W−a (Ω) for each α ∈ {1, . . . , M ' }.

*  1, p ∗" M
,
Collectively, these properties show that ' f ∈ W−a (Ω) . It is also clear from
*  "M ,
definitions that L A'u = 'f Ω in D (Ω) , in the sense of (8.5.13). Together with
Proposition 8.5.3, these observations show that the definition made in (8.5.23) is
indeed meaningful. Moreover, the fact that said definition is independent of the
cutoff function follows from the fact that, in the context of Proposition 8.5.3, the
conormal derivative ∂νA(u, f ) is the zero distribution whenever u vanishes near ∂Ω.
Finally, the generalized “half” Green’s formula (8.5.25) is seen from (8.5.23) and
(8.5.18), while the generalized “full” Green’s formula (8.5.26) is implied by (8.5.19)
and (8.5.23).
While the value p = ∞ is excluded from Proposition 8.5.3, we nonetheless have
the following more restrictive result corresponding to this end-point:

Proposition 8.5.5 Let Ω ⊆ Rn be a bounded (ε, δ)-domain with an Ahlfors regular

boundary. Abbreviate σ := H n−1 ∂Ω and denote by δ∂Ω the distance
  function to ∂Ω.
' M ∈ N, consider a coefficient tensor A = aαβ
Also, for some M, , with con-
jk 1≤α ≤ M
1≤β ≤M
1≤ j,k ≤n
stant (complex) entries, and associate with this coefficient tensor the second-order

homogeneous constant (complex) coefficient M ' × M system L A := aαβ ∂j ∂k ,
jk 1≤α ≤ M
1≤β ≤M
in Rn (with the summation convention over repeated indices understood throughout).
In this context, for each fixed s∗ ∈ (0, 1) define the conormal derivative operator

* "M
∂νA : u ∈ L ∞ (Ω, L n ) ∩ Wloc
1,1
(Ω) 1−s∗
: δ∂Ω |∇u| ∈ L ∞ (Ω, L n )
/ (8.5.28)
* "M, * "M
,
and L Au = 0 in D (Ω) −→ Bs∞,∞∗ −1
(∂Ω, σ)

associating to each u = (uβ )1≤β ≤M belonging to the domain of ∂νA in (8.5.28) the
functional
* "M
, ∗ * "M
,
∂νAu ∈ B1−s1,1
∗
(∂Ω, σ) = Bs∞,∞
∗ −1
(∂Ω, σ) (8.5.29)

defined according to
∫
- . αβ
1,1
([B1−s , ∗
(∂Ω,σ)] M )
∂νAu, ϕ [B1,1 (∂Ω,σ)] M
, := a jk (∂k uβ )(∂j Φα ) dL n
∗ 1−s∗ Ω (8.5.30)
* 1,1 "M
,
for all ϕ ∈ B1−s ∗
(∂Ω, σ) where (Φα )1≤α ≤ M
, := Ex∂Ω→Ω ϕ.

Above, the summation convention over repeated indices is in effect, and Ex∂Ω→Ω
denotes the extension operator from Theorem 8.4.1 (acting componentwise).
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 493

Then the conormal derivative operator ∂νA considered in (8.5.28)-(8.5.30) is well

defined, linear, and bounded in the sense that there exists a constant C ∈ (0, ∞) with
the property that
! A ! ! !
!∂ u! ∞,∞ ! 1−s∗ !
ν , ≤ C !δ
M ∇u ! ∞ (8.5.31)
[B (∂Ω,σ)]
s∗ −1
∂Ω n L (Ω, L )

for each function u belonging to the domain of ∂νA (cf. (8.5.28)).

Moreover, the conormal derivative operator defined in (8.5.28)-(8.5.30) is com-
patible with the conormal derivative operator introduced in Proposition 8.5.3 in
the following sense: Given any s ∈ (0, s∗ ) and p ∈ (1, ∞) it follows that any func-
* 1, p "M
tion u belonging to the domain of ∂νA in (8.5.28) satisfies u ∈ Wa (Ω) with
a := 1 − s − 1/p and
* "M
,
the conormal derivative ∂νAu ∈ Bs∞,∞ ∗ −1
(∂Ω, σ) defined as in
(8.5.30)-(8.5.30) coincides with the earlier conormal derivative (8.5.32)
* p, p "M
,
∂νA(u, 0) ∈ Bs−1 (∂Ω, σ) defined as in (8.5.16) (with f = 0).

It is worth noting that when M ' = M and A is weakly elliptic, the domain of ∂νA in
* s "M
(8.5.28) simply becomes C ∗ (Ω) ∩ Ker L A, i.e., the space of C M -valued Hölder
functions of order s∗ which are null-solutions of L A in Ω.
That the domain of ∂νA in (8.5.28) is, in the aforementioned circumstances, con-
* "M
tained in C s∗ (Ω) ∩ Ker L A may be seen by combining two observations. First,
by elliptic regularity, any distribution which is a null-solution of L A in Ω is smooth.
Second, in view of item (1) in [133, Proposition 5.11.14], the set Ω is a locally
uniform domain. Granted these, from [133, (5.11.75)] we conclude that
! !
! !
u[C s∗ (Ω)] M ≤ C !|∇u| · dist(·, ∂Ω)1−s∗ ! ∞ n
+ Cu[L ∞ (Ω, L n )] M , (8.5.33)
L (Ω, L )

* "M
hence ultimately u ∈ C s∗ (Ω) .
* "M
Conversely, given any null-solution u of L A belonging to C s∗ (Ω) , elliptic
regularity implies that u is smooth in Ω. Also, by interior estimates (cf. [133,
Theorem 6.5.7]), for each x ∈ Ω we have
⨏
−1
sup |(∇u)(z)| ≤ Cδ∂Ω (x) |u(x) − u(y)| dy
z ∈B(x,δ ∂Ω (x)/2) B(x,δ ∂Ω (x))

≤ Cδ∂Ω (x)s∗ −1, (8.5.34)

for some constant C = C(A, n) ∈ (0, ∞). In turn, this estimate readily implies that
1−s∗
δ∂Ω |∇u| ∈ L ∞ (Ω, L n ), and the desired conclusion follows.
We now turn to the task of presenting the proof of Proposition 8.5.5.
Proof of Proposition 8.5.5 From estimate (8.4.10) (presently used with p := 1,
* 1,1 "M
,
k := 0, and s := 1 − s∗ ) we know that given any ϕ ∈ B1−s ∗
(∂Ω, σ) , the function
494 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Φ := Ex∂Ω→Ω ϕ satisfies
! !
! s∗ −1 !
!δ∂Ω ∇Φ ! ≤ Cϕ[B1,1 ,
(∂Ω,σ)] M
, (8.5.35)
L 1 (Ω, L n ) 1−s∗

for some constant C ∈ (0, ∞) independent of ϕ. From (8.5.35) we then conclude that
whenever the function u = (uβ )1≤β ≤M belongs to the domain of ∂νA in (8.5.28) and
Φ = (Φα )1≤α ≤ M , is as above we may estimate
∫ ! ! ! !
αβ ! s∗ −1 ! ! 1−s∗ !
a jk |∂k uβ ||∂j Φα | dL n ≤ C !δ∂Ω ∇Φ ! 1 n
!δ ∂Ω
∇u ! ∞ n
Ω L (Ω, L ) L (Ω, L )
! !
! 1−s∗ !
≤ Cϕ[B1,1 (∂Ω,σ)] , !δ
M ∂Ω
∇u ! (8.5.36)
1−s∗ L ∞ (Ω, L n )

for some constant C = C(Ω, A, s∗ ) ∈ (0, ∞). This proves that the conormal derivative
operator introduced in (8.5.28)-(8.5.30) is indeed well defined, linear, and bounded.
To deal with the last claim in the statement, fix s ∈ (0, s∗ ), p ∈ (1, ∞), and suppose
the function u belongs to the domain of ∂νA in (8.5.28). Then [133, (8.7.3)] (presently
* 1, p "M
used with Σ := ∂Ω, N := 0, α := 1 − (s∗ − s)p) implies that u ∈ Wa (Ω) if
a := 1 − s − 1/p. With this in hand, the claim in (8.5.32) is now seen from (8.5.16),
(8.5.30), and (8.4.3) (used with p in place of p and 1 − s in place of s, a scenario in
which 1 − (1 − s) − 1/p becomes −a). 

It turns out that for smooth function we may actually compute the conormal
derivative, originally considered as in Proposition 8.5.3, in a pointwise sense, as
indicated below.

Proposition 8.5.6 Let Ω ⊆ Rn be a bounded (ε, δ)-domain with an Ahlfors regular

boundary, and such that Rn \ Ω is n-thick. Abbreviate σ := H n−1 ∂Ω and denote
by ν = (ν1, . . . , νn ) the geometric measure theoretic outward unit normal to Ω. Also,
fix p ∈ (1, ∞), denote by p its Hölder conjugate exponent, pick some s ∈ (0, 1),
and define a := 1 − s − p1 . Next, fix two integers M, ' M ∈ N, and consider a
 αβ 
coefficient tensor A = a jk 1≤α ≤ M , with constant (complex) entries. Associate with
1≤β ≤M
1≤ j,k ≤n
this coefficient tensor the
 second-order
 homogeneous constant (complex) coefficient
M' × M system L A := aαβ ∂j ∂k , in R (as usual, the summation convention
n
jk 1≤α ≤ M
1≤β ≤M
over repeated indices is understood throughout). Finally, consider an arbitrary
* "M * 1, p "M
vector valued function u = (uβ )1≤β ≤M ∈ C ∞ (Ω) ⊆ Wa (Ω) and define
 ∗ M
,
1, p
f ∈ W−a (Ω) by setting (bearing (8.3.35) in mind)
∫
- .
1, p  ,
[(W−a (Ω))∗ ] M
f, F 1, p  ,
[W−a (Ω)] M
:= (L Au)α Fα dL n
Ω
(8.5.37)
 M
,
1, p
for each F = (Fα )1≤α ≤ M
, ∈ W−a (Ω) .
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 495
* "M , * "M,
Then the function L Au ∈ C ∞ (Ω) satisfies L Au = f Ω in D (Ω) and,
* p, p "M
,
with the conormal derivative ∂ν (u, f ) ∈ Bs−1 (∂Ω, σ)
A considered in the sense of
Proposition 8.5.3, one has

αβ
∂νA(u, f ) = ν j a jk (∂k uβ ) ∂Ω . (8.5.38)
,
1≤α ≤ M

* "M
,
Proof From (8.5.37) and definitions it is clear that L Au = f Ω in D (Ω) . To
* "M
,
check (8.5.38), pick some arbitrary ϕ = (ϕα )1≤α ≤ M
, ∈ Lip (∂Ω) and consider
* " ,
M
Φ = (Φα )1≤α ≤ M
, ∈ Lipc (R )
n such that Φ ∂Ω = ϕ. Then
- .
p , p 
([B1−s , ∗
(∂Ω,σ)] M )
∂νA(u, f ), ϕ p , p 
[B1−s ,
(∂Ω,σ)] M
∫ ∫
αβ
= a jk (∂k uβ )(∂j Φα ) dL n + (L Au)α Φα dL n
Ω Ω
∫
αβ
= ν j a jk (∂k uβ ) ∂Ω ϕα dσ, (8.5.39)
∂Ω

thanks to (8.5.16), the definition of f in (8.5.37), and the De Giorgi-Federer ver-

sion of Divergence Formula from [133, Theorem 1.1.1] (bearing in mind that the
present geometric assumptions entail ∂∗ Ω = ∂Ω; cf. (A.0.89), [133, (5.11.35)], and
[133, Lemma 5.11.9)]). With this in hand, (8.5.38) now follows with the help of
Lemma 7.1.10. 

Proposition 8.5.7, stated below, deals with the definition and properties of what
we shall refer to as the principal symbol map. To state the stage, recall the convention
made in (8.5.13).

Proposition 8.5.7 Suppose Ω ⊆ Rn is an (ε, δ)-domain with a compact Ahlfors

regular boundary, and such that Rn \ Ω is n-thick. Abbreviate σ := H n−1 ∂Ω, and
fix two exponents p, p ∈ (1, ∞) satisfying 1/p + 1p = 1 along with some s ∈ (0, 1),
then set a := 1 − s − p1 ∈ − p1 , 1 − p1 . Also, let D be an N × M homogeneous
first-order system with constant (complex) coefficients, say

n
γβ
D= bk ∂k 1≤γ ≤ N . (8.5.40)
j=k 1≤β ≤M

In this setting, consider the map

 *  " M * 1, p ∗" N
ap
(−i)Sym(D; ν) : (u, f ) ∈ L p Ω, δ∂Ω L n ⊕ W−a (Ω) :

* "N * p, p "N
Du = f Ω in D (Ω) −→ Bs−1 (∂Ω, σ) (8.5.41)
496 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
*  ap " M *  1, p ∗" N
assigning to each pair (u, f ) with u ∈ L p Ω, δ∂Ω L n and f ∈ W−a (Ω)
*  "N
with the property that Du = f Ω in D (Ω) the functional
* "N ∗ * p, p "N
p, p
(−i)Sym(D; ν)(u, f ) ∈ B1−s (∂Ω, σ) = Bs−1 (∂Ω, σ) (8.5.42)

acting according to (with the summation convention over repeated indices in effect)
- .
([B
p , p  N
(∂Ω,σ)] )∗ (−i)Sym(D; ν)(u, f ), ϕ [B
p , p 
(∂Ω,σ)] N
1−s 1−s
∫ - .
:=  f , Φ − Ω
u, D Φ dL n
(8.5.43)
* p, p "N * 1, p "N
for all ϕ ∈ B1−s (∂Ω, σ) and Φ ∈ W−a (Ω)
with the property that TrΩ→∂Ω Φ = ϕ,

where the first set of brackets in the right-hand side of the first line of (8.5.43) is the
* 1, p ∗" N * 1, p "N
canonical duality pairing between W−a (Ω) and W−a (Ω) .
Then, in relation to this “symbol” map, the following statements are true.

(1) The operator (8.5.41)-(8.5.43) is well defined, linear, and bounded in the sense
that there exists a constant C ∈ (0, ∞) with the property that
! !
!(−i)Sym(D; ν)(u, f )! p, p (8.5.44)
[B (∂Ω,σ)] N s−1
 
≤ C u[L p (Ω,δ a p L n )] M +  f [(W 1, p (Ω))∗ ] N
∂Ω −a

for each pair (u, f ) belonging to the domain of (−i)Sym(D; ν) (cf. (8.5.41)).
*  ap " M * 1, p ∗" N
(2) For each pair (u, f ) with u ∈ L p Ω, δ∂Ω L n and f ∈ W−a (Ω) such
*  "N
that Du = f Ω in D (Ω) one has the following generalized integration by
parts formula:
- .
([B
p , p 
(∂Ω,σ)] N )∗
(−i)Sym(D; ν)(u, f ), TrΩ→∂Ω w [B p, p (∂Ω,σ)] N
1−s 1−s
∫
- . * 1, p "N
=  f , w − u, D w dL n for each w ∈ W−a (Ω) .
Ω
(8.5.45)

(3) In addition to the N × M system D from (8.5.40), consider a homogeneous,

' × N system D
constant (complex) coefficient, first-order M ' in Rn , say


n
'= 'αγ
D b j ∂j 1≤α ≤M , (8.5.46)
j=1 1≤γ ≤ N
8.5 Weighted Sobolev Spaces of Negative Order, Duality, and Conormal Derivatives 497

'
and define the homogeneous, constant (complex) coefficient, second-order M×M
system
'
L := DD. (8.5.47)
Also, define the coefficient tensor (with the summation convention over repeated
indices in effect)
 αβ  αβ
AD,D
' := a jk 1≤α ≤ M 'αγ γβ
, where each a jk := b j bk . (8.5.48)
1≤β ≤M
1≤ j,k ≤n

* 1, p "M * 1, p ∗" M
,
Then for each pair (u, f ) with u ∈ Wa (Ω) and f ∈ W−a (Ω) such
*  "M ,
that Lu = f Ω in D (Ω) one has
A D,
∂ν
,D
' ν)(Du, f ),
(u, f ) = (−i)Sym( D; (8.5.49)

where the conormal derivative in the left-hand side above is defined in the sense
of Proposition 8.5.3, and the principal symbol map in the right-hand side is
'
defined as in (8.5.41) (with u replaced by Du and D replaced by D).

Proof The claim in item (1) may be justified in a very similar fashion to the proof of
Proposition 8.5.3. Also, the claim in item (2) is implied by item (1), while the claim
in item (3) is clear from (8.5.43) and (8.5.16). 

We conclude this section by recording the following Green-type integration by

parts formula involving a null-solution of a system L and a null-solution of the
transpose system L  , belonging to certain suitably weighted Sobolev spaces with
dual integrability exponents.

Corollary 8.5.8 Assume Ω ⊆ Rn is an (ε, δ)-domain with a compact Ahlfors regular

boundary, and such that Rn \ Ω is n-thick. Abbreviate σ := H n−1 ∂Ω, and fix two
exponents p, p ∈ (1,
 ∞)

 1/p + 1p = 1 along with some s ∈ (0, 1), then set
satisfying
a := 1 − s − p ∈ − p , 1 − p . Next, consider a homogeneous, constant (complex)
1 1 1

coefficient, first-order M ' × N system D

' in Rn , along with a homogeneous, constant
(complex) coefficient, first-order N × M system D in Rn , and define the homogeneous,
constant (complex) coefficient, second-order M ' × M system

'
L := DD. (8.5.50)

Then for any two vector-valued functions,

* 1, p "M
u ∈ Wa (Ω) with Lu = 0 in Ω, and
* 1, p "M
,
(8.5.51)
w ∈ W−a (Ω) with L  w = 0 in Ω,

one has
498 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
0 1
p, p , ' ν)(Du, 0), TrΩ→∂Ω w
(−i)Sym( D; p , p  , (8.5.52)
[B s−1 (∂Ω,σ)] M [B1−s (∂Ω,σ)] M
0 1
' w, 0) p, p
= [Bsp, p (∂Ω,σ)] M TrΩ→∂Ω u, (−i)Sym(D ; ν)( D .
[B (∂Ω,σ)] M −s

Proof This follows from (8.5.19) and (8.5.49). 

8.6 Traces from Weighted Maximal Sobolev Spaces in Open

Subsets of R n

We begin by introducing a special brand of Sobolev spaces, requiring the membership

of distributional derivatives to solid maximal Lebesgue spaces defined as in (A.0.63)
with μ a power weighted Lebesgue measure. Specifically, we make the following
definition.

Definition 8.6.1 Let Ω be an open, nonempty, proper subset of Rn , and pick an

integrability exponent p ∈ (0, ∞) along with an integer k ∈ N0 and some power
k, p
a ∈ R. In this context, define the weighted maximal Sobolev space Wa, (Ω) as
 
p 
(Ω) : ∂ α u ∈ L Ω, δ∂Ω L n for all α ∈ N0n with |α| ≤ k
k, p k,1 ap
Wa, (Ω) := u ∈ Wloc
(8.6.1)
k, p
and equip it with the quasi-norm given for each u ∈ Wa, (Ω) by

uW k, p (Ω) := ∂ α u Lp (Ω,δ a p L n )
a, ∂Ω
|α | ≤k

 ∫ 1
p
(∂ α u),θ δ∂Ω dL n
p ap
≈ for θ ∈ (0, 1). (8.6.2)
|α | ≤k Ω

In the same setting as above, let us also consider the homogeneous weighted
. k, p
maximal Sobolev space homogeneous weighted maximal Sobolev Wa, (Ω) defined
as
. k, p  
p 
(Ω) : ∂ α u ∈ L Ω, δ∂Ω L n for all α ∈ N0n with |α| = k
k,1 ap
Wa, (Ω) := u ∈ Wloc
. k, p (8.6.3)
and for each u ∈ Wa, (Ω) set

uW. k, p (Ω) := ∂ α u Lp (Ω,δ a p L n ) . (8.6.4)
a, ∂Ω
|α |=k

In contrast to the standard generic L p -based Sobolev spaces in Ω which turn

out to have good functional analytic properties only when p ≥ 1, our weighted
maximal Sobolev spaces (of the sort introduced in Definition 8.6.1) turn out to
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 499

behave remarkably well even when p < 1. This point is underscored by the trace
result proved in Theorem 8.6.8, the regularity results established in Corollary 9.2.31
and Corollary 9.2.38, and the fact that basic singular integral operators, like the
double and single layer potential operators associated with an elliptic system, behave
naturally on this scale of spaces (a topic addressed at length in [136, Chapter 4]).
For now, we wish to make several remarks in relation to the brand of weighted
Sobolev spaces from Definition 8.3.4. First, given any open nonempty proper subset
Ω of Rn , from [133, (6.6.44)] we conclude that we have the continuous embedding
k, p k, p
Wa, (Ω) → Wa (Ω) for all p ∈ (0, ∞), k ∈ N0, and a ∈ R. (8.6.5)

The second remark concerning the weighted Sobolev spaces from Definition 8.6.1
and Definition 8.3.4 is contained in the lemma below. To facilitate its statement, we
make the following convention. Given a homogeneous constant (complex) coefficient
second-order weakly elliptic M × M system L in Rn along with an arbitrary open set
Ω ⊆ Rn , we agree to abbreviate
 * "M 
Ker L := u ∈ C ∞ (Ω) : Lu = 0 in Ω . (8.6.6)

The reader is also reminded that the characteristic matrix of L is defined in (A.0.64).

Lemma 8.6.2 With M, m ∈ N, let L be a constant (complex) coefficient homogeneous

M × M system of order 2m in Rn , with the property that det [L(ξ)]  0 for each
ξ ∈ Rn \ {0}. Also, assume Ω ⊆ Rn is an open, nonempty, proper subset of Rn , and
suppose
0 < p < ∞, k ∈ N0, a ∈ R. (8.6.7)
Then
* k, p "M * k, p "M
Wa, (Ω) ∩ Ker L = Wa (Ω) ∩ Ker L with equivalent norms. (8.6.8)

Proof This is a consequence of (8.6.5), [133, (6.5.40) in Theorem 6.5.7], and [133,
(6.6.91)]. 

Our third remark pertaining to the relation between the weighted Sobolev spaces
from Definition 8.6.1 and Definition 8.3.4 is discussed in the next proposition.

Proposition 8.6.3 Let Ω be an open, nonempty, proper subset of Rn . Suppose

 
n < p ≤ 1, (n − 1) p1 − 1 < s < 1, k ∈ N0,
n−1

  
then set a := 1 − s − p1 ∈ − p1 , −n p1 − 1 (8.6.9)
np
and q := n−1+p−sp ∈ (1, n).

Then one has the continuous embedding

k, p
Wa, (Ω) → W k,q (Ω). (8.6.10)
500 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Proof From the embedding in [133, (6.6.74)] used with α := (1 − p + sp)/(np)

(which, thanks to (8.6.9), is a number belonging to the interval (0, 1/p)) we see that
we have the continuous embedding
p p(1−s)−1 n 
L Ω, δ∂Ω L → L q (Ω, L n ). (8.6.11)

In turn, this readily implies (8.6.10), upon noting that ap = p(1 − s) − 1. 

The fourth (and last) remark we make is contained in the following lemma, where
we compare smooth truncations with restrictions to bounded sets (something which
is going to be relevant later on).
Lemma 8.6.4 Let Ω be an open (nonempty, proper) subset of Rn with compact
boundary, and for each R > 0 define ΩR := Ω ∩ B(0, R). Also, fix k ∈ N0 . Suppose
u ∈ C k (Ω) is a function whose derivatives of order ≤ k are all subaveraging in Ω
(in the sense of [133, Definition 6.5.1]). Finally, select some p ∈ (0, ∞) along with
a > −1/p. Then the following equivalence holds:

ψu ∈ Wa (Ω) for each ψ ∈ Cc∞ (Rn )

k, p
(8.6.12)

if and only if
k, p
u Ω R ∈ Wa, (ΩR ) for each sufficiently large R > 0. (8.6.13)

Proof The property formulated in (8.6.12) is readily seen to be equivalent to de-

manding that
 
∂ α u ∈ L p ΩR, δ∂Ω
a L n for each sufficiently large R > 0
(8.6.14)
and each multi-index α ∈ N0n with |α| ≤ k.

Note that for each sufficiently large R > 0 we have

 ∫
δ∂Ω R |∂ α u| p dL n < +∞
ap
(8.6.15)
|α | ≤k Ω R \Ω R/2

  ∫ ap
since u ∈ C k ΩR \ ΩR/2 and Ω δ∂Ω R dL n < +∞ given that ap > −1. Also,
R \Ω R/2

δ∂Ω R ≈ δ∂Ω on ΩR/2 . (8.6.16)

From (8.6.15)-(8.6.16) we then see that (8.6.14) is further equivalent to having

k, p
u Ω R ∈ Wa (ΩR ) for each sufficiently large R > 0. (8.6.17)

Finally, that this is equivalent to (8.6.13) is clear from [133, Lemma 6.6.7] (bearing
in mind [133, (6.5.2)]). 
Our principal result in this section is Theorem 8.6.8 dealing with the nature
of the trace operator acting from power weighted maximal Sobolev spaces, of the
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 501

sort introduced in Definition 8.6.1. The aforementioned trace theorem is stated and
proved later. To facilitated dealing with this, for the time being we discuss several
key technical results, starting with the following lemma.
 
Lemma 8.6.5 Suppose Φ : Rn × Rn \ diag → C is a function with the property
that there exist a constant A ∈ (0, ∞) along with two exponents, M ∈ (0, ∞) and
α ∈ (0, 1], such that
A
|Φ(x, z)| ≤ if x, z ∈ Rn with x  z, (8.6.18)
|x − z| M
A|x − y| α
Φ(x, z) − Φ(y, z) ≤ if x, y, z ∈ Rn with |x − y| < 12 |x − z|. (8.6.19)
|x − z| M+α
Also, assume

p ∈ (0, ∞), θ ∈ (0, 1), and

(8.6.20)
max n − 1, 2(n − 1) − M p < μ < n − 1 + αp.

Finally, having considered a closed upper Ahlfors regular set Σ ⊆ Rn , denote by δΣ

the distance function to Σ, and abbreviate σ := H n−1 Σ.
Then there exists a constant C ∈ (0, ∞), depending only on A, M, α, p, θ, μ, n, with
the property that
∫ ∫ ⨏ p
dσ(x) dσ(y)
Φ(x, z) − Φ(y, z) dz ≤ CδΣ (z0 )2(n−1)−μ−M p
Σ Σ B(z0,θ δΣ (z0 )) |x − y| μ
(8.6.21)

for each point z0 ∈ Rn \ Σ.

 
Proof Fix ε ∈ 0, (1 − θ)/4
 and consider
 an arbitrary point z0 ∈ Rn \ Σ. Then for
each x, y ∈ Σ and z ∈ B z0, θδΣ (z0 ) we have
 θ   θ 
|z − z0 | < θδΣ (z0 ) ≤ 1−θ δΣ (z) ≤ 1−θ |z − x|, (8.6.22)

so
 
|x − z0 | ≤ |x − z| + |z − z0 | < |x − z| + θ
1−θ |z − x| = (1 − θ)−1 |x − z|. (8.6.23)

In particular,

|x − y| < ε|x − z0 | =⇒ |x − y| < 12 |x − z|. (8.6.24)

Work now under the assumption that |x − y| < 12 |y − z0 | (which, given that ε < 1/2,
is the case if |x − y| < ε|y − z0 |). Then |y − z0 | ≤ |y − x| + |x − z0 | < 12 |y − z0 | + |x − z0 |,
hence |y − z0 | < 2|x − z0 | which, in combination with (8.6.22), gives
 2 
|y − z0 | < 2|x − z0 | ≤ 2|x − z| + 2|z − z0 | ≤ 1−θ |x − z|. (8.6.25)
502 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Thus,

|x − y| < ε|y − z0 | =⇒ |x − y| < 12 |x − z|. (8.6.26)

Moving on, consider

 
S := (x, y) ∈ Σ × Σ : |x − y| < ε max |x − z0 |, |y − z0 | . (8.6.27)

Then

S = S1 ∪ S2, (8.6.28)

where
  
S1 := (x, y) ∈ Σ × Σ : y ∈ B x, ε|x − z0 | ,
   (8.6.29)
S2 := (x, y) ∈ Σ × Σ : x ∈ B y, ε|y − z0 | ,

and
 
|x − y| < 12 |x − z| whenever (x, y) ∈ S and z ∈ B z0, θδΣ (z0 ) . (8.6.30)

Rely on (8.6.28) to write

∬ ⨏ p
dσ(x) dσ(y)
Φ(x, z) − Φ(y, z) dz ≤ I1 + I2, (8.6.31)
(x,y)∈S B(z0,θ δΣ (z0 )) |x − y| μ

where, for j ∈ {1, 2},

∬ ⨏ p
A|x − y| α dσ(x) dσ(y)
I j := dz . (8.6.32)
(x,y)∈S j B(z0,θ δΣ (z0 )) |x − z| M+α |x − y| μ

Note that
∬
|x − y| αp dσ(x) dσ(y)
I1 ≤ C
(x,y)∈S1 |x − z0 |
(M+α)p |x − y| μ
∫ ∫ 
1 dσ(y)
=C μ−αp
dσ(x)
y ∈Σ∩B(x,ε |x−z0 |) |x − y|
(M+α)p
x ∈Σ |x − z0 |
∫
dσ(x)
≤C ≤ CδΣ (z0 )2(n−1)−μ−M p . (8.6.33)
x ∈Σ |x − z0 | M p+μ−(n−1)
Above, the first inequality is based on (8.6.30) and (8.6.19), plus the observation
 that
|x − z| ≈ |x − z0 | uniformly for x ∈ Σ, z0 ∈ Rn \ Σ, and z ∈ B z0, θδΣ (z0 ) . The
subsequent equality uses the description of S1 given in (8.6.29). The penultimate
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 503

inequality is a consequence of the first estimate in [133, (7.2.5)] (since μ−αp < n−1),
while the last inequality in (8.6.33) is implied by [133, (8.7.92)] (presently used with
Ω := Rn \ Σ), bearing in mind that M p + μ − (n − 1) > n − 1. Going further, we write
∬
|x − y| αp dσ(x) dσ(y)
I2 ≤ C
(x,y)∈S2 |y − z0 |
(M+α)p |x − y| μ
∫ ∫ 
1 dσ(x)
=C μ−αp
dσ(y)
x ∈Σ∩B(y,ε |y−z0 |) |x − y|
(M+α)p
y ∈Σ |y − z0 |
∫
dσ(y)
≤C ≤ CδΣ (z0 )2(n−1)−μ−M p . (8.6.34)
y ∈Σ |y − z0 | M p+μ−(n−1)

The first estimate in (8.6.34) is seen from (8.6.30), (8.6.19), and (8.6.25). The
following equality in (8.6.34) is based on the description of S2 from (8.6.29). The
last two inequalities in (8.6.34) are justified as the case of (8.6.33), based on the first
estimate in [133, (7.2.5)] and [133, (8.7.92)].
Going further, make use of (8.6.28) to write, for some constant C ∈ (0, ∞)
depending only on the exponent p,
∬ ⨏ p
dσ(x) dσ(y)
Φ(x, z) − Φ(y, z) dz
(x,y)∈(Σ×Σ)\S B(z0,θ δΣ (z0 )) |x − y| μ
 
≤ C II1 + II2 , (8.6.35)

where
∬ ⨏ p
dσ(x) dσ(y)
II1 := |Φ(x, z)| dz (8.6.36)
B(z0,θ δΣ (z0 )) |x − y| μ
(x,y)∈Σ×Σ
|x−y | ≥ε |x−z0 |

and
∬ ⨏ p
dσ(x) dσ(y)
II2 := |Φ(y, z)| dz . (8.6.37)
B(z0,θ δΣ (z0 )) |x − y| μ
(x,y)∈Σ×Σ
|x−y | ≥ε |y−z0 |
 
In turn, since |x−z| ≈ |x−z0 | uniformly for x ∈ Σ, z0 ∈ Rn \Σ, and z ∈ B z0, θδΣ (z0 ) ,
we may invoke (8.6.18), the second inequality in [133, (7.2.5)] (keeping in mind that
μ > n − 1), and [133, (8.7.92)] to estimate
504 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
∫ ∫ 
1 dσ(y)
II1 ≤ C dσ(x)
x ∈Σ |x − z0 | M p y ∈Σ\B(x,ε |x−z0 |) |x − y| μ
∫
dσ(x)
≤C ≤ CδΣ (z0 )2(n−1)−μ−M p . (8.6.38)
x ∈Σ |x − z0 | M p+μ−(n−1)

 have |y − z| ≈ |y − z0 | uniformly for y ∈ Σ, z0 ∈ R \ Σ, and

Likewise, since we n

z ∈ B z0, θδΣ (z0 ) , we may rely on (8.6.18), the second inequality in [133, (7.2.5)],
and [133, (8.7.92)] to write
∫ ∫ 
1 dσ(x)
II2 ≤ μ
dσ(y)
y ∈Σ |y − z0 | x ∈Σ\B(y,ε |y−z0 |) |x − y|
Mp

∫
dσ(y)
≤C ≤ CδΣ (z0 )2(n−1)−μ−M p . (8.6.39)
y ∈Σ |y − z0 | M p+μ−(n−1)

At this stage, the estimate claimed in (8.6.21) follows by combining (8.6.31), (8.6.33),
(8.6.34), (8.6.35), (8.6.38), and (8.6.39). 
The proposition below prefigures the format of the main trace result, discuss later
in Theorem 8.6.8.
 
Proposition 8.6.6 Assume Φ : Rn × Rn \ diag → C (where n ∈ N, with n ≥ 2) is
a function satisfying, for some constant C ∈ (0, ∞),
C
|Φ(x, z)| ≤ if x, z ∈ Rn with x  z, (8.6.40)
|x − z| n−1
C|x − y|
Φ(x, z) − Φ(y, z) ≤ if x, y, z ∈ Rn with |x − y| < 12 |x − z|. (8.6.41)
|x − z| n
Also, consider an open, nonempty, proper subset Ω of Rn with an upper Ahlfors
regular boundary. Abbreviate σ := H n−1 ∂Ω, and recall that δ∂Ω denotes the
distance function to ∂Ω. Finally, suppose
 
n < p ≤ 1, (n − 1) p1 − 1 < s < 1,
n−1
(8.6.42)

and fix a function
p p(1−s)−1 n 
ϕ ∈ L Ω, δ∂Ω L . (8.6.43)
Then
∫
|Φ(x, z)||ϕ(z)| dz < +∞ for H n−1 -a.e. point x ∈ Rn, (8.6.44)
Ω

and the function (defined by means of an absolutely convergent integral) as

∫
u(x) := Φ(x, z)ϕ(z) dz for H n−1 -a.e. x ∈ Rn (8.6.45)
Ω
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 505

satisfies, for some constant C ∈ (0, ∞) independent of ϕ,

∫ ∫  p1
|u(x) − u(y)| p
dσ(x) dσ(y) ≤ Cϕ L p (Ω,δ p(1−s)−1 L n ) . (8.6.46)
∂Ω ∂Ω |x − y| n−1+sp  ∂Ω

Moreover, the limit

⨏
(TrRn →∂Ω u)(x) := [u]Rn (x) := lim+ u(y) dy
r→0 B(x,r) (8.6.47)
exists and equals u(x) at σ-a.e. point x ∈ ∂Ω.

To offer an example, work in the two-dimensional setting. Specifically, consider

an open, nonempty, proper subset Ω of R2 ≡ C with an upper Ahlfors regular
boundary, and pick a function
p p(1−s)−1 2 
ϕ ∈ L Ω, δ∂Ω L where 1
2 < p ≤ 1 and 1
p − 1 < s < 1. (8.6.48)

Then
∫
|ϕ(ζ)|
dL 2 (ζ) < +∞ for H 1 -a.e. point z ∈ C, (8.6.49)
Ω |ζ − z|
and the function
∫
ϕ(ζ)
u(z) := dL 2 (ζ) for H 1 -a.e. z ∈ C (8.6.50)
Ω ζ−z
(which is well defined, since it is given by an absolutely convergence integral)
satisfies, for some constant C ∈ (0, ∞) independent of ϕ,
∫ ∫  p1
|u(z) − u(w)| p
dH 1 (z) dH 1 (w) ≤ Cϕ L p (Ω,δ p(1−s)−1 L 2 ), (8.6.51)
∂Ω ∂Ω |z − w| 1+sp  ∂Ω

and has the property that

⨏
the limit (TrC→∂Ω u)(z) := lim+ u dL 2
r→0 B(z,r) (8.6.52)
exists and equals u(z) at H -a.e.
1 point z ∈ ∂Ω.

Here is the proof of Proposition 8.6.6.

Proof of Proposition 8.6.6 Introduce
np
q := n−1+p−sp ∈ (1, n), (8.6.53)

with the membership implied by (8.6.42). Then, if ϕ ' denotes the extension of ϕ to
Rn by zero outside Ω, based on (8.6.11) we see that
506 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

ϕ
' ∈ L q (Rn, L n ) and  ϕ
' L q (Rn, L n ) ≤ Cϕ L p (Ω,δ p(1−s)−1 L n ) (8.6.54)
 ∂Ω

for some constant C ∈ (0, ∞) independent of ϕ. We shall show next that

∫
|ϕ(z)|
dz < +∞ for H n−1 -a.e. point x ∈ Rn . (8.6.55)
Ω |x − z| n−1

To this end, [61, Theorem 4.1, p. 294] (used with α := 1 and p := 1) plus [49,
Remark on p. 156, Theorem 3 on p. 193] give
∫
|ϕ
'(z)|
dz < +∞ for H n−1 -a.e. point x ∈ Rn . (8.6.56)
B(x,1) |x − z|
n−1

See also [5, §4] and the comment in relation to [105, (7.8), p. 178] in this regard.
Let us next observe that q , the Hölder conjugate exponent of q ∈ (1, n), satisfies
q  > n/(n − 1). Consequently, for each x ∈ Rn we may write
∫ ∫ ∫
|ϕ(z)| |ϕ
'(z)| |ϕ
'(z)|
dz = dz + dz
Ω |x − z| n−1
B(x,1) |x − z| n−1 n
R \B(x,1) |x − z| n−1
∫ ∫
|ϕ
'(z)| dz 1
q
≤C dz + C ϕ
'  L q (R n , L n ) (n−1)q 
B(x,1) |x − z| R n \B(0,1) |z|
n−1

< +∞ for H n−1 -a.e. point x ∈ Rn, (8.6.57)

thanks to (8.6.56), Hölder’s inequality, (8.6.54), and the fact that (n − 1)q  > n. This
proves (8.6.44). In turn, (8.6.55) is a consequence of (8.6.44) and (8.6.40).
Turning our attention to (8.6.46), consider a Whitney decomposition of Ω into
Euclidean cubes Q j of side-length (Q j ) (cf. [133, Proposition 7.5.3]). Concretely,
suppose
 
j ∈N Q j =Ω= j ∈N 2Q j , Q̊ j ∩ Q̊ k =  if j  k,
δ∂Ω (x) ≈ (Q j ) uniformly for j ∈ N and x ∈ 2Q j , (8.6.58)
+
and j ∈N 12Q j ≤ C for a constant C = Cn ∈ (0, ∞).

Having fixed some θ ∈ (0, 1) we may then estimate

8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 507
∫ ∫
|u(x) − u(y)| p
dσ(x) dσ(y)
∂Ω |x − y|
n−1+sp
∂Ω

∫ ∫ ∫ /p
dσ(x) dσ(y)
≤ |ϕ(z)| Φ(x, z) − Φ(y, z) dz
∂Ω ∂Ω Ω |x − y| n−1+sp
∫ ∫ /p
∫ dσ(x) dσ(y)
= |ϕ(z)| Φ(x, z) − Φ(y, z) dz
∂Ω ∂Ω j Qj |x − y| n−1+sp

∫ ∫ ∫ /p
 dσ(x) dσ(y)
≤ sup |ϕ| Φ(x, z) − Φ(y, z) dz
∂Ω ∂Ω j Qj Qj |x − y| n−1+sp

∫ ∫  ⨏ 1
p p
≤C ϕ,θ dL n ×
∂Ω ∂Ω j 2Q j

∫ /p
dσ(x) dσ(y)
× Φ(x, z) − Φ(y, z) dz
Qj |x − y| n−1+sp
 ⨏ p
≤C ϕ,θ dL n ×
j 2Q j

∫ ∫ ∫ /p
dσ(x) dσ(y)
× Φ(x, z) − Φ(y, z) dz
∂Ω ∂Ω Qj |x − y| n−1+sp
 ∫ p n(p−1)
≤C ϕ,θ δ∂Ω dL n ×
j 2Q j

∫ ∫ ⨏ /p
dσ(x) dσ(y)
× Φ(x, z) − Φ(y, z) dz
∂Ω ∂Ω Qj |x − y| n−1+sp
 ∫
dL n (Q j )(n−1)(1−p)−sp
p n(p−1)
≤C ϕ,θ δ∂Ω
j 2Q j

∫ p p(1−s)−1
∫
p p(1−s)−1
≤C ϕ,θ δ∂Ω dL n ≤ C ϕ,θ δ∂Ω dL n
j 2Q j Ω

p
= Cϕ p p(1−s)−1 . (8.6.59)
L (Ω,δ ∂Ω Ln)

The first inequality above uses (8.6.44)-(8.6.45), and the subsequent equality is
based on the first equality in the first line of (8.6.58). Next, the second inequality in
(8.6.59) is obvious, while the third inequality in (8.6.59) is a consequence of [133,
Proposition 6.6.3], [133, Lemma 6.5.2], [133, (6.5.13)], and [133, (6.6.46)]. Going
508 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

further, the fourth inequality in (8.6.59) is implied by the fact that p ∈ (0, 1], while the
fifth inequality in (8.6.59) (where the integral average has been relocated) is justified
by the second line in (8.6.58). The sixth inequality in (8.6.59) is consequence of
Lemma 8.6.5 used here with Σ := ∂Ω and μ := n − 1 + sp, keeping in mind that
the current function Φ satisfies (8.6.18)-(8.6.19) with M := n − 1 and α := 1 (so
the inequalities in (8.6.20) are satisfied by these choices, thanks to (8.6.42)). Going
further, the seventh inequality in (8.6.59) originates from the second line in (8.6.58),
whereas the eighth inequality in (8.6.59) comes from the second equality in the first
line of (8.6.58) and the third line in (8.6.58). The final equality in (8.6.59) is a simple
consequence of the definition of the quasi-norm in the maximal Lebesgue space
p p(1−s)−1 n 
L Ω, δ∂Ω L .
To deal with the very last property in the statement of the proposition, we first
make the claim that
if α ∈ (0, n) and if Iα (x) := |x| α−n for each x ∈ Rn , then there
exists a finite constant C = C(n, α) > 0 with the property that (8.6.60)
1B(0,r) ∗ Iα ≤ C · Iα everywhere in Rn for each radius r ∈ (0, 1).

To prove (8.6.60), fix x ∈ Rn \ {0} along with r ∈ (0, 1) and observe that

∫
  dy
1B(0,r) ∗ Iα (x) = . (8.6.61)
B(x,r) |y|
n−α

In the case when |x| ≥ 2 any point  y ∈ B(x, r) satisfies |y| ≥ |x|/2 which, in
light of (8.6.61), further implies 1B(0,r) ∗ Iα (x) ≤ 2n−α Iα (x). Finally, in the case
when |x| ≤ ∫2 we have 2α−n ≤ Iα (x) as well as B(x, r) ⊆ B(0, 3). In view of
the formula B(0,3) |y| α−n dy = 3α n−1 ωn−1 , we ultimately conclude that we have
 
1B(0,r) ∗Iα (x) ≤ 2n−α 3α n−1 ωn−1 Iα (x) in this case. The claim in (8.6.60) is therefore
justified.
Going further, consider a point x ∈ Rn with the property that
∫
|ϕ(z)|
dz < +∞. (8.6.62)
Ω |x − z| n−1

With I1 (w) := |w| 1−n for each w ∈ Rn , it follows from (8.6.40) and (8.6.60) (presently
employed with α := 1) that there exists some C ∈ (0, ∞) with the property that for
each z ∈ Rn and r ∈ (0, 1) we have
∫ ∫
dy  
|Φ(y, z)| dy ≤ C = C 1B(0,r) ∗ I1 (x − z)
B(x,r) |y − z|
n−1
B(x,r)

C
≤ C · I1 (x − z) = . (8.6.63)
|x − z| n−1
Granted (8.6.62)-(8.6.63), Fubini’s Theorem allows us to write
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 509
∫ ∫ ∫
u(y) dy = Φ(y, z)ϕ(z) dz dy
B(x,r) B(x,r) Ω
∫ ∫
= Φ(y, z) dy ϕ(z) dz for each r ∈ (0, 1). (8.6.64)
Ω B(x,r)

Since (8.6.41) implies that for each z ∈ Rn the function Φ(·, z) is continuous in
Rn \ {z}, we have
⨏
lim+ Φ(y, z) dy = Φ(x, z) for each z ∈ Rn \ {x}. (8.6.65)
r→0 B(x,r)

Based on (8.6.64)-(8.6.65) and Lebesgue’s Dominated Convergence Theorem we

then conclude that
⨏ ∫ ⨏
lim+ u(y) dy = lim+ Φ(y, z) dy ϕ(z) dz
r→0 B(x,r) r→0 Ω B(x,r)
∫
= lim+ Φ(x, z)ϕ(z) dz = u(x). (8.6.66)
r→0 Ω

The above argument shows that

⨏
the limit [u]Rn (x) := lim+ u(y) dy exists and
r→0 B(x,r) (8.6.67)
equals u(x) at each point x ∈ Rn where (8.6.62) holds.

Then the claim in (8.6.47) becomes a consequence of this and (8.6.55). 

In order to wrap up the preparations for dealing with the main trace result in this
section, subsequently presented in Theorem 8.6.8, we need to revisit the arguments
from [103] leading up to Jones’ extension theorem (recalled in Theorem 8.3.2) and
explore the prospect of incorporating weights (in the form of powers of the distance
to the boundary function) in the estimates satisfied by Jones’ extension operator
(8.3.23). This task is accomplished in the following lemma.

Lemma 8.6.7 Let Ω be an (ε, δ)-domain in Rn with rad (Ω) > 0, and recall that δ∂Ω
denotes the distance function to ∂Ω. Fix some k ∈ N and bring in Jones’ extension
operator Λk from Theorem 8.3.2. Given a function u ∈ W k,q (Ω) with q ∈ [1, ∞],
define
+ +
f := |α | ≤k ∂ α u ∈ L q (Ω, L n ), G := |α | ≤k ∂ α Λk u ∈ L q (Rn, L n ),

and set g := G Rn \Ω ∈ L q (Rn \ Ω, L n ).

(8.6.68)
Finally, pick θ ∈ (0, 1), p ∈ (0, ∞), and γ ∈ R arbitrary.
Then there exists a constant C ∈ (0, ∞) independent of u with the property that
510 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces
∫ ∫
p γ p γ
g,θ δ∂Ω dL n ≤ C f,θ δ∂Ω dL n (8.6.69)
R n \Ω Ω

where the solid maximal functions f,θ and g,θ are considered (as in (A.0.111))
relative to the open sets Ω and Rn \ Ω where the functions f , g are defined.

Proof For starters, observe that simple connectivity arguments give that
     
dist x, ∂(Rn \ Ω) = dist x, ∂(Ω) = dist x, Ω = δ∂Ω (x)
(8.6.70)
for each point x ∈ Rn \ Ω.

To proceed, pick some Whitney

 decomposition W(Ω) of Ω, along with a Whitney
decomposition W Rn \ Ω of the open set (Rn \ Ω)◦ = Rn \ Ω (with the convention
   
that W Rn \ Ω =  if Ω = Rn ), and define Ws Rn \ Ω as in (8.3.19). Next, as in
[103, p. 77], abbreviate
   
W1 := W(Ω), W2 := W Rn \ Ω , W3 := Ws Rn \ Ω . (8.6.71)

Fix a large constant Co ∈ (1, ∞) and a large integer m ∈ N (depending only geometry).
For each Q ∈ W1 then define the cluster

'
[Q]m,θ to be the union of all dilated cubes of the form (1 + Co θ) · Q
'
where Q ∈ W1 is such that there exists a family {Q j }1≤ j ≤m ⊆ W1
(8.6.72)
with Q1 = Q, Q m = Q, ' and 4−1 ≤ (Q j )/ (Q j+1 ) ≤ 4 for each
j ∈ {1, . . . , m − 1}.

In addition, define the cluster [Q]m as [Q]m,θ in (8.6.72) with θ := 0. In this notation,
given any Q j , Q k ∈ W3 with Q j ∩ Q k   it follows that each chain Fj,k connecting
Q∗j with Q∗k as in [103, Lemma 2.8, p. 78] is contained in the cluster [Q∗j ]m . With
F(Q j ) defined for each Q j ∈ W3 as in [103, p. 80] we then have (recall that Q∗
denotes the reflexion of Q across the boundary of Ω; cf. Theorem 8.3.2)

F(Q) ⊆ [Q∗ ]m for each Q ∈ W3 . (8.6.73)

A geometric argument (based on (8.3.12)-(8.3.16); cf. [103, (3.1)-(3.2), p. 80]) also

gives that 
1[Q∗ ]m, θ ≤ M for some M ∈ (0, ∞). (8.6.74)
Q ∈W3

Granted (8.6.73), from [103, Lemma 3.2, p. 80] used with p := ∞ we conclude
(keeping in mind that (Q) ≤ εδ/(16n) for each Q ∈ W3 ) that there exists some
C ∈ (0, ∞) such that

g L ∞ (Q, L n ) ≤ C f  L ∞ ([Q∗ ]m, L n ) for each Q ∈ W3 . (8.6.75)

In turn, (8.6.75) implies that for each fixed λ ∈ (1, 3) (cf. (8.3.17)-(8.3.18)) we have
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 511

g L ∞ (λQ, L n ) ≤ C · max  f  L ∞ ([Q'∗ ]m, L n ) for each Q ∈ W3 . (8.6.76)

' ∈W3, Q∩Q
Q '
'
(Q)≈(Q)

In particular, if θ ∈ (0, 1) is small enough from (8.6.76) and (A.0.111) we conclude

that, on the one hand,

g,θ (x) ≤ C · max  f  L ∞ ([Q'∗ ]m, L n ) whenever x ∈ Q ∈ W3 . (8.6.77)

' ∈W3, Q∩Q
Q '
'
(Q)≈(Q)

On the other hand, given any Q ∈ W1 we may write

⨏ 1
p
p
 f  L ∞ (Q, L n ) ≤ sup f,θ dL n
x ∈Q B(x,θ δ ∂Ω (x))

⨏ 1
p
p
≤C f,θ dL n , (8.6.78)
(1+Co θ)·Q

thanks to [133, (6.6.6)], [133, (6.6.36)] (presently

 used with u := f and s := p), and
the observation that for each x ∈ Q we have B x, θδ∂Ω (x) ⊆ (1 + Co θ) · Q. As a
consequence of (8.6.78) and (8.6.72) we therefore have
⨏ 1
p
p
 f  L ∞ ([Q]m, L n ) ≤ C f,θ dL n for each Q ∈ W1 . (8.6.79)
[Q] m, θ

By combining (8.6.77) with (8.6.79) we arrive at the conclusion that

⨏
p
g,θ (x) p ≤ C · max f,θ dL n whenever x ∈ Q ∈ W3 . (8.6.80)
Q '
' ∈W3, Q∩Q '∗ ] m, θ
[Q
'
(Q)≈(Q)

Integrating over x ∈ Q then yields

∫ ∫
p p
g,θ dL n ≤ C · max f,θ dL n for each Q ∈ W3, (8.6.81)
Q Q '
' ∈W3, Q∩Q '∗ ] m, θ
[Q
'
(Q)≈(Q)

and since δ∂Ω ≈ (Q) on each Q ∈ W3 (cf. (8.6.70) and (8.3.14)), we also obtain
from (8.6.81) that
∫ ∫
p γ p γ
g,θ δ∂Ω dL ≤ C ·
n
max f,θ δ∂Ω dL n for each Q ∈ W3 .
Q Q '
' ∈W3, Q∩Q '∗ ] m, θ
[Q
'
(Q)≈(Q)
(8.6.82)
Summing over Q ∈ W3 and keeping (8.6.74) in mind finally produces
∫ ∫ 
p γ p γ
g,θ δ∂Ω dL n ≤ C f,θ δ∂Ω dL n where D3 := Q. (8.6.83)
D3 Ω Q ∈W3
512 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Moving on, [103, Lemma 3.3, p. 81] used with p := ∞ gives that

g L ∞ (Q, L n ) ≤ C  f  L ∞ (Q'∗, L n ) for each Q ∈ W2 \ W3 . (8.6.84)
' ∈W3, Q∩Q
Q '

Much as before, this implies (assuming θ ∈ (0, 1) is small) that

⨏ 1
p
p
g,θ (x) ≤ C · max f,θ dL n whenever x ∈ Q ∈ W2 \ W3, (8.6.85)
'
Q '
(1+Co θ)·Q

where the maximum is taken over all cubes Q ' ∈ W3 with the property that there
exists Q o ∈ W2 such that (Q o ) ≈ (Q), Q o ∩ Q  , and Q' ∩ Q o  . Ultimately,
this permits us to deduce that
∫ ∫ 
p γ p γ
g,θ δ∂Ω dL n ≤ C f,θ δ∂Ω dL n where D2 := Q. (8.6.86)
D2 Ω Q ∈W2 \W3

At this stage, given that D2 ∪ D3 = Rn \ Ω, the estimate claimed in (8.6.69) follows

from (8.6.83) and (8.6.86) assuming θ ∈ (0, 1) is small enough. Finally, the restriction
that θ is small may then be lifted by invoking [133, (6.6.28)] (bearing in mind
(8.6.70)). 

We are now in a position to present our main result concerning traces of functions
belonging to the power weighted maximal Sobolev spaces introduced in Defini-
tion 8.6.1, complementing the trace result from Theorem 8.3.6.

Theorem 8.6.8 Let Ω ⊆ Rn be an (ε, δ)-domain with rad (Ω) > 0 whose boundary
is an Ahlfors regular set, and abbreviate σ := H n−1 ∂Ω. Also, fix
 
n < p ≤ 1, (n − 1) p1 − 1 < s < 1,
n−1

   (8.6.87)
and set a := 1 − s − p1 ∈ − p1 , −n p1 − 1 .

1, p
Then for each function u ∈ Wa, (Ω) the boundary trace
⨏
(TrΩ→∂Ω u)(x) := [u]Ω (x) := lim+ u(y) dy
r→0 B(x,r)∩Ω (8.6.88)
exists at σ-a.e. point x ∈ ∂Ω,

and the following Gagliardo semi-norm estimate holds

∫ ∫ p  p1
(TrΩ→∂Ω u)(x) − (TrΩ→∂Ω u)(y)
dσ(x) dσ(y) ≤ CuW 1, p (Ω)
∂Ω ∂Ω |x − y| n−1+sp a,

(8.6.89)

for some constant C ∈ (0, ∞) independent of u.

8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 513
 
Moreover, if for each z ∈ ∂Ω and r ∈ 0, diam Ω some interior corkscrew point
Ar (z) for z at scale r has been selected, and if C >
 0 is a sufficiently
 large constant
so that B(Ar (z), r/C) ⊆ Ω for all z ∈ ∂Ω and r ∈ 0, diam Ω , then for each function
1, p
u ∈ Wa, (Ω) one has
⨏
(TrΩ→∂Ω u)(z) = lim+ u(y) dy for σ-a.e. z ∈ ∂Ω. (8.6.90)
r→0 B(Ar (z),r/C)

In addition, there exists some aperture parameter κ > 0 with the property that for
1, p
each function u ∈ Wa, (Ω) which has a κ-nontangential pointwise limit at σ-a.e.
point on ∂Ω it follows
κ−n.t.
TrΩ→∂Ω u = u ∂Ω at σ-a.e. point on ∂Ω. (8.6.91)

A few comments are in order here. First, thanks to (8.6.5), Theorem 8.3.6, and
(7.9.10)-(7.9.11), the above trace result is in fact also true in the range 1 < p < ∞.
Second, the Gagliardo semi-norm estimate (8.6.89) actually becomes a genuine
p, p
membership of the trace to the boundary Besov space Bs (∂Ω, σ) at least if Ω is
1, p
a Lipschitz domain; cf. (7.9.22). Hence, in such a setting, TrΩ→∂Ω maps Wa, (Ω)
p, p
boundedly into Bs (∂Ω, σ) for the range of indices specified in (8.6.87).
Third, when p, s are as in the first line of (8.6.87), i.e., n−1 < p ≤ 1 and
  n
(n − 1) p1 − 1 < s < 1, the disagreement parameter a := 1 − s − p1 is necessarily
negative (which
 is in  contrast to the case when p was allowed to be in (1, ∞)). In
 sharp
fact, a ∈ − p1 , −n p1 − 1 and this interval contains only strictly negative numbers.
Thus, in order to have a trace result when p is sub-unital we are naturally led to
considering genuinely weighted (maximal) Sobolev space.
We now turn to the task of providing the proof of Theorem 8.6.8.
1, p
Proof of Theorem 8.6.8 Given u ∈ Wa, (Ω), from Proposition 8.6.3 (used with
k := 1) we conclude that
np
u ∈ W 1,q (Ω) with q := n−1+p−sp ∈ (1, n)
(8.6.92)
and uW 1, q (Ω) ≤ CuW 1, p (Ω)
a,

for some C ∈ (0, ∞) independent of u. Bring in Jones’ extension operator Λ1 defined

as in Theorem 8.3.2 with k := 1. Then

Λ1 : W 1, p (Ω) −→ W 1, p (Rn ) linearly and boundedly, (8.6.93)

hence there exists a constant C ∈ (0, ∞), independent of u, such that

w := Λ1 u ∈ W 1,q (Rn ), w W 1, q (Rn ) ≤ CuW 1, q (Ω),

(8.6.94)
and w Ω = u at L n -a.e. point in Ω.
514 8 Boundary Traces from Weighted Sobolev Spaces into Besov Spaces

Moreover, from (8.3.40) in Theorem 8.3.6 (applied here with p := q and a := 0) we

see that the boundary trace TrΩ→∂Ω u, originally defined as in (8.6.88), satisfies
⨏
(TrΩ→∂Ω u)(x) = lim+ w(y) dy at σ-a.e. point x ∈ ∂Ω. (8.6.95)
r→0 B(x,r)

Going further, since Cc∞ (Rn ) embeds densely into the Sobolev space W 1,q (Rn ),
we may find a sequence {w j } j ∈N ⊆ Cc∞ (Rn ) such that w j → w in W 1,q (Rn ) as
j → ∞. By eventually passing to a sub-sequence, there is no loss of generality in
assuming that we also have w j → w at L n -a.e. point in Rn as j → ∞. Let En be
the standard radial fundamental solution for the Laplacian in Rn . Fix some arbitrary
point x ∈ Rn and denote by δx the Dirac distribution in Rn with mass at x. Then for
each j ∈ N we may write
- .
w j (x) = δx, w j  = Δ[En (x − ·)], w j
- . - .
= − ∇[En (x − ·)], ∇w j = (∇En )(x − ·), ∇w j
∫
= (∇En )(x − ·), ∇w j  dL n . (8.6.96)
Rn
* "n
Given that ∇w j → ∇w in L q (Rn, L n ) as j → ∞, the Fractional Integration
  −1
Theorem (cf. [133, (7.8.7)]) guarantees that if q∗ := q1 − n1 then
∫ ∫
(∇En )(· − y), (∇w j )(y) dy −→ (∇En )(· − y), (∇w)(y) dy
Rn Rn (8.6.97)
∗
in L q (Rn, L n ) as j → ∞.

In particular, by passing to a sub-sequence we may also ensure that the above

convergence takes place at L n -a.e. point in Rn . From this and (8.6.96) we then
conclude that
∫
w(x) = (∇En )(x − y), (∇w)(y) dy at L n -a.e. point x ∈ Rn . (8.6.98)
Rn

To proceed, consider the open set D := Rn \ ∂Ω. Since L n (∂Ω) = 0 given that
∂Ω is upper Ahlfors regular, it follows that ∂Ω has an empty interior. In turn, this
implies
∂D = ∂Ω hence also δ∂D = δ∂Ω in Rn . (8.6.99)
 n 
Keeping this in mind and observing that D = Ω ∪ R \ Ω (disjoint union), Lem-
ma 8.6.7 implies that
* p p(1−s)−1 n " n
∇w ∈ L D, δ∂D L and ∇w [L p (D,δ p(1−s)−1 L n )]n ≤ CuW 1, p (Ω)
 ∂D a,
(8.6.100)
8.6 Traces from Weighted Maximal Sobolev Spaces in Open Subsets of R n 515

for some C ∈ (0, ∞) independent of u. Since Rn \ D = ∂Ω and L n (∂Ω) = 0, we may

refashion (8.6.98) as
∫
w(x) = (∇En )(x − z), (∇w)(z) dz at L n -a.e. point x ∈ Rn . (8.6.101)
D

Upon noting that the function defined as Φ(x, z) := (∇En )(x − z) for each x, z ∈ Rn
with x  z satisfies (8.6.40)-(8.6.41), we may now invoke Proposition 8.6.6 (with
Ω := D and ϕ := ∇w which, thanks to (8.6.100), satisfies (8.6.43)) and conclude,
based on the integral representation formula in (8.6.101), that (cf. (8.6.47))
⨏
lim+ w(y) dy = w(x) at H n−1 -a.e. point x ∈ ∂D = ∂Ω, (8.6.102)
r→0 B(x,r)

and (cf. (8.6.46))

∫ ∫  p1
|w(x) − w(y)| p
dH n−1 (x) dH n−1 (y) ≤ C∇w [L p (D,δ p(1−s)−1 L n )]n .
∂D ∂D |x − y| n−1+sp  ∂D

(8.6.103)

At this stage, the Gagliardo semi-norm estimate claimed in (8.6.89) follows by

combining (8.6.103), (8.6.102), (8.6.95), (8.6.100), and (8.6.99).
Finally, the claims made in (8.6.90) and (8.6.91) are consequences of Propo-
sition 8.3.8, Corollary 8.3.9 (both used with a := 0), and Proposition 8.6.3 (cf.
(8.6.92)). 
Chapter 9
Besov and Triebel-Lizorkin Spaces in Open Sets

We begin by reviewing Besov and Triebel-Lizorkin spaces in the entire Euclidean

setting in §9.1. Besov and Triebel-Lizorkin in arbitrary open subsets of Rn are
then defined in §9.2 via restriction from the corresponding scales in the Euclidean
setting (in the sense of distributions). Certain quasi-Banach envelopes of Besov and
Triebel-Lizorkin spaces are concretely identified in §9.3.
The topic of traces of functions belonging to Besov and Triebel-Lizorkin spaces
is treated in §9.4. In this regard, we first consider said spaces in the entire Euclidean
setting and take traces on arbitrary Ahlfors regular sets in Rn , then we use these
results to establish mapping properties for trace operators from spaces defined in
open subsets of Rn onto theirs boundaries. This body of results is further augmented
in §9.5 by proving mapping properties for conormal derivative operators acting on
functions belonging to Besov and Triebel-Lizorkin spaces in certain open subsets of
Rn . Finally, in §9.6 we construct extension operators from boundary Besov spaces
into Besov and Triebel-Lizorkin spaces, which turn out to be inverses from the right
for the the trace operators acting from Besov and Triebel-Lizorkin spaces defined in
the open set in question considered earlier, in §9.4.

9.1 Besov and Triebel-Lizorkin Spaces in R n

Informative accounts pertaining to the topic of Besov and Triebel-Lizorkin spaces

in the Euclidean setting may be found in many references, including [189], [188],
[187], [59], [60], [15], [166]. One convenient point of view is offered by the classical
Littlewood-Paley theory (cf., e.g., [166], [187], [188]). More specifically, let Ξ be the
collection of all families {ζ j }∞
j=0 of Schwartz functions with the following properties:
(i) There exist finite positive constants A, B, C such that

supp (ζ0 ) ⊂ {x : |x| ≤ A} and
(9.1.1)
supp (ζ j ) ⊂ {x : B2 j−1 ≤ |x| ≤ C2 j+1 } if j ∈ N.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 517
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_9
518 9 Besov and Triebel-Lizorkin Spaces in Open Sets

(ii) For every multi-index α ∈ N0n there exists a positive finite constant Cα such
that
sup sup 2 j |α | |∂ α ζ j (x)| ≤ Cα . (9.1.2)
x ∈R n j ∈N

(iii) One has


∞
ζ j (x) = 1 for every x ∈ Rn . (9.1.3)
j=0

Fix some family {ζ j }∞  n

j=0 ∈ Ξ. Also, let F and S (R ) denote, respectively, the
Fourier transform and the class of tempered distributions in Rn . Then the Triebel-
p,q
Lizorkin space Fs (Rn ) is defined for 0 < p < ∞, 0 < q ≤ ∞, and s ∈ R as
 
Fs (Rn ) := f ∈ S (Rn ) :  f Fsp, q (Rn ) < +∞
p,q
(9.1.4)
where for each f ∈ S (Rn ) we have set
 
∞  1/q 
 
 f Fsp, q (Rn ) :=  |2s j F −1 (ζ j F f )| q  (9.1.5)
L p (R n, L n )
j=0

with the inner  q quasi-norm replaced by sup over j ∈ N0 when q = ∞. The case
p = ∞ is somewhat special, in that a suitable version of (9.1.5) needs to be used
(see, e.g., [166, p. 9]). First, for each k ∈ Z and l = (l1, . . . , ln ) ∈ Zn , we introduce
the notation
 
Q k,l := x = (x1, . . . , xn ) ∈ Rn : 2−k li ≤ xi ≤ 2−k (li + 1), i ∈ {1, . . . , n} . (9.1.6)

Then, corresponding to p = ∞, 0 < q ≤ ∞, and s ∈ R, for each f ∈ S (Rn ) we set

 ∫ 
∞  1/q
f ∞, q
Fs (R n ) := sup sup 2 jn
|2sk F −1 (ζ j F f )(x)| q dx . (9.1.7)
j ∈N0 l ∈Z n Q j, l k=j

p,q
The scale of Besov space Bs (Rn ), indexed by 0 < p, q ≤ ∞ and s ∈ R, is
defined as  
Bs (Rn ) := f ∈ S (Rn ) :  f Bsp, q (Rn ) < +∞
p,q
(9.1.8)
where, for each f ∈ S (Rn ),

∞  1/q
2s j F −1 (ζ j F f ) L p (Rn, L n )
q
 f Bsp, q (Rn ) := , (9.1.9)
j=0

with the outer  q quasi-norm replaced by sup over j ∈ N0 when q = ∞. Different

choices of the family {ζ j }∞
j=0 in the class Ξ yield the same spaces (9.1.4)-(9.1.8),
albeit equipped with equivalent quasi-norms. As is well known (see, e.g., [187,
p,q
§ 2.3.3]), Bs (Rn ) is a quasi-Banach space for s ∈ R, 0 < p, q ≤ ∞ (Banach space
if 1 ≤ p, q ≤ ∞) and
9.1 Besov and Triebel-Lizorkin Spaces in R n 519

S (Rn ) → Bs (Rn ) → S (Rn ).

p,q
(9.1.10)
p,q
in the sense of continuous (topological) embeddings. Similarly, Fs (Rn ) is a quasi-
Banach space for s ∈ R, 0 < p < ∞, 0 < q ≤ ∞ (Banach space if 1 ≤ p < ∞,
1 ≤ q ≤ ∞) and
S (Rn ) → Fs (Rn ) → S (Rn ).
p,q
(9.1.11)
Moreover, given s ∈ R,
p,q
S (Rn ) → Fs (Rn ) densely, if and only if q < ∞,
p,q
(9.1.12)
S (Rn ) → Bs (Rn ) densely, if and only if max {p, q} < ∞.

Furthermore, (cf., e.g., [166, p. 30] and [187, p. 47])

p,min{p,q } p,q p,max{p,q }
Bs (Rn ) → Fs (Rn ) → Bs (Rn )
(9.1.13)
for 0 < p, q ≤ ∞ and s ∈ R

For indices p, q, s such that n/(n + 1) < p, q ≤ ∞ and n(1/p − 1)+ < s < 1 one
has
∫ q
 f (· + t) − f (·) L p (Rn, L n )
1/q

 f Bsp, q (Rn ) ≈  f  L p (Rn, L n ) + dt . (9.1.14)

Rn |t| n+sq

This result is proved, for instance, in [187, § 2.5.12], where one can also find an anal-
ogous characterization of Triebel-Lizorkin spaces as well as intrinsic descriptions
of both of Besov and Triebel-Lizorkin spaces with larger amounts of smoothness,
involving higher order differences. Note that by specializing (9.1.14) to the case of
the diagonal Besov scale we obtain

if n
n+1 < p, q ≤ ∞ and n 1
p −1 + < s < 1 then
∫ ∫  1/p (9.1.15)
| f (x) − f (y)| p
f p, p
B s (R n ) ≈  f  L p (Rn, L n ) + dx dy .
Rn Rn |x − y| n+sp

In the next several theorems we collect other basic properties of Besov and Triebel-
Lizorkin spaces, starting with a theorem describing the monotonicity of the spaces
p,q
As (Rn ), with A := B or A := F, with respect to s and q (cf. [187, Proposition 2,
p. 47], [166, Proposition, p. 29]).

Theorem 9.1.1 Let s ∈ R, ε > 0, 0 < q0 ≤ q1 ≤ ∞, and 0 < p ≤ ∞. Then

p,q p,q0 p,q1
As+ε1 (Rn ) → As (Rn ) → As (Rn ), A ∈ {B, F}, (9.1.16)

with the convention that p < ∞ if A = F.

The next theorem describes lifting results for Besov and Triebel-Lizorkin spaces.
520 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Theorem 9.1.2 Assume that 0 < p, q ≤ ∞, s ∈ R, and A ∈ {B, F}. Then

As (Rn ) = (I − Δ)μ As+μ (Rn ) for each μ ∈ R.

p,q p,q
(9.1.17)

Also, for each m ∈ N one has

 
As (Rn ) = f ∈ S (Rn ) : ∂ α f ∈ As−m (Rn ) for all α ∈ N0n with |α| ≤ m
p,q p,q

 
= f ∈ As−m (Rn ) : ∂ α f ∈ As−m (Rn ) for all α ∈ N0n with |α| = m
p,q p,q

(9.1.18)

and

 f  Ap, q n ≈
s (R )
∂ α f  Ap, q n
s−m (R )
|α | ≤m

≈  f  Ap, q n +
s−m (R )
∂ α f  Ap, q n .
s−m (R )
(9.1.19)
|α |=m

In particular,
∂ α : As (Rn ) −→ As− |α | (Rn )
p,q p,q
(9.1.20)
is well defined, linear, and bounded for each multi-index α ∈ N0n .
We now collect the well-known embedding properties for Besov and Triebel-
Lizorkin spaces on Rn in the following theorem.
Theorem 9.1.3 (i) Let 0 < p, q, a, b ≤ ∞ and s ∈ R. Then the embeddings
p,a p,q p,b
Bs (Rn ) → Fs (Rn ) → Bs (Rn ) (9.1.21)

are continuous if and only if

0 < a ≤ min{p, q} and max{p, q} ≤ b ≤ ∞. (9.1.22)

(ii) Assume 0 < p0 < p < p1 ≤ ∞ and s0, s1, s ∈ R satisfy

s0 s1
1
p0 − n = 1
p − s
n = 1
p1 − n. (9.1.23)

Then, whenever 0 < a, b ≤ ∞,

p ,a p,q p ,b
Bs00 (Rn ) → Fs (Rn ) → Bs11 (Rn ) (9.1.24)

are well-defined continuous embeddings if and only if

0 < a ≤ p ≤ b ≤ ∞. (9.1.25)
s1
(iii) Let 0 < p < p1 ≤ ∞, s, s1 ∈ R, 0 < r, q ≤ ∞. Assume that 1
p − s
n = 1
p1 − n.
Then
p,r p ,q
Fs (Rn ) → Bs11 (Rn ). (9.1.26)
9.1 Besov and Triebel-Lizorkin Spaces in R n 521

Proof Part (i) of the theorem can be found in [174, Theorem 3.1.1]. See also [166,
Theorem, p. 30]. Parts (ii) and (iii) were first proved in [101, Theorem 2.1, pp. 95-96]
for homogeneous spaces, then in [174, Theorem 3.2.1] for inhomogeneous spaces.
See also [166, Theorem, p. 31]. 

We continue discussing embedding results, of the following tight nature.

Theorem 9.1.4 Given 0 < p0 ≤ p1 ≤ ∞, s0, s1 ∈ R, and 0 < q0 ≤ q1 ≤ ∞ with the

property that p10 − sn0 = p11 − sn1 , one has the continuous embedding
p ,q0 p ,q1
Bs00 (Rn ) → Bs11 (Rn ) (9.1.27)

Moreover, one has the continuous embedding

p ,q0 p ,q1
Fs00 (Rn ) → Fs11 (Rn ), (9.1.28)

provided either
1 s0 1 s1
0 < p0 < p1 < ∞, 0 < q0, q1 ≤ ∞, and − = − , (9.1.29)
p0 n p1 n
or
0 < p0 = p1 < ∞, 0 < q0 ≤ q1 ≤ ∞, and s0 = s1 . (9.1.30)

Proof The embedding in (9.1.27) follows from item (ii) in Theorem 9.1.3 and The-
orem 9.1.1, while the embedding claimed in (9.1.28) may be justified by combining
Theorem 9.1.1 with [166, Theorem 2.2.3(5), p. 31] (cf. also [166, Remark 2, p. 31]).

Turning to the issue of duality, we have the following result.

Theorem 9.1.5 For s ∈ R and 0 < p, q < ∞, one has

p,q ∗ p,q
Bs (Rn ) = B−s+n(1/p−1)+ (Rn ) (9.1.31)
p,q ∗ p,q
Fs (Rn ) = F−s (Rn ) if p > 1, (9.1.32)
p,q ∗ ∞,∞
Fs (Rn ) = B−s+n(1/p−1) (Rn ) if p < 1, (9.1.33)

where p and q  are, respectively, the Hölder conjugate exponents of p and q,

considered in the sense of (7.6.1).

Proofs and other related results may be found in, e.g., [187], [166] (cf. also the
references therein). Here we only want to point out that (9.1.32) with 0 < q < 1 is
not usually covered in the literature. An argument may be found in [127].
The reader is reminded that the version of the complex interpolation method
employed in the theorem below has been reviewed in §1.4.

Theorem 9.1.6 Fix n ∈ N and assume

522 9 Besov and Triebel-Lizorkin Spaces in Open Sets

θ ∈ (0, 1), α0, α1 ∈ R, α := (1 − θ)α0 + θα1,

0 < p j ≤ ∞, 0 < q j ≤ ∞, for j ∈ {0, 1}, (9.1.34)
θ −1 θ −1
p := 1−θ
p0 + p1 , q := 1−θ
q0 + q1 .

Then, for A ∈ {B, F},

p ,q0 p ,q1 p ,q0 p ,q1
Aα00 (Rn ), Aα11 (Rn ), and Aα00 (Rn ) + Aα11 (Rn )
(9.1.35)
are all analytically convex quasi-Banach spaces.

In addition,
 p ,q0 p ,q1  p,q
Fα00 (Rn ), Fα11 (Rn ) θ
= Fα (Rn ),
(9.1.36)
if either max{p0, q0 } < ∞, or max{p1, q1 } < ∞,

and  p ,q0 p ,q1  p,q

Bα00 (Rn ), Bα11 (Rn ) θ
= Bα (Rn ),
(9.1.37)
provided min {q0, q1 } < ∞.
Finally, similar results are valid for the homogeneous Besov and Triebel-Lizorkin
spaces in Rn , and also for the discrete versions of these scales of spaces.

Proof See [110, Theorem 9.1] and its proof. 

Regarding the real method of interpolation, we have the following classical result
(see, e.g., [187]).

Theorem 9.1.7 Assume that 0 < q0, q1, q ≤ ∞, 0 < θ < 1, α0, α1 ∈ R with α0  α1 ,
and define α := (1 − θ)α0 + θα1 . Then
p,q p,q p,q
Fα0 0 (Rn ), Fα1 1 (Rn ) θ,q = Bα (Rn ), 0 < p < ∞, (9.1.38)
p,q p,q p,q
Bα0 0 (Rn ), Bα1 1 (Rn ) θ,q = Bα (Rn ), 0 < p ≤ ∞. (9.1.39)
p,q p,q
A powerful tool in the study of the spaces Bs (Rn ) and Fs (Rn ) is the atomic
decomposition of these spaces. Here we review some of the main results from [57]
and [59] which, in turn, build on the work of many other people.
Turning to specifics, denote by J (Rn ) the collection of dyadic cubes in Rn and
fix some s ∈ R along with p ∈ (0, ∞]. Also, consider a parameter J ∈ R and pick
two integers K, L satisfying
 
K ≥ ([s] + 1)+ and L ≥ max [J − n − s], −1 . (9.1.40)

Given an arbitrary cube Q ∈ J (Rn ) with side-length (Q) ≤ 1, call a function

aQ ∈ W K,∞ (Rn ) a (s, p)-atom of type (K, L, J) if the following properties hold:
9.1 Besov and Triebel-Lizorkin Spaces in R n 523

(1) supp (aQ ) ⊆ 3Q, (9.1.41)

(2) ∂γ aQ  L ∞ (Rn, L n ) ≤ L n (Q)s/n−1/p− |γ |/n whenever |γ| ≤ K, (9.1.42)

∫
(3) xγ aQ (x) dx = 0 whenever |γ| ≤ L and (Q) < 1, (9.1.43)
Rn

with the convention that property (3) is omitted if L < 0. For a proof of the following
theorem see [57], [59].

Theorem 9.1.8 If s ∈ R, 0 < p, q ≤ ∞, J := min{1, n

p } , and K, L are two integers
as in (9.1.40), then there exists some finite constant C = C(s, p, n, K, L) > 0 with
the property that for any numerical sequence {λQ }Q ∈ J(Rn ),l(Q)≤1 and any family
{aQ }Q ∈ J(Rn ), (Q)≤1 of (s, p)-atoms of type (K, L, J) there holds
1/q
 
    ∞    q/p 
   
 λQ aQ  ≤ C |λQ | p  . (9.1.44)
  p, q  
Q ∈ J(R n ) Bs (R n ) j=0 Q ∈ J(R n )
(Q)≤1  (Q)=2− j 

Conversely, if s ∈ R, 0 < p, q ≤ ∞, and J := min{1, n

p } , then having fixed
two integers K, L satisfying (9.1.40), it follows that for each tempered distribution
p,q
f ∈ Bs (Rn ) one may find some family {aQ }Q ∈ J(Rn ), (Q)≤1 of (s, p)-atoms of type
(K, L, J), along with some sequence of complex numbers {λQ }Q ∈ J(Rn ), (Q)≤1 with
the property that
1/q
∞   q/p
  
 f Bsp, q (Rn ) ≈ |λQ | p  , (9.1.45)
 j=0 Q ∈ J(Rn ): (Q)=2− j 
such that f may be represented in the form

f = λQ aQ with convergence in S (Rn ). (9.1.46)
Q ∈ J(R n )
(Q)≤1

In particular, (9.1.44) implies that


p,q
λQ aQ convergence to f in Bs (Rn ) if max{p, q} < ∞. (9.1.47)
Q ∈ J(R n )
(Q)≤1

Results similar in spirit to those presented above are also valid for Triebel-
Lizorkin spaces. Concretely, if s ∈ R, 0 < p < ∞, 0 < q ≤ ∞ J := min{1,n p,q } ,
and K, L are two integers satisfying (9.1.40), then there exists some finite constant
C = C(s, p, n, K, L) > 0 with the property that
524 9 Besov and Triebel-Lizorkin Spaces in Open Sets
 1/q 
   
       q  
   
 λQ aQ  ≤ C  L n (Q)−1/p |λQ |1Q  
 n
 p, q n  n
 
Q ∈ J(R ) Fs (R ) ∈ J(R )
 Q(Q)≤1 
(Q)≤1   L p (R n, L n )
(9.1.48)
for any numerical sequence {λQ }Q ∈ J(Rn ), (Q)≤1 and family {aQ }Q ∈ J(Rn ), (Q)≤1 of
(s, p)-atoms of type (K, L, J).
In the converse direction, if s ∈ R, 0 < p < ∞, 0 < q ≤ ∞, J := min{1,n p,q } , and
K, L are two integers satisfying (9.1.40), it follows that each tempered distribution
p,q
f ∈ Fs (Rn ) may be written as

f = λQ aQ with convergence in S (Rn ), (9.1.49)
Q ∈ J(R n )
(Q)≤1

for some family {aQ }Q ∈ J(Rn ),l(Q)≤1 of (s, p)-atoms of type (K, L, J) and some nu-
merical sequence {λQ }Q ∈ J(Rn ), (Q)≤1 with the property that
 1/q 
 
   q  
  
 f Fsp, q (Rn ) ≈  L n (Q)−1/p |λQ |1Q   . (9.1.50)
Q∈J(Rn )  
 (Q)≤1  p n n
  L (R , L )

In particular, (9.1.48) implies that


p,q
λQ aQ convergence to f in Fs (Rn ) if q < ∞. (9.1.51)
Q ∈ J(R n )
(Q)≤1

It has long been known that many classical smoothness spaces are encompassed
by the Besov and Triebel-Lizorkin scales. For example,

Bs∞,∞ (Rn ) = C s (Rn ) if 0 < s  N, (9.1.52)

p,2
F0 (Rn ) = L p (Rn, L n ) if 1 < p < ∞, (9.1.53)
p,2 p
Fs (Rn ) = Ls (Rn ) if 1 < p < ∞ and s ∈ R, (9.1.54)
p,2
Fk (Rn ) = W k, p (Rn ) if 1 < p < ∞ and k ∈ Z, (9.1.55)
p,2
F0 (Rn ) = h p (Rn ) if 0 < p ≤ 1, (9.1.56)
∗
F0∞,2 (Rn ) = bmo(Rn ) = h1 (Rn ) . (9.1.57)

Let us comment on the nature of the spaces appearing above and, in the process,
discuss other identifications. First, there is a suitable notion of homogeneous Triebel-
Lizorkin space (see, e.g., [57]), and the standard John-Nirenberg of functions of
bounded mean oscillations fits on this scale, namely
9.1 Besov and Triebel-Lizorkin Spaces in R n 525
.
F0∞,2 (Rn ) = BMO(Rn ). (9.1.58)

Second, recall the classical Bessel potential spaces

Ls (Rn, L n ) := (I − Δ)−s/2 L p (Rn, L n )

p
(9.1.59)
for each s ∈ R and p ∈ (1, ∞)

(compare with the weighted version in (8.1.8)). Often, these are referred to as L p -
based fractional Sobolev spaces of order s. The case p = 2 is particularly ubiquitous.
In this scenario, the L 2 -based Sobolev space of fractional order s ∈ R in Rn are
sometimes simply denoted by H s (Rn ). These may be defined directly as
 
 ∈ L 2 (Rn, L n )
H s (Rn ) := U ∈ S (Rn ) : (1 + |ξ | 2 )s/2U (9.1.60)

(where “hat” denotes the Fourier transform in Rn ), with norm

 
U  H s (Rn ) := (1 + |ξ | 2 )s/2U
 2 n n .
L (R , L )
(9.1.61)

Hence,
Fs2,2 (Rn ) = H s (Rn ) for each s ∈ R. (9.1.62)
p,q
Third, we briefly discuss the scale Cα (Rn ) introduced in [172]. Concretely, for
each given m ∈ N0 set

Pm := polynomials in Rn of degree ≤ m. (9.1.63)

Also, given an integer m ∈ N0 , a point x ∈ Rn , a scale t > 0, and a Lebesgue-

measurable function u in B(x, t) ⊂ Rn , define
⨏
oscm (u, x, t) := inf |u − P| dL n . (9.1.64)
P ∈ Pm B(x,t)

Whenever 0 < p < ∞, 0 < q ≤ ∞, α > 0, and m ∈ N with m > α, set

p,q  
Cα (Rn ) := u ∈ Lloc
1
(Rn, L n ) : uCαp, q (Rn ) < +∞ , (9.1.65)

where the quasi-norm  · Cαp, q (Rn ) is defined as

 ∫ ∞
dt  1/q 

u p, q
Cα (R n ) := u L p (R n, L n ) + [t −α oscm−1 (u, ·, t)]q
 p n n
0 t L (R , L )
(9.1.66)
for each Lebesgue measurable function u in Rn . Whenever necessary to indicate the
p,q,m n p,q
dependence on the integer m we shall write Cα (R ) in place of Cα (Rn ). Then,
with this piece of notation, [172, Theorem 1, p. 393] implies that
p,q p,q,m
Fα (Rn ) = Cα (Rn ) whenever 0 < p < ∞, 0 < q ≤ ∞,
(9.1.67)
and m ∈ N with m > α > n 1
min {p,q } − 1 +,
526 9 Besov and Triebel-Lizorkin Spaces in Open Sets

with equivalence of quasi-norms. In particular, for the above range of indices, the
p,q,m n
space Cα (R ) is actually independent of m.
∫ Hardy spaces in R . Fix a test function
Fourth, recall the classical scale of n

ψ ∈ Cc∞ (Rn ) with the property that Rn ψ dL n = 1, and set ψt := t −n ψ(·/t) for
each t > 0. Given a tempered distribution u ∈ S (Rn ) we define its radial maximal
function and its truncated version, respectively, by setting

u++ := sup |ψt ∗ u|, u+ := sup |ψt ∗ u|. (9.1.68)

0<t<∞ 0<t<1

For p ∈ (0, ∞), the classical Hardy space H p (Rn ), and its local version, h p (Rn )
introduced in [67], are then defined as
 
H p (Rn ) := u ∈ S (Rn ) : u H p (Rn ) := u++  L p (Rn ) < ∞ , (9.1.69)
 
h p (Rn ) := u ∈ S (Rn ) : uh p (Rn ) := u+  L p (Rn ) < ∞ . (9.1.70)

Different choices of the function ψ yield equivalent quasi-norms so (9.1.69), (9.1.70)

viewed as topological spaces, are intrinsically defined. Also,

H p (Rn ) = h p (Rn ) = L p (Rn, L n ) if 1 < p < ∞. (9.1.71)

p
We then proceed to define the local Hardy-based Sobolev spaces hk (Rn ), with
0 < p < ∞ and k ∈ Z. When k ≥ 0 we set
 
hk (Rn ) := u ∈ S (Rn ) : ∂γ u ∈ h p (Rn ) for all γ ∈ N0n with |γ| ≤ k , (9.1.72)
p


and equip this space with the quasi-norm uh p (Rn ) := |γ | ≤k ∂γ uh p (Rn ) . When
k
k < 0 we set
  
hk (Rn ) := u ∈ S (Rn ) : u = ∂γ uγ where each uγ ∈ h p (Rn )
p
(9.1.73)
|γ | ≤−k

and equip this space with the quasi-norm uh p (Rn ) := inf |γ | ≤−k uγ h p (Rn ) ,
k
where the infimum is taken over all representations of the tempered distribution
p
u ∈ hk (Rn ) as in (9.1.73).
In relation to these spaces, we have the following natural identification result.

Proposition 9.1.9 For p ∈ (0, ∞) and k ∈ Z one has

 p n
p,2 n
hk (R ) if 0 < p ≤ 1,
Fk (R ) = (9.1.74)
W k, p (Rn ) if 1 < p < ∞.

Proof The case when 1 < p < ∞ has already been noted in (9.1.55) so we restrict
attention to the situation when 0 < p ≤ 1. Assume first that k ≥ 0. Then each
u ∈ hk (Rn ) has ∂γ u ∈ h p (Rn ) = F0 (Rn ) for every γ ∈ N0n with |γ| ≤ k. Lifting
p p,2

p,2
results for Triebel-Lizorkin spaces (cf. Theorem 9.1.2) then imply u ∈ Fk (Rn ). This
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 527
p p,2
establishes the continuous embedding hk (Rn ) → Fk (Rn ). In a similar fashion, if
u ∈ Fk (Rn ) then ∂γ u ∈ F0 (Rn ) = h p (Rn ) for every γ ∈ N0n with |γ| ≤ k. This
p,2 p,2

p p,2 p
goes to show that u ∈ hk (Rn ), hence Fk (Rn ) → hk (Rn ). At this point, the equality
in (9.1.74) is established in the case when k ≥ 0.
p
Consider now the case when k < 0. Then any u ∈ hk (Rn ) may be expressed as
 γ
u = |γ | ≤−k ∂ uγ with each uγ ∈ h (R ) = F (R ). The latter membership places
p n p n

each ∂γ uγ into Fk (Rn ) which ultimately shows that u ∈ Fk (Rn ). Hence, on the one
p p

p n p,2 n p
hand hk (R ) → Fk (R ). On the other hand, given any u ∈ Fk (Rn ) with k < 0 it
follows that for every γ ∈ N0n with |γ| ≤ −k we have
 
uγ := ∂γ (I − Δ)k u ∈ ∂γ (I − Δ)k Fk (Rn )
p,2

= ∂γ F−k (Rn ) → F−k− |γ | (Rn ) → F0 (Rn ) = h p (Rn ),

p,2 p,2 p,2
(9.1.75)

thanks to lifting, differentiation, and embedding results for Triebel-Lizorkin spaces

(recorded in Theorem 9.1.2 and Theorem 9.1.1). Since u belongs to the linear span
 
of ∂γ uγ : γ ∈ N0n, |γ| ≤ −k , we conclude that Fk (Rn ) → hk (Rn ). This finishes
p,2 p

the proof of (9.1.74). 

9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n

We begin by succinctly reviewing smoothness scales of spaces of Besov and Triebel-

Lizorkin type on arbitrary open subsets of the Euclidean  ambient. Given an arbitrary,
open, nonempty subset Ω of Rn , denote by U Ω the restriction of a distribution
U ∈ D (Rn ) to Ω. For 0 < p, q ≤ ∞ and s ∈ R we then define
  
As (Ω) := u ∈ D (Ω) : there exists U ∈ As (Rn ) such that U Ω = u ,
p,q p,q

  
u Ap, q := inf U  p, q
p,q n
n ) : U ∈ As (R ), U  = u , ∀u ∈ Asp,q (Ω),
s (Ω) A s (R Ω
(9.2.1)
where A ∈ {B, F} (with A = B corresponding to Besov spaces, and with A = F
corresponding to Triebel-Lizorkin spaces). In the same setting, for A ∈ {B, F} let us
also introduce
 
Ås (Ω) := the closure of Cc∞ (Ω) in As (Ω),  ·  Ap,
p,q p,q
s
q
(Ω) . (9.2.2)

p,q p,q
Hence, Ås (Ω) is a closed subspace of As (Ω) and we equip it with the quasi-norm
inherited from the latter space. We continue to assume that Ω is an arbitrary open
subset of Rn . First, for 0 < p, q ≤ ∞, s ∈ R, we set
p,q  p,q 
As,0 (Ω) := u ∈ As (Rn ) : supp u ⊆ Ω ,
p,q (9.2.3)
u Ap, q (Ω) := u Ap, q n ,
s (R )
∀u ∈ As,0 (Ω),
s,0
528 9 Besov and Triebel-Lizorkin Spaces in Open Sets
p,q
where either A = F and p < ∞, or A = B. Thus, As,0 (Ω) is a closed subspaces of
p,q
As,0 (Rn ). Second, for 0 < p, q ≤ ∞ and s ∈ R, we introduce
  
As,z (Ω) := u ∈ D (Ω) : there exists w ∈ As,0 (Ω) such that w Ω = u ,
p,q p,q

  
u Ap, q := inf w  p, q n) : w ∈ A
p,q
(Ω), w  = u , ∀u ∈ As,z
p,q
(Ω),
s, z (Ω) A s (R s,0 Ω
(9.2.4)
where, as before, either A = F and p < ∞, or A = B. It is then immediate from these
definitions that the restriction (in the sense of distributions)

R Rn →Ω : D (Rn ) −→ D (Ω),
 (9.2.5)
R Rn →Ω u := u for each u ∈ D (Rn ),
Ω

induces an operator
p,q p,q
R Rn →Ω : As (Rn ) −→ As (Ω), 0 < p, q ≤ ∞, s ∈ R, A ∈ {B, F}
(9.2.6)
which is well defined, linear, bounded, and surjective.

Moreover, for A ∈ {B, F},

  
Cc∞ (Ω) := φΩ : φ ∈ Cc∞ (Rn ) is dense in As (Ω)
p,q
(9.2.7)
whenever p, q ∈ (0, ∞) and s ∈ R.

Also, in the same setting, from definitions and the last property in Theorem 9.1.2 we
see that for every multi-index α ∈ N0n the operator

∂ α : As (Ω) −→ As− |α | (Ω)

p,q p,q
(9.2.8)
is well defined, linear, and bounded.

Definition (9.2.1) also implies that embeddings valid in the entire Euclidean space
continue to hold in arbitrary nonempty subset Ω of Rn . For example, Theorem 9.1.1
implies
p,q p,q
As0 0 (Ω) → As1 1 (Ω) continuously,
(9.2.9)
whenever 0 < p, q0, q1 ≤ ∞ if s0 > s1,
and p,q0 p,q1
As (Ω) → As (Ω) continuously,
(9.2.10)
whenever 0 < p ≤ ∞, 0 < q0 ≤ q1 ≤ ∞, and s ∈ R.
Also, (9.1.13) implies that the following inclusions are continuous:
p,min{p,q } p,q p,max{p,q }
Bs (Ω) → Fs (Ω) → Bs (Ω)
(9.2.11)
whenever 0 < p, q ≤ ∞ and s ∈ R.
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 529

Corollary 9.2.1 For any open set Ω ⊆ Rn , the same type of embedding results as in
p,q
Theorem 9.1.1, Theorem 9.1.3, and Theorem 9.1.4 are valid for the spaces Bs (Ω),
p,q p,q p,q p,q p,q
Fs (Ω), Bs,0 (Ω), Fs,0 (Ω), Bs,z (Ω), and Fs,z (Ω), respectively.

Proof This is immediate from Theorem 9.1.1, Theorem 9.1.3, Theorem 9.1.4, and
the definitions given in (9.2.1), (9.2.3), and (9.2.4). 

We shall also need embeddings for Besov and Triebel-Lizorkin spaces in arbitrary
bounded open subsets of the Euclidean space, of the sort presented in the next
theorem.

Theorem 9.2.2 Suppose Ω ⊆ Rn is an arbitrary bounded open set. Then whenever

0 < p0, p1 ≤ ∞, 0 < q0, q1, q ≤ ∞, and −∞ < s1 < s0 < +∞ it follows that one has
the continuous embeddings:

p ,q0 p ,q1 1 s0 1 s1
Fs00 (Ω) → Fs11 (Ω) if − ≤ − , (9.2.12)
p0 n p1 n

p ,q0 p ,q1 1 s0 1 s1
Bs00 (Ω) → Bs11 (Ω) if − < − , (9.2.13)
p0 n p1 n

p ,q p ,q 1 s0 1 s1
Bs00 (Ω) → Bs11 (Ω) if − ≤ − . (9.2.14)
p0 n p1 n
Moreover, whenever 0 < p0, p1 ≤ ∞, 0 < q ≤ ∞, and −∞ < s < +∞ the
following continuous embeddings hold:
p ,q p ,q
Fs 0 (Ω) → Fs 1 (Ω) if p0 ≥ p1, (9.2.15)

p ,q p ,q
Bs 0 (Ω) → Bs 1 (Ω) if p0 ≥ p1 . (9.2.16)

Proof This follows from (9.2.1) and [187, Theorem 3.3.1] (cf. also [166, Theorem 1,
p. 82]). 

The same blue-print for fashioning smoothness spaces in open sets out of their
global versions in the entire Euclidean space works in other concrete cases of interest.
For example, we define Bessel potential spaces in a given open set Ω ⊆ Rn by setting
     
Ls (Ω) := U Ω : U ∈ Ls (Rn ) = U Ω : U ∈ Fs (Rn )
p p p,2
(9.2.17)
for all p ∈ (1, ∞) and s ∈ R,
and equip each of them with the norm
  
u Lsp (Ω) := inf U  Lsp (Rn ) : U ∈ Ls (Rn ) such that u = U Ω
p

  
≈ inf U F p,2 (Rn ) : U ∈ Fs (Rn ) such that u = U Ω .
p,2
(9.2.18)
s
530 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Then from (9.2.17)-(9.2.18) and (9.1.54) we conclude that

p,2 p
Fs (Ω) = Ls (Ω) for each p ∈ (1, ∞) and s ∈ R. (9.2.19)

As in the past, the case p = 2 deserves special consideration. This corresponds to

L 2 -based fractional Sobolev spaces, defined in a given open set Ω ⊆ Rn as
     
H s (Ω) := U Ω : U ∈ H s (Rn ) = U Ω : U ∈ Fs2,2 (Rn ) , ∀s ∈ R, (9.2.20)

each equipped with the norm

  
u H s (Ω) := inf U  H s (Rn ) : U ∈ H s (Rn ) such that u = U Ω
  
≈ inf U F 2,2 (Rn ) : U ∈ Fs2,2 (Rn ) such that u = U Ω . (9.2.21)
s

It is important to remark that for any open set Ω ⊆ Rn we have

Fs2,2 (Ω) = Bs2,2 (Ω) = H s (Ω) for each s ∈ R. (9.2.22)

This is seen from (9.1.62) and (9.2.20)-(9.2.21). From (9.1.52)-(9.1.53) it is also

apparent that

Bs∞,∞ (Ω) = C s (Ω) if s ∈ (0, 1), and (9.2.23)

p,2
F0 (Ω) = L p (Ω, L n ) if 1 < p < ∞. (9.2.24)

Next we consider the version of the spaces (9.1.65) defined in open subsets of
Rn . Following [172], given an open set Ω ⊆ Rn abbreviate δ∂Ω (x) := dist(x, ∂Ω) for
each x ∈ Rn and, whenever 0 < p < ∞, 0 < q ≤ ∞, α > 0, and m ∈ N is such that
m > α, introduce
p,q  
Cα (Ω) := u ∈ Lloc 1
(Ω, L n ) : uCαp, q (Ω) < +∞ (9.2.25)

where, for each Lebesgue measurable function u defined in Ω, we have set

 ∫ 2 δ ∂Ω (x)
1
dt  1/q 

u p, q
Cα (Ω) := u L p (Ω, L n ) +  [t −α oscm−1 (u, x, t)]q
 p
0 t L x (Ω, L n )
(9.2.26)
with the subscript x in the last line above indicating that the outer L p -norm is taken
in the variable x ∈ Ω. To be pedantic, the notation should also reflect the perceived
p,q
dependence of Cα (Ω) on the parameter m. Whenever necessary, we shall therefore
p,q,m p,q
write Cα (Ω) in place of Cα (Ω). As a consequence of [172, Theorem 2, p. 397],

if Ω ⊆ Rn is an (ε, δ)-domain and 0 < p < ∞, 0 < q ≤ ∞,

m ∈ N with m > α > n min {p,q1
} − 1 + , then there exists a (9.2.27)
p,q,m p,q
bounded extension operator E : Cα (Ω) −→ Cα (Rn ).
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 531

Granted this, elementary considerations then imply that, in the same context as
p,q,m n p,q
above, the restriction operator R Rn →Ω : Cα (R ) → Cα (Ω) is well defined,
p,q
linear, bounded and onto. On the other hand, so is R Rn →Ω acting from Fα (Rn )
p,q
into Fα (Ω). In concert with (9.1.67), this analysis proves the following (which is
essentially [172, Corollary 1, p. 398]):

Proposition 9.2.3 Let Ω ⊆ Rn be an (ε, δ)-domain. Then, with equivalence of quasi-

norms,
p,q,m p,q
Cα (Ω) = Fα (Ω) whenever 0 < p < ∞, 0 < q ≤ ∞,
(9.2.28)
and m ∈ N with m > α > n 1
min {p,q } − 1 +.

The family of spaces (9.2.25)-(9.2.26) interfaces tightly with another scale intro-
duced by R. DeVore and R. Sharpley in [45]. Specifically, given an open set Ω ⊆ Rn
along with p ∈ (0, ∞) and α ≥ 0, they have considered
   
(Ω, L n ) : uCαp (Ω) := u L p (Ω, L n ) + uα#  L p (Ω, L n ) < ∞ ,
p
Cα (Ω) := u ∈ Lloc 1

(9.2.29)
where
 ∫ 
1
uα (x) :=
#
sup inf n+α |u − P| dL , n
∀x ∈ Ω. (9.2.30)
0<r <δ ∂Ω (x) P ∈ P[α] r B(x,r)

From [148, Theorem 4/(ii), p. 1035] we know that

if Ω ⊆ Rn is an (ε, δ)-domain and 0 < p, α < ∞, then there

p p (9.2.31)
exists a bounded extension operator E : Cα (Ω) −→ Cα (Rn ).

Proposition 9.2.4 Let Ω ⊆ Rn be an (ε, δ)-domain. Then

p,∞,m p
Cα (Ω) = Cα (Ω) (9.2.32)

whenever 0 < p < ∞ and m ∈ N is such that m > α > n 1

p − 1 +.

Proof It is useful to introduce uα,λ # , a variant of the sharp function (9.2.30), defined

for each fixed parameter λ ∈ (0, 1] much as before, except that the supremum is
now taken over 0 < t < λ · δ∂Ω  (x).Then from (9.2.26) with m = [α] + 1, we have
uCαp,∞ (Ω) ≈ u L p (Ω, L n ) + u# 1  L p (Ω, L n ) . As such, everything comes down to
α, 2
checking that  #   
u 1  p ≈ uα#  L p (Ω, L n ) . (9.2.33)
α, L (Ω, L n )
2

In turn, for the current range of indices, this is a direct consequence of the estimate
established in [149, Lemma 2.3, p. 65]. 

Combining Proposition 9.2.3 with Proposition 9.2.4 then yields the following
identification (for related results in smooth domains, see also [190, p. 50 and p. 248]).
 532 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Corollary 9.2.5 If Ω ⊆ Rn is an (ε, δ)-domain then

p p,∞
Cα (Ω) = Fα (Ω) whenever
(9.2.34)
0 < p < ∞ and α > n 1
p − 1 +.

Moving on, we shall now consider local Hardy spaces and local Hardy-based
Sobolev spaces in open subsets of the Euclidean ambient. To get started, in same
spirit as (9.2.1), given an open set Ω ⊆ Rn and p ∈ (0, ∞) define
  
h p (Ω) := u ∈ D (Ω) : there exists U ∈ h p (Rn ) such that U Ω = u ,
   (9.2.35)
uh p (Ω) := inf U h p (Rn ) : U ∈ h p (Rn ), U Ω = u , ∀u ∈ h p (Ω).

Then from (9.2.35), (9.2.1), and (9.1.56) we conclude that for each open set Ω ⊆ Rn
we have
p,2
F0 (Ω) = h p (Ω) for 0 < p < ∞ (9.2.36)

and
p,2
F0 (Ω) = h p (Ω) = L p (Ω, L n ) provided 1 < p < ∞. (9.2.37)

In order to be able to discuss a useful characterization of the local Hardy

space
∫ h p (Ω), we need some notation. Concretely, fix ψ ∈ Cc∞ B(0, 1) such that
B(0,1)
ψ dL n = 1 and set ψt := t −n ψ(·/t) for each t > 0. The radial maximal
function of a distribution u ∈ D (Ω) is defined as (see [145, p. 205])
+
uΩ (x) := sup |(ψt ∗ u)(x)|, ∀x ∈ Ω. (9.2.38)
0<t<dist(x,∂Ω)

Having fixed k ∈ N0 , for each u ∈ D (Ω) also associate the grand maximal
∗
function uk,Ω defined as (see [147, p. 171])

∗
 
uk,Ω (x) := sup | u, ψ| : ψ ∈ Ψx , ∀x ∈ Rn, (9.2.39)

where, for each x ∈ Rn , we have set


Ψx := ψ ∈ Cc∞ (Rn ) : there exists r > 0 such that supp ψ ⊆ B(x, r) ∩ Ω (9.2.40)

and sup |∂γ ψ| ≤ r −n− |γ | for each γ ∈ N0n with |γ| ≤ k
Rn

(a slight variant of (9.2.40) has been considered in [145, p. 209]). The result below
is implied by [145, Theorem 4, p. 226].

Theorem 9.2.6 Let Ω ⊆ Rn be an open set satisfying an exterior corkscrew condition

and fix some p ∈ (0, 1]. Then for any u ∈ D (Ω) one has
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 533

+
u ∈ h p (Ω) ⇐⇒ uΩ ∈ L p (Ω, L n ), (9.2.41)

with equivalence of quasi-norms. Furthermore, if k ∈ N is large enough so that

n+k < p, then
n
 +
∗
uk,Ω  L p (Rn, L n ) ≈ uΩ  p
L (Ω, L n )
. (9.2.42)
In particular, a different choice of the function ψ affects the size of u+ in L p (Ω, L n )
only up to a fixed multiplicative constant. Finally, similar results are valid in the
range 1 < p < ∞ provided h p (Ω) is replaced by L p (Ω, L n ).
In analogy with the Hardy-based Sobolev spaces in Rn introduced in (9.1.72)-
(9.1.73), for an open set Ω ⊂ Rn , k ∈ N0 , and p ∈ (0, 1], we define the local
p
Hardy-based Sobolev space hk (Ω) as
 
hk (Ω) := u ∈ D (Ω) : ∂γ u ∈ h p (Ω), for all γ ∈ N0n with |γ| ≤ k ,
p
(9.2.43)

equipped with the quasi-norm uh p (Ω) := |γ |=k ∂γ uh p (Ω) , and
k

  
h−k (Ω) := u ∈ D (Ω) : u = ∂γ uγ with each uγ ∈ h p (Ω)
p
(9.2.44)
|γ | ≤k

equipped with uh p (Ω) := inf |γ | ≤k uγ h p (Ω) where the infimum is taken over
−k
p
all representations of the distribution u ∈ hk (Ω) as in (9.2.44).
Proposition 9.2.7 Let Ω be a bounded an (ε, δ)-domain in Rn . Fix p ∈ (0, 1] and
assume k ∈ Z is either ≤ 0, or else satisfies k > n 1/p − 1 . Then
p p,2
hk (Ω) = Fk (Ω). (9.2.45)
p p,2
Proof Assume first that k < 0 and note that the inclusion hk (Ω) → Fk (Ω)
is immediate from definitions, (9.2.36), (9.2.8), and (9.2.9). To see the opposite
p,2 p,2
inclusion fix u ∈ Fk (Ω), say u = w|Ω for some w ∈ Fk (Rn ). Since, by (9.1.74), w
may be represented γ
in the form w = |γ | ≤−k ∂ wγ with each wγ ∈ h p (Rn ), it follows
 γ p
that u = |γ | ≤−k ∂ (wγ |Ω ) and wγ |Ω ∈ h p (Ω). Consequently, u ∈ hk (Ω), proving the
right-to-left inclusion in (9.2.45). At this stage, we may conclude that (9.2.45) holds
in the case when k < 0.
The case when k ∈ Z satisfies k > n 1/p − 1 makes essential use of Miyachi’s
work in [147], [146]. Specifically, as in [146, Remark (h), p. 80], let us introduce
 
Wpk (Ω) := u ∈ h p (Ω) : ∂γ u ∈ h p (Ω) for all γ ∈ N0n with |γ| = k (9.2.46)

and observe that, by virtue of the last remark of §4 on p. 80 in [146],

p
Wpk (Ω) = Ck (Ω), (9.2.47)
p
where Ck (Ω) is the DeVore-Sharpley space defined as in (9.2.29)-(9.2.30) (with
α := k). On account of the extension result recalled in (9.2.31) it follows that
534 9 Besov and Triebel-Lizorkin Spaces in Open Sets
  
Wpk (Ω) = uΩ : u ∈ Wpk (Rn ) . However, thanks to (9.1.18) and (9.1.56), we have
p,2
Wpk (Rn ) = Fk (Rn ) so, all together,
p,2
Wpk (Ω) = Fk (Ω). (9.2.48)
p p,2
Consequently, hk (Ω) → Wpk (Ω) = Fk (Ω), proving the left-to-right inclusion in
(9.2.45). The opposite inclusion is a consequence of (9.2.36), (9.2.8), (9.2.9), and
definitions. 

Suppose Ω ⊆ Rn is an arbitrary open set and fix p ∈ (1, ∞) along with k ∈ N0 .

Then the intrinsically defined L p -based Sobolev space of order k in Ω is defined as
 
W k, p (Ω) := u ∈ L p (Ω, L n ) : ∂γ u ∈ L p (Ω, L n ) for all γ ∈ N0n with |γ| ≤ k .
(9.2.49)
As is well known, this is a Banach space when equipped with the natural norm

uW k, p (Ω) := |γ | ≤k ∂γ u L p (Ω, L n ) for each u ∈ W k, p (Ω). We may also consider
L p -based Sobolev spaces of negative order. With Ω, p, k as above, these are intrin-
sically defined via
  
W −k, p (Ω) := u ∈ D (Ω) : u = ∂γ uγ with each uγ ∈ L p (Ω, L n ) , (9.2.50)
|γ | ≤k

and are equipped with the norm uW −k, p (Ω) := inf |γ | ≤k uγ  L p (Ω, L n ) where the
infimum is taken over all representations of the distribution u ∈ W −k, p (Ω) as in
(9.2.50). Based on these definitions and (9.2.37) it follows that for each open set
Ω ⊆ Rn we have
p
W k, p (Ω) = hk (Ω) if 1 < p < ∞ and k ∈ Z. (9.2.51)

Also, reasoning much as in Remark 8.5.1 we see that for each open set Ω ⊆ Rn we
have
  
W −k, p (Ω) = uΩ : u ∈ W −k, p (Rn )
(9.2.52)
whenever p ∈ (1, ∞) and k ∈ N0,

in a quantitative fashion. The above definitions and remarks set the stage for the
following natural identification result.

Corollary 9.2.8 Let Ω be a bounded (ε, δ)-domain in Rn . Then for each k ∈ Z and
p ∈ (0, ∞) one has
 p
p,2
hk (Ω) if 0 < p ≤ 1 and either k ≤ 0, or k > n p1 − 1 ,
Fk (Ω) = (9.2.53)
W k, p (Ω) if 1 < p < ∞ and k ∈ Z.

Proof This is a consequence of Proposition 9.1.9, (9.2.52), (9.2.1), Proposition 9.2.7,

and the fact that, in the current geometric setting, the space W k, p (Ω) may be described
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 535
  
as uΩ : u ∈ W k, p (Rn ) in a quantitative manner whenever p ∈ (1, ∞) and k ∈ N0
(which is established in [103]). 

There are geometric and analytic assumptions guaranteeing the coincidence of the
spaces defined in (9.2.1)-(9.2.2). A result of this favor is contained in the proposition
below.

Proposition 9.2.9 Let Ω ⊆ Rn be an open set with a compact Ahlfors regular

boundary. Then for each A ∈ {B, F} one has
p,q p,q
As (Ω) = Ås (Ω) provided p, q ∈ (0, ∞) and s < p.
1
(9.2.54)

Proof From [133, (8.6.97)] we know that ∂Ω is a porous set. Granted this property,
we may then invoke [22, Proposition 2.5, p. 52] to conclude (9.2.54). 

The gist of our next lemma is that sufficiently regular distributions supported on
“thin” sets necessarily vanish.

Lemma 9.2.10 Let Σ ⊆ Rn be a closed Ahlfors regular set. Then

 p,q 
u ∈ As (Rn ) : supp u ⊆ Σ = {0}, A ∈ {B, F}, (9.2.55)

if 1 
0 < p < ∞, 0 < q ≤ ∞, and max p − 1, n 1
p −1 < s. (9.2.56)

Proof In the case when 1 < p < ∞, 1

p − 1 < s, and 0 < q ≤ ∞ for A = B and
1 ≤ q ≤ ∞ for A = F, we may conclude (9.2.55) by invoking [191, Remark 3.20,
p. 87], whose applicability to the present case with Ω := Rn \ Σ is ensured by [133,
(8.6.97)], and Theorem 9.1.1. The more general case in the statement of the lemma
follows from this and standard embeddings for Besov and Triebel-Lizorkin spaces
in the entire Euclidean setting. More specifically, if p, q, s are as in (9.2.56) then
Theorem 9.1.4 implies the existence of p∗ ∈ (1, ∞), q∗ ∈ [1, ∞] and s∗ > p1∗ − 1 such
p,q p ,q
that As (Rn ) ⊆ As∗∗ ∗ (Rn ) and the desired conclusion follows from the first part of
the proof. 

A useful consequence of Lemma 9.2.10, identifying a context in which the re-

striction operator to an open set is actually an isomorphism, is given next.

Lemma 9.2.11 Let Ω ⊆ Rn be an open set with an Ahlfors regular boundary. Then
p,q p,q
R Rn →Ω : As,0 (Ω) −→ As,z (Ω), A ∈ {B, F}, isomorphically
  (9.2.57)
provided 0 < p < ∞, 0 < q ≤ ∞, max p1 − 1, n p1 − 1 < s.

As a consequence, for the above range of indices one may canonically identify
p,q p,q
As,0 (Ω) with As,z (Ω).
p,q
Proof From (9.2.3)-(9.2.4) we see that R Rn →Ω maps the space As,0 (Ω) onto
p,q
As,z (Ω), in a linear and bounded fashion whenever 0 < p, q ≤ ∞ and s ∈ R.
536 9 Besov and Triebel-Lizorkin Spaces in Open Sets

There remains to observe that assumptions on p, q, s made in (9.2.57) this opera-

p,q
tor is also injective. To this end, assume u ∈ As,0 (Ω) with p, q, s as in (9.2.57) is
such that R Rn →Ω u = 0. Since we know that supp u ⊆ Ω to begin with, this forces
supp u ⊆ Ω ∩ (Rn \ Ω) = ∂Ω. Having established this, Lemma 9.2.10 applies (with
Σ := ∂Ω) and gives u = 0. This proves the injectivity of the restriction operator in
the context of (9.2.57), so the desired conclusion follows. 

We next turn our attention to the issue of extension operators with preservation
of class. For starters, we collect some remarks of a general nature in the following
lemma.

Lemma 9.2.12 Let Ω ⊆ Rn be an open set and assume that 0 < p, q ≤ ∞, s ∈ R,

and A ∈ {B, F} are such that there exists a linear and bounded extension operator,
i.e., a linear and bounded mapping
p,q p,q
EΩ→Rn : As (Ω) −→ As (Rn ), (9.2.58)

with the property that

p,q
R Rn →Ω EΩ→Rn u = u for all u ∈ As (Ω). (9.2.59)

Then
 
EΩ→Rn u p, q ≈ u Ap, q uniformly for u ∈ As (Ω).
p,q
(9.2.60)
As (R n ) s (Ω)

As a consequence, EΩ→Rn has closed range and maps isomorphically onto its image,
p,q p,q
hence As (Ω) is isomorphic to a closed subspace of As (Rn ). In particular,
p,q
As (Ω) is a reflexive Banach space if 1 < p, q < ∞. (9.2.61)
p,q
Moreover, with I denoting the identity operator on As (Rn ), it follows that
p,q p,q
P := I − EΩ→Rn ◦ R Rn →Ω : As (Rn ) −→ As (Rn ) (9.2.62)

is a projection (i.e., P is linear, bounded, and satisfies P2 = P). Finally,

p,q p,q
Im P := P As (Rn ) = As,0 Rn \ Ω provided ∂Ω = ∂(Ω). (9.2.63)

As a consequence,
p,q
Im P = As,0 Rn \ Ω whenever
(9.2.64)
Ω satisfies an exterior corkscrew condition.

Proof Indeed, one inequality in (9.2.60) follows from the boundedness of the op-
p,q p,q
erator EΩ→Rn from As (Ω) into As (Rn ), while the opposite inequality is a con-
sequence of the definition in (9.2.1) and of the identity (9.2.59). Next, that P given
in (9.2.62) is a projection satisfying (9.2.63) is readily seen from definitions (the
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 537

demand that ∂Ω = ∂(Ω) ensures that Rn \ Ω = Rn \ Ω). Finally, (9.2.64) is a

consequence of (9.2.63) and [133, (5.1.10)]. 
Our next two lemmas establish that the Besov and Triebel-Lizorkin scales in open
sets (and related variants) consists of analytically convex quasi-Banach spaces (a
functional analytic concept introduced and studied in §1.4).

Lemma 9.2.13 Assume that Ω ⊆ Rn is an open set, and suppose 0 < p, q ≤ ∞,

s ∈ R, and A ∈ {B, F} are such that a linear and bounded extension operator as in
p,q p,q
(9.2.58) exists. Then the spaces As (Ω) and As,0 (Ω) are analytically convex.
p,q
Proof The claim regarding As (Ω) follows from the last part of Lemma 9.2.12,
p,q
Proposition 1.4.6, and the corresponding result for As (Rn ) (cf. Theorem 9.1.6).
p,q
The claim about As,0 (Ω) is a consequence of Proposition 1.4.6, definition (9.2.3)
and Theorem 9.1.6. 

Lemma 9.2.14 Let Ω ⊆ Rn be an open set. Fix pi, qi ∈ (0, ∞] along with si ∈ R, for
i ∈ {0, 1}, and let A ∈ {B, F}.

(1) If there exists a common linear and bounded extension operator

p ,qi p ,qi
EΩ→Rn : Asii (Ω) −→ Asii (Rn ), i ∈ {0, 1}, (9.2.65)
p ,q p ,q
then the sum space As00 0 (Ω) + As11 1 (Ω) is analytically convex.
(2) If ∂Ω = ∂(Ω) and there exists a common linear and bounded extension operator
p ,qi p ,qi
ERn \Ω→Rn : Asii (Rn \ Ω) −→ Asii (Rn ), i ∈ {0, 1}, (9.2.66)
p ,q p ,q
then the sum space As00,0 0 (Ω) + As11,0 1 (Ω) is analytically convex.
p ,q p ,q
Proof Let us abbreviate Xi := Aαii i (Ω) and Yi := Aαii ,0 i (Ω) for i ∈ {0, 1}. Granted
(9.2.65), the space X0 +X1 is analytically convex by the first part of Lemma 1.4.23 and
Theorem 9.1.6. This establishes the claim in part (1) of the lemma. The hypotheses
in part (2) together with Lemma 9.2.12 ensure that the operator
p ,qi p ,q
P := I − ERn \Ω→Rn ◦ R Rn →Rn \Ω : Aαii (Rn ) −→ Aαii ,0 i (Ω), i ∈ {0, 1}, (9.2.67)
p ,q
is a common projection for the spaces Aαii i (Rn ), i ∈ {0, 1}. Hence, in such a sce-
nario, the second part of Lemma 1.4.23 applies and, in concert with Theorem 9.1.6,
proves that Y0 + Y1 is analytically convex. 

The next theorem is a particular case of a more general extension result proved
by H. Triebel in [191].

Theorem 9.2.15 (i) Let Ω be a nonempty open subset of Rn satisfying an interi-

or corkscrew condition and with the property that Ω  Rn . Then for any given
smoothness threshold M > 0 there exists a common extension operator
538 9 Besov and Triebel-Lizorkin Spaces in Open Sets
p,q p,q
M
EΩ→R n : As (Ω) −→ As (Rn )
(9.2.68)
linearly and boundedly, for A ∈ {B, F},

whenever

0 < p ≤ ∞, 0 < q ≤ ∞, n 1
p −1 + < s < M if A = B, (9.2.69)

0 < p < ∞, 0 < q ≤ ∞, n 1

min{p,q } −1 + < s < M if A = F. (9.2.70)

(ii) Suppose Ω ⊆ Rn is a nonempty open set with an Ahlfors regular boundary

and such that Ω = Rn . Then there exists a common extension operator
p,q p,q
EΩ→Rn : As (Ω) −→ As (Rn )
(9.2.71)
linearly and boundedly, for A ∈ {B, F},

whenever
1 
0 < p < ∞, 0 < q ≤ ∞, and max p − 1, n 1
p −1 < s. (9.2.72)

Proof The claim in part (i) follows from [191, Theorem 4.4, p. 103] where such
an extension result is proved for nonempty open sets Ω ⊆ Rn satisfying Ω  Rn ,
L n (∂Ω) = 0, and which are I-thick. To elaborate on the nature of the latter condition,
recall the reflection technique Q → Q∗ used in the construction of Jones’ extension
operator from [103]. The key geometrical requirement for the open set Ω ⊆ Rn
◦
is the ability to associate to each cube Q ⊆ Ωc such that (Q) ≈ 2−j and
−j
dist(Q, ∂Ω) ≈ 2 where j ∈ N is larger than or equal to a suitably chosen threshold
jo ∈ N, a so-called reflected cube Q∗ ⊆ Ω with the property that (Q∗ ) ≈ 2−j ,
dist(Q∗, ∂Ω) ≈ 2−j , and dist(Q∗, Q) ≈ 2−j . Nonempty, open, proper subsets Ω of Rn
satisfying this condition are called I-thick (interior thick) by H. Triebel in [191,
Definition 3.1(iii), p. 70]. A routine argument shows that
any open, nonempty, proper subset of Rn satisfying the
(9.2.73)
interior corkscrew condition is automatically I-thick.
Granted this, and keeping in mind [133, (5.1.6)] as well as [133, Lemma 5.1.2], we
may invoke [191, Theorem 4.4, p. 103] which provides the conclusion in part (i).
Turning to part (ii), consider now the scenario when Ω ⊆ Rn is a nonempty open
set satisfying Ω = Rn and such that ∂Ω is Ahlfors regular. Then Σ := ∂Ω is a closed
Ahlfors regular set satisfying Σ = Rn \ Ω. Recall from (9.2.6) that the restriction to
Ω (in the sense of distributions) is a linear, bounded, surjective operator
p,q p,q
R Rn →Ω : As (Rn ) −→ As (Ω), A ∈ {B, F}, provided
  (9.2.74)
0 < p < ∞, 0 < q ≤ ∞, max p1 − 1, n p1 − 1 < s.

We claim that under the present assumptions this operator is also injective. Indeed,
p,q
if u ∈ As (Rn ), with the indices p, q, s as above, is such that R Rn →Ω u = 0 it follows
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 539

that supp u ⊆ Rn \ Ω = Σ. Granted this, Lemma 9.2.10 forces u = 0, proving that

R Rn →Ω in (9.2.74) is injective. Hence, ultimately, R Rn →Ω is a linear isomorphism
in the context of (9.2.74). Taking EΩ→Rn to be the inverse of this operator, yields a
mapping
p,q p,q
EΩ→Rn : As (Ω) −→ As (Rn ) linear and bounded, for A ∈ {B, F},
p,q
with the property that R Rn →Ω ◦ EΩ→Rn = I, the identity on As (Ω), (9.2.75)
 
provided 0 < p < ∞, 0 < q ≤ ∞, and max p1 − 1, n p1 − 1 < s.

Thus, EΩ→Rn is an extension operator with preservation of class. 

The next two theorems below describe extension operators in the negative (or
close to negative) regime of smoothness.

Theorem 9.2.16 Let Ω  Rn be a nonempty open set satisfying an exterior corkscrew

condition. Then for any ε ∈ (0, 1) there exists a common extension operator
ε p,q p,q
EΩ→R n : As (Ω) −→ As (Rn ) linearly and boundedly, (9.2.76)

whenever

ε < p ≤ ∞, 0 < q ≤ ∞, − ε1 < s < 0 if A = B (9.2.77)

ε < p < ∞, ε < q ≤ ∞, − ε1 < s < 0 if A = F. (9.2.78)

Proof This is implied by [191, Theorem 4.7, p. 105]. 

Theorem 9.2.17 Let Ω  Rn be an open set satisfying an exterior corkscrew condi-

tion and whose boundary is Ahlfors regular. Then for any ε ∈ (0, 1) there exists a
common extension operator
ε p,q p,q
EΩ→R n : As (Ω) −→ As (Rn ) linearly and boundedly, (9.2.79)

whenever
⎧
⎪ ε < p ≤ ∞, 0 < q ≤ ∞, − ε1 < s < 0,
⎪
⎨
⎪
or if A = B, (9.2.80)
⎪
⎪
⎪ 1 < p < ∞, 0 < q ≤ ∞, 0 ≤ s < 1 ,
⎩ p

and
⎧
⎪ ε < p < ∞, ε < q ≤ ∞, − ε1 < s < 0,
⎪
⎨
⎪
or if A = F. (9.2.81)
⎪
⎪
⎪ 1 < p < ∞, 1 ≤ q < ∞, 0 ≤ s < 1 ,
⎩ p

Proof This is implied by [191, Theorem 4.10 p. 106]. 

540 9 Besov and Triebel-Lizorkin Spaces in Open Sets

We now discuss the interpolation of the Besov and Triebel-Lizorkin spaces and
establish the following analogue of Theorem 9.1.6 in certain classes of open subsets
of the Euclidean ambient. We begin by considering the complex method for analyt-
ically convex quasi-Banach spaces developed in §1.4 (cf. Lemmas 9.2.13-9.2.14 in
this regard).

Theorem 9.2.18 Assume Ω ⊂ Rn is a nonempty open set with the property that
Ω  Rn . Also, suppose 0 < p0, p1 ≤ ∞, 0 < q0, q1 ≤ ∞ with min {q0, q1 } < ∞,
α0, α1 ∈ R, θ ∈ (0, 1), and define
θ −1 θ −1
p := 1−θ
p0 + p1 , q := 1−θ
q0 + q1 , α := (1 − θ)α0 + θα1 . (9.2.82)

Then, with A ∈ {B, F}, one has

 p0,q0 p ,q  p,q
Aα0 (Ω), Aα11 1 (Ω) θ = Aα (Ω) (9.2.83)

in each of the following scenarios:

(i) The set Ω satisfies an interior corkscrew condition and

n 1
pi −1 + < αi for i ∈ {0, 1}, if A = B,
(9.2.84)
0 < pi < ∞, n min{pi ,qi } − 1 +
1
< αi for i ∈ {0, 1}, if A = F.

(ii) The set Ω satisfies an exterior corkscrew condition and

αi < 0 for i ∈ {0, 1}, if A = B,

(9.2.85)
0 < pi < ∞ and αi < 0 for i ∈ {0, 1}, if A = F.

(iii) The set Ω satisfies an exterior corkscrew condition, has an Ahlfors regular
boundary, and

1 < pi < ∞ and 0 ≤ αi < 1

pi for i ∈ {0, 1}, if A = B,
(9.2.86)
1 < pi < ∞, 1 ≤ qi < ∞, 0 ≤ αi < 1
pi for i ∈ {0, 1}, if A = F.

Proof Consider first the scenario described in (i). Pick a number M > max{α0, α1 }
M
and bring in the extension operator EΩ→R n from part (i) of Theorem 9.2.15. Then
Lemma 1.4.23 used with E := EΩ→Rn and R := R Rn →Ω gives
M

    
p ,q0 p ,q1 p ,q0 p ,q1
Aα00 (Ω), Aα11 (Ω) θ
= R Rn →Ω Aα00 (Rn ), Aα11 (Rn ) θ
p,q p,q
= R Rn →Ω Aα (Rn ) = Aα (Ω), (9.2.87)

by Theorem 9.1.6 and the ontoness of the operator in (9.2.6). This establishes (9.2.83)
in the scenario described in item (i).
In the scenario described in item (ii), formula (9.2.83) is proved in a similar
fashion now using the extension operator from Theorem 9.2.16. Finally, the validity
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 541

of (9.2.83) in the setting from item (iii) is established via the same type of argument,
now relying on the extension operator from Theorem 9.2.17. 
In turn, Theorem 9.2.18 allows for the following generalization of Theorem 9.2.17.
Theorem 9.2.19 Let Ω  Rn be an open set satisfying an exterior corkscrew condi-
tion and whose boundary is Ahlfors regular. Also, as usual, let A ∈ {B, F}. Then for
each ε ∈ (0, 1) there exists a common extension operator
ε p,q p,q
EΩ→R n : As (Ω) −→ As (Rn ) linearly and boundedly, (9.2.88)

whenever

ε < p ≤ ∞, 0 < q ≤ ∞, − ε1 < s < 1

p − ε if A = B,
(9.2.89)
ε < p < ∞, 0 < q ≤ ∞, − ε1 < s < 1
p − ε if A = F.

Proof This follows from Theorem 9.2.18, Theorem 9.2.17, and complex interpola-
tion. 
The following corollary further expands on the interpolation result corresponding
to part (iii) of Theorem 9.2.18.
Corollary 9.2.20 Assume Ω ⊂ Rn is a nonempty open set satisfying an exterior
corkscrew condition and whose boundary is Ahlfors regular. In addition, suppose
0 < p0, p1 ≤ ∞, 0 < q0, q1 ≤ ∞ with min {q0, q1 } < ∞, α0, α1 ∈ R, θ ∈ (0, 1), and
define
θ −1 θ −1
p := 1−θ
p0 + p1 , q := 1−θ
q0 + q1 , α := (1 − θ)α0 + θα1 . (9.2.90)

Then, with A ∈ {B, F}, one has

 p0,q0 p ,q  p,q
Aα0 (Ω), Aα11 1 (Ω) θ = Aα (Ω) (9.2.91)

provided
αi < 1
pi for i ∈ {0, 1}, when A = B,
(9.2.92)
0 < pi < ∞, αi < 1
pi for i ∈ {0, 1}, when A = F.

Proof This is proved much as part (iii) of Theorem 9.2.18, now using the exten-
sion operator from Theorem 9.2.19 in place of the extension operator from Theo-
rem 9.2.17. 
The following result is the analogue of Theorem 9.2.18 for the scales of Besov
and Triebel-Lizorkin spaces in Rn with support in a fixed subset of the Euclidean
ambient.
Theorem 9.2.21 Assume Ω ⊂ Rn is a nonempty open set satisfying ∂Ω = ∂(Ω).
Suppose that 0 < p0, p1 ≤ ∞, 0 < q0, q1 ≤ ∞ with min {q0, q1 } < ∞, α0, α1 ∈ R,
θ ∈ (0, 1), and define
542 9 Besov and Triebel-Lizorkin Spaces in Open Sets

θ −1 θ −1
p := 1−θ
p0 + p1 , q := 1−θ
q0 + q1 , α := (1 − θ)α0 + θα1 . (9.2.93)

Then, with A ∈ {B, F}, one has

 p0,q0 p ,q  p,q
Aα0,0 (Ω), Aα11,0 1 (Ω) θ = Aα,0 (Ω) (9.2.94)

in each of the following scenarios:

(i) The set Ω satisfies an exterior corkscrew condition and

n 1
pi −1 + < αi for i ∈ {0, 1}, if A = B,
(9.2.95)
0 < pi < ∞, n min{pi ,qi } − 1 +
1
< αi for i ∈ {0, 1}, if A = F.

(ii) The set Ω satisfies an interior corkscrew condition and

αi < 0 for i ∈ {0, 1}, if A = B,

(9.2.96)
0 < pi < ∞ and αi < 0 for i ∈ {0, 1}, if A = F.

(iii) The set Ω satisfies an interior corkscrew condition, has an Ahlfors regular
boundary, and

αi < 1
pi for i ∈ {0, 1}, if A = B,
(9.2.97)
0 < pi < ∞ and αi < 1
pi for i ∈ {0, 1}, if A = F.

Proof Lemma 9.2.12 (used for Rn \ Ω in place of Ω) implies that, for Ω as in the
statement,
whenever there exists some linear and bounded extension operator
p,q p,q
ERn \Ω→Rn : Aα (Rn \ Ω) → Aα (Rn ), for some p, q ∈ (0, ∞], α ∈ R,
and A ∈ {B, F}, then P := I − ERn \Ω→Rn ◦ R Rn →Rn \Ω is a projection (9.2.98)
p,q p,q
on the space Aα (Rn ) whose image is precisely Aα,0 (Ω).

Granted this, all desired conclusions follow on account of the second part of
Lemma 1.4.23, Theorem 9.1.6, and the common extension operators (from Rn \ Ω)
provided by part (i) of Theorem 9.2.15 (for the current item (i)), Theorem 9.2.16 (for
the current item (ii)), and Theorem 9.2.19 (for the current item (iii)). 

As regards the real interpolation of Besov and Triebel-Lizorkin spaces defined in

open subsets of Rn , we note the following analogue of Theorem 9.1.7.

Theorem 9.2.22 Suppose Ω ⊂ Rn is a nonempty open set with the property that
Ω  Rn . Also, fix 0 < p < ∞, 0 < q0, q1, q ≤ ∞, 0 < θ < 1, α0, α1 ∈ R with α0  α1 ,
and set α := (1 − θ)α0 + θα1 . Then for every q ∈ (0, ∞] one has
p,q p,q p,q p,q p,q
Fα0 0 (Ω), Fα1 1 (Ω) θ,q = Bα0 0 (Ω), Bα1 1 (Ω) θ,q = Bα (Ω) (9.2.99)

in each of the following scenarios:

9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 543

(i) The set Ω satisfies an interior corkscrew condition and

p − 1 + < αi for i ∈ {0, 1}, if A = B,

1
n
(9.2.100)
min{p,qi } − 1 + < αi for i ∈ {0, 1}, if A = F.
1
n

(ii) The set Ω satisfies an exterior corkscrew condition and αi < 0 for i ∈ {0, 1}.
(iii) The set Ω satisfies an exterior corkscrew condition, has an Ahlfors regular
boundary, and αi < p1 for i ∈ {0, 1}.
Proof The approach used in the proof of Theorem 9.2.18 and Corollary 9.2.20 also
works in the context of the real method of interpolation, thanks to Theorem 9.1.7
and the fact that all basic ingredients as before are still in available. 
Finally, the theorem below contains the versions of the real interpolation results
from Theorem 9.1.7 for the scales of Besov and Triebel-Lizorkin spaces in Rn with
support in a fixed subset of the Euclidean ambient.
Theorem 9.2.23 Suppose Ω ⊂ Rn is a nonempty open set satisfying ∂Ω = ∂(Ω).
Also, fix 0 < p < ∞, 0 < q0, q1, q ≤ ∞, 0 < θ < 1, α0, α1 ∈ R with α0  α1 , and set
α := (1 − θ)α0 + θα1 . Then for every q ∈ (0, ∞] one has
p,q p,q p,q p,q p,q
Fα0,00 (Ω), Fα1,01 (Ω) θ,q = Bα0,00 (Ω), Bα1,01 (Ω) θ,q = Bα,0 (Ω) (9.2.101)

in each of the following scenarios:

(i) The set Ω satisfies an exterior corkscrew condition and

n 1
pi −1 + < αi for i ∈ {0, 1}, if A = B,
(9.2.102)
0 < pi < ∞, n min{pi ,qi } − 1 +
1
< αi for i ∈ {0, 1}, if A = F.

(ii) The set Ω satisfies an interior corkscrew condition and

αi < 0 for i ∈ {0, 1}, if A = B,

(9.2.103)
0 < pi < ∞ and αi < 0 for i ∈ {0, 1}, if A = F.

(iii) The set Ω satisfies an interior corkscrew condition, has an Ahlfors regular
boundary, and

αi < 1
pi for i ∈ {0, 1}, if A = B,
(9.2.104)
0 < pi < ∞ and αi < 1
pi for i ∈ {0, 1}, if A = F.

Proof All claims are dealt with much as in the proof of Theorem 9.2.21, now making
use of Theorem 9.1.7. 
The next topic pertains to the extension by zero from a set to the entire Eu-
clidean ambient. As a preamble, we discuss characteristics functions as multipliers.
The following proposition, which is closely related to work in [59], refines [191,
Proposition 3.19, p. 87].
544 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Proposition 9.2.24 Let Ω ⊆ Rn be an open set with an Ahlfors regular boundary.

Then the following statements are true.
p,q
(i) The characteristic function 1Ω is a multiplier for Fα (Rn ) in each of the follow-
ing cases:

(a) 0 < p < ∞, 0 < q ≤ ∞, and n 1

min{p,q } −1 + < α < p1 ,

(b) 1 < p < ∞, 1 < q < ∞, and − 1 + 1

p <α< p.
1

p,q
More generally, 1Ω is a multiplier for Fα (Rn ) if the point α, p1 , q1 belongs to R,
the interior of the convex hull in R3 of the sets of points with coordinates (0, 0, 0),
(−1, 0, 0), (0, 1, 0), (0, 0, 1), (0, 1, 1), (−1, 0, 1), n−1
n
, n−1
n
, n−1
n n
, n−1 , n−1
n
, 0 as
well as tothe interior of the  base of R which corresponds to q = ∞, 0 < p <∞
and max n p1 − 1 , p1 − 1 < α < p1 .
p,q
(ii) The characteristic function 1Ω is a multiplier for Bα (Rn ) whenever
 
0 < p < ∞, 0 < q ≤ ∞, max p1 − 1, n p1 − 1 < α < p1 . (9.2.105)

Proof In view of [133, (8.7.5) in Proposition 8.7.1] we may invoke [59, Theo-
p,q
rem 13.3, p. 139] and conclude that 1Ω is a multiplier for Fα (Rn ) whenever α, p, q
satisfy the conditions stated in (a) above. In particular,
p,q
1Ω is a multiplier for Fα (Rn ) whenever
(9.2.106)
1 < p < ∞, 1 < q < ∞, and 0 < α < p.
1

∗ p,q
By duality, since Fα (Rn ) = F−α (Rn ) with 1/p + 1/p = 1, 1/q + 1/q  = 1 (cf.
p,q
p,q
Theorem 9.1.5), from (9.2.106) we further obtain that 1Ω is a multiplier for Fα (Rn )
whenever
1 < p < ∞, 1 < q < ∞, −1 + p1 < α < 0. (9.2.107)
This takes care of the conditions formulated in (b).
Next, observe that

α= 1
p = n
n−1 and 0 < 1
q < n
n−1 (9.2.108)

is a limiting case of points satisfying (a), while

1 < p < ∞, 1 < q < ∞, −1 + 1

p <α<0 (9.2.109)

is a sub-region of (b). The values of α, p, q corresponding to the case when points of

coordinates α, p1 , q1 are one of the eight vertices (0, 0, 0), (−1, 0, 0), (0, 1, 0), (0, 0, 1),
(0, 1, 1), (−1, 0, 1), n−1
n
, n−1
n
, n−1
n n
and n−1 , n−1
n
, 0 are all contained in (9.2.108),
or (9.2.109). Hence, by complex interpolation (cf. (9.1.36)), we obtain the region R
specified in the statement of the theorem.
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 545
p,q
Finally, the fact that the characteristic function 1Ω is a multiplier for Bα (Rn )
whenever α, p, q satisfy the conditions stipulated in (9.2.105), follows from what we
proved so far via real interpolation. The proof is therefore complete. 

Here is the extension result (by zero from a given subset to the entire Euclidean
ambient) alluded to earlier.

Proposition 9.2.25 Let Ω ⊆ Rn be an open set with a compact Ahlfors regular

boundary. With A ∈ {B, F}, assume that

0 < p < ∞, min{1, p} ≤ q < ∞, and
  if A = F, (9.2.110)
max p1 − 1, (n − 1) p1 − 1 < s < p1
 
0 < p, q < ∞, max p1 − 1, (n − 1) p1 − 1 < s < p1 if A = B. (9.2.111)

In particular, (9.2.54) guarantees that Cc∞ (Ω) is dense in As (Ω).

p,q

Then the extension by zero Cc∞ (Ω)  ϕ → ϕ# ∈ Cc∞ (Rn ) induces a unique bounded
linear mapping in the context
p,q p,q
EΩ→R
zero
n : As (Ω) −→ As,0 (Ω) (9.2.112)
p,q
such that (with I denoting the identity operator on As (Ω))
p,q
R Rn →Ω ◦ EΩ→R
zero
n = I on As (Ω). (9.2.113)

As a corollary of (9.2.112)-(9.2.113) and (9.2.54) one therefore has

p,q p,q p,q
Ås (Ω) = As (Ω) = As,z (Ω), A ∈ {B, F},
(9.2.114)
whenever p, q, s are as in (9.2.110)-(9.2.111).

Proof Let A, p, q, s be as in (9.2.110)-(9.2.111). Define

p,q p,q
EΩ→R
zero
n : As (Ω) −→ As,0 (Ω) (9.2.115)

as follows. Given u ∈ As (Ω), there exists w ∈ As (Rn ) with w Ω = u. By Propo-
p,q p,q
p,q
sition 9.2.24, we have 1Ω w ∈ As (Rn ) plus an appropriate quasi-norm estimate.
Define
p,q
EΩ→R
zero
n u := 1Ω w ∈ As,0 (Ω). (9.2.116)
Checking that the definition of EΩ→R
zero
n u is independent of the choice of the extension
w of u comes down (by linearity) to showing that

w ∈ As (Rn ) and w Ω = 0 in D (Ω) =⇒ 1Ω w = 0 in D (Rn ).
p,q
(9.2.117)

However, since under the current hypotheses supp 1Ω w ⊆ ∂Ω, Lemma 9.2.10
applies and yields the desired conclusion. Thus, EΩ→R
zero
n is unambiguously defined.
In turn, this may be used to check that said operator is linear and bounded.
546 9 Besov and Triebel-Lizorkin Spaces in Open Sets

From definitions, it is also clear that for any test function ϕ ∈ Cc∞ (Ω) we have
EΩ→R
zero ϕ = ϕ
n # where tilde denotes the extension by zero to the entire Rn . In particular,
R Rn →Ω ◦ EΩ→R ∞ p,q
n = I on Cc (Ω) and since the latter is a dense subspace of As (Ω)
zero

(cf. (9.2.54)) we conclude that (9.2.113) holds. Finally, (9.2.114) is an immediate

consequence of what we have proved so far and definitions. 
As regards the issue of density of test functions into the spaces introduced in
(9.2.1)-(9.2.4), we first note the following result.

Proposition 9.2.26 Let Ω be a nonempty, open, proper subset of Rn satisfying an

exterior corkscrew condition. Then for A ∈ {B, F} one has

$
C ∞ p,q
c (Ω) ⊆ As,0 (Ω) densely, (9.2.118)

Cc∞ (Ω) ⊆ As,z (Ω) densely,

p,q
(9.2.119)
p,q p,q
As,z (Ω) → Ås (Ω) continuously and densely, (9.2.120)

where tilde denotes the extension by zero outside Ω, whenever

1 < p < ∞, 0 < q < ∞, and 0 < s if A = B, (9.2.121)

1 < p < ∞, 0 < q < ∞, and n 1

min{p,q } −1 + < s if A = F. (9.2.122)

Proof See [19, Theorem 8.6]. 

Turning to duality results, we first note the following proposition.

