Graduate Texts in Mathematics

234

Editorial Board
S. Axler
K.A. Ribet

Graduate Texts in Mathematics

1

TAKEUTI/ZARING. Introduction to

Axiomatic Set Theory. 2nd ed.

2 OxTOBY. Measure and Category. 2nd ed.
3 ScHAEFER. Topological Vector Spaces.
2nd ed.
4

6
7
8

TAKEUTI/ZARING. Axiomatic Set

Theory.
9

ANDERSON/FULLER. Rings and

Categories of Modules. 2nd ed.

14

GOLUBITSKY/GUILLEMIN. Stable

Mappings and Their Singularities.

15 BERBERIAN. Lectures in Functional
Analysis and Operator Theory.
16 WINTER. The Structure of Fields.
17 ROSENBLATT. Random Processes.
2nd ed.
18 HALMOS. Measure Theory.
19 HALMOS . A Hilbert Space Problem
Book. 2nd ed.
20 HusEMOLLER. Fibre Bundles. 3rd ed.
21 HUMPHREYS. Linear Algebraic Groups.
22

BARNES/MACK. An Algebraic

Introduction to Mathematical Logic.

23 GREUB. Linear Algebra. 4th ed.
24 HOLMES. Geometric Functional
Analysis and Its Applications.
25 HEWITT/STROMBERG. Real and Abstract
Analysis.
26 MANES. Algebraic Theories.
27 KELLEY. General Topology
28

36

ZARISKI/SAMUEL. Commutative

39
40

KEMENY/SNELL/KNAPP. Denumerable

37
38

41
42
43
44
45
46
47
48
49
50
51
52
53
54
55

56
57
58

Algebra. Vol.I.
29

ZARISKI/SAMUEL. Commutative

Algebra. Vol.11.
30 JACOBSON. Lectures in Abstract Algebra
I. Basic Concepts.
31 JACOBSON. Lectures in Abstract Algebra
II. Linear Algebra.
32 JACOBSON. Lectures in Abstract Algebra
III. Theory of Fields and Galois Theory.
33 HiRSCH. Differential Topology.

KELLEY/NAMIOKA et al. Linear

Topological Spaces.
MONK. Mathematical Logic.
GRAUERT/FRITZSCHE. Several Complex
Variables.
ARYESON. An Invitation to C*-Algebras.

HUMPHREYS. Introduction to Lie

Algebras and Representation Theory.

10 COHEN. A Course in Simple Homotopy
Theory.
11 CONWAY. Functions of One Complex
Variable I. 2nd ed.
12 BEALS. Advanced Mathematical
Analysis.
13

SPITZER. Principles of Random Walk.

2nd ed.
35 ALEXANDER/WERMER. Several Complex
Variables and Banach Algebras. 3rd ed.

HILTON/STAMMBACH. A Course in

Homological Algebra. 2nd ed.

MAC LANE. Categories for the Working
Mathematician. 2nd ed.
HUGHES/PIPER. Projective Planes.
J.-P. SERRE. A Course in Arithmetic.

34

59
60
61

Markov Chains. 2nd ed.

APOSTOL. Modular Functions and
Dirichlet Series in Number Theory.
2nded.
J.-P. SERRE. Linear Representations of
Finite Groups.
GILLMAN/JERISON. Rings of Continuous
Functions.
KENDIG. Elementary Algebraic
Geometry.
LoEVE. Probability Theory 1.4th ed.
LoEVE. Probability Theory II. 4th ed.
MoiSE. Geometric Topology in
Dimensions 2 and 3.
SACHS/WU. General Relativity for
Mathematicians.
GRUENBERG/WEIR. Linear Geometry.
2nd ed.
EDWARDS. Fermat's Last Theorem.
KLINGENBERG. A Course in Differential
Geometry.
HARTSHORNE. Algebraic Geometry.
MANIN. A Course in Mathematical
Logic.
GRAVER/WATKINS. Combinatorics with
Emphasis on the Theory of Graphs.
BROWN/PEARCY. Introduction to

Operator Theory I: Elements of

Functional Analysis.
MASSEY. Algebraic Topology:
An Introduction.
CROWELL/FOX. Introduction to Knot
Theory.
KoBLiTZ. ;7-adic Numbers, /7-adic
Analysis, and Zeta-Functions. 2nd ed.
LANG. Cyclotomic Fields.
ARNOLD. Mathematical Methods in
Classical Mechanics. 2nd ed.
WHITEHEAD . Elements of Homotopy
Theory.

62

KARGAPOLOV/MERLZIAKOV.

63

Fundamentals of the Theory of

Groups.
BoLLOBAS. Graph Theory
(continued after index)

Palle E.T. Jorgensen

Analysis and Probability

Wavelets, Signals, Fractals
With graphics by Brian Treadway
58 figures and illustrations

rmser
Spri
g^

Palle E.T. Jorgensen

Mathematics Department
The University of Iowa
Iowa City, lA 52242
USA
[email protected]

Editorial Board
S. Axler
Mathematics Department
San Francisco State University
San Francisco, CA 94132
USA
axler @sfsu. edu

K. A. Ribet
Mathematics Department
University of Cahfornia, Berkeley
Berkeley, CA 94720-3840
USA
ribe t@math. berkeley. edu

Mathematics Subject Classification (2000): 41A15,42C20,42A16,42A65,43A65,46L55,47C15,

60J15,94A11
Library of Congress Control Number: 2006921447
ISBN-10:0-387-29519-4

e-ISBN: 0-387-33082-8

ISBN-13:978-0-387-29519-0
Printed on acid-free paper.
2006 Springer Science+Business Media, LLC
All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer Science+Business Media LLC, 233 Spring Street,
New York, N Y 10013, U.S.A.), except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic
adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden.
The use in this publication of trade names, trademarks, service marks and similar terms, even if
they are not identified as such, is not to be taken as an expression of opinion as to whether or not
they are subject to proprietary rights.
9 8 7 6 5 4 321
springer.com

Dedicated to the memory ofShizuo Kakutani

About the cover figure

The figure in the lower right-hand comer of the book cover illustrates the three
themes fi'om inside the book, wavelets, signals, and fractals: Firstly, this graph is one
of the fimctions which are generated with the use of four magic numbers. Specifically, these same numbers also start an algorithmic generation of Daubechies wavelet
fimctions, and the reader is referred to Figures 7.5 and 7.16 on pages 120-121 and
134 for more images, for mathematical background and explanations, more figures,
as well as for theory and exercises in Chapter 7 itself
To the experts: Figure 7.5 includes the coverfigvire,and it represents the progression of fimctions which starts with one of the foiir-tap cases, and which is governed
by the so-called pyramid algorithm. The pyramid algorithm in turn is a delicate design used in the generation of sequences of basis fimctions. These fimctions are then
fiuther scaled and used in the representation of wavelet packets. And the representations and the choices leads to new bases, selectionsfi*om"libraries of bases" and
based on entropy considerations; hence probability!
More probability: The random-walk approach to the analysis of pyramid algorithms in turn is where the calculus of probability comes into the mix. And signals:
To begin with, the use of four magic numbers is by adaptation fi-om an algorithmic
design which was first used by engineers, and which is fimdamental in signal processing; now more recently adapted to image processing as well. But the particular
fimction on the cover also represents the kind of sound signals that feature a beat.
A quick glimpse of the figure finally reveals itsfi-actalnature: By this we mean that
shapes in a picture are repeated at different scales up to similarity; and which fiirther
display an underlying algorithm. Example: Large-scale shapes which envelop similar
shapes at smaller scales!

Preface

If people do not believe that mathematics is simple, it is only because they

John von Neumann
do not realize how complicated life is.
While this is a course in analysis, our approach departs from the beaten path in some
ways. Firstly, we emphasize a variety of connections to themes from neighboring
fields, such as wavelets, fractals and signals; topics typically not included in a graduate analysis course. This in turn entails excursions into domains with a probabilistic
flavor. Yet the diverse parts of the book follow a common underlying thread, and together they constitute a good blend; each part in the mix naturally complements the
other.
In fact, there are now good reasons for taking a wider view of analysis, for example the fact that several applied trends have come to interact in new and exciting
ways with traditional mathematical analysisas it was taught in graduate classes for
generations. One consequence of these impulses from "outside" is that conventional
boundaries between core disciplines in mathematics have become more blurred.
Fortunately this branching out does not mean that students will need to start out
with any different or additional prerequisites. In fact, the ideas involved in this book
are intuitive, natural, many of them visual, and geometric. The required background
is quite minimal and it does not go beyond what is typically required in most graduate
programs.
We believe that now is a good time to slightly widen the horizons of the subject
"analysis" as we teach it by stressing its relations to neighboring fields; in fact we
believe that analysis is thereby enriched.
Despite the inclusion of themes from probability and even from engineering,
the course still has an underlying core theme: A constmctive approach to building
bases in fiinction spaces. The word "constructive" here refers to our use of recursive
algorithms. As it turns out, the algorithmic ideas involved are commonly used in such
diverse areas as wavelets, fractals, signal and image processing. And yet they share
an underlying analysis core which we hope to bring to light.

viii

Preface

Our inclusion here of some applied topics (bordering probability theory and engineering) we believe is not only useful in itself, but more importantly, core mathematics, and analysis in particular have benefitedfi-omtheir many interconnections to
trends and influencesfi*omthe "outside" world.
Yet our wider view of the topic analysis only entails a minor adjustment in course
planning. Our branching out to some applications will be guided tours: to topics
fi:-om probability theory (e.g., to certain random-walk models), and to signal and
image processing. The ideas are presented fi'om scratch, are easy to follow, and they
do not require prior knowledge of probability or of engineering. But we will go a
little beyond the more traditional dose of measure theory and matrix algebra that is
otherwise standard or conventional fare in most first-year graduate courses.
For those reasons we believe the book may also be suitable for a "second analysis
course," and that it leaves the instructor a variety of good options for covering a
selection of neighboring disciplines and applications in more depth.

Iowa City,
June 2006

Palle E. T. Jorgensen

Contents

Preface

vii

Getting started
xv
An apology
xv
Glossary: function, random variable, signal, state, sequence (incl vectorvalued), random walk, time-series, measurement, nested subspaces,
refinement, multiresolution, scales of visual resolutions, operator,
process, black box, observable (if self adjoint), Fourier dual pair,
generating function, time/frequency, P/Q, convolution, filter, smearing,
decomposition (e.g., Fourier coefUcients in a Fourier expansion), analysis,
frequency components, integrate (e.g., inverse Fourier transform),
reconstruct, synthesis, superposition, subspace, resolution, (signals in a)
frequency band, Cuntz relations, perfect reconstruction from subbands,
subband decomposition, inner product, correlation, transition probability,
probability of transition from one state to another, /out = ^/in^
input/output, transformation of states, fractal, conditional expectation,
martingale, data mining (A translation guide!)
xvii
Multiresolutions
xxvi
Prerequisites and cross-audience
xxvii
Aim and scope
xxviii
Self-similarity
xxix
New issues, new tools
xxx
List of names and discoveries
xxx
General theory
xxxiv
A word about the graphics and the illustrations
xxxiv
Special features of the book
xxxv
Exercises: Overview
xxxvi
Figures. Read Me!

xxxix

Contents

Acknowledgments

xliii

Introduction: Measures on path space

Prerequisites
Prelude
1.1 Wavelets
1.2 Path space
1.3 Multiresolutions
1.4 Sampling
1.5 A convergence theorem for infinite products
1.6 A brief outline
1.7 From wavelets to
fi-actals
Exercises
History
References and remarks

1
1
1
2
6
9
17
18
21
22
27
33
35

Transition probabilities: Random walk

39
Prerequisites
39
Prelude
39
2.1 Standing assumptions
40
2.2 An example
41
2.3 Some definitions: The Ruelle operator, harmonic fimctions, cocycles 43
2.4 Existence of the measures P^
43
2.5 Kolmogorov's consistency condition
46
2.6 The probability space Q
47
2.7 A boundary representation for harmonic
fimctions
48
2.8 Invariant measures
52
Exercises
54
References and remarks
57

No vs. Z
Prerequisites
Prelude
3.1 Terminology
3.2 The unit interval
3.3 A sufficient condition iov Px{Z) = \
Exercises
References and remarks

59
59
59
60
62
64
66
67

A case study: Duality for Cantor sets

Prerequisites
Prelude

69
69
69

Contents

xi

4.1 Affine iterated function systems: The general case

4.2 The quarter Cantor set: The example W(x) co^ilnx)
4.3 The conjugate Cantor set, and a special harmonic function
4.4 A sufficient condition for P^ (No) = 1
Conclusions
Exercises
References and remarks

70
72
76
78
79
79
80

Infinite products
Prerequisites
Prelude
5.1 Riesz products
5.2 Random products
5.3 The general case
5.4 A uniqueness theorem
5.5 Wavelets revisited
Exercises
References and remarks

83
83
83
84
84
85
86
91
93
97

The minimal eigenfunction

Prerequisites
Prelude
6.1 A general construction of/zmin
6.2 A closed expression for /^min

99
99
99
100

102
Exercises
106
References and remarks
107
Generalizations and applications
109
Prerequisites
109
Prelude
109
7.1 Translations and the spectral theorem
110
7.2 Multiwavelets and generalized multiresolution analysis (GMRA) .. 114
7.3 Operator-coefficients
114
7.4 Operator-valued measures
115
7.5 Wavelet packets
122
7.6 Representations of the Cuntz algebra Oi
131
7.7 Representations of the algebra of the canonical anticommutation
relations (CARs)
138
Exercises
141
References and remarks
151

xii

Contents

Pyramids and operators

Prerequisites
Prelude
8.1 Why pyramids
8.2 Dyadic wavelet packets
8.3 Measures and decompositions
8.4 Multiresolutions and tensor products
Exercises
References and remarks

157
157
157
158
159
166
168
173
176

Pairs of representations of the Cuntz algebras O^, and their

application to multiresolutions
Prerequisites
Prelude
9.1 Factorization of unitary operators in Hilbert space
9.2 Generalized multiresolutions
9.3 Permutation of bases in Hilbert space
9.4 Tilings
9.5 Applications to wavelets
9.6 An application to
fractals
9.7 Phase modulation
Exercises
References and remarks

179
179
179
180
181
182
185
190
194
198
199
204

Appendices: Polyphase matrices and the operator algebra O^q

Prerequisites
Prelude
Appendix A: Signals and
filters
Appendix B: Hilbert space and systems of operators
Appendix C: A tale of two Hilbert spaces
Table C.l: Operations on two Hilbert spaces: The correspondence
principle
Appendix D: Signal processing, matrices, and programming diagrams . . .
References and remarks: Systems theory

205
205
205
206
210
212

Afterword
Comments on signal/image processing terminology
Introduction
JPEG 2000 vs. GIF
JPEG 2000
GIF
Grayscale

223
223
223
225
225
226
227

213
218
221

Contents
Quadrature-mirror
What is a frame?
To the mathematics student
To an engineer
AHas (ahasing)
Engineering
Mathematics
Computational mathematics

filter

xiii
227
228
228
229
229
229
229
230

Epigraphs

233

References

237

Symbols

251

Index

259

Drawing by the author, next page:

Wavelet algorithms are goodfor vast sets of numbers.
An engineering friend described the old approach to data mining as
"Just drop a computer down onto a gigantic set of unstructured numbers!"
(^data mining: see Section 6.2, pp. 102-105, and the Glossary, pp. xxiv-xxv).

n^ ^H

it.^fV.

.1J 5^^^^!l^^^l^^u%^v^^.^?^:^v^ir^ ^^

Getting started

From its shady beginnings devising gambling strategies and counting corpses in medieval London, probability theory and statistical inference now
emerge as better foundations for scientific models, especially those of the
process of thinking and as essential ingredients of theoretical mathematics,
even the foundations of mathematics itself
^David Mumford

An apology
You ask: "Why all the fiiss?" Wavelets, signals, fractals? Isn't all of this merely a
fad? Or a transient popularity trend? And what's the probability part in the book all
about? And non-commuting operators? As for bases in linear spaces, what's wrong
with Gram-Schmidt?
You may think: "Fourier has served us well for ages; so why do we need all the
other basis fiinctions?" Wavelets and so on? And why engineering topics in a
mathematics course? And the pictures? Are they really necessary?
And there are signal processing and image processing!? Yes, technology is
lovely, but why not leave it to the engineers?
Response: The links between mathematics and engineering are much deeper than
the fact that we mathematicians teach service courses for engineers. Our bread and
butter!
Mathematics draws ideas and strengths from the outside world, and the connections to parts of engineering have been a boon to mathematics: From signal processing to wavelet analysis! That is true even if we forget about all of the practical applications emerging from these connections. Without inspiration from the neighboring
sciences, mathematics would in all likelihood become rather sterile, and overly formal. I see opportunities at crossroads. In this book you will see the benefits mathematics is reaping from trends and topics in engineering. It is witnessed in a striking
way by exciting developments in wavelets. From wavelets we see how notions of

xvi

Getting started

scale-similarity can be exploited in basis computations that use tricks devised for
signal processing. Just open the book and glance at some of the wavelet fimctions.
At the same time, the key notion of self-similarity, such as the scale-similarity used
everywhere for wavelets, is essential to our understanding of fractals: Fern-like pictures that look the same at small and at large scales. One problem in the generation of
wavelet bases is selecting the "nice" (here this means differentiable) wavelets among
huge families of fractal-looking (non-smooth, or singular) functions. Z^-fiinctions
can be very "bad" indeed!! Computers generate the good and the bad, and we are
left with the task of sorting them out and making selections. We will see (directly
from large libraries of pictures) that mathematical wavelet machines are more likely
to spit out bad fimctions unless they are told where to concentrate the search from
the intrinsic mathematics.
These wavelets, signals, and fractals are things that have caught our attention
in recent decades, but the mathematical part of this has roots back at least a hundred years, for example, to Alfred Haar and to Oliver Heaviside at the turn of the
last century. From Haar we have the first wavelet basis, and with Heaviside we see
the beginning of signal analysis. It is unlikely that either one knew about the other.
Ironically, at the time (1909), Haar's paper had little impact and was hardly noticed,
even on the small scale of "notice" that is usually applied to mathematics papers.
Haar's wonderful wavelet only began to draw attention in the mid-nineteen-eighties
when the connections to modem signal processing became much better understood.
These connections certainly served as a main catalyst in what are now known as
wavelet tools in pure and applied mathematics. But at the outset, the pioneers in
wavelets had to "rediscover" a lot of stuff from signal processing: frequency bands,
high-pass, low-pass, analysis and synthesis using down-sampling, and up-sampling,
reconstruction of signals, resolution of images; all tools that have wonderfiil graphics
representations in the engineering literature.
But still, why would we think that Fourier's basis, and his lovely integral decomposition, are not good enough? Many reasons: Fourier's method has computational
drawbacks. This was less evident before computers became common and began to
play important roles in applied and theoretical work. But expansion of functions or
signals into basis decompositions (called "analysis" in signal processing) involves
basis coefficients (Fourier coefficients, and so on), and if we are limited to Fourier
bases, then the computation of the coefficients must by necessity rely on integration.
"Computers can't integrate!" Hmmm! Well, not directly. The problem must first be
discretized. And there is need for a more direct and algorithmic approach. Hence the
wavelet algorithm! In any case, algorithms are central in mathematics even if you
do not concern yourself with computers. And it is the engineering connections that
inspired the most successfiil algorithms in our subject.

Glossary

xvii

Glossary
function, random variable, signal, state, sequence (incl. vector-valued), random walk,
time-series, measurement, nested subspaces, refinement, multiresolution, scales of visual
resolutions, operator, process, black box, observable (if self adjoint), Fourier dual pair,
generating function, timefrequency, P/Q, convolution, filter, smearing, decomposition
(e.g., Fourier coefficients in a Fourier expansion), analysis, frequency components, integrate
(e.g., inverse Fourier transform), reconstruct, synthesis, superposition, subspace, resolution,
(signals in a) frequency band, Cuntz relations, perfect reconstruction from subbands,
subband decomposition, inner product, correlation, transition probability, probability of
transition from one state to another, /out = ^/in^ input/output, transformation of states,
fractal, conditional expectation, martingale, data mining (A translation guide!)

"The question is," said Alice, "whether you can make words mean so many
different things."
^Lewis Carroll
This glossary consists of a list of terms used inside the book in varied contexts of
mathematics, probability, engineering, and on occasion physics. To clarify the seemingly confusing use of up to four different names for the same idea or concept, we
have further added informal explanations spelling out the reasons behind the differences in current terminology from neighboring fields.
When sorting through the disparate variations of lingo in the sciences, the more
mundane problems of plain and "ordinary" languages might seem minor: There is the
variety of differences from Lido-European to the schizophrenia of Finno-Hungarian.
And inside the same building on campus, there are even the variations in lingo that
separate areas of mathematics: If you are a mathematician doing analysis in "plain
English" you might well learn to understand the "Portuguese" spoken by your colleagues in algebra; but when it comes to the Hungarian of an engineer, or the Bantu
of some physicists, you can get lost or confused. A former student just wrote me
about the language gulf between domestic mathematics and the jungle of industry.
He now works in a company and is doing applied wavelets. And he writes that the
language of implementation is quite different from that of our standard mathematics
books.
DISCLAIMER: This glossary has the structure of four columns. A number of terms
are listed line by line, and each line is followed by explanation. Some "terms" have
up to four separate (yet commonly accepted) names. The last four terms in the list,
"fractal," "conditional expectation," "martingale," and "data mining," are the only
ones where I could think of only one name.
It should be added that my "descriptions" for the various terms are meant to stress
intuitive aspects, as opposed to mathematical definitions. One reason for stressing the
intuition behind the concepts is that there is not quite agreement about the precise
meaning of some of the terms in the four fields: in mathematics, in probability, in
engineering, and in physics.

xviii

Getting started

MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

function
(measurable)

random variable

signal

state

Mathematically, fiinctions may map between any two sets, say, from X to
F; but if X is a probability space (typically called ^ ) , it comes with a cralgebra B of measurable sets, and probability measure P. Elements E inB
are called events, and P(E) the probability of E. Corresponding measurable
functions with values in a vector space are called random variables, a terminology which suggests a stochastic viewpoint. The fimction values of a
random variable may represent the outcomes of an experiment, for example
"throwing of a die." hi this simplest experiment, the range of the random
variable is the set of integers from 1 to 6, the six possible measurements
resulting from the experiment.
Yet, fimction theory is widely used also in engineering where fiinctions
are typically thought of as signal. In this case, X may be the real line for
time, or W^. And I noticed that engineers visualize fiinctions as signals. A
particular signal may have a stochastic component, and this feature simply
introduces an extra stochastic variable into the "signal," for example noise.
Turning to physics, in our present application, the physical fiinctions will
be typically be in some L^-space, and I^-fimctions with unit norm represent
quantum mechanical "states."
sequence (incl.
vector-valued)

random walk

time-series

measurement

Mathematically, a sequence is a fimction defined on the integers Z or on

subsets of Z, for example the natural numbers N. Hence, if time is discrete,
this to the engineer represents a time series, such as a speech signal, or any
measurement which depends on time. But we will also allow fiinctions on
lattices such as Z^.
In the case c/ = 2, we may be considering the grayscale numbers which
represent exposure in a digital camera. In this case, the fimction (grayscale)
is defined on a subset of Z^, and is then simply a matrix.
A random walk on Z^ is an assignment of a sequential and random motion
as a fimction of time. The randomness presupposes assigned probabilities.
But we will use the term "random walk" also in connection with random
walks on combinatorial trees.
nested
subspaces

refinement

multiresolution

scales of visual
resolutions

While finite or infinite families of nested subspaces are ubiquitous in mathematics, and have been popular in Hilbert-space theory for generations (at

Glossary
MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

least since the 1930s), this idea was revived in a different guise in 1986 by
Stephane Mallat, then an engineering graduate student; see [Mal89]. In its
adaptation to wavelets, the idea is now referred to as the multiresolution
method.
What made the idea especially popular in the wavelet community was that
it offered a skeleton on which various discrete algorithms in applied mathematics could be attached and turned into wavelet constructions in harmonic
analysis. In fact what we now call multiresolutions have come to signify
a crucial link between the world of discrete wavelet algorithms, which are
popular in computational mathematics and in engineering (signal/image processing, data mining, etc.) on the one side, and on the other side continuous
wavelet bases in function spaces, especially in 1-^(1^.^). Further, the multiresolution idea closely mimics how fractals are analyzed with the use of
finite fiinction systems.
But in mathematics, or more precisely in operator theory, the underlying
idea dates back to work of John von Neumann, Norbert Wiener, and Herman
Wold, where nested and closed subspaces in Hilbert space were used extensively in an axiomatic approach to stationary processes, especially for time
series. Wold proved that any (stationary) time series can be decomposed into
two different parts: The first (deterministic) part can be exactly described by
a linear combination of its own past, while the second part is the opposite
extreme; it is unitary, in the language of von Neumann.
von Neumann's version of the same theorem is a pillar in operator theory.
It states that every isometry in a Hilbert space H is the unique sum of a
shift isometry and a unitary operator, i.e., the initial Hilbert space H splits
canonically as an orthogonal sum of two subspaces Hs and Hu in H, one
which carries the shift operator, and the other Hu the unitary part. The shift
isometry is defined from a nested scale of closed spaces F, such that the
intersection of these spaces is Hu.
However, Stephane Mallat was motivated instead by the notion of scales
of resolutions in the sense of optics. This in turn is based on a certain
"artificial-intelligence" approach to vision and optics, developed earlier by
David Marr at MIT, an approach which imitates the mechanism of vision in
the human eye.
The connection from these developments in the 1980s back to von Neumann is this: Each of the closed subspaces Vn corresponds to a level of
resolution in such a way that a larger subspace represents a finer resolution.
Resolutions are relative, not absolute! In this view, the relative complement
of the smaller (or coarser) subspace in larger space then represents the visual

xix

XX

Getting started

MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

detail which is added in passing from a blurred image to a finer one, i.e., to
a finer visual resolution.
This view became an instant hit in the wavelet community, as it offered a
repository for the fimdamental father and the mother fimctions, also called
the scaling fimction cp, and the wavelet fimction y/. Via a system of translation and scaling operators, these fimctions then generate nested subspaces,
and we recover the scaling identities which initialize the appropriate algorithms. What results is now called the family of pyramid algorithms in
wavelet analysis. The approach itself is called the multiresolution approach
(MRA) to wavelets. And in the meantime various generalizations (GMRAs)
have emerged.
In all of this, there was a second "accident" at play: As it turned out,
pyramid algorithms in wavelet analysis now lend themselves via multiresolutions, or nested scales of closed subspaces, to an analysis based on frequency bands. Here we refer to bands of frequencies as they have already
been used for a long time in signal processing.
Even though J. von Neumann and H. Wold had been using nested or
scaled families of closed subspaces in representing past and fiature for time
series, S. Mallat found that this same idea applies successfiilly to the representation of visual resolutions. And even more importantly, it offers a variety
of powerfiil algorithms for processing of digital images.
Now parallel to all of this, pioneers in probability theory had in fact developed versions of the same refinement analysis. For example, in the theory
of martingales, consistency relations may naturally be reformulated in the
language of nested subspaces in Hilbert space.
One reason for the success in varied disciplines of the same geometric
idea is perhaps that it is closely modeled on how we historically have represented numbers in the positional number system; see, e.g., [Knu81]. Analogies to the Euclidean algorithm seem especially compelling.
operator

process

black box

observable
(if selfadjoint)

In linear algebra students are familiar with the distinctions between (linear)
transformations T (here called "operators") and matrices. For a fixed operator T: V -^ W, there is a variety of matrices, one for each choice of basis in
V and inJV.ln many engineering applications, the transformations are not
restricted to be linear, but instead represent some experiment ("black box,"
in Norbert Wiener's terminology), one with an input and an output, usually
fimctions of time. The input could be an external voltage fimction, the black

Glossary
MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

box an electric circuit, and the output the resulting voltage in the circuit.
(The output is a solution to a differential equation.)
This context is somewhat different from that of quantum mechanical
(QM) operators T.V ^ V where F is a Hilbert space. In QM, selfadjoint operators represent observables such as position Q and momentum P,
or time and energy.