Proposition 9.2.27 Let Ω be a nonempty, open, proper subset of Rn . Fix s ∈ R \ {0}

along with p, q ∈ (1, ∞) and denote by p, q  the Hölder conjugate exponents of p, q.
Finally, let A ∈ {B, F}.
(i) If ∂Ω = ∂(Ω) and Ω satisfies an interior corkscrew condition then
p,q ∗ p,q
As (Ω) = A−s,0 (Ω) if s > 0, (9.2.123)
p,q ∗ p,q
As,0 (Ω) = A−s (Ω) if s < 0, (9.2.124)

(ii) If Ω satisfies an exterior corkscrew condition then

p,q ∗ p,q
Bs (Ω) = B−s,z (Ω) if s < 0, (9.2.125)
p,q ∗ p,q
Bs,z (Ω) = B−s (Ω) if s > 0, (9.2.126)

with the duality pairings understood as extensions of the duality pairings between
D(Ω) and D (Ω).
(iii) If Ω satisfies an exterior corkscrew condition and ∂Ω is Ahlfors regular then
p,q ∗ p,q
B0 (Ω) = B0 (Ω) (9.2.127)
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 547

with the duality pairing understood as an extension of the duality pairings between
D(Ω) and D (Ω).
p,q ∗
Proof Note that if Λ ∈ As (Ω) then the composition Λ ◦ R Rn →Ω is a func-
p,q ∗ p,q
tional belonging to As (Rn ) = A−s (Rn ) and satisfies supp (Λ ◦ R Rn →Ω ) ⊆ Ω.
Ultimately, we conclude that the mapping
p,q ∗ p,q
Φ : As (Ω) −→ A−s,0 (Ω) given by
p,q ∗
(9.2.128)
Φ(Λ) := Λ ◦ R Rn →Ω for each Λ ∈ As (Ω)

is well defined, linear, and bounded. Next, define

p,q p,q ∗
Ψ : A−s,0 (Ω) −→ As (Ω) by setting
p,q
Ψ(u) := Λu for each u ∈ A−s,0 (Ω),
where
% & (9.2.129)
Λu (w) := Ap, q (Rn ) u, W Ap, q n for all
s (R )
−s

w ∈ As (Ω) and W ∈ As (Rn ) with W Ω = w.
p,q p,q

p,q
For a given w ∈ As (Ω), the definition of Λw above does not depend on the choice
p,q p,q
W ∈ As (R ) of w. Indeed, assume W1, W2 ∈ As (R ) are such
of the extension n n

that W1 Ω = W2 Ω . Given that we are presently assuming ∂Ω = ∂(Ω), guarantees
p,q
Rn \ Ω = Rn \ Ω, hence W1 − W2 ∈ As,0 Rn \ Ω . Since the fact that Ω satisfies
an interior corkscrew condition implies that Rn \ Ω satisfies an exterior corkscrew
condition (cf. [133, Definition 5.1.3] and [133, (5.1.9)]), we may invoke (9.2.118) to
conclude that there exists a sequence {φ j } j ∈N ⊆ Cc∞ Rn \ Ω with the property that
{ φ#j } j ∈N converges to W1 − W2 in As (Rn ). As such, we may compute
p,q

% %
A
p , q 
(R n )
u, W1  Ap, q n − p , q 
s (R ) A (R n )
u, W2  Ap, q n
s (R )
−s −s
%
= p , q 
A−s (R n )
u, W1 − W2  Ap, q n
s (R )

%
= lim p , q 
(R n )
u, φ#j  Ap, q n
s (R )
j→∞ A−s

= lim 0 = 0, (9.2.130)
j→∞

as wanted. Ultimately, Ψ is well defined, linear, and bounded. We next claim that
p,q ∗
Φ and Ψ are inverse to each other. To justify this claim, pick
 Λ ∈ As (Ω) and
w ∈ As (Ω), and choose some W ∈ As (Rn ) such that W Ω = w. Then
p,q p,q
548 9 Besov and Triebel-Lizorkin Spaces in Open Sets
% & % &
p, q
(A s (Ω))∗ (Ψ ◦ Φ)(Λ), w p, q
As (Ω) = (Ap, q
s (Ω))
∗ Ψ(Φ(Λ)), w A p, q (Ω)
s
% &
= Ap, q (Rn ) Φ(Λ), W Ap, q
s (R )
n
−s
% &
= p , q 
A−s (R n )
Λ ◦ R Rn →Ω, W p, q
As (R n )
%  &
= (Ap, q ∗ Λ, W
 p, q
s (Ω)) Ω A s (Ω)
% &
= (Ap, q
s (Ω))
∗ Λ, w A p, q (Ω),
s
(9.2.131)

p,q p,q
hence (Ψ ◦ Φ)(Λ) = Λ. In the opposite direction, if u ∈ A−s,0 (Ω) and W ∈ As (Rn )
we may write
% & % &
A
p , q 
(R n )
Φ(Ψ(u)), W Ap, q n =
s (R ) A
p , q 
(R n )
Ψ(u) ◦ R Rn →Ω, W Ap, q n
s (R )
−s −s
%  &
= (Ap, q ∗ Ψ(u), W
 p, q
s (Ω)) Ω A s (Ω)
% &
= Ap, q (Rn ) u, W Ap, q n ,
s (R )
(9.2.132)
−s

which goes to show that Φ(Ψ(u)) = u. All together, this proves that Φ and Ψ are
inverse to each other, so (9.2.123) is established at this point.
p,q ∗
As far as (9.2.124) is concerned, let Λ ∈ As,0 (Ω) be an arbitrary functional
and consider any
p,q
f ∈ A−s (Rn ) which satisfies
% & p,q
(9.2.133)
Λ(u) = Ap, q (Rn ) f , u Ap, q n for all u ∈ A
s (R ) s,0 (Ω).
−s

p,q p,q
Since As,0 (Ω) is a closed subspace of the Banach space As (Ω), the Hahn-Banach
p,q
 ∈ Asp,q (Rn ) ∗ = A−s
Theorem guarantees the existence of an extension Λ (Rn ) of
 
Λ p, q n = Λ p, q ∗ . Hence, some f as in (9.2.133)
Λ with the property that  A−s (R ) (A s,0 (Ω))
always exists with the additional property that  f  Ap, q (Rn ) ≈ Λ(Ap, q (Ω))∗ .
−s s,0
 p,q

Note that if f Ω = 0 then supp f ⊆ R \Ω = R \ Ω, hence f ∈ A−s,0 Rn \Ω . As
n n

such, (9.2.118) ensures the existence a sequence of functions {φ j } j ∈N ⊆ Cc∞ Rn \Ω

p,q
such that { φ#j } j ∈N converges to f in A−s (Rn ). Therefore, given any u ∈ As,0 (Ω)
p,q

p,q
(hence, u ∈ As (Rn ) with supp u ⊆ Ω) we may write
% & % &
A
p , q 
(R n )
f , u Ap, q n = lim
s (R ) A
p , q 
(R n )
φ#j , u Ap, q n = lim 0 = 0.
s (R )
(9.2.134)
−s j→∞ −s j→∞

The argument so far proves that the assignment

∗  p,q
As,0 (Ω)  Λ −→ f Ω ∈ A−s (Ω)
p,q
(9.2.135)
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 549

is well defined, linear, bounded, and injective. To prove that this is also surjective,
p,q p,q
 an arbitrary w ∈ A−s (Ω). Then there exists some W ∈ A−s (R ) such that
pick n

W Ω = w. If we then define
p,q ∗
Λw ∈ As,0 (Ω) by setting
% & p,q (9.2.136)
Λw (u) := Ap, q (Rn ) W, u Ap, q n for every u ∈ A
s (R ) s,0 (Ω),
−s


it follows from (9.2.133) that the assignment (9.2.135) maps Λw into W Ω = w.
This establishes the surjectivity of (9.2.135) which, ultimately, finishes the proof of
(9.2.124).
Finally, the assertion in part (ii) is a consequence of [191, Theorem 3.30, p. 98],
while the claim in part (iii) is implied by [191, Corollary 3.32, p. 99]. 

Corollary 9.2.28 (i) Let Ω ⊆ Rn be an open set which is n-thick, has an Ahlfors
regular boundary, and satisfies ∂Ω = ∂(Ω). Fix p, q ∈ (1, ∞), denote by p, q 
their Hölder conjugate exponents, and suppose s ∈ − 1 + p1 , p1 \ {0}. Finally, let
A ∈ {B, F}. Then
p,q p,q
the spaces As (Ω) and A−s (Ω) are dual to one another. (9.2.137)

(ii) Assume Ω ⊆ Rn is an open set satisfying an exterior corkscrew condition and

with an Ahlfors regular boundary. Pick p, q ∈ (1, ∞), denote by p, q  their Hölder
conjugate exponents, and select s ∈ − 1 + p1 , p1 . Then

p,q p,q
the spaces Bs (Ω) and B−s (Ω) are dual to one another. (9.2.138)

Proof Under the assumptions in part (i), [133, Proposition 8.6.12] ensures that Ω
satisfies an interior corkscrew condition. Granted this, the claim in (9.2.137) follows
by combining part (i) of Proposition 9.2.27, the last claim in Lemma 9.2.11, and
(9.2.114).
Under the assumptions made in the current part (ii), the claim in (9.2.138) becomes
a consequence of parts (ii)-(iii) in Proposition 9.2.27, the last claim in Lemma 9.2.11,
and (9.2.114). 

It is useful to further augment Proposition 9.2.27 and Corollary 9.2.28 with the
following duality result.

Proposition 9.2.29 Let Ω ⊆ Rn be an open set with a compact Ahlfors regular

boundary. Fix p, q ∈ (1, ∞), denote by p, q  their Hölder conjugate exponents, and
pick s ∈ − 1 + p1 , p1 . Also, let A ∈ {B, F}. Then the assignment

p,q p,q ∗
As (Ω)  u −→ Λu ∈ A−s (Ω) where
' (
Λu (w) := Ap,
s
q
(R n) E
zero
Ω→R n u, E zero w
Ω→R n
A
p , q 
(R n )
(9.2.139)
−s
p,q p,q
for every u ∈ As (Ω) and every w ∈ A−s (Ω),
550 9 Besov and Triebel-Lizorkin Spaces in Open Sets

is a well-defined, linear, bounded, and injective mapping. As such,

p,q p,q ∗
As (Ω) may be embedded into A−s (Ω)
(9.2.140)
(via u −→ Λu ).

In particular,
p,q p,q
for any two distributions,
% & u ∈ As (Ω) and w ∈ A−s (Ω), the
pairing Ap, q
s (Ω)
u, w p , q 
A−s (Ω) (
is well defined whenever interpreted
'
as Ap, q n
s (R )
EΩ→R
zero
n u, E Ω→R n w
zero
p , q 
A−s (R n )
and, thustly defined, the
% & (9.2.141)
p,q p,q
mapping As (Ω) ⊕ A−s (Ω)  (u, w) → Ap, q
s (Ω)
u, w Ap, q (Ω)
−s
happens % to & be bilinear, continuous, and enjoy the property that
∞
A s (Ω) u, ϕ Ap , q  (Ω) = D  (Ω) u, ϕ D(Ω) for all functions ϕ ∈ Cc (Ω).
p, q
−s

Proof The fact that the assignment u −→ Λu in (9.2.139) is a well-defined, linear,
p,q
bounded mapping is clear from Proposition 9.2.25. The fact that for each u ∈ As (Ω)
p,q
we have EΩ→R
zero
n u ∈ As (Rn ) and R Rn →Ω EΩ→R zero
n u = u (cf. Proposition 9.2.25)
implies that for each ϕ ∈ Cc∞ (Ω) we may write
' ( ' (
A s (R n ) EΩ→R n u, EΩ→R n ϕ Ap , q  (R n ) = A s (R n ) EΩ→R n u, ϕ
# Ap, q (Rn )
p, q
zero zero p, q
zero
−s −s

' (
= D  (R n ) EΩ→R
zero
n u, ϕ
# D(Rn )
' (
= D  (Ω) R Rn →Ω EΩ→R
zero
n u ,ϕ D(Ω)

= D  (Ω) u, ϕ D(Ω) . (9.2.142)

Thus, for each u ∈ As (Ω) and ϕ ∈ Cc∞ (Ω) we have

p,q

' (
Λu (ϕ) = Ap,
s
q
(R n ) E Ω→R n u, E Ω→R n ϕ
zero zero
A
p , q 
(R n )
= D  (Ω) u, ϕ D(Ω) . (9.2.143)
−s

p,q
In particular, if u ∈ As (Ω) is such that Λu = 0, then D  (Ω) u, ϕ D(Ω) for each
ϕ ∈ Cc∞ (Ω) hence, ultimately, u = 0 as a distribution in Ω. This proves that the
assignment in (9.2.139) is indeed injective. Finally, the claim in (9.2.141) is implicit
in the above argument. 
We now come to a basic regularity result. To facilitate its statement, recall the
notational convention adopted in (A.0.73). The reader is also reminded that the
notion of subaveraging function has been introduced in [133, Definition 6.5.1].
Theorem 9.2.30 Suppose Ω ⊆ Rn is an (ε, δ)-domain and suppose

k ∈ N0, 0 < p < ∞, 0 < q ≤ ∞,

(9.2.144)
and n 1
min {p,q } −1 + − k < α < 1.
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 551

k+1,1
Also, assume u ∈ Wloc (Ω) has the property that

|∇k+1 u| is subaveraging in Ω (9.2.145)

(cf. [133, (6.5.13)]) and, with δ∂Ω abbreviating dist(·, ∂Ω),

δ∂Ω |∇ u| + |u| ∈ L p (Ω, L n ).

1−α k+1
(9.2.146)

Then there exists a constant C = C(Ω, p, q, α, k) ∈ (0, ∞) such that

p,q p,max{p,q }
u ∈ Fα+k (Ω), u ∈ Bα+k (Ω), and
   1−α k+1 
max uF p, q (Ω), uB p,max{ p, q} (Ω) ≤ C δ∂Ω |∇ u|  L p (Ω, L n ) + Cu L p (Ω, L n ) .
α+k α+k
(9.2.147)

Proof From [172, Corollary 1, p. 398] (cf. also the discussion in [150, pp. 274–275])
we know that if

0 < p < ∞, 0 < q ≤ ∞, n 1

min {p,q } −1 + < s,
(9.2.148)
and m is an integer such that m > s
m,1
then for a given function w ∈ Wloc (Ω) we have (with oscm−1 (w, x, t) as defined in
(9.1.64))
 
 ∫ 1 δ∂Ω (·) 1/q 
 2  m−1  q dt 
w Fsp, q (Ω) ≈ w  L p (Ω, L n ) +  osc (w, ·, t) 1+sq

 .
 0 t  p n
L (Ω, L )
(9.2.149)
Let us momentarily digress to shed more light on Seeger’s work in [172]. M. Christ
p
has proved in [28] extension theorems a variant of the spaces Cα introduced in
(9.2.29)-(9.2.30) (see [28, Definition 2.1, p. 66]). One feature of this scale is that for
integer order of smoothness it yields classical Sobolev space (cf. [28, Lemma 2.2,
p. 66]). As such, M. Christ’s work generalizes P. Jones’s results from [103] which
deal precisely with the latter scale. Essentially, Seeger’s observation is that since
for non-integer values of the parameter α the scale considered in M. Christ’s work
p p,∞
[28] agrees with Cα hence ultimately with Fα in (ε, δ)-domains (cf. (9.2.34)),
this suggests that similar results should hold for all Triebel-Lizorkin spaces. Indeed,
it turns out that the Jones-Christ extension results do hold for this more general
scale, and this is one of the main points of [172]. The strategy is to re-work Christ’s
argument replacing the usual maximal theorem (appropriate for q = ∞) by vector
valued (Fefferman-Stein) maximal inequalities. The proof relies on the Jones idea
of reflecting Whitney cubes. One has a certain freedom in constructing Whitney
cubes; for example we may, of course, make the side-length of any such cube be
much smaller than its distance to the boundary (say, by a factor 2−10 ). In particular,
the coefficient 12 appearing in the upper limit of integration in (9.2.149) can be
harmlessly replaced by any fixed small constant c > 0.
552 9 Besov and Triebel-Lizorkin Spaces in Open Sets

After this digression, we now return to the main topic of conversation. In order
to estimate the (m − 1)-th order oscillation of u, we shall employ a higher-order
Poincaré-type inequality, proved in [45, Theorem 3.4, p. 18], to the effect that there
exists some finite constant C = C(n, m) > 0 with the property that for any ball BR of
radius R > 0 and any function w ∈ W m,1 (BR ) one can find a polynomial P ∈ Pm−1
(cf. (9.1.63)) with the property that
∫  ∫
|w − P| dL n ≤ C Rm |∂γ w| dL n . (9.2.150)
BR |γ |=m BR

Given p, q, α, k as in the statement of the theorem, choose m := k+1 and s := k+α.

From (9.1.64), (9.2.150), and (9.2.145) we then conclude that for any x ∈ Ω and
t ∈ 0, 12 δ∂Ω (x) we have
⨏
osck (u, x, t) ≤ Ct k+1 |∇k+1 u| dL n
B(x,t)
   
≤ C t k+1 sup |∇k+1 u| ≤ C t k+1 sup |∇k+1 u|
B(x,t) B(x,δ ∂Ω (x)/2)
⨏  1/q
≤ Ct k+1
|∇ k+1 q
u| dL n
. (9.2.151)
B(x,7δ ∂Ω (x)/8)

Based on (9.2.149) and (9.2.151) we may write

 
 ∫ 1 δ∂Ω (·) 1/q 
  k q dt 

2
osc (u, ·, t) 
 1+(k+α)q 
 0 t 
L p (Ω, L n )
 
 ∫ 1 δ∂Ω (x) )⨏ * 1/q 
 2 dt 

≤ C |∇ u(y)| dy 1+q(α−1)
k+1 q 

 0 B(x,7δ ∂Ω (x)/8) t  p
L x (Ω, L n )
 ⨏  1/q 

 
≤ C δ∂Ω (x) 1−α
|∇ u(y)| dy
k+1 q
 ,
 B(x,7δ ∂Ω (x)/8)  p
L x (Ω, L n )
(9.2.152)

where the subscript x in the last line above indicates that the outer L p -norm is taken
in the variable x ∈ Ω. By [133, Lemma 6.5.5], the last expression above is further
dominated by
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 553
 ⨏  1/p 

 
C δ∂Ω (x) 1−α
|∇ u(y)| dy
k+1 p
 (9.2.153)
 B(x,8δ ∂Ω (x)/9)  p
L x (Ω, L n )
∫ ∫  1/p
≤C δ∂Ω (x) p(1−α)−n |∇k+1 u(y)| p 1 |x−y | ≤8δ∂Ω (x)/9 dxdy .
Ω Ω

Next, let x ∈ Ω be arbitrary and pick x ∗ ∈ ∂Ω such that |x ∗ − x| = δ∂Ω (x). Then, for
each y ∈ Ω with |x − y| ≤ 8δ∂Ω (x)/9, we have

δ∂Ω (y) ≤ |y − x ∗ | ≤ |x − y| + |x − x ∗ |
≤ 8δ∂Ω (x)/9 + δ∂Ω (x) = 17δ∂Ω (x)/9. (9.2.154)

Moreover, if y ∈ Ω has |x − y| ≤ 8δ∂Ω (x)/9, pick y ∗ ∈ ∂Ω such that δ∂Ω (y) = |y − y ∗ |

and estimate

δ∂Ω (y) = |y − y ∗ | ≥ |x − y ∗ | − |x − y| ≥ δ∂Ω (x) − 8δ∂Ω (x)/9

= δ∂Ω (x)/9. (9.2.155)

Thus, altogether, |x − y| ≤ 8δ∂Ω (x)/9 implies 9δ∂Ω (y)/17 ≤ δ∂Ω (x) ≤ 9δ∂Ω (y).
Availing ourselves of this back in (9.2.153) yields
∫ ∫  1/p
δ∂Ω (x) p(1−α)−n |∇k+1 u(y)| p 1 |x−y | ≤8δ∂Ω (x)/9 dxdy
Ω Ω
∫ ∫  1/p
≤C δ∂Ω (y) p(1−α)−n
|∇ k+1 p
u(y)| 1 |x−y | ≤8δ∂Ω (y) dxdy
Ω Ω
∫  1/p
≤C δ∂Ω (y) p(1−α)
|∇ k+1 p
u(y)| dy . (9.2.156)
Ω

By putting together (9.2.149), (9.2.152), (9.2.153), and (9.2.156) we finally arrive at

the conclusion
p,q
u belongs to the Triebel-Lizorkin space Fα+k (Ω) and
 1−α k+1  (9.2.157)
uF p, q (Ω) ≤ C δ∂Ω |∇ u|  L p (Ω, L n ) + Cu L p (Ω, L n )
α+k

for some constant C ∈ (0, ∞) independent of u.

The remaining claims in (9.2.147) are consequences of (9.2.157) and (9.2.11).
This finishes the proof of Theorem 9.2.30. 
Recall the scale of homogeneous maximal weighted Sobolev spaces introduced
in the second part of Definition 8.6.1. Also, recall the piece of notation introduced in
(8.6.6). The main point of our next result is that null-solutions of an elliptic system
belonging to a weighted homogeneous maximal Sobolev space and also which satisfy
a mild integrability condition, exhibit regularity which may be measured optimally on
554 9 Besov and Triebel-Lizorkin Spaces in Open Sets

the Triebel-Lizorkin and Besov scales. For related results, in the opposite direction,
see also Corollary 9.2.38.

Corollary 9.2.31 With M, m ∈ N, let L be a constant (complex) coefficient homoge-

neous M × M system of order 2m in Rn , with the property that det [L(ξ)]  0 for each
ξ ∈ Rn \ {0}. Also, assume Ω ⊆ Rn is an (ε, δ)-domain and suppose k, p, s, q, a ∈ R
are such that
 
k ∈ N0, 0 < p < ∞, max (n − 1) p1 − 1 , 1 − p1 < s + k + 1,
(9.2.158)
n+k+s+1/p < q ≤ ∞, and a := 1 − s − p ≤ 1.
n 1

Then one has the continuous embeddings

 . k+1, p M  M
Wa, (Ω) ∩ Ker L ∩ L p (Ω, L n ) (9.2.159)
 p,q M  p,max{p,q }  M
→ F (Ω) → B (Ω) .
k+s+ p1 1
k+s+ p

In particular, if

k ∈ N0, 1 < p < ∞, 0 < s < 1,

(9.2.160)
n
n+k+s+1/p < q ≤ ∞, a := 1 − s − p1 ,

then
 k+1, p M  M
Wa (Ω) ∩ Ker L ∩ L p (Ω, L n ) (9.2.161)
 p,q M  p,max{p,q }  M
→ F (Ω) → B (Ω) .
k+s+ p1 1
k+s+ p

Proof Analyzing the cases when p ≤ q and p > q separately shows that if k, p, s, q, a
are as in (9.2.158) then

n 1
min {p,q } −1 + < k + s + p1 . (9.2.162)

Let [·] denote the integer part function. Since 1 − p1 < s + k + 1 it follows that
   
0 < s + k + p1 , hence 0 ≤ s + k + p1 = k + s + p1 . In turn, this implies
 
k o := k + s + p1 ∈ N0 and
  (9.2.163)
α := k + s + p1 − k o = s + p1 − s + p1 ∈ [0, 1).

Together, (9.2.158), (9.2.162), and (9.2.163) ensure that

k o ∈ N0, 0 < p < ∞, 0 < q ≤ ∞, and

(9.2.164)
n 1
min {p,q } −1 + − k o < α < 1.
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 555
 . k+1, p M  M
Next, pick an arbitrary function u ∈ Wa, (Ω) ∩ Ker L ∩ L p (Ω, L n ) .
Then [133, Theorem 6.5.7] implies that
 M
u ∈ C ∞ (Ω) and |∇ko +1 u| is subaveraging in Ω. (9.2.165)

Let us also observe that the assumption a ≤ 1 forces s+ p1 ≥ 0, hence k ≥ k o . As such,

having fixed some parameter θ ∈ (0, 1), we conclude from [136, Corollary 4.2.2]
that there exists some finite constant C = C(Ω, L, p, s, k, θ) > 0 such that
   
 1−α  ko +1   ko −k+a  ko +1 
δ∂Ω ∇ u p n
= δ ∂Ω
∇ u  p n
(9.2.166)
L (Ω, L ) L (Ω, L )
 
 a  k+1  
≤ C δ∂Ω ∇ u ,θ 
= CuW. k+1, p (Ω) .
L p (Ω, L n ) a,

Granted (9.2.164), (9.2.165), (9.2.166), and the fact that Ω is an (ε, δ)-domain,
Theorem 9.2.30 guarantees that
 p,q  M  p,q M
u ∈ Fα+ko (Ω) = F (Ω) and
k+s+ p1
(9.2.167)
u[F p, q (Ω)] M ≤ Cu[W. k+1, p (Ω)] M + Cu[L p (Ω, L n )] M .
k+s+1/p a,

This establishes the first embedding in (9.2.159). The second embedding in (9.2.159)
is a consequence of (9.2.11). Finally, (9.2.161) is a consequence of (9.2.159) and
Lemma 8.6.2. 

We are now in a position to obtain local embedding results of null-solutions of

a (higher-order) weakly elliptic system belonging to weighted Sobolev spaces in
(ε, δ)-domains into Besov and Triebel-Lizorkin spaces, of the sort discussed in the
next corollary.

Corollary 9.2.32 Fix M, m, n ∈ N with n ≥ 2. Let L be a constant (complex)

coefficient homogeneous M × M system of order 2m in Rn , with the property that
det [L(ξ)]  0 for each ξ ∈ Rn \ {0}. Also, assume Ω ⊆ Rn is an (ε, δ)-domain with
compact boundary and suppose k, p, s, q, a ∈ R are such that
 
k ∈ N0, 0 < p < ∞, max (n − 1) p1 − 1 , 1 − p1 < s + k + 1,
 (9.2.168)
n
n+k+s+1/p < q ≤ ∞, and a := 1 − s − 1
p ∈ − 1
p , 1 .

In this setting, consider a function

 p M
u ∈ Lbdd (Ω, L n ) ∩ Ker L such that
 k+1, p M (9.2.169)
ψu ∈ Wa (Ω) for each ψ ∈ Cc∞ (Rn ).

Then, in a quantitative fashion,

556 9 Besov and Triebel-Lizorkin Spaces in Open Sets
 p,q M  p,max{p,q }  M
ψu ∈ F (Ω) → B (Ω)
k+s+ p1 1
k+s+ p (9.2.170)
for each cutoff function ψ ∈ Cc∞ (Rn ).

Proof For each sufficiently large R > 0 define ΩR := Ω∩B(0, R). Then Lemma 8.6.4
(whose applicability in the present setting is ensured by [133, Theorem 6.5.7])
guarantees that
  k+1, p M  p M
uΩ R ∈ Wa, (ΩR ) ∩ L (ΩR, L n ) ∩ Ker L
(9.2.171)
for each sufficiently large R > 0.

Granted this, Corollary 9.2.31 implies (upon noting that ΩR continues to be an

(ε, δ)-domain whenever R > 0 is large enough) that
  p,q M
u Ω R ∈ F (ΩR ) for each sufficiently large R > 0. (9.2.172)
k+s+ p1

 p,q M
Hence (cf. (9.2.1)), for each R > 0 large enough there exists UR ∈ F (Rn )
k+s+ p1
with the property that
 
UR Ω R = uΩ R and
 (9.2.173)
UR [F p, q 
(R n )] M ≤ 2u Ω R [F (Ω R )] M .
p, q
1
k+s+ p 1
k+s+ p

As a consequence, given any cutoff function ψ ∈ Cc∞ (Rn ) we have

 
(ψu)Ω R = (ψUR )Ω R for each sufficiently large R > 0. (9.2.174)

Since ψu and (ψUR )Ω , viewed as distributions in Ω, also coincide in Ω \ supp ψ (as
 
they both vanish there), and since ΩR, Ω \ supp ψ constitutes an open covering of
Ω if R is large enough, we ultimately conclude that

ψu = (ψUR )Ω as distributions in Ω,
(9.2.175)
for each sufficiently large R > 0.
 p,q M
Given that each ψUR belongs to F (Rn ) we then deduce from (9.2.175) and
k+s+ p1
(9.2.1) that
 p,q M
ψu ∈ F 1 (Ω) for each ψ ∈ Cc∞ (Rn ). (9.2.176)
k+s+ p

From this, (9.2.170) follows (with the inclusion justified as in (9.2.159)). 

It turns out that for null-solutions of an elliptic system defined in an open set Ω,
a Gagliardo semi-norm closely related to Besov spaces in Ω controls the semi-norm
associated with the homogeneous weighted maximal Sobolev spaces in Ω introduced
in the second part of Definition 8.6.1. This is made precise in the proposition below.
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 557

Proposition 9.2.33 For M, m ∈ N, let L be a constant (complex) coefficient homo-

geneous M × M system of order 2m in Rn , with the property that det [L(ξ)]  0 for
each ξ ∈ Rn \ {0}. Let Ω be an arbitrary, open, nonempty, proper subset of Rn . Fix
p ∈ (0, ∞), k ∈ N0 , α > −k − np , and θ ∈ (0, 1). Then there exists a finite constant
C = C(Ω, L, p, k, α, θ) > 0 with the property that for any vector-valued function
 M
u ∈ C ∞ (Ω) satisfying Lu = 0 in Ω one has

 k+1  ∫ ∫ |u(x) − u(y)| p  p1

 |∇ u| ,θ
 (1−α)p n ≤C dx dy . (9.2.177)
L p (Ω,δ ∂Ω L ) |x − y| n+(k+α)p
Ω Ω

Moreover, whenever p ∈ (0, ∞), k ∈ N0 , α > − np , and θ ∈ (0, 1), there exists a
finite constant C = C(Ω, L, p, k, α, θ) > 0 with the property that for any vector-valued
 M
function u ∈ C ∞ (Ω) satisfying Lu = 0 in Ω one has
 k+1 
 |∇ u| 
,θ L p (Ω,δ (1−α)p L n ) (9.2.178)
∂Ω

  ∫ ∫ |(∂γ u)(x) − (∂γ u)(y)| p  p1

≤C dx dy .
Ω Ω |x − y| n+αp
|γ |=k

Proof Fix some ε ∈ (0, 1 − θ). By interior estimates (cf. [133, Theorem 6.5.7]), for
each x ∈ Ω we have

|∇k+1 u| ,θ (x) = sup |∇k+1 u(z)| (9.2.179)

z ∈B(x,θ δ ∂Ω (x))

⨏  p1
−k−1
≤ Cδ∂Ω (x) |u(x) − u(y)| dy
p
.
B(x,(θ+ε)δ ∂Ω (x))

Hence, since n + k p + αp > 0, we further obtain

δ∂Ω (x)(1−α)p |∇k+1 u| ,θ (x)

p

∫
−(n+αp+k p)
≤ Cδ∂Ω (x) |u(x) − u(y)| p dy
B(x,(θ+ε)δ ∂Ω (x))
∫
|u(x) − u(y)| p
≤C dy
B(x,(θ+ε)δ ∂Ω (x)) |x − y| n+αp+k p
∫
|u(x) − u(y)| p
≤C dy. (9.2.180)
Ω |x − y| n+αp+k p
Now (9.2.177) follows from (9.2.180) via integration in x over Ω.
 M
Finally, if α > − np , then given any u ∈ C ∞ (Ω) ∩ Ker L along with k ∈ N0 ,
writing the version of (9.2.177) corresponding to k := 0 for each ∂γ u with γ ∈ N0n
with |γ| = k and summing up the resulting estimates yields (9.2.178). 
558 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Our next two results contain intrinsic characterizations of some Besov spaces
defined in certain open subsets of the Euclidean ambient.
Proposition 9.2.34 Assume Ω is an arbitrary open subset in Rn and suppose that
0 < p < ∞ and that n p1 − 1 + < α < 1. Also, fix k ∈ N0 . Then there exists a finite
p, p
constant C > 0 such that for each function u ∈ Bα+k (Ω) one has
  ∫ ∫ |(∂γ u)(x) − (∂γ u)(y)| p  p1
u L p (Ω, L n ) + dx dy ≤ CuB p, p (Ω) .
Ω Ω |x − y| n+αp α+k
|γ |=k
(9.2.181)
Moreover, in the case when Ω is an (ε, δ)-domain in Rn and k = 0, the converse
inequality also holds. Consequently, in the situation when Ω is an (ε, δ)-domain in
Rn and under the assumption that 0 < p < ∞ and n p1 − 1 + < α < 1, one has

 ∫ ∫ |u(x) − u(y)| p  p1
uBαp, p (Ω) ≈ u L p (Ω, L n ) + dx dy , (9.2.182)
Ω Ω |x − y|
n+αp

uniformly in u ∈ Lloc
1 (Ω, L n ).

p, p p, p
Proof Fix an arbitrary u ∈ Bα+k (Ω). Then there exists a function w ∈ Bα+k (Rn )
with the property that w Ω = u and such that w B p, p (Rn ) ≤ 2uB p, p (Ω) . In concert
α+k α+k
with (9.1.15), this allows us to estimate
 ∫ ∫ |(∂γ u)(x) − (∂γ u)(y)| p
dx dy
Ω Ω |x − y| n+αp
|γ |=k

 ∫ ∫ |(∂γ w)(x) − (∂γ w)(y)| p

= dx dy
Ω Ω |x − y| n+αp
|γ |=k

∫ ∫
|(∂γ w)(x) − (∂γ w)(y)| p
≤ dx dy
Rn Rn |x − y| n+αp
|γ |=k

p p
≤ C∇k w B p, p (Rn ) ≤ Cw B p, p (Rn )
α α+k

p
≤ CuB p, p (Ω) . (9.2.183)
α+k

Since also u L p (Ω, L n ) ≤ uB p, p (Ω) , we conclude that (9.2.181) holds.
α+k
Consider now the opposite inequality in (9.2.181), under the assumptions that
the set Ω is an (ε, δ)-domain in Rn and k = 0. In such a scenario, from (9.2.148)-
(9.2.149) (presently used with w := u, m := 1, q := p ∈ n+1 n
, ∞ , and s := α) we
see that

∫  δ∂Ω
∫(x)/2
 0  p dt 
p
uB p, p (Ω) ≤C  osc (u, x, t)  dx + Cu p p , (9.2.184)
α
 t 1+αp  L (Ω, L n )
x ∈Ω  0 
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 559

where osc0 was introduced in (9.1.64). In turn, the double integral in the right-hand
side of (9.2.184) is successively bounded by

∫ ∫ ∞  ∫  dt 
 1
C  n
|u(x) − u(y)| dy 1+αp  dx
p
 0 t t 
x ∈Ω | x−y |≤t
 y ∈Ω 
∫ ∫ ∫ ∞ 
dt
≤C |u(x) − u(y)| p n+1+αp
dy dx
x ∈Ω y ∈Ω |x−y | t
∫ ∫
|u(x) − u(y)| p
≤C dx dy, (9.2.185)
Ω Ω |x − y|
n+αp

as desired. 

Corollary 9.2.35 Suppose L is a constant (complex) coefficient homogeneous M ×M

system of order 2m in Rn , where M, m, n ∈ N, with the property that det [L(ξ)]  0
for each ξ ∈ Rn \ {0}. Also, let Ω ⊆ Rn be an (ε, δ)-domain, and assume 0 < p < ∞
and n( p1 − 1)+ < α < 1. Then

 ∫ ∫ |u(x) − u(y)| p  p1
u[Bαp, p (Ω)] M ≈ u[L p (Ω, L n )] M + dx dy
Ω Ω |x − y|
n+αp

 1−α 
≈ δ∂Ω |∇u|  L p (Ω, L n ) + u[L p (Ω, L n )] M , (9.2.186)
 M
uniformly in u ∈ C ∞ (Ω) satisfying Lu = 0 in Ω.

Proof The first equivalence in (9.2.186) comes from (9.2.182). With this in hand, the
second equivalence in (9.2.186) is implied by Proposition 9.2.33 and Theorem 9.2.30
(also keeping in mind [133, (6.6.6)], and [133, (6.5.40) in Theorem 6.5.7]). 

It is also of interest to know whether results of a similar nature are valid in the
absence of a PDE, and under less stringent assumptions on the underlying domain.
This issue is addressed in the proposition below.

Proposition 9.2.36 Let Ω be an arbitrary open, nonempty, proper subset of Rn . Also,

assume that p ∈ (0, ∞), α ∈ (0, 1), and θ ∈ (0, 1). Then there exists a finite constant
1,1
C = C(Ω, p, α, θ) > 0 such that for each function u ∈ Wloc (Ω) one has
 ∬ |u(x) − u(y)| p  p1  
dx dy ≤ C  |∇u| 
,θ L p (Ω,δ (1−α)p L n )
|x − y| n+αp ∂Ω
Ω×Ω

+ Cu L p (Ω,δ −α p L n ), (9.2.187)

∂Ω

with the understanding that when p ≥ 1 the subscripts , θ may be omitted.

560 9 Besov and Triebel-Lizorkin Spaces in Open Sets

In particular, if Ω is also bounded, then

 ∬ |u(x) − u(y)| p  p1
u L (Ω, L ) +
p n dx dy
|x − y| n+αp
Ω×Ω
 
≤ C  |∇u| 
,θ L p (Ω,δ (1−α)p L n ) + Cu L p (Ω,δ −α p L n ), (9.2.188)
∂Ω ∂Ω

again, with the understanding that when p ≥ 1 the subscripts , θ may be omitted.
1,1
Proof Fix an arbitrary function u ∈ Wloc (Ω). First we claim that there exists a finite
constant C = C(Ω, p, α, θ) > 0 such that
∬ ∫
|u(x) − u(y)| p
dx dy ≤ C |u(x)| p δ∂Ω (x)−αp dx.
|x − y| n+αp Ω
(x,y)∈Ω×Ω
|x−y |> θ2 max{δ ∂Ω (x),δ ∂Ω (y)}
(9.2.189)
To justify (9.2.189), first observe that
∬ ∫  ∫ 
|u(x)| p dy
dx dy ≤ |u(x)| p dx
|x − y| n+αp
Ω |x − y| n+αp
(x,y)∈Ω×Ω R n \B(x,θ δ ∂Ω (x)/2)
θ δ ∂Ω (x)
|x−y |> 2
∫
≤C |u(x)| p δ∂Ω (x)−αp dx. (9.2.190)
Ω

For the last inequality in (9.2.190) we have used the fact that n + αp > n, thus
|x − y| −n−αp is integrable in y outside B(x, δ∂Ω (x)/8) and the corresponding integral
is bounded by Cδ∂Ω (x)−αp . Subsequently, (9.2.189) readily follows from (9.2.190).
Next, we make the claim that
∬
|u(x) − u(y)| p
dx dy (9.2.191)
|x − y| n+αp
(x,y)∈Ω×Ω
θ δ ∂Ω (x)
|x−y |< 2
∫  ∬ 
|u(x) − u(y)| p
≤C δ∂Ω (z)−n dx dy dz =: I.
Ω |x − y| n+αp
B(z,θ δ ∂Ω (z))×B(z,θ δ ∂Ω (z))

θ(1−θ)
To prove this claim, abbreviate ε := 2(1+θ) and use Fubini-Tonelli’s Theorem which
permits us to write
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 561
∬ ∫
|u(x) − u(y)| p  
I= δ∂Ω (z)−n dz dx dy
|x − y| n+αp
Ω×Ω z ∈Ω
|x−z |<θ δ ∂Ω (z), |y−z |<θ δ ∂Ω (z)
∬ ∫
|u(x) − u(y)| p  
≥ δ∂Ω (z)−n dz dx dy
|x − y| n+αp
(x,y)∈Ω×Ω z ∈Ω
θ δ ∂Ω (x) |x−z |<θ δ ∂Ω (z), |y−z |<θ δ ∂Ω (z)
|x−y |< 2

∬ ∫
|u(x) − u(y)| p  
≥ δ∂Ω (z)−n dz dx dy
|x − y| n+αp
(x,y)∈Ω×Ω B(x,εδ ∂Ω (x))
θ δ ∂Ω (x)
|x−y |< 2

∬
|u(x) − u(y)| p
=C dx dy. (9.2.192)
|x − y| n+αp
(x,y)∈Ω×Ω
θ δ ∂Ω (x)
|x−y |< 2

In (9.2.192) we also made use of the easily checked inclusion

 
B x, εδ∂Ω (x) ⊆ z ∈ Ω : |x − z| < θδ∂Ω (z), |y − z| < θδ∂Ω (z) (9.2.193)

valid for any pair of points (x, y) ∈ Ω × Ω satisfying |x − y| < θ δ∂Ω

2
(x)
, as well as of
the fact that we have δ∂Ω (z) ≈ δ∂Ω (x) uniformly for z ∈ B x, εδ∂Ω (x) . This proves
(9.2.191).
The next order of business is estimating the inner double integral in I. To this
end, fix z ∈ Ω and, for each x, y ∈ B(z, θδ∂Ω (z)), write
∫ 1
|u(x) − u(y)| ≤ |x − y| |(∇u)(x + t(y − x))| dt ≤ |x − y||∇u| ,θ (z). (9.2.194)
0

Hence,
∬
|u(x) − u(y)| p
dx dy
|x − y| n+αp
B(z,θ δ ∂Ω (z))×B(z,θ δ ∂Ω (z))
∫∫
|∇u| ,θ (z) p
≤ dx dy
|x − y| n+αp−p
B(z,θ δ ∂Ω (z))×B(z,θ δ ∂Ω (z))
∫  ∫ 
dy
≤ |∇u| ,θ (z)
p
dx
|x − y| n+αp−p
B(z,θ δ ∂Ω (z)) B(z,θ δ ∂Ω (z))

≤ C|∇u| ,θ (z)
p
δ∂Ω (z) n+p−αp
(9.2.195)

which, when used back in (9.2.191), yields

562 9 Besov and Triebel-Lizorkin Spaces in Open Sets
∬ ∫
|u(x) − u(y)| p
dx dy ≤ C |∇u| ,θ (z)
p
δ∂Ω (z) p−αp dz. (9.2.196)
|x − y| n+αp Ω
(x,y)∈Ω×Ω
θ δ ∂Ω (x)
|x−y |< 2

Thus, ultimately,
∬ ∫
|u(x) − u(y)| p
dx dy ≤ C |∇u| ,θ (z)
p
δ∂Ω (z) p−αp dz.
|x − y| n+αp Ω
(x,y)∈Ω×Ω
θ δ ∂Ω (x) θ δ ∂Ω (y)
|x−y |<max{ 2 , 2 }
(9.2.197)
Together, (9.2.197) and (9.2.189) prove (9.2.187).
There remains to show that the ‘star’ subscript in the right-hand side of (9.2.187)
may be omitted in the case when p ≥ 1. With this goal in mind, we return to
estimating the inner double integral in I and write
∬
|u(x) − u(y)| p
dx dy
|x − y| n+αp
B(z,θ δ ∂Ω (z))×B(z,θ δ ∂Ω (z))

∬ ∫1
|(∇u)(x + t(y − x))| p
≤ dt dx dy
|x − y| n+αp−p
B(z,θ δ ∂Ω (z))×B(z,θ δ ∂Ω (z)) 0

∫ 2θ∫
δ ∂Ω (z) ∫
= ×
B(z,θ δ ∂Ω (z)) 0 ∂B(x,r)∩B(z,θ δ ∂Ω (z))

∫1
|(∇u)(x + t(y − x))| p
× dt dH n−1 (y) dr dx
r n+αp−p
0

∫ 2θ∫
δ ∂Ω (z)∫1 ∫
|(∇u)(w)| p
= dH n−1 (w) dt dr dx
t n−1 r n+αp−p
B(z,θ δ ∂Ω (z)) 0 0 ∂B(x,tr)∩B(z,θ δ ∂Ω (z))

∫ 2θ∫
δ ∂Ω (z)∫r ∫
|(∇u)(w)| p
= dH n−1 (w) dτ dr dx
τ n−1 r 2+αp−p
B(z,θ δ ∂Ω (z)) 0 0 ∂B(x,τ)∩B(z,θ δ ∂Ω (z))

=: I I, (9.2.198)

where for the last two equalities in (9.2.198) we made the changes of variables
w = x − t(y − x) and τ = tr, respectively. Next, for s > 0 arbitrary fixed, we can
further write
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 563

∫ 2θ∫
δ ∂Ω (z)∫r

II ≤ ×
B(z,θ δ ∂Ω (z)) 0 0
∫
|(∇u)(w)| p
× dH n−1 (w) dτ dr dx
τ n−1+s r 2+αp−p−s
∂B(x,τ)∩B(z,θ δ ∂Ω (z))

∫ 2θ∫
δ ∂Ω (z)∫r

= ×
B(z,θ δ ∂Ω (z)) 0 0
∫
|(∇u)(w)| p
× dH n−1 (w) dτ dr dx
|x − w| n−1+s r 2+αp−p−s
∂B(x,τ)∩B(z,θ δ ∂Ω (z))

∫ 2θ∫
δ ∂Ω (z) ∫
|(∇u)(w)| p
= dw dr dx
|x − w| n−1+s r 2+αp−p−s
B(z,θ δ ∂Ω (z)) 0 B(x,r)∩B(z,θ δ ∂Ω (z))

∫ ∫ 2θ∫
δ ∂Ω (z)
|(∇u)(w)| p  dr 
≤ dw dx
|x − w| n−1+s r 2+αp−p−s
B(z,θ δ ∂Ω (z)) B(z,θ δ ∂Ω (z)) 0

=: I I I. (9.2.199)

Since α ∈ (0, 1) we may actually choose s to be a positive number in the interval

(1 − p(1 − α), 1). Such a choice allows us to conclude that
∫  ∫ 
−1−αp+p+s dx
I I I ≤ Cδ∂Ω (z) |(∇u)(w)| p
dw
|x − w| n−1+s
B(z,θ δ ∂Ω (z)) B(w,2θ δ ∂Ω (z))
∫
= Cδ∂Ω (z)−αp+p |(∇u)(w)| p dw. (9.2.200)
B(z,θ δ ∂Ω (z))

By returning with this in (9.2.191) we may therefore compute

564 9 Besov and Triebel-Lizorkin Spaces in Open Sets
∬
|u(x) − u(y)| p
dx dy
|x − y| n+αp
(x,y)∈Ω×Ω
θ δ ∂Ω (x)
|x−y |< 2

∫  ∫ 
−n −αp+p
≤C δ∂Ω (z) δ∂Ω (z) |(∇u)(w)| p dw dz
Ω B(z,θ δ ∂Ω (z))
∫  ∫ 
≤C δ∂Ω (w)−n−αp+p |(∇u)(w)| p dw dz
Ω B(z,θ δ ∂Ω (z))
∫  ∫ 
≤C |(∇u)(w)| p δ∂Ω (w)−n−αp+p 1 dz dw
Ω B(w,θ δ ∂Ω (w)/(1−θ))
∫
=C |(∇u)(w)| p δ∂Ω (w)−αp+p dw. (9.2.201)
Ω

using the fact that for each fixed z ∈ Ω we have δ∂Ω (w) ≈ δ∂Ω (z) uniformly for
w ∈ B(z, θδ∂Ω (z)). With this in hand, the version of (9.2.187) corresponding to the
case when p ≥ 1 and the subscripts , θ are omitted readily follows. Finally, if Ω is
bounded then (9.2.188) becomes a direct consequence of (9.2.187) and the fact that
α > 0. 

For null-solutions of an elliptic system L in an open set Ω, it is possible to

dominate the Gagliardo semi-norm typically associated with Besov spaces in Ω in
terms of a suitably weighted Lebesgue quasi-norm in Ω.

Corollary 9.2.37 For M, m ∈ N, suppose L is a constant (complex) coefficient

homogeneous M × M system of order 2m in Rn , with the property that det [L(ξ)]  0
for each ξ ∈ Rn \ {0}. Also, assume Ω is an arbitrary open, nonempty, proper subset
of Rn , and fix p ∈ (0, ∞) along with α ∈ (0, 1). Then there exists a finite constant
C = C(Ω, L, p, α) > 0 such that the estimate
 ∬ |u(x) − u(y)| p  p1
dx dy ≤ Cu[L p (Ω,δ −α p L n )] M (9.2.202)
|x − y| n+αp ∂Ω
Ω×Ω
 M
holds for any function u ∈ C ∞ (Ω) satisfying Lu = 0 in Ω.

Proof Estimate (9.2.202) follows by combining (9.2.187) with [133, (6.6.86)], the
latter written for k := 1 and s := −α. 

Our next result complements Corollary 9.2.31 (see also [136, Corollary 4.4.3]).
To state it, recall the scale of homogeneous maximal weighted Sobolev spaces
introduced in the second part of Definition 8.6.1, and the piece of notation from
(8.6.6).
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 565

Corollary 9.2.38 Assume M, m ∈ N, and let L be a constant (complex) coefficient

homogeneous M × M system of order 2m in Rn , with the property that det [L(ξ)]  0
for each ξ ∈ Rn \ {0}. Also, let Ω be an arbitrary, open, nonempty, proper subset of
Rn . Finally, suppose k, p, s, a ∈ R are such that

k ∈ N0, 0 < p < ∞, n 1

p −1 + < s+ 1
p < 1, a := 1 − s − p1 . (9.2.203)

Then one has the continuous embedding

 p, p M  . k+1, p M  p M
B 1 (Ω) ∩ Ker L → Wa, (Ω) ∩ L (Ω, L n ) . (9.2.204)
k+s+ p

Moreover, under the stronger assumption that Ω is an (ε, δ)-domain, then one
actually has1
 . k+1, p M   M  p, p M
Wa, (Ω) ∩ Ker L ∩ L p (Ω, L n ) = B (Ω) ∩ Ker L, (9.2.205)
k+s+ p1

with equivalence of quasi-norms.

Proof The fact that the embedding in (9.2.204) is well defined and continuous is seen
by combining (9.2.178) from Proposition 9.2.33 with (9.2.181) in Proposition 9.2.34
(both used with α := s + p1 ). If, in fact, Ω is an (ε, δ)-domain, then from (9.2.204)
and (9.2.159) in Corollary 9.2.31 we conclude that the equality in (9.2.205) holds,
with equivalence of quasi-norms. 

To formulate our next result, having fixed open nonempty proper subset Ω of Rn ,
where n ≥ 2, along with an aperture parameter κ > 0 and an integrability exponent
1,1
q ∈ (0, ∞), define the L q -based area-function of any given u ∈ Wloc (Ω) as (compare
with (4.4.32))
∫   1
A q,κ u (x) := (∇u)(y)q |x − y| q−n dy q
Γκ (x)

∫  q  q1
≈ δ∂Ω (y)q−n (∇u)(y) dy , ∀x ∈ ∂Ω. (9.2.206)
Γκ (x)

Theorem 9.2.39 Fix M, m ∈ N, and suppose L is a constant (complex) coefficient

homogeneous M × M system of order 2m in Rn (where n ≥ 2), with the property
that det [L(ξ)]  0 for each ξ ∈ Rn \ {0}. Also, assume Ω is an open, nonempty,
proper subset of Rn , whose boundary is upper Ahlfors regular, and abbreviate
σ := H n−1 ∂Ω. Finally, fix two integrability exponents, p ∈ (1, ∞) and q ∈ [p, ∞),
along with an aperture parameter κ > 0. Then there exists a finite constant C > 0
 M
with the property that for any function u ∈ C ∞ (Ω) satisfying Lu = 0 in Ω one
has

 k+1, p M  M
1 Lemma 8.6.2 implies that Wa, (Ω) ∩ Ker L ∩ L p (Ω, L n ) also coincides with the
spaces in (9.2.205).
566 9 Besov and Triebel-Lizorkin Spaces in Open Sets

   1− 1 
A q,κ u  
L p (∂Ω,σ)
≤ C δ∂Ω p |∇u|  ≤ Cu[B p, p (Ω)] M . (9.2.207)
L p (Ω, L n ) 1/p

From [42, Theorem 3, p. 1456] we know that in the class of null-solutions in

some Lipschitz domain Ω ⊆ Rn of a given constant coefficient elliptic system L the
L 2 -based area-function and the nontangential maximal function have comparable L p
quasi-norms (for 0 < p < ∞) on the boundary of Ω.  In concert with Theorem 9.2.39,
such result then permits one to dominate Nκ u L p (∂Ω,σ) by a fixed multiple of
u[B p, p (Ω)] M , for all null-solutions u of L in the Lipschitz domain Ω.
1/p

Proof of Theorem 9.2.39 Consider a Whitney decomposition of Ω into Euclidean

cubes Q j of side-length (Q j ) (cf., e.g., [175]; see also [133, Proposition 7.5.3]).
Specifically, assume
, ,
Ω= j ∈N Q j = j ∈N 2Q j , Q̊ j ∩ Q̊ k =  if j  k,
δ∂Ω (x) ≈ (Q j ) uniformly for j ∈ N and x ∈ 2Q j , (9.2.208)

and j ∈N 12Q j ≤ C for a constant C = Cn ∈ (0, ∞).

We may then write

∫
p
A q,κ u(x) dσ(x)
∂Ω
∫ ∫  qp
 q
≈ δ∂Ω (y)q−n (∇u)(y) 1 {y ∈Γκ (x)} dy dσ(x)
∂Ω Ω
∫ ∫  qp

q−n 
q
= δ∂Ω (y) (∇u)(y) 1 {y ∈Γ κ (x)} dy dσ(x). (9.2.209)
∂Ω j Qj

If y ∈ Q j and x ∈ ∂Ω such that y ∈ Γκ (x), then x ∈ Δ j , where Δ j is the “shadow”

of Q j on ∂Ω, i.e., Δ j := πκ (Q j ) = x ∈ ∂Ω : Γκ (x) ∩ Q j   (cf. (A.0.96)). As
a consequence of this, [133, (8.1.17)], and the upper Ahlfors regularity of ∂Ω, we
then have σ(Δ j ) ≤ C(Q j )n−1 for some C ∈ (0, ∞) independent of j. In turn, this
permits us to estimate (bearing in mind that 0 < p/q ≤ 1)
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 567
∫ ∫  qp
 q
δ∂Ω (y)q−n (∇u)(y) 1 {y ∈Γκ (x)} dy dσ(x)
∂Ω j Qj

∫ ∫  q  qp
≤ δ∂Ω (y)q−n (∇u)(y) 1 {y ∈Γκ (x)} dy dσ(x)
j ∂Ω Qj
- .p
∫ ⨏  q1
≤C (Q j ) |∇u| dL
q n
dσ
j Δj Qj
- .p
 ⨏  q1
=C σ(Δ j ) (Q j ) |∇u| dL
q n

j Qj

 ⨏
≤C (Q j )n−1+p |∇u| p dL n
j 2Q j
∫ / 0p
1− 1
≤C δ∂Ω p |∇u| dL n . (9.2.210)
Ω

For the penultimate inequality in (9.2.210) we have used the fact that, for each
i ∈ {1, . . . , n}, the function ∂i u continues to be a null-solution of L in Ω and, as
such, [133, Theorem 6.5.7] ensures that we have the reverse Hölder inequality
⨏ 1 ⨏  p1
n q
|∇u| dL
q
≤C |∇u| p dL n , (9.2.211)
Qj 2Q j

for some C ∈ (0, ∞) independent of u and j. Finally, the last inequality in (9.2.210)
relies on (9.2.208).
From (9.2.209)-(9.2.210) we conclude that the first inequality in (9.2.207) holds.
In concert with Proposition 9.2.33 and Proposition 9.2.34 (both used with α := p1
and k := 0), this ultimately proves that the second inequality in (9.2.207) holds as
well. 

We conclude this section with a technical result, pertaining to the membership

to Besov spaces (which is going to play a significant role later), whose proof makes
essential use of tools from interpolation theory.

Lemma 9.2.40 Let Ω be a bounded Lipschitz domain in Rn and assume that

θ ∈ (0, 1), p ∈ (1, ∞), and q ∈ [p, ∞). (9.2.212)

Also, for each r > 0 define

 
Or := x ∈ Ω : dist(x, ∂Ω) < r . (9.2.213)

Finally, pick an atlas (cf. [133, Definition 2.8.12]) with associated Lipschitz maps
{ϕi }1≤i ≤ N (whose graphs therefore cover ∂Ω). Fix some large geometric constant
568 9 Besov and Triebel-Lizorkin Spaces in Open Sets

λ ∈ (1, ∞) then pick some ro > 0 which is sufficiently small relative to λ and the
quantitative characteristics of the aforementioned atlas.
Then for each r ∈ (0, ro ) there exists a finite constant C = C(Ω, r, p, q, θ) > 0 with
the property that for every function u ∈ C 1 (O4λr ) one has

uB p, q (Or ) ≤ Cu L p (O2λr , L n ) (9.2.214)

θ

N /∫
 ∫ r  p/q 0 1/p
+C |(∇u)(x , ϕi (x ) + s)| q s q(1−θ+1/p)−1 ds dx 
i=1 |x  |<λr 0

(ignoring the effect of various systems of coordinates in which the graphs of the ϕi ’s
are considered, i.e., identifying all hyperplanes H in [133, Definition 2.8.12] with
Rn−1 and all points x0 with the origin).
As a corollary, for each p ∈ (1, ∞) and θ ∈ (0, 1) there exists C ∈ (0, ∞) such that
∫  1/p
uB p, p (Ω) ≤ C |(∇u)(x)| p dist(x, ∂Ω) p(1−θ) dx + Cu L p (Ω, L n )
θ
Ω
(9.2.215)

for each function u ∈ C 1 (Ω).

Proof As in [102], the proof relies on the so-called trace method of interpolation,
which we shall briefly recall. Specifically, given a compatible couple of Banach
spaces A0, A1 and 1 ≤ q < ∞, θ ∈ (0, 1), set (A0, A1 )θ,q for the intermediate space
obtained via the standard real interpolation method (see the discussion in §1.3).
Then, if 1 ≤ q0, q1 < ∞ are such that 1/q = (1 − θ)/q0 + θ/q1 , we have
 ∫ ∞  1/q0  ∫ ∞  1/q1 
q dt q dt
w (A0, A1 )θ, q ≈ inf t θ f (t) A00 + t θ f (t) A11 ,
0 t 0 t
(9.2.216)
uniformly for w ∈ (A0, A1 )θ,q , where the infimum is taken over all functions

f : (0, ∞) −→ A0 + A1 (9.2.217)

with the property that f is locally A0 -integrable, f  (taken in the sense of distri-
butions) is locally A1 -integrable, and such that lim+ f (t) = w in A0 + A1 ; see [15,
t→0
Theorem 3.12.2, p. 73].
To proceed in earnest, fix r > 0 sufficiently small relative to the atlas with
associated Lipschitz maps {ϕi }1≤i ≤ N . In particular, we may assume that Or is a
Lipschitz domain (cf. [131] for a proof). Assuming that this is the case, for every
p ∈ (1, ∞), q ∈ (0, ∞], and θ ∈ (0, 1) we have
 
p,q
Bθ (Or ) = W 1, p (Or ), L p (Or , L n ) (9.2.218)
1−θ,q

(cf., e.g., [102, Proposition 2.17(a), p. 173]). Pick an arbitrary function u ∈ C 1 (O4r ).
Also, for a sufficiently large geometric constant λ ∈ (1, ∞) and an arbitrary point
9.2 Besov and Triebel-Lizorkin Spaces in Open Subsets of R n 569

x∗ ∈ ∂Ω choose
η ∈ Cc∞ B(x∗, λr) with |η| ≤ 1. (9.2.219)
The aim is to show that the infimum of
∫ ∞ ∫ ∞
dt dt
t 1−θ f (t) L p (Or , L n )
q q
t 1−θ f (t)W 1, p (O ) + (9.2.220)
0 r t 0 t

taken over all functions f : (0, ∞) → W 1, p (Or ) + L p (Or , L n ) with the property that
f is locally W 1, p (Or )-integrable, f  (taken in the sense of distributions) is locally
L p (Or , L n )-integrable, and satisfying lim+ f (t) = ηu in W 1, p (Or ) + L p (Or , L n ),
t→0
may be controlled by the right-hand side of (9.2.214). Then choosing a family of
functions η’s to constitute a partition of unity on Or , and summing up such estimates
yields the desired conclusion. Since the domain Ω is Lipschitz, it suffices to do so
in the case when the boundary point x∗ is the origin and when B(0, λr) ∩ ∂Ω is part
of the graph of a Lipschitz function ϕ with ϕ(0) = 0 (where ϕ is one of the ϕi ’s
considered earlier). In addition, we may assume that

dist (x , xn ), ∂Ω ≈ xn − ϕ(x ), ∀x = (x , xn ) ∈ B(0, λr). (9.2.221)

Next, pick

ξ ∈ Cc∞ (−r, r) with |ξ | ≤ 1 and ξ(t) = 1 for |t| < r/2, (9.2.222)

and consider the function

f (t) (x) := η(x)u(x , xn + t)ξ(t) whenever x = (x , xn ). (9.2.223)

It is then apparent from this definition that f is locally W 1, p (Or )-integrable, f  is

locally L p (Or , L n )-integrable, and we have lim+ f (t) = ηu. Thus, it is enough to
t→0
bound (9.2.220) for this choice of f by the right-hand side of (9.2.214). In doing so,
there are two types of terms to consider. Terms of first type look like
∫ r ∫  q/p
q
I := |u(x)| p dx t q(1−θ)−1 dt ≤ Cu L p (O , L n ), (9.2.224)
2λr
0 B(0,2λr)∩Ω

where the last inequality uses the fact that q(1 − θ) − 1 > −1. Given the goal we have
in mind, the estimate just derived suits our purposes. A prototype for the second type
of terms we need to consider is
570 9 Besov and Triebel-Lizorkin Spaces in Open Sets
∫ ∞ ∫  p  q/p dt
 
II := |η(x , xn )| (∇u)(x , xn + t) |ξ(t)|t 1−θ dx dxn
0 Or t
∫ r  ∫ ∫ 2r   q/p dt

≤C t q(1−θ) (∇u)(x , ϕ(x ) + s) p ds dx 
0 |x  |<λr t t
∫ r  ∫ 2r  q/p
=C h(s) ds t q(1−θ)−1 dt, (9.2.225)
0 t

where we have set

∫
 
h(s) := (∇u)(x , ϕ(x ) + s) p dx  . (9.2.226)
|x  |<λr

To continue, recall Hardy’s inequality (cf. [175, p. 272])

∫ ∞ ∫ ∞ β  β β ∫ ∞
β
g(s) ds t α−1 dt ≤ sg(s) s α−1 ds, (9.2.227)
0 t α 0

which holds for for any measurable function g ≥ 0, as long as β ≥ 1 and α > 0.
When applied with g := h1[0,2r] , α := q(1 − θ), and β := q/p, in the case when
q ≥ p this inequality gives
∫ r ∫ 2r  q/p
h(s) ds t q(1−θ)−1 dt (9.2.228)
0 t
∫ ∞   q/p
≤C sh(s)1[0,2r] (s) s q(1−θ)−1 ds
0
∫ 2r  q/p
 
=C h(s)q/p s q/p+q(1−θ)−1 ds = C h(s)s1+p(1−θ)−p/q  q/p
0 L (0,2r),ds

/∫ ∫ 2r    p/q 0 q/p
≤C (∇u)(x , ϕ(x ) + s)q s q(1−θ+1/p)−1 ds dx  ,
|x  |<λr 0

where the last step in (9.2.228) uses Minkowski’s inequality, which once again
requires that q ≥ p. Since the format of the last term above suits our goals, this
completes the proof of the lemma. 

9.3 Envelopes of Besov and Triebel-Lizorkin Spaces in Open Sets

The goal here is to identify the envelopes of Besov and Triebel-Lizorkin Spaces
defined in open subsets of the Euclidean space. Our main result in this regard is the
following theorem.

Theorem 9.3.1 Let Ω ⊆ Rn be a nonempty open set with Ω  Rn . Then

9.3 Envelopes of Besov and Triebel-Lizorkin Spaces in Open Sets 571

p,q p ∗, p ∗
E p∗ Fs (Ω) = B (Ω), (9.3.1)
s−n( p1 − p1∗ )

p,q p ∗, p ∗
E p∗ Bs (Ω) = B (Ω), (9.3.2)
s−n( p1 − p1∗ )

hold provided s ∈ R, 0 < q < ∞, 0 < p < p∗ ≤ 1, and either

n 1
p∗ −1 + < s−n 1
p − 1
p∗ , n 1
min{p,q } −1 + < s,
(9.3.3)
and Ω satisfies an interior corkscrew condition,
or
s−n 1
p − 1
p∗ < p∗ ,
1
s < p1 ,
Ω satisfies an exterior corkscrew condition, (9.3.4)
and ∂Ω is an Ahlfors regular set.
Also, one has
p,q p, p ∗
E p∗ Bs (Ω) = Bs (Ω) (9.3.5)
provided s ∈ R, 0 < q ≤ p∗ ≤ p ≤ 1, and either

n 1
p − 1 < s and Ω satisfies an interior corkscrew condition, (9.3.6)

or
s < p1 , Ω satisfies an exterior corkscrew condition,
(9.3.7)
and ∂Ω is an Ahlfors regular set.

Proof To obtain the identification result from (9.3.1), apply Proposition 7.8.19 with
p,q p ∗, p ∗
X1 := Fs (Rn ), X2 := B (Rn ),
s−n( p1 − p1∗ )
p,q p ∗, p ∗
Y1 := Fs (Ω), Y2 := B (Ω),
s−n( p1 − p1∗ )
p,q p,q
(9.3.8)
R1 := R Rn →Ω : Fs (Rn ) −→ Fs (Ω),
p ∗, p ∗ p ∗, p ∗
R2 := R Rn →Ω : B (Rn ) →B (Ω),
s−n( p1 − p1∗ ) s−n( p1 − p1∗ )

and with p,q p,q

E1 : Fs (Ω) −→ Fs (Rn ),
p ∗, p ∗ p ∗, p ∗ (9.3.9)
E2 : B (Ω) −→ B (Rn ),
s−n( p1 − p1∗ ) s−n( p1 − p1∗ )

the extension operators from Theorem 9.2.15 if the conditions in (9.3.3) are as-
sumed, and the extension operators from Theorem 9.2.19 if the conditions in (9.3.4)
are assumed. Bearing in mind (7.8.168), the abstract identification result (7.8.129)
presently yields (9.3.1). Finally, formulas (9.3.2), (9.3.5) are justified in a similar
fashion. 
572 9 Besov and Triebel-Lizorkin Spaces in Open Sets

9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces

Motivated by the need in boundary value problems for partial differential equations,
here we shall study boundary traces of functions from Besov and Triebel-Lizorkin
spaces in rough domains.
The following theorem is a particular case of more general results, regarding
traces (and extensions) on (and from) d-sets proved by A. Jonsson and H. Wallin in
[106] for the scales of Besov spaces and K. Saka for the scales of Triebel-Lizorkin
spaces in [168].

Theorem 9.4.1 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, suppose A ∈ {B, F} and consider 1 ≤ p ≤ ∞, 0 < q ≤ ∞,
0 < s < 1 if A = B, and 1 < p < ∞, 1 ≤ q ≤ ∞, 0 < s < 1 if A = F. Then for each
p,q
u ∈ As+1/p (Rn ) the limit
⨏
TrRn →Σ u (x) := [u]Rn (x) = lim+ u dL n (9.4.1)
r→0 B(x,r)

exists at σ-a.e. point x ∈ Σ, and the induced trace operator

p,q p,q
TrRn →Σ : As+1/p (Rn ) −→ Bs (Σ, σ). (9.4.2)

well-defined, linear, and bounded.

Furthermore, there exists a bounded linear operator
p,q p,q
Σ→R n : Bs (Σ, σ) −→ As+1/p (R )
ExJWS n
(9.4.3)

with the property that

p,q
TrRn →Σ ◦ ExJWS
Σ→R n = I, the identity on Bs (Σ, σ). (9.4.4)

Proof The results in the range q ∈ [1, ∞] are obtained by specializing [106, Chap-
ter VI, Theorem 1, p. 141] and [168, Theorem 0.1, p. 4], [169, Theorem 0.2, p. 2], to
the case when d := n − 1, β := s, and α := s + 1/p. The extension to q ∈ (0, ∞] when
A = B then follows from this, Theorem 9.1.1, and real interpolation (cf. (7.4.2) in
Theorem 7.4.1). 

Assume next that Σ ⊆ Rn is an arbitrary closed Ahlfors regular set and abbreviate
σ := H n−1 Σ. In this context, a brand of higher-order Besov spaces on Σ has
been introduced by A. Jonsson and H. Wallin in [106] as follows. Given any two
integrability exponents p, q ∈ [1, ∞] and a smoothness index α ∈ (0, ∞)\N, define the
p,q .
higher-order Besov space Bα (Σ, σ) as the collection of families f := { fγ } |γ | ≤[α]
(where [α] denotes the integer part of α), whose components are functions from
L p (Σ, σ), with the property that if for each γ ∈ N0n with |γ| ≤ [α] we consider
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 573
 (x − y)β
Rγ (x, y) := fγ (x) − fγ+β (y) for σ-a.e. x, y ∈ Σ, (9.4.5)
β!
|β | ≤[α]− |γ |

then
. 
 f Bαp, q (Σ,σ) :=  fγ  L p (Σ,σ) (9.4.6)
|γ | ≤[α]

1
q
 ∫∫ q 
 ∞  p 
 
+ 2 j(s− |γ |)q 2 j(n−1) |Rγ (x, y)| p dσ(x) dσ(y)  < ∞,
 
 j=0 |γ | ≤[α] (x,y)∈Σ×Σ 
 |x−y |<2− j 
with a natural interpretation when max{p, q} = ∞. Hereafter, we shall always under-
p,q
stand that Bα (Σ, σ) is equipped with the norm (9.4.6), in which case this becomes
a Banach space. We wish to point out that the brand of higher-order Besov space
described above is compatible with the one considered in §7. Specifically, as is
apparent from (9.4.5)-(9.4.6) and Theorem 7.9.1 we have
p,q p,q
Bα (Σ, σ) = Bα (Σ, σ) whenever p, q ∈ [1, ∞] and α ∈ (0, 1). (9.4.7)

In relation to this variety of higher-order Besov spaces, we have the following

trace/extension theory for Besov and Triebel-Lizorkin spaces in thick open sets with
Ahlfors regular boundaries.

Theorem 9.4.2 Let Ω ⊆ Rn be a nonempty open set which is n-thick and has an
Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω and, given α ∈ (0, ∞) along
with p ∈ (1, ∞) and q ∈ [1, ∞] satisfying α − p1 ∈ (0, ∞)\N, let k := α − p1 +1 ∈ N.
Finally, let A ∈ {B, F} and set q∗ := q if A = B, and q∗ := p if A = F. Then the
following assertions hold.
p,q
(i) For every u ∈ Aα (Ω) the vector-valued function given by
1 ⨏ 2
(k)
RΩ→∂Ω u (x) := lim+ ∂γ w dL n at σ-a.e. x ∈ ∂Ω, (9.4.8)
r→0 B(x,r) |γ | ≤k−1

is meaningfully and unambiguously defined whenever


w ∈ Aα (Rn ) is such that w Ω = u
p,q
(9.4.9)

(the existence of such functions being guaranteed by (9.2.1)).

(ii) The induced higher-order restriction to the boundary operator
(k) p,q p,q
RΩ→∂Ω : Aα (Ω) −→ Bα−1/p
∗
(∂Ω, σ) (9.4.10)

is well defined, linear, bounded, and satisfies

574 9 Besov and Triebel-Lizorkin Spaces in Open Sets
  
(k)
RΩ→∂Ω u = ∂γ u ∂Ω |γ | ≤k−1 everywhere on ∂Ω
   (9.4.11)
for all u ∈ ϕΩ : ϕ ∈ S (Rn ) .

Also, the operator in (9.4.10) is surjective and, in fact, there exists a linear and
bounded operator
(k) p,q p,q
E∂Ω→Ω : Bα−1/p
∗
(∂Ω, σ) −→ Aα (Ω) (9.4.12)

with the property that

(k) (k) p,q
RΩ→∂Ω ◦ E∂Ω→Ω = I, the identity on Bα−1/p
∗
(∂Ω, σ). (9.4.13)
p,q
(iii) For every u ∈ Aα (Ω) the limit
 ⨏ 
γ
lim+ ∂ u dL n
exists, and
ρ→0 Uρ (x)
|γ | ≤k−1 (9.4.14)
(k)
equals RΩ→∂Ω u (x), for σ-a.e. point x ∈ ∂Ω,
 
where Uρ (x) x ∈∂Ω is any family of Lebesgue-measurable sets satisfying (for
0<ρ<ρo
some c > 0)

Uρ (x) ⊆ Ω ∩ B(x, ρ) and L n Uρ (x) ≥ c ρn

(9.4.15)
for all x ∈ ∂Ω and ρ ∈ (0, ρo ).

(iv) If Ω is actually an (ε, δ)-domain with rad (Ω) > 0 and whose boundary is Ahlfors
regular, and if q < ∞, it follows that
 
p,q (k) p,q
u ∈ Aα (Ω) : RΩ→∂Ω u = 0 at σ-a.e. point on ∂Ω = Åα (Ω). (9.4.16)

Hence, in this case,

 
(k)
Cc∞ (Ω) → u ∈ Aα (Ω) : RΩ→∂Ω
p,q
u = 0 at σ-a.e. point on ∂Ω densely.
(9.4.17)

Proof This is a particular case of a more general result regarding traces and ex-
tensions on (and from) d-Ahlfors regular closed subsets of Rn appearing in [19,
Theorem 8.7] (which, in turn, builds on and further refines fundamental work in
[28], [103], [106]). In [19, Theorem 8.7] it was assumed that Ω satisfies an interior
corkscrew condition (if A = F), but this is automatically satisfied in the present case
(in which we are assuming that Ω is n-thick and has an Ahlfors regular boundary)
thanks to [133, Proposition 8.6.12]. In [19, Theorem 8.7] it was also assumed that
Ω  Rn since the proof relied on part (i) of Theorem 9.2.15. However, the result in
part (ii) of Theorem 9.2.15 permits us to allow Ω = Rn as well. 
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 575

We next discuss boundary traces from Besov and Triebel-Lizorkin spaces in

the entire Euclidean space onto closed Ahlfors regular subsets of Rn for a larger
(essentially optimal) range of indices than considered earlier in Theorem 9.4.1 and
Theorem 9.4.2.

Theorem 9.4.3 Assume Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Then the trace operator u → TrRn →Σ u where
⨏
TrRn →Σ u (x) := lim+ u dL n for σ-a.e. point x ∈ Σ, (9.4.18)
r→0 B(x,r)

induces a well-defined bounded linear mapping both in the context

p,q p,q
TrRn →Σ : B 1 (Rn ) −→ Bs (Σ, σ) with
s+ p
 (9.4.19)
p ∈ n−1
n , ∞ , q ∈ (0, ∞], (n − 1) p1 − 1 + < s < 1,

as well as in the context

p,q p, p
TrRn →Σ : F (Rn ) −→ Bs (Σ, σ) with
s+ p1
(9.4.20)
p∈ n ,∞
n−1
, q ∈ (0, ∞), (n − 1) 1
p −1 + < s < 1.

To facilitate the proof of Theorem 9.4.3 we first discuss a useful re-labeling

lemma.

Lemma 9.4.4 Let Λ : J → D be an arbitrary function. Assume its domain, J , is a

nonempty finite set and introduce

N := max #Λ−1 {D} ∈ N. (9.4.21)

D ∈D

For each D ∈ Λ(J ) explicitly enumerate the elements of Λ−1 {D} , say
 j 
Λ−1 {D} = Q1D, Q2D, . . . , Q DD
(9.4.22)
j
where Q1D, Q2D, . . . , Q DD ∈ J and 1 ≤ jD ≤ N.

Finally, consider an Abelian group (A, +, 0), fix some finite family {bQ }Q ∈ J ⊆ A,
and for each j ∈ {1, . . . , N } and D ∈ D define

j bQ j if D ∈ Λ(J ) and 1 ≤ j ≤ jD,
bD := D (9.4.23)
0 otherwise.

Then
 N  
 
j
bQ = bD . (9.4.24)
Q∈J j=1 D ∈D

Proof Since
576 9 Besov and Triebel-Lizorkin Spaces in Open Sets
3
J= Λ−1 {D} , disjoint union, (9.4.25)
D ∈Λ( J)

we may write
  
bQ = bQ
Q∈J D ∈Λ( J) Q ∈Λ−1 ({D })
  
= bQ 1 + bQ 2 + · · · + bQ j D
D D D
D ∈Λ( J)

 
jD   
N 
j j
= bD = bD
D ∈Λ( J) j=1 D ∈D j=1

N  
 
j
= bD , (9.4.26)
j=1 D ∈D

as wanted. 

We are now prepared to present the proof of Theorem 9.4.3.

Proof of Theorem 9.4.3 To get started, we note that Theorem 9.4.1 and real inter-
polation (cf. (7.4.2) in Theorem 7.4.1) imply that the trace operator u → TrRn →Σ u
defined as in (9.4.18) induces a well-defined bounded linear mapping in the context
p,q p,q
TrRn →Σ : B (Rn ) −→ Bs (Σ, σ) with
s+ p1
(9.4.27)
p ∈ [1, ∞], q ∈ (0, ∞], 0<s<1

(the fact that the limit in (9.4.18) is meaningful in this setting is a consequence of
Theorem 9.4.1 and (9.1.16)).
To show that we may actually allow the larger range of indices as indicated in
(9.4.19), we first observe that

n , ∞ , q ∈ (0, ∞], and (n − 1) p − 1 + < s < 1, then

if p ∈ n−1 1

there exist p∗ ∈ (1, ∞), q∗ ∈ [1, ∞], and s∗ ∈ (0, 1) such that (9.4.28)
p,q p ,q
B 1 (Rn ) → B ∗ ∗1 (Rn ) is a continuous embedding.
s+ p s∗ + p∗

Indeed, if p > 1 then taking p∗ := p, q∗ := max{1, q}, and s∗ := s will do (thanks to

(9.1.16)). When p ≤ 1, the assumed lower bound on s implies

1 s+
1
1 1 0+ 1
1
p
− < 1− = − . (9.4.29)
p n n 1 n

Granted this, it follows that it is possible to select p0 ∈ (1, ∞) close to 1 and s0 ∈ (0, 1)
s+ p1 s0 + p1 b+ a1
close to 0 such that 1
p − n < 1
p0 − n
0
. Since 1
a − n  − n1 as a  ∞ and
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 577

s+ p1
b  1, and since the assumed upper bound on s implies that − n1 < 1
p − n , we
conclude that there exist

1 s+ s∗ +
1 1
p 1 p∗
p∗ ∈ (1, ∞) and s∗ ∈ (0, 1) such that − = − . (9.4.30)
p n p∗ n

Choosing q∗ := max{1, q} then ensures (see Theorem 9.1.4) that we have the con-
p,q p ,q
tinuous embedding B 1 (Rn ) → B ∗ ∗1 (Rn ). This finishes the proof of (9.4.28).
s+ p s∗ + p∗
For further reference let us also observe that for the choices of p∗, q∗, s∗ just made
we have
p,q p ,q
Bs (Σ, σ) → Bs∗∗ ∗ (Σ, σ). (9.4.31)
Indeed, from (9.4.30) and the choice of q we see that
s∗
p∗ ∈ (1, ∞), q∗ ∈ [q, ∞], s∗ ∈ (0, 1), and 1
p − s
n−1 = 1
p∗ − n−1 , (9.4.32)

so (9.4.31) follows from this on account of Theorem 7.7.4.

Let us record our progress so far: from (9.4.28) and the first claim in the statement
of Theorem 9.4.1 we conclude that
p,q
for each u ∈ Bs+1/p (Rn ) with p, q, s as in (9.4.19), the limit
⨏ (9.4.33)
TrR →Σ u (x) := lim+
n u dL n exists at σ-a.e. point x ∈ Σ.
r→0 B(x,r)

p,q
Thus, the assignment Bs+1/p (Rn )  u → TrRn →Σ u is well defined and linear when-
ever p, q, s are as in (9.4.19).
Moving on, fix p ∈ n−1 n , ∞ , q ∈ (0, ∞], and (n − 1) p − 1 + < s < 1. Our next
1

goal is to show that there exists a finite constant C = C(Σ, p, q, s) > 0 such that
 
TrRn →Σ u p, q p,q
≤ CuB p, q (Rn ), ∀u ∈ B 1 (Rn ). (9.4.34)
B s (Σ,σ) 1 s+ p
s+ p

To this end, consider first the diagonal case, i.e., when

n−1
n < p = q < ∞ and (n − 1) 1
p −1 + < s < 1. (9.4.35)
p, p
Fix some function u ∈ B (Rn ) which, according to Theorem 9.1.8, admits the
s+ p1
representation 
u= sQ aQ in S (Rn ), (9.4.36)
Q ∈ J(R n ): (Q)≤1

where the sequence of coefficients {sQ }Q belongs to  p and the atoms aQ ’s satisfy
(9.1.41)-(9.1.43) for some integers K ≥ 1, L ≥ 0, and with s replaced by s + 1/p. In
particular, for each Q ∈ J (Rn ) with (Q) ≤ 1 we have
578 9 Besov and Triebel-Lizorkin Spaces in Open Sets
1
s+ p
− p1
aQ ∈ Lip (Rn ), supp (aQ ) ⊆ 3Q, aQ  L ∞ (Rn, L n ) ≤ L n (Q) n ,
(9.4.37)
1
s+ p
− p1 − n1
and ∇aQ  L ∞ (Rn, L n ) ≤ L n (Q) n .

Furthermore, it may be assumed that (cf. (9.1.45))

   1/p
uB p, p (R n ) ≈ |sQ | p . (9.4.38)
1
s+ p
Q ∈ J(R n ): (Q)≤1
 
To proceed, consider a sequence J (m) (Rn ) m∈N of sets satisfying

J (m) (Rn ) is a finite subset of J (Rn ) for each m ∈ N,

J (m) (Rn ) ⊆ J (m+1) (Rn ) for each m ∈ N, (9.4.39)
, (m) (Rn ) = J (Rn ),
m∈N J

and define 
um := sQ aQ for each m ∈ N. (9.4.40)
Q ∈ J (m) (R n ): (Q)≤1

Then Theorem 9.1.8 (cf. (9.1.47)) implies

p, p
um −→ u in B (Rn ) as m → ∞, (9.4.41)
s+ p1

which, in concert with (9.4.28) and Theorem 9.4.1, guarantees that

p , p∗
TrRn →Σ um −→ TrRn →Σ u in Bs∗∗ (Σ, σ) as m → ∞. (9.4.42)

Also, from the first two properties in (9.4.37) together with (9.4.33) we see that
   

TrRn →Σ um = sQ aQ  for each m ∈ N. (9.4.43)
Σ
Q ∈ J (m) (R n ), (Q)≤1
3Q∩Σ

We wish to estimate the size of the sum in the right-hand side of (9.4.43) in the
quasi-norm  · Bsp, p (Σ,σ) . To get things started, assume that Q ∈ J (Rn ) has (Q) ≤ 1
and satisfies 3Q ∩ Σ  , and select some point xQ ∈ 3Q ∩ Σ. Then

the surface ball ΔQ := B xQ , 3 diam(Q) ∩ Σ

n−1
(9.4.44)
satisfies σ(ΔQ ) ≈ L n (Q) n ,

given that Σ is Ahlfors regular. Define


bQ := aQ Σ for each Q ∈ J (Rn ) with (Q) ≤ 1. (9.4.45)
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 579

Fix some Q ∈ J (Rn ) with (Q) ≤ 1. It is then clear from the definition of ΔQ and
the support condition on aQ that

supp bQ ⊆ ΔQ . (9.4.46)

Moreover, by (9.4.37),
1
s+ p
− p1
bQ  L ∞ (Σ,σ) ≤ aQ  L ∞ (Rn, L n ) ≤ L n (Q) n

1
s+ p
n s
− p1
= Cσ(ΔQ ) n−1 − p ,
1
≤ Cσ(ΔQ ) n−1 n (9.4.47)

while for each fixed η ∈ (0, 1] and any x, y ∈ 100ΔQ ,

 
bQ (x) − bQ (y) ≤ |x − y|∇aQ  L ∞ (Rn, L n )
1
s+ p
1−η n
− p1 − ηn
≤ |x − y|L n (Q)− n σ(ΔQ ) n−1 n

s η
≤ C|x − y| η σ(ΔQ ) n−1 − p − n−1 .
1
(9.4.48)

In the situation when we have x, y ∈ Σ \ ΔQ , both numbers bQ (x) and bQ (y) vanish.
Finally, in the scenarion in which one point is contained in the set ΔQ and the other
lies outside 100ΔQ , say x ∈ ΔQ and y ∈ Σ \ 100ΔQ , the idea is to select some
auxiliary point z ∈ ∂B xQ , 3 diam(Q) ∩ {(1 − t)x + t y : t ∈ (0, 1)} and, based on
the previous calculation (with x, z in place of x, y), write
      s η
bQ (x) − bQ (y) = bQ (x) = bQ (x) − bQ (z) ≤ C|x − z| η σ(ΔQ ) n−1 − p − n−1
1

s η
≤ C|x − y| η σ(ΔQ ) n−1 − p − n−1 .
1
(9.4.49)

Hence, ultimately,
  s η
bQ  .η ≤ Cσ(ΔQ ) n−1 − p − n−1 .
1

C (Σ)
(9.4.50)
Moving on, abbreviate
 
JΣ := Q ∈ J (Rn ) : (Q) ≤ 1 and 3Q ∩ Σ   . (9.4.51)

Also, bring in the family D∗ (Σ) of dyadic cubes on Σ defined as in (7.2.1)-(7.2.2). For
each Q ∈ JΣ it is then possible to choose some dyadic cube on Σ, call it Λ(Q) := Qτk,ν ,
which belongs to D∗ (Σ) and satisfies

3Q ∩ Qτk,ν   and (Q) ≈ 2−k , (9.4.52)

where the proportionality constants implicit in the above equivalence are independent
of Q. Hence, we have a function
580 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Λ : JΣ −→ D∗ (Σ) such that

(9.4.53)
3Q ∩ Λ(Q)   and (Q) ≈ (Λ(Q)) uniformly for Q ∈ JΣ .

From (9.4.46), (9.4.47), (9.4.50) we then conclude (taking the background constant
C0 ∈ (0, ∞) appearing in Definition 7.2.1 sufficiently large) that, for each fixed
η ∈ (0, 1],
for each Q ∈ JΣ the function bQ is, up to a fixed multiplicative constant,
an η-smooth block of type (p, s) for the dyadic cube Λ(Q) ∈ D∗ (Σ) on (9.4.54)
the Ahlfors regular set Σ (in the sense described in Definition 7.2.1).
From (9.4.53) we also see that

N := sup #Λ−1 {Qτk,ν } < +∞. (9.4.55)

Qτk,ν ∈D∗ (Σ)

Indeed, there exists some constant θ ∈ (0, 1) with the property that whenever the
cube Qτk,ν ∈ D∗ (Σ) and Q, Q  ∈ Λ−1 {Qτk,ν } then, with yτk,ν denoting the center of
Qτk,ν , we have
(Q) (Q )
θ< 2−k
< θ −1 and θ < 2−k
< θ −1,
(9.4.56)
dist(yτk,ν, Q) < θ −1 2−k and dist(yτk,ν, Q ) < θ −1 2−k ,

from which (9.4.55) follows.

We next claim that there exists a constant C ∈ (0, ∞) such that
1
   p
 
 sQ bQ  p, p
≤C |sQ | p for each finite set J ⊂ JΣ . (9.4.57)
Bs (Σ,σ)
Q∈J Q∈J

To prove this, given a finite set J ⊂ JΣ , the idea is to make use of Lemma 9.4.4
with D := D∗ (Σ), and the function Λ as the restriction of (9.4.53) to the given finite
set J . To be specific, for each D ∈ Λ(J ) let us explicitly enumerate the elements of
Λ−1 {D} as
 j 
Λ−1 {D} = Q1D, Q2D, . . . , Q DD
(9.4.58)
j
where Q1D, Q2D, . . . , Q DD ∈ J and 1 ≤ jD ≤ N,

and, for each j ∈ {1, . . . , N } and D ∈ D∗ (Σ), define


j bQ j if D ∈ Λ(J ) and 1 ≤ j ≤ jD,
bD := D (9.4.59)
0 otherwise,

as well as 
j sQ j if D ∈ Λ(J ) and 1 ≤ j ≤ jD,
sD := D (9.4.60)
0 otherwise.
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 581

Then (9.4.24) permits us to write

 N  
 
j j
sQ bQ = sD bD . (9.4.61)
Q∈J j=1 D ∈D∗ (Σ)

Also, from (9.4.54) and definitions it follows that, for each fixed η ∈ (0, 1],
j
for each j ∈ {1, . . . , N } and D ∈ D∗ (Σ) the function bD is, up to a fixed
multiplicative constant, an η-smooth block of type (p, s) for the dyadic (9.4.62)
cube D ∈ D∗ (Σ) on the Ahlfors regular set Σ (cf. Definition 7.2.1).
At this point we may estimate (with constants allowed to depend on Σ, p, s, and the
integer N from (9.4.55), which has a purely geometric nature)
  N   
   j j 
 sQ bQ  p, p
= sD bD  p, p
B s (Σ,σ) Bs (Σ,σ)
Q∈J j=1 D ∈D∗ (Σ)

N  
  N  
  p1
 j j  j
≤C  sD bD  p, p
≤C |sD | p
Bs (Σ,σ)
j=1 D ∈D∗ (Σ) j=1 D ∈D∗ (Σ)


N   p1    p1
j
≤C |sD | p =C |sQ | p , (9.4.63)
j=1 D ∈D∗ (Σ) Q∈J

where the second inequality above is a consequence of (9.4.62), (9.4.59), (7.2.10),

and the last claim in part (i) of Theorem 7.2.8 (cf. (7.2.38)), while the last equality
makes use once more of (9.4.24). This establishes (9.4.57).
Going further, from (9.4.43), (9.4.45), and (9.4.57) we then deduce that whenever
m , m  ∈ N are such that m  < m  we have
 
    
 n   
TrR →Σ um − TrRn →Σ um  p, p = sQ bQ 
B s (Σ,σ)   
 p, p
(m ) n (m ) n
Q∈J (R )\ J (R ) Bs (Σ,σ)
(Q)≤1 and 3Q∩Σ

1
 p

≤C |sQ | p
.
 
Q ∈ J (m ) (R n )\ J (m ) (R n )
(Q)≤1
(9.4.64)

Given that {sQ }Q ∈ J(Rn ): (Q)≤1 belongs to  p (cf. (9.4.38)), from (9.4.64) we further
conclude that
  p, p
TrRn →Σ um m∈N is a Cauchy sequence in Bs (Σ, σ). (9.4.65)
582 9 Besov and Triebel-Lizorkin Spaces in Open Sets
p, p
Since Bs (Σ, σ) is a quasi-Banach space, hence complete, it follows that the se-
p, p
quence TrRn →Σ um m∈N converges in Bs (Σ, σ). From this, (9.4.42), and (9.4.31)
we then see that actually
p, p
TrRn →Σ um −→ TrRn →Σ u in Bs (Σ, σ) as m → ∞. (9.4.66)

Based on (9.4.66) (while also bearing [138, (3.289), p. 109] in mind), (9.4.43),
(9.4.45), (9.4.57), and (9.4.38) we may now estimate
   
TrRn →Σ u p, p ≤ C · lim sup TrRn →Σ um B p, p (Σ,σ)
B (Σ,σ)
s s
m→∞
 
  
 
≤ C · lim sup  sQ bQ 
m→∞   p, p
Q ∈ J (m) (R n ), (Q)≤1 Bs (Σ,σ)
3Q∩Σ

1
 p

≤ C · lim sup |sQ | p

m→∞
Q ∈ J (m) (R n ), (Q)≤1
3Q∩Σ

1
 p

=C |sQ | p

Q ∈ J(R n ), (Q)≤1
3Q∩Σ

≤ CuB p, p (R n ), (9.4.67)
1
s+ p

for some constant C ∈ (0, ∞) independent of u. In summary, the above analysis

proves that
p, p p, p
TrRn →Σ : B 1 (Rn ) −→ Bs (Σ, σ) with
s+ p
(9.4.68)
p ∈ n−1
n , ∞ and (n − 1) p − 1 + < s < 1,
1

p, p
defined for each u ∈ B (Rn ) as in (9.4.18), is a well-defined linear and bounded
s+ p1
operator. Via real interpolation (cf. (7.4.2) in Theorem 7.4.1, used with q0 := p,
q1 := p) we may then conclude that (9.4.34) holds. This concludes the proof of the
claim that the trace operator (9.4.18) is a well-defined linear and bounded mapping
in the context of (9.4.19).
Dealing with Triebel-Lizorkin spaces turns out to be more delicate as this case
brings into focus the peculiarity that the image of the trace of Triebel-Lizorkin
spaces does not depend on index the q. The proof presented below builds on an
idea first used in [59] where the authors handled the case of the upper half-space.
To get started, based on (9.4.91) and the fact that we have a continuous embedding
p,q p,max{p,q } n
F 1 (Rn ) → B 1 (R ) (cf. (9.1.13)) we conclude that the trace operator is
s+ p s+ p
well defined, linear, and bounded in the context
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 583

p,q p,max{p,q }
TrRn →Σ : F 1 (Rn ) −→ Bs (Σ, σ),
s+ p
 (9.4.69)
p ∈ n−1
n , ∞ , q ∈ (0, ∞], (n − 1) 1
p −1 + < s < 1.

p,max{p,q } p, p
Since Bs (Σ, σ) = Bs (Σ, σ) in the case in which 0 < q ≤ p, it remains to
show that the operator in question actually maps boundedly into the smaller space
p, p
Bs (Σ, σ) when p < q, a condition we assume from now on.
To this end, fix p ∈ n−1n , ∞ , q ∈ (p, ∞), and assume (n − 1) p − 1 + < s < 1,
1
p,q
then pick some arbitrary u ∈ F 1 (R ). Let J := min{1, p,q } = min{1,
n n n
p } and fix two
s+ p
integers K, L satisfying (9.1.40). From Theorem 9.1.8 we know that the function u
has the atomic decomposition

u= sQ aQ in S (Rn ), (9.4.70)
Q ∈ J(R n ): (Q)≤1

where {aQ }Q ∈ J(Rn ),l(Q)≤1 are (s, p)-atoms of type (K, L, J) (i.e., functions sat-
isfying the conditions listed in (9.1.41)-(9.1.43)) and the numerical sequence
{sQ }Q ∈ J(Rn ), (Q)≤1 is such that
 
   1/q 

uF p, q (Rn ) ≈  L n (Q)−1/p |sQ |1Q 
q
 . (9.4.71)
1
s+ p  Q ∈ J(Rn ), (Q)≤1  p n n
L (R , L )

We claim that
there is a function u∗ ∈ F
p, p p, p
(Rn ) =B (Rn ) with the property
s+ p1 s+ p1
that u∗ F p, p (Rn ) ≤ CuF p, q and TrRn →Σ u∗ = TrRn →Σ u. (9.4.72)
1 1
(R n )
s+ p s+ p

To prove this claim, observe that (as is apparent from (9.1.41)-(9.1.43) and the fact
that J, K, L currently do not depend on q since min{1, p} = min{1, p, q}) the aQ ’s
p, p p, p
are (s, p)-atoms of type (K, L, J) for the space F 1 (Rn ) = B 1 (Rn ). Introduce
s+ p s+ p


u∗ := sQ aQ , (9.4.73)
Q ∈ J(R n ), (Q)≤1
3Q∩Σ

and observe that, since

584 9 Besov and Triebel-Lizorkin Spaces in Open Sets
 
 
  1/q 
 
 L n (Q)−1/p |sQ |1Q
q 
 
 Q ∈ J(R n ), (Q)≤1 
 3Q∩Σ

L p (R n, L n )
 
  1/q 
 
≤  L n (Q)−1/p |sQ |1Q 
q

 Q ∈ J(R n ), (Q)≤1 
L p (R n, L n )

≈ uF p, q (R n ) < ∞, (9.4.74)

1
s+ p

we may invoke Theorem 9.1.8 to conclude that the series in the right-hand side of
p,q
(9.4.73) converges in F 1 (Rn ). Hence,
s+ p

u∗ ∈ F
p,q p,q
(Rn ) → B (Rn ). (9.4.75)
s+ p1 s+ p1

To proceed, associate with the present function  {um }m∈N as in (9.4.39)-

 u a sequence
(9.4.40). Likewise, for the same family of sets J (m) (Rn ) m∈N (as in (9.4.39)), define

u∗m := sQ aQ for each m ∈ N. (9.4.76)
Q ∈ J (m) (R n ), (Q)≤1
3Q∩Σ

From the continuity of the embedding in (9.4.75) and Theorem 9.1.8 (cf. (9.1.51))
we deduce that

um −→ u and u∗m −→ u∗ in B
p,q
(Rn ) as m → ∞ (9.4.77)
s+ p1

which, in light of (9.4.19), implies

TrRn →Σ um −→ TrRn →Σ u and TrRn →Σ u∗m −→ TrRn →Σ u∗

p,q
(9.4.78)
in Bs (Σ, σ) as m → ∞.

Upon noting that, thanks to the first two properties in (9.4.37) and (9.4.33), for each
m ∈ N we have
   

TrRn →Σ um = sQ aQ  = TrRn →Σ u∗m, (9.4.79)
Σ
Q ∈ J (m) (R n ), (Q)≤1
3Q∩Σ

we ultimately conclude from (9.4.78) and (9.4.79) that

TrRn →Σ u = TrRn →Σ u∗ in Bs (Σ, σ),

p,q
(9.4.80)

where both traces are understood in the sense of (9.4.19).

Next, we wish to show that, in fact,
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 585

u∗ ∈ F and u∗ F p, p (Rn ) ≤ CuF p, q

p, p
(Rn ) (R n ) . (9.4.81)
s+ p1 1
s+ p 1
s+ p

With this goal in mind, for a fixed m ∈ N, sufficiently large, and for every cube
Q ∈ J (Rn ) with (Q) ≤ 1 and such that 3Q ∩ Σ  , we define
 √ 
EQm
:= x ∈ Q : dist(x, Σ) > n 2−m (Q) . (9.4.82)

We now claim that if m is large enough then there exists a constant c = c(Σ, m) > 0
such that

Q ∈ J (Rn ), (Q) ≤ 1, 3Q ∩ Σ   =⇒ L n (EQ

m
) ≥ cL n (Q). (9.4.83)

To justify this claim, pick xQ ∈ Q ∩ Σ and note that, on the one hand,
 
Q \ EQ m
⊆ x ∈ B xQ, 3 diam(Q) : dist(x, Σ) < 2−m diam(Q) . (9.4.84)

On the other hand, by relying on [133, Corollary 8.6.11] we may estimate the (n-
dimensional) Lebesgue measure of the bigger set above by C 2−m diam(Q) , with
n

C > 0 a finite constant depending only on the Ahlfors regularity constants of Σ. With
this in hand, the claim made in (9.4.83) readily follows. Moving on, it is clear that
one can choose m  ∈ N, depending exclusively on Σ, n, and m, with the property
that if Q, Q  ∈ J (Rn ) satisfy 3Q ∩ Σ  , 3Q  ∩ Σ   and are such that Q  ⊆ Q

and (Q ) ≤ 2−m (Q) then Q  ∩ EQ m = . The upshot of this observation is that the
 m
family of sets EQ Q ∈ J(Rn ), (Q)≤1 has finite overlap and, hence, there exists a finite
3Q∩Σ
constant N = N(Σ, n, m) > 0 such that

1EQm ≤ N. (9.4.85)
Q ∈ J(R n ), (Q)≤1
3Q∩Σ

At this stage we may rely on (9.4.71), [59, Proposition 2.7 on p. 51], the elementary
 α 
fact that if α > 0 then i ∈I ai 1Ei ≈ i ∈I aiα 1Ei , uniformly for ai ≥ 0, granted
that the sets {Ei }i ∈I have finite overlap, (9.4.85), and (9.4.83) to write
586 9 Besov and Triebel-Lizorkin Spaces in Open Sets
 
 
  1/q 
 
≥  L n (Q)−1/p |sQ |1Q 
q
CuF p, q (R n ) 
1
s+ p  Q ∈ J(R n ), (Q)≤1 
 3Q∩Σ

L p (R n, L n )
 
 
  1/q 
 
≈  L n (Q)−1/p |sQ |1EQm 
q

 Q ∈ J(R n ), (Q)≤1 
 3Q∩Σ

L p (R n, L n )
 1
 p
  
 

≈ L (Q)
n −1/p
|sQ | 1EQ 
p
m
Q ∈ J(Rn ), (Q)≤1 
 3Q∩Σ
 1
L (R n, L n )
  
 
≈  sQ Q ∈ J(Rn ), (Q)≤1  p . (9.4.86)
3Q∩Σ 

 
Thus, sQ Q ∈ J(Rn ), (Q)≤1 belongs to the space  p hence, by Theorem 9.1.8, we
3Q∩Σ
have u∗ ∈ B
p, p p, p
1 (R ) = F (Rn )
n and
s+ p s+ p1

  
 
u∗ F p, p (Rn ) ≤ C  sQ Q ∈ J(Rn ), (Q)≤1  ≤ CuF p, q (R n ), (9.4.87)
1
s+ p 3Q∩Σ p 1
s+ p

finishing the proof of (9.4.81), and completing the justification of (9.4.72). Having
established (9.4.72), we conclude that the trace operator (9.4.18) is well defined,
linear, and bounded in the context of (9.4.20). 
We are now in a position to formulate and prove our main result describing traces
from Besov and Triebel-Lizorkin spaces defined in open subsets of Rn onto theirs
boundaries.

Theorem 9.4.5 Let Ω ⊆ Rn be an open set which is n-thick and has an Ahlfors
regular boundary. Abbreviate σ := H n−1 ∂Ω and let A ∈ {B, F}. Then the following
statements are true.
p,q
(i) For each function u ∈ Aα (Ω) with

0 < p ≤ ∞, 0 < q ≤ ∞, α> 1

p + (n − 1) 1
p − 1 +,
(9.4.88)
assuming p < ∞ when A = F

the limit
⨏
TrΩ→∂Ω u (x) := lim+ w dL n at σ-a.e. x ∈ ∂Ω (9.4.89)
r→0 B(x,r)

is meaningfully and unambiguously defined whenever (cf. (9.2.1))

9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 587

w ∈ Aα (Rn ) is such that w Ω = u.
p,q
(9.4.90)

(ii) The boundary trace operator u → TrΩ→∂Ω u induced by (9.4.89) is well defined,
linear, and bounded in the context
p,q p,q
TrΩ→∂Ω : B 1 (Ω) −→ Bs (∂Ω, σ) with
s+ p
 (9.4.91)
p ∈ n , ∞ , q ∈ (0, ∞], (n − 1) p1 − 1 + < s < 1.
n−1

Moreover,
when p ∈ [1, ∞], q ∈ (0, ∞], and s ∈ (0, 1), the trace
p,q p,q
operator TrΩ→∂Ω maps B 1 (Ω) onto Bs (∂Ω, σ). (9.4.92)
s+ p

Also, the boundary trace operator u → TrΩ→∂Ω u induced by (9.4.89) is well

defined, linear, and bounded in the context
p,q p, p
TrΩ→∂Ω : F (Ω) −→ Bs (∂Ω, σ) with
s+ p1
(9.4.93)
p∈ n , ∞),
n−1
q ∈ (0, ∞), (n − 1) 1
p −1 + < s < 1.

In addition,
when p ∈ (1, ∞), q ∈ [1, ∞], and s ∈ (0, 1), the trace
p,q p, p
operator TrΩ→∂Ω maps F 1 (Ω) onto Bs (∂Ω, σ). (9.4.94)
s+ p

p,q
(iii) For each function u ∈ Aα (Ω) with p, q, α as in (9.4.88) the limit
⨏
lim+ u dL n exists and equals TrΩ→∂Ω u (x),
ρ→0 Uρ (x) (9.4.95)
at σ-a.e. point x ∈ ∂Ω,
 
where Uρ (x) x ∈∂Ω is any family of Lebesgue-measurable sets satisfying (for
0<ρ<ρo
some c > 0)

Uρ (x) ⊆ Ω ∩ B(x, ρ) and L n Uρ (x) ≥ c ρn

(9.4.96)
for all x ∈ ∂Ω and ρ ∈ (0, ρo ).

Two significant consequences of this are as follows. First, one may choose (for
some background geometric constant ρo > 0)

Uρ (x) := Ω ∩ B(x, ρ) for all x ∈ ∂Ω and ρ ∈ (0, ρo ) (9.4.97)

p,q
and conclude from (9.4.95) that for each u ∈ Aα (Ω) with p, q, α as in (9.4.88)
one has
⨏
TrΩ→∂Ω u (x) = lim+ u dL n at σ-a.e. point x ∈ ∂Ω. (9.4.98)
r→0 Ω∩B(x,r)
588 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Second, (9.4.95) implies that there exists an aperture parameter κ > 0 with the
p,q
property that whenever u ∈ Aα (Ω) with p, q, α as in (9.4.88) has a pointwise
κ−n.t.
nontangential trace u∂Ω at σ-a.e. point on ∂Ω then actually
κ−n.t.
TrΩ→∂Ω u = u∂Ω at σ-a.e. point on ∂Ω. (9.4.99)

(iv) Assume Ω is actually an (ε, δ)-domain with rad (Ω) > 0 and whose boundary is
Ahlfors regular, p, q ∈ (1, ∞), and s ∈ (0, 1). Also, let A ∈ {B, F}. Then
 
p,q p,q
u ∈ A 1 (Ω) : TrΩ→∂Ω u = 0 at σ-a.e. point on ∂Ω = Å 1 (Ω),
s+ p s+ p
(9.4.100)
hence
 
Cc∞ (Ω) → u ∈ A 1 (Ω) : TrΩ→∂Ω u = 0 at σ-a.e. point on ∂Ω densely.
p,q
s+ p
(9.4.101)

Before presenting the proof of Theorem 9.4.5 we wish to make three remarks.
Remark 1. As is apparent from (8.3.37) and (9.4.98) (while also keeping [133,
Lemma 5.11.9)] in mind),
whenever Ω ⊆ Rn is an (ε, δ)-domain with rad (Ω) > 0 and whose
boundary is an Ahlfors regular set, the boundary trace operator TrΩ→∂Ω
(9.4.102)
from Theorem 8.3.6 acts in a compatible fashion with the boundary
trace operator TrΩ→∂Ω from Theorem 9.4.5.
In particular, this justifies employing the same notation for said boundary trace
operators, even though they act in two different functional analytic settings (whenever
relevant, the actual functional analytic setting will be made clear each time the
boundary trace operator is used).
Remark 2. If Mψ denotes the operator of pointwise multiplication by ψ then

TrΩ→∂Ω ◦ Mψ = ψ ∂Ω · TrΩ→∂Ω for each ψ ∈ Cc∞ (Rn ), (9.4.103)

where the trace operator TrΩ→∂Ω is considered either as in (9.4.91), or as in (9.4.92).

Indeed, by the continuity of said trace operators
 and the density result in (9.2.7), it
suffices to check that TrΩ→∂Ω (ψu) = ψ ∂Ω · TrΩ→∂Ω u for each u ∈ Cc∞ (Ω) and
ψ ∈ Cc∞ (Rn ). This, however, is readily implied by (9.4.89)-(9.4.90).
Remark 3. Retain the setting of Theorem 9.4.5 and make the additional assumption
that ∂Ω is compact. In this case, we may naturally extend the action of the boundary
trace operator TrΩ→∂Ω from (9.4.91) to the larger space
 
u ∈ D (Ω) : ψu ∈ B 1 (Ω) for all ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near ∂Ω
p,q
s+ p
(9.4.104)
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 589
p,q
by defining TrΩ→∂Ω u in the “new” sense, to be TrΩ→∂Ω (ψu) ∈ Bs (∂Ω, σ) in the
“old” sense, whenever u and ψ as in (9.4.104). It is then clear that this definition is
meaningful and unambiguous (here (9.4.103) is also helpful). In view of Conven-
tion 8.3.7, this yields a well defined, linear, and continuous boundary trace operator
in the context
p,q p,q
TrΩ→∂Ω : B 1 (Ω)bdd −→ Bs (∂Ω, σ) with
s+ p
 (9.4.105)
p ∈ n−1
n , ∞ , q ∈ (0, ∞], (n − 1) p1 − 1 + < s < 1.

Moreover, a similar extension works in the case of the trace operator in (9.4.93),
thus producing a well defined, linear, and continuous boundary trace operator in the
context
p,q p, p
TrΩ→∂Ω : F (Ω)bdd −→ Bs (∂Ω, σ) with
s+ p1
(9.4.106)
p∈ n , ∞),
n−1
q ∈ (0, ∞), (n − 1) 1
p −1 + < s < 1.

We now turn to the task of providing the proof of Theorem 9.4.5.

p,q
Proof of Theorem 9.4.5 Let u ∈ Aα (Ω) with p, q, α as in (9.4.88) plus  the addi-
tional assumption that p < ∞, and consider w ∈ Aα (Rn ) such that w Ω = u. Then
p,q

Corollary 9.2.1 and (9.2.9) guarantee the existence of p∗ ∈ (1, ∞), q∗ ∈ [1, ∞] and
α∗ ∈ p1∗ , 1 + p1∗ such that
p,q p ,q∗ p,q p ,q∗
Aα (Ω) ⊆ Aα∗∗ (Ω) and Aα (Rn ) ⊆ Aα∗∗ (Rn ). (9.4.107)

Hence, u ∈ Aα∗∗ ∗ (Ω) and w ∈ Aα∗∗ ∗ (Rn ) are such that w Ω = u, so we may invoke
p ,q p ,q

item (i) of Theorem 9.4.2 to conclude that the limit (9.4.89) is meaningfully and
unambiguously defined when p < ∞. There remains to treat the scenario when
p = ∞, 0 < q ≤ ∞, α > 0, and A = B. Of course, it suffices to restrict attention to
the situation when α ∈ (0, 1). In this case, all desired conclusions follow by observing
from (9.1.16) and (9.1.52) that
∞,q
Bα (Rn ) ⊆ Bα∞,∞ (Rn ) = C α (Rn ). (9.4.108)

Moving on, fix p ∈ n−1 n , ∞ along with q ∈ (0, ∞], and choose s such that
p,q
(n − 1) p1 − 1 + < s < 1. Consider next an arbitrary u ∈ B 1 (Ω) and recall from
s+ p
(9.2.1) that there exists

such that w Ω = u and w B p, q
p,q
w∈B (Rn ) (R n ) ≤ 2uB p, q (Ω) . (9.4.109)
s+ p1 1
s+ p 1
s+ p

Then at σ-a.e. point x ∈ ∂Ω we may write

⨏
TrΩ→∂Ω u (x) = lim+ w dL n = TrRn →∂Ω w (x) (9.4.110)
r→0 B(x,r)

thanks to (9.4.89) with α := s + 1

p and (9.4.18) with Σ := ∂Ω. Consequently,
590 9 Besov and Triebel-Lizorkin Spaces in Open Sets
   
TrΩ→∂Ω u p, q = TrRn →∂Ω w B p, q (∂Ω,σ)
Bs (∂Ω,σ) s

≤ Cw B p, q (R n ) ≤ CuB p, q (Ω) (9.4.111)

1
s+ p 1
s+ p

by (9.4.110), the boundedness of the trace operator in (9.4.19), and (9.4.109). This
shows that the boundary trace operator u → TrΩ→∂Ω u induced by (9.4.89) is well
defined, linear, and bounded in the context of (9.4.91). That similar results are valid
for the boundary trace operator in the context of (9.4.93) may be justified in a
completely analogous fashion, this time relying on the boundedness of the operator
(9.4.20).
∞,q
Let us now consider the case p = ∞ in (9.4.91). Hence, fix u ∈ Bs (Ω) with
0 < q ≤ ∞ and s ∈ (0, 1). From (9.4.89), (9.4.108), and (9.4.18) we see that

TrΩ→∂Ω u = w ∂Ω = TrRn →∂Ω w
 (9.4.112)
whenever w ∈ Bs (Rn ) ⊆ C s (Rn ) is such that w Ω = u.
∞,q

Since (9.2.1) guarantees that we may take


w ∈ Bs (Rn ) such that w Ω = u and w B∞,
∞,q
(R n ) ≤ 2u B s (Ω), (9.4.113)
q ∞, q
s

we conclude from (9.4.112)-(9.4.113) that

   
TrΩ→∂Ω u ∞, q = TrRn →∂Ω w B∞, q (∂Ω,σ)
B s(∂Ω,σ) s

≤ Cw B∞,
s
q
(R n ) ≤ Cu B s (Ω)
∞, q (9.4.114)

thanks to the boundedness of the trace operator in (9.4.19) with p = ∞. Ultimately,

this shows that the boundary trace operator u → TrΩ→∂Ω u induced by (9.4.89) is
well defined, linear, and bounded in the context of (9.4.91) when p = ∞ as well.
Finally, the claims made in (9.4.92) and (9.4.94) are seen from Theorem 9.4.1
and (9.4.89)-(9.4.90). To conclude the treatment of item (ii).
Turning attention to item (iii), the claims in (9.4.95) with p < ∞ follow from item
(iii) of Theorem 9.4.2 and embeddings, much as we have reasoned in the proof of
the current item (i). The claims in (9.4.95) when p = ∞ and A = B are clear from
(cf. (9.2.10) and (9.2.23))
∞,q
Bα (Ω) ⊆ Bα∞,∞ (Ω) = C α (Ω) = C α (Ω). (9.4.115)

Next, since Ω is assumed to be n-thick, the choice in (9.4.97) satisfies the properties
demanded in (9.4.96). Corresponding to this choice, the equality in (9.4.97) becomes
(9.4.98).
Going further, consider the claim made in (9.4.99). From the current assump-
tions and [133, Proposition 8.6.12] it follows that Ω satisfies an interior corkscrew
condition with some constant θ ∈ (0, 1) (cf. [133, (5.1.5)]). Granted this, fix an
aperture parameter κ > 2(θ −1 − 1) and recall from [133, (8.8.66)] that (with notation
introduced in [133, Definition 5.1.3]) we have
9.4 Traces of Functions from Besov and Triebel-Lizorkin Spaces 591

B zr (x), θr/2 ⊆ Γκ (x) ∩ B(x, r) ⊆ Ω ∩ B(x, r)

(9.4.116)
for all x ∈ ∂Ω and r ∈ (0, 2 diam Ω).

If we then take

Uρ (x) := B zρ (x), θ ρ/2 for all x ∈ ∂Ω and ρ ∈ (0, 2 diam Ω), (9.4.117)

then (9.4.116) ensures that all conditions in (9.4.96) are satisfied.

 As
 such, we may
conclude that (9.4.95) holds for this choice of the family Uρ (x) x ∈∂Ω . On
0<ρ<2 diam Ω
κ−n.t.
the other hand, the assumption that the pointwise nontangential trace u∂Ω exists
σ-a.e. on ∂Ω implies, in light of [133, Definition 8.9.1] (cf. (A.0.82)) and the first
inclusion in (9.4.116), that
 κ−n.t.  ⨏

u ∂Ω (x) = lim+ u dL n at σ-a.e. point on ∂Ω. (9.4.118)
ρ→0 Uρ (x)

In concert with (9.4.95), this finishes the proof of (9.4.99).

Finally, the claims in item (iv) become consequences of (9.4.16)-(9.4.17) upon
(1)
observing that we presently have RΩ→∂Ω = TrΩ→∂Ω (cf. (9.4.8) and (9.4.89)). 
Our next result is an embodiment of the intuitive idea that the two traces of a
function belonging to a Besov or Triebel-Lizorkin space in Rn from either side of a
surface should agree with one another.
Proposition 9.4.6 Let Ω ⊆ Rn be an open set which is two-sided n-thick and has an
Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω and define

Ω+ := Ω, Ω− := Rn \ Ω. (9.4.119)

Also, let A ∈ {B, F} and assume

p∈ n ,∞
n−1
, q ∈ (0, ∞), (n − 1) 1
p −1 + < s < 1. (9.4.120)

Then both Ω+ and Ω− are open subset of Rn which are n-thick and have Ahlfors
regular boundaries; in fact,

∂(Ω+ ) = ∂(Ω− ) = ∂Ω. (9.4.121)

In particular, the boundary trace operators TrΩ± →∂Ω are meaningfully defined on
p,q
functions from A 1 (Ω± ) (as in items (i)-(ii) of Theorem 9.4.5), and one has
s+ p

 
TrΩ+ →∂Ω uΩ+ = TrΩ− →∂Ω uΩ− for each u ∈ A 1 (Rn ).
p,q
(9.4.122)
s+ p

Proof Since Ω ⊆ Rn is Lebesgue measurable and two-sided n-thick, [133, (5.2.4)]

implies ∂∗ Ω = ∂Ω. Given that Ω is also open and has an Ahlfors regular boundary,
we conclude that Ω is an Ahlfors regular domain (cf. (A.0.1)). Granted this, [133,
592 9 Besov and Triebel-Lizorkin Spaces in Open Sets

(5.10.52)] applies and gives that ∂(Ω) = ∂Ω. In turn, this readily implies (9.4.121).
Clearly, Ω+ = Ω is n-thick. To show that Ω− is also n-thick, observe first that

diam(Ω− ) = diam(Ω− ) = diam(Rn \ Ω). (9.4.123)

Hence, whenever x ∈ ∂(Ω− ) = ∂Ω and 0 < r < 2 diam(Ω− ) = 2 diam(Rn \ Ω) we

may write, for some c > 0 independent of x and r,

L n B(x, r) ∩ Ω− = L n B(x, r) ∩ (Rn \ Ω) ≥ c r n, (9.4.124)

where the equality is implied by [133, Lemma 5.1.2] and the inequality is a conse-
quence of the fact that the set Rn \ Ω is n-thick (cf. [133, Definition 5.1.1]). This
shows that, indeed, Ω− is n-thick.
Having proved that Ω± are n-thick open subset of Rn whose boundaries are Ahlfors
regular and satisfy (9.4.121), we may conclude from item (i) of Theorem 9.4.5 that
the boundary trace operators TrΩ± →∂Ω are meaningfully defined when acting on
p,q
functions belonging to A 1 (Ω± ).
s+ p
Let us now justify (9.4.122) in the case when A = B. There are two basic
observations to make. First, from (9.2.6), what we have proved up to this point, and
(9.4.91) we deduce that the operators

B 1 (Rn )  u −→ TrΩ± →∂Ω uΩ± ∈ Bs (∂Ω, σ)
p,q p,q
s+ p
(9.4.125)
are well-defined, linear, and bounded.

Second, with S (Rn ) denoting the class of Schwartz functions in Rn , from (9.4.89)-
(9.4.90) we see that
 
TrΩ± →∂Ω uΩ± = u∂Ω whenever u ∈ S (Rn ). (9.4.126)

By combining (9.4.125), (9.4.126), and (9.1.12) we arrive at the conclusion that

(9.4.122) holds when A = B. Finally, the case of (9.4.122) when A = F is dealt with
similarly, this time using (9.4.93) in place of (9.4.91). 

9.5 The Conormal Derivative Operator on Besov and

Triebel-Lizorkin Spaces

The goal here is to define the action of the conormal (associated with some second-
order system) in the context of Besov and Triebel-Lizorkin spaces. To set the stage,
recall that given a nonempty open subset Ω of Rn , we denote by D (Ω) the space of
distributions in Ω, and by D  (Ω) ·, · D(Ω) the pairing between distributions and test
p,q ∗
functions in Ω. Also, if f ∈ As (Ω) , where A ∈ {B, F}, 1 ≤ p, q ≤ ∞ and s ∈ R,
we shall denote by f Ω the distribution induced by f on Ω according to
9.5 The Conormal Derivative Operator on Besov and Triebel-Lizorkin Spaces 593
% & % &
D  (Ω) f Ω, ϕ D(Ω)
:= (Ap, q
s (Ω))
∗ f, ϕ p, q
As (Ω), ∀ϕ ∈ Cc∞ (Ω), (9.5.1)

where the brackets in the right-hand side of (9.5.1) stand for the duality pairing
∗
between functionals from As (Ω) and functions from As (Ω) (with ϕ ∈ Cc∞ (Ω)
p,q p,q

viewed, this time, as a function in the latter space).

As a preamble to introducing the conormal derivative operator in Proposi-
tion 9.5.2, we prove the following lemma.

Lemma 9.5.1 Assume Ω is a nonempty, open, proper subset of Rn , satisfying an

interior corkscrew condition and such that ∂Ω = ∂(Ω). Also, suppose s > 0,
p, q ∈ (1, ∞), and denote by p, q  the Hölder conjugate exponents of p, q. In this
setting, recall from (9.2.123) that
p,q ∗ p,q
As (Ω) = A−s,0 (Ω) (9.5.2)

via the isomorphic identification (cf. (9.2.129))

p,q p,q ∗
A−s,0 (Ω)  f → Λ f ∈ As (Ω) where
% &
Λ f (w) := Ap, q (Rn ) f , W Ap, q n for all
s (R )
(9.5.3)
−s

w ∈ As (Ω) and W ∈ As (Rn ) with W Ω = w.
p,q p,q

p,q p,q ∗
Then for any f ∈ A−s,0 (Ω) the distribution Λ f Ω induced by Λ f ∈ As (Ω) in
 p,q
the sense of (9.5.1) agrees with the distribution f  ∈ A−s,z (Ω), i.e., Ω

Λ f Ω = f Ω in D (Ω). (9.5.4)

Proof Given any test function ϕ ∈ Cc∞ (Ω), if ϕ # denotes its extension by zero outside
Ω to the entire Rn , we may write
% & % &
D  (Ω) Λ f Ω, ϕ D(Ω) = (Ap, q
s (Ω))
∗ Λ f Ω, ϕ A p, q (Ω)
s

% & % &
= p , q 
A−s (R n )
f, ϕ
# Ap, q n =
s (R ) D  (R n ) f, ϕ
# D(Rn )
%  &
= D  (Ω) f Ω, ϕ D(Ω) (9.5.5)
% &
thanks to (9.5.1), (9.5.3), the compatibility of Ap, q (Rn ) ·, · Ap, q n with the distri-
s (R )
−s
butional pairing in R , and the manner in which a distribution in Rn is restricted to
n

an open set. In view of the arbitrariness of ϕ ∈ Cc∞ (Ω) (9.5.5) establishes (9.5.4). 

We are ready to introduce the conormal derivative operator acting from Besov
and Triebel-Lizorkin spaces, and study some of its most basic properties.

Proposition 9.5.2 Let Ω ⊆ Rn be an (ε, δ)-domain with a compact Ahlfors regu-

lar boundary, and abbreviate σ := H n−1 ∂Ω. Also, fix p, q ∈ (1, ∞) along with
−1 −1 # M ∈ N,
s ∈ (0, 1), and denote p := 1 − p1 , q  := 1 − q1 . Next, for some M,
594 9 Besov and Triebel-Lizorkin Spaces in Open Sets
αβ
consider a coefficient tensor A = a jk 1≤α ≤ M4 with constant (complex) entries, and
1≤β ≤M
1≤ j,k ≤n
associate with this coefficient tensor the second-order homogeneous constant (com-
plex) coefficient M # × M system L A := aαβ ∂j ∂k 4 in R (with the summation
n
jk 1≤α ≤ M
1≤β ≤M
convention over repeated indices understood throughout). Finally, let A ∈ {B, F},
set q∗ := q if A = B and q∗ := p if A = F, and denote by q∗ the Hölder conjugate
exponent of q∗ .
In this context, define the conormal derivative operator ∂νA as the mapping from
the space
 
 p,q  M / p,q 04
∗ M   M 4
(u, f ) ∈ As+1/p (Ω) ⊕ A1−s+1/p (Ω) : L Au = f Ω in D (Ω)
 p,q M
4
with values in Bs−1∗ (∂Ω, σ)
(9.5.6)

(where the convention introduced in (9.5.1) has been used) associating to each pair
 p,q M / 04
p,q ∗ M
(u, f ) with u = (uβ )β ∈ As+1/p (Ω) and f ∈ A1−s+1/p (Ω) satisfying
  M 4
L Au = f Ω in D (Ω) the functional
 M 
4 ∗  p,q M
4
p,q
∂νA(u, f ) ∈ B1−s ∗ (∂Ω, σ) = Bs−1∗ (∂Ω, σ) (9.5.7)

acting according to (recall that the summation convention over repeated indices is
in effect)
% A & % αβ &
p , q∗ 4 ∗ ∂ν (u, f ), ϕ
M p , q∗ 4 := a jk ∂k uβ , ∂ j Φα + f , Φ
M
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]

 p,q M
4  p,q M
4
for all ϕ ∈ B1−s ∗ (∂Ω, σ) and Φ = (Φα )α ∈ A1−s+1/p (Ω)
satisfying TrΩ→∂Ω Φ = ϕ,
(9.5.8)
where the first set of brackets in the right side of the first line above is understood
p,q p,q
as the duality pairing between As+1/p−1 (Ω) and A−s+1/p (Ω) (cf. (9.2.141)) and the
second set of brackets in the right side of the first line above is the canonical duality
/ 04  p,q M
4
p,q ∗ M
pairing between A1−s+1/p (Ω) and A1−s+1/p (Ω) .
Then the conormal derivative operator ∂νA considered in (9.5.6)-(9.5.8) is well
defined, linear, and bounded in the sense that there exists a constant C ∈ (0, ∞) with
the property that
 A   
∂ (u, f ) p, q∗ 4 ≤ C u p, q M +  f  p  , q  (9.5.9)
ν [B (∂Ω,σ)] M [A (Ω)] s+1/p [(A 4
(Ω))∗ ] M
s−1 1−s+1/p 

for each pair (u, f ) belonging to the domain of ∂νA (cf. (9.5.6)).
9.5 The Conormal Derivative Operator on Besov and Triebel-Lizorkin Spaces 595
 p,q M
Moreover, for any given pair (u, f ) with u = (uβ )β ∈ As+1/p (Ω) and with
/  
0M
4   4
p ,q ∗ M
f ∈ A1−s+1/p (Ω) such that L Au = f Ω in D (Ω) one has the following
generalized “half” Green’s formula:
% A &
p , q∗ M4 ∗ ∂ν (u, f ), TrΩ→∂Ω w p , q∗ M4 (9.5.10)
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]
% αβ &
= A p, q (Ω) a jk ∂k uβ, ∂j wα p , q 
A−s+1/p  (Ω)
s+1/p−1

% &
+ [(A p, q 4
(Ω))∗ ] M
f, w p , q  4
[A1−s+1/p  (Ω)] M
1−s+1/p 

 p,q M
4
for each w = (wα )α ∈ A1−s+1/p (Ω) .
 p,q  M / p,q 04
∗ M
Furthermore, for any two pairs, (u, f ) ∈ As+1/p (Ω) ⊕ A1−s+1/p (Ω)
 M
4  p,q M
4
/ 0
∗ M
with L Au = f Ω in D (Ω)
p,q
and (w, g) ∈ A1−s+1/p (Ω) ⊕ As+1/p (Ω)
 M
with L A w = gΩ in D (Ω) (where A is the transpose of A), one has the
following generalized “full” (or “symmetric”) Green’s formula:
% A &
p , q∗ 4 ∗ ∂ν (u, f ), TrΩ→∂Ω w
M p , q∗ 4
M
(9.5.11)
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]

%  &
− ([Bsp, q∗ (∂Ω,σ)] M )∗ ∂νA (w, g), TrΩ→∂Ω u [Bsp, q∗ (∂Ω,σ)] M
% & % &
4 f, w
= [(A p, q (Ω))∗ ] M [A
p , q  4 − [(A p, q (Ω))∗ ] M g, u [A p, q
(Ω)] M (Ω)] M .
1−s+1/p  1−s+1/p  s+1/p s+1/p

Proof Granted the present geometric assumptions, [133, (5.11.35)] guarantees that
rad (Ω) > 0 which, in concert with [133, Lemma 5.11.9)], further implies that Ω
satisfies an interior corkscrew condition. In particular, Ω is n-thick (cf. [133, (5.1.6)]).
 p,q M
4
We proceed to show that, given any ϕ ∈ B1−s ∗ (∂Ω, σ) , the action of the
functional (9.5.7) described in (9.5.8) does not depend on the particular choice of
 p,q M
4
Φ ∈ A1−s+1/p (Ω) with the property that TrΩ→∂Ω Φ = ϕ. To this end, let us
 p,q M4
suppose Φ j ∈ A1−s+1/p (Ω) is such that TrΩ→∂Ω Φ j = ϕ, for j ∈ {1, 2}. It
follows that
 p,q M
4
Φ := Φ1 − Φ2 belongs to A1−s+1/p (Ω) and TrΩ→∂Ω Φ = 0. (9.5.12)

We may then invoke item (iv) of Theorem 9.4.5 to conclude that there exists a
 M
sequence of (vector-valued) test functions, say Ψ(i) = (Ψα(i) )α ∈ Cc∞ (Ω) indexed
(i)
 p,q M
4
by i ∈ N, for which lim Ψ = Φ in the space A1−s+1/p (Ω) . Hence, with (Φα )α
i→∞
denoting the scalar components of Φ, we may write
596 9 Besov and Triebel-Lizorkin Spaces in Open Sets
% αβ & % &
p, q
A s+1/p−1 (Ω) a jk ∂k uβ, ∂j Φα p , q 
A−s+1/p  (Ω)
+ [(A p, q 4
(Ω))∗ ] M
f, Φ p , q  4
[A1−s+1/p  (Ω)] M
1−s+1/p 


% αβ &
= lim p, q
A s+1/p−1 (Ω) a jk ∂k uβ, ∂j Ψα(i) p , q 
A−s+1/p  (Ω)
i→∞

% (i)
&
+ [(A p, q 4
(Ω))∗ ] M
f, Ψ p , q  4
[A1−s+1/p  (Ω)] M
1−s+1/p 

 
%
αβ (i) & % (i)
&
= lim D  (Ω) a jk ∂k uβ, ∂ j Ψα D(Ω) + [D  (Ω)] M
4 LA u, Ψ 4
[D(Ω)] M
i→∞

 
% αβ & % &
= lim D  (Ω) − a jk ∂j ∂k uβ, Ψα(i) D(Ω) + [D  (Ω)] M
4 L A u, Ψ
(i)
4
[D(Ω)] M
i→∞

= lim 0 = 0, (9.5.13)
i→∞

on account of (9.2.141) and the convention made in (9.5.1). This proves that the
definition made in (9.5.8) is indeed coherent. Once this has been established, we
may then conclude that the conormal operator is well defined and linear in the
context of (9.5.6). To prove its boundedness, fix a an arbitrary pair (u, f ) belonging
 p,q  M / p,q 04
∗ M  M
4
to As+1/p (Ω) ⊕ A1−s+1/p (Ω) and satisfying L Au = f Ω in D (Ω) .
 p,q M
4
Also, pick an arbitrary ϕ ∈ B1−s ∗ (∂Ω, σ) . Then, as a consequence of (9.4.92),
(9.4.94), and the Open Mapping Theorem, it follows that there exists a function Φ
which satisfies
 p,q M
4
Φ ∈ A1−s+1/p (Ω) , TrΩ→∂Ω Φ = ϕ,
(9.5.14)
and Φ[A p, q 4
(Ω)] M
≤ Cϕ p , q  4
1−s+1/p  [B1−s ∗ (∂Ω,σ)] M

for some constant C ∈ (0, ∞) independent of ϕ. Relying on the definition from (9.5.8),
the properties listed in (9.5.14), the continuity of the bilinear form in (9.2.141), and
(9.2.8) we may then estimate
 % A & 
   
 p , q∗ 4 ∗
M
∂ν (u, f ), ϕ p  , q 
∗ 4
M
(9.5.15)
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]
 αβ 
≤ a jk  ∂k uβ A p, q (Ω) ∂ j Φα A p , q  (Ω)
s+1/p−1 −s+1/p 

+  f [(A p, q 4
(Ω))∗ ] M
Φ[A p, q 4
(Ω)] M
1−s+1/p  1−s+1/p 

 
≤ C u[A p, q (Ω)] M +  f [(A p, q 4
(Ω))∗ ] M
ϕ p , q∗ 4 .
s+1/p 1−s+1/p  [(B1−s (∂Ω,σ)] M
9.5 The Conormal Derivative Operator on Besov and Triebel-Lizorkin Spaces 597
 p,q M4
From this, the arbitrariness of ϕ ∈ B1−s ∗ (∂Ω, σ) , and Proposition 7.6.1 we then
conclude that (8.5.17) holds. Hence, the conormal operator is bounded in the context
of (9.5.6). Finally, the generalized “half” Green’s formula (9.5.10) is a consequence
of (9.5.8) used with ϕ := TrΩ→∂Ω w and Φ := w, while the “full” Green’s formula
(9.5.11) follows by subtracting two versions of (9.5.10) (one written for A, and one
written for A ). 

On occasions, we shall need the notion of conormal derivative in the more general
setting described in the next remark (and subsequent comments).

Remark 9.5.3 Retain the assumptions on Ω, A, p, q, s from Proposition 9.5.2, along

with the conventions made there. In addition, select some arbitrary cutoff function
ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near ∂Ω.
 p,q M
For each u = (uβ )1≤β ≤M ∈ As+1/p (Ω)bdd (cf. Convention 8.3.7) and each
 p,q ∗ M
4  M
4
f ∈ A1−s+1/p (Ω) with the property that L Au = f Ω in D (Ω) define

 p,q M
4
u, #
∂νA(u, f ) := ∂νA # f ∈ Bs−1∗ (∂Ω, σ) (9.5.16)

with
 p,q M  p,q ∗ M
4
#
u := ψu ∈ As+1/p (Ω) and #f ∈ A1−s+1/p (Ω) given by
 
# αβ αβ αβ
f := a jk (∂j ∂k ψ)uβ + a jk (∂j ψ)(∂k uβ ) + a jk (∂k ψ)(∂j uβ ) +ψf.
4
1≤α ≤ M
(9.5.17)
Then the above definition is meaningful, and does not depend on the particular cut-
 p,q M
4
off function ψ. Furthermore, for each function w = (wα )1≤α ≤ M 4 ∈ A1−s+1/p (Ω)
which vanishes outside a bounded subset of Ω the generalized “half” Green’s formula
% A &
p , q∗ 4 ∗ ∂ν (u, f ), TrΩ→∂Ω w
M p , q∗ M4
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]
% αβ &
= A p, q (Ω) a jk ψ∂k uβ, ∂j wα p , q 
A−s+1/p  (Ω)
s+1/p−1

% &
+ [(A p, q 4
(Ω))∗ ] M
f, w p , q  4
[A1−s+1/p  (Ω)] M
(9.5.18)
1−s+1/p 

holds for each cutoff function ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near both ∂Ω and the support
of w.
Finally, for each pair
 p,q M
4
/ 0
p,q ∗ M
(w, g) ∈ A1−s+1/p (Ω) ⊕ As+1/p (Ω)
 M (9.5.19)
with L A w = gΩ in D (Ω)

(where A is the transpose of A) and such that w vanishes outside a bounded subset
of Ω, the generalized “full” Green’s formula
598 9 Besov and Triebel-Lizorkin Spaces in Open Sets
% &
p , q  4 ∗ ∂νA(u, f ), TrΩ→∂Ω w p , q∗ 4
([B1−s ∗ (∂Ω,σ)] M ) [B1−s (∂Ω,σ)] M

%  &
− ([Bsp, q∗ (∂Ω,σ)] M )∗ ∂νA (w, g), TrΩ→∂Ω u [Bsp, q∗ (∂Ω,σ)] M
% &
4 f, w
= [(A p, q (Ω))∗ ] M [A
p , q  4
(Ω)] M
1−s+1/p  1−s+1/p 

% &
− [(A p, q (Ω))∗ ] M g, ψu p, q
[A s+1/p (Ω)] M (9.5.20)
s+1/p

holds for each cutoff function ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near both ∂Ω and the support
of w.

A few clarifications regarding Remark 9.5.3 are in order. First, the fact that u
 p,q M  p,q M
belongs to the space As+1/p (Ω)bdd puts the function #
u := ψu in As+1/p (Ω)
p,q
(cf. Convention 8.3.7). Second, since A1−s+1/p (Ω) is a module over Cc∞ (Rn ), so
p,q ∗  p,q ∗ M
4
is A1−s+1/p (Ω) , via duality. As such, we have ψ f ∈ A1−s+1/p (Ω) . In
addition, since from (9.2.7) and (9.2.9) we know that
p,q p,q
A1−s+1/p (Ω) → A−s+1/p (Ω) continuously and densely, (9.5.21)

Lemma 1.2.1 guarantees that

p,q ∗ p,q ∗
A−s+1/p (Ω) → A1−s+1/p (Ω) is a continuous embedding. (9.5.22)

Then, on account of (9.2.8), (9.2.140), and (9.5.22), we see that for each index
# } we have
α ∈ {1, . . . , M
αβ αβ αβ
a jk (∂j ∂k ψ)uβ + a jk (∂j ψ)(∂k uβ ) + a jk (∂k ψ)(∂j uβ ) belongs to
p,q p,q ∗ p,q ∗
(9.5.23)
As+1/p−1 (Ω) → A−s+1/p (Ω) → A1−s+1/p (Ω) .

 p,q ∗ M
4
Collectively, these properties show that we have # f ∈ A1−s+1/p (Ω) . It is
  M 4
also clear from definitions that L A#u= # f Ω in D (Ω) , in the sense of (8.5.13).
Together with Proposition 9.5.2, these observations show that the definition made in
(8.5.23) is indeed meaningful. Moreover, the fact that said definition is independent
of the cutoff function follows from the fact that, in the context of Proposition 9.5.2,
the conormal derivative ∂νA(u, f ) is the zero distribution whenever u vanishes near
∂Ω. Finally, the generalized “half” Green’s formula (9.5.18) is seen from (9.5.10)
and (9.5.16), while the generalized “full” Green’s formula (9.5.20) is implied by
(9.5.11) and (9.5.16).
The conormal derivative introduced in Proposition 9.5.2 may actually be com-
puted in a pointwise sense if the function in question is smooth. A concrete result to
this effect is given below.
9.5 The Conormal Derivative Operator on Besov and Triebel-Lizorkin Spaces 599

Proposition 9.5.4 Let Ω ⊆ Rn be a bounded (ε, δ)-domain with an Ahlfors regular

boundary. Set σ := H n−1 ∂Ω and denote by ν = (ν1, . . . , νn ) the geometric mea-
sure theoretic outward unit normal to Ω. Also, fix p, q ∈ (1, ∞), denote by p, q  the
Hölder conjugate exponents of p, q, pick some s ∈ (0, 1), and let A ∈ {B, F}. Next,
# M ∈ N and consider a coefficient tensor A = aαβ
let M, 4 with constant
jk 1≤α ≤ M
1≤β ≤M
1≤ j,k ≤n
(complex) entries. Associate with this coefficient tensor the second-order homoge-
neous constant (complex) coefficient M # × M system L A := aαβ ∂j ∂k 4 in R
n
jk 1≤α ≤ M
1≤β ≤M
(as usual, the summation convention over repeated indices is understood through-
 M / 0M
out). Finally, consider u = (uβ )1≤β ≤M ∈ C ∞ (Ω)
p,q
⊆ A 1 (Ω) and define
s+ p
/ 04
p,q ∗ M
f ∈ A1−s+1/p (Ω) by setting
∫
% &
p , q  4
[(A1−s+1/p  (Ω))∗ ] M
f, F p , q  4
[A1−s+1/p  (Ω)] M
:= (L Au)α Fα dL n
Ω
/ 0M
4
(9.5.24)
p,q
for each F = (Fα )1≤α ≤ M
4 ∈ A (Ω) .
1−s+ p1

 M
4  M
4
Then the function L Au ∈ C ∞ (Ω) satisfies L Au = f Ω in D (Ω) and
  
∂νA(u, f ) = 1∂∗ Ω · ν j a jk (∂k uβ )∂Ω
αβ
. (9.5.25)
4
1≤α ≤ M
 M
Proof That L Au = f Ω in D (Ω) is clear from (9.5.24) and definitions. To justify
(9.5.25), let us agree to set q∗ := q if A = B, and q∗ := p if A = F. Also, denote
by q∗ the Hölder conjugate exponent of q∗ . Finally, pick some arbitrary function
 M4  M
4
ϕ = (ϕα )1≤α ≤M ∈ Lip (∂Ω) and consider Φ = (Φα )1≤α ≤M ∈ Lipc (Rn )
such that Φ∂Ω = ϕ. We may then write
% &
p , q∗ 4 ∗ ∂νA(u, f ), ϕ p , q∗ 4
([B1−s (∂Ω,σ)] M ) [B1−s (∂Ω,σ)] M
∫ ∫
αβ
= a jk (∂k uβ )(∂j Φα ) dL n + (L Au)α Φα dL n
Ω Ω
∫

ν j a jk (∂k uβ )∂Ω ϕα dσ,
αβ
= (9.5.26)
∂∗ Ω

thanks to (9.5.8) in Proposition 9.5.2, the definition of f in (9.5.24), and the De

Giorgi-Federer version of the Divergence Formula from [133, Theorem 1.1.1]. Now
(9.5.25) follows from (9.5.26) and Lemma 7.1.10. 
Next we shall introduce and study the principal symbol map in the setting of Besov
and Triebel-Lizorkin spaces. In this regard, it is useful to recall the convention made
in (9.5.1).
600 9 Besov and Triebel-Lizorkin Spaces in Open Sets

Proposition 9.5.5 Suppose Ω ⊆ Rn is an (ε, δ)-domain with a compact Ahlfors

regular boundary, and abbreviate σ := H n−1 ∂Ω. Fix a number s ∈ (0, 1) along
with two exponents p, q ∈ (1, ∞), and denote by p , q  the Hölder conjugate exponents
of p, q. In addition, let A ∈ {B, F}, set q∗ := q if A = B and q∗ := p if A = F,
and denote by q∗ the Hölder conjugate exponent of q∗ . Finally, let D be an N × M
homogeneous first-order system with constant (complex) coefficients, say

n 
γβ
D= bk ∂k 1≤γ ≤ N . (9.5.27)
k=1 1≤β ≤M

In this setting, consider the map

  p,q M  p,q ∗ N
(−i)Sym(D; ν) : (u, f ) ∈ A 1 (Ω) ⊕ A (Ω) :
s+ p −1 1−s− p1

 N  p,q N
Du = f Ω in D (Ω) −→ Bs−1∗ (∂Ω, σ) (9.5.28)
 p,q M  p,q ∗ N
associating to each pair (u, f ) with u ∈ A 1 (Ω) and f ∈ A 1 (Ω)
s+ p −1 1−s− p 
 N
with the property that Du = f Ω in D (Ω) the functional
  N ∗  p,q N
p,q
(−i)Sym(D; ν)(u, f ) ∈ B1−s ∗ (∂Ω, σ) = Bs−1∗ (∂Ω, σ) (9.5.29)

acting according to
% & % &
p , q  (−i)Sym(D; ν)(u, f ), ϕ p , q  := f , Φ − u, D Φ
([B1−s ∗ (∂Ω,σ)] N )∗ [B1−s ∗ (∂Ω,σ)] N
 p,q N  p,q N
for all ϕ ∈ B1−s ∗ (∂Ω, σ) and Φ ∈ A (Ω) with TrΩ→∂Ω Φ = ϕ,
1−s+ p1
(9.5.30)
where the first set of brackets in the right-hand side of the first line above is the
 p,q ∗ N  p,q N
canonical duality pairing between A (Ω) and A (Ω) , while the
1−s+ p1 1−s+ p1
 p,q M
second set of brackets is understood as the duality pairing between A 1 (Ω)
s+ p −1
 p,q M
and A 1 (Ω) (cf. (9.2.141)).
−s+ p 
Then, in relation to this “symbol” map, the following statements are true.

(1) The operator (9.5.28)-(9.5.30) is well defined, linear, and bounded in the sense
that there exists a constant C ∈ (0, ∞) with the property that
 
(−i)Sym(D; ν)(u, f ) p, q∗ (9.5.31)
[B (∂Ω,σ)] Ns−1
 
≤ C u[A p, q (Ω)] M +  f [(A p, q (Ω))∗ ] N
1 −1
s+ p 1−s+ 1
p
9.5 The Conormal Derivative Operator on Besov and Triebel-Lizorkin Spaces 601

for each pair (u, f ) belonging to the domain of (−i)Sym(D; ν) (cf. (9.5.28)).
 p,q M  p,q ∗ N
(2) For each pair (u, f ) with u ∈ A 1 (Ω) and f ∈ A 1 (Ω) with
s+ p −1 1−s+ p 
  N
the property that Du = f Ω in D (Ω) one has the following generalized
integration by parts formula
% &
p , q∗ N ∗
(−i)Sym(D; ν)(u, f ), TrΩ→∂Ω w p, q∗ N
([B1−s (∂Ω,σ)] ) [B1−s (∂Ω,σ)]

% &  p,q N
= f , w − u, D w for each w ∈ A (Ω) ,
1−s+ p1
(9.5.32)

adopting the same conventions for the pairings appearing in the right-hand side
as those in effect for (9.5.30).
(3) In addition to the N × M system D from (9.5.27), consider a homogeneous,
# × N system D
constant (complex) coefficient, first-order M # in Rn , say


n 
#= #αγ
D b j ∂j 4
1≤α ≤ M
, (9.5.33)
j=1 1≤γ ≤ N

#
and define the homogeneous, constant (complex) coefficient, second-order M×M
system
#
L := DD. (9.5.34)
Also, define the coefficient tensor (with the summation convention over repeated
indices in effect)
αβ αβ αγ γβ
AD,D
# := a jk 1≤α ≤ M4 where each a jk := #
b j bk . (9.5.35)
1≤β ≤M
1≤ j,k ≤n

 p,q M  p,q ∗ M
4
Then for any pair (u, f ) with u ∈ A 1 (Ω) and f ∈ A (Ω) and
s+ p 1−s+ p1
 M
4
such that Lu = f Ω in D (Ω) one has
A D,
∂ν
4D
# ν)(Du, f ),
(u, f ) = (−i)Sym( D; (9.5.36)

where the conormal derivative in the left-hand side of (9.5.36) is defined in the
sense of Proposition 9.5.2, and the principal symbol map in the right-hand side
is defined as in (9.5.28) (with u replaced by Du and D replaced by D).#
Proof The claim in item (1) may be established analogously to the proof of Propo-
sition 9.5.2, while the claim in item (2) is implied by item (1). Finally, the claim in
item (3) may be justified based on (9.5.30) and (9.5.8). 
We close this section by presenting a Green-type integration by parts formula
involving a null-solution of a system L along with a null-solution of the transpose
602 9 Besov and Triebel-Lizorkin Spaces in Open Sets

system L  , belonging to certain Besov and Triebel-Lizorkin spaces involving dual

integrability exponents.

Corollary 9.5.6 Assume Ω ⊆ Rn is an (ε, δ)-domain with a compact Ahlfors regular

boundary, and abbreviate σ := H n−1 ∂Ω. Fix a number s ∈ (0, 1) along with two
exponents p, q ∈ (1, ∞), and denote by p, q  the Hölder conjugate exponents of p, q.
In addition, let A ∈ {B, F}, set q∗ := q if A = B and q∗ := p if A = F, and denote
by q∗ the Hölder conjugate exponent of q∗ . Next, consider a homogeneous, constant
(complex) coefficient, first-order M # × N system D # in Rn , along with a homogeneous,
constant (complex) coefficient, first-order N × M system D in Rn , and define the
homogeneous, constant (complex) coefficient, second-order M # × M system

#
L := DD. (9.5.37)

Then for any two vector-valued functions,

 p,q M
u ∈ A 1 (Ω) with Lu = 0 in Ω, and
s+ p
 p, q M
4 (9.5.38)
w∈ A 1 (Ω) with L  w = 0 in Ω,
1−s+ p 

one has
' (
p, q 4 # ν)(Du, 0), TrΩ→∂Ω w
(−i)Sym( D; p , q∗ (9.5.39)
[B s−1 ∗ (∂Ω,σ)] M [B1−s 4
(∂Ω,σ)] M
' (
# w, 0)
= [Bsp, q∗ (∂Ω,σ)] M TrΩ→∂Ω u, (−i)Sym(D ; ν)( D p , q∗ .
[B−s (∂Ω,σ)] M

Proof This is a consequence of (9.5.11) and (9.5.36). 

9.6 Extension from Boundary Besov Spaces into Smoothness

Spaces in Open Sets

The trace operator acting from Besov and Triebel-Lizorkin spaces defined in open
subsets of Rn onto theirs boundaries, introduced in Theorem 9.4.5, has a linear and
bounded inverse from the right in the geometric/analytic setting described below.
The reader is alerted to the fact that the proof of Theorem 9.6.1 makes essential
use of tools/results which are developed/proved later on (independently the present
considerations, of course).