Fourier dual
pair

generating
function

time/frequency

PIQ

The following dual pairs position g/momentum P, and time/energy may be

computed with the use of Fourier series or Fourier transforms; and in this
sense they are examples of Fourier dual pairs. If for example time is discrete,
then frequency may be represented by numbers in the interval [ 0,2;r); or in
[ 0, 1) if we enter the number In into the Fourier exponential. Functions of
the frequency are then periodic, so the two endpoints are identified. In the
case of the interval [ 0, I), 0 on the left is identified with I on the right. So
a low frequency band is an interval centered at 0, while a high frequency
band is an interval centered at 1/2. Let a function ^ on [0, 1) represent
a probability assignment. Such fiinctions W are thought of as "filters" in
signal processing. We say that W is low-pass if it is 1 at 0, or if it is near
1 for frequencies near 0. Low-pass filters pass signals with low frequencies,
and block the others.
If instead some filter ^ is 1 at 1/2, or takes values near 1 for frequencies near 1/2, then we say that W is high-pass; it passes signals with high
frequency.

convolution

filter

smearing

Pointwise multiplication of fiinctions of frequencies corresponds in the

Fourier dual time-domain to the operation of convolution (or of Cauchy
product if the time-scale is discrete.) The process of modifying a signal with
a fixed convolution is called a linear filter in signal processing. The corresponding Fourier dual frequency fianction is then referred to as "frequency
response" or the "frequency response fiinction."
More generally, in the continuous case, since convolution tends to improve smoothness of fiinctions, physicists call it "smearing."

xxi

xxii

Getting started

MATHEMATICS

PROBABILITY

decomposition

(e.g., Fourier
coefficients in a
Fourier expansion)

ENGINEERING

PHYSICS

analysis

frequency
components

Calculating the Fourier coefficients is "analysis," and adding up the pure

frequencies (i.e., summing the Fourier series) is called synthesis. But this
view carries over more generally to engineering where there are more operations involved on the two sides, e.g., breaking up a signal into its frequency
bands, transforming fiirther, and then adding up the "banded" fiinctions in
the end. If the signal out is the same as the signal in, we say that the analysis/synthesis yields perfect reconstruction.

integrate
reconstruct
(e.g., inverse
Fourier transform)

synthesis

superposition

Here the terms related to "synthesis" refer to the second half of the kind of
signal-processing design outlined in the previous paragraph.

subspace

resolution

(signals in a)
frequency band

For a space of fiinctions (signals), the selection of certain frequencies serves

as a way of selecting special signals. When the process of scaling is introduced into optics of a digital camera, we note that a nested family of
subspaces corresponds to a grading of visual resolutions.

Cuntz relations

perfect
reconstruction
from subbands

subband
decomposition

The operator relations of Joachim Cuntz described in detail in Chapter 7

serve to make precise the kind of signal-processing design outlined in the
previous paragraph. Although anticipated by engineers, the Cuntz relations
give a particularly elegant way for us to understand the geometry of iterated
subdivision schemes which include a variety of models with perfect reconstruction from signal and image processing.

Glossary
MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

inner product

correlation

transition
probability

probability
of transition
from one state
to another

xxiii

In many applications, a vector space with inner product captures perfectly

the geometric and probabilistic features of the situation. This can be axiomatized in the language of Hilbert space; and the inner product is the most
crucial ingredient in the familiar axiom system for Hilbert space.
/o^i = TfiYi

input/output

transformation
ofstates

Systems theory language for operators T\V -^ W. Then vectors in V are

input, and in the range of T output.
fractal

Intuitively, think of a fractal as reflecting similarity of scales such as is seen

in fern-like images that look "roughly" the same at small and at large scales.
While there may not be agreement about a rigorous mathematical definition,
Mandelbrot originally d^QfiwtA fractals as sets whose Hausdorff-Besicovich
dimension exceeded their topological dimension, but later accepted all selfsimilar, self-afiine, or quasi-self-similar sets as fractals. Moreover, even
more generally, the self-similarity could refer alternately to space, and to
time. And frirther versatility was added, in that flexibility is allowed into the
definition of "similar."
In this book, our focus is more narrowly on the self-afiine variant (where
computations are relatively simple); but in addition, we encounter the fractal
concept in other contexts, e.g., for measures, and for probability processes
(such as fractal Brownian motion). Further, we have stressed examples more
than the general theory.
The fractal concept for measures is especially agreeable in the self-affine
case, since affine maps act naturally on measures. So for each fractal dimension s, there is a corresponding ^-fractal probability measure jj. = fjts
(see Chapter 4) which is the unique solution to a natural fixed-point equation, one which depends on s. Moreover, we may then recover this way the
spatial fractal set itself as the support of this measure /JL.
As for ^-fractal Brownian motion (fBm), the "^-fractal" feature there
refers to how the position Xf at time / of the f Bm-process transforms under
scaling off. If time t scales by c, then the respective distributions before and
after scaling are related by the power-law c\ Specifically, for all t, the finite
distributions calculated for Xct and for c^Xt coincide.

xxiv

Getting started

MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

conditional
expectation

Let (Q, A, P) be a probability space.

Recall that a ftinction X: ^ ^ E (the reals) is an ^-random variable if
it is measurable with respect to the two cr-algebras on the domain and the
target, i.e., A on Q, and B := the standard Borel cr-algebra on E. For an
^-random variable X, let E {X) := J^XdP, or simply {X) denote the
corresponding mean (or expected value). If r> is a sub-cr-algebra of ^ , define a measure ju on V by ju(S) := J^XclP for S e V. Then denote by
E[X \T>]thQ Radon-Nikodym derivative of ju with respect to P.
Let 7 be a second ^-random variable, and let V :=VY = Y~^ (B) be the
smallest cr-algebra with respect to which Y is measurable. Then clearly V
is a sub-(j-algebra of A mdwQ set E (X \ Y) := E[X \V] = {X \ Y).
This extends naturally to the case when there is a finite family of ^-random
variables in place of Y.

martingale

We use the term here only in its simplest form, the discrete case, called
"discrete-time:" Let (Q, A, P) be a probability space. The following identity defines a martingale. Consider a sequence of .4-random variables XQ,
X\, . . . each with finite mean. We say it is a martingale if for each n,
the conditional expectation of X+i given XQ, X \ , . . . , Xn is X, i.e.,
(X+i I X o , . . . , X ) = Xn (see Feller's book [Fel71, p. 210]). The term
was first used to describe a type of wagering in which the bet is doubled or
halved after a loss or win, respectively. The concept of martingales is due to
Levy, and it was developed extensively by Doob [Doo94].
A random walk on the integers with each successive step left or right
equally likely, i.e., with equal transition probabilities {pi = p^ = 1/2), is
an example of a martingale.

data mining

The problem of how to handle and make use of large volumes of data is a
corollary of the digital revolution. As a result, the subject of data mining itself changes rapidly. Digitized information (data) is now easy to capture automatically and to store electronically [HuTKOS]. hi science, in commerce,
and in industry, data represents collected observations and information: In
business, there is data on markets, competitors, and customers [AgKu04a,
AgKu04b]. In manufacturing, there is data for optimizing production opportunities, and for improving processes [Kus02, Kus05]. A tremendous

Glossary
MATHEMATICS

PROBABILITY

ENGINEERING

PHYSICS

potential for data mining exists in medicine [KuLDOl, KuDS05], genetics

[ShKu04], and energy [KuBuOS]. But raw data is not always directly usable, as is evident by inspection. A key to advances is our ability to extract
information and knowledge from the data (hence "data mining"), and to understand the phenomena governing data sources. Data mining is now taught
in a variety of forms in engineering departments, as well as in statistics and
computer science departments.
One of the structures often hidden in data sets is some degree of scale.
The goal is to detect and identify one or more natural global and local scales
in the data. Once this is done, it is often possible to detect associated similarities of scale, much like the familiar scale-similarity from multidimensional
wavelets, and from fractals. Indeed, various adaptations of wavelet-like algorithms have been shown to be useful. These algorithms themselves are
useful in detecting scale-similarities, and are applicable to other types of
pattern recognition. Hence, in this context, generalized multiresolutions offer another tool for discovering structures in large data sets, such as those
stored in the resources of the hitemet. Because of the sheer volume of data
involved, a strictly manual analysis is out of the question. Instead, sophisticated query processors based on statistical and mathematical techniques
are used in generating insights and extracting conclusions from data sets.
But even such an approach breaks down as the quantity of data grows and
the number of dimensions increases. Instead there is a new research area
(knowledge discovery in databases (KDD)) which develops various tools
for automated data analysis.
However, statistics is still at the heart of the problem of inference from
the data. The widespread use of statistics, pattern recognition, and machinelearning algorithms is somewhat hindered in many areas by our ability to
collect large volumes of data. The next limitation in the subject arises when
the data is too large to fit in the main computer memory. As a result, we are
faced with new issues, e.g., quality of data, creative data analysis, and data
transformation [KusOl]. Theory and hypothesis formation now becomes
critical in our task of deriving insights into underlying phenomena from the
raw data. Various adaptations of wavelet-like algorithms have again proved
usefiil in detecting scale-similarities, and in other types of pattern recognition. Hence in this context wavelet ideas offer another tool for discovering
structures in vast data sets, such as those in the resources of the Internet.
And there are now a variety of such effective Web mining tools in use.
Areas of data mining include problems of representation, search complexity, and automated use of prior knowledge to help in a data search. Thus we
see the beginnings of a new science for efiicient inference from massive data
sets.

xxv

xxvi

Getting started

Multiresolutions
If you stop to think about it, it is really ironic that Haar's wavelet basis (Figures
1.2 and 7.4, pp. 13 and 118-119) was missed for so long. It is especially ironic
since Haar's work in 1909-1910 had in it implicitly the key idea which got wavelet
mathematics started on a roll 75 years later with Yves Meyer, Ingrid Daubechies,
Stephane Mallat, and others^namely the idea of a multiresolution. In that respect
Haar was ahead of his time. In 1909, Haar's measure on non-abelian groups became
much more widely known.
Yet, returning to wavelets, the word "multiresolution" suggests a connection to
optics from physics. So that should have been a hint to mathematicians to take a
closer look at trends in signal and image processing! Moreover, even staying within
mathematics, it turns out that as a general notion this same idea of a "multiresolution" has long roots in mathematics, even in such modem and pure areas as operator theory and Hilbert-space geometry. And in probability theory and in dynamics,
A.N. Kolmogorov and J. Doob had already long ago identified martingales, again
close cousins of multiresolutions. Looking even closer at these interconnections, we
can now recognize scales of subspaces (so-called multiresolutions) in classical algorithmic construction of orthogonal bases in inner-product spaces, now taught in lots
of mathematics courses under the name of the Gram-Schmidt algorithm. Indeed, a
closer look at good old Gram-Schmidt reveals that it is a matrix algorithm. Hence
new mathematical tools involving non-commutativity! Obviously, fiinction spaces
are infinite-dimensional. Since Gram-Schmidt is recursive, it does not stop, at least
not until we tell it to stop. We do that when the basis expansion has achieved a "good
enough" approximation to the true fimction, or the true signal or image which is
being analyzed.
Approximation? So we must retain the "significant" terms in an analysis expansion and throw out the other terms! To know which is "significant," thresholds must
be assigned, and probabilities, and even entropy, from information theory must be
used.
A Wiener process, carried to the limit, gives a nowhere differentiable continuous
Brownian path, parametrized by time. This is a model for a physical Brownian trajectory of an actual particle. Actual Brownian particles do not follow paths that are
precisely of this nature.
If the signal to be analyzed is an image, then why not select a fixed but suitable resolution (or a subspace of signals corresponding to a selected resolution), and
then do the computations there? Of course, the selection of a fixed "resolution" is
dictated by practical concerns. That idea was key in turning computation of wavelet coefiicients into iterated matrix algorithms. As the matrix operations get large,
the computation is carried out in a variety of paths arising from big matrix products.
Such paths have been studied in probability since Kolmogorov in the 1930s, but paths
are perhaps better known in their continuous variants. Yet, what we know about the

Prerequisites and cross-audience

xxvii

continuous case is the result of limit considerations arising from the discrete case.
The dichotomy, continuous vs. discrete, is quite familiar to engineers. The industrial
engineers typically work with huge volumes of numbers.
Numbers! So why wavelets? Well, what matters to the industrial engineer
is not really the wavelets, but the fact that special wavelet functions serve as an
eflicient way to encode large data setsI mean encode for computations. And the
wavelet algorithms are computational. They work on numbers. Encoding numbers
into pictures, images, or graphs of functions comes later, perhaps at the very end of
the computation. But without the graphics, I doubt that we would understand any of
this half as well as we do now. The same can be said for the many issues that relate
to the crucial mathematical concept of self-similarity, as we know it from fractals,
and more generally from recursive algorithms.
Prerequisites and cross-audience
I have used preliminary versions of this book in my courses. These courses fit the bill:
"a second course in analysis," in one form or the other. Some of my courses were
more traditional, and for mathematics students, while others served a quite mixed
audience, including students from engineering. At my university, there is a serious
demand for interdisciplinary mathematics courses, not only involving engineers, but
also students from physics, statistics, computer science, finance, and more. And I
expect that this is a national trend. However, we found that many traditional mathematics texts are too narrowly specialized and not well suited for cross-audiences.
With that in mind, we have attempted a wider focus even within mathematics proper;
specifically, we feel that a graduate text in analysis should be relevant to computational mathematics and to numerical analysis. Our present emphasis on wavelets,
signals and fractals invites such a wider focus.
The students in my courses typically had some familiarity with function theory
and measures. But their background was always diverse and varied. To accommodate
diverse audiences (referring to the level of mathematical maturity, specialization,
etc.), I included in my courses, and in the book, facts from a variety of topics.
How? Some of the exercises serve this purpose, assimilation. For example. Exercises 1.11 through 1.14 have the flavor of tutorials', they let students pick up on
some quite basic but central points from Fourier series, iimer-product spaces, linear
transformations, matrices (finite and infinite), and Hilbert space. Working the exercises is the best way for students to learn and review fundamentals! In these multipart
exercises, the reader is then guided step by step through the issues, but as they are
needed in the text. As other basic topics are needed later in the book, there are then
other multipart exercises that accomplish the same goal: e.g.. Exercise 2.6 (integral
operators). Exercises 2.7 through 2.10, and 9.8-9.9 (Brownian motion, including
the fractional variant). Exercises 7.2 through 7.7 (matrix theory), and Exercises 7.9
through 7.11 (tensor product of Hilbert space, and product measure). An advantage

xxviii

Getting started

of this approach is that students then see and learn these preliminaries precisely in
the form in which they are used in the course.
When I was teaching mixed groups of students, the engineers in my class typically came from the following departments, EE (electrical engineers), electrical and
computer engineering, communications, mechanical, industrial engineering, data
mining, sensorsamong others! Generally, the engineers looked for an agreeable
mathematical presentation of central ideas from signal and image processing. ("Why
then wavelets?" you may say. Reason: Wavelet algorithms share a lot in common
with signal/image processing schemes! And they are usedhy engineers!)
Some of the engineers in my course worked at our University Hospital, a large
teaching hospital. Their projects involve designing and improving technology in
medical imaging.
I recently gave one of my thesis students (with a good background in mathematics and engineering) the following assignment: "Work out the mathematics used in
the processing of color images in a digital camera!" Reason: This is actually lovely
algorithmic matrix theory, and it is based directly on wavelet algorithms.
Challenge! There is a huge difference in jargon used by engineers and by mathematicians (not to mention the other disciplines!); and often there is a call for some
serious translation of technical lingo before you discover that the two groups are talking about the same thing. At times, my course would begin with first overcoming the
cultural and the communication barriers; hence the system of interrelated appendices
on polyphase matrices in the back of the book, and the list of names and discoveries
on pages xxx-xxxiii below.
Aim and scope
The aim of this book is to show how to use processes from probability, random walks
on branches, and their path-space measures in the study of convergence questions
from harmonic analysis, with particular emphasis on the infinite products that arise
in the analysis of wavelets.
The focus is by nature interdisciplinary, and it is motivated by some new mathematical trends that are still somewhat in a state of flux. We outline how they combine
diverse areas from mathematics and engineering in unexpected ways. As a result, we
aim to address a diverse audience, perhaps unusually diverse, ranging from pure to
applied (engineering and physics), from probability (random walk) to analysis (infinite products), from wavelets to fractals, from linear to non-linear, from fimction
theory to non-abelian operator algebras, from Lebesgue to Hausdorff measure, and
from classical (Fourier) to modem (wavelets and more generally scale-similarity in
time and space). This diversity presents us with a challenge, and we have taken pains
to articulate the interconnections, describe a coherent unity, and address the union
(and not the intersection) of the various groups of readers who have an interest in
this area of mathematics. Our exposition is designed to reach the beginning graduate

Self-similarity

xxix

student, having in mind both students and workers meeting at the crossroads in a
variety of fields.
Self-similarity
The idea of self-similarity is common to the construction of wavelets, various fractals, and graph systems. These subjects may seem diverse, and they are often thought
of as disparate. However, we hope to outline how the notion of self-similarity runs
through these subjects as a red thread connecting central themes: themes from harmonic analysis, discrete mathematics, probability, and operator algebras, hi addition,
we list the two papers [BoNe03] and [Is[ek04] as sources that cover this from a somewhat different but related angle.
Indeed, the concept of self-similarity has proved ubiquitous as well as fundamental in mathematics and in diverse applications, perhaps because it serves to renormalize a rich variety of structures on nested families of scales (for example, similarity
scales in time and/or in space). In wavelet theory, the scales may be represented in
resolutions (a name deriving from optics) taking the form of nested systems of linear
spaces. Similar such systems occur in fractal theory and in dynamics. And in quantum field theory [Nel73], the self-similarity idea underlies the notion of "renormalization group," while in C*-algebra theory [Dav96], it gives rise to representations
of algebras on relations and generators, such as operator algebras on infinite particle
systems [PoStTO], or the algebras of Cuntz, or Cuntz-Krieger. In fact, the pioneering
paper [Bat87] early on stressed a fascinating connection between the renormalization group in quantum physics on the one hand, and a certain fundamental wavelet
construction on the other.
To help the reader better see the bigger picture which we wish to present, we have
included a system of interrelated appendices outlining a certain geometric approach
to subband filtering which is both more operator-theoretic and more general than
is usual. Our appendices build a bridge between two ways of doing wavelet/fractal
subdivision algorithms, one based on quadrature conditions and their generalizations,
and the other based on unitary operator functions and on representations of a certain
C*-algebra.
Recent developments in operator theory, approximation theory, orthogonal expansions, and wavelets demonstrate the need for a combination of techniques from
analysis and probability. Since the relevant tools and techniques for these new applications are often interdisciplinary, they are not always readily available in the
standard textbook literature. Or if they are, they are scattered over a number of older
books or papers which are written for other purposes and aimed at other applications.
The analysis problems that we focus on here are typically not presented together with
their counterparts from probability theory: random walk, path space, infinite products, martingales, Kolmogorov extension techniques in their classical form, and in
some of their modem non-commutative versions.

XXX

Getting started

This book aims to fill an apparent gap in the literature, or at least to help bridge
a gap. It gives a rigorous treatment of a number of convergence questions, and it
also includes some new results. Our use of analysis in this book relies heavily on
probabilistic tools, and the book offers a presentation of them.
New issues, new tools
This book is primarily about wavelets, but it takes a new direction in the subject:
While it is about convergence issues, the questions come fi-om applications with a
twist that is different fi-om the one that has dominated the literature.
The aim of this book is broader. We wish to show how some methods from probability, operator algebras, and random-walk theory can be used to prove theorems in
analysis. The ideas from probability that we will use include processes, random walk
on branches, and path-space measures associated with them. The recurring theme
will be the use of these ideas from probability in the study of convergence questions
from harmonic analysis. Our emphasis is on the kinds of infinite products that arise
in the analysis of wavelets and more generally in the harmonic analysis of iterated
fimction systems, as well as in dynamical systems.
Many modem wavelet constructions from the past few years are by necessity
frequency-localized; and because of that fact, the convergence issues and the tools
that must be employed to resolve them are completely different from the more traditional ones that were developed and tailored in the 1980s to time-localized fiinctions,
i.e., the wavelets that have the scaling fimction (the father fimction) and the wavelet fimction (the mother fimction) both of compact support. Those are the wavelets
where multiresolution tools have been especially successfial.
By contrast, the frequency-localized wavelets have compact support only in the
Fourier-dual variable, and their resolution subspaces are typically not singly generated, so the more traditional multiresolution methods have fallen short, at least in the
form in which they originally were given.
What this means is that some of the fiindamental issues concerning pointwise
convergence in the theory must be attacked with quite different tools: in this case,
with tools from random walk, probability, and the theory of difftisions, processes,
and martingales.
So the book is on the interface and applications of probability, random walk, and
path-space measures to convergence questions in harmonic analysis and dynamics.
List of names and discoveries
Many of the main discoveries summarized below are now lore, and they are often
used inside the book without explicit mention of the pioneers by name. (For page
numbers of the indicated chapters and figures, refer to Contents on pages ix-xiii
above and Figures on pages xxxix-xlii below.)

List of names and discoveries

xxxi

1807
Jean Baptiste Joseph
Fourier
mathematics, physics
(heat conduction)

Expressing fiinctions as sums of sine

and cosine waves of frequencies in arithmetic progession (now called Fourier series). Received somewhat late recognition for his work.

Section
6.2
[Zyg32]

1909
Alfred Haar
mathematics

Discovered, while a student of David

Hilbert, an orthonormal basis consisting of step fiinctions, applicable both to
fionctions on an interval, and functions
on the whole real line. While it was not
realized at the time, Haar's construction
was a precursor of what is now known as
the Mallat subdivision, and multiresolution method, as well as the subdivision
wavelet algorithms.

Chapters
5, 6, 7,
see also
Figures
1.2, 6.1,
7.4
[Haa 10,
Waln02]

1946
Denes Gabor
(Nobel Prize): physics
(optics, holography)

Discovered basis expansions for what

might now be called time-frequency
wavelets, as opposed to time-scale wavelets, the subject of this book.

[JaMROl,
GroOl]

1948
Claude Elwood Shannon
mathematics, engineering
(information theory)

A rigorous formula used by the phone

company for sampling speech signals. Quantizing information, entropy,
founder of what is now called the mathematical theory of communication.

Section
1.4
[Sha49,
AlGrO 1 ]

1935
Andrei Nikolaevich
Kolmogorov
information theory,
probability, statistics

Pioneered the use of probability theory

in analysis, especially the use of measures on path space, the Kolmogorov
consistency relation.

Chapters
1,2,5
[Kol77,
Wil91,
Nev65]

1976
Claude Garland, Daniel
Esteban (both)
signal processing

Discovered subband coding of digital

transmission of speech signals over the
telephone

Chapter 7,
Figures
7.7,7.14
[JaMROl,
StNg96]

xxxii

Getting started

1981
Jean Morlet
petroleum engineer

Suggested the term "ondelettes." J.M.

decomposed reflected seismic signals
into sums of "wavelets (Fr.: ondelettes)
of constant shape," i.e., a decomposition of signals into wavelet shapes, selected from a library of such shapes
(now called wavelet series). Received
somewhat late recognition for his work.
Due to contributions by A. Grossman
and Y. Meyer, Morlet's discoveries have
now come to play a central role in the
theory.

Section
7.2
[JaMROl,
GrMo84]

1984
Alex Grossman
mathematics,
quantum physics

Pioneered the theory of coherent

(quantum-mechanical) states, precursors for dual wavelet systems. Mentor
for, and coauthor with, J. Morlet and
I. Daubechies.

Sections
8.1-8.2
[Dau92,
MeCo97,
GrMo84]

1985
Yves Meyer
mathematics,
applications

Mentor for A. Cohen, S. Mallat, and

other of the wavelet pioneers, YM. discovered infinitely often differentiable
wavelets.

[JaMROl,
Mey89,
MeCo97,
Mey97]

1989
Albert Cohen
mathematics (orthogonality relations),
numerical analysis

Discovered the use of wavelet filters in

the analysis of wavelets^the so-called
Cohen condition for orthogonality.

Chapters
5,6
[Coh90,
Dau92,
Law91a,
Law91b]

1986
Stephane Mallat
mathematics, signal
and image processing

While still a graduate student in engineering, working on vision, and on

the Littlewood-Paley octaves, formalized and discovered what is now known
as the subdivision, and multiresolution
method, as well as the subdivision wavelet algorithms. This allowed the effective use of operators in the Hilbert space
Z^(R), and of the parallel computational
use of recursive matrix algorithms.

Chapters
5,6,7
[Dau92,
Mal89,
Mal98]

List of names and discoveries

xxxiii

1987
Ingrid Daubechies
mathematics, physics,
and communications

Discovered differentiable wavelets, with

the number of derivatives roughly half
the length of the support interval. Further found polynomial algorithmic for
their construction (with coauthor Jeff
Lagarias; joint spectral radius formulas).

Chapters
1, 7,
see also
Figures
7.16,7.5
[Dau92]

1991
Wayne Lawton
mathematics
(the wavelet
transfer operator)

Discovered the use of a transfer operator

in the analysis of wavelets: orthogonality
and smoothness.

Chapters
8,9
[Law91a,
Law91b,
Dau92]

1999
Richard Gundy
use of probability in
the analysis of filters

Discovered the use of martingales and of

Kolmogorov's 0-1 law for the rigorous
analysis of wavelet filters in the analysis
of wavelets^refined and corrected the
so-called Cohen condition, orthogonality condition for wavelet filters.

Chapters
1, 5, 6
[Gun99,
Gun04]

1992
The FBI
using wavelet algorithms in digitizing and
compressing
fingerprints

C. Brislawn and his group at Los Alamos

created the theory and the codes which
allowed the compression of the enormous FBI fingerprint file, creating A/D,
a new database of fingerprints.

Chapter 7
[Bri95,
BrRo91,
JaMROl]

2000
The International
Standards
Organization

A wavelet-based picture compression

standard, called JPEG 2000, for digital
encoding of images.

Chapters 3,
7, especially
the figures
[JaMROl]

1994
David Donoho
statistics,
mathematics

Pioneered the use of wavelet bases and

toolsfi*omstatistics to "denoise" images
and signals.

Chapter 7
[DoJo94,
JaMRO 1 ]

xxxiv

Getting started

General theory
We consider a measurable space X with an endomorphism a, mapping onto X, such
that the inverse image of every point is of the same finite cardinaUty. The branches of
the inverse are assigned probabiUties by some positive fimction W, and we study the
corresponding transition operator, also called the Perron-Frobenius-Ruelle operator
R = Rff^.By iteration, the branches determine a tree, and we study an associated
random walk on this tree, and the transition measures indexed by the points in X.
While there is a considerable literature on this setup already, the various papers
place some kind of regularity condition on W, or on the system, (X, a), and the
branches of the inverse. Our setting, assumptions and conclusions are in the measurable category, hi contrast, when X is assumed to have a dififerentiable structure,
and W is assumed Lipschitz, then, by Ruelle's theorem, there are solutions to the
equations vR = v, Rh = h, and v(h) = 1, where v is a Borel probability measure
on X, and his a non-negative measurable fimction on X.
While the measure v may not exist in the general measurable setting, we develop
formulas for solutions h to the equation Rh = h, the so-called 7^-harmonic fimctions, and we give applications to the case when R is the wavelet transition operator
defined fi-om some measurable low-pass filter. In that setting, there are existence
questions, and we show that the random-walk properties determine a variety of notions of wavelet orthogonality properties.
A word about the graphics and the illustrations
We owe much to the professional skill of Brian Treadway in programming and creating the computer generated renditions of the numerous figures in this book. They
are an essential part of the exposition. A list of our numbered figures, giving their
captions and the page nimibers where they appear, follows this "Getting started"
section. This list includes the text for our figure captions. Readers are encouraged
to preview the figures themselves to appreciate our use of trees, graphs, programming diagrams, subband-filtering schemesfi*omengineering (signal/image processing), and a number of other visual tools and presentations. The author has himself
found the figures, the graphs, and visualization of the algorithms exceptionally helpfiil in learning, teaching and discovering some of the material in this book; but more
importantly, also in understanding the deeper connections in the subject.
The Mathematica program used in the production of Figure 7.5 (pp. 120-121)
by Brian Treadway is given in fiill in the narrative in the "References and remarks"
section at the end of Chapter 7 (pp. 152-153). The two figures 7.4-7.5 (pp. 118-121,
each figure spanning two pages) serve to illustrate a particular aspect of the so-called
pyramid algorithm for wavelets, i.e., the recursive algorithm which is used in among
other things in the creation of wavelet packets; see, e.g.. Figure 7.3 (p. 113). For this
purpose, the algorithm is used in the two figures with two different initializations.

Special features of the book

xxxv

The first is the simpler one, the Haar scaling function (see Figure 7.4, pp. 118-119),
and it makes the dyadic subdivision especially transparent. The second initializes
with Daubechies' scaling fimction (see Figures 7.5 and 7.16, pp. 120-121 and 134).
The reason for the name "wavelet packet" (which is the subject of Chapter 7) is especially transparent in Figure 7.5. Notice especially two aspects of the progression of
functions in the sequence of graphs inside the respective figures: In moving from one
graph to the next, the numerical frequency appears to increase with each subdivisionstep. But in addition, an enveloping shadow emerges in the progression through the
graphs, and a shape of a wave with a lower frequency (i.e., longer wavelength) appears to "capture" and group the functions themselves into packets, much like the enveloping "beats" from music composition. We mention these figures already now, as
the geometry of the steps that go into the diverse recursive algorithms are especially
transparent to the naked eye in Figures 7.4-7.5. See also the sequence of figures
7.6-7.13 (pp. 123-128) for the programming aspects of the same idea. Specifically,
Figure 7.13 stresses the matrix steps that are indicated by the recursions.
We fijrther stress that these figures play a central role in our presentation in this
book: Some figures illustrate the kind of self-similarity in time and in space coordinates that is typical both in fractal analysis, and in the study of wavelets; while others
illustrate decision trees; and yet others make clear the kinds of arrow-flow diagrams
which are popular in building of recursive algorithms, and in progranmiing more
generally. See, e.g., [Knu84] as well as Knuth's monumental volumes [Knu81]. We
mention Knuth's article [Knu84] as it emphasizes in a special case both the stochastic
and the algorithmic features of the fundamental subdivision/filter algorithm.
For explanation of the cover figure, see "About the cover figure" at the front of
the book (p. vi).
Special features of the book
A main aim of this book is to show how these ideas have proved fixiitful in both
the study of iterated function systems (IFS), see Chapter 4, and of wavelets, see
Chapter 7. While the pyramid algorithm (in its diverse incarnations) is now typically identified with wavelets, it in fact has a long history in engineering, information theory [Sha49], and symbolic dynamics. The coimection to signal and image
processing from engineering was emphasized in our survey article [Jor03]; but see
also [D0MSSO3, Mal98, StNg96] for much more detail. The connections to symbolic dynamics are manifold, but we wish to especially recommend the beautiful and
inspiring invitation [Rad99] addressed to students, and authored by Charles Radin.
The book contains four separate items which we hope will help readers reconcile current terminology used simultaneously in mathematics and in applications
(especially in signal-processing engineering, and in physics^notably in optics!)
These four elements are as follows: (i) a system of interrelated appendices (pp. 205222), (ii) a list of comments for mathematicians on signal/image

xxxvi

Getting started

processing jargon ("Afterword," pp. 223-230), and (iii) a list of names (with historical comments) of mathematicians and scientists, both past and current era mathematicians/engineers/scientists who made pioneering contributions to the main ideas
presented in the book (pp. xxx-xxxiii above).
Finally, in item (iv) we include a glossary consisting of a list of terms occurring
in the book in varied contexts of mathematics, probability, engineering, and on occasion physics (pp. xvii-xxv above). To reduce the apparent confusion created by
the same concept having up to four different names, the glossary includes informal
explanations spelling out the reasons behind the differences in current terminology
from neighboring fields.
hi item (i) we attempt to translate the various engineering terms and constructs
into mathematical formulas. This is a continuation of a theme we started in an AMS
Notices article [Jor03]. For (iii), we apologize for the subjective nature of our comments; and we readily acknowledge the diflficulties in writing history of events that
are rather modem on a mathematical scale. All three concluding additions to the
book are motivated by the nature of the subject at hand, specifically by the many
connections between ideasfi-omsignal/image processing and the more mathematical
themesfi*omwavelets,fi-actals,and dynamics.
Exercises: Overview
The exercises are essential, and they serve several pedagogical purposes: They are
there to help students and users practice the fimdamental concepts in the book, and
to test his/her hand at computations, and at sketching fimctions or iterative schemes.
But they also serve to expand horizons! We hope students will acquire a handson feeling for the various basis fimctions already discussed in the book; and more
importantly, get started at building bases of his/her own. Some readers might even
want to begin with the exercises, and then read the text as they work along in the
exercises, hideed, in designing the exercises, I have been inspired, at least in part by
books with a philosophy like this: "Learn Hilbert space by doing problems!" e.g.,
[Hal67]. Or: "Operator algebras by example!" e.g., [Dav96]. Finally, several of our
exercises have been put in to expand on themes inside the text, pointing the reader
to new developments in the subject, and especially stressing links to neighboring
subjects and to applications; pointing to connections between modem trends and
classical subjects, highlighting an especially powerfiil (and beautifiil) idea, or even
suggesting unorthodox applications.
One exercise in Chapter 7 is different fi*om the others; it is a multipart exercise
involving a search on the Intemet. The student is asked to browse images (fi-om
image-processing engineering) on the web, and to follow visually how a certain (geometric and algorithmic) cascading process is made concrete. The mathematics of
the cascade is worked out in Chapter 7 itself

Exercises: Overview xxxvii

The exercises vary, with variation both within a given chapter, and variation from
chapter to chapter. The longer chapters are supported by proportionally more exercises. Some of the exercises are short, and some are long. Some are conceptual, and
others are computational. Some are hard, and a few are relatively easy. The latter
are meant mainly for the student to practice the use of one definition or another. Yet
others serve to extend material covered in the book, and again others are there to help
the student review background. Then there is a sliding scale up to the more difiicult
exercises, but the reader will then find hints. Finally, some exercises are built up in
steps as multipart exercises with many parts. So what is a single numbered exercise,
for example Exercise 7.9, may have many parts, (a), (b), etc., and they progress in
logical steps, often with one part building on the previous, hi all, the exercises make
up more than 40 pages.
We have presented the material so that different readers can select the parts of it
that are closest to his/her own interests; and in particular, it is not necessary to begin
with Chapter 1. Li fact, for some it might be better to begin with the Appendices,
or with the Afterword containing the special sections "Comments on signal/image
processing terminology" and "Computational mathematics," or for some insight into
the history of the subject, the "List of names and discoveries" on pages xxx-xxxiii
above. While some chapters build on earlier ones, non-sequential reading is possible,
especially with the use of two extensive index lists, an index of "Symbols" and a
general "Lidex," which are the last items in the book. The following quote may be
appropriate for interdisciplinary mathematics books.
Det er ganske sandt, hvad Philosophien siger, at Livet maa forstaaes
baglcends. Men derover glemmer man den anden Scetning, at det maa leves
forlcends. Hvilken Scetning, jo meer den gjennemtcenkes, netop ender med,
at Livet i Timeligheden aldrig ret bliver forstaaeligt, netop fordi jeg intet
Oieblik kanfaaefuldelig Ro til at indtage Stillingen: baglcends. ^
Soren Kierkegaard
Iowa City,
June 2006

Palle E. T. Jorgensen

^ "It is really true what philosophy tells us, that life must be understood backwards. But
with this, one forgets the second proposition, that it must be lived forwards. A proposition
which, the more it is subjected to careful thought, the more it ends up concluding precisely
that life at any given moment cannot really ever be fully understood; exactly because there
is no single moment where time stops completely in order for me to take position [to do
this]: going backwards." Often shortened to "Life can only be understood backwards; but
it must be lived forwards" (Livet skalforstaas baglcens, men leves forlcens).

Figures. Read Me!

Le plus court chemin entre deux verites dans le domaine reel passe par le
domaine complexe.
^Jacques Hadamard

1.1

1.3

N = 2, simple dyadic branch points; a fimction W:X -^ [0, 1 ]

is given. A path in the random-walk model. Assignment of
probability to a particular length-five path: Pi^^ ({(0 1 10 1)}) =
W{T()X) W{T:\TOX) W{TITQX) ^(ror^rox) W{T\TQT^TOX)
Haar wavelets: The two fimctions cp and y/ are the father, resp., the
mother fimctions for Haar's construction in cases (a) on the left, and
the stretched Haar (b), on the right. In each case, we use y/ to create
a double-indexed basis system (1.3.16). For (a) this system will be
an orthonormal basis (ONB) in X^(M); but for (b), the fimctions
(1.3.16) will only represent a Parseval fi-ame, i.e., a fimction system
which yields the Parseval identity (1.3.22) for every function in
Z^(R). As is apparent from (b), this Parseval system will not be an
ONB for the stretched version. We resume the discussion of this in
the text following Figure 6.1 (p. 103).
The slanted matrices F and G of (1.7.2)

13
24

2.1

The Farey tree

42

4.1
4.2
4.3
4.4

Cantor sets
The Cantor set X3
The quarter Cantor set X4
Alternate limiting approach to the quarter Cantor set X4

73
74
74
75

1.2

xl

Figures. Read Me!

4.5
4.6

X4, the conjugate or left-handed quarter Cantor set

Alternate limiting approach to the conjugate quarter Cantor set X4

75
76

5.1
5.2

The 2-adic
Powers of ^

89
90

6.1

Haar's wavelets (compare Figure 1.2, p. 13): Haar's twofimctions(p

(6.2.3) and y/ (6.2.4) are illustrated in the two columns: The system
in the column on the left in the figure illustrates Haar's orthonormal
(ONB) wavelet basis, and the scaling identities are visually
apparent. The system of fimctions ^3 (6.2.5) and y/3 (6.2.6), for
/? = 3, in the right-hand column illustrates the two stretched Haar
fimctions. The two scaling identities (6.2.1) and (6.2.2) are visually
apparent for both function systems, on the left and on the right.
Moreover, for both systems, the y/ fimction yields a double-indexed
Parseval basis (see (1.3.16)), i.e., a fimction system for which the
Parseval identity (6.2.12) holds for every fimction / in Z^(IR). If
p = I (the left-hand-side case), then the inequality in (6.2.11) is an
equality, i.e., is an =; while for /? = 3 (the right-hand-side column
in the figure), the term on the left in (6.2.11) is strictly smaller than
||/||^2(m fo^ non-zero / . The reason for this is the overlap for the
Z-translated fimctions, which is graphically illustrated too
Solutions to the eigenvalue problem (6.2.19)

6.2
7.1

7.2

7.3
7.4
7.5
7.6

fractions

The Fourier transform of the basis sequence (before scaling and

selection) generated from the scaling fimction ^0, using the fiiU
combinatorial tree
The generation of the basis fimctions ^0? ^1? ^2? . . . by use
of the algorithm (7.5.5) initiated with (po = (p such that
(p{t) = 2 ^ aj(p{2t 7), the fimctional equation by which, in the
first (A) branch, (po "re-generates" itself.
The generation of the basis fimctions ^ 1 , (p2, ^3, . . . by use of the
algorithm (7.5.5) initiated with (p\ = y/
The first thirty-two fimctions in the sequence cpn in the case of the Haar
wavelet construction
The first thirty-two fimctions in the sequence cpn in the case of the
Daubechies wavelet construction
A decreasing family of resolution subspaces FQ D V\ D F2 D ,
and corresponding detail spaces W\^ W2,
Dyadic wavelets.
Vn = Vn^x + Wn^un = OA,

103
107

Ill

112
113
118
120

123

Figures. Read Mel

7.7

7.8
7.9

7.10

7.11
7.12
7.13
7.14
7.15
7.16
8.1
8.2
8.3

Subband filtering for dyadic wavelets: mo = low-pass filter,

m\ = high-pass filter, down-sampling, (J) up-sampling. The
figure refers to input (fi*om left) and output (to the right) of signals
fi-om standard signal processing; for example, of speech signals.
From the left, we start with (signal) input. The signal is then split
into two frequency bands with two filters mo and m \. The first filter,
mo, passes low frequencies, and the second, wi, passes high. The
split into frequency bands is called "analysis." We get two signal
strands which are now each followed by signal down-sampling. As
we move right, the two bands are then up-sampled. The processed
signal bands then pass dual filters, and are finally merged again with
H- into an output signal. "Perfect reconstruction" means that the
output signal to the right is matched up perfectly with the input from
the left. From engineering, we know that it is possible to find filters
mo and m\ that achieve perfect reconstruction; i.e., recovering the
input signal by synthesis of the bands. Magically, and by hindsight,
these filter systems serve at the same time to give us orthogonal
wavelets, subject to a technical condition discussed in Chapter 5.
(There can be more than two frequency bands, and we refer to
Section 7.6 for the general case.)
Wavelet algorithm in the dyadic case: = Vn+\ + ^w+i, w = 0, 1,
Wavelet algorithm in the 7V-adic case: Vn-\ = Vn -\- Wn,
Wn = WP + WP + .. + wP-^\ n = l,2,....
Multiple
subbands mi,..., mjsi-i
A decreasing family of resolution subspaces FQ D Fi D F2 D ,
iV-adic case, N = 4, Vn-\ = F + Wn. Wn = wP + wP + wP,
n = l,2,..
A selection of a dyadic wavelet packet. (An example of a path in a
pyramid!)
Illustration of a path in the pyramid: wo -^ m\ -^ m\ ^ mo-^ m\ ^>,
orVo-^ Vi^ W2-> W3^ V3^ W5
Combined filter/down-sampling in matrix form
Perfect reconstruction of signals
Iteration of the filter/down-sampling operation as it arises in wavelet
algorithms
Daubechies wavelet fianctions and series of cascade approximants
The fimdamental dyadic pyramid
Branching with step size 3
The pyramid on E2 = {2,-1} (may be seen as a subdiagram of
Figure 8.2, p. 161)

xli

124
125

126

126
127
127
128
132
133
134
159
161
173

xlii

Figures. Read Me!

C.l

Representation of S = TU*. This and the next diagram illustrate

the use of the operator relations from Appendix C in synthesis of
subbands of a signal, or of a subdivision of an image. The two
diagrams form a pair. The present first one represents a matrix
action transforming one band-configuration into another. This is
followed by up-sampling applied to each strand. There is then a
system of translations before the strands are merged and added. This
diagram is a natural dual version of the next one, Figure C.2, i.e.,
the analysis step which begins with subdivision, or breaking apart
an input signal
C.2 Representation of 5* = UT*. The present two figures together.
Figures C.l and C.2, represent equivalent formulations of the
respective left-hand and right-hand sides in Figure 7.7 (p. 124). But
the present figures in fact represent signal-processing algorithms
(or equivalently wavelet algorithms) that are more versatile than
the more traditional polyphase matrices: specifically, the matrix
fimctions in our present diagrams may be arbitrary, and be of
arbitrary size. The size of a polyphase matrix equals the number
of subbands which is used. But we also allow non-unitary matrix
fimctions. Of course, more general choices of matrices in Figures
C.1-C.2 will then correspond to more general choices of filters in
Figure 7.7. These choices are highly relevant, for example for the
algorithms based on lifting, see [DaSw98]. Our present Figures
C.1-C.2 are multiband versions of diagrammatic representations of
subsampling/subband-filtering algorithms which appear frequently
in the signal-processing engineering literature; see for example
[WWWl] with text, and [WWW2] and [WWW3] with pictures

214

215

Acknowledgments

an apt comment on how science, and indeed the whole of civilization, is a

series of incremental advances, each building on what went before.
Stephen Hawking
While working on the book, we had numerous discussions with Dorin Dutkay,
Wayne Lawton, Wai Shing Tang, and with the members of the 2004 Singapore Workshop on "Functional and harmonic analyses of wavelets and frames," 4-7 August
2004, at the National University of Singapore. Li particular, Larry Raggett, Kathy
Merrill, and Judy Packer made helpful suggestions.
We are pleased to acknowledge many helpftil conversations with Roger Nussbaum and Richard Gundy while we visited Rutgers in May 2003. We had very helpful e-mail discussions in March 2004 with Richard Gundy. We also acknowledge
numerous helpful discussions with Anilesh Mohari.
We are especially grateftil to Brian Treadway. With his usual professionalism he
produced this beautifiil typeset text, provided numerous corrections, and created the
figures and the graphics, including the graphic design and Mathematica programming of iterative algorithms. He also made a number of extremely helpfiil suggestions.
We had constructive conversations with the members of our NSF Focused
Research Group (FRG), especially with Akram Aldroubi, Chris Heil, David Larson,
and Gestur Olafsson.
David Larson kindly read our final manuscript and offered encouragements and
suggestions.
The key idea of using a transfer operator in the sense of random walks and of Perron, Frobenius and Ruelle for the analysis of wavelet basis systems was originally
pioneered in the two papers [Law91a, Law91b] by Wayne Lawton. The reader will
notice that this idea runs like a red thread through our present exposition. Indeed, the
two papers of Lawton have influenced not only our present approach, but over the

xliv

Acknowledgments

years, they have served to bring closer central ideas from probability and analysis in
the study of classes of basis problems in ftinction theory and in applications. These
include the admittedly best-known such bases, the wavelets. By now, variants of the
transfer operator, alias the Perron-Frobenius-Ruelle operator, serve as key tools in
diverse areas of mathematics: in ergodic theory and dynamical systems, the analysis of fractals, iterated fiinction systems (IFSs), and geometric spectral theory; see,
e.g., [BalOO, BaCM02, Coh90, Coh03, Dau92, DaHRSOS, DuJo06b, Dut02, Fal90,
FaLa99, FeLa02, Gun99, Gun04, JiJSOl, JoPe98, LaNg98, PaPo90, StNg96], which
is but a partial list. Our present exposition in this book owes a great deal to Lawton's
pioneering insight, and we are further pleased to acknowledge numerous discussions
with him during the period of preparation of this book.
The author is also pleased to thank his colleagues and students at the University
of Iowa; he benefited especially from our joint seminars and from helpful discussions
with Tom Branson and Paul Muhly (in mathematics), and with Wayne Polyzou (in
physics).
We thank our students in the course 22M:321, "Wavelets," taught in the spring
of 2005, especially Myung-Sin Song, who was finishing her Ph.D. thesis in image
processing at the time. We extend our thanks to Chris Brislawn of the Los Alamos
National Laboratory for illuminating discussions and for guest lectures in 22M:321,
and to two colleagues in the University of Iowa School of Engineering, Andrew Kusiak and Albert Ratner, of the Department of Mechanical and Industrial Engineering.
We benefited enormously from many discussions with these colleagues, and with
others, and we hope in the book to convey some of the excitement we experienced in
communicating across disparate disciplines.
Several recent Springer authors have been generous with encouragement and
with tips for preparation of the manuscript. We are pleased to thank William Arveson. Ward Cheney, Weimin Han, and Bruce Sagan.
A number of friends have been kind and generous with valuable advice. Here
Peter Renz's wise suggestions stand out.
Finally, we thank Ann Kostant and the series editors for constructive suggestions.
One notable suggestion which I took to heart: I was persuaded that it is poor taste
for textbook authors to sprinkle the narrative with citations to journal articles and
advanced books. The presentation of main ideas is self-contained from first principles
and not distracted! Yet I want to assure researchers that their work does get cited. Be
patient! Citations are concentrated in sections at the end of each chapter, "References
and remarks;" and there is a "History" section at the end of Chapter 1. In addition,
in the "Getting started" section above, we have a section called "List of names and
discoveries" in which a number of pioneers in wavelets and signal processing are
celebrated. Many undoubtedly have been omitted; and we apologize for that.
At least this principle of postponing citations has been followed within reason!
This material is based upon work supported by the U.S. National Science Foundation under Grant No. DMS-0139473 (FRG).

Analysis and Probability

The image collage on the next page illustrates central themes in the book:
The three oscillatory graphs placed down diagonally in the collage represent the kind of functions generated algorithmically by tree-like, or so-called
pyramid, algorithms. These recursive algorithms are structured around
pyramid shapes as in the top right corner of the collage, and they come
from successive repetitions of dyadic branching steps from the sketch in the
lower left corner of the collage. See also ''About the coverfigure " (page vi).

>0

SYNTHESIS

-
^dual low-paps filtM
I

tsampling

SIGNAL O U T

dual high-pass filter

-(D^1

Introduction: Measures on path space

We believe that all curious people can enjoy and understand great mathematical ideas without having to brush up on garden-variety school math or
^Edward B. Burger, Michael Starbird
relive their painful algebra daze.

A sense of curiosity! Algebra of the integers and the real numbers; rudimentary facts about functions and measures; limits; compactness; Cantor;
matrices; inner-product spaces.
Question: "What if I don't have the prerequisites for reading the prerequisites?"
A: "Doesn't matter! Skip Chapter 1!"
PREREQUISITES:

Prelude
In the first section of this chapter, we introduce the fundamental concepts of wavelet
fimctions and wavelet constructions, but in a form which stresses scale-similarity.
This scale-similarity is a special case of a more general notion of self-similarity.
We will later see that the more general idea of self-similarity is needed for our understanding offi*actalsand symbolic dynamics.
One of the difiiculties in moving fi-om the traditional setting of wavelets to the
more general iterative models is the distinction between linear and non-linear: Recall
that wavelets are built in the linear space E^, which ui turn comes equipped with its
canonical (i-dhnensional Lebesgue measure. By contrast,fi*actalsare typically nonImear structures, not even groups; and one of the first issues confi-onting us will be
that of finding a substitute for Lebesgue measure. As it turns out, such substitute measures exist, having certain mtrinsic properties quite analogous to those of Lebesgue
measure; but the new measures depend on the particular fi-actal X under discussion.
Once a particular X is specified (by a finite system of affine mappings), we will
see that it acquires a natural and intrinsic fi*actal dimension (called the Hausdorfif

1 Introduction: Measures on path space

dimension, say s), and a corresponding fractal measure called the Hausdorff measure. However, note that there is an 5--dependence. Specifically, we have the existence
of a Hausdorflf measure jUs on X for each value of 5. And for the traditional case of
R^, we have s = d. Later we shall outline iterative schemes which yield JULSFurther into the book, we will then extend both Fourier tools and wavelets to
these fractals. The class of fractals we consider is somewhat restricted. But these are
the fractal classes where a reasonable harmonic analysis is feasible.
To make the ideas more concrete, we concentrate initially on simple and very
explicit fractals such as the middle-third Cantor set, and other Cantor sets on the line.
But the various more general points and constructions will only emerge gradually as
the subjects in the book progress in a natural order. Hence, Section 1.1 will include
also a brief chapter-by-chapter outline of the central ideas in the book, stressing
throughout how our random-walk model is used.
To help highlight wavelets, fractals, and signals, throughout the book, the narrative is illustrated with figures. Each figure is there to help students visualize main
ideas, and many of the figures are created from recursive Mathematica programs, by
Brian Treadway.
And in addition to the figures inside the text, each chapter contains little graphical
ornaments or "dingbats." They should help readers navigate the book, and they also
serve as reminders of the three themes. Some dingbats mark separation between the
prelude sections and the rest of the chapter, while others are placed near the end of
each chapter to mark the start of the exercises.
A closer inspection reveals that the dingbats that follow the chapter preludes are
all different. Some dingbats represent certain selections from families of images.
Within each family though, the dingbat images are different.

1.1 Wavelets
Wavelets are families of fimctions of one or several real variables, i.e., fimctions on
R or on E^, satisfying three properties:
(i) They form a basis for Z^(M) (or X^(R^) in the several-variable case) and have
suitable orthogonality properties which we spell out below.
(ii) They are indexed by integer translations, i.e., by the operation x -> x + A:, for
A: G Z, and by integer powers of a scaling. Often the scaling is dyadic, i.e.,
scaling is by powers 2^, for j e Z.

1.1 Wavelets

(iii) There are two generating fiinctions (p and ^ (called father and mother functions)
for the whole double-indexed wavelet which relate the two operations of scaling
and translation in the following way:

^ i \ ^ ( 2 . -K), ^ = X ^^^(^ * ~*)'

where the coefficients i \ and Q}^ are carefully chosen.
Comments on (iii). The choice of the coefficients is essential, and it has both
an algebraic and a geometric flavor. We will see that various admissible classes for
the coefficient systems {Pi, Qt} may be thought of as quadratic varieties, and there
are solutions both in finite and infinite dimensions. The two relations in (iii) above
represent only the dyadic case of what is called scaling relations. The first is an
equation for a function ^, and is called the dyadic scaling equation. The function cp is
called a scalingfunction. The context and the ambient function space dictate choices
for the Pi's, which in turn yield non-trivial solutions cp. Moreover, the possibilities for
solutions (p are extremely sensitive to these choices, and we shall develop algorithms
and discuss stability for solutions (p and y/.
The coefficients Pi are called masking coefficients, and conditions on these coefficients must always be imposed. This is a crucial point which will be taken up in
Section 1.3 below, and in more detail again within the book itself.
The context for solutions: Our context will include both standard Lebesgue
measure and some classes offi*actalmeasures as well. Specifically, we shall consider
higher dimensions, such as E^, and more general scaling operations (e.g., matrix
scaling). We will further introduce a variety of fi-actal structures which come with
associated scaling relations. But in (iii) we have listed the simplest scaling relations
only to illustrate the general idea. Special cases of classes of conditions for the coefficient systems {P/, Qi] are known as quadrature conditions, or quadrature-mirror
conditions. The ambient fionction space might be simply the familiar Hilbert space
L?- (M), or it may be a Hilbert space built on some affine fi*actal, which in turn is
designedfi-oma matrix scaling and an iterated function system of affine maps in R^
(definitions in Section 1.3 and Chapter 4). However, depending on the context, the
admissibility conditions for the coefficients Pi are different.
The various problems connected with classes of scaling relations have traditionally been addressed with only standard tools fi-om analysis. Here we will instead
employ a suitable mix of analysis and path-space methodsfi*omprobability.
In these problems, solutions are constructed fi*om an algorithm which relates a
certain function (pit), t E:^, called a scaling fimction, to its scaled version (p{Nt)
where TV is a fixed integer, 2 or more, or an expansive integral d-hy-d matrix ^ if ^
is in R^.
For d = \, the formula which determines (p is (1.3.1) below. It is called the
scaling identity. (Note that the first relation in (iii) is a special case of (1.3.1).) It

1 Introduction: Measures on path space

is not at all clear a priori that this identity should have L^ solutions at all, or even
solutions that are given by some kind of convergent algorithm. The coefficients in
(1.3.1) are called masking coefficients, filters, or filter coefficients. For special values
of the coefficients, it turns out that there is a normalized Z^ solution cp, and that
pointwise convergence of a good approximation can be established in a meaningfiil
way. Moreover, we show that the convergence behavior is dictated by properties of
an associated transfer operator i?, or Ruelle operator.
The expression on the right-hand side in equation (1.3.1) is also called a subdivision because of the function values (p{Nt A:). An iteration of the operations on the
right-hand side in (1.3.1) on some initial fimction is called a cascade approximation.
This approximation is one approach to the fiinction ^, and the other is the infinite
product formula (1.3.5) for the Fourier transform of ^.
The relation between cp and its scaled refinement is well understood when we
pass from the time domain to the frequency domain via the Fourier transform. In that
case the relation is multiplicative, and involves a certain periodic matrix fimction m,
called the low-pass filter. For fiirther details, see formula (1.3.4) below.
The study of the filters m is part of signal processing (see, e.g., [Jor03, StNg96]).
But by a "mathematical miracle," they have become one of the most useful tools in
wavelet constructions; and at the same time, they have pointed to a host of exciting
applications of wavelet mathematics. To get a path-space measure for some of the
wavelet problems, we use the quadrature-mirror properties, or their generalizations,
which are assumed for the filter fiinction m, also called a frequency response fimction. It is periodic, in one or several variables. In R^, there is a variety of choices of
a period lattice for the problems at hand.
Since the first question is to decide when ^ is in L^, we iterate and get an infinite
matrix product involving the matrix W := m*m. Since W is positive semidefinite,
we may create a positive path measure of a random walk starting at x in some period interval. In several dimensions, x starts in a fimdamental domain D for some
fixed lattice, for example Z^. The paths starting at x arise by iteration of the inverse
branches of x -^ ^x mod l/. There are N = |det^| distinct branches. These N
branches may be viewed as endomorphisms of D.
We construct our random walks in a general framework which includes both
wavelets, wavelet packets (see Chapter 7), and some of the other more classical
problems. We fiirther show how some of the classical questions may be phrased
and solved with the use of path-space measures.
It is interesting to contrast our proposed approach with the more traditional one
used in wavelet analysis: see, e.g., [Dau92]. Traditionally, some kind of Lipschitz
or Dini regularity condition must be assumed for the filter. Then the corresponding
infinite product may be made precise, and we can turn to the question of when the
wavelet generators are in I^(M^). As it turns out, both of these issues have natural
formulations and solutions in terms of the path-space measures. The results allow
a wider generality. In a variety of wavelet questions for band-limited wavelets, it is

1.1 Wavelets

just an unavoidable fact that the regularity conditions are not satisfied for the filters
m that are dictated by the setting and the applications.
Before turning to the details, we sketch a brief outline. We resume the summary
in more detail in Section 1.6 below. The remaining sections in this chapter contain
more definitions, motivation and examples, which will be used in the overview at the
conclusion of Chapter 1.
Li Section 1.2, we introduce the basic concept of a probability space built on
covering mappings with a fixed number of branches. These systems give rise to random walks on trees, paths, and transition probabilities. We discuss these notions in
Section 1.2.
hi Section 1.3, we illustrate the relevance of the path-space measures to some
questions in wavelets, and in Section 1.4, to sampling theory. To make the questions
more concrete, we outline four examples at the end of Section 1.3, two variants of
Haar's wavelet construction, Daubechies' wavelet, and finally we sketch an analogous harmonic analysis question for the middle-third Cantor set.
Li Section 1.5, we prove a theorem on pointwise convergence of a class of infinite
products. Our theorem uses the transition probabilities which are associated with the
random walks that we introduce in Section 1.2 below.
Chapter 2 develops general theory, transition probabilities, branching processes,
and corresponding probability spaces. This material has a more technical flavor.
hi Chapter 3 we specialize the measures, constructed in a general fi-amework in
Chapter 2, to the concrete context of wavelets in the Hilbert space Z^(E). hi Chapter
2, we introduced transition probabilities Px on a probability space Q; and in the
wavelet application, we outline how the group of integers Z is naturally embedded in
Q. And we show that the orthogonality issue for wavelets is equivalent to Z having
fiill measure in Q.
Li Chapter 4 we turn to a class of aflBne fi-actals, iterated fimction systems (IPS)
by afiine maps, but with focus on certain Cantor sets, mainly the middle-quarter
Cantor set. For the affine IFSs we show that a natural basis question is equivalent to
the natural numbers No having fiiU measure in Q. For this formulation we outline a
conjugate system of Cantor sets.
Chapter 5 addresses the general case of measurable systems constructed fi-om a
given 7V-to-one mapping. Li this general context, we give a sufiicient condition for
Z havingfiiUmeasure in Q.
hi Chapter 6 we resume the discussion of wavelets, but in a non-commutative
setting motivated by multiwavelets.
hi Chapters 7 and 8 we turn to two measure-theoretic issues fi-om the general
fi*amework of Chapter 5: There is a class of representations of two familiar C*algebras which produce IFSs in the measure category, the Cuntz algebra, and the
algebra of the canonical anticommutation relations. Among other things, we show
that the determinantal measures studied by Russell Lyons and others in combinatorial
probability theory are special cases of this.