Theorem 9.6.1 Suppose Ω ⊆ Rn (where n ∈ N, n ≥ 2) is a two-sided NTA domain

whose boundary is compact and Ahlfors regular, and abbreviate σ := H n−1 ∂Ω.
In particular Ω is an n-thick open set with an Ahlfors regular boundary, so it makes
sense to consider the boundary trace operators in the settings described in item (ii)
of Theorem 9.4.5. Then, in relation to these, the following statements are true.
9.6 Extension from Boundary Besov Spaces into Smoothness Spaces in Open Sets 603

(1) The boundary trace operator

p,q p,q
TrΩ→∂Ω : B (Ω) −→ Bs (∂Ω, σ) (9.6.1)
s+ p1

(cf. (9.4.91)) has a bounded linear right-inverse, i.e.,

p,q p,q
there exists Ex∂Ω→Ω : Bs (∂Ω, σ) → B (Ω), bounded linear
s+ p1 (9.6.2)
p,q
operator, so that TrΩ→∂Ω ◦ Ex∂Ω→Ω = I, the identity on Bs (∂Ω, σ),
whenever
n−1
n < p < ∞, (n − 1) 1
p −1 + < s < 1, 0 < q < ∞. (9.6.3)

(2) The boundary trace operator

p,q p, p
TrΩ→∂Ω : F (Ω) −→ Bs (∂Ω, σ) (9.6.4)
s+ p1

(cf. (9.4.93)) has a bounded linear right-inverse, i.e.,

p, p p,q
there exists Ex∂Ω→Ω : Bs (∂Ω, σ) → F (Ω), bounded linear op-
s+ p1 (9.6.5)
p,q
erator, so that TrΩ→∂Ω ◦ Ex∂Ω→Ω = I, the identity on Bs (∂Ω, σ),
whenever
n−1
n < p < ∞, (n − 1) 1
p −1 + < s < 1, n
n+s+1/p < q < ∞. (9.6.6)

Proof Define
Ω+ := Ω and Ω− := Rn \ Ω. (9.6.7)
Then [133, (5.11.66)] and the present assumptions imply that Ω± are (ε, δ)-domains
for some ε, δ > 0. Also, from [133, Lemma 5.10.9] and [133, Lemma 5.10.10] we
see that one of the sets Ω± is bounded, the other has a bounded complement (hence
is an exterior domain), and
∂(Ω+ ) = ∂Ω = ∂(Ω− ), and the geometric measure theoretic
(9.6.8)
outward unit normal to Ω± is ±ν at σ-a.e. point on ∂Ω,
where ν denotes the geometric measure theoretic outward unit normal to Ω. To fix
ideas, assume Ω+ is bounded (the case when Ω− is bounded is handled similarly).
Also, pick a cutoff function ψ ∈ Cc∞ (Rn ) with ψ ≡ 1 near ∂Ω, and denote by D±
the boundary-to-domain versions of the double layer potential operators associated
with the Laplacian in Ω+ and Ω− , respectively.
We are ready to address the claim made in item (1) in the statement of the theorem.
To get started, fix p, q, s as in (9.6.3) and define M := 2n−1
n−1 . Then

0 < p ≤ ∞, 0 < q ≤ ∞, n 1
p −1 + < s+ 1
p < M, (9.6.9)
604 9 Besov and Triebel-Lizorkin Spaces in Open Sets

so the conditions in (9.2.69) are valid with s replaced by s + p1 . This permits us to

bring in the linear and bounded extension operator
p,q p,q
EΩM− →Rn : B (Ω− ) −→ B (Rn ) (9.6.10)
s+ p1 s+ p1

as in part (i) of Theorem 9.2.15 with A := B, Ω replaced by Ω− , and s replaced by

s + p1 . The idea now is to define

Ex∂Ω→Ω := D+ − R Rn →Ω+ ◦ EΩM− →Rn ◦ (ψD− ) (9.6.11)

where ψ acts as pointwise multiplier, and R Rn →Ω+ is the restriction operator to Ω+

defined as in (9.2.5). From (9.6.11), item (1) in [136, Theorem 4.3.1] (used with
Ω replaced by both Ω+ and Ω− ), part (i) of Theorem 9.2.15 (used with A := B, Ω
replaced by Ω− , and s replaced by s + p1 ), and (9.2.6) (used with A := B, Ω replaced
by Ω− , and s replaced by s + p1 ), we conclude that
p,q p,q
Ex∂Ω→Ω : Bs (∂Ω, σ) −→ B (Ω)
s+ p1
(9.6.12)
is a well-defined, linear, and bounded mapping.

To check that the aforementioned mapping is indeed an extension operator, pick an

p,q
arbitrary function f ∈ Bs (∂Ω, σ) and abbreviate

u := EΩM− →Rn ψD− f . (9.6.13)

Then, as is apparent from the discussion in the build-up to (9.6.12), we have

p,q
u ∈ B 1 (Rn ). As such, with K denoting the boundary-to-boundary double layer
s+ p
potential operator associated with the Laplacian in Ω, we may write
 
TrΩ→∂Ω R Rn →Ω u = TrΩ →∂Ω u
+ + = TrΩ →∂Ω u
Ω+ − Ω−
  
= TrΩ− →∂Ω R Rn →Ω− EΩM− →Rn ψD− f

= TrΩ− →∂Ω ψD− f = ψ − 12 I + K f

= − 12 I + K f , (9.6.14)

where the second equality comes from Proposition 9.4.6, the penultimate equality
is implied by item (1) in [136, Theorem 4.3.2] (invoked with Ω− in place of Ω; the
last property in (9.6.8) is responsible for the change in the sign of the jump-term),
and last equality takes into account the fact that ψ ≡ 1 on ∂Ω. Having established
(9.6.14), we may now compute

TrΩ→∂Ω Ex∂Ω→Ω f = TrΩ→∂Ω (D+ f ) − TrΩ→∂Ω u

= 1
2I + K f − − 12 I + K f = f . (9.6.15)
9.6 Extension from Boundary Besov Spaces into Smoothness Spaces in Open Sets 605

This finishes the proof of (9.6.2). The proof of (9.6.5) is similar (most notably,
now item (2) of [136, Theorem 4.3.1] is involved, and the assumptions in (9.6.6)
guarantee that the conditions in (9.2.70) are satisfied) and this concludes the proof
of the theorem. 
Chapter 10
Strong and Weak Normal Boundary Traces of
Vector Fields in Hardy and Morrey Spaces

The notion of normal boundary trace of a vector field F defined in an open set Ω ⊆ Rn
 n.t. 
is a phrase of relative meaning. For example, we may interpret this as ν · F∂Ω
assuming that Ω is an open set of locally finite perimeter, with geometric outward unit
normal ν, and assuming that the nontangential boundary trace of F is meaningfully
defined. We shall refer to this interpretation as the “strong” (or “pointwise”) normal
boundary trace of F.  On the other hand, under suitable integrability assumptions on

F, we may consider the bullet product ν • F, in the sense of [133, Proposition 4.2.3],
and refer to this as the the “weak” (or “distributional”) normal boundary trace of F. 
In this chapter we shall study both the strong and weak version of the normal trace,
seeking conditions that guarantee membership to Hardy spaces on ∂Ω.

10.1 Strong Normal Boundary Traces of Vector Fields in Hardy

Spaces

Let Ω ⊆ Rn be a nonempty open set with the property that

 
0 < H n−1 ∂Ω ∩ B(x, r) < +∞ for each x ∈ ∂Ω and r > 0, (10.1.1)

and abbreviate σ := H n−1 ∂Ω. In this measure-geometric setting, we shall associate

with every given L n -measurable function u : Ω → C what we shall refer to as its
P-maximal function, defined at every point x ∈ ∂Ω as
 ∫ 
1
(Pu)(x) := sup   |u| dL n ∈ [0, ∞]. (10.1.2)
0<r <2 diam(∂Ω) σ ∂Ω ∩ B(x, r) Ω∩B(x,r)

For example, having Pu bounded on ∂Ω amounts to saying that |u|L n is a Carleson

measure in Ω. The maximal operator P is related to the Carleson maximal operator
defined in [33, (2.5), p. 308] when Ω = R+n , and subsequently considered on spaces
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 607
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_10
608 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

of homogeneous type in [95]. A few of its basic properties are recorded below (the
p
reader is reminded that the notation Lbdd has been introduced in (A.0.55)).

Lemma 10.1.1 Let Ω ⊆ Rn be a nonempty open set satisfying (10.1.1) and define
σ := H n−1 ∂Ω. Then
for any given L n -measurable u : Ω → C, the
(10.1.3)
function Pu : ∂Ω → [0, ∞] is σ-measurable.
Also,
 
if u ∈ Lloc
1 (Ω, L n ) is such that Pu (x ) < ∞ for some
0
(10.1.4)
point x0 ∈ ∂Ω, then necessarily u lies in Lbdd
1 (Ω, L n ).

Moreover, in a quantitative fashion,

Pu ∈ L ∞ (∂Ω, σ) whenever ∂Ω is a lower

(10.1.5)
Ahlfors regular set and u is in L n (Ω, L n ),
and
if ∂Ω is lower Ahlfors regular and u ∈ L 1 (Ω, L n ) vanishes outside a
bounded subset of Ω, then
 there exist x0 ∈ ∂Ω along with R > 0 and (10.1.6)
C ∈ (0, ∞) so that Pu (x) ≤ C(1 + |x|)−(n−1) for all x ∈ ∂Ω \ B(x0, R).

Henceforth, strengthen (10.1.1) by assuming that σ is a doubling measure on ∂Ω,

and fix an aperture parameter κ > 0. Then there exists some finite constant C > 0
with the property that for each given L n -measurable function u : Ω → C one has
 
Pu ≤ C · M ∂Ω, n−1 , 1 Nκ u pointwise on ∂Ω, (10.1.7)
n n−1

where M ∂Ω, n−1 , 1 is the L (n−1)/n -based fractional Hardy-Littlewood maximal op-
n n−1
erator of order 1/(n−1) defined as in (A.0.68) relative to the space space of homoge-
neous type (∂Ω, | · − · |, σ). Consequently, in such a setting, for each L n -measurable
function u : Ω → C satisfying

Nκ u ∈ L p,q (∂Ω, σ) with n−1

n < p < n − 1 and 0 < q ≤ ∞, (10.1.8)

one has, for some constant C ∈ (0, ∞) independent of u,

∗
Pu ∈ L p ,q (∂Ω, σ) and Pu L p ∗ , q (∂Ω,σ) ≤ C Nκ u L p, q (∂Ω,σ),
 1 −1
 (10.1.9)
where p∗ := p1 − n−1 ∈ (1, ∞).

Also, corresponding to the limiting case p = n−1

n ,

n−1
if Nκ u ∈ L n (∂Ω, σ) then Pu ∈ L 1,∞ (∂Ω, σ)
(10.1.10)
and Pu L 1,∞ (∂Ω,σ) ≤ C Nκ u n−1 .
L n (∂Ω,σ)
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 609

Finally, another consequence of (10.1.7) is that for each L n -measurable function

u : Ω → C such that
Nκ u belongs to the Morrey space M p,λ (∂Ω, σ)
(10.1.11)
(n−1)2
with 0 < λ < n and n−1
n < p < n − 1 − λ,

there exists a constant C ∈ (0, ∞) independent of u with the property that

Pu ∈ M p,λ (∂Ω, σ) and Pu M p, λ (∂Ω,σ) ≤ C Nκ u M p, λ (∂Ω,σ),

  −1 (10.1.12)
where p := p1 − n−1−λ
1
∈ (1, ∞).

Proof The claim in (10.1.3) may be established by reasoning much as in the case
of the Hardy-Littlewood maximal function, treated in [9, pp. 82-84]. The claim
in (10.1.4) is an easy consequence of (10.1.1)-(10.1.2) and (A.0.55). Let us now
temporarily work under the assumption that ∂Ω is a lower Ahlfors regular set and
that u ∈ L n (Ω, L n ). Then, at every point x ∈ ∂Ω, Hölder’s inequality permits us to
estimate
 ∫ 
1
0 ≤ (Pu)(x) ≤ C · sup n−1
|u| dL n
0<r <2 diam(∂Ω) r Ω∩B(x,r)

 ∫ 1/n  ∫ 1−1/n 
1
≤C· sup |u| n dL n dL n
0<r <2 diam(∂Ω) r n−1 Ω B(x,r)

=C· u L n (Ω, L n ), (10.1.13)

for some finite constant C > 0 which depends only on n and the lower Ahlfors
regularity character of ∂Ω. In view of (10.1.3) and the arbitrariness of the point
x ∈ ∂Ω, we ultimately conclude that Pu belongs to L ∞ (∂Ω, σ) and

Pu L ∞ (∂Ω,σ) ≤C u L n (Ω, L n ) , for some C = C(Ω) ∈ (0, ∞), (10.1.14)

proving (10.1.5). Next, to justify the claim made in (10.1.6), suppose

∂Ω is lower Ahlfors regular and u ∈ L 1 (Ω, L n ) is such that u ≡ 0 at

(10.1.15)
L n -a.e. point in Ω \ B(x0, R/4), for some x0 ∈ ∂Ω and R ∈ (0, ∞).

The goal is to prove that there exists C = C(Ω, n, u, R) ∈ (0, ∞) such that
  C
Pu (x) ≤ for each x ∈ ∂Ω \ B(x0, R). (10.1.16)
1 + |x − x0 | n−1

Indeed, given any x ∈ ∂Ω \ B(x0, R) we have

∫
1
  |u| dL n = 0 if 0 < r < |x − x0 | − R/4, (10.1.17)
σ ∂Ω ∩ B(x, r) Ω∩B(x,r)
610 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

and, if r ≥ |x − x0 | − R/4,
∫ ∫
1
  |u| dL n ≤ Cr −(n−1) |u| dL n
σ ∂Ω ∩ B(x, r) Ω∩B(x,r) Ω∩B(x0,R)

  −(n−1)
≤ C 1 + |x − x0 | u L 1 (Ω, L n ) . (10.1.18)
 
Taking the supremum over all r ∈ 0, 2 diam(∂Ω) , ultimately establishes (10.1.16),
from which the claim in (10.1.6) follows.
Next, consider the task of proving (10.1.7) under the assumption that σ is a
doubling measure on ∂Ω. Having fixed an aperture parameter κ > 0, and having
picked some L n -measurable
 function u : Ω → C, for an arbitrary point x ∈ ∂Ω and
an arbitrary radius r ∈ 0, 2 diam(∂Ω) we may then estimate
∫
1
  |u| dL n
σ ∂Ω ∩ B(x, r) Ω∩B(x,r)
∫
1   n−1 n

≤   Nκ u n dσ n−1
σ ∂Ω ∩ B(x, r) πκ (Ω∩B(x,r))

  n ⨏
σ ∂Ω ∩ B(x, (2 + κ)r) n−1   n−1 n

≤   Nκ u n dσ n−1
σ ∂Ω ∩ B(x, r) B(x,(2+κ)r)∩∂Ω
⨏
  1   n−1 n

≤ C · σ ∂Ω ∩ B(x, (2 + κ)r) n−1 Nκ u n dσ n−1

B(x,(2+κ)r)∩∂Ω

 
≤ C · M ∂Ω, n−1 , 1 Nκ u (x), (10.1.19)
n n−1

thanks to [133, Proposition 8.6.3], [133, (8.1.17)], the doubling property of σ, and
(A.0.67). Taking the supremum over r and keeping (10.1.2) in mind then yields
 
(Pu)(x) ≤ C · M ∂Ω, n−1 , 1 Nκ u (x), from which (10.1.7) follows. Going further,
n n−1
the claims in (10.1.8)-(10.1.9) are consequences of (10.1.7), (10.1.3), and [133,
Theorem 7.6.1], bearing [133, (6.2.16)] in mind.
Finally, the claim in (10.1.11)-(10.1.12) is a consequence of (10.1.7) and the
mapping properties of the Hardy-Littlewood maximal operator on Morrey spaces
established in Corollary 6.2.15. 

The above considerations are relevant in the statement of our first major result in
this section, presented below.

Theorem 10.1.2 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an Ahlfors regular domain

(i.e., Ω is a nonempty open subset of Rn such that ∂Ω is an Ahlfors regular set and
H n−1 (∂Ω \ ∂∗ Ω) = 0). As a consequence, Ω is a set of locally finite perimeter whose
geometric measure theoretic outward unit normal ν is defined almost everywhere on
∂Ω with respect to the measure σ := H n−1 ∂Ω.
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 611

In this context, consider a vector field F : Ω → Cn , having L n -measurable

components, with the property that for some aperture parameter κ ∈ (0, ∞) one has
κ−n.t.
Nκ F ∈ Lloc
1
(∂Ω, σ) and F∂Ω exists σ-a.e. on ∂Ω. (10.1.20)

In particular, F ∈ Lloc1 (Ω, L n ) n (cf. [133, Lemma 8.3.1]), and one also assumes

that the divergence of F, computed in the sense of distributions in Ω, is of function

type, i.e., it satisfies
divF ∈ Lloc
1
(Ω, L n ). (10.1.21)
Finally, fix some γ ∈ (0, 1) and select some power
n−1+γ
α ∈ 0, . (10.1.22)
n−1
 κ−n.t. 
Then the Fefferman-Stein grand maximal function of ν · F∂Ω

 κ−n.t.  
ν · F∂Ω : ∂Ω −→ [0, ∞] is lower-semicontinuous, (10.1.23)
γ

(when ∂Ω is equipped with the topology inherited from Rn ), and the following
pointwise inequality, also involving the nontangential maximal operator, the Hardy-
Littlewood maximal operator M ∂Ω on ∂Ω (cf. (A.0.71)), and the P-maximal operator
defined in (10.1.2), holds:
 κ−n.t.     
1  α  
ν · F∂Ω ≤ C M ∂Ω |Nκ F | α + C P divF
γ (10.1.24)
pointwise on ∂Ω, for some C = (∂Ω, κ, n, γ, α) ∈ (0, ∞).

It is worth pointing out that for any

n−1
s∈ ,∞ (10.1.25)
n−1+γ

we may recast (10.1.24) (with α := 1/s) as

 κ−n.t.      
ν · F∂Ω ≤ CM ∂Ω,s Nκ F + C P divF on ∂Ω, (10.1.26)
γ

where M ∂Ω,s is the L s -based Hardy-Littlewood maximal operator on ∂Ω (see

(A.0.69)). The proof of Theorem 10.1.2 is going to be presented shortly. For the
time being, we are going to deal with the extension result described in the lemma
below, which plays an important role in this endeavor.

Lemma
 10.1.3 Let Ω be an arbitrary open set in Rn . Then for each x ∈ ∂Ω and
r ∈ 0, 2 diam(∂Ω) fixed there exists a linear mapping C 0 (∂Ω)  ψ → ψ  ∈ C 0 (Ω)
with the property that if
612 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
.γ
ψ ∈ C (∂Ω) for some γ ∈ (0, 1] and supp ψ ⊆ ∂Ω ∩ B(x, r) (10.1.27)

then
. 
 ∈ C ∞ (Ω) ∩ C γ (Ω),
ψ 
ψ = ψ,  ⊆ B(x, 2r) ∩ Ω,
supp ψ
∂Ω
  c
1−γ | ≤ c ψ .γ
sup δ∂Ω |∇ψ + γ · sup |ψ| (10.1.28)
C (∂Ω) r
Ω ∂Ω
| ≤ sup |ψ|,
and sup | ψ
Ω ∂Ω

where the constant c ∈ (0, ∞) depends only on n and γ.

 
Proof Having fixed
 a point
 x ∈ ∂Ω and a scale r ∈ 0, 2 diam(∂Ω) , select a cutoff
function ϕ ∈ Cc∞ B(x, 2r) satisfying
cn
0 ≤ ϕ ≤ 1, ϕ ≡ 1 on B(x, r), and sup |(∇ϕ)(x)| ≤ , (10.1.29)
x ∈R n r

where cn ∈ (0, ∞) is a purely dimensional finite constant. Then, given any function
ψ ∈ C 0 (∂Ω) define 
 := ϕ(Eψ) ∈ C 0 (Ω)
ψ (10.1.30)
Ω
where E is Whitney’s extension operator [133, (6.1.9)] associated with the closed
subset F := ∂Ω of Rn . It is then clear from [133, Theorem
. 6.1.3] that whenever the
function ψ is an in (10.1.27) we have ψ  ⊆ B(x, 2r) ∩ Ω,
 ∈ C ∞ (Ω) ∩ C γ (Ω), supp ψ

and ψ = ψ. Also, by virtue of (10.1.30), (10.1.29), [133, Theorem 6.1.3], and
∂Ω
(10.1.27) we have
| ≤ sup |Eψ| ≤ sup |ψ|.
sup | ψ (10.1.31)
Ω Ω ∂Ω

It remains to observe that, thanks to [133, Theorem 6.1.3] and (10.1.27), in a point-
wise sense in Ω we have
1−γ | ≤ δ
δ∂Ω |∇ψ
1−γ
|ϕ||∇(Eψ)| + δ∂Ω |∇ϕ||Eψ|
1−γ
∂Ω

cn 1−γ
≤ c ψ C.γ (∂Ω) + · sup δ∂Ω · sup |Eψ|
r Ω∩B(x,2r) Rn

c
≤ c ψ C.γ (∂Ω) + · sup |ψ| , (10.1.32)
rγ ∂Ω

where the constant c ∈ (0, ∞) intervening above depends only on n and γ. This
 as in (10.1.28). Clearly, the
finishes the proof of the existence of a function ψ
 is linear.
assignment ψ → ψ 

We are now ready to present the actual proof of Theorem 10.1.2.

Proof of Theorem 10.1.2 For starters we note that the current assumptions ensure
 κ−n.t. 
(in view of [133, Corollary 8.9.6] and [133, (5.6.23)]) that ν · F∂Ω belongs
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 613
 
1 (∂Ω, σ) ⊆ Lip (∂Ω)  . As a consequence, it makes sense to consider the
to Lloc c
 κ−n.t. 
Fefferman-Stein grand maximal function of ν · F∂Ω , regarded as a functional in
 
Lipc (∂Ω) . In addition, from (4.1.22) we know that the claim made in (10.1.23) is
true.  
The inequality recorded in (10.1.24) is trivially true if P divF (x) = ∞ for each
x ∈ ∂Ω. There remains to treat the case when there exists a point x0 ∈ ∂Ω such that
P divF (x0 ) < ∞. In such a scenario, (10.1.4) implies that

divF ∈ Lbdd
1
(Ω, L n ). (10.1.33)

With the goal of proving the veracity of the inequality in (10.1.24) everywhere
on ∂Ω, fix a point x ∈ ∂Ω and consider an arbitrary function
 ψ ∈ Tγ (x). Hence,
ψ ∈ Lipc (∂Ω) and there exists some r ∈ 0, 2 diam(∂Ω) such that (cf. (4.1.1))

1
supp ψ ⊆ ∂Ω ∩ B(x, r), sup |ψ| ≤  ,
∂Ω σ ∂Ω ∩ B(x, r)
(10.1.34)
r −γ
and ψ C.γ (∂Ω) ≤  .
σ ∂Ω ∩ B(x, r)

Granted these, Lemma 10.1.3 ensures the existence of a function


ψ  = ψ, supp ψ
 ∈ C ∞ (Ω) ∩ C 0 (Ω) with ψ  ⊆ B(x, 2r) ∩ Ω,
∂Ω
  c r −γ
| ≤
1−γ
sup δ∂Ω |∇ψ  , (10.1.35)
Ω σ ∂Ω ∩ B(x, r)
| ≤  1
and sup | ψ ,
Ω σ ∂Ω ∩ B(x, r)

where the constant c ∈ (0, ∞) depends only on n and γ.

Going forward, we write
   κ−n.t.     ∫  κ−n.t.   ∫  κ−n.t.  
         
 F)
 ν · F ∂Ω , ψ  =  ν · F ∂Ω ψ dσ  =  ν · (ψ ∂Ω
dσ 
∂Ω ∂Ω
 ∫    
  dL n 
= divF ψ + F · ∇ψ
Ω
∫ ∫
   
≤ divF | ψ
| dL n + F |∇ψ
| dL n =: I + I I. (10.1.36)
Ω Ω

Above, the left-most side involves the duality pairing between the test function
 κ−n.t.   
1 (∂Ω, σ) ⊆ Lip (∂Ω)  . Since
ψ ∈ Lipc (∂Ω) and the functional ν · F∂Ω ∈ Lloc c
this pairing is compatible with the ordinary integral pairing on ∂Ω, with respect
to the measure σ, the first equality in (10.1.36) follows. The second equality in
(10.1.36) is clear from the first two properties in (10.1.35).
614 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

The third equality in (10.1.36) is in many ways the most delicate step, as it
ultimately rests on the Divergence Formula from [133, Theorem 1.2.1]. To justify it,
 F belongs to L 1 (Ω, L n ) n and vanishes outside
observe first that the vector field ψ loc
of a bounded subset of Ω. Also, in view of (10.1.35), [133, (8.2.14)], and [133,
(8.1.17)], we may estimate

 F)
Nκ (ψ | · 1πκ (B(x,2r)∩Ω) · Nκ F
 ≤ sup | ψ
Ω

1 
≤   · 1B(x,(4+2κ)r)∩∂Ω · Nκ F. (10.1.37)
σ ∂Ω ∩ B(x, r)

In turn, from this, the first property in (10.1.20), and [133, (8.2.28)], we conclude
that
Nκ (ψ  ∈ L 1 (∂Ω, σ).
 F) (10.1.38)
As regards the divergence of the vector field ψ  since ψ
 F,  ∈ C ∞ (Ω), we may compute
 
div(ψ  = divF ψ
 F)  + F · ∇ψ
 in D (Ω). (10.1.39)

Thanks to (10.1.33) and the first line in (10.1.35) (which implies that ψ  is
continuous
  and compactly
 supported in Ω) it follows that, on the one hand,
divF ψ ∈ L 1 Ω, L n . On the other hand, the last property in (10.1.35) implies
that, for some finite constant C(Ω, x, r) > 0, we have
γ−1
| F · ∇ψ
| ≤ C(Ω, x, r)1Ω∩B(x,2r) · δ
∂Ω
· | F | in Ω. (10.1.40)

As such, in order to ensure that [133, Theorem 1.2.1] is presently applicable, it

suffices to check that
γ−1  
δ∂Ω · | F | belongs to L 1 Ω ∩ B(x, 2r), L n . (10.1.41)

However, [133, Proposition 8.6.15] applied with Σ := ∂Ω, E := Ω ∩ B(x, 2r),

u := | F |, β := n−1, α := 1 (which forces the parameter n−αβ to be 1), λ := 1−γ < 1,
and p := 1, allows us to estimate (bearing in mind [133, (8.1.17)])
∫ ∫
γ−1  
γ
δ ∂Ω
· F dL n ≤ sup δ∂Ω · Nκ F dσ
Ω∩B(x,2r) Ω∩B(x,2r) B(x,(4+2κ)r)∩∂Ω

 
≤ (2r)γ · Nκ F L 1 (B(x,(4+2κ)r)∩∂Ω,σ) < +∞. (10.1.42)
   
This finishes the proof of (10.1.41), hence div ψ F ∈ L 1 Ω, L n . Having established
this, the justification of the third equality in (10.1.36) is complete. The subsequent
inequality in (10.1.36) is obvious, while the last equality is merely a definition.
To estimate term I defined in the last part of (10.1.36), we write
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 615
∫ ∫
   
I= divF | ψ
| dL n = divF | ψ
| dL n
Ω Ω∩B(x,2r)
∫
1    
≤   divF dL n ≤ C P divF (x), (10.1.43)
σ ∂Ω ∩ B(x, r) Ω∩B(x,2r)

for some finite constant C = C(Ω) > 0, thanks to (10.1.35) and (10.1.2).
To estimate term I I defined in the last part of (10.1.36), the idea is to invoke [133,
Proposition 8.6.15], for the closed set Σ := ∂Ω, the measure σ := H n−1 ∂Ω, the
aperture parameter κ as in the current statement, the powers β := n − 1 and
 n−1+γ
α ∈ 1, , (10.1.44)
n−1
the parameter λ := 1 − γ < n − α(n − 1), the integrability exponent p := 1,
the Lebesgue measurable set E := Ω ∩ B(x, 2r) ⊆ Rn \ Σ, and the Lebesgue
measurable function u := | F | restricted to E. Let us recall from [133, (8.1.17)] that
the aforementioned choice of E implies
 
πκ (E) ⊆ ∂Ω ∩ B x, (4 + 2κ)r . (10.1.45)

With M ∂Ω denoting the Hardy-Littlewood maximal operator on ∂Ω (cf. (A.0.71)),

we then have
∫
c γ−1
II ≤ γ   δ · | F | dL n
r · σ ∂Ω ∩ B(x, r) Ω∩B(x,2r) ∂Ω
∫
c γ−1  
=   δ∂Ω · 1Ω∩B(x,2r) F dL n
r γ · σ ∂Ω ∩ B(x, r) Ω

c n−α(n−1)−1+γ
≤   sup δ∂Ω ×
r γ · σ ∂Ω ∩ B(x, r) Ω∩B(x,2r)
∫
 Ω∩B(x,2r)  1 α
× Nκ F α dσ
∂Ω∩B(x,(4+2κ)r)
⨏
 Ω∩B(x,2r)  1 α
≤C Nκ F α dσ
∂Ω∩B(x,(4+2κ)r)

 1 
α
≤ C M ∂Ω |Nκ F | α (x) . (10.1.46)

The first inequality in (10.1.46) is a direct consequence of (10.1.35), while the

subsequent equality is obvious. The second inequality in (10.1.46) is provided by
[133, Proposition 8.6.15] (applied here in the manner specified in the build up to
(10.1.46)). This step also takes into account the inclusion in (10.1.45). The penul-
timate inequality in (10.1.46) uses the Ahlfors regularity of ∂Ω together with the
fact that we have n − α(n − 1) − 1 + γ > 0, and the last inequality is clear from the
616 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

definition of the Hardy-Littlewood maximal operator on ∂Ω (cf. (A.0.71)) and [133,

(8.2.14)]. At this point, all steps in (10.1.46) have been justified.
In light of the arbitrariness of ψ ∈ Tγ (x) and x ∈ ∂Ω, from (4.1.6), (10.1.36),
(10.1.43), and (10.1.46) we ultimately deduce that the pointwise inequality claimed
in (10.1.24) holds if α is as in (10.1.44). Finally, that (10.1.24) continues to be true
when α belongs to the larger range specified in (10.1.22) is a consequence of what
of what we have just proved, (A.0.71), and Hölder’s inequality. 

One of the basic consequences of Theorem 10.1.2 is the following strong trace
result for the normal component of a vector field, in the Hardy space H 1 .

Theorem 10.1.4 Assume Ω ⊆ Rn (where n ∈ N with n ≥ 2) is an Ahlfors regular

domain (that is, Ω is a nonempty open subset of Rn such that ∂Ω is an Ahlfors regular
set and H n−1 (∂Ω\∂∗ Ω) = 0). In particular, Ω is a set of locally finite perimeter whose
geometric measure theoretic outward unit normal ν is defined almost everywhere on
∂Ω with respect to the measure σ := H n−1 ∂Ω.
Having fixed some κ ∈ (0, ∞), consider a vector field F : Ω → Cn , having
L -measurable components, with the property that
n

κ−n.t.
Nκ F ∈ L 1 (∂Ω, σ) and F∂Ω exists σ-a.e. on ∂Ω. (10.1.47)

As a consequence, F ∈ Lloc1 (Ω, L n ) n (cf. [133, Lemma 8.3.1]), and one also

 computed in the sense of distributions in Ω, belongs to L 1 (Ω, L n )

assumes that div F, loc
and satisfies  
P divF ∈ L 1 (∂Ω, σ). (10.1.48)
Then
 κ−n.t. 
ν · F∂Ω belongs to the Hardy space H 1 (∂Ω, σ) (10.1.49)

and there exists a constant CΩ,κ ∈ (0, ∞) such that

  κ−n.t.       
 
ν · F∂Ω  1 ≤ CΩ,κ Nκ F L 1 (∂Ω,σ) + P(divF)
 1
L (∂Ω,σ)
. (10.1.50)
H (∂Ω,σ)

As a corollary, the conclusions in (10.1.49)-(10.1.50) hold if in place of (10.1.48)

one assumes that F is actually divergence-free in the sense of distributions in Ω (a
scenario in which the last norm in (10.1.50) may be dropped).
 κ−n.t. 
The conditions in (10.1.47) imply that the function ν · F∂Ω is well defined at
σ-a.e. point on ∂Ω, and that
 κ−n.t.    κ−n.t.    
 
ν · F∂Ω ∈ L 1 (∂Ω, σ) and ν · F∂Ω  1 ≤ Nκ F L 1 (∂Ω,σ) . (10.1.51)
L (∂Ω,σ)

Compared to these, the membership in (10.1.49) and the estimate in (10.1.50) are
definite improvements, given that H 1 (∂Ω, σ) is a proper subspace of L 1 (∂Ω, σ). A
remarkable byproduct of (10.1.49) and (4.4.9) is the fact that
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 617
∫
 κ−n.t. 
ν · F∂Ω dσ = 0 whenever ∂Ω is unbounded. (10.1.52)
∂Ω

When F is actually divergence-free in Ω, the vanishing property recorded in (10.1.52)

is a direct consequence of the Divergence Formula from [133, Theorem 1.2.1]. In
the more general case treated in Theorem 10.1.4 there are no guarantees that [133,
(1.2.2)] works, as our assumptions on divF (leading up to, and including, (10.1.48)),
are insufficient in this regard.
Several special cases of Theorem 10.1.4 worth discussing separately are as fol-
lows. First, with Ω ⊆ Rn an Ahlfors regular domain, consider the scenario in which
the vector field has the form F := ∇u for some scalar-valued function
u ∈ C ∞ (Ω) which is harmonic in Ω and, for some κ > 0, has the
property that Nκ (∇u) ∈ L 1 (∂Ω, σ) and the nontangential trace (10.1.53)
κ−n.t.
(∇u)∂Ω exists at σ-a.e. point on the topological boundary ∂Ω.

Then (10.1.49) gives that the normal derivative

 κ−n.t. 
∂ν u := ν · (∇u)∂Ω belongs to the Hardy space H 1 (∂Ω, σ), (10.1.54)

plus a naturally accompanying estimate. The particular case of (10.1.53)-(10.1.54)

when Ω is a Lipschitz domain has been treated by B. Dahlberg and C. Kenig in [41,
Lemma 2.10, p. 446], using a conceptually different approach (based on duality and
Varopoulos’ extension theorem from [196]), which does not necessarily extend to
the present, more general, setting.
Second, continue to retain the assumption that Ω ⊆ Rn is an Ahlfors regular
domain, and consider a scalar-valued function u which, for some κ > 0, satisfies
1,1
u ∈ Wloc (Ω) such that Nκ (∇u) ∈ L 1 (∂Ω, σ)
κ−n.t. (10.1.55)
and (∇u)∂Ω exists at σ-a.e. point on ∂Ω.

Then for each pair of indices j, k ∈ {1, . . . , n}, Theorem 10.1.4 applied to the
divergence-free vector field Fjk
u := (∂ u)e − (∂ u)e ∈ L 1 (Ω, L n ) n gives that, in
k j j k loc
a quantitative fashion,
κ−n.t. κ−n.t.
ν j (∂k u)∂Ω − νk (∂j u)∂Ω belongs to the Hardy space H 1 (∂Ω, σ). (10.1.56)

The vector field in Theorem 10.1.4 may be allowed to take values in a Clifford
algebra. Here is a case of interest. Consider two functions u, w ∈ C ∞ (Ω) ⊗ Cn
satisfying, for some κ > 0,
618 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Nκ u < ∞, Nκ w < ∞ at σ-a.e. point on ∂Ω,

Nκ u · Nκ w belongs to the space L 1 (∂Ω, σ),
κ−n.t. κ−n.t. (10.1.57)
u∂Ω and w ∂Ω exist σ-a.e. on ∂Ω,

DR u = 0 in Ω, and D L w = 0 in Ω,

where DR denotes the Dirac operator acting from the right (as in (A.0.33)), and
D L denotes the Dirac operator acting from the left (as in (A.0.32)). Granted these
properties, Theorem 10.1.4 applied to F = (Fj )1≤ j ≤n with each Fj := u  e j  w
then implies that there exists some C ∈ (0, ∞) independent of u, w such that
 κ−n.t.   κ−n.t. 
u∂Ω  ν  w ∂Ω ∈ H 1 (∂Ω, σ) ⊗ Cn and
  κ−n.t.   κ−n.t.     (10.1.58)
 
 u ∂Ω  ν  w ∂Ω  1 ≤ C Nκ u · Nκ w  L 1 (∂Ω,σ) .
H (∂Ω,σ)⊗ C n

Other applications are discussed later in this section. We now present the proof
of Theorem 10.1.4.
Proof of Theorem 10.1.4 Fix some γ ∈ (0, 1). In view of (4.2.6) and the present
assumptions, all desired claims follow once we show that
 κ−n.t.  
ν · F∂Ω ∈ L 1 (∂Ω, σ) (10.1.59)
γ

and there exists a constant CΩ,κ,γ ∈ (0, ∞) such that

  κ−n.t.        
 
 ν · F∂Ω  ≤ CΩ,κ,γ Nκ F L 1 (∂Ω,σ) + P(divF)
 1
L (∂Ω,σ)
.
γ L 1 (∂Ω,σ)
(10.1.60)
To this end, fix some
n−1+γ
α ∈ 1, . (10.1.61)
n−1
Since α > 1, from (10.1.24), (10.1.23), and [133, Corollary 7.6.3] we conclude that
  κ−n.t.   
 
 ν · F∂Ω 
γ L 1 (∂Ω,σ)

      
1  α

≤ C P divF  L 1 (∂Ω,σ) + C  M ∂Ω |Nκ F | α 
L 1 (∂Ω,σ)

     
1  α

= C P divF  L 1 (∂Ω,σ) + C M ∂Ω |Nκ F | α  α
L (∂Ω,σ)

     1 α

≤ C P divF  L 1 (∂Ω,σ) + C Nκ F α  α
L (∂Ω,σ)
    
= C P divF  L 1 (∂Ω,σ) + C Nκ F L 1 (∂Ω,σ) < +∞, (10.1.62)
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 619

for some constant C = CΩ,κ,γ ∈ (0, ∞). At this stage, the claims in (10.1.59)-(10.1.60)
become consequences of (10.1.23) and (10.1.62). 

Theorem 10.1.4 yields new, non-trivial results, even in mundane geometric set-
tings such as the upper half-space. In such a scenario, due to the specific geometry
of the environment, there is a version of Theorem 10.1.4 which places less stringent
demands when it comes to the existence of the nontangential boundary trace in
(10.1.47). Here is the actual statement.

Corollary 10.1.5 Fix an aperture parameter κ > 0 and consider a vector field

F = (F1, . . . , Fn ) : R+n −→ Cn
(10.1.63)
with Lebesgue measurable components

satisfying the following properties:

κ−n.t.
Fn ∂Rn exists at L n−1 -a.e. point on Rn−1 ≡ ∂R+n,
+ (10.1.64)
Nκ F ∈ L 1 (Rn−1, L n−1 ), and divF = 0 in R+n,

with all derivatives taken in the sense of distributions in R+n .

κ−n.t.
Then the nontangential boundary trace Fn ∂Rn belongs to the Hardy space
+
H 1 (Rn−1, L n−1 ) and there exists a constant C = C(n, κ) ∈ (0, ∞) with the prop-
erty that
 κ−n.t.   
Fn  n  1 n−1 n−1 ≤ C Nκ F 1 n−1 n−1 . (10.1.65)
∂R + H (R , L ) L (R , L )

Proof The fact that Nκ F ∈ L 1 (Rn−1, L n−1 ) implies (cf. [133, Lemma 8.3.1])

F ∈ Lloc
∞ n n
(R+n, L n ) ⊆ Lloc
1
(R+n, L n ) . (10.1.66)

In particular, considering the derivatives ∂j Fj in the sense of distributions is mean-

ingful. To proceed, observe from simple geometric considerations that there exists a
large constant K = K(n, κ) ∈ (1, ∞) such that
 
B x + εen, ε/K ⊆ Γκ (z) for each ε > 0,
(10.1.67)
each z ∈ ∂R+n, and each x ∈ Γκ (z).
∫
Pick a function θ ∈ Cc∞ (Rn ) with supp θ ⊆ B(0, 1/K) and with Rn θ dL n = 1. For
each ε > 0 set θ ε (x) := ε −n θ(x/ε) for all x ∈ Rn , then define G ε := F(·
 + εen ) ∗ θ ε
in R+n . That is,
∫
G ε (x) =  − y + εen )θ ε (y) dy for each x ∈ R+n .
F(x (10.1.68)
R+n
620 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

From this and (10.1.66) we see that G ε is a well-defined vector field which belongs
to C ∞ (R+n ) . Moreover, the last property in (10.1.64) ensures that divG ε = 0 in
n

R+n . Finally, (10.1.67) ensures that Nκ G ε ≤ Nκ F at each point on Rn−1 ≡ ∂R+n . In

concert with the membership in (10.1.64) and [133, (8.2.28)], this guarantees that
Nκ G ε ∈ L 1 (Rn−1, L n−1 ) and, in fact,
   
sup Nκ G ε  L 1 (Rn−1, L n−1 ) ≤ Nκ F L 1 (Rn−1, L n−1 ) . (10.1.69)
ε>0

Granted these properties, we may invoke Theorem 10.1.4 (with Ω := R+n ) to

κ−n.t.
conclude that (G ε )n ∂Rn , the nontangential boundary trace of the n-th component
of G ε , belongs to the Hardy space H 1 (Rn−1, L n−1 ) and there exists a constant
C = C(n, κ) ∈ (0, ∞) with the property that
 κ−n.t.   
sup (G ε )n ∂Rn  H 1 (Rn−1, L n−1 ) ≤ C Nκ F L 1 (Rn−1, L n−1 ) . (10.1.70)
+
ε>0

Since for ε > 0 we have

∫
κ−n.t.
(G ε )n ∂Rn (x) = Fn (x − y + εen )θ ε (y) dy for each x ∈ ∂R+n, (10.1.71)
R+n

we may conclude from (10.1.67) and the first line in (10.1.64) that
κ−n.t. κ−n.t.
lim+ (G ε )n ∂Rn = Fn ∂Rn at L n−1 -a.e. point on Rn−1 ≡ ∂R+n . (10.1.72)
ε→0 +

As this stage, a combination of (10.1.70), (10.1.71), and Proposition 4.6.3 gives all
desired conclusions. 

There is also a version of Corollary 10.1.5 for the unit ball in place of the upper
half-space, of the sort described below.

Proposition 10.1.6 Let n ∈ N and fix an aperture parameter κ ∈ (0, ∞). Consider a

vector field F = (F1, . . . , Fn ) : B(0, 1) → Cn with Lebesgue measurable components
and define

n

f (x) := x · F(x) = x j Fj (x) for each x = (x1, . . . , xn ) ∈ B(0, 1). (10.1.73)
j=1

With S n−1 = ∂B(0, 1), abbreviate σ := H n−1 S n−1 , and assume that the following
properties are satisfied:
κ−n.t.
the nontangential trace f ∂B(0,1) exists σ-a.e. on S n−1,
  (10.1.74)
Nκ F ∈ L 1 (S n−1, σ), and divF = 0 in D  B(0, 1) .
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 621
κ−n.t.
Then the nontangential boundary trace f ∂B(0,1) belongs to the Hardy space
H 1 (S n−1, σ) and there exists a constant C = C(n, κ) ∈ (0, ∞) with the property that
 κ−n.t.   
f  ≤ C Nκ F L 1 (S n−1,σ) . (10.1.75)
∂B(0,1) H 1 (S n−1,σ)

Proof We shall retain notation and results from the proof of [133, Proposi-
tion 2.8.21]. For each ε ∈ (0, 1) bring back the regularized vector field

G ε ∈ C ∞ (B(0, 1))
n
(10.1.76)

defined as
∫
 
G ε (x) := F (1 + ε)−1 (x − y) θ ε (y) dy for each x ∈ B(0, 1). (10.1.77)
Rn

From [133, (2.8.233)] and the last property in (10.1.74) we see that

divG ε = 0 in B(0, 1). (10.1.78)

Granted these properties, we may invoke Theorem 10.1.4 (with Ω := B(0, 1)) and
conclude that the function fε : S n−1 → C defined as fε (x) := x · G ε (x) for each
x ∈ S n−1 belongs to the Hardy space H 1 (S n−1, σ) for each ε ∈ (0, 1) and there exists
a constant C = C(n, κ) ∈ (0, ∞) with the property that
 
fε H 1 (S n−1,σ) ≤ C Nκ G ε  L 1 (S n−1,σ) . (10.1.79)

Moving on, we claim that

Nκ G ε ≤ Nκ F at each point on S n−1 ≡ ∂B(0, 1). (10.1.80)

Indeed, [133, (2.8.225)] implies for each z ∈ S n−1 , each

 x ∈ Γκ (z), and each
y ∈ supp θ ε ⊆ B(0, c ε) we have (1 + ε)−1 (x − y) ∈ B (1 + ε)−1 x, c ε ⊆ Γκ (z).
Keeping this in mind, we conclude from [133, (2.8.230)] that Nκ G ε (z) ≤ Nκ F(z)

which, in view of the arbitrariness of z ∈ S , establishes (10.1.80). In particular,
n−1

from (10.1.79) and (10.1.80) we conclude that

each fε belongs to H 1 (S n−1, σ) and

  (10.1.81)
sup fε H 1 (S n−1,σ) ≤ C Nκ F L 1 (S n−1,σ) .
0<ε<1

Finally, recall from [133, (2.8.248)] that

κ−n.t.
lim fε = f ∂B(0,1) at σ-a.e. point on S n−1 ≡ ∂B(0, 1). (10.1.82)
ε→0+
622 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
κ−n.t.
Combining (10.1.81), (10.1.82), and Proposition 4.6.3 then proves that f ∂B(0,1)
belongs to the Hardy space H 1 (S n−1, σ) and that the estimate claimed in (10.1.75)
holds. 
It turns out that Theorem 10.1.4 self-extends as to allow homogeneous, constant
(complex) coefficient, first-order differential operators of a completely general nature
to play the role of the divergence operator. To state such a generalization, recall that
given an N × M homogeneous first-order system (where N, M ∈ N are arbitrary)
with constant complex coefficients

n
αβ
D= a j ∂j 1≤α ≤ N (10.1.83)
j=1 1≤β ≤M

its principal symbol is defined as the N × M matrix


n
αβ
Sym(D; ξ) := i aj ξj 1≤α ≤ N for each ξ = (ξ1, . . . , ξn ) ∈ Rn . (10.1.84)
j=1 1≤β ≤M

Corollary 10.1.7 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an Ahlfors regular

domain. In particular, Ω is a set of locally finite perimeter whose geometric measure
theoretic outward unit normal ν is defined almost everywhere on ∂Ω with respect to
the measure σ := H n−1 ∂Ω.
Also, pick two integers N, M ∈ N and consider an arbitrary N × M homogeneous
first-order system D with constant complex coefficients in Rn , along with a vector-
valued function F : Ω → C M , having L n -measurable components, with the property
that for some κ ∈ (0, ∞) one has
κ−n.t.
Nκ F ∈ L 1 (∂Ω, σ) and F ∂Ω exists in C M at σ-a.e. point on ∂Ω. (10.1.85)
M
In particular, F ∈ Lloc
1 (Ω, L n ) (cf. [133, Lemma 8.3.1]), and one also assumes
1 (Ω, L n )
that DF, computed in the sense of distributions in Ω, has components in Lloc
and satisfies
P(DF) ∈ L 1 (∂Ω, σ). (10.1.86)
Then  κ−n.t. 
Sym(D; ν) F ∂Ω ∈ H 1 (∂Ω, σ)
N
(10.1.87)
and there exists a constant C = C(Ω, D, κ) ∈ (0, ∞) such that
  κ−n.t.    

Sym(D; ν) F ∂Ω  1 N
≤ C Nκ F  1
L (∂Ω,σ)
+ C P(DF) L 1 (∂Ω,σ) .
[H (∂Ω,σ)]
(10.1.88)
Finally, with the scale of homogeneous Hardy spaces defined as in (4.2.12),
 κ−n.t. 
if Ω is bounded and DF = 0 in Ω then in fact Sym(D; ν) F ∂Ω
. N (10.1.89)
belongs to the homogeneous Hardy space H 1 (∂Ω, σ) .
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 623

To offer a concrete example, consider the case when n = 2 and Ω := D, the

standard open unit disk in the plane. Set σ := H n−1 ∂D and fix an aperture parameter
κ > 0. Then, given any holomorphic function u : D → C with the property that
Nκ u ∈ L 1 (∂D, σ) Fatou’s theorem implies (keeping in mind [133, (8.9.44)]) that
κ−n.t.
the function u∂D belongs to L 1 (∂D, σ). On the other hand, with ∂ := 12 (∂x + i∂y )
denoting the Cauchy-Riemann operator in the plane, the membership (10.1.87) in
Corollary 10.1.7 presently gives that
 κ−n.t.   κ−n.t. 
∂D  ζ → ζ u∂D (ζ) = (−2i)Sym(∂; ν) u∂Ω (ζ) ∈ C
(10.1.90)
belongs to the Hardy space H 1 (∂D, σ)  L 1 (∂D, σ).

In addition, Cauchy’s vanishing formula guarantees that

∫ ∫
 κ−n.t.  1  κ−n.t. 

ζ u ∂D dσ(ζ) = u∂D (ζ) dζ = 0, (10.1.91)
∂D i ∂D

hence (recall that the scale of homogeneous Hardy spaces has been introduced in
(4.2.12))
 κ−n.t. 
the function ∂D  ζ −→ ζ u∂D (ζ) ∈ C
. (10.1.92)
belongs to the homogeneous Hardy space H 1 (∂D, σ).

As regards the original nontangential boundary trace of u, we wish to note that since
H 1 (∂D, σ) is a module over the ring of smooth functions (cf. Proposition 4.4.8), we
may further conclude from (10.1.90) that
 κ−n.t.   κ−n.t. 
∂D  ζ −→ u∂D (ζ) = ζ ζ u∂D (ζ) ∈ C
(10.1.93)
belongs to the Hardy space H 1 (∂D, σ)  L 1 (∂D, σ).

However, generally speaking, there is no reason to expect that the integral of this
latter function vanishes. In fact, since dζ = iζ dσ(ζ) on ∂D, Cauchy’s reproducing
formula yields
∫ ∫  κ−n.t. 
 κ−n.t.  1 u∂D (ζ)
u∂D dσ(ζ) = dζ = 2πu(0). (10.1.94)
∂D i ∂D ζ
κ−n.t.
Thus, the function u∂D does not necessarily belong to the homogeneous Hardy
.1
space H (∂D, σ).
Proof of Corollary 10.1.7 Suppose D is expressed as in (10.1.83) and that the vector-
valued function F has scalar components (Fβ )1≤β ≤M in Ω. For each fixed index
α ∈ {1, . . . , N }, consider then the vector field Fα = (Fjα )1≤ j ≤n : Ω → Cn with
components given by
624 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .


M
αβ
Fjα := a j Fβ in Ω, for j ∈ {1, . . . , n}. (10.1.95)
β=1

Then, on account of (10.1.95) and the properties recorded in (10.1.85), it follows

that the vector field Fα has L n -measurable component and satisfies
κ−n.t.
Nκ Fα ∈ L 1 (∂Ω, σ) and Fα ∂Ω exists σ-a.e. on ∂Ω. (10.1.96)

In addition,


n 
M 
n
αβ
divFα = ∂j Fjα = a j ∂j Fβ = (DF)α in D (Ω). (10.1.97)
j=1 β=1 j=1

In particular, in a pointwise sense we have


N
 
P divFα ≈ P(DF) uniformly on ∂Ω (10.1.98)
α=1

which, thanks to (10.1.3) and (10.1.86), goes to show that


N
  
P divFα  1 ≈ P(DF) L 1 (∂Ω,σ) < +∞. (10.1.99)
L (∂Ω,σ)
α=1

Granted these properties, Theorem 10.1.4 applies to each vector field Fα and proves
that  κ−n.t. 
ν · Fα ∂Ω ∈ H 1 (∂Ω, σ) (10.1.100)
and that there exists a constant CΩ,κ ∈ (0, ∞) such that
  κ−n.t.        
   
ν· Fα ∂Ω  1 ≤ CΩ,κ Nκ Fα  L 1 (∂Ω,σ) +P divFα  L 1 (∂Ω,σ) . (10.1.101)
H (∂Ω,σ)

Upon noting that, for each α ∈ {1, . . . , N }, at σ-a.e. on point on ∂Ω we have

 κ−n.t.  
n
 κ−n.t.  
n 
M
αβ  κ−n.t. 
ν · Fα ∂Ω = ν j Fjα ∂Ω = ν j a j Fβ ∂Ω
j=1 j=1 β=1

 κ−n.t. 
= (−i)Sym(D; ν) F ∂Ω , (10.1.102)
α

the claims in (10.1.87)-(10.1.88) now readily follow from (10.1.99), (10.1.100),

(10.1.101), and (10.1.102).
Finally, the claim in (10.1.89) is a consequence of (4.2.12), (10.1.87), and the
integration by parts formula from [133, Theorem 1.7.2] (currently used with w a
constant vector in C M ). 
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 625

A significant particular case of Corollary 10.1.7 is as follows. Start by fixing an

Ahlfors regular domain Ω ⊆ Rn , with geometric measure theoretic outward unit
normal ν and surface measure σ := H n−1 ∂Ω. Work in the Clifford algebra context
and consider a function F ∈ C ∞ (Ω) ⊗ Cn which, for some κ > 0, satisfies (with D
denoting the associated Dirac operator)

DF = 0 in Ω, Nκ F ∈ L 1 (∂Ω, σ) and
κ−n.t. (10.1.103)
F∂Ω
exists (in Cn ) at σ-a.e. point on ∂Ω.

Then Corollary 10.1.7 implies that, for some C ∈ (0, ∞) independent of F, we have
 κ−n.t. 
ν  F ∂Ω ∈ H 1 (∂Ω, σ) ⊗ Cn and
  κ−n.t.   (10.1.104)
ν  F   1 ≤ C Nκ F L 1 (∂Ω,σ) .
∂Ω H (∂Ω,σ)⊗ C n

It turns out that there is a bi-linear version of Corollary 10.1.7, as discussed

in our next result. To state it, recall that ·, · denotes the (real) inner product in
higher-dimensional complex spaces (defined in (A.0.46)).

Corollary 10.1.8 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an Ahlfors regular

domain. In particular, Ω is a set of locally finite perimeter whose geometric measure
theoretic outward unit normal ν is defined almost everywhere on ∂Ω with respect to
the measure σ := H n−1 ∂Ω.
In this context, pick two integers N, M ∈ N and consider an arbitrary N × M
homogeneous first-order system D with constant complex coefficients in Rn , together
with two vector-valued functions
M N
F ∈ C 1 (Ω) and G ∈ C 1 (Ω) , (10.1.105)

satisfying, for some κ > 0,

Nκ F < ∞, Nκ G < ∞ at σ-a.e. point on ∂Ω,

Nκ F · Nκ G belongs to the space L 1 (∂Ω, σ),
κ−n.t. κ−n.t. (10.1.106)
F ∂Ω and G∂Ω exist at σ-a.e. point on ∂Ω,
 
P DF, G − F, D G belongs to L 1 (∂Ω, σ),

where D denotes the (real) transpose of the differential operator D.

Then   κ−n.t.  κ−n.t. 
Sym(D; ν) F ∂Ω , G∂Ω ∈ H 1 (∂Ω, σ) (10.1.107)

and there exists a constant C = C(Ω, D, κ) ∈ (0, ∞) such that

626 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
  κ−n.t.  κ−n.t.  

 Sym(D; ν) F ∂Ω ,G∂Ω  (10.1.108)
H 1 (∂Ω,σ)
 
≤ C Nκ F · Nκ G L 1 (∂Ω,σ)
  
+ C P DF, G − F, D G  L 1 (∂Ω,σ) .

In particular, the conclusions in (10.1.107)-(10.1.108) hold if in place of the last

condition in (10.1.106) one merely assumes that DF = 0 and D G = 0 in Ω (in
which scenario the last norm in (10.1.108) may be dropped).
For example, the conditions in the first two lines of (10.1.106) are satisfied if

Nκ F ∈ L p (∂Ω, σ) and Nκ G ∈ L p (∂Ω, σ)
(10.1.109)
for some p, p ∈ [1, ∞] with 1/p + 1/p = 1.

Here is the proof of Corollary 10.1.8.

Proof of Corollary 10.1.8 Assuming the operator D is written as in (10.1.83), its
transpose D is the M × N homogeneous first-order system given by (10.2.93). Let
us denote by (Fβ )1≤β ≤M and (Gα )1≤β ≤ N , respectively, the scalar components of the
vector valued functions F and G, and use these to define the vector field


N 
M
αβ
H :=
n
a j Fβ Gα ∈ C 1 (Ω) . (10.1.110)
1≤ j ≤n
α=1 β=1

On account of (10.1.110) and (10.2.93) we then have


N 
M 
n 
N 
M 
n
αβ αβ
divH = a j (∂j Fβ )Gα + a j Fβ (∂j Gα )
α=1 β=1 j=1 α=1 β=1 j=1

= DF, G − F, D G in Ω. (10.1.111)

From (10.1.110), (10.1.111), [133, (8.2.28)], and the assumptions imposed on F, G

in (10.1.106) we then conclude that
κ−n.t.
Nκ H ∈ L 1 (∂Ω, σ), H ∂Ω exists σ-a.e. on ∂Ω,
  (10.1.112)
and P divH belongs to L 1 (∂Ω, σ).

Granted these, Theorem 10.1.4 applies to H and, since

κ−n.t. 
N 
M 
n
κ−n.t. κ−n.t.
ν · H ∂Ω = a j ν j Fβ ∂Ω Gα ∂Ω
αβ

α=1 β=1 j=1

  κ−n.t.  κ−n.t. 
= (−i)Sym(D; ν) F ∂Ω , G∂Ω , (10.1.113)
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 627

the desired conclusions follow.

Parenthetically, we wish to note that there is a more effective algebraic formalism
for the above argument, based on the observation made in [133, (1.7.21)]. Specifi-
cally, if we define H = (H j )1≤ j ≤n with
 
H j := (−i)Sym(D; e j )F, G for each j ∈ {1, . . . , n}, (10.1.114)

then thanks to [133, (1.7.21)] and [133, (1.7.17)] we may write


n n 
  n  
divH = ∂j H j = (−i)Sym(D; e j )∂j F, G + (−i)Sym(D; e j )F, ∂j G
j=1 j=1 j=1


n   n 
= (−i)Sym(D; e j )∂j F, G − F, (−i)Sym(D ; e j )∂j G
j=1 j=1
   
= DF, G − F, D G in Ω. (10.1.115)

Also, at σ-a.e. point on ∂Ω we have

κ−n.t. 
n   κ−n.t.   κ−n.t.  
ν · H ∂Ω = ν j (−i)Sym(D; e j ) F ∂Ω , G∂Ω
j=1
  κ−n.t.  κ−n.t. 
= (−i)Sym(D; ν) F ∂Ω , G∂Ω . (10.1.116)

With these in hand, the rest of the argument is then as before. 

We continue by presenting a result in which the conclusion is the (quantitative)

membership of the conormal derivative to the Hardy space H 1 .

Corollary 10.1.9 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an Ahlfors regular

domain. Abbreviate σ := H n−1 ∂Ω and denote by ν = (ν j )1≤ j ≤n the geometric
measure theoretic outward unit normal to Ω. Also, fix M, N ∈ N and consider a
coefficient tensor of type (n × n, M × N),
αβ
A = ar s 1≤r,s ≤n . (10.1.117)
1≤α ≤M
1≤β ≤ N

Associate with this coefficient tensor the M × N second-order system (with the
summation convention over repeated indices is in effect)
αβ
L A :=: divA∇ := ar s ∂r ∂s 1≤α ≤M . (10.1.118)
1≤β ≤ N

Finally, consider a C N -valued function u satisfying, for some aperture parameter

κ > 0,
628 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
N  αβ  n·M
u ∈ Lloc
1 (Ω, L n ) , A∇u := ar s ∂s uβ α,r ∈ Lloc
1 (Ω, L n ) ,
M
Nκ (A∇u) ∈ L 1 (∂Ω, σ), L Au ∈ Lloc
1 (Ω, L n ) , P(L Au) ∈ L 1 (∂Ω, σ),
κ−n.t.
the nontangential boundary trace (A∇u)∂Ω exists at σ-a.e. point on ∂Ω.
(10.1.119)
Then
 αβ  κ−n.t.
νr ar s ∂s uβ ∂Ω
M
belongs to the Hardy space H 1 (∂Ω, σ)
1≤α ≤M
(10.1.120)
and there exists a constant C = C(Ω, κ) ∈ (0, ∞) such that
M 
   αβ  κ−n.t. 
νr ar s ∂s uβ ∂Ω  (10.1.121)
H 1 (∂Ω,σ)
α=1
    
≤ C Nκ (A∇u) L 1 (∂Ω,σ) + P(L Au) L 1 (∂Ω,σ) .

Proof This follows by applying Theorem 10.1.4 to the vector field

αβ
Fα := ar s ∂s uβ (10.1.122)
1≤r ≤n

for each fixed α ∈ {1, . . . , M }. 

Our next theorem may be regarded as a version of Corollary 10.1.8 adapted to the
standard Euclidean Dirac operator.

Theorem 10.1.10 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an Ahlfors regular

domain. Abbreviate σ := H n−1 ∂Ω and denote by ν = (ν j )1≤ j ≤n the geometric
measure theoretic outward unit normal to Ω. Consider two Clifford algebra-valued
functions
u ∈ C 1 (Ω) ⊗ Cn and w ∈ C 1 (Ω) ⊗ Cn, (10.1.123)
satisfying, for some κ > 0,

Nκ u < ∞, Nκ u < ∞ at σ-a.e. point on ∂Ω,

Nκ u · Nκ w belongs to the space L 1 (∂Ω, σ),
κ−n.t. κ−n.t. (10.1.124)
u∂Ω and w ∂Ω exist at σ-a.e. point on ∂Ω,
 
P (DR u)  w + u  (D L w) belongs to L 1 (∂Ω, σ),

where D L and DR denote the Dirac operator acting from the right and from the left,
respectively (cf. (A.0.32)-(A.0.33)). 
Then, with the identification ν = nj=1 ν j e j , one has
 κ−n.t.   κ−n.t. 
u∂Ω  ν  w ∂Ω ∈ H 1 (∂Ω, σ) ⊗ Cn (10.1.125)
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 629

and there exists a constant C = C(Ω, κ) ∈ (0, ∞) such that

  κ−n.t.   κ−n.t.  
 
 u ∂Ω  ν w ∂Ω  1 (10.1.126)
H (∂Ω,σ)⊗ C n
 
≤ C Nκ u · Nκ w  L 1 (∂Ω,σ)
  
+ C P (DR u)  w + u  (D L w)  L 1 (∂Ω,σ) .

In particular, the conclusions in (10.1.125)-(10.1.126) hold if in place of the last

condition in (10.1.124) one merely assumes that DR u = 0 and D L w = 0 in Ω (in
which scenario the last norm in (10.1.126) may be dropped).

Proof This follows from Corollary 10.1.8. More directly, this is seen by applying
Theorem 10.1.4 to the vector field (with Clifford algebra-valued components)
 
F := u  e j  w 1≤ j ≤n in Ω. (10.1.127)

Upon noting that

divF = (DR u)  w + u  (D L w) in Ω, (10.1.128)
that Nκ F ≤ Nκ u · Nκ w, and that
 κ−n.t.   κ−n.t.   κ−n.t. 
ν · F∂Ω = u∂Ω  ν  w ∂Ω , (10.1.129)

all desired conclusions follow from (10.1.49)-(10.1.50). 

In turn, Theorem 10.1.10 is the main ingredient in the proof of the Corol-
lary 10.1.12, stated a little further below. As a preamble, we briefly elaborate on
the nature of the Cauchy-Clifford integral operators employed in the statement of the
aforementioned corollary.

Remark 10.1.11 Given an arbitrary set Ω ⊆ Rn of locally finite perimeter, abbre-

viate σ∗ := H n−1 ∂∗ Ω and denote by ν the geometric measure theoretic outward
unit normal to Ω. In this context, we define the “left-handed” Cauchy-Clifford in-
tegral operators (boundary-to-domain and, respectively,
 boundary-to-boundary)
 as
the mappings acting on each function f ∈ L 1 ∂∗ Ω , 1+ |σ· |∗n−1 ⊗ Cn according to
∫
1 x−y
CL f (x) :=  ν(y)  f (y) dσ∗ (y), ∀ x ∈ Ω̊, (10.1.130)
ωn−1 ∂∗ Ω |x − y| n

and, for σ∗ -a.e. point x ∈ ∂∗ Ω,

∫
1 x−y
C L f (x) := lim+  ν(y)  f (y) dσ∗ (y). (10.1.131)
ε→0 ωn−1 |x − y| n
y ∈∂∗ Ω
|x−y |>ε
630 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

These should be contrasted with the “right-handed” Cauchy-Clifford integral op-

erators (boundary-to-domain and, respectively,
 boundary-to-boundary),
 defined as
the mappings acting on each function f ∈ L 1 ∂∗ Ω , 1+ |σ· |∗n−1 ⊗ Cn according to
∫
1 x−y
CR f (x) := f (y)  ν(y)  dσ∗ (y), ∀ x ∈ Ω̊, (10.1.132)
ωn−1 ∂∗ Ω |x − y| n

and, for σ∗ -a.e. point x ∈ ∂∗ Ω,

∫
1 x−y
C R f (x) := lim+ f (y)  ν(y)  dσ∗ (y). (10.1.133)
ε→0 ωn−1 |x − y| n
y ∈∂∗ Ω
|x−y |>ε
 σ∗ 
It is apparent from the above definitions that for each f ∈ L 1 ∂∗ Ω , 1+ | · | n−1
⊗ Cn
we have (with ‘bar’ denoting Clifford conjugation)

CR f = CL ( f ) in Ω̊, (10.1.134)

and
C R f = C L ( f ) at σ∗ -a.e. point on ∂∗ Ω. (10.1.135)
 
Also, for each function f ∈ L 1 ∂∗ Ω , 1+ |σ· |∗n−1 ⊗ Cn we have

DR CR f = 0 and D L CL f = 0 in Ω̊, (10.1.136)

where D L and DR denote the Dirac operator acting from the right and from the left,
respectively (cf. (A.0.32)-(A.0.33)).

Here is the corollary alluded to before the remark.

Corollary 10.1.12 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is a UR domain.

Denote by ν the geometric measure theoretic outward unit normal to Ω and abbrevi-
ate σ := H n−1 ∂Ω. In this setting, let C L , C R be the left-handed and, respectively,
right-handed boundary-to-boundary Cauchy-Clifford operators defined as in Re-
mark 10.1.11. Then, with some naturally accompanying estimates, the memberships
1
4 f  ν  g + (C R f )  ν  (C L g) ∈ H 1 (∂Ω, σ) ⊗ Cn, (10.1.137)

f  ν  (C L g) + (C R f )  ν  g ∈ H 1 (∂Ω, σ) ⊗ Cn, (10.1.138)

are valid in each of the following scenarios:


(1) f ∈ L p (∂Ω, w) ⊗ Cn and g ∈ L p (∂Ω, w ) ⊗ Cn with p, p ∈ (1, ∞) satisfying

1/p + 1/p = 1, w ∈ Ap (∂Ω, σ), and where w  := w 1−p ∈ Ap (∂Ω, σ) is
the conjugate weight of w (hence, in particular, for f ∈ L (∂Ω, σ) ⊗ Cn and
p

g ∈ L p (∂Ω, σ) ⊗ Cn with p, p ∈ (1, ∞) satisfying 1/p + 1/p = 1);
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 631

(2) f ∈ M p,λ (∂Ω, σ) ⊗ Cn and g ∈ B q,λ (∂Ω, σ) ⊗ Cn with p, q ∈ (1, ∞) satisfying
1/p + 1/q = 1 and λ ∈ (0, n − 1);
(3) f ∈ B q,λ (∂Ω, σ) ⊗ Cn and g ∈ M p,λ (∂Ω, σ) ⊗ Cn with p, q ∈ (1, ∞) satisfying
1/p + 1/q = 1 and λ ∈ (0, n − 1).
Proof Suppose f , g are as in any of the scenarios described in items (1)-(3). Applying
Theorem 10.1.10 to

u := CR f ∈ C ∞ (Ω) ⊗ Cn and w := CL g ∈ C ∞ (Ω) ⊗ Cn, (10.1.139)

then yields
1   
2I + C R f  ν  12 I + C L g ∈ H 1 (∂Ω, σ) ⊗ Cn (10.1.140)

and
     
 1
 2 I + C R f  ν  12 I + C L g 
H 1 (∂Ω,σ)⊗ C n (10.1.141)
is less than a fixed multiple of the product of the norms of f and g,

on account of Remark 10.1.11, the mapping properties of generic double layer po-
tential operators (including the Cauchy-Clifford operators of interest here) discussed
at length in [136] (bearing in mind that Ω is currently assumed to be a UR domain),
and Proposition 6.2.8.
Going further, recall from item (6) in [133, Lemma 5.10.9] that Ω− := Rn \ Ω is
also an Ahlfors regular domain, whose topological boundary coincides with that of
Ω, whose geometric measure theoretic boundary agrees with that of Ω, and whose
geometric measure theoretic outward unit normal is −ν at σ-a.e. point on ∂Ω.
Bearing this in mind, the versions of (10.1.140)-(10.1.141) produced by working
with Ω− in place of Ω read
   
− 12 I + C R f  ν  − 12 I + C L g ∈ H 1 (∂Ω, σ) ⊗ Cn (10.1.142)

and
     

 − 12 I + C R f  ν  − 12 I + C L g 
H 1 (∂Ω,σ)⊗ C n (10.1.143)
is less than a fixed multiple of the product of the norms of f and g.

Then the claims made in relation to (10.1.137)-(10.1.138) now readily follow by

combining (10.1.140)-(10.1.141) with (10.1.142)-(10.1.143). 
Untangling the memberships established in Corollary 10.1.12 into scalar compo-
nents then yields the following type of result:
Corollary 10.1.13 Let Ω ⊆ Rn , with n ∈ N with n ≥ 2, be a UR domain. Denote
by ν = (ν j )1≤ j ≤n the geometric measure theoretic outward unit normal to Ω and
abbreviate σ := H n−1 ∂Ω. Bring in the family of Riesz transforms {R j }1≤ j ≤n
632 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

defined as in (A.0.102) with Σ := ∂Ω, and recall the transpose harmonic double
layer potential operator KΔ# from (A.0.53). Then, with some naturally accompanying
estimates, the memberships

f (R j g) + (R j f )g ∈ H 1 (∂Ω, σ) for each j ∈ {1, . . . , n}, (10.1.144)


n
−ν j f g + ν j (Rk f )(Rk g) − 2(KΔ# f )(R j g) − 2(R j f )(KΔ# g)
k=1
(10.1.145)
belongs to the Hardy space H 1 (∂Ω, σ) for each j ∈ {1, . . . , n},
and
  
abc
sgn (Ri f )ν j (Rk g) belongs to H 1 (∂Ω, σ)
i j k (10.1.146)
{i, j,k }={a,b,c }
for each triplet of mutually distinct numbers a, b, c ∈ {1, . . . , n},

are all valid in each of the following scenarios:


(1) f ∈ L p (∂Ω, w) and g ∈ L p (∂Ω, w ) for p, p ∈ (1, ∞) with 1/p + 1/p = 1,

w ∈ Ap (∂Ω, σ), and w  := w 1−p ∈ Ap (∂Ω, σ) (hence, in particular, for f in

L (∂Ω, σ) and g in L (∂Ω, σ), if p, p ∈ (1, ∞) satisfy 1/p + 1/p = 1);
p p

(2) f ∈ M p,λ (∂Ω, σ) and g ∈ B q,λ (∂Ω, σ) for integrability exponents p, q ∈ (1, ∞)
satisfying 1/p + 1/q = 1, and λ ∈ (0, n − 1).

Some of these results are known; see, e.g., [34], [70, p. 565].
Proof of Corollary 10.1.13 Let f , g be as in any of the scenarios described in items
(1)-(3) in Corollary 10.1.12. Then

n
CR f = 1
2 R j ( f  ν)  e j (10.1.147)
j=1

and

n
CL g = 1
2 e j  R j (ν  g). (10.1.148)
j=1

After replacing f by f  ν and g by ν  g (which leaves the nature of f , g unaffected;

cf. (6.2.5) and (6.2.74)), and taking into account the fact that ν  ν = −1 (cf. [133,
(6.4.1)]), the membership in (10.1.138) implies
n 
 
f  e j  (R j g) + (R j f )  e j  g ∈ H 1 (∂Ω, σ) ⊗ Cn . (10.1.149)
j=1

Assuming now that f , g are two scalar-valued functions as in any of the items (1)-(2)
in the statement, (10.1.149) readily implies (10.1.144).
10.1 Strong Normal Boundary Traces of Vector Fields in Hardy Spaces 633

As regards (10.1.145)-(10.1.146), we once again start with f , g as in any of the

scenarios described in items (1)-(3) in Corollary 10.1.12. Use (10.1.147)-(10.1.148)
back in (10.1.137) to conclude that the Clifford algebra-valued function

n 
n
f νg+ R j ( f  ν)  e j  ν  ek  Rk (ν  g) (10.1.150)
j=1 k=1

belongs to the Hardy space H 1 (∂Ω, σ) ⊗ Cn . After replacing f by f  ν and g by

ν  g (which, as noted earlier, leaves the nature of f , g unaffected) and, once again,
taking into account the fact that ν  ν = −1, we arrive at the conclusion that

n 
n
−f  ν  g + (R j f )  e j  ν  ek  (Rk g) (10.1.151)
j=1 k=1

belongs to the Hardy space H 1 (∂Ω, σ) ⊗ Cn . At this stage, assume that f , g are two
scalar-valued functions as in any of the items (1)-(2) in the present statement. With

R := nj=1 e j R j , this membership implies

− f gν + (R f )  ν  (Rg) ∈ H 1 (∂Ω, σ) ⊗ Cn . (10.1.152)

Finally, from this and [133, (6.4.87)], fact that the scalar components of the vector
part of the Clifford algebra valued function in (10.1.152) belong to the Hardy space
H 1 (∂Ω, σ) yields (10.1.145), whereas writing that the scalar components of the tri-
vector part of the Clifford algebra valued function in (10.1.152) belong to the Hardy
space H 1 (∂Ω, σ) yields (10.1.146). 

Particularizing Corollary 10.1.13 to the special case when the underlying domain
is the upper half-space yields the following corollary, of independent interest.

Corollary 10.1.14 Fix n ∈ N and consider the standard family of Riesz transforms
{R j }1≤ j ≤n in Rn , i.e.,
 where each R j with j ∈ {1, . . . , n} acts on an arbitrary function
f ∈ L 1 Rn, 1+dx |x | n according to
∫
2 xj − yj
R j f (x) := lim+ f (y) dy for L n -a.e. x ∈ Rn . (10.1.153)
ε→0 ωn |x − y| n+1
|x−y |>ε
y ∈R n

Then, with some naturally accompanying estimates, the memberships

f (R j g) + (R j f )g ∈ H 1 (Rn, L n ) for each j ∈ {1, . . . , n}, (10.1.154)


n
fg − (R j f )(R j g) belongs to the Hardy space H 1 (Rn, L n ), (10.1.155)
j=1

and
634 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

(R j f )(Rk g) − (Rk f )(R j g) belongs to H 1 (Rn, L n )

(10.1.156)
for each j, k ∈ {1, . . . , n},
are all valid in each of the following scenarios:

(1) f ∈ L p (Rn, w) and g ∈ L p (Rn, w ) for p, p ∈ (1, ∞) with 1/p + 1/p = 1,

w ∈ Ap (Rn, L n ), and w  := w 1−p ∈ Ap (Rn, L n ) (hence, in particular, if

f ∈ L (R , L ) and g ∈ L (R , L n ) whenever the exponents p, p ∈ (1, ∞)
p n n p n

satisfy 1/p + 1/p = 1);

(2) f ∈ M p,λ (Rn, L n ) and g ∈ B q,λ (Rn, L n ) where λ ∈ (0, n − 1) and the exponents
p, q ∈ (1, ∞) satisfy 1/p + 1/q = 1.
Proof This is seen by specializing the version of Corollary 10.1.13, written with n
replaced by n + 1, to the case when Ω := R+n+1 . In this scenario we have ν = −en+1 ,
KΔ# = 0, and the (n + 1)-th Riesz transform associated (in the (n + 1)-dimensional
Euclidean setting) as in (A.0.102) with Σ := ∂R+n+1 and j := n + 1 is identically
zero. Bearing these special features in mind, (10.1.144)-(10.1.146) reduce precisely
to (10.1.154)-(10.1.156). 

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy

and Morrey Spaces

It is possible to adopt a more general point of view which allows us to produce trace
results in Hardy spaces H p , with exponents p in the range ( n−1 n , ∞], and even the
more inclusive scale of Lorentz-based Hardy spaces H p,q with indices p ∈ ( n−1 n , ∞]
and q ∈ (0, ∞] arbitrary. Since for p < 1 the aforementioned Hardy spaces no longer
consist of functions (but rather involve functionals), we are necessarily led to consid-
ering a notion of boundary trace which is different from the nontangential pointwise
trace used in the formulation of Theorem 10.1.4. Specifically, we shall work with the
weak normal boundary trace of a vector field, as defined in [133, Proposition 4.2.3]
(the reader is also advised to recall the space introduced in (A.0.55)).
Theorem 10.2.1 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an
Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. Also, fix some  aperture
parameter κ ∈ (0, ∞) along with two integrability exponents p ∈ n−1 n ∞ and
,
q ∈ (0, ∞]. Consider a vector field F : Ω → Cn , having L n -measurable components,
with the property that1
Nκ F ∈ L p,q (∂Ω, σ). (10.2.1)
In particular, F ∈ Lloc
1 (Ω, L n ) n (cf. [133, Lemma 8.3.1]) and, with the divergence

taken in the sense of distributions in Ω, assume

 
divF ∈ Lloc
1
(Ω, L n ) and P divF ∈ L p,q (∂Ω, σ). (10.2.2)

1 making the convention that L ∞, q (∂Ω, σ) = L ∞ (∂Ω, σ) for any q ∈ (0, ∞]

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 635

Then, first,

F ∈ Lbdd and divF ∈ Lbdd

n
1
(Ω, L n ) 1
(Ω, L n ), (10.2.3)

hence
 it is meaningful to consider the “bullet product” ν • F as a functional in

Lipc (∂Ω) (cf. [133, Proposition 4.2.3]). Second, if p < ∞ then

ν • F belongs to the Lorentz-based Hardy space H p,q (∂Ω, σ) (10.2.4)

and there exists a constant CΩ,κ, p,q ∈ (0, ∞), independent of F,  such that
      
ν • F p, q ≤ CΩ,κ, p,q
Nκ F p, q + P(divF)
  p, q . (10.2.5)
H (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ)

Third, corresponding to the limiting case p := ∞ and q := ∞, one has

ν • F ∈ L ∞ (∂Ω, σ) and there exists CΩ,κ ∈ (0, ∞) such that

      
ν • F ∞ ≤ CΩ,κ
Nκ F ∞ + P(divF)
 ∞
L (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ) (10.2.6)
   
= CΩ,κ F L ∞ (Ω, L n ) + P(divF)
 ∞
L (∂Ω,σ)
.
 n−1
As a corollary, corresponding to the choice q := p ∈ n , 1 the conclusions in
(10.2.4)-(10.2.5) become

ν • F belongs to the (ordinary) Hardy space H p (∂Ω, σ) and

       (10.2.7)
ν • F p ≤ CΩ,κ, p Nκ F L p (∂Ω,σ) + P(divF)
 p ,
H (∂Ω,σ) L (∂Ω,σ)

while corresponding to the choice p ∈ (1, ∞) the conclusions in (10.2.4)-(10.2.5)

become

ν • F belongs to the Lorentz space L p,q (∂Ω, σ) and

       (10.2.8)
ν • F p, q ≤ CΩ,κ, p,q
Nκ F p, q + P(divF)
  p, q .
L (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ)

Finally, corresponding to the choice q := p ∈ (1, ∞) the conclusions in (10.2.4)-

(10.2.5) become

ν • F belongs to the Lebesgue space L p (∂Ω, σ) and

       (10.2.9)
ν • F p ≤ CΩ,κ, p
Nκ F p + P(divF)
 p .
L (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ)

Of course, the conclusions in (10.2.4)-(10.2.5), as well as (10.2.7), (10.2.8), and

(10.2.9), hold if in place of (10.2.2) one assumes that F is actually divergence-
free in the sense of distributions in Ω, a scenario in which the last quasi-norm in
(10.2.5), (10.2.7), (10.2.8), and (10.2.9) may be dropped. Moreover, thanks to [133,
636 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Theorem 1.9.1] and [133, Theorem 1.9.2],

under the additional assumptions that Ω is bounded

 and divF = 0,
corresponding to the choice q := p ∈ n , ∞ the membership in
n−1
. (10.2.10)
(10.2.4) improves to ν • F ∈ H p (∂Ω, σ), the homogeneous Hardy
space defined as in (4.2.12) with Σ := ∂Ω.
A concrete example of this flavor is offered below.

Example 10.2.2 Assume Ω ⊆ Rn is an open set with an Ahlfors regular boundary

and abbreviate σ := H n−1  ∂Ω. Also, fix an aperture parameter κ ∈ (0, ∞) along
with two exponents p ∈ n−1 n , ∞ and q ∈ (0, ∞]. Start with any scalar function
1,1
u ∈ Wloc (Ω) which satisfies Nκ (∇u) ∈ L p,q (∂Ω, σ) and fix two arbitrary indices
j, k ∈ {1, . . . , n}. Finally, consider the divergence-free. vector field
Fjk
u := (∂ u)e − (∂ u)e in Ω. Then the weak tangential derivative ∂
k j j k τ j k u, defined in
[133, Example 4.2.4] (cf. (A.0.87)-(A.0.88) in the Glossary) as the distribution
.    
∂τ j k u := ν • Fjk
u
= ν • (∂k u)e j − (∂j u)ek ∈ Lipc (∂Ω) , (10.2.11)

satisfies
∫
.    

(Lip c (∂Ω)) ∂τ j k u, Ψ ∂Ω Lip c (∂Ω) = (∂k u)(∂j Ψ) − (∂j u)(∂k Ψ) dL n, (10.2.12)
Ω

for each Ψ ∈ Lipc (Rn ). Also, Theorem 10.2.1 applied to Fjk

u shows that

.
∂τ j k u ∈ H p,q (∂Ω, σ) and there exists a finite C > 0, depending only
.    (10.2.13)
on Ω, p, q, κ, such that ∂τ j k u H p, q (∂Ω,σ) ≤ C Nκ (∇u) L p, q (∂Ω,σ) .

As a consequence of this, (4.2.25), and [133, (6.2.25)],

 
corresponding to the particular case when q := p ∈ n−1 n , ∞ one has
.
∂τ j k u ∈ H p (∂Ω, σ) and there exists C ∈ (0, ∞), depending only on (10.2.14)
.   
Ω, p, κ, such that ∂τ j k u H p (∂Ω,σ) ≤ C Nκ (∇u) L p (∂Ω,σ) .

In addition,

if q := p ∈ n−1 , ∞) and under the additional assumptions that Ω
. n
is bounded, ∂τ j k u actually belongs to the homogeneous Hardy space (10.2.15)
.
H p (∂Ω, σ), defined as in (4.2.12) with Σ := ∂Ω.
In particular, in the range p = q ≥ 1 we have the following result: With the set Ω as
1,1
before, if u ∈ Wloc (Ω) has Nκ (∇u) ∈ L p (∂Ω, σ) for some κ ∈ (0, ∞) and p ∈ [1, ∞)
.
then for any indexes j, k ∈ {1, . . . , n} there exists a unique function, denoted by ∂τ j k u,
which belongs to L p (∂Ω, σ) and satisfies (for some constant C ∈ (0, ∞) independent
of u)
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 637
.   
∂τ u p ≤ C Nκ (∇u) L p (∂Ω,σ), as well as
jk L (∂Ω,σ)
∫ ∫
.   (10.2.16)
(∂τ j k u)Ψ dσ = (∂k u)(∂j Ψ) − (∂j u)(∂k Ψ) dL n
∂Ω Ω
for every complex-valued function Ψ ∈ Lipc (Rn ).

The uniqueness part is a consequence of [133, Corollary 3.7.3]. Moreover, a closer

inspection shows that this result continues to be true when p = ∞ (in which scenario,
we may rephrase the original assumptions on u by the demands that u ∈ Lloc 1 (Ω, L n )
∞ n
and ∇u ∈ L (Ω, L ) ; cf. [133, Lemma 8.3.1]).
n

For a more in-depth analysis we shall return to this important special case later on.
Another special case of Theorem 10.2.1 which naturally presents itself is as
follows. Given an open set Ω ⊆ Rn with an Ahlfors regular boundary, abbreviate
σ := H n−1 ∂Ωand fix an aperture parameter κ ∈ (0, ∞) along with two integrability
exponents p ∈ n−1 n , ∞ and q ∈ (0, ∞]. Then we have the following companion to
the result recorded in (10.1.53):
if u ∈ C ∞ (Ω) is harmonic and satisfies. Nκ (∇u) ∈ L p,q (∂Ω, σ) it fol-
lows that the weak normal derivative ∂ν u := ν • (∇u) belongs to the
(10.2.17)
Lorentz-based Hardy space H p,q (∂Ω, σ), and a naturally accompany-
ing estimate holds.
We shall introduce more general weak conormal derivatives in Definition 10.2.18
and, subsequently, study them in detail (see, in particular, Theorem 10.2.24).
For now, we shall record a version of Theorem 10.2.1 in which all hypotheses
of a global quantitative nature are formulated entirely in terms of the nontangential
maximal operator. Its proof simply combines Theorem 10.2.1 with the last part of
Lemma 10.1.1.

Corollary 10.2.3 Suppose Ω ⊆ Rn (where n ∈ N with n ≥ 2) is an open set with an

Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. Also, fix some aperture
parameter κ ∈ (0, ∞) along with three integral exponents
 n−1  1 
1 −1
p∈ n ,n−1 , p∗ := p − n−1 ∈ (1, ∞) and q ∈ (0, ∞]. (10.2.18)

In this setting, consider a vector field

F ∈ Lloc with divF ∈ Lloc

n
1
(Ω, L n ) 1
(Ω, L n ) (10.2.19)

(where all derivatives taken in the sense of distributions in Ω) satisfying

∗
Nκ F ∈ L p ,q (∂Ω, σ) and Nκ (divF)
 ∈ L p,q (∂Ω, σ). (10.2.20)

Then F belongs to Lbdd

1 (Ω, L n ) n while div F belongs to L 1 (Ω, L n ) and, for
bdd

some constant C = C(Ω, κ, p, q) ∈ (0, ∞) independent of F,
638 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

ν • F belongs to the Lorentz-based Hardy space H p ,q (∂Ω, σ) and

∗

       (10.2.21)
ν • F p ∗, q ≤ C Nκ F p ∗ , q + Nκ (divF)
  p, q .
H (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ)

Moreover, corresponding to the limiting case p = n−1 n , for each vector field as in
(10.2.19) one has (in a quantitative sense)

Nκ F ∈ L 1,∞ (∂Ω, σ) and
n =⇒ ν • F ∈ H 1,∞ (∂Ω, σ). (10.2.22)

Nκ (divF) ∈ L n−1 (∂Ω, σ)

In the theorem below we isolate a pointwise estimate of independent interest,

which is going to be a key ingredient in the proof of Theorem 10.2.1, presented just
a little later.

Theorem 10.2.4 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an

Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. In this setting, consider
a vector field satisfying (with all derivatives taken in the sense of distributions in Ω)

F ∈ Lbdd and divF ∈ Lbdd

n
1
(Ω, L n ) 1
(Ω, L n ). (10.2.23)

Then, with the bullet product considered as in [133, Proposition 4.2.3] (cf. also
(A.0.79) for a brief review),
 
ν • F is well defined in Lipc (∂Ω) (10.2.24)

and for each choice of parameters

n−1+γ
κ ∈ (0, ∞), γ ∈ (0, 1), α ∈ 0, , (10.2.25)
n−1
the following pointwise inequality holds, involving the Fefferman-Stein grand max-
imal function (cf. (4.1.6)), the nontangential maximal operator (cf. (A.0.74)), the
Hardy-Littlewood maximal function on ∂Ω (cf. (A.0.71)), and the P-maximal oper-
ator defined in (10.1.2):
    1 
α  
ν • F γ (x) ≤ C M ∂Ω |Nκ F | α (x) + C P divF (x)
(10.2.26)
at each x ∈ ∂Ω, for some C = (Ω, n, κ, γ, α) ∈ (0, ∞).

We wish to note that for each

n−1
s∈ ,∞ (10.2.27)
n−1+γ

we may reformulate (10.2.26) (written for α := 1/s) simply as

     
ν • F γ ≤ CM ∂Ω,s Nκ F + C P divF on ∂Ω, (10.2.28)
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 639

where M ∂Ω,s is the L s -based Hardy-Littlewood maximal operator on ∂Ω (cf.

(A.0.69)).
Proof of Theorem 10.2.4 In view of (10.2.23) we may invoke [133, Proposi-
tion 4.2.3] to conclude that (10.2.24) holds. To proceed, suppose κ, γ, α are as in
(10.2.25). Fix an arbitrary point x ∈ ∂Ω along with an arbitrary function ψ ∈ Tγ (x).
Recall
 from (4.1.2)-(4.1.1)
 that this means that ψ ∈ Lipc (∂Ω) and there exists some
r ∈ 0, 2 diam(∂Ω) such that

1
supp ψ ⊆ ∂Ω ∩ B(x, r), sup |ψ| ≤  ,
∂Ω σ ∂Ω ∩ B(x, r)
(10.2.29)
and ψ C.γ (∂Ω) ≤  .
r γ · σ ∂Ω ∩ B(x, r)

 which, for some constant

Then Lemma 10.1.3 guarantees the existence of a function ψ
c ∈ (0, ∞) depending only on n and γ, satisfies

 ∈ C ∞ (Ω) ∩ Lip(Ω) with ψ
ψ  = ψ, supp ψ  ⊆ B(x, 2r) ∩ Ω,
∂Ω
  c
| ≤
1−γ
sup δ∂Ω |∇ψ  , (10.2.30)
Ω r γ · σ ∂Ω ∩ B(x, r)
1
| ≤ 
and sup | ψ .
Ω σ ∂Ω ∩ B(x, r)

Thanks to (10.2.30) and (A.0.79)-(A.0.80) in [133, Proposition 4.2.3], the functional
ν • F ∈ Lipc (∂Ω) acts on the given function ψ ∈ Lipc (∂Ω) according to
∫ ∫
 
ψ =
ν • F, F · ∇ψ
 dL n + (divF) ψ dL n . (10.2.31)
Ω Ω

To estimate the first term in the right-hand side of (10.2.31), the plan is to use
[133, Proposition 8.6.15], for the closed set Σ := ∂Ω, the measure σ := H n−1 ∂Ω,
the power β := n − 1, the aperture parameter κ as in (10.2.25), the power
 n−1+γ
α ∈ 1, , (10.2.32)
n−1
the parameter λ := 1 − γ < n − αβ, the integrability exponent p := 1, the Lebesgue
measurable set E := Ω ∩ B(x, 2r) ⊆ Rn \ Σ, and the Lebesgue measurable function
u := | F | restricted to E. Let us also observe that this choice of E implies (cf. [133,
(8.1.17)])  
πκ (E) ⊆ ∂Ω ∩ B x, (4 + 2κ)r . (10.2.33)
Upon recalling that M ∂Ω denotes the Hardy-Littlewood maximal operator on ∂Ω
(cf. (A.0.71)), much as in (10.1.46) we obtain
640 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
∫  ∫   n
  dL n  ≤
 F · ∇ψ | F | ∇ψ
 dL
Ω Ω
∫
c γ−1
≤ γ
  δ∂Ω | F | dL n
r · σ ∂Ω ∩ B(x, r) Ω∩B(x,2r)
∫
c γ−1  
= γ
  δ∂Ω 1Ω∩B(x,2r) F dL n
r · σ ∂Ω ∩ B(x, r) Ω

c n−α(n−1)−1+γ
≤   sup δ∂Ω ×
rγ · σ ∂Ω ∩ B(x, r) Ω∩B(x,2r)
∫
 Ω∩B(x,2r)  1 α
× Nκ F α dσ
∂Ω∩B(x,(4+2κ)r)
⨏
 Ω∩B(x,2r)  1 α
≤C Nκ F α dσ
∂Ω∩B(x,(4+2κ)r)

 1 
α
≤ C M ∂Ω |Nκ F | α (x) . (10.2.34)

Specifically, the first inequality above is clear, the second inequality is implied
by (10.2.30), and the subsequent equality is obvious. Next, the third inequality in
(10.2.34) is a consequence of [133, Proposition 8.6.15] (used here in the manner
specified in the build up to (10.2.34)), keeping in mind (10.2.33). The penultimate
inequality in (10.2.34) relies on the upper Ahlfors regularity of ∂Ω and that our
choice of α entails n − α(n − 1) − 1 + γ > 0, while the last inequality in (10.2.34) is
clear from (A.0.71) and [133, (8.2.14)]. This completes the justification of (10.2.34).
To estimate the second term in the right-hand side of (10.2.31), we write
∫  ∫  
 ψ dL n  ≤ divF | ψ
| dL n
 (divF)
Ω Ω
∫
1  
≤   divF dL n
σ ∂Ω ∩ B(x, r) Ω∩B(x,2r)
 
≤ C P divF (x), (10.2.35)

for some finite constant C = C(Ω) > 0, thanks to (10.2.30) and (10.1.2).
Keeping in mind the arbitrariness of ψ ∈ Tγ (x) and x ∈ ∂Ω, from (4.1.6),
(10.2.31), (10.2.35), and (10.2.34) we ultimately conclude that (10.2.26) holds if α
is as in (10.2.32). Finally, that (10.2.26) continues to be valid when α belongs to the
larger range specified in (10.2.25) follows from what we have just proved, (A.0.71),
and Hölder’s inequality. 

Let us now turn to the proof of Theorem 10.2.1.

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 641

Proof of Theorem 10.2.1 Given F as in (10.2.1)-(10.2.2), for each x ∈ ∂Ω and each

r > 0 we may invoke [133, (8.6.50) in Proposition 8.6.3] and the embedding in [133,
(6.2.38)] to conclude that, for some Cr ∈ (0, ∞) which is allowed to depend on r, we
have
   
F 1 ≤ Cr F L p n/(n−1), q (Ω∩B(x,r), L n )
L (Ω∩B(x,r), L n )
 
≤ Cr Nκ F L p, q (∂Ω,σ) < +∞. (10.2.36)

Hence, F belongs to the space Lbdd

n
1 (Ω, L n ) . Since (10.2.2) also implies


(on account of the first line in [133, (6.2.39)]) that P(divF)(x) < +∞ for σ-a.e.
x ∈ ∂Ω, it follows that the hypotheses in (10.2.23) are presently satisfied. As such,
Theorem 10.2.4 implies that ν • F is a well-defined distribution on ∂Ω (cf. (10.2.24))
whose grand maximal function satisfies the estimate recorded in (10.2.26) for each
choice of parameters κ, γ, α as in (10.2.25). In addition, from (4.1.22) we know that
the function
 
ν • F γ : ∂Ω −→ [0, ∞] is lower-semicontinuous, (10.2.37)

when ∂Ω is equipped with the topology inherited from Rn . To proceed, we distinguish

two cases.

Case I: Assume p ∈ n−1 n , 1 . In such a scenario, the parameters κ, γ, α from
(10.2.25) may be chosen so that
1  1 n−1+γ
γ ∈ (n − 1) − 1 , 1 and α ∈ , . (10.2.38)
p p n−1
Since by design αp > 1, from (10.2.26), (10.2.37), [133, (6.2.16)], [133, (6.2.19)],
and [133, (7.2.20)] we may therefore conclude that
  
 ν • F  
γ L p, q (∂Ω,σ)

      
1  α

≤ C P divF  L p, q (∂Ω,σ) + C  M ∂Ω |Nκ F | α 
L p, q (∂Ω,σ)

     
1  α

= C P divF  L p, q (∂Ω,σ) + C M ∂Ω |Nκ F | α  α p, αq
L (∂Ω,σ)

     1 α

≤ C P divF  L p, q (∂Ω,σ) + C Nκ F α  α p, αq
L (∂Ω,σ)
    
= C P divF  L p, q (∂Ω,σ) + C Nκ F L p, q (∂Ω,σ), (10.2.39)
 
for some constant C = CΩ,n,κ, p,q,γ,α ∈ (0, ∞). Hence, ν • F ∈ Lipc (∂Ω) satisfies
       
 ν • F   p, q ≤ C Nκ F p, q + P(divF)
  p, q (10.2.40)
γ L (∂Ω,σ) L (∂Ω,σ) L (∂Ω,σ)
642 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

for some constant C = C(Ω, n, κ, p, q, γ) ∈ (0, ∞) independent of F.  In view of

(4.2.23)-(4.2.24), the claims in (10.2.4)-(10.2.5) follow in this case.
Case II: Assume p ∈ (1, ∞). This time we pick α := 1 (which is a permissible
choice in (10.2.25)), a scenario in which (10.2.26) becomes
     
ν • F γ (x) ≤ C · M ∂Ω Nκ F (x) + C · P divF (x)
(10.2.41)
at every point x ∈ ∂Ω,

where C = (Ω, n, κ, γ) > 0 is a finite constant. Based on this, [133, (6.2.16)],

(10.2.37), and [133, (6.2.20)] we then conclude that (10.2.40) once again holds,
from which the desired conclusions follow.
Case III: Assume p = q = ∞. This case is handled similarly, now starting with
(10.2.41), then relying on Lemma 4.1.3 and Lemma 4.1.2, while keeping in mind
that the Hardy-Littlewood maximal operator M ∂Ω is bounded on L ∞ (∂Ω, σ). The
final equality in (10.2.6) is a direct consequence of [133, Lemma 8.3.2]. The proof
of Theorem 10.2.1 is therefore complete. 

Theorem 10.2.4 continues to be a key ingredient in the proof of the following

variant of Theorem 10.2.1, in which certain weighted Lebesgue spaces are employed
in place of Lorentz spaces.

Theorem 10.2.5 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an

Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω, fix an aperture parameter
κ > 0, and consider a Muckenhoupt weight
 
w ∈ A n p (∂Ω, σ) for some p ∈ n−1
n ,∞ . (10.2.42)
n−1

In addition, suppose F : Ω → Cn is a vector field with Lebesgue-measurable

components satisfying (with all derivatives taken in the sense of distributions in Ω)

Nκ F ∈ L p (∂Ω, wσ), divF ∈ Lloc

1 (Ω, L n ),
(10.2.43)
 ∈ L p (∂Ω, wσ).
and P(divF)

Then
 
F ∈ Lbdd divF ∈ Lbdd ν • F ∈ Lipc (∂Ω)
n
1
(Ω, L n ) , 1
(Ω, L n ), (10.2.44)

(with the bullet product defined as in (A.0.79)), and there exists γ ∈ (0, 1) such that,
in a quantitative fashion,
 
ν • F γ ∈ L p (∂Ω, wσ). (10.2.45)

Furthermore, if the hypotheses in (10.2.42)-(10.2.43) are strengthened to

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 643
 
Nκ F ∈ L p (∂Ω, wσ), divF ∈ Lloc  ∈ L p (∂Ω, wσ)
1 (Ω, L n ) and P div F

for some exponent p ∈ (1, ∞) and Muckenhoupt weight w ∈ Ap (∂Ω, σ),

(10.2.46)
then

ν • F belongs to the weighted Lebesgue space L p (∂Ω, wσ) (10.2.47)

and there exists a constant C = C(Ω, κ, p, w) ∈ (0, ∞), independent of F,  such that
      
ν • F p ≤ C Nκ F p + P(divF)  p . (10.2.48)
L (∂Ω,wσ) L (∂Ω,wσ) L (∂Ω,wσ)

Proof From (10.2.42) and the self-improving property of Muckenhoupt weights

recorded in item (9) in [133, Lemma 7.7.1] (whose present validity is guaranteed by
[133, (3.6.26)]) we see that
 
np np
there exists ε ∈ 0, min n−1 − 1, n−1 −p
(10.2.49)
np
such that if r := n−1 − ε then w ∈ Ar (∂Ω, σ).

If we also
(n − 1)ε r
pick γ ∈ (0, 1) such that 1 − γ < and set α := , (10.2.50)
p p
then such choices ensure that
n−1+γ n
α ∈ 1, and > 1. (10.2.51)
n−1 (n − 1)α
Next, from the first membership in (10.2.43), (10.2.49), and [133, (7.7.105)] we
conclude that
 1
Nκ F α ∈ L αp (∂Ω, wσ) = L r (∂Ω, wσ) ⊆ L 1 (∂Ω, σ) (10.2.52)
loc

hence, ultimately,
Nκ F ∈ Lloc
1/α
(∂Ω, σ). (10.2.53)
In concert with [133, (8.6.51)] (used with p := 1/α) and the second property in
(10.2.51), this further implies (cf. also [133, (8.1.17)]) that
n
F ∈ Lbdd
(n−1)α n n
(Ω, L n ) ⊆ Lbdd
1
(Ω, L n ) . (10.2.54)

Hence, the first condition in (10.2.44) has been established. The second condition
in (10.2.44) is a consequence of the last two properties in (10.2.43) together with
(10.1.4). Granted the first two conditions in (10.2.43), [133,
 Proposition
  4.2.3] applies
and gives that ν • F is a well-defined distribution in Lipc (∂Ω) . The proof of
(10.2.44) is therefore complete.
644 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

As regards (10.2.45), from (4.1.22) we know that the grand maximal function
 
ν • F γ is lower-semicontinuous and, for some constant C = CΩ,n,κ, p,γ,α ∈ (0, ∞),
we may estimate
  
 ν • F  
γ L p (∂Ω,wσ)

      
1  α

≤ C P divF  L p (∂Ω,wσ) + C  M ∂Ω |Nκ F | α 
L p (∂Ω,wσ)

     
1  α

= C P divF  L p (∂Ω,wσ) + C M ∂Ω |Nκ F | α  r
L (∂Ω,wσ)

     1 α

≤ C P divF  L p (∂Ω,wσ) + C Nκ F α  r
L (∂Ω,wσ)
    
= C P divF  L p (∂Ω,wσ) + C Nκ F L p (∂Ω,wσ) . (10.2.55)

Above, the first inequality is implied by the pointwise inequality (10.2.26) in The-
orem 10.2.4 (note that (10.2.25) are satisfied, thanks to the current choices of γ
and α in (10.2.50)-(10.2.51)). The subsequent equality uses αp = r. The second
inequality in (10.2.55) is a consequence of (10.2.49) and the boundedness of the
Hardy-Littlewood maximal operator on Muckenhoupt weighted Lebesgue spaces
(cf. item (1) in [133, Lemma 7.7.1]). The final equality once again uses the fact that
αp = r. This proves that the membership in (10.2.45) holds in a quantitative fashion.
Moving on, work under the assumptions made in (10.2.46). For starters, note
that these hypotheses are stronger that (10.2.42)-(10.2.43) since w ∈ Ap (∂Ω, σ) for
some p ∈ (1, ∞) implies w ∈ A n p (∂Ω, σ) by the monotonicity of the Muckenhoupt
n−1
classes (cf. item (5) in [133, Lemma 7.7.1]). As such, the conclusions in (10.2.44),
(10.2.45), (10.2.55) remain valid in the current setting. Hence,
 
ν • F γ belongs to the weighted Lebesgue space L p (∂Ω, wσ) (10.2.56)

and there exists a constant C = C(Ω, κ, p, w) ∈ (0, ∞), independent of F, such that
       
 ν • F   p ≤ C Nκ F p + P(divF)
 p . (10.2.57)
γ L (∂Ω,wσ) L (∂Ω,wσ) L (∂Ω,wσ)

Next, a combination of (10.2.56), [133, (7.7.105)], and (4.1.29) gives

ν • F ∈ Lloc
1
(∂Ω, σ). (10.2.58)

In concert with Lemma 4.1.2, this further permits us to conclude that there exists
some C ∈ (0, ∞) independent of F such that
   
ν • F ≤ C ν • F  at σ-a.e. point on ∂Ω. (10.2.59)
γ

Finally, from (10.2.58), (10.2.59), (10.2.56), and (10.2.57) we conclude that the
claims in (10.2.47) and (10.2.48) hold. 
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 645

It turns out that there the results established in Theorem 10.2.1, Theorem 10.2.5,
are the manifestation of a more general abstract principle, brought to the forefront in
the proposition below.

Proposition 10.2.6 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an

Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. With L 0 (∂Ω, σ) denoting
the linear space consisting of (equivalence classes of) σ-measurable functions which
are finite σ-a.e. on ∂Ω, let L+0 (∂Ω, σ) denote the cone consisting of functions
f ∈ L 0 (∂Ω, σ) satisfying f (x) ≥ 0 for σ-a.e. x ∈ ∂Ω. Suppose

· : L+0 (∂Ω, σ) −→ [0, +∞] (10.2.60)

is a non-degenerate homogeneous mapping satisfying the quasi-triangle inequality,

i.e.,

for each f ∈ L+0 (∂Ω, σ) one has f = 0 ⇐⇒ f = 0 at σ-a.e. point on ∂Ω,

there is C ∈ (0, ∞) so that f + g ≤ C( f + g ) for all f , g ∈ L+0 (∂Ω, σ),
λf = λ f for each function f ∈ L+0 (∂Ω, σ) and each scalar λ ∈ (0, ∞),
(10.2.61)
which is also assumed to be monotone in the sense that
whenever the functions f , g ∈ L+0 (∂Ω, σ) are such that f ≤ g at
(10.2.62)
σ-a.e. point on ∂Ω then one necessarily has f ≤ g .
Define  
X := f ∈ L 0 (∂Ω, σ) : f X := | f | < +∞ (10.2.63)
and for each power θ ∈ (0, ∞) consider the quasi-normed space
 
X θ := f ∈ L 0 (∂Ω, σ) : | f | 1/θ ∈ X
 θ (10.2.64)
f X θ :=  | f | 1/θ X , ∀ f ∈ X θ .

(Hence, in particular, X 1 = X.) Finally, with M ∂Ω denoting the Hardy-Littlewood

maximal function on ∂Ω, make the assumption that
 θ
there exists some θ ∈ n−1n , 1 such that X ⊆ Lloc (∂Ω, σ)
1
(10.2.65)
and M ∂Ω : X θ −→ X θ is well defined and bounded.

In this setting, suppose F : Ω → Cn is a vector field with Lebesgue-measurable

components satisfying, for some aperture parameter κ > 0 and with all derivatives
taken in the sense of distributions in Ω,

Nκ F ∈ X, divF ∈ Lloc
1  ∈ X.
(Ω, L n ), and P(divF) (10.2.66)

Then
646 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
 
F ∈ Lbdd divF ∈ Lbdd ν • F ∈ Lipc (∂Ω)
n
1
(Ω, L n ) , 1
(Ω, L n ), (10.2.67)

(with the bullet product defined as in (A.0.79)), and there exists γ ∈ (0, 1) such that,
in a quantitative fashion,
 
ν • F γ ∈ X. (10.2.68)

Furthermore, under the additional assumption that X ⊆ Lloc1 (∂Ω, σ) one actually

has ν • F ∈ X and there exists some C ∈ (0, ∞), independent of F,

 such that
       
ν • F ≤ C Nκ F + P(divF)  . (10.2.69)
X X X

Proof
 The
 functional analytic assumptions made in (10.2.61)-(10.2.62) imply that
X, · X is a quasi-normed space which is an order-ideal in L 0 (∂Ω, σ), in the sense
that
if f ∈ X and g ∈ L+0 (∂Ω, σ) are such that |g| ≤ | f | at
(10.2.70)
σ-a.e. point on ∂Ω then g ∈ X and g X ≤ f X .
The first and last lines in (10.2.61) also imply that if f∞ (x) := +∞ for each x ∈ ∂Ω,
then f∞ = +∞. As a consequence of this and (10.2.63) we then see that

if f ∈ X then there exists x0 ∈ ∂Ω such that | f (x0 )| < +∞. (10.2.71)

In concert with the last property in (10.2.66) and (10.1.4), this allows us to conclude
that
divF ∈ Lbdd
1
(Ω, L n ). (10.2.72)
 n−1
To proceed, with θ ∈ n , 1 as in (10.2.65), set

1  n n−1+γ
α := ∈ 1, and pick γ ∈ (0, 1) such that α < . (10.2.73)
θ n−1 n−1

Then the fact that Nκ F belongs to X forces, in view of of the definition made in
 θ
(10.2.64), Nκ F ∈ X θ ⊆ Lloc 1 (∂Ω, σ), hence N F
κ
 ∈ L θ (∂Ω, σ) = L 1/α (∂Ω, σ).
loc loc
Granted this, much as in (10.2.54) we get F ∈ Lbdd 1 (Ω, L n ) n which, together with
 
(10.2.72), goes to show that ν • F is a well-defined distribution in Lipc (∂Ω) (see
[133, Proposition 4.2.3]). In turn, this permits us to conclude that the Fefferman-Stein
 
grand maximal function ν • F γ is well defined and σ-measurable, while (10.2.70)
together with Theorem 10.2.4 and (10.2.64)-(10.2.65) allow us to estimate
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 647
          α 
 ν • F   ≤ C P divF  + C  M ∂Ω |Nκ F | α1 
γ X X X
     
1  α

= C P divF X + C M ∂Ω |Nκ F | α  θ
X
     1 α

≤ C P divF X + C Nκ F α  θ
X
    
= C P divF X + C Nκ FX , (10.2.74)

 This establishes (10.2.68).

for some constant C ∈ (0, ∞) independent of F.
1 (∂Ω, σ), from (10.2.68)
Finally, under the additional assumption that X ⊆ Lloc
 
we conclude that ν • F ∈ L 1 (∂Ω, σ). In concert with [133, (7.7.105)] and
γ loc
(4.1.29), this further implies that ν • F belongs to Lloc
1 (∂Ω, σ). Having proved this,

Lemma 4.1.2 guarantees the existence of a constant C ∈ (0, ∞) independent of F

   
such that ν • F ≤ C ν • F γ at σ-a.e. point on ∂Ω. From this, (10.2.70), and
(10.2.74) we then conclude that we actually have ν • F ∈ X and there exists some
C ∈ (0, ∞), independent of F,  such that (10.2.69) holds. 

A significant particular case of Proposition 10.2.6, involving the scale of Morrey

spaces, is as follows. Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an
Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. Recall from §6 that the
Morrey space M p,λ (∂Ω, σ) is defined for each integrability exponent p ∈ (0, ∞) and
parameter λ ∈ (0, n − 1) as
 
M p,λ (∂Ω, σ) := f ∈ L 0 (∂Ω, σ) : f M p, λ (∂Ω,σ) < +∞ (10.2.75)

where, for each function f ∈ L 0 (∂Ω, σ),

 ⨏ 
n−1−λ 1
p
f M p, λ (∂Ω,σ) := sup R p | f | dσ
p
. (10.2.76)
x ∈∂Ω and ∂Ω∩B(x,R)
0<R<2 diam(∂Ω)

Thus, if we consider

· : L+0 (∂Ω, σ) −→ [0, +∞] defined for each f ∈ L 0 (∂Ω, σ) by

 
n−1−λ ⨏ 1
(10.2.77)
p
f := sup R p | f | dσ
p
,
x ∈∂Ω and ∂Ω∩B(x,R)
0<R<2 diam(∂Ω)

it follows that the properties stated in (10.2.61)-(10.2.62) are currently satisfied. Also,
with X defined as in (10.2.63) and X θ defined as in (10.2.64) for each θ ∈ (0, ∞),
we have

X = M p,λ (∂Ω, σ) and X θ = M p/θ,λ (∂Ω, σ) for each θ ∈ (0, ∞). (10.2.78)
648 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Consequently, from Corollary 6.2.13

 we see that the hypothesis made in (10.2.65) is
presently satisfied whenever p ∈ n−1 n , ∞ . In light of these observations, the analysis
in (10.2.66)-(10.2.69) currently proves the following result, placing the bullet product
in a Morrey space.
Corollary 10.2.7 Suppose Ω ⊆ Rn (where n ∈ N with n ≥ 2) is an openset with
an Ahlfors regular boundary and abbreviate σ := H n−1 ∂Ω. Fix p ∈ n−1 n ,∞

along with λ ∈ (0, n − 1) and assume F : Ω → C is a vector field with Lebesgue-
n

measurable components satisfying, for some aperture parameter κ > 0 and with all
derivatives taken in the sense of distributions in Ω,

Nκ F ∈ M p,λ (∂Ω, σ), divF ∈ Lloc

1 (Ω, L n ),
(10.2.79)
 ∈ M p,λ (∂Ω, σ).
and P(divF)

Then, with the bullet product defined as in (A.0.79),

 
F ∈ Lbdd
1 (Ω, L n ) n, divF ∈ Lbdd
1 (Ω, L n ), ν • F ∈ Lipc (∂Ω) ,
  (10.2.80)
and there exists γ ∈ (0, 1) such that ν • F γ ∈ M p,λ (∂Ω, σ).

Moreover, if in fact p ∈ (1, ∞), one actually has ν • F ∈ M p,λ (∂Ω, σ) and there
exists some constant C ∈ (0, ∞), independent of F,  such that
      
ν • F p, λ ≤ C Nκ F p, λ + P(divF)  p, λ . (10.2.81)
M (∂Ω,σ) M (∂Ω,σ) M (∂Ω,σ)

The statement of Corollary 10.2.7 naturally streamlines in the special case when
the vector field F is actually divergence-free (in the sense of distributions in Ω). A
concrete example of this flavor is offered below (compare with Example 10.2.2).
Example 10.2.8 Suppose Ω ⊆ Rn is an open set with an Ahlfors regular boundary
and abbreviate σ := H n−1 ∂Ω. Also, fix an aperture parameter κ ∈ (0, ∞) along
with an integrability exponent p ∈ (1, ∞) and a parameter λ ∈ (0, n − 1). Assume
1,1
the function u ∈ Wloc (Ω) has the property that Nκ (∇u) ∈ M p,λ (∂Ω, σ) and fix two
arbitrary indices j, k ∈ {1, . . . , n}. Then Corollary 10.2.7 applied to the divergence-
free vector field Fjk
u := (∂ u)e − (∂ u)e shows that the weak tangential derivative
.
k j j k
∂τ j k u, originally defined in [133, Example 4.2.4] (cf. (A.0.87)-(A.0.88)) as the
   
distribution ν • Fjku = ν • (∂ u)e − (∂ u)e
k j j k ∈ Lipc (∂Ω) , satisfies

.
∂τ j k u ∈ M p,λ (∂Ω, σ) and there exists a finite C > 0, depending only
.    (10.2.82)
on Ω, p, λ, κ, such that ∂τ j k u M p, λ (∂Ω,σ) ≤ C Nκ (∇u) M p, λ (∂Ω,σ) .

To solidify the link between Theorem 10.2.1 and Theorem 10.1.4 we need a
companion compatibility result guaranteeing that, under the assumptions of Theo-
rem 10.1.4, the weak normal trace ν • F coincides (as a functional) with the normal
 κ−n.t. 
trace of F computed in the nontangential pointwise sense, i.e., with ν · F∂Ω . This
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 649

is accomplished in Proposition 10.2.9 below (which, together with Theorem 10.2.1

specialized to p = 1, yields another, albeit more circuitous, proof of Theorem 10.1.4).

Proposition 10.2.9 Let Ω ⊆ Rn be an Ahlfors regular domain. Consequently, Ω

is a set of locally finite perimeter, and its geometric measure theoretic outward
unit normal ν is defined almost everywhere on ∂Ω with respect to the measure
σ := H n−1 ∂Ω. Fix an aperture parameter κ ∈ (0, ∞) and consider a vector field
F : Ω → Cn , having L n -measurable components, with the property that
κ−n.t.
F∂Ω exists σ-a.e. on ∂Ω and Nκ F ∈ Lloc
1
(∂Ω, σ). (10.2.83)

In particular, the (10.2.83) implies F ∈ Lloc

1 (Ω, L n ) n (cf. [133, Lemma 8.3.1]),

and one also assumes that divF, computed in the sense of distributions in Ω, belongs
to Lbdd (Ω, L ). Then
1 n

κ−n.t.  
F ∈ Lbdd ν · F∂Ω ∈ Lloc
n
1
(Ω, L n ) , 1
(∂Ω, σ) → Lipc (∂Ω) (10.2.84)

and κ−n.t.  
ν • F = ν · F∂Ω as functionals in Lipc (∂Ω) . (10.2.85)

Proof Given that we are presently assuming Nκ F ∈ Lloc 1 (∂Ω, σ), from [133,

(8.6.51)] and [133, (8.1.17)] it follows that the vector field F belongs to the space
n/(n−1) n 1 (Ω, L n ) n . Also, from (10.2.83), [133, (8.9.8)], [133,
Lbdd (Ω, L n ) ⊆ Lbdd
(8.9.44)], [133, Proposition 8.8.6], and the present assumptions on Ω, we conclude
κ−n.t.
that ν · F∂Ω belongs to Lloc1 (∂Ω, σ). This justifies the memberships in (10.2.84).

As a consequence, since we are also assuming that div F ∈ Lbdd 1 (Ω, L n ), [133,
 
Proposition 4.2.3] ensures that ν • F is a well-defined functional in Lipc (∂Ω) .
Going further, pick an arbitrary ψ ∈ Lipc (∂Ω) and consider a function Ψ satisfy-
ing (A.0.80). Making use of [133, Lemma 4.2.1] it is straightforward to check that

 = F · ∇Ψ + (divF)Ψ
div(Ψ F)  in D (Ω). (10.2.86)
In particular, keeping in mind that F ∈ Lbdd and divF ∈ Lbdd
1 (Ω, L n ) n 1 (Ω, L n ), we

conclude from (10.2.86) and the properties of Ψ that

 ∈ L 1 (Ω, L n ).
div(Ψ F) (10.2.87)
 
Recall from [133, Proposition 4.2.3] that the functional ν • F ∈ Lipc (∂Ω) acts on
the given ψ according to
∫ ∫ ∫
 
ψ =
ν • F, F · ∇Ψ dL n + (divF)Ψ  dL n =  dL n, (10.2.88)
div(Ψ F)
Ω Ω Ω

where the last equality is a seen from (10.2.86). The idea is now to apply the
version of Divergence Theorem presented in [133, Theorem 1.2.1] to the vector field
650 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Ψ F ∈ Lloc
n
1 (Ω, L n ) . In this regard, notice that

  
0 ≤ Nκ (Ψ F) ≤ Nκ Ψ Nκ F ∈ Lcomp ∞
(∂Ω, σ) · Lloc
1
(∂Ω, σ) ⊆ L 1 (∂Ω, σ),
(10.2.89)
thanks to [133, (8.2.10), (8.2.26), (8.1.17), (8.3.17)], as well as the assumptions
made on the function on Ψ and the vector field F.  In view of [133, (8.2.26)], this
κ−n.t.
ultimately implies Nκ (Ψ F)  ∈ L (∂Ω, σ). In addition, (Ψ F)
1  exists σ-a.e. on ∂Ω
∂Ω
κ−n.t.     κ−n.t.   κ−n.t. 

and, in fact, (Ψ F) = Ψ∂Ω F∂Ω = ψ F∂Ω at σ-a.e. point on ∂Ω, as
∂Ω
seen from [133, (8.9.10), (8.9.11), Proposition 8.8.6], and the current assumptions
on Ψ and F. Finally, we note that the vector field Ψ F vanishes identically outside of
a bounded subset of Ω. Granted all these properties, the Divergence Formula [133,
(1.2.2)] applies and presently gives
∫ ∫
 κ−n.t. 

div(Ψ F) dL =
n
ν · F ∂Ω ψ dσ. (10.2.90)
Ω ∂Ω

From (10.2.90) and (10.2.88), the claim in (10.2.85) now follows. 

Example 10.2.10 Let Ω ⊆ Rn be an Ahlfors regular domain. Set σ := H n−1 ∂Ω and
denote by ν = (ν1, . . . , νn ) the geometric measure theoretic outward unit normal
 to Ω.
Also, fix an aperture parameter κ ∈ (0, ∞) along with two exponents p ∈ n−1 n ,∞
1,1
and q ∈ (0, ∞]. In this setting, suppose u ∈ Wloc (Ω) is a scalar function which
satisfies

Nκ (∇u) ∈ L p,q (∂Ω, σ) ∩ Lloc1 (∂Ω, σ) and

κ−n.t. (10.2.91)
the nontangential trace (∇u)∂Ω exists σ-a.e. on ∂Ω.

Fix an arbitrary pair of indices j, k ∈ {1, . . . , n}. Then (11.10.62), [133, (8.9.8),
κ−n.t. κ−n.t.
Proposition 8.8.4, and Corollary 8.9.6] imply that ν j (∂k u)∂Ω − νk (∂j u)∂Ω be-
1 (∂Ω, σ). In particular, [133, Proposition 4.1.4] ensures that this locally
longs to Lloc
integrable function induces a distribution on ∂Ω. In turn, said distribution belongs
to the Lorentz-based Hardy space H p,q (∂Ω, σ). More precisely, Proposition 10.2.9
and Example 10.2.2 imply that there exists a constant C = C(Ω, κ, p, q) ∈ (0, ∞)
such that
κ−n.t. κ−n.t. .
ν j (∂k u)∂Ω − νk (∂j u)∂Ω = ∂τ j k u ∈ H p,q (∂Ω, σ) and
 κ−n.t. κ−n.t.    (10.2.92)
ν j (∂k u) − ν k (∂ j u)   p, q ≤ C Nκ (∇u) p, q .
∂Ω ∂Ω H (∂Ω,σ) L (∂Ω,σ)

Our next goal is to discuss

 how Corollary 10.1.7 may be naturally extended to
Hardy spaces H p , with p ∈ n−1n , ∞ arbitrary, and even to the more inclusive scale
 n−1 
of Lorentz-based Hardy spaces H , with p ∈ n , ∞ and q ∈ (0, ∞] arbitrary,
p,q

by adopting a suitable notion of weak boundary trace, canonically associated with a

given first-order differential operator. The latter concept, of “bullet product” between
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 651

the principal symbol of a first-order differential operator and a suitable vector field,
is made precise in the proposition below.

Proposition 10.2.11 Let D be an N × M homogeneous first-order system D with

constant complex coefficients as in (10.1.83), and recall that its (real) transpose is
defined as (cf. (A.0.29))

n
αβ
D = − a j ∂j 1≤β ≤M . (10.2.93)
j=1 1≤α ≤ N

M
Also, let Ω be an open subset of Rn and suppose F ∈ Lbdd
1 (Ω, L n ) is a C M -valued
function with the property that DF, considered in the sense of distributions in Ω,
has components in Lbdd1 (Ω, L n ). Consider a functional, denoted by Sym(D; ν) • F,
N
which acts on each ψ ∈ Lipc (∂Ω) in such way that
∫ ∫
 
(−i)Sym(D; ν) • F, ψ = DF, Ψ dL − F, D Ψ dL n,
n
(10.2.94)
Ω Ω

whenever Ψ is a C N -valued function satisfying


Ψ ∈ Lip(Ω) , Ψ∂Ω = ψ, and Ψ ≡ 0
N
(10.2.95)
outside of some compact subset of Ω.

Then the functional Sym(D; ν) • F is meaningfully and unambiguously defined

   M
and, in fact, belongs to the space Lipc (∂Ω) .
Strengthen the hypotheses on the underlying set by assuming that Ω is actually an
Ahlfors regular domain. Denote by ν its geometric measure theoretic outward unit
normal, and and abbreviate σ := H n−1 ∂Ω. Then for any function F : Ω → C M
having L n -measurable components and satisfying (for some aperture parameter
κ ∈ (0, ∞))
κ−n.t.
F ∂Ω exists σ-a.e. on ∂Ω, Nκ F ∈ Lloc
1 (∂Ω, σ),
(10.2.96)
N
and DF ∈ Lbdd
1 (Ω, L n ) ,

one has κ−n.t.

Sym(D; ν) • F = Sym(D; ν) F ∂Ω
   M (10.2.97)
as functionals in Lipc (∂Ω) .

These considerations may be naturally adapted to the setting of Clifford algebras.

Here is an example to that effect.

Example 10.2.12 Let Ω ⊆ Rn be an arbitrary open set and consider a Clifford

algebra-valued function u : Ω → Cn satisfying
652 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

u ∈ Lbdd
1
(Ω, L n ) ⊗ Cn and Du ∈ Lloc
1
(Ω, L n ) ⊗ Cn, (10.2.98)
where D is the Dirac operator

n
D := e j  ∂j . (10.2.99)
j=1

In this setting, we agree to abbreviate

ν
• u := (−i)Sym(D; ν) • u (10.2.100)

where the distribution in the right-hand side is defined as in Proposition 10.2.11

used with D as in (10.2.99). Pairing this distribution to the left (in the sense of [133,
(6.4.45)-(6.4.46)]) with a test function ψ ∈ Lipc (∂Ω) ⊗ Cn gives, on account of
(10.2.94)-(10.2.95), that
∫
 
Lip c (∂Ω)⊗ C n ψ, ν 
• u 
Lip c (∂Ω) ⊗ C n = (DR Ψ)  u dL n
Ω
∫
+ Ψ  (D L u) dL n, (10.2.101)
Ω

where D L, DR are as in (A.0.32)-(A.0.33), and Ψ is a Cn -valued function satisfying


Ψ ∈ Lip(Ω) ⊗ Cn , Ψ∂Ω = ψ, and Ψ ≡ 0
(10.2.102)
outside of some compact subset of Ω.
Moreover, if Ω is actually an Ahlfors regular domain, and if u additionally satisfies
κ−n.t.
u∂Ω exists σ-a.e. on ∂Ω and Nκ u ∈ Lloc
1
(∂Ω, σ) (10.2.103)

for some aperture parameter κ ∈ (0, ∞) and with σ := H n−1 ∂Ω then, with ν
denoting the geometric measure theoretic outward unit normal to Ω, we have the
following pointwise formula for the Clifford bullet product originally introduced in
(10.2.100):
κ−n.t.  
• u = ν  u∂Ω
ν as functionals in Lipc (∂Ω) ⊗ Cn . (10.2.104)

We continue by presenting the proof of Proposition 10.2.11.

Proof of Proposition 10.2.11 The very first claim in the conclusion may be justified
by reasoning much as in the proof of [133, Proposition 4.2.3] with natural alterations
(of purely algebraic nature). Then the claim in (10.2.97) is proved in a parallel fashion
to Proposition 10.2.9, using the version of the Divergence Theorem presented in [133,
Theorem 1.2.1]. 
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 653

Here is the promised version of Corollary 10.1.7, corresponding to the scale of

Lorentz-based Hardy spaces H p,q , with p ∈ n−1
n , ∞ and q ∈ (0, ∞] arbitrary, using
the notion of weak trace from Proposition 10.2.11.

Corollary 10.2.13 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an open set with an

 Abbreviate σ := H ∂Ω and select two integrability
Ahlfors regular boundary. n−1

exponents, p ∈ n , ∞ and q ∈ (0, ∞]. Also, pick two integers N, M ∈ N and con-
n−1

sider an arbitrary N × M homogeneous first-order system D with constant complex

coefficients in Rn , as in (10.1.83), along with a vector-valued function F : Ω → C M ,
having L n -measurable components, with the property that for some κ ∈ (0, ∞) one
has
Nκ F ∈ L p,q (∂Ω, σ). (10.2.105)
M
1 (Ω, L n )
In particular, F ∈ Lloc (cf. [133, Lemma 8.3.1]), and one also assumes
that DF, computed in the sense of distributions in Ω, has components in Lloc
1 (Ω, L n )

and satisfies
P(DF) ∈ L p,q (∂Ω, σ). (10.2.106)
Then
N
Sym(D; ν) • F ∈ H p,q (∂Ω, σ) (10.2.107)
and there exists a constant C = C(Ω, D, p, q, κ) ∈ (0, ∞) such that
   
 
Sym(D; ν) • F  p, q N
≤ C Nκ F  L p, q (∂Ω,σ) + C P(DF) L p, q (∂Ω,σ) .
[H (∂Ω,σ)]
(10.2.108)
Moreover, if in place of (10.2.105) and (10.2.106) one now assumes

Nκ F ∈ L p (∂Ω, wσ) and P(DF) ∈ L p (∂Ω, wσ)

(10.2.109)
for some exponent p ∈ (1, ∞) and weight w ∈ Ap (∂Ω, σ),

then actually

Sym(D; ν) • F belongs to
N (10.2.110)
the weighted Lebesgue space L p (∂Ω, wσ)

and there exists a constant C = C(Ω, κ, p, w) ∈ (0, ∞), independent of F, such that
      
Sym(D; ν) • F  p N ≤ C Nκ F  p + P(DF) p .
[L (∂Ω,wσ)] L (∂Ω,wσ) L (∂Ω,wσ)
(10.2.111)
Finally, if in lieu of (10.2.105)-(10.2.106) one now asks that

Nκ F ∈ M p,λ (∂Ω, σ) and P(DF) ∈ M p,λ (∂Ω, σ)

(10.2.112)
for some exponent p ∈ (1, ∞) and parameter λ ∈ (0, n − 1),

then in fact
654 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
N
Sym(D; ν) • F belongs to the Morrey space M p,λ (∂Ω, σ) (10.2.113)

and there exists a constant C = C(Ω, κ, p, λ) ∈ (0, ∞), independent of F, such that
      
Sym(D; ν) • F  p, λ ≤ C N κ F  + P(DF)  .
[M (∂Ω,σ)] N M p, λ (∂Ω,σ) M p, λ (∂Ω,σ)
(10.2.114)
As an example, specializing Corollary 10.2.13 to the case when D is the Dirac
operator (cf. (A.0.28)) yields the following result.
Example 10.2.14 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an open set with an
 and abbreviate σ := H ∂Ω. Also, fix two integrability
Ahlfors regular boundary n−1

exponents, p ∈ n , ∞ and q ∈ (0, ∞], along some aperture parameter κ ∈ (0, ∞).
n−1

Then, given a Clifford algebra-valued function u : Ω → Cn with L n -measurable

components and satisfying

Nκ u ∈ L p,q (∂Ω, σ), Du ∈ Lloc

1 (Ω, L n ) ⊗ C ,
n

n (10.2.115)
and P(DF) ∈ L p,q (∂Ω, σ), where D := e j  ∂j ,
j=1

Corollary 10.2.13 implies that, with the Clifford bullet product defined as in
(10.2.100),

ν
• u belongs to the space H p,q (∂Ω, σ) ⊗ Cn and
(10.2.116)
ν
•u H p, q (∂Ω,σ)⊗ C n ≤ C Nκ u L p, q (∂Ω,σ) + C P(Du) L p, q (∂Ω,σ),

for some constant C ∈ (0, ∞) independent of u. Moreover, similar results are valid
on Muckenhoupt weighted Lebesgue spaces, as well as on Morrey spaces.
We now turn to the proof of Corollary 10.2.13.
Proof of Corollary 10.2.13 First, let us work under the assumptions made in
(10.2.105) and (10.2.106). Write (Fβ )1≤β ≤M for the scalar components of the vector-
valued function F in Ω. Then, if for each α ∈ {1, . . . , N } we define the vector field
Fα = (Fjα )1≤ j ≤n as in (10.1.95), our current assumptions imply that


N
Nκ Fα ≈ Nκ F uniformly on ∂Ω, (10.2.117)
α=1

and, since divFα = (DF)α in D (Ω) for each α ∈ {1, . . . , N },


N
 
P divFα ≈ P(DF) uniformly on ∂Ω. (10.2.118)
α=1

As a consequence,
 each index α ∈ {1, . . . , N } we have Nκ Fα ∈ L p,q (∂Ω, σ) as
 forp,q
well as P divFα ∈ L (∂Ω, σ). Granted these properties, Theorem 10.2.1 applies
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 655

to each vector field Fα and proves that, in a quantitative sense,

ν • Fα ∈ H p,q (∂Ω, σ) for each α ∈ {1, . . . , N }. (10.2.119)

At this stage in the proof, there remains to observe that given any test vector
N
field ψ = (ψα )1≤α ≤ N ∈ Lipc (∂Ω) , if we let Ψ = (Ψα )1≤α ≤ N be any C N -valued
function as in (10.2.95) then according to (10.2.94) we have
∫  
  αβ αβ
(−i)Sym(D; ν) • F, ψ = a j Fβ (∂j Ψα ) + a j (∂j Fβ )Ψα dL n
Ω
∫ ∫
= Fα · ∇Ψα dL n + (divFα )Ψα dL n
Ω Ω
 
= ν • Fα, ψα , (10.2.120)

hence, in a quantitative fashion,

 
(−i)Sym(D; ν) • F = ν • Fα 1≤α ≤ N ∈ H p,q (∂Ω, σ)
N
. (10.2.121)

Going further, that we actually have (10.2.110)-(10.2.111) if the conditions from

(10.2.109) are now imposed in place of (10.2.105)-(10.2.106) is established in the
same fashion, by relying now on Theorem 10.2.5 (in place of Theorem 10.2.1).
Finally, the claims in (10.2.113)-(10.2.114) may be justified in a similar fashion,
under the assumptions made in (10.2.112), this time invoking Corollary 10.2.7 in
place of Theorem 10.2.1. 
In the next two propositions we elaborate on relationship between the bullet
product and the principal symbol map.

Proposition 10.2.15 Let Ω ⊆ Rn be a bounded (ε, δ)-domain with an Ahlfors regular

boundary, and such that Rn \ Ω is n-thick. Abbreviate σ := H n−1 ∂Ω, and fix two
exponents p, p ∈ (1, ∞) satisfying
 1/p + 1p = 1 along with some s ∈ (0, 1), then set
a := 1 − s − p ∈ − p , 1 − p . Finally, let D be an N × M homogeneous first-order
1 1 1

system with constant (complex) coefficients.

 ap  M
Then for each pair (u, f ), with u belonging to L p Ω, δ∂Ω L n and f belonging
 1, p ∗ N  N
to W−a (Ω) , satisfying Du = f Ω in D (Ω) and with the additional
N
property that Du ∈ L 1 (Ω, L n ) , it follows that
   N
(−i)Sym(D; ν) • u = (−i)Sym(D; ν)(u, f ) in Lip (∂Ω) , (10.2.122)

with the left-hand side defined in the sense of (10.2.94)-(10.2.95), and with the right-
hand side originally defined as in (8.5.41) then regarded as a vector distribution on
∂Ω.

Proof This is a consequence of Proposition 8.5.7 and Proposition 10.2.11. 

656 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Here is the second proposition alluded to above.

Proposition 10.2.16 Suppose Ω ⊆ Rn is an (ε, δ)-domain with a compact Ahlfors
regular boundary, and abbreviate σ := H n−1 ∂Ω. Fix s ∈ (0, 1) along with two
exponents p, q ∈ (1, ∞), and denote by p, q  the Hölder conjugate exponents of p, q.
Also, let A ∈ {B, F}, and let D be an N × M homogeneous first-order system with
constant (complex) coefficients.
p,q M  p,q ∗ N
Then for any pair (u, f ) with u ∈ A 1 (Ω) and f ∈ A 1 (Ω)
s+ p −1 1−s− p 
N
such that Du = f Ω in D (Ω) plus the additional property that Du belongs to
1 (Ω, L n ) N
Lbdd , it follows that
   N
(−i)Sym(D; ν) • u = (−i)Sym(D; ν)(u, f ) in Lip (∂Ω) , (10.2.123)

with the left-hand side defined in the sense of (10.2.94)-(10.2.95) and with the right-
hand side originally defined as in (9.5.28)-(9.5.29) then subsequently viewed as a
vector distribution on ∂Ω.
Proof This follows from Proposition 9.5.5 and Proposition 10.2.11. 
In Corollary 10.2.17 below we present a bi-linear version of Corollary 10.2.13,
in the spirit of Corollary 10.1.8 in which, this time, the trace is taken in the sense

of the bullet-product in the Lorentz-based Hardy space H p,q with p ∈ n−1 n , ∞ and
q ∈ (0, ∞] arbitrary.
Corollary 10.2.17 Fix n ∈ N with n ≥ 2, and suppose Ω is an open subset of Rn with
Ahlfors regular boundary. In particular, Ω is a set of locally finite perimeter whose
geometric measure theoretic outward unit normal ν is defined almost everywhere on
∂Ω with respect to the measure σ := H n−1 ∂Ω.
In this context, pick two integers N, M ∈ N and consider an arbitrary N × M
homogeneous first-order system D with constant complex coefficients in Rn , and
denote by D its (real) transpose. Also, pick two vector-valued functions
M N
F ∈ C 1 (Ω) and G ∈ C 1 (Ω) , (10.2.124)

with the property that for some aperture parameter κ > 0 and some integrability
exponents
 
p0, p1, q0, q1 ∈ (0, ∞] along with p ∈ n−1
n , ∞ and q ∈ (0, ∞]
(10.2.125)
satisfying 1/p0 + 1/p1 = 1/p and 1/q0 + 1/q1 = 1/q

one has
Nκ F ∈ L p0,q0 (∂Ω, σ), Nκ G ∈ L p1,q1 (∂Ω, σ),
    (10.2.126)
P DF, G , P F, D G ∈ L p,q (∂Ω, σ).
Define Sym(D; ν)F • G as the functional acting on each ψ ∈ Lipc (∂Ω) in such a
way that
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 657
∫ ∫
 
(−i)Sym(D; ν)F • G, ψ = DF, GΨ dL n − F, D GΨ dL n
Ω Ω
∫
+ (−i)Sym(D; ∇Ψ)F, G dL n (10.2.127)
Ω

whenever Ψ is a function satisfying


Ψ ∈ Lip(Ω), Ψ∂Ω = ψ, as well as Ψ ≡ 0
(10.2.128)
outside of some compact subset of Ω.
 
Then Sym(D; ν)F • G is well defined as a functional in Lipc (∂Ω) and, in fact,
belongs to the Lorentz-based Hardy space H p,q (∂Ω, σ) in a quantitative fashion;
specifically, there exists some finite constant C = C(Ω, D, κ, p0, p1, q0, q1 ) > 0 such
that
 
 
Sym(D; ν)F • G p, q (10.2.129)
H (∂Ω,σ)
   
≤ C Nκ F  L p0 , q0 (∂Ω,σ)
· Nκ G L p1, q1 (∂Ω,σ)
     
+ C P DF, G  L p, q (∂Ω,σ) + C P F, D G  L p, q (∂Ω,σ) .

Proof Suppose D is written as in (10.1.83), in which case its transpose D is

as in (10.2.93). Denote by (Fβ )1≤β ≤M and (Gα )1≤β ≤ N , respectively, the scalar
components of the vector valued functions F and G, and define the vector field
H = (H j )1≤ j ≤n as in (10.1.110). Then the computation in (10.1.111) continues to be
   
valid and 0 ≤ Nκ H ≤ C Nκ F · Nκ G pointwise on ∂Ω. Granted these, the current
assumptions on F and G together with the version of Hölder’s inequality on the scale
of Lorentz spaces recorded in [133, (6.2.61)] imply (bearing in mind (10.2.125))
 
Nκ H ∈ L p,q (∂Ω, σ) and P divH ∈ L p,q (∂Ω, σ). (10.2.130)

Consequently, Theorem 10.2.1 applies to H and gives that

 
ν • H is well defined as a functional in Lipc (∂Ω) and,
(10.2.131)
in a quantitative sense, ν • H belongs to H p,q (∂Ω, σ).
We next claim that
 
 ψ
(−i)Sym(D; ν)F • G, ψ = ν • H, for every ψ ∈ Lipc (∂Ω). (10.2.132)

To prove this claim, fix ψ ∈ Lipc (∂Ω) and consider some Ψ as in (10.2.128). Then
recalling (A.0.79)-(A.0.80) in [133, Proposition 4.2.3], (10.1.110), (10.1.111), and
(10.1.84) we may write
658 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
∫ ∫
 ψ =
ν • H, H · ∇Ψ dL n +  dL n
(divH)Ψ
Ω Ω
∫ ∫ ∫
αβ
= a j Fβ Gα ∂j Ψ dL n + DF, GΨ dL n − F, D GΨ dL n
Ω Ω Ω
∫ ∫
= (−i)Sym(D; ∇Ψ)F, G dL n + DF, GΨ dL n
Ω Ω
∫
− F, D GΨ dL n . (10.2.133)
Ω

Now (10.2.132) follows from (10.2.133) and (10.2.127). Finally, a combination of

(10.2.131) and (10.2.132) proves that Sym(D; ν)F • G ∈ H p,q (∂Ω, σ) and that
(10.2.129) holds. 

Moving on, we introduce the notion of weak conormal derivative associated with
a given coefficient tensor, as a distribution given by a suitable bullet product.

Definition 10.2.18 Fix two integers M, N ∈ N and consider a coefficient tensor of

type (n × n, M × N), i.e., a block of the form
αβ
A = ar s 1≤r,s ≤n (10.2.134)
1≤α ≤M
1≤β ≤ N

αβ
with each entry ar s a complex number. Associate with this coefficient tensor the
M × N second-order system (with the summation convention over repeated indices
in effect)
αβ
L A := ar s ∂r ∂s 1≤α ≤M . (10.2.135)
1≤β ≤ N

Next, let Ω ⊆ Rn be an open set and consider a vector-valued function u satisfying

(with all derivatives taken in the sense of distributions in Ω)
N M
u ∈ Lloc
1 (Ω, L n ) , L Au ∈ Lbdd
1 (Ω, L n ) ,
 αβ  M×n
(10.2.136)
and A∇u := ar s ∂s uβ α,r ∈ Lbdd
1 (Ω, L n ) .

Finally, introduce the family of vector fields

 αβ 
Fα := (A∇u)α = ar s ∂s uβ 1≤r ≤n for each α ∈ {1, . . . , M }. (10.2.137)

In this setting, define the weak conormal derivative of u associated with the
coefficient tensor A as the distribution
.      M
∂νAu := ν • Fα 1≤α ≤M ∈ Lipc (∂Ω) . (10.2.138)

Since (10.2.136) implies that for each α ∈ {1, . . . , M } we have

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 659

Fα ∈ Lbdd and divFα = (L Au)α ∈ Lbdd

n
1
(Ω, L n ) 1
(Ω, L n ), (10.2.139)

     M
we conclude from [133, Proposition 4.2.3] that ν • Fα 1≤α ≤M ∈ Lipc (∂Ω) .
Thus, Definition 10.2.18 is meaningful.
As illustrated by our next lemma, conormal derivatives are inherently linked to
cancellation properties.

Lemma 10.2.19 Fix n ∈ N with n ≥ 2 and let Ω ⊆ Rn be an open set with compact
boundary. With the coefficient tensor A as in (10.2.134), the system L as in (10.2.135),
assume u is a function satisfying
N  αβ  M×n
1 (Ω, L n )
u ∈ Lloc , A∇u := ar s ∂s uβ α,r ∈ Lbdd
1 (Ω, L n )
(10.2.140)
M
and L Au = 0 in D (Ω) .

Then
 .A 
[Lip(∂Ω) ] M ∂ν u, c [Lip(∂Ω)] M = 0 for each c ∈ C M (10.2.141)

provided
either Ω is bounded,⨏ or Ω is an exterior domain and there exists
λ ∈ (1, ∞) such that B(0,λ R)\B(0,R) |∇u| dL n = o(R1−n ) as R → ∞. (10.2.142)

Proof In the case when Ω is bounded this is an immediate consequence of

(10.2.137)-(10.2.138), (10.2.140), (A.0.79)-(A.0.80), and the fact that we may take
the function Ψ to be constant in (A.0.80). In the case when Ω is an exterior domain,
fix a bump-function θ ∈ Cc∞ (Rn ) with θ ≡ 1 on B(0, 1) and define Ψ j := θ(·/ j)
for each j ∈ N. Then, given any c = (cα )1≤α ≤M ∈ C M , use (10.2.137)-(10.2.138),
(10.2.140), (A.0.79)-(A.0.80) to write for each j ∈ N sufficiently large
 .A     
[Lip(∂Ω) ] M ∂ν u, c [Lip(∂Ω)] M = ν • Fα, cα = cα ν • Fα, 1
∫
= cα Fα · ∇Φ j dL n
Ω
∫
αβ
= cα ar s ∂s uβ ∂r Φ j dL n
Ω

= o(1) as j → ∞, (10.2.143)

where the last equality uses the decay condition stipulated in (10.2.142). Upon letting
j → ∞, we obtain (10.2.141). 

The main point of our next proposition is to express the weak conormal derivative
as a bullet product, of the sort indicated in (10.2.148) below.
660 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Proposition 10.2.20 Let Ω ⊆ Rn be an open set. Fix M, N ∈ N and consider a

coefficient tensor of type (n×, M × N),
αβ
A = ar s 1≤r,s ≤n . (10.2.144)
1≤α ≤M
1≤β ≤ N

Associate with this coefficient tensor the M × N second-order system (as always, the
summation convention over repeated indices is in effect)
αβ
L A := divA∇ := ar s ∂r ∂s 1≤α ≤M (10.2.145)
1≤β ≤ N

as well as the M × (N · n) first-order system

α(β,s) α(β,s) αβ
D A := 
ar ∂r 1≤α ≤M with each 
ar := ar s . (10.2.146)
(β,s)∈ {1,..., N }×{1,...,n}

Assume u is a C N -valued function in Ω satisfying

N n·N
u = (uβ )1≤β ≤ N ∈ Lloc
1 (Ω, L n ) , ∇u ∈ Lbdd
1 (Ω, L n ) ,
(10.2.147)
M
L Au ∈ Lbdd
1 (Ω, L n ) .

Then the weak conormal derivative of u satisfies

.      M
∂νAu = (−i)Sym D A; ν • (∇u) in Lipc (∂Ω) , (10.2.148)

where the distribution in the right-hand side of (10.2.148) is defined as in Propo-

sition 10.2.11 (presently used with D := D A), regarding ∇u as the C N ·n -valued
function in Ω whose scalar component with index (β, s) ∈ {1, . . . , N } × {1, . . . , n} is
∂s uβ .

Proof Note that the assumptions in (10.2.147) are stronger than those in (10.2.136),
.    M
so Definition 10.2.18 ensures that ∂νAu is well-defined in Lipc (∂Ω) . To elab-
orate on the nature of the right-hand side of (10.2.148), consider the vector-valued
function
 
F = F(β,s) (β,s)∈ {1,..., N }×{1,...,n} : Ω −→ C N ·n (10.2.149)

with components

F(β,s) := ∂s uβ for each (β, s) ∈ {1, . . . , N } × {1, . . . , n}. (10.2.150)

1 (Ω, L n ) N ·n
Then F belongs to the space Lbdd and we have D A F = L Au in
the sense of distributions in Ω. In particular, Proposition 10.2.11 guarantees that
       M
(−i)Sym D A; ν • (∇u) = (−i)Sym D A; ν • F is well defined in Lipc (∂Ω) .
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 661

To proceed, observe that, since generally speaking,

 αβ 
DA w = − ar s ∂r wα (β,s) if w = (wα )1≤α ≤M . (10.2.151)

M
Also, fix an arbitrary ψ ∈ Lipc (∂Ω) and consider any C M -valued function Ψ
satisfying

Ψ = (Ψα )1≤α ≤M ∈ Lip(Ω) , Ψ∂Ω = ψ on ∂Ω,
M
(10.2.152)
and Ψ ≡ 0 outside of some compact subset of Ω.
Then from (10.2.94)-(10.2.95) and (10.2.138) it follows that
∫ ∫
 
(−i)Sym(D A; ν) • (∇u), ψ = D A F, Ψ dL − F, DA Ψ dL n
n
Ω Ω
∫ ∫
αβ
= L Au, Ψ dL n + (∂s uβ )ar s ∂r Ψα dL n
Ω Ω
∫ ∫
= (divFα )Ψα dL n + Fα · ∇Ψα dL n
Ω Ω
. 
= ∂νAu, ψ . (10.2.153)

In view of the arbitrariness of ψ, this establishes (10.2.148). 

Proposition 10.2.20 is a particular manifestation of a more general principle, of
the sort described in our next result.
 ∈ N. Suppose L is
Proposition 10.2.21 Let Ω ⊆ Rn be an open set and fix M, N, N
a homogeneous, constant (complex) coefficient, second-order M × N system in Rn .
Consider a factorization of L of the form

L = DD, (10.2.154)

where D is a homogeneous, constant (complex) coefficient, first-order M × N

 system
in R , and D is a homogeneous, constant (complex) coefficient, first-order N
n × N
system in R , say
n


n 
n
= αγ γβ
D b j ∂j 1≤α ≤M and D = bk ∂k 
1≤γ ≤ N
. (10.2.155)
j=1 
1≤γ ≤ N k=1 1≤β ≤ N

Define (with the summation convention over repeated indices in effect)

 αβ  αβ αγ γβ
AD,D
 := a jk 1≤ j,k ≤n where each a jk := 
b j bk , (10.2.156)
1≤α ≤M
1≤β ≤ N

so that
L = L A D,
D
. (10.2.157)
662 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Finally, assume u = (uβ )1≤β ≤ N is a C N -valued function in Ω satisfying

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lbdd
1 (Ω, L n ) ,
(10.2.158)
M
and Lu ∈ Lbdd
1 (Ω, L n ) .

Then the weak conormal derivative of u with respect to the coefficient tensor AD,D

satisfies
. AD,     M
∂ν
D
 ν • (Du) in Lipc (∂Ω)  ,
u = (−i)Sym D; (10.2.159)

where the distribution in the right-hand side of (10.2.159) is defined as in Proposi-


tion 10.2.11 (used here with D := D).

Proof For starters, observe that since

 αβ   αγ γβ 

 ∇u = a jk ∂k uβ 1≤α ≤M = b j bk ∂k uβ 1≤α ≤M
AD,D
1≤ j ≤n 1≤ j ≤n
 αγ 
= b j (Du)γ 1≤α ≤M , (10.2.160)
1≤ j ≤n

1 (Ω, L n ) 
N
the membership of Du to Lbdd entails
M×n
 ∇u ∈ Lbdd (Ω, L ) .
1 n
AD,D (10.2.161)

Granted this, from (10.2.157)-(10.2.158) and Definition 10.2.18 we conclude that

. A    M
∂ν D, D u is well defined in Lipc (∂Ω) . To further describe this vector distribu-
tion, for each α ∈ {1, . . . , M } define the vector field
   αγ 
Fα := AD,D 
 ∇u α = b j (Du)γ 1≤ j ≤n . (10.2.162)

Then, according to (10.2.138),

. AD,      M
u := ν • Fα 1≤α ≤M ∈ Lipc (∂Ω)
D
∂ν . (10.2.163)

M
Let us now fix an arbitrary ψ ∈ Lipc (∂Ω) and consider any C M -valued function
Ψ satisfying

Ψ = (Ψα )1≤α ≤M ∈ Lip(Ω) , Ψ∂Ω = ψ on ∂Ω,
M
(10.2.164)
and Ψ ≡ 0 outside of some compact subset of Ω.
Then (10.2.138) and (10.2.94)-(10.2.95) permit us to compute
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 663
∫ ∫
 .A 
∂ν u, ψ = (divFα )Ψα dL n + Fα · ∇Ψα dL n
Ω Ω
∫ ∫
αγ
= L Au, Ψ dL n + b j (Du)γ ∂j Ψα dL n
Ω Ω
∫ ∫
= Lu, Ψ dL n −  Ψ dL n
Du, D
Ω Ω
 
 ν) • (Du), ψ ,
= (−i)Sym( D; (10.2.165)

which ultimately proves (10.2.159). 

We also have the following result, asserting that the weak conormal derivatives
associated as in Definition 10.2.18 with any two coefficient tensors producing the
same second-order system actually differ by a linear combination of weak tangential
derivatives.

Proposition 10.2.22 Let Ω be an open set in Rn and consider a homogeneous,

constant (complex) coefficient, second-order M × N system L in Rn . Also, suppose
u = (uβ )1≤β ≤ N is a C N -valued function in Ω satisfying
N n·N
1 (Ω, L n )
u ∈ Lloc , ∇u ∈ Lbdd
1 (Ω, L n ) ,
(10.2.166)
M
and Lu ∈ Lbdd
1 (Ω, L n ) .

Then for any two complex coefficient tensors of type (n × n, M × N),

 αβ   αβ 
A = ar s 1≤r,s ≤n and B = br s 1≤r,s ≤n (10.2.167)
1≤α ≤M 1≤α ≤M
1≤β ≤ N 1≤β ≤ N

with the property that

αβ αβ
L = ar s ∂r ∂s 1≤α ≤M and L = br s ∂r ∂s 1≤α ≤M , (10.2.168)
1≤β ≤ N 1≤β ≤ N

one has
. .  αβ αβ  .   M
∂νAu − ∂νB u = 1
2 ar s − br s ∂τr s uβ in Lipc (∂Ω) . (10.2.169)
1≤α ≤M

As a particular manifestation ofthe property recorded in (10.2.169) in the case

 αβ
when B := 0, whenever A = ar s 1≤r,s ≤n is a complex coefficient tensor of type
1≤α ≤M
1≤β ≤ N
(n×n, M × N) with
 αβ the property
 that its associated second-order homogeneous M × N
system L A := ar s ∂r ∂s 1≤α ≤M is identically zero, then for each C N -valued function
1≤β ≤ N
N n·N
u = (uβ )1≤β ≤ N ∈ Lloc
1 (Ω, L n ) with ∇u ∈ Lbdd
1 (Ω, L n ) one has
664 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
. 1 αβ
.   M
∂νAu = 2 ar s ∂τr s uβ 1≤α ≤M in Lipc (∂Ω) . (10.2.170)

αβ αβ αβ
Proof From property (10.2.168) we conclude that if cr s := ar s − br s for each
α, β, r, s then
αβ αβ
cr s = −csr whenever 1 ≤ r, s ≤ n, 1 ≤ α ≤ M, and 1 ≤ β ≤ N. (10.2.171)

For each α ∈ {1, . . . , M } define the vector fields

 αβ  αβ
Fα := ar s ∂s uβ 1≤r ≤n = ar s (∂s uβ )er ,
 αβ  αβ
(10.2.172)
G α := br s ∂s uβ 1≤r ≤n = br s (∂s uβ )er .

Then, based on (10.2.138), (10.2.171), (10.2.172), and (A.0.87) we may write

. .    
∂νAu − ∂νB u = ν • Fα 1≤α ≤M − ν • G α 1≤α ≤M
   αβ 
= ν • Fα − G α = ν • cr s (∂s uβ )er
1≤α ≤M 1≤α ≤M
 αβ   αβ 
= 2ν
1
• cr s (∂s uβ )er + 2ν
1
• cr s (∂s uβ )er
1≤α ≤M 1≤α ≤M
 αβ   αβ 
= 2ν
1
• cr s (∂s uβ )er − 2ν
1
• cr s (∂r uβ )es
1≤α ≤M 1≤α ≤M

1 αβ
 
= 2 cr s ν • (∂s uβ )er − (∂r uβ )es
1≤α ≤M

αβ .
= 1
2 cr s ∂τr s uβ . (10.2.173)
1≤α ≤M

From (10.2.173), the claim made in (10.2.169) follows. 

 D are two
As a particular embodiment of Proposition 10.2.22, we see that if D,
 
first-order systems satisfying DD = 0, then the bullet product Sym D; ν • D turns
out to be a linear combination of weak tangential derivatives. This is made precise
in the corollary below.
 ∈ N. Suppose D
Corollary 10.2.23 Let Ω ⊆ Rn be an open set, and fix M, N, N  is a

homogeneous, constant (complex) coefficient, first-order M × N system in R , and
n

D is a homogeneous, constant (complex) coefficient, first-order N  × N system in Rn ,

say
 n  n
= αγ γβ
D b j ∂j 1≤α ≤M and D = bk ∂k 1≤γ ≤ N , (10.2.174)
j=1 
1≤γ ≤ N k=1 1≤β ≤ N

with the property that

 = 0.
DD (10.2.175)
Then for each C N -valued function
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 665
N n·N
u = (uβ )1≤β ≤ N ∈ Lloc
1
(Ω, L n ) with ∇u ∈ Lbdd
1
(Ω, L n ) (10.2.176)

one has
 
 ν • (Du) = 1  αγ γβ .
(−i)Sym D; 2 br bs ∂τr s uβ 1≤α ≤M
   M (10.2.177)
in Lipc (∂Ω) ,

where the distribution in the left-hand side of (10.2.177) is defined as in Proposi-

 and the distribution in the right-hand side of
tion 10.2.11 (used here with D := D),
(10.2.177) is defined as in [133, Example 4.2.4] (cf. (A.0.87)-(A.0.88)).

Proof This is a consequence of (10.2.159), (10.2.170), and (10.2.156). 

If Ω ⊆ Rn is an open set with an Ahlfors regular boundary and the function

 n−1  listed in (10.2.166) as well as Nκ (∇u) ∈ L (∂Ω, σ) for
u satisfies the properties p,q

some κ > 0, p ∈ n , ∞ , and q ∈ (0, ∞], then from Example 10.2.2 and (10.2.169)
we conclude that
. . M
∂νAu − ∂νB u belongs to ∈ H p,q (∂Ω, σ) (10.2.178)

whenever the coefficient tensors A, B as in (10.2.167)-(10.2.168). Such an observa-

tion brings into focus the issue of establishing the membership of each individual
weak conormal derivative to Lorentz-based Hardy spaces, a task accomplished in
our next theorem.

Theorem 10.2.24 Fix n ∈ N with n ≥ 2, and suppose Ω ⊆ Rn is an open set with an

Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω and denote by ν = (ν j )1≤ j ≤n
the geometric measure theoretic outward unit normal to Ω. Also, fix M, N ∈ N and
consider a coefficient tensor of type (n × n, M × N),
αβ
A = ar s 1≤r,s ≤n . (10.2.179)
1≤α ≤M
1≤β ≤ N

Associate with this coefficient tensor the M × N second-order system (as always, the
summation convention over repeated indices is in effect)
αβ
L A :=: divA∇ := ar s ∂r ∂s 1≤α ≤M . (10.2.180)
1≤β ≤ N

Finally,
 n−1consider
 a C N -valued function u satisfying, for some integrability exponents
p ∈ n , ∞ and q ∈ (0, ∞], and some aperture parameter κ > 0,
N n·M
1 (Ω, L n )
u ∈ Lloc , A∇u ∈ Lloc
1 (Ω, L n ) ,
M
Nκ (A∇u) ∈ L p,q (∂Ω, σ), L Au ∈ Lloc
1 (Ω, L n ) (10.2.181)
and P(L Au) ∈ L p,q (∂Ω, σ).
666 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Then the following statements are true.

(1) There exists a constant C = C(Ω, A, κ, p, q) ∈ (0, ∞) with the property that the
weak conormal derivative of u associated with the coefficient tensor A (in the
sense of Definition 10.2.18) satisfies
. M
∂νAu belongs to the Lorentz-based Hardy space H p,q (∂Ω, σ) and
 .A   
∂ u p, q ≤ C Nκ (A∇u) L p, q (∂Ω,σ) + C P(L Au) L p, q (∂Ω,σ) .
ν [H (∂Ω,σ)] M
(10.2.182)
(2) Whenever
Ω is actually an Ahlfors regular domain and, in addition to (10.2.181),
one assumes that Nκ (A∇u) belongs to Lloc
1 (∂Ω, σ) and the nontangen-
κ−n.t.
tial pointwise trace (A∇u)∂Ω exists (in C N ·n ) σ-a.e. on ∂Ω,
(10.2.183)
it follows that the weak conormal derivative of u agrees with the distribution
associated (as in [133, Proposition 4.1.4]) with the conormal derivative of u
taken in a pointwise sense, i.e.,
.  αβ  κ−n.t.    M
∂νAu = νr ar s ∂s uβ ∂Ω in Lipc (∂Ω) . (10.2.184)
1≤α ≤M

(3) If, in fact,

Ω is actually an Ahlfors regular domain, the conditions in (10.2.181)
hold with p ∈ [1, ∞) and q ∈ (0, p], and one additionally asks that the
κ−n.t.
nontangential boundary trace (A∇u)∂Ω exists σ-a.e. on ∂Ω,
(10.2.185)
.A M
then the weak conormal derivative ∂ν u ∈ H (∂Ω, σ)p,q from (10.2.182)
M
actually belongs to the Lebesgue space L p (∂Ω, σ) and one has the pointwise
formula
.  αβ  κ−n.t.
∂νAu = νr ar s ∂s uβ ∂Ω at σ-a.e. point on ∂Ω. (10.2.186)
1≤α ≤M

(4) If in place of (10.2.181) one assumes that, for some exponent p ∈ (1, ∞), weight
w ∈ Ap (∂Ω, σ), and aperture parameter κ > 0,
N n·M
1 (Ω, L n )
u ∈ Lloc , A∇u ∈ Lloc
1 (Ω, L n ) ,
M
Nκ (A∇u) ∈ L p (∂Ω, wσ), L Au ∈ Lloc
1 (Ω, L n ) , (10.2.187)
and P(L Au) ∈ L p (∂Ω, wσ),

then the weak conormal derivative of u associated with the coefficient tensor A
(in the sense of Definition 10.2.18) satisfies
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 667
. M
∂νAu belongs to the weighted Lebesgue space L p (∂Ω, wσ) and
 .A   
 ∂ u p ≤ C Nκ (A∇u) L p (∂Ω,wσ) + C P(L Au) L p (∂Ω,wσ)
ν [L (∂Ω,wσ)] M
(10.2.188)
for some constant C = C(Ω, A, κ, p, w) ∈ (0, ∞) independent of u.
(5) If in lieu of (10.2.181) one now asks that, for some integrability exponent
p ∈ (1, ∞), parameter λ ∈ (0, n − 1), and aperture parameter κ > 0,
N n·M
u ∈ Lloc
1 (Ω, L n ) , A∇u ∈ Lloc
1 (Ω, L n ) ,
M
Nκ (A∇u) ∈ M p,λ (∂Ω, σ), L Au ∈ Lloc
1 (Ω, L n ) , (10.2.189)
and P(L Au) ∈ M p,λ (∂Ω, σ),

then the weak conormal derivative of u associated with the coefficient tensor A
(in the sense of Definition 10.2.18) satisfies
. M
∂νAu belongs to the Morrey space M p,λ (∂Ω, σ) and
 .A   
∂ u p, λ ≤ C Nκ (A∇u) M p, λ (∂Ω,σ) + C P(L Au) M p, λ (∂Ω,σ)
ν [M (∂Ω,σ)] M
(10.2.190)
for some constant C = C(Ω, A, κ, p, λ) ∈ (0, ∞) independent of u.

As an example, assume Ω ⊆ Rn is an Ahlfors regular domain with geometric

measure theoretic outward unit normal ν and surface measure σ := H n−1 ∂Ω. In
this setting, consider a harmonic function u in Ω which,
 for some
 aperture parameter
κ > 0, satisfies Nκ (∇u) ∈ L p,q (∂Ω, σ) for some p ∈ n−1 n , ∞ and q ∈ (0, ∞]. Then
the weak normal derivative
.
∂ν u := ν • ∇u belongs to
(10.2.191)
the Lorentz-based Hardy space H p,q (∂Ω, σ).
κ−n.t.
Moreover, in the case when p ∈ [1, ∞), q ∈ (0, p], and (∇u)∂Ω exists σ-a.e. on ∂Ω,
then said weak normal derivative may actually be computed in a pointwise sense as
 κ−n.t. 
the inner product ν, (∇u)∂Ω at σ-a.e. point on ∂Ω.
Proof of Theorem
  10.2.24 With the vector-valued function u as in α(10.2.181) for
some p ∈ n−1n , ∞ , q ∈ (0, ∞], and κ > 0, each of the vector fields F defined as in
(10.2.137) satisfies

Nκ Fα ∈ L p,q (∂Ω, σ), divFα ∈ Lloc

1 (Ω, L n ),

    (10.2.192)
and P divFα = P (L Au)α ∈ L p,q (∂Ω, σ).

Granted these, Theorem 10.2.1 applies and from (10.2.4)-(10.2.5) we conclude

that (10.2.182) holds. Next, the claim in item (2) is implied by (10.2.138) and the
compatibility property established in Proposition 10.2.9. In turn, the claim in item
668 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

(3) is a consequence of (2) in view of the fact that the present conditions on p, q
entail L p,q (∂Ω, σ) ⊆ L p (∂Ω, σ) ⊆ Lloc
1 (∂Ω, σ); cf. [133, (6.2.25)-(6.2.26)].

Next, if we now assume that the vector-valued function u is as in (10.2.181) for

some p ∈ (1, ∞), w ∈ Ap (∂Ω, σ), and κ > 0, then each of the vector fields Fα
defined as in (10.2.137) satisfies

Nκ Fα ∈ L p (∂Ω, wσ), divFα ∈ Lloc

1 (Ω, L n ),

    (10.2.193)
and P divFα = P (L Au)α ∈ L p (∂Ω, wσ).

This, in turn, makes it possible to invoke (10.2.109)-(10.2.111) to conclude the

claims in (10.2.188). Finally, the claims in item (5) are handled similarly, making
use of the results in the very last part of Corollary 10.2.13. 
In the next corollary we look at the case when the conormal derivative is produced
by a factorization of an arbitrary second-order system into two first-order systems.
As was with Theorem 10.2.24, the goal is to establish memberships of the conormal
derivative, expressed in this fashion, to (Lorentz-based) Hardy spaces.

Corollary 10.2.25 Suppose Ω ⊆ Rn (with n ∈ N satisfying n ≥ 2) is an open set

with an Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω and denote by ν the
 ∈ N and
geometric measure theoretic outward unit normal to Ω. Also, fix M, N, N
assume L is a homogeneous, constant (complex) coefficient, second-order M × N
system in Rn . Consider a factorization of L of the form

L = DD, (10.2.194)

where D is a homogeneous, constant (complex) coefficient, first-order M × N  system

in R , and D is a homogeneous, constant (complex) coefficient, first-order N
n × N
 in R . Finally, consider a C -valued function u satisfying, for some exponents
system n N

p ∈ n−1n , ∞ , q ∈ (0, ∞], and some aperture parameter κ > 0,

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lloc
1 (Ω, L n ) ,
M (10.2.195)
Nκ (Du) ∈ L p,q (∂Ω, σ), Lu ∈ Lloc
1 (Ω, L n ) ,
and P(Lu) ∈ L p,q (∂Ω, σ).

Then the following statements are true.

 
(a) The distribution (−i)Sym D;  ν • (Du) (where the bullet product taken in
the sense of Proposition 10.2.11) belongs to the Lorentz-based Hardy space
H p,q (∂Ω, σ)
M
 D, κ, p, q) ∈ (0, ∞)
and there exists a constant C = C(Ω, D,
with the property that
   
(−i)Sym D; ν • (Du) p, q (10.2.196)
[H (∂Ω,σ)] M
 
≤ C Nκ (Du) L p, q (∂Ω,σ) + C P(Lu) L p, q (∂Ω,σ) .
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 669

(b) Whenever
Ω is actually an Ahlfors regular domain and, in addition to (10.2.195),
one assumes that Nκ (Du) belongs to Lloc
1 (∂Ω, σ) and the nontangential
κ−n.t.
pointwise trace (Du)∂Ω exists (in C N ) σ-a.e. on ∂Ω,


(10.2.197)
it follows that
    κ−n.t.
(−i)Sym D; ν • (Du) = (−i)Sym D;  ν (Du) .
∂Ω
  M (10.2.198)
as distributions in Lipc (∂Ω) .

(c) Whenever
Ω is actually an Ahlfors regular domain, the conditions in (10.2.195)
are assumed to hold with p ∈ [1, ∞) and q ∈ (0, p], and one additionally
κ−n.t.
asks that the nontangential trace (Du)∂Ω to exist σ-a.e. on ∂Ω,
(10.2.199)
 

it follows that the distribution (−i)Sym D; ν • (Du) ∈ H (∂Ω, σ)
p,q M
from
M
item (a) actually belongs to the Lebesgue space L p (∂Ω, σ) and one has the
pointwise formula
    κ−n.t.
 ν (Du)
 ν • (Du) = (−i)Sym D;
(−i)Sym D; ∂Ω
(10.2.200)
at σ-a.e. point on ∂Ω.

(d) If in place of (10.2.195) one assumes that, for some exponent p ∈ (1, ∞), weight
w ∈ Ap (∂Ω, σ), and aperture parameter κ > 0,

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lloc
1 (Ω, L n ) ,
M (10.2.201)
Nκ (Du) ∈ L p (∂Ω, wσ), Lu ∈ Lloc
1 (Ω, L n ) ,
and P(Lu) ∈ L p (∂Ω, wσ),
 
then distribution (−i)Sym D;  ν • (Du) (originally considered as in Proposi-
M
tion 10.2.11) belongs to the weighted Lebesgue space L p (∂Ω, wσ) and
   
(−i)Sym D;
 ν • (Du) p (10.2.202)
[L (∂Ω,wσ)] M
 
≤ C Nκ (Du) L p (∂Ω,wσ) + C P(Lu) L p (∂Ω,wσ),

 D, κ, p, w) ∈ (0, ∞) independent of u.
for some constant C = C(Ω, D,
(e) If in lieu of (10.2.195) one now asks that, for some integrability exponent
p ∈ (1, ∞), parameter λ ∈ (0, n − 1), and aperture parameter κ > 0,
670 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lloc
1 (Ω, L n ) ,
M (10.2.203)
Nκ (Du) ∈ M p,λ (∂Ω, σ), Lu ∈ Lloc
1 (Ω, L n ) ,
and P(Lu) ∈ M p,λ (∂Ω, σ),
 
then distribution (−i)Sym D;  ν • (Du) (originally considered as in Proposi-
M
tion 10.2.11) belongs to the Morrey space M p,λ (∂Ω, σ) and
   
(−i)Sym D;
 ν • (Du) p, λ (10.2.204)
[M (∂Ω,σ)] M
 
≤ C Nκ (Du) M p, λ (∂Ω,σ) + C P(Lu) M p, λ (∂Ω,σ)

 D, κ, p, λ) ∈ (0, ∞) independent of u.
for some constant C = C(Ω, D,

Before presenting the proof of this result we wish to note that a significant
feature of Corollary 10.2.25 is the fact that this is formulated in a coordinate-free,
invariant form. In view of the local, purely real-variable nature of the proof, this
ultimately allows us to naturally adapt Corollary 10.2.25 to the setting of subdomains
on manifolds, and differential operators acting between vector bundles over said
manifold.
Proof of Corollary 10.2.25 Express D, D as in (10.2.155) and define the coefficient
tensor AD,D
 as in (10.2.156). Then, as noted in (10.2.157), we have L = L A D,
D
.
In view of this, (10.2.160), and (10.2.159), all desired conclusions follow from
 ).
Theorem 10.2.24 (applied with A := AD,D 

The special case of Corollary 10.2.25 corresponding to the scenario in which the
second-order system is actually zero deserves special mention.

Corollary 10.2.26 Let Ω ⊆ Rn (with n ∈ N satisfying n ≥ 2) be an open set with an

Ahlfors regular boundary. Abbreviate σ := H n−1 ∂Ω and denote by ν the geometric
measure theoretic outward unit normal to Ω. Also, fix M, N, N ∈ N and consider
a homogeneous, constant (complex) coefficient, first-order M × N  system D in Rn ,
along with a homogeneous, constant (complex) coefficient, first-order N × N system
D in Rn . In relation to these, make the assumption that
 = 0.
DD (10.2.205)

Finally,
 n−1consider
 a C N -valued function u satisfying, for some integrability exponents
p ∈ n , ∞ and q ∈ (0, ∞], and some aperture parameter κ > 0,

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lloc
1 (Ω, L n ) ,
(10.2.206)
Nκ (Du) ∈ L p,q (∂Ω, σ).

Then the following statements are true.

10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 671
 
(i) The distribution (−i)Sym D;  ν • (Du) (where the bullet product taken in
the sense of Proposition 10.2.11) belongs to the Lorentz-based Hardy space
H p,q (∂Ω, σ)
M
and there exists a constant C = C(Ω, D,  D, κ, p, q) ∈ (0, ∞)
with the property that
     
(−i)Sym D;  ν • (Du) p, q ≤ C Nκ (Du) L p, q (∂Ω,σ) . (10.2.207)
[H (∂Ω,σ)] M

Moreover, the claims in items (b)-(c) of Corollary 10.2.25 continue to be valid

in the present setting.
(ii) If in place of (10.2.206) one assumes that, for some exponent p ∈ (1, ∞), weight
w ∈ Ap (∂Ω, σ), and aperture parameter κ > 0,

N 
N
u ∈ Lloc
1 (Ω, L n ) , Du ∈ Lloc
1 (Ω, L n ) ,
(10.2.208)
Nκ (Du) ∈ L p (∂Ω, wσ),
 
then distribution (−i)Sym D;  ν • (Du) (originally considered as in Proposi-
M
tion 10.2.11) actually belongs to the weighted Lebesgue space L p (∂Ω, wσ)
and, for some finite constant C = C(Ω, D,  D, κ, p, w) > 0 independent of u,
     
(−i)Sym D;  ν • (Du) p ≤ C Nκ (Du) L p (∂Ω,wσ) . (10.2.209)
[L (∂Ω,wσ)] M

(iii) If in lieu of (10.2.206) one now asks that, for some integrability exponent
p ∈ (1, ∞), parameter λ ∈ (0, n − 1), and aperture parameter κ > 0,
N 
N
u ∈ Lloc
1
(Ω, L n ) , Du ∈ Lloc
1
(Ω, L n ) , Nκ (Du) ∈ M p,λ (∂Ω, σ),
(10.2.210)
 
then distribution (−i)Sym D;  ν • (Du) (originally considered as in Proposi-
M
tion 10.2.11) belongs to the Morrey space M p,λ (∂Ω, σ) and
     
(−i)Sym D;
 ν • (Du) p, λ ≤ C Nκ (Du) M p, λ (∂Ω,σ) (10.2.211)
[M (∂Ω,σ)] M

 D, κ, p, λ) ∈ (0, ∞) independent of u.
for some constant C = C(Ω, D,

For example, if D := div and if for some fixed j, k ∈ {1, . . . , n} we take D to be the
differential operator mapping a distribution u into Du := (∂j u)ek − (∂k u)e j , then the
cancellation property (10.2.205) is satisfied. In such a scenario, Corollary 10.2.26
yields the results described in Example 10.2.2, Example 10.2.8, Example 10.2.10.
Other significant cases when the cancellation property (10.2.205) is satisfied is
offered by the choice D  := d and D := d, where d is the exterior derivative operator
acting on differential forms, or D := δ and D := δ, where δ is the formal transpose of
d (hence, in particular, one may take D  := curl and D := ∇, etc). Once again, similar
results hold in the context of differential operators acting between vector bundles on
manifolds.
672 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Proof of Corollary 10.2.26 This follows by specializing Corollary 10.2.25 to the

trivial case when L = 0. 

We next consider the membership of the normal component of a divergence-free

vector field to Hardy spaces defined on the boundary of an Ahlfors regular domain.

Proposition 10.2.27 Let Ω ⊆ Rn (where n ∈ N satisfies n ≥ 2) be an Ahlfors regular

domain. Abbreviate σ := H n−1 ∂Ω and denote by ν the geometric measure theoretic
outward unit normal to Ω. Also, consider a vector-valued function

k ∈ C 1 (Rn \ {0}) satisfying div k = 0 in Rn \ {0} and with the

n


property that there exists some C ∈ (0, ∞) such that | k(x)| ≤ C|x| 1−n

and |(∇ k)(x)| −n
≤ C|x| at each point x ∈ R \ {0}.
n

(10.2.212)

Then the following statements are true.

(1) When regarded as a distribution on ∂Ω,

  
 − ·)
the (locally integrable) function ν, k(x belongs to
∂Ω
(10.2.213)
the weak Hardy space H 1,∞ (∂Ω, σ) for each point x ∈ Ω,

in a uniform fashion, i.e.,

   
 
sup  ν, k(x − ·)∂Ω  < +∞. (10.2.214)
x ∈Ω H 1, ∞ (∂Ω,σ)

Moreover, the assignment

  
Ω  x −→ ν, k(x − ·) ∈ H 1,∞ (∂Ω, σ) is continuous. (10.2.215)
∂Ω

(2) The difference

     
 0 − ·)
ν, k(x  1 − ·)
− ν, k(x
∂Ω ∂Ω
(10.2.216)
belongs to n−1 H p (∂Ω, σ) for each x0, x1 ∈ Ω,
n <p<∞

though the uniformity of the membership is now lost.

Proof If Ω = Rn there is nothing to prove, so assume Ω  Rn . Fix an aperture

parameter κ > 0 and a point xo ∈ Ω. From item (6) in [133, Lemma 5.10.9] we know
that Ω− := Rn \Ω is also an Ahlfors regular domain, whose topological and geometric
measure theoretic boundaries agree with those of Ω, and whose geometric measure
theoretic outward unit normal is −ν at σ-a.e. point on ∂Ω. If we now consider the
vector field
 o − x),
Fxo (x) := k(x ∀x ∈ Ω−, (10.2.217)
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 673

then Fxo ∈ C 1 (Ω− ) and divFxo = 0 in Ω− . In addition, from [133, Lemma 8.3.7]
n

we see that there exists a constant C = C(Ω, k, n, κ) ∈ (0, ∞), independent of xo ,
with the property that
  C
NκΩ− Fxo (x) ≤ for σ-a.e. x ∈ ∂Ω. (10.2.218)
|x − xo | n−1
In turn, from (10.2.217), (10.2.218), [133, (6.2.22)] (used with d := n − 1), and [133,
(6.2.16)] we conclude that

NκΩ− Fxo ∈ L 1,∞ (∂Ω, σ) ∩ Lloc 1 (∂Ω, σ) and

  (10.2.219)
supxo ∈Ω NκΩ− Fxo  L 1, ∞ (∂Ω,σ) < +∞.

Let us also note that from [133, (8.9.10)], [133, (8.8.45)], and (A.0.1), it follows that
 κ−n.t. 
the pointwise nontangential trace Fxo ∂Ω (x) exists and is equal to k(x  o − x) at
σ-a.e. point x ∈ ∂Ω. Granted these properties, Theorem 10.1.4 applies (in Ω− ) and,
from (10.1.49), (10.1.50) and (10.2.219) we conclude that the claims in (10.2.213)-
(10.2.214) hold.
As regards the continuity of the assignment in (10.2.215), consider a sequence
{x j } j ∈N ⊆ Ω convergent to a point xo ∈ Ω. For each j ∈ N introduce

 j − x) − k(x
Fj (x) := k(x  o − x) for each x ∈ Ω−, (10.2.220)

then set
f j := NκΩ− Fj at each point on ∂Ω− = ∂Ω. (10.2.221)
In relation to this sequence of functions, we claim that
 
lim f j = 0 in L p (∂Ω, σ) for each p ∈ n−1n ,∞ . (10.2.222)
j→∞

To justify (10.2.222), first observe that (10.2.220), (10.2.212), and the Mean Value
Theorem ensure the existence of a rank a constant jo = j(Ω, xo ) ∈ (0, ∞) and a
 n, xo ) ∈ (0, ∞) with the property that
constant C = C(Ω, k,
C
| Fj (x)| ≤ for each x ∈ Ω−, whenever j ≥ jo . (10.2.223)
1 + |x| n
In concert with (10.2.221) and [133, (8.3.47)], this implies
C
0 ≤ f j (x) ≤ for each x ∈ ∂Ω, whenever j ≥ jo . (10.2.224)
1 + |x| n
We next make the sub-claim that

lim f j (x) = 0 for each x ∈ ∂Ω. (10.2.225)

j→∞
674 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

To prove this, fix an arbitrary threshold ε > 0 and choose R ∈ (0, ∞) large enough
so that
C
< ε for all x ∈ Rn \ B(0, R). (10.2.226)
1 + |x| n
Together with (10.2.223) this shows that

| Fj (x)| < ε for each x ∈ Ω− \ B(0, R), whenever j ≥ jo . (10.2.227)

Next, the fact that k is continuous in Rn \ {0} implies that there exists jε ∈ N with
the property that
 
k(x j − x)− k(x
 o − x) < ε for each x ∈ B(0, R)∩Ω− whenever j ≥ jε . (10.2.228)

On account of (10.2.220), this yields

| Fj (x)| < ε for each x ∈ Ω− \ B(0, R), whenever j ≥ jε . (10.2.229)

All together, (10.2.227), (10.2.229), (10.2.221), and [133, (8.2.15)] give that
 
f j (x) < ε for each x ∈ ∂Ω, whenever j ≥ max jo, jε . (10.2.230)

In view of the arbitrariness of ε > 0, this establishes (10.2.225). With (10.2.224)

and (10.2.225) in hand, the claim made in (10.2.222) is implied by Lebesgue’s
Dominated Convergence Theorem (keeping in mind [133, (7.2.5)]).
In turn, from (10.2.222), (1.3.44), and [133, (6.2.48)] we conclude that

lim f j = 0 in L 1,∞ (∂Ω, σ). (10.2.231)

j→∞

At this stage, we may compute

      
    
lim  ν, k(x j − ·) ∂Ω − ν, k(xo − ·) ∂Ω 
j→∞ H 1,∞ (∂Ω,σ)
  κ−n.t.  
 
= lim ν · Fj ∂Ω 
j→∞ H 1,∞ (∂Ω,σ)
 
≤ C · lim sup NκΩ− Fj  L 1,∞ (∂Ω,σ)
j→∞

= C · lim sup f j L 1,∞ (∂Ω,σ) = 0, (10.2.232)

j→∞

where the first equality is a consequence of (10.2.220), [133, (8.9.10), (8.8.45)],

and (A.0.1), the first inequality is implied by (10.2.212) and Theorem 10.1.4 (used
in Ω− ), the penultimate equality comes from (10.2.221), and the final equality is
guaranteed by (10.2.231). Finally, (10.2.232) finishes the proof of (10.2.215).
To prove (10.2.216), one reasons in a similar fashion. Specifically, having fixed
two arbitrary points x0, x1 ∈ Ω, consider the vector field
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 675

 0 − x) − k(x
Fx0,x1 (x) := k(x  1 − x), ∀x ∈ Ω− . (10.2.233)

Then Fx0,x1 belongs to C 1 (Ω− ) and divFx0,x1 = 0 in Ω− . As before, the nontan-

n
 κ−n.t. 
gential boundary trace Fx0,x1 ∂Ω (x) exists and is equal to k(x  1 − x) at
 0 − x) − k(x
σ-a.e. point x ∈ ∂Ω. In addition, from (10.2.233), (11.9.12), the Mean Value Theo-
rem, and [133, Lemma 8.3.7] we see that there exists some constant Cx0,x1 ∈ (0, ∞),
 n, and κ, such that for σ-a.e. x ∈ ∂Ω we have
which depends on x0 , x1 , Ω, k,
  Cx0,x1
NκΩ− Fx0,x1 (x) ≤ for σ-a.e. x ∈ ∂Ω. (10.2.234)
1 + |x| n
As a consequence of (10.2.234) and [133, (7.2.5)],
!
NκΩ− Fx0,x1 ∈ L p (∂Ω, σ). (10.2.235)
n−1
n <p<∞

Having established this, Theorem 10.1.4 applies (in Ω− ) and proves (10.2.216). 
A version of Proposition 10.2.27 in which the singularity now occurs at boundary
points is discussed later, in Proposition 11.9.2. For now we wish to note that, among
other things, Proposition 10.2.27 implies that the conormal derivative together with
all the tangential derivative of the fundamental solution E of a weakly elliptic system
belong to the weak Hardy space H 1,∞ on the boundary of an Ahlfors regular domain.

Corollary 10.2.28 Let Ω ⊆ Rn , where n ∈ N, n ≥ 2, be an Ahlfors regular domain.

Abbreviate σ := H n−1 ∂Ω and denote by ν = (ν1, . . . , νn ) the geometric measure
αβ
theoretic outward unit normal to Ω. Also, for some M ∈ N, let A = ar s 1≤r,s ≤n
1≤α,β ≤M
be a constant (complex) coefficient tensor with the property that the associated
αβ
homogeneous second-order M × M system L A := ar s ∂r ∂s 1≤α,β ≤M (with the
summation convention over repeated indices in effect) is weakly elliptic in Rn . Finally,
denote by E = (Eαβ )1≤α,β ≤M the matrix-valued fundamental solution associated
with L A as in [130, Theorem 11.1, pp. 393-395]. Then

the function νr ar s (∂s Eγβ )(x − ·)∂Ω
αγ
belongs to
1≤α,β ≤M
(10.2.236)
M×M
the weak Hardy space H 1,∞ (∂Ω, σ) for each x ∈ Ω,

in a uniform fashion, i.e.,


M
 αγ 
sup νr ar s (∂s Eγβ )(x − ·) < +∞. (10.2.237)
H 1, ∞ (∂Ω,σ)
x ∈Ω α,β=1

Furthermore, for any two pairs of indices, α, β ∈ {1, . . . , M } and j, k ∈ {1, . . . , n},
one has
676 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .
 
ν j (∂k Eαβ )(x − ·)∂Ω − νk (∂j Eαβ )(x − ·)∂Ω ∈ H 1,∞ (∂Ω, σ) for all x ∈ Ω
 
and supx ∈Ω ν j (∂k Eαβ )(x − ·) − νk (∂j Eαβ )(x − ·) H 1, ∞ (∂Ω,σ) < +∞.
(10.2.238)
Finally, for each fixed α, β ∈ {1, . . . , M } the difference
 
νr ar s (∂s Eγβ )(x0 − ·)∂Ω − νr ar s (∂s Eγβ )(x1 − ·)∂Ω
αγ αγ

(10.2.239)
belongs to H p (∂Ω, σ) for each x , x ∈ Ω,
n−1
n <p<∞ 0 1

and also
 
ν j (∂k Eαβ )(x0 − ·)∂Ω − νk (∂j Eαβ )(x0 − ·)∂Ω
 
− ν j (∂k Eαβ )(x1 − ·)∂Ω + νk (∂j Eαβ )(x1 − ·)∂Ω
!
belongs to H p (∂Ω, σ) for each x0, x1 ∈ Ω, (10.2.240)
n−1
n <p<∞

though the uniformity of the membership is now lost.

Proof Given any α, β ∈ {1, . . . , M }, the claims in (10.2.236)-(10.2.237) are seen

from (10.2.213)-(10.2.214) for the vector field
 αγ 
k := ar s ∂s Eγβ 1≤r ≤n, (10.2.241)

whereas for each α, β ∈ {1, . . . , M } and j, k ∈ {1, . . . , n} the claims in (10.2.238)

are seen from (10.2.213)-(10.2.214) for the vector field

k := (∂k Eαβ )e j − (∂j Eαβ )ek . (10.2.242)

Finally, given any α, β ∈ {1, . . . , M }, the claims made in (10.2.239) and (10.2.240)
are implied by (10.2.216) used for k as in (10.2.241) and (10.2.242), respectively. In
all cases, k satisfies the properties demanded in (10.2.212), so the desired conclusions
follow. 

We continue by making several comments pertaining to Proposition 10.2.27 and

Corollary 10.2.28.
Comment 1: Let us study in detail what (10.2.238) says in the case when Ω is
the upper half-space R+n+1 (hence, we work in Rn+1 ), and the system L is simply
the Laplacian in Rn+1 (i.e., M = 1 and the coefficient tensor A is the identity
matrix). In particular, ∂Ω = ∂R+n+1 ≡ Rn , σ = L n , and ν = (0, . . . , 0, −1) = −en+1 .
Also, E is the standard fundamental solution for the Laplacian in Rn+1 . Hence, we
have (∇E)(X) = ωn−1 X/|X | n+1 for each X ∈ Rn+1 \ {0}. Fix j ∈ {1, . . . , n} and
take k := n + 1. For these choices, (10.2.238) specialized to points of the form
X = (0, . . . , 0, δ) ∈ R+n+1 with δ > 0 then implies (compare with (4.2.35))
10.2 Weak Normal Boundary Traces of Vector Fields in Hardy and Morrey Spaces 677
xj
n+1
∈ H 1,∞ (Rn, L n ) uniformly in δ > 0. (10.2.243)
(|x| 2 + δ2 ) 2

To be able to prove (10.2.243) directly, bring in the harmonic Poisson kernel and
its conjugates defined at points x ∈ Rn as
2 1 2 xj
P(x) := and Q j (x) := , 1 ≤ j ≤ n. (10.2.244)
ωn (|x| 2 + 1) 2
n+1
ωn (|x| 2 + 1) n+1
2

Then, as is well known, in the sense of tempered distributions in Rn we have (cf.

[130, (4.8.9), p. 178] and [130, (4.8.24), p. 180])
ξ j − |ξ |
P(ξ) #j (ξ) = −i
" = e− |ξ | and Q e , ξ ∈ Rn, (10.2.245)
|ξ |
which, in view of the fact that the j-th Riesz transform R j in Rn (defined as in
(A.0.102) with n replaced by n + 1 and with Σ := Rn × {0} ≡ Rn ) is the multiplier
operator corresponding to the symbol −iξ j /|ξ | (compare with [130, (4.9.12), p. 183])
imply (cf. [130, (4.9.29), p. 186])

R j (P) = Q j for each j ∈ {1, . . . , n}. (10.2.246)

Thus, if Pδ (x) := δ−n P(x/δ) and Q j,δ (x) := δ−n Q j (x/δ) for each δ > 0 and x ∈ Rn ,
we also have

R j (Pδ ) = Q j,δ for each j ∈ {1, . . . , n} and δ > 0. (10.2.247)

Now, recall that each Riesz transform R j maps the Hardy space H p (Rn, L n ) linearly
and boundedly into itself for each p ∈ (0, ∞) (cf. [52, Remark 2, p. 191]). Hence,
by real interpolation, we further have R j : H 1,∞ (Rn, L n ) → H 1,∞ (Rn, L n ). Since
L 1 (Rn, L n ) → H 1,∞ (Rn, L n ) and since Pδ ∈ L 1 (Rn, L n ) uniformly in δ > 0, we
may therefore conclude that Q j,δ = R j (Pδ ) ∈ H 1,∞ (Rn, L n ) uniformly in δ > 0,
proving (10.2.243) via classical Fourier analysis.
In the same vein, let us also observe that the version that (10.2.236)-(10.2.237)
acquire in the above setting is simply δ(|x| 2 + δ2 )−(n+1)/2 ∈ H 1,∞ (Rn, L n ) uniformly
in δ > 0, which follows at once from the fact that Pδ ∈ L 1 (Rn, L n ) uniformly in
δ > 0.
Comment 2: It turns out that (10.2.238) is a purely real-variable result, which does
not make use of the PDE’s nature of E. In fact, as is apparent from the proof of
(10.2.238), a more general phenomenon is at play here, according to which if a
function b ∈ C 1 (Rn \ {0}) has the property that there exists C ∈ (0, ∞) such that
|(∇b)(z)| ≤ C|z| 1−n for each z ∈ Rn \ {0} then for every j, k ∈ {1, . . . , n} we have
 
ν j (∂k b)(x − ·)∂Ω − νk (∂j b)(x − ·)∂Ω ∈ H 1,∞ (∂Ω, σ) for all x ∈ Ω
  (10.2.248)
and supx ∈Ω ν j (∂k b)(x − ·) − νk (∂j b)(x − ·) H 1,∞ (∂Ω,σ) < +∞.
678 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

Comment 3: Retain the context of Corollary 10.2.28. Writing (10.2.236)-(10.2.237)

for the coefficient tensor A in place of A and bearing in mind the manner in which
the fundamental solution from [130, Theorem 11.1, pp. 393-395] behaves under
transposition yields the conclusion that

the function νr asr (∂s Eβγ )(x − ·)∂Ω
γα
belongs to
1≤α,β ≤M
M×M
the weak Hardy space H 1,∞ (∂Ω, σ) for each x ∈ Ω, (10.2.249)
 M  γα 
and supx ∈Ω α,β=1 νr asr (∂s Eβγ )(x − ·) H 1, ∞ (∂Ω,σ) < +∞.

Then from (10.2.249) it follows that we have a linear and bounded mapping in the
context
D : H 1,∞ (∂Ω, σ)∗ M −→ L ∞ (Ω, L n ) M (10.2.250)
defined via
 
    
− ·)∂Ω, fα
 f (x) := − γα
D H 1,∞ (∂Ω,σ) νr asr (∂s Eβγ )(x
(H 1,∞ (∂Ω,σ))∗
1≤β ≤M
M
for all f = ( fα )1≤α ≤M ∈ H 1,∞ (∂Ω, σ)∗ and all x ∈ Ω.
(10.2.251)
Moreover, from (10.2.251) and Lemma 4.7.1 we see that for each vector-valued
. M
function f = ( fα )1≤α ≤M ∈ L p (∂Ω, σ) ∩ C η (∂Ω) with p ∈ (1, ∞) and η ∈ (0, 1)
we may write
 
    
 f (x) = − H 1,∞ (∂Ω,σ) νr asr (∂s Eβγ )(x − ·) , fα
D
γα
∂Ω (H 1,∞ (∂Ω,σ))∗
1≤β ≤M
 ∫
γα
= − νr asr (∂s Eβγ )(x − ·) fα dσ
∂Ω
1≤β ≤M
 
= D f (x) for all x ∈ Ω, (10.2.252)

where D (defined by the last equality) is the double layer potential operator associated
with the coefficient tensor A and the set Ω as discussed at length in [136]. Hence,
D is compatible with D when acting on C M -valued Hölder functions of some
exponent η ∈ (0, 1) which are also p-th power integrable on ∂Ω for some p ∈ (1, ∞).
In particular,
 
sup (D f )(x) ≤ C f [(H 1,∞ (∂Ω,σ))∗ ] M
x ∈Ω
 
≤ C · min f [L p (∂Ω,σ)] M , f .η (10.2.253)
[C (∂Ω)] M

for each function

10.3 The Manifold Setting 679
. M
f ∈ L p (∂Ω, σ) ∩ C η (∂Ω)
(10.2.254)
with p ∈ (1, ∞) and η ∈ (0, 1),
where the constant C ∈ (0, ∞) depends only on the Ahlfors regularity constants of
∂Ω, n, L, p, and η. In this regard, we parenthetically note that since every uniformly
continuous function belonging to a Lebesgue space necessarily vanishes at infinity,
.
it follows that L p (∂Ω, σ) ∩ C η (∂Ω) with p ∈ (0, ∞) and η > 0 is a subspace
of L ∞ (∂Ω, σ); in fact, given any p ∈ (0, ∞) and η > 0 there exists a constant
C = C(∂Ω, p, η) ∈ (0, ∞) with the property that
(1+(n−1)/(pη))−1 1−(1+(n−1)/(pη))−1
f L ∞ (∂Ω,σ) ≤C f L p (∂Ω,σ)
f .η
.ηC (∂Ω) (10.2.255)
for each f ∈ L p (∂Ω, σ) ∩ C (∂Ω).

10.3 The Manifold Setting

Here we wish to remark that all results discussed so far in this chapter have natural
versions in the context of differential operators acting between vector bundles on a
Riemannian manifold (M, g) assumed to satisfy:
M is a connected, compact, boundaryless, oriented manifold of class
C 1 , of real dimension n, equipped with a continuous Riemannian (10.3.1)

metric tensor g = 1≤ j,k ≤n g jk dx j ⊗ dxk .

Suppose Ω ⊆ M is an Ahlfors regular domain, and abbreviate σg := Hgn−1 ∂Ω. In

this geometric setting, associate with every Lgn -measurable function u : Ω → R its
Pg -maximal function
 ∫ 
1
(Pg u)(x) := sup   |u| dLg ,
n
x ∈ ∂Ω, (10.3.2)
r >0 σg ∂Ω ∩ Bg (x, r) Ω∩Bg (x,r)

where Bg (x, r) denotes the geodesic ball of radius r centered at x. For example, the
manifold version of Corollary 10.1.8 reads as follows.

Theorem 10.3.1 Let the Riemannian manifold (M, g) be as in (10.3.1) and suppose
D : E → F is a first-order differential operator, acting between the sections of
two Hermitian vector bundles E, F → M, whose top coefficients are of class C 1
and the lower order coefficients are continuous. Having fixed an Ahlfors regular
domain Ω ⊆ M (in particular, Ω is a set of locally finite perimeter), abbreviate
σg := Hgn−1 ∂Ω and denote by νg its geometric measure theoretic outward unit
normal (which is defined almost everywhere on ∂Ω with respect to σg ). In this
context, consider two sections

F ∈ C 1 (Ω, E) and G ∈ C 1 (Ω, F ) (10.3.3)

680 10 Strong and Weak Normal Boundary Traces of Vector Fields . . .

with the property that for some κ > 0 and p, p ∈ [1, ∞] with 1/p + 1/p = 1 one has

Nκ F ∈ L p (∂Ω, σg ), Nκ G ∈ L p (∂Ω, σg ),
κ−n.t. κ−n.t.
F ∂Ω , G∂Ω exist σg -a.e. on ∂Ω, and (10.3.4)
 
Pg DF, G F − F, D G E ∈ L 1 (∂Ω, σg ),

where D : F → E denotes the (real) transpose of the differential operator D.

Then   κ−n.t.  κ−n.t. 
Sym(D; νg ) F ∂Ω , G∂Ω ∈ H 1 (∂Ω, σg ) (10.3.5)
F

and there exists a constant C ∈ (0, ∞) such that

  κ−n.t.  κ−n.t.  

 Sym(D; νg ) F ∂Ω , G∂Ω  1 (10.3.6)
F H (∂Ω,σg )
   
≤ C Nκ F  L p (∂Ω,σg ) Nκ G L p (∂Ω,σg )
  
+ C Pg DF, G F − F, D G E  L 1 (∂Ω,σg ) .

As a consequence, the conclusions in (10.3.5)-(10.3.6) hold if in place of the last

condition in (10.3.4) one simply assumes that DF = 0 and D G = 0 in Ω (in which
scenario the last norm in (10.3.6) may be dropped).

Proof All claims are seen by routinely adapting to the manifold setting the real
variable argument from the proof of [133, Corollary 1.11.5], going back all the way
to the proof of Theorem 10.1.4, with [133, Corollary 1.11.5] now substituting [133,
Theorem 1.2.1]. 
As an example, suppose (M, g) is a Riemannian manifold as in (10.3.1) and
Ω ⊆ M is an Ahlfors regular domain. Denote by νg the geometric measure theoretic
outward unit normal to Ω and abbreviate σg := Hgn−1 ∂Ω. Having fixed some degree
 ∈ {0, 1, . . . , n}, pick two differential forms
−1 +2
u ∈ C 1 (Ω, Λ T M) and w ∈ C 1 (Ω, Λ T M) (10.3.7)

such that 
Nκ (du) ∈ L p (∂Ω, σg ), Nκ (δw) ∈ L p (∂Ω, σg )
κ−n.t. κ−n.t. (10.3.8)
and (du) ∂Ω
, (δw) ∂Ω
exist σg -a.e. on ∂Ω,
for some κ > 0 and p, p ∈ [1, ∞] with 1/p + 1/p = 1. Then Theorem 10.3.1 gives
that, in a quantitative sense,
  κ−n.t.  κ−n.t. 
νg ∧ (du)∂Ω , (δw)∂Ω +1
∈ H 1 (∂Ω, σg ). (10.3.9)
Λ TM

In closing, we wish to reiterate that both Corollary 10.2.25 and Corollary 10.2.26
have natural versions in the context of differential operators acting between vector
bundles on manifolds.
Chapter 11
Sobolev Spaces on the Geometric Measure
Theoretic Boundary of Sets of Locally Finite
Perimeter

While there is an ample literature on Sobolev spaces on open sets (see, e.g., [4], [3],
[49], [125], [130], [166], [183] and the references cited therein), here the goal is to
introduce a scale of Sobolev spaces on the geometric measure theoretic boundaries
of sets of locally finite perimeter in the Euclidean setting and on Riemannian man-
ifolds. This builds and expands on the work in [97], [139], and [141]. Our brand
of “boundary” Sobolev spaces are analytic and geometric in nature, in the sense
that they are defined using “weak derivatives” (and integration by parts along the
boundary) which, in turn, are manufactured using the scalar components of the geo-
metric measure theoretic outward unit normal to the set of locally finite perimeter in
question. Corresponding to a domain with flat boundary (i.e., a half-space) our def-
inition reduces precisely to the classical definition of Sobolev spaces (of order one)
in the Euclidean setting, based on ordinary weak derivatives. Such a compatibility
reinforces the idea that this is indeed a natural generalization of the standard scale
of Sobolev spaces from the (entire) Euclidean ambient to sets exhibiting a much
more intricate geometry (both in the Euclidean setting and that of manifolds). In
contrast to other types of Sobolev spaces on generic measure metric spaces which
have been introduced and studied elsewhere in the literature (see, e.g., [79], [81]
and the references there) our brand of Sobolev spaces allows for integration by parts
(on the boundary). Such a feature is critical in light of the applications to singular
integral operators and boundary value problems we have in mind.
The list of topics covered in this chapter includes: weak tangential derivatives
and the very definition of our boundary Sobolev spaces (in §11.1), the definition of
distributional tangential derivatives and the manner in which they compare to weak
tangential derivatives (in §11.2 and §11.3), the tangential gradient (in §11.4), Sobolev
spaces on boundaries of Ahlfors regular domains satisfying a two-sided local John
condition (in §11.5), boundary Sobolev spaces on Riemannian manifolds (in §11.6),
a general recipe for concocting Sobolev-like spaces (in §11.7), boundary Sobolev
spaces with a negative order of smoothness (in §11.8), principal-value distributions
on countably rectifiable sets (in §11.9), Hardy-based boundary Sobolev spaces (in
§11.10), embedding results involving boundary Sobolev spaces (in §11.11), real

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 681
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1_11
682 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

interpolation for boundary Sobolev spaces (in §11.12), Morrey-based (and block-
based) homogeneous Sobolev spaces (in §11.13).

11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces

To fix notation, let Ω ⊆ Rn be a set of locally finite perimeter, and abbreviate

σ∗ := H n−1 ∂∗ Ω. In particular, the geometric measure theoretic outward unit normal
ν = (ν1, . . . , νn ) to Ω is defined σ∗ -a.e. on ∂∗ Ω. Also, from [133, (5.2.6)] and [133,
Lemma 3.6.4] we conclude that
σ∗ is a complete, locally finite (hence also sigma-finite), sep-
arable, Borel-regular measure on ∂∗ Ω, where the latter set is
endowed with the topology canonically inherited from Rn , and (11.1.1)
for each integrability exponent p ∈ (0, ∞) the Lebesgue space
L p (∂∗ Ω, σ∗ ) is separable.
In addition, (11.1.1), the Riesz Duality Theorem (cf., e.g., [55, Theorem 6.15,
p. 190]), and the Sequential Banach-Alaoglu Theorem (cf. [133, (3.6.22)]) ensure
that
 
 p p ∈ [1, ∞)
whenever  ∗ and pp ∈ (1, ∞] satisfy 1/p + 1/p = 1 it follows
that L (∂∗ Ω, σ∗ ) = L (∂∗ Ω, σ∗ ) (via an isometric isomorphism),
 (11.1.2)
and any bounded sequence in L p (∂∗ Ω, σ∗ ) has a sub-sequence weak-∗

convergent to a function in L p (∂∗ Ω, σ∗ ).
p
Recall that for each p ∈ (0, ∞] we may turn Lloc (∂∗ Ω, σ∗ ) into a topological
vector space1 by defining a topology which has as open sets all subsets O of
p
Lloc (∂∗ Ω, σ∗ ) with the property that for each given f ∈ O there exist some number
p
ε > 0 and finite set J ⊆ N such that any function g ∈ Lloc (∂∗ Ω, σ∗ ) satisfying
max j ∈J  f − g L p (B(0, j)∩∂∗ Ω,σ∗ ) < ε necessarily belongs to O. In relation to this,
[133, Proposition 3.1.1] shows that

for each integrability exponent p ∈ (0, ∞], the topological vec-

p
tor space Lloc (∂∗ Ω, σ∗ ) is metrizable, via a translation invariant (11.1.3)
metric which renders it a complete metric space.

To proceed, given any open set O ⊆ Rn , for each j, k ∈ {1, . . . , n} consider

the first-order tangential derivative operator ∂τ j k acting on an arbitrary function
ϕ ∈ C 1 (O) according to
   
∂τ j k ϕ := ν j ∂k ϕ  O∩∂∗ Ω − νk ∂j ϕ  O∩∂∗ Ω
(11.1.4)
at σ∗ -a.e. point on O ∩ ∂∗ Ω.

1 which is actually locally convex if p ∈ [1, ∞]

11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 683

In particular, (11.1.4) implies that any function ϕ ∈ C 1 (O) satisfies

∂τk j ϕ = −∂τ j k ϕ at σ∗ -a.e. point on O ∩ ∂∗ Ω, (11.1.5)

and the following Leibniz product rule holds:

∂τ j k (ϕψ) = (∂τ j k ϕ)ψ + ϕ(∂τ j k ψ) at σ∗ -a.e. point on O ∩ ∂∗ Ω,

(11.1.6)
for every ϕ, ψ ∈ C 1 (O) and every j, k ∈ {1, . . . , n}.

Lemma 11.1.1 Assume Ω ⊆ Rn is an arbitrary set of locally finite perimeter, and

abbreviate σ∗ := H n−1 ∂∗ Ω. Also, consider an open set O ⊆ Rn and pick two
scalar-valued functions, ϕ ∈ C 1 (O) and ψ ∈ Cc1 (O). Then for every pair of indices
j, k ∈ {1, . . . , n} one has
∫ ∫
(∂τ j k ϕ)ψ dσ∗ = − ϕ(∂τ j k ψ) dσ∗ (11.1.7)
∂∗ Ω ∂∗ Ω

and 
if ϕ∂∗ Ω∩O = 0 at σ∗ -a.e. point on ∂∗ Ω ∩ O then
(11.1.8)
∂τ j k ϕ = 0 at σ∗ -a.e. point on the set ∂∗ Ω ∩ O.

Proof Having fixed j, k ∈ {1, . . . , n} arbitrary, define the vector field

F := (∂k ϕ)ψe j − (∂j ϕ)ψek + ϕ(∂k ψ)e j − ϕ(∂j ψ)ek

= ∂k (ϕψ)e j − ∂j (ϕψ)ek . (11.1.9)

Then
 n
F ∈ Cc0 (Rn ) satisfies divF = 0 in D (Rn ), and
(11.1.10)
ν · F = (∂τ j k ϕ)ψ + ϕ(∂τ j k ψ) at σ∗ -a.e. point on ∂∗ Ω.

With these in hand, identity (11.1.7) now follows from the version of the Gauss-Green
formula recorded in [133, (2.8.1)]. 
To justify the claim in (11.1.8), assume ϕ∂∗ Ω∩O = 0 at σ∗ -a.e. point on ∂∗ Ω ∩ O.
Then (11.1.7) gives
∫ ∫
(∂τ j k ϕ)ψ dσ∗ = − ϕ(∂τ j k ψ) dσ∗
∂∗ Ω ∂∗ Ω

= 0 for each ψ ∈ Cc1 (O). (11.1.11)

In concert with the equivalence [133, (3.7.9)] (which holds in the current setting
thanks to (11.1.1)), this proves ∂τ j k ϕ = 0 at σ∗ -a.e. point on ∂∗ Ω ∩ O, finishing the
proof of (11.1.8). 

In turn, Lemma 11.1.1 suggests making the following definition.

684 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Definition 11.1.2 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω.
(i) Given a function f ∈ Lloc 1 (∂ Ω, σ ) along with two indices j, k ∈ {1, . . . , n},
∗ ∗
1 (∂ Ω, σ ) if there exists a
say that ∂τ j k f exists in (or, belongs to) the space Lloc ∗ ∗
function f jk ∈ Lloc1 (∂ Ω, σ ) such that
∗ ∗
∫ ∫
(∂τ j k ϕ) f dσ∗ = − ϕ f jk dσ∗ for all ϕ ∈ Cc1 (Rn ). (11.1.12)
∂∗ Ω ∂∗ Ω

(ii) For each p ∈ [1, ∞] define the (global) L p -based Sobolev space of order one on
∂∗ Ω as

p
L1 (∂∗ Ω, σ∗ ) := f ∈ L p (∂∗ Ω, σ∗ ) : ∂τ j k f exists in L p (∂∗ Ω, σ∗ )

for each j, k ∈ {1, . . . , n} (11.1.13)

where the last membership in (11.1.13) is understood in the above sense (i.e.,
there exists some function f jk ∈ L p (∂∗ Ω, σ∗ ) such that (11.1.12) holds).
More generally, for each p, q ∈ [1, ∞] define the off-diagonal Sobolev space2

p,q
L1 (∂∗ Ω, σ∗ ) := f ∈ L p (∂∗ Ω, σ∗ ) : ∂τ j k f exists in L q (∂∗ Ω, σ∗ )

for each j, k ∈ {1, . . . , n} . (11.1.14)

(iii) For each p, q ∈ [1, ∞] consider the local version of the off-diagonal Sobolev
spaces introduced above, i.e.,

p,q p q
L1,loc (∂∗ Ω, σ∗ ) := f ∈ Lloc (∂∗ Ω, σ∗ ) : ∂τ j k f exists in Lloc (∂∗ Ω, σ∗ )

for each j, k ∈ {1, . . . , n} , (11.1.15)

and abbreviate
p p, p
L1,loc (∂∗ Ω, σ∗ ) := L1,loc (∂∗ Ω, σ∗ ). (11.1.16)

Regarding the above considerations, we wish to note that since (up to a nullset)
the largest environment where the geometric measure theoretic outward unit normal
ν is defined for a given set Ω ⊆ Rn of locally finite perimeter is ∂∗ Ω, and since
the definition of the elementary tangential differential operators ∂τ j k involve the
components ν j , νk of said outward unit normal vector ν (cf. (11.1.4)), it follows
that ∂∗ Ω is, up to a nullset, the largest environment where one may hope to define
boundary Sobolev spaces in the spirit of Definition 11.1.2.
In the context of Definition 11.1.2 we have (11.1.1). Granted this, [133, Propo-
sition 3.7.2] applies and gives that f jk is unambiguously defined by the demand in
p, p p
2 clearly, this is a more inclusive scale since L1 (∂∗ Ω, σ∗ ) = L1 (∂∗ Ω, σ∗ ) for each p ∈ [1, ∞]
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 685

(11.1.12). Henceforth we shall favor the notation

∂τ j k f := f jk (11.1.17)

which, in particular, allows us to recast (11.1.12) more in line with (11.1.7), namely
as
∫ ∫
f (∂τ j k ϕ) dσ∗ = − (∂τ j k f )ϕ dσ∗ for all ϕ ∈ Cc1 (Rn ). (11.1.18)
∂∗ Ω ∂∗ Ω

In analogy with the classical flat, Euclidean case, it is natural to refer to ∂τ j k f as a

weak (tangential) derivative of the function f .
Under the additional assumption that H n−1 (∂Ω \ ∂∗ Ω) = 0, it turns out that we
may systematically replace ∂∗ Ω by ∂Ω and σ∗ by σ := H n−1 ∂Ω in the formulation
of Definition 11.1.2. For this reason, it is natural to adopt the convention that
whenever Ω ⊆ Rn is a set of locally finite perimeter with the addi-
tional property that H n−1 (∂Ω \ ∂∗ Ω) = 0, and if σ∗ := H n−1 ∂∗ Ω,
then for each p ∈ [1, ∞] we shall re-denote the global Sobolev
p p
space L1 (∂∗ Ω, σ∗ ) simply as L1 (∂Ω, σ), and the local Sobolev space (11.1.19)
p p
L1,loc (∂∗ Ω, σ∗ ) simply as L1,loc (∂Ω, σ), where σ := H n−1 ∂Ω; we also
make similar conventions for the off-diagonal versions of these Sobolev
scales.
Our definition of global boundary Sobolev spaces in (11.1.13) is natural from
a multitude of viewpoints. As a first example, consider the special case when Ω
is the open upper half-space R+n . Then ∂∗ Ω = ∂R+n may be canonically identified
with Rn−1 , and under this identification we have σ∗ = H n−1 ∂R+n ≡ L n−1 in
Rn−1 . As such, having fixed some exponent p ∈ [1, ∞], along with some function
f belonging to L p (∂∗ Ω, σ∗ ) ≡ L p (Rn−1, L n−1 ), the demand that ∂τ j k f exists in
L p (Rn−1, L n−1 ), formulated as in (11.1.13) with Ω := R+n , is presently automatically
satisfied whenever j, k ∈ {1, . . . , n − 1} or j = k = n (simply taking f jk = 0 in either
scenario). Next, in the case when j ∈ {1, . . . , n − 1} and k = n, said demand may
currently be re-phrased as asking that the existence of a function f j ∈ L p (Rn−1, L n−1 )
with the property that
∫ ∫
(∂j φ) f dL n−1 = − φ f j dL n−1 for each φ ∈ Cc1 (Rn−1 ). (11.1.20)
R n−1 R n−1

This, of course, is precisely the condition that ∂j f , considered in the sense of

distributions in Rn−1 , actually belongs to L p (Rn−1, L n−1 ). Likewise, when j = n
and k ∈ {1, . . . , n − 1} the demand that ∂τ j k f exists in L p (Rn−1, L n−1 ) reduces
precisely to asking that ∂k f , taken in the sense of distributions in Rn−1 , belongs
to L p (Rn−1, L n−1 ). Collectively, these observations ultimately show that in the case
p
when Ω := R+n it follows that L1 (∂∗ Ω, σ∗ ) may be canonically identified the classical
Sobolev space W (R ).
1, p n−1
686 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

A second example, which is more general in scope, is seen from [141, Proposi-
tion 2.9, p. 33] which implies (with the convention made in (11.1.19) in effect) that
if Ω ⊂ Rn is the open set lying above the graph of a Lipschitz
function φ : Rn−1 → R, then for each given p ∈ (1, ∞) one has (11.1.21)
p
f ∈ L1 (∂Ω, σ) ⇐⇒ f (·, φ(·)) ∈ W 1, p (Rn−1 ), quantitatively.
In concert with a natural localization result discussed later (in item (v) of Propo-
sition 11.1.9), this also implies that, in the case when Ω ⊂ Rn is a bounded Lip-
p
schitz domain, our brand of Sobolev spaces L1 (∂Ω, σ) with p ∈ (1, ∞) agrees
with the standard scale of Sobolev spaces defined on ∂Ω via localization (involv-
ing a smooth, finite, partition of unity) and pull-back via a bi-Lipschitz map to the
((n − 1)-dimensional) Euclidean model.
Our third example sheds light on the relationship between the present brand
of Sobolev spaces and other varieties, defined on abstract abstract measure metric
spaces. Specifically, [97, Theorem 4.27, p. 2720] gives that
if Ω ⊂ Rn is an open set satisfying a two-sided local John condition
and whose boundary is compact and Ahlfors regular, then the brand
p
of Sobolev spaces L1 (∂Ω, σ) with p ∈ (1, ∞), described in (11.1.13)
(11.1.22)
(cf. also (11.1.19)), agrees with the scale of Sobolev spaces defined
generically on ∂Ω, viewed as an abstract measure metric space as in
[79] (see item (v) in Theorem 11.5.18).
Some basic properties of the scale of boundary Sobolev spaces introduced earlier
are collected in the next lemma.

Lemma 11.1.3 Assume Ω ⊆ Rn is an arbitrary set of locally finite perimeter. Ab-

breviate σ∗ := H n−1 ∂∗ Ω and denote by ν = (ν1, . . . , νn ) the geometric measure
theoretic outward unit normal to Ω. Then the following properties hold:

(1) If 1 ≤ r ≤ p ≤ ∞ and 1 ≤ s ≤ q ≤ ∞ then

p,q p,q
L1 (∂∗ Ω, σ∗ ) ⊆ L1,loc (∂∗ Ω, σ∗ )
r,s
⊆ L1,loc (∂∗ Ω, σ∗ ) ⊆ L1,loc
1
(∂∗ Ω, σ∗ ). (11.1.23)

(2) If p, q ∈ [1, ∞], then for each j, k ∈ {1, . . . , n} one has

p,q
∂τ j k f = −∂τk j f at σ∗ -a.e. point on Σ, for every f ∈ L1,loc (∂∗ Ω, σ∗ ).
(11.1.24)
(3) For each p, q ∈ [1, ∞] one has

 p,q
Cc1 (Rn ) → L1 (∂∗ Ω, σ∗ ) (11.1.25)
∂∗ Ω

in the precise sense that

11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 687

if f := ϕ∂∗ Ω with ϕ ∈ Cc1 (Rn ) then f ∈ L1 (∂∗ Ω, σ∗ ) and,
p,q

for every j, k ∈ {1, . . . , n}, at σ∗ -a.e. point on ∂∗ Ω (11.1.26)

   
one has ∂τ j k f = ν j ∂k ϕ ∂∗ Ω − νk ∂j ϕ ∂∗ Ω .

(4) For each p, q ∈ [1, ∞] one has


 p,q
C 1 (Rn ) → L1,loc (∂∗ Ω, σ∗ ) (11.1.27)
∂∗ Ω

with a similar precise interpretation as above.

(5) Fix two integrability exponents p, q ∈ [1, ∞]. Then any constant function f on
p,q
∂∗ Ω belongs to L1,loc (∂∗ Ω, σ∗ ) and ∂τ j k f = 0 for every j, k ∈ {1, . . . , n}.

Proof The inclusions in item (1) are clear from definitions. To deal with the claim
p,q
made in item (2) fix j, k ∈ {1, . . . , n} arbitrary and pick some f ∈ L1,loc (∂∗ Ω, σ∗ )
with p, q ∈ [1, ∞]. From (11.1.5) and (11.1.18) (used twice), it follows that for each
function ϕ ∈ Cc1 (Rn ) we have
∫ ∫
(∂τ j k f )ϕ dσ∗ = − f (∂τ j k ϕ) dσ∗
∂∗ Ω ∂∗ Ω
∫ ∫
= f (∂τk j ϕ) dσ∗ = − (∂τ j k f )ϕ dσ∗ . (11.1.28)
∂∗ Ω ∂∗ Ω

In view of the arbitrariness of the function ϕ in Cc1 (Rn ), [133, Proposition 3.7.2]
(whose applicability in the present setting is ensured by (11.1.1)) then forces
∂τ j k f = −∂τk j f at σ∗ -a.e. point on ∂∗ Ω, finishing the proof of (11.1.24). Next,
the claims in (11.1.26) and (11.1.27) are direct consequences of Lemma 11.1.1 and
Definition 11.1.2.
As far as item (5) is concerned, suppose f = c on ∂∗ Ω, for some constant c ∈ C.
Pick j, k ∈ {1, . . . , n} along with an arbitrary test function ϕ ∈ Cc1 (Rn ). We need to
check that (11.1.12) holds for the choice f jk = 0, i.e., show that
∫
(∂τ j k ϕ) f dσ∗ = 0. (11.1.29)
∂∗ Ω

Mollify ϕ to obtain a sequence {ϕ } ∈N ⊆ Cc∞ (Rn ) with the property that ϕ → ϕ
and ∇ϕ → ∇ϕ uniformly in Rn as → ∞. For each ∈ N define the vector field
 n
F := (∂k ϕ )e j − (∂j ϕ )ek ∈ Cc∞ (Rn ) (11.1.30)

and write
688 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ ∫
(∂τ j k ϕ) f dσ∗ = c · lim ∂τ j k ϕ dσ∗
∂∗ Ω →∞ ∂∗ Ω
∫

= c · lim ν · F ∂∗ Ω dσ∗ = 0, (11.1.31)
→∞ ∂∗ Ω

with the last equality provided by the De Giorgi-Federer Divergence Theorem

recorded in [133, Theorem 1.1.1] (bearing in mind that, as seen from (11.1.30),
each F is divergence-free). This establishes (11.1.29), and finishes the proof of the
lemma. 

There is also a local version of (11.1.26) of the following sort.

Lemma 11.1.4 Suppose Ω ⊆ Rn is a set of locally finite perimeter. Abbreviate

σ∗ := H n−1 ∂∗ Ω and denote by ν = (ν1, . . . , νn ) the geometric measure theoretic
outward unit normal to Ω. Fix a function f ∈ L1,loc1 (∂ Ω, σ ) and assume O ⊆ Rn is
∗ ∗
an open set with the property that
 
f  O∩∂∗ Ω = ϕ O∩∂∗ Ω for some ϕ ∈ C 1 (Rn ). (11.1.32)

Then for every j, k ∈ {1, . . . , n} one has

    
(∂τ j k f ) O∩∂∗ Ω = ν j ∂k ϕ  O∩∂∗ Ω − νk ∂j ϕ  O∩∂∗ Ω
(11.1.33)
at σ∗ -a.e. point in O ∩ ∂∗ Ω.

Proof Fix j, k ∈ {1, . . . , n} and choose an arbitrary function ψ ∈ Cc1 (O). Denote by
ψ the extension of ψ by zero outside its support to the entire Rn . Then ψ ∈ Cc1 (Rn )
and (11.1.18), (11.1.32), (11.1.7), (11.1.4) permit us to compute
∫ ∫ ∫
(∂τ j k f )ψ dσ∗ = (∂τ j k f )ψ dσ∗ = − f (∂τ j k ψ) dσ∗
O∩∂∗ Ω ∂∗ Ω ∂∗ Ω
∫ ∫
=− f (∂τ j k ψ) dσ∗ = − ϕ(∂τ j k ψ) dσ∗
O∩∂∗ Ω O∩∂∗ Ω
∫ ∫
=− ϕ(∂τ j k ψ) dσ∗ = (∂τ j k ϕ)ψ dσ∗
∂∗ Ω ∂∗ Ω
∫     
=− ν j ∂k ϕ  O∩∂∗ Ω − νk ∂j ϕ  O∩∂∗ Ω ψ dσ∗ .
O∩∂∗ Ω
(11.1.34)

In view of the arbitrariness of the function ψ in Cc1 (O), [133, Corollary 3.7.3] (whose
applicability in the present setting is ensured by (11.1.1)) then yields (11.1.33). 

Further examples of Sobolev functions are offered by the lemma below.

11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 689

Lemma 11.1.5 Let Ω ⊆ Rn be a set of locally finite perimeter. Next denote by

ν = (ν1, . . . , νn ) the geometric measure theoretic outward unit normal to Ω and
abbreviate σ∗ := H n−1 ∂∗ Ω. Also, pick an open neighborhood O of ∂Ω and fix two
integrability exponents p, q ∈ [1, ∞]. In this setting, consider a function

ψ ∈ C 1 (O) such that ψ ∂∗ Ω ∈ L p (∂∗ Ω, σ∗ )
  n (11.1.35)
and (∇ψ)∂∗ Ω ∈ L q (∂∗ Ω, σ∗ ) .

Then the function f := ψ ∂∗ Ω belongs to the off-diagonal boundary Sobolev space
p,q
L1 (∂∗ Ω, σ∗ ) and for any two indices j, k ∈ {1, . . . , n} one has
   
∂τ j k f = ν j ∂k ψ ∂∗ Ω − νk ∂j ψ ∂∗ Ω at σ∗ -a.e. point on ∂∗ Ω. (11.1.36)

Proof The claim in the statement is a consequence of Definition 11.1.2 and

Lemma 11.1.1 (keeping (11.1.4) in mind). 

The next lemma establishes the locality of the tangential differential operators
∂τ j k , which is remarkable since they lack a pointwise definition.

Lemma 11.1.6 Let Ω ⊆ Rn be a set of locally finite perimeter, and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, fix two arbitrary indices j, k ∈ {1, . . . , n}. Then the operator
∂τ j k is local in the sense that for every f ∈ L1,loc
1 (∂ Ω, σ ) one has
∗ ∗

supp (∂τ j k f ) ⊆ supp f . (11.1.37)

As a consequence, for every f ∈ L1,loc

1 (∂ Ω, σ ) one has
∗ ∗

∂τ j k f = 0 at σ∗ -a.e. point in ∂∗ Ω \ supp f . (11.1.38)

Moreover, if the function f ∈ L1,loc

1 (∂ Ω, σ ) is locally constant on U ⊆ ∂ Ω (in
∗ ∗ ∗


the sense that for each x ∈ U there exists r > 0 with the property that f B(x,r)∩∂∗ Ω is
constant) then
∂τ j k f = 0 at σ∗ -a.e. point in U. (11.1.39)
1 (∂ Ω, σ ) be arbitrary. From item (1) in [133, Lemma 3.8.4]
Proof Let f ∈ L1,loc ∗ ∗
we know that supp f is a relatively closed subset of ∂∗ Ω, so there exists an open set
O ⊆ Rn such that
∂∗ Ω \ supp f = O ∩ ∂∗ Ω. (11.1.40)
Since ∂∗ Ω is second-countable when equipped with the topology inherited from
the ambient Euclidean space (given that being second-countable is a hereditary
property), from (11.1.40) and item (9) in [133, Lemma 3.8.4] we conclude that

f  O∩∂∗ Ω = 0 at σ∗ -a.e. point in O ∩ ∂∗ Ω. (11.1.41)

Granted this, from (11.1.32)-(11.1.33) (used with ϕ = 0) we deduce that

690 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

(∂τ j k f ) O∩∂∗ Ω = 0 at σ∗ -a.e. point in O ∩ ∂∗ Ω. (11.1.42)

With this in hand, item (5) in [133, Lemma 3.8.4] then implies that

O ∩ ∂∗ Ω ⊆ ∂∗ Ω \ supp (∂τ j k f ). (11.1.43)

Together with (11.1.40), this proves that ∂∗ Ω \ supp f ⊆ ∂∗ Ω \ supp (∂τ j k f ), which
ultimately establishes the inclusion in (11.1.37).
As regards (11.1.38), observe that O := ∂∗ Ω \ supp f is a relatively open subset of
∂∗ Ω which is disjoint from supp (∂τ j k f ) (as seen from item (1) in [133, Lemma 3.8.4]
and (11.1.37)). Having noticed this, from item (9) in [133, Lemma 3.8.4] we may
then conclude that ∂τ j k f vanishes σ∗ -a.e. on O, finishing the proof of (11.1.38).
Turning our attention to the last claim in the statement, assume f ∈ L1,loc
1 (∂ Ω, σ )
∗ ∗
is locally constant on U ⊆ ∂∗ Ω. Then for each x ∈ U there exists rx > 0 and a constant
cx such that 
f B(x,rx )∩∂∗ Ω = cx on B(x, rx ) ∩ ∂∗ Ω. (11.1.44)
With this in hand, Lemma 11.1.4 (applied with O := B(x, rx ) and ϕ ≡ cx ) forces

∂τ j k f = 0 at σ∗ -a.e. point in B(x, rx ) ∩ ∂∗ Ω. (11.1.45)

Use the fact that Rn is strongly Lindelöf to find a sequence of points {xi }i ∈N ⊆ U
with the property that
 
U⊆ B(x, rx ) = B xi, rxi . (11.1.46)
x ∈U i ∈N
 
From (11.1.45) used for each B xi, rxi we conclude that
  
∂τ j k f = 0 at σ∗ -a.e. point in B xi, rxi ∩ ∂∗ Ω. (11.1.47)
i ∈N

At this stage, the claim in (11.1.39) is clear from (11.1.47) and (11.1.46). 

A version of the integration by parts formula on the boundary recorded in (11.1.18)

also holds if the compact support condition for the intervening test functions is
replaced by suitable integrability conditions.

Lemma 11.1.7 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Pick some f ∈ L1,loc
1 (∂ Ω, σ ), fix two indices j, k ∈ {1, . . . , n},
∗ ∗
and consider a function ϕ satisfying

ϕ ∈ C 1 (U) for some open neighborhood U of ∂Ω,

∫ ∫
 
| f ||∇ϕ| dσ∗ < +∞, ∂τ f  |ϕ| dσ∗ < +∞,
jk (11.1.48)
∂∗ Ω ∫ ∂∗ Ω

as well as | f (x)||ϕ(x)|(1 + |x|)−1 dσ∗ (x) < +∞.

∂∗ Ω
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 691

Then ∫ ∫
f (∂τ j k ϕ) dσ∗ = − (∂τ j k f )ϕ dσ∗ . (11.1.49)
∂∗ Ω ∂∗ Ω
p,q
As a corollary, if f ∈ L1 (∂∗ Ω, σ∗ ) with p, q ∈ [1, ∞] and ϕ satisfies
   n
ϕ ∈ C 1 (U) for some neighborhood U of ∂Ω, (∇ϕ)∂∗ Ω ∈ L p (∂∗ Ω, σ∗ ) ,
 
ϕ∂∗ Ω ∈ L q (∂∗ Ω, σ∗ ) and (1 + |x|)−1 · ϕ∂∗ Ω ∈ L p (∂∗ Ω, σ∗ ),
 

(11.1.50)
where p, q  ∈ [1, ∞] are the Hölder conjugate exponents of p, q, then the integration
by parts formula (11.1.49) holds for each j, k ∈ {1, . . . , n}. In particular, if ∂∗ Ω is
bounded then (11.1.49) holds for every f ∈ L11 (∂∗ Ω, σ∗ ) and every ϕ ∈ C 1 (U) for
some neighborhood U of ∂Ω.
reg
Proof Recall that δ∂Ω denotes the regularized distance to ∂Ω (cf. [133, Proposi-
tion 6.1.1]). With C0, C1 denoting the constants appearing in [133, (6.1.2)] with
F := ∂Ω, pick some N > C1 /C0 then consider a function ψ ∈ C ∞ (R), 0 ≤ ψ ≤ 1,
with ψ ≡ 1 on (−∞, C1 /N) and ψ ≡ 0 on (C0, ∞). For each ε > 0 define
 
Uε := x ∈ Rn : dist(x, ∂Ω) < ε (11.1.51)
reg
δ ∂Ω (x)
and introduce Ψε (x) := ψ ε for all x ∈ Rn . Then for each ε > 0 we have

Ψε ∈ C ∞ (Rn ), supp Ψε ⊆ Uε, 0 ≤ Ψε ≤ 1,

Ψε ≡ 1 on Uε/N , and
 
for each α ∈ N0n there is Cα ∈ (0, ∞) so that sup (∂ α Ψε )(x) ≤ Cα ε − |α | .
x ∈R n
(11.1.52)
 
Let us also select a function θ ∈ Cc∞ B(0, 2) with the property that θ ≡ 1 on B(0, 1)
and, for each R > 0, define θ R (x) := θ(x/R) for every x ∈ Rn .
Consider now a function ϕ as in (11.1.48). For each fixed R > 0 there exists
ε = ε(R) > 0 with the property that Uε ∩ supp θ R ⊆ U ∩ supp θ R . Then ϕ · Ψε/2 · θ R
belongs to C 1 (U) and is supported in a compact subset of U. If we denote by ϕR its
extension by zero outside its support to the entire Rn , then (11.1.52) and Leibniz’
product rule recorded in (11.1.6) ensure that for each R > 0 we have
 
ϕR ∈ Cc1 (Rn ), ϕR ∂∗ Ω = θ R ϕ∂∗ Ω, and
      (11.1.53)
∂τ j k ϕR = θ R ∂∗ Ω ∂τ j k ϕ + ϕ∂∗ Ω ∂τ j k θ R .

For each R > 0 fixed, we may then use (11.1.53), the assumptions on f , and (11.1.18),
to write
692 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ ∫ ∫
f (∂τ j k ϕ)θ R dσ∗ = f ∂τ j k ϕR dσ∗ − f ϕ ∂τ j k θ R dσ∗
∂∗ Ω ∂∗ Ω ∂∗ Ω
∫ ∫
=− (∂τ j k f )ϕR dσ∗ − f ϕ ∂τ j k θ R dσ∗
∂∗ Ω ∂∗ Ω
∫ ∫
=− (∂τ j k f )ϕθ R dσ∗ − f ϕ ∂τ j k θ R dσ∗ . (11.1.54)
∂∗ Ω ∂∗ Ω

Passing to the limit as R → ∞ in the most extreme sides of (11.1.54) then yields
∫ ∫
f ∂τ j k ϕ dσ∗ = lim f (∂τ j k ϕ)θ R dσ∗
∂∗ Ω R→∞ ∂∗ Ω
∫ ∫
= − lim (∂τ j k f )ϕθ R dσ∗ − lim f ϕ ∂τ j k θ R dσ∗
R→∞ ∂∗ Ω R→∞ ∂∗ Ω
∫
=− (∂τ j k f )ϕ dσ∗ . (11.1.55)
∂∗ Ω

Above, the first equality is a consequence∫ of Lebesgue’s Dominated

∫ Convergence
Theorem (taking into account that ∂ Ω | f ||∂τ j k ϕ| dσ∗ ≤ 2 ∂ Ω | f ||∇ϕ| dσ∗ < +∞,
∗ ∗
as seen from (11.1.4) and (11.1.48)). The second equality comes from (11.1.54). The
third equality uses two observations. First, since for each R ≥ 1 we have the σ∗ -a.e.
pointwise estimate
    
 f ϕ ∂τ θ R  ≤ C| f |(1 + |x|)−1 · ϕ  · 1[B(0,2R)\B(0,R)]∩∂ Ω on ∂∗ Ω, (11.1.56)
jk ∂∗ Ω ∗

∫   
and since we are assuming ∂ Ω | f |(1 + |x|)−1 · ϕ∂∗ Ω  dσ∗ < +∞ (cf. (11.1.48)),
∗
Lebesgue’s Dominated Convergence Theorem gives
∫
lim f ϕ ∂τ j k θ R dσ∗ = 0. (11.1.57)
R→∞ ∂∗ Ω

Second, Lebesgue’s
∫  Dominated  Convergence Theorem also gives (this time, keeping
in mind that ∂ Ω ∂τ j k f  |ϕ| dσ∗ < +∞; cf. (11.1.48)) that
∗

∫ ∫
lim (∂τ j k f )ϕθ R dσ∗ = (∂τ j k f )ϕ dσ∗ . (11.1.58)
R→∞ ∂∗ Ω ∂∗ Ω

This finishes the proof of (11.1.55). Hence, (11.1.49) is established, and the remain
claims in the statement are directly implied by it. 

Two useful consequences of Lemma 11.1.7 are as follows. First, given a set of
locally finite perimeter Ω ⊆ Rn and with σ∗ := H n−1 ∂∗ Ω, specializing (11.1.49)-
(11.1.50) to the case when p = q = 1 and ϕ ≡ 1 yields
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 693
∫
∂τ j k f dσ∗ = 0 for each f ∈ L11 (∂∗ Ω, σ∗ ) and j, k ∈ {1, . . . , n}. (11.1.59)
∂∗ Ω

A second useful consequence of Lemma 11.1.7 is recorded in the corollary below.

Corollary 11.1.8 Let Ω ⊆ Rn , where n ≥ 3, be a Lebesgue measurable set with the

property that ∂∗ Ω is an upper Ahlfors regular set. In particular (cf. [133, (5.9.16)]),
the measure σ∗ := H n−1 ∂∗ Ω is locally finite and Ω is a set of locally finite perimeter.
In this context, pick some
p,q
f ∈ L1 (∂∗ Ω, σ∗ ) with p ∈ [1, ∞) and q ∈ [1, n − 1), (11.1.60)

and consider a function ϕ satisfying, for some C ∈ (0, ∞) and some open neighbor-
hood U of ∂Ω,
 
ϕ ∈ C 1 (U), (∇ϕ)(x) ≤ C(1 + |x|)1−n for each x ∈ U
(11.1.61)
and |ϕ(x)| ≤ C(1 + |x|)2−n for each x ∈ U.

Then for each j, k ∈ {1, . . . , n} one has the boundary integration by parts formula
∫ ∫
f (∂τ j k ϕ) dσ∗ = − (∂τ j k f )ϕ dσ∗ . (11.1.62)
∂∗ Ω ∂∗ Ω

Proof This is an immediate consequence of Lemma 11.1.7 and [133, Lemma 7.2.1].

Other basic properties of the brand of boundary Sobolev scale introduced earlier
are summarized in the next proposition.

Proposition 11.1.9 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, fix two integrability exponents p, q ∈ [1, ∞]. Then the
following statements are true.
p,q
(i) The space L1 (∂∗ Ω, σ∗ ) is complete, hence a Banach space, when equipped
with the norm
p,q
L1 (∂∗ Ω, σ∗ )  f →  f  L p, q (∂∗ Ω,σ∗ ) :=  f  L p (∂∗ Ω,σ∗ ) (11.1.63)
1

+ ∂τ j k f  L q (∂∗ Ω,σ∗ ) .
1≤ j,k ≤n

p
In particular, L1 (∂∗ Ω, σ∗ ) is a Banach space when equipped with the norm
p
L1 (∂∗ Ω, σ∗ )  f →  f  L p (∂∗ Ω,σ∗ ) :=  f  L p (∂∗ Ω,σ∗ ) (11.1.64)
1

+ ∂τ j k f  L p (∂∗ Ω,σ∗ ) .
1≤ j,k ≤n
694 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

(ii) The following holds:

p,q
the space L1 (∂∗ Ω, σ∗ ) is reflexive whenever p, q ∈ (1, ∞); in
p (11.1.65)
particular L1 (∂∗ Ω, σ∗ ) is also reflexive if p ∈ (1, ∞).

(iii) One has

p,q
L1 (∂∗ Ω, σ∗ ) → L p (∂∗ Ω, σ∗ ) continuously, as well as densely if p < ∞.
(11.1.66)
(iv) For each j, k ∈ {1, . . . , n}, the tangential differential operator
p,q
∂τ j k : L1 (∂∗ Ω, σ∗ ) −→ L q (∂∗ Ω, σ∗ ) is linear and bounded. (11.1.67)
p,q
(v) Whenever 1 ≤ q ≤ p ≤ ∞ it follows that the space L1 (∂∗ Ω, σ∗ ) is a module
 
over ϕ|∂∗ Ω : ϕ ∈ Cc1 (Rn ) and a natural Leibniz rule holds. More specifically,
if 1 ≤ q ≤ p ≤ ∞, then for each j, k ∈ {1, . . . , n} one has
p,q
ϕ f ∈ L1 (∂∗ Ω, σ∗ ) and ∂τ j k (ϕ f ) = (∂τ j k ϕ) f + ϕ(∂τ j k f ),
p,q
(11.1.68)
for every ϕ ∈ Cc1 (Rn ) and every f ∈ L1 (∂∗ Ω, σ∗ ).
p,q
Proof The right-hand side of (11.1.63) makes L1 (∂∗ Ω, σ∗ ) a normed linear
space which embeds continuously into L p (∂∗ Ω, σ∗ ). To check completeness, suppose
p,q
{ f } ∈N is a Cauchy sequence in L1 (∂∗ Ω, σ∗ ) equipped with the norm (11.1.63),
and fix j, k ∈ {1, . . . , n} arbitrary. Then there exist functions f ∈ L p (∂∗ Ω, σ∗ ) and
f jk ∈ L q (∂∗ Ω, σ∗ ) such that

f → f in L p (∂∗ Ω, σ∗ ) and ∂τ j k f → f jk
(11.1.69)
in L q (∂∗ Ω, σ∗ ), as → ∞.

Since for each ϕ ∈ Cc1 (Rn ) we have

∫ ∫
(∂τ j k ϕ) f dσ∗ = − ϕ ∂τ j k f dσ∗, (11.1.70)
∂∗ Ω ∂∗ Ω

upon letting → ∞ we arrive at the conclusion that

∫ ∫
(∂τ j k ϕ) f dσ∗ = − ϕ f jk dσ∗ for each ϕ ∈ Cc1 (Rn ). (11.1.71)
∂∗ Ω ∂∗ Ω

In turn, (11.1.71) proves that

p,q
f ∈ L1 (∂∗ Ω, σ∗ ) and ∂τ j k f = f jk for each j, k ∈ {1, . . . , n}. (11.1.72)

With this in hand, (11.1.69) may be rephrased as saying

p,q
f → f in L1 (∂∗ Ω, σ∗ ) as → ∞. (11.1.73)
p,q
This shows that L1 (∂∗ Ω, σ∗ ) is indeed a Banach space. Going further, the applica-
tion
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 695

p,q   n(n−1)/2
J : L1 (∂∗ Ω, σ∗ ) −→ L p (∂∗ Ω, σ∗ ) ⊕ L q (∂∗ Ω, σ∗ )
  p,q
(11.1.74)
J f := f , (∂τ j k f )1≤ j<k ≤n for each f ∈ L1 (∂∗ Ω, σ∗ ),

has the property that Range J is closed, and J an isomorphism onto its range.
p,q
Consequently, when p, q ∈ (1, ∞), the off-diagonal Sobolev space L1 (∂∗ Ω, σ∗ ) is
isomorphic to a closed subspace of a reflexive Banach space. Standard functional
p,q
analysis then shows that L1 (∂∗ Ω, σ∗ ) is itself a reflexive Banach space, whenever
1 < p, q < ∞. Next, the claims in (11.1.66) are clear from (11.1.13), (11.1.63),
(11.1.25), and [133, Corollary 3.7.3] (cf. (11.1.77)). Also, (11.1.67) is immediate
from definitions.
As regards (11.1.68), for each ϕ, ψ ∈ Cc1 (Rn ), each j, k ∈ {1, . . . , n}, and each
p,q
f ∈ L1 (∂∗ Ω, σ∗ ) with 1 ≤ q ≤ p ≤ ∞, we have
∫ ∫ ∫
(ϕ f )(∂τ j k ψ) dσ∗ = f ∂τ j k (ϕψ) dσ∗ − f (∂τ j k ϕ)ψ dσ∗
∂∗ Ω ∂∗ Ω ∂∗ Ω
∫
 
=− (∂τ j k f )ϕ + f (∂τ j k ϕ) ψ dσ∗, (11.1.75)
∂∗ Ω

thanks to Leibniz’ product rule (11.1.6), used in the first equality. In view of the
arbitrariness of the function ψ ∈ Cc1 (Rn ), and the fact that (∂τ j k f )ϕ + f (∂τ j k ϕ)
p
belongs to L q (∂∗ Ω, σ∗ ) + Lcomp (∂∗ Ω, σ∗ ) ⊆ L q (∂∗ Ω, σ∗ ), given that p ≥ q, all
conclusions in (11.1.68) follow. 

Moving on, in the range p ∈ (1, ∞] it is possible to equivalently characterize

the L p -based Sobolev space of order one from (11.1.13) in the intrinsic manner
described in our next result, for the more general scale of off-diagonal boundary
Sobolev spaces.

Lemma 11.1.10 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Then, for each p ∈ [1, ∞] and q ∈ (1, ∞], if q  ∈ [1, ∞) is such
that 1/q + 1/q  = 1 one has

p,q
L1 (∂∗ Ω, σ∗ ) := f ∈ L p (∂∗ Ω, σ∗ ) : there exists c ∈ (0, ∞) so that for

n ∫
  
   
all ϕ ∈ Cc1 (Rn ) one has    
f (∂τ j k ϕ) dσ∗  ≤ c ϕ|∂∗ Ω L q (∂∗ Ω,σ∗ ) ,

j,k=1 ∂∗ Ω

(11.1.76)

in a quantitative fashion.

Proof For starters observe that [133, Corollary 3.7.3] (whose applicability in the
current setting is ensured by [133, (5.2.6)] and [133, (5.6.35)]) guarantees that
  
ϕ|∂∗ Ω : ϕ ∈ Cc1 (Rn ) is a dense subspace of L q (∂∗ Ω, σ∗ ). (11.1.77)
696 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

To proceed, associated with each f ∈ Lloc

1 (∂ Ω, σ ) and j, k ∈ {1, . . . , n} consider
∗ ∗
jk
the functional Λ f , defined as follows:
∫
jk   jk  
Λf : ϕ|∂∗ Ω : ϕ ∈ Cc1 (Rn ) −→ R, Λ f ϕ|∂∗ Ω := f (∂τ j k ϕ) dσ∗ .
∂∗ Ω
(11.1.78)
jk
That Λ f is unambiguously defined is clear from (11.1.8).
Suppose now that f ∈ L p (∂∗ Ω, σ∗ ) has the property that there exists a constant
c ∈ (0, ∞) for which
n ∫ 
   
 f (∂τ j k ϕ) dσ∗  ≤ cϕ|∂∗ Ω  L q (∂∗ Ω,σ∗ ) for each ϕ ∈ Cc1 (Rn ). (11.1.79)

j,k=1 ∂∗ Ω

jk
From (11.1.77)-(11.1.79) and (11.1.2) we then conclude that each Λ f extends
  ∗
uniquely to a functional in L q (∂∗ Ω, σ∗ ) = L q (∂∗ Ω, σ∗ ). Hence, for each pair
of indices j, k ∈ {1, . . . , n} there exists a function g jk ∈ L q (∂∗ Ω, σ∗ ) with the
property that g jk  L q (∂∗ Ω,σ∗ ) ≤ c and
∫
jk  
Λ f ϕ|∂∗ Ω = g jk ϕ dσ∗ for every ϕ ∈ Cc1 (Rn ). (11.1.80)
∂∗ Ω

In turn, from (11.1.78), (11.1.80), and (11.1.13) we eventually conclude that f

p,q
belongs to L1 (∂∗ Ω, σ∗ ) (as the role of f jk is presently played by g jk ∈ L q (∂∗ Ω, σ∗ )).
This establishes the right-to-left inclusion in (11.1.76). Finally, the left-to-right
inclusion in (11.1.76) is seen from (11.1.18) and Hölder’s inequality, completing the
proof of the lemma. 
For a subset E of a given metric space, recall that Lip (E) denotes the space of
(scalar-valued) Lipschitz functions defined on E, and that  · Lip (E) stands for the
associated semi-norm (cf. (A.0.57)-(A.0.56)).

Proposition 11.1.11 Let Ω ⊆ Rn be a set of locally finite perimeter and define

σ∗ := H n−1 ∂∗ Ω. Then
 
f ∈ Lip (∂∗ Ω) : f bounded ⊆ L1∞ (∂∗ Ω, σ∗ ). (11.1.81)

Moreover, this inclusion is quantitative, in the sense that there exists some purely
dimensional constant Cn ∈ (0, ∞) with the property that for every bounded Lipschitz
function f on ∂∗ Ω one has

n
 
∂τ f  ∞ ≤ Cn  f Lip (∂∗ Ω) . (11.1.82)
jk L (∂∗ Ω,σ∗ )
j,k=1

Later on, in item (xii) of Theorem 11.5.18, we shall see that under stronger
hypotheses, specifically if Ω ⊆ Rn is an open set satisfying a two-sided local John
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 697

condition and whose boundary is compact and Ahlfors regular, we actually have
equality in (11.1.81).
Proof of Proposition 11.1.11 Denote by ν = (ν1, . . . , νn ) the geometric measure
theoretic outward unit normal to Ω. Throughout, fix j, k ∈ {1, . . . , n} arbitrary.
Suppose some bounded function f ∈ Lip (∂∗ Ω) has been given. By first canonically
extending f to a uniformly continuous function on the closure of ∂∗ Ω, and then
invoking [133, Theorem 6.1.3], we may assume that

f = ψ ∂∗ Ω for some bounded function ψ ∈ Lip (Rn ). (11.1.83)

Next, we can use a mollifier to construct a sequence {ψ } ∈N ⊆ C ∞ (Rn ) such that

ψ → ψ uniformly on compact subsets of Rn , and

(11.1.84)
supx ∈Rn |(∇ψ )(x)| ≤ Cn ψLip (Rn ) for each ∈ N.

Introduce 
f := ψ ∂∗ Ω for every ∈ N. (11.1.85)

Then, for each ϕ ∈ Cc1 (Rn ),

∫ ∫
(∂τ j k ϕ) f dσ∗ = lim (∂τ j k ϕ) f dσ∗ (11.1.86)
∂∗ Ω →∞ ∂∗ Ω

while for every ∈ N we have

∫ ∫

(∂τ j k ϕ) f dσ∗ = − ϕ f jk dσ∗ (11.1.87)
∂∗ Ω ∂∗ Ω

with
  

f jk := ∂τ j k f = ∂τ j k ψ ∂∗ Ω
   
= ν j ∂k ψ ∂∗ Ω − νk ∂j ψ ∂∗ Ω at σ∗ -a.e. point on ∂∗ Ω. (11.1.88)

As a result, for each ϕ ∈ Cc1 (Rn ) we have

∫ ∫

(∂τ j k ϕ) f dσ∗ = − lim ϕ f jk dσ∗
∂∗ Ω →∞ ∂∗ Ω
(11.1.89)
and c := sup    ∞
f jk < ∞.
L (∂∗ Ω,σ∗ )
 ∈N

Hence,
∫    
 
 (∂τ j k ϕ) f dσ∗  ≤ cϕ∂∗ Ω  L 1 (∂∗ Ω,σ∗ ), ∀ϕ ∈ Cc1 (Rn ). (11.1.90)
∂∗ Ω
698 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

On account of (11.1.90), Lemma 11.1.10 gives that f ∈ L1∞ (∂∗ Ω, σ∗ ), proving

(11.1.81). Finally, the estimate in (11.1.82) is implicit in what we have done so far.

Proposition 11.1.11 tells us that bounded Lipschitz functions on the geometric

measure theoretic boundary belong to the boundary Sobolev space of essentially
bounded functions with essentially bounded tangential derivatives. While we are
lacking a pointwise formula for the tangential derivatives of bounded Lipschitz
functions, our next lemma points to the fact that, at least in the weak-∗ topology,
these tangential derivatives may be thought of as limits of pointwise tangential
derivatives of suitable sequences of smooth functions.

Lemma 11.1.12 Let Ω ⊆ Rn be a set of finite perimeter and define σ∗ := H n−1 ∂∗ Ω.
Also, fix a function f ∈ Lipc (∂∗ Ω) (so that, in particular, f ∈ L1∞ (∂∗ Ω, σ∗ ) by
(11.1.81)), and consider a sequence

{ϕ } ∈N ⊆ Cc1 (Rn ) with the property that

 
sup ∈N supx ∈∂∗ Ω (∇ϕ )(x) < +∞ and (11.1.91)

ϕ ∂∗ Ω → f in Lloc
1 (∂ Ω, σ ) as
∗ ∗ → ∞.

Then for each pair of indices j, k ∈ {1, . . . , n} one has

∂τ j k ϕ −→ ∂τ j k f weak-∗ in L ∞ (∂∗ Ω, σ∗ ) as → ∞. (11.1.92)

Remark 11.1.13 In the context of Lemma 11.1.12, a sequence with the properties
listed in (11.1.91) always exists. For example, we may extend the given function
f ∈ Lipc (∂∗ Ω) to some Lipschitz
∫ compactly supported function F in Rn , and, having
fixed θ ∈ Cc∞ (Rn ) with Rn θ dL n = 1, define ϕ := F ∗ θ  for each ∈ N where
θ  (x) := n · θ( · x) for each x ∈ Rn . Then the following properties are standard
consequences of this definition:
  
{ϕ } ∈N ⊆ Cc∞ (Rn ), lim supx ∈∂∗ Ω ϕ (x) − f (x) = 0,
→∞
there exists a compact set K ⊂ Rn such that supp ϕ ⊆ K for all ∈ N,
   (11.1.93)
sup ∈N supx ∈Rn (∇ϕ )(x) ≤ C f Lip(∂∗ Ω),

and sup ∈N supx ∈Rn |ϕ (x)| ≤ C supx ∈∂∗ Ω | f (x)|.

In concert with [133, (7.3.25)] and the Mean Value Theorem, the last three properties
above further entail
  1−α α
sup ϕ C α (Rn ) ≤ sup | f (x)| + 21−α sup | f (x)|  f Lip(∂∗ Ω)
(11.1.94)
 ∈N x ∈∂∗ Ω x ∈∂∗ Ω

for each α ∈ (0, 1).

11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 699

Proof of Lemma 11.1.12 Fix some arbitrary j, k ∈ {1, . . . , n}. Since thanks to
(11.1.4) and (11.1.91) we have
 
sup ∂τ j k ϕ  L ∞ (∂∗ Ω,σ∗ ) < +∞, (11.1.95)
 ∈N

we may invoke (11.1.2) (with p := 1) to conclude the existence of some function

g ∈ L ∞ (∂∗ Ω, σ∗ ) such that

∂τ j k ϕ −→ g weak-∗ in L ∞ (∂∗ Ω, σ∗ ) as → ∞. (11.1.96)

Granted this, there remains to show that

g = ∂τ j k f at σ∗ -a.e. point on ∂∗ Ω. (11.1.97)

To this end, pick an arbitrary function ψ ∈ Cc∞ (Rn ) and write

∫ ∫ ∫
ψ ∂τ j k f dσ∗ = − f ∂τ j k ψ dσ∗ = − lim ϕ ∂τ j k ψ dσ∗
∂∗ Ω ∂∗ Ω →∞ ∂∗ Ω
∫ ∫
= lim ψ ∂τ j k ϕ dσ∗ = ψ g dσ∗ . (11.1.98)
→∞ ∂∗ Ω ∂∗ Ω

Here, the first equality is a consequence of (11.1.18) (bearing in mind that, according
to (11.1.81), we have f ∈ L1∞ (∂∗ Ω, σ∗ )), the second equality is implied by the last
condition in (11.1.91), the third equality comes from (11.1.7), and the last equality
is guaranteed by (11.1.96).
Having proved (11.1.98), the claim in (11.1.97) now follows by invoking [133,
Corollary 3.7.3]. 

The next result extends the scope of (11.1.18) by relaxing the demands on the test
function.

Proposition 11.1.14 Assume Ω ⊆ Rn is a set of finite perimeter and abbreviate

p
σ∗ := H n−1 ∂∗ Ω. Consider two functions, f ∈ Lipc (∂∗ Ω) and g ∈ L1,loc (∂∗ Ω, σ∗ )
with p ∈ [1, ∞], and fix a pair of indices j, k ∈ {1, . . . , n}. Then

f ∈ L1∞ (∂∗ Ω, σ∗ ), ∂τ j k f ∈ L ∞ (∂∗ Ω, σ∗ ), supp(∂τ j k f ) ⊆ supp f , (11.1.99)

and ∫ ∫
(∂τ j k f )g dσ∗ = − f (∂τ j k g) dσ∗ . (11.1.100)
∂∗ Ω ∂∗ Ω

Proof The first membership in (11.1.99) is clear from Proposition 11.1.11, the
second membership in (11.1.99) is implied by the first (bearing in mind (11.1.13)),
and the last property in (11.1.99) originates in (11.1.37). To proceed, associate with
the given f a sequence of functions {ϕ } ∈N as in (11.1.93). Then, having fixed
some j, k ∈ {1, . . . , n}, we may write
700 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ ∫
(∂τ j k f )g dσ∗ = lim (∂τ j k ϕ )g dσ∗
∂∗ Ω →∞ ∂∗ Ω
∫ ∫
= − lim ϕ (∂τ j k g) dσ∗ = − f (∂τ j k g) dσ∗ (11.1.101)
→∞ ∂∗ Ω ∂∗ Ω

where the first equality is based on (11.1.92) (also taking into account the second line
in (11.1.93)), the second equality is implied by (11.1.18), and the third equality is
seen with the help of the last property in the first line of (11.1.93). Thus, (11.1.100)
is established. 
The next goal is to show that the Sobolev spaces introduced in Definition 11.1.2
are modules over space of compactly supported Lipschitz functions.
Proposition 11.1.15 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate
σ∗ := H n−1 ∂∗ Ω.
p
Then, given any pair of functions f ∈ Lipc (∂∗ Ω) and g ∈ L1,loc (∂∗ Ω, σ∗ ) with
p
p ∈ [1, ∞], it follows that f g ∈ L1 (∂∗ Ω, σ∗ ), and for each j, k ∈ {1, . . . , n} the
following product rule holds:

∂τ j k ( f g) = (∂τ j k f )g + f (∂τ j k g) at σ∗ -a.e. point on ∂∗ Ω. (11.1.102)

Proof Fix j, k ∈ {1, . . . , n} and pick ϕ ∈ Cc1 (Rn ) arbitrary. We may then write
∫ ∫
 
(∂τ j k ϕ) f g dσ∗ = ∂τ j k ( f ϕ) − ϕ(∂τ j k f ) g dσ∗
∂∗ Ω ∂∗ Ω
∫ ∫
=− f ϕ ∂τ j k g dσ∗ − ϕ(∂τ j k f )g dσ∗
∂∗ Ω ∂∗ Ω
∫
 
=− ϕ f ∂τ j k g + (∂τ j k f )g dσ∗ (11.1.103)
∂∗ Ω

where the first equality is provided by (11.1.68) (with p := ∞) and (11.1.99), while
the second equality is implied by (11.1.100). In view of (11.1.13), from (11.1.103)
p
we may now conclude that f g ∈ L1 (∂∗ Ω, σ∗ ) and that (11.1.102) holds. 
In turn, Proposition 11.1.15 contains the following useful observation.
Corollary 11.1.16 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate
σ∗ := H n−1 ∂∗ Ω. Then

p
Lipc (∂∗ Ω) ⊆ L1 (∂∗ Ω, σ∗ ). (11.1.104)
1≤p ≤∞

Proof This is an immediate consequence of the first claim in Proposition 11.1.15,

specialized to the case when g ≡ 1. 
Proposition 11.1.15 also allows us to establish the following intrinsic characteri-
p
zation of the scale of local boundary Sobolev spaces L1,loc (∂∗ Ω, σ∗ ) with p ∈ [1, ∞].
11.1 Weak Tangential Derivatives and Boundary Sobolev Spaces 701

Theorem 11.1.17 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Then for each p ∈ [1, ∞] one has
 
L1,loc (∂∗ Ω, σ∗ ) = f : ∂∗ Ω → C : ψ f ∈ L1 (∂∗ Ω, σ∗ ) for each ψ ∈ Cc∞ (Rn ) .
p p

(11.1.105)

Proof The left-to-right implication is clear from Proposition 11.1.15, so we focus

on the opposite one. To this end, fix a function f : ∂∗ Ω → C with the property that

ψ f ∈ L1 (∂∗ Ω, σ∗ ) for each ψ ∈ Cc∞ (Rn ).

p
(11.1.106)
p
In particular, this implies that f ∈ Lloc (∂∗ Ω, σ∗ ). To proceed, for each number
R ∈ (0, ∞) pick a function ψR ∈ Cc∞ (Rn ) such that ψR ≡ 1 near B(0, R). Fix
j, k ∈ {1, . . . , n} and define h jk : ∂∗ Ω → C by setting

h jk := ∂τ j k (ψR f ) on B(0, R) ∩ ∂∗ Ω for each R > 0. (11.1.107)

Thanks to (11.1.106) and the locality property for the operator ∂τ j k established in
Lemma 11.1.6, it follows that h jk is an unambiguously defined function, which
p
belongs to Lloc (∂∗ Ω, σ∗ ). Moreover, given any function ϕ ∈ Cc1 (Rn ), if R > 0 is
chosen sufficiently large so that supp ϕ ⊆ B(0, R) we may compute
∫ ∫
f (∂τ j k ϕ) dσ∗ = ψR f (∂τ j k ϕ) dσ∗
∂∗ Ω ∂∗ Ω
∫ ∫
=− ϕ ∂τ j k (ψR f ) dσ∗ = − ϕ h jk dσ∗ . (11.1.108)
∂∗ Ω ∂∗ Ω

p
In turn, (11.1.108) proves that ∂τ j k f = h jk ∈ Lloc (∂∗ Ω, σ∗ ). Since j, k ∈ {1, . . . , n}
p
have been arbitrarily chosen, this ultimately shows that f belongs to L1,loc (∂∗ Ω, σ∗ ).

Theorem 11.1.17 permits us to prove the following local version of Proposi-

tion 11.1.11.

Corollary 11.1.18 Let Ω ⊆ Rn be an arbitrary set of locally finite perimeter and

define σ∗ := H n−1 ∂∗ Ω. Then

∞ p
Liploc (∂∗ Ω) ⊆ L1,loc (∂∗ Ω, σ∗ ) ⊆ L1,loc (∂∗ Ω, σ∗ ). (11.1.109)
1≤p ≤∞

Proof If f ∈ Liploc (∂∗ Ω) and ψ ∈ Cc∞ (Rn ) then ψ f is a bounded Lipschitz function
on ∂∗ Ω, so the first inclusion in (11.1.109) follows by combining Theorem 11.1.17
(used with p = ∞) and Proposition 11.1.11. The second inclusion in (11.1.109) is
provided by item (1) in Lemma 11.1.3. 

Theorem 11.1.17 also allows us to establish the following local version of Propo-
sition 11.1.15.
702 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Corollary 11.1.19 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω.
p
Then, given any pair of functions f ∈ Liploc (∂∗ Ω) and g ∈ L1,loc (∂∗ Ω, σ∗ ) with
p ∈ [1, ∞], it follows that
p
f g ∈ L1,loc (∂∗ Ω, σ∗ ), (11.1.110)
and for each j, k ∈ {1, . . . , n} the following product rule holds:

∂τ j k ( f g) = (∂τ j k f )g + f (∂τ j k g) at σ∗ -a.e. point on ∂∗ Ω. (11.1.111)

Proof Fix j, k ∈ {1, . . . , n} and pick ψ ∈ Cc1 (Rn ) arbitrary. Choose R > 0 large
enough so that B(0, R) contains the support of ψ and select ϕ ∈ Cc1 (Rn ) with the
property that ϕ ≡ 1 in B(0, R). Then since f := ϕ f ∈ Lipc (∂∗ Ω), Proposition 11.1.15
applies and gives that
p
f g ∈ L1 (∂∗ Ω, σ∗ ) (11.1.112)
and for each j, k ∈ {1, . . . , n} we have

∂τ j k ( f g) = (∂τ j k f )g + f (∂τ j k g) at σ∗ -a.e. point on ∂∗ Ω. (11.1.113)

In turn, from (11.1.112), the fact that ψϕ = ψ in Rn , and item (v) in Proposition 11.1.9
we conclude that
p
ψ( f g) = ψ( f g) ∈ L1 (∂∗ Ω, σ∗ ) (11.1.114)
which, in light of Theorem 11.1.17, goes to show that (11.1.110) holds. Next, since
f = f and f g = f g everywhere in B(0, R) ∩ ∂∗ Ω, the locality of the operators ∂τ j k
established in Lemma 11.1.6 implies (bearing in mind (11.1.110) and (11.1.112))
that for each j, k ∈ {1, . . . , n} we have

∂τ j k f = ∂τ j k f and ∂τ j k ( f g) = ∂τ j k ( f g)
(11.1.115)
at σ∗ -a.e. point on B(0, R) ∩ ∂∗ Ω.

Together with (11.1.113), this proves that for each j, k ∈ {1, . . . , n} we have

∂τ j k ( f g) = (∂τ j k f )g + f (∂τ j k g)
(11.1.116)
at σ∗ -a.e. point on B(0, R) ∩ ∂∗ Ω.

The fact that R > 0 can be taken arbitrarily large in (11.1.116) then proves
(11.1.111). 

11.2 Distributional Tangential Derivatives

The point of view that we shall adopt here is that even for arbitrary locally integrable
functions on the geometric measure theoretic boundary of a domain it is still possible
11.2 Distributional Tangential Derivatives 703

to speak of tangential derivatives but in a distributional sense. This is made precise

in the definition below.

Definition 11.2.1 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Given an arbitrary function f ∈ Lloc 1 (∂ Ω, σ ) along with a pair
∗ ∗
of indices j, k ∈ {1, . . . , n}, define the distribution
 
∂τ j k f ∈ Lipc (∂∗ Ω) (11.2.1)

by setting
∫
 
∂τ j k f , ψ := − f (∂τ j k ψ) dσ∗ for all ψ ∈ Lipc (∂∗ Ω). (11.2.2)
∂∗ Ω
 
Henceforth, call ∂τ j k f ∈ Lipc (∂∗ Ω) , with 1 ≤ j, k ≤ n, defined as above, the
distributional tangential derivatives of f .

A few comments are in order here. First, each ψ ∈ Lipc (∂∗ Ω) belongs to the
boundary Sobolev space L1∞ (∂∗ Ω, σ∗ ) (cf. (11.1.81)), so the tangential derivative
∂τ j k ψ appearing in the right-hand side of (11.2.2) is considered in the sense of
boundary Sobolev spaces (cf. (11.1.13)). Second, since ∂τ j k ψ is essentially bounded,
with bounded support (cf. (11.1.99)), and f is locally integrable, the integral in
the right-hand side of (11.2.2) is absolutely convergent. Third, from (11.2.2) and
(11.1.82) it follows that there exists a purely dimensional constant Cn ∈ (0, ∞) with
the property that for each compact set K ⊂ Rn we have
∫
 
 ∂τ f , ψ  ≤ Cn | f | dσ∗ ψLip(∂∗ Ω)
jk
∂∗ Ω∩K (11.2.3)
for each ψ ∈ Lipc (∂∗ Ω) which vanishes identically on ∂∗ Ω \ K.

On account of the equivalence (1) ⇔ (5) in [133, Proposition 4.1.2], this implies that
∂τ j k f is indeed a distribution on ∂∗ Ω, thus justifying (11.2.1). Fourth, as is apparent
from Proposition 11.1.14,
p,q
if the function f belongs to the local Sobolev space L1,loc (∂∗ Ω, σ∗ )
for some  p, q ∈ [1, ∞], then the distributional tangential derivative

∂τ j k f ∈ Lipc (∂∗ Ω) , defined as in (11.2.2), agrees with the distri-
(11.2.4)
bution associated (as described in [133, Proposition 4.1.4]) with the
q
Sobolev tangential derivative ∂τ j k f ∈ Lloc (∂∗ Ω, σ∗ ) ⊆ Lloc
1 (∂ Ω, σ ),
∗ ∗
now considered in the sense of (11.1.16).
Conversely, from (11.2.2), Lemma 11.1.4, and (11.1.12) we see that
704 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

if f ∈ Lloc
1 (∂ Ω, σ ) has the property that all its distributional tangential
∗ ∗
 
derivatives ∂τ j k f ∈ Lipc (∂∗ Ω) with 1 ≤ j, k ≤ n actually belong
1 (∂ Ω, σ ) (in the sense of [133, Proposition 4.1.4]) then, in
to Lloc ∗ ∗ (11.2.5)
1 (∂ Ω, σ ), and the Sobolev
fact, f lies in the local Sobolev space L1,loc ∗ ∗
tangential derivatives ∂τ j k f (in the sense of (11.1.16)) agree with the
aforementioned distributional tangential derivatives of f .

Fifth, for each pair of indices j, k ∈ {1, . . . , n}, the distributional tangential
derivative operator (considered as in (11.2.2))

1 (∂ Ω, σ ) −→ Lip (∂ Ω) 

∂τ j k : Lloc ∗ ∗ c ∗
(11.2.6)
is well defined, linear, and continuous,
 
where Lipc (∂∗ Ω) is regarded as a (locally convex) topological vector space,
equipped with the weak-∗ topology (cf. [133, (4.1.37)]). Indeed, from (11.2.2) and
[133, (4.1.38)] it follows that this is a well-defined linear mapping, which is sequen-
tially continuous. Since from (11.1.3) we know that the topological vector space
1 (∂ Ω, σ ) is metrizable, we may invoke [165, Theorem 1.32(d) ⇒ (a), p. 24]
Lloc ∗ ∗
to conclude that the distributional tangential derivative operator ∂τ j k is actually a
continuous mapping in the context of (11.2.6).
Finally, in the case when Ω ⊆ Rn is a set of locally finite perimeter with the
additional property that H n−1 (∂Ω \ ∂∗ Ω) = 0, all the above considerations continue
to be valid with the geometric measure theoretic boundary ∂∗ Ω replaced by the
topological boundary ∂Ω and with the measure σ∗ replaced by σ := H n−1 ∂Ω.
Our next proposition provides a natural description of the boundary Sobolev
spaces introduced in Definition 11.1.2 in terms of distributional derivatives.

Proposition 11.2.2 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Then for each p, q ∈ [1, ∞] one has
p,q  
L1 (∂∗ Ω, σ∗ ) = f ∈ L p (∂∗ Ω, σ∗ ) : ∂τ j k f ∈ L q (∂∗ Ω, σ∗ ), 1 ≤ j, k ≤ n ,
(11.2.7)
where the tangential derivatives ∂τ j k f in the right-hand side of (11.2.7) are taken
in the sense of distributions (cf. Definition 11.2.1). Moreover, with a similar inter-
pretation,
p,q  p q 
L1,loc (∂∗ Ω, σ∗ ) = f ∈ Lloc (∂∗ Ω, σ∗ ) : ∂τ j k f ∈ Lloc (∂∗ Ω, σ∗ ), 1 ≤ j, k ≤ n .
(11.2.8)

Proof The left-to-right inclusions in (11.2.7) and (11.2.8) are implied by (11.2.4),
while the right-to-left inclusions in (11.2.7) and (11.2.8) are implied by (11.2.5). 
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 705

11.3 Distributional Derivatives Versus Weak Tangential

Derivatives

Our first result elaborates on the manner in which distributional tangential derivatives
interact with nontangential traces on boundaries.

Proposition 11.3.1 Let Ω ⊆ Rn be an open set with a lower Ahlfors regular bound-
ary, such that σ := H n−1 ∂Ω is a doubling measure on ∂Ω, and

H n−1 (∂Ω \ ∂∗ Ω) = 0. (11.3.1)

In particular, Ω is a set of locally finite perimeter, and its geometric measure theoretic
outward unit normal ν is defined σ-a.e. on ∂Ω. Fix an arbitrary aperture parameter
κ ∈ (0, ∞), a truncation parameter ε > 0, and consider a scalar-valued function u
in Ω satisfying
 1 n
u ∈ Lloc
1 (Ω, L n ), ∇u ∈ Lbdd (Ω, L n ) , Nκε u ∈ Lloc1 (∂Ω, σ),
κ−n.t. (11.3.2)
and the nontangential trace u∂Ω exists σ-a.e. on ∂Ω.
κ−n.t.
Then u∂Ω ∈ Lloc 1 (∂Ω, σ) and its distributional tangential derivatives (defined

as in (11.2.2)) satisfy
κ−n.t. .  
∂τ j k u∂Ω = ∂τ j k u in Lipc (∂Ω) for each j, k ∈ {1, . . . , n}, (11.3.3)
.  
where the weak tangential derivative ∂τ j k u ∈ Lipc (∂Ω) is defined as in [133,
Example 4.2.4] (cf. (A.0.87)-(A.0.88)).
κ−n.t.
Proof The fact that u∂Ω belongs, as claimed, to Lloc 1 (∂Ω, σ) is a consequence of

(11.3.2), [133, (8.9.8)], [133, Proposition 8.8.4], [133, Corollary 8.9.6], and (11.3.1).
κ−n.t.
Once the membership of u∂Ω to Lloc 1 (∂Ω, σ) has been established, Definition 11.2.1
κ−n.t.  
guarantees that ∂τ u
jk ∂Ω
is also well-defined in Lip (∂Ω) .
c
To proceed, define the vector fields
u
Fjk := (∂k u)e j − (∂j u)ek in Ω, for each j, k ∈ {1, . . . , n}. (11.3.4)

Note that the hypotheses formulated in (11.3.2) imply that for each pair of indices
j, k ∈ {1, . . . , n} the vector field Fjk
u defined in (11.3.4) satisfies

 1 n
u
Fjk ∈ Lbdd (Ω, L n ) and divFjk
u
= 0 in D (Ω). (11.3.5)
 
u ∈ Lip (∂Ω)  is meaningfully defined as in (A.0.79). Also, from
As such, ν • Fjk c
[133, Example 4.2.4] (cf. (A.0.87)-(A.0.88) in the Glossary) we know that
.  
∂τ j k u := ν • Fjk
u
in Lipc (∂Ω) for each j, k ∈ {1, . . . , n}, (11.3.6)
706 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

where the bullet product is defined as in [133, Proposition

 4.2.3]. 
Going further, for each ρ > 0 define Uρ := x ∈ Rn : dist(x, ∂Ω) < ρ . Given
an arbitrary ψ ∈ Lipc (∂Ω), consider a function Ψ satisfying

Ψ ∈ Lip(Ω), Ψ∂Ω = ψ, as well as Ψ ≡ 0
(11.3.7)
outside of some compact subset of Ω.
Without loss of generality, we may also assume that

supp Ψ ⊆ Uε/2 . (11.3.8)

Apply Remark 11.1.13 with ∂∗ Ω replaced by ∂Ω, f replaced by ψ, and F replaced

by Ψ, to obtain a sequence {ϕ } ∈N satisfying (for C ∈ (0, ∞), a purely dimensional
constant):
  
{ϕ } ∈N ⊆ Cc∞ (Rn ), lim supx ∈∂Ω ϕ (x) − ψ(x) = 0,
→∞
supp ϕ ⊆ K for all ∈ N and some compact set K ⊆ Uε,
   (11.3.9)
sup ∈N supx ∈Rn (∇ϕ )(x) ≤ CψLip(∂Ω),

and sup ∈N supx ∈Rn |ϕ | ≤ C · supx ∈∂Ω |ψ|.

In addition, since the ϕ ’s are constructed in Remark 11.1.13 by mollifying Ψ, we

may assume

ϕ −→ Ψ uniformly on compact sets in Rn , as → ∞. (11.3.10)

In particular, this entails

(supp Ψ) ∪ (supp ψ) ∪ supp ϕ ⊆ K ⊆ Uε . (11.3.11)

 ∈N

To proceed, fix j, k ∈ {1, . . . , n} and use (11.1.92) in Lemma 11.1.12 to conclude

that

∂τ j k ϕ converges to ∂τ j k ψ weak-∗ in L ∞ (∂Ω, σ) as → ∞. (11.3.12)

In addition, we claim that

     n
∇ϕ Ω −→ ∇Ψ Ω weak-∗ in L ∞ (Ω, L n ) as → ∞. (11.3.13)

To justify this, observe first from [133, (3.6.27)]

 (used here  ∗ with s := n and X := Ω)
that the space L 1 (Ω, L n ) is separable. Since L 1 (Ω, L n ) = L ∞ (Ω, L n ) (given that
the Lebesgue
 measure
  is sigma-finite)
 and since,  n as seen from (11.3.9), the se-
quence ∇ϕ Ω  ∈N is bounded in L ∞ (Ω, L n ) , the Sequential Banach-Alaoglu
Theorem (recorded in [133, (3.6.22)]) implies that there exists some vector-valued
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 707
 n
function Θ ∈ L ∞ (Ω, L n ) for which
   n
∇ϕ Ω −→ Θ weak-∗ in L ∞ (Ω, L n ) as → ∞. (11.3.14)
 n  n
In particular, for each vector-valued function ξ ∈ Cc∞ (Ω) ⊂ L 1 (Ω, L n ) , the
weak-∗ convergence property in (11.3.14) implies
∫ ∫
lim (∇ϕ ) · ξ dL =
n
Θ · ξ dL n . (11.3.15)
→∞ Ω Ω

On the other hand, standard integration by parts (recall that ξ is smooth and compactly
supported in Ω) allows us to write
∫ ∫ ∫
lim (∇ϕ ) · ξ dL n = − lim ϕ divξ dL n = − Ψ divξ dL n, (11.3.16)
→∞ Ω →∞ Ω Ω

where the last equality  uses n(11.3.10). In turn,  from

  (11.3.15),
 (11.3.16),
n and the
arbitrariness of ξ ∈ Cc∞ (Ω) , it follows that ∇Ψ Ω = Θ in D (Ω) . On account
of (11.3.14), this establishes (11.3.13).
κ−n.t.
Based on (11.2.2), (11.3.9), the fact that 1K u∂Ω belongs to L 1 (∂Ω, σ), and
the weak-∗ convergence result from (11.3.12), we may now write (bearing in mind
(11.3.11))
  ∫
κ−n.t. κ−n.t.
(Lip c (∂Ω)) ∂τj k u  , ψ Lip c (∂Ω) = − u∂Ω ∂τ j k ψ dσ
∂Ω
∂Ω
∫
κ−n.t.
=− 1K u∂Ω ∂τ j k ψ dσ
∂Ω
∫
κ−n.t.
= − lim 1K u∂Ω ∂τ j k ϕ dσ
→∞ ∂Ω
∫
κ−n.t.
= − lim u∂Ω ∂τ j k ϕ dσ. (11.3.17)
→∞ ∂Ω

Next, for each ∈ N define the vector field

 1 n
G := u(∂k ϕ )e j − u(∂j ϕ )ek ∈ Lloc (Ω, L n ) . (11.3.18)

Then (11.3.18), (11.3.2), (11.3.9), and (11.3.11) imply (also keeping in mind the last
line in [133, (8.2.26)]) that for each fixed ∈ N we have
708 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
κ−n.t.
G ∂Ω exists at σ-a.e. point on ∂Ω,
Nκ G belongs to the space L 1 (∂Ω, σ),
(11.3.19)
G vanishes outside of a bounded set, and
divG = (∂j u)(∂k ϕ ) − (∂k u)(∂j ϕ ) ∈ L 1 (Ω, L n ).

Consequently, upon observing from (11.3.18), (11.1.4), and (11.3.1) that for each
∈ N we have
κ−n.t. κ−n.t.
ν · G ∂Ω = u∂Ω ∂τ j k ϕ at σ-a.e. point on ∂Ω, (11.3.20)

we may apply the Divergence Formula from [133, Theorem 1.2.1] to write, for each
∈ N,
∫ ∫ ∫
κ−n.t. κ−n.t.
u∂Ω ∂τ j k ϕ dσ = ν · G ∂Ω dσ = divG dL n
∂Ω ∂Ω Ω
∫
 
= (∂j u)(∂k ϕ ) − (∂k u)(∂j ϕ ) dL n
Ω
∫
=− u
Fjk · ∇ϕ dL n, (11.3.21)
Ω

where the last equality comes from (11.3.4). On  the other hand, using (11.3.13),
u ∈ L 1 (Ω, L n ) n and div F u = 0 (cf. (11.3.5)), we
(A.0.79), plus the fact that 1K Fjk jk
may write (also keeping in mind (11.3.11))
∫ ∫
lim Fjk · ∇ϕ dL = lim
u n
1K Fjku
· ∇ϕ dL n
→∞ Ω →∞ Ω
∫ ∫
= u
1K Fjk · ∇Ψ dL n = u
Fjk · ∇Ψ dL n
Ω Ω
 
= (Lipc (∂Ω)) ν • Fjk
u
, ψ Lipc (∂Ω) . (11.3.22)

A combination of (11.3.17), (11.3.21), (11.3.22), and (11.3.6) then yields

 κ−n.t.   

(Lip c (∂Ω)) ∂τ j k u ∂Ω , ψ Lipc (∂Ω) = (Lipc (∂Ω)) ν • Fjk
u
, ψ Lipc (∂Ω)
. 
= (Lipc (∂Ω)) ∂τ j k u , ψ Lipc (∂Ω) . (11.3.23)

In view of the arbitrariness of ψ ∈ Lipc (∂Ω), this completes the proof of (11.3.3).
In the proposition below we study the manner in which weak tangential derivatives
interact with pointwise nontangential traces. The reader is reminded that the truncated
nontangential maximal operator Nκε has been defined in (A.0.75).
Proposition 11.3.2 Assume Ω is an open nonempty proper subset of Rn with a lower
Ahlfors regular boundary and such that σ:=H n−1 ∂Ω is a doubling measure on
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 709

∂Ω; in particular, Ω is a set of locally finite perimeter. Denote by ν = (ν1, . . . , νn ) the

geometric measure theoretic outward unit normal to Ω, and additionally abbreviate
σ∗ := H n−1 ∂∗ Ω. Also, fix two integrability exponents p, q ∈ [1, ∞], an aperture
parameter κ ∈ (0, ∞), and a truncation parameter ε > 0.
1,1
In this context, assume the function u ∈ Wloc (Ω) satisfies

Nκε u ∈ Lloc (∂Ω, σ), Nκε (∇u) ∈ Lloc (∂Ω, σ),

p q
(11.3.24)

and the nontangential traces

κ−n.t. κ−n.t.
u∂Ω and (∂j u)∂Ω for j ∈ {1, . . . , n}
(11.3.25)
exist at σ-a.e. point on ∂nta Ω.
κ−n.t.
Then, when considered on ∂∗ Ω, the function u∂Ω belongs to L1,loc (∂∗ Ω, σ∗ ),
p,q
κ−n.t. κ−n.t.
the functions (∂1 u)∂Ω , . . . , (∂n u)∂Ω belong to Lloc (∂∗ Ω, σ∗ ) and, for each pair of
q

indices j, k ∈ {1, . . . , n}, one has

κ−n.t. κ−n.t. κ−n.t.
∂τ j k u∂Ω = ν j (∂k u)∂Ω −νk (∂j u)∂Ω at σ∗ -a.e. point on ∂∗ Ω. (11.3.26)

Moreover, if (11.3.24) is strengthened to

Nκε u ∈ L p (∂Ω, σ) and Nκε (∇u) ∈ L q (∂Ω, σ), (11.3.27)

κ−n.t.
then actually the function u∂Ω belongs to the off-diagonal (boundary) Sobolev
κ−n.t. κ−n.t.
space L1 (∂∗ Ω, σ∗ ), while the functions (∂1 u)∂Ω , . . . , (∂n u)∂Ω belong to
p,q

L q (∂∗ Ω, σ∗ ), and there exists a finite constant C > 0, independent of u and ε > 0,
such that
 κ−n.t.     
  
u ∂Ω  p, q ≤ C Nκε u L p (∂∗ Ω,σ∗ ) + Nκε (∇u) L q (∂∗ Ω,σ∗ ) . (11.3.28)
L1 (∂∗ Ω,σ∗ )

Proof The fact that the functions

κ−n.t. κ−n.t.
f := u∂Ω and f j := (∂j u)∂Ω for j ∈ {1, . . . , n} (11.3.29)
p q
belong to Lloc (∂∗ Ω, σ∗ ) and Lloc (∂∗ Ω, σ∗ ), respectively, is clear from [133, (8.9.8)]
and [133, (8.9.44)]. To proceed, pick an arbitrary test function ϕ ∈ Cc1 (Rn ). Given
the goals we have in mind, there is no loss of generality in assuming that
 
supp ϕ ⊆ x ∈ Rn : dist(x, ∂Ω) < ε . (11.3.30)

Having fixed j, k ∈ {1, . . . , n} arbitrary, then define the following vector field in Ω:
 1 n
F := (∂j ϕ)u ek − (∂k ϕ)u e j − ϕ(∂k u)e j + ϕ(∂j u)ek ∈ Lloc (Ω, L n ) , (11.3.31)
710 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

where {e }1≤ ≤n is the standard orthonormal basis in Rn . Upon observing that we
may express F as ∂j (ϕu)ek − ∂k (ϕu)e j in Ω, it follows that

divF = 0 in D (Ω). (11.3.32)

∞ (∂Ω, σ), from (11.3.30), (11.3.31), [133, (8.2.26)],
Also, since Nκ ϕ, Nκ (∇ϕ) ∈ Lcomp
and (11.3.24) we conclude that
p q
Nκ F ∈ Lcomp (∂Ω, σ) + Lcomp (∂Ω, σ) ⊆ L 1 (∂Ω, σ). (11.3.33)

In addition, (11.3.31), (11.3.25), and (11.3.29) imply that

κ−n.t.
F ∂Ω exists at σ-a.e. point on ∂nta Ω (11.3.34)

and
 κ−n.t. 
ν · F ∂Ω = (∂τ j k ϕ) f + ϕ(ν j fk − νk f j ) at σ∗ -a.e. point on ∂∗ Ω. (11.3.35)

Finally,
in the case when Ω is an exterior domain, the vector field
(11.3.36)
F vanishes identically in a neighborhood of infinity.
Granted (11.3.31)-(11.3.36), [133, Theorem 1.2.1] applies and formula [133, (1.2.2)]
presently yields
∫ ∫
 κ−n.t.   
0= ν · F ∂Ω dσ = (∂τ j k ϕ) f + ϕ(ν j fk − νk f j ) dσ, (11.3.37)
∂∗ Ω ∂∗ Ω

hence ∫ ∫
(∂τ j k ϕ) f dσ∗ = − ϕ(ν j fk − νk f j ) dσ∗ . (11.3.38)
∂∗ Ω ∂∗ Ω
p,q
This proves that the function f belongs to L1,loc (∂∗ Ω, σ∗ ), and also establishes
(11.3.26). Furthermore, if (11.3.24) is strengthened to (11.3.27), then from (11.3.25),
(11.3.26), and [133, (8.9.8)] it follows that there exists a finite constant C > 0,
independent of u and ε, such that (11.3.28) holds. 

To illustrate the scope of Proposition 11.3.2, suppose Ω ⊆ Rn is an Ahlfors regular

domain and consider a scalar-valued function u which, for some κ > 0, satisfies
1,1
u ∈ Wloc (Ω), Nκ u, Nκ (∇u) ∈ L 1 (∂Ω, σ),
κ−n.t. κ−n.t. (11.3.39)
u∂Ω and (∇u)∂Ω exist at σ-a.e. point on ∂Ω,

where σ := H n−1 ∂Ω. Then from (10.1.56) and Proposition 11.3.2 we conclude
that
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 711
κ−n.t.
u∂Ω belongs to the Sobolev space L11 (∂Ω, σ) and
 κ−n.t.  (11.3.40)
∂τ j k u∂Ω ∈ H 1 (∂Ω, σ) for each j, k ∈ {1, . . . , n}.
Other consequences of Proposition 11.3.2 are discussed in the next corollary.

Corollary 11.3.3 Let Ω be an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary and such that σ := H n−1 ∂Ω is a doubling measure on
∂Ω. In particular, Ω is a set of locally finite perimeter. Denote by ν = (ν1, . . . , νn )
the geometric measure theoretic outward unit normal to Ω, and fix an aperture
parameter κ ∈ (0, ∞) along with a truncation parameter ε > 0. Suppose the function
1,1
u ∈ Wloc (Ω) satisfies
Nκε u, Nκε (∇u) ∈ Lloc
1
(∂Ω, σ), (11.3.41)
and the nontangential traces
κ−n.t. κ−n.t.
u∂Ω and (∂j u)∂Ω for j ∈ {1, . . . , n}
(11.3.42)
exist at σ-a.e. point on ∂nta Ω.
κ  −n.t. κ  −n.t.
Then for any other κ  > 0 the nontangential traces u∂Ω , (∇u)∂Ω exist σ-a.e.
on ∂nta Ω and are independent of κ . Moreover, with the dependence on κ  dropped,
for each j, k ∈ {1, . . . , n} one has
n.t. n.t.  n.t. 
ν j (∂k u)∂Ω = νk (∂j u)∂Ω at σ-a.e. point on ∂∗ Ω \ supp u|∂Ω . (11.3.43)

In particular, with the normal derivative of u defined as

 n.t. 
∂ν u := ν, (∇u)∂Ω at σ-a.e. point on ∂∗ Ω, (11.3.44)

one has
n.t.  n.t. 
(∇u)∂Ω = (∂ν u)ν at σ-a.e. point on ∂∗ Ω \ supp u|∂Ω . (11.3.45)

Consequently, for any vector field t : ∂∗ Ω → Cn which is tangential, in the sense

that ν, t  = 0 at σ-a.e. point on ∂∗ Ω, one has
 n.t.   n.t. 
t, (∇u)∂Ω = 0 at σ-a.e. point on ∂∗ Ω \ supp u|∂Ω . (11.3.46)

Proof The first claim in the statement is a consequence of [133, Corollary 8.9.9].
κ−n.t.
Next, Proposition 11.3.2 (used with p = 1) ensures that u∂Ω belongs to
1 (∂ Ω, σ ) which, in concert with (11.1.38), permits us to write
L1,loc ∗ ∗

κ−n.t.  κ−n.t. 
∂τ j k u∂Ω = 0 at σ∗ -a.e. point in ∂∗ Ω \ supp u∂Ω . (11.3.47)

Formula (11.3.43) now follows from this and (11.3.26).

712 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Going further, (11.3.45) is proved by making use of (11.3.43), the fact that
n
k=1 νk νk = 1 at σ-a.e. point on ∂∗ Ω, and (11.3.44), by writing

n.t. n.t. 
n
n.t.
(∇u)∂Ω = (∂j u)∂Ω = νk νk (∂j u)∂Ω
1≤ j ≤n 1≤ j ≤n
k=1


n
n.t.  n.t. 
= νk ν j (∂k u)∂Ω = ν, (∇u)∂Ω ν j
1≤ j ≤n 1≤ j ≤n
k=1
 n.t. 
= (∂ν u)ν at σ-a.e. point on ∂∗ Ω \ supp u|∂Ω . (11.3.48)

Finally, the claim in (11.3.46) is an obvious consequence of (11.3.45). 

Next we present a version of Proposition 11.3.2 which does not make any as-
sumptions on the nontangential boundary traces of the gradient of the function
in question, at the expense of imposing a mild additional condition (of geometric
measure theoretic nature) on the underlying domain.

Proposition 11.3.4 Suppose Ω is an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary and such that σ := H n−1 ∂Ω is a doubling measure on
∂Ω. Also, assume that

H n−1 ∂nta Ω ∩ (∂∗ Ω \ ∂∗ Ω =0 (11.3.49)
 
(a condition satisfied if, in particular, H n−1 ∂nta Ω \ ∂∗ Ω = 0, or if ∂∗ Ω is closed).
Abbreviate σ∗ := H n−1 ∂∗ Ω, and select two integrability exponents p ∈ [1, ∞],
q ∈ (1, ∞] along with an aperture parameter κ ∈ (0, ∞) and a truncation parameter
1,1
δ > 0. In this setting, suppose that some complex-valued function u ∈ Wloc (Ω) has
been given which satisfies the following conditions:
κ−n.t.
u∂Ω exists at σ-a.e. point on ∂nta Ω and
(11.3.50)
Nκδ u ∈ L p (∂Ω, σ), Nκδ (∇u) ∈ L q (∂Ω, σ).
κ  −n.t.
Then for any κ  > 0 the nontangential trace u∂Ω exists σ-a.e. on ∂nta Ω and is
actually independent of κ . When regarding it as a function defined σ-a.e. on ∂∗ Ω
(which, up to a σ-nullset, is contained in ∂nta Ω), this nontangential trace satisfies
(with the dependence on κ  dropped)
n.t.
u∂Ω ∈ L1 (∂∗ Ω, σ∗ )
p,q
(11.3.51)

and there exists a finite constant C > 0, independent of u and δ > 0, such that
 n.t.     
  
u  ≤ C N δ u p
∂Ω L p, q (∂ Ω,σ ) κ + N δ (∇u) q
L (∂∗ Ω,σ∗ )
. κ (11.3.52)
L (∂Ω,σ)
1 ∗ ∗
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 713

Proof The current hypotheses imply that Ω is a set of locally finite perimeter, and we
agree to denote by ν the geometric measure theoretic outward unit normal to Ω. That
κ  −n.t.
for any other aperture κ  > 0 the nontangential trace u∂Ω exists at σ-a.e. point on
∂nta Ω and is actually independent of κ  is a consequence of [133, Proposition 8.9.8].
Also, from [133, (8.9.8)], [133, (8.9.44)], and [133, (8.8.52)] we conclude that
n.t.
if f denotes the restriction of u∂Ω to ∂∗Ω then f belongs to the
(11.3.53)
space L p (∂∗ Ω, σ∗ ) and  f  L p (∂∗ Ω,σ∗ ) ≤ Nκδ u L p (∂∗ Ω,σ∗ ) .