1 Introduction: Measures on path space

Chapter 8 contains some examples of wavelet packets and permutative representations. We give a systematic presentation of this subject, and we show that the
concrete examplesfitwithin a general measure-theoretic fi"amework. While the examples have been studied in research papers, they have so far not found a unified
formulation. In Chapter 9 we show that the underlying geometric questions for these
examples may be phrased in the context of operators in Hilbert space.
Our selection in Chapters 8-9 and in the appendices includes certain geometric
topicsfi*omoperator algebras and Hilbert space. They have been carefiilly selected
with a view to overlapping with signal and image processing, and keeping in mind
the kind of scale-similarity relations that are used in the analysis of wavelets, fi*actals,
and discrete dynamics. For example, we stress how problems in "fi-actal noise," in
images, and more generally in higher dimension may be unified with the use of
operators, tensor products of Hilbert space, and states on C*-algebras.
A case in point is the serendipity involved in the observation (see [Jor03]) to the
effect that representations of the so-called Cuntz relations and Cuntz algebras from
C*-algebra theory have been extensively used in the early days of signal processing
(before they were "discovered" in C*-algebras!), and that they continue to be relevant
to the kind of recursive algorithms still used to this day in wireless communication
and in image processing.

1.2 Path space

A well-tested tool in analysis and mathematical physics centers around the application of path-space measures. This tool is used in attacking a variety of singular convergence, or approximation, problems. We will adopt this viewpoint in our study of
wavelet approximations. Traditionally, the setting for wavelet questions has included
assumptions concerning continuity, or some kind of differentiability. In contrast, we
shall work almost entirely in the measurable category. One advantage of this approach is that we stay in the measurable category when addressing problems from
multiresolution analysis (MRA). Earlier work on the use of probability in wavelets
includes that of R.F. Gundy et al.
Our present viewpoint is more general than [DoGHOO], and it starts with the
random walks naturally associated with a measure space (X, B) and a given measurable onto map a:X ^ X such that #cr~^ ({x}) = N for all x e X, where N,
2 < iV < oo, is fixed. Iteration of the branches
a-\{x]):={yeX\a{y)=x}

(1.2.1)

then yields a combinatorial tree. If

a> = ( c o i , a ; 2 , . . . ) e ^ : = { 0 , l , . . . , i V - l } N ,
an associated path may be thought of as an infinite extension of the finite walks

1.2 Path space

^COn

1 ^
' '^ny)
''0)2 7ft)
^CO\-^

Starting at jc, where (r^), / = 0, 1, . . . , # 1, is a system of inverses, i.e., where

a oTi= idx,

0 <i < N.

(1-2.2)

We now turn to the path-space measures Px alluded to above, but for now only in
a special case. More general cases emerge later on, especially in Chapters 2, 5, and 7.
These will be different applications, and Chapters 7 and 9 will even be in a non-commutative setting. The definition of Px which we give below is quite standard, but it
may not look simple or natural to someone unfamiliar with infinite products, or with
the present viewpoint. This should not be alarming. We believe that the fundamentals
of Px will allow the reader to already now see a number of quite familiar and basic
wavelet constructions in a new light. This will be usefiil later, both for wavelets
and for a class of similar fi*actal constructions given in Chapter 4, as well as for
frequency-localized multiwavelets in Chapter 7. Each measure Px depends on the
initial point x and on a prescribed and fixed weight fimction W on X.
We shall think of Px as a measure on (infinite) paths originating at x, but its
construction rests on a fiindamental idea of Kolmogorov which states that Px is determined by its value on paths that are only fixed at a finite number n of places. It
is prescribed for each n, and it extends to all paths, i.e., to Q, if and only if there is
consistency from ^ to + 1 for all n. This consistency condition is defined from the
given fijnction W, in that W assigns the transition probabilities; see Lemma 2.5.1
(Kolmogorov consistency, p. 46) for precise details. Our probability space Q will
be the space of all infinite words co, Q will be given its standard Tychonofif infiniteproduct topology, and C (Q) will denote the space of all continuous complex-valued
fijnctions on Q.
IfW'.X-^ [ 0,1 ] is a given measurable fimction such that

Y^W(y) = l,

(1.2.3)

y:a(y)=x

then an associated measure Px on Q may be defined as follows. Suppose some fimction f e C (Q) depends only on a finite number of coordinates, say co\,... ,con', then
set
/
'^

fdPx

f (Q)U ' ' ' , COn) W {TOJ^X) W {TCO2T(OIX) ' " W {TCO * ' ^ X ) .

(cou...,co)

(1.2.4)
Extensions of this formula to Q can be done in a number of ways: see Chapter 2,
Figure 1.1, and the cited references.
We shall need the concept of random walk in a slightly restricted context. Normally, the concept of random walk is confined to a lattice, i.e., Z^ for some d. The
qualitative features of such a walk depend both on the lattice dimension d and on

1 Introduction: Measures on path space

>i
TQX

y\

rjjc

TITQJJC

TQTlTQX

/\

T^T{)X n=3 t^TlX

/\ / \
TQtfTQX

"-"^

/\

T\TQTIX

/\

TQT^X

/\

ifx

/v
^1

/
\.
TlTOTj^TQ^

Fig. 1.1. N = 2, simple dyadic branch points; a function W:X -^ [0,1] is given. A
path in the random-walk model. Assignment of probability to a particular length-five path:
P f ^ ({(0 1 1 0 1)}) = W(TOX^ ^ ( T I T O ^ ) W(T^TOX) ^^(roTfTo^) ^(rirorfTQ^).

suitable transition probabilities. In our present context, we shall use a combinatorial tree (see Figures 1.1 and 2.1 (the Farey tree, p. 42)) in place of a lattice, and
we shall use a given function W to prescribe transition probabilities. But the various paths within our tree can originate at points chosen from the set X. Here X is
given (typically compact), and X carries a fixed finite-to-one endomorphism a. Our
construction is easier to visualize when a finite set of branches of the "inverse" for
G is assigned in concrete terms: hence the maps T^y. If we assume property (1.2.3)
for the fiinction W, and if x and y are points in X such that a {y) = x, then the
number W (y) from (1.2.3) represents the probability of a transition from x to y.
Figure 1.1 above illustrates how step-by-step conditional probabilities are then used
in assigning probabilities to a finite path which originates at some chosen point in x.
What if the path is infinite? We shall see in Chapter 2 (Sections 2.4 and 2.5)
how a fundamental idea of Kolmogorov then allows us to assign probabilities also to
infinite paths.
In Section 1.1 above we already used this idea for the case when X is the circle,
also called the one-torus, T. For each N, we may then consider a (z) := z^. And in
the context of wavelet constructions, we akeady noted how to concretely parametrize
the branches of the inverse of z^ when z is complex and restricted to T.
A special feature of this construction, which will be explored in this work, is that
of attractive convergence properties for infinite products of the form
(1.2.5)
n,co\,...,G)n

over certain subsets of Q. As it turns out, these infinite products are determined by
the measures {Px)x^x-> ^i^d by the Ruelle transition operator

1.3 Multiresolutions
iRwg)(x):=

^(>^)^(>^)'

geL^(X).

9
(1.2.6)

y:(T(y)=x

The operator R in (1.2.6) is called the transition operator, the Ruelle operator, or
the Ruelle-Perron-Frobenius operator, and it will play a major role in what follows.
Many problems in dynamics are governed by transition probabilities W, and Px,
and by an associated transition operator Rjv as in (1.2.6). Wavelet theory is a case in
point, and we show that fundamental convergence questions for wavelets and properties of the solutions to (1.3.1) depend on the positive solutions h to the eigenvalue
problem Rw(h) = h. We refer to(1.3.10)-(1.3.11) below. The solutions// are called
harmonic, and the function h in (1.3.10) is a special harmonic function which we
will study in detail later in Chapter 3.
The nature of these solutions is a key to the link between the analysis and probability of path space. Since the early days when use of measures on infinite path space
was first suggested for problems in mathematical analysis, a key question was: what
is the support of the measure under discussion? And our use of the ^-measures Px
for problems in wavelets and fractals is a case in point.
hi particular applications of this viewpoint, the measures Px are initially defined
naturally in connection with certain random-walk models. And each Px, for x e X,
is then a Borel measure on a certain rather large and unwieldy probability space Q.
But for computations it is essential to know that the measures are in fact supported
on (or carried by) much smaller subsets of the initial space Q. These smaller subsets
in Q can be made quite explicit, and they are closely related to dynamics and general
classes of multiresolutions, as we proceed to describe.

1.3 Multiresolutions
All this time the Guard was looking at her, first through a telescope, then
through a microscope, and then through an opera-glass. ^Lewis Carroll
The notion of wavelet refers to a specific basis construction in a fiinction space which
is fixed at the outset and which carries an inner product. This construction is popular, perhaps because it is so computationally attractive. It is based on a certain selfsimilarity (or scaling-similarity) which gives rise to a cascading or nested family of
closed subspaces. These subspaces begin with a fixed resolution which is determined
by a scaling equation such as (1.3.1) below; but as we will see, there are more general
ways of identifying scaling-self-similarity.
The idea is that there is a scaling transformation which lets us move in discrete steps, up and down an associated scale of subspaces, hence the concept of a
multiresolution, a concept derived from optics; see, e.g., [JaMROl], [Mal89], and the
references given there. One end of the resolution scale refers to "coarse," and the

10

1 Introduction: Measures on path space

other to "fine." Comparing two subspaces, the space of the coarse scale is contained
in that of the fine resolution. The traditional wavelet multiresolutions are by now
very nicely presented in the literature, and we especially recommend Daubechies'
book [Dau92].
As noted for example in [Mey89] and [Mey97], a fixed scaling transformation
is also the feature which most efficiently introduces such tools as Hilbert space, operator theory, martingales, conditional expectations, and computational algorithms
into wavelet analysis. Since we are in Hilbert space, closed subspaces correspond in
a one-to-one fashion to (orthogonal) projections, and the notion of multiresolution
may thus be rewritten, as we will see (Chapter 7) in the language of projections. Thus
the multiresolution structure, understood this way, offers a telescope for looking at
fimctions or at signals in digitized form, and it makes a crucial link to signal and
image processing. (See the appendices for fiuther details!) We will meet a number of
incarnations (some special, and some rather general) of multiresolutions in several
chapters throughout the book, especially in Chapters 7-9; but they will always be
based on the geometry of scaling projections in Hilbert space.
The multiresolution approach to wavelets involves fimctions on E. It begins with
the fixed-point problem
(p{t)=N'^ak(p{Nt-k),

^GE,

(1.3.1)

where a given sequence (ak)keZ is chosen with special filtering properties, e.g.,
quadrature-mirror filters; see [BrJo02b, JorOla, DuJo06b]. The equation (1.3.1) is
called the scaling identity, and the ak's the response coefficients, or the masking
coefficients. Introducing the Fourier series
m

(^) = Y,ake-'^''^'

(1.3.2)

and Fourier transform

0(x)=

I ^-'^^^^^ (0 dt,

(1.3.3)

we get the relation

^(x)=m(~)^(^),

X G E ,

(1.3.4)

which suggests a closer inspection of the irffinite products

oo

Since we want solutions (p to (1.3.1) which are in L^ (E), (1.3.4)-(1.3.5) suggest the
corresponding convergence questions for the fimction W := |mp.

1.3 Multiresolutions

11

When (Px)xe[0,\] is the family of measures on Q corresponding to W = \m\^,

then the formal infinite product
oo

.TV"
=i

is Px ({(0, 0,0,...)}), i.e., the measure of the singleton ( 0 , 0 , . . . ) (an infinite string
of zeroes) in { 0 , . . . , TV 1}^. There is a natural way (based on Euclid's algorithm)
of embedding Z into
Q = [0,..,,N-lf

x{0,...,N-lf

(1.3.6)

A(Z) = ^ | ^ ( x + ^)|'.

(1.3.7)

such that
keZ

Remark 1.3.1. Note that, in general, it is not at all clear that the measures (Px),
X G X, on Q should even have atoms. Typically, they do not have them! But if
atoms exist, i.e., when there are points co e Q such that Px ({co}) > 0, we note that
this yields convergence of an associated infinite product. Let No := {0, 1,2,...} =
N U {0}. Using Euclid, and the TV-adic expansion
yt = /I H- /2TV + + /TV"-^

f o r i e No,

(1.3.8)

we see that the points

co(k) = {iu . . . , in, 0, 0, 0, . . . )
00 string of zeroes

represent a copy of No sitting in Q. With the identification k <r^ co (k), we set

00

Px(m = J^Px({co(m.

(1.3.9)

k=0

But in general, the fimction Px (No) might be zero. Our first observation (Proposition
5.3.2) about
h(x):=Px(No),
xeX,
(1.3.10)
is that it solves the eigenvalue problem
Rwh = h.

(1.3.11)

We say that /i is a minimal harmonic fimction relative to Rw- See Chapter 7 for the
justification of the term "minimal". Note that h = 0 may happen!

12

1 Introduction: Measures on path space

Remark 1.3.2. Let X, B, cr, ro, . . . , r^r-i, and W be as described above, and let
{Px)x&x be the corresponding transition probabilities. Let
0 = ( 0, 0, 0, . . . ) e Q = {0, 1 , . . . , TV - 1}^ .
oo string of zeroes

While in general, often Px ({0}) = 0, the case when Px ({0}) > 0 is important. The
condition Px ({0}) > 0 is a way of making precise sense of the infinite product
oo

Pxm)

= Y{^{rEx).

(1.3.12)

If, for example, lim^oo ^ ( ^ o ^ ) < 1? then it is immediate fi-om (1.3.12) that
Px ({0}) = 0.
Suppose Px ({0}) > 0. Then it follows that
Px ( { ( / i , . . . , in. 0)}) = W {zi,x) --W {n, . . . r,^x) P,,^...,,^x ({0}) .

(1.3.13)

Using (1.3.8), we shall identify A: e No with the point co (k) e Q, and write Px {{k})
for the expression in (1.3.13).
Examples 1.3.3. Here are four examples of the equation (1.3.1) which may be understood with the use of transition probabilities for a certain random-walk model:
(a) (p{t) = (p (2t) + ^ (2/ - 1), Haar's wavelet. Figure 1.2(a);
(b) (p{t) = (p {It) + (p{2t - 3), the stretched Haar wavelet. Figure 1.2(b);
(c) (p{t) = cp (30 -\-(pOt 2), Cantor's example;
and
(d) ^ ( 0 = ^ ^ ( 2 0 + ^ ^ ( 2 / - 1) + ^ ^ ( 2 / ~ 2) + ^-^(p{2t - 3),
Daubechies' scaling fimction.
We shall be interested in solutions (p, called scaling functions, to (a)-(d) which
satisfy the fiirther normalization

(p{t)dt = \.

(1.3.14)

A direct verification shows that (a) and (d) have normalized solutions (p e L^ (M).
We will meet several notions of "normalization," starting with (1.3.14). In addition to (1.3.14), we will consider Z^-normalization, defined by condition (1.3.24) below. Z^-normalization means "unit L^-norm." The stretched fimction ^ fi*om example (b), henceforth denoted ^b, ie., the function in Figure 1.2(b), is Z^-normalized,
but not L^-normalized. Specifically, one checks that ^b satisfies (1.3.14) but not
(1.3.24).

1.3 Multiresolutions
(a) The Haar wavelet system

13

(b) The stretched Haar wavelet system

i_

_.,:

"-"-f;"'

''^'

H T

^-"

"-^

1 /

-1

Fig. 1.2. Haar wavelets: The two functions <p and ^ are the father, resp., the mother functions
for Haar's construction in cases (a) on the left, and the stretched Haar (b), on the right. In each
case, we use y/ to create a double-indexed basis system (1.3.16). For (a) this system will be
an orthonormal basis (ONB) in L^(R); but for (b), the functions (1.3.16) will only represent
a Parseval frame, i.e., a function system which yields the Parseval identity (1.3.22) for every
function in L^(R). As is apparent from (b), this Parseval system will not be an ONB for the
stretched version. We resume the discussion of this in the text following Figure 6.1 (p. 103).
So the stretching, i.e., passing from (a) to (b) in Figiire 1.2, leaves (1.3.14)
stable, while the other tw^o conditions (1.3.23) and (1.3.24) get lost. Specifically,
the stretched scaling function ^b also does not satisfy the orthogonality relations
(1.3.23).
In contrast, the tw^o distinct scaling functions
^a=/[o,i]

(Haar)

and
(pd

(Daubechies)

both satisfy all three conditions (1.3.14), (1.3.23), and (1.3.24).

14

1 Introduction: Measures on path space

For (a), we may take
(P = ^ a = X [ 0 , l ] '

and for (b),

1
(P=^(ph= 3/[0,3]'
where / denotes the indicator function of the respective intervals. Readers familiar
with wavelets will recognize (d) as the scaling identity for the Daubechies wavelet:
it has a normalized solution cp = cp^ = ^Dau in Z^ (E) which is supported in the
compact interval [0, 3 ], see Figure 7.16 (p. 134), and diflferentiable, see [Dau92].
The solution cp in (a) is the scaling function for the Haar wavelet; see Figure 1.2(a).
Equation (c) does not have any normalized solution ^ in I ^ (R), but still cp = Xx^
solves (c) when X3 is the "middle-third Cantor set": see Chapter 4 and Figure 4.2
(p. 74). Moreover, there is a unique probability measure // = //3 supported on X^
such that
lf(t)dfi{t) = ^^(^lf(^^^d^it)^lf(^^^

(1.3.15)

for all bounded continuous functions / on M. The Cantor set X3, with associated
measure //, is an example of an iteratedfunction system (IFS) of affine type, and the
harmonic analysis of these systems is the subject of Chapter 4.
Cantor's scaling identity, example (c), admits a normalized solution (p as follows.
The Hilbert space will not be defined from Lebesgue measure, but rather from Hausdorflf measure of fractal dimension 5, in this case s = log3(2). As our Hilbert space
Hs, we use in [DuJo06b] a separable Z^-space defined from the Hausdorfif measure
of fractal dimension log3(2), and consisting of all L^-fiinctions on a certain set 3,
extending X3, and built from X3, using scaling in the large, and certain gap-fiUing
steps; see [DuJo06b]. With this construction, the scaling fimction ^ will be the indicator fimction of X3, having unit norm in the Hilbert space Hs.
The only reasonable solution to (c) is the indicator function of X3, ^ = Xx This fimction is not normalized when referring to Lebesgue measure. The Lebesgue
measure of X3 is actually zero, while the Hausdorff measure hs,s = log3(2), of X3
gives us the right normalization, hsiX^) = 1. Or stated differently, the analogues
of the classical results are true if we modify the measure and the Hilbert space,
using instead the Hilbert space Hs in place of L^(M). (Recall Z^(E) is defined from
Lebesgue measure.)
Rather, the new Hilbert space Hs is defined as an Z^ space with respect to Hausdorflfmeasure hs,s = log3 (2). For (c), we have the following true version of (1.3.14),
which we can call (1.3.14)^:

(p(t)dhs{t) = 1,

(p = Xx-x [indicator function].

(1.3.14)^

1.3 Multiresolutions

15

As stated, when the integral in (1.3.14) is computed on the solution to (c), referring
to Lebesgue measure, it is = 0.
The notion of "normalization" which is independent of the choice of Hilbert
space involves the fimction /? in (1.3.10). Of the four examples, (a)-(d), the three (a),
(c), and (d) have h = 1, the constant function, while (b) has h non-constant, and h
for this stretched wavelet is given by formula (6.2.17), p = 3, and sketched in Figure
6.2 (p. 107).
We postpone a full discussion of (c) to Chapter 4. This fractal case, i.e., the
middle-third Cantor example, and a related much wider class of affine fractals, will
be treated in detail and in a wider context in Sections 4.1 and 4.2; see also the paper
[DuJo06b].
So in the list of four, (a)-(d) above, the only one that stands out as being different
is the stretched Haar wavelet (b). Of course, all the fractal wavelets admit stretched
versions. But we do not discuss those in the present book. However, it is true more
generally that all the Cantor sets that arise as affine iterated function systems (IFSs)
admit orthonormal wavelet bases; see [DuJo06b].
The fractal wavelets we treat in Chapters 4 and 6 are the ones that are orthonormal
bases, and in particular, they are normalized.
As a result, the function h from (1.3.10), called the minimal eigenfimction, is the
constant 1 in the examples from cases (a), (c), and (d), but not (b).
In the "stretched" example, case (b), we calculate the minimal eigenfimction h
in (6.2.17) by a closed expression, the case p = 3; see also Figure 6.2 (p. 107).
An important question for dyadic wavelets in Z^ (R) is the issue of when these
wavelets form orthonormal bases (ONBs). A dyadic wavelet function y/ e L^ (R)
generates an ONB if the double-indexed family

J2"/V(2"^-A:) \n,keZ\

(1.3.16)

satisfies (i) and (ii) below:

(i) / Wn,k ( 0 mj ( 0 dt = dn,m\u

with

m
y/n,k(t):=2''/^yf{2''t-k),

(1.3.17)

and
(ii) the closed linear span of { y/n^k \ n,k eZ}is

L-^ (R).

In our analysis of the scaling identity

(p (t) =2"^ak(p

{It - k)

(1.3.18)

(a special case of (1.3.1)), we will be looking at two functions cp and ^ ; the second
one may be taken to be

16

1 Introduction: Measures on path space

^ ( 0 = 2 ^ (-1)^+1 ax.k(p (2t - k) .

(1.3.19)

keZ

This analysis is the approach to wavelets which goes under the name of multiresolution analysis. The function y/ which is used in (1.3.16)-(1.3.17) is the solution
to (1.3.19). The two standing conditions which are placed on the numbers (ak)keZ^
called masking coefficients, are
^^kak-{-2n = -^o,,

n eZ,

(1.3.20)

keZ

and
Y,ak

= L

(1.3.21)

keZ

These conditions in themselves do not imply orthonormality in (1.3.17), but only the
following much weaker property:
Z Z l ( ^ ' ^ l ^ > l ' = ll^ll'= / 1 / ( 0 1 ' ^ ^
n,keZ

felHW).

(1.3.22)

-^^

A system of functions {y/n,k) satisfying (1.3.22) is called a Parseval frame, or a

normalized tight frame.
It turns out that the ONB property for the wavelet is equivalent to either one of
the following two conditions for the normalized scalingfimctioncp:

(p{t)(p(t-k)dt

= So,k,

(1.3.23)

or
II^IIL2(M) = 1.

(1-3.24)

Using this, the reader may check that the ONB property is satisfied for the two examples (a) and (d), but not for (b). However, the Parseval-fi-ame property is evidently satisfied for example (b), i.e., for the stretched Haar wavelet. While in (b),
(p = i;^rQ ^., the corresponding wavelet fimction y/ is
1

r
it) =

r.

o<.<-,

1
- r,
3

3
- <^ <3.
2

otherwise;

see Figure 1.2(b).

The relevance of random-walk probabilities for understanding of the distinction
between Parseval frames and ONBs is treated systematically in Chapters 2 and 6.

1.4 Sampling

17

Summary of the conclusions from Examples 1.3.3.

Condition (1.3.14): yes for (a), (b), and (d). No for (c)!
Condition (1.3.14)^, s = log3(2), yes for (c). The integral with respect to the
Hausdorff measure hs is oo for ^a, ^b? and ^dCondition (1.3.24): yes for (a), (d). No for (b) and (c)!
Condition (1.3.24)5, ^ = log3(2), i.e., referring to the Hilbert space Tig. Now yes
for (c)! For s = log3(2), the Hilbert T^^-norm of ^a? ^b? and cpd is oc.
Condition (1.3.23), orthogonality: Yes for (a), and (d); but no for (b).
Condition (1.3.23)5, s = log3(2). Yes for (c), and no for the other three!

Remark 1.3.4. In an earlier textbook [BrJo02b], we consistently used the notational

convention mo(0) = y/N for the so-called frequency response function mo; but in
engineering, the alternative convention, mo(0) = 1, is less confusing. This is because
the convention mo(0) = 1 captures the probabilistic meaning of "low-pass" better
than the one from [BrJo02b]; see the "References and remarks" section at the end
of Chapter 9 for a discussion of this point. (Here we have W = |mop.) Also notice
that our current convention, mo(0) = 1, is consistent with the pointwise convergence
of the infinite products (1.3.5) and (1.3.12). Various versions of, and approaches to,
these and related infinite products will play a central role in the next two sections,
and in fact throughout the book.

1.4 Sampling
In this section and the next, we study an intriguing relationship between the following
three problems:
(1) When does the scaling identity (1.3.1) have Z^ solutions?
(2) How may the transition probabihties Px be used in sampling certain functions at
the points x + A: as A: runs over a set of integers?
(3) When is the infinite product (1.3.5) pointwise convergent?
In the wavelet applications, X = [0, 1 ], and the system tr, TQ, ..., z^-i is as
follows:
a (x) = Nx mod 1,
(1.4.1)
r y ( x ) - ^ ,
7=0,1,...,iV-l.
See also Figure 4.1 (p. 73) in Chapter 4.
Lemma 1.4.1. Setting F(x):=

Px ({0}) and

A: = /I + /27V + + /A^-i

(e No),

(1.4.2)

18

1 Introduction: Measures on path space

we get the formula

PA{k]) = F(x-hk),
where we have identified k with the point co{k) := (i\,...

(1.4.3)
,in,0) in Q..

Proof. To see this, identify fimctions on [0, 1 ] with 1-periodic functions on R, and
note that the second formula in (1.4.1) yields T/ zt^ (x) = (x -\- k) /N^ where k
is given by (1.4.2). Hence, if I < s < n, then
W

('..'.i'^)--{''""',r'''"~'y-{^)

It follows that (1.4.3) is really just a rewrite of (1.3.13). The right-hand side of
(1.3.13) yields

which is the desired conclusion.

Remark 1.4.2. To extend Pxi) from No to Z, recall that A: e No is identified with

the singleton co (k) = ( / i , . . . , /, 0) via (1.4.2). Now, if-TV" < A: < 0, then set
Px ({k}) := P, [[co (7V+l + ^ ) } ) .

(1.4.5)

To help the reader gain some intuitive feeling for the conclusion in Lemma 1.4.1,
observe that the right-hand side of (1.4.3) represents a sampling of the function F
at the integral translates on E, starting at x, i.e., x -\- k. Obviously, different subsets
of Q would yield different sets of sampling points for F, including non-uniformly
distributed sampling points; see [AlGrOl].
Starting with Shannon [Sha49], the theory of sampling has emerged as a significant tool in signal processing; see, e.g., the beautifully written survey [AlGrOl] as
well as the references cited therein. Thus, in a general context, our formula (1.4.3)
offers a probabilistic prescription for the sampling of functions on the real line, and
at the same time it stresses the "random" feature of sampling.