Next, fix an arbitrary pair of indices j, k ∈ {1, . . . , n} and pick an arbitrary test

function ϕ ∈ Cc1 (Rn ). Given the goals we have in mind, there is no loss of generality
in assuming that  
supp ϕ ⊆ x ∈ Rn : dist(x, ∂Ω) < δ . (11.3.54)
Recall that {e }1≤ ≤n denotes the standard orthonormal basis in Rn and bring in the
family of functions {Φε }ε>0 ⊂ C ∞ (Ω) introduced in [133, Lemma 6.1.2]. Extend
each Φε by zero the entire space Rn to a function Φε ∈ C ∞ (Rn ) and then define

Ψε := 1 − Φε ∈ C ∞ (Rn ) for each ε > 0. (11.3.55)

Then [133, (6.1.5)-(6.1.6)] imply that there exists N ∈ (1, ∞) such that the following
properties are satisfied for each ε > 0:

0 ≤ Ψε ≤ 1, supp (∇Ψε ) ⊆ Oε \ Oε/N ,

Ψε (x) = 1 if dist(x, ∂Ω) < ε/N, Ψε (x) = 0 if x ∈ Ω \ Oε, (11.3.56)
 
and there exists C ∈ (0, ∞) such that supx ∈Ω (∇Ψε )(x) ≤ Cε −1 .

For each ε > 0 consider next the vector field in Ω given by

 1 n
Fε := u ∂j (Ψε ϕ) ek − u ∂k (Ψε ϕ) e j ∈ Lloc (Ω, L n ) . (11.3.57)

Then

divFε = (∂k u)∂j (Ψε ϕ) − (∂j u)∂k (Ψε ϕ) in D (Ω), (11.3.58)

and [133, (8.6.51)] implies (bearing in mind (11.3.54) and (11.3.50)) that divFε
belongs to L nq/(n−1) (Ω, L n ) and has bounded support. In particular,

divFε ∈ L 1 (Ω, L n ). (11.3.59)

Moreover, since Nκ ϕ, Nκ (∇ϕ) ∈ Lcomp∞ (∂Ω, σ), from (11.3.54), (11.3.57), [133,

(8.2.26)], and (11.3.50) we conclude that

p
Nκ Fε ∈ Lcomp (∂Ω, σ) ⊆ L 1 (∂Ω, σ). (11.3.60)

In addition, from (11.3.57) and (11.3.50) we see that

714 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
κ−n.t.
Fε ∂Ω exists at σ-a.e. point on ∂nta Ω, (11.3.61)

and the abbreviation in (11.3.53) implies

 κ−n.t. 
ν · Fε ∂Ω = f ∂τ j k (Ψε ϕ) at σ∗ -a.e. point on ∂∗ Ω. (11.3.62)

Finally,
in the case when Ω is an exterior domain, the vector field
(11.3.63)
Fε also vanishes identically in a neighborhood of infinity.
Granted (11.3.57)-(11.3.63), [133, Theorem 1.2.1] applies and formula [133, (1.2.2)]
presently yields
∫ ∫
f ∂τ j k ϕ dσ = f ∂τ j k (Ψε ϕ) dσ (11.3.64)
∂∗ Ω ∂∗ Ω
∫ ∫
 κ−n.t. 
= ν · Fε ∂Ω dσ = divFε dL n
∂∗ Ω Ω
∫
 
= (∂k u)∂j (Ψε ϕ) − (∂j u)∂k (Ψε ϕ) dL n
Ω
∫ 
 
= (∂k u) (∂j Ψε )ϕ + Ψε (∂j ϕ)
Ω
 
− (∂j u) (∂k Ψε )ϕ − Ψε (∂k ϕ) dL n .

Together with (11.3.56), this allows us to estimate

∫  ∫
 
 
 f ∂τ j k ϕ dσ∗  ≤ C |∇u| ε −1 |ϕ| + |∇ϕ| dL n . (11.3.65)
∂∗ Ω Oε

Collectively, [133, (8.6.51)], (11.3.54), and (11.3.50) imply that

∫
lim+ |∇u||∇ϕ| dL n = 0. (11.3.66)
ε→0 Oε

Also, assuming that ε ∈ (0, δ), based on [133, Proposition 8.6.10] we may estimate
∫
  
ε −1
|∇u||ϕ| dL n ≤ C Nκε |∇u||ϕ|  L 1 (∂Ω,σ)
Oε
   
≤ C Nκε (∇u)Nκε ϕ L 1 (∂Ω,σ) ≤ C Nκδ (∇u)Nκε ϕ L 1 (∂Ω,σ)
   
≤ C Nκδ (∇u) L q (∂Ω,σ) Nκε ϕ L q (∂Ω,σ), (11.3.67)

for some constant C ∈ (0, ∞) depending only on ∂Ω and κ, where q  ∈ [1, ∞) is the
Hölder conjugate exponent of q. From [133, (8.9.45), (8.9.43)] we know that
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 715
    
lim+ Nκε ϕ L q (∂Ω,σ) = ϕ∂  
L q (∂nta Ω,σ)
. (11.3.68)
ε→0 nta Ω

All together, (11.3.65)-(11.3.68) imply that

∫      
 
 f ∂τ j k ϕ dσ∗  ≤ C Nκδ (∇u) L q (∂Ω,σ) ϕ∂ q
  , (11.3.69)
∂∗ Ω nta Ω L (∂nta Ω,σ)

for some constant C ∈ (0, ∞) depending only on ∂Ω and κ. Since this inequality
involves only the behavior of ϕ near ∂Ω, the assumption made in (11.3.54) becomes,
in retrospect, redundant. Simply put, (11.3.69) is valid for every function ϕ ∈ Cc1 (Rn ),
with a constant C ∈ (0, ∞) depending only on ∂Ω and κ.
Moving on, let us recall from item (i) of [133, Proposition 8.8.6] that ∂nta Ω is
σ-measurable.
 Also, the
 current
  that the measure σ is doubling implies
assumption
that σ B(x, r) ∩ ∂nta Ω ≤ σ B(x, r) ∩ ∂Ω < ∞ for each x ∈ ∂Ω and each r > 0 (cf.
[133, (7.4.1)]). Granted these properties, [133, Corollary 3.7.3] applies and gives
that
  
ϕ|∂nta Ω : ϕ ∈ Cc1 (Rn ) is a dense subspace of L q (∂nta Ω, σ). (11.3.70)

Consider next the functional Λ defined as follows:

∫
   
Λ : ϕ|∂nta Ω : ϕ ∈ Cc1 (Rn ) −→ R, Λ ϕ|∂nta Ω := f ∂τ j k ϕ dσ∗ . (11.3.71)
∂∗ Ω

Then (11.1.8) and [133, (8.8.52)] imply that Λ is unambiguously defined, whereas

(11.3.69) guarantees that Λ is bounded in the norm in L q (∂nta Ω, σ). Bearing these in
mind, it follows from (11.3.70)
  ∗ Riesz’s Representation Theorem that Λ extends
and
to a functional in L q (∂nta Ω, σ) = L q (∂nta Ω, σ). As such, there exists a function
h ∈ L q (∂nta Ω, σ) with the property that
 
h L q (∂nta Ω,σ) ≤ C Nκδ (∇u) L q (∂Ω,σ) (11.3.72)

and such that

∫
 
Λ ϕ|∂nta Ω = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.73)
∂nta Ω

In view of (11.3.71) the latter identity implies

∫ ∫
f ∂τ j k ϕ dσ∗ = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.74)
∂∗ Ω ∂nta Ω

Bring in the open set O := Rn \ ∂∗ Ω and specialize the identity in (11.3.74) to the
case when ϕ ∈ Cc∞ (O). Since in such a scenario the left-hand side vanishes (given
that ∂τ j k ϕ = 0 on ∂∗ Ω), [133, Corollary 3.7.3] implies that

h = 0 at σ-a.e. point on O ∩ ∂nta Ω = ∂nta Ω \ ∂∗ Ω. (11.3.75)

716 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

In concert with (11.3.49) and [133, (8.8.52)], this forces

h = 0 at σ-a.e. point on ∂nta Ω \ ∂∗ Ω. (11.3.76)


If we now define g := h∂∗ Ω then from (11.3.76), (11.3.72), and (11.3.74) we see that
 
g ∈ L q (∂∗ Ω, σ∗ ), g L q (∂∗ Ω,σ∗ ) ≤ C Nκδ (∇u) L q (∂Ω,σ), (11.3.77)

and ∫ ∫
f ∂τ j k ϕ dσ∗ = g ϕ dσ∗ for every ϕ ∈ Cc1 (Rn ). (11.3.78)
∂∗ Ω ∂∗ Ω

Having established (11.3.77)-(11.3.78), based on (11.1.13) we may now conclude

(keeping in mind (11.3.53) and the arbitrariness of j, k ∈ {1, . . . , n}) that function
n.t.
u∂Ω belongs to L1 (∂∗ Ω, σ∗ ), and that (11.3.52) holds.
p,q


Next we discuss a variant of [133, Proposition 8.9.22] in which stronger assump-

tions on the domain allow us to place the nontangential pointwise trace in a Sobolev
space.

Corollary 11.3.5 Suppose Ω ⊆ Rn is a bounded NTA domain with the property

that σ := H n−1 ∂Ω is a doubling measure on ∂Ω. Fix an aperture parameter
κ ∈ (0, ∞) along with an integrability exponent p ∈ (1, ∞). Then there exists a
constant C ∈ (0, ∞) such that each function u ∈ C 1 (Ω) with Nκ (∇u) ∈ L p (∂Ω, σ)
satisfies
κ−n.t.
Nκ u ∈ L p (∂Ω, σ), u∂Ω exists and belongs to L1 (∂Ω, σ),
p

 κ−n.t.      (11.3.79)
and u∂Ω  L p (∂Ω,σ) ≤ C Nκ (∇u) L p (∂Ω,σ) + Nκ u L p (∂Ω,σ) .
1

Proof From [133, (5.11.4)] we know that ∂Ω is a lower Ahlfors regular set which
coincides with ∂∗ Ω. Then all desired results follow by combining [133, Proposi-
tion 8.9.22] with Proposition 11.3.4. 

Even though formula (11.3.26) is not expected to hold in the setting of Propo-
sition 11.3.4 (as no provisions are made here in relation to the existence of the
 κ−n.t. 
nontangential trace of ∇u), we may nonetheless identify each ∂τ j k u∂Ω with a
κ−n.t.
weak tangential derivative of u . This is made precise in the corollary below.
∂Ω

Corollary 11.3.6 Suppose Ω is an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary, with the property that σ := H n−1 ∂Ω is a doubling
measure on ∂Ω, and such that

H n−1 (∂Ω \ ∂∗ Ω) = 0. (11.3.80)

Also, fix p ∈ [1, ∞] and q ∈ (1, ∞], along with an aperture parameter κ > 0 and
1,1
a truncation parameter ε > 0. Finally, suppose u ∈ Wloc (Ω) is a complex-valued
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 717

function which satisfies the following conditions:

κ−n.t.
u∂Ω exists at σ-a.e. point on ∂Ω and
(11.3.81)
Nκε u ∈ L p (∂Ω, σ), Nκε (∇u) ∈ L q (∂Ω, σ).
κ  −n.t.
Then for any other κ  > 0 the nontangential trace u∂Ω exists σ-a.e. on ∂Ω and
is actually independent of κ . With the dependence on κ  dropped, this nontangential
trace satisfies
n.t.
u∂Ω belongs to L1 (∂Ω, σ), in a quantitative fashion, and
p,q

n.t. .   (11.3.82)
∂τ j k u∂Ω = ∂τ j k u in Lipc (∂Ω) for each j, k ∈ {1, . . . , n},

where the weak tangential derivatives are defined as in [133, Example 4.2.4] (cf. also
. explanation in (A.0.87)-(A.0.88)). In particular, it follows that one currently has
the
∂τ j k u ∈ L q (∂Ω, σ) for each j, k ∈ {1, . . . , n}.
1,1
Proof From the fact that the function u belongs to Wloc (Ω), the last membership in
(11.3.81), and
 1 [133, (8.6.76)]
n (used for the individual components of ∇u), it follows
that ∇u ∈ Lbdd (Ω, L n ) . Granted this, Proposition 11.3.1 presently applies and,
in concert with Proposition 11.3.4 and [133, Proposition 8.8.6], yields all desired
conclusions. 

Proposition 11.3.4 also permits us to derive the following useful membership

criterion to a local Sobolev spaces.

Corollary 11.3.7 Assume Ω is an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary such that σ := H n−1 ∂Ω is a doubling measure on ∂Ω
and such that  
H n−1 ∂Ω \ ∂∗ Ω = 0. (11.3.83)
Fix an integrability exponent p ∈ (1, ∞], an aperture parameter κ ∈ (0, ∞), and
a truncation parameter δ > 0. In this setting, suppose that some complex-valued
1,1
function u ∈ Wloc (Ω) has been given which satisfies the following conditions:
κ−n.t.
u∂Ω exists at σ-a.e. point on ∂Ω and
(11.3.84)
Nκδ u ∈ Lloc (∂Ω, σ), Nκδ (∇u) ∈ Lloc (∂Ω, σ).
p p

κ  −n.t.
Then for any κ  > 0 the nontangential trace u∂Ω exists σ-a.e. on ∂Ω and is
actually independent of κ . Also, with the dependence on κ  dropped, this function
satisfies
n.t.
u∂Ω ∈ L1, loc (∂Ω, σ).
p
(11.3.85)

Proof Fix an arbitrary cutoff function ψ ∈ Cc∞ (Rn ) and define w := ψu. Then
Nκδ w ∈ L p (∂Ω, σ) and Nκδ (∇w) ∈ L p (∂Ω, σ). Bearing in mind that, up to a
σ-nullset, the set ∂∗ Ω is contained in ∂nta Ω, Proposition 11.3.4 applies to the function
718 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
 n.t. 
w and gives that ψ u∂Ω ∈ L1 (∂Ω, σ). With this in hand, we may invoke Theo-
p

rem 11.1.17 to conclude that the membership claimed in (11.3.85) holds. 

There is also a version of Proposition 11.3.4 with Muckenhoupt weights, of the

sort described below.

Corollary 11.3.8 Assume Ω is an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary such that σ := H n−1 ∂Ω is a doubling measure on ∂Ω
and such that  
H n−1 ∂Ω \ ∂∗ Ω = 0. (11.3.86)
Fix an aperture parameter κ ∈ (0, ∞), a truncation parameter δ > 0, and an inte-
grability exponent p ∈ (1, ∞). Also, assume w : ∂Ω → [0, +∞] is a σ-measurable
p
function satisfying 0 < w(x) < +∞ for σ-a.e. x ∈ ∂Ω and w −1/p ∈ Lloc (∂Ω, σ),
where p ∈ (1, ∞) denotes the Hölder conjugate exponent of p. In this setting, sup-
1,1
pose that some complex-valued function u ∈ Wloc (Ω) has been given which satisfies
the following conditions:
κ−n.t.
u∂Ω exists at σ-a.e. point on ∂Ω and
(11.3.87)
Nκδ u ∈ Lloc
1 (∂Ω, σ), N δ (∇u) ∈ L p (∂Ω, wσ).
κ

κ  −n.t.
Then for any κ  > 0 the nontangential trace u∂Ω exists σ-a.e. on ∂Ω, is actually
independent of κ , and belongs to Lloc
1 (∂Ω, σ). Moreover, with the dependence on κ 

dropped, this function satisfies

n.t.
∂τ j k u∂Ω ∈ L p (∂Ω, wσ) for each j, k ∈ {1, . . . , n}
n 
  n.t.    (11.3.88)
and ∂τ j k u∂Ω  ≤ C Nκδ (∇u) L p (∂Ω,wσ)
L p (∂Ω,wσ)
j,k=1

for some constant C ∈ (0, ∞) independent of u.

Proof The argument proceeds largely as in the proof of Proposition 11.3.4, so we will
only indicate the main differences. For starters, [133, (3.6.26)] and the assumptions

on the weight function w imply that μ := w −p /p σ is a locally finite Borel-regular
measure on ∂Ω. Also,

∞ (∂Ω, σ) → L p (∂Ω, μ), and
L p (∂Ω, wσ) → Lloc
1 (∂Ω, σ), Lcomp
g1 g2  L 1 (∂Ω,σ) ≤ g1  L p (∂Ω,wσ) g2  L p (∂Ω,μ) (11.3.89)
for any two σ-measurable functions g1, g2 on ∂Ω.
  ∗
In addition, for any functional Λ ∈ L p (∂Ω, μ) Riesz’ representation theorem
ensures the existence of a function h ∈ L p (∂Ω, wσ) with the property that
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 719

h L p (∂Ω,wσ) = Λ(L p (∂Ω,μ))∗ and

∫ (11.3.90)

Λ(g) = gh dσ for each g ∈ L p (∂Ω, μ).
∂Ω

Next, in place of (11.3.67) we now have

∫
  
ε −1
|∇u||ϕ| dL n ≤ C Nκε |∇u||ϕ|  L 1 (∂Ω,σ)
Oε
 
≤ C Nκε (∇u)Nκε ϕ L 1 (∂Ω,σ)
   
≤ C Nκδ (∇u) L p (∂Ω,wσ) Nκε ϕ L p (∂Ω,μ) (11.3.91)

for some constant C ∈ (0, ∞) depending only on ∂Ω and κ, while in place of (11.3.68)
we now have     
lim+ Nκε ϕ L p (∂Ω,μ) = ϕ∂Ω  L p (∂Ω,μ) . (11.3.92)
ε→0

Since from [133, Proposition 3.7.1] we know that

  
ϕ|∂Ω : ϕ ∈ Cc1 (Rn ) is a dense subspace of L p (∂Ω, μ), (11.3.93)

the same type of argument that has produced (11.3.72)-(11.3.74) currently proves,
in view of (11.3.90), the existence of a function h ∈ L p (∂Ω, wσ) with the property
that  
h L p (∂Ω,wσ) ≤ C Nκδ (∇u) L p (∂Ω,wσ) (11.3.94)
and such that
∫ ∫
n.t.
u∂Ω ∂τ j k ϕ dσ = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.95)
∂Ω ∂Ω

Granted this, the claims in (11.3.88) follow on account of (11.1.12) (bearing in mind
(11.1.17)). 

We continue by discussing a version of Proposition 11.3.4 for Morrey spaces.

Corollary 11.3.9 Let Ω ⊆ Rn be an Ahlfors regular domain. Set σ := H n−1 ∂Ω and
pick two integrability exponents, p ∈ (1, ∞) and q ∈ (1, ∞], along with a parameter
λ ∈ (0, n − 1). Also, fix an aperture parameter κ ∈ (0, ∞) and a truncation parameter
1,1
δ > 0. In this context, assume some complex-valued function u ∈ Wloc (Ω) has been
given satisfying the following conditions:
κ−n.t.
u∂Ω exists at σ-a.e. point on ∂Ω and
(11.3.96)
Nκδ u ∈ Lloc (∂Ω, σ), Nκδ (∇u) ∈ M p,λ (∂Ω, σ).
q

κ  −n.t.
Then for any other aperture parameter κ  > 0 the nontangential trace u∂Ω
exists at σ-a.e. point on ∂Ω, is actually independent of κ , and belongs to the local
720 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

r
Sobolev space L1,loc (∂Ω, σ) where r := min{p, q} ∈ (1, ∞). Moreover, with the
dependence on κ  dropped, this function satisfies
n.t.
∂τ j k u∂Ω ∈ M p,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  n.t.    (11.3.97)
and ∂τ j k u∂Ω  ≤ C Nκδ (∇u) M p, λ (∂Ω,σ)
M p, λ (∂Ω,σ)
j,k=1

for some constant C = C(Ω, p, λ) ∈ (0, ∞) independent of u.

Proof Since Nκδ (∇u) ∈ M p,λ (∂Ω, σ) → Lloc (∂Ω, σ) (cf. (11.3.96) and (6.2.7)),
p

Corollary 11.3.7 applies and gives that for any other given κ  > 0 the nontangential
κ  −n.t.
trace u∂Ω exists σ-a.e. on ∂Ω, is actually independent of κ , and belongs to the
r
local Sobolev space L1,loc (∂Ω, σ). Henceforth, drop the dependence on the aperture
n.t.
parameter and abbreviate f := u . ∂Ω
To proceed, fix j, k ∈ {1, . . . , n} arbitrary. Also, pick a test function ϕ ∈ Cc1 (Rn )
satisfying  
supp ϕ ⊆ x ∈ Rn : dist(x, ∂Ω) < δ . (11.3.98)
As in the proof of Proposition 11.3.4 we then conclude (see (11.3.65) and (11.3.66);
cf. also (A.0.1)) that
∫  ∫
 
 
 f ∂τ j k ϕ dσ  ≤ C |∇u| ε −1 |ϕ| + |∇ϕ| dL n (11.3.99)
∂Ω Oε

and ∫
lim+ |∇u||∇ϕ| dL n = 0. (11.3.100)
ε→0 Oε

Let p ∈ (1, ∞) denote the Hölder conjugate exponent of p. From (6.2.69) we then

see that for each ε > 0 we have Nκε ϕ ∈ B p ,λ (∂Ω, σ). Keeping this in mind and
assuming that ε ∈ (0, δ), we may rely on [133, Proposition 8.6.10] and (6.2.78) to
estimate
∫
  
ε −1 |∇u||ϕ| dL n ≤ C N ε |∇u||ϕ|  1
κ L (∂Ω,σ)
Oε
   
≤ C Nκε (∇u)Nκε ϕ L 1 (∂Ω,σ) ≤ C Nκδ (∇u)Nκε ϕ L 1 (∂Ω,σ)
   
≤ C Nκδ (∇u) M p, λ (∂Ω,σ) Nκε ϕ B p, λ (∂Ω,σ), (11.3.101)

for some constant C ∈ (0, ∞) depending only on ∂Ω, κ, p, and λ.

Continue to assume that ε ∈ (0, δ) and consider now x ∈ ∂Ω and y ∈ Γκ (x) ∩ Oε
arbitrary. With C := supRn |∇ϕ| ∈ (0, ∞) we then have

|ϕ(y) − ϕ(x)| ≤ C|y − x| < C(1 + κ)dist(y, ∂Ω) < C(1 + κ)ε, (11.3.102)
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 721

hence

|ϕ(y)| ≤ |ϕ(y) − ϕ(x)| + |ϕ(x)| < C(1 + κ)ε + |ϕ(x)|. (11.3.103)

Consequently,
 ε 
Nκ ϕ (x) = sup |ϕ(y)| ≤ C(1 + κ)ε + |ϕ(x)|. (11.3.104)
y ∈Γκ (x)∩Oε

In view of the fact there exists a compact set K = K(ϕ, δ, κ) ⊆ ∂Ω with the property
that both Nκε ϕ and |ϕ| vanish on ∂Ω \ K, from (11.3.104) we ultimately conclude
that
 ε 
N ϕ − |ϕ|  ≤ C(1 + κ)ε · 1K pointwise on ∂Ω, for each ε ∈ (0, δ). (11.3.105)
κ

In particular, for each ε ∈ (0, δ) this permits us to estimate

 ε        
N ϕ p, λ ≤ Nκε ϕ − ϕ∂Ω  B p, λ (∂Ω,σ) + ϕ∂Ω  B p, λ (∂Ω,σ)
κ B (∂Ω,σ)
  
≤ C(1 + κ)ε1K  B p, λ (∂Ω,σ) + ϕ∂Ω  B p, λ (∂Ω,σ), (11.3.106)

where the last inequality also uses (6.2.75). Since 1K  B p, λ (∂Ω,σ) < +∞ (cf.
(6.2.69)), from (11.3.106) we then conclude that
    
lim+ Nκε ϕ B p, λ (∂Ω,σ) = ϕ∂Ω  B p, λ (∂Ω,σ) . (11.3.107)
ε→0

All together, (11.3.99)-(11.3.107) imply that

∫      
 
 f ∂τ j k ϕ dσ  ≤ C Nκδ (∇u) M p, λ (∂Ω,σ) ϕ∂Ω  B p, λ (∂Ω,σ), (11.3.108)
∂Ω

for some constant C ∈ (0, ∞) depending only on ∂Ω, κ, p, and λ. Given that
this inequality involves only the behavior of ϕ near ∂Ω, the assumption made in
(11.3.98) becomes, in retrospect, redundant. Hence, (11.3.108) is valid for every
function ϕ ∈ Cc1 (Rn ), with a constant C ∈ (0, ∞) depending only on ∂Ω, κ, p, and λ.
Going further, observe from (6.2.69) that
  
ϕ|∂Ω : ϕ ∈ Cc1 (Rn ) is a dense subspace of B p ,λ (∂Ω, σ). (11.3.109)

If we now consider the functional

∫
   
Λ : ϕ|∂Ω : ϕ ∈ Cc1 (Rn ) −→ R, Λ ϕ|∂Ω := f ∂τ j k ϕ dσ, (11.3.110)
∂Ω

then (11.1.8) and (A.0.1) imply that Λ is unambiguously defined, whereas (11.3.108)

guarantees that Λ is bounded in the norm in B p ,λ (∂Ω, σ). Bearing these in
mind,  ∗ from (11.3.109) and (6.2.80) that Λ extends to a functional in
 p,λ it follows
B (∂Ω, σ) = M p,λ (∂Ω, σ). On account of (6.2.79) we may then conclude that
722 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

there exists a function h ∈ M p,λ (∂Ω, σ) with the property that

 
h M p, λ (∂Ω,σ) ≤ C Nκδ (∇u) M p, λ (∂Ω,σ) (11.3.111)

and such that

∫
 
Λ ϕ|∂Ω = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.112)
∂Ω

In concert with (11.3.71) the latter identity implies

∫ ∫
f ∂τ j k ϕ dσ = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.113)
∂Ω ∂Ω

Having established (11.3.113), from (11.1.12) and (11.1.17) we may now conclude
that ∂τ j k f = −h. At this stage, all claims made in (11.3.97) are seen from the
definition of f and the properties of h. 

Finally, here is a variant of Proposition 11.3.4 for block spaces.

Corollary 11.3.10 Let Ω ⊆ Rn be an Ahlfors regular domain. Set σ := H n−1 ∂Ω

and pick two integrability exponents, p ∈ (1, ∞] and q ∈ (1, ∞), along with a
q(n−1)
parameter λ ∈ (0, n − 1). Define qλ := n−1+λ(q−1) ∈ (1, q). Also, pick an aperture
parameter κ ∈ (0, ∞) and a truncation parameter δ > 0. In this context, suppose
1,1
some complex-valued function u ∈ Wloc (Ω) has been given satisfying the following
conditions:
κ−n.t.
u∂Ω exists at σ-a.e. point on ∂Ω and
(11.3.114)
Nκδ u ∈ Lloc (∂Ω, σ), Nκδ (∇u) ∈ B q,λ (∂Ω, σ).
p

κ  −n.t.
Then for any other aperture parameter κ  > 0 the nontangential trace u∂Ω
exists at σ-a.e. point on ∂Ω, is actually independent of κ , and belongs to the local
r
Sobolev space L1,loc (∂Ω, σ) where r := min{p, qλ } ∈ (1, ∞). Furthermore, with the
dependence on κ  dropped, this function satisfies
n.t.
∂τ j k u∂Ω ∈ B q,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  n.t.    (11.3.115)
and ∂τ j k u∂Ω  ≤ C Nκδ (∇u) B q, λ (∂Ω,σ)
B q, λ (∂Ω,σ)
j,k=1

for some constant C = C(Ω, q, λ) ∈ (0, ∞) independent of u.

Proof The argument here largely proceeds along the lines of the earlier proof of
Corollary 11.3.9 with several key differences, singled out below. For one thing,
since Nκδ (∇u) ∈ B q,λ (∂Ω, σ) → L qλ (∂Ω, σ) (cf. (11.3.114) and (6.2.71)), Corol-
κ  −n.t.
lary 11.3.7 guarantees that for any κ  > 0 the nontangential trace u exists
∂Ω
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 723

σ-a.e. on ∂Ω, is actually independent of κ , and belongs to the local Sobolev space
r
L1,loc (∂Ω, σ). To proceed, we agree to drop the dependence on the aperture param-
n.t.
eter and abbreviate f := u . Also, fix j, k ∈ {1, . . . , n} and pick an arbitrary test
∂Ω
function ϕ ∈ Cc1 (Rn ). As before, there is no loss of generality in assuming that
ϕ satisfies (11.3.98). As in the proof of Proposition 11.3.4 we then conclude (see
(11.3.65) and (11.3.66); cf. also (A.0.1)) that
∫  ∫
 
 
 f ∂τ j k ϕ dσ  ≤ C |∇u| ε −1 |ϕ| + |∇ϕ| dL n (11.3.116)
∂Ω Oε

and ∫
lim+ |∇u||∇ϕ| dL n = 0. (11.3.117)
ε→0 Oε

Let q  ∈ (1, ∞) denote the Hölder conjugate exponent of q. From (6.2.15) we then
∞ (∂Ω, σ) ⊆ M̊ q,λ (∂Ω, σ). Keeping this
see that for each ε > 0 we have Nκε ϕ ∈ Lcomp
in mind and assuming that ε ∈ (0, δ), we may rely on [133, Proposition 8.6.10] and
(6.2.78) to estimate
∫
  
ε −1
|∇u||ϕ| dL n ≤ C Nκε |∇u||ϕ|  L 1 (∂Ω,σ)
Oε
   
≤ C Nκε (∇u)Nκε ϕ L 1 (∂Ω,σ) ≤ C Nκδ (∇u)Nκε ϕ L 1 (∂Ω,σ)
   
≤ C Nκδ (∇u) B q, λ (∂Ω,σ) Nκε ϕ M q, λ (∂Ω,σ), (11.3.118)

for some constant C ∈ (0, ∞) depending only on ∂Ω, κ, q, and λ. Next, based on
(11.3.105), for each ε ∈ (0, δ) we may estimate
 ε         
N ϕ  q  , λ ≤ N ε ϕ − ϕ  q, λ +   ϕ   q  , λ
κ M (∂Ω,σ) κ ∂Ω M (∂Ω,σ) ∂Ω M (∂Ω,σ)
  
≤ C(1 + κ)ε1K  M q, λ (∂Ω,σ) +  ϕ 
∂Ω M q
, λ , (11.3.119)
(∂Ω,σ)

where the last inequality also uses (6.2.3). Since 1K  M q, λ (∂Ω,σ) < +∞ (cf. (6.2.7)),
from (11.3.119) we then conclude that
    
lim+ Nκε ϕ M q, λ (∂Ω,σ) = ϕ∂Ω  M q, λ (∂Ω,σ) . (11.3.120)
ε→0

All together, (11.3.116)-(11.3.120) imply that

∫      
 
 f ∂τ j k ϕ dσ  ≤ C Nκδ (∇u) B q, λ (∂Ω,σ) ϕ∂Ω  M q, λ (∂Ω,σ), (11.3.121)
∂Ω

for some constant C ∈ (0, ∞) depending only on ∂Ω, κ, q, and λ. Going further,
observe from (6.2.15) that
  
ϕ|∂Ω : ϕ ∈ Cc1 (Rn ) is a dense subspace of M̊ q ,λ (∂Ω, σ). (11.3.122)
724 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

If we now consider the functional

∫
   
Λ : ϕ|∂Ω : ϕ ∈ Cc1 (Rn ) −→ R, Λ ϕ|∂Ω := f ∂τ j k ϕ dσ, (11.3.123)
∂Ω

then (11.1.8) and (A.0.1) imply that Λ is unambiguously defined, whereas (11.3.121)

guarantees that Λ is bounded in the norm in M̊ q ,λ (∂Ω, σ) (inherited from
,λ
M (∂Ω, σ)). Bearing these in mind,
q
  it follows ∗ from (11.3.122) and (6.2.155)
that Λ extends to a functional in M̊ q ,λ (∂Ω, σ) = B q,λ (∂Ω, σ). On account of
(6.2.79) we may then conclude that there exists a function h ∈ B q,λ (∂Ω, σ) with the
property that  
h B q, λ (∂Ω,σ) ≤ C Nκδ (∇u) B q, λ (∂Ω,σ) (11.3.124)
and such that
∫
 
Λ ϕ|∂Ω = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.125)
∂Ω

In concert with (11.3.71) the latter identity implies

∫ ∫
f ∂τ j k ϕ dσ = h ϕ dσ for every ϕ ∈ Cc1 (Rn ). (11.3.126)
∂Ω ∂Ω

Having established (11.3.126), from (11.1.12) and (11.1.17) we may now conclude
that ∂τ j k f = −h. At this stage, all claims made in (11.3.115) are seen from the
definition of f and the properties of h. 

Among other things, Proposition 11.3.4 is relevant in the formulation of the

(off-diagonal) Regularity Problem for some give second-order M × M system of
differential operators
 n
L= ∂j A jk ∂k (11.3.127)
j,k=1

in an open set Ω ⊆ Rn ,
where each A jk is an M × M matrix with complex-valued
∞ (Ω, L n ). Specifically, make the assumption that
entries from Lloc

the set ∂Ω is lower Ahlfors regular and the mea-

(11.3.128)
sure σ := H n−1 ∂Ω is doubling on ∂Ω.

In this setting, having fixed two integrability exponents p, q ∈ (1, ∞) along with
some aperture parameter κ ∈ (0, ∞), we formulate said boundary value problem as
the task of finding some C M -valued function u in Ω satisfying
⎧  1,1  M
⎪
⎪ u ∈ Wloc (Ω) , Lu = 0 in Ω,
⎪
⎨
⎪
Nκ u ∈ L (∂Ω, σ), Nκ (∇u) ∈ L q (∂Ω, σ),
p
(11.3.129)
⎪
⎪ κ−n.t.
⎪
⎪ u
⎩ ∂Ω = f at σ∗ -a.e. point on ∂∗ Ω,
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 725

where σ∗ := H n−1 ∂∗ Ω and the boundary datum f is a given C M -valued func-
tion defined σ∗ -a.e. on ∂∗ Ω. Above, all partial derivatives are taken in the sense of
distributions in Ω. Also, the fact that the boundary condition is meaningfully for-
mulated is a consequence item (iii) in [133, Proposition 8.8.6] (whose applicability
is ensured by (11.3.128)). The main point of this discussion is that, as seen from
Proposition 11.3.4,
if Ω also satisfies (11.3.49), then a necessary condition for the
solvability of the off-diagonal Regularity Problem (11.3.129) is that
the boundary datum f belongs to the off-diagonal Sobolev space (11.3.130)
 p,q M
L1 (∂∗ Ω, σ∗ ) .

The result in the proposition below sheds further light on the issue of regularity
of the nontangential boundary trace.

Proposition 11.3.11 Suppose Ω ⊂ Rn is a locally pathwise nontangentially acces-

sible set with a lower Ahlfors regular boundary, such that σ := H n−1 ∂Ω is a
doubling measure on ∂Ω, and for which
 
H n−1 ∂Ω \ ∂∗ Ω = 0. (11.3.131)

Then there exists κ ∈ (0, ∞) with the property that if a function u ∈ C 1 (Ω)
satisfies3

Nκε (∇u) ∈ L p (∂Ω, σ) for some κ > 0, ε > 0, p ∈ (1, ∞], and
κ−n.t. (11.3.132)

u exists σ-a.e. on ∂nta Ω and belongs to L p (∂Ω, σ),
∂Ω

it follows that, for any other given aperture parameter κ  ∈ (0, ∞), the nontangential
boundary trace
κ  −n.t.

u exists at σ-a.e. point on ∂Ω, agrees with the trace from (11.3.132),
∂Ω
p
and belongs to the boundary Sobolev space L1 (∂Ω, σ).
(11.3.133)

Proof We may rely on [133, Propositions 8.9.16-8.9.17] to presently guarantee the

existence of some aperture parameter κ ∈ (0, ∞), some small θ ∈ (0, 1), and some
small ε > 0, such that
  κ−n.t.  
 
(Nκθoε u)(x) ≤  u∂Ω (x) + ε Nκε (∇u) (x) at σ-a.e. x ∈ ∂Ω. (11.3.134)

We may also invoke [133, Proposition 8.4.1] to ensure that Nκε (∇u) ∈ L p (∂Ω, σ). In
concert with [133, (8.2.28)] and the second line in (11.3.132), this ultimately implies

3 recall that any LPNA set Ω ⊆ R n for which σ := H n−1 ∂Ω is a doubling measure on ∂Ω has
the property that H n−1 ∂Ω \ ∂nta Ω = 0; cf. [133, Proposition 8.9.16]
726 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Nκδo u ∈ L p (∂Ω, σ) for some small δ > 0 and κo > 0. (11.3.135)

With this in hand, the claim in the first line of (11.3.133) now follows from [133,
Corollary 8.9.9], while the claim in the second line of (11.3.133) is seen from what
we have just proved and Proposition
 11.3.4 (whose
 applicability is ensured by the
fact that (11.3.131) implies H n−1 ∂nta Ω \ ∂∗ Ω = 0). 
We conclude this section by proving an integration by parts formula along the
geometric measure theoretic boundary of an open set, of the sort made precise in the
proposition below.

Proposition 11.3.12 Let Ω be an open nonempty proper subset of Rn with a lower

Ahlfors regular boundary and such that σ := H n−1 ∂Ω is a doubling measure on
∂Ω. Define σ∗ := H n−1 ∂∗ Ω. Also, fix two pairs of exponents p, q, p, q  ∈ [1, ∞]
satisfying 1/p+1/p = 1/q +1/q  = 1, along with an aperture parameter κ ∈ (0, ∞).
1,1
In this context, assume the functions u, w ∈ Wloc (Ω) have satisfy

Nκ u ∈ L p (∂Ω, σ), Nκ w ∈ L q (∂Ω, σ),

(11.3.136)
Nκ (∇u) ∈ L q (∂Ω, σ), Nκ (∇w) ∈ L p (∂Ω, σ),

and the nontangential traces

κ−n.t. κ−n.t. κ−n.t. κ−n.t.
u∂Ω , w ∂Ω , (∇u)∂Ω , (∇w)∂Ω
(11.3.137)
exist at σ-a.e. point on ∂nta Ω.
κ−n.t.
Then, when considered on ∂∗ Ω, the function u∂Ω belongs to L1 (∂∗ Ω, σ∗ ), the
p,q
κ−n.t.  
function w ∂Ω belongs to L1 (∂∗ Ω, σ∗ ) and, for each j, k ∈ {1, . . . , n}, one has
q ,p

the integration by parts formula on the (geometric measure theoretic) boundary:

∫ ∫
 κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 
 
∂τ j k u ∂Ω w ∂Ω dσ∗ = − u∂Ω ∂τ j k w ∂Ω dσ∗, (11.3.138)
∂∗ Ω ∂∗ Ω

when either Ω is bounded, or ∂Ω is unbounded. In the remaining case, i.e., when Ω

is unbounded and ∂Ω is bounded, formula (11.3.138) continues to hold under the
additional assumption that there exists λ ∈ (1, ∞) such that
∫
 
|u||∇w| + |∇u||w| dL n = o(R) as R → ∞. (11.3.139)
[B(0,λ R)\B(0,R)]∩Ω

Proof The assumptions imply that Ω is a set of locally finite perimeter and we let
ν = (ν1, . . . , νn ) denote the geometric measure theoretic outward unit normal to
κ−n.t. κ−n.t. q, p
Ω. That u∂Ω ∈ L1 (∂∗ Ω, σ∗ ) and w ∂Ω ∈ L1 (∂∗ Ω, σ∗ ) is a consequence of
p,q

Proposition 11.3.2. From (11.3.136), [133, Lemma 8.3.1], and the identification
1,∞
Wloc (O) = Liploc (O) for any open set O ⊆ Rn, (11.3.140)
11.3 Distributional Derivatives Versus Weak Tangential Derivatives 727

we also conclude that

u, w ∈ Liploc (Ω). (11.3.141)
To proceed, fix two arbitrary indices j, k ∈ {1, . . . , n} and introduce the vector
field

F := (∂k u)we j − (∂j u)wek + u(∂k w)e j − u(∂j w)ek (11.3.142)

where, as usual, {e }1≤ ≤n is the standard orthonormal basis in Rn . From (11.3.141)-
(11.3.142) we see that  1 n
F ∈ Lloc (Ω, L n ) (11.3.143)

and after expressing F as ∂k (uw)e j − ∂j (uw)ek in Ω, we also conclude that

divF = 0 in D (Ω). (11.3.144)

In addition, (11.3.142) and (11.3.136) ensure that

Nκ F ∈ L 1 (∂Ω, σ), (11.3.145)

while (11.3.142) and (11.3.137) guarantee that

κ−n.t.
F ∂Ω exists at σ-a.e. point on ∂nta Ω (11.3.146)

and, in fact,
κ−n.t.  κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 
F ∂Ω = (∂k u)∂Ω w ∂Ω e j − (∂j u)∂Ω w ∂Ω ek
 κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 
+ u∂Ω (∂k w)∂Ω e j − u∂Ω (∂j w)∂Ω ek (11.3.147)

at σ-a.e. point on ∂nta Ω. Also mindful of [133, (8.8.52)], from (11.3.147) we see that
at σ∗ -a.e. point on ∂∗ Ω we may write
 κ−n.t.    κ−n.t.   κ−n.t.   κ−n.t. 
ν · F ∂Ω = ν j (∂k u)∂Ω − νk (∂j u)∂Ω w ∂Ω
  κ−n.t.   κ−n.t.   κ−n.t. 
+ ν j (∂k w)∂Ω − νk (∂j w)∂Ω u∂Ω

 κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 

= ∂τ j k u∂Ω w ∂Ω + u∂Ω ∂τ j k w ∂Ω , (11.3.148)

where the last equality is provided by (11.3.26). Finally,

in the case when Ω is unbounded and ∂Ω is

(11.3.149)
bounded, the vector field F satisfies [133, (1.2.3)].
Having established (11.3.143)-(11.3.149), [133, Theorem 1.2.1] applies and formula
[133, (1.2.2)] presently yields
728 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫
 κ−n.t. 
0= ν · F ∂Ω dσ
∂∗ Ω
∫   κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 
= ∂τ j k u∂Ω w ∂Ω + u∂Ω ∂τ j k w ∂Ω dσ, (11.3.150)
∂∗ Ω

from which (11.3.138) readily follows. 

11.4 The Tangential Gradient

Fix n ∈ N and assume Ω ⊆ Rn is a given set of locally finite perimeter. As in the

past, we abbreviate σ∗ := H n−1 ∂∗ Ω and agree to denote by ν = (ν1, . . . , νn ) the
geometric measure theoretic outward unit normal to Ω. In this context, we define the
tangential gradient of an arbitrary function f ∈ L1,loc 1 (∂ Ω, σ ) as the vector
∗ ∗
given by

n
∇tan f := νk ∂τk j f at σ∗ -a.e. point on ∂∗ Ω. (11.4.1)
1≤ j ≤n
k=1

That this
 definition is natural may be seen by considering the special case when
f := ϕ∂∗ Ω for some ϕ ∈ C 1 (Rn ). Then from (11.1.26) and (11.1.105) we deduce
1 (∂ Ω, σ ) and
that f ∈ L1,loc ∗ ∗


n 
n
     
∇tan f = νk ∂τk j f = νk νk ∂j ϕ ∂∗ Ω − ν j ∂k ϕ ∂∗ Ω
1≤ j ≤n 1≤ j ≤n
k=1 k=1
   
= ∇ϕ ∂∗ Ω − ν · ∇ϕ ∂∗ Ω ν at σ∗ -a.e. point on ∂∗ Ω, (11.4.2)

which is in agreement
 with what one would expect the tangential gradient of the
function f = ϕ∂∗ Ω to be in this case.
The piece of terminology used in connection to (11.4.1) is justified in light of the
following result.

Lemma 11.4.1 Let Ω ⊆ Rn be a set of locally finite perimeter. Denote by ν the geo-
metric measure theoretic outward unit normal to Ω and abbreviate σ∗ := H n−1 ∂∗ Ω.
1 (∂ Ω, σ ) one has
Then for each given function f ∈ L1,loc ∗ ∗

 
ν, ∇tan f = 0 at σ∗ -a.e. point on ∂∗ Ω, (11.4.3)

and
supp (∇tan f ) ⊆ supp f . (11.4.4)
In fact,
11.4 The Tangential Gradient 729

if the function f ∈ L1,loc

1 (∂ Ω, σ ) is locally constant on U ⊆ ∂ Ω (in
∗ ∗ ∗
the sense that for each x ∈ U there exists r > 0 with the property that (11.4.5)
f B(x,r)∩∂∗ Ω is constant) then ∇tan f = 0 at σ∗ -a.e. point in U.

Moreover, for every g ∈ Liploc (∂∗ Ω) one has f g ∈ L1,loc

1 (∂ Ω, σ ) and the follow-
∗ ∗
ing product rule holds:

∇tan ( f g) = f ∇tan g + g∇tan f at σ∗ -a.e. point on ∂∗ Ω. (11.4.6)

Proof Let (ν1, . . . , νn ) be the components of ν. Then (11.4.1) implies that at σ∗ -a.e.
point on ∂∗ Ω we have

  
n
ν, ∇tan f = ν j νk ∂τk j f = 0, (11.4.7)
j,k=1

where the last equality is based on the antisymmetry property recorded in (11.1.24).
This proves (11.4.3). Next, (11.4.4) is a direct consequence of (11.4.1) and (11.1.37),
whereas (11.4.5) is implied by (11.4.1) and (11.1.39). Finally, Corollary 11.1.19
1 (∂ Ω, σ ), while from (11.1.111) and
(used here with p = 1) ensures that f g ∈ L1,loc ∗ ∗
(11.4.1) we conclude that (11.4.6) holds. 

As is apparent from (11.4.1), the scalar components of the tangential gradient

operator ∇tan , associated with any given set Ω of locally finite perimeter, are linear
combinations of the elementary tangential derivative operators ∂τ j k whose coeffi-
cients are components of the geometric measure theoretic outward unit normal ν to
Ω. In the case when the set Ω in question is actually a UR domain, the converse is
true, namely each tangential derivative operator ∂τ j k may be recovered from the tan-
gential gradient operator and the geometric measure theoretic outward unit normal,
as indicated in the proposition below.

Proposition 11.4.2 Fix n ∈ N with n ≥ 2 and assume Ω ⊆ Rn is a UR domain.

Denote its geometric measure theoretic outward unit normal by ν = (ν1, . . . , νn ), and
abbreviate σ := H n−1 ∂Ω. Also, fix an integrability exponent p ∈ (1, ∞). Then for
p
each function f ∈ L1,loc (∂Ω, σ) and each pair of indices j, k ∈ {1, . . . , n} one has

∂τ j k f = ν j (∇tan f )k − νk (∇tan f ) j at σ-a.e. point on ∂Ω. (11.4.8)

Consequently,
   
(∇tan f )(x) ≈ n (∂τ f )(x), uniformly for
j,k=1 jk
p (11.4.9)
f ∈ L1,loc (∂Ω, σ) and σ-a.e. x ∈ ∂Ω.

In particular,
730 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter


n
p
∇tan f  L p (∂Ω,σ) ≈ ∂τ j k f  L p (∂Ω,σ), uniformly for f ∈ L1 (∂Ω, σ),
j,k=1
(11.4.10)
hence
 f  L p (∂Ω,σ) ≈  f  L p (∂Ω,σ) + ∇tan f  L p (∂Ω,σ),
1
p (11.4.11)
uniformly for f ∈ L1 (∂Ω, σ).

To facilitate the discussion in the proof of Proposition 11.4.2, we shall review

some basic properties of the boundary-to-domain harmonic double layer operator D
associated with an open set Ω ⊆ Rn having a UR boundary. Define σ := H n−1 ∂Ω,
and let ν be the geometric measure theoretic outward unit normal to Ω. With EΔ
denoting the standard fundamental solution for the Laplacian in Rn (cf. (A.0.38)),
 
for each function f ∈ L 1 ∂∗ Ω , 1+σ(y)
|y | n consider
∫
1 ν(y), y − x
D f (x) := f (y) dσ(y) (11.4.12)
ωn−1 |x − y| n
∂∗ Ω

at each x ∈ Ω. Here are some properties established in [136, §1.5] which are
significant for us. For one thing, with the summation convention in effect we have
∫
 
∂ D f (x) = (∂j EΔ )(x − y)(∂τ j f )(y) dσ(y) (11.4.13)
∂∗ Ω

for each index ∈ {1, . . . , n}, each point x ∈ Ω, and each function

σ(x)
f ∈ L 1 ∂∗ Ω , with the property that
1 + |x| n−1
(11.4.14)
σ(x)
∂τ j k f ∈ L ∂∗ Ω ,1
for all j, k ∈ {1, . . . , n}.
1 + |x| n−1
Also, if ∂Ω is actually a UR set, then
  κ−n.t.
∂ D f ∂Ω exists at σ-a.e. point on ∂∗ Ω,
(11.4.15)
for each f as in (11.4.14) and each ∈ {1, . . . , n},

and for each exponent p ∈ (1, ∞) and aperture parameter κ ∈ (0, ∞) there exists
constant C = C(Ω, n, κ, p) ∈ (0, ∞) such that
 
Nκ (D f ) p ≤ C f  L p (∂∗ Ω,σ),
L (∂Ω,σ)
      (11.4.16)
Nκ ∇D f  p ≤ C nj,k=1 ∂τ j k f  L p (∂∗ Ω,σ) .
L (∂Ω,σ)

Lastly, whenever ∂Ω is a UR set then, for each fixed aperture parameter κ ∈ (0, ∞),
it follows that and at σ-a.e. point on ∂∗ Ω we have
11.4 The Tangential Gradient 731
  κ−n.t.  
D f ∂Ω = 12 I + KΔ f , (11.4.17)

for each function f ∈ L 1 ∂∗ Ω , 1+σ(x)

|x | n−1
, where KΔ is the principal-value (boundary-
to-boundary) harmonic double layer operator, defined in relation to Ω as in (A.0.52).
We are now prepared to present the proof of Proposition 11.4.2
Proof of Proposition 11.4.2 If Ω is empty or coincides with the entire Rn there is
nothing to prove, so we exclude these cases in what follows. In dealing with formula
(11.4.8), thanks to Theorem 11.1.17, Lemma 11.1.6, and Lemma 11.4.1, there is no
loss of generality in assuming that the function f actually belongs to the (global)
p
Sobolev space L1 (∂Ω, σ). Assuming this is the case, bring in the boundary-to-
boundary harmonic double layers D ± associated with Ω+ := Ω and Ω− := Rn \ Ω as
above. If we now define u± := ±D ± f in Ω± , then from (11.4.13)-(11.4.17) we see
(with κ ∈ (0, ∞) denoting a fixed background aperture parameter) that

u± ∈ C ∞ (Ω± ),
Nκ u± ∈ L p (∂Ω, σ), Nκ (∇u± ) ∈ L p (∂Ω, σ), (11.4.18)
κ−n.t.   κ−n.t.
u± ∂Ω and ∇u± ∂Ω exist σ-a.e. on ∂Ω,

as well as
κ−n.t. κ−n.t.
f = u+ ∂Ω − u− ∂Ω at σ-a.e. point on ∂Ω. (11.4.19)

Next, for two arbitrary indices j, k ∈ {1, . . . , n} write

ν j (∇tan f )k − νk (∇tan f ) j
 κ−n.t.   κ−n.t. 
= ν j ∇tan u+ ∂Ω − νk ∇tan u+ ∂Ω
k j
 κ−n.t.   κ−n.t. 
− ν j ∇tan u− ∂Ω + νk ∇tan u− ∂Ω . (11.4.20)
k j

Granted this, (11.4.8) is proved as soon as we show that

 κ−n.t.   κ−n.t.   κ−n.t. 
ν j ∇tan u± ∂Ω − νk ∇tan u± ∂Ω± = ∂τ j k u± ∂Ω . (11.4.21)
k j

However, by (11.4.1) and Proposition 11.3.2, for each j, k ∈ {1, . . . , n} we have (with
the summation convention over repeated indices in effect)
732 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
 κ−n.t.   κ−n.t.   κ−n.t.   κ−n.t. 
ν j ∇tan u± ∂Ω k − νk ∇tan u± ∂Ω j = ν j νr ∂τr k u± ∂Ω − νk νs ∂τs j u± ∂Ω
  κ−n.t.   κ−n.t.
= ν j νr νr ∂k u± ∂Ω − ν j νr νk ∂s u± ∂Ω
  κ−n.t.   κ−n.t.
− νk νs νs ∂j u± ∂Ω + νk νs ν j ∂s u± ∂Ω
  κ−n.t.   κ−n.t.
= ν j ∂k u± ∂Ω − νk ∂j u± ∂Ω
 κ−n.t. 
= ∂τ j k u± ∂Ω , (11.4.22)

at σ-a.e. point on ∂Ω. Thus, (11.4.21) holds, completing the proof of (11.4.8).
Together, (11.4.1) and (11.4.8) then justify (11.4.9). Since (11.4.9) then readily
implies (11.4.10)-(11.4.11), the proof of Proposition 11.4.2 is complete. 
Other useful properties of the tangential gradient are described in the following
proposition.
Proposition 11.4.3 Pick n ∈ N with n ≥ 2 and assume Ω ⊆ Rn is an open set
satisfying a two-sided local John condition and with an Ahlfors regular boundary.
Define σ := H n−1 ∂Ω and fix some p ∈ (1, ∞). Then for each function

σ(x) p
f ∈ L 1 ∂Ω, ∩ Lloc (∂Ω, σ) (11.4.23)
1 + |x| n
with the property that

σ(x) p
∂τ j k f ∈ L 1 ∂Ω, ∩ Lloc (∂Ω, σ) for all j, k ∈ {1, . . . , n}, (11.4.24)
1 + |x| n−1
the following equivalences are true:

f is locally constant on ∂Ω ⇐⇒ ∇tan f = 0 at σ-a.e. point on ∂Ω

(11.4.25)
⇐⇒ for all j, k ∈ {1, . . . , n} one has ∂τ j k f = 0 at σ-a.e. point on ∂Ω.

Essentially the same proof works with the two-sided local John condition replaced
by the demand that Ω is a UR domain satisfying a local connectivity condition.
As a preamble to the proof of the above proposition, we briefly discuss the
modified boundary-to-domain harmonic double layer operator Dmod associated with
an open set Ω ⊆ Rn whose topological boundary is a UR set, as well as its principal-
value version, Kmod , on ∂∗ Ω. Specifically, define σ := H n−1 ∂Ω and denote by ν
the geometric measure theoretic outward unit normal to Ω. Also, recall the standard
fundamental solution EΔ for the Laplacian in Rn (cf. (A.0.38)). Then for each function
 
f ∈ L 1 ∂∗ Ω , 1+σ(y)
|y | n we define
∫ 
1 ν(y), y − x ν(y), y
Dmod f (x) := − · 1Rn \B(0,1) (y) f (y) dσ(y)
ωn−1 |x − y| n |y| n
∂∗ Ω
(11.4.26)
11.4 The Tangential Gradient 733

at each x ∈ Ω, and
∫ !
1 ν(y), y − x
Kmod f (x) := lim+ · 1Rn \B(x,ε) (y)
ε→0 ωn−1 |x − y| n
y ∈∂∗ Ω
|x−y |>ε
"
ν(y), y
− · 1Rn \B(0,1) (y) f (y) dσ(y)
|y| n
(11.4.27)

at σ-a.e. point x ∈ ∂Ω. These integral operators have been studied at length in [136,
§1.8], and here we record several properties which are particularly relevant for the
present goals. First, with the summation convention in effect we have
∫
 
∂ Dmod f (x) = (∂j EΔ )(x − y)(∂τ j f )(y) dσ(y) (11.4.28)
∂∗ Ω

for each index ∈ {1, . . . , n}, each point x ∈ Ω, and each function

σ(x)
f ∈ L 1 ∂∗ Ω , with the property that
1 + |x| n
(11.4.29)
σ(x)
∂τ j k f ∈ L ∂∗ Ω ,
1
for all j, k ∈ {1, . . . , n},
1 + |x| n−1
and, with D as in (11.4.12),
σ(x)
∇Dmod f = ∇D f in Ω for each f ∈ L 1 ∂∗ Ω , . (11.4.30)
1 + |x| n−1
Moreover, if ∂Ω is actually a UR set, then
  κ−n.t.
∂ Dmod f ∂Ω exists at σ-a.e. point on ∂∗ Ω,
(11.4.31)
for each f as in (11.4.29) and each ∈ {1, . . . , n},

and for each exponent p ∈ (1, ∞) and aperture parameter κ ∈ (0, ∞) there exists
constant C = C(Ω, n, κ, p) ∈ (0, ∞) such that

   
n
 
Nκ ∇D f  p ≤ C ∂τ f  p . (11.4.32)
mod L (∂Ω,σ) jk L (∂∗ Ω,σ)
j,k=1

In addition, for each truncation parameter ε ∈ (0, ∞) we have

Nκε (Dmod f ) ∈ Lloc (∂Ω, σ) for each function

p

  (11.4.33)
f ∈ L 1 ∂∗ Ω, 1+σ(x)
p
|x | n ∩ Lloc (∂∗ Ω, σ) with p ∈ (1, ∞).
734 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Also, retaining the hypothesis that ∂Ω is a UR set it follows that

σ(x) 1,∞
Kmod : L 1 ∂∗ Ω , −→ Lloc (∂Ω, σ) (11.4.34)
1 + |x| n
is a well-defined, linear and continuous operator, which is compatible with KΔ from
(A.0.52) in the sense that
 
for any function f ∈ L 1 ∂∗ Ω , 1+σ(x)
|x | n−1
the
(11.4.35)
difference Kmod f − KΔ f is constant on ∂Ω.

Finally, in the latter geometric setting (and having fixed an arbitrary aperture param-
eter κ ∈ (0, ∞)), it follows that and at σ-a.e. point on ∂∗ Ω we have
  κ−n.t.  
Dmod f ∂Ω = 12 I + Kmod f , (11.4.36)

where f ∈ L 1 ∂∗ Ω , 1+σ(x)
|x | n is an arbitrary function.
We now turn to the proof of Proposition 11.4.3.
Proof of Proposition 11.4.3 We begin by observing that, thanks to [133, (5.11.27)]
and the current hypotheses, Ω is actually a UR domain. Granted this, the equivalence
in the second line of (11.4.25) becomes a direct consequence of (11.4.9). Consider
next the equivalence in the first line of (11.4.25). The left-to-right implication is a
consequence of (11.4.5). In the opposite direction, pick a function f as in (11.4.23)-
(11.4.24) and satisfying ∇tan f = 0 at σ-a.e. point on ∂Ω. Also, introduce Ω+ := Ω
and Ω− := Rn \ Ω, then bring in the modified boundary-to-domain harmonic double
± associated with Ω as in (11.4.26). Define u± := D ± f in Ω .
layer operators Dmod ± mod ±
In view of (11.4.28) and (11.4.8) we then see that u± are locally constant in Ω± . Pick
an arbitrary x ∈ ∂Ω and consider some small r > 0. As in [133, Definition 5.11.7]
(cf. also (A.0.54)) let xr ∈ B(x, r) ∩ Ω be the John center relative to the surface ball
Δ(x, r) := B(x, r) ∩ ∂Ω and for each fixed y ∈ Δ(x, r) consider a rectifiable path
γy : [0, 1] → Ω joining y with xr as in (A.0.54). Since u+ is locally constant   in
Ω and {γy (t) : 0 < t ≤ 1} is a connected subset of Ω, it follows that u+ γy (t) is
independent of t ∈ (0, 1]. Having fixed some sufficiently large aperture parameter
 κ−n.t.   
κ > 0, this ultimately implies that u+ ∂Ω (y) = lim+ u+ γy (t) = u+ (xr ) for σ-a.e.
t→0
κ−n.t. κ−n.t.
y ∈ Δ(x, r). Hence, u+ ∂Ω is locally constant on ∂Ω. Likewise, u− ∂Ω is locally
constant on ∂Ω, and since the jump formula (11.4.36) implies
κ−n.t. κ−n.t.
f = u+ ∂Ω − u− ∂Ω at σ-a.e. point on ∂Ω, (11.4.37)

we finally conclude that f is locally constant on ∂Ω. 

The tangential gradient allows for a natural decomposition of the nontangential

boundary trace of the “full” gradient of a function defined in an open set into
tangential and normal components, on the boundary of that set. Geometric and
11.4 The Tangential Gradient 735

analytic conditions guaranteeing the validity of such a decomposition are given in

the proposition below.

Proposition 11.4.4 Assume Ω ⊆ Rn (where n ∈ N with n ≥ 2) is a UR domain.

Denote its geometric measure theoretic outward unit normal by ν, and abbreviate
σ := H n−1 ∂Ω. Also, fix an integrability exponent p ∈ (1, ∞), along with an
aperture parameter κ ∈ (0, ∞) and a truncation parameter ε > 0. Suppose the
1,1
function u ∈ Wloc (Ω) has the property that

Nκε u, Nκε (∇u) ∈ Lloc (∂Ω, σ),

p
(11.4.38)

and the nontangential traces

κ−n.t. κ−n.t.
u∂Ω and (∂j u)∂Ω for j ∈ {1, . . . , n}
(11.4.39)
exist at σ-a.e. point on ∂Ω.
κ−n.t.
Then the function u∂Ω belongs to L1,loc (∂Ω, σ) and, with the normal derivative
p

of u defined as  n.t. 
∂ν u := ν, (∇u)∂Ω at σ-a.e. point on ∂Ω, (11.4.40)

at σ-a.e. point on ∂Ω one has

κ−n.t. κ−n.t.
(∇u)∂Ω = ∇tan u∂Ω + (∂ν u)ν (11.4.41)

and 
 κ−n.t. 2  κ−n.t. 2  2
(∇u)∂Ω  = ∇tan u∂Ω  + ∂ν u . (11.4.42)
κ−n.t.
Proof From Proposition 11.3.2 we know that u∂Ω belongs to L1,loc (∂Ω, σ). Based
p

on this, (11.4.1), (11.3.26), and (11.4.40), at σ-a.e. point on ∂Ω we may then compute

κ−n.t. 
n
κ−n.t.
∇tan u∂Ω = νk ∂τk j u∂Ω
1≤ j ≤n
k=1


n
κ−n.t. 
n
κ−n.t.
= νk νk (∂j u)∂Ω − νk ν j (∂k u)∂Ω
1≤ j ≤n
k=1 k=1
κ−n.t.
= (∇u)∂Ω − (∂ν u)ν, (11.4.43)

proving (11.4.41). Lastly, (11.4.42) is a consequence of (11.4.41) and (11.4.3). 

736 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

11.5 Sobolev Spaces on Boundaries of Ahlfors Regular Domains

Satisfying a Two-Sided Local John Condition

The Poincaré inequality discussed below refines [97, Proposition 4.13, p. 2692] by
allowing a significantly larger class of functions. Recall the local John condition,
introduced in [133, Definition 5.11.7] (cf. also (A.0.54)).
Proposition 11.5.1 Let Ω ⊆ Rn be an open set satisfying a two-sided local John
condition and whose boundary is Ahlfors regular. Abbreviate σ := H n−1 ∂Ω and
fix some integrability exponent p ∈ (1, ∞). Then there exist Ro ∈ (0, ∞] (which may
actually be taken ∞ if ∂Ω is unbounded) along with M ∈ (1, ∞) and C ∈ (0, ∞), all
depending only on the local John constants of Ω and the Ahlfors regularity constants
of ∂Ω, such that for each point xo ∈ ∂Ω, each scale R ∈ (0, Ro ), and each function
f such that

σ(x)
f ∈ L 1 ∂Ω , and
1 + |x| n+1
(11.5.1)
σ(x) p
∂τ j k f ∈ L ∂Ω ,
1
∩ Lloc (∂Ω, σ) for all j, k ∈ {1, . . . , n},
1 + |x| n
p
it follows that f ∈ L1,loc (∂Ω, σ) and, with Δ := B(xo, R) ∩ ∂Ω,
⨏  1/p ⨏  1/p
| f − fΔ | p dσ ≤ CR |∇tan f | p dσ
Δ MΔ

∞ ∫
2−j
+ CR |∇tan f | dσ. (11.5.2)
j=1
σ(2 j MΔ) 2 j MΔ\2 j−1 MΔ

In particular, the above estimate holds whenever Ω is a two-sided NTA domain

with an Ahlfors regular boundary.
A comment is in order here. In [173, Lemma 1.1, p. 406] Semmes derives a
Poincaré inequality of the type
⨏ ⨏  1/2
| f − fΔ | dσ ≤ CR |∇tan f | 2 dσ , (11.5.3)
Δ 5Δ

in the case when the quantity

νBMO (∂Ω,σ) + sup sup sup R−1 |x − y, νΔ(x,R) | (11.5.4)

x ∈∂Ω R>0 y ∈Δ(x,R)

is sufficiently small to begin with, and when both f and ∂Ω are smooth (for a constant
C which is independent of smoothness). The result described in Proposition 11.5.1
differs from this in several fundamental aspects. First, the smallness condition im-
posed on (11.5.4) is, in effect, an a priori flatness assumption on ∂Ω, a hypothesis
conspicuously absent Proposition 11.5.1. Second, the smoothness assumptions on
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 737

f and ∂Ω play a crucial role in Semmes’ proof, (as his approach clearly fails if the
smoothness assumptions on f or ∂Ω are dropped), and there is no clear mechanism
for removing them a posteriori. By way of contrast, both f and ∂Ω are allowed to
be “rough” in the formulation of Proposition 11.5.1.
The proof given below is conceptually different and follows the lines of the proof
of [97, Proposition 4.13, pp. 2693-2696]. We also make crucial use of the Calderón-
Zygmund theory for singular integrals of layer potential type, developed in [136]
(independently of the present considerations).
Proof of Proposition 11.5.1 Assume that Ω+ := Ω and Ω− := Rn \ Ω satisfy a
local John condition with constants θ ∈ (0, 1), M ∈ (1, ∞), and Ro > 0 (see [133,
Definition 5.11.7]; cf. also (A.0.54)). From [133, Proposition 5.10.4] and [133,
(5.2.4)] we know that Ω± are UR domains. Fix an integrability exponent p ∈ (1, ∞),
a scale R ∈ (0, Ro ), a reference point xo ∈ ∂Ω. In a first stage we shall show that
there exists C ∈ (0, ∞), depending only on the local John constants of Ω and the
Ahlfors regularity constants of ∂Ω, such that
⨏ 1/p ⨏ 
n  1/p
| f − fΔ | p dσ ≤ CR |∂τ j k f | p dσ
Δ 4MΔ j,k=1

∫ 
n
R2
+C |∂τ j k f (y)| dσ(y).
∂Ω\B(x o ,4M R) |xo − y| n j,k=1
(11.5.5)

for each function

σ(x)
f ∈ L 1 ∂Ω , with the property that
1 + |x| n
(11.5.6)
σ(x)
∂τ j k f ∈ L ∂Ω ,
1
for all j, k ∈ {1, . . . , n}.
1 + |x| n−1
To this end, fix f as in (11.5.6) and, with E denoting the standard (radial)
fundamental solution for the Laplacian in Rn , define g := (g j )1≤ j ≤n where, using
the summation convention over repeated indices, we have set
∫
g j (x) := − (∂k E)(x − y)(∂τ j k f )(y)1∂Ω\4MΔ (y) dσ(y),
∂Ω (11.5.7)
 
for all j ∈ {1, . . . , n} and x ∈ Rn \ ∂Ω ∪ B(xo, 4M R).

Then, as is apparent from (11.5.7) and the second line in (11.5.6), the vector-valued
function g is well defined and
   n
g ∈ C ∞ Rn \ ∂Ω ∪ B(xo, 4M R) . (11.5.8)

In particular, it makes sense to consider

738 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
⨏
gΔ := g dσ ∈ Rn . (11.5.9)
Δ

Going further, bring in the modified harmonic double layer operators Dmod ±

(associated with Ω± and the Laplacian as in (11.4.26)), mapping functions from

 
L 1 ∂Ω , 1+σ(x) ∞
|x | n into functions in C (Ω± ), and define

u+ (x) := Dmod
+ f (x) − x, g  for each x ∈ Ω ,
Δ +
(11.5.10)
u− (x) := Dmod
− f (x) + x, g  for each x ∈ Ω .
Δ −

Also, fix an aperture parameter κ ∈ (0, ∞). Then, with Kmod denoting the modified
principal value version of the harmonic double layer on ∂Ω (associated with Ω and
the Laplacian as in (11.4.27)), we may then write
1     + κ−n.t.   − κ−n.t. 
f = 2I + Kmod f + 12 I − Kmod f = Dmod f ∂Ω + Dmod f ∂Ω
 κ−n.t.   κ−n.t. 
= u+ ∂Ω + u− ∂Ω , (11.5.11)

thanks to (11.4.36) and (11.5.10). As such,

⨏ 1/p
| f − fΔ | p dσ
Δ
#⨏ $ 1/p
  κ−n.t.  ⨏  κ−n.t.   − κ−n.t. 
⨏
 − κ−n.t.   p
 +
=  u ∂Ω − u ∂Ω dσ + u ∂Ω −
+
u ∂Ω dσ  dσ
Δ Δ Δ
#⨏ $ 1/p
  κ−n.t.  ⨏  κ−n.t.   p
 + + 
≤  u ∂Ω − u ∂Ω dσ  dσ
Δ Δ
#⨏ $ 1/p
  κ−n.t.  ⨏  κ−n.t.   p
 − − 
+  u ∂Ω − u ∂Ω dσ  dσ . (11.5.12)
Δ Δ

Let xΔ± ∈ Ω± be John centers relative to Δ(xo, R) (cf. [133, Definition 5.11.7]; see
also (A.0.54)) and note that
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 739
#⨏ ⨏ $ 1/p
  κ−n.t.   ± κ−n.t.   p
 ± 
 u ∂Ω (x) − u ∂Ω dσ  dσ(x)
Δ Δ
#⨏ ⨏ $ 1/p
  κ−n.t.   κ−n.t.  p
 ± ± 
=  u ∂Ω
(x) − u ∂Ω
(y) dσ(y) dσ(x)
Δ Δ
#⨏ ⨏ $ 1/p
  κ−n.t.   ± κ−n.t.   p
 ± 
≤  u ∂Ω (x) − u ∂Ω (y) dσ(y) dσ(x)
Δ Δ
#⨏ ⨏ $ 1/p
  κ−n.t.  p
 ± ± ± 
≤  u ∂Ω (x) − u (xΔ ) dσ(y) dσ(x)
Δ Δ
#⨏ ⨏ $ 1/p
  κ−n.t.  p
 ± ± ± 
+  u ∂Ω (y) − u (xΔ ) dσ(y) dσ(x)
Δ Δ
#⨏ $ 1/p
  κ−n.t.  p
 ± ± ± 
=2  u ∂Ω − u (xΔ ) dσ , (11.5.13)
Δ

by Hölder’s inequality. Thus, matters are reduced to estimating

#⨏ $ 1/p #⨏ $ 1/p
  κ−n.t.  p   κ−n.t.  p
 + + +   − − − 
 u ∂Ω − u (xΔ ) dσ ,  u ∂Ω − u (xΔ ) dσ .
Δ Δ
(11.5.14)
Let us deal with the first expression above. Recall from [133, Definition 5.11.7] (cf.
also (A.0.54)) that for each point x ∈ Δ there exists a rectifiable path γx joining x with
xΔ+ , of length ≤ M · R. In addition, there exists some geometric constant κ ∈ (0, ∞),
independent of x, with the property that γx is contained in the nontangential approach
region for Ω+ with vertex at x ∈ ∂Ω and aperture κ, i.e., γx ⊆ Γκ+ (x). Let L stand
for the length of γx , and denote by [0, L]  s → γx (s) ∈ Γκ+ (x) the arc-length
parametrization of γx . Then since for each s ∈ [0, L] we have

|γx (s) − xo | ≤ |γx (s) − x| + |x − xo | ≤ L + R ≤ (M + 1)R, (11.5.15)

hence γx (s) ∈ B(xo, (M + 1)R) for each s ∈ [0, L], we may employ the Fundamental
Theorem of Calculus to estimate
740 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
#⨏ $ 1/p
  κ−n.t.  p
 + + + 
 u ∂Ω (x) − u (xΔ ) dσ(x)
Δ
#⨏ ∫ $ 1/p
 L   p
 . +
=  γx (s), ∇u γx (s) ds dσ(x)
Δ 0
#⨏ $ 1/p
    p
 +
≤ CR Nκ |∇u |1B(xo,(M+1)R)∩Ω  dσ . (11.5.16)
Δ

On the other hand, for each x ∈ Ω we may use (11.4.28) (bearing in mind (11.5.6))
and (11.5.10) to write

∇u+ (x) = ∇Dmod

+
f (x) − gΔ = a(x) + b(x), (11.5.17)

where
∫
a(x) := − (∂k EΔ )(x − y)(∂τ j k f )(y)14MΔ (y) dσ(y) (11.5.18)
∂Ω 1≤ j ≤n

and

b(x) := g(x) − gΔ (11.5.19)

∫
=− (∂k EΔ )(x − y)(∂τ j k f )(y)1∂Ω\4MΔ (y) dσ(y)
∂Ω
⨏ ∫
− (∂k EΔ )(z − y)(∂τ j k f )(y)1∂Ω\4MΔ (y) dσ(y) dσ(z) .
Δ ∂Ω 1≤ j ≤n

Observe that
#⨏ $ 1/p #∫ $ 1/p
 p 1  
Nκ (a) dσ ≤ Nκ (a) p dσ (11.5.20)
Δ σ(Δ)1/p ∂Ω

#∫ $ 1/p
C 
n p
≤ |∂τ j k f |14MΔ dσ
σ(Δ)1/p ∂Ω j,k=1

⨏ 
n  1/p
≤C |∂τ j k f | p dσ ,
4MΔ j,k=1

thanks to the nontangential maximal function estimates on singular integrals from

[135, Chapter 2] (this step does not require knowing a priori that each function
(∂τ j k f ) · 1∂Ω\4MΔ belongs to L p (∂Ω, σ), since otherwise (11.5.20) is trivially true),
and the Ahlfors regularity of ∂Ω. Moreover, for each x ∈ Ω we have
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 741
⨏ ∫  
 
| b(x)| ≤ (∇EΔ )(x − y) − (∇EΔ )(z − y) × (11.5.21)
Δ ∂Ω\4MΔ


n
× |∂τ j k f (y)| dσ(y) dσ(z),
j,k=1

so using the Mean Value Theorem we obtain

∫ 
n
R
sup | b(x)| ≤ C |∂τ j k f (y)| dσ(y).
x ∈B(x o ,2M R)∩Ω ∂Ω\B(x o ,4M R) |xo − y| n j,k=1
(11.5.22)

In turn, this readily implies

#⨏ $ 1/p
    p

Nκ b1B(xo,2M R)∩Ω  dσ
Δ
∫ 
n
R
≤C |∂τ j k f (y)| dσ(y). (11.5.23)
∂Ω\B(x o ,4M R) |xo − y| n j,k=1

At this stage, from (11.5.16), (11.5.17), (11.5.20), and (11.5.23) we obtain

#⨏ $ 1/p
  κ−n.t.  p
 + 
 u ∂Ω − u+ (xΔ+ ) dσ (11.5.24)
Δ

⨏ 
n  1/p
≤ CR |∂τ j k f | p dσ
4MΔ j,k=1

∫ 
n
R2
+C |∂τ j k f (y)| dσ(y).
∂Ω\B(x o ,4M R) |xo − y| n j,k=1

In a completely similar fashion,

#⨏ $ 1/p
  κ−n.t.  p
 − − − 
 u ∂Ω − u (xΔ ) dσ (11.5.25)
Δ

⨏ 
n  1/p
≤ CR |∂τ j k f | p dσ
4MΔ j,k=1

∫ 
n
R2
+C |∂τ j k f (y)| dσ(y),
∂Ω\B(x o ,4M R) |xo − y| n j,k=1
742 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

since the difference in sign (for u− compared with u+ in (11.5.10)) is accounted

for by the fact that the geometric measure theoretic outward unit normal to Ω− is
the opposite of that for Ω+ (cf. [133, Lemma 5.10.9]). Combining (11.5.12) with
(11.5.24)-(11.5.25) proves (11.5.5) for any function f as in (11.5.6).
Next we shall establish (11.5.5) in the larger class of functions satisfying (compare
with (11.5.1))
σ(x)
f ∈ L 1 ∂Ω , and such that
1 + |x| n+1
(11.5.26)
σ(x)
∂τ j k f ∈ L 1 ∂Ω , for all j, k ∈ {1, . . . , n}.
1 + |x| n
This builds on what we have proved so far via a truncation argument using a cutoff
function, of the following sort. Start with some φ : Rn → R satisfying

φ ∈ C ∞ (Rn ), 0 ≤ φ ≤ 1, φ ≡ 1 on B(0, 1),

(11.5.27)
as well as φ ≡ 0 on Rn \ B(0, 2),

and for each ρ ∈ (0, ∞) set

x − xo
φρ (y) := φ for all y ∈ Rn . (11.5.28)
ρ
1 (∂Ω, σ). If we define
Let f be as in (11.5.26). In particular, f belongs to L1,loc

fρ := φρ · f on ∂Ω, (11.5.29)

then for each ρ ∈ (0, ∞) we have (thanks to (11.5.27)-(11.5.29) and Proposi-

tion 11.1.15)

fρ ∈ L11 (∂Ω, σ), fρ = f on B(x, ρ) ∩ ∂Ω, and

    (11.5.30)
∂τ j k fρ = ∂τ j k φρ · f + φρ · ∂τ j k f for 1 ≤ j, k ≤ n,

hence for each j, k ∈ {1, . . . , n} we have

C  
|∂τ j k fρ (y)| ≤ 1ρ ≤ |y−xo | ≤2ρ | f (y)| + ∂τ j k f (y) (11.5.31)
ρ
| f (y)|  
≤ C1ρ ≤ |y−xo | ≤2ρ + ∂τ j k f (y) for σ-a.e. y ∈ ∂Ω.
|y − xo |
Assume ρ > 4M R and write (11.5.5) for fρ (which satisfies all properties stipulated
in (11.5.6)). On account of (11.5.30)-(11.5.31) we obtain
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 743
⨏ 1/p ⨏ 
n  1/p
| f − fΔ | p dσ ≤ CR |∂τ j k f | p dσ
Δ 4MΔ j,k=1

∫ 
n
R2
+C |∂τ j k f (y)| dσ(y)
∂Ω\B(x o ,4M R) |xo − y| n j,k=1
∫
R2
+C 1ρ ≤ |y−xo | ≤2ρ | f (y)| dσ(y).
∂Ω\B(x o ,4M R) |xo − y| n+1
(11.5.32)

The second property in (11.5.26) ensures that we may rely on Lebesgue’s Dominated
Convergence Theorem to pass to limit ρ → ∞ in (11.5.32) and ultimately obtain
that (11.5.5) holds for all functions f as in (11.5.26).
In summary, (11.5.5) is valid for any function f satisfying (11.5.26). In the case
p
when f is as in (11.5.1), we also have ∂τ j k f ∈ Lloc (∂Ω, σ) for 1 ≤ j, k ≤ n which,
p
together with (11.5.5), allows us to conclude that f ∈ L loc (∂Ω, σ) hence, further,
p
f ∈ L 1,loc (∂Ω, σ). In turn, this membership guarantees the validity of (11.4.9).
Keeping this in mind, we conclude from (11.5.5) that for each function f as in
(11.5.1) we have
⨏ 1/p ⨏  1/p
| f − fΔ | p dσ ≤ CR |∇tan f | p dσ
Δ 4MΔ
∫
R2
+C |∇tan f (y)| dσ(y)
∂Ω\B(x o ,4M R) |xo − y|
n

⨏  1/p
≤ CR |∇tan f | p dσ
4MΔ

∞ ∫
2−j
+ CR |∇tan f | dσ.
j=1
σ(2 j 4MΔ) 2 j 4MΔ\2 j−1 4MΔ

(11.5.33)

After adjusting notation, this establishes (11.5.2), so the proof of Proposition 11.5.1
is complete. 

The Poincaré inequality from Proposition 11.5.1 permits us to establish the result
stated in Theorem 11.5.2 below, which may be regarded as a quantitative version of
[133, Proposition 5.6.10]. Indeed, at each fixed scale we are now able to obtain a
quantitative control of the inner product between the average of unit normal and the
(normalized) chord in terms of the corresponding local mean oscillations of the unit
normal. Theorem 11.5.2 refines and corrects the result proved in [97, Theorem 4.14,
p. 2697]. Before stating it, the reader is advised to recall the piece of notation
introduced in (A.0.7).
744 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Here we also want to mention that a companion result to Theorem 11.5.2, valid
for arbitrary Ahlfors regular domains with compact boundary, is obtained in [137,
§3.1].

Theorem 11.5.2 Assume Ω ⊂ Rn is an open set satisfying a two-sided local John

condition and whose boundary is an Ahlfors regular set (hence, any two-sided NTA
domain with Ahlfors regular boundary will do). Abbreviate σ := H n−1 ∂Ω and
denote by ν the geometric measure theoretic outward unit normal to Ω.
Then there exist a threshold R ∈ (0, ∞] (which may be taken ∞ if ∂Ω is unbounded)
and two numbers, C ∈ [1, ∞) and λ ∈ (32, ∞), all of which depend only on the local
John constants of Ω and the Ahlfors regularity constants of ∂Ω, with the property
that for each point z ∈ ∂Ω, each scale R ∈ (0, R), each threshold ε ∈ (0, 1), and
each dilation factor γ ∈ [1, ∞) one has

sup R−1 |x − y, νΔ(z,R) | ≤ Cγ 1 + log2 γ)ν∗ (Δ(z, γλε −1 R)) + Cγε,
x,y ∈Δ(z,γR)
(11.5.34)
⨏
where νΔ(x,R) := Δ(x,R)
ν dσ. Moreover, for each dilation factor γ ∈ [1, ∞) one has
  
lim sup sup sup R−1  x − y, νΔ(z,R)  
R→0+ z ∈∂Ω x,y ∈Δ(z,γR)
    n
≤ Cγ 1 + log2 γ) · dist ν , VMO(∂Ω, σ) , (11.5.35)
 n
where the distance in the right-hand side is considered in the space BMO(∂Ω, σ) .

An estimate in the spirit of (11.5.34) (see also [137, §3.1.6] for related matters)
is the main result in [173] and Semmes establishes this in the case when ∂Ω is a
C ∞ smooth surface using his Poincaré type inequality ([173, Lemma 1.1, p. 406]).
The necessity that ∂Ω is smooth in Semmes’ argument is inherited from here (see
the comments following the statement of Proposition 11.5.1). In the more general
geometric setting considered here we shall follow [97] and employ the version of
Poincaré inequality from Proposition 11.5.1.
Proof of Theorem 11.5.2 Let R ∈ (0, ∞] be such that the Poincaré inequality from
Proposition 11.5.1 applies on each surface ball of radius ≤ C R, where C ∈ (1, ∞)
is a suitably large geometric constant. To justify (11.5.34), fix x ∈ ∂Ω, R ∈ (0, R),
ε ∈ (0, 1), and abbreviate Δ := Δ(x, R). Consider next

f (y) := x − y, νΔ  for all y ∈ ∂Ω. (11.5.36)

We claim that for each exponent α ∈ (0, 1) there exist C = C(Ω, n, α) ∈ (0, ∞) and
some geometric constant λ ∈ (2, ∞) with the property that

| f (y) − f (y )| ≤ CR1−α |y − y  | α ν∗ (Δ(x, 8λε −1 M R)) + ε , ∀ y, y  ∈ Δ.
(11.5.37)
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 745

Granted this, choosing y  := x yields

| f (y)| ≤ CRν∗ (Δ(x, 8λε −1 M R)) + CR ε, ∀ y ∈ Δ, (11.5.38)

which, after adjusting notation, readily gives (now with λ ∈ (16, ∞))
 
sup R−1  x − y, νΔ(x,R)   ≤ Cν∗ (Δ(x, λε −1 R)) + Cε. (11.5.39)
y ∈Δ(x,2R)

We may now use this to prove (11.5.34) as follows. Fix γ ∈ [1, ∞) along with z ∈ ∂Ω,
R ∈ (0, R/γ), and x, y ∈ Δ(z, γR). Also, pick an arbitrary ε ∈ (0, 1). Then

y ∈ Δ(x, 2γR), Δ(x, 2γR) ⊆ Δ(z, 3γR),

(11.5.40)
Δ(x, γλε −1 R) ⊆ Δ(z, γ(λε −1 + 1)R),

so there exists some C ∈ (0, ∞), which depends only on the local John constants of
Ω and the Ahlfors regularity constants of ∂Ω, such that
 
 x − y, νΔ(z,R)   ≤ |x − y, νΔ(x,γR) | + |x − y||νΔ(x,γR) − νΔ(z,R) |

≤ CγRν∗ (Δ(x, γλε −1 R)) + CγRε

+ 2γR|νΔ(x,2γR) − νΔ(z,3γR) | + 2γR|νΔ(z,3γR) − νΔ(z,R) |

≤ CγRν∗ (Δ(z, γ(λε −1 + 1)R)) + CγRε


+ CγR 1 + log2 γ)ν∗ ((Δ(z, 3γR))

≤ CγR 1 + log2 γ)ν∗ (Δ(z, γ(λε −1 + 1)R)) + CγRε, (11.5.41)

thanks to (11.5.39) (written with γR in place of R), (11.5.40), [133, (7.4.56),

(7.4.63)], and the fact that σ is a doubling measure. Work with the most extreme sides
of (11.5.41). Dividing by R, and taking the supremum with respect to x, y ∈ Δ(z, γR)
and z ∈ ∂Ω establishes (11.5.34) (after estimating λε −1 + 1 ≤ 2λε −1 and relabeling
2λ as λ).
At this stage, there remains to prove (11.5.37). To this end, first note that
 %
f = ϕ∂Ω ∈ 1≤p ≤∞ L 1,loc (∂Ω, σ),
p
(11.5.42)
with ϕ ∈ C ∞ (Rn ) given by ϕ(y) := x − y, νΔ  for all y ∈ Rn .

Keeping this in mind, from (11.4.2) we deduce that, at σ-a.e. point on ∂Ω,
   
∇tan f = ∇ϕ ∂Ω − ν · ∇ϕ ∂Ω ν

= −νΔ + νΔ, νν

= −νΔ + ν + νΔ − ν, νν, (11.5.43)
746 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

since |ν| 2 = 1 at σ-a.e. point on ∂Ω. Consequently,

|∇tan f (y)| = |νΔ − ν(y) − νΔ − ν(y), ν(y)ν(y)|

≤ 2|νΔ − ν(y)| for σ-a.e. point y ∈ ∂Ω. (11.5.44)

The idea is to apply the Poincaré inequality (11.5.2) to the function f which, thanks
to [133, Lemma 7.2.1] (cf. [133, (7.2.5)]) and (11.5.42), satisfies the properties
demanded
 in (11.5.1).
 To set the stage, fix an arbitrary surface ball Δr of radius
r ∈ 0, 2 diam(∂Ω) such that Δr ⊆ Δ. In view of the Ahlfors regularity of ∂Ω this
entails r ≤ CR, for some geometric constant C ∈ (0, ∞). Then (11.5.2) implies
⨏ 1/p
⨏ 1/p
1
| f − fΔr | dσ
p
≤C |∇tan f | p dσ
r Δr MΔr

∞ ⨏
+C 2−j |∇tan f | dσ
j=1 2 j MΔr


∞ ⨏ 1/p
≤C 2−j |∇tan f | p dσ . (11.5.45)
j=1 2 j MΔr

To proceed, set N := [−log2 ε] + 1. In particular,

2 N ∼ ε −1 and 2−N ∼ ε. (11.5.46)

Then (11.5.45), (11.5.44), the Ahlfors regularity of ∂Ω, the fact that we have used
the abbreviation Δ := Δ(x, R), [133, (7.4.121)], John-Nirenberg’s inequality from
[133, Lemma 7.4.10], the property that |ν| = 1 at σ-a.e. point on ∂Ω, and (11.5.46)
yield
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 747
⨏ 1/p
1
| f − fΔr | p dσ
r Δr


∞ ⨏ 1/p
≤C 2−j |ν − νΔ | p dσ
j=1 2 j MΔr


∞ ∫ 1/p
j (1−n)/p −j
≤C (2 r) 2 |ν − νΔ | p dσ
j=1 2 j MΔ

(n−1)/p 
∞ ⨏ 1/p
R
≤C 2−j |ν − νΔ | p dσ
r j=1 2 j MΔ

(n−1)/p  
N ⨏ 1/p
R
≤C 2−j |ν − νΔ | p dσ + 2−N
r j=1 2 j MΔ

R (n−1)/p  
∞
≤C j 2−j ν∗ (Δ(x, 2 N +3 M R)) + 2−N
r j=1

R (n−1)/p 
≤C ν∗ (Δ(x, 8ε −1 M R)) + ε . (11.5.47)
r
Given α ∈ (0, 1), if p ∈ (n − 1, ∞) is chosen such that α = 1 − (n − 1)/p, then the
above estimate gives
⨏ 1/p 
r −α | f − fΔr | p dσ ≤ CR1−α ν∗ (Δ(x, 8ε −1 M R)) + ε (11.5.48)
Δr

for some C = C(Ω, n, α) ∈ (0, ∞). Bearing in mind [133, Lemma 3.6.4], the criterion
for (local) Hölder continuity from [133, Proposition 7.4.9] then gives that, for some
geometric constant λ ∈ (2, ∞), we have
⨏
| f (y) − f (y )| −α
1/p
sup ≤ C sup r | f − fΔ | p
dσ
y,y  ∈Δ |y − y  | α Δr ⊆λΔ Δr
r

yy 

≤ CR1−α ν∗ (Δ(x, 8λε −1 M R)) + ε . (11.5.49)

Estimate (11.5.49) justifies (11.5.37), and finishes the proof of (11.5.34).

To finish the proof of the theorem, there remains to show (11.5.35). Working with
(11.5.34), take the supremum with respect to z ∈ ∂Ω, send R → 0+ , invoke (3.1.55)
and, finally, send ε → 0+ . Doing so establishes (11.5.35). 

Moving on, given an open set with an Ahlfors regular boundary, we introduce the
L p -based homogeneous Sobolev space of order one on the boundaries of Euclidean
sets of a very general geometric nature (as indicated in Definition 11.5.3, given
below). Since this is going to play a major role in subsequent considerations, we find
748 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

it appropriate to elaborate on the genesis of this brand of Sobolev space. The starting
point is to note that, in the context of the entire (n − 1)-dimensional Euclidean space
(regarded as the boundary of the upper half-space in Rn ), a natural definition is to
consider, for each p ∈ (1, ∞),
.p 
L1 (Rn−1, L n−1 ) := f ∈ Lloc1
(Rn−1, L n−1 ) : ∂j f ∈ L p (Rn−1, L n−1 )
for each j ∈ {1, . . . , n − 1} .
(11.5.50)
.p
In this regard, we wish to remark that any function in L1 (Rn−1, L n−1 ) automatically
enjoys a global integrability property, stemming from the inclusion

.p L n−1 (x )
L1 (Rn−1, L n−1 ) ⊆ L 1 Rn−1, . (11.5.51)
1 + |x  | n
.p 1,1 n−1
To justify (11.5.51), observe that, by design, L1 (Rn−1, L n−1 ) ⊆ Wloc (R ). As such,
. p n−1 n−1
for any function f ∈ L1 (R , L ) Poincaré’s inequality (cf., e.g., [49, Theorem 2,
p. 141]) implies that for any λ ∈ (0, ∞) we have
⨏ ⨏
  1/p
 f − fB (0,λ)  dL n−1 ≤ Cn λ |∇  f | p dL n−1
n−1
B n−1 (0,λ) B n−1 (0,λ)

≤ Cn λ1−(n−1)/p ∇  f [L p (Rn−1, L n−1 )]n−1 ,

(11.5.52)
⨏
where fBn−1 (0,λ) := B n−1 (0,λ)
f dL n−1 and ∇  denotes the gradient in Rn−1 . In
concert with this, the first inequality in [133, (7.4.115)] (used here with X := Rn−1 ,
μ := L n−1 , ρ := | · − · |, d := n − 1, x0 := 0 ∈ Rn−1 , ε := 1, p := 1, q := 1, and
r := 1) presently gives
∫  
f (x ) − fBn−1 (0,1) 
dL n−1 (x )
1 + |x  | n
R n−1
∫ ∞
 dλ
≤ Cn ∇ f [L p (Rn−1, L n−1 )]n−1 λ1−(n−1)/p
1 λ2

≤ Cn, p ∇ f [L p (Rn−1, L n−1 )]n−1 , (11.5.53)

for some constant Cn, p ∈ (0, ∞). The inclusion claimed in (11.5.51) now readily
follows from (11.5.53).
In light of (11.5.51), the definition given in (11.5.50) self-improves to
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 749
!
.p L n−1 (x )
L1 (Rn−1, L n−1 ) = f ∈ L 1 Rn−1, : ∂j f ∈ L p (Rn−1, L n−1 )
1 + |x  | n
"
for each j ∈ {1, . . . , n − 1} .

(11.5.54)

This is the point of view we shall adopt in the definition below, since we shall consider
geometric settings so general that no reasonable version of Poincaré’s inequality may
be available.

Definition 11.5.3 Let Ω ⊆ Rn be an open set with an Ahlfors regular boundary and
abbreviate σ := H n−1 ∂Ω. For each given integrability exponent p ∈ (1, ∞), define
the L p -based homogeneous Sobolev space of order one on ∂Ω as
!
.p σ(x)
L1 (∂Ω, σ) := f ∈ L 1 ∂Ω, : ∂τ j k f ∈ L p (∂∗ Ω, σ)
1 + |x| n
"
for each j, k ∈ {1, . . . , n} (11.5.55)

and equip this space with the semi-norm4

.p 
n
L1 (∂Ω, σ)  f →  f  L. p (∂Ω,σ) := ∂τ j k f  L p (∂∗ Ω,σ) . (11.5.56)
1
j,k=1

.p
This definition is natural since, as seen from the earlier discussion, L1 (∂Ω, σ)
.p
becomes precisely the space L1 (Rn−1, L n−1 ) recalled in (11.5.50) in the situation
when Ω := R+n .
Let us also note
. p that, in the context of Definition 11.5.3, all constant functions
on ∂Ω belong to L1 (∂Ω, σ) and their respective semi-norms vanish. It is also clear
from [133, Lemma 7.2.1] that we have the continuous embeddings

p .p σ(x)
L1 (∂Ω, σ) → L1 (∂Ω, σ) ∩ L 1 ∂Ω, (11.5.57)
1 + |x| n−1
and, more generally,

q, p .p σ(x)
L1 (∂Ω, σ) → L1 (∂Ω, σ) ∩ L 1 ∂Ω, for each q ∈ [1, ∞). (11.5.58)
1 + |x| n−1
We augment this list of embeddings with the following result, valid in a more
restrictive geometric setting.

4 It may happen that, say, ∂∗ Ω =  in which case the semi-norm (11.5.56) is identically zero. We
will, however, typically work in the setting of Ahlfors regular domains where ∂∗ Ω and ∂Ω coincide
up to a σ-nullset.
750 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Lemma 11.5.4 Let Ω ⊆ Rn be an open set satisfying a two-sided local John condi-
tion, whose boundary is an Ahlfors regular set. Abbreviate σ := H n−1 ∂Ω. and fix
some integrability exponent p ∈ (1, ∞). Then
.p p p
L1 (∂Ω, σ) ⊆ L1,loc (∂Ω, σ) ⊆ Lloc (∂Ω, σ). (11.5.59)

In particular,
.p p
L1 (∂Ω, σ) = L1 (∂Ω, σ) as sets, if ∂Ω is bounded, (11.5.60)

Proof The first inclusion claimed in (11.5.59) is a consequence of Definition 11.5.3

and the first conclusion in Proposition 11.5.1. The second inclusion in (11.5.59)
is clear from item (iii) in Definition 11.1.2. Finally, (11.5.60) is immediate from
(11.5.59) and (11.5.57). 

In turn, Lemma 11.5.4 allows us to re-phrase Proposition 11.4.3 in terms of

homogeneous Sobolev spaces.

Corollary 11.5.5 Suppose Ω ⊆ Rn (where n ∈ N with n ≥ 2) is an open set satisfying

.p
a two-sided local John condition and with an Ahlfors regular boundary. Define
σ := H n−1 ∂Ω and fix some p ∈ (1, ∞). Then for each function f ∈ L1 (∂Ω, σ) one
has the following equivalences:

f is locally constant on ∂Ω ⇐⇒ ∇tan f = 0 at σ-a.e. point on ∂Ω

⇐⇒ for all j, k ∈ {1, . . . , n} one has ∂τ j k f = 0 at σ-a.e. point on ∂Ω (11.5.61)
⇐⇒  f  L. p (∂Ω,σ) = 0.
1

Proof This is a direct consequence of Proposition 11.4.3, Lemma 11.5.4, and Defi-
nition 11.5.3. 

Following a series of preliminary results (cf. Lemma 11.5.6-Lemma 11.5.10

below), in Propositions 11.5.12-11.5.14 and Theorems 11.5.16-11.5.17 we study
basic properties of the scale of homogeneous boundary Sobolev spaces introduced
in Definition 11.5.3. Here is the first such preliminary result.

Lemma 11.5.6 Let Σ ⊆ Rn be a closed, upper Ahlfors regular set. Abbreviate

σ := H n−1 Σ, pick p ∈ [1, ∞) along with α ∈ R satisfying α < 1 + (n − 1)/p, and
fix an arbitrary reference point x0 ∈ Σ. Also, suppose f ∈ Lloc
1 (Σ, σ) is such that

∫ 1/p
1
C∗ := sup | f − fΔ(x0,r) | p dσ < +∞, (11.5.62)
r >0 rα Δ(x0,r)

where, for each r ∈ (0, ∞),

⨏
Δ(x0, r) := Σ ∩ B(x0, r) and fΔ(x0,r) := f dσ. (11.5.63)
Δ(x0,r)
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 751

Then there exists some constant C = C(Σ, n, p) ∈ (0, ∞) with the property that for
each r ∈ (0, ∞) one has
∫
| f (x) − fΔ(x0,r) | C · C∗
dσ(x) ≤ 1−α+(n−1)/p . (11.5.64)
Σ (r + |x − x0 |) n r
 
In particular, f belongs to the space L 1 Σ, 1+σ(x)
|x | n .

Proof The estimate claimed in (11.5.64) is a direct consequence of (11.5.62), the up-
per Ahlfors regularity of Σ, and the very first inequality in [133, (7.4.115)] (presently
used with X := ∂Ω, μ := σ, d := n − 1, ε := 1, p := 1, and q := p). 

Our second preliminary result reads as follows:

Lemma 11.5.7 Let Σ ⊆ Rn be a closed upper Ahlfors regular set and abbreviate
σ := H n−1 Σ. Assume f is a complex-valued σ-measurable function on Σ with the
property that there exists some non-negative function g ∈ L p (Σ, σ) with p ∈ [1, ∞)
and some σ-measurable set N ⊆ Σ with σ(N) = 0 such that
 
| f (x) − f (y)| ≤ |x − y| · g(x) + g(y) for all x, y ∈ Σ \ N. (11.5.65)

Then
 
f ∈ L 1 Σ, 1+σ(x)
p
|x | n ∩ Lloc (Σ, σ) (11.5.66)

and there exists some constant C = C(Σ, p) ∈ (0, ∞) with the property that for each
reference point x0 ∈ Σ and each radius r ∈ (0, ∞) one has
∫  
f (x) − fΔ(x0,r)  C
  n dσ(x) ≤ (n−1)/p g L p (Σ,σ), (11.5.67)
Σ r + |x − x0 | r
⨏
where Δ(x0, r) := Σ ∩ B(x0, r) and fΔ(x0,r) := Δ(x0,r)
f dσ.
p
Proof That f ∈ Lloc (Σ, σ) is clear from (11.5.65). To justify (11.5.67), fix a reference
point x0 ∈ Σ. Then for each r ∈ (0, ∞) we may rely on Hölder’s inequality and
(11.5.65) to estimate
752 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ 1/p
1
| f (x) − fΔ(x0,r) | p dσ(x)
r Δ(x0,r)
∫  ⨏ p
1   1/p
=  f (x) − f (y) dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏ p 1/p
1
≤ | f (x) − f (y)| dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏ 1/p
1
≤ | f (x) − f (y)| p dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏
1  p 1/p
≤ |x − y| p · g(x) + g(y) dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏
  1/p
≤4 g(x) p + g(y) p dσ(y) dσ(x)
Δ(x0,r) Δ(x0,r)

≤ 8g L p (Σ,σ) . (11.5.68)

Hence,
∫ 1/p
1
C∗ := sup | f − fΔ(x0,r) | p dσ ≤ 8g L p (Σ,σ) < +∞, (11.5.69)
r >0 r Δ(x0,r)

so (11.5.67) is implied by (11.5.69) and (11.5.64) with α := 1. Finally, the fact

 
that the function f belongs to the space L 1 Σ, 1+σ(x)
|x | n is a direct consequence of
(11.5.67). 

The third preliminary result makes the object of our next lemma.

Lemma 11.5.8 Let Ω ⊆ Rn be an open set satisfying a local John condition and
whose boundary is Ahlfors regular. Abbreviate σ := H n−1 ∂Ω and fix some aperture
parameter κ ∈ (0, ∞).
Then there exists C ∈ (0, ∞) with the property that for each function u ∈ C 1 (Ω)
 κ−n.t. 
for which the nontangential trace u∂Ω (x) exists at σ-a.e. point x ∈ ∂Ω one has
  κ−n.t.   κ−n.t.    
 
 u ∂Ω (x) − u∂Ω (y) ≤ C|x − y| · g(x) + g(y)
(11.5.70)
for σ-a.e. points x, y ∈ ∂Ω,

where
⎧
⎨ Nκ (∇u) if ∂Ω is unbounded,
⎪
⎪
g :=  κ−n.t.  (11.5.71)
⎪
⎪ u∂Ω  + Nκ (∇u) if ∂Ω is bounded.
⎩
As a consequence of (11.5.70), Lemma 11.5.7, [133, (8.2.28)], and [133, Corol-
lary 8.9.6], for each integrability exponent p ∈ [1, ∞) there exists some constant
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 753

C = C(Ω, p) ∈ (0, ∞) with the property that for each reference point x0 ∈ ∂Ω and
each radius r ∈ (0, ∞) one has
∫  
f (x) − fΔ(x0,r)  C
  n dσ(x) ≤ (n−1)/p g L p (∂Ω,σ),
∂Ω r + |x − x0 | r
κ−n.t.
where f := u∂Ω , the function g is as in (11.5.71), (11.5.72)
⨏
Δ(x0, r) := ∂Ω ∩ B(x0, r), and fΔ(x0,r) := f dσ.
Δ(x0,r)

Also, (11.5.66) implies that, in a quantitative fashion,

if g from (11.5.71) belongs to L p (∂Ω, σ) for some p ∈ [1, ∞)

κ−n.t. σ(x) (11.5.73)
then u∂Ω ∈ L 1 ∂Ω,
p
∩ Lloc (∂Ω, σ).
1 + |x| n
Proof Fix an arbitrary point x∗ ∈ ∂Ω. The interior local John condition ensures
that there exist C ∈ (1, ∞) and R ∈ (0, ∞) (in fact, we may take R = ∞ if ∂Ω
is unbounded) such that for any scale r ∈ (0, R) one can find a (corkscrew) point
zr ∈ B(x∗, Cr) ∩ Ω which may be joined with any point ξ ∈ B(x∗, r) ∩ ∂Ω via a
rectifiable curve γξ of length ≈ r, and contained in a nontangential approach region
Γκ (ξ), for some κ ∈ (0, ∞) independent of x∗ and r. See [133, Definition 5.11.7]
(cf. also (A.0.54)) which shows that we may take κ := θ −1 − 1. In turn, for σ-a.e.
points x, y ∈ B(x∗, R/2) ∩ ∂Ω, this geometric condition allows us to estimate (with
r := |x − y|)
  κ−n.t.   κ−n.t.  
 
 u ∂Ω (x) − u∂Ω (y)
  κ−n.t.     κ−n.t.  
   
≤  u∂Ω (x) − u(zr ) +  u∂Ω (y) − u(zr )

∫ d
1    1 d 
∫
 

= u(γx (t)) dt  +  u(γy (t)) dt 
0 dt 0 dt
∫ 1 ∫ 1
   
≤ (∇u)(γx (t)) | γ x (t)| dt + (∇u)(γy (t)) | γ y (t)| dt
0 0

≤ length(γx ) · Nκ (∇u)(x) + length(γy ) · Nκ (∇u)(y)

 
≤ Cr Nκ (∇u)(x) + N (∇u)(y)
 
= C|x − y| · Nκ (∇u)(x) + Nκ (∇u)(y) . (11.5.74)

The claim made in (11.5.70)-(11.5.71) now readily follows from (11.5.74). 

Our next preliminary result is discussed in the lemma below.
754 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Lemma 11.5.9 Suppose Σ ⊆ Rn is a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Assume f is a complex-valued σ-measurable function on Σwith the
property that there exists some non-negative function g ∈ L p (Σ, σ) with p ∈ n−1
n ,∞
such that
 
| f (x) − f (y)| ≤ |x − y| · g(x) + g(y) for σ-a.e. x, y ∈ Σ. (11.5.75)

Then the following conclusions are valid:

 
(1) Whenever p ∈ n−1 n , n − 1 it follows that there exist two constants, c = c( f ) ∈ C
along with C = C(Σ, n, p) ∈ (0, ∞), such that the function f − c belongs to the
space L p∗ (Σ, σ), where
1 
1 −1
p∗ := p − n−1 ∈ (1, ∞), (11.5.76)

and
 f − c L p∗ (Σ,σ) ≤ Cg L p (Σ,σ) . (11.5.77)
As a consequence of (11.5.77), in this case one has
 
f ∈ L 1 Σ, 1+σ(x)
p∗
|x | n ∩ Lloc (Σ, σ). (11.5.78)

(2) In the special case when p = n − 1, it follows that

f belongs to the Sarason space VMO(Σ, σ) (hence, f also

(11.5.79)
belongs to the John-Nirenberg space BMO(Σ, σ))

and there exists a constant C = C(Σ, n) ∈ (0, ∞) independent of f such that (cf.
(A.0.8))
.
 f BMO(Σ,σ) ≤ Cg L n−1 (Σ,σ) . (11.5.80)
(3) In the case when p ∈ (n − 1, ∞), if

α := 1 − n−1
p ∈ (0, 1), (11.5.81)
.
it follows that, up to re-defining f on a σ-nullset in Σ, one has f ∈ C α (Σ) and

 f C.α (Σ) ≤ Cg L p (Σ,σ) (11.5.82)

for some constant C = C(Σ, n, p) ∈ (0, ∞) independent of f .

 
Proof To deal with item (1), suppose p ∈ n−1 n , n − 1 and fix a reference point
x0 ∈ ∂Ω. The assumption made in (11.5.75) permits us to invoke [80, Theorem 8.7,
p. 197; cf. (8.3)] (see also [80, Remark 8.8, p. 198; cf. (8.7)]), according to which
p
f ∈ Lloc∗ (Σ, σ) ⊆ Lloc
1
(Σ, σ) (11.5.83)

and there exists a constant C ∈ (0, ∞) independent of f such that for each radius
r ∈ 0, 2 diam Σ we have
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 755
⨏ 1/p∗
⨏ 1/p
| f − cr | p∗ dσ ≤ Cr g p dσ , (11.5.84)
Σ∩B(x0,r) Σ∩B(x0,2r)

where we have set

⨏
cr := f dσ for each r ∈ (0, ∞). (11.5.85)
Σ∩B(x0,r)

In the case when Σ is bounded, all desired conclusions follow readily from (11.5.83)-
(11.5.84) taking ⨏
c := f dσ. (11.5.86)
Σ
There remains to treat the case when Σ is unbounded. In relation to the entity
introduced in (11.5.85) we make the claim that the sequence {c j } j ∈N (consisting of
terms constructed as in (11.5.85) with r := j ∈ N) is convergent, i.e.,

c := lim c j exists in C. (11.5.87)

j→∞

This follows as soon as we show that {c j } j ∈N is a Cauchy sequence. A proof of

this fact which works in the case when p ∈ (1, n − 1) goes as follows. From [133,
Lemma 7.4.15] (cf. the very first estimate in [133, (7.4.128)]) we then see that
there exists some constant C = C(Σ, n, p) ∈ (0, ∞) with the property that whenever
0 < r < R < ∞ we may estimate (using Hölder’s inequality, (11.5.75), and the
Ahlfors regularity of Σ)
756 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

|cR − cr |
∫ 2R ⨏  ⨏ p
  1/p dt
≤C  f (x) − f (y) dσ(y) dσ(x)
r Σ∩B(x0, t) Σ∩B(x0, t) t
∫ 2R ⨏ ⨏ p 1/p dt
≤C | f (x) − f (y)| dσ(y) dσ(x)
r Σ∩B(x0, t) Σ∩B(x0, t) t
∫ 2R ⨏ ⨏ 1/p dt
≤C | f (x) − f (y)| p dσ(y) dσ(x)
r Σ∩B(x0, t) Σ∩B(x0, t) t
∫ 2R ⨏ ⨏
≤C |x − y| p ×
r Σ∩B(x0, t) Σ∩B(x0, t)

  1/p dt
× g(x) p + g(y) p dσ(y) dσ(x)
t
∫ ⨏ ⨏
2R   1/p dt
≤C t g(x) p + g(y) p dσ(y) dσ(x)
r Σ∩B(x0, t) Σ∩B(x0, t) t
∫ 2R ⨏ 1/p
≤C g p dσ dt
r Σ∩B(x0, t)
∫ 2R   −1/p
≤ Cg L p (Σ,σ) σ Σ ∩ B(x0, t) dt
r
∫ 2R
dt
≤ Cg L p (Σ,σ) . (11.5.88)
r t (n−1)/p
Given that (n − 1)/p > 1, we conclude from (11.5.88) that the sequence {c j } j ∈N is
indeed Cauchy, hence (11.5.87) holds if p ∈ (1, n − 1).
To prove (11.5.87) for arbitrary p ∈ n−1 n , n − 1 , assume 0 < r < R < ∞ and
write (based on the triangle inequality, the Ahlfors regularity of Σ, (11.5.84) used
twice, and the definition of p∗ )
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 757
⨏ 1/p∗
|cR − cr | = |cR − cr | p∗ dσ
Σ∩B(x0,r)
⨏ 1/p∗
⨏ 1/p∗
≤ | f − cR | p∗ dσ + | f − cr | p∗ dσ
Σ∩B(x0,r) Σ∩B(x0,r)

(n−1)/p∗
⨏ 1/p∗
R
≤C | f − cR | p∗ dσ
r Σ∩B(x0,R)
⨏ 1/p
+Cr g p dσ
Σ∩B(x0,2r)

(n−1)/p∗
⨏ 1/p
R
≤C R g p dσ
r Σ∩B(x0,2R)
⨏ 1/p
+Cr g p dσ
Σ∩B(x0,2r)
! "
R (n−1)/p∗
≤C R 1−(n−1)/p
+r 1−(n−1)/p
g L p (Σ,σ)
r

= C r −(n−1)/p∗ g L p (Σ,σ) . (11.5.89)

This readily implies that the sequence {c j } j ∈N (consisting of terms constructed as in

(11.5.85) with r := j ∈ N) is Cauchy, so (11.5.87) is established in full generality.
Going further, with c as in (11.5.87) we may then compute
∫ ∫
| f − c| dσ =
p∗
lim inf 1Σ∩B(x0, j) · | f − c j | p∗ dσ
Σ Σ j→∞
∫
≤ lim inf 1Σ∩B(x0, j) · | f − c j | p∗ dσ
j→∞ Σ
⨏
 
= lim inf σ Σ ∩ B(x0, j) | f − c j | p∗ dσ
j→∞ Σ∩B(x0, j)
⨏ p∗ /p
≤ C lim inf j n−1 · j p∗ g p dσ
j→∞ Σ∩B(x0,2j)
∫ p∗ /p
≤ C lim inf g p dσ
j→∞ Σ∩B(x0,2j)

p
= Cg L∗p (Σ,σ), (11.5.90)

thanks to (11.5.87), Fatou’s Lemma, (11.5.84), the Ahlfors regularity of Σ, and the
fact that we presently have 1/p∗ = 1/p − 1/(n − 1). This shows that the function
∗
f − c belongs to L p (Σ, σ) and that the estimate claimed in (11.5.77) holds. Finally,
758 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

(11.5.78) is justified by writing

 
f = ( f − c) + c ∈ L p∗ (Σ, σ) + C ⊆ L 1 Σ, 1+σ(x)
p∗
|x | n ∩ Lloc (Σ, σ), (11.5.91)

where the inclusion uses [133, (7.7.106)] together with the fact that p∗ ∈ (1, ∞), and
the observation that Σ  x → (1 + |x| n )−1 ∈ R is in L 1 (Σ, σ) (cf. [133, (7.2.5)]).
This finishes the treatment of item (1).
Turning to item (2), consider the case when p = n − 1. Having g = 0 forces f
to be constant, in which case all desired conclusions are trivially true. Henceforth,
assume g ∈ L n−1 (Σ, σ) is a non-negative function which is not zero σ-a.e. on Σ.
Then (11.5.75) implies that f ∈ Lloc1 (Σ, σ) and [80, Theorem 8.7, p. 197; cf. (8.4)]

ensures the existence of some small constant c ∈ (0, ∞) along with some large
constant C ∈ (0, ∞), both of which depend only on Σ, with the property that for each
surface ball Δ ⊆ Σ we have
⨏ ! "
c | f − fΔ |
exp dσ ≤ C (11.5.92)
Δ g L n−1 (Σ,σ)
⨏
where fΔ := Δ f dσ. Note that, trivially, for each surface ball Δ ⊆ Σ and each
λ ∈ (0, ∞) we have
 
c | f (x)− fΔ |
1 ≤ exp − g  cλ · exp g  L n−1 (Σ, σ)
L n−1 (Σ, σ) (11.5.93)
for every x ∈ Δ with | f (x) − fΔ | > λ.

This shows that (11.5.92) implies the following level set estimate with exponential
decay, for each surface ball Δ ⊆ Σ and each λ ∈ (0, ∞):
 
σ x ∈ Δ : | f (x) − fΔ | > λ
∫ ! "
 c | f − fΔ |
≤ exp − g  cλ exp dσ
n−1
L (Σ, σ) g L n−1 (Σ,σ)
Δ

≤ C · exp − cλ
g  L n−1 (Σ, σ) σ(Δ). (11.5.94)

After integrating in λ ∈ (0, ∞), this gives

⨏ ∫ ∞
1  
| f (x) − fΔ | dσ(x) = σ x ∈ Δ : | f (x) − fΔ | > λ dλ
Δ σ(Δ) 0
∫ ∞ 
≤C exp − g  cλ
n−1
dλ
0 L (Σ, σ)

= c−1 g L n−1 (Σ,σ), (11.5.95)

for each surface ball Δ ⊆ Σ. Then (11.5.80) follows from this and (A.0.8).
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 759

Another proof of (11.5.80) which also gives (11.5.79) proceeds

 as follows. Given
any surface ball Δ = Δ(z, r) ⊆ Σ, with z ∈ Σ and r ∈ 0, 2 diam Σ , we may use
(11.5.75), the lower Ahlfors regularity of Σ, and Hölder’s inequality to estimate
⨏  ⨏ 
  
  f (x) − f (y) dσ(y) dσ(x)
Δ(z,r) Δ(z,r)
⨏ ⨏
≤ | f (x) − f (y)| dσ(y) dσ(x)
Δ(z,r) Δ(z,r)
⨏ ⨏
 
≤C |x − y| · g(x) + g(y) dσ(y) dσ(x)
Δ(z,r) Δ(z,r)
⨏ ⨏
  1/(n−1)
≤ Cr g(x)n−1 + g(y)n−1 dσ(y) dσ(x)
Δ(z,r) Δ(z,r)
⨏ 1/(n−1)
≤ Cr g n−1 dσ ≤ Cg L n−1 (Δ(z,r),σ) . (11.5.96)
Δ(z,r)

Together with (A.0.8) and the observation that

sup g L n−1 (Δ(z,r),σ) ≤ g L n−1 (Σ,σ) (11.5.97)

z ∈Σ,r >0

this proves (11.5.80). Finally, from (11.5.80), (3.1.54), (11.5.96) and the fact that

lim sup g L n−1 (Δ(z,r),σ) = 0 (11.5.98)

R→0+ z ∈Σ and
r ∈(0,R)

(itself a consequence of the Ahlfors regularity of Σ and the observation that g n−1 dσ
is a finite measure which is absolutely continuous with respect to σ; see also [164,
Exercise 12, p. 32]) we conclude that (11.5.79) holds.
Finally, consider the claim made in item (3), now
 assuming p ∈ (n − 1, ∞). Define
α ∈ (0, 1) as in (11.5.81). Then for each r ∈ 0, 2 diam Σ and each surface ball
Δr ⊆ Σ of radius r we may use (11.5.75), the lower Ahlfors regularity of Σ, Hölder’s
inequality, and (11.5.81) to estimate
760 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
⨏  ⨏ 
  
r −α   f (x) − f (y) dσ(y) dσ(x)
Δr Δr
⨏ ⨏
−α
≤r | f (x) − f (y)| dσ(y) dσ(x)
Δr Δr
⨏ ⨏
 
≤ r −α |x − y| · g(x) + g(y) dσ(y) dσ(x)
Δr Δr
⨏ ⨏
  1/p
≤ Cr 1−α
g(x) p + g(y) p dσ(y) dσ(x)
Δr Δr
⨏ 1/p
≤ Cr (n−1)/p g p dσ ≤ Cg L p (Σ,σ) . (11.5.99)
Δr

From [133, Lemma 3.6.4] we know that σ is a locally finite Borel regular measure
on Σ. Moreover, σ is doubling, since Σ is Ahlfors regular. Granted these properties,
[133, Proposition 7.4.7] applies (with X := Σ) and, together with (11.5.99), yields
all desired conclusions in this case. 
In anticipation of dealing with our last preliminary result, the reader is advised to
recall the definition of the homogeneous Sobolev space of order one from (11.5.55)-
(11.5.56).

Lemma 11.5.10 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is an Ahlfors regular set. Abbreviate σ := H n−1 ∂Ω
and fix some integrability exponent p ∈ (1, ∞).
. p there exists some constant C = C(Ω, n, p) ∈ (0, ∞) such that for each
Then
f ∈ L1 (∂Ω, σ) one may find some non-negative function g ∈ L p (∂Ω, σ) such that
 
| f (x) − f (y)| ≤ |x − y| · g(x) + g(y) for σ-a.e. x, y ∈ ∂Ω, (11.5.100)

as well as

g L p (∂Ω,σ) ≤ C f  L. p (∂Ω,σ) if ∂Ω is unbounded. (11.5.101)

1

As a corollary of (11.5.101), (11.5.55)-(11.5.56), and item (5) in Lemma 11.1.3

one has (compare with Corollary 11.5.5)
.p
if ∂Ω is unbounded then for each f ∈ L1 (∂Ω, σ) one has
(11.5.102)
 f  L. p (∂Ω,σ) = 0 ⇐⇒ f is a constant function on ∂Ω.
1

Also, as a corollary of (11.5.100)-(11.5.101) and (11.5.66)-(11.5.67), in the case

when ∂Ω is unbounded it follows that for each reference point x0 ∈ ∂Ω and each
radius r ∈ (0, ∞) there exists some constant Cx0,r ∈ (0, ∞) with the property that for
.p
each f ∈ L1 (∂Ω, σ) one has
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 761
 
 f − fΔ(x ,r)  
0 σ(x)  ≤ Cx0,r ·  f  L. p (∂Ω,σ) (11.5.103)
1
L ∂Ω, 1+ |x | n 1

⨏
where Δ(x0, r) := ∂Ω ∩ B(x0, r) and fΔ(x0,r) := Δ(x0,r)
f dσ.
In the context of Lemma 11.5.10, from Corollary 11.5.5 and (11.5.102) (or
(11.5.103)) we see that if Ω has an unbounded boundary then any locally constant
function on ∂Ω is actually (globally) constant. In turn, this implies that
if Ω ⊆ Rn is an open set satisfying a two-sided local John condition
and whose boundary is an unbounded Ahlfors regular set, then actually (11.5.104)
∂Ω is a connected set.
This is an illustration of how hypotheses of a purely geometric measure theoretic
nature have topological implications.
Proof of Lemma 11.5.10 To set the stage, define Ω+ := Ω and Ω− := Rn \ Ω. The
present hypotheses then imply that both Ω+ and Ω− are UR domains satisfying a
local John condition, and with ∂Ω+ = ∂Ω = ∂Ω− (see [133, (5.11.27)] and item (7)
in [133, Lemma 5.10.9]).
Bring back the modified principal-value harmonic double layer operator Kmod
.p
from (11.4.27). Having fixed an arbitrary function f ∈ L1 (∂Ω, σ), i.e.,

f ∈ L 1 ∂Ω, 1+σ(x)
|x | n ⊆ Lloc (∂Ω, σ) satisfying
1
(11.5.105)
∂τ j k f ∈ L p (∂Ω, σ) for all j, k ∈ {1, . . . , n},

the idea is to decompose5

 
f = f+ − f− where f± := ± 12 I + Kmod f , (11.5.106)

and treat f+, f− separately. To deal with f+ , define u+ := Dmod f in Ω+ , where Dmod
is the modified boundary-to-domain harmonic double layer operator associated with
Ω+ as in (11.4.26). Then from (11.4.26)-(11.4.36) we learn that u+ : Ω+ → C is a
well-defined (harmonic) function which belongs to C ∞ (Ω+ ), and for any aperture
parameter κ > 0 satisfies
κ−n.t.  
u+ ∂Ω = 12 I + Kmod f = f+ at σ-a.e. point on ∂Ω,
Nκ (∇u+ ) belongs to the space L p (∂Ω, σ), and
(11.5.107)
  
n
 
Nκ (∇u+ ) p ≤ C ∂τ f  p ,
L (∂Ω,σ) jk L (∂Ω,σ)
j,k=1

for some constant C = C(Ω, n, κ, p) ∈ (0, ∞).

5 to some extent, here we are emulating the principle espoused by E.M. Stein and G. Weiss in [179,
Chapter II, p. 37] where they write that “A basic tool in the study of a function, f , on En is obtained
by passage to a harmonic function, defined on the (n + 1)-dimensional upper half-space, which has
f as its boundary value. ”
762 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

At this stage, the discussion branches out, depending on whether ∂Ω is bounded

or unbounded. We shall first treat the latter eventuality. Specifically, when ∂Ω is
unbounded, Lemma 11.5.8 gives that
 
| f+ (x) − f+ (y)| ≤ |x − y| · g+ (x) + g+ (y) for σ-a.e. x, y ∈ ∂Ω,
(11.5.108)
where g+ := C · Nκ (∇u+ ) ∈ L p (∂Ω, σ).

Reason in a similar fashion for f− . Specifically, define u− := −Dmod f in Ω− ,

where Dmod is now the modified boundary-to-domain harmonic double layer op-
erator associated with Ω− as in (11.4.26). As before, (11.4.26)-(11.4.36) tell us that
u− : Ω− → C is a well-defined (harmonic) function, belonging to C ∞ (Ω− ), and
with the property that for any aperture parameter κ > 0 there exists some constant
C = C(Ω, n, κ, p) ∈ (0, ∞) such that
κ−n.t.  
u− ∂Ω = − 12 I + Kmod f = f− at σ-a.e. point on ∂Ω,
Nκ (∇u− ) belongs to the space L p (∂Ω, σ), and
(11.5.109)
  
n
 
Nκ (∇u− ) p ≤ C ∂τ f  p ,
L (∂Ω,σ) jk L (∂Ω,σ)
j,k=1

and
 
| f− (x) − f− (y)| ≤ |x − y| · g− (x) + g− (y) for σ-a.e. x, y ∈ ∂Ω,
(11.5.110)
where g− := C · Nκ (∇u− ) ∈ L p (∂Ω, σ).

If we now set g := g+ + g− then all desired conclusions in the case when ∂Ω is

unbounded follow on account of (11.5.107), (11.5.108), (11.5.109), (11.5.110).
Finally, consider the case when ∂Ω is bounded, a scenario in which we shall
present two proofs, one which uses Lemma . p11.5.4 and one which does not. The first
proof starts by selecting an arbitrary f ∈ L1 (∂Ω, σ) and observing that (11.5.60) in
Lemma 11.5.4 guarantees that, in fact,
p
f ∈ L1 (∂Ω, σ). (11.5.111)

As such, if we define u± as before, then these functions continue to enjoy the proper-
ties described in (11.5.107), (11.5.109). In addition, the membership in (11.5.111)
implies that the nontangential traces
κ−n.t.
u± ∂Ω are well-defined functions in L p (∂Ω, σ). (11.5.112)

Then, with f± as in (11.5.106), Lemma 11.5.8 presently gives

 
| f± (x) − f± (y)| ≤ |x − y| · g± (x) + g± (y) for σ-a.e. x, y ∈ ∂Ω,
 κ−n.t.  (11.5.113)
where g± := C u± ∂Ω  + Nκ (∇u± ) ,
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 763

while from (11.5.112) and the memberships in the second lines of (11.5.107) and
(11.5.109) we see that
g± ∈ L p (∂Ω, σ). (11.5.114)
If we then set
g := g+ + g−, (11.5.115)
it follows from (11.5.114) that g ∈ L p (∂Ω, σ) and all other desired conclusions once
again follow as before. This completes the first proof, mentioned earlier.
The second proof6 proceeds along similar lines, with one major difference. This
time, we shall not rely on Lemma 11.5.4, which was used in (11.5.111), then in
(11.5.112), and, ultimately, in (11.5.115). Instead, we shall give an alternative argu-
ment justifying. the membership in (11.5.115). Turning to specifics, fix an arbitrary
p
function f ∈ L1 (∂Ω, σ) and define u± together with f± and g± as before. We may
use the mapping property recorded in (11.4.34) to conclude that
κ−n.t. 
u± ∂Ω = f± ∈ L 1,∞ (∂Ω, σ) ⊆ L q (∂Ω, σ), (11.5.116)
0<q<1

bearing in mind the formulas in the first lines of (11.5.107) and (11.5.109), the
membership in the first line of (11.5.105), the fact that ∂Ω is bounded, and the
embedding in [133, (6.2.38)]. As a consequence of (11.5.116), the definitions of g±
from (11.5.113), the memberships in the second lines of (11.5.107) and (11.5.109),
we then see that 
g± ∈ L q (∂Ω, σ). (11.5.117)
0<q<1

If we once again define

g := g+ + g−, (11.5.118)
in place of (11.5.115) we presently arrive at the weaker conclusion that

g∈ L q (∂Ω, σ). (11.5.119)
0<q<1

Nonetheless, in view of (11.5.113) and (11.5.106), the function g satisfies (11.5.100).

This allows us to invoke Lemma 11.5.9 and deduce that
  1 
1 −1
whenever q0 ∈ n−1 n , 1 and q1 := q0 − n−1 then
 n−1  (11.5.120)
q1 ∈ 1, n−2 and f belongs to the space L q1 (∂Ω, σ).

In turn, from (11.5.120) and (11.4.36) we see that

κ−n.t.  
u± ∂Ω = ± 12 I + Kmod f ∈ L q1 (∂Ω, σ), (11.5.121)

6 which we will revisit and expand upon when dealing with homogeneous Hardy-based Sobolev
spaces in [136, §2.3]
764 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

which, in light of (11.5.118) and the definitions in (11.5.113), implies

g ∈ L q1 (∂Ω, σ) + L p (∂Ω, σ). (11.5.122)

This puts g in L p (∂Ω, σ) if q1 ≥ p, since ∂Ω is presently assumed to be bounded,

so we are done in this case. In the remaining situation, when q1 ∈ (1, p), this implies
g ∈ L q1 (∂Ω, σ) and we once more invoke Lemma 11.5.9 (bearing in mind that the
function g satisfies (11.5.100) and (11.5.122)) to now conclude that

f ∈ L q2 (∂Ω, σ), (11.5.123)

where 1 
⎧
⎪ 1 −1
− n−1 if q1 ∈ (1, n − 1),
⎪
⎪ q
⎨
⎪ 1

q2 := anything in (1, ∞) if q1 = n − 1, (11.5.124)

⎪
⎪
⎪
⎪∞
⎩ if q1 ∈ (n − 1, ∞).
This is an improvement over (11.5.120) and, much as in (11.5.122), leads to the
conclusion that
g ∈ L q2 (∂Ω, σ) + L p (∂Ω, σ). (11.5.125)
This scheme may be iterated and in finitely many steps we arrive at the conclusion
that g ∈ L p (∂Ω, σ), as wanted. This finishes the second proof. 

Remark 11.5.11 (1) Lemma 11.5.10 provides an alternative justification for the
claims made earlier in Lemma 11.5.4.
(2) In the proof of Lemma 11.5.10, in the case when ∂Ω is bounded one may
be tempted to invoke [133, Proposition 8.9.22] to conclude that the nontangential
κ−n.t.
traces u± ∂Ω ∈ L p (∂Ω, σ), since this would directly imply that g± ∈ L p (∂Ω, σ)
and, ultimately, that g ∈ L p (∂Ω, σ). However, such a line of reasoning would require
the stronger assumption that Ω is actually a two-sided NTA domain.

En route to Theorem 11.5.16, in the next two propositions we record some

remarkable trace results.

Proposition 11.5.12 Let Ω ⊆ Rn be an NTA domain such that ∂Ω is an Ahlfors

regular set and abbreviate σ := H n−1 ∂Ω. Fix an aperture parameter κ ∈ (0, ∞) and
an integrability exponent p ∈ (1, ∞). In this setting, consider a function u : Ω → C
satisfying
u ∈ C 1 (Ω) and Nκ (∇u) ∈ L p (∂Ω, σ). (11.5.126)
Then
κ−n.t. .p
u∂Ω exists σ-a.e. on ∂Ω, belongs to L1 (∂Ω, σ) ∩ L1,loc (∂Ω, σ),
p

 κ−n.t.    (11.5.127)
and u∂Ω  L. p (∂Ω,σ) ≤ C Nκ (∇u) L p (∂Ω,σ)
1

for some constant C ∈ (0, ∞) independent of u.

11.5 Sobolev Spaces on Boundaries of AR Domains . . . 765

Proof Observe that the present hypotheses on Ω ensure (cf. [133, (5.11.28)]) that
Ω is an open set satisfying a two-sided corkscrew condition, as well as a local John
condition. In addition, [133, (5.2.4)] implies that ∂∗ Ω = ∂Ω. In particular, Ω is an
Ahlfors regular domain.
Based on these properties, the current assumptions, and [133, Proposition 8.9.22],
we then see that
 κ−n.t. 
the nontangential trace u∂Ω (x) exists at σ-a.e. x ∈ ∂Ω,
κ−n.t. (11.5.128)
and the function u
p
∂Ω
belongs to the space L (∂Ω, σ).
loc

From [133, Proposition 8.4.9] we also know that there exists some ε ∈ (0, ∞) such
that
Nκε u ∈ Lloc (∂Ω, σ).
p
(11.5.129)
Granted this, we may invoke Corollary 11.3.8 (with w ≡ 1) to conclude that
κ−n.t.
∂τ j k u∂Ω ∈ L p (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.5.130)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) L p (∂Ω,σ)
L p (∂Ω,σ)
j,k=1

for some constant C ∈ (0, ∞) independent of u. Lemma 11.5.8 also gives that
κ−n.t. σ(x)
u∂Ω ∈ L 1 ∂Ω, . (11.5.131)
1 + |x| n
From (11.5.128), (11.5.130)-(11.5.131), item (iii) in Definition 11.1.2, and Defini-
tion 11.5.3 we then conclude that the claims made in (11.5.127) are indeed true. 

Here is a companion to Proposition 11.5.12 containing a similar boundary trace

result, valid for a larger class of domains but under additional assumptions on the
functions involved.

Proposition 11.5.13 Suppose Ω ⊆ Rn is an Ahlfors regular domain satisfying a

local John condition, and abbreviate σ := H n−1 ∂Ω. Pick an aperture parameter
κ ∈ (0, ∞) along with a truncation parameter ε ∈ (0, ∞), and fix two integrability
exponents, p ∈ (1, ∞) and q ∈ [1, ∞]. In such a setting, suppose some complex-
valued function u has been given, which satisfies the following conditions:

Nκε u ∈ Lloc (∂Ω, σ), Nκ (∇u) ∈ L p (∂Ω, σ),

1,1 q
u ∈ Wloc (Ω),
κ−n.t. (11.5.132)
and u∂Ω exists at σ-a.e. point on ∂Ω.

Then there exists a constant C ∈ (0, ∞) independent of u such that

766 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
κ−n.t. .p
u∂Ω belongs to L1 (∂Ω, σ) ∩ L1,loc (∂Ω, σ)
q, p

 κ−n.t.    (11.5.133)
and u∂Ω  L. p (∂Ω,σ) ≤ C Nκ (∇u) L p (∂Ω,σ) .
1

Proof Corollary 11.3.8 (applied with w ≡ 1) guarantees that there exists a constant
C ∈ (0, ∞) independent of u such that
κ−n.t.
∂τ j k u∂Ω ∈ L p (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.5.134)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) L p (∂Ω,σ)
L p (∂Ω,σ)
j,k=1

Together, (11.5.132), (11.5.134), [133, (8.9.8)], [133, Corollary 8.9.6], and item (iii)
in Definition 11.1.2 imply that
κ−n.t.
u∂Ω belongs to L1,loc (∂Ω, σ).
q, p
(11.5.135)

We next claim that

κ−n.t. σ(x)
u∂Ω ∈ L 1 ∂Ω, . (11.5.136)
1 + |x| n
Indeed, if ∂Ω is bounded this is a direct consequence of (11.5.135), while if ∂Ω is
unbounded this follows from the current assumptions and Lemma 11.5.8. Collec-
tively (11.5.134)-(11.5.136) and Definition 11.5.3 then prove that the claims made
in (11.5.133) hold. 
Going further, assume Ω ⊆ Rn is an Ahlfors regular domain, and abbreviate
σ := H n−1 ∂Ω. We shall find it useful to consider homogeneous Sobolev spaces
of order one, modulo constants, on ∂Ω, defined for each integrability exponent
p ∈ (1, ∞) as .p &  .p 
L1 (∂Ω, σ) ∼ := [ f ] : f ∈ L1 (∂Ω, σ) (11.5.137)
where [ f ] denotes
. p the equivalence
& class, modulo constants, of the function f . We
agree to equip L1 (∂Ω, σ) ∼ with the semi-norm
.p &  
L1 (∂Ω, σ) ∼  [ f ] −→ [ f ] L. p (∂Ω,σ)/∼ :=  f  L. p (∂Ω,σ) (11.5.138)
1 1


n
 
= ∂τ f  p .
jk L (∂Ω,σ)
j,k=1

In a favorable geometric setting, this semi-norm becomes a genuine norm, and the
homogeneous Sobolev space of order one, modulo constants, becomes a Banach
space. This is made precise in the proposition below.
Proposition 11.5.14 Suppose Ω ⊆ Rn is an open set satisfying a two-sided local
John condition and whose boundary is an unbounded Ahlfors regular set. Abbreviate
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 767
.p &
σ := H n−1 ∂Ω and pick some integrability exponent p ∈ (1, ∞). Let L1 (∂Ω, σ) ∼
. p the quotient space of classes [·] of equivalence modulo constants of functions
denote
in L1 (∂Ω, σ) (cf. (11.5.137)), equipped with the semi-norm (11.5.138).
.p & .p &
Then (11.5.138) is a genuine norm on L1 (∂Ω, σ) ∼, and L1 (∂Ω, σ) ∼ is a
Banach space when equipped with the norm (11.5.138).

Proof
. p The&fact that the semi-norm (11.5.138) is actually a norm . on the space
&
p
L1 (∂Ω, σ) ∼ follows readily from (11.5.102). Next, to prove that L1 (∂Ω, σ) ∼
.p
is complete when equipped with the norm (11.5.138), let { fα }α∈N ⊆ L1 (∂Ω, σ) be
  .p &
such that [ fα ] α∈N is a Cauchy sequence in the quotient space L1 (∂Ω, σ) ∼. Then
 
for each fixed j, k ∈ {1, . . . , n} it follows that ∂τ j k fα α∈N is a Cauchy sequence in
L p (∂Ω, σ). Since the latter is complete, it follows that there exists g jk ∈ L p (∂Ω, σ)
such that
∂τ j k fα −→ g jk in L p (∂Ω, σ) as α → ∞. (11.5.139)
Fix a reference point⨏x0 ∈ ∂Ω and, for each r ∈ (0, ∞), set Δr := B(x0, r) ∩ ∂Ω. Also,
abbreviate fα,r := Δ fα dσ for each radius r ∈ (0, ∞) and each number α ∈ N.
r
From (11.5.103) (written for f := fα − fβ ) it follows that for each r ∈ (0, ∞) there
exists a constant Cr ∈ (0, ∞) which depends on Ω, p, and r such that that for each
α, β ∈ N we have
   
 fα − fα,r − fβ − fβ,r   σ(x) 
L 1 ∂Ω, 1+ |x | n

n
 
≤ Cr ∂τ fα − ∂τ fβ  p . (11.5.140)
jk jk L (∂Ω,σ)
j,k=1

In view of (11.5.139), this implies that for each fixed r ∈ (0, ∞) the sequence
   
fα − fα,r α∈N is Cauchy in the Banach space L 1 ∂Ω, 1+σ(x)|x | n . Hence, for each fixed
 
r ∈ (0, ∞) there exists a function hr ∈ L 1 ∂Ω, 1+σ(x)
|x | n such that

 
fα − fα,r −→ hr in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.5.141)

From (11.5.141) we see that for each fixed r1, r2 ∈ (0, ∞) we have
 
fα,r2 − fα,r1 −→ hr1 − hr2 in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.5.142)

This forces hr1 − hr2 to be a constant, which ultimately goes to show that actually
 
hr ∈ L 1 ∂Ω, 1+σ(x)
|x | n for each r ∈ (0, ∞). (11.5.143)

Henceforth, we agree to simply write h for hr with r = 1, and cα for fα,r with r = 1.
Then (11.5.143), (11.5.141) tell us that the function
 
h belongs to L 1 ∂Ω, 1+σ(x)
|x | n (11.5.144)

and the sequence {cα }α∈N ⊆ C is such that

768 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
 
fα − cα −→ h in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.5.145)

For each j, k ∈ {1, . . . , n} and each test function ϕ ∈ Cc∞ (Rn ) we may then write
∫ ∫
h(∂τ j k ϕ) dσ = lim ( fα − cα )(∂τ j k ϕ) dσ
∂Ω α→∞ ∂Ω
∫
= − lim ∂τ j k ( fα − cα )ϕ dσ
α→∞ ∂Ω
∫
= − lim (∂τ j k fα )ϕ dσ
α→∞ ∂Ω
∫
=− g jk ϕ dσ, (11.5.146)
∂Ω

thanks to (11.5.145), (11.1.18), and (11.5.139). From this and (11.1.12)-(11.1.17)

we then conclude that

∂τ j k h = g jk ∈ L p (∂Ω, σ) for each j, k ∈ {1, . . . , n}. (11.5.147)

.p
Collectively, (11.5.144), (11.5.147), and (11.5.55) prove that h ∈ L1 (∂Ω, σ). Finally,
 
from (11.5.139), (11.5.147), and (11.5.138) we conclude that the sequence [ fα ] α∈N
.p &
converges to [h], the class of h, in the quotient space L1 (∂Ω, σ) ∼. This proves
.p &
that L1 (∂Ω, σ) ∼ is complete when equipped with the norm (11.5.138). 
We are now in a position to state and prove the following result, pertaining to
the nature of the tangential derivative operator acting from homogeneous Sobolev
spaces on boundaries of chord-arc domains.

Proposition 11.5.15 Let Ω ⊆ R2 be a chord-arc domain with unbounded boundary

and abbreviate σ := H 1 ∂Ω. Also, let ∂τ stand for the tangential derivative operator
∂τ12 on ∂Ω (cf. (11.1.4)). Then for each p ∈ (1, ∞) the operator
.p &
∂τ : L1 (∂Ω, σ) ∼ −→ L p (∂Ω, σ)
.p (11.5.148)
∂τ [ f ] := ∂τ f for each f ∈ L1 (∂Ω, σ),

is a bounded linear isomorphism, with bounded inverse.

The fact that the operator ∂τ is unambiguously defined in (11.5.148) is a conse-

quence of item (5) in Lemma 11.1.3.
Proof of Proposition 11.5.15 From the above observation and Definition 11.5.3 it
follows that (11.5.148) is a well-defined, linear, and bounded operator. Also, from
Corollary 11.5.5 and [133, (5.9.93)] we see that ∂τ is injective in the context of
(11.5.148). Let us show that the operator in question is also surjective. To this end,
denote by R  s → z(s) ∈ ∂Ω the arc-length parametrization of ∂Ω. Having picked
an arbitrary function f ∈ L p (∂Ω, σ), define
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 769
∫ t
g : R → R, g(t) := f (z(s)) ds for each t ∈ R. (11.5.149)
0

From (11.5.149) and [133, (5.9.97), (5.9.100)] we see that g is meaningfully defined
and satisfies:
g ∈ Lloc (R, L 1 ) and g  = f ◦ z ∈ L p (R, L 1 ),
p
(11.5.150)
with g   L p (R, L 1 ) =  f  L p (∂Ω,σ) .

In concert with (11.5.50)-(11.5.51), these properties further imply that

L 1 (t)
g ∈ L 1 R, . (11.5.151)
1 + |t| 2
To proceed, introduce
h := g ◦ z −1 : ∂Ω −→ R, (11.5.152)
and notice that (11.5.151)-(11.5.152) together with [133, (5.9.100), (5.9.95)] entail
∫ ∫ ∫
|h(ζ)| |(h ◦ z)(s)| |(h ◦ z)(s)|
dσ(ζ) = ds ≈ ds
∂Ω 1 + |ζ | R 1 + |z(s)| R 1 + |z(s) − z(0)|
2 2 2

∫
|g(s)|
≈ ds = g 1  L1 (t )  < +∞, (11.5.153)
R 1 + |s| L R,
2
1+|t | 2

which goes to show that

σ(ζ)
h ∈ L 1 ∂Ω, ⊂ Lloc
1
(∂Ω, σ). (11.5.154)
1 + |ζ | 2

Next, if ν = (ν1, ν2 ) ≡ ν1 + iν2 denotes the geometric measure theoretic outward unit
normal to Ω, then [133, (5.9.101)] tells us that
 
ν(z(s)) = −iz (s) = −i z1 (s) + iz2 (s) = −iz1 (s) + z2 (s)
  (11.5.155)
≡ z2 (s), −z1 (s) at L 1 -a.e. point s ∈ R,

where z1, z2 are the (real) components of the function z. Fix an arbitrary test function
ϕ ∈ Cc∞ (R2 ) and observe that for L 1 -a.e. s ∈ R we have

(∂τ ϕ)(z(s)) = ν1 (z(s))(∂2 ϕ)(z(s)) − ν2 (z(s))(∂1 ϕ)(z(s))

= z2 (s)(∂2 ϕ)(z(s)) + z1 (s)(∂1 ϕ)(z(s))

d  
= ϕ(z(s)) , (11.5.156)
ds
thanks to (11.1.4), (11.5.155), and Chain Rule. Based on [133, (5.9.100)], (11.5.152),
(11.5.156), and (11.5.150) we may write
770 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ ∫ ∫
d  
h ∂τ ϕ dσ = h(z(s))(∂τ ϕ)(z(s)) ds = g(s) ϕ(z(s)) ds
∂Ω R R ds
∫ ∫
=− g (s)ϕ(z(s)) ds = − f (z(s))ϕ(z(s)) ds
R R
∫
=− f ϕ dσ. (11.5.157)
∂Ω

In view of the arbitrariness of the test function ϕ ∈ Cc∞ (R2 ), we conclude from
this, (11.5.154), and Definition 11.1.2 that h ∈ L1,loc1 (∂Ω, σ) and that, in fact, we

have ∂τ h = f ∈ L (∂Ω, σ). From

p
. pthis and (11.5.154) we deduce that h belongs to
the homogeneous Sobolev space L1 (∂Ω, σ) (cf. Definition 11.5.3), which ultimately
proves that the operator ∂τ is surjective in the context of (11.5.148). Hence, ∂τ
is an isomorphism, and the fact that its inverse is bounded follows from the Open
Mapping Theorem and Proposition 11.5.14 (whose applicability in the present setting
is guaranteed by [133, (5.9.93)]). 
Our next theorem is focused on the embedding properties enjoyed by the scale of
homogeneous boundary Sobolev spaces defined in (11.5.55).
Theorem 11.5.16 Let Ω ⊆ Rn be an open set satisfying a two-sided local John
condition and whose boundary is an Ahlfors regular set. Abbreviate σ := H n−1 ∂Ω
and fix some integrability exponent p ∈ (1, ∞).
 