1.5 A convergence theorem for infinite products

We now show how this viewpoint from sampling theory is closely related to some
fimdamental properties of the measures Px> In particular, our Theorem 1.5.2 gives a
necessary and sufficient condition for pointwise convergence of the infinite product
(1.3.5), or more generally (1.2.5), with the condition for convergence stated in terms
of the harmonic fixnction h of (1.3.10). The relationship between h and the measure
family Px is studied more systematically in Sections 2.7 and 3.3 below.

1.5 A convergence theorem for infinite products

19

Let ^ c Z. Returning to (1.4.3), we set

P, {A) :=Y,P, m) = Y.^{x+k).

keA

(1.5.1)

keA

As in Lemma 1.4.1, the number N, N > 2, and the function W are given. The
measures Px are constructed from these data using (1.2.4), and we have the two
functions
F(x):=PA{(0.
0, 0, . . . ) } )
(1.5.2)
oo string of zeroes

and
h(x):=Px(Z),

xeR.

(1.5.3)

We shall meet this function h at several places later in the book. It is a distinguished non-negative eigenfunction (corresponding to eigenvalue 1) for the transfer
operator Rw associated with the weight fimction W, and our next result. Theorem
1.5.2, shows that a property ofh at points x determines the convergence of a crucial
infinite product built from W at the same point x. The infinite products are studied
systematically in Chapter 5, while Section 2.7 describes a wider class of eigenfianctions for Rw- We show in Theorem 2.7.1 that the 1-eigenspace E\ (Rw) for Pjv
admits a boundary representation which mimics a classical boundary representation
from harmonic analysis, and we outline the role h plays: It is characterized by a
certain minimality property among functions in E\ (Rw)- This in turn is motivated
by a certain Perron-Frobenius context for the operator Rw which is studied more
generally in Chapter 6.
Finally, for k eN, set
N^Z:=IN^J

\j e Z \ .

(1.5.4)

Using (1.4.5), we see that N^Z is represented in Q = {0, 1 , . . . , TV 1}^ as

( O ^ ^ . ^ q , CPU c^, . . . , .; 0, 0, 0, . . , )
k zeroes

afinitestring
of symbols
coie{0,...,N-l}

(1.5.5)

oo string of zeroes

An infinite string of zeroes will be denoted 0.

Lemma 1.5.1. Let N, W, F, Px, and h be as described above; see (1.5.2)-(1.5.3).
Then h satisfies the following cocycle identity:
h (x) W (x) = PNX (NZ) ,

X G M.

(1.5.6)

In particular, ifh(x) = 1 a.e.x e R, then we recover the fiinction Wfi'omthe

transition probabilities Px, a.e.x e R.

20

1 Introduction: Measures on path space

Proof. We calculate the left-hand side in (1.5.6), using the earlier equations:
by (1.4.3)

t^

and (1.5.3)

^^'^

=.

YW(x+J)F(,X+j)

by periodicity -*^
by (1.4.4)

Y,F{N(x+J))
jeZ

^F(Nx

+ Nj)

JeX
by (74.3)
=

by (1.5.1)

JeZ

PNxiNZ).

This is the desired identity (1.5.6), and the proof is completed.

Theorem 1.5.2. Let N, W, F, P^, and h be as described above. Let x e M, and

suppose that P^ ({0}) > 0. Then the following two conditions are equivalent.
(a) The limit on the right-hand side below exists, and

Fix)^X^f[w{-^).
(b) The limit on the left-hand side below exists, and

lim h(4r) = ^'

Proof (a) =^ (b). An iteration of the identity (1.5.6) in Lemma 1.5.1 above yields
P.('z)=(nr(^)).(^).

0.5.7,

Using (1.5.5), and working in Q, we find that

p|A^^Z = {0}.

(1.5.8)

A:eN

An application of a standard result in measure theory [Rud87, Theorem 1.19(e),

p. 16] now yields existence of the following limit:
lim P,
A:-^oo

(N'^Z)
V
/

= P, ({0})

F {X).

(1.5.9)

by (1.5.2)

Since (a) is assumed, and F {x) > 0, we conclude that the limit h (x/N^), for k -^
cxD, must exist as well, and fijrther that

1.6 A brief outline

21

Using again F (x) > 0, we finally conclude that (b) holds.

(b) =^ (a). Recall that formula (1.5.7) holds in general. If (b) is assumed, we then
conclude that the limit Ylj=\ W (x/N-^) exists as A: ^ cx). The limit on the left-hand
side in (1.5.7) exists and is F (x). The conclusion (a) follows as a result:
k

F ( x ) = lim n ^ ( ^ ) lim / 2 ( - ^ )
A:-^ooAl
7=1

\NJ

) k-^oo

\ m )

limfT^f^).

using (b) A:-^oo A 1

^'
7=1

\NJ

1.6 A brief outline

In Chapter 2, we study random walk as a Markov process, and we compute the
Markov transition measures Px as x varies over the set X. The construction of Px
and the probability space Q. begins with a sequence of measures Px \ ^ = 1,2,...,
corresponding to finite paths of length n. Then we show that the infinite-path-space
measure. Lemma 2.4.1, results fi*om an application of the Kolmogorov extension
principle. Our technical analysis involves a certain transition operator R which generalizes Lawton's wavelet transition operator; see [Law91a, Law91b].
We already mentioned how the measures Px serve to prescribe sampling of
functions on the real line, Lemma 1.4.1. However, a deeper understanding of this
sampling viewpoint is facilitated by the introduction of the Perron-Frobenius-Ruelle
operator R, Definitions 2.3.1, and an associated family of harmonic functions. In fact,
in Theorem 2.7.1, we make precise the concepts of boundary and of boundary values for these harmonic fimctions. Our proof of Theorem 2.7.1 uses the convergence
theorem for martingales, and we outline the martingales associated with our general
transition operator R.
In Chapter 3, a fimdamental tightness condition for the random-walk model is
introduced. The transition probabilities Px live in a universal probability space Q,
but the essential convergence questions for the infinite products (see Chapter 2) depend on the the Px's being supported on a certain copy of No (the natural numbers),
or of Z (the integers). This refers to the natural embedding of Z in Q which we
described above (and further in Chapter 2). These concepts and ideas are presented
systematically in Chapter 3.
In Chapter 4, we study a family of examples of random walk on fi-actals, focusing
on simple examples of afiine iteration on the line, especially the middle-third Cantor
set, X3 (Figures 4.1 and 4.2), and its cousin X4 (Figures 4.1 and 4.3) arising fi-om
quarter divisions. We recallfi*om[JoPe98] that X^ has a natural orthonormal Fourier
basis, while X3 has no such basis of Fourier fi*equencies. The difference between
the two cases may be understood fi'om an analysis of the random walks and the
associated infinite products and transition probabilities.

22

1 Introduction: Measures on path space

We also recall from [DuJo06b] that both X3 and a wider class of affine fractals
have orthonormal wavelet bases, the so-called gap-filling wavelets.
For these fractal examples and also for the wavelet applications, the basis properties are determined by the size of the integers Z naturally embedded in our probability
space Q: the question is when P^ (Z) = 1, i.e., when Z has full measure in Q.
In Chapter 5, we prove a general theorem, Theorem 5.4.1, which captures all the
examples. We prove that if a certain family of tail-sets in Q is not negligible, then
the condition P^ (Z) = 1 is satisfied.
hi general, the fionction x ^^ Px (Z) is not constant. In fact, we show in Chapter
6 that this fianction is harmonic. Moreover, it is minimal in a sense we make precise.
In Chapters 7 and 8 we give a number of applications: wavelets, wavelet packets,
generalized multiresolutions, wavelet filters that take matrix values, and we consider
applications which depend on a certain family of infinite random products of matrix
fimctions.
In Chapter 9, we study pairs of representations of the Cuntz algebras. While this
is a subject from operator algebras, it turns out to apply to the examples from Sections
8.2, 8.3, and 8.4, i.e., to wavelet packets, and to the corresponding measure-theoretic
issues.

1.7 From wavelets to fractals

The FBI has a database consisting of some 200 million fingerprint records
... As part of a modernization program, the FBI is digitizing these records
as 8-bit grayscale images, with a spatial resolution of 500 dots per inch. This
results in some 10 megabytes per card, making the current archive about
2,000 terabytes in size.
CM. Brislawn 1995
One of the reasons wavelets have found so many uses and applications is that
they are especially attractive from the computational point of view. Traditionally,
scale/translation wavelet bases are used in fimction spaces on the real line, or on the
Euclidean space R^. Since we have Lebesgue measure, the Hilbert space Z^(R'^)
offers the natural setting for harmonic analysis with wavelet bases. These bases can
be made orthonormal in Z^(R^), and they involve only a fixed notion of scaling, for
example by a given expansive d-hy-d matrix A over Z, and translation by the integer
lattice Z^. But this presupposes an analysis that is localized in a chosen resolution
subspace, say Fb in Z^(M^). That this is possible is one of the successes of wavelet
computations. Indeed, it is a non-trivial fact that a rich variety of such subspaces FQ
exist; and fiuther that they may be generated by one, or a finite set of fimctions cp in
Z^(E^) which satisfy a certain scaling equation [Dau92].
The determination of this equation might only involve a finite set of numbers (four-tap, six-tap, etc.), and it is of central importance for computation. The

1.7 From wavelets to fractals

23

solutions to a scaling equation are called scaling functions, and are usually denoted
(p. Specifically, the scaling equation relates in a well-known way the ^-scaling of the
fianction(s) cp to their Z^-translates.
The fact that there are solutions in Z^(]R^) is not at all obvious; see [Dau92]. In
application to images, the subspace VQ may represent a certain resolution, and hence
there is a choice involved, but we know by standard theory, see, e.g., [Dau92], that
under apropriate conditions such choices are possible. As a result there are extremely
usefiil and computationally efficient wavelet bases in Z^(M^). A resolution subspace
Fo within Z^CR^) can be chosen to be arbitrarily fine: Finer resolutions correspond
to larger subspaces.
As noted for example in [BrJo02b], a variant of the scaling equation is also used
in computer graphics: there data is successively subdivided and the refined level
of data is related to the previous level by prescribed masking coefficients. The latter
coefficients in turn induce generating Sanctions which are direct analogues of wavelet
filters; see the discussion in Section 1.3, Chapters 7 and 9, and the Appendices.
One reason for the computational efficiency of wavelets lies in the fact that
wavelet coefficients in wavelet expansions for fimctions in VQ may be computed using matrix iteration rather than by a direct computation of inner products: the latter
would involve integration over E^, and hence would be computationally inefficient,
if feasible at all. The deeper reason why we can compute wavelet coefficients using
matrix iteration is an important connection to the subband-filtering methodfi-omsignal/image processing involving digital filters, down-sampling and up-sampling. In
this setting, filters may be realized as fiinctions mo on a (i-torus, e.g., quadraturemirror filters.
As emphasized for example in [Jor05] and [Bri95], because of down-sampling,
the matrix iteration involved in the computation of wavelet coefficients involves socalled slanted Toeplitz matrices Ffi*omsignal processing. The slanted matrices F are
immediately available; they simply record the numbers (masking coefficients) fi'om
the ^-scaling equation. Further, these matrices have the computationally attractive
property that the iterated powers F^ become sucessively more sparse as k increases,
i.e., the matrix representation of F^ has mostly zeroes, and the non-zero terms have
an especially attractive geometric configuration. In fact subband signal processing
yields a finite family, F , G, etc., of such slanted matrices: for example, with L for
"lowfi*equency"and H for "high fi-equency,"
cp = Y.^k(p{2.

-k),

w = Y.Qk'p^^'

-^)'

(i-'^-i)

k
T

NT-

(1.7.2)

24

1 Introduction: Measures on path space

P-2 P-X Po Pi

Pi

P-2 P-1 Po Pi

Pi

P-2 P-\ Po Pi

Pi

P-2 P-l Po Pi

Q-1 Q-\\Qo

Q\

Qi

Q-IQ-\\QO

Q-i

QI

QI

Q-\ Qo Qi

Pi

P-i

P-i Po Pi

Pi

Q2

Q-1 Q-i Qo Qi

Q2

2-2 2-1 Qo Qi

22

Fig. 1.3. The slanted matrices F and G of (1.7.2).

The associated pair of slanted matrices F and G are sketched in Figure 1.3.
Before getting to the theory behind the slanted matrices, we will need some
preparation, and the matrices will then be motivated and studied systematically in
Chapter 7 and again in Chapter 9.
In brief outline, the numbers Pj are entered into the rows of the matrix F via
a specific slant-pattern. The placement is initiated in row/column position 0,0 as

1.7 From wavelets to fractals

25

follows: The number PQ is used in the row/column position 0, 0, and as we move to

the next row one number down, the same row is simply repeated, but with a shift to
the right by two places.
Similarly, we shift two places to the left for each one row-move up in the matrix.
The resulting slanting is seen in Figure 1.3. Moreover, the matrix G is created fi-om
the numbers Qt in exactly the same way.
The wavelet coefiicients at scaling level k of a numerical signal s fi-om VQ are then
simply the coordinates of GF^s. By this we mean that a signal in VQ is represented
by a vector s via a fixed choice of scaling function; see [Dau92, BrJo02b, Jor03].
Then the matrix product GF^ is applied to s; and the matrices GF^ get more slanted
as k increases.
Our approach begins with the observation that the computational feature of this
engineering device can be said to begin with an endomorphism r^ of the rf-torus
jd _ ^ jjd ^ ^jj endomorphism which results from simply passing matrix multiplication by A on M^ to the quotient by Z^. It is then immediate that the inverse images
r^^(x) are finite for all X inT^, in fact #rj^(x) = jdet^l. From this we recover the
scaling identity, and we note that the wavelet scaling equation is a special case of a
more general identity known in computational fractal theory and symbolic dynamics
[Jor04a]. We show that wavelet algorithms and harmonic analysis naturally generalize to afline iterated fimction systems. Moreover, in this general context, we are able
to build the ambient Hilbert spaces for a variety of dynamical systems which arise
from the iterated dynamics of endomorphisms of compact spaces [DuJo06b].
As a consequence, the fact that the ambient Hilbert space in the traditional wavelet setting is the more familiar Z^(R^) is merely an artifact of the choice of filters
mo. As we fiirther show, by enlarging the class of admissible filters, there is a variety
of other ambient Hilbert spaces possible with corresponding wavelet expansions: the
most notable are those which arise from iterated fimction systems (IFSs) of fractal
type, for example for the middle-third Cantor set and scaling by 3.
With examples, theorems, and graphics, we hope to bring these threads to light:
The journey from wavelets to fractals via signal processing.
More generally (see Chapters 4 and 5), there is a variety of other natural dynamical settings (afiine IFSs) that invite the same computational approach.
The two most striking examples which admit such a harmonic analysis are perhaps complex dynamics and subshifts. Both will be worked out in detail. In the first
case, consider a given rational fimction r{z) of one complex variable. We then get
an endomorphism r acting on an associated Julia set X in the complex plane C as
follows: This endomorphism r\X ^> X results by restriction to X [Bea91]. (Details: Recall that X is by definition the complement of the points in C where the
sequence of iterations r" is a normal family. Specifically, the Fatou set F of r(z) is
the largest open set in C where r" is a normal sequence of functions, and we let X
be the complement of F. Here r" denotes the w'th iteration of the rational function
r(z)?^ The induced endomorphism r of X is then simply the restriction to X of r (z).

26

1 Introduction: Measures on path space

If r then denotes the resulting endomorphism, r\X -^ X, it is known [DuJo06a] that

#r ~^ (x) = degree of r, for every x'mX (except for a finite set of singular points).
In the second case, for a particular one-sided subshift, we may take X as the
corresponding state space, and again we have a naturally induced finite-to-one endomorphism of X of geometric and computational significance.
But in the general framework, there is no natural candidate for the ambient
Hilbert space. That is good in one sense, as it means that the subband filters mo
which are feasible will constitute a richer family of fimctions on X.
In all cases, the analysis is governed by a random-walk model with successive
iterations where probabilities are assigned on the finite sets #r ~^ (x) and are given by
the fimction W := \mo\^. This leads to a transfer operator Rw (Section 2.3) which
has features in common with the classical operator considered first by Perron and
Frobenius for positive matrices; in particular, it has a Perron-Frobenius eigenvalue,
and positive Perron-Frobenius eigenvectors, one on the right, a fimction, and one
on the left, a measure; see [Rue89]. As we show in Chapters 2 and 6, this PerronFrobenius measure, also sometimes called the Ruelle measure, is an essential ingredient for our construction of an ambient Hilbert space. All of this, we show, applies
to a variety of examples, and as we show, has the more traditional wavelet setup as a
special case, in fact the special case when the Ruelle measure on T^ is the Dirac mass
corresponding to the point 0 in T*^ (additive notation) representing zero frequency in
the signal-processing setup.
There are two more ingredients entering in our construction of the ambient
Hilbert space: a path-space measure governed by the l^-probabilities, and certain
finite cycles for the endomorphism r. For each x in X, we consider paths by infinite
iterated tracing back with r ~^ and recursively assigning probabilities with W. Hence
we get a measure Px on a space of paths for each x (Section 2.4). These measures
are in turn integrated in x using the Ruelle measure on X. The resulting measure will
now define the inner product in the ambient Hilbert space.
Our present harmonic analysis for these systems is governed by a certain class
of geometric cycles for r, i.e., cycles under iteration by r. We need cycles where
the fimction W attains its maximum, and we call them W-cycles. They are essential, and we include a discussion of ^-cycles for particular examples, including their
geometry, and a discussion of their significance for the computation in an orthogonal harmonic analysis; see especially Chapter 9. We give a necessary and sufiicient
condition for a certain class of affine fractals in W^ to have an orthonormal Fourier
basis, and we even give a recipe for what these orthonormal bases look like. We
believe that this throws new light on a rather fimdamental question: which fractals
admit complete sets of Fourier frequencies? Our results further extend earlier work
by a number of authors, and in particular, clarify the scale-4 Cantor set on the line,
considered earlier by the author and S. Pedersen [JoPe98], and also by R. Strichartz
[StrOO, Str05], and I. Laba and Y. Wang [LaWa02].

Exercises

27

Exercises
Problems worthy
of attack
prove their worth
by hitting back.
Piet Hein
1.1. Let Z2 = {0, 1}, and Q := W^ Z2.
(a) For w G N, ij e {0, 1}, 1 < 7 < n, set
^ (/i, . . . , / ) : = { CO e Q I coj =ij,

I <J

<n}.

Show that 0 and these sets generate a topology T and a cr-algebra S, and moreover
that
where 7^ denotes the set of all subsets of Q. We shall let T and 8 denote topology
and cr-algebra, respectively, whenever continuity or measurability is needed, unless
otherwise specified.
(b) Show that Q is uncountable.
(c) Show that (Q, T) is a compact space.
(d) Show that Q and the Cantor set X3 (with its usual topology) are homeomorphic.
(e) Show that the fimctions XAG / ) separate points in Q.
(f) Using the Stone-WeierstraB theorem, show that the algebra generated by the
constant fimction 11 on Q and the fimctions XAU / ) is dense in C (Q).
(g) Let p e (0,1) be fixed, and set L (t) = I and

i (/^ow))-/f^ I'^='J (!-;'/{ *!'*=''JUsing (f), show that L extends uniquely by linearity to C (Q), and that there is a
measure ju = jup on B such that
L(f)=

/ fdju

forall/GC(Q).

(The measure jUp is called the /?-Bemoulli-product measure.)

1.2. Let the measures jUp be as in Exercise 1.1(g). Show that for p ^ p\ the two
measures /Xp and jip' are mutually singular.
1.3. (a) Show that Z2 is an abelian group under addition mod 1 in Z.
(b) Show that Q in Exercise 1.1 becomes a compact abelian group under the
infinite-"product" operation, i.e., with

28

1 Introduction: Measures on path space

where cof + co\ is the addition in Z2.

(c) The Haar measure // of ^ is defined on Bfi*omExercise 1.1(a) and satisfies:

// is a measure (cr-additive) defined on the cr-algebra B;

/^(Q) = i;
IJL{E-^(O) = fx{E) for 2i\\E e f i a n d o ^ e Q, where jE:+cy := { f + cy \i e E).

Find the Haar measure JJL of Q . Explicit formula!

Hint: Take p = \/2'm Exercise 1.1(g), and show that /^l/2 satisfies the three
listed conditions. You may use the fact that Haar measure is unique subject to these
conditions.
(d) Set
A := { A = (>^/)/eNo I ^i ^ ^2J ^i =^ except for a finite number of places } ,
and set
00

and then show that Xx is a well-defined fimction on Q. for all A e A.

(e) Show that each Xi is continuous and satisfies:

Xi[^ + <^0 = Xx (<^) /yi (<^0 ^*^^ all /I e A and co, co^

G Q.

(Note: ;^;L is called a character on the group Q.)

(f) Show that the linear span of the characters { /;^ | /I G A } is dense in C (Q).
(g) Show that

Xx M Xx' ifo) dfi (co) = bx^v

a
for A, A^ G A, where ju is Haar measure.
1.4. Definition: An indexed family {hx \ ^ e A] im, Hilbert space H is said to be
SLframe(or a frame basis) if there are positive constants A and B such that
A Wff

<Y,\{hx

I f)\^

< B Wff

for all / G H.

(lE.l)

/leA

Returning to the Cantor group Q and its dual A, we examine the family

in the Hilbert space L^ (Q, jup) for each p,0 < p < I.
(a) Show that { / ^ | A G A } is an orthonormal basis in Z^ (Q, jup) if and only if
p = 1/2.

Exercises

29

(b) Show that { / ^ | 2 e A } is ^o/ a frame basis in I ^ (Q, fip) when p ^ 1/2.
Hint: For each n, consider the subset A c A of points X = (/I/) such that
A/ = 0 for / > n; and for each k,0<k<n,
the subset A (k) c A consisting of
X = (Xi) in An such that # {/ | 2/ = 1} = A:. Let / be the constant function 1, or
equivalently, f = XQ. Using #A (A:) = Q), justify the following calculation:

ziui/)r= z
leAn

IUAI/>I'

k=OXeAik)

f : ( ^ ) | l - 2 ; , | 2 * = (l + | l - 2 p P ) " .

A:=0

Since this limit is cx) for w -^ oo whenever /? 7?^ 1/2, the conclusion follows.
(c) Is there a positive lower frame bound for { / ^ | A e A } in the Hilbert space
L^ (Q, jup), p / 1/2? In other words, is there some A > 0 such that the lower
estimate in the relation (lE.l) is valid for all f e L^ (Q, fip)l
1.5. Using the idea from Exercise 1.1(a), show that there is a natural orthonormal
basis (ONB) for Z^ (Q, fii/i).
1.6. Can you find a countable ONB for Z^ (Q, /^ 1/2) ?
1.7. Let (pa and cpb denote the fimctions in the top part of Figure 1.2 (p. 13). Which
of the two families of functions is orthogonal in Z^ (M),
{(Pa{' -k)\keZ}

or

{(pb{' -k)

\k

eZ}l

1.8. Let y/a and yjb denote the two functions in the bottom part of Figure 1.2 (p. 13).
(a) Which of the two families of fimctions
{ 2>/Va h-^t -k\\j,keZ\

or

{ 2>/V^ {^^t

-k\\j,kez\

form ONBs in Z^ (M)?

(b) Give a direct argument with pictures, showing that the double-indexed family
of fimctions

^fl{t)

2J^^yfa{2Jt-k)

forms an ONB in Z^ (R).

(c) Give a direct argument with pictures, showing that the double-indexed family
of functions

y,fl{t) =

2il\j,{2Jt-k)

satisfies

ii/iii^-ZZ|(^]5l/)
for all / e I^

(K)

30

1 Introduction: Measures on path space

1.9. Consider the four versions of the scaling identity on E listed in Section 1.3,
Example 1.3.3, under (a), (b), (c), and (d).
Fill in the missing arguments and check that the conclusions listed in the summary at the end of Section 1.3 are correct.
1.10. Let (p e L^ (R), and consider the ansatz, for some operator W = Wfp,

vAY.Skcp{'--k)\
where {sk) is some sequence.
(a) Give a necessary and sufficient condition on the function cp for when
defines a bounded linear operator fi-om the natural closed subspace K^ c L?into ^2 = ^2 (2)
(b) Give a necessary and sufficient condition on the fimction cp for when
defines an invertible operator fi-om V(p onto ^.
(c) Give a necessary and sufficient condition on the fimction cp for when
defines an isometric isomorphism of F^ onto ^.

W(p
(R)
W(p
W(p

The remaining exercises in this chapter are most likely review for mathematics students: We hope that they make it easier for the reader to build up (or more
likely review!) prerequisites that first-year grad students might need to refresh, as
they move from one chapter to the next: Fourier, Hilbert, bases, a little on linear
operators, etc. All standard material first year grad students have seen but perhaps
might want to brush up on: Learning by doing!
1.11. Hilbert space (the separable case)
Definition (the list of axioms)

7i: vector space over the complex field C.

{ | > : W x H - ^ C , satisfying
(i) X h^ (X I j^ > is conjugate linear for all y
(ii) J i-> (X I J ) is linear for all x eH,
(iii) {x \y) = {y \x),x,y
eH,
(iv) (x I X ) > 0 for all x eH, and
(v) (x |x> = : 0 = x = 0 .

eH,

Setting ||x|| := (x | x >^/^ turns H into a complete normed space, i.e., if some
0, then there is some x e H such
sequence (x)eN satisfies ||x XmW -^
t h a t | | x - x | | -^

0.

Review your real-variables book, and then show that the following are examples
of Hilbert spaces:

Exercises

31
..,Xk),

7Y := C^ = the A:-dimensional complex vector space with x = (x\,.

Xj e C for 1 < 7 < A:, and
k

7^ := ^2 _ ^jj sequences x = (xj) .^^ such that X y l i \^J \ < ^^' where we set
oo

{X

\y)'=^xjyj.
7=1

Ji := ^ (A) where A is some fixed countable set = all

satisfying ^aeA 1^ (^)l^ ^ ^^' ^^^ where we set
{x \y)

fiinctions

x:A^C

:=^x(a)y(a).
aeA

Does this construction work also if A is not countable? What is ^ (A) if A is not
countable?

Let (X, B, ju) be a cr-finite measure space, and set H := Z^ (X, B, ju), or Z^ (ju)
for short, i.e., all measurable fimctions / on X satisfying
/ \f(x)\^dju(x)
Jx

<oo,

where we set
[f\g)'=

f7J^g(x)dju(x),
Jx
and where we identify two fimctions / i and /2 if there is a subset iS with ju (S) =

0,mdfi=f2onX\S.
1.12. Give your own proof of the following basic facts about Hilbert space H.

l(^ I >^)l < \\x\\ \\yhx,y

A linear mapping L:H -^ C (fimctional) satisfies \L (x)| < Const. ||x|| if and
only if it has the form L(x) = {y \x) for some y eH, (Riesz's theorem).

6 H (Schwarz's inequality).

1.13. Operators
Let Hi,i = 1,2, be two Hilbert spaces and let T:li\ -> 7^2 be a linear transformation. Show that the following conditions are equivalent.

r is continuous relative to the respective norms on Tii and H2-

32

1 Introduction: Measures on path space

II Tx II2 < C ||x II1 \/x eH\ holds for some finite constant C.

|( Tx\ I X2 >| < C \\x\ 111 IIX2II2 Vx/ G Hi, i = 1,2, holds for some finite constant C.

Show that the best constant C is the same in the two estimates.
1.14. Fourier series
(a) We shall identify the interval / = [0, 1) and T := {z e C | |z| = 1} via
^ _ ^z2;r/ si^Q^ that this identification respects the usual structures of topology and
measurable subsets for / and for T.
(b) Show that the restriction of the usual Lebesgue measure on R to / induces a
unique normalized measure ju on T, and that this measure satisfies
ju (zE) = ju (E)
whenever z e T and E isameasurable subset of T. (Here zE := [zw \ w e E}.)
(c) For n eZ, set e (z) := z". Show that the fimctions {e j w Z} on T are in
L-^ (T), and satisfy the two conditions

{en\em)

= Sn,m, n,m eZ; and

/ e Z ^ ( T ) & {en I / > =:Oforallw e Z ^ / = 0.

(A system of fimctions with these properties is called an orthonormal basis (ONB).)