1 −1
(1) Assume that7 p ∈ (1, n − 1) and define p∗ := p1 − n−1 ∈ (1, ∞). Then

.p σ(x) p
L1 (∂Ω, σ) ⊆ L 1 ∂Ω, ∩ Lloc∗ (∂Ω, σ). (11.5.158)
1 + |x| n
.p
Also, for each given function f ∈ L1 (∂Ω, σ) there exists a number c = c( f ) ∈ C
such that

the function f − c belongs to the space L p∗ (∂Ω, σ) (11.5.159)

and
whenever ∂Ω is an unbounded set it follows that

n
 
 f − c L p∗ (∂Ω,σ) ≤ C ∂τ f  p
jk L (∂Ω,σ)
= C f  L. p (∂Ω,σ), (11.5.160)
1
j,k=1

for some constant C = C(Ω, n, p) ∈ (0, ∞) independent of f . Moreover, in this

range of p’s, .p p ,p
L1 (∂Ω, σ) = C + L1 ∗ (∂Ω, σ) (11.5.161)
(where C is identified with the space of constant functions on ∂Ω, while the
off-diagonal Sobolev space is defined as in (11.1.14)), and one also has
7 this implicitly requires n ≥ 3
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 771

.p σ(x)
L1 (∂Ω, σ) ∩ L 1 ∂Ω, ⊆ L p∗ (∂Ω, σ), (11.5.162)
1 + |x| n−1
as well as:
if ∂Ω is unbounded then there exists C = C(Ω, n, p) ∈ (0, ∞)
with the property that  f  L p∗ (∂Ω,σ) ≤ C f  L. p (∂Ω,σ)
1 (11.5.163)
.p σ(x)
for all f ∈ L1 (∂Ω, σ) ∩ L ∂Ω, 1
.
1 + |x| n−1

In particular, for each q ∈ [1, ∞) the embeddings

q, p p ,p
L1 (∂Ω, σ) → L1 ∗ (∂Ω, σ) → L p∗ (∂Ω, σ) (11.5.164)

are well defined, linear, and continuous, hence so are

p p ,p
L1 (∂Ω, σ) → L1 ∗ (∂Ω, σ) → L p∗ (∂Ω, σ). (11.5.165)

(2) Corresponding to the case when n ≥ 3, p = n − 1, and ∂Ω is unbounded, one

has the continuous embeddings
. 
q
L1n−1 (∂Ω, σ) → VMO(∂Ω, σ) → BMO(∂Ω, σ) → Lloc (∂Ω, σ).
0<q<∞
(11.5.166)
As a consequence of this and the John-Nirenberg inequality, there exist some
small constant c ∈ (0, ∞) along with some large constant C ∈ (0, ∞), both of
which depend only on Ω, with the property that
⨏ ! "
c | f − fΔ |
exp dσ ≤ C (11.5.167)
Δ  f  L. n−1 (∂Ω,σ)
1

. n−1
for each non-constant function f ∈ L1 (∂Ω, σ) and each surface ball Δ ⊆ ∂Ω.
(3) Assume p ∈ (n − 1, ∞) and define α := 1 − n−1
p ∈ (0, 1). If ∂Ω is unbounded then
one has the continuous embedding
.p .
L1 (∂Ω, σ) → C α (∂Ω). (11.5.168)

(4) One has

.p σ(x) p
L1 (∂Ω, σ) ⊆ L 1 ∂Ω, ∩ Lloc (∂Ω, σ)
1 + |x| n−1+ε (11.5.169)
for each ε > 0 if p ≤ n − 1, and each ε > 1 − n−1
p if p > n − 1.

Proof Let us justify the claims made in item (1). For one thing, (11.5.158) and
(11.5.159)-(11.5.160) are direct consequences of Lemma 11.5.9 and Lemma 11.5.10.
772 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

As regards (11.5.161), the right-to-left inclusion is seen from (11.1.14), (11.5.55),

item (5) in Lemma 11.1.3, and [133, (7.2.5)]. The left-to-right inclusion in (11.5.161)
also uses (11.5.159). Next, if ∂Ω is bounded then (11.5.162) is clear from (11.5.158).
To check (11.5.162)-(11.5.163) in the case when ∂Ω is unbounded, pick some
.p  
arbitrary function f ∈ L1 (∂Ω, σ) ∩ L 1 ∂Ω, 1+σ(x)|x | n−1
. From (11.5.159)-(11.5.160)
we know that there exists c ∈ C such that
 
f − c ∈ L p∗ (∂Ω, σ) ⊆ L 1 ∂Ω, 1+σ(x) |x | n−1
(11.5.170)
and  f − c L p∗ (∂Ω,σ) ≤ C f  L. p (∂Ω,σ)
1

for some C ∈ (0, ∞) independent of f . Then

 
c = f − ( f − c) ∈ L 1 ∂Ω, 1+σ(x)
|x | n−1
(11.5.171)

which, in view of the fact that ∂Ω is presently assumed to be an unbounded Ahlfors

regular set, forces c = 0. Having established this, it follows that the embedding in
(11.5.162) is indeed well defined, linear, and continuous in the case when ∂Ω is
unbounded.
Next, the claims in items (2)-(3) are directly implied by Lemma 11.5.9 and
Lemma 11.5.10. As regards item (4), if ∂Ω is unbounded then the claim in (11.5.169)
follows from (11.5.159), (11.5.166), (11.5.168), [133, (7.4.118)-(7.4.119)], and
(11.5.59). Finally, (11.5.169) becomes a simple consequence of (11.5.59) if ∂Ω
is bounded. 

Other basic properties of Sobolev spaces on the boundaries of Ahlfors regular

domains satisfying a two-sided local John condition are presented in our next two
theorems. The first such theorem focuses on the case when the underlying domain
has an unbounded boundary.

Theorem 11.5.17 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is an unbounded Ahlfors regular set. Abbreviate
σ := H n−1 ∂Ω and fix p ∈ (1, ∞). Then the following properties are true.
(1) For each index q ∈ (1, p] there exists some finite. p constant C = C(Ω, p, q) > 0
with the property that for each function f ∈ L1 (∂Ω, σ), each location x ∈ ∂Ω,
and each scale r ∈ (0, ∞) one has
⨏ 1/q  
| f − fx,r | q dσ ≤ C r · M ∂Ω,q |∇tan f | (x) (11.5.172)
∂Ω∩B(x,r)
⨏
where fx,r := ∂Ω∩B(x,r) f dσ, and M ∂Ω,q is the L q -based Hardy-Littlewood
maximal operator on ∂Ω (cf. (A.0.69)).
(2) Having fixed q ∈ [1, p), for each f ∈ Lloc
1 (∂Ω, σ) consider the following version

of Calderón’s maximal operator:

11.5 Sobolev Spaces on Boundaries of AR Domains . . . 773
'⨏   ( 1/q
   f (y) − fx,r q
Λ∗,q f (x) := sup   dσ(y) (11.5.173)
 r 
r ∈(0,∞) ∂Ω∩B(x,r)
⨏
at each x ∈ ∂Ω, where fx,r := ∂Ω∩B(x,r)
f dσ. Then

 f  L. p (∂Ω,σ) ≈ Λ∗,q f  L p (∂Ω,σ)

1
.p (11.5.174)
uniformly for f ∈ L1 (∂Ω, σ).
p
(3) Fix a Muckenhoupt weight w ∈ Ap (∂Ω, σ). Then for each f ∈ L1 (∂Ω, wσ) there
exist some non-negative function g ∈ L p (∂Ω, wσ) and some σ-measurable set
N ⊆ ∂Ω with σ(N) = 0 such that
 
| f (x) − f (y)| ≤ |x − y| g(x) + g(y) , ∀x, y ∈ ∂Ω \ N. (11.5.175)

In addition, matters may be arranged so that for some C = C(∂Ω, p, w) ∈ (0, ∞)

one has

n
 
∂τ f  ≤ C · g at σ-a.e. point on ∂Ω
jk
j,k=1
 (11.5.176)
n
 
and g L p (∂Ω,wσ) ≤ C · ∂τ f  p .
jk L (∂Ω,wσ)
j,k=1

Also, for every point x ∈ ∂Ω and every scale r ∈)(0, ∞) the following Poincaré
∫ ∫
inequality holds (with fw,x,r := ∂Ω∩B(x,r) f w dσ ∂Ω∩B(x,r) w dσ):
∫ 1/p
∫ 1/p
| f (y) − fw,x,r | w dσ(y)
p
≤ Cr g p w dσ .
∂Ω∩B(x,r) ∂Ω∩B(x,r)
(11.5.177)
(4) The inclusion

⎧ L p (∂Ω, σ) if p ∈ (1, n − 1) and p∗ := (n−1)p

∗
⎪
⎪
⎪ n−1−p ,
⎨
⎪
L1 (∂Ω, σ) → VMO(∂Ω, σ) if p = n − 1,
p
⎪
⎪ .α
⎪
⎪
⎩ C (∂Ω) if p ∈ (n − 1, ∞) and α := 1 − n−1
p ,
(11.5.178)
is well defined, linear, and continuous.

Proof As regards item (1), the estimate claimed in (11.5.172) is a direct consequence
of the Poincaré inequality discussed in Proposition 11.5.1 with p := q (whose present
applicability is ensured by [133, (7.7.104)] with w ≡ 1), and (A.0.69).
Consider next the claim made in item (2). Given. q ∈ [1, p), pick qo ∈ (1, p)
p
with qo ≥ q. Also, select an arbitrary function f ∈ L1 (∂Ω, σ). Based on Hölder’s
inequality and (11.5.172) (used with q := qo ) we see that there exists some finite
constant C = C(Ω, p, qo ) > 0, independent of f , with the property that
774 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
'⨏   ( 1/qo
 f − fx,r qo
Λ∗,q f (x) ≤   dσ
 r 
∂Ω∩B(x,r)
 
≤ C · M ∂Ω,qo |∇tan f | (x) for all x ∈ ∂Ω. (11.5.179)

In concert with the boundedness of M ∂Ω,qo on L p (∂Ω, σ) (cf. [133, (7.6.11)]) we

then conclude that

Λ∗,q f  L p (∂Ω,σ) ≤ C∇tan f  L p (∂Ω,σ) . (11.5.180)

In view of (11.5.56) and (11.4.9) (whose current applicability is ensured by

(11.5.59) and [133, Proposition 5.10.4]), this justifies the right-pointing inequali-
ty in (11.5.174).
To justify the left-pointing inequality in (11.5.174), denote

Ω+ := Ω and Ω− := Rn \ Ω. (11.5.181)

Next, associate with f the functions u± ∈ C ∞ (Ω± ) as in the proof of Lemma 11.5.10.
In particular, given some arbitrary aperture parameter κ ∈ (0, ∞), it follows that
the functions u± satisfy (11.5.107) and (11.5.109), respectively. In addition, from
(11.4.31) we know that the nontangential boundary traces
  κ−n.t.
∂ u± ∂Ω exist at σ-a.e. point on ∂Ω,
(11.5.182)
for each index ∈ {1, . . . , n}.

Also, as seen from (11.4.33) and (11.5.59), for each truncation parameter ε ∈ (0, ∞)
we have
Nκε u± ∈ Lloc (∂Ω, σ).
p
(11.5.183)
Granted these properties, Proposition 11.3.2 applies and, for each j, k ∈ {1, . . . , n},
at σ-a.e. point on ∂Ω allows us to write
 κ−n.t.   κ−n.t. 
∂τ j k f = ∂τ j k u+ ∂Ω − ∂τ j k u− |∂Ω
κ−n.t. κ−n.t.
= ν j (∂k u+ )∂Ω − (∂k u− )∂Ω
κ−n.t. κ−n.t.
− νk (∂j u+ )∂Ω − (∂j u+ )∂Ω , (11.5.184)

where ν = (ν1, . . . , νn ) is the geometric measure theoretic outward unit normal to Ω.

To proceed, let N ⊆ ∂Ω be a σ-measurable set with σ(N) = 0 such that at
each point in ∂Ω \ N all the nontangential boundary traces in (11.5.182) exist
and the pointwise identity (11.5.184) is valid. Given any x ∈ ∂ ∗ Ω ∩ ∂nta Ω \ N,
consider two rectifiable paths in Ω± , of lengths L± , with arc-length parametrizations
[0, L± ]  s → γ ± (s), both of which emerge from x (i.e., γ ± (0) = x). Moreover,
assume that these are nontangential, in the sense that dist(γ ± (s), ∂Ω) ≈ s uniformly
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 775

for s ∈ [0, L± ]. The existence of such paths is ensured by [133, Definition 5.11.7] (cf.
also (A.0.54)). Exercising the freedom of choosing the aperture parameter κ ∈ (0, ∞)
sufficiently large to begin with, we may then interpret the left-most side of (11.5.184)
evaluated at x as

lim+ ν j (x) (∂k u+ )(γ + (s)) − (∂k u− )(γ − (s))
s→0

− νk (x) (∂j u+ )(γ + (s)) − (∂j u− )(γ − (s)) . (11.5.185)

Fix a boundary point x along with two nontangential paths γ ± as above, and fix
also s ∈ (0, L), where L = Lx := min{L+, L− }. We claim that for each index
∈ {1, . . . , n},
 
(∂ u+ )(γ + (s)) − (∂ u− )(γ − (s)) ≤ CΛ∗,1 f (x). (11.5.186)

In turn, (11.5.186), (11.5.184), and (11.4.1) entail (bearing in mind [133, (8.8.52)],
[133, (5.2.4)], and [133, (5.6.21)]) the pointwise bound

∇tan f (x)| ≤ CΛ∗,1 f (x) for σ-a.e. x ∈ ∂Ω, (11.5.187)

from which the left-pointing inequality in (11.5.174) follows immediately.

To prove the claim made in (11.5.186), fix an arbitrary index ∈ {1, . . . , n} along
with an arbitrary point x ∈ ∂ ∗ Ω ∩ ∂nta Ω \ N. Also, pick some s ∈ (0, L). Observe
that, for the present goal, we may replace f by f − fx,s in the definitions of u± since
∂ Dmod annihilates constants (cf. (11.4.28)). Assume this is the case and expand


∞
f − fx,s = fi, (11.5.188)
i=0

where

f0 := ( f − fx,s )1Δ(x,s), fi := ( f − fx,s )1Δ(x,2i s)\Δ(x,2i−1 s), i ≥ 1. (11.5.189)

Then (11.5.59) guarantees that fi ∈ L p (∂Ω, σ) for each i ∈ N0 , so (11.4.30) permits

us to write
 
(∂ u+ )(γ + (s)) − (∂ u− )(γ − (s))
   
≤ (∂ D f0 )(γ + (s)) + (∂ D f0 )(γ − (s))

∞
 
+ (∂ D fi )(γ + (s)) − (∂ D fi )(γ − (s)) =: I + II + III,
i=1
(11.5.190)

where D is the harmonic double layer potential operator mapping functions from
L p (∂Ω, σ) into functions in C ∞ (Rn \ ∂Ω). Standard estimates for derivatives of the
776 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

fundamental solution for the Laplacian in Rn and the specific nature of the parametric
paths γ ± ensure that
∫
1
I + II ≤ C ±
| f (y) − fx,s | dσ(y)
Δ(x,s) |γ (s) − y|
n

⨏  
 f (y) − fx,s 
≤C   dσ(y) ≤ CΛ∗,1 f (x). (11.5.191)
 s 
Δ(x,s)

Using also the Mean Value Theorem, for each i ∈ N0 large enough, depending only
on geometry, we obtain that
 
(∂ D fi )(γ + (s)) − (∂ D fi )(γ − (s))
∫
 2 
≤C· (∇ EΔ )(γ + (s) − y) − (∇2 EΔ )(γ − (s) − y) | f (y) − fx,s | dσ(y)
Δ(x,2i s)\Δ(x,2i−1 s)
⨏  
 f (y) − fx,s 
≤ C 2−i   dσ(y), (11.5.192)
 2i s 
Δ(x,2i s)
 
since |γ + (s) − γ − (s)| ≤ Cs, and if y ∈ Δ(x, 2i s) \ Δ(x, 2i−1 s) and z ∈ γ + (s), γ − (s)
then for any index i ∈ N0 large enough, depending only on geometry, we may write
|z − y| ≥ |x − y| − |x − z| ≥ 2i−1 s − Cs ≥ C2i s. Going further, we estimate
⨏  
 f (y) − fx,s 
  dσ(y)
 2i s 
Δ(x,2i s)
⨏  
 f (y) − fx,2i s 
≤   dσ(y)
 2i s 
Δ(x,2i s)

i ⨏  
 fx,2 j s − fx,2 j−1 s 
+   dσ(y)
i s)
 2is 
j=1 Δ(x,2

  i ⨏  
 f (y) − fx,2i s   f (y) − fx,2 j s 

≤    dσ(y)
2i s  dσ(y) + C j s)
 2 js 
j=1 Δ(x,2

≤ Ci · Λ∗,1 f (x). (11.5.193)

The terms in the series defining III corresponding to i small may be estimated much
as we have done in (11.5.191). In conjunction with (11.5.192), this ultimately yields
the bound

∞
III ≤ C i 2−i Λ∗,1 f (x) ≤ CΛ∗,1 f (x), (11.5.194)
i=1

for some constant C ∈ (0, ∞) independent of f and x. At this stage, (11.5.186) is

established, thus concluding the proof of (11.5.174).
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 777

Let us now treat item (3). Recall that KΔ denotes the principal-value harmonic
p
double layer operator associated with Ω as in (A.0.52). Given f ∈ L1 (∂Ω, wσ),
decompose  
f = f+ − f− where f± := ± 12 I + K f . (11.5.195)
p
From [136, §1.5] we know that KΔ maps L1 (∂Ω, wσ) boundedly into itself, hence
p
f± ∈ L1 (∂Ω, wσ). We shall show that the claims made in item (v) hold for f+
and f− separately. To deal with f+ , define u± := ±D ± f in Ω± , where D ± are
the boundary-to-domain harmonic double layer operators associated with Ω± as in
(11.4.12). For any aperture parameter κ > 0 the functions u± ∈ C ∞ (Ω) then satisfy
(cf. (11.4.13)-(11.4.17) and (11.4.26)-(11.4.36))

Nκ u±, Nκ (∇u± ) ∈ L p (∂Ω, wσ),

 
Nκ u±  p ≤ C f  L p (∂Ω,wσ),
L (∂Ω,wσ)
  
n
 
Nκ (∇u± ) p ≤ C ∂τ f  p ,
L (∂Ω,wσ) jk L (∂Ω,wσ)
j,k=1 (11.5.196)
κ−n.t.  
u± ∂Ω = ± 12 I + K f = f± at σ-a.e. point on ∂Ω, and

n
 
∂τ f  ≤ C Nκ (∇u+ ) + Nκ (∇u− ) at σ-a.e. point on ∂Ω.
jk
j,k=1

From Lemma 11.5.8 we also know that

 
| f± (x) − f± (y)| ≤ |x − y| · g(x) + g(y) for σ-a.e. x, y ∈ ∂Ω,
  (11.5.197)
where g := C Nκ (∇u+ ) + Nκ (∇u− ) .

At this stage, the claims made in (11.5.175)-(11.5.176) follow from (11.5.195)-

(11.5.197).
To prove the Poincaré inequality claimed in (11.5.177), fix some point x ∈ ∂Ω
along with some scale r ∈ (0, ∞). Also, define the measure dw := w dσ. Then using
(11.5.212) and Hölder’s inequality we may estimate
778 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ 1/p
| f (y) − fw,x,r | p w dσ(y)
∂Ω∩B(x,r)
∫  ⨏ p
  1/p
=  f (y) − f (z) dw(z) dw(y)
∂Ω∩B(x,r) ∂Ω∩B(x,r)
∫ ⨏ 1/p
≤ | f (y) − f (z)| p dw(z) dw(y)
∂Ω∩B(x,r) ∂Ω∩B(x,r)
∫ ⨏
 p 1/p
≤C |y − z| p g(y) + g(z) dw(z) dw(y)
∂Ω∩B(x,r) ∂Ω∩B(x,r)
∫ 1/p
≤ Cr g p dw
∂Ω∩B(x,r)
∫ 1/p
= Cr g p w dσ , (11.5.198)
∂Ω∩B(x,r)

since y, z ∈ ∂Ω ∩ B(x, r) forces |z − y| < 2r. This finishes the treatment of item (3).
Finally, consider the claim made in item (4). The version of (11.5.178) corre-
sponding to the case when p ∈ (1, n − 1) has been already noted in (11.5.165). The
version of (11.5.178) corresponding to p = n − 1 is a consequence of (11.5.57)
and (11.5.166). The version of (11.5.178) corresponding to p > n − 1 is clear from
(11.5.57) and (11.5.168). 
We further augment Theorem 11.5.17, now working on domains with compact
boundaries.

Theorem 11.5.18 Assume Ω ⊆ Rn (where n ∈ N with n ≥ 2) is an open set satisfying

a two-sided local John condition and whose boundary ∂Ω is compact and Ahlfors
regular. Abbreviate σ := H n−1 ∂Ω and fix p ∈ (1, ∞). Then the following statements
are true.
(i) The space
  
ψ ∂Ω : ψ ∈ Cc∞ (Rn ) is dense in L1 (∂Ω, σ).
p
(11.5.199)

Moreover, the inclusion

p
Lip (∂Ω) → L1 (∂Ω, σ)
(11.5.200)
is well defined, continuous, with dense range.

(ii) In a quantitative sense, one has

⎧ L p (∂Ω, σ) for p∗ := (n−1)p

∗
⎪
⎪
⎪ n−1−p if p ∈ (1, n − 1),
p
⎨
⎪
L1 (∂Ω, σ) → L q (∂Ω, σ) for all q ∈ (1, ∞) if p = n − 1, (11.5.201)
⎪
⎪
⎪
⎪ C (∂Ω)
α
⎩ for α := 1 − p if p ∈ (n − 1, ∞).
n−1
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 779

In addition, corresponding to n ≥ 3 and p = n − 1, the following inclusions are

well defined, linear, and continuous:

L1n−1 (∂Ω, σ) → VMO(∂Ω, σ) → BMO(∂Ω, σ). (11.5.202)

As a corollary of (11.5.202), (11.1.25), and (3.1.51), in the present setting one

therefore has

if n ≥ 3 then the closure of L1n−1 (∂Ω, σ) in the John-Nirenberg

space BMO(∂Ω, σ) is the Sarason space VMO(∂Ω, σ).
(11.5.203)
(iii) The following inclusions are well defined, linear, and compact:
p (n−1)p
L1 (∂Ω, σ) → L q (∂Ω, σ) if 1 < p < n − 1 and 0 < q < n−1−p , (11.5.204)

L1 (∂Ω, σ) → C α (∂Ω) if n − 1 < p < ∞ and 0 < α < 1 −

p
p .
n−1
(11.5.205)

As a consequence, the following inclusion operators are well defined, linear, and
compact:

p   −1
L1 (∂Ω, σ) → L q (∂Ω, σ) if 0 < q < max 0, 1
p − 1
n−1 , (11.5.206)

L1 (∂Ω, σ) → L ∞ (∂Ω, σ) if n − 1 < p < ∞.

p
(11.5.207)
p
(iv) Select f ∈ L1 (∂Ω, σ) along with some g ∈ L1 (∂Ω, σ) where p ∈ (1, ∞)
p

satisfies 1/p + 1/p = 1. Then for each j, k ∈ {1, . . . , n} the following boundary
integration by parts formula holds:
∫ ∫
(∂τ j k f )g dσ = − f (∂τ j k g) dσ. (11.5.208)
∂Ω ∂Ω
p q
(v) Fix q ∈ (n−1, ∞) satisfying q ≥ p. Then L1 (∂Ω, σ) is a module over L1 (∂Ω, σ),
i.e.,
q p p
f ∈ L1 (∂Ω, σ), g ∈ L1 (∂Ω, σ) =⇒ f · g ∈ L1 (∂Ω, σ). (11.5.209)
p
In particular, L1 (∂Ω, σ) is an algebra whenever p ∈ (n − 1, ∞). Furthermore,
the Leibniz formula

∂τ j k ( f g) = f ∂τ j k g + (∂τ j k f )g, ∀ j, k ∈ {1, . . . , n}, (11.5.210)

q p
holds for every f ∈ L1 (∂Ω, σ) and g ∈ L1 (∂Ω, σ), and there exists a finite
constant C > 0 with the property that

 f · g L p (∂Ω,σ) ≤ C f  L q (∂Ω,σ) g L p (∂Ω,σ), (11.5.211)

1 1 1

q p
for any f ∈ L1 (∂Ω, σ) and g ∈ L1 (∂Ω, σ).
 780 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

(vi) A σ-measurable function f defined on ∂Ω actually belongs to the Sobolev space

p
L1 (∂Ω, σ) if and only if there exist some non-negative function g ∈ L p (∂Ω, σ)
and some σ-measurable set N ⊆ ∂Ω with σ(N) = 0 such that
 
| f (x) − f (y)| ≤ |x − y| g(x) + g(y) for all x, y ∈ ∂Ω \ N. (11.5.212)

Moreover, if G f denotes the collection of all functions g with the above properties
(referred to as “generalized gradients” of f ), then

 f  L p (∂Ω,σ) ≈  f  L p (∂Ω,σ) + inf g ∈G f g L p (∂Ω,σ)

1
p
(11.5.213)
uniformly for f ∈ L1 (∂Ω, σ).

(vii) Assume p ∈ (1, n − 1) and define p∗ := (n−1)pn−1−p . Then there exists a constant
p
C ∈ (0, ∞) such that for every f ∈ L1 (∂Ω, σ) one has
⨏ 1/p ∗
⨏ 1/p
∗
| f − fx,r | p dσ ≤ Cr g p dσ , (11.5.214)
∂Ω∩B(x,r) ∂Ω∩B(x,2r)

where x ∈ ∂Ω is ⨏an arbitrary point, r ∈ (0, ∞) is an arbitrary scale, the symbol

fx,r abbreviates ∂Ω∩B(x,r) f dσ, and where g ∈ L p (∂Ω, σ) is associated with
the function f as in (11.5.212).
(viii) There exist a threshold R ∈ (0, diam ∂Ω) and a finite constant C = C(Ω, p) > 0
p
with the property that for each f ∈ L1 (∂Ω, σ), x ∈ ∂Ω, and r ∈ (0, R) one has
⨏ 1/p  
| f − fx,r | p dσ ≤ C r · M ∂Ω, p |∇tan f | (x) (11.5.215)
∂Ω∩B(x,r)
⨏
where fx,r := ∂Ω∩B(x,r) f dσ, and M ∂Ω, p is the L p -based Hardy-Littlewood
maximal operator on ∂Ω (cf. (A.0.69)). In particular, there exists a constant
p
C = C(Ω, p) > 0 such that, for each function f ∈ L1 (∂Ω, σ) and each radius
r > 0, at every point x ∈ ∂Ω one has
⨏ 
1/p    
| f − fx,r | p dσ ≤ C r · M ∂Ω, p |∇tan f | (x) + M ∂Ω, p f (x) .
∂Ω∩B(x,r)
(11.5.216)
(ix) Having fixed q ∈ [1, p), for each f ∈ Lloc
1 (∂Ω, σ) consider the following version

of Calderón’s maximal operator:

'⨏   ( 1/q
   f (y) − fx,r q
Λ∗,q f (x) := sup   dσ(y) (11.5.217)
 r 
0<r <diam ∂Ω ∂Ω∩B(x,r)
⨏
at each x ∈ ∂Ω, where fx,r := ∂Ω∩B(x,r) f dσ. Then
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 781

 f  L p (∂Ω,σ) ≈ Λ∗,q f  L p (∂Ω,σ) +  f  L p (∂Ω,σ) .

1
p
(11.5.218)
uniformly for f ∈ L1 (∂Ω, σ).
p
(x) For each Lipschitz function Φ : C → C and each f ∈ L1 (∂Ω, σ) it follows that
p
the composition Φ ◦ f belongs to L1 (∂Ω, σ).
∞
(xi) One has L1 (∂Ω, σ) = Lip (∂Ω), in the sense that these vector spaces coincide
as sets and their respective norms are equivalent.

Proof Items (i)-(vi) amount to a collection of slight variations of results established

p
in [97], [139]. As regards item (vii), pick f ∈ L1 (∂Ω, σ) with p ∈ (1, n − 1). Then the
current item (vi) permits us to invoke [80, Theorem 8.7, p. 197; cf. (8.3)] (see also
[80, Remark 8.8, p. 198; cf. (8.7)]), according to which (11.5.214) holds. Incidentally,
p ∗
the inclusion L1 (∂Ω, σ) → L p (∂Ω, σ) readily follows from (11.5.214), providing
another proof of the version of (11.5.178) when p ∈ (1, n − 1) (the version of
(11.5.178) in the case when p ∈ (n − 1, ∞) is seen directly from [80, Theorem 8.7,
pp. 197-198; cf. (8.6)]).
As far as item (viii) is concerned, estimate (11.5.215) is a direct consequence
of the Poincaré-type inequality proved in [97, Proposition 4.13, p. 2692] (cf. also
[97, (4.3.20)-(4.3.21), p. 2717]) and (A.0.69), while (11.5.216) readily follows from
(11.5.215). Next, the claim in item (ix) follows from [97, Proposition 4.26, p. 2716].
Moving on, the claim in item (x) may be easily justified based on the characterization
of boundary Sobolev spaces given in item (vi).
Finally, we shall prove the claim made in item (xi) of the statement of the theorem
via double inclusion. In one direction, given an arbitrary function f ∈ L1∞ (∂Ω, σ)
p
it follows that f ∈ L1 (∂Ω, σ) which, on account of (11.5.216) and the fact that
 |∇tan f |  ∞ +  f  L ∞ (∂Ω,σ) ≤ cn  f  L1∞ (∂Ω,σ) , further implies that
L (∂Ω,σ)
⨏ 1/p
| f − fx,r | p dσ ≤ C r ·  f  L1∞ (∂Ω,σ) (11.5.219)
∂Ω∩B(x,r)

for each scale r > 0 and each point x ∈ ∂Ω. We therefore have
! ⨏ "
1 1/p
sup sup | f − fΔr | dσ
p
≤ C f  L1∞ (∂Ω,σ), (11.5.220)
r >0 Δr r Δr

where the second supremum is taken over all surface balls Δr ⊆ ∂Ω of radius r.
Granted this, [133, Proposition 7.4.9] then implies (in view of [133, Lemma 3.6.4]
presently used with s := n − 1) that, after eventually redefining the function f on a
set of σ-measure zero, for all x, y ∈ ∂Ω we have

| f (x) − f (y)|
sup ≤ C f  L1∞ (∂Ω,σ) . (11.5.221)
x,y ∈∂Ω |x − y|
xy
782 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

As such, f belongs to the space Lip(∂Ω) and  f Lip(∂Ω) ≤ C f  L1∞ (∂Ω,σ) , hence
L1∞ (∂Ω, σ) → Lip(∂Ω). The opposite inclusion follows from Proposition 11.1.11,
completing the proof of the claim in item (xi). 

To describe a useful application of the above result, work in the two-dimensional

setting and let Σ be a chord-arc curve in the plane (cf. [133, Definition 5.9.5]).
If Ω ⊆ R2 is the inner domain of Σ, then Ω is a chord-arc domain (see [133,
Definition 5.9.8]). Recall from [133, (5.9.76)] and [133, (5.11.28)] that
Ω is an Ahlfors regular domain satisfying a two-sided local
(11.5.222)
John condition and with a compact boundary ∂Ω = Σ.

In particular, for each p ∈ [1, ∞] it makes sense to define

p p
L1 (Σ, σ) := L1 (∂Ω, σ) where σ := H 1 Σ = H 1 ∂Ω. (11.5.223)

For this brand of Sobolev space on chord-arc curves, we have the following natural
characterization.

Proposition 11.5.19 Suppose Σ is a chord-arc curve in the plane. Let L ∈ (0, ∞) be

its length, and consider an arc-length parametrization z(·) : [0, L] → Σ of the curve
Σ. Then for each integrability exponent p ∈ (1, ∞) one has the equality
  
p p 
L1 (Σ, σ) = f ∈ C 0 (Σ) : f z(·) ∈ L1 (0, L), L 1 , (11.5.224)

in a quantitative fashion.

Proof Let Ω ⊆ R2 be the inner domain of Σ. Throughout, we shall keep in mind

(11.5.222). To prove the left-to-right inclusion in (11.5.224), start with an arbitrary
f ∈ L1 (Σ, σ). Then (11.5.201) guarantees that f ∈ C α (Σ) where α := 1 − 1/p ∈
p

(0, 1). In particular, f ∈ C 0 (Σ). Let g ∈ L p (Σ, σ) be an arbitrary generalized gradient

of f , i.e., g ∈ G f with the latter class defined in item (vi) of Theorem 11.5.18. From
[133, (5.8.46)] we see that z(·) is Lipschitz. Keeping this in mind, the change of
variable formula [49, Theorem 2, p. 99] applies and presently gives that for each
σ-measurable set E ⊆ Σ and each non-negative σ-measurable function h on E we
have ∫ ∫
h dσ = h(z(s)) ds. (11.5.225)
E z −1 (E)

In turn, this ensures that composition

  with z(·) is an isometry on Lebesgue scales,
which
  further implies that g z(·) ∈ L p (0, L), L 1 is a generalized gradient for
f z(·) . Granted this, from [79, Theorem 1, p. 405] and (11.5.213) we conclude that
  p    
f z(·) ∈ L1 (0, L), L 1 and  f z(·)  L p (0, L), L 1  ≤ C f  L p (Σ,σ), (11.5.226)
1 1

for some constant C ∈ (0, ∞) which depends only on Σ and p. This takes care of the
left-to-right inclusion in (11.5.224).
11.5 Sobolev Spaces on Boundaries of AR Domains . . . 783

 consider a function f ∈ C (Σ)

To prove the right-to-left 0
 inclusion
 in (11.5.224),
p
with the property that f z(·) ∈ L1 (0, L), L 1 . Via standard embeddings (and
keeping mind that z(0) = z(L)), the latter membership then self-improves to
  p       
f z(·) ∈ L1 (0, L), L 1 ∩ C 0 [0, L] and f z(0) = f z(L) . (11.5.227)

To proceed, pick an arbitrary ϕ ∈ Cc1 (R2 ). If x1 (·), x2 (·) are the scalar components
(or the real and imaginary parts) of z(·), then from [133, (5.8.44)] we know that the
geometric measure theoretic outward unit normal to Ω is given by
   
ν z(s) = −iz (s) ≡ x2 (s), −x1 (s) for L 1 -a.e. s ∈ [0, L]. (11.5.228)

Consequently, with ∂τ abbreviating ∂τ12 (cf. (11.1.4)) we have

          
∂τ ϕ z(s) = ν1 z(s) ∂x2 ϕ z(s) − ν2 z(s) ∂x1 ϕ z(s)
     
= x2 (s) ∂x2 ϕ z(s) + x1 (s) ∂x1 ϕ z(s)

d   
= ϕ z(s) for L 1 -a.e. s ∈ [0, L]. (11.5.229)
ds
Using (11.5.229) and the change of variable formula from [49, Theorem 2, p. 99] we
may then compute
∫ ∫ L   d   
f ∂τ ϕ dσ = f z(s) ϕ z(s) ds
∂Ω 0 ds
∫ L
d     
=− f z(s) ϕ z(s) ds (11.5.230)
0 ds
where in the last step we have integrated by parts. That this is possible, and that
no boundary terms arise in the process, is visible from (11.5.227) (plus a limiting
argument using the fact that Sobolev functions in a finite interval on the real line
may be approximated by smooth functions simultaneously in the Sobolev norm and
in an everywhere pointwise sense). In turn, (11.5.230) and the change of variable
formula from [49, Theorem 2, p. 99] imply that
∫       
 
 f ∂τ ϕ dσ  ≤  f z(·)  L p (0, L), L 1  ϕ∂Ω  L p (∂Ω,σ), (11.5.231)
∂Ω 1

where 1/p + 1/p = 1. From this and Lemma 11.1.10 we ultimately see that
  
f ∈ L1 (Σ, σ) and  f  L p (Σ,σ) ≤ C  f z(·)  L p (0, L), L 1  ,
p
(11.5.232)
1 1

for some constant C ∈ (0, ∞) which depends only on Σ and p. This proves the
right-to-left inclusion in (11.5.224). 
784 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Our last result in this section shows that our brand of boundary Sobolev spaces is
bi-Lipschitz invariant in the sense made precise below (which should be compared
with Proposition 11.6.3, stated a little later, in the next section).

Proposition 11.5.20 Let O, O be two open sets in Rn and suppose F : O → O is a

bi-Lipschitz homeomorphism. Assume Ω ⊆ Rn is an open set satisfying a two-sided
local John condition, whose boundary is compact and Ahlfors regular, and such
that Ω ⊆ O. Then Ω := F(Ω) is also an open set in Rn satisfying a two-sided local
John condition and whose boundary is compact and Ahlfors regular. Moreover, if
σ := H n−1 ∂Ω and σ := H n−1 ∂ Ω, then for each integrability exponent p ∈ (1, ∞)
and each σ-measurable function f : ∂ Ω → R one has
p p
f ∈ L1 (∂ Ω, σ) ⇐⇒ f ◦ F ∈ L1 (∂Ω, σ), (11.5.233)

in a quantitative fashion.

Proof The first claim, pertaining to the nature of Ω is a consequence of the transfor-
mational properties of Euclidean domains established in [96]. From [96, (3.10)] and
[96, Theorem 3.4] we also know that for each σ-measurable function f : ∂ Ω → R
we have the “surface-to-surface” change of variable formula
∫ ∫
| f | dσ =
p
| f ◦ F | p J∂Ω F dσ (11.5.234)
∂Ω ∂Ω

where the “Jacobian” J∂Ω F is a function satisfying (cf. [96, Lemma 3.9])
 
cn (Lip F −1 )1−n ≤ J∂Ω F (x) ≤ Cn (Lip F)n−1
(11.5.235)
at H n−1 -a.e. point x ∈ ∂Ω,

where cn, Cn ∈ (0, ∞) are purely dimensional constants, and Lip F, Lip F −1 denote
the Lipschitz constants of F and F −1 , respectively. In particular, from (11.5.234)-
(11.5.235) we deduce that composition with F distorts norms in Lebesgue spaces
(defined on ∂Ω and ∂ Ω) by at most fixed multiplicative constants. With this in hand,
the claim in (11.5.233) readily follows making use of the (quantitative) criterion for
membership to Sobolev spaces given in item (vi) of Theorem 11.5.18. 

11.6 Boundary Sobolev Spaces on Manifolds

To facilitate discussing Sobolev spaces on the boundaries of sets of locally finite

perimeter contained in a given differentiable manifold, it is useful to establish the
invariance of the Euclidean scale of our Sobolev spaces introduced earlier under
C 1 diffeomorphisms of the Euclidean ambient. In the process, we need to know
how surface integrals transform under C 1 diffeomorphisms. This aspect has been
addressed in [96], where the following result appears.
11.6 Boundary Sobolev Spaces on Manifolds 785

Proposition 11.6.1 Let Ω ⊆ Rn be an arbitrary set of locally finite perimeter, let U

be an open neighborhood of Ω, and let F : U → Rn be an orientation preserving
C 1 -diffeomorphism onto its image.
Then Ω := F(Ω) is a set of locally finite perimeter and if ν, ν are the geometric
measure theoretic outward unit normals to Ω and Ω, respectively, then at H n−1 -a.e.
point one has  
DF −1 (ν ◦ F −1 )
ν =    (11.6.1)
 DF −1  (ν ◦ F −1 )

(with the convention that the right side of (11.6.1) is zero whenever ν ◦ F −1 = 0). In
addition, one has ∂∗ Ω = F(∂∗ Ω) and

if σ∗ := H n−1 ∂∗ Ω and σ∗ := H n−1 ∂∗ Ω then

   (11.6.2)
σ∗ =  DF −1 (ν ◦ F −1 ) |(det DF) ◦ F −1 | F∗ σ∗,

where F∗ σ∗ is the push-forward of σ∗ via F, defined by the requirement that

(F∗ σ∗ )(E) = σ∗ (F −1 (E)) for each H n−1 -measurable set E ⊆ ∂∗ Ω. (11.6.3)

Finally, if the set ∂Ω is upper (or lower) Ahlfors regular then so is ∂ Ω, and if
H n−1 (∂Ω \ ∂∗ Ω) = 0 then H n−1 (∂ Ω \ ∂∗ Ω) = 0.

Observe that (11.6.2) entails that for every g ∈ L 1 (∂∗ Ω, σ∗ ) we have the following
“surface-to-surface” change of variable formula
∫ ∫
  
g dσ∗ = (g ◦ F) DF −1 ◦ F ν  |det(DF)| dσ∗ . (11.6.4)
∂∗ Ω ∂∗ Ω

We now turn to the result that permits us to consider Sobolev spaces on bound-
aries of locally finite perimeter subsets of C 1 manifolds. As a preamble, we alter
Definition 11.1.2 as to produce a brand of local Sobolev spaces on just a portion of
the (geometric measure theoretic) boundary.

Definition 11.6.2 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, fix a relatively open subset Σ of ∂∗ Ω. Given a function
1 (Σ, σ ) along with two indices j, k ∈ {1, . . . , n}, say that ∂
f ∈ Lloc ∗ τ j k f exists in (or,
1 (Σ, σ ) if there exists a function f ∈ L 1 (Σ, σ ) such that
belongs to) the space Lloc ∗ jk loc ∗

for every open set O ⊆ Rn with the property that

O ∩ ∂∗ Ω = Σ and each function ϕ ∈ Cc1 (O) one has
∫ ∫ (11.6.5)
(∂τ j k ϕ) f dσ∗ = − ϕ f jk dσ∗ .
Σ Σ
p
In the same setting as above, for each p ∈ [1, ∞] define L1 (Σ, σ∗ ), the L p -based
Sobolev space of order one on Σ, as the collection of all functions f ∈ L p (Σ, σ∗ ) with
the property that ∂τ j k f belongs to L p (Σ, σ∗ ) (in the above sense, i.e., there exists
786 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

some function f jk ∈ L p (Σ, σ∗ ) such that (11.1.12) holds) for each pair of indices
j, k ∈ {1, . . . , n}.

The spaces considered in Definition 11.6.2 turn out to be stable under pointwise
multiplication by C 1 functions. Their invariance under C 1 diffeomorphisms of the
ambient space is studied in the proposition below (which should be compared with
Proposition 11.5.20).

Proposition 11.6.3 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, let U an open neighborhood of Ω in Rn and suppose
F : U → Rn is an orientation preserving C 1 -diffeomorphism onto its image. Pick
an open set O ⊆ Rn and define Ω := F(Ω), O := F(O ∩ U). Also, fix p ∈ [1, ∞].
Then Ω is a set of locally finite perimeter, and if σ∗ := H n−1 ∂∗ Ω, then for each
scalar-valued function f defined on O ∩ ∂∗ Ω one has

f ∈ L1,loc (O ∩ ∂∗ Ω, σ∗ ) ⇐⇒ f ◦ F −1 ∈ L1,loc (O ∩ ∂∗ Ω, σ∗ ).
p p
(11.6.6)

Proof That Ω is a set of locally finite perimeter has been noted in Proposition 11.6.1.
Also, (11.6.4) readily implies that

f ∈ Lloc (O ∩ ∂∗ Ω, σ∗ ) ⇐⇒ f ◦ F −1 ∈ Lloc (O ∩ ∂∗ Ω, σ∗ ).
p p
(11.6.7)

Handling tangential derivatives is more delicate. To set the stage, abbreviate

 
A := (Ar s )1≤r,s ≤m := (DF −1 ◦ F) = (DF)−1
(11.6.8)
and J := |det(DF)|.
p
Also, fix a function f ∈ L1,loc (O ∩ ∂∗ Ω, σ∗ ) along with numbers j, k ∈ {1, . . . , n},
and pick some arbitrary ϕ ∈ Cc1 (O). Let ν = (νi )1≤i ≤n and ν = (νi )1≤i ≤n be the
geometric measure  unit normals to Ω and Ω, respectively. If we
 theoretic outward
now set ∂τ j k ϕ := ν j ∂k ϕ − νk ∂j ϕ on O ∩ ∂∗ Ω, then based on (11.6.4) and (11.6.1)
we may write
∫ ∫
−1
 
( f ◦ F )∂τ j k ϕ dσ∗ = ( f ◦ F −1 ) ν j ∂k ϕ − νk ∂j ϕ dσ∗ (11.6.9)
∂∗ Ω ∂∗ Ω
∫
 
= f (Aν) j (∂k ϕ) ◦ F − (Aν)k (∂j ϕ) ◦ F J dσ∗ .
∂∗ Ω

Since chain rule gives that for every r ∈ {1, . . . , n},

(∂r ϕ) ◦ F = ∂s (ϕ ◦ F)(DF)−1
sr = ∂s (ϕ ◦ F)Ar s, (11.6.10)

the expression in the last set of curly brackets in (11.6.9) may be written as
11.6 Boundary Sobolev Spaces on Manifolds 787

(Aν) j (∂k ϕ) ◦ F − (Aν)k (∂j ϕ) ◦ F

= A j ν ∂s (ϕ ◦ F)Aks − Aks νs ∂ (ϕ ◦ F)A j
= A j Aks ∂τ s (ϕ ◦ F). (11.6.11)

Collectively, (11.6.9) and (11.6.11) imply

∫ ∫
( f ◦ F −1 )∂τ j k ϕ dσ∗ = f A j Aks ∂τ s (ϕ ◦ F) J dσ∗ . (11.6.12)
∂∗ Ω ∂∗ Ω

In relation to (11.6.12) we wish to observe that specializing this identity to the

case when f = 1 yields
∫
A j Aks ∂τ s (ϕ ◦ F) J dσ∗ = 0. (11.6.13)
∂∗ Ω

Ultimately, this proves that

∂τ s (A j Aks J) = 0 on ∂∗ Ω. (11.6.14)

In order to be able to take advantage of (11.6.14) we need to regularize F. As

a preamble, we recall the elementary fact that any C 1 function is locally Lipschitz
(i.e., Lipschitz when restricted to any compact subset of its original open domain of
definition). Applying this to F and F −1 it follows that for each compact set K ⊂ U
the function F : K → F(K) is bi-Lipschitz, i.e., there exists λK ∈ (1, ∞) such that
−1
λK |x − y| ≤ |F(x) − F(y)| ≤ λK |x − y|, ∀x, y ∈ K. (11.6.15)
∫
To proceed, pick θ ∈ Cc∞ (Rn ) such that supp θ ⊂ B(0, 1), θ ≥ 0, and θ = 1, and
for each ε > 0 set θ ε (x) := ε −n θ(x/ε), x ∈ Rn . Also, introduce

Ωε := {x ∈ Rn : dist(x, Ω) < ε}, ε > 0. (11.6.16)

 1 
Then for each ε ∈ 0, 2 dist(Ω, ∂U) and each x ∈ Ωε , set
∫ ∫
F ε (x) := θ ε (x − z)F(z) dz = θ ε (z)F(x − z) dz. (11.6.17)
U B(0, ε)

Hence F ε ∈ C ∞ (Ωε ) and the upper-inequality in (11.6.15) ensures that for each
compact set K ⊂ U there exists CK ∈ (0, ∞) such that

sup |F ε − F | ≤ εCK . (11.6.18)

K

Together, (11.6.15) and (11.6.18) then readily imply that for each compact set K ⊂ U
there exists εK ∈ (0, ∞) such that F ε : K → F ε (K) is bi-Lipschitz provided
0 < ε < εK . Ultimately, this shows that there exists εo > 0 with the property that
788 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

ε ∈ (0, εo ) =⇒ F ε is a C ∞ diffeomorphism from Ωε onto F ε (Ωε ). (11.6.19)

Next, denote by Aε , J ε the objects associated to F ε as in (11.6.8) and observe

that Aε → A, J ε → J uniformly on ∂Ω as ε → 0+ . Also, Aε , J ε are C ∞ in a
neighborhood of ∂Ω. Based on these observations, as well as (11.6.19), (11.6.14),
and (11.6.4), we may then compute
∫
f A j Aks ∂τ s (ϕ ◦ F) J dσ∗
∂∗ Ω
∫
= lim+ f Aεj Aks
ε
∂τ s (ϕ ◦ F ε ) J ε dσ∗
ε→0 ∂∗ Ω
∫
= − lim+ (ϕ ◦ F ε )∂τ s ( f Aεj Aks
ε
J ε ) dσ∗
ε→0 ∂∗ Ω
∫
= − lim+ (ϕ ◦ F ε )(∂τ s f )Aεj Aks
ε
J ε dσ∗
ε→0 ∂∗ Ω
∫
− lim+ (ϕ ◦ F ε ) f ∂τ s (Aεj Aks
ε
J ε ) dσ∗
ε→0 ∂∗ Ω
∫
= − lim+ (ϕ ◦ F ε )(∂τ s f )Aεj Aks
ε
J ε dσ∗
ε→0 ∂∗ Ω
∫
=− (ϕ ◦ F)(∂τ s f )A j Aks J dσ∗
∂∗ Ω
∫
A j Aks ∂τ s f
=− ◦ F −1 ϕ dσ∗ . (11.6.20)
∂∗ Ω | Aν|

This proves that

A j Aks ∂τ s f
∂τ j k ( f ◦ F −1 ) = ◦ F −1
| Aν|
   
(DF −1 ) j (DF −1 ) ks (∂τ s f ) ◦ F −1
= . (11.6.21)
|(DF −1 ) (ν ◦ F −1 )|
Since, as seen from (11.6.7), the last expression in (11.6.21) belongs to the space
p p
Lloc (O ∩ ∂∗ Ω, σ∗ ), given that ∂τ s f ∈ Lloc (O ∩ ∂∗ Ω, σ∗ ), we conclude from this
analysis and (11.6.7) that left-to-right implication in (11.6.6) holds. Since the op-
posite implication is a consequence of what we have just proved used for the C 1
diffeomorphism F −1 , the equivalence in (11.6.6) is established. 

Suppose M is a differentiable manifold of class C 1 equipped with a continuous

Riemannian metric g. Based on the module property established earlier in (11.1.68)
and the diffeomorphism invariance presented in Proposition 11.6.3 we may define
L p -based local Sobolev space of order one on the geometric measure theoretic
11.6 Boundary Sobolev Spaces on Manifolds 789

boundary of any given set Ω ⊂ M of locally finite perimeter via localization involving
a smooth partition of unity and pull-back to the Euclidean model using coordinate
p
charts. This scale of spaces, denoted by L1,loc (∂∗ Ω, σg∗ ) with 1 ≤ p ≤ ∞, where we
have set σg∗ := Hgn−1 ∂∗ Ω, continues to enjoy similar properties as in the Euclidean
setting, treated earlier in this section.
The scale of L p -based local Sobolev space of order one on the geometric measure
theoretic boundary of a set Ω of locally finite perimeter in the manifold M adapts
naturally to the case when the functions involved take values in a given Hermitian
vector bundle E → M. In such a scenario, we denote this brand of Sobolev space by
p
L1,loc (∂∗ Ω, σg∗ ) ⊗ E, where 1 ≤ p ≤ ∞.
In fact, it is possible to provide an intrinsic description of these Sobolev space.
To state such a result, we need the following piece of notation. Given two continuous
vector fields X, Y ∈ T M, for every section ϕ ∈ Cc1 (M, E) define
     
∂τXY ϕ := ν(X) ∇Y ϕ  − ν(Y ) ∇X ϕ  on ∂∗ Ω, (11.6.22)
∂∗ Ω ∂∗ Ω

where ν : ∂ ∗ Ω → T ∗ M is the geometric measure theoretic outward unit conormal to

Ω, and ∇ is a connection (or covariant derivative) on E, with continuous connection
coefficients. In particular, if in local coordinates the outward unit conormal ν is

n
expressed as ν = ν dx , then we agree to re-denote ∂τXY defined as in (11.6.22)
=1
corresponding to the choice X := ∂j and Y := ∂k simply as ∂τ j k . That is, for every
section ϕ ∈ Cc1 (M, E) define
     
∂τ j k ϕ := ν j ∇∂k ϕ  − νk ∇∂j ϕ  locally on ∂∗ Ω, (11.6.23)
∂∗ Ω ∂∗ Ω

which is in agreement with the Euclidean case considered in (11.1.4).

Proposition 11.6.4 Let M be a C 2 manifold of dimension n, equipped with a C 1
Riemannian metric g. Assume E → M is a C 1 Hermitian vector bundle over M
and denote by ·, · E the pointwise (real ) pairing in the fibers of E. Also, suppose
Ω ⊂ M is a set of locally finite perimeter and define σg∗ := Hgn−1 ∂∗ Ω. Finally, fix
an integrability exponent p ∈ [1, ∞].
p
In the geometric context considered above, a function f ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E is
p
said to belong to the local Sobolev space L1,loc (∂∗ Ω, σg∗ ) ⊗ E if and only if for any
p
two C 1 vector fields X, Y ∈ T M there exists a function hXY ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E
with the property that
∫ ∫
    
f , ∂τXY ϕ E dσg∗ = hXY , ϕ∂∗ Ω E dσg∗, ∀ϕ ∈ Cc1 (M, E).
∂∗ Ω ∂∗ Ω
(11.6.24)
p
Then the function hXY ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E doing the job in (11.6.24) is unique, and
naturally accompanying estimates are valid for each implication.
Proof With the summation convention over repeated indices used throughout, write
ν = ν j dx j for the geometric measure theoretic outward unit conormal to Ω and
790 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

assume that X = X j ∂j and Y = Yk ∂k are two arbitrary C 1 vector fields on M. Select

a coordinate patch U ⊂ M over which E may be trivialized using a smooth local
orthonormal frame {eα }α . In particular, for each j ∈ {1, . . . , n} we locally have
 αβ 
∇∂j u = ∂j uα + γ j uβ eα for each section u = uα eα ∈ E, (11.6.25)

αβ
where the coefficients γ j ’s are determined by the metric on E. Lastly, fix an open
set O ⊆ Rn such that O is a compact subset of U, and pick a scalar-valued cut-off
function ψ ∈ Cc1 (U) which is identically one in O. Then, for every C 1 section
ϕ = ϕα eα of E with compact support in O, from (11.6.22), (11.6.25), and [133,
(1.11.10)] we deduce that on U ∩ ∂∗ Ω we have
 αβ   αβ 
∂τXY ϕ = ν j X j Yk ∂k ϕα + Yk γk ϕβ eα − νk Yk X j ∂j ϕα + X j γ j ϕβ eα
 αβ 
= G−1/2 ν Ej X j Yk ∂k ϕα + Yk γk ϕβ eα
 αβ 
− G−1/2 νkEYk X j ∂j ϕα + X j γ j ϕβ eα
 αβ αβ 
= G−1/2 X j Yk (ν Ej ∂k − νkE ∂j )ϕα eα + G−1/2 X j Yk ν Ej γk − νkE γ j ϕβ eα
   αβ αβ 
= G−1/2 X j Yk ∂τEj k ϕα eα + G−1/2 X j Yk ν Ej γk − νkE γ j ϕβ eα, (11.6.26)

with ∂τEj k := ν Ej ∂k − νkE ∂j , tangential first-order operator to ∂∗ Ω in the Euclidean

setting.
p
Assume now that some function f ∈ L1,loc (∂∗ Ω, σg∗ ) ⊗ E, 1 ≤ p ≤ ∞, has been
given. Hence, f = f α eα with each component f α satisfying ψ f α ∈ L1 (∂∗ Ω, σg∗ ).
p

Also, define σ∗ := H ∂∗ Ω where H

E n−1 n−1 denotes the standard (n−1)-dimensional
Hausdorff measure in Rn . Then, on the one hand, from (11.6.26) and [133, (1.11.10)]
we conclude that
∫ ∫
    √
f , ∂τXY ϕ E dσg∗ = X j Yk ∂τEj k ϕα ψ f α g dσ∗E (11.6.27)
∂∗ Ω ∂∗ Ω
∫
 αβ αβ  √
+ X j Yk ν Ej γk − νkE γ j ϕβ f α g dσ∗E .
∂∗ Ω

On the other hand, granted the current assumptions on the vector fields X, Y and the
metric g, there exist functions hα ∈ L p (∂∗ Ω, σ∗E ) with the property that
∫ ∫
 E α  α√
X j Yk ∂τ j k ϕ ψ f g dσ∗ =
E
hα ϕα dσ∗E . (11.6.28)
∂∗ Ω ∂∗ Ω

Thus,
∫ ∫
  O
f , ∂τXY ϕ E
dσg∗ = hXY , ϕ E dσg∗ for each ϕ ∈ Cc1 (O, E), (11.6.29)
∂∗ Ω ∂∗ Ω
11.7 A General Recipe for Sobolev-Like Spaces 791

if we take

O 1  βα βα 
hXY := √ hα eα +ψ X j Yk ν j γk −νk γ j f β eα ∈ L p (∂∗ Ω, σg∗ )⊗E. (11.6.30)
gG1/2
O
Appropriately patching the functions hXY using a C 1 partition of unity then yields
p
a global function hXY ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E for which (11.6.24) holds.
p
Conversely, assume that f ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E has the property that for any
p
two C 1 vector fields X, Y ∈ T M there exists a function hXY ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E
such that (11.6.24) holds. Much as before, work in local coordinates and, given
j, k ∈ {1, . . . , n}, pick X := ∂j , Y := ∂k . Then (11.6.27) gives
∫ ∫
 E α  α√
hXY , ϕ E dσg∗ = ∂τ j k ϕ ψ f g dσ∗E
∂∗ Ω ∂∗ Ω
∫
 αβ αβ  √
+ ν j γk − νk γ j ϕβ f α g dσg∗ (11.6.31)
∂∗ Ω
√
for each function ϕ ∈ Cc1 (O, E). In turn, this proves that ψ f α g ∈ L1 (∂∗ Ω, σ∗E )
p
α p
hence, further, that f ∈ L1,loc (∂∗ Ω, σ∗ ), granted the regularity assumption
E

on the metric. This suffices to conclude that f belongs to the Sobolev space
p
L1,loc (∂∗ Ω, σg∗ ) ⊗ E.
p
Finally, [133, Proposition 3.7.2] guarantees that hXY ∈ Lloc (∂∗ Ω, σg∗ ) ⊗ E is the
unique function doing the job in (11.6.24), while the fact that there are naturally
accompanying estimates for each implication is implicit in the above calculations.

Of course, in the context of Proposition 11.6.4, we can define the global Sobolev
space
p
L1 (∂∗ Ω, σg∗ ) ⊗ E (11.6.32)
p
as the linear subspace of L1,loc (∂∗ Ω, σg∗ ) ⊗ E consisting of all functions f belonging
to L p (∂∗ Ω, σg∗ ) ⊗ E with the property that hXY ∈ L p (∂∗ Ω, σg∗ ) ⊗ E for any two C 1
vector fields X, Y ∈ T M.

11.7 A General Recipe for Sobolev-Like Spaces

We wish to remark that many of the considerations in this chapter may be naturally
generalized by using a more flexible template for constructing Sobolev-like spaces
of the following flavor. Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate
σ∗ := H n−1 ∂∗ Ω. Then, given any two subspaces A, B of Lloc 1 (∂ Ω, σ ), define
∗ ∗
 
W 1 (A, B) := f ∈ A : ∂τ j k f ∈ B for each j, k ∈ {1, . . . , n} . (11.7.1)
792 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Note that, by design, W 1 (A, B) ⊆ L1,1,1loc (∂∗ Ω, σ∗ ). Whenever A, B are quasi-normed

spaces we agree to endow W 1 (A, B) with the quasi-norm

n
W 1 (A, B)  f −→  f  A + ∂τ j k f B . (11.7.2)
j,k=1

Of course, in the case when A = B = L p (∂∗ Ω, σ∗ ) with p ∈ [1, ∞] the space

p
W 1 (A, B) becomes the boundary L p -based Sobolev space L1 (∂∗ Ω, σ∗ ), while taking
A = L p (∂∗ Ω, σ∗ ) and B = L q (∂∗ Ω, σ∗ ) with p, q ∈ [1, ∞] corresponds to the
p,q
off-diagonal Sobolev spaces L1 (∂∗ Ω, σ∗ ) Both these scales have been studied at
length here, but spaces corresponding to other choices of A, B are of considerable
interest as well. For example, taking A = B = L p,q (∂∗ Ω, σ∗ ) with p ∈ (1, ∞)
and q ∈ (0, ∞] corresponds to Lorentz-based Sobolev spaces, and choosing
different integrability exponents (for the space in which the tangential derivatives
belong) yields off-diagonal Lorentz-based Sobolev spaces, etc.
Let us fix p ∈ [1, ∞] and consider a non-negative σ∗ -measurable function w on
∂∗ Ω which, if p is the Hölder conjugate exponent of p, satisfies
p
w −1/p ∈ Lloc (∂∗ Ω, σ∗ ). (11.7.3)
p
(In particular, this condition ensures that Lloc (∂∗ Ω, wσ∗ ) embeds into Lloc 1 (∂ Ω, σ ).)
∗ ∗
Then specializing our earlier general recipe to the scenario in which the spaces
A := B := L p (∂∗ Ω, wσ∗ ) gives rise to the weighted Sobolev space
p 
L1 (∂∗ Ω, wσ∗ ) := f ∈ L p (∂∗ Ω, wσ∗ ) : ∂τ j k f ∈ L p (∂∗ Ω, wσ∗ )

for each j, k ∈ {1, . . . , n} , (11.7.4)

which we equip with the natural norm


n
p
L1 (∂∗ Ω, wσ∗ )  f −→  f  L p (∂∗ Ω,wσ∗ ) + ∂τ j k f  L p (∂∗ Ω,wσ∗ ) . (11.7.5)
j,k=1

For example, the case when w(x) := (1 + |x| a )−1 for x ∈ ∂∗ Ω with a ∈ [0, ∞) yields
the scale of “power” weighted Sobolev spaces given for each p ∈ [1, ∞] by

p σ∗ (x)   σ∗ (x)  p σ∗ (x) 
L1 ∂∗ Ω, 1+ |x | a = f ∈ L p ∂∗ Ω, 1+ |x | a : ∂τ j k f ∈ L ∂∗ Ω, 1+ |x | a (11.7.6)

for each j, k ∈ {1, . . . , n} .

This brand of boundary Sobolev spaces is going to be important in future endeavors.

To present another basic special case, assume Ω ⊆ Rn is a Lebesgue measurable
set whose topological boundary is Ahlfors regular. Introduce σ∗ := H n−1 ∂∗ Ω and
σ := H n−1 ∂Ω. Also, recall the classes of Muckenhoupt weights on (∂Ω, | · − · |, σ),
regarded as a space of homogeneous type (cf. (A.0.3)). Then an important special
11.7 A General Recipe for Sobolev-Like Spaces 793

case corresponds to the recipe in (11.7.4) for the choice w ∈ Ap (∂Ω, σ) with
p ∈ (1, ∞), giving rise to the Muckenhoupt weighted (boundary) Sobolev space
p
L1 (∂∗ Ω, wσ∗ ). In this regard, we may adopt a more general point of view and define
off-diagonal Muckenhoupt weighted (boundary) Sobolev spaces of the type
 
f ∈ L p0 (∂∗ Ω, w0 σ) : ∂τ j k f ∈ L p1 ∂∗ Ω, w1 σ) (11.7.7)
for each j, k ∈ {1, . . . , n} ,

where p0, p1 ∈ (1, ∞) and w0 ∈ Ap0 (∂Ω, σ), w1 ∈ Ap1 (∂Ω, σ).
A basic special case is as follows. Consider an Ahlfors regular domain Ω ⊆ Rn
and abbreviate σ := H n−1 ∂Ω. Having fixed an integrability exponent p ∈ (1, ∞)
along with a Muckenhoupt weight w ∈ Ap (∂Ω, σ), the (boundary) weighted Sobolev
space from (11.7.7) with p0 = p1 = p and with w0 = w1 = w becomes (with the
convention adopted in [133, (7.7.1)])
p  
L1 (∂Ω, w) := f ∈ L p (∂Ω, w) : ∂τ j k f ∈ L p (∂Ω, w), 1 ≤ j, k ≤ n . (11.7.8)

This is a Banach space when equipped with the norm


n
p
L1 (∂Ω, w)  f →  f  L p (∂Ω,w) :=  f  L p (∂Ω,w) + ∂τ j k f  L p (∂Ω,w) . (11.7.9)
1
j,k=1

q
Since there exists q ∈ (1, ∞) such that L p (∂Ω, w) → Lloc (∂Ω, σ) (cf. [133,
p q
Lemma 7.7.13]), we see that L1 (∂Ω, w) → L1,loc (∂Ω, σ) for such an exponent
p
q (in particular, the equality in (11.4.8) holds for every function f ∈ L1 (∂Ω, w)
whenever Ω is a UR domain).
In the same geometric setting, recall that L p,q (∂Ω, σ) with p, q ∈ (0, ∞] stands
for the scale of Lorentz spaces on ∂Ω, with respect to the measure σ. In relation
to this scale of spaces, it is also of interest to consider (boundary) Lorentz-based
Sobolev spaces. Specifically, for each p ∈ (1, ∞) and q ∈ (0, ∞] we set
p,q  
L1 (∂Ω, σ) := f ∈ L p,q (∂Ω, σ) : ∂τ j k f ∈ L p,q (∂Ω, σ), 1 ≤ j, k ≤ n
(11.7.10)
which is a quasi-Banach space when equipped with the quasi-norm

n
p,q
L1 (∂Ω, σ)  f →  f  L p, q (∂Ω,σ) :=  f  L p, q (∂Ω,σ) + ∂τ j k f  L p, q (∂Ω,σ) .
1
j,k=1
(11.7.11)
We may also consider Morrey-based Sobolev spaces along the lines of blue-
print described in (11.7.1)-(11.7.2). Specifically, if Ω ⊆ Rn is an Ahlfors regular
domain and σ := H n−1 ∂Ω, then for each p ∈ (1, ∞) and λ ∈ (0, n − 1) we define
794 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

p,λ
M1 (∂Ω, σ) := f ∈ M p,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ M p,λ (∂Ω, σ) , (11.7.12)

equipped with the natural norm


n
 
p,λ
M1 (∂Ω, σ)  f −→  f  M p, λ (∂Ω,σ) + ∂τ f  p, λ . (11.7.13)
jk M (∂Ω,σ)
j,k=1

More generally, in the same setting we may consider off-diagonal Morrey-based

p,q,λ
Sobolev space M1 (∂Ω, σ) defined for p, q ∈ (1, ∞) and λ ∈ (0, n − 1) as

p,q,λ
M1 (∂Ω, σ) := f ∈ M p,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ M q,λ (∂Ω, σ) , (11.7.14)

and endowed with the norm


n
 
M1
p,q,λ
(∂Ω, σ)  f −→  f  M p, λ (∂Ω,σ) + ∂τ f  q, λ . (11.7.15)
jk M (∂Ω,σ)
j,k=1

p,q,λ
A significant (closed) subspace of M1 (∂Ω, σ) is

p,q,λ
M̊1 (∂Ω, σ) := f ∈ M̊ p,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ M̊ q,λ (∂Ω, σ) . (11.7.16)

p,λ p, p,λ
Henceforth, we agree to abbreviate M̊1 (∂Ω, σ) := M̊1 (∂Ω, σ), i.e.,

p,λ
M̊1 (∂Ω, σ) := f ∈ M̊ p,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ M̊ p,λ (∂Ω, σ) . (11.7.17)

In the same vein, let us also define


p,q,λ
B1 (∂Ω, σ) := f ∈ B p,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ B q,λ (∂Ω, σ) , (11.7.18)

and endowed with the norm


n
 
B1
p,q,λ
(∂Ω, σ)  f −→  f  B p, λ (∂Ω,σ) + ∂τ f  q, λ . (11.7.19)
jk B (∂Ω,σ)
j,k=1
11.7 A General Recipe for Sobolev-Like Spaces 795
q,λ q,q,λ
Also, abbreviate B1 (∂Ω, σ) := B1 (∂Ω, σ), i.e., set


q,λ
B1 (∂Ω, σ) := f ∈ B q,λ (∂Ω, σ) : for each j, k ∈ {1, . . . , n}

one has ∂τ j k f ∈ B q,λ (∂Ω, σ) . (11.7.20)

In relation to the scales of spaces introduced above we wish to note that, given any
Lebesgue measurable set Ω ⊆ Rn whose topological boundary is Ahlfors regular,
then [133, (7.7.104)] implies that for each p ∈ (1, ∞) and w ∈ Ap (∂Ω, σ) the
p
Muckenhoupt weighted Sobolev space L1 (∂∗ Ω, wσ∗ ) embeds continuously into the
 
“power” weighted Sobolev space L11 ∂∗ Ω, 1+σ|x∗ (x)
| n−1
(cf. (11.7.6)), i.e.,
    
L1 (∂∗ Ω, wσ∗ ) → f ∈ L 1 ∂∗ Ω, 1+σ|x∗ (x) : ∂τ j k f ∈ L 1 ∂∗ Ω, 1+σ|x∗ (x)
p
| n−1 | n−1

for each j, k ∈ {1, . . . , n} continuously,

(11.7.21)

while from [133, (7.7.106)] we conclude that

if p, q ∈ [1, ∞), the off-diagonal boundary Sobolev space
 
L1 (∂∗ Ω, σ∗ ) embeds continuously into L11 ∂∗ Ω, 1+σ|x∗ (x) (11.7.22)
p,q
| n−1
.

Also, from [133, (7.7.105)] it is clear that

if ∂Ω is compact and Ahlfors regular then for each p ∈ (1, ∞) and
each w ∈ Ap (∂Ω, σ) there exists some q ∈ (1, p) with the property (11.7.23)
p q
that L1 (∂∗ Ω, wσ∗ ) → L1 (∂∗ Ω, σ∗ ) continuously.

Finally, if Ω ⊆ Rn is an Ahlfors regular domain and σ := H n−1 ∂Ω, then

(6.2.25) and (6.2.7) guarantee that for each p, q ∈ (1, ∞) and λ ∈ (0, n − 1) we have
the continuous embeddings
 
(∂Ω, σ) → L11 ∂Ω, 1+σ(x)
p,q,λ p,q,λ
M̊1 (∂Ω, σ) → M1 |x | n−1
, (11.7.24)

p,λ p,λ p(n−1)

L1s (∂Ω, σ) → M̊1 (∂Ω, σ) → M1 (∂Ω, σ) with s := n−1−λ ∈ (p, ∞),
(11.7.25)

while from (6.2.71) and [133, (7.7.106)] we see that, for each p, q ∈ (1, ∞) and
λ ∈ (0, n − 1), we have the continuous embeddings
 
(∂Ω, σ) → L11 ∂Ω, 1+σ(x)
p,q,λ
B1 |x | n−1
, (11.7.26)

q,λ q(n−1)
B1 (∂Ω, σ) → L1r (∂Ω, σ) with r := n−1+λ(q−1) ∈ (1, q). (11.7.27)
796 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

11.8 Boundary Sobolev Spaces of Negative Smoothness

We first make the following definition, introducing the scale of boundary Sobolev
spaces of negative smoothness.
Definition 11.8.1 Given an arbitrary set Ω ⊆ Rn of locally finite perimeter, abbre-
viate σ∗ := H n−1 ∂∗ Ω and for each p ∈ (1, ∞) define the (L p -based) negative
Sobolev space
∗
p p
L−1 (∂∗ Ω, σ∗ ) := L1 (∂∗ Ω, σ∗ ) where 1
p + 1
p = 1. (11.8.1)

In particular, from (11.1.66), Lemma 1.2.1, and (11.1.65) we see that in the setting
of Definition 11.8.1 the mapping
p
ι : L p (∂∗ Ω, σ∗ ) −→ L−1 (∂∗ Ω, σ∗ )
∫
 
p
L (∂∗ Ω,σ∗ ) ι( f ), g p 
L (∂ Ω,σ )
:= f g dσ∗ (11.8.2)
−1 ∗ ∗
1 ∂∗ Ω
p 
∀g ∈ L1 (∂∗ Ω, σ∗ ) ⊆ L p (∂∗ Ω, σ∗ ),

is well defined, linear, continuous, and injective. As a consequence of this, via the
identification
p
L p (∂∗ Ω, σ∗ )  f ≡ ι( f ) ∈ L−1 (∂∗ Ω, σ∗ ), (11.8.3)
we have the embedding
p
L p (∂∗ Ω, σ∗ ) → L−1 (∂∗ Ω, σ∗ )
(11.8.4)
continuously and densely, for each p ∈ (1, ∞).

In this vein, let us also point out that, in light of the identification made in (11.8.3),
we have
∫
 
L (∂∗ Ω,σ∗ ) f , g L p (∂ Ω,σ ) = f g dσ∗ for each
p 
−1 ∗ ∗
1 ∂∗ Ω
p p 
f ∈ L p (∂∗ Ω, σ∗ ) → L−1 (∂∗ Ω, σ∗ ) and g ∈ L1 (∂∗ Ω, σ∗ )
→ L p (∂∗ Ω, σ∗ ).
(11.8.5)
Finally, we wish to note here that, thanks to (11.8.1) and (11.1.65), we have
∗
p p
L−1 (∂∗ Ω, σ∗ ) = L1 (∂∗ Ω, σ∗ ) where 1
p + 1
p = 1. (11.8.6)

It turns out that the negative Sobolev space (11.8.1) may also be characterized
explicitly. Specifically, we have the following result.
Proposition 11.8.2 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate
σ∗ := H n−1 ∂∗ Ω. Also, fix 1 < p, p < ∞ such that 1/p + 1/p = 1. Then there exists
p
a finite constant C > 0 with the property that for each given f ∈ L−1 (∂∗ Ω, σ∗ ) there
exist f0, f jk ∈ L (∂∗ Ω, σ∗ ), 1 ≤ j < k ≤ n, satisfying
p
11.8 Boundary Sobolev Spaces of Negative Smoothness 797

 f0  L p (∂∗ Ω,σ∗ ) +  f jk  L p (∂∗ Ω,σ∗ ) ≤ C f  L p (∂∗ Ω,σ∗ ) (11.8.7)
−1
1≤ j<k ≤n

as well as
∫ 
 
p
L−1 (∂∗ Ω,σ∗ ) f, g p
L1 (∂∗ Ω,σ∗ )
= f0 g + f jk ∂τ j k g dσ∗
∂∗ Ω 1≤ j<k ≤n (11.8.8)
p
for every function g ∈ L1 (∂∗ Ω, σ∗ ).

Proof Recall the mapping J defined in (11.1.74) and observe that

p
J −1 : Range J −→ L1 (∂∗ Ω, σ∗ ) (11.8.9)
p  p ∗
is an isomorphism. For each fixed functional f ∈ L−1 (∂∗ Ω, σ∗ ) = L1 (∂∗ Ω, σ∗ ) ,
consider the composition

f ◦ J −1 : Range J −→ R. (11.8.10)
   1+n(n−1)/2
Since Range J is a closed subspace of L p (∂∗ Ω, σ∗ ) , Hahn-Banach’s
Extension Theorem in concert with Riesz’s Representation Theorem ensure the
existence of a family of functions f0, f jk ∈ L p (∂∗ Ω, σ∗ ), with 1 ≤ j < k ≤ n, such
that all properties listed in the statement hold. 

Proposition 11.8.2 is going to be useful throughout. For now, we use it to prove

the following weak-star convergence result.

Lemma 11.8.3 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, fix 1 < p, p < ∞ such that 1/p + 1/p = 1 and select a
p p
function φ ∈ L1 (∂∗ Ω, σ∗ ). Suppose {φi }i ∈N ⊆ L1 (∂∗ Ω, σ∗ ) is a sequence with the
property that
  ∗
φi → φ weak-∗ in L p (∂∗ Ω, σ∗ ) = L p (∂∗ Ω, σ∗ ) as i → ∞ (11.8.11)

and, for each j, k ∈ {1, . . . , n},

  ∗
∂τ j k φi → ∂τ j k φ weak-∗ in L p (∂∗ Ω, σ∗ ) = L p (∂∗ Ω, σ∗ ) as i → ∞. (11.8.12)

Then8
p  p ∗
φi → φ weak-∗ in L1 (∂∗ Ω, σ∗ ) = L−1 (∂∗ Ω, σ∗ ) as i → ∞. (11.8.13)
p
Proof To set the stage, fix an arbitrary functional f ∈ L−1 (∂∗ Ω, σ∗ ). Let the func-

tions f0, f jk ∈ L p (∂∗ Ω, σ∗ ), with 1 ≤ j < k ≤ n, be associated with f as in
Proposition 11.8.2 (presently employed with the roles of p and p reversed). Then
based on (11.8.8) (used twice), (11.8.11), and (11.8.11) we may compute

8 recall the duality result from (11.8.6)

798 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ 
 
lim p
L−1 (∂∗ Ω,σ∗ )
f , φi p
L1 (∂∗ Ω,σ∗ )
= lim f0 φi + f jk ∂τ j k φi dσ∗
i→∞ i→∞ ∂∗ Ω 1≤ j<k ≤n
∫ 
= f0 φ + f jk ∂τ j k φ dσ∗
∂∗ Ω 1≤ j<k ≤n
 
= p
L−1 (∂∗ Ω,σ∗ )
f, φ p
L1 (∂∗ Ω,σ∗ )
(11.8.14)

from which (11.8.13) follows. 

The manner in which boundary Sobolev spaces with negative smoothness embed
into the space of distributions is clarified in the proposition below.

Proposition 11.8.4 Assume Ω ⊆ Rn is an open set satisfying a two-sided local

John condition and whose boundary is compact and Ahlfors regular. Abbreviate
σ := H n−1 ∂Ω and fix p, p ∈ (1, ∞) such that 1/p + 1/p = 1. Then the mapping
p  p ∗
which assigns to each f ∈ L−1 (∂Ω, σ), regarded as a functional in L1 (∂Ω, σ)
p 
(cf. (11.8.1)), its restriction to Lip (∂Ω), viewed as a subspace of L1 (∂Ω, σ) (cf.
(11.1.109)), induces
 a well-defined,
 linear, injective, and continuous mapping from
p
L−1 (∂Ω, σ) into Lip (∂Ω) . As a consequence, one has a continuous embedding
p  
L−1 (∂Ω, σ) → Lip (∂Ω) (11.8.15)
p  
and for each f ∈ L−1 (∂Ω, σ) ⊆ Lip (∂Ω) (with the inclusion interpreted as in
p
(11.8.15)) and each φ ∈ Lip (∂Ω) ⊆ L1 (∂Ω, σ) (with the inclusion provided by
(11.1.109)) one has
   
L (∂Ω,σ) f , φ L p  (∂Ω,σ) = (Lip (∂Ω)) f , φ Lip (∂Ω) . (11.8.16)
p
−1 1

Proof The only delicate issue is the injectivity of the mapping assigning to each
p  p ∗
f ∈ L−1 (∂Ω, σ), viewed as a functional in L1 (∂Ω, σ) , the distribution obtained
by restricting said functional to Lip (∂Ω). This, however, is a consequence of the
density result in (11.5.200). 

Fix p, p ∈ (1, ∞) such that 1/p + 1/p = 1, along with j, k ∈ {1, . . . , n}. If
Ω ⊆ Rn is a set of locally finite perimeter and σ∗ := H n−1 ∂∗ Ω, then taking the
transpose of the linear and bounded operator (cf. (11.1.67))
p 
∂τ j k : L1 (∂∗ Ω, σ∗ ) −→ L p (∂∗ Ω, σ∗ ) (11.8.17)

implies that

∂τj k : L p (∂∗ Ω, σ∗ ) −→ L−1 (∂∗ Ω, σ∗ )

p
(11.8.18)
is a well-defined, linear, and bounded operator.
11.8 Boundary Sobolev Spaces of Negative Smoothness 799

Under stronger assumptions on the underlying set Ω, the above operator turns out
p
to act in a consistent fashion with −∂τ j k considered on L1 (∂∗ Ω, σ∗ ) ⊆ L p (∂∗ Ω, σ∗ ).
Specifically, we have the following result.

Proposition 11.8.5 Assume Ω ⊆ Rn is an open set satisfying a two-sided local

John condition and whose boundary is compact and Ahlfors regular. Abbreviate
σ := H n−1 ∂Ω and fix p ∈ (1, ∞). Then, for each j, k ∈ {1, . . . , n}, the tangential
derivative operator (cf. (11.1.67))
p
∂τ j k : L1 (∂Ω, σ) −→ L p (∂Ω, σ) (11.8.19)

extends uniquely to a well-defined, linear, and bounded mapping

p
∂τ j k : L p (∂Ω, σ) −→ L−1 (∂Ω, σ). (11.8.20)

This extension of the tangential derivative operator ∂τ j k has the property that for
p  p ∗
each pair of functions, f ∈ L p (∂Ω, σ) and g ∈ L1 (∂Ω, σ) = L−1 (∂Ω, σ) with
p ∈ (1, ∞) such that 1/p + 1/p = 1, one has the following “integration by parts on
the boundary” formula:
∫
 
p
L (∂Ω,σ) ∂τ jk f , g p 
L (∂Ω,σ)
= − f ∂τ j k g dσ. (11.8.21)
−1 1 ∂Ω

As a consequence, in the current geometric context one has the (quantitative)

identification
 
p
L−1 (∂Ω, σ) = f0 + ∂τ j k f jk : f0, f jk ∈ L p (∂Ω, σ) . (11.8.22)
1≤ j<k ≤n

Proof Taking transpose of the version of (11.8.18) written for p in place of p shows
that the operator

∂τj k : L p (∂Ω, σ) → L−1 (∂Ω, σ) is well defined, linear, and bounded.

p
(11.8.23)

In relation to this, we claim that

 p ∗
∂τ j k f = −∂τj k f in L−1 (∂Ω, σ) = L1 (∂Ω, σ)
p

p
(11.8.24)
for each f ∈ L1 (∂Ω, σ) ⊆ L p (∂Ω, σ).

To prove (11.8.24), on account of the density result recorded in (11.5.199) it suffices

to check that
      
 
L−1 (∂Ω,σ) ∂τ j k f , ψ ∂Ω L p  (∂Ω,σ) = − L−1 (∂Ω,σ) ∂τ j k f , ψ ∂Ω L p  (∂Ω,σ)
p p
1 1 (11.8.25)
for each f ∈ L1 (∂Ω, σ) and ψ ∈ Cc∞ (Rn ).
p

p
To see that this is indeed the case, fix an arbitrary f ∈ L1 (∂Ω, σ) along with some
ψ ∈ Cc∞ (Rn ) and note that thanks to (11.8.5), (11.1.18), and (11.8.23) we have
800 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
  
p
L−1 (∂Ω,σ) ∂τ j k f , ψ ∂Ω L p (∂Ω,σ)
1
∫ ∫
   
= ∂τ j k f ψ dσ = − f ∂τ j k ψ dσ
∂Ω ∂Ω
  
= − L p (∂Ω,σ) ∂τj k f , ψ ∂Ω L p (∂Ω,σ)
. (11.8.26)
−1 1

This establishes (11.8.25) which, in turn, finishes the proof of (11.8.24).

So far, we have shown that −∂τj k : L p (∂Ω, σ) → L−1 (∂Ω, σ) is well defined, lin-
p
p
ear, and bounded operator which agrees with (11.8.19) on L1 (∂Ω, σ) ⊆ L p (∂Ω, σ).
In view of (11.8.4), we conclude that this is the unique extension of (11.8.19) to a
p
linear and bounded mapping from L p (∂Ω, σ) into L−1 (∂Ω, σ). Retaining the symbol
p
∂τ j k for said extension, it follows that for each f ∈ L p (∂Ω, σ) and g ∈ L1 (∂Ω, σ)
we have
    
L (∂Ω,σ) ∂τ j k f , g L p  (∂Ω,σ) = − L (∂Ω,σ) ∂τ j k f , g p
p p
−1 1 −1 L1 (∂Ω,σ)
∫
=− f ∂τ j k g dσ, (11.8.27)
∂Ω

proving (11.8.21). Finally, (11.8.22) is a consequence of Proposition 11.8.2 and

(11.8.21). 

Similar considerations apply to the off-diagonal Sobolev spaces. Specifically, let

Ω ⊆ Rn be a set of locally finite perimeter and abbreviate σ∗ := H n−1 ∂∗ Ω. For
each p, q ∈ (1, ∞) define the off-diagonal negative Sobolev spacespace
∗
p,q p,q
L−1 (∂∗ Ω, σ∗ ) := L1 (∂∗ Ω, σ∗ )
(11.8.28)
here 1
p + 1
p = 1 and 1
q + 1
q = 1.

Then
p, p p
L−1 (∂∗ Ω, σ∗ ) = L−1 (∂∗ Ω, σ∗ ) for each p ∈ (1, ∞), (11.8.29)
and Proposition 11.1.9 implies that
p,q
L p (∂∗ Ω, σ∗ ) → L−1 (∂∗ Ω, σ∗ )
(11.8.30)
continuously and densely, for each p, q ∈ (1, ∞).

In a completely analogous fashion to Proposition 11.8.2, we also have the following

characterization of the spaces just introduced.

Proposition 11.8.6 Let Ω ⊆ Rn be a set of locally finite perimeter and abbreviate

σ∗ := H n−1 ∂∗ Ω. Also, fix 1 < p, q, p, q  < ∞ such that 1/p + 1/p = 1 and
1/q + 1/q  = 1. Then there exists a finite constant C > 0 with the property that
p,q
for each given f ∈ L−1 (∂∗ Ω, σ∗ ) there exist f0 ∈ L p (∂∗ Ω, σ∗ ), along with some
f jk ∈ L (∂∗ Ω, σ∗ ) for 1 ≤ j < k ≤ n, satisfying
q
11.8 Boundary Sobolev Spaces of Negative Smoothness 801

 f0  L p (∂∗ Ω,σ∗ ) +  f jk  L q (∂∗ Ω,σ∗ ) ≤ C f  L p, q (∂∗ Ω,σ∗ ) (11.8.31)
−1
1≤ j<k ≤n

and also
∫ 
 
p, q
L−1 (∂∗ Ω,σ∗ ) f, g p , q 
L1 (∂∗ Ω,σ∗ )
= f0 g + f jk ∂τ j k g dσ∗
∂∗ Ω 1≤ j<k ≤n (11.8.32)
p,q
for every function g ∈ L1 (∂∗ Ω, σ∗ ).

Moving on, we introduce Muckenhoupt weighted negative Sobolev spaces on

boundaries of Ahlfors regular domains.

Definition 11.8.7 Let Ω ⊆ Rn be an Ahlfors regular domain and set σ := H n−1 ∂Ω.
Fix an exponent p ∈ (1, ∞) along with a Muckenhoupt weight w ∈ Ap (∂Ω, σ)
and recall (11.7.8). In this setting, define the Muckenhoupt weighted negative
Sobolev space
∗
p
L−1 (∂Ω, w) := L1 (∂Ω, w )
p
(11.8.33)

where p := (1 − 1/p)−1 ∈ (1, ∞) is the Hölder conjugate exponent of p, and


where w  := w 1−p ∈ Ap (∂Ω, σ) is the conjugate weight of w (cf. item (2) in [133,
Lemma 7.7.1]).

In relation to this, it is worth making the following observation:

Remark 11.8.8 For the brand of Sobolev spaces considered in Definition 11.8.7,
similar properties to (11.8.2), (11.8.4), (11.8.5), and Proposition 11.8.2 are valid as
well.

Going further, we introduce Morrey-based, and block-based, negative Sobolev

spaces, of the following sort:

Definition 11.8.9 Let Ω ⊆ Rn be an Ahlfors regular domain and set σ := H n−1 ∂Ω.
Fix p, q ∈ (1, ∞) with 1/p + 1/q = 1 along with some λ ∈ (0, n − 1). In this setting,
define the Morrey-based negative Sobolev space
∗
p,λ q,λ
M−1 (∂Ω, σ) := B1 (∂Ω, σ) , (11.8.34)

and the block-based negative Sobolev space

∗
q,λ p,λ
B−1 (∂Ω, σ) := M̊1 (∂Ω, σ) . (11.8.35)

It is useful to have the following characterization of the Morrey-based negative

Sobolev space (11.8.34).

Proposition 11.8.10 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be an Ahlfors regular

domain and abbreviate σ := H n−1 ∂Ω. Also, fix p, q ∈ (1, ∞) with 1/p + 1/q = 1
along with some λ ∈ (0, n − 1). Then there exists a finite constant C > 0 with
802 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
p,λ
the property that for each given f ∈ M−1 (∂Ω, σ) there exist a family of functions
f0, f jk ∈ M p,λ (∂Ω, σ), indexed by 1 ≤ j < k ≤ n, satisfying

 f0  M p, λ (∂Ω,σ) +  f jk  M p, λ (∂Ω,σ) ≤ C f  M p, λ (∂Ω,σ) (11.8.36)
−1
1≤ j<k ≤n

as well as
∫ 
 
p, λ
M−1 (∂Ω,σ)
f, φ q, λ
B1 (∂Ω,σ)
= f0 φ + f jk ∂τ j k φ dσ
∂Ω 1≤ j<k ≤n (11.8.37)
q,λ
for every function φ ∈ B1 (∂Ω, σ).
q,λ
Also, there exists a constant C ∈ (0, ∞) such that for each g ∈ B−1 (∂Ω, σ) there
exist g0, g jk ∈ B q,λ (∂Ω, σ), indexed by 1 ≤ j < k ≤ n, satisfying

g0  B q, λ (∂Ω,σ) + g jk  B q, λ (∂Ω,σ) ≤ Cg B q, λ (∂Ω,σ) (11.8.38)
−1
1≤ j<k ≤n

as well as
∫ 
 
q, λ
B−1 (∂Ω,σ)
g, ψ p, λ
M̊1 (∂Ω,σ)
= g0 ψ + g jk ∂τ j k ψ dσ
∂Ω 1≤ j<k ≤n (11.8.39)
p,λ
for every function ψ ∈ M̊1 (∂Ω, σ).

Proof All claims in the statement may be justified using the Hahn-Banach Extension
Theorem plus the duality results from Proposition 6.2.8 and Proposition 6.2.16, much
as in the proof of Proposition 11.8.2. 

Our next proposition sheds light on the manner in which Lebesgue-based negative
Sobolev spaces embed into Morrey-based negative Sobolev spaces.

Proposition 11.8.11 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is compact and Ahlfors regular. Set σ := H n−1 ∂Ω.
Fix p, q ∈ (1, ∞) with 1/p + 1/q = 1 along with some λ ∈ (0, n − 1), then consider
the Hölder conjugate exponents

p(n − 1) q(n − 1)
s := ∈ (p, ∞) and r := ∈ (1, q). (11.8.40)
n−1−λ n − 1 + λ(q − 1)
 
s (∂Ω, σ) = L r (∂Ω, σ) ∗ define the functional
For each f ∈ L−1 1

q,λ
Λ f : B1 (∂Ω, σ) −→ C given by
  q,λ
(11.8.41)
Λ f (g) := L−1
s (∂Ω,σ) f , g
L r (∂Ω,σ)
for each g ∈ B1 (∂Ω, σ).
1

Then the assignment

11.8 Boundary Sobolev Spaces of Negative Smoothness 803
∗
q,λ
s
L−1 (∂Ω, σ)  f −→ Λ f ∈ B1 (∂Ω, σ) (11.8.42)

is well defined, linear, continuous, and injective. In other words, one has the following
continuous embedding:
∗
q,λ p,λ
s
L−1 (∂Ω, σ) → B1 (∂Ω, σ) = M−1 (∂Ω, σ) (11.8.43)

Moreover, the following compatibility result is valid:

   
s (∂Ω,σ) f , g = M p, λ (∂Ω,σ) f , g B q, λ (∂Ω,σ)
L−1 L r (∂Ω,σ) 1 −1 1

s (∂Ω, σ) → M p,λ (∂Ω, σ) (cf. (11.8.43))

for all f ∈ L−1 (11.8.44)
−1
q,λ
and all g ∈ B1 (∂Ω, σ) → L1r (∂Ω, σ) (cf. (11.7.27)).

Proof Essentially, this follows from Lemma 1.2.1. Specifically, from (11.8.41) and
(11.7.27) it follows that there exists some constant C = C(Ω, n, p, λ) ∈ (0, ∞) with
∗
q,λ
the property that for each f ∈ L−1
s (∂Ω, σ) one has Λ ∈ B
f 1 (∂Ω, σ) and

Λ f (B q, λ (∂Ω,σ))∗ ≤ C f  L−1

s (∂Ω,σ) . (11.8.45)
1

Hence, the assignment (11.8.42) is indeed well defined, linear, and continuous. To
s (∂Ω, σ) is such that
show that this is also injective, assume f ∈ L−1
  q,λ
s (∂Ω,σ)
L−1 f, g L1r (∂Ω,σ)
= 0 for each g ∈ B1 (∂Ω, σ). (11.8.46)

In view of (6.2.70), this implies

  
s (∂Ω,σ) f , ψ  = 0 for each ψ ∈ Cc∞ (Rn ) (11.8.47)
L−1 ∂Ω L r (∂Ω,σ) 1

s (∂Ω, σ).
which, thanks to (11.5.199) (used with p replaced by r), forces f = 0 in L−1
Consequently, the assignment (11.8.42) is also injective. Finally, (11.8.44) is seen
from (1.2.25). 

In turn, Proposition 11.8.11 makes it possible for us to make the following defi-
nition, introducing the scale of negative Sobolev spaces based on vanishing Morrey
spaces.

Definition 11.8.12 Assume Ω ⊆ Rn is an open set satisfying a two-sided local

John condition and whose boundary is compact and Ahlfors regular. Abbreviate
σ := H n−1 ∂Ω. Fix an exponent p ∈ (1, ∞) along with some λ ∈ (0, n − 1) and
p,λ
consider s := p(n−1)
n−1−λ ∈ (p, ∞). In this setting, define M̊−1 (∂Ω, σ) to be the closure
s (∂Ω, σ) in M p,λ (∂Ω, σ) (cf. (11.8.43)).
of L−1 −1

p,λ p,λ
Hence, by design, M̊−1 (∂Ω, σ) is a closed subspace of M−1 (∂Ω, σ), and we shall
endow the former with the norm naturally inherited from the latter. In particular,
p,λ
M̊−1 (∂Ω, σ) becomes a Banach space.
804 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

11.9 Principal-Value Distributions on Countably Rectifiable Sets

Here we discuss the notion of principal-value distribution on a given countably

rectifiable set. In this regard, it is useful to recall the discussion in [133, § 4.1].

Proposition 11.9.1 Let Σ ⊆ Rn (where n ≥ 2) be a Borel set which is countably

rectifiable (of dimension n − 1) and upper Ahlfors regular (e.g., any UR set will
p
do). Abbreviate σ := H n−1 Σ and fix a function b ∈ Lloc (Σ, σ) for some exponent
p ∈ (n − 1, ∞]. Finally, consider an odd function k ∈ C 2 (Rn \ {0}) which satisfies,
for some C ∈ (0, ∞),

|k(x)| ≤ C|x| 1−n, |(∇k)(x)| ≤ C|x| −n, |(∇2 k)(x)| ≤ C|x| −1−n, (11.9.1)

for all x ∈ Rn \ {0}. Then for σ-a.e. point x ∈ Σ the functional

  
P.V. b k(x − ·)Σ : Lipc (Σ) −→ C (11.9.2)

acting on each φ ∈ Lipc (Σ) according to

 ∫
   

P.V. b k(x − ·) Σ , φ := lim+ b(y)k(x − y)φ(y) dσ(y), (11.9.3)
ε→0
y ∈Σ
|y−x |>ε

is a well-defined distribution on Σ.

Proof Since for each number m ∈ N the function b · 1B(0,m)∩Σ belongs to the space
p
Lcomp (Σ, σ) ⊆ L 1 (Σ, σ), [133, Corollary 5.3.6] guarantees the existence of some
σ-nullset Am ⊆ Σ with the property that the limit
∫
lim+ b(y)k(x − y) dσ(y) exists for each x ∈ Σ \ Am . (11.9.4)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε
*
Hence, A := m∈N Am is a σ-measurable subset of Σ satisfying σ(A) = 0 and such
that the limit
∫
lim+ b(y)k(x − y) dσ(y) exists for all m ∈ N and all x ∈ Σ \ A. (11.9.5)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε

Given any m ∈ N, any φ ∈ Lipc (Σ) with supp φ ⊆ Σ ∩ B(0, m), and any x ∈ Σ \ A,
we may then write
11.9 Principal-Value Distributions on Countably Rectifiable Sets 805
∫
lim b(y)k(x − y)φ(y) dσ(y)
ε→0+
y ∈Σ
|y−x |>ε
∫
= lim+ b(y)k(x − y)φ(y) dσ(y)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε
∫
 
= lim+ b(y)k(x − y) φ(y) − φ(x) dσ(y)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε
+ ∫ ,
+ φ(x) lim+ b(y)k(x − y) dσ(y) . (11.9.6)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε

  −1
Let p := 1 − p1 be the Hölder conjugate exponent of p. Note that for each m ∈ N
and each x ∈ Σ we have
∫
 
|b(y)||k(x − y)| φ(y) − φ(x) dσ(y) (11.9.7)
Σ∩B(0,m)
∫
|b(y)|
≤ CφLip (Σ) dσ(y)
|y − x| n−2
Σ∩B(0,m)

+ ∫ , 1/p
dσ(y)
≤ CφLip (Σ) b L p (Σ∩B(0,m),σ)  ,
|y − x| (n−2)p
Σ∩B(0,m)

and that thanks to [133, (7.2.5)] we have

+ ∫ , 1/p
dσ(y)
 < +∞ (11.9.8)
|y − x| (n−2)p
Σ∩B(0,m)

since p > n − 1 forces (n − 2)p < n − 1. As such, we may invoke Lebesgue’s

Dominated Convergence Theorem to conclude that, for each m ∈ N, each x ∈ Σ, and
each φ ∈ Lipc (Σ) with supp φ ⊆ Σ ∩ B(0, m) we have
806 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫
 
lim+ b(y)k(x − y) φ(y) − φ(x) dσ(y) (11.9.9)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε
∫
 
= b(y)k(x − y) φ(y) − φ(x) dσ(y)
Σ∩B(0,m)

and
∫
 
|b(y)||k(x − y)| φ(y) − φ(x) dσ(y) ≤ CφLip (Σ) (11.9.10)
Σ∩B(0,m)

for some constant C = C(Σ, k, b, m, x) ∈ (0, ∞) independent of φ. At this stage, from

(11.9.6), (11.9.5), and (11.9.9) we conclude that for each fixed point x ∈ Σ \ A the
functional in (11.9.2)-(11.9.2) is meaningfully defined. In fact, its action on each
φ ∈ Lipc (Σ) may be expressed as
 ∫
     
P.V. b k(x − ·)Σ , φ = b(y)k(x − y) φ(y) − φ(x) dσ(y)
Σ∩B(0,m)
+ ∫ ,
+ φ(x) lim+ b(y)k(x − y) dσ(y) (11.9.11)
ε→0
y ∈Σ∩B(0,m)
|y−x |>ε

where m ∈ N is any integer such that supp φ ⊆ Σ ∩ B(0, m). In light of this, (11.9.10),
and [133, Proposition
   we ultimately conclude that for σ-a.e. point x ∈ Σ the
4.1.2]
functional P.V. b k(x − ·)Σ is indeed a well-defined distribution on Σ. 
In more special circumstances, like the ones described in our next proposition, it
turns out that the principal-value distribution defined earlier actually belongs to the
weak Hardy space H 1,∞ .
Proposition 11.9.2 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be a UR domain.
Abbreviate σ := H n−1 ∂Ω and denote by ν the geometric measure theoretic outward
unit normal to Ω. Also, consider
 n
a vector field k ∈ C 2 (Rn \ {0}) which is
odd, positive homogeneous of degree 1 − n, (11.9.12)
and which satisfies div k = 0 in Rn \ {0}.

Also, recall the brand of principal-value distribution defined in Proposition 11.9.1

(currently used with Σ := ∂Ω). Then one has the following results:
(1) For σ-a.e. point xo ∈ ∂Ω the functional
  
P.V. ν, k(xo − ·)∂Ω : Lipc (∂Ω) −→ C, (11.9.13)
11.9 Principal-Value Distributions on Countably Rectifiable Sets 807

acting on each φ ∈ Lipc (∂Ω) according to

 ∫
     
P.V. ν, k(xo − ·)∂Ω , φ := lim+ ν(y), k(xo − y) φ(y) dσ(y),
ε→0
y ∈∂Ω
|y−x |>ε
(11.9.14)
is a well-defined distribution on ∂Ω which belongs to the weak Hardy space
H 1,∞ (∂Ω, σ) in a uniform fashion, i.e., there exists C ∈ (0, ∞) such that
    

P.V. ν, k(xo − ·)∂Ω  1, ∞ ≤ C for σ-a.e. xo ∈ ∂Ω. (11.9.15)
H (∂Ω,σ)

(2) The difference

     
P.V. ν, k(x0 − ·)∂Ω − P.V. ν, k(x1 − ·)∂Ω
% (11.9.16)
<p<1 H (∂Ω, σ) for σ-a.e. x0, x1 ∈ ∂Ω.
belongs to p
n−1
n

Proof Consider the claims made in item (1). To get things started, first observe that
Proposition 11.9.1 (whose applicability in the present setting is ensured by (11.9.12))
implies that there exists a σ-measurable subset A of ∂Ω satisfying σ(A) = 0 and
such that
  
P.V. ν, k(xo − ·)∂Ω is a well-defined distribution on ∂Ω, in (11.9.17)
the sense of (11.9.2)-(11.9.3), for each point xo ∈ ∂Ω \ A.

To proceed, fix an arbitrary point xo ∈ ∂Ω \ A and bring in the vector field

F := k(xo − ·) in Ω. (11.9.18)

Then (11.9.18) and (11.9.12) permit us to conclude that

 n
F ∈ C 2 (Ω \ {xo }) ∩ Lbdd
1
(Ω, L n ) and divF = 0 in Ω. (11.9.19)

In addition, having picked some aperture parameter κ > 0, from [133, Lemma 8.3.7]
we see that there exists a constant C = C(Ω, n, k, κ) ∈ (0, ∞), independent of xo ,
with the property that
  C
Nκ F (x) ≤ for σ-a.e. x ∈ ∂Ω. (11.9.20)
|x − xo | n−1
In turn, from (11.9.20), [133, (6.2.22)] (used with d := n−1), [133, Proposition 8.2.3]
and [133, (6.2.16)] we conclude that there exists a constant C ∈ (0, ∞), independent
of xo , such that
 
Nκ F ∈ L 1,∞ (∂Ω, σ) and Nκ F  L 1, ∞ (∂Ω,σ) ≤ C. (11.9.21)
808 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Granted this, we may invoke Theorem 10.2.1 (with p := 1 and q := ∞) and conclude
that there exists a constant C = C(Ω, k, κ) ∈ (0, ∞) with the property that

ν • F belongs to the weak Hardy space H 1,∞ (∂Ω, σ)

  (11.9.22)
and satisfies the estimate ν • F  1,∞ ≤ C.H (∂Ω,σ)

Going further, pick an arbitrary test function ψ ∈ Lipc (∂Ω) and consider any

Ψ ∈ Lip(Ω) with Ψ∂Ω = ψ on ∂Ω, as
well as Ψ ≡ 0 outside of some compact (11.9.23)
subset of Ω.
Also, define the vector field
 n
G := Ψ F ∈ Liploc (Ω \ {xo }) , (11.9.24)

and consider the L 1 -nullset N ⊂ (0, ∞) associated with the point xo and the vector
field G as in [133, Lemma 5.7.2]. We may then write
∫
 
ν • F, ψ = F · ∇Ψ dL n
Ω
∫ ∫
= lim F · ∇Ψ dL n = lim divG dL n
ε ∈(0,∞)\N Ω\B(x o ,ε) ε ∈(0,∞)\N Ω\B(x o ,ε)
ε→0+ ε→0+
∫
= lim ν · G dσ
ε ∈(0,∞)\N ∂Ω\B(x o ,ε)
ε→0+
∫
y − xo
− lim · G(y) dH n−1 (y)
ε ∈(0,∞)\N Ω∩∂B(x o ,ε) ε
ε→0+
∫
 
= lim ν, k(xo − ·) ψ dσ
ε ∈(0,∞)\N ∂Ω\B(x o ,ε)
ε→0+
∫ y − x 
o
− lim , k(xo − y) ψ(y) dH n−1 (y)
ε ∈(0,∞)\N Ω∩∂B(x o ,ε) ε
ε→0+

=: I − II. (11.9.25)

Above, the first equality is implied by (10.2.138) (bearing in mind (11.9.23) and
the last property in (11.9.19)), the second equality is a consequence of Lebesgue’s
Dominated Convergence Theorem (cf. (11.9.19) and (11.9.23) in this regard), the
third equality is seen from (11.9.24), the fourth equality is a consequence of [133,
(5.7.27)] in [133, Lemma 5.7.2] (written with B(x, r) replaced by B(x, r), and keeping
11.9 Principal-Value Distributions on Countably Rectifiable Sets 809

in mind that σ(∂Ω\ ∂∗ Ω) = 0), the fifth equality uses (11.9.24), and the final equality
defines I, II. Next, the fact that (11.9.13)-(11.9.14) is a well-defined distribution on
∂Ω ensures that
∫
  
I = lim+ ν, k(xo − ·)∂Ω ψ dσ
ε→0 ∂Ω\B(x o ,ε)
    
= P.V. ν, k(xo − ·)∂Ω , ψ . (11.9.26)

If we set
∫
 
ϑ := ω, k(ω) dH n−1 (ω) ∈ C, (11.9.27)
S n−1

then work in [135, §2.5] shows that for H n−1 -a.e. point x ∈ ∂∗ Ω (hence for H n−1 -a.e.
point x ∈ ∂Ω in the present setting) there exists some L 1 -nullset Nx ⊂ (0, ∞) with
the property that for each ε ∈ (0, ∞) \ Nx the set Ω ∩ ∂B(x, ε) is H n−1 -measurable
and
∫ y − x  ϑ
lim , k(x − y) dH n−1 (y) = − . (11.9.28)
ε ∈(0,∞)\N x Ω∩ ∂B(x,ε) ε 2
ε→0

Next, if N := N ∪ Nxo where Nxo ⊂ (0, ∞) is the L 1 -nullset associated with the
point xo as above then
∫ y − x 
o
II = lim , k(xo − y) ψ(y) dH n−1 (y)
ε ∈(0,∞)\ N Ω∩∂B(x o ,ε) ε
ε→0+
+ ∫ ,
y − x 
o
= ψ(xo ) lim , k(xo − y) dH n−1
(y)
ε ∈(0,∞)\ N Ω∩∂B(x o ,ε) ε
ε→0+

ϑ ϑ 
= − ψ(xo ) = − δxo , ψ , (11.9.29)
2 2
where the first equality is simply the definition of II, the second equality follows
from the fact that ψ(y) = ψ(xo ) + O(ε) on the domain of integration, while the third
equality is implied by (11.9.28) (keeping in mind that Nxo ⊆ N). At this stage, from
(11.9.25), (11.9.26), (11.9.29) we conclude that
ϑ   
ν•F = δx + P.V. ν, k(xo − ·)∂Ω
2 o (11.9.30)
 
as distributions in Lipc (∂Ω) ,

and the claim in (11.9.15) now becomes a consequence of this, (11.9.22), and
Example 4.2.4.
810 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

To deal with the claim in item (2), recall that there exists a σ-measurable subset
A of ∂Ω satisfying σ(A) = 0 and such that (11.9.17) holds. Fix two arbitrary points
x0, x1 ∈ ∂Ω \ A and consider the vector field

Fx0,x1 (x) := k(x − x0 ) − k(x − x1 ), ∀x ∈ Ω. (11.9.31)

Then (11.9.12) ensure that

 
1 (Ω, L n ) n
Fx0,x1 ∈ C 2 (Ω \ {x0, x1 }) ∩ Lbdd
(11.9.32)
and divFx0,x1 = 0 in Ω.

Also, having picked some aperture parameter κ > 0, from (11.9.31), (11.9.12), the
Mean Value Theorem, and [133, Lemma 8.3.7] we see that there exists some constant
Cx0,x1 ∈ (0, ∞), which depends on x0 , x1 , Ω, k, n, and κ, such that for σ-a.e. x ∈ ∂Ω
we have
!
 Ω−  |x − x0 | −n if x is away from x0, x1,
Nκ Fx0,x1 (x) ≤ Cx0,x1
|x − x0 | 1−n − |x − x1 | 1−n if x is near x0, x1 .
(11.9.33)
Together with [133, (7.2.5)], this implies

Nκ Fx0,x1 ∈ L p (∂Ω, σ). (11.9.34)
n−1
n <p<1

In view of (11.9.32)-(11.9.34) we may invoke Theorem 10.2.1  which

  presently
gives that the “bullet product” ν • Fx0,x1 as a functional in Lipc (∂Ω) (cf. [133,
Proposition 4.2.3]) satisfies

ν • Fx0,x1 ∈ H p (∂Ω, σ). (11.9.35)
n−1
n <p<1

On the other hand, if ϑ is as in (11.9.27), then a direct application of (11.9.30) (with

xo replaced by x0 and x1 ) gives
ϑ       
ν • Fx0,x1 = − δx0 − δx1 + P.V. ν, k(x0 − ·)∂Ω − P.V. ν, k(x1 − ·)∂Ω
2
 
as distributions in Lipc (∂Ω) .
(11.9.36)

Collectively, (11.9.35), (11.9.36), and (4.2.17) now prove (11.9.16). 

In turn, Proposition 11.9.2 readily implies the following result, pertaining to the
membership of certain principal-value distributions to Hardy spaces H p with p < 1.

Proposition 11.9.3 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be a UR domain and set

σ := H n−1 ∂Ω. Let b ∈ C ∞ (Rn \ {0}) be a complex-valued function which is even
11.10 Hardy-Based Boundary Sobolev Spaces 811

and positive homogeneous of degree 2 − n. Finally, fix j, k ∈ {1, . . . , n}. Then for

σ-a.e. x0, x1 ∈ ∂Ω the distribution
     
P.V. ν j (∂k b)(x0 − ·)∂Ω − P.V. νk (∂j b)(x0 − ·)∂Ω
     
− P.V. ν j (∂k b)(x1 − ·)∂Ω + P.V. νk (∂j b)(x1 − ·)∂Ω (11.9.37)
%
belongs to n−1 <p<1 H p (∂Ω, σ).
n

Proof This follows from item (2) of Proposition 11.9.2 applied to

k := (∂k b)e j − (∂j b)ek (11.9.38)

which is a vector field as in (11.9.12). 

11.10 Hardy-Based Boundary Sobolev Spaces

In this section we discuss an important brand of Hardy spaces defined in the spirit of
the general recipe given in (11.7.1) but where now the space B, where the tangential
derivatives live, is allowed to contained distributions which are not locally integrable
functions. We begin by presenting a number of auxiliary results.
Lemma 11.10.1 Suppose Ω is an Ahlfors regular domain in Rn and abbreviate
σ := H n−1 ∂Ω. Also, consider a function f ∈ Lloc 1 (∂Ω, σ) and fix some exponent
 n−1 
p ∈ n , 1 . Then for each pair if indices j, k ∈ {1, . . . , n} there could be at most
one distribution g jk ∈ H p (∂Ω, σ) with the property that
  
∫ ⎧ 
⎨ − φ ∂Ω, g jk
⎪
⎪ for ∂Ω bounded,
f ∂τ j k φ dσ =     ∀φ ∈ Cc∞ (Rn ),
∂Ω ⎪
⎪ − φ , g jk for ∂Ω unbounded,
⎩ ∂Ω
(11.10.1)

where the pairings in (11.10.1) are understood in the sense of Theorem 4.6.1 (used
with Σ := ∂Ω).
Proof If there were two g jk ’s in H p (∂Ω, σ) doing the job in (11.10.1) then their
difference, call it h ∈ H p (∂Ω, σ), would have the property that for each φ ∈ Cc∞ (Rn )
one has   
φ∂Ω, h = 0 if ∂Ω is bounded,
    (11.10.2)
and φ∂Ω , h = 0 if ∂Ω is unbounded.
 
In turn, if p ∈ n−1
n , 1 then Lemma 4.8.12 (used with Σ := ∂Ω) implies that (11.10.2)
.α  
actually holds for every φ ∈ C (∂Ω) with α := (n − 1) p1 − 1 ∈ (0, 1). According
to Theorem 4.6.1, this ultimately forces h to be zero in H p (∂Ω,
∫ σ).  Corresponding

to the case when p = 1, since for each φ ∈ Cc∞ (Rn ) we have ∂Ω φ∂Ω  |h| dσ < +∞,
812 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫   
from (11.10.2) and Proposition 4.8.7 we conclude that ∂Ω φ∂Ω h dσ = 0 for all
φ ∈ Cc∞ (Rn ). Given that H 1 (∂Ω, σ) ⊆ L 1 (∂Ω, σ), [133, Corollary 3.7.3] then applies
and gives h = 0 at σ-a.e. point on ∂Ω, thus h = 0 in H 1 (∂Ω, σ). 

We wish to augment Lemma 11.10.1 with the “integration by parts on the bound-
ary” formula established in the proposition below.

Proposition 11.10.2 Suppose Ω is an Ahlfors regular domain in Rn and abbreviate

σ := H n−1 ∂Ω. Consider a function f ∈ Lloc 1 (∂Ω, σ) and fix some integrability
 n−1 
exponent p ∈ n , 1 along with a pair of indices j, k ∈ {1, . . . , n}.
Then there exists a distribution g jk ∈ H p (∂Ω, σ) with the property that (11.10.1)
holds if and only if ∂τ j k f ∈ H p (∂Ω, σ) (cf. Definition 11.2.1). In the latter scenario
one then has
  
∫ ⎧ 
⎨ − φ ∂Ω, ∂τ j k f
⎪
⎪ for ∂Ω bounded,
f ∂τ j k φ dσ =     ∀φ ∈ Cc1 (Rn ),
∂Ω ⎪
⎪ − φ , ∂τ f for ∂Ω unbounded,
⎩ ∂Ω jk
(11.10.3)

where the pairings in (11.10.3) are understood in the sense of Theorem 4.6.1 (used
with Σ := ∂Ω).

Proof Assume for now that (n − 1)/n < p < 1 and that ∂Ω is unbounded. Suppose
there exists some g jk ∈ H p (∂Ω, σ) such that (11.10.1) holds. Pick an arbitrary
ψ ∈ Lipc (∂Ω). Then the distributional derivative ∂τ j k f is well defined and satisfies
(cf. Definition 11.2.1)
∫
 
(Lip c (∂Ω)) ∂τ j k f , ψ Lip c (∂Ω) = − f ∂τ j k ψ dσ. (11.10.4)
∂Ω

Apply Remark 11.1.13 with ∂∗ Ω replaced by ∂Ω and f replaced by ψ to obtain a

sequence {ϕ } ∈N satisfying (for some purely dimensional constant C ∈ (0, ∞)):
  
{ϕ } ∈N ⊆ Cc∞ (Rn ), lim supx ∈∂Ω ϕ (x) − ψ(x) = 0,
→∞
supp ϕ ⊆ K for all ∈ N and some compact set K ⊂ Rn,
   (11.10.5)
sup ∈N supx ∈Rn (∇ϕ )(x) ≤ CψLip(∂Ω),

and sup ∈N supx ∈Rn |ϕ | ≤ C · supx ∈∂Ω |ψ|.

Thus, invoking (11.1.92) from Lemma 11.1.12 we may conclude that ∂τ j k ϕ con-
verges to ∂τ j k ψ weak-∗ in L ∞ (∂Ω, σ) as → ∞. Since (11.10.5) also implies
supp ψ ⊆ K and since 1K · f ∈ L 1 (∂Ω, σ), the latter convergence allows us to write
11.10 Hardy-Based Boundary Sobolev Spaces 813
∫ ∫
f ∂τ j k ψ dσ = 1K · f ∂τ j k ψ dσ
∂Ω ∂Ω
∫ ∫
= lim 1K · f ∂τ j k ϕ dσ = lim f ∂τ j k ϕ dσ
→∞ ∂Ω →∞ ∂Ω
   
= − lim H p (∂Ω,σ) g jk , ϕ ∂Ω C.α (∂Ω)/∼ (11.10.6)
→∞

for α := (n − 1)(1/p − 1) ∈ (0, 1), where the last equality in (11.10.6) uses (11.10.1).
Note that (11.10.5) entails
  
sup ϕ ∂Ω C.α (∂Ω) < +∞ and lim ϕ (x) = ψ(x) for each x ∈ ∂Ω. (11.10.7)
 ∈N →∞

  
These allow us to apply Lemma 4.8.4 to obtain that lim ϕ ∂Ω = [ψ] weak-∗ in
→∞
. &
C α (∂Ω) ∼. Consequently,
     
lim H p (∂Ω,σ) g jk , ϕ ∂Ω C.α (∂Ω)/∼ = H p (∂Ω,σ) g jk , [ψ] C.α (∂Ω)/∼
→∞
 
= (Lipc (∂Ω)) g jk , ψ Lipc (∂Ω), (11.10.8)

where the last equality is provided by Lemma 4.6.4. Combining (11.10.4), (11.10.6),
and (11.10.8) yields
   
(Lip c (∂Ω)) ∂τ j k f , ψ Lip c (∂Ω) = (Lip c (∂Ω)) g jk , ψ Lip c (∂Ω) (11.10.9)

which, in view of the arbitrariness of ψ ∈ Lipc (∂Ω), allows us to conclude that the
distributional derivative ∂τ j k f is equal to g jk , hence ∂τ j k f ∈ H p (∂Ω, σ).
To deal with the opposite implication, first retain the same assumptions on p and
∂Ω. Concretely, assume the distributional derivative ∂τ j k f belongs  to H p (∂Ω, σ).
1 n 
Also, select an arbitrary function φ ∈ Cc (R ) and note that φ ∂Ω ∈ Lipc (∂Ω). From
(4.6.31) in Lemma 4.6.4 and the definition of the distributional derivative ∂τ j k f (cf.
Definition 11.2.1) we have (again, with α := (n − 1)(1/p − 1))
      
H p (∂Ω,σ) ∂τ f , φ
 .α = (Lip (∂Ω)) ∂τ f , φ Lip (∂Ω)
jk ∂Ω C (∂Ω)/∼ c jk ∂Ω c

∫
  
=− f ∂τ j k φ∂Ω dσ. (11.10.10)
∂Ω

In addition, from (11.1.26) and (11.1.4) we know that

      
∂τ j k φ∂Ω = ν j ∂k φ ∂Ω − νk ∂j φ ∂Ω

= ∂τ j k φ at σ-a.e. point on ∂Ω. (11.10.11)

A combination of (11.10.11) and (11.10.10) now shows that (11.10.1) holds with
g jk := ∂τ j k f ∈ H p (∂Ω, σ).
814 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

The case when ∂Ω is bounded is handled analogously, without having to consider

.
equivalence classes of functions modulo constants, and with C α (∂Ω)/∼ replaced by
the inhomogeneous Hölder space C α (∂Ω).
In the case p = 1 the arguments are very similar, with appropriate adjustments
for the duality pairings (which now involve the space BMO). The most significant
difference is that, this time, in place of (11.10.7) we have
     
sup ∈N ϕ ∂Ω BMO(∂Ω,σ) ≤ C · sup ∈N ϕ ∂Ω  L ∞ (∂Ω,σ) < +∞
 (11.10.12)
and ϕ ∂Ω −→ ψ in Lloc 1 (∂Ω, σ) as → ∞,

and these properties, in turn, ensure the applicability of Lemma 4.8.1 (which is now
used in lieu of Lemma 4.8.4). 

In turn, Proposition 11.10.2 may be used to further generalize the integration by

parts formula (11.10.1), by now allowing the “test” function to be merely Hölder
and satisfy suitable integrability properties when restricted to the boundary of the
Ahlfors regular domain involved.

Lemma 11.10.3 Let Ω be an Ahlfors regular domain  in Rn (n ≥ 2) and abbreviate


σ := H ∂Ω. For each ε > 0 define Uε := x ∈ R : dist(x, ∂Ω) < ε . Also, fix
n−1 n
 n−1 
p ∈ n , 1 together with q, q  ∈ [1, ∞] such that 1/q + 1/q  = 1, and select a pair
of indices j, k ∈ {1, . . . , n}. Finally, pick some f ∈ L q (∂Ω, σ) with the property that
∂τ j k f ∈ H p (∂Ω, σ) and consider a function ϕ satisfying

ϕ∂Ω ∈ L q ∂Ω, 1+σ(x)

ϕ ∈ C 1 (Uε ) for some ε > 0, |x | q
 ∩ L ∞ (∂Ω, σ),
  q n
and (∇ϕ)∂Ω ∈ L (∂Ω, σ) .
(11.10.13)
In the case when p < 1 make the additional assumption that
 .  
ϕ∂Ω ∈ C η (∂Ω) where η := (n − 1) p1 − 1 ∈ (0, 1). (11.10.14)

Then
  
∫ ⎧ 
⎨ − ϕ ∂Ω, ∂τ j k f
⎪
⎪ for ∂Ω bounded,
f (∂τ j k ϕ) dσ =     (11.10.15)
∂Ω ⎪
⎪ − ϕ , ∂τ f for ∂Ω unbounded,
⎩ ∂Ω jk

where the pairings in (11.10.15) are understood in the sense of Theorem 4.6.1 (used
with Σ := ∂Ω).

Proof To fix ideas, suppose first that the set ∂Ω is unbounded. Retaining notation
from the proof of Lemma 11.1.7, then write
11.10 Hardy-Based Boundary Sobolev Spaces 815
∫ ∫
f ∂τ j k ϕ dσ = lim f (∂τ j k ϕ)θ R dσ
∂Ω R→∞ ∂Ω
∫ ∫
= lim f ∂τ j k ϕR dσ − lim f ϕ ∂τ j k θ R dσ
R→∞ ∂Ω R→∞ ∂Ω
       
= − lim ϕR ∂Ω , ∂τ j k f = − lim (ϕθ R )∂Ω , ∂τ j k f
R→∞ R→∞
   
= − ϕ∂Ω , ∂τ j k f . (11.10.16)

Above, the first equality is a consequence of Lebesgue’s Dominated Convergence


Theorem (taking into account that f ∈ L q (∂Ω, σ) and ∂τ j k ϕ ∈ L q (∂Ω, σ), as seen
from (11.1.4) and the last condition in (11.10.13)). The second equality is implied
by (11.1.53). The third equality uses two observations. First, since for each R > 0
we have ϕR ∈ Cc1 (Rn ) (cf. (11.1.53)), we may invoke (11.10.3) to conclude that
∫    
f ∂τ j k ϕR dσ = − ϕR ∂Ω , ∂τ j k f for each R > 0. (11.10.17)
∂Ω

Second, since for each R ≥ 1 we have the pointwise estimate

    
 f ϕ ∂τ θ R  ≤ C| f |(1 + |x|)−1 · ϕ  · 1[B(0,2R)\B(0,R)]∩∂Ω (11.10.18)
jk ∂Ω

at σ-a.e. point on ∂Ω, and since our assumptions imply

∫
  
| f |(1 + |x|)−1 · ϕ∂Ω  dσ
∂Ω
∫
   q  1/q
≤ f L q (∂Ω,σ) (1 + |x|)−q · ϕ∂Ω  dσ < +∞ (11.10.19)
∂Ω

(cf. (11.10.13)), Lebesgue’s Dominated Convergence Theorem gives

∫
lim f ϕ ∂τ j k θ R dσ = 0. (11.10.20)
R→∞ ∂Ω

Hence, the last limit in the second line of (11.10.16) vanishes. Next, the fourth
equality in (11.10.16) is a consequence of the fact that, as noted in (11.1.53), the
functions ϕR and ϕθ R agree on ∂Ω. The crux of the matter is the last equality in
(11.10.16) which, in view of the fact that ∂τ j k f ∈ H p (∂Ω, σ) and the duality result
from Theorem 4.6.1, follows as soon as we check that
⎧ . &
      ⎪ C η (∂Ω) ∼ if p < 1,
⎨
⎪
lim (ϕθ R )∂Ω = ϕ∂Ω weak-∗ in (11.10.21)
R→∞ ⎪ -
⎪ BMO(∂Ω,
⎩ σ) if p = 1.
816 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

This, however, is implied by the general weak-∗ convergence results established in

Lemma 4.8.4 and, respectively, Lemma 4.8.1 (in the latter case also bearing in mind
the trivial bounded embedding L ∞ (∂Ω, σ) → BMO(∂Ω, σ)).
Finally, the case when ∂Ω is bounded is handled in a very similar fashion. 

Later on, we shall also need the following version of Lemma 11.10.3.

Lemma 11.10.4 Let Ω be an Ahlfors  regular domain Rn . Abbreviate

 σ := H n−1 ∂Ω
and for each ε > 0 define Uε := x ∈ R : dist(x, ∂Ω) < ε . Next, fix p ∈ n−1
n
n ,1
along with j, k ∈ {1, . . . , n} and consider a function

σ(x)
f ∈ L 1 ∂Ω, with ∂τ j k f ∈ H p (∂Ω, σ). (11.10.22)
1 + |x| n

Finally, for some ε > 0 and C ∈ (0, ∞), suppose

ϕ ∈ C 1 (Uε ) satisfies, for all x ∈ Uε,

C C (11.10.23)
|ϕ(x)| ≤ and |(∇ϕ)(x)| ≤ .
1 + |x| n−1 1 + |x| n

Then
  
∫ ⎧ 
⎨ − ϕ ∂Ω, ∂τ j k f
⎪
⎪ for ∂Ω bounded,
f (∂τ j k ϕ) dσ =     (11.10.24)
∂Ω ⎪
⎪ − ϕ , ∂τ f for ∂Ω unbounded,
⎩ ∂Ω jk

where the pairings in (11.10.15) are understood in the sense of Theorem 4.6.1 (used
with Σ := ∂Ω).

Proof Bring in the function Ψε from (11.1.52). Bearing in mind (11.10.23) we see
that the extension to Rn of the function ϕ · Ψε ∈ C 1 (Uε ) by zero outside its support
belongs to C 1 (Rn ), is bounded, and has bounded gradient. In particular, said function
is bounded and Lipschitz. Since its restriction to ∂Ω coincides with that of ϕ, we
ultimately conclude that

ϕ∂Ω ∈ C η (∂Ω) for each η ∈ (0, 1). (11.10.25)

Granted this, the same argument as in the proof of Lemma 11.10.3 works (with only
minor, natural adjustments) and gives (11.10.24). 

Moving on, we define Hardy-based homogeneous Sobolev spaces of order one on

the boundary of an Ahlfors regular domain. These may be considered as the natural
continuation of the scale of L p -based homogeneous Sobolev spaces of order one
(introduced in Definition 11.5.3) by now allowing p < 1.

Definition 11.10.5 Let Ω ⊆ Rn be an Ahlfors  regular domain and abbreviate

σ := H n−1 ∂Ω. For each given exponent p ∈ n−1n , ∞ , define the Hardy-based
homogeneous Sobolev space of order one on ∂Ω as
11.10 Hardy-Based Boundary Sobolev Spaces 817
!
.p σ(x)
H1 (∂Ω, σ) := f ∈ L 1 ∂Ω, : ∂τ j k f ∈ H p (∂Ω, σ)
1 + |x| n
"
for each j, k ∈ {1, . . . , n} (11.10.26)

and equip this space with the semi-norm

.p 
n
H1 (∂Ω, σ)  f →  f  H. p (∂Ω,σ) := ∂τ j k f  H p (∂Ω,σ) . (11.10.27)
1
j,k=1

In the context of Definition 11.10.5, for each j, k ∈ {1, . . . , n} the tangential

derivative operator
.p
∂τ j k : H1 (∂Ω, σ) −→ H p (∂Ω, σ) (11.10.28)

is well defined, linear, and bounded. Also, as is apparent from definitions,

.
H11 (∂Ω, σ) ⊆ L1,loc
1
(∂Ω, σ). (11.10.29)

Let us also note that, as seen from Definition 11.10.5, Definition 11.5.3, [133,
(3.6.27)], and (A.0.1), we have
.p .p
H1 (∂Ω, σ) = L1 (∂Ω, σ) if p ∈ (1, ∞), (11.10.30)
 .p 
while (4.2.21) implies that the scale H1 (∂Ω, σ) n−1 <p<∞ is nested if ∂Ω is bounded.
.p n
 
Examples of functions in H1 (∂Ω, σ) with p ∈ n−1 n , 1 are offered by functions
q
in L1 (∂Ω, σ) with q ∈ (1, ∞] having bounded support. Indeed, if the function
q
f ∈ L1 (∂Ω, σ) with q ∈ (1, ∞] has supp f ⊆ Δ := B(x0, r) ∩ ∂Ω for some x0 ∈ ∂Ω
and r ∈ (0, ∞), then (11.1.37) and (11.1.59) imply that, for each j, k ∈ {1, . . . , n},

σ(Δ)1/q−1/p ∂τ j k f  L−1q (∂Ω,σ) ∂τ j k f is a (p, q)-atom on ∂Ω. (11.10.31)

This ultimately implies that

q
whenever f ∈ L1 (∂Ω, σ) with q ∈ (1, ∞] has supp f ⊆ Δ for some
.p
surface ball Δ ⊂ ∂Ω then f belongs to the space H1 (∂Ω, σ) for each (11.10.32)
 
p ∈ n−1 .
n , 1 and  f  H p (∂Ω,σ) ≤ C · σ(Δ)
1/p−1/q  f  .
q .
1
L (∂Ω,σ) 1

Given an Ahlfors regular domain Ω ⊆ Rn , where

 n ∈ N with n ≥ 2, abbreviate
σ := H n−1 ∂Ω and pick some exponent p ∈ n−1 n ∞ . For later purposes, it is
,
useful to consider
.p &  .p 
H1 (∂Ω, σ) ∼ := [ f ] : f ∈ H1 (∂Ω, σ) , (11.10.33)
818 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
.p &
that is, H1 (∂Ω, σ) ∼ denotes the quotient space of classes [·] of equivalence modulo
.p
constants of functions in H1 (∂Ω, σ). We equip this space with the semi-quasinorm

.p &   
n
 
H1 (∂Ω, σ) ∼  [ f ] −→ [ f ] H. p (∂Ω,σ)/∼ := ∂τ f  p
jk H (∂Ω,σ)
. (11.10.34)
1
j,k=1

Moving on, we introduce the inhomogeneous version of the brand of Hardy-based

Sobolev space from Definition 11.10.5.

Definition 11.10.6 Let Ω be an Ahlfors regular domain in Rn and abbreviate

σ := H n−1 ∂Ω. For each p ∈ n−1 , ∞ and q ∈ [1, ∞] define the Hardy-based
n q, p
inhomogeneous Sobolev space H1 (∂Ω, σ) as
q, p .p
H1 (∂Ω, σ) := L q (∂Ω, σ) ∩ H1 (∂Ω, σ)

= f ∈ L q (∂Ω, σ) : ∂τ j k f ∈ H p (∂Ω, σ) for 1 ≤ j, k ≤ n ,
(11.10.35)

and equip it with the quasi-norm

 f  H q, p (∂Ω,σ) :=  f  L q (∂Ω,σ) +  f  H. p (∂Ω,σ)

1
q, p
1 (11.10.36)
for each function f ∈ H1 (∂Ω, σ).

As seen in [136, Chapter 2], this scale of spaces is both relevant and natural in
the context of singular integral operators of boundary layer type. For now, the goal
is to study this scale for its own sake. We begin by making several observations, in
the setting considered in Definition 11.10.6. For one thing, (11.10.32) implies that
  q
given p ∈ n−1 n , 1 and q ∈ (1, ∞], any q,
f ∈ L1 (∂Ω, σ) with
p (11.10.37)
bounded support belongs to the space H1 (∂Ω, σ),

hence also (cf. (4.4.11) and [133, (3.6.27)]),

q q, p
L1 (∂Ω, σ) → H1 (∂Ω, σ) continuously, whenever
(11.10.38)
∂Ω is bounded, n−1
n < p < ∞ and max{p, 1} < q ≤ ∞.

For another thing, it is clear from (11.1.104), (11.1.59), and definitions that
q, p  
Lipc (∂Ω) ⊆ H1 (∂Ω, σ) for all p ∈ n−1 n , ∞ and q ∈ [1, ∞]. (11.10.39)

More sophisticated examples are discussed in Proposition 11.10.13. For now we

wish to remark that, together, (11.10.35)-(11.10.36), (11.10.39), and [133, (3.7.22)]
imply that
11.10 Hardy-Based Boundary Sobolev Spaces 819
 n−1 
for all p ∈ n , ∞ and q ∈ [1, ∞] one has
q, p
H1 (∂Ω, σ) → L q (∂Ω, σ) continuously, (11.10.40)
and also densely if q < ∞.

We also wish to observe here that, as is apparent from Definition 11.10.5 and
Definition 11.10.6,
.p 1, p  
H1 (∂Ω, σ) = H1 (∂Ω, σ) for all p ∈ n−1
n ,∞
(11.10.41)
whenever ∂Ω is bounded.

It also useful to characterize explicitly the dual of Hardy-based inhomogeneous

Sobolev space (11.10.35) with p = 1 and 1 ≤ q < ∞, in the manner described below.

Proposition 11.10.7 Let Ω be an Ahlfors regular domain in Rn with the property

that ∂Ω is unbounded. Abbreviate σ := H n−1 ∂Ω, and suppose 1 ≤ q < ∞ and
1 < q  ≤ ∞ are such that 1/q + 1/q  = 1. Then there exists a constant C ∈ (0, ∞)
∗
q,1
with the property that for each f ∈ H1 (∂Ω, σ) there exist some functions,

f0 ∈ L q (∂Ω, σ) and f jk ∈ BMO(∂Ω, σ) with 1 ≤ j < k ≤ n, satisfying (with [·]
denoting the equivalence class modulo constants)
  
 f0  L q (∂Ω,σ) + [ f jk ] ≤ C f (H q,1 (∂Ω,σ))∗ (11.10.42)
BMO(∂Ω,σ)/∼ 1
1≤ j<k ≤n

q,1
and with the property that for each g ∈ H1 (∂Ω, σ) one has
∫
 
q,1
(H (∂Ω,σ))∗
f , g q,1
H (∂Ω,σ)
= f0 g dσ (11.10.43)
1 1 ∂Ω
  
+ BMO(∂Ω,σ)/∼ [ f jk ], ∂τ j k g H 1 (∂Ω,σ)
.
1≤ j<k ≤n

Moreover, a similar result is valid when ∂Ω is bounded, this time without having
to consider equivalence classes of functions modulo constants.

Proof Work under the assumption that ∂Ω is unbounded. Consider the application
q,1   n(n−1)/2
J : H1 (∂Ω, σ) −→ L q (∂Ω, σ) ⊕ H 1 (∂Ω, σ)
  (11.10.44)
q,1
J f := f , (∂τ j k f )1≤ j<k ≤n for each f ∈ H1 (∂Ω, σ).
  n(n−1)/2
Then Range J is a closed subspace of L q (∂Ω, σ) ⊕ H 1 (∂Ω, σ) since J is
an isometry. The Open Mapping Theorem then implies that

J −1 : Range J −→ H1 (∂Ω, σ)
q,1
(11.10.45)
820 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∗
q,1
is an isomorphism. For each given functional f ∈ H1 (∂Ω, σ) , consider the
composition
f ◦ J −1 : Range J −→ R. (11.10.46)
Then the Hahn-Banach Extension Theorem in concert with Riesz’s Representation
Theorem and the duality result from Theorem 4.6.1 guarantee the existence of a

family of functions f0 ∈ L q (∂Ω, σ) and f jk ∈ BMO(∂Ω, σ) with 1 ≤ j < k ≤ n,
satisfying (11.10.42) and (11.10.43).
Finally, the case when ∂Ω is bounded is handled analogously, without having to
consider equivalence classes of functions modulo constants (cf. (4.6.2)). 

The next goal is to introduce the negative Sobolev space BMO−1 on the boundary
of an Ahlfors regular domain. We shall do so in Definition 11.10.9 and, as a preamble,
we first discuss the embedding result described in the lemma below.

Lemma 11.10.8 Fix n ∈ N with n ≥ 2. Let Ω ⊆ Rn be an Ahlfors regular domain

and abbreviate σ := H n−1 ∂Ω. Then the following is a well-defined continuous
embedding:
 ∗
q,1
L q (∂Ω, σ) → H1 (∂Ω, σ)
(11.10.47)
if q, q  ∈ (1, ∞) with 1/q + 1/q  = 1.
As a consequence, the following is a well-defined continuous embedding:
n−1
,1 ∗
L n−1 (∂Ω, σ) → H1n−2 (∂Ω, σ) if n ≥ 3. (11.10.48)

Proof Since from (11.10.40) and (11.10.39) we know that

q,1
H1 (∂Ω, σ) → L q (∂Ω, σ) continuously
(11.10.49)
and densely for each q ∈ (1, ∞),

the fact that (11.10.47) is a well-defined continuous embedding becomes a conse-

quence of (11.10.49) and Lemma 1.2.1. Finally, (11.10.48) is a particular case of
(11.10.47) (specialized to the case when q := (n − 1)/(n − 2), which is an admissible
choice if n ≥ 3). 

Here is the definition we propose for the negative Sobolev space BMO−1 on the
boundary of an Ahlfors regular domain.

Definition 11.10.9 Suppose n ∈ N satisfies9 n ≥ 3. Also, assume Ω is an Ahlfors

regular domain in Rn and abbreviate σ := H n−1 ∂Ω. In this setting, set
n−1
,1 ∗
BMO−1 (∂Ω, σ) := H1n−2 (∂Ω, σ) (11.10.50)

and define

9 the case when n = 2 is discussed in [136, §2.3]

11.10 Hardy-Based Boundary Sobolev Spaces 821

VMO−1 (∂Ω, σ) := the closure of L n−1 (∂Ω, σ) in BMO−1 (∂Ω, σ). (11.10.51)

In the context of Definition 11.10.9, Lemma 11.10.8 guarantees that we have a

well-defined continuous embedding

L n−1 (∂Ω, σ) → BMO−1 (∂Ω, σ). (11.10.52)

In particular, (11.10.52) ensures that it is meaningful to define VMO−1 (∂Ω, σ) as

we did in (11.10.51). Let us also note that, as is apparent from Definition 11.10.9,

VMO−1 (∂Ω, σ) is a closed subspace of BMO−1 (∂Ω, σ), and the space
(11.10.53)
L n−1 (∂Ω, σ) embeds continuously and densely into VMO−1 (∂Ω, σ).
 
In fact, since φ|∂Ω : φ ∈ Cc∞ (Rn ) is dense L n−1 (∂Ω, σ) (cf. [133, (3.7.22)]), we
conclude from (11.10.53) that
 
φ|∂Ω : φ ∈ Cc∞ (Rn ) is dense in VMO−1 (∂Ω, σ); as a
consequence, VMO−1 (∂Ω, σ) is the closure of Lipc (∂Ω) (11.10.54)
in the space BMO−1 (∂Ω, σ).
n−1
,1 n−1
We also wish to remark that, since H1n−2 (∂Ω, σ) → L n−2 (∂Ω, σ) continuously
and densely, (11.10.50) and (1.2.25) imply that the following compatibility result is
valid:
∫
   
f g dσ = L n−1 (∂Ω,σ) f , g n−1 = BMO−1 (∂Ω,σ) f , g n−1 ,1
∂Ω L n−2 (∂Ω,σ) H1n−2 (∂Ω,σ)
for all f ∈ L n−1 (∂Ω, σ) → BMO−1 (∂Ω, σ)
n−1
,1 n−1
and all g ∈ H1n−2 (∂Ω, σ) → L n−2 (∂Ω, σ).
(11.10.55)
The following result refines the observation made in (11.3.39)-(11.3.40). In [136,
§2.2] a result similar in spirit is established for more restrictive class of sets but
without making assumptions on the nontangential maximal operator of the function
itself.

Theorem 11.10.10 Let Ω ⊆ Rn be an Ahlfors regular domain. Set σ := H n−1 ∂Ω

and fix an arbitrary aperture parameter κ ∈ (0, ∞). Assume u is a scalar-valued
function in Ω satisfying
1,1
u ∈ Wloc (Ω), Nκε u ∈ Lloc
1 (∂Ω, σ) for some ε > 0,
 
Nκ (∇u) ∈ L p,q (∂Ω, σ) for some p ∈ n−1 n , 1 and q ∈ (0, ∞], (11.10.56)
κ−n.t.
and the nontangential trace u∂Ω exists σ-a.e. on ∂Ω.
κ−n.t.
Then the function u∂Ω belongs to Lloc 1 (∂Ω, σ), its distributional tangential

derivatives (defined as in (11.2.2)) satisfy

822 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
κ−n.t.
∂τ j k u∂Ω ∈ H p,q (∂Ω, σ) for each j, k ∈ {1, . . . , n}, (11.10.57)

and there exists C = CΩ,κ, p,q ∈ (0, ∞) such that


n
  κ−n.t.  
∂τ u  p, q ≤ CNκ (∇u) L p, q (∂Ω,σ) . (11.10.58)
jk ∂Ω H (∂Ω,σ)
j,k=1

In fact, as distributions on ∂Ω,

κ−n.t. .
∂τ j k u∂Ω = ∂τ j k u for each j, k ∈ {1, . . . , n}. (11.10.59)

Proof Suppose u satisfies (11.10.56). We claim that

 1 n
∇u ∈ Lbdd (Ω, L n ) . (11.10.60)

To justify the membership claimed in (11.10.60), let E ⊆ Ω be an arbitrary bounded

Lebesgue measurable set. In particular, L n (E) < ∞. Then [133,
 (8.6.50) in Propo- n
sition 8.6.3] (applied with u replaced by ∇u) implies ∇u ∈ L np/(n−1),q (E, L n ) .
Since np/(n − 1) > 1, we may now invoke  [133, Lemma
n 6.2.4] (with X := E
and μ := L n ) to conclude that ∇u ∈ L 1 (E, L n ) . This proves (11.10.60).
Granted this and the current assumptions, Proposition 11.3.1 applies and gives that
κ−n.t.
u∂Ω ∈ Lloc 1 (∂Ω, σ) and (11.10.59) holds. On the other hand, (10.2.13) implies that

for each j, k ∈ {1, . . . , n} we have

.
∂τ j k u ∈ H p,q (∂Ω, σ) and
.    (11.10.61)
∂τ u p, q ≤ C Nκ (∇u) L p, q (∂Ω,σ),
jk H (∂Ω,σ)

for some constant C = CΩ,κ, p,q ∈ (0, ∞). Now (11.10.57)-(11.10.58) become conse-
quences of (11.10.59) and (11.10.61). 
Recall the Hardy-based inhomogeneous Sobolev spaces introduced in Defini-
tion 11.10.6.
Corollary 11.10.11 Let Ω ⊆ Rn be an Ahlfors regular domain. Set σ := H n−1 ∂Ω
and fix an arbitrary aperture parameter κ ∈ (0, ∞). Assume u is a scalar-valued
function in Ω satisfying
1,1
u ∈ Wloc (Ω), Nκ u ∈ L q (∂Ω, σ) for some q ∈ [1, ∞],
 
Nκ (∇u) ∈ L p (∂Ω, σ) for some p ∈ n−1n ,1 , (11.10.62)
κ−n.t.
and the nontangential trace u∂Ω exists σ-a.e. on ∂Ω.
κ−n.t.
Then u∂Ω ∈ H1 (∂Ω, σ) and there exists C = CΩ,κ, p,q ∈ (0, ∞) such that
q, p

 κ−n.t. 
u  q, p ≤ CNκ u L q (∂Ω,σ) + CNκ (∇u) L p (∂Ω,σ) . (11.10.63)
∂Ω H
1
(∂Ω,σ)
11.10 Hardy-Based Boundary Sobolev Spaces 823
κ−n.t.
Proof The fact that u∂Ω belongs to L q (∂Ω, σ) is a consequence of (11.10.62),
[133, (8.9.8)], [133, Proposition 8.8.4], and [133, Corollary 8.9.6]. In addition,
(11.10.62) allows us to invoke Theorem 11.10.10 (corresponding to p = q) to see
κ−n.t.
that u∂Ω ∈ H1 (∂Ω, σ) (cf. Definition 11.10.6) and that (11.10.63) holds.
q, p

We continue by producing a formula expressing tangential derivatives of a func-
tion exhibiting a singularity at a boundary point in terms of principal-value distribu-
tions.

Proposition 11.10.12 Suppose Ω ⊆ Rn (where n ∈ N with n ≥ 2) is a UR domain.

Denote the geometric measure theoretic outward unit normal to Ω by ν = (ν1, . . . , νn )
and set σ := H n−1 ∂Ω. Consider a complex-valued function b ∈ C ∞ (Rn \ {0})
which is even and positive homogeneous of degree 2 − n. Also, fix some arbitrary
indices j, k ∈ {1, . . . , n}. Then for σ-a.e. xo ∈ ∂Ω one has (cf. Proposition 11.9.1)
     
∂τ j k b(xo − ·)∂Ω = −P.V. ν j (∂k b)(xo − ·)∂Ω
  
+ P.V. νk (∂j b)(xo − ·)∂Ω (11.10.64)

as distributions on ∂Ω (i.e., in (Lipc (∂Ω))).

Proof Let B stand for the (formal convolution-type) singular integral operator on ∂Ω
whose kernel is b(x − y). Then for σ-a.e. xo ∈ ∂Ω we may write, with φ ∈ Lipc (∂Ω)
arbitrary,
 ∫
   

(Lip c (∂Ω)) ∂τ j k b(xo − ·) ∂Ω , φ Lip c (∂Ω) = − b(xo − ·)(∂τ j k φ) dσ
∂Ω

= −B(∂τ j k φ)(xo ) = −(Tjk φ)(xo )

∫
 
= − lim+ ν j (y)(∂k b)(xo − y) − νk (y)(∂j b)(xo − y) φ(y) dσ(y)
ε→0
y ∈∂Ω
|x−y |>ε
    
= − P.V. ν j (∂k b)(xo − ·)∂Ω , φ
    
+ P.V. νk (∂j b)(xo − ·)∂Ω , φ . (11.10.65)

Above, the first equality is implied by Definition 11.2.1, the second equality comes
from the definition of B, the third equality is a consequence of work in [136, §1.2]
on tangential singular integral operators, the fourth equality is simply the definition
of Tjk , and the fifth equality is based on Proposition 11.9.1 (currently used with
Σ := ∂Ω). 
Nontrivial examples of functions belonging to inhomogeneous Hardy-based
Sobolev spaces on the boundary of an arbitrary UR domain are contained in the
proposition below.
824 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Proposition 11.10.13 Let Ω ⊆ Rn (where n ∈ N with n ≥ 2) be a UR domain and

set σ := H n−1 ∂Ω. Also, let b ∈ C ∞ (Rn \ {0}) be a complex-valued function which
is even and positive homogeneous of degree 2 − n. Then for σ-a.e. x0, x1 ∈ ∂Ω the
difference
 
b(x0 − ·)∂Ω − b(x1 − ·)∂Ω belongs to
  .
q, p p (11.10.66)
H1 (∂Ω, σ) ⊆ H1 (∂Ω, σ).
n−1 n−1 n−1
n <p<1<q< n−2 n <p<1

Proof From Proposition 11.10.12 we know that for σ-a.e. x0, x1 ∈ ∂Ω we have
   
∂τ j k b(x0 − ·)∂Ω − b(x1 − ·)∂Ω (11.10.67)
     
= −P.V. ν j (∂k b)(x0 − ·)∂Ω + P.V. νk (∂j b)(x0 − ·)∂Ω
     
+ P.V. ν j (∂k b)(x1 − ·)∂Ω − P.V. νk (∂j b)(x1 − ·)∂Ω

as distributions on ∂Ω. Thanks to this and Proposition 11.9.3,

    
∂τ j k b(x0 − ·)∂Ω − b(x1 − ·)∂Ω belongs to H p (∂Ω, σ). (11.10.68)
n−1
n <p<1

Also, since for each x0, x1 ∈ ∂Ω there exists a constant Cx0,x1 ∈ (0, ∞), which
depends on x0 , x1 , such that for σ-a.e. x ∈ ∂Ω we have
!
  |x − x0 | 1−n if x is away from x0, x1,
b(x0 − x) − b(x1 − x) ≤ Cx ,x
|x − x0 | − |x − x1 | if x is near x0, x1,
0 1 2−n 2−n

(11.10.69)
with the help of [133, (7.2.5)] we conclude that
   σ(x)
b(x0 − ·)∂Ω − b(x1 − ·)∂Ω ∈ L q (∂Ω, σ) ⊆ L 1 ∂Ω, . (11.10.70)
1 + |x| n
1<q< n−1
n−2

Then (11.10.66) follows from (11.10.26), (11.10.35), (11.10.68), and (11.10.70). 

11.11 Embedding Results Involving Boundary Sobolev Spaces

We start by recalling the following definition.

Definition 11.11.1 Let Σ ⊆ Rn be an arbitrary closed Ahlfors regular set, and

abbreviate σ := H n−1 Σ. For each p ∈ [1, ∞] and s > 0 define the fractional
Hajłasz-Sobolev space W s, p (Σ, σ) as the set of all functions f ∈ L p (Σ, σ) for which
there exists a non-negative function g ∈ L p (Σ, σ) some σ-measurable set N ⊆ Σ
11.11 Embedding Results Involving Boundary Sobolev Spaces 825

with σ(N) = 0 such that

 
| f (x) − f (y)| ≤ |x − y| s g(x) + g(y) for all x, y ∈ Σ \ N. (11.11.1)

Moreover, for each f ∈ W s, p (Σ, σ) define

 f W s, p (Σ,σ) :=  f  L p (Σ,σ) (11.11.2)

 
+ inf g L p (Σ,σ) : g ∈ L p (Σ, σ) so that (11.11.1) holds .

The Sobolev space W 1, p (Σ, σ) appears in [79, Definition, § 3], while the space
W s, p (Σ, σ)
with s > 0 may be found in [201, Definition 1.4, p 678]. See also [64,
Definition 1.3] for a more inclusive range of indices.

Proposition 11.11.2 Let Σ ⊆ Rn be a closed Ahlfors regular set and abbreviate

σ := H n−1 Σ. Also, suppose 0 < s < s1 < 1 and 1 ≤ p ≤ ∞. Then

W s1, p (Σ, σ) → Bs (Σ, σ) for 0 < q ≤ ∞,

p,q
(11.11.3)
p, p p,∞
Bs (Σ, σ) → W s, p (Σ, σ) → Bs (Σ, σ), (11.11.4)

W s1, p (Σ, σ) → Fs (Σ, σ) for

p,q n−1
n < q ≤ ∞. (11.11.5)

Proof The inclusions in (11.11.3) and (11.11.5) are implied by [152, Proposition 5.1,
p. 293]. To prove the first inclusion in (11.11.4), consider an arbitrary function
p, p
f ∈ Bs (Σ, σ). According to [64, Theorem 1.2] there exists a family {gk }k ∈Z of
non-negative σ-measurable functions on Σ along with a σ-nullset N ⊆ Σ such that
 1/p
p
gk  L p (Σ,σ) ≤ C f Bsp, p (Σ,σ) (11.11.6)
k ∈Z

(for some C ∈ (0, ∞) independent of f ) and with the property that

 
| f (x) − f (y)| ≤ |x − y| s gk (x) + gk (y) whenever
(11.11.7)
k ∈ Z and x, y ∈ Σ \ N satisfy 2−k−1 ≤ |x − y| < 2−k .

Hence, if we define
 1/p
g := |gk | p , (11.11.8)
k ∈Z

then g is a non-negative σ-measurable functions on Σ satisfying

 1/p
p
gk  L p (Σ,σ) = gk  L p (Σ,σ) ≤ C f Bsp, p (Σ,σ) < +∞, (11.11.9)
k ∈Z

as well as
826 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

g ≥ max gk . (11.11.10)
k ∈Z

In particular, (11.11.9) implies that g ∈ L p (Σ, σ), while from (11.11.10) and
(11.11.7) we conclude that (11.11.1) holds. Consequently, f ∈ W s, p (Σ, σ) and
(11.11.2) together with (7.9.10)-(7.9.11) imply that

 f W s, p (Σ,σ) ≤  f  L p (Σ,σ) + g L p (Σ,σ) ≤ C f Bsp, p (Σ,σ) . (11.11.11)

This establishes the first inclusion in (11.11.4).

Finally, the second inclusion in (11.11.4) is a consequence of Definition 11.11.1,
Theorem 7.9.1, and [64, Theorem 1.2] (presently used with gk := g for each k ∈ Z).

We are ready to discuss loose embeddings of our brand of boundary Sobolev

spaces into Besov spaces.

Proposition 11.11.3 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is compact as well as Ahlfors regular, and abbreviate
σ := H n−1 ∂Ω. Then
p p,q p,q p
L1 (∂Ω, σ) → Bs (∂Ω, σ) → L p (∂Ω, σ) → B−s (∂Ω, σ) → L−1 (∂Ω, σ)
continuously and densely, whenever 1 < p < ∞, 0 < q < ∞, 0 < s < 1.
(11.11.12)

Proof From item (vi) in Theorem 11.5.18 and Definition 11.11.1 it follows that
p
L1 (∂Ω, σ) → W s, p (∂Ω, σ), 1 < p < ∞, 0 < s < 1. (11.11.13)

In concert with Proposition 11.11.2 and Lemma 7.1.10, this further implies
p p,q
L1 (∂Ω, σ) → Bs (∂Ω, σ) continuously and densely
(11.11.14)
whenever 1 < p < ∞, 0 < s < 1, 0 < q < ∞.

Via duality (cf. Proposition 7.6.1 and (11.8.1)) we then obtain from this that
p,q p
B−s (∂Ω, σ) → L−1 (∂Ω, σ), 1 < p, q < ∞, 0 < s < 1. (11.11.15)

Via real interpolation (cf. Theorem 7.4.2), we may now allow arbitrary exponents
q ∈ (0, ∞) in (11.11.15). As may be seen with the help of Theorem 7.9.1, we
p,q p,q
also have Bs (∂Ω, σ) → L p (∂Ω, σ). Lastly, that L p (∂Ω, σ) → B−s (∂Ω, σ) is a
well-defined and continuous embedding is clear from (7.1.55), (7.7.5), and (7.7.3).
Bearing (7.7.3) in mind, we may now conclude that all inclusions in (11.11.12)
are well-defined, linear, and continuous. Finally, that said inclusions also have dense
ranges is a consequence of the fact that Lip (∂Ω) is a dense subset of all spaces
involved in (11.11.12) (cf. (11.5.200), (7.1.62), (11.8.4), and [133, (3.7.22)]). The
proof of (11.11.12) is therefore complete. 

In fact, the following tight embedding also holds.

11.11 Embedding Results Involving Boundary Sobolev Spaces 827

Proposition 11.11.4 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is compact as well as Ahlfors regular, and abbreviate
σ := H n−1 ∂Ω. Then
p, p p∗
Bs−1 (∂Ω, σ) → L−1 (∂Ω, σ) if n−1
n < p ≤ 1,
1   
s −1
(11.11.16)
(n − 1) p − 1 < s < 1, and p∗ := p1 − n−1 .

Proof Let (p∗ ) be the conjugate exponent of p∗ and fix some η ∈ (1 − s, 1). In a

first stage we shall prove that there exists a constant C ∈ (0, ∞) such that for each
η-smooth atoms a of type (p, s − 1) on ∂Ω we have
∫ 
 
 a f dσ  ≤ C f  L (p ∗ ) (∂Ω,σ)
∂Ω 1 (11.11.17)
(p ∗ )  p∗ ∗
for all f ∈ L1 (∂Ω, σ) = L−1 (∂Ω, σ) .

Denote by Δ(xa, r) := B(xa, r) ∩ ∂Ω the surface ball with center xa ∈ ∂Ω and radius
r > 0 where the atom a is known to be supported. Also, fix an arbitrary function
(p ∗ )
f ∈ L1 (∂Ω, σ). Then, with α := 1 − (p ∗ ) ∈ (0, 1), the embedding in (11.5.205)
n−1
∗
(p ) ∫
α
presently implies L1 (∂Ω, σ) → C (∂Ω). Since Δ(x ,r) a dσ = 0 we may then
a
write ∫ ∫
a f dσ = a(x)( f (x) − f (xa )) dσ(x), (11.11.18)
Δ(x a ,r) Δ(x a ,r)

hence
∫  ∫
 
 a f dσ  ≤ |a(x)|| f (x) − f (xa )| dσ(x)
∂Ω Δ(x a ,r)
∫
n−1
≤ r s− p −1 |x − xa | α dσ(x)  f C α (∂Ω)
Δ(x a ,r)
n−1
≤ C r s− p −1 r n−1 r α  f  L (p ∗ ) (∂Ω,σ)
1

= C f  L (p ∗ ) (∂Ω,σ) . (11.11.19)
1

(p ∗ )
The second inequality in (11.11.19) uses f ∈ L1 (∂Ω, σ) ⊂ C α (∂Ω) and the
s− n−1
p −1
size estimate of the atom a, i.e., |a(x)|
 ≤r for x ∈ Δ(xa, r). The third
inequality takes into account that σ Δ(xa, r) ≈ r n−1 and |x − xa | α ≈ r α , uniformly
for x ∈ Δ(xa, r). Finally, the last inequality in (11.11.19) is based on the observation
p − 1 + n − 1 + α = 0. This establishes (11.11.17).
that s − n−1
In fact, (11.11.17) remains true in the case in which a is an η-smooth block of
type (p, s − 1) on ∂Ω supported in a surface ball of radius r, although the constant C
intervening in (11.11.17) is allowed to depend on r. Specifically, we now have
828 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ 
 
 a f dσ  ≤ a L p ∗ (∂Ω,σ)  f  L (p ∗ ) (∂Ω,σ)
∂Ω
  1/p∗
≤ a L ∞ (∂Ω,σ) σ Δ(xa, r)  f  L (p ∗ ) (∂Ω,σ)
1

−1
≤ Cr  f  L (p ∗ ) (∂Ω,σ) . (11.11.20)
1

p, p
Next, consider an arbitrary f ∈ Bs (∂Ω, σ). Then Theorem 7.2.7 ensures the
existence of a sequence {a j } j ∈N of functions which are either smooth η-blocks of
type (p, s−1) on ∂Ω supported in surface balls of a fixed radius ro which depends only
on the geometry, or smooth η-atoms of type (p, s − 1) on ∂Ω, along with a numerical
sequence {λ j } j ∈N belonging to p (a space with which bp, p (∂Ω) identifies; cf.
(7.2.24) in Definition 7.2.5) with the property that

∞ 1/p
|λ j | p ≤ C f B p, p (∂Ω,σ) < +∞, (11.11.21)
s−1
j=1

for some constant C ∈ (0, ∞) independent of f , and such that if


m
fm := λ j a j for each m ∈ N (11.11.22)
j=1

then
p, p
fm −→ f in Bs−1 (∂Ω, σ) as m → ∞. (11.11.23)
Bearing in mind that we are presently assuming 0 < p ≤ 1, based on (11.11.17) and
(11.11.20) used with r := ro > 0, a fixed geometric constant, we may estimate
∫ 
 
 fm  L p (∂Ω,σ) =
∗ sup  fm g dσ 
−1 g  ≤1 ∂Ω
(p ∗ )
L (∂Ω, σ)
1

+ ,

m ∫ 
 
≤ sup |λ j |  a j g dσ 
g  (p ∗ ) ≤1 j=1 ∂Ω
L (∂Ω, σ)
1


m 
m 1/p
≤C |λ j | ≤ C |λ j | p
j=1 j=1

≤ C f B p, p (∂Ω,σ), (11.11.24)
s−1

for some constant C ∈ (0, ∞) which is independent of f and m. Likewise, given

any m , m  ∈ N with m  ≤ m , we may estimate (again, mindful of the fact that
0 < p ≤ 1)
11.11 Embedding Results Involving Boundary Sobolev Spaces 829

m
 1/p
fm −fm  p∗
L−1 (∂Ω,σ)
≤C |λ j | p , (11.11.25)
j=m

for some constant C ∈ (0, ∞) which is independent of f , m, m . In particular, from

p∗
this and (11.11.21) we conclude that { fm }m∈N is a Cauchy sequence in L−1 (∂Ω, σ).
p∗
Since the latter is a Banach space, there exists f ∈ L−1 (∂Ω, σ) with the property that
p ∗
the sequence { fm }m∈N converges in L−1 (∂Ω, σ) to f . Then (11.11.24) shows that
 
f  p∗ ≤ C f B p, p (∂Ω,σ), (11.11.26)
L−1 (∂Ω,σ) s−1

for some constant C ∈ (0, ∞) independent of f . The argument so far shows that we
p, p p∗
have a bounded mapping Bs−1 (∂Ω, σ)  f → f ∈ L−1 (∂Ω, σ). We shall show that
this mapping is also injective and linear. To this end, observe that since p1 − n−1
s
<1
it is possible to select
1 
s −1
s ∈ (0, s) ⊆ (0, 1) such that p := p − s
n−1 + n−1 ∈ (1, ∞). (11.11.27)

Then −1 < s −1 < s −1 < 0 and 1/p−(s −1)/(n−1) = 1/ p−(s −1)/(n−1). As such,
Theorem 7.7.4 (presently employed with Σ := ∂Ω) together with Proposition 11.11.3
give
p, p p, p p
Bs−1 (∂Ω, σ) → Bs−1 (∂Ω, σ) → L−1 (∂Ω, σ) continuously. (11.11.28)

Moreover, since 1/p∗ = 1/ p − s/(n − 1) < 1/ p, we conclude that 1 < p < p∗ , hence
p∗ p
L−1 (∂Ω, σ) → L−1 (∂Ω, σ) continuously. (11.11.29)
p∗
Recall that the sequence { fm }m∈N converges to f in L−1 (∂Ω, σ) hence also in
p
L−1 (∂Ω, σ), by (11.11.29). On the other hand, f is the limit of { fm }m∈N in
p, p p
Bs−1 (∂Ω, σ) hence also in L−1 (∂Ω, σ), by (11.11.28). Ultimately, this proves that
p
f = f in L−1 (∂Ω, σ), and all desired conclusions follow from this. 
We next discuss how Hardy spaces embed into negative Sobolev spaces, on the
boundary of open set satisfying a two-sided local John condition when said boundary
is compact and Ahlfors regular.
Proposition 11.11.5 Suppose Ω ⊆ Rn , where n ∈ N with n ≥ 2, is an open set
satisfying a two-sided local John condition and whose topological boundary is a
compact Ahlfors regular set. Abbreviate σ := H n−1 ∂Ω. Then
p∗
H p (∂Ω, σ) → L−1 (∂Ω, σ) continuously and densely
∗
1 
1 −1
if n ≥ 3, n−1n < p ≤ 1, and p := p − n−1 , (11.11.30)
  −1
or if n = 2, 12 < p < 1, and p∗ := p1 − 1 .
830 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

The case when n = 2 and p = 1 is excluded from the above proceedings (naturally
so, since in this scenario p∗ becomes ∞, which is not permissible in the context of
negative Sobolev spaces; cf. Definition 11.8.1). We shall revisit this issue in [136,
p∗
§2.3], where a suitable candidate for what L−1 (∂Ω, σ) should be when p∗ = ∞ is
identified.
Proof of Proposition 11.11.5 When n ≥ 2 and n−1 n < p < 1 the continuity of the
inclusion in (11.11.30) is a direct consequence of (7.7.54) and (11.11.16). The fact
that said inclusion has dense range is seen from (4.2.13) and (11.8.4). There remains
to treat the case when n ≥ 3 and p = 1. That
(n−1)/(n−2)
H 1 (∂Ω, σ) → L−1 (∂Ω, σ) continuously if n ≥ 3 (11.11.31)

may be justified starting from the fact that, as seen from (11.5.202), (11.1.25), and
(3.1.51),

L1n−1 (∂Ω, σ) → VMO(∂Ω, σ) continuously and densely, (11.11.32)

then invoking Lemma 1.2.1 and, finally, using (11.8.1) and (4.6.22). Finally, the
embedding in (11.11.31) has dense range, thanks to (11.8.4) and the fact that
L (n−1)/(n−2) (∂Ω, σ) ⊆ H 1 (∂Ω, σ) (cf. (4.4.11) and [133, (3.6.27)]). 

In light of the embedding result established in Proposition 11.11.5, it is both mean-

ingful and important to consider the issue of compatibility between the Hardy/Hölder,
or Hardy/BMO, paring and the paring involving a Sobolev space and its dual. This
issue is addressed in the proposition below.

Proposition 11.11.6 Suppose Ω ⊆ Rn , where n ∈ N with n ≥ 2, is an open set

satisfying a two-sided local John condition and whose topological boundary is a
compact Ahlfors regular set. Abbreviate σ := H n−1 ∂Ω and, having picked some
 
p ∈ n−1
n ,1 , (11.11.33)

consider the pair of Hölder conjugate exponents

1 
1 −1
   
1 −1
p∗ := p − n−1 ∈ 1, n−1
n−2 and (p∗ ) := n
n−1 − p ∈ [n − 1, ∞). (11.11.34)
1 
If n ≥ 2 and n−1
n < p < 1, then with α := (n − 1) p − 1 ∈ (0, 1) one has
   
H p (∂Ω,σ) f, g C α (∂Ω)
= p∗
L−1 (∂Ω,σ)
f, g (∂Ω,σ)
(p ∗ )
L1
p∗
for every f ∈ H p (∂Ω, σ) → L−1 (∂Ω, σ) (11.11.35)
(p ∗ )
and every g ∈ L1 (∂Ω, σ) → C α (∂Ω),

where the above embeddings are provided by (11.11.30) and (11.5.205). Also, cor-
responding to the case when n ≥ 3 and p = 1, one has
11.12 Real Interpolation Involving Boundary Sobolev Spaces 831
   
H 1 (∂Ω,σ) f, g BMO(∂Ω,σ)
= (n−1)/(n−2)
L−1 (∂Ω,σ)
f, g
L1n−1 (∂Ω,σ)
(n−1)/(n−2)
for every f ∈ H 1 (∂Ω, σ) → L−1 (∂Ω, σ) (11.11.36)
and every g ∈ L1n−1 (∂Ω, σ) → BMO(∂Ω, σ),

where the above embeddings are provided by (11.11.30) and (11.5.202).

As was the case with Proposition 11.11.5, here we once again are forced to exclude
the case when n = 2 and p = 1. A version of Proposition 11.11.6 in this scenario is
presented in [136, §2.3].
Proof of Proposition 11.11.6 Suppose first that n ≥ 2 and n−1 n < p < 1. The
fact that both sides of the equality in (11.11.35) depend on continuously on the
distribution f ∈ H p (∂Ω, σ) means that it is enough to prove this equality when f
belongs to a dense subspace of H p (∂Ω, σ). In view of (4.4.114), it therefore suffices
∗
to treat the case when f ∈ L p (∂Ω, σ). In this scenario, Lemma 4.6.6 gives that
∫
 
p
H (∂Ω,σ) f , g C α (∂Ω)
= f g dσ for every
∂Ω
∗ (11.11.37)
f ∈ L p (∂Ω, σ) → H p (∂Ω, σ) and
(p ∗ )
every g ∈ L1 (∂Ω, σ) → C α (∂Ω),

while (11.8.5) gives that

∫
 
p∗
L−1 (∂Ω,σ)
f, g (p ∗ )
L1 (∂Ω,σ)
= f g dσ for every
∂Ω
∗ p∗ (11.11.38)
f ∈ L p (∂Ω, σ) → L−1 (∂Ω, σ) and
(p ∗ ) ∗ 
every g ∈ L1 (∂Ω, σ) → L (p ) (∂Ω, σ).

From (11.11.37) and (11.11.38) we then conclude that (11.11.35) holds. Finally, the
case when n ≥ 3 and p = 1 is handled similarly, now using Lemma 4.6.5. 

11.12 Real Interpolation Involving Boundary Sobolev Spaces

Recall that if Σ ⊆ Rn is a closed Ahlfors regular set and σ := H n−1 Σ then,

as a special case of (9.4.19), the trace operator TrRn →Σ from (9.4.18) induces a
well-defined bounded linear mapping
p,1 p,1
TrRn →Σ : B (Rn ) −→ Bs (Σ, σ)
s+ p1 (11.12.1)
for p ∈ (1, ∞) and 0 < s < 1.

In the lemma below we deal with the end-point cases s = 0 and s = 1 of this result,
assuming Σ is the boundary of an Ahlfors regular domain Ω ⊆ Rn .
832 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Lemma 11.12.1 Assume Ω ⊆ Rn is an Ahlfors regular domain and abbreviate

σ := H n−1 ∂Ω. Then for every p ∈ (1, ∞), the boundary trace operators
p,1
TrRn →∂Ω : B1/p (Rn ) −→ L p (∂Ω, σ) (11.12.2)

and
p,1 p
TrRn →∂Ω : B1+1/p (Rn ) −→ L1 (∂Ω, σ) (11.12.3)
are well-defined, linear, and bounded.

Proof The first conclusion (11.12.2) is a consequence of [191, Proposition 6.64,

p. 220]). Turning attention to (11.12.3), we us start by using (11.12.2) and the
p,1 p,1
embedding B1+1/p (Rn ) → B1/p (Rn ) to conclude that the operator

p,1
TrRn →∂Ω : B1+1/p (Rn ) −→ L p (∂Ω, σ)
(11.12.4)
is well defined, linear, and bounded.
p,1 p,1
Going further, suppose u ∈ B1+1/p (Rn ). Then u, ∇u ∈ B1/p (Rn ) hence, (11.12.4)
p,1 p,1
and the fact that B1+1/p (Rn ) → B1/p (Rn ) continuously ensure that there exists a
constant C ∈ (0, ∞) such that
 n
TrRn →∂Ω u ∈ L p (∂Ω, σ), TrRn →∂Ω (∇u) ∈ L p (∂Ω, σ) ,
 
TrRn →∂Ω u p ≤ CuB p,1 (Rn ), and (11.12.5)
L (∂Ω,σ)
  1+1/p
TrRn →∂Ω (∇u) p ≤ CuB p,1 (Rn ) .
[L (∂Ω,σ)] n 1+1/p

p,1
Also, since the space of Schwartz functions is dense in B1+1/p (Rn ) (cf. (9.1.12)),
there exists a sequence {φi }i ∈N of Schwartz functions which converges to u in
p,1
B1+1/p (Rn ). As a consequence of this, (11.12.4), (8.1.12), and the continuity of the
p,1  p,1 n
operator ∇ : B1+1/p (Rn ) → B1/p (Rn ) , we have

φi ∂Ω = TrRn →∂Ω φi −→ TrRn →∂Ω u in L p (∂Ω, σ) as i → ∞, and
  n
∇φi ∂Ω = TrRn →∂Ω (∇φi ) −→ TrRn →∂Ω (∇u) in L p (∂Ω, σ) as i → ∞.
(11.12.6)
To proceed, observe that the present hypotheses imply that Ω is of locally
finite perimeter and that the geometric measure theoretic outward unit normal
ν = (ν1, . . . , νn ) to Ω is defined at σ-a.e. point on ∂Ω. Let us also fix two in-
dices j, k ∈ {1, . . . , n} and pick an arbitrary ψ ∈ Cc∞ (Rn ). On account of (11.12.6),
(11.1.7), and (11.1.4) we may then write
11.12 Real Interpolation Involving Boundary Sobolev Spaces 833
∫
 
(TrRn →∂Ω u) ∂τ j k ψ dσ
∂Ω
∫
 
= lim (TrRn →∂Ω φi ) ∂τ j k ψ dσ
i→∞ ∂Ω
∫ ∫
= lim φi ∂τ j k ψ dσ = lim ∂τk j φi ψ dσ
i→∞ ∂Ω i→∞ ∂Ω
∫
= lim (νk ∂j φi − ν j ∂k φi )ψ dσ
i→∞ ∂Ω
∫
 
= lim νk TrRn →∂Ω (∂j φi ) − ν j TrRn →∂Ω (∂k φi ) ψ dσ
i→∞ ∂Ω
∫
 
= νk TrRn →∂Ω (∂j u) − ν j TrRn →∂Ω (∂k u) ψ dσ. (11.12.7)
∂Ω

In light of Definition 11.1.2 and the convention made in (11.1.17), this proves that

∂τ j k (TrRn →∂Ω u) = ν j TrRn →∂Ω (∂k u) − νk TrRn →∂Ω (∂j u) ∈ L p (∂Ω, σ). (11.12.8)

In addition,
    
∂τ TrRn →∂Ω u  p = ν j TrRn →∂Ω (∂k u) − νk TrRn →∂Ω (∂j u) L p (∂Ω,σ)
jk L (∂Ω,σ)
 
≤ C TrRn →∂Ω (∇u)[L p (∂Ω,σ)]n

≤ CuB p,1 ,
(R n )
(11.12.9)
1+1/p

where the equality in (11.12.9) is just (11.12.8),

 the first
 n inequality follows from the
triangle inequality and the fact that ν ∈ L ∞ (∂Ω, σ) , and the second inequality
comes from (11.12.5).
From (11.12.8)-(11.12.9), (11.12.5), (11.1.13), (11.1.64), and the convention
made in (11.1.19) we then conclude that
p
TrRn →∂Ω u ∈ L1 (∂Ω, σ) and
(11.12.10)
TrRn →∂Ω u L p (∂Ω,σ) ≤ CuB p,1 ,
(R n )
1 1+1/p

for some constant C ∈ (0, ∞) independent of u. This establishes (11.12.3). 

It is remarkable that our brand of boundary Sobolev space produces, when in-
terpolated with Lebesgue spaces via the real method, Besov spaces. This is made
precise in the theorem below.

Theorem 11.12.2 Assume Ω ⊆ Rn is an open set satisfying a two-sided local John

condition and whose topological boundary is compact and Ahlfors regular. Abbre-
viate σ := H n−1 ∂Ω and fix p ∈ (1, ∞) along with q ∈ (0, ∞] and θ ∈ (0, 1). Then
the following real interpolation result holds:
834 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
 p  p,q
L p (∂Ω, σ), L1 (∂Ω, σ) θ,q = Bθ (∂Ω, σ). (11.12.11)
p
Proof Let us first deal with the left-to-right inclusion in (11.12.11). Since L1 (∂Ω, σ)
embeds continuously into L p (∂Ω, σ), we have
 p p 
L (∂Ω, σ), L1 (∂Ω, σ) θ, p → L p,q (∂Ω, σ). (11.12.12)
 p 
For any function f ∈ L p (∂Ω, σ), L1 (∂Ω, σ) θ,q we have f ∈ L p,q (∂Ω, σ) by
(11.12.12), and our goal is to show that there exists a constant C ∈ (0, ∞) independent
of f such that

 f B p, q (∂Ω,σ) ≤ C f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q . (11.12.13)

θ 1

Let the integer κ∂Ω be defined as in (7.1.24) with Σ := ∂Ω. In particular, (7.1.23)
gives that 12 diam(∂Ω) < 2−κ∂Ω −1 ≤ diam(∂Ω). Given that L1 (∂Ω, σ) embeds con-
p

tinuously into L (∂Ω, σ), it has been noted

p
 in (1.3.49) that we may
 take as norm in
p
the real interpolation intermediate space L p (∂Ω, σ), L1 (∂Ω, σ) θ,q the quantity
∫
⎧
⎪
diam(∂Ω)   q dt 1
⎪
⎪ t −θ K(t, f )
q
if q < ∞,
⎨
⎪ t
 f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q = 0
 −θ 
1 ⎪
⎪
⎪
⎪ sup t K(t, f ) if q = ∞,
⎩ t ∈(0,diam(∂Ω))
(11.12.14)
 p 
for each f ∈ L p (∂Ω, σ), L1 (∂Ω, σ) θ,q where
 p 
K(t, f ) := K t, f , L p (∂Ω, σ), L1 (∂Ω, σ) (11.12.15)

= inf g L p (∂Ω,σ) + t h L p (∂Ω,σ) : f = g + h with
1

p
g ∈ L p (∂Ω, σ) and h ∈ L1 (∂Ω, σ) .
 p 
Until further notice, assume q ∈ [1, ∞]. Fix f ∈ L p (∂Ω, σ), L1 (∂Ω, σ) θ,q
arbitrary. If q < ∞, then on account of (1.3.36)-(1.3.37) we may estimate
11.12 Real Interpolation Involving Boundary Sobolev Spaces 835
∫ 
∞ ∫ 2−k
diam(∂Ω)   q dt 1
q   q dt 1
q
t −θ K(t, f ) ≥ t −θ K(t, f )
0 t 2−(k+1) t
k=κ∂Ω +1


∞ ∫ 2−k 1
1 q
≥ K(2−k , f )q t −θq−1 dt
2 2−(k+1)
k=κ∂Ω +1

1 2θq − 1 1
q

∞
 q 1
q
= 2kθ K(2−k , f ) .
2 θq
k=κ∂Ω +1
(11.12.16)

On the other hand, if q = ∞ then (1.3.36)-(1.3.37) permit us to write

 −θ   −θ 
sup t K(t, f ) ≥ sup sup t K(t, f )
t ∈(0,diam(∂Ω)) k ∈Z 2−(k+1) <t ≤2−k
k ≥ κ∂Ω +1
 
≥C sup 2kθ K(2−k , f ) . (11.12.17)
k ∈Z
k ≥ κ∂Ω +1

The intention now is to prove that there exists a constant C ∈ (0, ∞) with the
property that

K(2−k , f ) ≥ CEk f  L p (∂Ω,σ) for each k ∈ Z with k ≥ κ∂Ω + 1, (11.12.18)

where the family of conditional expectation operators {Ek }k ∈Z,k ≥κ∂Ω is an in Defi-
nition 7.1.5 (cf. (7.1.30)) with Σ := ∂Ω. In particular,

Ek = Sk − Sk−1 for each k ∈ Z with k ≥ κ∂Ω + 1, (11.12.19)

where the family of operators {St }0<t<diam(∂Ω) is an approximation of identity, with

integral kernels {St (·, ·)}0<t<diam(∂Ω) as in (4.2.14) (with Σ := ∂Ω).
p
Let g ∈ L p (∂Ω, σ) and h ∈ L1 (∂Ω, σ) be such that g + h = f . Also, fix k ∈ Z
with k ≥ κ∂Ω + 1. Then Ek f = Ek g + Ek h, which implies

Ek f  L p (∂Ω,σ) ≤ Ek g L p (∂Ω,σ) + Ek h L p (∂Ω,σ) . (11.12.20)

The term Ek g L p (∂Ω,σ) may be estimated as follows. From (4.2.14) we see that
there exists some positive finite constant C0 such that for each x ∈ ∂Ω we may write
836 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫   
 
|(Ek g)(x)| =  Sk (x, y) − Sk−1 (x, y) g(y) dσ(y)
∂Ω
 ∫ 
   
= Sk (x, y) − Sk−1 (x, y) g(y) dσ(y)
y ∈∂Ω
|x−y | ≤C0 2−k+1
∫
≤C (2−k )1−n |g(y)| dσ(y)
y ∈∂Ω
|x−y | ≤C0 2−k+1
⨏
≈ |g| dσ ≤ C(M ∂Ω g)(x), (11.12.21)
B(x,C0 2−k )∩∂Ω

where M ∂Ω denotes the Hardy-Littlewood maximal operator on ∂Ω. Since this is

bounded on L p (∂Ω, σ) (see [133, Corollary 7.6.3]), we ultimately obtain

Ek g L p (∂Ω,σ) ≤ Cg L p (∂Ω,σ) . (11.12.22)

We now turn to estimating the term Ek h L p (∂Ω,σ) in (11.12.20). Fix x ∈ ∂Ω and
for each r > 0 consider ⨏
hx,r := h dσ. (11.12.23)
B(x,C0 r)∩∂Ω
∫  
Thanks to the fact that ∂Ω
Sk (x, y) − Sk−1 (x, y) dσ(y) = 0 (cf. (4.2.14)), we have
∫   
 
|(Ek h)(x)| =  Sk (x, y) − Sk−1 (x, y) h(y) dσ(y)
∂Ω
 ∫ 
    
= Sk (x, y) − Sk−1 (x, y) h(y) − hx,2−k dσ(y)
y ∈∂Ω
|x−y | ≤C0 2−k+1
⨏  h(y) − h −k 
 x,2 
≤ C 2−k  −k+1
 dσ(y)
B(x,C0 2−k+1 )∩∂Ω C0 2

≤ C 2−k (Λ∗,1 h)(x) (11.12.24)

where we have set (cf. (11.5.173))

⨏  h(y) − h 
 x,r 
(Λ∗,1 h)(x) := sup   dσ(y). (11.12.25)
0<r <diam(∂Ω) B(x,r)∩∂Ω r

From (11.5.218) we know that

Λ∗,1 h L p (∂Ω,σ) ≤ Ch L p (∂Ω,σ), (11.12.26)

1
11.12 Real Interpolation Involving Boundary Sobolev Spaces 837

therefore
Ek h L p (∂Ω,σ) ≤ C2−k h L p (∂Ω,σ) . (11.12.27)
1

Together with (11.12.20) and (11.12.22), this leads to the conclusion that
 
Ek f  L p (∂Ω,σ) ≤ C g L p (∂Ω,σ) + 2−k h L p (∂Ω,σ) . (11.12.28)

Taking the infimum over all possible decompositions of f as g + h with g, h as above

then finally yields (11.12.18).
Pressing on, recall from Definition 7.1.5 that the norm of f in the space
p,q
Bθ (∂Ω, σ) is given by
    p1
p
N (
κ∂Ω,τ)

f p, q
B θ (∂Ω,σ) = σ(Qτκ∂Ω,ν ) mQ κ ∂Ω,ν |Eκ∂Ω f |
τ
τ ∈Iκ ∂Ω ν=1

! " q1
  q
+ 2kθ Ek f  L p (∂Ω,σ) , (11.12.29)
k ∈Z
k ≥ κ∂Ω +1

with a natural modification in the last term above when q = ∞.

As regards the second term in the right-hand side of (11.12.29), note that if q < ∞
we may estimate
! " q1 ! " q1
  q   q
−k
2 Ek f  L p (∂Ω,σ)
kθ
≤C 2 K(2 , f )
kθ

k ∈Z k ∈Z
k ≥ κ∂Ω +1 k ≥ κ∂Ω +1

∫ diam(∂Ω)   q dt 1
q
≤C t −θ K(t, f )
0 t

≤ C f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q (11.12.30)

1

by (11.12.18), (11.12.16), and (11.12.14). This, of course, suits our purposes. Also,
if q = ∞, then (11.12.18), (11.12.17), and (11.12.14) imply
   
sup 2kθ Ek f  L p (∂Ω,σ) ≤ C · sup 2kθ K(2−k , f )
k ∈Z k ∈Z
k ≥ κ∂Ω +1 k ≥ κ∂Ω +1
 
≤C· sup t −θ K(t, f )
t ∈(0,diam(∂Ω))

≤ C f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q (11.12.31)

1

which, again, is a bound of the right order.

It remains to take care of the first term in the right-hand side of (11.12.29). By
Hölder inequality, we have
838 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
    p1
p
N (
κ∂Ω,τ)

σ(Qτκ∂Ω,ν ) m κ ,ν |Eκ∂Ω f |
Qτ ∂Ω
τ ∈Iκ ∂Ω ν=1

  N (
κ∂Ω,τ) ⨏  p1
 p
≤ σ(Qτκ∂Ω,ν )  Eκ f  dσ
κ ,ν ∂Ω
τ ∈Iκ ∂Ω ν=1 Qτ ∂Ω

∫ ∫
 p 1
  1

≤  Eκ f  dσ
p
≤C M ∂Ω f  p dσ p
∂Ω
∂Ω ∂Ω
 
≤ C f  L p (∂Ω,σ) ≤ C g L p (∂Ω,σ) + h L p (∂Ω,σ)
 
≤ C g L p (∂Ω,σ) + h L p (∂Ω,σ) (11.12.32)
1

 
since the dyadic cubes Qτκ∂Ω,ν : τ ∈ Iκ∂Ω , 1 ≤ ν ≤ N( κ∂Ω, τ) are mutually disjoint.
By taking the infimum over all g, h as before, we therefore arrive at
    p1
p
N (
κ∂Ω,τ)
  
σ(Qτκ∂Ω,ν ) m κ ,ν |Eκ∂Ω f | ≤ C K 2−κ∂Ω −1, f (11.12.33)
Qτ ∂Ω
τ ∈Iκ ∂Ω ν=1

for some constant C = C(Ω, p) ∈ (0, ∞) independent of f . On the other hand, from
(11.12.16) and (11.12.14) it is clear that in the case when q < ∞ we have
∫
  diam(∂Ω)   q dt 1
q
K 2−κ∂Ω −1, f ≤ C t −θ K(t, f )
0 t

= C f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q . (11.12.34)

1

Also, when q = ∞, from (11.12.17) and (11.12.14) we see that

   −θ 
K 2−κ∂Ω −1, f ≤ C sup t K(t, f )
t ∈(0,diam(∂Ω))

≤ C f (L p (∂Ω,σ), L p (∂Ω,σ))θ,∞ . (11.12.35)

1

Combining (11.12.33) and (11.12.34)-(11.12.35) gives

    p1
p
N (
κ∂Ω,τ)

σ(Qτκ∂Ω,ν ) m κ ,ν |Eκ∂Ω f |
Qτ ∂Ω
τ ∈Iκ ∂Ω ν=1

≤ C f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q (11.12.36)

1

which, in concert with (11.12.30)-(11.12.31) and (11.12.29), establishes (11.12.13).

At this point, the left-to-right inclusion in (11.12.11) is proved.
11.12 Real Interpolation Involving Boundary Sobolev Spaces 839

Next we turn our attention to proving the right-to-left inclusion in (11.12.11), i.e.,
p,q  p 
Bθ (∂Ω, σ) → L p (∂Ω, σ), L1 (∂Ω, σ) θ,q . (11.12.37)
p,q
In this regard, the goal is to show that for each f ∈ Bθ (∂Ω, σ) we have the estimate

 f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q ≤ C f B p, q (∂Ω,σ), (11.12.38)

1 θ

for some finite constant C = C(Ω, p, q, θ) > 0 independent of f . To set the stage, we
recall that an equivalent norm in the intermediate real interpolation space may be
given in terms of J-functional (cf. [15, § 3.2-§ 3.3, pp. 42–46]). Specifically, in our
case we have

 f (L p (∂Ω,σ), L p (∂Ω,σ))θ, q (11.12.39)

1
! ∫ ∫ "
∞   q dt 1 ∞
q dt
≈ inf t −θ J(t, u(t; ·)) : f = u(t; ·) ,
0 t 0 t
p
(naturally altered when q = ∞) where (0, ∞)  t → u(t; ·) ∈ L1 (∂Ω, σ) is a
measurable function and, for each t > 0,
  
J t, u(t; ·) := max u(t; ·) L p (∂Ω,σ), t u(t; ·) L p (∂Ω,σ) . (11.12.40)
1

p,q
Thus, given a fixed, arbitrary function f ∈ Bθ (∂Ω, σ),
it suffices to construct a
p
family of functions {u(t; ·)}t>0 ⊂ L1 (∂Ω, σ), depending in a measurable fashion on
the parameter t, and satisfying the following properties:
∫ ∞
dt
f (x) = u(t; x) for σ-a.e. x ∈ ∂Ω, (11.12.41)
0 t
∫ ∞
 −θ  q dt q1
t u(t; ·) L p (∂Ω,σ) ≤ C f B p, p (∂Ω,σ), (11.12.42)
0 t θ

∫ ∞
 1−θ  q dt q1
t u(t; ·) L p (∂Ω,σ) ≤ C f B p, q (∂Ω,σ), (11.12.43)
0 1 t θ

with natural alterations in the case when q = ∞.

To this end, recall the extension operator ExJWS
∂Ω→R n from Theorem 9.4.1, used
with Σ := ∂Ω. Also, fix some background cutoff function η ∈ Cc∞ (Rn ) with the
property that η ≡ 1 near ∂Ω, and define F := η · ExJWS∂Ω→R n f . It follows that F is
compactly supported and satisfies
p,q
F ∈ Bθ+1/p (Rn ), F B p, q (R n ) ≈  f B p, q (∂Ω,σ),
θ +1/p θ
(11.12.44)
and TrRn →∂Ω F = f .

Let now ϕ be a Schwartz function in Rn such that

840 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

supp ϕ ⊆ {ξ ∈ Rn : 2−1 ≤ |ξ | ≤ 2}, ϕ(ξ) > 0 if 2−1 < |ξ | < 2,

 (11.12.45)
and ∞ −k
k=−∞ ϕ(2 ξ) = 1 for each ξ ∈ R \ {0}.
n

The existence of such a function is well known (see, e.g., [15, Lemma 6.1.7, p. 135]).
Also, consider the Schwartz functions ψ and ϕk , k ∈ Z, such that

∞
/(ξ) = 1 −
.k (ξ) = ϕ(2−k ξ) for k ∈ N and ψ
ϕ ϕ
.k (ξ). (11.12.46)
k=1

Thus, since supp ϕ /k ⊆ {ξ ∈ Rn : 2k−1 ≤ |ξ | ≤ 2k+1 } for every k ∈ N, it follows that

.k ϕ/j = 0 for every j, k ∈ N such that | j − k | ≥ 2. Hence, ultimately,
ϕ

ϕ j ∗ ϕk = 0 in Rn if | j − k | ≥ 2. (11.12.47)

Also, from definitions,


∞
ψ+ ϕk = δ, Dirac’s distribution in Rn, (11.12.48)
k=1

and from Lemma 6.2.1 (and its proof) on p. 140 in [15], for every j, k ∈ N we have

ϕ j ∗ ϕk ∗ F  L p (Rn, L n ) ≤ Cϕk ∗ F  L p (Rn, L n ) if | j − k | ≤ 2, (11.12.49)

where the constant C ∈ (0, ∞) is independent of j, k. For the uniformity of notation,

set ϕ0 := ψ. It follows that

∞
p,q
F= ϕk ∗ F with convergence in Bθ+1/p (Rn ), and (11.12.50)
k=0

+ ∞ 
, q1
 q
(θ+1/p)k
F B p, q (R n ) ≈ 2 ϕk ∗ F  L p (Rn, L n ) , (11.12.51)
θ +1/p
k=0

(with a natural alteration in the case when q = ∞) where the equivalence in (11.12.51)
is a well-known by-product of the Littlewood-Paley characterization of Besov spaces
in the Euclidean ambient (see, e.g., [15, Definition 6.2.2, pp. 140–141]). Let us also
note here that, as a consequence of (11.12.50), the last formula in (11.12.44), and
the continuity of the trace operator (cf. Theorem 9.4.1) we have

∞
f = TrRn →∂Ω (ϕk ∗ F). (11.12.52)
k=0

After this preamble, we now define

11.12 Real Interpolation Involving Boundary Sobolev Spaces 841
!
0 if t > 1,
u(t; ·) := (11.12.53)
1
ln 2 TrRn →∂Ω (ϕk ∗ F) if 2−(k+1) < t < 2−k , k ∈ N0 .

Since (11.12.41) is readily verified from (11.12.53) and (11.12.52), there remains to
prove that the estimates claimed in (11.12.42)-(11.12.43) hold as well. To check the
validity of (11.12.42) in the case when q < ∞ we write
∫ ∞ ∞ ∫ 2−k
dt
t −qθ u(t; ·) L p (∂Ω,σ) t −qθ−1 ϕk ∗ F  p,1 n dt
q q
≤C
0 t 2−(k+1) B1/p (R )
k=0


∞ 
∞ q
≤C 2kq θ 2 j/p ϕ j ∗ ϕk ∗ F  L p (Rn )
k=0 j=0


∞
q
≤C 2kq(1/p+θ) ϕk ∗ F  L p (Rn, L n )
k=0
q q
≤ CF B p, q (R n )
≤ C f B p, q (∂Ω,σ) . (11.12.54)
θ +1/p θ

where the first inequality is implied by (11.12.53) and (11.12.2), the second one uses
p,1
the definition of the norm in the Besov space B1/p (Rn ) in Rn (cf. [15, Definition 6.2.2,
pp. 140–141]), the third one is based on (11.12.47) and (11.12.49), the fourth one
comes from (11.12.51) and, finally, the last inequality is justified by the equivalence
in (11.12.44). That (11.12.42) also holds in the case when q = ∞ is verified in an
analogous fashion.
Similarly, when q < ∞ we have
∫ ∞ ∞ ∫ 2−k
q dt q
t q−qθ u(t; ·) L p (∂Ω,σ) ≤C t q−qθ−1 ϕk ∗ F  p,1 n dt
0 1 t 2−(k+1) B1+1/p (R )
k=0


∞ 
∞ q
≤C 2kq(θ−1) 2 j(1+1/p) ϕ j ∗ ϕk ∗ F  L p (Rn, L n )
k=0 j=0


∞
q
≤C 2kq(θ+1/p) ϕk ∗ F  L p (Rn, L n )
k=0
q q
≤ CF B p, q (R n )
≤ C f B p, q (∂Ω,σ), (11.12.55)
θ +1/p θ

where the first inequality is a consequence of (11.12.3) and the rest follows as in
(11.12.54). The case when q = ∞ is treated in an analogous manner. Thus (11.12.38)
holds, which establishes (11.12.37). At this stage, the proof of (11.12.11) is complete
in the case q ∈ [1, ∞].
An alternative proof of (11.12.38) (which implies (11.12.37)) in the case when
q ∈ [1, ∞] goes as follows. First, Lemma 11.12.1 and real interpolation ensure that
842 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
p,q  p,1 p,1 
TrRn →∂Ω : Bθ+1/p (Rn ) = B1/p (Rn ) , B1+1/p (Rn ) θ,q
 p 
−→ L p (∂Ω, σ), L1 (∂Ω, σ) θ,q (11.12.56)

is a well-defined, linear, and bounded operator. Hence, there exists C ∈ (0, ∞) with
the property that
 
TrRn →∂Ω F  p p ≤ CF B p, q (Rn )
(L (∂Ω,σ), L1 (∂Ω,σ)) θ, q θ +1/p
p,q (11.12.57)
for each F ∈ Bθ+1/p (Rn ).
p,q
Now pick an arbitrary f ∈ Bθ (∂Ω, σ) and specialize (11.12.57) to the case when
F := ExJWS
∂Ω→R n f which, according to Theorem 9.4.1, is a function that belongs to
p,q
Bθ+1/p (Rn ) and satisfies

TrRn →∂Ω F = f and F B p, q (R n ) ≤ C f B p, q (∂Ω,σ) (11.12.58)

θ +1/p θ

for some constant C ∈ (0, ∞) independent of f . Then (11.12.38) becomes a conse-

quence of (11.12.58) and (11.12.57).
Moving on, to deal with (11.12.11) in the more general case when q ∈ (0, ∞]
we reason as follows. Fix two smoothness indices s0, s1 ∈ (0, 1) satisfying s0  s1 ,
along with some parameter θ ∈ (0, 1), and set s := (1 − θ)s0 + θs1 . Then for each
q0, q1 ∈ [1, ∞], q ∈ (0, ∞], and p ∈ (1, ∞) we may write
p,q  p,q p,q 
Bs (∂Ω, σ) = Bs0 0 (∂Ω, σ), Bs1 1 (∂Ω, σ) θ,q
 p   p 
= L p (∂Ω, σ), L1 (∂Ω, σ) s0,q0 , L p (∂Ω, σ), L1 (∂Ω, σ) s1,q1 θ,q

 p 
= L p (∂Ω, σ), L1 (∂Ω, σ) (1−θ)s0 +θs1,q
 p 
= L p (∂Ω, σ), L1 (∂Ω, σ) s,q (11.12.59)

thanks to (7.4.2), (11.12.11), and the Reiteration Formula for the real method of
interpolation (cf. (1.3.80) in Theorem 1.3.9). The proof of Theorem 11.12.2 is now
complete. 

In turn, Theorem 11.12.2 has a number of useful consequences, some of which

are collected in the next corollary.

Corollary 11.12.3 Let Ω ⊆ Rn is an open set satisfying a two-sided local John con-
dition and whose topological boundary is compact and Ahlfors regular. Abbreviate
σ := H n−1 ∂Ω. Then
 p p  p,q
L (∂Ω, σ), L−1 (∂Ω, σ) θ,q = B−θ (∂Ω, σ)
(11.12.60)
for each p ∈ (1, ∞), q ∈ (0, ∞], and θ ∈ (0, 1),
11.12 Real Interpolation Involving Boundary Sobolev Spaces 843
 p,q1 p  p,q
Bs (∂Ω, σ), L1 (∂Ω, σ) θ,q2 = B(1−θ)s+θ
2
(∂Ω, σ)
(11.12.61)
for each p ∈ (1, ∞), q1, q2 ∈ (0, ∞], and s, θ ∈ (0, 1),

 p,q1  p,q
L p (∂Ω, σ), Bs (∂Ω, σ) θ,q2 = Bθs 2 (∂Ω, σ)
(11.12.62)
for each p ∈ (1, ∞), q1, q2 ∈ (0, ∞], and s, θ ∈ (0, 1).

Proof Let us first deal with (11.12.60) when q ∈ (1, ∞]. Assume this is the case and
fix p ∈ (1, ∞) along with θ ∈ (0, 1), then pick p ∈ (1, ∞) and q  ∈ [1, ∞) such that
1/p + 1/p = 1, 1/q + 1/q  = 1. We then have

p,q  p,q ∗   p  ∗
B−θ (∂Ω, σ) = Bθ (∂Ω, σ) = L p (∂Ω, σ), L1 (∂Ω, σ) θ,q

   ∗  p ∗
= L p (∂Ω, σ) , L1 (∂Ω, σ)
θ,q
 p 
= L p (∂Ω, σ), L−1 (∂Ω, σ) θ,q (11.12.63)

where the first equality comes from Proposition 7.6.1, the second equality is provided
by Theorem 11.12.2, the third is a consequence of the Duality Theorem for the real
method of interpolation (cf. [15, Theorem 3.7.1, p. 54]), and the fourth equality uses
(11.8.1). To establish (11.12.60) for the full range q ∈ (0, ∞] we then reason as in
(11.12.59), based on what we have proved so far and the Reiteration Formula for the
real method of interpolation.
As regards (11.12.61), for each p ∈ (1, ∞), q1, q2 ∈ (0, ∞], and θ, s ∈ (0, 1) we
may write
 p,q1 p   p  p
Bs (∂Ω, σ), L1 (∂Ω, σ) θ,q2 = L p (∂Ω, σ), L1 (∂Ω, σ) s,q2 , L1 (∂Ω, σ)
θ,q
 p 
= L p (∂Ω, σ), L1 (∂Ω, σ) (1−θ)s+θ,q2

p,q
= B(1−θ)s+θ (∂Ω, σ) (11.12.64)

thanks to (11.12.11) (used twice) and the Reiteration Formula (1.3.82). Finally,
(11.12.62) is proved in a similar fashion, making use of (11.12.11) and the Reiteration
Formula (1.3.81). 

Theorem 11.12.2 has a multitude of other useful consequences, some of which

we will explore later, in connection with singular integral operators. For now, we
wish to note that by combining (11.12.11), (11.5.201), and [133, (6.2.48)] it follows
that, in the geometric context of Theorem 11.12.2,
∗
Bs (∂Ω, σ) → L p ,q (∂Ω, σ) whenever
p,q

 
s −1
(11.12.65)
1 < p < n − 1, 0 < q ≤ ∞, 0 < s < 1, and p∗ = p1 − n−1 .
844 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Moving on, fix two arbitrary indices j, k ∈ {1, . . . , n} and consider the first-order
system in Rn given by
 n
D jk u := (∂k u)e j − (∂j u)ek ∈ D (Rn ) , ∀u ∈ D (Rn ), (11.12.66)

so that
 n
Djk w = −∂k w j + ∂j wk ∈ D (Rn ), ∀w ∈ D (Rn ) . (11.12.67)

Note that, in the sense of distributions in Rn ,

divD jk = 0 and Djk ∇ = 0. (11.12.68)

Proposition 11.12.4 Assume Ω ⊆ Rn is a bounded (ε, δ)-domain with an Ahlfors

regular boundary. Abbreviate σ := H n−1 ∂Ω, and denote by ν the geometric
measure theoretic outward unit normal to Ω. Finally, fix j, k ∈ {1, . . . , n} and recall

the first-order system D jk from (11.12.66). Then, as distributions in Lip (∂Ω) , for
a given function u one has
   
∂τ j k TrΩ→∂Ω u = (−i)Sym(div; ν) D jk u, 0
 
= (−i)Sym(Djk ; ν) ∇u, 0 (11.12.69)

in any of the following scenarios:

 
(1) assume H n−1 ∂Ω \ ∂∗ Ω = 0, and the function u belongs  to the Besov space
p,q
B 1 (Ω) with n−1 n < p < ∞, 0 < q ≤ ∞, (n − 1) 1
p − 1 + < s < 1;
s+ p
 
(2) assume H n−1 ∂Ω \ ∂∗ Ω = 0, and suppose the functionu belongs  to the Triebel-
p,q
Lizorkin space F 1 (Ω) with n−1 n < p < ∞, (n − 1) 1
p − 1 + < s < 1, and
s+ p
n
< q ≤ ∞;
n+s+1/p
(3) assume Rn \ Ω is n-thick and the function u belongs to the weighted Sobolev
1, p  
space Wa (Ω) with p ∈ (1, ∞) and a ∈ − p1 , 1 − p1 .

Proof Consider the scenario described in item (1). Via embeddings (cf. Corol-
lary 9.2.1), there is no loss of generality in assuming that p = q ∈ (1, ∞) and
s ∈ (0, 1). Suppose this is the case and denote by p the Hölder conjugate exponent
of p. Also, denote by (ν1, . . . , νn ) the components of ν and pick an arbitrary function
φ ∈ C ∞ (Ω). On account of (8.5.49), (8.5.38), (11.12.66), and (11.1.4) we may write
  Adiv, D j k
(−i)Sym(div; ν) D jk φ, 0 = ∂ν (φ, 0)

= (−i)Sym(div; ν) (D jk φ)∂Ω
 
= ν j (∂k φ)∂Ω − νk (∂j φ)∂Ω = ∂τ j k φ. (11.12.70)

In addition, Proposition 9.5.5 (used with D := Djk , D := −∇, and L := 0) implies

11.12 Real Interpolation Involving Boundary Sobolev Spaces 845
 
(−i)Sym(−∇; ν)(Djk u, 0) ∈ Bs−1 (∂Ω, σ) → Lip (∂Ω) ,
p, p
(11.12.71)

while item (ii) in Theorem 9.4.5 (whose applicability in the present geometric setting
is ensured by [133, Lemma 5.11.9)]) in concert with (7.9.10) give
p, p
TrΩ→∂Ω u ∈ Bs (∂Ω, σ) → L 1 (∂Ω, σ). (11.12.72)

Bearing these memberships in mind and combining (7.6.9), (11.12.70), together with
(9.5.39) (written for D := D jk , D := div, u := φ, w := u, L := 0) yields
∫  
 
TrΩ→∂Ω u ∂τ j k φ dσ = L p (∂Ω,σ) TrΩ→∂Ω u, ∂τ j k φ L p (∂Ω,σ)
∂Ω
 
= p, p
B s (∂Ω,σ) TrΩ→∂Ω u, (−i)Sym(div; ν)(D jk φ, 0) p , p 
B−s (∂Ω,σ)
  
= p, p
B s−1 (∂Ω,σ) (−i)Sym(−∇; ν)(Djk u, 0), φ∂Ω p , p 
B1−s (∂Ω,σ)
  
= − (Lip(∂Ω)) (−i)Sym(∇; ν)(Djk u, 0), φ∂Ω Lip(∂Ω) . (11.12.73)

In turn, from this and Definition 11.2.1 we conclude that the first equality in
(11.12.69) holds. The second equality in (11.12.69) is established in a similar fashion,
now using in place of (11.12.70) the fact that for each φ ∈ C ∞ (Ω) we have

  A D  ,∇
(−i)Sym(Djk ; ν) ∇φ, 0 = ∂ν j k (φ, 0)

= (−i)Sym(Djk ; ν) (∇φ)∂Ω
 
= −νk (∂j φ)∂Ω + ν j (∂k φ)∂Ω = ∂τ j k φ, (11.12.74)

thanks to (8.5.49), (8.5.38), (11.12.67), and (11.1.4).

Finally, the same type of argument yields (11.12.69) in the scenarios described
in items (2), (3) of the statement (in the latter case appealing to (8.5.52)). 
Theorem 11.12.2 and Proposition 11.12.4 are the main ingredients in the proof of
the result below which elaborates on the manner in which our boundary tangential
derivative operators ∂τ j k act on Besov spaces in suitable geometric settings.
Proposition 11.12.5 Let Ω ⊆ Rn be an Ahlfors regular domain whose boundary is
compact. In addition, assume that

either Ω is a bounded (ε, δ)-domain,

(11.12.75)
or Ω satisfies a two-sided local John condition.

Abbreviate σ := H n−1 ∂Ω and fix j, k ∈ {1, . . . , n}.

Then the tangential derivative operator ∂τ j k , originally considered as in Propo-
sition 11.8.5, induces a well-defined, linear, and bounded mapping
846 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
p,q p,q
∂τ j k : Bs (∂Ω, σ) −→ Bs−1 (∂Ω, σ)
(11.12.76)
for each p ∈ (1, ∞), q ∈ (0, ∞], and s ∈ (0, 1).

Moreover, one has the integration by parts formula on the boundary

   
B s−1 (∂Ω,σ) ∂τ j k f , g B p , q  (∂Ω,σ) = B s (∂Ω,σ) f , ∂τk j g B−s
p, q p, q p , q 
1−s
(∂Ω,σ)
p,q p,q  p,q ∗
for all f ∈ Bs (∂Ω, σ) and g ∈ B1−s (∂Ω, σ) = Bs−1 (∂Ω, σ) (11.12.77)
if s ∈ (0, 1) and p, p q, q  ∈ (1, ∞) satisfy 1
p + 1
p = 1
q + 1
q = 1.

Proof Work first under the assumption that Ω is a bounded (ε, δ)-domain. Also,
assume for now that p, q ∈ (1, ∞) and s ∈ (0, 1). Then from (9.4.92) and the Open
Mapping Theorem we deduce that there exists C ∈ (0, ∞) with the property that for
p,q p,q
each f ∈ Bs (∂Ω, σ) we may find u ∈ B 1 (Ω) such that
s+ p

TrΩ→∂Ω u = f and uB p, q (Ω) ≤ C f Bsp, q (∂Ω,σ) . (11.12.78)

1
s+ p

On the other hand, from (11.12.69), (9.5.29), and (9.5.31)  we see

  that, with
∂τ j k TrΩ→∂Ω u originally considered as a distribution in Lip (∂Ω) (cf. Defini-
tion 11.2.1), we have
    p,q
∂τ j k TrΩ→∂Ω u = (−i)Sym(div; ν) D jk u, 0 ∈ Bs−1 (∂Ω, σ)
    (11.12.79)
and (−i)Sym(div; ν) D jk u, 0 B p, q (∂Ω,σ) ≤ CuB p, q (Ω) .
s−1 1
s+ p

Combining (11.12.78) with (11.12.79) then shows that

 
∂τ j k f ∈ Bs−1 (∂Ω, σ) and ∂τ j k f B p, q (∂Ω,σ) ≤ C f Bsp, q (∂Ω,σ) .
p,q
(11.12.80)
s−1

This proves that the operator (11.12.76) is well defined, linear, and bounded when
q ∈ (1, ∞). The extension to q ∈ (0, ∞] is then achieved using real interpolation (cf.
(7.4.2)). Finally, in the current geometric setting, the integration by parts formula on
the boundary claimed in (11.12.77) follows from (9.5.39) (used with D := D jk as in
(11.12.66) and D := div), (11.12.69), and (9.4.92).
Moving on, work under the assumption that Ω satisfies a two-sided local John
condition. Then the claims about (11.12.76) follow from Proposition 11.8.5 and the
real interpolation results established in (11.12.11) and (11.12.60). Next, observe that
(11.12.76) implies that for all f , g as in (11.12.77) we have
     
 p, q 
B (∂Ω,σ) ∂τ j k f , g B p, q (∂Ω,σ)  ≤ ∂τ j k f B p, q (∂Ω,σ) gB p, q (∂Ω,σ)
s−1 1−s s−1 1−s

≤ C f Bsp, q (∂Ω,σ) gB p, q (∂Ω,σ) (11.12.81)

1−s

and
11.12 Real Interpolation Involving Boundary Sobolev Spaces 847
     
 p, q 
Bs (∂Ω,σ) f , ∂τk j g B p, q (∂Ω,σ)  ≤  f Bsp, q (∂Ω,σ) ∂τk j g B p, q (∂Ω,σ)
−s −s

≤ C f Bsp, q (∂Ω,σ) gB p, q (∂Ω,σ) . (11.12.82)

1−s

   
Hence both B p, q (∂Ω,σ) ∂τ j k f , g B p, q (∂Ω,σ) and Bsp, q (∂Ω,σ) f , ∂τk j g B p, q (∂Ω,σ) de-
s−1 1−s −s
p,q p,q
pend on a continuous bilinear fashion on ( f , g) ∈ Bs (∂Ω, σ) ⊕ B1−s (∂Ω, σ).
Granted this, it suffices to check the equality in (11.12.77) in the special case when
( f , g) belongs to Lip (∂Ω) ⊕ Lip (∂Ω) which, by Lemma 7.1.10, is known to be a
p,q p,q
dense subspace of Bs (∂Ω, σ) ⊕ B1−s (∂Ω, σ). However, when f , g ∈ Lip (∂Ω)
the equality in (11.12.77) is a consequence of Proposition 11.1.11 and Proposi-
tion 11.1.14. This establishes the integration by parts formula on the boundary
claimed in (11.12.77) in the present geometric setting. 
We conclude this section by discussing several versatile integration by parts
formulas in the context of Besov, Triebel-Lizorkin, and weighted Sobolev spaces in
(ε, δ)-domains.
Proposition 11.12.6 Let Ω ⊆ Rn be a bounded (ε, δ)-domain. Set σ := H n−1 ∂Ω
and suppose that  
σ ∂Ω \ ∂∗ Ω = 0. (11.12.83)
(i) Fix s ∈ (0, 1) and p, p, q, q  ∈ (1, ∞) such that p1 + p1 = 1 = q1 + q1 . Then
for each pair of indices j, k ∈ {1, . . . , n} the following integration by parts formula
holds:
 
(∂Ω,σ) ∂τ j k (TrΩ→∂Ω u), TrΩ→∂Ω w B p , q  (∂Ω,σ)
p, q
B s−1 1−s
   
= B
p, q
(Ω) ∂k u, ∂j w p , q 
B (Ω)
− B p, q (Ω) ∂j u, ∂k w B
p , q 
(Ω)
s+ p1 −1 1 −1
s+ p
−s+ 1 −s+ 1
p p

p,q p,q
for each u ∈ B (Ω) and w ∈ B (Ω), (11.12.84)
s+ p1 1−s+ p1

with the pairings in the right-hand side understood in the sense of Proposition 9.2.29.
(ii) Fix s ∈ (0, 1) and p, p, q, q  ∈ (1, ∞) such that p1 + p1 = 1 = q1 + q1 . Then
for each pair of indices j, k ∈ {1, . . . , n} the following integration by parts formula
holds:
 
(∂Ω,σ) ∂τ j k (TrΩ→∂Ω u), TrΩ→∂Ω w B p , p  (∂Ω,σ)
p, p
B s−1 1−s
   
= F p, q (Ω) ∂k u, ∂j w F
p , q 
(Ω)
− F p, q (Ω) ∂j u, ∂k w p , q 
F (Ω)
1 −1
s+ p 1 −1
s+ p
−s+ 1 −s+ 1
p p

p,q p,q
for each u ∈ F (Ω) and w ∈ F (Ω), (11.12.85)
s+ p1 1−s+ p1

with the pairings in the right-hand side understood in the sense of Proposition 9.2.29.
848 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

(iii) Strengthen (11.12.83) by assuming that the set Rn \ Ω is n-thick. Fix s ∈ (0, 1)
along with p, p ∈ (1, ∞) such that p1 + p1 = 1, and define a := 1 − s − p1 . Then for
each j, k ∈ {1, . . . , n} the following integration by parts formula holds:
 
(∂Ω,σ) ∂τ j k (TrΩ→∂Ω u) , TrΩ→∂Ω w B p , p  (∂Ω,σ)
p, p
B s−1 1−s
∫ 
= (∂k u)(∂j w) − (∂j u)(∂k w) dL n
Ω

1, p 1, p
for each u ∈ Wa (Ω) and w ∈ W−a (Ω). (11.12.86)

Proof To justify the claims in items (i)-(ii), write the generalized “half” Green’s
formula (9.5.10) with f := 0 and L := 0 written as DD with D := div and D := D jk
defined as in (11.12.66) (also taking A := AD,D defined as in (8.5.48)). By also
taking into account (8.5.49) and (11.12.69) (in the settings described in items (1)-(2)
of Proposition 11.12.4) we arrive at (11.12.84)-(11.12.85).
Finally, specializing the generalized “half” Green’s formula (8.5.18) to the case
when f := 0 and L := 0 is factored out as DD with D := div and D := D jk
defined as in (11.12.66) (hence, A := AD,D as in (8.5.48)) yields (11.12.86) on
account of (8.5.49) and (11.12.69) (considered in the setting described in item (3)
of Proposition 11.12.4). 

11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev

Spaces

We begin by making the following basic definition.

Definition 11.13.1 Assume Ω ⊆ Rn (where n ≥ 2) is an Ahlfors regular domain

and set σ := H n−1 ∂Ω. Also, select an exponent p ∈ (1, ∞) along with a parameter
λ ∈ (0, n − 1). In this context, define the Morrey-based homogeneous Sobolev
space of order one on ∂Ω as
!
. p,λ σ(x) p
M1 (∂Ω, σ) := f ∈ L 1 ∂Ω, ∩ Lloc (∂Ω, σ) : ∂τ j k f ∈ M p,λ (∂Ω, σ)
1 + |x| n
"
for all j, k ∈ {1, . . . , n}

(11.13.1)

and equip this space with the semi-norm

. p,λ 
n
M1 (∂Ω, σ)  f →  f  M. p, λ (∂Ω,σ) := ∂τ j k f  M p, λ (∂Ω,σ) . (11.13.2)
1
j,k=1
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 849

Hence, in the context of Definition 11.13.1, all constant functions on ∂Ω belong to

. p,λ
M1 (∂Ω, σ) and their respective semi-norms vanish (see item (5) in Lemma 11.1.3).
From (11.13.1), (6.2.7), and item (iii) in Definition 11.1.2 we see that
. p,λ p p
M1 (∂Ω, σ) ⊆ L1,loc (∂Ω, σ) ⊆ Lloc (∂Ω, σ). (11.13.3)

It is also clear from (6.2.25) that we have the continuous embedding

p,λ . p,λ σ(x)

M1 (∂Ω, σ) → M1 (∂Ω, σ) ∩ L 1 ∂Ω, . (11.13.4)
1 + |x| n−1
After a series of preliminary results (cf. Lemma 11.13.2-Lemma 11.13.6 below),
in Propositions 11.13.7-11.13.10 we study basic properties of the scale of homoge-
neous Morrey-based Sobolev spaces introduced in Definition 11.13.1. The first such
preliminary result is formulated next.

Lemma 11.13.2 Let Σ ⊆ Rn be a closed upper Ahlfors regular set and abbreviate
σ := H n−1 Σ. Assume f is a complex-valued σ-measurable function on Σ with the
property that there exists some non-negative function g ∈ M p,λ (Σ, σ) with p ∈ [1, ∞)
and λ ∈ (0, n − 1) along with some σ-measurable set N ⊆ Σ satisfying σ(N) = 0
such that
 
| f (x) − f (y)| ≤ |x − y| · g(x) + g(y) for all x, y ∈ Σ \ N. (11.13.5)

Then
 
f ∈ L 1 Σ, 1+σ(x)
p
|x | n ∩ Lloc (Σ, σ) (11.13.6)

and there exists some constant C = C(Σ, p) ∈ (0, ∞) with the property that for each
reference point x0 ∈ Σ and each radius r ∈ (0, ∞) one has
∫  
f (x) − fΔ(x0,r)  C
  n dσ(x) ≤ (n−1−λ)/p g M p, λ (Σ,σ), (11.13.7)
Σ r + |x − x0 | r
⨏
where Δ(x0, r) := Σ ∩ B(x0, r) and fΔ(x0,r) := Δ(x0,r)
f dσ.
p
Proof That f ∈ Lloc (Σ, σ) is clear from (11.13.5) and (6.2.7). To justify (11.13.7),
fix a reference point x0 ∈ Σ. Then for each r ∈ (0, ∞) we may rely on Hölder’s
inequality, (11.13.5), and (6.2.2) to estimate
850 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
∫ 1/p
1
| f (x) − fΔ(x0,r) | p dσ(x)
r Δ(x0,r)
∫  ⨏ p
1   1/p
=  f (x) − f (y) dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏ p 1/p
1
≤ | f (x) − f (y)| dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏ 1/p
1
≤ | f (x) − f (y)| p dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏
1  p 1/p
≤ |x − y| p · g(x) + g(y) dσ(y) dσ(x)
r Δ(x0,r) Δ(x0,r)
∫ ⨏
  1/p
≤4 g(x) p + g(y) p dσ(y) dσ(x)
Δ(x0,r) Δ(x0,r)

≤ Cr λ/p g M p, λ (Σ,σ), (11.13.8)

for some C ∈ (0, ∞) independent of r, f , and g. Hence,

∫ 1/p
1
C∗ := sup 1+λ/p | f − fΔ(x0,r) | p dσ
r >0 r Δ(x0,r)

≤ Cg M p, λ (Σ,σ) < +∞, (11.13.9)

so (11.13.7) is implied by (11.13.9) and (11.5.64) with α := 1 + λ/p. Finally, the

 
fact that the function f belongs to the space L 1 Σ, 1+σ(x)
|x | n is a direct consequence of
(11.13.7). 

Here is the second preliminary result alluded to earlier.

Lemma 11.13.3 Let Ω ⊆ Rn be an open set satisfying a local John condition and
whose boundary is Ahlfors regular. Abbreviate σ := H n−1 ∂Ω and fix some aperture
parameter κ ∈ (0, ∞).
Then there exists C ∈ (0, ∞) with the property that for each function u ∈ C 1 (Ω)
 κ−n.t. 
for which the nontangential trace u∂Ω (x) exists at σ-a.e. point x ∈ ∂Ω one has
  κ−n.t.   κ−n.t.    
 
 u ∂Ω (x) − u∂Ω (y) ≤ C|x − y| · g(x) + g(y)
(11.13.10)
for σ-a.e. points x, y ∈ ∂Ω,

where
⎧
⎨ Nκ (∇u) if ∂Ω is unbounded,
⎪
⎪
g :=  κ−n.t.  (11.13.11)
⎪  
⎪ u∂Ω  + Nκ (∇u) if ∂Ω is bounded.
⎩
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 851

As a consequence of (11.13.10)-(11.13.11), Lemma 11.13.2, [133, (8.2.28)], as

well as [133, Corollary 8.9.6], for each integrability exponent p ∈ [1, ∞) and each
parameter λ ∈ (0, n − 1) there exists some constant C = C(Ω, p, λ) ∈ (0, ∞) with the
property that for each reference point x0 ∈ ∂Ω and each radius r ∈ (0, ∞) one has
∫  
f (x) − fΔ(x0,r)  C
  n dσ(x) ≤ (n−1−λ)/p g M p, λ (∂Ω,σ),
∂Ω r + |x − x0 | r
κ−n.t.
where f := u∂Ω , the function g is as in (11.5.71), (11.13.12)
⨏
Δ(x0, r) := ∂Ω ∩ B(x0, r), and fΔ(x0,r) := f dσ.
Δ(x0,r)

Also, (11.13.6) implies that, in a quantitative fashion,

if g from (11.13.11) belongs to M p,λ (∂Ω, σ),

for some p ∈ [1, ∞) and some λ ∈ (0, n − 1),
(11.13.13)
κ−n.t. σ(x)
then u∂Ω ∈ L 1 ∂Ω,
p
∩ Lloc (∂Ω, σ).
1 + |x| n
Proof The claim made in (11.13.10) is a direct consequence of (11.5.70), and
everything else follows from this. 

To state our next preliminary result, the reader is advised to recall (11.13.1)-
(11.13.2).

Lemma 11.13.4 Let Ω ⊆ Rn be an open set satisfying a two-sided local John

condition and whose boundary is an unbounded Ahlfors regular set. Abbreviate
σ := H n−1 ∂Ω and fix some integrability exponent p ∈ (1, ∞) along with some
parameter λ ∈ (0, n − 1).
. p,λthere exists some constant C = C(Ω, n, p, λ) ∈ (0, ∞) such that for each
Then
f ∈ M1 (∂Ω, σ) one may find some non-negative function g ∈ M p,λ (∂Ω, σ) such
that
 
| f (x) − f (y)| ≤ |x − y| · g(x) + g(y) for σ-a.e. x, y ∈ ∂Ω, (11.13.14)

as well as
g M p, λ (∂Ω,σ) ≤ C f  M. p, λ (∂Ω,σ) . (11.13.15)
1

In particular, as. seen from (11.13.14)-(11.13.15) and (11.13.1)-(11.13.2), for

p,λ
each function f ∈ M1 (∂Ω, σ) one has

 f  M. p, λ (∂Ω,σ) = 0 ⇐⇒ f is constant on ∂Ω. (11.13.16)

1

Also, as a corollary of (11.13.14)-(11.13.15) and (11.13.6)-(11.13.7) it follows

that for each reference point x0 ∈ ∂Ω and each. radius r ∈ (0, ∞) there exists some
p,λ
constant Cx0,r ∈ (0, ∞) such that for each f ∈ M1 (∂Ω, σ) one has
852 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
 
 f − fΔ(x ,r)  
0 σ(x)  ≤ Cx0,r ·  f  M. p, λ (∂Ω,σ) (11.13.17)
1L ∂Ω, 1+ |x | n 1

and
∫
| f (x) − fΔ(x0,r) |
dσ(x) ≤ Cx0,r ·  f  M. p, λ (∂Ω,σ), (11.13.18)
∂Ω 1 + |x| n 1

⨏
where Δ(x0, r) := ∂Ω ∩ B(x0, r) and fΔ(x0,r) := Δ(x ,r) f dσ.
0

. p,λ as a corollary of (11.13.14)-(11.13.15) and (11.13.8), for each function

Finally,
f ∈ M1 (∂Ω, σ), each reference point x0 ∈ ∂Ω and each radius r ∈ (0, ∞) one has
∫ 1/p
| f − fΔ(x0,r) | p dσ ≤ Cr 1+λ/p  f  M. p, λ (∂Ω,σ), (11.13.19)
Δ(x0,r) 1

for some C ∈ (0, ∞) independent of x0 , r, and f .

Proof To get started, recall that Kmod is the modified principal-value harmonic double
layer operator associated with Ω as in (11.4.27). Given any function

f ∈ L 1 ∂Ω, 1+σ(x)
p
|x | n ∩ Lloc (∂Ω, σ) satisfying
(11.13.20)
∂τ j k f ∈ M p,λ (∂Ω, σ) for all j, k ∈ {1, . . . , n},

the idea is to decompose

 
f = f+ − f− where f± := ± 12 I + Kmod f , (11.13.21)

and treat f+, f− separately. To deal with f+ , define u+ := Dmod f in Ω, where Dmod is the
modified boundary-to-domain harmonic double layer operator associated with Ω±
as in (11.4.26). Then from (11.4.26)-(11.4.36) plus mapping properties of singular
integral operators on Morrey spaces from [135, §2.6] we learn that u+ : Ω+ → C is
a well-defined (harmonic) function which belongs to C ∞ (Ω), and for any aperture
parameter κ > 0 satisfies
κ−n.t.  
u+ ∂Ω = 12 I + Kmod f = f+ at σ-a.e. point on ∂Ω,
Nκ (∇u+ ) belongs to the space M p,λ (∂Ω, σ), and
(11.13.22)
  
n
 
Nκ (∇u+ ) p, λ ≤ C ∂τ f  p, λ ,
M (∂Ω,σ) jk M (∂Ω,σ)
j,k=1

for some constant C = C(Ω, n, κ, p, λ) ∈ (0, ∞). From Lemma 11.13.3 we also know
that
 
| f+ (x) − f+ (y)| ≤ |x − y| · g+ (x) + g+ (y) for σ-a.e. x, y ∈ ∂Ω,
(11.13.23)
where g+ := C · Nκ (∇u+ ) ∈ M p,λ (∂Ω, σ).
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 853

Reasoning in a similar fashion for f− (now working in Ω− := Rn \ Ω, which

continues to be an open set satisfying a two-sided local John condition and whose
boundary ∂Ω− = ∂Ω is an unbounded Ahlfors regular set) we also conclude that
u− : Ω− → C is a well-defined (harmonic) function which belongs to C ∞ (Ω), and
for any aperture parameter κ > 0 there exists some C = C(Ω, n, κ, p, λ) ∈ (0, ∞) such
that
κ−n.t.  
u− ∂Ω = − 12 I + Kmod f = f− at σ-a.e. point on ∂Ω,
Nκ (∇u− ) belongs to the space M p,λ (∂Ω, σ), and
(11.13.24)
  
n
 
Nκ (∇u− ) p, λ ≤ C ∂τ f  p, λ ,
M (∂Ω,σ) jk M (∂Ω,σ)
j,k=1

and
 
| f− (x) − f− (y)| ≤ |x − y| · g− (x) + g− (y) for σ-a.e. x, y ∈ ∂Ω,
(11.13.25)
where g− := C · Nκ (∇u− ) ∈ M p,λ (∂Ω, σ).

If we now set g := g+ + g− then all desired conclusions follow on account of

(11.13.22), (11.13.23), (11.13.24), and (11.13.25). 

The following trace result plays a role in the proof of Lemma 11.13.6, stated a
little later.

Lemma 11.13.5 Let Ω ⊆ Rn be an NTA domain such that ∂Ω is a compact Ahlfors

regular set. Abbreviate σ := H n−1 ∂Ω, and fix an aperture parameter κ ∈ (0, ∞)
along with an integrability exponent p ∈ (1, ∞) and a number λ ∈ (0, n − 1). Let
u : Ω → C be a function satisfying

u ∈ C 1 (Ω) and Nκε (∇u) ∈ M p,λ (∂Ω, σ) (11.13.26)

for some truncation parameter ε ∈ (0, ∞). 

Then there exists a threshold ρ ∈ 0, 2 diam(∂Ω) with the property that
ρ
Nκ u ∈ M p,λ (∂Ω, σ). (11.13.27)

Proof In the present setting, the space M p,λ (∂Ω, σ) is contained in L p (∂Ω, σ) (see
the second inclusion in (6.2.7)). As such, [133,
 (8.4.109) in Proposition 8.4.9] enures
that there exists some small threshold ρ ∈ 0, 2 diam(∂Ω) such that
ρ
Nκ u ∈ L p (∂Ω, σ). (11.13.28)

For some small number r ∈ (0, ∞) to be determined below, this permits us to estimate
(also using the Ahlfors regularity of ∂Ω)
854 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
! ⨏ "
n−1−λ  ρ p 1 −λ  
sup R p Nκ u dσ p
≤ Cr p N ρ u 
κ .
L p (∂Ω,σ)
x ∈∂Ω and ∂Ω∩B(x,R)
r ≤R<2 diam(∂Ω)
(11.13.29)
ρ
As such, in order to conclude that Nκ u ∈ M p,λ (∂Ω, σ) there remains to prove that
! "
n−1−λ ⨏  ρ p 1

sup R p Nκ u dσ p
< +∞. (11.13.30)
x ∈∂Ω and ∂Ω∩B(x,R)
0<R<r

To justify (11.13.30) observe that, according to [133, Proposition 8.4.9] we may

pick another (possibly larger) aperture parameter κ ∈ (0, ∞) and choose some con-
stant C ∈ (0, ∞)
 with the property
 that, after possibly decreasing the value of the
threshold ρ ∈ 0, 2 diam(∂Ω) , for every point xo ∈ ∂Ω we can find a compact subset
Kρ,xo of Ω such that
 ρ 
Nκ u (x) ≤ C ρ · Nκε (∇u)(x) + sup |u|, ∀x ∈ B(xo, ρ) ∩ ∂Ω. (11.13.31)
K ρ, x o

Since ∂Ω is presently assumed  to be a compact set, there exists a finite family of points
{x j }1≤ j ≤ N in ∂Ω such that B(x j , ρ) ∩ ∂Ω 1≤ j ≤ N covers ∂Ω. Then the Lebesgue
Number Theorem ensures the existence of some r ∈ (0, ∞) with the property that for
each xo ∈ ∂Ω there exists jo ∈ {1, . . . , N } such that
 
B(xo, r) ∩ ∂Ω ⊆ B x jo , ρ ∩ ∂Ω. (11.13.32)

This is the choice of r we make in (11.13.29).

Going further, pick xo ∈ ∂Ω arbitrary and consider jo ∈ {1, . . . , N } such that
(11.13.32) holds. Then for each scale R ∈ (0, r) and σ-a.e. point x ∈ B(xo, r) ∩ ∂Ω
we may estimate
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 855
⨏
n−1−λ   ρ  p 1

R p  Nκ u (x) dσ(x) p

∂Ω∩B(x o ,R)

n−1−λ ⨏ 1
p
≤ Cρ · R p Nκε (∇u)(x) p dσ(x)
∂Ω∩B(x o ,R)

n−1−λ
+R p · sup |u|
K ρ, x j o

  n−1−λ
≤ C ρ · Nκε (∇u) M p, λ (∂Ω,σ) + r p · sup |u|
K ρ, x j o

  n−1−λ
≤ C ρ · Nκε (∇u) M p, λ (∂Ω,σ) + r p · max sup |u| .
1≤ j ≤ N K ρ, x j
(11.13.33)

Above, the first estimate uses the fact that B(xo, R) ∩ ∂Ω ⊆ B(xo, r) ∩ ∂Ω together
with (11.13.32) and (11.13.31). The second estimate in (11.13.33) is implied by the
definition of the Morrey norm (cf. (6.2.2)), the fact that (n − 1 − λ)/p > 0, and that
R ∈ (0, r). The third estimate in (11.13.33) is provided by [133, Corollary 8.4.5] (see
also item (5) of [133, Corollary 8.4.8]).
Finally, since the very last expression in (11.13.33) is a finite number, independent
of xo and R, (thanks to the memberships in the middle lines of (11.13.22) and
(11.13.24)), we conclude from (11.13.33) that (11.13.30) holds. This ultimately
establishes (11.13.27). 
Here is our last preliminary result, mentioned earlier.

Lemma 11.13.6 Suppose Ω ⊆ Rn (where n ∈ N, n ≥ 2) is a two-sided NTA domain

whose boundary is Ahlfors regular, and abbreviate σ := H n−1 ∂Ω. Also, fix some
integrability exponent p ∈ (1, ∞) together with some parameter λ ∈ (0, n − 1). Then
. p,λ p,λ
M1 (∂Ω, σ) = M1 (∂Ω, σ) if ∂Ω is bounded. (11.13.34)

Proof The right-to-left inclusion has been noted earlier, in (11.13.4) (this does
not require ∂Ω to be compact). To prove the. left-to-right inclusion, assume ∂Ω is
p,λ
bounded and pick an arbitrary function f ∈ M1 (∂Ω, σ). The crux of the matter is
establishing that f belongs to M p,λ (∂Ω, σ). To this end, we shall employ notation and
results from the proof of Lemma 11.13.4 in relation to the function f . Specifically,
given any ρ > 0, for σ-a.e. x ∈ ∂Ω we may estimate
 κ−n.t.   κ−n.t. 
   
| f (x)| ≤  u+ ∂Ω (x) +  u− ∂Ω (x)
 ρ   ρ 
≤ Nκ u+ (x) + Nκ u− (x), (11.13.35)
856 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

thanks to (11.13.21), the first lines of (11.13.22) and (11.13.24), and [133, (8.9.8)]. In
view of (11.13.35), (11.13.22), (11.13.24), Lemma 11.13.5, and (6.2.3) we conclude
that f ∈ M p,λ (∂Ω, σ), from which the desired conclusion follows. 

Next, we note the following remarkable trace result.

Proposition 11.13.7 Let Ω ⊆ Rn be an NTA domain such that ∂Ω is an Ahlfors

regular set. Abbreviate σ := H n−1 ∂Ω, and fix an aperture parameter κ ∈ (0, ∞).
Also, select an integrability exponent p ∈ (1, ∞) and a number λ ∈ (0, n − 1). In this
setting, consider a function u : Ω → C satisfying

u ∈ C 1 (Ω) and Nκ (∇u) ∈ M p,λ (∂Ω, σ). (11.13.36)

Then
κ−n.t. . p,λ
u∂Ω exists σ-a.e. on ∂Ω, belongs to M1 (∂Ω, σ),
 κ−n.t.    (11.13.37)
and u∂Ω  M. p, λ (∂Ω,σ) ≤ C Nκ (∇u) M p, λ (∂Ω,σ)
1

for some constant C ∈ (0, ∞) independent of u.

Proof The present hypotheses on Ω ensure (cf. [133, (5.11.28)]) that Ω is an open
set satisfying a two-sided corkscrew condition, as well as a local John condition.
Additionally, [133, (5.2.4)] implies that ∂∗ Ω = ∂Ω, so in particular Ω is an Ahlfors
regular domain. From the current assumptions and [133, Proposition 8.9.22] we see
(bearing in mind the second inclusion in (6.2.7)) that
 κ−n.t. 
the nontangential trace u∂Ω (x) exists at σ-a.e. x ∈ ∂Ω,
κ−n.t. (11.13.38)
and the function u∂Ω belongs to the space Lloc (∂Ω, σ).
p

From [133, Proposition 8.4.9] and the second embedding in (6.2.7) we also see that
there exists some ε ∈ (0, ∞) such that

Nκε u ∈ Lloc (∂Ω, σ).

p
(11.13.39)

Granted these, we may invoke Corollary 11.3.9 to conclude that

κ−n.t.
∂τ j k u∂Ω ∈ M p,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.13.40)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) M p, λ (∂Ω,σ)
M p, λ (∂Ω,σ)
j,k=1

for some constant C ∈ (0, ∞) independent of u. We next claim that

κ−n.t. σ(x)
u∂Ω ∈ L 1 ∂Ω, . (11.13.41)
1 + |x| n
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 857

On the one hand, when ∂Ω is bounded, the space in (11.13.41) reduces to L 1 (∂Ω, σ),
a scenario in which the claim is a simple consequence of (11.13.38). On the other
hand, if ∂Ω is unbounded, we may invoke Lemma 11.13.3 (also keeping in mind
(11.13.36)) and conclude that (11.13.41) holds in this case as well. This establishes
(11.13.41).
Finally, from (11.13.40)-(11.13.41) and Definition 11.13.1 we then conclude that
the claims made in (11.13.37) are indeed true. 

We augment Proposition 11.13.7 by including the following boundary trace result,

of a similar flavor, valid for a larger class of domains but under additional assumptions
on the functions involved.

Proposition 11.13.8 Assume Ω ⊆ Rn is an Ahlfors regular domain satisfying a

local John condition, and denote σ := H n−1 ∂Ω. Select an aperture parameter
κ ∈ (0, ∞), a truncation parameter ε ∈ (0, ∞), an integrability exponent p ∈ (1, ∞),
and a number λ ∈ (0, n − 1). Suppose some complex-valued function u has been
given, satisfying:

Nκε u ∈ Lloc (∂Ω, σ), Nκ (∇u) ∈ M p,λ (∂Ω, σ),

1,1 p
u ∈ Wloc (Ω),
κ−n.t. (11.13.42)
and u∂Ω exists at σ-a.e. point on ∂Ω.

Then there exists a constant C ∈ (0, ∞) independent of u such that

κ−n.t. . p,λ
u∂Ω belongs to the space M1 (∂Ω, σ) and
 κ−n.t.    (11.13.43)
u  . p, λ ≤ C Nκ (∇u) M p, λ (∂Ω,σ) .
∂Ω M1
(∂Ω,σ)

Proof Corollary 11.3.9 guarantees that there exists a constant C ∈ (0, ∞) indepen-
dent of u such that
κ−n.t.
∂τ j k u∂Ω ∈ M p,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.13.44)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) M p, λ (∂Ω,σ)
M p, λ (∂Ω,σ)
j,k=1

We also claim that

κ−n.t. σ(x)
u∂Ω ∈ L 1 ∂Ω,
p
∩ Lloc (∂Ω, σ). (11.13.45)
1 + |x| n
When ∂Ω is bounded, this follows from (11.13.42), [133, (8.9.8)], and [133, Corol-
lary 8.9.6]. When ∂Ω is unbounded, we may rely on (11.13.42), the current geometric
assumptions, and Lemma 11.5.8 to reach the same conclusion. Together, (11.13.44)-
(11.13.45) and Definition 11.13.1 then justify the claims made in (11.13.43). 
858 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

Our next proposition contains several equivalent characterizations of the mem-

bership to the homogeneous Morrey-based Sobolev space we have introduced in
Definition 11.13.1.

Proposition 11.13.9 Suppose Ω ⊆ Rn is a two-sided NTA domain such that ∂Ω

is an unbounded Ahlfors regular set, and abbreviate σ := H n−1 ∂Ω. Fix some
reference point x0 ∈ ∂Ω, along with some integrability exponent p ∈ (1, ∞) and
some parameter λ ∈ (0, n − 1). Finally, assume that
p
f is a function belonging to Lloc (∂Ω, σ) with
(11.13.46)
∂τ j k f ∈ M p,λ (∂Ω, σ) for all j, k ∈ {1, . . . , n}.

Then the following statements are equivalent:

 
(i) The function f belongs to the space L 1 ∂Ω, 1+σ(x)
|x | n .
(ii) There exists a constant C = C(Ω, p, λ, x0 ) ∈ (0, ∞) which is independent of
⨏ that if for each scale r ∈ (0, ∞) one defines
the function f , with the property
Δr := B(x0, r) ∩ ∂Ω and fΔr := Δ f dσ then
r

∫ 
1   1/p n
 
sup  f − fΔ  p dσ ≤C ∂τ f  p, λ . (11.13.47)
r 1+λ/p r jk M (∂Ω,σ)
r >0 Δr j,k=1

(iii) For each r ∈ (0, ∞) there exists a constant Cr ∈ (0, ∞) which depends only on
Ω, p, λ, x0 , and r such that, with fΔr as before, one has
∫ 
| f (x) − fΔr |
n
 
dσ(x) ≤ Cr
∂τ f  p, λ . (11.13.48)
1 + |x| n jk M (∂Ω,σ)
∂Ω j,k=1

(iv) There exists a constant C = C(Ω, p, λ, x0 ) ∈ (0, ∞) independent of f , and some

constant c f ∈ C which is allowed to depend on f , such that


n
 
 f − cf  1  σ(x)  ≤C ∂τ f  p, λ . (11.13.49)
L ∂Ω, 1+ |x | n
jk M (∂Ω,σ)
j,k=1

. p,λ
(v) The function f belongs to the space M1 (∂Ω, σ).

Proof The equivalence (i) ⇔ (v) is clear from (11.13.1). The implication (v) ⇒ (ii) is
a consequence of (11.13.19), bearing in mind (11.13.2). The implication (v) ⇒ (iii)
follows from (11.13.18) while the implication (v) ⇒ (iv) is justified by (11.13.17)
again, keeping in mind (11.13.2). Next, the implication (ii) ⇒ (i) is seen from
Lemma 11.5.6 used with α := 1 + λ/p (which satisfies α < 1 + (n − 1)/p). Given that
the implications (iii) ⇒ (i) and (iv) ⇒ (i) are obvious (once one realizes that constant
 
functions on ∂Ω belong to the space L 1 ∂Ω, 1+σ(x) |x | n ; cf. [133, (7.2.5)]), the proof of
the proposition is complete. 
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 859

Moving on, suppose Ω ⊆ Rn is an Ahlfors regular domain, and abbreviate

σ := H n−1 ∂Ω. For future purposes, we find it useful to consider homogeneous
Morrey-based Sobolev spaces of order one, modulo constants, on ∂Ω, defined for
each integrability exponent p ∈ (1, ∞) and each parameter λ ∈ (0, n − 1) as
. p,λ &  . p,λ 
M1 (∂Ω, σ) ∼ := [ f ] : f ∈ M1 (∂Ω, σ) (11.13.50)

where [ f ] denotes
. p,λthe equivalence
& class, modulo constants, of the function f . We
agree to equip M1 (∂Ω, σ) ∼ with the semi-norm
. p,λ &  
M1 (∂Ω, σ) ∼  [ f ] −→ [ f ] M. p, λ (∂Ω,σ)/∼ :=  f  M. p, λ (∂Ω,σ) (11.13.51)
1 1


n
 
= ∂τ f  p, λ .
jk M (∂Ω,σ)
j,k=1

In the geometric setting described in the proposition below, the above semi-norm
turns out to be a genuine norm and the homogeneous Morrey-based Sobolev space
(11.13.50) is a Banach space.

Proposition 11.13.10 Suppose Ω ⊆ Rn is an open set satisfying a two-sided local

John condition and whose boundary is an unbounded Ahlfors regular set. Abbreviate
σ := H n−1 ∂Ω and pick some. integrability& exponent p ∈ (1, ∞) along with some
p,λ
parameter λ ∈ (0, n − 1). Let M1 (∂Ω, σ) ∼ denote the quotient space of classes
. p,λ
[·] of equivalence modulo constants of functions in M1 (∂Ω, σ), equipped with the
semi-norm (11.13.51). . p,λ & . p,λ &
Then (11.13.51) is a genuine norm on M1 (∂Ω, σ) ∼, and M1 (∂Ω, σ) ∼ is
a Banach space when equipped with the norm (11.13.51).

Proof
. p,λ That the& semi-norm (11.13.51) is in fact an actual a norm . p,λ on the &space
M1 (∂Ω, σ) ∼ is clear from (11.13.16). Next, to prove that M1 (∂Ω, σ) ∼ is
. p,λ
complete when equipped with the norm (11.13.51), let { fα }α∈N ⊆ M1 (∂Ω, σ) be
  . p,λ &
such that [ fα ] α∈N is a Cauchy sequence in the quotient space M1 (∂Ω, σ) ∼.
 
Then for each fixed j, k ∈ {1, . . . , n} it follows that ∂τ j k fα α∈N is a Cauchy sequence
in M p,λ (∂Ω, σ). Since the latter is complete, it follows that there exists
g jk ∈ M p,λ (∂Ω, σ) such that

∂τ j k fα −→ g jk in M p,λ (∂Ω, σ) as α → ∞. (11.13.52)

Fix some reference point x0 ∈ ∂Ω and, ⨏ for each r ∈ (0, ∞), define the surface ball
Δr := B(x0, r) ∩ ∂Ω. Also, set fα,r := Δ fα dσ for each r ∈ (0, ∞) and each α ∈ N.
r
From (11.13.17) (written for f := fα − fβ ) it follows that for each r ∈ (0, ∞) there
exists a constant Cr ∈ (0, ∞) which depends on Ω, p, and r such that that for each
α, β ∈ N we have
860 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
   
 fα − fα,r − fβ − fβ,r   σ(x) 
1 L ∂Ω, 1+ |x | n

n
 
≤ Cr ∂τ fα − ∂τ fβ  p, λ . (11.13.53)
jk jk M (∂Ω,σ)
j,k=1

In view of (11.13.52), this implies that for each fixed r ∈ (0, ∞) the sequence
   
fα − fα,r α∈N is Cauchy in the Banach space L 1 ∂Ω, 1+σ(x)|x | n . Hence, for each fixed
 
r ∈ (0, ∞) there exists a function hr ∈ L 1 ∂Ω, 1+σ(x)
|x | n such that

 
fα − fα,r −→ hr in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.13.54)

Next, the estimate recorded in (11.13.19) (written for f := fα − fβ ) implies that

there exists some constant C = C(Ω, p, λ, x0 ) ∈ (0, ∞) with the property that for each
fixed r ∈ (0, ∞) we have
∫
    1/p
 fα − fα,r − fβ − fβ,r  p dσ
Δr

n
 
≤ C · r 1+λ/p ∂τ fα − ∂τ fβ  p, λ .
jk jk M (∂Ω,σ)
j,k=1
(11.13.55)

 we also conclude that for each fixed scale r ∈ (0, ∞)

Again relyingon (11.13.52),
the sequence fα Δr − fα,r α∈N is Cauchy in the Banach space L p (Δr , σ). As such,

for each r ∈ (0, ∞) there exists some kr ∈ L p (Δr , σ)

 (11.13.56)
such that fα  − fα,r → kr in L p (Δr , σ) as α → ∞.
Δr

Since convergence in Lebesgue spaces implies, after eventually passing to a sub-

sequence, pointwise a.e. convergence, from (11.13.54) and (11.13.56) we see that,
in fact, 
hr Δr = kr ∈ L p (Δr , σ) for each r ∈ (0, ∞). (11.13.57)
From (11.13.54) we also see that for each fixed r1, r2 ∈ (0, ∞) we have
 
fα,r2 − fα,r1 −→ hr1 − hr2 in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.13.58)

This forces hr1 − hr2 to be a constant, which ultimately goes to show that actually
 
hr ∈ L 1 ∂Ω, 1+σ(x)
|x | n for each r ∈ (0, ∞). (11.13.59)

Henceforth, we agree to simply write h for hr with r = 1, and cα for fα,r with r = 1.
Then (11.13.59), (11.13.54), and (11.13.57) tell us that the function
 
h belongs to L 1 ∂Ω, 1+σ(x)
p
|x | n ∩ Lloc (∂Ω, σ) (11.13.60)
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 861

and the sequence {cα }α∈N ⊆ C is such that

 
fα − cα −→ h in L 1 ∂Ω, 1+σ(x)
|x | n as α → ∞. (11.13.61)

For each j, k ∈ {1, . . . , n} and each test function ϕ ∈ Cc∞ (Rn ) we may then write
∫ ∫
h(∂τ j k ϕ) dσ = lim ( fα − cα )(∂τ j k ϕ) dσ
∂Ω α→∞ ∂Ω
∫
= − lim ∂τ j k ( fα − cα )ϕ dσ
α→∞ ∂Ω
∫
= − lim (∂τ j k fα )ϕ dσ
α→∞ ∂Ω
∫
= g jk ϕ dσ, (11.13.62)
∂Ω

thanks to (11.13.61), (11.1.18), and (11.13.52). From this and (11.1.12)-(11.1.17)

we then conclude that

∂τ j k h = g jk ∈ M p,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}. (11.13.63)

. p,λ
Collectively, (11.13.60), (11.13.63), and (11.13.1) prove that h ∈ M1 (∂Ω, σ). Fi-
nally,
 from (11.13.52), (11.13.63), and (11.13.51) we conclude that. p,λthe sequence
&
[ fα ] α∈N converges to [h], the class of h, in the quotient space M1 (∂Ω, σ) ∼.
. p,λ &
This proves that M1 (∂Ω, σ) ∼ is complete when equipped with the norm
(11.13.51). 

On a related topic, we now introduce the scale of block-based homogeneous

Sobolev spaces of order one on the boundary of a given Ahlfors regular domain.

Definition 11.13.11 Assume Ω ⊆ Rn (where n ≥ 2) is an Ahlfors regular domain

and abbreviate σ := H n−1 ∂Ω. Also, fix an integrability exponent q ∈ (1, ∞) along
with a parameter λ ∈ (0, n − 1) then introduce

q(n − 1)
qλ := ∈ (1, q). (11.13.64)
n − 1 + λ(q − 1)
In this context, define the block-based homogeneous Sobolev space of order one
on ∂Ω as
!
. q,λ σ(x) q
B1 (∂Ω, σ) := f ∈ L 1 ∂Ω, ∩ Llocλ (∂Ω, σ) : ∂τ j k f ∈ B q,λ (∂Ω, σ)
1 + |x| n
"
for all j, k ∈ {1, . . . , n}

(11.13.65)
862 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

and equip this space with the semi-norm

. q,λ 
n
B1 (∂Ω, σ)  f →  f  B. q, λ (∂Ω,σ) := ∂τ j k f  B q, λ (∂Ω,σ) . (11.13.66)
1
j,k=1

In the context
. q,λof Definition 11.13.11 it follows that all constant functions on
∂Ω belong to B1 (∂Ω, σ) and their respective semi-norms vanish. It is also appar-
ent from (11.7.20), (6.2.71), and Definition 11.13.11 that we have the continuous
embeddings

q,λ . q,λ σ(x)

B1 (∂Ω, σ) → B1 (∂Ω, σ) ∩ L 1 ∂Ω, , (11.13.67)
1 + |x| n−1
and
. q,λ .q
B1 (∂Ω, σ) → L1 λ (∂Ω, σ). (11.13.68)

Going further, assume Ω ⊆ Rn is an Ahlfors regular domain, and abbreviate

σ := H n−1 ∂Ω. For each integrability exponent q ∈ (1, ∞) and each parameter
λ ∈ (0, n − 1) define the homogeneous block-based Sobolev spaces of order one,
modulo constants, on ∂Ω, by setting
. q,λ &  . q,λ 
B1 (∂Ω, σ) ∼ := [ f ] : f ∈ B1 (∂Ω, σ) (11.13.69)

where [ f ] denotes the equivalence class, modulo constants, of the function f . We

. q,λ &
agree to equip B1 (∂Ω, σ) ∼ with the semi-norm
. q,λ &  
B1 (∂Ω, σ) ∼  [ f ] −→ [ f ] B. q, λ (∂Ω,σ)/∼ :=  f  B. q, λ (∂Ω,σ) (11.13.70)
1 1


n
 
= ∂τ f  q, λ .
jk B (∂Ω,σ)
j,k=1

In a suitable geometric setting, the above semi-norm becomes a genuine norm and
the homogeneous block-based Sobolev space modulo constants (11.13.69) becomes
a Banach space. Here is a formal statement to this effect:

Proposition 11.13.12 Assume Ω ⊆ Rn is an open set satisfying a two-sided local

John condition and whose boundary is an unbounded Ahlfors regular set. Abbreviate
σ := H n−1 ∂Ω and pick some integrability
. q,λ exponent
& q ∈ (1, ∞) along with some
parameter λ ∈ (0, n − 1). Denote by B1 (∂Ω, σ) ∼ the quotient space of classes
. q,λ
[·] of equivalence modulo constants of functions in B1 (∂Ω, σ), equipped with the
semi-norm (11.13.70). . q,λ & . q,λ &
Then (11.13.70) is a genuine norm on B1 (∂Ω, σ) ∼, and B1 (∂Ω, σ) ∼ is a
Banach space when equipped with the norm (11.13.70).
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 863
. q,λ &
Proof That the semi-norm (11.13.70) is in fact a norm on the space B1 (∂Ω, σ) ∼
. q,λ &
follows readily from (11.5.102) and (11.13.68). To prove that B1 (∂Ω, σ) ∼ is
. q,λ
complete when equipped with the norm (11.13.70), let { fα }α∈N ⊆ B1 (∂Ω, σ) be
  . q,λ &
such that [ fα ] α∈N is a Cauchy sequence in the quotient space B1 (∂Ω, σ) ∼.
Then,
 on the one hand, for each fixed j, k ∈ {1, . . . , n} it follows that the sequence
∂τ j k fα α∈N is Cauchy, hence convergent, in the Banach space B q,λ (∂Ω, σ). Hence,
there exists g jk ∈ B q,λ (∂Ω, σ) such that

∂τ j k fα −→ g jk in B q,λ (∂Ω, σ) as α → ∞. (11.13.71)

 
On the other hand, (11.13.68) implies that [ fα ] α∈N is also a Cauchy sequence in
.q &
the quotient space L1 λ (∂Ω, σ) ∼. Granted this, the same type of argument as in the
proofs of Proposition 11.5.14 and Proposition 11.13.10 then shows that there exists
a function
σ(x) q
h ∈ L 1 ∂Ω, ∩ Llocλ (∂Ω, σ) (11.13.72)
1 + |x| n
such that
∫ ∫
h(∂τ j k ϕ) dσ = g jk ϕ dσ, (11.13.73)
∂Ω ∂Ω

for each j, k ∈ {1, . . . , n} and each test function ϕ ∈ Cc∞ (Rn ). In turn, from (11.13.73)
and (11.1.12)-(11.1.17) we then conclude that

∂τ j k h = g jk ∈ B q,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}. (11.13.74)

. q,λ
Together, (11.13.72), (11.13.74), and (11.13.65) prove that h ∈ B1 (∂Ω, σ). Finally,
 
from (11.13.71), (11.13.74), and (11.13.70) we conclude that the sequence [ fα ] α∈N
. q,λ &
converges to [h], the class of h, in the quotient space B1 (∂Ω, σ) ∼. This ultimately
. q,λ &
shows that the space B1 (∂Ω, σ) ∼ is complete when equipped with the norm
(11.13.70). 
The following trace result is of significance.
Proposition 11.13.13 Let Ω ⊆ Rn be an NTA domain such that ∂Ω is an Ahlfors
regular set. Abbreviate σ := H n−1 ∂Ω, and fix an aperture parameter κ ∈ (0, ∞).
Also, select an integrability exponent q ∈ (1, ∞) and a parameter λ ∈ (0, n − 1). In
this setting, consider a function u : Ω → C satisfying

u ∈ C 1 (Ω) and Nκ (∇u) ∈ B q,λ (∂Ω, σ). (11.13.75)

Then
κ−n.t. . q,λ
u∂Ω exists σ-a.e. on ∂Ω, belongs to B1 (∂Ω, σ),
 κ−n.t.    (11.13.76)
and u∂Ω  B. q, λ (∂Ω,σ) ≤ C Nκ (∇u) B q, λ (∂Ω,σ)
1
864 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter

for some constant C ∈ (0, ∞) independent of u.

Proof Since, with qλ as in (11.13.64), we have

Nκ (∇u) ∈ B q,λ (∂Ω, σ) → L qλ (∂Ω, σ), (11.13.77)

Proposition 11.5.12 applies and gives that

κ−n.t.
u∂Ω exists σ-a.e. on ∂Ω and, as a function, belongs to
.q (11.13.78)
L1 λ (∂Ω, σ) ∩ Llocλ (∂Ω, σ) → L 1 ∂Ω, 1+σ(x)
q q
|x | n ∩ Lloc (∂Ω, σ),
λ

with the last inclusion originating in (11.5.55). Based on assumptions, (6.2.71),

(11.13.77), and [133, Proposition 8.4.9] we also conclude that there exists some
ε > 0 such that Nκε u ∈ Llocλ (∂Ω, σ). Granted this, we may invoke Corollary 11.3.10
q

to conclude that
κ−n.t.
∂τ j k u∂Ω ∈ B q,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.13.79)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) B q, λ (∂Ω,σ)
B q, λ (∂Ω,σ)
j,k=1

for some constant C ∈ (0, ∞) independent of u. From (11.13.78)-(11.13.79) and

Definition 11.13.11 we then conclude that the claims made in (11.13.76) are indeed
true. 

We next discuss a companion to Proposition 11.13.13 containing a similar bound-

ary trace result, which holds for a larger class of domains but under additional
assumptions on the functions involved.

Proposition 11.13.14 Let Ω ⊆ Rn is an Ahlfors regular domain satisfying a local

John condition, and set σ := H n−1 ∂Ω. Pick an aperture parameter κ ∈ (0, ∞), a
truncation parameter ε ∈ (0, ∞), an integrability exponent q ∈ (1, ∞), and a number
λ ∈ (0, n − 1). Define qλ as in (11.13.64). Finally, assume u is some complex-valued
function satisfying:

Nκε u ∈ Llocλ (∂Ω, σ), Nκ (∇u) ∈ B q,λ (∂Ω, σ),

1,1 q
u ∈ Wloc (Ω),
κ−n.t. (11.13.80)
and u∂Ω exists at σ-a.e. point on ∂Ω.

Then there exists some constant C ∈ (0, ∞) independent of u such that

κ−n.t. . q,λ
the function u∂Ω belongs to the space B1 (∂Ω, σ)
 κ−n.t.    (11.13.81)
and u∂Ω  B. q, λ (∂Ω,σ) ≤ C Nκ (∇u) B q, λ (∂Ω,σ) .
1
11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 865

Proof In view of the fact that Nκ (∇u) ∈ B q,λ (∂Ω, σ) → L qλ (∂Ω, σ), Proposi-
tion 11.5.12 together with (11.5.55) guarantee that
κ−n.t. . q σ(x)
u∂Ω ∈ L1 λ (∂Ω, σ)∩ Llocλ (∂Ω, σ) → L 1 ∂Ω,
q q
∩ Llocλ (∂Ω, σ). (11.13.82)
1 + |x| n
In the current setting we may also invoke Corollary 11.3.10 to conclude that
κ−n.t.
∂τ j k u∂Ω ∈ B q,λ (∂Ω, σ) for each j, k ∈ {1, . . . , n}
n 
  κ−n.t.    (11.13.83)
and ∂τ j k u∂Ω  ≤ C Nκ (∇u) B q, λ (∂Ω,σ)
B q, λ (∂Ω,σ)
j,k=1

for some constant C ∈ (0, ∞) independent of u. The claims made in (11.13.81) then
follow from (11.13.82)-(11.13.83) and Definition 11.13.11. 

Finally, we introduce the scale of vanishing Morrey-based homogeneous Sobolev

spaces of order one on the boundary of an arbitrary Ahlfors regular domain (compare
with Definition 11.13.1).

Definition 11.13.15 Suppose Ω ⊆ Rn be an Ahlfors regular domain and abbreviate

σ := H n−1 ∂Ω. Also, pick an integrability exponent p ∈ (1, ∞) along with a
parameter λ ∈ (0, n − 1). In this context, define the vanishing Morrey-based
homogeneous Sobolev space of order one on ∂Ω as
!
. p,λ σ(x) p
M1 (∂Ω, σ) := f ∈ L 1 ∂Ω, ∩ Lloc (∂Ω, σ) : ∂τ j k f ∈ M̊ p,λ (∂Ω, σ)
1 + |x| n
"
for all j, k ∈ {1, . . . , n}

(11.13.84)

and equip this space with the semi-norm

. p,λ 
n
M1 (∂Ω, σ)  f →  f M. p, λ (∂Ω,σ) := ∂τ j k f  M p, λ (∂Ω,σ) . (11.13.85)
1
j,k=1

As seen from of Definition 11.13.15, all constant functions on ∂Ω belong to

. p,λ
M1 (∂Ω, σ) and their respective semi-norms vanish. It is also apparent from
(11.13.84)-(11.13.85) and (11.13.1)-(11.13.2) that
. p,λ  . p,λ
M1 (∂Ω, σ) = f ∈ M1 (∂Ω, σ) : ∂τ j k f ∈ M̊ p,λ (∂Ω, σ)

for all j, k ∈ {1, . . . , n} (11.13.86)

and
866 11 Sobolev Spaces on the GMT Boundary of Sets of Locally Finite Perimeter
. p,λ . p,λ
M1 (∂Ω, σ) is a closed subspace of M1 (∂Ω, σ). (11.13.87)

Moreover, we have the continuous embedding

p,λ . p,λ σ(x)

M̊1 (∂Ω, σ) → M1 (∂Ω, σ) ∩ L 1 ∂Ω, . (11.13.88)
1 + |x| n−1
Much as in Proposition 11.13.10, if Ω also satisfies a two-sided local John condition
then
. p,λ &   
n
 
M1 (∂Ω, σ) ∼  [ f ] −→ [ f ]M. p, λ (∂Ω,σ)/∼ := ∂τ f  p, λ
jk M (∂Ω,σ)
1
j,k=1
. p,λ & . p,λ & (11.13.89)
is a genuine norm on M1 (∂Ω, σ) ∼, and M1 (∂Ω, σ) ∼ is a Banach space when
equipped with the norm (11.13.89).
Vanishing Morrey spaces are not dual spaces, so the proof of Proposition 11.13.7
(making use of Corollary 11.3.9 which, in turn, employs duality) no longer applies.
We nonetheless have the following useful trace result.
Proposition 11.13.16 Let Ω ⊆ Rn be an NTA domain such that ∂Ω is an Ahlfors
regular set. Abbreviate σ := H n−1 ∂Ω, and fix an aperture parameter κ ∈ (0, ∞).
Also, pick an integrability exponent p ∈ (1, ∞) and a parameter λ ∈ (0, n − 1). In
this setting, consider a function u ∈ C 1 (Ω) satisfying
κ−n.t.
Nκ (∇u) ∈ M̊ p,λ (∂Ω, σ) and (∇u)∂Ω exists σ-a.e. on ∂Ω. (11.13.90)

Then
κ−n.t. . p,λ
u∂Ω exists σ-a.e. on ∂Ω, belongs to M1 (∂Ω, σ),
 κ−n.t.    (11.13.91)
and u∂Ω M. p, λ (∂Ω,σ) ≤ C Nκ (∇u) M p, λ (∂Ω,σ)
1

for some constant C ∈ (0, ∞) independent of u.

Proof Thanks to (6.2.15), Proposition 11.13.7 applies and gives that the nontangen-
κ−n.t.
tial trace u∂Ω exists at σ-a.e. point on ∂Ω,
κ−n.t. . p,λ
the function u∂Ω belongs to M1 (∂Ω, σ), (11.13.92)

and  κ−n.t.   
u  . p, λ ≤ C Nκ (∇u) M p, λ (∂Ω,σ) (11.13.93)
∂Ω M1
(∂Ω,σ)

for some constant C ∈ (0, ∞) independent of u. To proceed, observe from (6.2.15)

p p
and (6.2.7) that M̊ p,λ (∂Ω, σ) → Lloc (∂Ω, σ), hence Nκ (∇u) ∈ Lloc (∂Ω, σ). Then
[133, Proposition 8.4.9] guarantees that there exists some ε ∈ (0, ∞) such that
Nκε u ∈ Lloc (∂Ω, σ). Granted these, we may invoke Proposition 11.3.2 and conclude
p

from (11.3.26) that for each j, k ∈ {1, . . . , n} we have

11.13 Morrey-Based, and Block-Based, Homogeneous Sobolev Spaces 867
κ−n.t. κ−n.t. κ−n.t.
∂τ j k u∂Ω = ν j (∂k u)∂Ω − νk (∂j u)∂Ω at σ-a.e. point on ∂Ω. (11.13.94)

In particular, from (11.13.94) and [133, (8.9.8)] we see that

 κ−n.t. 

∂τ j k u∂Ω  ≤ 2 Nκ (∇u) at σ-a.e. point on ∂Ω, (11.13.95)

for all j, k ∈ {1, . . . , n}. In concert with (6.2.18) and the membership in (11.13.90),
this implies that
κ−n.t.
∂τ j k u∂Ω ∈ M̊ p,λ (∂Ω, σ) for all j, k ∈ {1, . . . , n}. (11.13.96)

κ−n.t.
From (11.13.92), (11.13.96), and (11.13.86) we then conclude that u∂Ω belongs to
. p,λ
the space M1 (∂Ω, σ). Together with (11.13.93), this establishes (11.13.91). 

It is also useful to augment Proposition 11.13.16 with a similar boundary trace

result which is valid for a larger class domains, while placing stronger demands on
the functions involved.

Proposition 11.13.17 Suppose Ω ⊆ Rn is an Ahlfors regular domain satisfying a

local John condition, and abbreviate σ := H n−1 ∂Ω. Pick an aperture parameter
κ ∈ (0, ∞), a truncation parameter ε ∈ (0, ∞), an integrability exponent p ∈ (1, ∞),
and a number λ ∈ (0, n − 1). Also, assume u is some complex-valued function
satisfying

(Ω), Nκε u ∈ Lloc (∂Ω, σ), Nκ (∇u) ∈ M̊ p,λ (∂Ω, σ),

1,1 p
u ∈ Wloc
κ−n.t. κ−n.t. (11.13.97)
u∂Ω
and (∇u) ∂Ω
exist at σ-a.e. point on ∂Ω.

Then there exists some constant C ∈ (0, ∞) independent of u such that

κ−n.t. . p,λ
the function u∂Ω belongs to the space M1 (∂Ω, σ)
 κ−n.t.    (11.13.98)
and u∂Ω M. p, λ (∂Ω,σ) ≤ C Nκ (∇u) M p, λ (∂Ω,σ) .
1

Proof Since M̊ p,λ (∂Ω, σ) is a subspace of M p,λ (∂Ω, σ), we may invoke Propo-
sition 11.13.8 to conclude that (11.13.92)-(11.13.93) continue to hold in the cur-
rent setting. We may also call upon Proposition 11.3.2 and, reasoning much as in
(11.13.94)-(11.13.95), deduce that (11.13.96) is presently valid. Much as in the end-
game of the proof of Proposition 11.13.16, these properties ultimately permit us to
justify the claims made in (11.13.98). 
Appendix A
Terms and Notation used in Volume II

A
Ahlfors regular domain (cf. [133, Definition 5.9.15]):

a nonempty open subset Ω of Rn such that ∂Ω is an

(A.0.1)
Ahlfors regular set and H n−1 (∂Ω \ ∂∗ Ω) = 0

Aκ (∂Ω), the κ-accessible (from within Ω) subset of ∂Ω (cf. [133, (8.8.2)]):

 
Aκ (∂Ω) := x ∈ ∂Ω : x ∈ Γκ (x) (A.0.2)

[w] A p , the characteristic of the weight w on a space of homogeneous type (X, ρ, μ)

(cf. [133, §7.7]):
⨏ ⨏  p−1
[w] A p := sup w dμ w −1/(p−1) dμ (A.0.3)
B ρ -ball B B

Ap (X, ρ, μ), the Muckenhoupt Ap -class on a space of homogeneous type (X, ρ, μ)

(cf. [133, §7.7]):
 
Ap (X, ρ, μ) := w weight function : [w] A p < ∞ (A.0.4)

A∞ (X, ρ, μ), the Muckenhoupt A∞ -class on a space of homogeneous type (X, ρ, μ)

(cf. [133, §7.7]): 
A∞ (X, ρ, μ) := Ap (X, μ) (A.0.5)
1≤p<∞
 
V ⊥ := Λ ∈ X ∗ : Λ(x) = 0 for all x ∈ V , the annihilator of a subspace V of a
Banach space X (2.1.17)
 
⊥ W := x ∈ X : Λ(x) = 0 for all Λ ∈ W , the annihilator of a subspace W of X ∗ ,

where X is Banach (2.1.18)

© The Editor(s) (if applicable) and The Author(s), under exclusive license to 869
Springer Nature Switzerland AG 2022
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1
870 A Terms and Notation used in Volume II

T ∗ , the adjoint of the operator T ∈ L(X → Y ) defined by T ∗ : Y ∗ → X ∗ with

T ∗ (Λ) := Λ ◦ T for each Λ ∈ Y ∗ (1.1.20)
A p , the absolutely p-convex hull of a subset A of a vector space X, with p ∈ (0, 1],
defined as (see (7.8.3)):

M 
M
A p := λ j v j : M ∈ N, {v j }1≤ j ≤M ⊆ A, {λ j }1≤ j ≤M ⊆ C with |λ j | p ≤ 1
j=1 j=1

p,q
As (Ω), the Besov/Triebel-Lizorkin space in the open set Ω ⊆ Rn (with A := B
corresponding to Besov spaces, and with A := F corresponding to Triebel-Lizorkin
spaces), with 0 < p, q ≤ ∞ and s ∈ R, defined as (see (9.2.1)):
p,q  p,q 
As (Ω) := u ∈ D (Ω) : there exists U ∈ As (Rn ) such that U Ω = u

and equipped with the quasi-norm


p,q
u Ap, q
s (Ω)
:= inf U p, q
As (R n ) : U ∈ As (Rn ), U Ω = u

 
Ås (Ω), the closure of Cc∞ (Ω) in As (Ω), · Ap,
p,q p,q
q
s (Ω)
, with 0 < p, q ≤ ∞ and
s ∈ R (9.2.2)
p,q  p,q 
As,0 (Ω) := u ∈ As (Rn ) : supp u ⊆ Ω , with 0 < p, q ≤ ∞, s ∈ R, where
either A := F and p < ∞, or A := B, equipped with the quasi-norm given by
u Ap, q (Ω) := u Ap, q n (9.2.3)
s (R )
s,0
p,q  p,q 
As,z (Ω) := u ∈ D (Ω) : there exists w ∈ As,0 (Ω), w Ω = u , with 0 < p, q ≤ ∞
and s ∈ R, where either A := F and p < ∞, or A := B, equipped  with the quasi-norm
p,q
u Ap, q
s, z (Ω) := inf w A
p, q
s (R n) : w ∈ A
s,0 (Ω), w Ω
= u (9.2.4)
∫ q 1
A q,κ u (x) := (∇u)(y) |x − y| q−n dy , the L q -based area-function with
q

Γκ (x)
q ∈ (0, ∞) (9.2.206)
B
Bρ (x, r), the ρ-ball with center at x ∈ X and radius r > 0 in the quasi-metric space
(X, ρ) (cf. [133, §7.1]):

Bρ (x, r) := {y ∈ X : ρ(x, y) < r } (A.0.6)

 
Bn−1 (x , r) := y ∈ Rn−1 : |y − x | < r , the (n − 1)-dimensional (open) ball in
Rn−1 centered at x ∈ Rn−1 and of radius r ∈ (0, ∞)
BMO1 , the BMO-based Sobolev spaces of order one (locally integrable functions
with distributional first-order partial derivatives in BMO)
f ∗ (Δ), the local BMO norm of the function f on the surface ball Δ (cf. [133,
§7.4]):
A Terms and Notation used in Volume II 871
⨏
f ∗ (Δ) := sup | f − fΔ | dμ (A.0.7)
Δ ⊆Δ Δ

.
f BMO(X,μ) , the homogeneous BMO semi-norm of the function f in the context of
a space of homogeneous type (X, ρ, μ) (cf. [133, §7.4]):
⨏
.
f BMO(X,μ) := sup f − fBρ (x,r) dμ (A.0.8)
x ∈X, r >0 B ρ (x,r)

· BMO(X,μ) , the inhomogeneous BMO “norm” in the context of a space of homo-

geneous type (X, ρ, μ) (cf. [133, §7.4]):
⎧
⎪ f . if X is unbounded
⎪
⎨
⎪ BMO(X,μ)
f := ∫ (A.0.9)
BMO(X,μ)
⎪
⎪ . +
⎪ f f dμ if X is bounded
⎩ BMO(X,μ)
X

BMO X, μ , the space of functions of bounded mean oscillations for a space of

homogeneous type (X, ρ, μ) (cf. [133, §7.4]):

BMO X, μ := f ∈ Lloc 1
(X, μ) : f BMO(X,μ) < +∞ (A.0.10)


BMO(X, μ), the space BMO modulo constants for a space of homogeneous type
(X, ρ, μ) (cf. [133, (7.4.96)]):
  

BMO(X, μ) := BMO(X, μ) ∼ = [ f ] : f ∈ BMO(X, μ) (A.0.11)

∂E, the topological boundary of the set E

Borelτ (X), the Borelians of the topological space (X, τ)
BL(Σ), the space of all bounded Lipschitz functions defined on the set Σ
BV(O), the space of functions of bounded variation in the set O (cf. [133, (5.5.5)]):
 
BV(O) := f ∈ L 1 (O, L n ) : V( f ; O) < +∞ (A.0.12)

BVloc (O), the space of functions of locally bounded variation in the set O (cf. [133,
(5.5.6)]):

BVloc (O) := f ∈ Lloc1
(O, L n ) : V( f ; U) < +∞ for each open set (A.0.13)

U in Rn with U compact subset of O

[A; B] := [A, B] := AB − BA, the commutator of A and B

{ A; B} := AB + BA, the anti-commutator of A and B
872 A Terms and Notation used in Volume II
 
Bd(X → Y ) := T : X → Y : T linear mapping with T X→Y < +∞ , where
X, · X as well as Y, · Y are quasi-normed vector spaces (1.2.16)
Bd(X) := Bd(X → X), the space of bounded linear operators from the quasi-normed
vector space X into itself (1.2.18)
B(X → Y ), the space of linear and bounded operators from X to Y (1.1.9)
B q,λ (Σ, σ), the block space on the Ahlfors regular set Σ ⊆ Rn , defined for q ∈ (1, ∞)
and λ ∈ (0, n − 1) as the collection of f ∈ Lipc (Σ) for which there exist a sequence
∞
{λ j } j ∈N ∈ 1 (N) and a family {b j } j ∈N of B q,λ -blocks on Σ so that f = λj bj
j=1
with convergence in (Lipc (Σ)) , equipped with the norm f B q, λ (Σ,σ) defined as the

∞
infimum of |λ j | over all such possible writings of f (6.2.64), (6.2.65)
j=1
. p,q
Bs (Σ, σ), the homogeneous Besov space on the Ahlfors regular set Σ ⊆ Rn , defined
β,γ ∗
as the collection of functionals f ∈ G̊0 (Σ) for which
   q  1/q
f B. p, q (Σ,σ) := 2ks Ek f L p (Σ,σ) <∞
s
k ∈Z

(with natural alterations when p = ∞, or q = ∞) (7.1.13), (7.1.14)

p,q
Bs (Σ, σ), the inhomogeneous Besov space on the Ahlfors regular set Σ ⊆ Rn ,
β,γ ∗
defined as the collection of functionals f ∈ G0 (Σ) with a finite inhomogeneous
Besov quasi-norm f Bsp, q (Σ,σ) (of Littlewood-Paley type) (7.1.31), (7.1.32)
bmo(Σ, σ), the local space of functions of bounded mean oscillations on the Ahlfors
regular set Σ ⊆ Rn , defined as the collection of all functions f belonging to Lloc
1 (Σ, σ)

and with the property that (see (7.1.53))

⨏ ⨏
f bmo(Σ,σ) := sup f− f dσ dσ
x ∈Σ, 0<r <rΣ Δ(x,r) Δ(x,r)
⨏
+ sup | f | dσ < ∞
x ∈Σ, r ≥rΣ Δ(x,r)

bp,q (Σ), the discrete Besov space on the Ahlfors regular set Σ ⊆ Rn , i.e., the
collection of all numerical sequences λ = {λQ }Q ∈D∗ (Σ) with
    N
(k,τ)  q/p  1/q
λ b p, q (Σ) := |λQ k,ν | p
<∞
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

(with natural modifications when p = ∞, or q = ∞) (7.2.24), Definition 7.2.5

p,q
bs (Σ), the discrete Besov space on the Ahlfors regular set Σ ⊆ Rn , i.e., the
collection of all sequences λ = {λQ }Q ∈D∗ (Σ) with
A Terms and Notation used in Volume II 873
    N
(k,τ)   1/q
k,ν p1 − 12 p q/p
λ p, q
b s (Σ) := 2 ks
σ(Qτ ) |λQ k,ν | <∞
τ
k ∈Z τ ∈Ik ν=1
k ≥κΣ

(with natural modifications when p = ∞, or q = ∞) (7.2.28), Definition 7.2.6

p,q
Bs (Rn ), the (inhomogeneous) Besov space in Rn with 0 < p, q ≤ ∞ and s ∈ R,
equipped with the quasi-norm · Bsp, q (Rn ) (9.1.8), (9.1.9)
p,q  p,q 
Bs (Ω) := u ∈ D (Ω) : there exists U ∈ Bs (Rn ) such that U Ω = u , the Besov
space in the open set Ω ⊆ Rn with 0 < p, q ≤ ∞ and s ∈ R, equipped with the
quasi-norm (see (9.2.1))

p,q
u Bsp, q (Ω) := inf U Bsp, q (Rn ) : U ∈ Bs (Rn ), U Ω = u

 
B̊s (Ω), the closure of Cc∞ (Ω) in Bs (Ω), · Bsp, q (Ω) , for 0 < p, q ≤ ∞ and s ∈ R
p,q p,q

(9.2.2)
p,q  p,q 
Bs,0 (Ω) := u ∈ Bs (Rn ) : supp u ⊆ Ω , with 0 < p, q ≤ ∞ and s ∈ R, equipped
with the quasi-norm u B p, q (Ω) := u Bsp, q (Rn ) (9.2.3)
s,0
p,q  p,q 
Bs,z (Ω) := u ∈ D (Ω) : there exists w ∈ Bs,0 (Ω), w Ω = u , with 0 < p, q ≤ ∞
and s ∈ R, equipped with the quasi-norm (see (9.2.4))
 p,q 
u Bs,p,zq (Ω) := inf w Bsp, q (Rn ) : w ∈ Bs,0 (Ω), w Ω = u

 n−1 ,1 ∗
BMO−1 (∂Ω, σ) := H1n−2 (∂Ω, σ) , the BMO-based negative Sobolev space of
order minus one on ∂Ω (11.10.50), Definition 11.10.9
q,λ
B1 (∂Ω, σ), the block-based Sobolev space of order one on ∂Ω (11.7.20)
 ∗
q,λ p,λ
B−1 (∂Ω, σ) := M̊1 (∂Ω, σ) , the block-based negative Sobolev space of order
minus one on ∂Ω, with p, q ∈ (1, ∞), 1/p + 1/q = 1 and λ ∈ (0, n − 1), (11.8.35)
. q,λ
B1 (∂Ω, σ), the block-based homogeneous Sobolev space of order one on ∂Ω,
defined as the collection of functions f ∈ L 1 ∂Ω, 1+σ(x)
qλ
|x | n ∩ Lloc (∂Ω, σ) satisfying
the condition ∂τ j k f ∈ B q,λ (∂Ω, σ) for all pairs of indices j, k ∈ {1, . . . , n}, where
q(n−1)
q ∈ (1, ∞), λ ∈ (0, n − 1), and qλ := n−1+λ(q−1) , and equipped with the natural semi-

n
norm, defined as f B. q, λ (∂Ω,σ) := ∂τ j k f B q, λ (∂Ω,σ) (11.13.65), (11.13.66),
1 j,k=1
Definition 11.13.11
C
U, the closure of the set U ⊆ Rn
C k (Ω), the space of functions of class C k in an open neighborhood of Ω
Cck (Ω), the space of functions of class C k with compact support in the open set Ω
Cbk (Ω), the space of bounded functions of class C k in Ω
874 A Terms and Notation used in Volume II
 ∗
Cb∞ (Ω) , the algebraic dual of Cb∞ (Ω)
CBM(Ω), the collection of complex Borel measures in the open set Ω ⊆ Rn
CBM(X, τ), the collection of complex Borel measures in the topological space (X, τ)

Cθ,b (x, h), the cone with vertex at x ∈ Rn , symmetry axis along h ∈ S n−1 and full
aperture θ ∈ (0, π) (cf. [133, (5.6.93)]):

Cθ,b (x, h) := {y ∈ Rn : cos(θ/2) |y − x| < (y − x) · h < b} (A.0.14)

alt
Cmax , the  “altered” Cauchy integral operator on Σ, acting on any function
 maximal
H 1 (ζ )
f ∈ L 1 Σ, 1+ |ζ | according to (cf. [133, (5.9.27)]):
∫
alt f (ζ)
Cmax f (z) := sup dH 1 (ζ) for all z ∈ Σ (A.0.15)
ε>0 ζ−z
ζ ∈Σ
|z−ζ |>ε

Cmax , the
 maximal Cauchy-Clifford
 integral operator whose action on each function
f ∈ L 1 ∂Ω, 1+σ(x)
|x | n−1
⊗ C n is defined as (cf. [133, (5.10.12)]):

∫
1 x−y
Cmax f (x) := sup  ν(y)  f (y) dσ(y) , ∀x ∈ ∂Ω
ε>0 ωn−1 |x − y| n
y ∈∂Ω
|x−y |>ε
(A.0.16)
Cε , the truncated Cauchy-Clifford
 integral
 operator which, for each ε > 0, acts on
σ(x)
every function f ∈ L ∂Ω, 1+ |x | n−1 ⊗ C n according to (cf. [133, (5.10.13)]):
1

∫
1 x−y
Cε f (x) :=  ν(y)  f (y) dσ(y), ∀x ∈ ∂Ω (A.0.17)
ωn−1 |x − y| n
y ∈∂Ω
|x−y |>ε

C, the boundary-to-boundary
  Cauchy-Clifford integral operator acting on each func-
σ(x)
tion f ∈ L ∂Ω, 1+ |x | n−1 ⊗ C n according to (cf. [133, (5.10.14)]):
1

∫
1 x−y
C f (x) := lim+  ν(y)  f (y) dσ(y) for σ-a.e. x ∈ ∂Ω
ε→0 ωn−1 |x − y| n
y ∈∂Ω
|x−y |>ε
(A.0.18)
C n = (C n, +, ), the Clifford algebra generated by n imaginary units defined as
the minimal enlargement of Rn to a unitary real algebra which is not generated (as
A Terms and Notation used in Volume II 875

an algebra) by any proper subspace of Rn , and such that x  x = −|x| 2 for each
x ∈ Rn → C n (cf. [133, §6.4])
Cρ , the triangle inequality “penalty” constant associated with the quasi-distance ρ
as follows (cf. [133, (7.1.3)]):
ρ(x, y)
Cρ := sup ∈ [1, ∞) (A.0.19)
x,y,z ∈X max{ρ(x, z), ρ(z, y)}
not all equal

ρ , the symmetry “penalty” constant associated with the quasi-distance ρ as follows

C
(cf. [133, (7.1.4)]):
ρ := sup ρ(y, x) ∈ [1, ∞)
C (A.0.20)
x,y ∈X ρ(x, y)
xy

· . , the homogeneous Hölder space semi-norm of order α > 0 in the set

C α (U,ρ)
U ⊆ X, in the context of a quasi-metric space (X, ρ), defined for each function
f : U → R as (cf. [133, (7.3.2)]):

| f (x) − f (y)|
f . := sup (A.0.21)
C α (U,ρ)
x,y ∈U ρ(x, y)α
xy

.
C α (U, ρ), the homogeneous Hölder space of order α > 0 in the set U ⊆ X, defined
in the context of a quasi-metric space (X, ρ) as (cf. [133, (7.3.1)]):
. 
C α (U, ρ) := f : U → R : f C.α (U,ρ) < +∞ (A.0.22)

.
C α (U, ρ)/∼, the homogeneous Hölder space of order α > 0 modulo constants, in
the set U ⊆ X, defined in the context of a quasi-metric space (X, ρ) as (cf. [133,
(7.3.6)]): .  . 
C α (U, ρ)/∼ := [ f ] : f ∈ C α (U, ρ) (A.0.23)
.α (U, ρ), the local homogeneous Hölder space of order α > 0 in the set U ⊆ X,
Cloc
defined in the context of a quasi-metric space (X, ρ) as (cf. [133, (7.3.7)]):
.α  .
Cloc (U, ρ) := f : U → C : f Bρ (x,r)∩U ∈ C α Bρ (x, r) ∩ U, ρ

for each x ∈ U and r ∈ (0, ∞) (A.0.24)

· C α (U,ρ) , the inhomogeneous Hölder space norm of order α > 0 in the set U ⊆ X,
in the context of a quasi-metric space (X, ρ), defined for each function f : U → R
as (cf. [133, (7.3.20)]):
876 A Terms and Notation used in Volume II

f C α (U,ρ) := sup | f | + f .
C α (U,ρ)
, ∀ f ∈ C α (U, ρ) (A.0.25)
U

C α (U, ρ), the inhomogeneous Hölder space of order α > 0 in the set U ⊆ X, defined
in the context of a quasi-metric space (X, ρ) as (cf. [133, (7.3.19)]):
 . 
C α (U, ρ) := f ∈ C α (U, ρ) : f is bounded in U (A.0.26)

Ccα (U, ρ), the space of Hölder functions of order α > 0 with ρ-bounded support in
the set U ⊆ X, defined in the context of a quasi-metric space (X, ρ) as (cf. [133,
(7.3.26), (7.3.27)]):
 . 
Ccα (U, ρ) := f ∈ C α (U, ρ) : f vanishes outside of a ρ-bounded subset of U
(A.0.27)
.γ  .γ
Cvan (Σ) := f ∈ C (Σ) : lim+ sup f C.γ (B(x,r)∩Σ) = 0 , the homogeneous “van-
r→0 x ∈Σ
ishing” Hölder space of order γ on the set Σ (3.2.5)

γ (Σ) :=
Cvan f ∈ C γ (Σ) : lim+ sup f C.γ (B(x,r)∩Σ) = 0 , the inhomogeneous
r→0 x ∈Σ
“vanishing” Hölder space of order γ on the set Σ (3.2.8)
X01−θ X1θ , the Calderón product of two quasi-Banach lattices of functions X0 and X1 ,
defined for 0 < θ < 1 by (see (1.4.45))
 
X01−θ X1θ := h ∈ L0 : there exist f ∈ X0, g ∈ X1 such that |h| ≤ | f | 1−θ |g| θ

Cp(X) := Cp(X → X), the space of compact linear operators from X into itself
(1.1.14)
CMO(Σ, σ), the closure in BMO(Σ, σ) of C00 (Σ) (the latter being the space of all
continuous functions on Σ which vanish at infinity) (4.6.11)
 
Cc∞ (Ω) := φ Ω : φ ∈ Cc∞ (Rn ) (9.2.7)
p,q p,q,m  
Cα (Ω) := Cα (Ω) := u ∈ Lloc
1 (Ω, L n ) : u
Cα (Ω) < +∞ , for 0 < p < ∞,
p, q

0 < q ≤ ∞, α > 0, and m ∈ N such that m > α, with

 ∫ 2 δ ∂Ω (x)
1
dt  1/q 

u p, q
Cα (Ω) := u L p (Ω, L n ) + [t −α oscm−1 (u, x, t)]q  p
0 t L x (Ω, L n )

where oscm−1 (u, x, t) is the oscillation of order m − 1 of the function u at location x

and scale t (9.2.25), (9.2.26)
  #
p
Cα (Ω) := u ∈ Lloc 1 (Ω, L n ) : u  
Cα (Ω) := u L p (Ω, L n ) + uα L p (Ω, L n ) < ∞ , for
p
 ∫
p ∈ (0, ∞) and α ≥ 0, where uα# (x) := sup 1
inf r n+α B(x,r)
|u − P| dL n
0<r <δ ∂Ω (x) P ∈ P[α]
for x ∈ Ω (9.2.29), (9.2.30)
A Terms and Notation used in Volume II 877

D
u · w = u, w , the dot product of two vectors u, w ∈ Rn
 the the divergence of the vector field F
divF,
divg , the differential geometric divergence (associated with the metric tensor g; cf.
[133, §1.11])
dg (x, y), the geodesic distance (induced by the metric tensor g) between x and y (cf.
[133, §1.11])
D (Ω), the space of distributions in the open set Ω
D (Ω) ·, · D(Ω) , the distributional pairing in Ω
Δ := ∂12 + · · · + ∂n2 , the Laplace operator in Rn
δx , the Dirac distribution with mass at x
D, the classical (homogeneous) Dirac operator in Rn defined as (cf. [133, (6.4.139)]):

n
D= e j  ∂j (A.0.28)
j=1
 
αβ
D= n
j=1 a j ∂j + bαβ 1≤α ≤ N , a (generic) N × N first-order system
1≤β ≤ N
D , the (real) transpose of the first-order system D:
 n 
αβ
D := − a j ∂j + bαβ 1≤β ≤ N (A.0.29)
j=1 1≤α ≤ N

D, the complex conjugate of the first-order system D

D∗ , the Hermitian adjoint of the first-order system D

d, the exterior derivative operator acting on the differential form u = J uJ dx J
according to (cf. [133, (1.11.32)], and also [133, (6.4.140)-(6.4.141)] for the Clifford
algebra context):
 n 
∂uJ
du = dx j ∧ dx J (A.0.30)
j=1 J
∂ x j

δ, the formal adjoint of the exterior derivative operator d on differential forms (see
also [133, (6.4.142)] for the Clifford algebra context)
δ jk , the Kronecker symbol, i.e., δ jk := 1 if j = k and δ jk := 0 if j  k
Dist [E, F], the Pompeiu-Hausdorff distance between the sets E and F defined as
(cf. [133, (2.8.131)]):

Dist [E, F] := max sup inf |x − y|, sup inf |x − y| (A.0.31)
x ∈E y ∈F y ∈F x ∈E
878 A Terms and Notation used in Volume II

δF (·), the distance function to the set F

δ∂Ω (·), the distance function to the boundary of Ω
UV := (U \ V) ∪ (V \ U), the symmetric difference of the sets U and V
Δ(x, r) := B(x, r) ∩ ∂Ω, the surface ball on ∂Ω with center at x ∈ ∂Ω and radius
r>0
Δ(x, r) := B(x, r) ∩ Σ, the surface ball on the set Σ, with center at x ∈ Σ and radius
r>0
D L , the Dirac operator acting from the left on each Clifford algebra-valued distribu-
tion u ∈ D (Ω) ⊗ C n according to (cf. [133, (6.4.48)]):

n
D L u := e j  (∂j u) (A.0.32)
j=1

DR , the Dirac operator acting from the right on each Clifford algebra-valued distri-
bution u ∈ D (Ω) ⊗ C n according to (cf. [133, (6.4.49)]):

n
DR u := (∂j u)  e j (A.0.33)
j=1

diamρ (A), the ρ-diameter of the set A ⊆ X, in the context of a quasi-metric space
(X, ρ), defined as (cf. [133, (7.1.6)]):

diamρ (A) := sup {ρ(x, y) : x, y ∈ A} (A.0.34)

Dk (X), the k-th generation of dyadic cubes in the geometrically doubling quasi-
metric space X, defined as in [133, Proposition 7.5.4]:

Dk (X) := {Q αk }α∈Ik (A.0.35)

D(X), the dyadic grid on the geometrically doubling quasi-metric space X, defined
as in [133, Proposition 7.5.4]:

D(X) := Dk (X) (A.0.36)
k ∈Z, k ≥κ X

dim X, the dimension of a vector space X

D∗ (Σ), the family of suitably small dyadic cubes on the Ahlfors regular set Σ ⊆ Rn
(7.2.1)
E
(ε, δ)-domain (cf. [133, Definition 5.11.8]):
A Terms and Notation used in Volume II 879

a nonempty, open, proper subset Ω of Rn such that for any x, y ∈ Ω

with |x − y| < δ there exists γ : [0, 1] → Ω rectifiable curve satisfying
γ(0) = x, γ(1) = y, as well as the conditions length(γ) ≤ ε1 |x − y| and (A.0.37)
|z−x | |z−y |
|x−y | ≤ ε1 dist(z, ∂Ω) for each z ∈ γ([0, 1])

e j := (δ jk )1≤k ≤n ∈ Rn , where δ jk is the Kronecker symbol, standard unit vector in

the j-th direction in Rn
{e j }1≤ j ≤n , the standard orthonormal basis in Rn
EΔ , the standard fundamental solution for the Laplacian, defined for each x ∈ Rn \{0}
as
⎧
⎪ 1 1
⎪
⎨ ωn−1 (2 − n) |x| n−2 if n ≥ 3,
⎪
⎪
EΔ (x) := (A.0.38)
⎪
⎪ 1
⎪
⎪ ln |x| if n = 2
⎩ 2π
ext∗ (E), the measure theoretic exterior of the set E ⊆ Rn , defined as (cf. [133,
Definition 2.8.3]):
 L n B(x, r) ∩ E
ext∗ (E) := x ∈ Rn : lim+ =0 (A.0.39)
r→0 L n B(x, r)

EK (Ω), the space of distributions in Ω supported in the compact set K ⊂ Ω

E (Ω), the space of distributions compactly supported in the open set Ω ⊆ Rn
εBA, the generalized Kronecker symbol defined for any two arrays A, B as (cf. [133,
(6.4.116)]): 
det (δab )a ∈ A,b ∈B if | A| = |B|,
εBA := (A.0.40)
0 otherwise