(d)If/El2(T),set
(Uf) (n) := f(n):={en\f)=

e^/dfi,
JT

and show that U'.L-^ (T) -> ^^ (Z) is linear, isometric, and maps onto (?- (Z).
(e) Deduce from (d) that the identity
X 1/(^^)1^= / l / P ^ y "

(Parseval)

holds for all f e L? (T), and then make precise the following representation:

eZ

(f) Review your real-variables book and check that the last identity (1E.2) holds
a.e. on T, i.e., the series (1E.2) is convergent on T except possibly on a subset of
measure zero. (This is a difiicult theorem due to Lenard Carleson.)

History

33

History
Naturally infinite products have a rich history in mathematics and its applications;
but in this book, the use of infinite products and applications to wavelet algorithms
and harmonic analysis are emphasized. This serves as a modem key between analysis
and probability: Perhaps somewhat surprisingly, the wavelet connection started with
signal processing. In fact, the use of digital filters in wavelet theory dates from the
1980s, with the pioneering work of S. Mallat [Mal89], A. Cohen [Coh90], W. Lawton
[Law91a, Law91b], among others. And fiondamental ideas of Y. Meyer [Mey79] and
I. Daubechies [Dau92] loom in the background. For this interdisciplinary connection,
we fiarther refer to the papers cited in [Dau92] and in [BrJo02b]. Another element
is the transfer operator, or the wavelet transfer operator. It has many names, e.g., the
Perron-Frobenius-Ruelle operator. Or by now this transfer operator is ofl;en called
the Ruelle operator because of D. Ruelle's use of it in the 1960s on phase-transition
problems in statistical mechanics [Rue69]. As we hope will become clear in the
first two chapters below, the transfer operator (in any one of its many incarnations)
serves as a crucial mathematical link connecting diverse interdisciplinary trends. The
probability aspect of this endeavor was stressed by R. Gundy [GunOO] and others,
e.g., [CoRa90]. Our presentation here relies in crucial ways on Gundy's viewpoint,
as it seems to unify the different threads in the subject that came before it.
It turns out that the marriage of signal processing and wavelets inspired the generation of new algorithms that now go under the name of wavelet algorithms. The
algorithms have fiuther served to make wavelets usefiil to engineers and the medical
community, among others.
Mathematically, the choice of closed subspaces (F) in Hilbert space to model
resolutions is inspired by optics and image processing. So this inspiration came from
physics, and from very practical applications. (And hence the terms resolution, pixels, and level of detail have now entered mathematics.) In the Hilbert-space context,
the relative complement Wn of two successive spaces Vn from the nested scale of
spaces represent a level of detail, and the reader may find it helpfiil to think of the
elements in some space Wn as representing the detail level in an image. But at the
same time, this is also the way to think of algorithms, much the same way we think of
numerical algorithms based on the positional number system. The positional number
representation closely mirrors the wavelet context. In the wavelet algorithm, there
is a scaling operation which implements a kind of similarity (to be made precise in
Chapter 5) between level n and the next level + 1 in the finer scale of resolutions.
Hence we have Vn realized as a subspace of F+i. The intersection of the spaces F
is {0} and the union is dense in Z^(E).
Mathematically, this viewpoint has inspired the use of pyramid algorithms in
wavelets (see Chapters 7, 8, and 9). The viewpoint is versatile, and applies equally
well to one and several dimensions, as we shall see.

34

1 Introduction: Measures on path space

Infinite products have played a central role in a number of classical problems for
more than a hundred years. Recall some very familiar cases:
(1) The factorization over the primes of the Riemann zeta fiinction [EdwOl], and its
many generalizations in dynamics [May91, Rue02, Rue94, SeCh67].
(2) The solution of second order differential equations with the use of a 2 x 2 propagator matrix [PaPo90].
(3) Complex dynamics, iterated substitution of rational fimctions [Schl 871, FrLM83].
(4) Symbolic dynamics: Iterated substitutionfi*omwords to letters, random products
of Perron-Frobenius matrices [JorOla, PerOT, WalOl, BalOO, Rad99].
(5) Statistical mechanics and phase-transition problems [Rue69, BalOO].
(6) The solution of diffusion equations using Wiener's path measure [Sim79].
(7) Analytic continuation of Wiener's solution in (6) to the solution of Schrodinger's
equation; again based on path measures, but now the (ill-defined) Feynman measure [Nel64, Nel69].
More recent applications of infinite random products include:
(8) piecewise linear iterated fiinction systems (IFS) [BrJo99a, DeSh04, Shu04,
ShuOS], fractals and chaos [Rue94]; and
(9) wavelets [Dau92, BrJo02b, CoHR97, CoRa90, GunOO, DoGHOO, GuKaOO, Gun04,
PaSW99].
Even though general methods from probability, random walk, transition probabilities, and path space have a long history in analysis and in applications, their use
in wavelet analysis is of rather more recent vintage.
Ubiquitous to our present approach is a certain "transfer operator" R, see [BalOO]
and [BrJo02b]. This operator in fact has many incarnations (and many names). It has
emerged and re-emerged, over the years, in a variety of applications. The underlying
idea behind it is clear from matrix theory, in the guise of a positive matrix P and the
familiar Perron-Frobenius theorem on the spectrum of P. But we now address an
infinite-dimensional setting.
Our approach is motivated by David Ruelle's use of an infinite-dimensional variant of 7? in the 1960s. Ruelle used R in his study of phase-transition problems of
quantum statistical mechanics. Since then, other variants of R have re-emerged in a
variety of different applications: in dynamics, continuous and discrete, experimental
and symbolic; and in our (limited) imderstanding of fractals and other attractors! The
operator R is used in such applications as wavelets and fractals, as well as in pure
mathematics (zeta fimctions, trace formulas, etc.).
This book is actually focused around (9) from the list of topics above, i.e., the
kind of random-walk problems (see, e.g., [Spi76, DiFr99]) that are associated with
analysis of wavelets [Dau92, BrJo02b] and with iterative algorithms for wavelet
packets [Wic93, Wic94].

References and remarks

35

References and remarks

Historically the meeting ground of probability and mathematical analysis has been in
the areas of potential theory and stochastic differential equations; see, e.g., [Bas95],
[Bas03], [Doo94], and [DooOl]. A classic textbook in probability theory for the beginner is Feller's [Fel71], and it is still a pleasure to read.
Recent papers which build on the interconnections between wavelets and fractals
on the one side and probability and signals on the other include [HeJL04], [JorOla],
[i:)uJo06b], [BaJMP04], and [Gun66, Gun99, GunOO, DoGHOO, GuKaOO, Gun04].
Li contrast to traditional textbooks and research articles, there is now a new information system, Wikipedia, which offers a delightfiil and student-friendly invitation to some fimdamental wavelet ideas and wavelet packets. One of the entries
from Wikipedia for the keyword "wavelet" [WWW/wiki], also available as the paper
[ViMu95], contains both delightfiil exposition and Mathematica do-it-yourself procedures. In addition it reaches out to applications (e.g., to hands-on experiments with
thresholding), and to neighboring fields such as probability. The reader is likely to
find other helpful and relevant material in [WWW/wiki]. Since it is a "community
site," not everything is up to date; but if used with care, it can be really helpfiiland
free!
Turning to some fundamentals of probability theory, the novice might find the
lovely survey article [Stro96] helpfiil. It is a bird's-eye view on path-space measures
from the Gaussian viewpoint and their applications. See also [StroOO]. Both treatments include many useful and modem references. Further, the same author D.W.
Stroock has an attractive new book [StroOS] giving a concise introduction to Markov
processes.
'Dur present focus is different: We stress certain algorithmic and potentialtheoretic features of the theory of wavelets, fractals, and iterated fimction systems
(IFS). Our emphasis is more on the discrete side and on developments that have
taken place in the past two decades, but they have been motivated by the more classical approaches to continuous martingales and stochastic analysis that grew out of
probabilistic methods for the heat equation and its cousins in applied mathematics.
One approach to discrete harmonic analysis is via quadratic forms and resistance inequalities; see, e.g., [KigOl]. Our viewpoint is related to this, but different in that it is
based on the analysis of a certain transfer operator rather than on energy forms as in
[KigOl], i.e., on certain quadratic forms with resistance numbers relating transitions
between states on a graph configuration.
As outlined in Section 1.3 in this chapter, Mallat's approach to multiresolutions
[Mal89] involves three mathematical tools: Fourier transform, infinite products, and
the analysis of periodic fimctions. In several variables, i.e., fimctions on M.^ this last
step in turn relies on classical Fourier duality between the familiar rank-J lattice Z^
on the (multi)frequency side and the compact torus R ^ / Z ^ on the other. Now the
quotient R ^ / Z ^ may be represented as the rf-torus T^, but also in a variety of other

36

1 Introduction: Measures on path space

ways as well. Specifically, we may represent it as a suitably selected subset of R^,

for example as the usual J-cube B. The other choices for B are cdXl^A fundamental
domains for the lattice Z^. Whichever the choice of fimdamental domain, we arrive
at the Hilbert space Z^ (E)) and the following fact: The fimctions ei {x) := e^'^^^'^
with the vector index X'ml/ form an orthogonal basis for the Hilbert space Z^ (B);
i.e., an ONB when normalized.
There is a variation of this theme which starts with the observation that the pair
of sets (B, Z^) is a spectral pair in the sense of [JoPe93]. Specifically, consider
two subsets B and A of M.^ with B having positive and finite Lebesgue measure.
We say that it is a spectral pair if the fimctions { ex (x) := e^'^^^'^ | >l e A } form
an orthogonal basis for I ^ (B). The ex fimctions are restricted to B, and L?- (B)
is defined relative to Lebesgue measure, but there are important variations of the
definition (see, e.g., [JoPe96]) when other measures are used, and we shall return to
this theme in Chapter 4.
The joint papers by S. Pedersen and the present author, e.g., [JoPe92], were motivated in part by an influential paper by B. Fuglede [Fug74]. An important observation
in [Fug74] is that the spectral pairs (B, A) for which A is a group are precisely those
when B is necessarily a fimdamental domain for a rank-(i lattice F. In that case, A
may be taken to be the lattice which is dual to F.
There are a number of other papers which follow up on this general theme. Of
these perhaps [ACM04] and [CaHM04] and the references cited therein are especially relevant.
Li a series of papers (starting with [GaNa98b]) J.-P. Gabardo and his coauthors
have taken up the idea of using spectral pairs as the basis for wavelet constructions
which are then based on "non-imiform multiresolutions." The main theme in this
approach is the use of general spectral pairs (possibly with irregularities, and with
inhomogeneities) in the same way Mallat's traditional multiresolution approach uses
Fourier's spectral pair ((i-cube, Z^). This theory follows the general theme in this
book, but the details are outside the scope of a moderate-size book; so interested
readers are referred to [GaNa98b, GaYu05]. Here we only emphasize that the approach is especially well suited to the analysis offi*actalswith multiresolutions. They
are taken up in Chapters 8 and 9.
Before embarking on the details in the chapters to follow, the reader may want to
first consult some relevant references (books and papers) covering measures [Bil99,
Jor04b, FrLM83], probability theory [Kol77] [Wil91], sampling [Sha49, AlGrOl],
and wavelets [AyTa03, Coh90, BrJo02a, BrJo02b, Dau92, Jor05], and [CoRa90].
As for the general theory of random processes, we recommend the reader study
the books by Skorohod [Sko61, Sko65]. A more modem approach, stressing martingales, may be found in the two books by J. Neveu [Nev75, Nev65]. A (small) sample
of the rich variety of applications includes the papers [Hug95, GlZu80, Mon64].
The fimction h in (1.3.10) and variations of it depend on the chosen space X
and the given weight fimction W. We outlined this fimction above, and it will again

References and remarks

37

play a central role in Chapter 3. Here we stress its connection to Kolmogorov's zeroone law [Sat99, Theorem 1.14]. Recall that Kolmogorov's 0-1 law applies to a given
probability space (Q, P,J^ where P is a probability measure defined on a cr-algebra
^ in Q: Let (^) be an independent and countable family of sub-cr-algebras in Q,
and let A he a, subset of Q. If, for each , A belongs to the tail-cr-algebra generated
by fn, -^w+i, ?>z+2, , then it follows that P (A) must be zero, or one.
There are three features of this:
(1) We view the integers Z as canonically embedded in the probability space Q of
(1.3.6).
(2) When our weight fimction W is given, we then get an associated family of measures Px on Q. (Li our applications to wavelets, W will be Imof' where mo is
some low-pass wavelet filter.)
(3) For every x in X, we may apply the zero-one law of Kolmogorov to the probability space (Q, Px). Let x be given. If there is an independent and countable
family of cr-algebras (^) on Q such that Z belongs to all the tail-cr-algebras of
(J^n), then it follows that h (x) = Px (Z) must be zero or one.
We now turn to a closer study of the measures Px in Chapter 2, and we outline
their dependence on the prescribed weight function W.
In addition to the papers and books on wavelets, sampling, and signal processing,
cited inside Chapter 1, readers might wish to consult one or more of the following
treatments. They serve to supplement, in one way or the other, the point of view
taken in this book: [GroOl, HeWe96, JaMe96, JaMROl, Mal98, MeCo97, StNg96,
Wahi02].
In the chapters to follow, we will stress interconnections between several trends
in the subject, i.e., we emphasize how toolsfi*omanalysis (including operator theory
and operator algebras) and probability theory are used in wavelets, fi-actals, and dynamics. The part of operator algebras that we have in mind is often called non-commutative probability, and it is concerned with representations of algebraic structures
by operators in Hilbert space: indeed, this is the key to their usefulness in wavelets,
fi*actals, dynamics, and other areas as well.
In the non-commutative setting, the analogue of a standard probability measure
is called a state. In mathematical terms, a state is a positive linear functional on an
algebra of operators. A normalization condition is usually added to the definition.
This viewpoint (see especially Chapter 9) is of course motivated by Riesz's theorem,
see [Rud87], and quantum physics. Getting a concrete problem represented in Hilbert
space is the key to the use of such tools as spectral theory; see, for example, the
following papers where the present author has had a hand: [BaJMP06, DuJoOSb,
DuJoOSa, BrJO04, Jor04a, BaJMPOS, JoKr03, JorOlc, JorOlb, BrEJOO, BrJKWOO,
BrJo99b, Jor99, BrJo97]. But, as will become clear later, each of our themes will
have its connection to Hilbert space in one form or another, and more citations will
follow.

Transition probabilities: Random walk

The Cat only grinned when it saw Alice. ...

"Cheshire Puss," she began, rather timidly, as she did not at all know
whether it would like the name: however, it only grinned a little wider
''Come, it's pleased so far," thought Alice, and she went on. 'Would you
tell me, please, which way I ought to go from here? "
"That depends a good deal on where you want to get to," said the Cat.
"I don't much care where'' said Alice.
"Then it doesn 't matter which way you go," said the Cat.
"so long as I get somewhere," Alice added as an explanation.
"Oh, you're sure to do that," said the Cat, ''if you only walk long
enough."
^Lewis Carroll

PREREQUISITES: Curiosity about an idea of Kolmogorov; a vague recollection of

the Stone-WeierstraB theorem; a rough idea about probabilities, and measurable sets;
comparing measures; having encountered the spectral theorem in its simplest form.

Prelude
A key link between wavelets and fractals on the analysis side and random walk on
the probability side is to be found in the use of filters from signal processing. For
the standard dyadic wavelets on the real line, we already sketched this approach
in Chapter 1. Stepping back and taking a more general and systematic view of the
underlying idea, one sees that in a real sense it is (almost) ubiquitous in both pure
and applied mathematics.
Originally "filters" were introduced in their most primitive form by Norbert
Wiener and Andrei Kobnogorov in the 1930s and 1940s for use in a variety of applied
problems involving information and time series. Since then and up to the present, this

40

2 Transition probabilities: Random walk

has been further refined into an art form by engineers with a view to transmission of
speech signals. The first step in the refinement is an identification of so-called fi*equency bands, for example a subdivision into two bands, say high or low, or more
generally, a subdivision of the whole frequency range into a fixed finite niunber of
bands, say N bands.
In image processing (as used in digital cameras), the same idea, suitably modified, yields instead arrays of visual resolutions. Then it is natural to use four filters,
i.e., N = 4, corresponding to recursive and iterated subdivision of squares (called
pixels).
One approach to the analysis of the signal/image at hand may be understood via a
random-walk model, i.e., random walk on a combinatorial tree with a suitable A^-fold
branching. Thought about this way, one sees that the basic idea is indeed ubiquitous,
and in particular that it is fiirther relevant for the kind of geometric self-similarity
notions which go into our understanding of fractals.
We will adopt this view here, and note that these more general filters are then
prescribed by certain fimctions which "assign probabilities" to branch points, or the
bands, i.e., assign probabilities to the A^ "choices" at each step in the "walk." In the
case of speech, the good filters from engineering are those that produce perfect reconstruction of output signals from a synthesis of subbands. Surprisingly, these carefiilly
designed subband filters are the very same ones which may be adapted successfiiUy
to mathematics problems, and which, among other things, produce efficient wavelets. If there are only two bands, the filters are known as quadrature-mirror filters. It
is intriguing to take an even wider view and to compare the subband study of wavelets and fractals to the familiar positional nimiber system dating back to the Arabs
of ancient time: We all know that the expansion of a number in base N amounts to
specifying a string of "digits," in this case the "digits" are selected from the possibilities 0, 1, . . . , TV 1. In the most elementary case, that of the natural numbers,
this is accomplished by a repeated application of Euclid's algorithm; and we are here
attempting to imitate the idea, now with a probabilistic twist, and in a varied array
of applications, ranging from wavelets to fractals and from speech signals to image
processing.

2.1 Standing assumptions

We now turn to the construction of the random walk and the transition probabilities
in the general context of a measure space (X, B) and a fixed A^-to-1 measurable
mappings.
Let X be a set, and let B be a fixed cr-algebra of subsets of X Let

2.2 An example

a:X-^

41

be an onto mapping which is assumed S-measurable, i.e., for all B e B, the inverse
image
(7-^(B):={x
eX\(T(x)e
B]
is again in B. Let jubea, probability measure defined on B. We shall assume that the
singletons {x}, for x e X, are in B, but the measure ju need not be atomic. We shall
fiuther assume that
(2.1.1)
for allx G X, #a-^ ({x}) = N,
where N >2is fixed (and finite).
Let W:X -^ [ 0, oo) be given, and assume that
(i) W is measurable, i.e., that W~^ (J) e B for all intervals J in [ 0, oo),
and fiirther that
(ii)

W(y) < l,//a.e.jc e X.

yeX,(T(y)=x

Jn view of (2.1.1), the sets a~^ ({x}) may be labeled by

ZN = {0,l,...,N-l}

Z/NZ.

Since the singletons are in B, we may pick measurable branches of the (set-theoretic)
inverse a~^, i.e., measurable maps tf-.X ^> X,i =0,1,...,
N I, such that
(7oTi= idx,

/ = 0, 1 , . . . , TV - 1.

(2.1.2)

Using W, we may then define a probability of a transition, or walk,fi-omx to one of

the points (r/ (x))o</<Ar in cr~^ (x), by
P(x,Ti(x)):=W(Ti(x)),

(2.1.3)

Note that by assumption (ii),

Y,P(x,Ti(x))=
i

W(y)<l.

(T(y)=x

The fact that we allow " < " rather than " = " in (ii) is a way of including the possibility
of dissipation in our random-walk model. For details, see Remark 2.8.3 below.

2.2 An example
Example 2.2.1. Farey trees. Let X = I J ^ i ^n where Xn is the set of monotonically increasing sequences of those continued fi-actions

42

2 Transition probabilities: Random walk

1
1 +1

2+1

3+ 1

1+

2+

2+ 2+T

1+

1+1

2+1+ 1+1

1 +3+T

1 +1

1+
1+
1+1

2+1

1+2+T+T

1+1+2+T

1+1+1+T+T

Fig. 2.1. The Farey tree.

[aua2,...,ak]

=
ai +
^2 +

between 0 and 1 such that at G N for 1 < / < A:, a^ > 2, and Xf=i cit =n-\- 2. The
number of terms in Xn is 2", and two branches TQ and T\ are defined by
T o ( [ . . . , ^ ] ) = [ . . . , - 1,2],
n ( [ . . . , a ] ) = [...,a + l ] .
Note that each ry maps Xn into X^+i. If x e X, we say that rox and x\x are the
two daughters in the binary tree. As an example, consider Figure 2.1. These systems
define random walks, and are studied in connection with circle maps, dynamics, and
gap labeling in statistical mechanics; see [Hal83] and [CvSS85].

2.4 Existence of the measures Px

43

2.3 Some definitions: The Ruelle operator, harmonic functions,

cocycles
The next definitions prepare us for the statements and proofs of Lemma 2.4.1, giving
the existence and uniqueness of the measures Px, and Theorem 2.7.1, giving cocycles
as boundary values of bounded harmonic fimctions.
Definitions 2.3.1. (a) The Ruelle operator R = Rw is defined by

(Rf)(x)=

W(y)f(y),

xeX,

feL'^iX),

(2.3.1)

yeX,a(y)=:x

and maps L^ (X) into itself.

(b) Let Q be the compact Cartesian product
oo

Q = Z^ = {0,...,N^lf

= ll{0,...,N-l}.

(2.3.2)

(c) A bounded measurable fimction V:X x Q,^> Cis said to be a cocycle if

V(x,

{co\,co2,''))=

for all CO = (coi, co2,...)

V [TCO^ (X) , {co2, c y 3 , . . . ) )

(2.3.3)

G Q.

(d) A fimction h:X ^ C is said to be harmonic, or Rw-harmonic, if

Rwh = h.

(2.3.4)

(e) Let n GN, and let / i , . . . , / ^ "^N- Then the subset

A(iu...,in):={w

eQ\o}i

= /'i, . . . , con=in}

(2.3.5)

is called a cylinder set.

2.4 Existence of the measures Px

The cylinder sets generate the topology of Q and its Borel cr-algebra. In determining
Radon measures on Q, it is therefore convenient to first specify them on cylinder
sets. This approach was initiated by Kolmogorov [Kol77]; see also Nelson [N[el691.
Recall that Q is compact in the Tychonoflf topology, and that we may use the StoneWeierstrafi theorem on C (Q) = the algebra of all continuous fimctions on Q.
Lemma 2.41. Let X, B, ju, a, TQ, . . . , TN-I, ^^d W be given as described above.
We make the following more restrictive assumption on W:
^

W(y) = l

yeX,(T(y)=x

a.e. x e X.

(2.4.1)

44

2 Transition probabilities: Random walk

Then for every x e X there is a unique positive Radon probability measure Px on Q

such that
Px {A ( / i , . . . , in)) = W {n.x) W {n.Ti.x) 'W

(r, . . . Ti,x).

(2.4.2)

Remark 2.4.2. It turns out that the general case ^

< 1 (assumption (ii) above)
may be reduced to (2.4.1). So (2.4.1) is not really a restriction. This is discussed in
Remark 2.8.3 below.
Proof of Lemma 2.4.1. If P is a Radon measure on ft, we set
P[f]:=

f f(co)dP(cD)

for all / e C (ft) .

(2.4.3)

Set
Cfin (ft) = { / G C (ft) I 3n such that /(a>) = f {cou . . . , a>),

(2.4.4)

i.e., / depends only on the first n coordinates in ft}

= U^Note that (2.4.4) defines Cfin (ft), but may also be regarded as an implicit definition
of 2t, fimctions depending only on the first n coordinates. The algebra Cfin (ft) is
defined as a union over , which as such is independent ofn, and the argument is that
this union is dense in C (ft). Considering 2t in the definition (2.4.4) for increasing
values of the index , we get an ascending nest of subalgebras of C (ft),
2li C 2t2 C C 2l C 2t+i C .

(2.4.5)

An immediate application of Stone-WeierstraB shows that Cfin (ft) is uniformly

dense in C (ft), i.e.,
oo

[j^n =C(ft),
=1

where ~ stands for norm-closure. Let x e X, and / G Cfin (ft). Suppose

/ ( ^ ) = /(coi,...,co),
and set
Px[f]=

^{^co,x)'"W{Ta,--'ra,,x)f(cDu...,COn).

Note that if there is some h e L^ (X) such that

f(cOi,'",COn)=h{To)"-

TcoiX) ,

(2.4.6)

2.4 Existence of the measures Px

45

^m^) ^ i^con ' ra,,x)

(2.4.7)

then

Px[f]=

^ (''^i'^) '"^{^co,'"

yeX,

a^y=x

= {Rlh){x).
We now show that Px[f]is well defined. This is the Kolmogorov consistency:
we must check that the number i \ [ / ] is the same when some / G 2l ( c 2l+i) is
viewed also as an element in 21-|_i. Then
f{(0)

f(cO\,...,COn)

and
^x[/+l]==

W{T^^X)'''W

{To,^^^'''Xa,^x)

f((0\,...,(Dn^\)

C0i,...,C0rj

W
{T^COr,+
iT^(Dr,- - - rco,x)
T^COiX) j I /
^ {^OJr,+
lT^CD

(cOl, . . . , C0)

= 1 by (2.4.1)

W (TCO^X) ''W{TCO'-'

rcoix) f(cou

. . . , ft>)

(Di,...,COn

Px[fn],

as claimed.
The consistency conditions may be stated differently in terms of conditional
probabilities: for / e C (Q), set
Pi''Hf]

= Px[f\^n]
=

X !

(2.4.8)
(^ft)l^) "-W

(Xcon'"

Tco^x)

f{(0\,...,COn).

We proved that
^i"^[/] = ^i"+'^[/]

forall/e2l.

Using now the theorems of Stone-WeierstraB and Riesz, we get the existence of
the measure Px on Q. It is clear that it has the desired properties. In particular, the
property (2.4.2) resultsfi-omapplying (2.4.6) to the function

46

2 Transition probabilities: Random walk

f(co) := di,^^, .. .^,,,,

coea,

(2.4.9)

when the point ( / i , . . . , z) is fixed.

These fimctions, in turn, span a dense subalgebra in C (Q) (by Stone-WeierstraB), so Px is determined uniquely by (2.4.2).
n

2.5 Kolmogorov's consistency condition

For general reference, we now make explicit the extension principle of Kolmogorov
[Kol77] in its fimction-theoretic form.
Lemma 2.5.1. (Kolmogorov) Let N >2 be fixed, and let
Q = {0,1,...,A^-1}^.
Forn = 1,2,...,

let
P^>: 2l - ^ C

be a sequence of linear fimctionals such that (i)-(iii) hold:

(i) P^"^ [ 11 ] = 1, where t denotes the constant fiinction 1 on Q,
(ii) f e^n, f> Opointwise =^ P^) [ / ] > 0,
and
(iii) P ( " > [ / ] = P ^ " + ^ > [ / ] / o r a / / / G 2 t .
Then there is a unique Borel probability measure P onQ. such that
P[f]-^P^"Hf],

fe^n.

(2.5.1)

Specifically, fi)r P, we have the implication result

f eC{a),

f > Opointwise =^ P [ / ] > 0.

(2.5.2)

Remark 2.5.2. Here we have identified positive linear fimctionals P on C (Q) with
the corresponding Radon measures P on Q, i.e..

[/]= f fdP.

(2.5.3)

This identification P <r^ P is based on an implicit application of Riesz's theorem;

see [Rud87, Chapter 1].
Proof of Lemma 2.5.1. The proof of Kolmogorov's extension result may be given
several forms, but we note that the argument we used above (in a special case), based
on an application of the Stone-WeierstraB theorem, also works in general.
n

2.6 The probability space Q.

47

2.6 The probability space Q

We note that the probability space Q itself carries mappings a and r/, / = 0, 1 , . . . ,
N I, satisfying the conditions (2.1.2): stressing the Q-dependence, we write
(T^(co) = (co2,co3,"')
for cy = (coi,co2,...)

and rf^ (co) = (i,coi,co2, - --)

(2.6.1)

e Q.

Remark 2.6.1. The connection between the cylinder sets in (2.3.5) and the iterated
function systems (IFS) (X, cr, r o , . . . , r^v-i) may be spelled out as follows: the cylinder sets in Q generate the cr-algebra of measurable subsets of ^ , and similarly the
subsets r/j r/ (X) c X generate a cr-algebra of measurable subsets of X. When
nothing further is specified, these will be the cr-algebras which we refer to when
discussing the measurable functions on Q and X. In particular, we will denote by
M (Q) and M (X) the respective algebras of all bounded measurable fimctiond on
Q, respectively X.
Note that i f X = [ 0 , 1 ] and
1^fX

X -\- i
= ,
N

0 < i < N -1,

then we recover the familiar N-adic subintervals:

^/l ^in W =

/'l

in

i\

in

(2.6.2)

Lemma 2.6.2. There is a unique mapping p\ M (Q) -^ M (X) which satisfies

Pifg)^P(f)p(g)

(2.6.3)

and
(2.6.4)

P\XA(iu:..in)) - Xxu-TiAX)'
The mapping p is an isomorphism ofM (Q) onto M (X).
Proof. Recalling (2.4.9), we note that

As a result,
XA{h,...4n)^A(Ju...Jn)

^nJ\

' ' ' KjnXAiju-.Jn)

'

We then define p first on 2l by

--CiniX)'
\iu...,in

i\,-,in

(2.6.5)

48

2 Transition probabilities: Random walk

where a/^,...,/ e C, and note that (2.6.3) is satisfied.