[x]X/Y := x + Y , the equivalence class of the vector x ∈ X in the quotient space X/Y
{Ek }k ∈Z, k ≥κΣ , the conditional expectation operators defined on the Ahlfors regular
set Σ ⊆ Rn (cf. Definition 7.1.5)
E p (X), the p-envelope (with p ∈ (0, 1]) of a quasi-normed space X whose dual
separates points, defined as the completion of X in the quasi-norm · p (cf. Defi-
nition 7.8.4)
Ex∂Ω→Ω , the extension operator from ∂Ω to Ω (cf. (8.4.1), and Theorem 8.4.1)
p,q p,q
EΩ→Rn : As (Ω) → As (Rn ), the extension operator with preservation of class
from the open set Ω to Rn (9.2.58)
F
 ∞ , the contribution of the vector field F at infinity, defined for any family {φ R }R>0
[F]
(referred to as a system of auxiliary functions) consisting of smooth compactly
supported functions in Rn which are globally bounded and progressively equal to 1
on compact sets in a uniform fashion, as (cf. [133, (1.3.2)-(1.3.3)]):
880 A Terms and Notation used in Volume II
∫
 ∞ := − lim
[F] ∇φ R · F dL n (A.0.41)
R→∞ Ω

fE∗ , the non-increasing rearrangement of f : E → R defines as (cf. [133, (6.2.2)]):


fE∗ (t) := inf λ ≥ 0 : μ {x ∈ E : | f (x)| > λ} ≤ t , ∀ t ∈ [0, ∞) (A.0.42)

Φ(X → Y ), the space of Fredholm operators from the linear topological space X
into the linear topological space Y (cf. (2.1.1), Definition 2.2.1)
 
ρFred (T; X) := inf r > 0 : zI − T ∈ Φ(X → X) for each z ∈ C \ B(0, r) , the
Fredholm (or essential spectral) radius of the operator T ∈ Bd(X), with X a quasi-
Banach space (2.2.12)
fγ (x) := sup f , ψ , the Fefferman-Stein grand maximal function (at the point
ψ ∈ Tγ (x)
x ∈ Σ) of the “distribution” f ∈ Lipc (Σ) on the Ahlfors regular set Σ (4.1.6)
Φ ∈ Δ2 , the membership indicating that the Young function Φ is doubling (5.3.7)
. p,q
Fs (Σ, σ), the homogeneous Triebel-Lizorkin space on the Ahlfors regular set
Σ ⊆ Rn , defined as the collection of all functionals f belonging to the dual space
∗     ks  q  1/q 
G̊0 (Σ) satisfying f F. p, q (Σ,σ) :=   p
β,γ
2 |Ek f | L (Σ,σ)
< ∞ if the
s k ∈Z
index p < ∞ (with a natural alteration when q = ∞) and, corresponding to the case
⨏ ∞ 
  q  1/q
when p = ∞, f F. ∞, q (Σ,σ) := sup sup Q 2ks |Ek f | dσ < ∞ (with a
s τ
 ∈Z τ ∈I k=
natural alteration when q = ∞), where the Ek ’s are conditional expectation operators
on Σ (7.1.15), (7.1.16), (7.1.17)
p,q
Fs (Σ, σ), the inhomogeneous Triebel-Lizorkin space on the Ahlfors regular set
β,γ ∗
Σ ⊆ Rn , defined as the collection of functionals f ∈ G0 (Σ) with the property
that f Fsp, q (Σ,σ) < +∞ (7.1.33), (7.1.34), (7.1.35)
f p,q (Σ), the discrete Triebel-Lizorkin space on the Ahlfors regular set Σ ⊆ Rn ,
defined as the collection of numerical sequences λ = {λQ }Q ∈D∗ (Σ) with the property
that the following quasi-norm is finite:

    N
(k,τ)  q 
 1/q 
σ(Qτk,ν )− p λQ k,ν 1Q k,ν
1
λ f p, q (Σ) :=   <∞
τ τ L p (Σ,σ)
k ∈Z τ ∈Ik ν=1
k ≥κΣ

when p < ∞, and with suitable adaptations when q = ∞, or when p = ∞ and

q ∈ (0, ∞] (7.2.25), (7.2.26), Definition 7.2.5
p,q
fs (Σ), the discrete Triebel-Lizorkin space on the Ahlfors regular set Σ ⊆ Rn ,
defined as the collection of sequences λ = {λQ }Q ∈D∗ (Σ) with the property that the
following quasi-norm is finite:
A Terms and Notation used in Volume II 881

    N
(k,τ)  q 
 1/q 
2ks σ(Qτk,ν )− 2 λQ k,ν 1Q k,ν
1
λ p, q
fs (Σ) :=   < ∞,
τ τ L p (Σ,σ)
k ∈Z τ ∈Ik ν=1
k ≥κΣ

when p < ∞, and with natural adaptations when q = ∞, or when p = ∞ and

q ∈ (0, ∞] (7.2.29), (7.2.30), Definition 7.2.6
[ f ]E (x),
⨏ the strictly defined version of a function f at x ∈ E defined as
lim+ E∩B(x,r) f dL n if the limit exists, and zero otherwise (8.1.11)
r→0
p,q
Fs (Rn ), the (inhomogeneous) Triebel-Lizorkin space in Rn for 0 < p, q ≤ ∞ and
s ∈ R, equipped with the quasi-norm · Fsp, q (Rn ) (9.1.4), (9.1.5)
p,q  p,q 
Fs (Ω) := u ∈ D (Ω) : there exists U ∈ Fs (Rn ) such that U Ω = u , the
Triebel-Lizorkin space in the (arbitrary) open set Ω ⊆ Rn for 0 < p, q ≤ ∞ and
s ∈ R, equipped with the quasi-norm (see (9.2.1))

p,q
u Fsp, q (Ω) := inf U Fsp, q (Rn ) : U ∈ Fs (Rn ), U Ω = u

 
F̊s (Ω), the closure of Cc∞ (Ω) in Fs (Ω), · Fsp, q (Ω) for 0 < p, q ≤ ∞ and s ∈ R
p,q p,q

(9.2.2)
p,q  p,q 
Fs,0 (Ω) := u ∈ Fs (Rn ) : supp u ⊆ Ω , for 0 < p, q < ∞, s ∈ R, equipped with
the quasi-norm u F p, q (Ω) := u Fsp, q (Rn ) (9.2.3)
s,0
p,q  p,q 
Fs,z (Ω) := u ∈ D (Ω) : there exists w ∈ Fs,0 (Ω), w Ω = u , for 0 < p, q < ∞
and s ∈ R, equipped with the quasi-norm (see (9.2.4))
 p,q 
u Ap, q
s, z (Ω)
:= inf w Fsp, q (Rn ) : w ∈ Fs,0 (Ω), w Ω = u

G

g= g jk dx j ⊗ dxk , the Riemannian metric tensor
1≤ j,k ≤n

∇u, the gradient (Jacobian matrix) of a C M -valued function u = (uα )1≤α ≤M defined
in an open subset of Rn , defined as:
⎡ ∂1 u1 · · · ∂n u1 ⎤⎥
⎢
⎢ .. ⎥
∇u := ∂j uα 1≤α ≤M = ⎢ ... ..
. . ⎥⎥ (A.0.43)
1≤ j ≤n ⎢
⎢ ∂1 u M · · · ∂n u M ⎥⎦
⎣

∇ , the gradient operator in Rn−1

 
∇tan f := n
k=1 νk ∂τ kj f , the tangential gradient of the function f (11.4.1)
1≤ j ≤n
Γκ (x) = ΓΩ,κ (x), the (κ-)nontangential approach region with vertex at x ∈ ∂Ω,
defined as (cf. [133, (8.1.2)]):
882 A Terms and Notation used in Volume II
 
Γκ (x) = ΓΩ,κ (x) := y ∈ Ω : |x − y| < (1 + κ)δ∂Ω (y) , ∀x ∈ ∂Ω (A.0.44)

X, a Generalized Banach Function Space on the measure space (X, M, μ) equipped

with the norm · X (5.1.5), (5.1.6)
∞ (X, μ) · X ∞
X̊ := Lcomp , the closure of Lcomp in the Generalized Banach Function Space
X, · X (5.2.81)
GΣ (x0, r, β, γ), the collection of all test functions of type (x0, r, β, γ) on the Ahlfors
regular set Σ ⊆ Rn , equipped with the norm · GΣ (x0,r,β,γ) Definition 7.1.1
β,γ
G0 (Σ), the closure of G 1,1 (Σ) in G β,γ (Σ) whenever 0 < β, γ < 1 (7.1.5)
β,γ
G̊0 (Σ), the closure of G̊ 1,1 (Σ) in G̊ β,γ (Σ) whenever 0 < β, γ < 1 (7.1.10)
β,γ
Gb (Σ) := { f ∈ G β,γ (Σ) : f has bounded support} (7.1.61)
Gα (x) := Fξ→x (1 + |ξ | 2 )−α/2 (x), for x ∈ Rn and α > 0, the Bessel kernel of order
α (8.1.5)
H
H n−1 , the (n − 1)-dimensional Hausdorff measure in Rn
Hgn−1 , the (n − 1)-dimensional Hausdorff measure induced by the metric tensor g
H s , the s-dimensional Hausdorff measure in Rn
H∗s , the s-dimensional Hausdorff outer-measure in Rn
%
μ, the Cauchy-Clifford transform of the measure μ, defined at each point x ∈ Ω as
(cf. [133, (6.4.96)]):
∫
1 x−y
%
μ(x) :=  dμ(y) (A.0.45)
ωn−1 Ω |x − y| n

H p (Σ, σ) := f ∈ Lipc (Σ) : fγ ∈ L p (Σ, σ) , for n−1
n < p < ∞ and for parameter
γ ∈ (n − 1) 1
p − 1 +, 1 , the Lebesgue-based Hardy space on the Ahlfors regular set
 
Σ⊆ Rn , equipped with the quasi-norm f H p (Σ,σ) :=  fγ  p (4.2.5), (4.2.6)
L (Σ,σ)
.
H. p (Σ, σ), the homogeneous Hardy space
 on Σ, defined .when Σ is bounded as
H (Σ, σ) := f ∈ H (Σ, σ) :
p p f , 1 = 0 , and simply as H (Σ, σ) := H (Σ, σ) if
p p

Σ is unbounded (4.2.12)

H p,q (Σ, σ) := f ∈ Lipc (Σ) : fγ ∈ L p,q (Σ, σ) , for p ∈ n , ∞ , q ∈ (0, ∞],
n−1

and the (background) parameter γ belonging to the interval (n − 1) p1 − 1 +, 1 , the

Lorentz-based Hardy space on the Ahlfors regular set Σ ⊆ Rn , equipped with the
 
quasi-norm f H p, q (Σ,σ) :=  fγ  L p, q (Σ,σ) (4.2.23), (4.2.24)
p,q
Hat (Σ, σ), for p ∈ (0, 1) and q ∈ [1, ∞], the atomic Hardy space on the Ahlfors
∗
regular set Σ ⊆ Rn , defined as the collection of f ∈ L (n−1)(1/p−1) (Σ) for which
A Terms and Notation used in Volume II 883


∞
there exist {λ j } j ∈N ∈ p (N) and (p, q)-atoms {a j } j ∈N such that f = λ j a j in
j=1
∗
L (n−1)(1/p−1) (Σ) (4.4.4)
1,q
Hat (Σ, σ), for q ∈ (1, ∞], the atomic Hardy space on the Ahlfors regular subset Σ of
Rn , defined as the collection of f ∈ L 1 (Σ, σ) for which there exist {λ j } j ∈N ∈ 1 (N)

∞
and (1, q)-atoms {a j } j ∈N such that f = λ j a j in L 1 (Σ, σ) (4.4.5)
j=1
p,q
Hfin (Σ, σ), the vector space of all finite linear combinations of (p, q)-atoms on
Σ equipped with the quasi-norm f H p, q (Σ,σ) defined as the infimum of all
N  1/p 
N
fin

|λ j | p such that f = λ j a j for {λ j }1≤ j ≤ N ⊆ C and (p, q)-atoms {a j }1≤ j ≤ N

j=1 j=1
(4.4.113)
H, the L p -filtering operator, acting on each given distribution
& f ∈' H p (Σ, σ) with
n < p < ∞ according to (H f )(x) := lim+ (H (Σ,σ)) St (x, ·), f H (Σ,σ) at each
n−1 p ∗ p
t→0
x ∈ Σ, where {St (·, ·)}t are the integral kernels of a suitable approximation to the
identity (4.9.1)
H q,λ (Σ, σ), for q ∈ (1, ∞) and λ ∈ (0, n − 1), the space of f ∈ Lipc (Σ) for which
here exist a sequence {λ j } j ∈N ∈ 1 (N) and a family {a j } j ∈N of H q,λ -atoms on Σ
∞
so that f = λ j a j with convergence in (Lipc (Σ)) equipped with f H q, λ (Σ,σ)
j=1

∞
defined as the infimum of |λ j | over all such writings of f (6.1.16), (6.1.17)
j=1
h p (Σ, σ), the local Hardy space on the Ahlfors regular set Σ ⊆ Rn , defined as the col-
β,γ ∗
lection of functionals f ∈ G0 (Σ) for which there exist two numerical sequences,
 
call them {λ j } j ∈N and {μ j } j ∈N , satisfying |λ j | p < +∞ and | μ j | p < +∞,
j ∈N j ∈N
along with a sequence {a j } j ∈N of p-atoms and a sequence {b j } j ∈N of p-blocks

∞ ∞
β,γ
such that f = λj aj + μ j b j in (G0 (Σ))∗ equipped with the quasi-norm
j=1 j=1
   1/p    1/p
j ∈N |λ j | + j ∈N | μ j |
f h p (Σ,σ) := inf p p with the infimum taken
over all corresponding decompositions of f (7.1.49), (7.1.50)

H s (Ω) := U Ω : U ∈ H s (Rn ) for s ∈ R, the L 2 -based fractional Sobolev spaces
defined in an open set Ω ⊆ Rn equipped with the norm defined for each u ∈ H s (Ω)
as u H s (Ω) := inf U H s (Rn ) : U ∈ H s (Rn ) such that u = U Ω (9.2.20), (9.2.21)
h p (Ω), the local Hardy space in the set Ω, defined as the restriction to the open set
Ω ⊆ Rn of distributions belonging to the local Hardy space h (R ), equipped with
p n

the quasi-norm u h p (Ω) := inf U h p (Rn ) : U ∈ h p (Rn ), U Ω = u (9.2.35)

884 A Terms and Notation used in Volume II
 
hk (Ω) := u ∈ D (Ω) : ∂γ u ∈ h p (Ω), for all γ ∈ N0n with |γ| ≤ k , for k ∈ N0
p

and p ∈ (0, 1], the local Hardy-based Sobolev space equipped with the quasi-norm

u h p (Ω) := ∂γ u h p (Ω) (9.2.43)
k
|γ |=k
 
∂γ uγ with each uγ ∈ h p (Ω) , for k ∈ N0 and
p
h−k (Ω) := u ∈ D (Ω) : u =
|γ | ≤k
p ∈ (0, 1], the localHardy-based Sobolev space of negative order in Ω, equipped with
u h p (Ω) := inf uγ h p (Ω) where the infimum is taken over all corresponding
−k
|γ | ≤k
p
representations of the distribution u ∈ hk (Ω) (9.2.44)
.p
H1 (∂Ω, σ) for p ∈ n ,∞
n−1
, the Hardy-based homogeneous Sobolev space of
 
order one on ∂Ω, defined as the collection of all functions f ∈ L 1 ∂Ω, 1+σ(x) |x | n

satisfying ∂τ j k f ∈ H p(∂Ω, σ) for j, k ∈ {1, . . . , n} and equipped with the semi-

norm f H. p (∂Ω,σ) := nj,k=1 ∂τ j k f H p (∂Ω,σ) (11.10.26)
1
.p 
H1 (∂Ω, σ) ∼, the quotient space of classes [·] of equivalence modulo constants
.p
of functions in H1 (∂Ω, σ) equipped with the semi-quasinorm defined for each
.p      
f ∈ H1 (∂Ω, σ) ∼ as [ f ] H. p (∂Ω,σ)/∼ := nj,k=1 ∂τ j k f  H p (∂Ω,σ) (11.10.33),
1
(11.10.34)
q, p .p
H1 (∂Ω, σ) := L q (∂Ω, σ) ∩ H1 (∂Ω, σ), for p ∈ n−1 n , ∞ and q ∈ [1, ∞], the
Hardy-based inhomogeneous Sobolev space of order one on the set ∂Ω equipped
with the natural quasi-norm f H q, p (∂Ω,σ) := f L q (∂Ω,σ) + f H. p (∂Ω,σ) (11.10.35),
1 1
(11.10.36)
I
·, · , the (real) inner product in C M defined for any vectors u = (uk )1≤k ≤M ∈ C M
and w = (wk )1≤k ≤M ∈ C M as:


M
u, w := u k wk (A.0.46)
k=1

√
i := −1 ∈ C, the complex imaginary unit
ι∗ , the pull-back map induced by the canonical inclusion ι
ι∗# , the sharp pull-back to the set ∂∗ Ω, mapping any (n−1)-form ω ∈ C 0 (M, Λn−1T M)
into the Radon measure (cf. [133, (1.11.58)]):
& '
ι∗# ω := (−1)n−1 (∗ω) ∂∗ Ω, νg ∗ σg on ∂∗ Ω (A.0.47)
T M

ι∗n.t. , the nontangential pull-back to ∂∗ Ω, mapping any given measurable (n − 1)-form

defined in Ω into the Radon measure (cf. [133, (1.11.67)]):
& n.t. '
ι∗n.t. ω := (−1)n−1 (∗ω) ∂Ω, νg σg on ∂∗ Ω (A.0.48)
T∗M
A Terms and Notation used in Volume II 885

int∗ (E), the measure theoretic interior of the set E ⊆ Rn , defined as (cf. [133,
(2.8.18)]):
 L n B(x, r) \ E
int∗ (E) := x ∈ Rn : lim+ =0
r→0 L n B(x, r)
 L n B(x, r) ∩ E
= x ∈ Rn : lim+ =1 (A.0.49)
r→0 L n B(x, r)
⨏ ∫
1
f dμ := μ(E)
E E
f dμ, the integral average of the function f on the set E ⊆ X, in
a measure space (X, μ)
fBρ (x,r) , the integral average of f over the ρ-ball Bρ (x, r), in the context of a space
of homogeneous type (X, ρ, μ), defined as (cf. [133, (7.4.9)]):
⨏ ∫
1
fBρ (x,r) := f dμ := f (y) dμ(y) (A.0.50)
B ρ (x,r) μ Bρ (x, r) Bρ (x,r)
⨏
fΔ := Δ
f dσ, the integral average of the function f in the “surface ball” Δ
IE,α , the fractional integral operator of order α on the set E contained in a metric
space (X, ρ) equipped with upper d-dimensional Borel measure μ on (X, τρ ), acting
μ(x)
on functions f ∈ L 1 E, 1+ρ(x,x ) d−α
according to (cf. [133, (7.8.3)]):
0

∫
f (y)
IE,α f (x) := dμ(y) for μ-a.e. x ∈ E (A.0.51)
E ρ(x, y)d−α

Im(T : X → Y ) := {T x : x ∈ X }, the image (or range) of the operator T (1.1.4)

ind T := dim(Ker T) − codim(Im T), the index of a (semi-)Fredholm operator T
(2.1.5), (2.2.2)
[X0, X1 ]θ , the intermediate space between X0 and X1 (a compatible couple of quasi-
Banach spaces with X0 + X1 analytically convex), corresponding to θ ∈ (0, 1) for the
complex method of interpolation, equipped with the complex interpolation quasi-
norm · [X0,X1 ]θ (1.4.42), (1.4.43)
 
(X0, X1 )θ,q := x ∈ X0 + X1 : x (X0,X1 )θ, q < +∞ , the intermediate space for the
real method of interpolation between the compatible pair of quasi-Banach spaces
X0 and X1 , equipped with the real interpolation quasi-norm · (X0,X1 )θ, q (1.3.39),
(1.3.38)
ln hΦ (t) ln hΦ (t)
i(Φ) := sup ln t = lim+ ln t where hΦ (t) := sup Φ(st)
Φ(s) for t ∈ (0, ∞), the lower
0<t<1 t→0 s>0
dilation index of the Young function Φ (5.3.14)
ln hΦ (t) ln hΦ (t)
I(Φ) := inf ln t = lim ln t where hΦ (t) := sup Φ(st)
Φ(s) for t ∈ (0, ∞), the
1<t<∞ t→∞ s>0
upper dilation index of the Young function Φ (5.3.15)
886 A Terms and Notation used in Volume II

K
KΔ , the boundary-to-boundary harmonic double layer potential, defined as (cf. [133,
(1.1.32)]):
∫
1 ν(y), y − x
KΔ f (x) := lim+ f (y) dH n−1 (y), x ∈ ∂Ω
ε→0 ωn−1 ∂Ω\B(x,ε) |x − y| n
(A.0.52)
KΔ# , the transpose harmonic double layer potential, defined as (cf. [133, (1.1.33)]):
∫
1 ν(x), x − y
KΔ# f (x) := lim+ f (y) dH n−1 (y), x ∈ ∂Ω
ε→0 ωn−1 ∂Ω\B(x,ε) |x − y| n
(A.0.53)
Ker(T : X → Y ) := {x ∈ X : T x = 0}, the kernel (or null-space) of the operator T
(1.1.4)
  M
Ker L := u ∈ C ∞ (Ω) : Lu = 0 in Ω , the null-space of the M × M system L,
in an open set Ω (8.6.6)
L
local John condition, satisfied by an open set Ω ⊆ Rn (cf. [133, Definition 5.11.7]):

there exist θ ∈ (0, 1), Mo ∈ (1, ∞), and R ∈ 0, diam ∂Ω (the
latter required to be ∞ if ∂Ω is unbounded) such that for every
x ∈ ∂Ω and r ∈ (0, R∗ ) one can find a point xr ∈ B(x, r) ∩ Ω
with the property that B(xr , θr) ⊆ Ω and for each y ∈ Δ(x, r) it (A.0.54)
is possible to find a rectifiable path γy : [0, 1] → Ω whose length
is ≤ Mo · r, which satisfies γy (0) = y, γy (1) = xr , and such that
dist γy (t), ∂Ω > θ · |γy (t) − y| for every t ∈ (0, 1]

L n , the (n-dimensional) Lebesgue measure in Rn

∞ , the space of essentially bounded functions with compact support
Lcomp
√
dLgn := g dL n , the Lebesgue measure induced by the metric tensor g
Λ T M, the -th exterior power of the vector bundle on the manifold M
L 0 (X, μ), the space of measurable functions which are pointwise finite μ-a.e. on X
p
Lbdd (Ω, L n ), the space of functions which are p-th power integrable on bounded
subsets of Ω (cf. [133, (4.2.4)]):
p
Lbdd (Ω, L n ) be the collection of all L n -measurable functions defined
in Ω which are p-th power absolutely integrable with respect to the (A.0.55)
Lebesgue measure on each bounded L n -measurable subset of the set Ω.

Lip(X), the space of Lipschitz functions on the (quasi-)metric space X, defined as

(cf. [133, (3.7.2)]):
 
Lip(X) := f : X → C : f Lip(X) < +∞ (A.0.56)
A Terms and Notation used in Volume II 887

· Lip(X) , the natural semi-norm on Lip(X), defined in the context of a metric space
(X , d) as (cf. [133, (3.7.1)]):

| f (x) − f (y)|
f Lip(X) := sup (A.0.57)
x,y ∈X, xy d(x, y)

Lipc (X) the space of Lipschitz functions with bounded support in the (quasi-)metric
space X
Lipc (Σ) the space distributions on a given set Σ ⊆ Rn , defined as (cf. [133,
(4.1.34)]):
the topological dual of Lipc (Σ), τD (A.0.58)

(Lip c (Σ)) ·, · Lip c (Σ) , or simply ·, · , the distributional pairing on the set Σ
L p,q (X,μ), the Lorentz space on X with respect to the measure μ defined as (cf.
[133, (6.2.13)]):

L p,q (X, μ) := f : X → R μ-measurable : f L p, q (X,μ) < +∞ (A.0.59)

· L p, q (X,μ) ,the Lorentz space quasi-norm, defined as (cf. [133, (6.2.14)]):

⎧ ∫ ∞  
⎪ q dt 1/q
⎪
⎪ t 1/p fX∗ (t) if 0 < p, q < ∞,
⎪
⎨ 0
⎪ t
f L p, q (X,μ) := sup  1/p f ∗ (t)
 (A.0.60)
⎪
⎪ t>0 t if 0 < p ≤ ∞, q = ∞,
⎪
⎪
X
⎪ f ∞ if p = ∞, 0 < q ≤ ∞
⎩ L (X,μ)

r (X, μ), the space of L r -integrable functions on subsets of X of finite μ-measure

Lfin
(cf. [133, (6.2.35)]):
 ∫
r
Lfin (X, μ) := f μ-measurable on X : | f | r dμ < ∞ for each
E

μ-measurable set E ⊆ X with μ(E) < ∞ (A.0.61)

p,q
L (Ω, μ), the maximal Lorentz space with respect to the Borel measure μ in the
open set Ω ⊆ Rn , defined as (cf. [133, (6.6.41)]):
p,q  
L (Ω, μ) := u : Ω → C : u is L n -measurable and u,θ ∈ L p,q (Ω, μ) (A.0.62)

p
L (Ω, μ), the maximal Lebesgue space with respect to the Borel measure μ in the
open set Ω ⊆ Rn , defined as (cf. [133, (6.6.43)]):
p p, p
L (Ω, μ) := L (Ω, μ) (A.0.63)
 
= u : Ω → C : u is L n -measurable and u,θ ∈ L p (Ω, μ)
888 A Terms and Notation used in Volume II

L(ξ), the characteristic matrix of the homogeneous (constant-coefficient) higher-


order system L = ∂ α Aαβ ∂ β , defined as (cf. [133, (6.5.39)]):
|α |= |β |=m

L(ξ) := (−1)m ξ α+β Aαβ, ∀ξ ∈ Rn (A.0.64)
|α |= |β |=m

log+ , the positive part of ln, defined for each t ∈ [0, ∞) as (cf. [133, (7.6.68)]):

0 if t ∈ [0, 1],
log+ t := (A.0.65)
ln t if t ∈ [1, ∞)
 
L ρ := f measurable : f := ρ(| f |) < +∞ , the Köthe function space associated
with a function norm ρ defined on M+ , the collection of all non-negative measurable
functions (1.5.1)
L(X → Y ), the space of linear and continuous operators from X to Y (1.1.10)
L β (Σ), the space defined on Σ ⊆ Rn , for β ∈ (0, ∞), as L β (Σ) := C β (Σ) if Σ is
.
bounded, and as L β (Σ) := C β (Σ)/∼ if Σ is unbounded (4.4.1)
 
L Φ (X, μ) := f ∈ M (X, μ) : f L Φ (X,μ) < ∞ , the Orlicz space on the sigma-finite
measure space (X, M, μ) associated with
 a Young function Φ and equipped with the
∫
Luxemburg norm f L Φ (X,μ) := inf λ > 0 : X Φ(| f (x)|/λ) dμ(x) ≤ 1 (5.3.27),
(5.3.23), (5.3.24)
L p exp(logθ L)(X, μ), the Orlicz space (5.3.149), for p ∈ (1, ∞) and θ ∈ [0, 1), defined
as the collection:
 ∫
 θ
f ∈ M (X, μ) : | f (x)| p exp ln(e + | f (x)|) dμ(x) < ∞
X

L p exp(logθ L)(X, μ), the Orlicz space (5.3.149), for p ∈ (1, ∞) and θ ∈ [0, 1), defined
as the collection:
 ∫
 θ
f ∈ M (X, μ) : | f (x)| p exp ln(e + | f (x)|) dμ(x) < ∞
X

L p (logL)α (X, μ), the Zygmund space (5.3.138), for p ∈ (1, ∞) and α ∈ R, defined
as the collection:
 ∫
 α
f ∈ M (X, μ) : | f (x)| p ln(e + | f (x)|) dμ(x) < +∞
X

.
L p,λ (Σ, σ), the homogeneous Morrey-Campanato space on an Ahlfors regular set
Σ ⊆ Rn (6.1.1), (6.1.2), for p ∈ (1, ∞) and λ ∈ (0, n − 1), equipped with the
Morrey-Campanato semi-norm
A Terms and Notation used in Volume II 889
 n−1−λ  ⨏  p1
p
f L. p, λ (Σ,σ) := sup R p f − fΔ(x,R) dσ
x ∈Σ and Δ(x,R)
0<R<2 diam(Σ)

L p,λ (Σ, σ), the inhomogeneous Morrey-Campanato space (6.1.11), (6.1.12), for
each p ∈ (1, ∞) and λ ∈ (0, n − 1), equipped with the Morrey-Campanato norm
f L p, λ (Σ,σ) := f L p (Σ,σ) + f L. p, λ (Σ,σ)
p
w (D∗ (Σ)),the (weighted) quasi-Banach space (7.5.19), (7.5.20), defined as the
collection of sequences of real numbers {λQ }Q ∈D∗ (Σ) satisfying

    1
{λQ }Q ∈D  σ(Q) p − 2 |λQ |
p
1 1 p
∗ (Σ) w
p
(D∗ (Σ))
:= <∞
Q ∈D∗ (Σ)

p
Lα (Rn, wL n ), the weighted Bessel potential space, for p ∈ (1, ∞) and α ∈ (0, ∞),
defined as the collection of Lebesgue-measurable functions u : Rn → C satisfying
u = Gα ∗ f for some f ∈ L p (Rn, wL n ), where w ∈ Ap (Rn, L n ), α ∈ (0, ∞),
p ∈ (1, ∞), equipped with the norm u Lαp (Rn, w L n ) := f L p (Rn, w L n ) (8.1.8), (8.1.9)
L p (Ω, wL n ), the weighted L p Lebesgue space on the set Ω ⊆ Rn , equipped with
∫  1/p
the natural norm u L p (Ω, w L n ) := Ω |u| p w dL n (8.3.1)
(Q), the side-length of a cube Q
λQ, the concentric dilate of the cube Q by the factor λ > 0
Λk , the P. Jones extension operator (8.3.23), (8.3.24), (8.3.25)

p p
Ls (Ω) := U Ω : U ∈ Ls (Rn ) , the Bessel potential space in an open set Ω ⊆ Rn
(9.2.17), (9.2.18), for p ∈ (1, ∞) and s ∈ R, equipped with the norm

p
u Lsp (Ω) := inf U Lsp (Rn ) : U ∈ Ls (Rn ), u = U Ω

p
L1 (∂∗ Ω, σ∗ ), the L p -based (boundary) Sobolev space of order one on ∂∗ Ω (11.1.13),
(11.1.64), for p ∈ [1, ∞], equipped with the norm

f L p (∂∗ Ω,σ∗ ) := f L p (∂∗ Ω,σ∗ ) + ∂τ j k f L p (∂∗ Ω,σ∗ )
1
1≤ j,k ≤n

p,q
L1 (∂∗ Ω, σ∗ ) for p, q ∈ [1, ∞], the off-diagonal (boundary) Sobolev space of or-
der one defined as the collection of f ∈ L p (∂∗ Ω, σ∗ ) for which ∂τ j k f exists in
L q (∂∗ Ω, σ∗ ) for j, k ∈ {1, . . . , n} (11.1.14)
p,q
L1,loc (∂∗ Ω, σ∗ ), for p, q ∈ [1, ∞], the local off-diagonal (boundary) Sobolev space
p q
defined as the collection of f ∈ Lloc (∂∗ Ω, σ∗ ) for which ∂τ j k f exists in Lloc (∂∗ Ω, σ∗ )
for j, k ∈ {1, . . . , n} (11.1.15)
890 A Terms and Notation used in Volume II
p p, p
L1,loc (∂∗ Ω, σ∗ ) := L1,loc (∂∗ Ω, σ∗ ), for p ∈ [1, ∞], the local L p -based (boundary)
Sobolev space of order one (11.1.16)
.p 
L1 (∂Ω, σ) := f ∈ L 1 ∂Ω, 1+σ(x) |x | n : ∂τ j k f ∈ L (∂∗ Ω, σ), 1 ≤ j, k ≤ n , the L -
p p

based homogeneous (boundary)Sobolev space of order one on ∂Ω equipped with

the semi-norm f L. p (∂Ω,σ) := nj,k=1 ∂τ j k f L p (∂∗ Ω,σ) (11.5.55), (11.5.56)
1
p
L1 (∂∗ Ω, σg∗ ) ⊗ E, for p ∈ [1, ∞], the global L p -based (boundary) Sobolev space on
∂∗ Ω, when Ω is an open subset of a manifold equipped with a Riemannian metric g
p
L1,loc (∂∗ Ω, σg∗ ) ⊗ E, for p ∈ [1, ∞], the local L p -based (boundary) Sobolev space
on ∂∗ Ω, when Ω is an open subset of a manifold equipped with a Riemannian metric
g
p
L1 (∂∗ Ω, wσ∗ ), for p ∈ [1, ∞], the weighted (boundary) Sobolev space on ∂∗ Ω
(11.7.4)
p σ∗ (x)
L1 ∂∗ Ω, 1+ |x | a , for p ∈ [1, ∞] and a ∈ [0, ∞), the power weighted (boundary)
σ∗ (x)
Sobolev space on ∂∗ Ω defined as the collection of functions f ∈ L p ∂∗ Ω, 1+ |x | a
σ∗ (x)
with ∂τ j k f ∈ L p (∂∗ Ω, 1+ |x | a ) for j, k ∈ {1, . . . , n} (11.7.6)
p,q
L1 (∂Ω, σ), the Lorentz-based (boundary) Sobolev space of order one on ∂Ω
(11.7.10), (11.7.11), for p ∈ (1, ∞) and q ∈ (0, ∞], equipped with the quasi-norm

n
f p, q
L1 (∂Ω,σ) := f L p, q (∂Ω,σ) + ∂τ j k f L p, q (∂Ω,σ)
j,k=1

 ∗
p p
L−1 (∂∗ Ω, σ∗ ) := L1 (∂∗ Ω, σ∗ ) , for p ∈ (1, ∞), the negative (boundary) Sobolev
space of order minus one on ∂∗ Ω (11.8.1), Definition 11.8.1
 ∗
p,q p ,q
L−1 (∂∗ Ω, σ∗ ) := L1 (∂∗ Ω, σ∗ ) , for p, q ∈ (1, ∞), the off-diagonal negative
(boundary) Sobolev space of order minus one on ∂∗ Ω (11.8.28)
 ∗
L−1 (∂Ω, w) := L1 (∂Ω, w ) , for p ∈ (1, ∞), p := (1 − 1/p)−1 ∈ (1, ∞), and
p p

w := w 1−p , the Muckenhoupt weighted negative (boundary) Sobolev space of

order minus one on ∂Ω (11.8.33)
Λ∗,q , the L q -styled Calderón’s maximal operator defined on ∂Ω, acting on each
⨏  1/q
f − fΔ(x, r ) q
given function f according to Λ∗,q f (x) := sup Δ(x,r) r dσ ,
⨏ 0<r <diam ∂Ω
where fΔ(x,r) := Δ(x,r) f dσ (11.5.217)
M
Mγ∗ (F) upper γ-dimensional Minkowski content of the set F, defined as (cf. [133,
Definition 4.5.1]):

L n {x ∈ Rn : dist(x, F) < r }
Mγ∗ (F) := lim sup ∈ [0, +∞] (A.0.66)
r→0+ r n−γ
A Terms and Notation used in Volume II 891

mE (λ, f ) := μ {x ∈ E : | f (x)| > λ} measure of the level set of f at height λ > 0

M A,s,α , the L s -styled fractional Hardy-Littlewood maximal operator associated with
the family A, defined for each μ-measurable function f on X as (cf. [133, (6.3.9)]):
( ⨏  s1 )
α
M A,s,α f (x) := sup μ A(x, r) | f | s dμ , ∀x ∈ X (A.0.67)
r >0 A(x,r)

M X,s,α , the L s -based fractional Hardy-Littlewood maximal operator of order α in

the space of homogeneous type (X, ρ, μ), defined for each μ-measurable function f
on X as (cf. [133, (7.6.1)]):
* +
⨏  s1
α
M X,s,α f (x) := sup μ(Bρ (x, r)) | f | dμ
s
, ∀x ∈ X (A.0.68)
r >0 B ρ (x,r)

M X,s , the L s -based Hardy-Littlewood maximal operator in the space of homoge-

neous type (X, ρ, μ), defined for each μ-measurable function f on X as (cf. [133,
(7.6.7)]):
⨏  s1
M X,s f (x) := sup | f | s dμ , ∀ x ∈ X (A.0.69)
r >0 B ρ (x,r)

M X,s
R , the local L s -based Hardy-Littlewood maximal operator in the space of homo-

geneous type (X, ρ, μ), defined for each μ-measurable function f on X as (cf. [133,
(7.6.12)]):
⨏  s1
M X,s
R
f (x) := sup | f | s dμ , ∀ x ∈ X (A.0.70)
0<r ≤R B ρ (x,r)

M X , the Hardy-Littlewood maximal operator of the function f on the space of

homogeneous type (X, ρ, μ), defined for each μ-measurable function f on X as (cf.
[133, (7.6.16)]):
∫
1
M X f (x) := sup | f | dμ, ∀ x ∈ X (A.0.71)
r ∈(0,∞) μ Bρ (x, r) B ρ (x,r)

M , the tangential maximal function of u (with exponent M), defined at each x ∈ ∂Ω

umax
as (cf. [133, (8.5.2)]):

M : ∂Ω −→ [0, +∞] defined by

umax
  δ (y)  M  (A.0.72)
 ∂Ω 
M (x) := u(y)
umax  ∞ for each x ∈ ∂Ω
|x − y| L y (Ω, L n )

M (X, μ), the collection of μ-measurable functions defined on an arbitrary measure

space (X, M, μ)
892 A Terms and Notation used in Volume II

M+ (X, μ), the collection of non-negative μ-measurable functions defined on an

arbitrary measure space (X, M, μ) (5.1.1)
M p,λ (Σ, σ), for each integrability exponent p ∈ (0, ∞) and each λ ∈ (0, n − 1), the
Morrey space on the Ahlfors regular set Σ ⊆ Rn , equipped with the Morrey norm
 n−1−λ  ⨏ 1
f M p, λ (Σ,σ) := sup R p | f | p dσ p (6.2.1), (6.2.2)
Δ(x,R)
x ∈Σ and
0<R<2 diam(Σ)

M̊ p,λ (Σ, σ), for p ∈ (0, ∞) and λ ∈ (0, n − 1), the vanishing Morrey space on the
Ahlfors regular set Σ ⊆ Rn , defined as the closure of L s (Σ, σ) with s := p(n−1)
n−1−λ in
M p,λ (Σ, σ) (6.2.14)
M q,λ (Σ, σ), for q ∈ (1, ∞) and λ ∈ (0, n − 1), the (q, λ)-mid space on the Ahlfors
regular set Σ ⊆ Rn , defined as the collection of f ∈ Lipc (Σ) for which there
exist {λ j } j ∈N ∈ 1 (N) and a family {m j } j ∈N of (q, λ)-dull molecules on Σ so that
∞
f = λ j m j with convergence in Lipc (Σ) , equipped with the norm f M q, λ (Σ,σ)
j=1

∞
defined as the infimum of |λ j | over all such possible writings of f (6.2.249),
j=1
(6.2.250)
p,λ
M1 (∂Ω, σ), for each integrability exponent p ∈ (1, ∞) and each λ ∈ (0, n − 1), the
Morrey-based (boundary) Sobolev space of order one on ∂Ω, equipped with the norm

n
f M p, λ (∂Ω,σ) := f M p, λ (∂Ω,σ) + ∂τ j k f M p, λ (∂Ω,σ) (11.7.12), (11.7.13)
1 j,k=1
p,q,λ
M1 (∂Ω, σ), for p, q ∈ (1, ∞) and λ ∈ (0, n − 1), the off-diagonal Morrey-based
(boundary) Sobolev space of order one on ∂Ω, equipped with the natural norm
(11.7.14), (11.7.15)

p,q,λ
M̊1 (∂Ω, σ) := f ∈ M̊ p,λ (∂Ω, σ) : ∂τ j k f ∈ M̊ q,λ (∂Ω, σ), 1 ≤ j, k ≤ n , for all
p, q ∈ (1, ∞) and all λ ∈ (0, n − 1) (11.7.16)
p,λ
M̊1 (∂Ω, σ), for p ∈ (1, ∞) and λ ∈ (0, n − 1), the vanishing Morrey-based (bound-
ary) Sobolev space of order one on ∂Ω, defined as the collection of functions
f ∈ M̊ p,λ (∂Ω, σ) with ∂τ j k f ∈ M̊ p,λ (∂Ω, σ) for j, k ∈ {1, . . . , n} (11.7.17)
 ∗
p,λ q,λ
M−1 (∂Ω, σ) := B1 (∂Ω, σ) , for p, q ∈ (1, ∞), 1/p + 1/q = 1 and λ ∈ (0, n − 1),
the Morrey-based negative (boundary) Sobolev space of order minus one on ∂Ω
(11.8.34)
. p,λ
M1 (∂Ω, σ), for all p ∈ (1, ∞) and λ ∈ (0, n − 1), the homogeneous Morrey-based
(boundary) Sobolev space of order one on ∂Ω defined as the collection
 σ(x) p
f ∈ L 1 ∂Ω, ∩ Lloc (∂Ω, σ) : ∂τ j k f ∈ M p,λ (∂Ω, σ), 1 ≤ j, k ≤ n
1 + |x| n
A Terms and Notation used in Volume II 893


n
equipped with the semi-norm f M. p, λ (∂Ω,σ) := ∂τ j k f M p, λ (∂Ω,σ) (11.13.1),
1 j,k=1
(11.13.2), Definition 11.13.1
. p,λ
M1 (∂Ω, σ), for p ∈ (1, ∞) and λ ∈ (0, n − 1), the homogeneous vanishing Morrey-
based (boundary) Sobolev
 space  of order one on ∂Ω, defined as the collection of
all functions f ∈ L 1 ∂Ω, 1+σ(x)
p
|x | n ∩ Lloc (∂Ω, σ) satisfying ∂τ j k f ∈ M̊
p,λ (∂Ω, σ) for

all indices .j, k ∈ {1, . . . , n} and equipped with the semi-norm given for each func-
p,λ 
tion f ∈ M1 (∂Ω, σ) by f M. p, λ (∂Ω,σ) := nj,k=1 ∂τ j k f M p, λ (∂Ω,σ) (11.13.84),
1
(11.13.85), Definition 11.13.15
N
N0 := N ∪ {0} = {0, 1, 2, . . . }
|∇k u|, the (pointwise) norm of the k-order homogeneous partial derivatives of a
k,1
function u ∈ Wloc (Ω): 
|∇k u| := |∂γ u| (A.0.73)
γ ∈N0n, |γ |=k

NTA domain: a nonempty open subset Ω of Rn satisfying a two-sided corkscrew

condition as well as a Harnack chain condition (cf. [133, Definition 5.11.1])
Nκ u, the (κ-)nontangential maximal operator acting on the function measurable
u : Ω → Rn according to (cf. [133, (8.2.1)]):

Nκ u : ∂Ω −→ [0, +∞], (Nκ u)(x) := u L ∞ (Γκ (x), L n ) for all x ∈ ∂Ω (A.0.74)

Nκε u, the (κ-)nontangential maximal operator truncated at height ε > 0, acting on

the function u : Ω → Rn according to (cf. [133, (1.5.5)]):
 
Nκε u := Nκ (u · 1 Oε ) where Oε := x ∈ Ω : dist(x, ∂Ω) < ε . (A.0.75)

NκE u, the restricted nontangential maximal function of u : Ω → R, relative to the

set E (cf. [133, (8.2.4)]):

NκE u : ∂Ω −→ [0, +∞] defined as

(A.0.76)
(NκE u)(x) := u L ∞ (Γκ (x)∩E, L n ) for each x ∈ ∂Ω

κ,θ,r u, the averaged (κ-)nontangential maximal operator of the measurable function

N
u : Ω → R (cf. [133, (8.10.1)]):

κ,θ,r u : ∂Ω −→ [0, +∞] defined at each x ∈ ∂Ω by

N
 ⨏  1/r 
 (A.0.77)
κ,θ,r u)(x) := Γκ (x)  y →
(N |u| r
dL n 
 
B(y,θ δ ∂Ω (y)) L y∞ (Γκ (x), L n )
894 A Terms and Notation used in Volume II

ν, the geometric measure theoretic outward unit normal to a set Ω ⊆ Rn of locally

finite perimeter, defined via the requirement that (cf. [133, (5.6.2)-(5.6.3)]):

 nmeasure σ∗ in R and a
there exist a locally finite Borel-regular n

vector-valued function ν ∈ L ∞ (Rn, σ∗ ) satisfying |ν(x)| = 1 at (A.0.78)
 n
σ∗ -a.e. x ∈ Rn and with such that ∇1Ω = −νσ∗ in D (Rn )

νg , the geometric measure theoretic outward unit normal induced by the metric
tensor g
ν E , the geometric measure theoretic outward unit normal induced by the standard
Euclidean metric
 
ν • F,  the “bullet” product involving a vector field F ∈ L 1 (Ω, L n ) n (where Ω
bdd
is an arbitrary open subset of Rn ) whose divergence, considered in the sense of
distributions in Ω, satisfies divF ∈ Lbdd 1 (Ω, L n ), defined as a functional acting on

each ψ ∈ Lipc (∂Ω) according to (cf. [133, Proposition 4.2.3]):

∫ ∫
& '
 ψ :=
ν • F, F · ∇Ψ dL n + (divF)Ψ  dL n, (A.0.79)
Ω Ω

where Ψ is any complex-valued function satisfying

Ψ ∈ Lip(Ω), Ψ ∂Ω = ψ, and Ψ ≡ 0 outside

(A.0.80)
of some compact subset of Ω
p
Nκ (Ω; μ), the space of measurable functions in Ω with a p-th power integrable
non-tangential maximal function on ∂Ω with respect to the measure μ (cf. [133,
(8.3.31)]):
p 
Nκ (Ω; μ) := u : Ω → C : u is L n -measurable, and

u Nκp (Ω;μ) := Nκ u L p (∂Ω, μ) < +∞ (A.0.81)

κ−n.t.
u|∂Ω (x), the nontangential trace of the function u : Ω → R at the point x ∈ ∂Ω
such that x ∈ Γκ (x), defined as (cf. [133, Definition 8.9.1]):
κ−n.t.
u|∂Ω (x) is the number a ∈ R with the property that for every
ε > 0 there exists some r > 0 such that |u(y) − a| < ε for L n -a.e. (A.0.82)
point y ∈ Γκ (x) ∩ B(x, r)
 
T X→Y := sup Tu Y : u ∈ X, u X = 1 ∈ [0, +∞], the operator “norm” of
a positively homogeneous mapping T acting from the quasi-normed vector space
X, · X into the quasi-normed vector space Y, · Y (1.2.11)
T Bd(X→Y) := T X→Y
A Terms and Notation used in Volume II 895
ess  
T X→Y := inf T −K X→Y : K ∈ Cp(X → Y ) , the essential norm of the operator
T ∈ Bd(X → Y ), where X, Y are quasi-normed spaces (1.2.54)
x (X0,X1 ) θ, q , the real interpolation quasi-norm (1.3.38)

[X]r := f measurable : | f | 1/r ∈ X , the r-convexification of the quasi-Banach
lattice of functions (X, · X ), equipped with the norm f [X]r := | f | 1/r Xr (1.4.39)

∞  1/p
f Hatp, q (Σ,σ) := inf |λ j | p , with infimum taken over all atomic decomposi-
j=1

∞
p,q
tions f = λ j a j , the quasi-norm on Hat (Σ, σ) (4.4.6)
j=1
∫
N Φ ( f ) := X Φ | f (x)| dμ(x), the modular size of the function f ∈ M (X, μ) for
some Young function Φ (5.3.38)
ν
• u := (−i)Sym(D; ν) • u, the Clifford bullet product of ν with u (10.2.100)
O
1E , the characteristic function of a given set E
ωn−1 := H n−1 (S n−1 ), the surface area of S n−1
 
Oε := x ∈ Ω : δ∂Ω (x) < ε , the one-sided collar neighborhood of ∂Ω with width
ε>0
osc p ( f ; R), the L p -based mean oscillation of the function f at scales up to R, in a
space of homogeneous type (X, ρ, μ), defined as (cf. [133, (7.4.107)]):
⨏  p1
p
osc p ( f ; R) := sup f (y) − fBρ (x,r) dμ(y) ∈ [0, +∞] (A.0.83)
x ∈X, r ∈(0,R) B ρ (x,r)

⨏
oscm (u, x, t) := inf P ∈ Pm B(x,t) |u − P| dL n the oscillation of order m of the function
u at location x and scale t (where Pm is the set of polynomials in Rn of degree ≤ m)
(9.1.64)
P
(a)+ := max{a, 0}, the positive part of the number a ∈ R
& '
·, · E pointwise (real) pairing in the fibers of the Hermitian vector bundle E
&& ''
E (Ω) ·, · E (Ω) , the pairing between a compactly supported distribution u in Ω and a
smooth function f ∈ C ∞ (supp u), say f ∈ C ∞ (O) with O ⊆ Ω open set containing
supp u, defined for each F ∈ C ∞ (Ω) with the property that F = f near supp u as (cf.
[133, (2.2.33)]): && '' & '
E (Ω) u, f E (Ω) := E (Ω) u, F E (Ω) (A.0.84)

·, · Λ TM, the (real) pointwise pairing on Λ T M

∧, the exterior product of differential forms
∨, the interior product of differential forms
896 A Terms and Notation used in Volume II

∂τXY , the tangential derivative operator on the boundary of a subset Ω (of locally
finite perimeter) of a Riemannian manifold (M, g), associated with two vector fields
p
X, Y ∈ T M, acting on any function f ∈ L1 (∂∗ Ω, σg ) ⊗ E according to (cf. [133,
Definition 1.12.7]):

∂τXY f := −hXY − νg (X) divgY − νg (Y ) divg X + νg ([X, Y ]) f (A.0.85)

at σg -a.e. point on ∂∗ Ω, where hXY ∈ L p (∂∗ Ω, σg ) ⊗ E is the unique function with

the property that for each section ϕ ∈ Cc1 (M, E) one has
∫ ∫
& ' & '
f , ∂τXY ϕ E dσg = hXY , ϕ ∂∗ Ω E dσg (A.0.86)
∂∗ Ω ∂∗ Ω

.
∂τ j k u, the weak tangential derivative of a given function u ∈ Lloc
1 (Ω, L n ) with the
 1 n
property that ∇u ∈ Lbdd (Ω, L ) , defined as (cf. [133, Example 4.2.4]):
n

.
∂τ j k u := ν • (∂k u)e j − (∂j u)ek , (A.0.87)

i.e., the “bullet” product ν • Fjk

u , where

Fjk
u
:= (∂k u)e j − (∂j u)ek (A.0.88)

∂∗ E, the measure theoretic boundary of a Lebesgue measurable set E ⊆ Rn , defined

as (cf. [133, (5.2.1)]):
 L n (B(x, r) ∩ E)
∂∗ E := x ∈ Rn : lim sup > 0 and
r→0+ rn
L n (B(x, r) \ E)
lim sup >0 (A.0.89)
r→0+ rn

∂ ∗ E, the reduced boundary of a set E ⊆ Rn of locally finite perimeter, defined as

(cf. [133, (5.6.13)]):
∂ ∗ E consists of all points x ∈ ∂E satisfying the following three
properties:
⨏ 0 < σ∗ (B(x, r)) < +∞ for each r ∈ (0, ∞), formula (A.0.90)
lim+ B(x,r) ν dσ∗ = ν(x) is valid, and |ν(x)| = 1.
r→0

∂T E, the set of points in ∂∗ E at which an approximate tangent plane exists to the set
E ⊆ Rn of locally finite perimeter (cf. [133, (5.6.65)]):
 
∂T E := x ∈ ∂∗ E : there exists an approximate tangent (n − 1)-plane to ∂∗ E at x
(A.0.91)
∂ N E, the collection of points at which a given set E ⊆ Rn of locally finite perimeter
possesses a reasonable unit normal vector (cf. [133, (5.6.69)]):
A Terms and Notation used in Volume II 897

∂ N E := x ∈ Rn : there exists a vector N(x) ∈ S n−1 having the property that

L n B(x, r) ∩ {y ∈ E : (y − x) · N(x) > 0}

lim+ = 0 and
r→0 L n B(x, r)

L n B(x, r) ∩ {y ∈ Rn \ E : (y − x) · N(x) < 0}

lim+ =0 (A.0.92)
r→0 L n B(x, r)

∂lfp Ω, the “locally finite perimeter” boundary of the L n -measurable set Ω ⊆ Rn ,

defined as (cf. [133, (5.7.47)]):

∂lfp Ω := x ∈ ∂Ω : there exists rx > 0 such that

Ω ∩ B(x, rx ) has locally finite perimeter (A.0.93)

∂nta Ω, the nontangentially accessible boundary of an open proper subset Ω of Rn ,

defined as (cf. [133, Definition 8.8.5]):
,  
∂nta Ω := Aκ (∂Ω) = x ∈ ∂Ω : x ∈ Γκ (x) for each κ > 0 (A.0.94)
κ>0

Πm , the projection map of differential forms of mixed degree m onto Λ , the space
m
of differential forms of degree , defined as (cf. [133, (6.4.112)]):
 
m 

Πm u := u I eI if u = uI eI ∈ C m (A.0.95)
|I |= =0 |I |=

(a)+ := max{a, 0} the positive part of the number a ∈ R

πκ (E) = πΩ,κ (E), the “shadow” (or projection) of a given set E ⊆ Ω onto ∂Ω,
defined as (cf. [133, (8.1.15)]):
 
πκ (E) = πΩ,κ (E) := x ∈ ∂Ω : Γκ (x) ∩ E   (A.0.96)

pX , the lower Boyd index of a rearrangement invariant Banach function space X on

a non-atomic sigma-finite measure space (X, M, μ)
 ∫
(Pu)(x) := sup σ(∂Ω∩B(x,r)) Ω∩B(x,r) |u| dL , the P-maximal function
1 n
0<r <2 diam(∂Ω)
(10.1.2)
 ∫ 
(Pg u)(x) := sup 1
σg (∂Ω∩Bg (x,r)) Ω∩Bg (x,r) |u| dLgn , the Pg -maximal function of
r >0
Carleson type on manifolds (10.3.2)
898 A Terms and Notation used in Volume II
& ' ∫
P.V. b k(x − ·) Σ , φ := lim+ b(y)k(x − y)φ(y) dσ(y), the principal-value
ε→0 y ∈Σ
|y−x |>ε
distribution associated with the kernel k on the countably rectifiable, upper Ahlfors
regular, set Σ (11.9.2)
∂νA(·, ·), the conormal derivative operator with respect to the coefficient tensor A
acting from weighted Sobolev spaces (8.5.14), and from Besov/Triebel-Lizorkin
spaces (9.5.7)
.
∂νAu := ν • (A∇u)α 1≤α ≤M , the weak conormal derivative of u with respect to the
coefficient tensor A (10.2.138)
   
∂τ j k ϕ := ν j ∂k ϕ O∩∂∗ Ω − νk ∂j ϕ O∩∂∗ Ω , the pointwise tangential derivative oper-
ator acting of a function ϕ ∈ C 1 (O) (11.1.4)
∂τ j k , the tangential derivative operator defined in a weak (distributional) sense
(11.1.17), (11.2.2)
Q
X/Y , the quotient space of a vector space X and a linear subspace Y of X
qX , the upper Boyd index of a rearrangement invariant Banach function space X on
a non-atomic sigma-finite measure space (X, M, μ)
R
[w]RHq , the reverse Hölder constant of the weight w in the context of a space of
homogeneous type (X, ρ, μ), defined as (cf. [133, (7.7.17)]):
 1 ∫  q1  1 ∫  −1
[w]RHq := sup w q dμ w dμ (A.0.97)
B ρ -ball μ(B) B μ(B) B

RHq (X, ρ, μ), the L q reverse Hölder class on a space of homogeneous type (X, ρ, μ),
defined as (cf. [133, item (11) in Lemma 7.7.1]):
 
RHq (X, ρ, μ) := w weight function : [w]RHq < ∞ (A.0.98)

regsupp u, the smallest closed set outside of which the distribution u is a locally
integrable function (cf. [133, (1.5.4)]):
regsupp u := the smallest relatively closed subset of Ω
(A.0.99)
outside of which u is a locally integrable function

R+n , the (open) upper half-space in Rn

R−n , the (open) lower half-space in Rn
RRn →∂Ω , the restriction operator from Rn to ∂Ω
R j,max , the maximal j-th
 Riesz transform
 defined on an Ahlfors regular set Σ ⊆ Rn ,
acting on each f ∈ L 1 Σ, 1+σ(x)
|x | n−1
according to (cf. [133, (5.10.15)]):
A Terms and Notation used in Volume II 899

∫
2 xj − yj
(R j,max f )(x) := sup f (y) dσ(y) , ∀x ∈ Σ (A.0.100)
ε>0 ωn−1 |x − y| n
y ∈Σ
|x−y |>ε

R j,ε , the truncated j-th Riesz transform

 defined on an Ahlfors regular set Σ ⊆ Rn ,
acting on each f ∈ L 1 Σ, 1+σ(x)|x | n−1
according to (cf. [133, (5.10.16)]):
∫
2 xj − yj
R j,ε f (x) := f (y) dσ(y), ∀x ∈ Σ (A.0.101)
ωn−1 |x − y| n
y ∈Σ
|x−y |>ε

R j , the boundary-to-boundary (or principal-value) j-th Riesz transform

 defined on
σ(x)
an Ahlfors regular set Σ ⊆ R , acting on each f ∈ L Σ, 1+ |x | n−1 according to (cf.
n 1

[133, (5.10.17)]):
∫
2 xj − yj
(R j f )(x) := lim+ f (y) dσ(y) for σ-a.e. x ∈ Σ (A.0.102)
ε→0 ωn−1 |x − y| n
y ∈Σ
|x−y |>ε

j , the distributional j-th Riesz transform defined on an Ahlfors regular set Σ in

Rweak
Rn , defined as the mapping from Lipc (∂Ω) into Lipc (∂Ω) , whose action is described
for each f , g ∈ Lipc (∂Ω) as (cf. [133, (5.10.19)]):
- .
weak
Lip c (∂Ω) R j f ,g (A.0.103)
Lip c (∂Ω)
∫ ∫
2 xj − yj  
:= f (y)g(x) − f (x)g(y) dσ(y) dσ(x)
ωn−1 |x − y| n
∂Ω ∂Ω

rad(Ω), the number associated with any nonempty open set Ω ⊆ Rn as (cf. [133,
(5.11.31)]):

rad(Ω) := inf inf sup |x − y|, where

j x ∈Ω j y ∈Ω j
(A.0.104)
{Ω j } j are the connected components of Ω

R Rn →Ω : D (Rn ) → D (Ω) restriction operator from Rn to the open set Ω (9.2.5)

(k)
RΩ→∂Ω multi-trace operator of order k from Ω to its boundary ∂Ω, defined for
 
(k)
⨏
p,q γ
each u ∈ Aα (Ω) according to RΩ→∂Ω u (x) := lim+ B(x,r) ∂ w dL n at
r→0 |γ | ≤k−1
p,q
σ-a.e. x ∈ ∂Ω, where w ∈ Aα (Rn ) is an extension of u (9.4.8)
900 A Terms and Notation used in Volume II

S
σ := H n−1 ∂Ω, the surface measure on ∂Ω
σ∗ := H n−1 ∂∗ Ω, the surface measure on ∂∗ Ω
σg , the surface measure induced by the metric tensor g on the boundary of a set of
locally finite perimeter in a Riemannian manifold (M, g)
σ E surface measure induced by the standard Euclidean metric
 
αβ αβ
Sym(D; ξ), the principal symbol of a system D = j=1 a j ∂ j + b
n
1≤α ≤ N ,
1≤β ≤ N
defined as (cf. [133, (1.7.16)]):

n 
αβ
Sym(D; ξ) := i aj ξj 1≤α ≤ N for all ξ = (ξ1, . . . , ξn ) ∈ Rn (A.0.105)
j=1 1≤β ≤ N

∗, the Hodge star operator

ξ → ξ  , X → X  , the musical isomorphisms (metric identifications via lowering
and rising indices), defined as (cf. [133, (1.12.139)]):

T M  X = X j ∂j −→ X  := g jk Xk dx j ∈ T ∗ M,
(A.0.106)
T ∗ M  ξ = ξ j dx j −→ ξ  := g jk ξ j ∂k ∈ T M

S n−1 := ∂B(0, 1), the unit sphere in Rn

S±n−1 := S n−1 ∩ R±n , the upper and lower (open) hemispheres
C k -singsup u, the smallest closed set outside of which the distribution u is of class
Ck
S(X, μ), the collection of simple functions on the measure space (X, μ)
Sfin (X, μ), the collection of simple functions on the measure space (X, μ) with support
of finite measure, defined as (cf. [133, (3.1.10)]):
 
Sfin (X, μ) = s ∈ S(X, μ) : μ({s  0}) < ∞ (A.0.107)

suppμ, the support of the Borel measure measure μ on a topological space (X, τ),
defined as (cf. [133, (3.8.1) in Definition 3.8.1]):
 
supp μ := x ∈ X : μ(O) > 0 for each open set O ⊆ X such that x ∈ O .
(A.0.108)
supp f , the support of the μ-measurable function f defined on a topological space
(X, τ) (where μ is a Borel measure measure on (X, τ)), defined as (cf. [133, (3.8.7)
in Definition 3.8.3]):
A Terms and Notation used in Volume II 901
 ∫
supp f := x ∈ X : | f | dμ > 0 for each open set O ⊆ X with x ∈ O
O
(A.0.109)
S (Rn ), the space of (smooth, rapidly decreasing) Schwartz functions in Rn
S (Rn ), the space of tempered distributions in Rn

uscal , the scalar part of the Clifford element u = I u I eI ∈ C n ∈ C n , defined as (cf.
[133, (6.4.26)]):
uscal := u e = u ∈ R (A.0.110)

u,θ , the solid maximal function of u : Ω → C, defined at each point x ∈ Ω according

to (cf. [133, (6.6.2)]):

u,θ (x) := u L ∞ (B(x,θ δ ∂Ω (x)), L n ) ∈ [0, ∞] (A.0.111)

E , the local solid maximal function of u : E → C (where E ⊆ Ω), defined at each

u,θ
point x ∈ Ω according to (cf. [133, (6.6.79)]):
E
u,θ (x) := u L ∞ (E∩B(x,θ δ ∂Ω (x)), L n ) ∈ [0, +∞] (A.0.112)

ρsym , the symmetrized version of the quasi-distance ρ for a quasi-metric space (X, ρ),
defined as (cf. [133, (7.1.15)]):

ρsym : X × X −→ [0, ∞],

(A.0.113)
ρsym (x, y) := max{ρ(x, y), ρ(y, x)}, ∀x, y ∈ X

ρ# , the regularized version of the quasi-distance ρ for a quasi-metric space (X, ρ),
defined as (cf. [133, (7.1.17) in Theorem 7.1.2]):

ρ# := (ρsym )α for α := (log2 Cρ )−1 ∈ (0, ∞]. (A.0.114)

fp# , the L p -based Fefferman-Stein sharp maximal function of f ∈ Lloc

1 (X, μ), defined

as (cf. [133, (7.4.110)]):

⨏  1/p
p
fp# (x) := sup f (y) − fBρ (x,r) dμ(y) , ∀x ∈ X (A.0.115)
r >0 B ρ (x,r)

Φ+ (X → Y ), the space of finite-dim kernel semi-Fredholm operators from the

Banach space X into the Banach space Y (2.1.3)
Φ− (X → Y ), the space of finite-dim cokernel semi-Fredholm operators from the
Banach space X into the Banach space Y (2.1.4)
 
ρinv (T; X) := inf r > 0 : zI − T : X → X homeomorphism for all z ∈ C \ B(0, r) ,
the spectral radius of T ∈ Bd(X), with X a quasi-Banach space (2.2.11)
902 A Terms and Notation used in Volume II

T
Tγ (x), the family of “bump” (i.e., localized, and normalized in the Hölder norm)
functions centered at x ∈ Σ (4.1.2)
τt , the dilation by a factor of t mapping S (Rn ) into itself, acting on each ϕ ∈ S (Rn )
according to (cf. [133, (4.5.36)]):

(τt ϕ)(x) := ϕ(t x), ∀x ∈ Rn (A.0.116)

τρ , the topology induced by the quasi-distance ρ in a quasi-metric space (X, ρ),

defined as (cf. [133, (7.1.7)]):
def
O is open in τρ ⇐⇒ for any x ∈ O there is r > 0 with Bρ (x, r) ⊆ O (A.0.117)
⨏
(TrRn →Σ u)(x) := [u]Rn (x) := lim+ B(x,r)
u(y) dy, the trace operator from Rn to the
r→0
set Σ (8.1.12)
⨏
(TrΩ→∂Ω u)(x) := [u]Ω (x) := lim+ B(x,r)∩Ω
u(y) dy, the trace operator from the set
r→0
Ω ⊆ Rn to its boundary ∂Ω (8.3.37)
TrΩ→∂Ω , the trace operator from the open set Ω ⊆ Rn to its ⨏boundary ∂Ω, acting
p,q
on each u ∈ Aα (Ω) according to TrΩ→∂Ω u (x) := lim+ B(x,r) w dL n , where
r→0
p,q
w ∈ Aα (Rn ) is an extension of u (9.4.89)
U
UR set (cf. [133, Definition 5.10.1]): a closed set Σ ⊂ Rn which is (upper) Ahlfors
regular and has Big Pieces of Lipschitz Images (in a uniform, scale-invariant fashion)
UR domain (cf. [133, Definition 5.10.6]): a nonempty open subset Ω of Rn such that
∂Ω is a UR set and H n−1 (∂Ω \ ∂∗ Ω) = 0
U # V, the union of two disjoint sets U, V
UC(X, ρ), the space of uniformly continuous functions on the metric space (X, ρ)
+ (x) :=
uΩ sup |(ψt ∗ u)(x)|, for x ∈ Ω, the radial maximal function of
0<t<dist(x,∂Ω)
u ∈ D (Ω) (9.2.38)
∗ (x) := sup{| u, ψ | : ψ ∈ Ψ } for x ∈ Rn and k ∈ N , the grand maximal
uk,Ω x 0
function of u ∈ D (Ω), where Ψx is the collection of all test functions ψ ∈ Cc∞ (Rn )
for which there exits r > 0 such that supp ψ ⊆ B(x, r) ∩ Ω and sup |∂γ ψ| ≤ r −n− |γ |
Rn
for all γ ∈ N0n with |γ| ≤ k (9.2.39), (9.2.40)
V
√
dVg := g dx1 ∧ · · · ∧ dxn , the (expression in local coordinates of the) volume
element induced by the metric tensor g on a Riemannian manifold (M, g)
VarF, the pointwise variation of the function F : (a, b) → C, defined as (cf. [133,
(2.6.12)]):
A Terms and Notation used in Volume II 903

b
VarF (A.0.118)
a


N
:= sup |F(x j ) − F(x j−1 )| : N ∈ N and a < x0 < · · · < x N < b ∈ [0, +∞]
j=1

V( f ; O), the variation of the function f in the set O, defined as (cf. [133, (2.6.12)]):
 ∫  n
V( f ; O) := sup f divϕ dL n : ϕ ∈ Cc1 (O) with sup ϕ ≤ 1 (A.0.119)
O O


uvect , the vector part of the Clifford element u = I u I eI ∈ C n defined as (cf. [133,
(6.4.25)]):

n
uvect := u j e j ∈ Rn (A.0.120)
j=1

VMO(X, μ) := the closure of UC(X, ρ) ∩ BMO(X, μ) in BMO(X, μ), the Sarason

space of functions of vanishing mean oscillations on the measure metric space
(X, ρ, μ) (3.1.1)
VMO−1 (∂Ω, σ), the VMO-based negative (boundary) Sobolev space of order minus
one on ∂Ω, defined as the closure of L n−1 (∂Ω, σ) in BMO−1 (∂Ω, σ) (11.10.51),
Definition 11.10.9
W
weakly elliptic system: a system L with invertible characteristic matrix, i.e.,
det [L(ξ)]  0 for each ξ ∈ Rn \ {0}
W k, p (Ω), the L p -based Sobolev space of order k in Ω (intrinsically defined)
k, p
Wloc (Ω), the local L p -based Sobolev space of order k in Ω
k, p
Wbdd (Ω), the space of Sobolev functions on any bounded measurable subset of Ω
(cf. [133, (3.0.4)]):
k, p k, p
Wbdd (Ω) denote the space of functions u ∈ Wloc (Ω) with the property
that ∂ α u ∈ L p (O, L n ) for each α ∈ N0n with |α| ≤ k and each bounded (A.0.121)
Lebesgue measurable subset O of Ω.

W k, p (Rn, wL n ), for p ∈ (0, ∞) and k ∈ N0 , the weighted Sobolev space in Rn

 ∫  1/p
equipped with the norm f W k, p (Rn,w L n ) := Rn
|∂ β f | p w dL n (8.1.1),
|β | ≤k
(8.1.2)
 
k,1
W k, p (Ω, wL n ) := u ∈ Wloc (Ω) : ∂ β u ∈ L p (Ω, wL n ) for all β ∈ N0n, | β| ≤ k ,
for each p ∈ (0, ∞) and k ∈ N, the weighted L p -based Sobolev space of order k
904 A Terms and Notation used in Volume II

in the open set Ω ⊆ Rn , equipped with the (intrinsic) quasi-norm defined for every
 ∫ β u| p w dL n
 1/p
u ∈ W k, p (Ω, wL n ) as u W k, p (Ω,w L n ) := Ω
|∂ (8.3.5), (8.3.6)
|β | ≤k
k, p
Wa (Ω) for k ∈ N0 , p ∈ (0, ∞), a ∈ R, the weighted Sobolev space in the open
ap
set Ω ⊆ Rn , corresponding to the weight w := δ∂Ω , equipped with the quasi-norm
 ∫  1/p
u W k, p (Ω) := Ω
|(∂ α u)(x)| p δ∂Ω (x)ap dx (8.3.32), Definition 8.3.4
a
|α | ≤k

for p ∈ (1, ∞) and a ∈ (−1/p, 1 − 1/p), the closure of Cc∞ (Ω) in the
1, p
W̊a (Ω),
1, p
weighted Sobolev space Wa (Ω) (8.3.65)
 
−1, p ap
Wa (Ω) := f = f0 + nj=1 ∂j f j ∈ D (Ω) : f j ∈ L p Ω, δ∂Ω L n , 0 ≤ j ≤ n , for
p ∈ (1, ∞) and a ∈ (−1/p, 1 − 1/p), the weighted Sobolev space of order minus one

n+1
equipped with the norm f W −1, p (Ω) given by the infimum of f j L p (Ω,δ a p L n )
a ∂Ω
j=0
taken over all the corresponding decompositions of f (8.5.1), (8.5.2)
k, p
Wa, (Ω), for each p ∈ (0, ∞), k ∈ N0 , and a ∈ R, the weighted maximal Sobolev
(Ω) satisfying ∂ α u ∈ L Ω, δ∂Ω L n for α ∈ N0n
k,1 p ap
space, the collection of all u ∈ Wloc

with |α| ≤ k, equipped with the quasi-norm u W k, p (Ω) := ∂ α u L p (Ω,δ a p L n )
a, ∂Ω
|α | ≤k
(8.6.1), (8.6.2)
. k, p
Wa, (Ω), for p ∈ (0, ∞), k ∈ N0 , and a ∈ R, the homogeneous weighted max-
k,1
imal Sobolev space, defined as the collection of all u ∈ Wloc (Ω) satisfying
α p
∂ u ∈ L Ω, δ∂Ω
ap n
L for α ∈ N n with |α| = k, equipped with the quasi-norm
 0
u W. k, p (Ω) := ∂ α u L p (Ω,δ a p L n ) (8.6.3), (8.6.4)
a, ∂Ω
|α |=k
W s, p (Σ, σ), for p ∈ [1, ∞] and s > 0, the fractional Hajłasz-Sobolev space defined
as the set of all functions f ∈ L p (Σ, σ) for which there exists g ∈ L p (Σ, σ) such that
| f (x) − f (y)| ≤ |x − y| s g(x) + g(y) for σ-a.e. x, y ∈ Σ (11.11.1)
X
X∗ ·, · X, the duality pairing between a vector space X and its algebraic dual X ∗
X , the associated space of the Generalized Banach Function Space X, equipped
with the norm · X (5.1.34), (5.1.35)
 & '
x p := · X, p := inf λ > 0 : λ−1 x ∈ BX (0, 1) p , for all x ∈ X, the Minkowski
functional associated with the absolutely p-convex hull of the unit ball in X (7.8.6)
Xbdd (Ω) = X (Ω)bdd the collection of distributions u in Ω satisfying ψ Ω u ∈ X (Ω)
for each cutoff function ψ ∈ Cc∞ (Rn ) (8.3.42)
References

1. D.R. Adams, A note on Riesz potentials, Duke Math. J., 42 (1975), no. 4, 765–778.
2. D.R. Adams, Morrey Spaces, Lecture Notes in Applied and Numerical Harmonic Analysis,
Birkhäuser, 2015.
3. D.R. Adams and J.J.F. Fournier, Sobolev Spaces, Pure and Applied Mathematics Series,
Vol. 140, Academic Press, 2-nd edition, 2003.
4. D.R. Adams and L.I. Hedberg, Function Spaces and Potential Theory, Springer-Verlag, Berlin,
Heidelberg, 1996.
5. D.R. Adams and N.G. Meyers, Bessel potentials. Inclusion relations among classes of excep-
tional sets, Indiana Univ. Math. J., 22 (1973), 873–905.
6. D.R. Adams and J. Xiao, Morrey spaces in harmonic analysis, Ark. Mat., 50 (2012), 201–230.
7. E. Albrecht, Spectral interpolation, pp. 13–37 in “Operator Theory: Advances and Applica-
tions,” Vol. 14, Birkhäuser, Basel, 1984.
8. A.B. Aleksandrov, Essays on nonlocally convex Hardy classes, Complex Analysis and Spec-
tral Theory (Leningrad, 1979/1980), pp. 1–89, Lecture Notes in Math., No. 864, Springer,
Berlin-New York, 1981.
9. R. Alvarado and M. Mitrea, Hardy Spaces on Ahlfors-Regular Quasi-Metric Spaces. A Sharp
Theory, Lecture Notes in Mathematics, Vol. 2142, Springer, 2015.
10. J. Alvarez, Continuity of Calderón-Zygmund type operators on the predual of a Morrey space,
pp. 309–319 in “Clifford Algebras in Analysis and Related Topics”, Stud. Adv. Math. CRC,
Boca Raton, FL, 1996.
11. T. Aoki, Locally bounded topological spaces, Proc. Japan Acad. Tokyo, 18 (1942), 588–594.
12. S. Axler, N. Jewell, and A. Shields, The essential norm of an operator and its adjoint, Trans.
Amer. Math. Soc., 261 (1980), no. 1, 159–167.
13. O. Bakas, A variant of Yano’s extrapolation theorem on Hardy spaces, Arch. Math., 113
(2019), 537–549.
14. C. Bennett and R. Sharpley, Interpolation of Operators, Pure and Applied Mathematics,
Vol. 129, Academic Press, 1988.
15. J. Bergh and J. Löfström, Interpolation Spaces. An Introduction, Springer-Verlag, Berlin/New
york, 1976.
16. A. Bernal and J. Cerdà, Complex interpolation of quasi-Banach spaces with an A-convex
containing space, Ark. Mat., 29 (1991), 183–201.
17. D.D. Bleecker and B. Booss-Bavnbek, Index Theory with Applications to Mathematics and
Physics, International Press, Boston, 2013.
18. M. Bownik, Boundedness of operators on Hardy spaces via atomic decompositions, Proc.
Amer. Math. Soc., 133 (2005), no. 12, 3535–3542.

© The Editor(s) (if applicable) and The Author(s), under exclusive license to 905
Springer Nature Switzerland AG 2022
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1
906 References

19. K. Brewster, D. Mitrea, I. Mitrea, and M. Mitrea, Extending Sobolev functions with partially
vanishing traces from locally (ε, δ)-domains and applications to mixed boundary problems,
J. Funct. Anal., 266 (2014), 4314–4421.
20. Y. Brudnyı̆ and N. Krugljak, Interpolation Functors and Interpolation Spaces, Vol. I, North-
Holland Mathematical Library, No. 47, North-Holland Publishing Co., Amsterdam, 1991.
21. Yu.D. Burago, V.G. Maz’ya, and V.D. Sapozhnikova, On the double layer potential for non-
regular regions, Sov. Math. Dokl., 3 (1962), 1640–1642 (translation from Dokl. Akad. Nauk
SSSR 147 (1962), 523–525).
22. A.M. Caetano, Approximation by functions of compact support in Besov-Triebel-Lizorkin
spaces on irregular domains, Studia Math., 142 (2000), no. 1, 47–63.
23. A.P. Calderón, Intermediate spaces and interpolation, the complex method, Studia Math., 24
(1964), 113–190.
24. A.P. Calderón, Commutators, singular integrals on Lipschitz curves and applications, pp. 85–
96 in “Proceedings of the International Congress of Mathematicians” (Helsinki, 1978), Acad.
Sci. Fennica, Helsinki, 1980.
25. L. Carleson, Two remarks on H 1 and BMO, Advances in Math., 22 (1976), no. 3, 269–277.
26. M.J. Carro, On the range space of Yano’s extrapolation theorem and new extrapolation
estimates at infinity, Publ. Mat., 46 (2002), 27–37.
27. F. Chiarenza and M. Frasca, Morrey spaces and Hardy-Littlewood maximal function, Rend
Mat., 7 (1987), 273–279.
28. M. Christ, The extension problem for certain function spaces involving fractional orders of
differentiability, Arkiv för Matematik, 22 (1984), no. 1-2, 63–81.
29. S.-K. Chua, Extension Theorems on Weighted Sobolev Spaces, Indiana Univ. Math. J., 41
(1992), no. 4, 1027–1076.
30. F. Cobos, P. Fernández-Matrinez, and A. Martinez, Interpolation of measure of non-
compactness by the real method, Studia Math., 135 (1999), no. 1, 25–38.
31. R.R. Coifman, A real variable characterization of H p , Studia Math., 51 (1974), 269–274.
32. R. Coifman, A. McIntosh, and Y. Meyer, L’intégrale de Cauchy definit un opérateur borné
sur L 2 pour les courbes lipschitziennes, Ann. Math., 116 (1982), 361–388.
33. R. Coifman, Y. Meyer, and E.M. Stein, Some new function spaces and their applications to
Harmonic Analysis, J. Funct. Analysis, 62 (1985), 304–335.
34. R.R. Coifman, R. Rochberg and G. Weiss, Factorization theorems for Hardy spaces in several
variables, Ann. of Math., 103 (1976), no. 3, 611–635.
35. R.R. Coifman and G. Weiss, Analyse Harmonique Non-Commutative sur Certains Espaces
Homogènes, Lecture Notes in Mathematics No. 242, Springer-Verlag, 1971.
36. R.R. Coifman and G. Weiss, Extensions of Hardy spaces and their use in analysis, Bull. Amer.
Math. Soc., 83 (1977), no. 4, 569–645.
37. D. Cruz-Uribe, J. Martell and C. Pérez, Extensions of Rubio de Francia’s extrapolation theo-
rem, pp. 195–231 in Collect. Math., Vol. Extra, 2006.
38. D. Cruz-Uribe, J.M. Martell, and C. Pérez, Weights, extrapolation and the theory of Rubio de
Francia, Operator Theory: Advances and Applications, Vol. 215, Birkhäuser/Springer Basel
AG, Basel, 2011.
39. M. Cwikel and P. Nilsson, Interpolation of weighted Banach lattices, Technion Preprint Series
No. MT-834.
40. G. Dafni and H. Yue, Some characterizations of local bmo and h 1 on metric measure spaces,
Anal. Math. Phys., 2 (2012), no. 3, 285–318.
41. B.E.J. Dahlberg and C.E. Kenig, Hardy spaces and the Neumann problem in L p for Laplace’s
equation in a Lipschitz domain, Ann. of Math., 125 (1987), 437–465.
42. B.E. Dahlberg, C.E. Kenig, J. Pipher, and G.C. Verchota, Area integral estimates for higher
order elliptic equations and systems, Ann. Inst. Fourier, (Grenoble) 47 (1997), no. 5, 1425–
1461.
43. G. David, J.L. Journé, and S. Semmes, Opérateurs de Calderón-Zygmund, fonctions para-
accrétives et interpolation, Rev. Mat. Iberoamericana, 1 (1985), 1–56.
44. W.J. Davis, D.J. H. Garling, and N. Tomczak-Jaegermann, The complex convexity of quasi-
normed linear spaces, J. Funct. Anal., 55 (1984), 110–150.
References 907

45. R.A. DeVore and R.C. Sharpley, Maximal Functions Measuring Smoothness, Memoirs of the
Amer. Math. Soc., Vol. 47, No. 293, AMS, Providence, 1984.
46. N. Dunford and J.T. Schwartz, Linear Operators. Part I: General Theory, Pure and Applied
Mathematics, Vol. VII, Interscience Publishers, Inc., New York, 1957.
47. D.E. Edmunds and M. Krbec, Variations on Yano’s extrapolation theorem, Rev. Mat. Com-
plut., 18 (2005), no. 1, 111–118.
48. D.E. Edmunds and H. Triebel, Function Spaces, Entropy Numbers, Differential Operators,
Cambridge University Press, Cambridge, 1996.
49. L.C. Evans and R.F. Gariepy, Measure Theory and Fine Properties of Functions, Studies in
Advanced Mathematics, CRC Press, Boca Raton, FL, 1992.
50. E.B. Fabes, M. Jodeit Jr., and N.M. Rivière, Potential techniques for boundary value problems
on C 1 -domains, Acta Math., 141 (1978), no. 3–4, 165–186.
51. C. Fefferman, N. Rivière, and Y. Sagher, Interpolation between H p spaces: the real method,
Trans. Amer. Math. Soc., 191 (1974), 75–81.
52. C. Fefferman and E.M. Stein, H p spaces of several variables, Acta Math., 129 (1972), no. 3-4,
137–193.
53. R. Fefferman and F. Soria, The space Weak H 1 , Studia Math., 85 (1987), 1–16.
54. A. Fiorenza and M. Krbec, Indices of Orlicz spaces and some applications, Comment. Math.
Univ. Carolin., 38 (1997), no. 3, 433–451.
55. G.B. Folland, Real Analysis, Modern Techniques and Their Applications, 2-nd edition, John
Wiley and Sons, Inc., New York, 1999.
56. G.B. Folland and E.M. Stein, Hardy Spaces on Homogeneous Groups, Mathematical Notes,
Vol. 28, Princeton University Press, Princeton, N.J., 1982.
57. M. Frazier and B. Jawerth, Decomposition of Besov spaces, Indiana Univ. Math. J., 34 (1985),
777–799.
58. M. Frazier and B. Jawerth, The φ-transform and applications to distribution spaces, Function
spaces and applications, pp. 223–246 in Lecture Notes in Math., Vol. 1302, Springer, Berlin,
1988.
59. M. Frazier and B. Jawerth, A discrete transform and decompositions of distribution spaces, J.
Funct. Anal., 93 (1990), no. 1, 34–170.
60. M. Frazier, B. Jawerth and G. Weiss, Littlewood-Paley Theory and the Study of Function
Spaces, CBMS Regional Conference Series in Mathematics, Vol. 79, AMS, Providence, RI,
1991.
61. B. Fuglede, On generalized potentials of functions in the Lebesgue classes, Math. Scandinav-
ica, 8 (1960), 287–304.
62. J. García-Cuerva and J.L. Rubio de Francia, Weighted Norm Inequalities, North Holland,
Mathematics Studies Vol. 116, 1985.
63. M. Giaquinta, Multiple Integrals in the Calculus of Variations and Nonlinear Elliptic Systems,
Annals of Mathematics Studies, Vol. 105, Princeton University Press, Princeton, NJ, 1983.
64. A. Gogatishvili, P. Koskela, and Y. Zhou, Characterizations of Besov and Triebel-Lizorkin
spaces on metric measure spaces, Forum Mathematicum, 25 (2013), no. 4, 787–819.
65. I. Gohberg, S. Goldberg, and M.A. Kaashoek, Classes of Linear Operators, Vol. I, Operator
Theory: Advances and Applications, Vol. 49, Birkhäuser, Basel, 1990.
66. I. Gohberg and N. Krupnik, One-Dimensional Linear Singular Integral Equations, Vol. I-II,
Operator Theory: Advances and Applications Vol. 53-54, Birkhäuser Verlag Basel, 1992.
67. D. Goldberg, A local version of real Hardy spaces, Duke Math. J., 46 (1979), 27–42.
68. V. Gol’dshtein and A. Ukhlov, Weighted Sobolev spaces and embedding theorems, Trans.
Amer. Math. Soc., 361 (2009), no. 7, 3829–3850.
69. M. Gomez and M. Milman, Complex interpolation of H p -spaces on product domains, Ann.
Math. Pura Appl., 155 (1989), 103–115.
70. L. Grafakos, Classical and Modern Fourier Analysis, Pearson Education, Inc., New Jersey,
2004.
71. L. Grafakos, L. Liu, and D. Yang, Maximal function characterizations of Hardy spaces on RD-
spaces and their applications, Science in China Series A: Mathematics, 51 (2008), no. 12,
2253–2284.
908 References

72. L. Grafakos, L. Liu and D. Yang, Vector-valued singular integrals and maximal functions on
spaces of homogeneous type, Math. Scand., 104 (2009), no. 2, 296–310.
73. G. Green, An Essay on the Application of Mathematical Analysis to the Theories of Electricity
and Magnetism, Nottingham, England, T. Wheelhouse, 1828.
74. P. Grisvard, Elliptic problems in nonsmooth domains, Monographs and Studies in Mathemat-
ics, Vol. 24 Pitman Boston, MA, 1985.
75. R.F. Gundy and R.L. Wheeden, Weighted integral inequalities for the nontangential maximal
function, Lusin area integral and Walsh-Paley series, Studia Math., 49 (1974), 107–124.
76. M. Grüter and K.-O., Widman, The Green function for uniformly elliptic equations,
Manuscripta Math., 37 (1982), no. 3, 303–342.
77. N.M. Günter, Potential Theory and its Applications to Basic Problems of Mathematical
Physics, F. Ungar, 1967.
78. G. Hardy and J. Littlewood, Some properties of conjugate functions, J. Reine Angew. Math.,
167 (1932), 405–423.
79. P. Hajłasz, Sobolev spaces on an arbitrary metric space, Potential Anal., 5 (1996), no. 4,
403–415.
80. P. Hajłasz, Sobolev spaces on metric-measure spaces, pp. 173–218 in Contemp. Math.,
Vol. 338, Amer. Math. Soc., Providence, RI, 2003.
81. P. Hajłasz and P. Koskela, Sobolev Met Poincaré, Memoirs of the American Mathematical
Society, No. 688, 2000.
82. Yanchang Han, T 1 Theorems for inhomogeneous Besov and Triebel-Lizorkin spaces over
space of homogeneous type, Vietnam Journal of Mathematics, 36 (2008), no. 2, 125–136.
83. Y. Han, Calderón-type reproducing formula and the Tb theorem, Rev. Mat. Iberoamericana,
10 (1994), 51–91.
84. Y. Han, Inhomogeneous Calderón reproducing formula on spaces of homogeneous type, J.
Geom. Anal., 7 (1997), 259–284.
85. Y.-S. Han and C.-C. Lin, Embedding theorem on spaces of homogeneous type, J. Fourier
Anal., Appl. 8 (2002), no. 3, 291–307.
86. Y. Han, S. Lu, and D. Yang, Inhomogeneous Besov and Triebel-Lizorkin spaces on spaces of
homogeneous type, Approx. Theory Appl. (N.S.), 15 (1999), no. 3, 37–65.
87. Y. Han, S. Lu, and D. Yang, Inhomogeneous discrete Calderón reproducing formulas for
spaces of homogeneous type, J. Fourier Anal. Appl., 7 (2001), 571–600.
88. Y. Han, D. Müller, and D. Yang, Littlewood-Paley characterizations for Hardy spaces on
spaces of homogeneous type, Math. Nachr., 279 (2006), no. 13–14, 1505–1537.
89. Y. Han, D. Müller, and D. Yang, A theory of Besov and Triebel-Lizorkin spaces on metric
measure spaces modeled on Carnot-Carathéodory spaces, Abstr. Appl. Anal., 2008, 1–250.
90. Y.-S. Han and E.T. Sawyer, Littlewood-Paley Theory on Spaces of Homogeneous Type and
the Classical Function Spaces, Mem. Amer. Math. Soc., Vol. 110, No. 530, 1994.
91. Y.-S. Han and D. Yang, New characterizations and applications of inhomogeneous Besov
and Triebel-Lizorkin spaces on homogeneous type spaces and fractals, Dissertationes Math.
(Rozprawy Mat.), Vol. 403, 2002.
92. Y. Han and D. Yang, Some new spaces of Besov and Triebel-Lizorkin type on homogeneous
spaces, Studia Math., 156(1) (2003), 67–97.
93. W. Hansen, Uniform boundary Harnack principle and generalized triangle property, J. Funct.
Anal., 226 (2005), no. 2, 452–484.
94. E. Hille and R.S. Phillips, Functional Analysis and Semigroups, American Mathematical
Society Colloquium Publications, Vol. XXXI, Providence, 1957. (Third printing of Revised
Edition, 1974).
95. S. Hofmann, D. Mitrea, M. Mitrea, and A.J. Morris, L p -Square Function Estimates on
Spaces of Homogeneous Type and on Uniformly Rectifiable Sets, Memoirs of the Ameri-
can Mathematical Society, Vol. 245, No. 1159, 2017.
96. S. Hofmann, M. Mitrea, and M. Taylor, Geometric and transformational properties of Lips-
chitz domains, Semmes-Kenig-Toro domains, and other classes of finite perimeter domains,
J. Geom. Anal., 17 (2007), no. 4, 593–647.
References 909

97. S. Hofmann, M. Mitrea, and M. Taylor, Singular Integrals and Elliptic Boundary Problems on
Regular Semmes-Kenig-Toro Domains, Int. Math. Res. Not. IMRN (2010), no. 14, 2567–2865.
98. T. Holmstedt, Interpolation of quasi-normed spaces, Math. Scand., 26 (1970), 177–199.
99. Ha Duy Hung and Luong Dang Ky, On weak∗ -convergence in the Hardy space H 1 over
spaces of homogeneous type, arXiv:1510.01019 [math.CA] (2015).
100. G. Hu, D. Yang, and Y. Zhou, Boundedness of singular integrals in Hardy spaces on spaces
of homogeneous type, Taiwanese Journal of Mathematics, 133 (2009), no. 1, 91–135.
101. B. Jawerth, Some observations on Besov and Lizorkin-Triebel spaces, Math. Scand., 40 (1977),
no. 1, 94–104.
102. D. Jerison and C.E. Kenig, The inhomogeneous Dirichlet problem in Lipschitz domains, J.
Funct. Anal., 130 (1995), 161–219.
103. P.W. Jones, Quasiconformal mappings and extendability of functions in Sobolev spaces, Acta
Math., 47 (1981), 71–88.
104. P.W. Jones and J.-L. Journé, On weak convergence in H 1 (R d ), Proc. Amer. Math. Soc., 120
(1994), no. 1, 137–138.
105. A. Jonsson and H. Wallin, A Whitney extension theorem in L p and Besov spaces, Ann. Inst.
Fourier, Grenoble, 28, (1978), no. 1, 139–192.
106. A. Jonsson and H. Wallin, Function Spaces on Subsets of R n , Math. Rep., Vol. 2, No. 1, 1984.
107. N.J. Kalton, Convexity conditions for non-locally convex lattices, Glasgow Math. J., 25 (1984),
141–152.
108. N.J. Kalton, Analytic functions in non-locally convex spaces, Studia Math., 83 (1986), no. 3,
275–303.
109. N.J. Kalton, Plurisubharmonic functions on quasi-Banach spaces, Studia Math., 84 (1986),
297–324.
110. N. Kalton, S. Mayboroda, and M. Mitrea, Interpolation of Hardy-Sobolev-Besov-Triebel-
Lizorkin spaces and applications to problems in partial differential equations, pp. 121–177 in
“Interpolation Theory and Applications,” AMS Contemporary Mathematics, Vol. 445, 2007.
111. N. Kalton and M. Mitrea, Stability of Fredholm properties on interpolation scales of quasi-
Banach spaces and applications, Trans. Amer. Math. Soc., 350, (1998), 3837–3901.
112. N.J. Kalton, N.T. Peck and J.W. Roberts, An F-space Sampler, London Math. Society Lecture
Notes Series, Vol. 89, Cambridge University Press, Cambridge, 1984.
113. C.E. Kenig, Harmonic Analysis Techniques for Second Order Elliptic Boundary Value Prob-
lems, CBMS Regional Conference Series in Mathematics, Vol. 83, AMS, Providence, RI,
1994.
114. J. Kinnunen, The Hardy-Littlewood maximal function of a Sobolev function, Israel Journal of
Mathematics, 100 (1997), 117–124.
115. G. Köthe, Topological Vector Spaces, Vol. I, Springer-Verlag, New York, 1969.
116. M.A. Krasnoselskiı̆ and Ya.B. Rutitskiı̆, Convex Functions and Orlicz Spaces, Noordhof,
Groningen, 1961; English transl. from the first Russian edition Gos. Izd. Fiz. Mat. Lit.,
Moskva, 1958.
117. A. Kufner, O. John, and S. Fucik, Function Spaces, De Gruyter, 2012.
118. R.H. Latter, The atomic decomposition of Hardy spaces, pp. 275–279 in Harmonic Analysis
in Euclidean Spaces (Proc. Sympos. Pure Math., Williams Coll., Williamstown, Mass.), Part
1, Proc. Sympos. Pure Math., Vol. XXXV, Amer. Math. Soc., Providence, R.I., 1979.
119. J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces, Springer-Verlag, Berlin Heidelberg,
New York, 1977.
120. J.-L. Lions and E. Magenes, Problèmes aux Limites non Homogènes et Applications, Vol. 1,
Travaux et Recherches Mathématiques, No. 17 Dunod, Paris 1968.
121. R.A. Macías and C. Segovia, A Decomposition into atoms of distributions on spaces of ho-
mogeneous type, Adv. in Math., 33 (1979), no. 3, 271–309.
122. L. Maligranda, Orlicz spaces and interpolation, Vol. 5 of “Seminars in Mathematics,”
IMECC, Universidad Estadual de Campinas, Campinas, Brazil, 1989.
123. J.J. Marín, J. María Martell, D. Mitrea, I. Mitrea, and M. Mitrea, Singular Integrals, Quan-
titative Flatness, and Boundary Problems, Progress in Mathematics, Vol. 344, Birkhäuser,
2022.
910 References

124. J.M. Martell, D. Mitrea, I. Mitrea, and M. Mitrea, The BMO-Dirichlet problem for elliptic
systems in the upper half-space and quantitative characterizations of VMO, Analysis and
PDE, 12 (2019), no. 3, 605–720.
125. V. Maz’ya, Sobolev Spaces with Applications to Elliptic Partial Differential Equations, 2nd,
revised and augmented edition, Grundlehren der mathematischen Wissenschaften, Vol. 342,
Springer, 2011.
126. V. Maz’ya, M. Mitrea, and T. Shaposhnikova, The Dirichlet problem in Lipschitz domains for
higher order elliptic systems with rough coefficients, Journal d’Analyse Mathématique, 110
(2010), no. 1, 167–239.
127. O. Mendez and M. Mitrea, The Banach envelope of Besov and Triebel-Lizorkin spaces and
applications, J. Four. Anal. Appl., 6 (2000), 503–531.
128. S.G. Mikhlin and S. Prössdorf, Singular Integral Operators, Springer, Berlin, 1986.
129. N. Miller, Weighted Sobolev spaces and pseudodifferential operators with smooth symbols,
Trans. Amer. Math. Soc., 269, (1982), no. 1, 91–109.
130. D. Mitrea, Distributions, Partial Differential Equations, and Harmonic Analysis, Second
edition, Springer Nature Switzerland, 2018.
131. D. Mitrea, I. Mitrea, and M. Mitrea, On the Geometry of Sets Satisfying Uniform Ball Condi-
tions, preprint, (2014).
132. D. Mitrea, I. Mitrea, and M. Mitrea, A sharp divergence theorem with nontangential traces,
Notices of the American Mathematical Society, 67 (2020), no. 9, 1295–1305.
133. D. Mitrea, I. Mitrea, and M. Mitrea, Geometric Harmonic Analysis I: A Sharp Divergence The-
orem with Nontangential Pointwise Traces, Developments in Mathematics Vol. 72, Springer
Nature, Switzerland, 2022.
134. D. Mitrea, I. Mitrea, and M. Mitrea, Geometric Harmonic Analysis II: Function Spaces Mea-
suring Size and Smoothness on Rough Sets, Developments in Mathematics Vol. 73, Springer
Nature, Switzerland, 2022.
135. D. Mitrea, I. Mitrea, and M. Mitrea, Geometric Harmonic Analysis III: Integral Representa-
tions, Calderón-Zygmund Theory, Fatou Theorems, and Applications to Scattering, Develop-
ments in Mathematics Vol. 74, Springer Nature, Switzerland, 2022.
136. D. Mitrea, I. Mitrea, and M. Mitrea, Geometric Harmonic Analysis IV: Boundary Layer Poten-
tials in Uniformly Rectifiable Domains, and Applications to Complex Analysis, Developments
in Mathematics Vol. 75, Springer Nature, Switzerland, 2022.
137. D. Mitrea, I. Mitrea, and M. Mitrea, Geometric Harmonic Analysis V: Fredholm Theory and
Finer Estimates for Integral Operators, with Applications to Boundary Problems, Develop-
ments in Mathematics Vol. 76, Springer Nature, Switzerland, 2022.
138. D. Mitrea, I. Mitrea, M. Mitrea, and S. Monniaux, Groupoid Metrization Theory with Appli-
cations to Analysis on Quasi-Metric Spaces and Functional Analysis, Birkhäuser, 2013.
139. D. Mitrea, I. Mitrea, M. Mitrea, and M. Taylor, The Hodge-Laplacian: Boundary Value Prob-
lems on Riemannian Manifolds, Studies in Mathematics, Vol. 64, De Gruyter, 2016.
140. D. Mitrea, I. Mitrea, M. Mitrea, and E. Ziadé, Abstract capacitary estimates and the complete-
ness and separability of certain classes of non-locally convex topological vector spaces, J.
Funct. Anal., 262 (2012), no. 11, 4766–4830.
141. I. Mitrea and M. Mitrea, Multi-Layer Potentials and Boundary Problems for Higher-Order
Elliptic Systems in Lipschitz Domains, Lecture Notes in Mathematics, Vol. 2063, Springer,
Berlin, 2013.
142. I. Mitrea and M. Mitrea, Fredholm Theory in Topological Vector Spaces, book manuscript,
2020.
143. M. Mitrea and M. Taylor, Potential theory on Lipschitz domains in Riemannian manifolds:
L p , Hardy, and Hölder space results, Communications in Analysis and Geometry, 9 (2001),
no. 2, 369–421.
144. M. Mitrea and M. Wright, Boundary Value Problems for the Stokes System in Arbitrary
Lipschitz Domains, Astérisque, Soc. Math. de France, Vol. 344, 2012.
145. A. Miyachi, H p spaces over open subsets of R n , Studia Math., 95 (1990), no. 3, 205–228.
146. A. Miyachi, Hardy-Sobolev spaces and maximal functions, J. Math. Soc. Japan, 42 (1990),
no. 1, 73–90.
References 911

147. A. Miyachi, Extension theorems for real variable Hardy and Hardy-Sobolev spaces, Harmonic
Analysis (Sendai, 1990), pp. 170–182, ICM-90 Satell. Conf. Proc., Springer, Tokyo, 1991.
148. A. Miyachi, Extension theorems for the function spaces of DeVore and Sharpley, Math. Japon.,
38 (1993), 1033–1049.
149. A. Miyachi, Atomic decompositions for Sobolev spaces and for the C pα spaces on general
domains, Tsukura J. Math., 21 (1997), no. 1, 59–96.
150. A. Miyachi, On the extension properties of Triebel-Lizorkin spaces, Hokkaido Mathematical
Journal, 27 (1998), 273–301.
151. A. Mukherjee, Atiyah-Singer Index Theorem. An Introduction, Hindustan Book Agency, New
Delhi, 2013.
152. D. Müller and D. Yang, A difference characterization of Besov and Triebel-Lizorkin spaces
on RD-spaces, Forum Math., 21 (2009), 259–298.
153. V. Müller, Spectral Theory of Linear Operators and Spectral Systems in Banach Algebras,
2-nd edition, Operator Theory: Advances and Applications, Birkhäuser, 2007.
154. P. Nilsson, Reiteration theorems for real interpolation and approximation spaces, Ann. Mat.
Pura ed Appl., 132 (1982), 291–330.
155. J. Peetre, On the theory of L p, λ spaces, J. Funct. Anal., 4 (1969), 71–87.
156. J. Radon, Über lineare Funktionaltransformationen und Funktionalgleichungen, Sitzungs-
berichte Akad. Wiss., Abt. 2a, Wien, 128 (1919), 1083–1121.
157. H. Rafeiro, N. Samko, and S. Samko, Morrey-Campanato spaces: an overview, pp. 293–323
in Operator Theory: Advances and Applications, Vol. 228, Springer, Basel, 2013.
158. A.W. Roberts and D.E. Varberg, Another proof that convex functions are locally Lipschitz,
The American Mathematical Monthly, 81 (1974), no. 9, 1014–1016.
159. P. Rocha, A note on Hardy spaces and bounded operators, Georgian Mathematical Journal,
25 (2018), no. 1, 73–76.
160. R.T. Rockafellar,Convex Analysis, Princeton Landmarks in Mathematics, Princeton Univer-
sity Press, Princeton, New Jersey, 1997.
161. L.G. Rogers, Degree-independent Sobolev extension on locally uniform domains, J. Funct.
Anal., 235 (2006), 619–665.
162. S. Rolewicz, On a certain class of linear metric spaces, Bull. Acad. Polon. Sci. Sér. Sci. Math.
Astronom. Phys., 5 (1957), 471–473.
163. M. Rosenthal and H. Triebel, Morrey spaces, their duals and preduals, Rev. Mat. Complut.,
28 (2015), 1–30.
164. W. Rudin, Real and Complex Analysis, 3-rd edition, McGraw-Hill, Boston, Massachusetts,
1987.
165. W. Rudin, Functional Analysis, 2nd edition, McGraw-Hill, Inc., New York, 1991.
166. T. Runst and W. Sickel, Sobolev Spaces of Fractional Order, Nemytskij Operators, and Non-
linear Partial Differential Operators, de Gruyter, Berlin, New york, 1996.
167. V. Rychkov, On restrictions and extensions of the Besov and Triebel-Lizorkin spaces with
respect to Lipschitz domains, J. London Math. Soc. (2), 60 (1999), no. 1, 237–257.
168. K. Saka, The trace theorem for Triebel-Lizorkin spaces and Besov spaces on certain fractal
sets I - The restriction theorem, Memoirs of the College of Education Akita University
(Natural Science), 48 (1995), 1–17.
169. K. Saka, The trace theorem for Triebel-Lizorkin spaces and Besov spaces on certain fractal
sets II - The extension theorem, Memoirs of the College of Education Akita University (Natural
Science), 49 (1996), 1–27.
170. D. Sarason, Functions of vanishing mean oscillation, Trans. Amer. Math. Soc., 207 (1975),
391–405.
171. M. Schechter, Principles of Functional Analysis, 2-nd edition, Graduate Studies in Math.,
Vol. 36, Amer. Math. Soc., Providence, RI, 2002.
172. A. Seeger, A note on Triebel-Lizorkin spaces, Approximation and Function Spaces, pp. 391–
400, Banach Center Publ., No. 22, PWN, Warsaw, 1989.
173. S. Semmes, Hypersurfaces in R n whose unit normal has small BMO norm, Proc. Amer.
Math. Soc., 112 (1991), no. 2, 403–412.
912 References

174. W. Sickel and H. Triebel, Hölder inequalities and sharp embeddings in function spaces of
s
B p, s
q and Fp, q type, Z. Anal. Anwendungen, 14 (1995), no. 1, 105–140.
175. E.M. Stein, Singular Integrals and Differentiability Properties of Functions, Princeton Math-
ematical Series, No. 30, Princeton University Press, Princeton, NJ, 1970.
176. E.M. Stein, Harmonic Analysis: Real-Variable Methods, Orthogonality, and Oscillatory In-
tegrals, Princeton Mathematical Series, Vol. 43, Monographs in Harmonic Analysis, III,
Princeton University Press, Princeton, NJ, 1993.
177. E.M. Stein and G. Weiss, On the theory of harmonic functions of several variables, I. The
theory of H p spaces, Acta Math., 103 (1960), 25–62.
178. E.M. Stein and G. Weiss, Generalization of the Cauchy-Riemann equations and representa-
tions of the rotation group, Amer. J. Math., 90 (1968), no. 1., 163–196.
179. E.M. Stein and G. Weiss, Introduction to Fourier Analysis on Euclidean Spaces, Princeton
University Press, Princeton, N.J., 1971.
180. E. Straube, Interpolation between Sobolev and between Lipschitz spaces of analytic functions
on star-shaped domains, Trans. Amer. Math. Soc., 316 (1989), no. 2, 653–671.
181. M.H. Taibleson and G. Weiss, The molecular characterization of Hardy spaces, pp. 281–
287 in Harmonic Analysis in Euclidean Spaces (Proc. Sympos. Pure Math., Williams Coll.,
Williamstown, Mass., 1978), Part 1, Proc. Sympos. Pure Math., Vol. XXXV, Amer. Math.
Soc., Providence, R.I., 1979.
182. M.H. Taibleson and G. Weiss, The molecular characterization of certain Hardy spaces,
pp. 67–149 in Representation Theorems for Hardy Spaces, Astérisque, Vol. 77, Soc. Math.
France, Paris, 1980.
183. M. Taylor, Partial Differential Equations, Vol. I, Springer-Verlag, 1996 (2nd ed., 2011).
184. M.E. Taylor, Tools for PDE: Pseudodifferential Operators, Paradifferential Operators, and
Layer Potentials, Vol. 81, Math. Surveys and Monographs, AMS, Providence, R.I.,2000.
185. A. Torchinsky, Real-Variable Methods in Harmonic Analysis, Dover Publications, 2004.
186. H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, North-Holland,
Amsterdam, 1978 (2nd ed., Barth, Heidelberg, 1995).
187. H. Triebel, Theory of Function Spaces, Birkhäuser, Berlin, 1983.
188. H. Triebel, Theory of Function Spaces, II, Monographs in Mathematics, Vol. 84, Birkhäuser
Verlag, Basel, 1992.
189. H. Triebel, Interpolation Theory, Function Spaces, Differential Operators, 2-nd edition,
Heidelberg-Leipzig, Barth, 1995.
190. H. Triebel, Theory of function spaces, III, Monographs in Mathematics, Vol. 100, Birkhäuser
Verlag, Basel, 2006.
191. H. Triebel, Function Spaces and Wavelets on Domains, Tracts in Mathematics, Vol. 7, Euro-
pean Mathematical Society, Zürich, Switzerland, 2008.
192. B.O. Turesson, Nonlinear Potential Theory and Weighted Sobolev Spaces, Lecture Notes in
Mathematics, Vol. 1736, Springer-Verlag, 2000.
193. P. Turpin, Convexités dans les espaces vectoriels topologiques generaux, Dissertationes Math.,
Vol. 131, 1974.
194. S.V. Uspenskiı̆, Imbedding theorems for classes with weights, Trudy Mat. Inst. Steklov., 60
(1961), 282–303.
195. A. Uchiyama, A maximal function characterization of H p on the space of homogeneous type,
Trans. Amer. Math. Soc., 262 (1980), no. 2, 579–592.
196. N. Varopoulos, A remark on BMO and bounded harmonic functions, Pacific J. Math., 73
(1977), 257–259.
197. X. Wu and X. Wu, Weak Hardy spaces H p,∞ on spaces of homogeneous type and their
applications, Taiwanese J. Math., 16 (2012), no. 6, 2239–2258.
198. T. Xiangxing, Noncontinuous data boundary value problems for Schrödinger equation in
Lipschitz domains, Front. Math. China, 4 (2006), 589–603.
199. D. Yang, Frame characterizations of Besov and Triebel-Lizorkin spaces on spaces of homo-
geneous type and their applications, Georgian Math. J., 9 (2002), no. 3, 567–590.
200. D. Yang, Embedding theorems of Besov and Triebel-Lizorkin spaces on spaces of homoge-
neous type, Sci. China Ser. A, 46 (2003), 187–199.
References 913

201. D. Yang, New characterizations of Hajłasz-Sobolev spaces on metric spaces, Sci. China Ser.
A 46 (2003), no. 5, 675–689.
202. D. Yang, Real interpolations for Besov and Triebel-Lizorkin spaces on spaces of homogeneous
type, Math. Nachr., 273 (2004), 96–113.
203. D. Yang, Some new inhomogeneous Triebel-Lizorkin spaces on metric measure spaces and
their various characterizations, Studia Math., 167 (2005), 63–98.
204. D. Yang and Y. Lin Spaces of Lipschitz type on metric spaces their applications, Proceedings
of the Edinburgh Mathematical Society, 47 (2004), 709–752.
205. D. Yang and Y. Zhou, Radial maximal function characterizations of Hardy spaces on RD-
spaces and their applications, Math. Ann., 346 (2010), 307–333.
206. D. Yang and Y. Zhou, New properties of Besov and Triebel-Lizorkin spaces on RD-spaces,
Manuscripta Math., 134 (2011), 59–90.
207. S. Yano, Notes on fourier analysis (XXIX): an extrapolation theorem, J. Math. Soc. Japan, 3
(1951), no. 2, 296–305.
208. A.C. Zaanen, Integration, North-Holland Publishing Company, Amsterdam, 1967.
209. C.T. Zorko, Morrey space, Proc. Amer. Math. Soc., 98 (1986), no. 4, 586–592.
Subject Index

(ε, δ)-domain, 878 Banach function space (classical),

p-convex set, 417 219
p-envelope, 421 Besov space
p-norm, 5 homogeneous, 365
r-convexification, 47 in Rn , 518
s-norm, 59 inhomogeneous, 368
P-maximal function, 607 block
Pg -maximal function, 679 η-smooth of type (p, s), 374
H q,λ,θ -dome, 300 p, 371
B q,λ , 326
block-based homogeneous Sobolev
absolutely p-convex hull (of a set), space, 861
417 block-based negative Sobolev space,
absolutely p-convex set, 417 801
absolutely continuous norm, 59 boundary
admissible family of functions, 37 measure theoretic, 896
Ahlfors regular boundary-to-boundary
domain, 869 Cauchy-Clifford integral operator,
analytic function 874
one variable, 36 Boyd indices, 277
approximation to the identity bullet product, 894
(ATTI), 84, 115
associated space, 225 Calderón product, 48
atom Calderón’s maximal operator, 780
(p, q), 143 Cauchy sequence (in a topological
η-smooth of type (p, s), 373 group), 56
p, 371 Cauchy-Clifford integral operator
H q,λ , 297 boundary-to-boundary,
of type (K, L, J), 522 left-handed, 629
© The Editor(s) (if applicable) and The Author(s), under exclusive license to 915
Springer Nature Switzerland AG 2022
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1
916 SUBJECT INDEX

boundary-to-boundary, complementary, 261

right-handed, 630 strictly defined, 457
boundary-to-domain, left-handed, Young, 253
629 function norm, 218
boundary-to-domain,
right-handed, 630 Generalized Banach Function Space,
truncated, 874 219
charge, 166 grand maximal function, 532
compatible quasi-normed spaces, 25
complementary function, 261 Hardy inequalities, 132
complete topological group, 56 Hardy space
complex interpolation atomic, 143
intermediate space, 48 ionic characterization of, 166
quasi-norm, 48 Lebesgue-based, 114
conditional expectation operators, local, 371
382 Lorentz-based, 117
conormal derivative from weighted Hardy-based
Sobolev spaces, 488 homogeneous Sobolev space, 816
countably simple (function), 39 inhomogeneous Sobolev space,
818
dilation indices of a Young function, harmonic double layer
255 (boundary-to-domain), 730
distribution Hodge star operator, 900
on an arbitrary set, 887
homogeneous
distributional tangential derivatives,
Besov space, 365
703
Hardy space, 115
domain
Triebel-Lizorkin space, 365
(ε, δ), 878
homogeneous Sobolev space
Ahlfors regular, 869
block-based, 861
doubling Young function, 254
Hardy-based, 816
essential norm, 14 Morrey-based, 848
extension operator, 536
exterior derivative operator, 877 index function, 64, 74
extrapolation theorem indices
for Generalized Banach Function Boyd, 277
Spaces, 233 dilation, 255
inhomogeneous Besov space, 368
Fefferman-Stein inhomogeneous Sobolev space
grand maximal function, 108 Hardy-based, 818
filtering operator, 212 inhomogeneous Triebel-Lizorkin
fractional Hajłasz-Sobolev space, space, 368
824 integration by parts on the boundary
Fredholm radius, 77 both functions are smooth, 683
function both functions belong to Sobolev
analytic (one variable), 36 spaces, 779
SUBJECT INDEX 917

both functions in Besov spaces, (q, λ)-dull, 359

846 H p , 174
one Hardy-based Sobolev H q,λ,θ , 301
function, one smooth function, Morrey space, 314
812, 814, 816 Morrey-based
one Lebesgue function, one homogeneous Sobolev space, 848
Sobolev function, 799 negative Sobolev space, 801
one Sobolev function, one smooth Sobolev space, 793
function, 685 Morrey-Campanato space
interpolation space homogeneous, 295
inner complex, 48 inhomogeneous, 297
outer complex, 48 Muckenhoupt weighted
ion, 166 negative Sobolev space, 801
Sobolev space, 793
Jones extension operator, 470
nontangential
lattice of functions maximal operator, 893
r-convex, 46 pull-back, 884
local (boundary) Sobolev space, 684 nontangential maximal function
local Hardy space, 371 truncated, 893
local John condition, 886 norm
locally p-convex (topological vector Luxemburg, on an Orlicz space,
space), 417 256
Luxemburg norm on an Orlicz space,
256 operator
L p -filtering, 212
maximal “altered” Cauchy integral bounded, 3, 61
operator, 874 finite-rank, 2
maximal Cauchy-Clifford integral Fredholm on Banach spaces, 63
operator, 874 Fredholm on TVS, 73
maximal Riesz transform, 898 quasi-subadditive, 33
measure semi-Fredholm, 64
feeble, 56 Orlicz space, 256
push-forward, 785
support of, 900 pointwise variation, 902
measure theoretic exterior, 879 Pompeiu-Hausdorff distance, 877
measure theoretic interior, 885 principal symbol, 622
modified (boundary-to-boundary) principal symbol of a first-order
harmonic double layer, 733 system, 900
modified (boundary-to-domain) pseudo-quasi-Banach space, 59
harmonic double layer, 732 pseudo-quasi-norm, 59
modulus of concavity, 5 pseudo-quasi-normed space, 59
molecule push-forward of a measure, 785
(η, ε)-smooth of type (p, s), 374
(p, q, ε), 174 quasi-Banach analytically convex, 38
(p, qo, ε, s)-rough, 375 quasi-isometry, 347
918 SUBJECT INDEX

quasi-normed lattice of functions, 33 p-Banach, 5

quotient topology, 9 associated, 225
atomic Hardy, 143
radial maximal function, 532 block, 326
real interpolation Hardy (Lebesgue-based), 114
intermediate space, 29 Hardy (Lorentz-based), 117
quasi-norm, 29 homogeneous Besov, 365
rearrangement invariant Banach homogeneous Hardy, 115
function space, 257 homogeneous Morrey-Campanato,
retract, 52 295
Riemannian metric tensor, 679 homogeneous Triebel-Lizorkin,
Riesz transform 365
boundary-to-boundary, 899 inhomogeneous Besov, 368
in Rn , 633 inhomogeneous
truncated, 899 Morrey-Campanato, 297
Sarason space VMO, 82 inhomogeneous Triebel-Lizorkin,
simple functions with support of 368
finite measure, 900 local Hardy, 371
Sobolev space local off-diagonal (boundary)
L p -based homogeneous, 749 Sobolev, 684
block-based, 795 local BMO, 372
block-based homogeneous, 861 locally bounded, 3
homogeneous, 749 Lorentz-based Sobolev, 792
vanishing Morrey-based, 865 Morrey, 314
local, 684 Morrey-based Sobolev, 793
local off-diagonal (boundary), 684 Muckenhoupt weighted Sobolev,
Lorentz-based, 792, 793 793
Morrey-based homogeneous, 848 negative Sobolev, 796
negative of functions of bounded variation,
block-based, 801 871
definition, 796 of functions of locally bounded
Morrey-based, 801 variation, 871
Muckenhoupt weighted, 801 off-diagonal Morrey-based
off-diagonal, 800 Sobolev, 794
vanishing Morrey-based, 803 off-diagonal negative Sobolev, 800
off-diagonal (boundary), 684 Orlicz, 256
off-diagonal Lorentz-based, 792 pseudo-quasi-Banach, 59
off-diagonal Morrey-based, 794 pseudo-quasi-normed, 59
vanishing Morrey-based, 794 quasi-Banach, 5
weighted, 792 rearrangement invariant, 257
space Sarason, VMO, 82
(q, λ)-midway, 360 weighted Bessel potential, 457
L p -based Sobolev, 684 weighted maximal Sobolev, 498
L p -based homogeneous Sobolev, weighted Sobolev, 792
749 Zygmund, 274
SUBJECT INDEX 919

spectral radius Cauchy-Clifford integral operator,

definition, 76 874
standard fundamental solution for the Riesz transform, 899
Laplacian, 879
strictly defined version of a function, unit, (η, ε)-smooth of type (p, s), 374
457
strongly μ-measurable (function), 39 vanishing Hölder space, 98
support of a measurable function, vanishing mean oscillations, 82
901
weak conormal derivative
surface-to-surface change of variable
definition, 658
formula, 785
weak normal derivative, 667
system of auxiliary function, 879
weak tangential derivative, 896
weakly compatible (TVS), 61
tangential derivative
weighted
distributional, 703
L p Lebesgue space over Ω, 467
weak, 896
maximal Sobolev space, 498
tangential derivative operator, 896
Sobolev space, 792
in a pointwise sense, 682 Sobolev space in Ω, for the weight
tangential gradient, 728 ap
w := δ∂Ω , 472
test functions Sobolev spaces in Rn , 456
of type (x0, r, β, γ), 363 Whitney
topologically bounded set, 2 decomposition lemma, 469
Triebel-Lizorkin space
homogeneous, 365 Young function, 253
in Rn , 518 doubling, 254
inhomogeneous, 368
truncated Zygmund’s space, 274
Symbol Index

∗ Hodge star operator, 900 u · w = u, w dot product of two

∧ exterior product of differential vectors u, w ∈ Cn , 884
forms, 895 u · w = u, w dot product of two
∨ interior product of differential vectors u, w ∈ Rn , 877
forms, 895  ∞ contribution of F at infinity,
[F]
Δ Laplace operator, 877 879
∇u gradient (Jacobian matrix) of u, · X→Y operator norm, 6
ess
881 · X→Y essential norm, 14
∇ gradient operator in Rn−1 , 881 · p Minkowski functional of
∇tan tangential gradient, 728 & '
BX (0, 1) p , 417
Δ = Δ(x, r) surface ball, 878
· (X0,X1 )θ, q real interpolation
UV symmetric difference of U and quasi-norm, 29
V, 878
· [X0,X1 ]θ complex interpolation
U # V the union of two disjoint sets
quasi-norm, 48
U, V, 902
(a)+ := max{a, 0}, 897
D (Ω) ·, · D(Ω) distributional pairing
in Ω, 877 X/Y quotient space, 1
·, · [x]X/Y equivalence class of x ∈ X in
X ∗ X , 904
X/Y , 1
(Lip c (Σ)) ·, · Lip c (Σ) (or simply ·, · )
1E characteristic function of E, 895
& ' distributional pairing, 887
·, · E pointwise (real) pairing in the ⨏fΔ integral⨏ average of f in Δ, 758
fibers of Hermitian vector E
f dμ, E f dμ integral average of
bundle E, 895 f on E, 885
·, · Λ T M (real) pointwise pairing on U closure of the set U, 873
Λ T M, 895 V ⊥ annihilator of a subspace V of a
Lip c (∂Ω) ·, · Lip c (∂Ω) pairing between Banach space X, 66
Lipc (∂Ω) and its dual ⊥ W annihilator of a subspace W of

Lipc (∂Ω) , 899 X ∗ , where X is Banach, 66

© The Editor(s) (if applicable) and The Author(s), under exclusive license to 921
Springer Nature Switzerland AG 2022
D. Mitrea et al., Geometric Harmonic Analysis II, Developments in Mathematics 73,
https://doi.org/10.1007/978-3-031-13718-1
922 SYMBOL INDEX
√
i = −1 ∈ C complex imaginary ωn−1 surface area of S n−1 , 895
unit, 884, 900 ρinv (T; X) spectral radius of
[A; B] := [A, B] := AB − BA the T ∈ Bd (X), 76
commutator of A and B, 871 ρFred (T; X) Fredholm (or essential
{ A; B} := AB + BA the spectral) radius of T ∈ Bd (X),
anti-commutator of A and B, 77
871 ρsym the symmetrized version of ρ,
d exterior derivative operator, 877 901
δ formal adjoint of the exterior ρ# the regularized version of ρ, 901
derivative operator d, 877 σg surface measure induced by the
δ jk Kronecker symbol, 877 metric tensor g, 900
δx Dirac distribution with mass at x, σ E surface measure induced by the
877 standard Euclidean metric, 900
δF distance function to the set F, 877 σ∗ = H n−1 ∂∗ Ω surface measure,
δ∂Ω (·) distance function to the 900
boundary, 878 σ = H n−1 ∂Ω, the surface measure
εBA generalized Kronecker symbol, on ∂Ω, 900
879 ∂nta Ω nontangentially accessible
Φ(X → Y ) Fredholm operators from boundary of Ω, 897
X into Y , 63, 74 ∂lfp Ω, 897
Φ+ (X → Y ) finite-dim kernel
∂∗ E measure theoretic boundary of
semi-Fredholm operators from
E, 896
X into Y , 64
∂ ∗ E reduced boundary of E, 896
Φ− (X → Y ) finite-dim cokernel
∂T E, 896
semi-Fredholm operators from
X into Y , 64 ∂ N E, 896
Γκ (x) nontangential approach region, ∂νA conormal derivative operator
881 with respect to the coefficient
λQ concentric dilate of the cube Q tensor A acting from Besov and
by the factor λ, 469 Triebel-Lizorkin spaces, 594
Λk Jones extension operator, 470 ∂νA(·, ·) conormal derivative operator
Λ∗,q Calderón’s maximal operator, with respect to the coefficient
780 tensor A acting from weighted
Λ T M the -th exterior power of the . A Sobolev spaces, 488
vector bundle on M, 886 ∂ν weak conormal derivative
%
μ Cauchy-Clifford transform of the operator with respect to the
measure μ, 882 coefficient tensor A, 658
νg GMT unit normal induced by the ∂τ j k pointwise tangential derivative
metric tensor g, 894 operator, 682
ν E GMT unit normal induced by the ∂τ j k tangential derivative operator,
standard Euclidean metric, 894 . 685, 703
ν • F the bullet product of ν with F,
 ∂τ j k weak tangential derivative, 896
894 ∂τXY tangential derivative operator
ν • u the Clifford bullet product of ν on manifolds, 896
with u, 652 Πm  projection map onto Λ , 897
m
SYMBOL INDEX 923

πκ (E), πΩ,κ (E) “shadow” (or BMO(X, μ) space of functions of

projection) of E ⊆ Ω onto ∂Ω, bounded mean oscillations, 871
897 .
· BMO(X,μ) , homogeneous BMO
τt dilation by a factor of t, 902 semi-norm, 871
τρ topology induced by the · BMO(X,μ) inhomogeneous BMO
quasi-distance ρ, 902 “norm”, 871
Aκ (∂Ω) accessibility set, 869 bmo (Σ, σ) local BMO space on the
Ap (X, ρ, μ) Muckenhoupt class, 869 set Σ, 372
[w] A p characteristic of the · bmo (Σ,σ) local BMO semi-norm,
Muckenhoupt weight w, 869 372
A∞ (X, ρ, μ) Muckenhoupt class, 869 f ∗ (Δ) local BMO norm of f on Δ,
A p absolutely p-convex hull of the 870
set A, 417 
BMO(X, μ) the space BMO modulo
p,q constants, 871
As,z (Ω) restrictions to Ω of
p,q BV(O) space of functions of
distributions from As,0 (Ω),
528 bounded variation in O, 871
· Ap, q quasi-norm in BVloc (O) space of functions of
s, z (Ω)
Besov/Triebel-Lizorkin space locally bounded variation in O,
of restrictions to Ω of 871
p,q
distributions from As,0 (Ω), Bn−1 (x , r) open ball with center x
528 and radius r in Rn−1 , 870
p,q
As (Ω) Besov/Triebel-Lizorkin Bρ (x, r) ρ-ball with center at x and
space in Ω, 527 . p,q radius r, 870
· Ap, q quasi-norm in Bs (Σ, σ) homogeneous Besov
s (Ω)
Besov/Triebel-Lizorkin space space on the set Σ, 365
in Ω, 527 · B. p, q (Σ,σ) homogeneous Besov
s

Ås (Ω) closure of Cc∞ (Ω) in

p,q quasi-seminorm, 365
p,q
p,q
As (Ω), 527 Bs (Σ, σ) (inhomogeneous) Besov
p,q p,q
As,0 (Ω) distributions from As (Rn ) space on the set Σ, 368
· Bsp, q (Σ,σ) (inhomogeneous)
supported in Ω, 527 Besov quasi-norm, 368
· p, q quasi-norm in
A s,0 (Ω)
p,q
bs (Σ) discrete Besov space on Σ,
Besov/Triebel-Lizorkin space 377
p,q
of distributions from As (Rn ) · bsp, q (Σ) quasi-norm on discrete
supported in Ω, 527 Besov space, 377
A q,κ L q -based area-function, 565 bp,q (Σ), 376
B(X → Y ) linear and (topologically) · b p, q (Σ) , 376
p,q
bounded operators from X to Bs (Rn ) Besov space in Rn , 518
Y, 3 · Bsp, q (Rn ) quasi-norm in the Besov
Bd(X) linear and bounded operators space in Rn , 518
q,λ
on X, 7 B1 (∂Ω, σ) block-based Sobolev
Bd X → Y linear and (norm) . q,λ space, 795
bounded operators from X to B1 (∂Ω, σ) block-based
Y, 7 homogeneous Sobolev space,
BMO−1 (∂Ω, σ), 820 861
924 SYMBOL INDEX

B q,λ (Σ, σ) block space, 326 Cc∞ (Ω) restrictions to Ω of functions

· B q, λ (Σ,σ) norm on block space, in Cc∞ (Rn ), 528
∞ ∗
327 Cb (Ω) the algebraic dual of
· M q, λ (Σ,σ) norm on midway Cb∞ (Ω), 874
space, 360 .
q,λ C α (U, ρ) homogeneous Hölder
B−1 (∂Ω, σ) block-based negative space, 875
Sobolev space, 801
· C.α (U,ρ) homogeneous Hölder
Borelτ (X) Borelians of the
topological space (X, τ), 871 space semi-norm, 875
.α
BL(Σ) bounded Lipschitz functions C (U, ρ)/∼ homogeneous Hölder
on Σ, 871 .α
space modulo constants, 875
Bψ ( f , r), 56 Cloc (U, ρ) local homogeneous
CL left-handed boundary-to-domain Hölder space, 875
Cauchy-Clifford integral C α (U, ρ) inhomogeneous Hölder
operator, 629 space, 876
CR right-handed · C α (U,ρ) inhomogeneous Hölder
boundary-to-domain space norm, 875
Cauchy-Clifford integral α
Cc (U, ρ) Hölder functions with
operator, 630 ρ-bounded support, 876
Cmax maximal Cauchy-Clifford .γ
Cvan (Σ) homogeneous vanishing
integral operator, 874
alt Hölder space, 98
Cmax maximal “altered” Cauchy γ
Cvan (Σ) inhomogeneous vanishing
integral operator, 874
Hölder space, 99
Cε truncated Cauchy-Clifford
integral operator, 874 CBM(Ω) complex Borel measures in
C boundary-to-boundary Ω, 874
Cauchy-Clifford integral CBM(X, τ) complex Borel measures
operator, 874 in the topological space (X, τ),
C L left-handed 874
boundary-to-boundary CMO(Σ, σ), 184
p,q
Cauchy-Clifford integral Cα (Ω), 530
operator, 629 · Cαp, q (Ω) , 530
C R right-handed Cα
p,q,m
(Ω), 530
boundary-to-boundary p
Cα (Ω), 531
Cauchy-Clifford integral
C n Clifford algebra generated by n
operator, 630
imaginary units, 875
C k (Ω) functions of class C k in an
Cp(X → Y ) space of compact linear
open neighborhood of Ω, 873
operators from X into Y , 3
Cck (Ω) functions of class C k with
compact support in the open Cp(X) space of compact linear
set Ω, 873 operators from X into itself, 3
Cbk (Ω) bounded functions of class Cθ,b (x, h), 874
C k in Ω, 873 Cρ , 875
C -singsup u singular support of u,
k ρ , 875
C
900 D(X) dyadic grid on X, 878
SYMBOL INDEX 925

D harmonic double layer potential E (Ω) distributions compactly

operator supported in Ω, 879
(boundary-to-domain), 730 EK (Ω) distributions in Ω supported
Kmod boundary-to-boundary modified in K, 879
harmonic double layer E p (X) p-envelope of X, 421
potential operator, 733 Ex∂Ω→Ω extension operator from ∂Ω
Dmod boundary-to-domain modified to Ω, 481
harmonic double layer EΩ→Rn extension operator from Ω to
potential operator, 732 Rn , 536
Dk (X), 878 ext∗ (E) measure theoretic exterior of
D (Ω) space of distributions in Ω, E, 879

877 e j standard j-th unit vector in Rn ,
D = nj=1 e j  ∂j Dirac operator in 879
Rn , 877 {e j }1≤ j ≤n standard orthonormal
D first-order system, 877 basis in Rn , 879
D (real) transpose of the first-order
p,q n
Fs (R ) Triebel-Lizorkin space in
system D, 877 Rn , 518
D complex conjugate of the · Fsp, q (Rn ) quasi-norm in the
first-order system D, 877 Triebel-Lizorkin space in Rn ,
D∗ Hermitian adjoint of the 518
first-order system D, 877 f p,q (Σ), 377
D∗ (Σ), 373 · p, q , 377
. p,qf (Σ)
D L Dirac operator acting from the Fs (Σ, σ) homogeneous
left, 878 Triebel-Lizorkin space on the
DR Dirac operator acting from the set Σ, 365
right, 878 · F. p, q (Σ,σ) homogeneous
s
Dist [E, F] Pompeiu-Hausdorff Triebel-Lizorkin
distance between E and F, 877 quasi-seminorm, 365
diamρ (A) ρ-diameter of the set A, p,q
Fs (Σ, σ) (inhomogeneous)
878 Triebel-Lizorkin space on the
dg (x, y) geodesic distance between x set Σ, 368
and y, 877 · Fs q (Σ,σ) (inhomogeneous)
p,

divF the divergence of the vector Triebel-Lizorkin quasi-norm,

field F, 877 368
p,q
divg differential geometric fs (Σ) discrete Triebel-Lizorkin
divergence (associated with the space on Σ, 378
metric tensor g), 877 · fsp, q (Σ) quasi-norm on discrete
dim X dimension of X, 1 Triebel-Lizorkin space, 378
dVg volume element on M induced fBρ (x,r) integral average of f over
by the metric tensor g, 902 Bρ (x, r), 885
{Ek }k ∈Z, k ≥κΣ conditional fE∗ non-increasing rearrangement of
expectation operators, 382 f : E → R, 880
EΔ standard fundamental solution for [ f ]E strictly defined version of the
the Laplacian, 879 function f on the set E, 457
926 SYMBOL INDEX

fγ Fefferman-Stein grand maximal · H p, q (Σ,σ) quasi-norm on

function of f , 108 Lorentz-based Hardy space,
fp# L p -based Fefferman-Stein sharp 117
p,q
maximal function, 901 Hat (Σ, σ) atomic Hardy space, 143
p,q
Gα Bessel kernel of order α, 457 Hfin (Σ, σ) finite linear combinations
GΣ (x0, r, β, γ), 364 of atoms, 164
β,γ · H p, q (Σ,σ) quasi-norm on
G̊0 (Σ), 364 fin
p,q
β,γ
G0 (Σ) test functions on Σ, 364 Hfin (Σ, σ), 164
β,γ
Gb (Σ), 373 H (Σ, σ), 297
q,λ
 · H q, λ (Σ,σ) , 297
g = 1≤ j,k ≤n g jk dx j ⊗ dxk
h p (Σ, σ) local Hardy space, 371
Riemannian metric tensor, 881
 · h p (Σ,σ) quasi-norm on local
g = 1≤ j,k ≤n g jk dx j ⊗ dxk
Hardy space, 371
Riemannian metric tensor, 679
h p (Ω) local Hardy space in Ω, 532
H n−1 , the (n − 1)-dimensional
· h p (Ω) quasi-norm in the local
Hausdorff measure in Rn , 882
Hardy space in Ω, 532
Hg (n − 1)-dimensional Hausdorff
n−1
p
hk (Ω) local Hardy-based Sobolev
measure associated with the
space in Ω, 533
metric g, 882
IE,α fractional integral operator of
H s s-dimensional Hausdorff order α on E, 885
measure in Rn , 882 Im T : X → Y image of T : X → Y ,
H∗ s-dimensional Hausdorff
s
2
outer-measure in Rn , 882 index index function, 64, 74
H L -filtering operator, 212
p
int∗ (E) measure theoretic interior of
H s (Ω) L 2 -based fractional Sobolev E, 885
space in Ω, 530 i(Φ) lower dilation index of Φ, 255
· H s (Ω) norm in the L 2 -based I(Φ) upper dilation index of Φ, 255
fractional Sobolev space in Ω, ι∗ pull-back map induced by the
. p 530 canonical inclusion ι, 884
H1 (∂Ω, σ) Hardy-based ι∗# sharp pull-back, 884
homogeneous Sobolev space, ι∗n.t. nontangential pull-back, 884
. p 816  KΔ boundary-to-boundary harmonic
H1 (∂Ω, σ) ∼ classes of double layer potential, 886
equivalence, modulo
. p constants, KΔ# transpose harmonic double layer
of functions in H1 (∂Ω, σ), 817 potential, 886
p,q
H1 (∂Ω, σ) Hardy-based Ker T : X → Y kernel of
inhomogeneous Sobolev space, T : X → Y, 2
818 Ker L null-space of the system L, 499
H. p (Σ, σ) Hardy space, 114 L ρ Köthe function space associated
H p (Σ, σ) homogeneous Hardy with ρ, 55
space, 115 L(ξ) characteristic matrix of L, 888
· H p (Σ,σ) quasi-norm on Hardy L(X → Y ) linear and continuous
space, 114 operators from X to Y , 3
H p,q (Σ, σ) Lorentz-based Hardy L β (Σ), 143
space, 117 L n Lebesgue measure in Rn , 886
SYMBOL INDEX 927
p,q
Lgn measure associated with the L1 (∂Ω, σ) Lorentz-based Sobolev
n-form dVg , 886 space, 793
p
Lip(X) space of Lipschitz functions Ls (Ω) Bessel potential space in Ω,
on X, 886 529
Lipc (X) space of Lipschitz functions · Lsp (Ω) norm in the Bessel
with bounded support in X, potential space in Ω, 529
p
887 Lα (Rn, wL n ) weighted Bessel
Lipc (Σ) distributions on Σ, 887 potential space in Rn , 457
L 0 (X, μ) measurable functions which · Lαp (Rn,w L n ) norm on
p
are a.e. pointwise finite, 886 Lα (Rn, wL n ), 457
r (X, μ) L r -integrable functions on
Lfin L exp (logθ L) Orlicz space, 276
p
sets of finite μ-measure, 887 L Φ (X, μ) Orlicz space, 256
∞
Lcomp essentially bounded functions · L Φ (X,μ) Luxemburg norm on the
with compact support, 886 Orlicz space L Φ (X, μ), 256
p
Lbdd (Ω, L n ) p-th power integrable L (log L)α Zygmund’s space, 274
p
functions over bounded subsets p,q
L1 (∂∗ Ω, σ∗ ) off-diagonal
of Ω, 886
(boundary) Sobolev space, 684
L p (Ω, wL n ) weighted L p Lebesgue p,q
L1,loc (∂∗ Ω, σ∗ ) local off-diagonal
space over Ω, 467
p (boundary) Sobolev space, 684
L (Ω, μ) maximal Lebesgue space, p
L−1 (∂∗ Ω, σ∗ ) negative Sobolev
887
p space, 796
L1 (∂∗ Ω, σ∗ ) L p -based (boundary) p
L−1 (∂Ω, w) Muckenhoupt weighted
Sobolev space, 684
negative Sobolev space, 801
· L p (∂∗ Ω,σ∗ ) norm on Sobolev p,q
1 L−1 (∂∗ Ω, σ∗ ) off-diagonal negative
space, 693 Sobolev space, 800
p
L1,loc (∂∗ Ω, σ∗ ) local (boundary)
log+ positive ln, 888
Sobolev space, 684 .
p
L1 (∂∗ Ω, σg∗ ) ⊗ E global (boundary) L p,λ (Σ, σ) homogeneous
Sobolev space on manifolds, Morrey-Campanato space, 295
791 · L. p, λ (Σ,σ) Morrey-Campanato
p semi-norm, 295
L1,loc (∂∗ Ω, σg∗ ) ⊗ E local (boundary)
Sobolev space on manifolds, L (Σ, σ) inhomogeneous
p,λ

789 Morrey-Campanato space, 297

p
L1 (∂Ω, w) Muckenhoupt weighted · L p, λ (Σ,σ) Morrey-Campanato
(boundary) Sobolev space, 793 norm, 297
p
L1 (∂∗ Ω, wσ∗ ) weighted Sobolev (Q) side-length of the cube Q, 469
p
w (D∗ (Σ)), 397
. p space, 792
L1 (∂Ω, σ) homogeneous Sobolev · wp (D∗ (Σ)) , 397
space, 749 Mγ∗ (F) upper γ-dimensional
L p,q (X, μ) Lorentz space on X with Minkowski content of F, 890
respect to the measure μ, 887 M p,λ (Σ, σ) Morrey space, 314
· L p, q (X,μ) Lorentz space · M p, λ (Σ,σ) norm on Morrey space,
quasi-norm, 887 314
p,q
L (Ω, μ) maximal Lorentz space, M̊ p,λ (Σ, σ) vanishing Morrey space,
887 316
928 SYMBOL INDEX

Nκε the nontangential maximal

p,λ
M1 (∂Ω, σ) Morrey-based Sobolev
space, 793 function truncated at height ε,
p,q,λ
M1 (∂Ω, σ) off-diagonal 893
Morrey-based Sobolev space, Nκ,θ,r averaged nontangential
794 maximal function, 893
M̊1
p,q,λ
(∂Ω, σ) off-diagonal osc (u, x, t) oscillation of order m of
m

vanishing Morrey-based the function u at location x and

scale t, 525
. p,λ Sobolev space, 794 Oε one-sided collar neighborhood of
M1 (∂Ω, σ) homogeneous
Morrey-based Sobolev space, ∂Ω, 895
848 osc p ( f ; R) L p -based mean
p,λ oscillation of f at scales up to
M̊1 (∂Ω, σ) vanishing
Morrey-based Sobolev space, R, 895
pX lower Boyd index, 277
. p,λ 794 qX upper Boyd index, 277
M1 (∂Ω, σ) homogeneous
vanishing Morrey-based Φ ∈ Δ2 doubling Young function,
Sobolev space, 865 254
p,λ
M−1 (∂Ω, σ) Morrey-based negative P maximal function of Carleson
Sobolev space, 801 type, 607
M+ (X, μ) non-negative Pg maximal function of Carleson
μ-measurable functions on X, type on manifolds, 679
218 P.V. b k(x − ·)|Σ principal-value
M (X, μ) μ-measurable functions on distribution on the set Σ, 804
X, 218 R+n upper half-space in Rn , 898
R−n lower half-space in Rn , 898
M X Hardy-Littlewood maximal
operator on X, 891 R j boundary-to-boundary Riesz
transform, 899
M A,s,α fractional Hardy-Littlewood
R j,ε truncated Riesz transform, 899
maximal operator, 891
R j,max maximal Riesz transform, 898
M X,s L s -based Hardy-Littlewood
R Rn →Ω restriction operator from Rn
maximal operator, 891
to Ω, 528
M X,s
R local L s -based
RRn →∂Ω restriction operator from
Hardy-Littlewood maximal
Rn to ∂Ω, 898
operator, 891
RΩ→∂Ω higher-order restriction
M X,s,α fractional Hardy-Littlewood operator from Ω to ∂Ω, 573
maximal operator, 891 RHq (X, ρ, μ) reverse Hölder class,
mE (λ, f ), 891 898
M q,λ (Σ, σ) (q, λ)-mid space, 360 [w]RHq reverse Hölder constant of a
N0 = N ∪ {0}, 893 weight in RHq (X, ρ, μ), 898
p
Nκ (Ω; μ), 894 rad(Ω), 899
N Φ ( f ) modular size of f , 258 regsupp u regular support of a
Nκ nontangential maximal operator, distribution u ∈ D (Ω), 898
893 S n−1 unit sphere in Rn , 900
NκE the nontangential maximal S±n−1 upper/lower hemispheres of
operator restricted to E, 893 S n−1 , 900
SYMBOL INDEX 929

k, p
S(X, μ) simple functions on (X, μ), Wbdd (Ω), 903
900 k, p
Wloc (Ω) local L p -based Sobolev
Sfin (X, μ) simple functions on (X, μ) space of order k in Ω, 903
with support of finite measure, W (Ω, wL n ) weighted Sobolev
k, p
900 space in Ω, 468
Sym(D; ξ) principal symbol of the
· W k, p (Ω,w L n ) norm in weighted
first-order system D, 900
Sobolev space in Ω, 468
Sym(D; ν) • F bullet product of F
W k, p (Rn, wL n ) weighted Sobolev
with the principal symbol of
spaces in Rn , 456
the first-order system D, 651
S (Rn ) Schwartz functions, 901 · W k, p (Rn,w L n ) norm in the
S (Rn ) tempered distributions, 901 weighted Sobolev spaces in
supp μ support of the measure μ, 900 Rn , 456
k, p
supp f support of the measurable Wa (Ω) weighted Sobolev space in
ap
function f , 900 Ω, for the weight w := δ∂Ω , 472
T ∗ adjoint of T, 4 · W k, p (Ω) quasi-norm in the
a
TrΩ→∂Ω trace operator from Ω to weighted Sobolev space
k, p
∂Ω, 586 Wa (Ω), 472
1, p
TrRn →Σ trace operator from Rn to Σ, Wa (Ω)bdd , 475
W̊a (Ω) closure of Cc∞ (Ω) in
458 1, p
TrΩ→∂Ω trace operator from Ω to 1, p
Wa (Ω), 479
∂Ω, 473 k, p
Wa, (Ω) weighted maximal Sobolev
Tγ (x) bump functions centered at x,
space, 498
107
κ−n.t. · W k, p (Ω) quasi-norm in weighted
u|∂Ω (x) nontangential trace of u a,

at x ∈ ∂Ω, 894 maximal Sobolev space, 498

. k, p
uα# , 531 Wa, (Ω) homogeneous weighted
uΩ+ radial maximal function of u in maximal Sobolev space, 498
Ω, 532 · W. k, p (Ω) homogeneous weighted
∗ a,
uk,Ω grand maximal function of u in maximal Sobolev space, 498
Ω, 532 −1, p
Wa (Ω) weighted Sobolev space of
u,θ solid maximal function of u, 901
E local solid maximal function of order −1 in Ω, 485
u,θ
· W −1, p (Ω) norm on weighted
u, 901 a
Sobolev space of order −1 in
umax tangential maximal function of
M Ω, 485
u, 891
uscal scalar part of u, 901 W s, p (Σ, σ) fractional
uvect vector part of u, 903 Hajłasz-Sobolev space, 824
VarF pointwise variation of F, 902 X01−θ X1θ Calderón product, 48
V( f ; O) variation of f in O, 903 (X0, X1 )θ,q real interpolation
VMO(X, μ) space of functions of intermediate space, 29
vanishing mean oscillations, 82 [X0, X1 ]θ complex interpolation
VMO−1 (∂Ω, σ), 820 intermediate space, 48
W k, p (Ω) L p -based Sobolev space of [X]r r-convexification of X, 47
order k in Ω, 903 Xbdd (Ω), X (Ω)bdd , 474
930 SYMBOL INDEX

X Generalized Banach Function · Xnorm on the associated space

Space on the measure space of X , 225
(X, M, μ), 219 ∞
X̊ closure of Lcomp in the Generalized
Banach Function Space X, 247
· X norm on the Generalized
Banach Function Space X, 219
X  , 900
X associated space of X, 225 ξ  , 900