It is easy to check that the extension of p fi-om 2l to 2t+i is consistent. The
final extensionfi*om|J 2t to M (Q) is done by Kolmogorov's lemma, and it can be
checked that p has the properties stated in the conclusion of the present lemma, n
Lemma 2.6.3. Let X, W, andN be as described in the beginning of this chapter, and
let {Px \x e X] be the process obtained in the conclusion of Lemma 2.4.1. Then
N-\

W{Tix) P,,x [fit, ')] = Px[f]

forallfeCiQ).

(2.6.6)

Remark 2.6.4. Stated informally, formula (2.6.6) is an assertion about the random
walk: it says that if the walk starts at x, then with probability one, it makes a transition
to one of the N points TQ ( x ) , . . . , Tiv-i (x). The probability of the move x H^ r/x is
W (r,x). Recall (2.4.1) asserts that X / ^ i'^i^) = 1Proof of Lemma 2.6.3. It follows fi-om (2.4.6) and the arguments in the proof of
Lemma 2.4.1 that it is enough to verify (2.6.6) for / e Cfin (^), or for / e 2t.
L e t / G 2 1 . Then

J^W(Ttx)Pr,Af(h-)]
i
= ^

] ^ ^ (^zX) W (Tft,i T/x) 'W{TCO'"

i

Tcoi Tix) f(i,(D\,"-,

COn)

COi,-,0)n

by (2.4.6)

which is the desired conclusion. Recall Cfin (Q) is norm-dense in C (Q).

Remark 2.6.5. Note that the formula (2.6.6) generalizes the familiar notion of selfsimilarity for measures introduced by Hutchinson in [HutSl]; see also (4.1.4) below.
In fact, (2.6.6) may be restated as
N-l

^W(Tix)Pr,xo{r[')

=Px.

(2.6.7)

/=0

2.7 A boundary representation for harmonic functions

The next result. Theorem 2.7.1, is a variant of the classical Fatou-Primalov theorem.
But in the present context, it has cocycles as boundary values of certain bounded
harmonicfimctions^thenotion "harmonic" is here defined from the Ruelle operator
R of Definition 2.3.1(a). We obtain our cocycles as an application of the familiar
convergence theorem for bounded martingales.
We then resume our study of harmonic fimctions in Chapter 6 below.

2.7 A boundary representation for harmonic functions

Theorem 2.7.1. Let X,
and suppose in addition
on L^ {X), and let {Px
is a 1-1 correspondence
V as follows: If

49

W, and N be as described in the beginning of this chapter,

that (2.4.1) is satisfied. Let R = Rw bethe Ruelle operator
\ x e X} be the process on Qfi-om Lemma 2.4.1. Then there
between the bounded harmonic functions h and the cocycles
F: X X Q -> C

is a cocycle {bounded and measurable), then

h(x) = hv(x) = PAV(x,

.)]

(2.7.1)

is harmonic, i.e.,
Rh = h.

(2.7.2)

Conversely, V may be recoveredfrom h as a martingale limit.

Proof. (2.7.1) => (2.7.2). This follows by Lemma 2.6.3 and the cocycle property.
Let r be a cocycle, and define h = hy by (2.7.1). Then for x e X,
(Rh) (x) = Y,W (^/^) h (Tfx) = Y,W (Ttx) Pr,x [ V (r,x, ) ]
(the cocycle property) ^T^
by (2.6.6)

by (2.7.1)

where the summations are over i e Z^v- This proves (2.7.2).

The converse, (2.7.2) = (2.7.1). Let /z be a given harmonic function on X. Let
e N, and define
LAf]=

I f{co)h{Tco,'-'

r^jx) dPx (CO) ,

/ G 2l.

(2.7.3)

An application of (2.7.2) and the argument above show that L^ is consistent, i.e.,
using /(co) = f{co\,.. , cy) = f {(o\, '--.con, co+i), we get
/ fifo)h{
Jo.

dPx ()

x) dPx {CO)
u .Tn..

/ / H ^ (^6> Tft^jx) dPx (co) =

Lx[f],

by (2.7.2) JQ

which shows that Kolmogorov's consistency condition holds.

Hence there is a Radon measure on Q, also denoted Lx, such that
LAf)^

Ja

fdL

feCiO).

(2.7.4)

50

2 Transition probabilities: Random walk

Absolute continuity is immediate, i.e.,

dLx < dPx.
Let V (x, ) be the Radon-Nikodym derivative, i.e.,
^(^, O ^ ^

onQ.

(2.7.5)

It then follows from (2.7.3) and (2.7.4) that F is a cocycle, i.e., that it satisfies the
D
conditions in Definition 2.3.1 (c).
Remark 2.7.2. It is usefiil to think of Theorem 2.7.1 as an analogue of the classical
Fatou-Primalov theorem [Rud87, Chapter 11] about the existence of boundary fimctions for bounded harmonic functions. Recall that eveiy bounded harmonic fimction
h in the disk X = {x e C\\x\ < 1} is the Poisson integral of a fimction v on the
boundary of X, viz., the circle S ' = : { x 6 C | | x | = l } . Specifically, if Px for x e X
denotes the Poisson measure, then the analogue of (2.7.1) is
h(x) =

Px[v],

or
h(x)=

v(co)dPx(co),

or more explicitly in polar coordinates,

2;ri_/r

\-^2rcos(9-co)^r^'

where v (e^^) = lim^-^i h (re^^) a.e. 6.

Moreover, see [Rud87, Theorem 11.22], v (co) is the non-tangential limit of the
given harmonic fimction, for points x tending non-tangentially to co.
Hence in the present analogy, our probability space Q is the analogue of the
boundary, and our cocycle V is the analogue of the limit fimction v.
Using the notion of martingales, see [Wil91], we may recover V from h the
following way. Let
for CO e Q.,
Xk (co) = cok
and set
Zn (X, co) =

TX(Q)) ' ' ' 7XI(G>)^

and note that Z is a Markov chain with transition probability Px at jc. If/z satisfies
(2.7.2), i.e., is i?-harmonic, then h (Zn (x, ')),n = 1, 2 , . . . , is a bounded martingale, and so it converges pointwise PX-^L.Q. on Q, i.e., the limit

2.7 A boundary representation for harmonic functions

V(x, ):^

lira h(Z(x,

))

onQ

51
(2.7.6)

>00

is well defined. It is immediate fi*om (2.4.7) that this F is a cocycle, and that it
satisfies (2.7.1).
A good reference on the standard facts about martingales, including the martingale convergence theorem, is [RoWiOO, 11.48, pp. 146-148]. The martingale property
for h (Zn (x, co)) amounts to the identity
P, (h {Zn (X, .)) I 2t_i) = h {Zn-x (X, .)) .

(2.7.7)

This is the defining property of a martingale, and to verify it we proceed as follows.

Using the identity Rwh = h, we get
Px (h (Zn (X, ) ) I 2 t _ l ) = ^ W

{TO,X) h {TcoT^(Dr,-i T^coiX)

(On

= {Rwh) [TCO_^ '' Tcj^x)

= h (Ta,_i Tco^x) = h {Zn-\ {X, )) ,

which is the desired identity (2.7.7).
Corollary 2.7.3. The 1-1 correspondence between the harmonic functions h in Theorem 2.7.1 and the cocycles V is an order-isomorphism when the ordering of both
families of functions is defined in thepointwise sense.
Proof We noted that when F is a given cocycle, then the corresponding fimction h
is defined by (2.7.1); and conversely, V may be computed from h via the formula
(2.7.6).
Since each Px is a positive measure, (2.7.1) yields the implication
Vi < V2 =^ hvi < hv2'
Conversely, if/? i, /?2 are given harmonic fimctions and Vi, V2 are the corresponding cocycles, then
0 < hi < h2 => Vi < V2.
This second implication is immediate from (2.7.6).

Corollary 2.7.4. Let 0: Q > C be a bounded measurable function, and let X, W,

and R = Rw beas stated in the theorem. Then
V{x,c6)\=v{co),

cue Q,

(2.7.8)

is a cocycle if and only ifv is invariant, i.e.,

?;ot7" = i;.
In particular, ifv is invariant under the one-sided shift on Q, then

(2.7.9)

52

2 Transition probabilities: Random walk

h(x) = P,[v]

(2.7.10)

is harmonic.
Proof. Let u: Q > C be given, and suppose V in (2.7.8) is a cocycle. Then
v(co)= V (x, co)=V {TCO^X, (CO2, m,"-))

=V {(02, a;3,...) = t (tr" {co)) ,

which is the desired conclusion (2.7.9). The converse implication is immediate.

\ih is defined by (2.7.10) for some shift-invariant t>, then it followsfi*omTheon
rem 2.7.1 that h is harmonic; see the proof of (2.7.1) => (2.7.2).
Definition 2.7.5. We say that a subset c Q is shift-invariant if (cr^)~ E = E.
In view of (2.6.1),

{a^)-'EJ\jrp{E),
so shift-invariant sets E satisfy ZN x E = E. For example,
E :=[j

(ZAT

X X

ZN^

X {0})

n times

is shift-invariant. Recall that there is a canonical bijection, see (1.3.8), between this
set E and the natural numbers No = {0, 1,...}.
Corollary 2.7.6. Let E d Q>, and let X, a, r^, N, and W be as described in the
beginning of the chapter. Set
hE{x):^Px[xE],

(2-7.11)

where XE denotes the indicator function ofE.

Then hE is harmonic, i.e., satisfies RwhE = ^E if and only ifE is shift-invariant.
Proof. This follows directly from Corollary 2.7.4 above, if we note that

2.8 Invariant measures

Definition 2.8.1. One of the uses of the fimctions h from (2.7.2), i.e., the harmonic
fiinctions, is that they determine invariant measures v on X, i.e., measures on X
invariant under the endomorphism (T.X -^ X. We say that v is a-invariant, or
simply invariant, if v o cr~^ = v, or
V (a-^ (B)\ =v(B),

B eB.

(2.8.1)

2.8 Invariant measures

53

We say that v is R-invariant if

[ Rfdv=
Jx

[ fdv
jx

(2.8.2)

for all bounded measurable functions / on X.

Proposition 2.8.2. Let v be R-invariant, and let h satisfy Rh = h. Then the measure
dvh:=hdv

(2.8.3)

is invariant.
Conversely, ifv\ is a B-measure on X which is assumed a-invariant, and if
dv\
vi ^ V, then the Radon-Nikodym derivative h = - satisfies
dv
Rh = h

v-a.Q.onX.

(2.8.4)

Proof. (Part one!)

f OCT dvh = / f oa hdv
= f R(foah)dv=
=

fhdv

fRhdv

fdvh,

which implies that Vh is cr-invariant.

To prove (2.8.4) under the assumptions in the second part of the proposition, let
/ be a bounded B-measurable fiinction on X. Then
f fRhdv

= f R[(foa)h]dv=
=

(f o (T) dvi =

j
fdvi

{foa)hdv
= /

fhdv,

where all integrals are over X. As / is arbitrary, the conclusion (2.8.4) follows,

Remark 2.8.3. Suppose there is some probability measure v on X satisfying

vRw = V.

(2.8.5)

Then the assumption (ii),

may be reduced to the special normalization (2.4.1) by the following argument. Assuming (2.8.6), then the sequence

54

2 Transition probabilities: Random walk

is monotone decreasing. If a probability measure v exists satisfying (2.8.5), then

V (hn) = / hndv = \
Jx

for all n,

and the limit function

h {x) := inf hn (x) = lim hn (x)
n

>oo

is measurable, and satisfies

v{h) = \

and

Rwh = h.

(2.8.8)

Since W (x) h(x) < h {a (x)), it follows that the modified ^-fianction
^ . W : = ^ ^ ^ ^ ,

x.X,

(2.8.9)

is well defined, and satisfies the special normalization rule (2.4.1), i.e.,
^

W{y) = \

a.e. X

G X

y&X,a{y)=x

To see this, recall that

^ /2(X) = 1,
/2(X)

which is the desired identity.

Exercises
2.1. Verify the details in the argument in the proof of Lemma 2.4.1 for why U 2l
is an algebra of continuous fimctions, and for why it is dense in C (Q).

Exercises

55

2.2. Give a direct and geometric argument for the identity (2.6.7) in Remark 2.6.5.
2.3. Give three different ways to see that a cocycle V = Vh may be obtained from
every solution h as specified by (2.7.2) in Theorem 2.7.1.
2.4. Can the assumption in Theorem 2.7.1 of boundedness on the harmonic fimction
h be omitted?
2.5. Give a version of Theorem 2.7.1 which holds when h is not assumed bounded.
2.6. Compact operators
Let / := [0, 1], and let 7?: / x / -> C be a continuous fimction. Suppose that

Y,Y.^iR{xi,Xj)^j>0
J

(2E.1)

holds for all finite sequences ((f/) and all point configurations xi, X2,... in / .
(a) Then show that there is a monotone sequence i i , ^2) ? 0 < A+i < /l <
< ^i, such that In -^ 0; and an ONB (g) in Z^ (/), with Lebesgue measure,
satisfying

//O

^ (X,y)gn

(y) dy = Xngn (x) .

JO

(b) The spectral theorem: Show that the operator TR, defined as

{TRf){x):=

R(x,y)f(y)dy

Jo
in L^ (I), satisfies
oo

TR=Y,^n\gn){gn\,
n=\

where \gn) {gn I is Dirac notation for the rank-one projection onto Cg. (We say that
TR is a positive (or non-negative) compact operator.)
2.7. Karhunen-Loeve [Ash90]
Let (Q, B, v) be a probability space, i.e., with v (Q) = 1; and set
E(Y):=

Jo.

Y((jo)dv(co)

for random variables Y on Q.

Let X: / ^ L^ (Q, v) be a random process with X(x) := X(x, ) varying
continuously in X^ (Q). Suppose the following two conditions hold:
(i) E (X(x)) = 0, for all X e / , and
(ii) R(x,y) := E (^X(x)X(y)) is continuous on / x / .

56

2 Transition probabilities: Random walk

(a) Then show that condition (2E.1) is satisfied, and let (g) be an associated
ONB in I^ (/).
(b) Tensor product: Show that there is a sequence (7) such that
(i)

YeLHn,v),
oo

(ii) X(x, co) = ^y^gn (x) Yn (co) is convergent in L^ (!) (g) I ^ (Q, y), and
=1

(iii) E {YkYn) = Sk^nhHint: Set

Yk '= / gk(x)X(x,
Jo

) dx.

2.8. Brownian motion

Notation as in Exercises 2.6 and 2.7: / = [0, 1], and (Q, B, v) is a fixed probabiUty space.
Suppose X:I -> L^ (Q) is given, and auppose also that it is Gaussian, that is,
that the family {X{x) \x e 1} of random variables has joint Gaussian distributions.
(a) Under the stated condition, prove that the random variables { r | G N} in
Exercise 2.7(b) are automatically independent Gaussian.
(b) Suppose in addition to the above condition on {X (x) | x G / } that R{x,y) =
min (x,y) for (x,y) e I x I. Then show that (X(x)) is Brownian motion and has
the following representation:

X(x,co) = ^/2^

oo sin{ ( A) ;rx)
y
r^
^-Zn (co)
=i
\n-^J7r

for X el,

CO e Q,

(2E.2)

where {Z | w G N} is an orthonormal family of random variables in L^ (Q). In

particular, E (Zn) = 0 and (Z^) = 1 holds for all neN.
(c) With the same assumptions as in (b) above, prove that the expansion (2E.2)
for a.e. CO G Q in fact converges uniformly for x G / .
(d) Show that Xwli ( ('^ "" I ) ^ )

I^ (*)I converges with probability 1.

2.9. Consider the setting in Exercise 2.8, i.e., / = [0,1], and (Q, B, v) a fixed
probability space.
(a) Let X[0,x) denote the indicator fimction of the subinterval [0, x) c / . Show
that
R(x,y)

:= min (x, >;) = (X[0,x) I X[0,y) )jr2(/)

for

(b) Find the expansion of/[o,;c) in the ONB ( v ^ sin (w/r/))

show that

all x, >; G /.
, and use this to

References and remarks

X{x,

) = v 2 >

57

Z(-)

where (X(x)) denotes the usual Brownian motion, and where {Zn)n^^ is an independent family of Gaussian random variables with E (Zn) = 0 and E (Z^) = 1; in
particular, the random variables (Z) form an orthonormal family, different from the
one in Exercise 2.8.
2.10. Fractional Brownian motion
Let H e (0, 1), and define
RH (X,

y) := ^ (x^^ + y^^ - \ x - y\^^^

for x, y e [0, oo).

(a) Show that RH(- , ) is positive definite, i.e., that

^^^nR(xn,Xm)^m>0
n

(finite sum)

for all <fi, (^2,... e C, and all xi, X2,... e [0, oc).
(b) Show that H = j reduces to the case of the standard Brownian motion
considered in Exercises 2.8-2.9.

References and remarks

Readers who want to brush up on some of the tools used in this chapter might wish
to consult one or more of the following references (books and papers), for example,
covering path-space measures [IS[el59, Nel64, Nel67, Nel69, Nel73, Sim79], random
walk [BaNiTO, HuMS82, Spi76, RoWiOO, DiFr99], conditional expectations [Jam64,
Rev84], and martingale theory [CoRa90, Gun66, HuSM81].
The idea of using the Kolmogorov extension principle in wavelet constructions
was suggested in three papers by Dorin Dutkay [Dut02, Dut04b, Dut04a].
The use of the Ruelle operator R (alias the Perron-Frobenius-Ruelle operator)
in wavelet analysis was first suggested by Lawton; see [Law91a, Law91b]. The idea
is that the part of the spectrum of R lying on the circle {2 G C | |2| = 1} determines the wavelet orthogonality relations. This idea, and the results of Lawton, were
extended by Bratteli and Jorgensen in [BrJo02b]. The papers by Jorgensen et al.
[JoPe98, DuJo06b] and by Lau et al. [LaNg98, FaLa99, FeLa02] contain other applications of spectral theory of R to iterated function systems. The study of the transfer
operator (alias Ruelle operator) Rw for the case when W is assumed to be in the
Lipschitz class was initiated by C.T. lonescu Tulcea et al. in the seminal Aimals
paper [loMaSO]. The results from [loMaSO] were extended in Chapters 4 and 5 of
[BrJo02b].

58

2 Transition probabilities: Random walk

and within a week or two I saw that the noncommutation was really the
^P.A.M. Dirac
dominant characteristic of Heisenberg's new theory.
While the Karhunen-Loeve theorem in Exercise 2.7 (pp. 55-56) may be viewed
as a result in probability theory, we will see in Chapters 8 and 9 that it is also a
theorem in operator theory. In fact, it may be viewed as "pure" operator theory. A
similar observation applies to our consideration of Schmidt's theorem (Exercise 7.11,
p. 150) and to tensor products in general from Chapter 7. One reason for this double
life is the familiar mathematical distinction between Hilbert spaces in their concrete
form (say Z^-function spaces, or ^^-sequence spaces) and in their axiomatic form.
We will see that the use of operator theory in its more axiomatic form helps clarify
the use of tensor products, and this in turn is vital for all the basis constructions we
shall encounter, wavelets, fractals, and time series from signal analysis.
Abstract considerations of Hilbert space are facilitated by Dirac's elegant bra-ket
notation, which we shall adopt. It is a terminology which makes basis considerations
fit especially nicely into an operator-theoretic framework: If 7-^ is a (complex) Hilbert
space with vectors x,y,z, etc., then we denote the inner product as a Dirac bra-ket,
thus (x I ;;) e C. In contrast, the rank-one operator defined by the two vectors x, y
will be written as a ket-bra, thus E = \x)\y\- Hence E is the operator in H which
sendsz into [y | z)x.
The reader will notice from Exercise 2.6(b) that the conclusion of the spectral
theorem for compact operators takes an especially nice form when expressed with
Dirac's formalism.

No vs. Z

Every axiomatic (abstract) theory admits, as is well known, an unlimited

number of concrete interpretations besides those from which it was derived.
Thus we find applications infields of science which have no relation to the
concepts of random event and of probability in the precise meaning of these
words.
^A.N. Kolmogorov 1933

Periodic functions; power series; Fourier series; examples of

bases in some function spaces, especially in X^.
PREREQUISITES:

Prelude
The title of this chapter calls for an explanation: Given a base space X, an endomorphism a of X, and a prescribed weight fimction W on X, we saw in Section
1.2 that there is an associated measure Px on the space of infinite paths rooted at
x\ see Figure 1.1 (p. 8) for an illustration. As noted, the function W determines the
transition probabilities that go into the probability measure Px as follows: For two
"successive" points y and z on such a path in X, a transition is possible if cr (z) = y,
and the transition probability is then W (z). If points on the path are further apart, we
use a natural formula for conditional probabilities.
In the chapter that follows this prelude we will show that this construction is
key to our understanding of both geometric and computational aspects of the kind of
multiresolutions that can be built on X.
More generally, when path-space measures such as Px are used in analysis, it
is essential to have detailed knowledge about the "size" of their support. Typically
path-space measures are supported on a rather "small" subset of the fiill space of all
infinite paths. An example of this is illustrated by the canonical embedding of the
real line M in the dyadic solenoid. Analogously, Section 1.5 shows that our Px path

60

3 NQVS. Z

space contains a copy of both No and Z. This takes place via a specific encoding
where No corresponds to the set of paths which after a finite number of bits finish
with an infinite string of zeroes.
hi the dyadic case, similarly Z corresponds to the paths which terminate with an
infinite repetition of the pair 01, i.e., paths starting with an arbitrary finite bit-word
followed by an infinite string 010101
(A systematic study of the encodings into
pathsfi*omgeneral N-adic trees will be resumed in Chapter 8 below.)
Perhaps surprisingly, the measures Px tend to have "small" support, and moreover their support properties are significant for the analysis of wavelets and fractals
associated with systems (X, W, Px).ln this and the next two chapters, we will give
conditions for when the support of Px is No or Z. And we will show that these cases
of "small support" are "responsible" for a rather nice harmonic analysis.
hi Chapters 1 and 2, we saw that the convergence of the infinite product (1.3.5)
depends on the support of the measures Px (see Lemma 2.4.1). In particular, the
measure Px applied to the two subsets No and Z in Q is crucial. Recall fi-om Remark
1.4.2 that the integers Z are naturally embedded in the probability space Q. Hence,
in this chapter, we will study the two fimctions Px (No) and Px (Z).

3.1 Terminology
Let TV G N, TV > 2, be given. The cyclic group Zjv = Z/NZ will be identified with
{0,1,.. .,N l], and the operations of multiplication and addition are modulo N;
or we use the fact that Z^ is a ring, and a cyclic group under addition.
Similarly, we will work with the circle, or one-torus, T, as R / Z , and M will
serve as a covering
R-^M/Z^T.
(3.1.1)
In this case, we can take advantage of Fourier duality and use the realization
(3.1.2)
Functions on T may be viewed, or realized, in either one of the following equivalent ways (i)-(ii).
(i) Functions / on E which are 1-periodic, i.e.,
/ ( x + l) = / ( x ) ,

xeR.

If f is measurable, then (3.1.3) is understood to hold only a.e.

(3.1.3)

3.1 Terminology

61

(ii) If / is absolutely integrable on T = R / Z with respect to Haar measure, then it

has a Fourier expansion
(3.1.4)
/ ( x ) = ;^c,^^2;^^^
or
/(z)=:;^c^z^
A:eZ

z = e'2^^

(3.1.5)

where
ck= I f ix) ^-'^^^^ dx= [ f (z) z-^ J// (z) ;

(3.1.6)

the second formula uses (3.1.2), where T is taken to be

(3.1.7)

T = {ZGC||Z| = 1},

and where (3.1.7) may serve to define the Haar measure ju on T, i.e.,
dju (e'^^'A ^ dx

on [ 0 , 1 ) .

(3.1.8)

If PF is a bounded measurable fimction on T, then the corresponding Ruelle operator R = Rw may be written in either one of the following equivalent forms:

(is)

<^/)=Z'-(^)/(^).

P.,.9)

or
{Rf){z)^

W{w)f{w),

zeT,

(3.1.10)

where / is a bounded measurable Sanction on T. The fimction which is constant

1 on T will be denoted 1. Our normalization condition (1.2.3) may therefore be
restated as
(3.1.11)
Rwt = 1.
The equivalence of the two formulations (i) and (ii), or of (\R) and (ii^?), also
yields the following result.
Lemma 3.1.1. Let W and f be bounded l-periodic functions on M. (If the two
functions are measurable, they will be assumed only essentially bounded, and a.e.
\-periodic.)
Then the function

is again bounded and l-periodic.

62

3 NQVS. Z

Proof, Note that the expression in (3.1.12) is the right-hand side in (3.1.9), or equivalently in (3.1.10), and so it is the fimction i?pf/. We now give the explicit argument
as to why it is 1-periodic. Note that the individual terms in the sum (3.1.12) have
period N and not 1. Let

.=|'^(^)/(T)Then
N-\

k=0

3.2 The unit interval

As before, N >2,N eN.is given and fixed throughout. We now revert to the setting
of Chapter 2, but specializing to X = [ 0, 1 ] = the imit interval. Then the mappings
cr, TO,..., TN-\ may be specified as follows.
cT'.x \-^ Nx mod 1,
x-^k
zk'.x ^ TT"'

(3.2.1)
A: = 0, ! , . . . , # - 1,

or alternatively, in complex form,

(7:zh^z^,

zeT,

Tk: z i-> the A:'th of the N roots,

(3.2.2)

k = 0,.. .,N - I.

The second line in (3.2.2) may be spelled out as follows:

Lemma 3.2.1. Let W: [0, \] -^ [0, I] be a given measurable function and extend
W from [ 0, 1 ] ^o R Z)j; periodicity. Then the densities for the probabilities Px in
{\2A)are

W ( r . , x ) . . . ^ (T. . . T^,x) = W ( ^ ) ... ^ {^^

(3.2.4)

3.2 The unit interval

63

where
k = (Di-^mN

+ +cy7V-^

(3.2.5)

andcoi e {0, I,..., N I}, I < i < n.

Proof. A direct calculation, using (3.2.1) and the periodicity of W, yields
W (V^cos-i T^ix) = r f ^ ^ j

for

1< s < n.

It is understood that the integer A: e No is given by (3.2.5). Conversely, every A: e No

has a unique such representation if we agree to let con be the last (if any) non-zero
term. The result (3.2.4) follows.
n
If n is given, and 0 < k < N" I, then the terms coi, ..., con in (3.2.5) are
determined. Let k eZ. Then determine k\l e Z such that
0 < ye < TV" - 1

and

k = k'-^ IN"".

(3.2.6)

This may be done with the Euclidean algorithm. If

k' = coi-\- mN + + conN""'^,
it follows by Lemmas 1.4.1 and 3.1.1 that

As a result, the probability densities Px ({^}) are now defined for all A: G Z, and
not just for 0 <k<m
-\.
Proposition 3.2.2. Let N, X =[0,\l
W, G, TO, ..., TN-X, Q = { 0 , . . . , TV - 1}^,
and Px ({^}) b^ cis described above. In particular, assume that
N-\

X ^ ( ^ )

= l

a.e.xe[0,l).

(3.2.7)

Then it follows that

{N-\)N"-\

^
k=-N"

Pi"+') {{k}) = 1

a.e. X.

(3.2.8)

64

3 NQVS. Z

Proof. We claim that

^

Pi+l) (W) =

;^

k=-m

/>i+i) [{k + 7V}).

(3.2.9)

k=-N"

Once this is proved, the result follows since the sum on the right-hand side in (3.2.9)
is
Ar+i_i

^ (^">') ^ (^"~ V) ^ (y)

= 7^^+1(1) ( x ) - 1 1 (x) = l,
where we used Lemmas 3.1.1 and 3.2.1, and formula (2.4.7).
We now prove (3.2.9) by induction. The case n = 0 reduces to the assumption
(3.2.7). Suppose (3.2.9) holds up to ?? 1. Then the next term is
{N-l)N"~\
k.

= 1 by the induction hypothesis

= ^^(11) = 1,
and the induction step is completed.

3.3 A sufficient condition for Px (Z) = 1

In applications to wavelets, we noted in Chapter 1 that the scaling function (p of an
A^-scale wavelet satisfies
\^(x+k)\' = P,({k}).

(3.3.1)

It is well known that orthonormality of the corresponding A^-adic wavelet system

is equivalent to the normalization
^|^(x+A:)|^ = l

a.e.x G [ 0 , 1 ] .

keZ

As a result, it is of interest to decide when the normalization property

(3.3.2)

3.3 A sufficient condition for Px (Z) = 1

65

holds, or does not.

Recall [BrJo02b] that a scale-A/^ orthonormal wavelet is a system
^ 1 , . . . , y/N-i e L^(R)
such that
(NJ^^y/i(NJx-k))
\

//j,kEZ,i=l,...,N-l

is an orthonormal basis for Z^ (M).

Theorem 3.3.1. Let N, X =[0,l], W, a, r^, ..., T:N-\> andQ. be as in Proposition
3.2.2. In particular, we assume that ^ : [0, 1 ] > [0, 1 ] w measurable and satisfies
(3.2.7). For x e [0, 1 ], let Px be the corresponding measure on Q determined in
Lemma 2.4.1.
Suppose in addition that there are no ^f^, b e R+ such that

all integers k with -

N"'>+P

<k<

(N

- I)

N"''+P.

(3.3.3)

Then it follows that

Px (Z) = 1;
i.e., that
keZ

Proof Let x e [0,1 ] be given. Suppose the numbers no and b have been chosen
such that (3.3.3) holds. Then set

fn (k) = Xl-Mr^iN-m") W ^ ( ^ ) ^ ( ^ )
Let Px be the measures on Q from Lemma 2.4.1. Then it follows from Proposition
3.2.2 that XA:eZ fn W = 1- From Lemmas 2.4.1 and 3.2.1, we have
lim fn (k) = Px ({k})
and
keZ

But, using (3.3.3), we get

fn(k)<b-^Pxm),

n>no.

(3.3.4)

The conclusion Px (Z) = 1 now follows from the dominated convergence theorem,
and a second application of Proposition 3.2.2.
D

66

3 NQVS. Z

Remarks 3.3.2. (a) If W is continuous near x = 0, and if ^ (0) = 1, then condition

(3.3.3) will be satisfied provided W does not vanish on the dyadic rationals kN~",
n > no.
(b) Even though W is initially only defined on [0, 1 ], we extend it to E as a
1-periodic fiinction. We already noted in Lemma 3.1.1 that 7?^ is acting on the 1periodic functions on M.

Exercises
3.1. Let TV be a (fixed) natural number, N > \. For A: e {0, 1 , 2 , . . . , TV 1}, let Tk
denote the transformation x -^ Nx + k, viewed as endomorphisms in Z.
Set C := {0, 1}, and let D be the smallest subset of Z which contains C and is
invariant under all the maps Tk.
(a)ShowthatZ) = Z.
(b) What other two-element subsets C of Z have the property that the smallest
subset of Z which contains C and is invariant under all the maps Tk is all of Z?
(c) (Note that the transformations Tk: x -> Nx + A:, for A: G {0, 1 , 2 , . . . , A^ 1},
leave No invariant.) Show that the smallest subset of No which contains {0} and is
invariant under all the maps Tk is all of No3.2. (a) Consider fimctions / and g as in (3.L5):
/ (z) = Y.'^kz^

g (z) = Y^dkz^-

keZ

Show that
f(z)g(z)

= Y,lkz'

(3E.1)

where
lk = ^Cpdk-p.

(3E.2)

(b) Show that if (ck) and (dk) e f\ then (k) in (3E.2) is also in ^.
(c) Is the conclusion in (b) valid if ^^ is replaced with ^^?
3.3. Let (ck) e i^, and let a fiinction / be defined by (3.1.5). Then show that / e
L^ (T), where T is given the Haar measure; and moreover that the Parseval identity
I|C||,2 = ||/|L2(T)

holds.

References and remarks

67

3.4. Let f e L^ (T) and consider the Fourier correspondence (3.1.5). Set

Show that lim^oo ^ " / exists in L^ (T).

3.5. Let f eL^ (T) be as in Exercise 3.4, and consider (3.L5).
(a) Show that then

(f/)(z) = 2Lc2AZ^
keZ

(b) Let iS = F* be determined by

{Sf\g)L2^{f\Fg)L2
Show that then

forall/,gel2

(-S/)(z) = / ( z 2 ) ,

ZGT,

and further that S is isometric in Z^ (T).

(c) Show that
{Sf) (z) = Y. ^^^^^'

z e T, / e Z^ (T) .

A:eZ

References and remarks

Some relevant references covering infinite products, as used in this chapter, are the
books [Rud87, Zyg32], and papers which aim close to the present theme are [JoPe98,
BeBKOS, BeBe95, CvSS85].
The idea of using harmonic analysis of iterated fiinction systems (IFS) and more
general fi-actals in the study of similarity structures may originate with the paper [JoPe96] by Steen Pedersen and the author; see also [JoPe98], [BrJo99a], and
[Str98].
The next chapter turns to harmonic analysis of afiine IFSs. We will see that the
kind of Fourier bases that can be devised for afiine IFSs and wavelet bases have one
thing in common: they are both better localized than Fourier waves. As a result (see
[JoPe98], [Str05], and [Tao96]) both classes of the new bases have better convergence properties than traditional Fourier series; and the improvements come almost
entirely from the fact that these new bases are better localized: if you try to adjust
locally in a basis expansion with a Fourier wave, you get errors at far distances. Not
so for localized bases!

A case study: Duality for Cantor sets

Closer to this book and equally illuminating are the many problems triggered by a sound or a picture. Only afterwards is a formula devised, and
then proclamed....
^Benoit B. Mandelbrot

PREREQUISITES: The Euclidean algorithm, recursion; the positional number representation; rudimentary examples of abeiian groups; spectrum; Lebesgue; Fatou;
Cantor

Prelude
A well-known principle in Fourier series (reviewed in Section 3.1) for functions
on a finite interval states that an orthogonal trigonometric basis exists and will be
indexed by an arithmetic progression of (Fourier)fi*equencies,i.e., by integers times
the inverse wave length. Similarly, in higher dimensions d, we define periodicity
in terms of a lattice of rank d. The principle states that for rf-periodic functions on
E^, the appropriate Fourierfi*equenciesmay then be realized by a certain dual rankd lattice. In this case, the inverse relation is formulated as a duality principle for
lattices; see, for example, [JoPe93] for a survey of this point.
The purpose of this chapter is to extend, in a self-contained presentation, this
duality principle to a class of fractals.
In order to capture the essence of the idea, we have restricted the exposition
here to those compact fi*actals that are realized by simple arithmetic on the real
line, and which were considered first by Cantor. Hence, the middle-third Cantor
set is an example. But in addition to scaling by 3, we shall consider more general
notions of scaling, including (later in the book) matrix scaling for the multivariable
case. Specifically, here we shall aim for compact Cantor constructions which take
place in M^.

70

4 A case study: Duality for Cantor sets

Our first general observation is that a fixed such affinefi*actalX has an associated
and canonical probability measure fi (= fix)- We define (Section 4.1) our measure
/^ by a precise invariance property which is induced directly by the afiine structure
oiX. The issue of the measure is a delicate one, as X, being afi*actal,is a non-linear
object and does not carry the structure of a group, or even anything close to that. (So
no Haar measure!)
Nonetheless the measure // allovs^s us to formulate a natural Fourier principle,
and to ask for "fi-actal" Fourier bases. Having //, v^e may then imitate Fourier's construction, i.e., we may ask for an orthogonal Fourier basis for L^ {X, ju). But since X
is constructed fi-om recursively leaving out "fi-actions" (in E^), we will then expect
that there is a dual "fi-actal-like" thinning of fi*equencies at infinity, i.e., a thinning
relative to some rank-J lattice. Intuitively, we will expect a set A of Fourier fi*equencies to come fi'om "fi*actals in the large." So we ask for the complex exponentials
^iix indexed by A e A to form an orthogonal basis for Z^ (X, ju). We shall refer to
such an orthogonal basis as a Fourier basis, or a complete orthogonal set of Fourier
fi-equencies. (Our choice of complex exponentials, as opposed to Fourier's traditional
wave-fi*ames of sinusoids, is mainly convenience.)
In the context of affine fi-actals X, the surprise is that such orthogonal Fourier
bases exist at all for some fi-actals, and not for others. In fact, it was previously
believed that only asymptotic basis formulas were possible in the "fi*actal world."
lfd= 1, we show that the basis principle works when Cantor's middle-interval
construction is done with scale 4, but not with scale 3. So, surprisingly, the familiar middle-third Cantor set (the first example that comes to mind!) does not have a
Fourier basis, but if it is modified a little, changing the scale number fi*om 3 to 4,
then we show that Fourier's basis principle holds.

4.1 Affine iterated function systems: The general case

We considered in Chapter 2 a general class of iterated fimction systems (IFS) X
with N branches, N >2. They are defined in a general measure-theoretic category.
We specified probabilities for random walks on the corresponding branches. They
are determined by a fixed measurable fimction W on X and a normalization (see
(1.2.3)). What results (Lemma 2.4.1) is a family of path-space measures Px on the
probability space Q = {0, 1 , . . . , A^ - 1}^, for x e X

4.1 Affine iterated function systems: The general case

71

In Chapter 3 we saw that both the natural numbers No and the integers Z embed naturally as subsets of Q. As we saw in Chapter 1, the traditional integrability
properties of wavelets in the Hilbert space L^ (M) turn out to be closely related to
the fimction x \-^ Px (Z). For example, we proved that the orthonormal basis (ONB)
property for the standard wavelets is equivalent to the identity
Px (Z) = 1

a.e. X e X

In other words, we are concerned with deciding when Z has full measure in Q.
In this chapter, we consider instead a family of Cantor sets X, totally disconnected compact subsets of M of Hausdorfif dimension s,0 < s < 1. As we show
below, there is a natural basis question for these Cantor sets which is similarly related to the fimction x -^ Px (No), i.e., to the size of No in Q.
Each Cantor set is an affine iterated fimction system (IFS), and it will have a
natural realization X in [ 0, 1 ], and a conjugate one X in [ 1, 0 ]. We will refer to
them as the right-handed, resp., the left-handed version of X.
The question we raise is when X has an ONB consisting of Fourier frequencies
(see the definitions below). Let Px,x e X, be the path-space measures. We then show
(see also Theorem 5.4.1) that the ONB property for X is equivalent to the identity
Px(m = h

xeX.

In other words, in this case, the requirement is that No have fiill measure as a subset
ofQ.
The function
W(x):=cos^(27rx)
(4.1.1)
came up in an earlier study by Jorgensen and Pedersen [JoPe98]. In that paper, we ask
which Cantor sets have orthonormal bases (ONB) {e;^ | A e A} for some A c M.
Here
ex(t):=e'^''^^, t eR.
(4.1.2)
While wavelets have fractal features, it turns out that a class of affine fractals,
such as Cantor sets with division scale 3 or 4, have wavelet-like features. We illustrate
this with two examples, and we then cover the general theory later in the chapter; see
Sections 4.3^.4.
One of the results in [JoPe98] states that a certain class of Cantor sets X do admit
orthonormal bases of this form for suitable choices of sets A. If some X admits an
ONB {e^ I A e A}, we say that (X, A) is a spectral pair and that (ex)xeA is a Fourier
basis.
The following two examples (Figure 4.1) illustrate this point: in Figure 4.1(a),
we sketch the middle-third Cantor set X3. It is constructed from
cr:x 1-^ 3x mod Z

72

4 A case study: Duality for Cantor sets

and the two branches of cr ~ ^,

X

70 (x) = -

X -\-2

and

Ti (x) =

The Cantor set X4 in Figure 4.1b is constructed from

a:x \-^ 4x mod Z
and the two branches of cr ~ ^,
X

^0 (x) = -

X -\-2

and

n (x) = .

(4.1.3)

We prove in [JoPe98] that X3 does not have a Fourier basis, while X4 does.
Even so, Dutkay and Jorgensen showed in [DuJo06b] that all the aflBne IFS fractals
X admit orthonormal wavelet bases. These wavelet bases are constructed from the
ambient Hausdorflf measure "(dxy of dimension s (= the Hausdorff dimension
of X), and they are realized in a separable Hilbert space built from (dxY. We call
them "gap-filling wavelets." To help appreciate the examples, we recall the Hausdorff
measure (dx)\ 0 < 5 < 1, and its restriction to the corresponding Cantor sets; see
also [Fal85] and [Huts 1].
Let N e N, N > 2,hQ given. Pick a subset B (Z R such that the points in B
represent distinct residue classes in Z/7VZ, i.e., such that N does not divide b b^
when b and b^ are distinct points in B. Then there is a unique measure ju = JU(N,B)
on M such that

where zb (z) := (JC + b) /N. Let p :=# (B), Then the support of JU(N,B) is a Cantor
set X(^N,B) of Hausdorflf dimension s = \np/lnN = log^ (p). Hence the Hausdorflf
dimension of X3 is logs (2), and for X4 in Figure 4.1(b), it is log4 (2) = 1/2.
For these fractals, the Hausdorflf dimension equals the scaling dimension. We will
not go into details here, but refer instead to [Fal85, Fal90]. The scaling dimension s
is typically easier to compute; the formula is
log (number of replicas)
log (magnification factor)
To be precise, we say that X(N,B) has a Fourier basis if, for some A, the family
{ei I 1 A} is anONB for Z^ {/ii(N,B))'

4.2 The quarter Cantor set: The example W(x) = cos^(2;rx)

Returning to X4 in Figure 4.1(b), we recall the following lemma from [JoPe98].

4.2 The quarter Cantor set: The example W(x) = cos^(2n:jc)

1/3

2/3

1/3

2/3

(a) Xy. middle-third Cantor set.

1/4

0
1

1/2

1/4

3/4
I

1/2

1
I

3/4

(b) X4: the quarter Cantor set.

-1

_J

-1

-3/4

-1/2

-3/4

-1/4

-1/2

(c) X4 := -

-1/4

// G {0,1} [, the conjugate or

left-handed quarter Cantor set.

Fig. 4.1. Cantor sets.

73

74

4 A case study: Duality for Cantor sets

2
3

I I I I

J I I I

^3

Fig. 4.2. The Cantor set X3.

I I I I I

I I I I I

X4

Fig. 4.3. The quarter Cantor set X4.

Lemma 4.2.1. Let the set X4 with associated measure ju denote the Cantor construction of Figure 4.1(b) a(i(4.1.3)-(4.1.4). Specifically, ju is the Hausdorff measure on
thefi^actal X4 with Hausdorff dimension 1/2. Let
A = A4 := j / o + / i 4 + 724^ + . . . | /. e {0, 1} where the sums are finite \.
Then {ex \ X e A] is orthonormal in L^ (X4, ju). Equivalently, (X4, A) is a spectral
pair
Remark 4.2.2. The main result in this chapter is that {e;t I >^ ^ A} is in fact an
ONB.

4.2 The quarter Cantor set: The example W(x) = cos^(2;r;c)

75

'

II

X4

Fig. 4.4. Alternate limiting approach to the quarter Cantor set X4.

3
-1 - 4

1
"1

1
~4

0 -1

0-1

(T

I I I I I I I I I

X4
Fig. 4.5. X4, the conjugate or left-handed quarter Cantor set.

Proof of Lemma 4.2.1. Setting

(4.2.1)
JX4

we get the scaling relation

e K.

(4.2.2)

Since the inner products are

{ei\ex')^=

ex (x) e^' (x) J/^ (x) = C (^' - i ) ,

(4.2.3)

76

4 A case study: Duality for Cantor sets

3

-1 ~?

-1

-?

0-1

0-1

(J

IIIII

Jf4
Fig. 4.6. Alternate limiting approach to the conjugate quarter Cantor set X4.

we need only show that, if 1' 7!^ 1 in A, then the product

(4.2.4)
vanishes for some n. If A' ^^ X, there is a first term where // ^ l\. If it is n,
then A' - A G 4" . ( 1 + 4Z), and the last factor in (4.2.4) is then cos ( f f ^ )

cos ( f ) = 0 .

4.3 The conjugate Cantor set, and a special harmonic function

Our next lemma is also fi*om [JoPe98]. For an alternative to the argument fi*om
[JoPe98], see also an appendix to [JoPe98] w^ritten by R. Strichartz [Str98]. We state
the lemma here without proof. The reader will easily be able to supply the details.
Lemma 4.3.1.
(a) To 'verify the O^B property, it is enough to verify that the function
hUx):=^\C{x-X)\^

(4.3.1)

AeA

is constant and equal to\,x G M.

(b) Setting W {x):= cos^ {Inx), see (4.1.1), we get
RwhA = AA,

(4.3.2)

4.3 The conjugate Cantor set, and a special harmonic function

77

where
(R^f)

(X) := W Q

/ (J) + r ( ^ )

/ ( ^ )

(4.3.3)

Remark 4.3.2. We establish the fact that {^i \ X e A} is an ONB for L^ (X4, ft) by
showing that the eigenvalue problem (4.3.2) has only one solution /ZA which satisfies
hjs^ (0) = 1, and which is Lipschitz continuous.
We now turn to the random walk and the measures Px onQ. {0, 1 }^.
Now the measures Px will be indexed by the conjugate fi*actal X4 in Figure
4.1(c), i.e., the one constructedfi-oma.x \-^ Ax mod Z, and the pair
^o(x) = J ,

Ti(x) = ^ ^ .

(4.3.4)

We say that the Cantor set X4 in the interval [ 1,0] constructedfi*omthe IFS
(4.3.4), see Figures 4.5^.6, is the conjugate of the quarter Cantor set X4 fi-om Figures 4.3^.4. Both sets are afiine fractals with Hausdorff dimension 1/2: see formula
(4.1.5).
Let Xt be the Cantor set of Figures 4.3-4.4, i.e., based on (4.1.3). The Cantor
set of the alternative system (4.3.4) will be denoted by X4, and we refer to it as the
conjugate Cantor set to X4. It is a little harder to visualize than the fractals in Figures
4.1(a) and 4.1(b). It is a little different from the fractals in Figures 4.1(a) and 4.1(b),
in that the gap configuration appears to have less separation. However, that is just a
"visual effect," as one can make the gaps appear the same size as in the first one by
just plotting the image on [ 1/2,0 ] instead of [ 1, 0 ]. This fractal X4 in Figure
4.1 (c) may be written as
00

X4 =

7.

^^|/,e{0,l}

(4.3.5)

and we call it the conjugate or left-handed quarter Cantor set (or the "back side" or
"flip side" of the quarter Cantor set). It has the same Hausdorff dimension 5 = 1/2
as does X4 from Figure 4.1(b).
Applying Lemma 2.4.1 to W (x) = cos^ (2;rx) and the fractal XA in (4.3.5), or
Figure 4.1(c), we get the following formulas for the measure Px onl = {0, 1}^:
p . {A ( / , , . . . , /)) = n

cos2 { ^ ^ ^ )

(4.3.6)

and
Px {{CO (A)}) = n '^os' i^t-J^)

= I< ^^ I ^^ >^^(^4./.)I' '

(4-3-7)

78

4 A case study: Duality for Cantor sets

where A(i\,...,

in) denotes the cylinder set (2.3.5), and

co(X) = {iu

. . . , /, 0, 0, 0, . . . )
^
,
'

(4.3.8)

oo string of zeroes

for
2 ^ / 1 + /24 + i34^ + .. + /4-i.

(4.3.9)

In view of (4.3.8)-(4.3.9), we may identify No and A. Both are uniquely represented

with points in Q which terminate with an infinite string of zeroes, i.e., one of the
form (/'i,..., /, 0,0, 0,...) for some n.
0

It follows in particular that A, or equivalently No, has measure 1 if and only if

Px (No) = hA(x) = Y.\^^^\^^ h2(x,,J^

= 1

for

all x e X.

(4.3.10)

4.4 A sufficient condition for Px (No) = 1

Proposition 4.4.1. [JoPe98] With the stated properties, we have Px (No) = \for all
X eX.
Proof. Let

and set

1
ifAe Aand/l < - (4" - l ) ,

otherwise,
n-\

Then we get
;^Fi")(l)=l.

(4.4.1)

It is clear from (4.3.7) that

lim F^"H>.) = Px{{(om).

(4.4.2)

To prove that Px (No) = 1, we must establish that the convergence in (4.4.2) is

dominated.
For n large, the terms ;r (x A) / (2 4^) are close to 0 for /? > n\ and we may
pick some b > 0 and ^o such that

Exercises

79

We conclude that
Fi) (1) <b-'P,

({CO (m,

n>no,

^e A.

(4.4.4)

This is the desired domination, and we therefore may apply the dominated convergence theorem to the sum XieA ' * * The conclusion is

PAm =

X\{e.\e,)^2^j,^^^/^l,
AeA

which is the desired result.

Remark 4.4.2. Li the next chapter, we will prove a theorem, Theorem 5.4.1, for
general path-space measures, which shows when No has fiill measure in Q,. As in the
proof of Proposition 4.4.1 above, our reasoning there will be based on a domination
argument.

Conclusions
When Lemma 4.3.1 and Proposition 4.4.1 are combined we conclude that the set
{^^ I A 6 A} is an orthonormal basis (ONB) in the Hilbert space L-^ (X4, ju); i.e.,
that the Cantor set X4 has a Fourier basis. The orthogonality of the functions ex is
the easier part of the argument. It follows from (4.3.2) in the lemma. But only by also
proving that the special function HA in (4.3.1), the minimal 7^^p-harmonic fimction,
is in fact the constant function 1 are we able to infer that the set { ^;L | >i e A} is total
in L^ (X4, ju), i.e., that it is an ONB.
However, this step is quite analogous to a key argument which we already encountered in Chapter 1 for wavelets. The argument for the Z^-density of the linear
span of {^2 I >^ e A} hinges on (4.3.10) as follows: First use (4.3.10) to show that
every ex can be approximated in L^ by functions in the span of the ^^'s. Then use
Stone-WeierstraB to infer that the span of the ex's is dense in Z^.

Exercises
4.1. Let TO and TI be the transformations in (4.1.3), and let X4 be the quarter Cantor
set X4 in Figure 4.3 (p. 74). Show that
X4 = TO (X4) u n

(X4).

80

4 A case study: Duality for Cantor sets

4.2. Formulate and prove the analogous result for the usual middle-third Cantor set
4.3. Let TO, Ti, and a be the transformations of X4 introduced in coimection with
(4.1.3).
(a) Compare the following two sets of Borel probability measures on X4:
Mc = I // I // o cr ^ = yW I

and

Mx = 1/^1/^ = 5 ( / ^ o ^ o ^ + ^ ^ ^ i / i

(b) Estimate the cardinality of the two sets Mo and Mx in (a).

4.4. Give a direct verification of the scaling law {A.22) in Lemma 4.2. L
4.5. Try to carry out the arguments in Lemma 4.2.1 for the middle-third Cantor set
X^ in place of X4, and check what goes wrong.

References and remarks

Readers not already familiar withfi*actalsmight wish to consult some relevant references covering analysis on afiinefi*actals,e.g., the very readable books [Fal85, Fal90,
BalOO], and/or the papers [LaNg98, FaLa99, LaNROl, FeLa02, BrMo75, HutSl,
JoPe98, Hal83]. As for a student-fi-iendly presentation offi*actalsin many parts of
mathematics, we can recommend Devaney's lovely little book [Dev92].
The regularity issues that are the focus of this chapter are related to harmonic
analysis questionsfi-omthe theory of affine iterated fimction systems (IFS), and there
is a substantial amount of work by Strichartz and others that spells out more coimections to diverse areas of mathematics and applications. The treatment we gave above
is especially inspired by [Str98], and [StrOO]. In the context of our Chapters 4-5,
these papers address the questionfi*omProposition 4.4.1 above; see also Theorem
5.4.1 below. The question and the general theme were central to the paper [JoPe98].
In fact [Str98] came about fi*om Strichartz's suggested alternative approach to the
result in [JoPe98]. And other papers followed [JoPe98], for example [LaWa02] by
Laba and Wang.
At the conclusion of our work on this book, we received a preprint [Str05] fi-om
Strichartz which continues the harmonic analysis theme from [JoPe98] and [Str98],
but in a different direction, viz., convergence: The new preprint [Str05] addresses a
degree to which fractal Fourier series from [JoPe98] and the papers following it tend
to be more localized, and as a result have better convergence properties than do the
classical Fourier series. This localization of the Mock Fourier series is interesting as
it is analogous to a strong localization property of wavelet bases, and we turn to that
in detail in Chapter 7 below.
There is a vast diversity of geometric structures with self-similarity, or with an inherent consistency of scale. It includes wavelets, fractals and classes of models from

References and remarks

81

dynamics. However, within this variety, the fractal structures that are most amenable
to mathematical analysis have their self-similarity defined by afiine mappings (in
ambient Euclidean space). But even this more narrow focus, restricting to afiine
mappings (or rather afiine fimction systems), encompasses both standard and nonstandard wavelets, as well as some of the best-known fractals. Our present viewpoint
is to study these geometries in the light of Fourier duality.
But there are lots of other viewpoints: A delightful and student-friendly presentation of the class of fractals we have in mind here is the little book [YaHK97] by
Yamaguti, Hata, and Kigami. It is one of the few available books which aims to unify
wavelets and fractals, but its aim is potential-theoretic: Laplace operators, resistance
inequalities, and so on.

Infinite products

Albert Einstein to Oscar Veblen (a math professor):

''The Lord God is subtle, but malicious he is not."
(Raffiniert ist der Herr Gott, aber boshaft ist Er nicht.)
And ten years later:
''Ihave second thoughts. Maybe God is malicious."
^Albert Einstein
PREREQUISITES:

Riesz; random; Fubini; square-integrable functions.

Prelude
In the first two chapters, we introduced the random walks that are used throughout
the book. We outlined this in the context of endomorphisms of compact spaces X and
combinatorial trees; and we showed in Chapters 3 and 4 how this applies to wavelets
and fractals. The combinatorial trees to keep in mind for illustration are sketched in
Figures 1.1 (p. 8) and 2.1 (the Farey tree, p. 42). Recall fiirther that the transition
probabilities in the random-walk model are assigned via a prescribed function W on
X which is assumed to satisfy a certain normalization condition. Within the context
of signals, W is the absolute square of some fi'equency fimction m, or of a wavelet
filter. The various paths within our tree can originate at points x chosen fi-om the
set X. As before, X carries a fixed finite-to-one endomorphism cr. If x and y are
points in X such that cr (y) = x, then the number W (y) represents the probability
of a transitionfi-omxioy. Step-by-step conditional probabilities and finite products
are used in assigning probabilities to finite paths which originate at x. (The simplest
instance of this idea is for the case when X is the circle, i.e., the one-torus T. For each
TV, we may then consider a {z) := z^. And in the context of wavelet constructions,
we introduced the additive formulation of the distinct branches of the inverse of
z -^ z^ when z is complex and restricted to T.)

84

5 Infinite products

Here we are concerned with the case of infinite paths. For this we must use infinite
products coupled with a fimdamental idea of Kohnogorov. These infinite products are
the subject of the present chapter. The measure which assigns probabihties to subsets
of paths starting at x is called Px. To understand the measures Px and their support,
convergence issues for infinite products come into play.

5.1 Riesz products

The study of infinite products has a long history in analysis, see especially [BeBKOS]
and the references given there; in number theory [Mey79], lacunary trigonometric
series [Zyg32]; and in ergodic theory [BrMo75, Kak48, Kat87], as well as in other
applications, e.g., [BeBe95, Rit79]. A special case of the infinite products we discuss in the present chapter includes the so-called Riesz products [Kiel8], and their
many variants. While the treatment in [BeBKOS] is centered around certain multiscale properties of a class of generalized Riesz products, our present discussion is
motivated instead by our probabilistic viewpoint, i.e., the use of random walk, iterated fimction systems (IFS), and probability theory.

5.2 Random products

In this chapter, we consider a general class of infinite products which is connected to
the random-walk systems fi-om Chapter 2. A special case of these infinite products
arises in the analysis of wavelets, to which we shall turn in the next chapter.
We begin with a lemma.
Lemma 5.2.1. Let {X, B) be a measure space as described in Chapter 2. Leta\ X -^
X and N be as described, and let TQ, . . . , TTV^-I be a choice of branches ofG~^. Let
W'.X ^ \^,\\be
measurable and satisfy (2.4.1), i.e., Y^aiy)=x W iy) \for all
X e X Then the infinite product
F(x):===^Um^fjPr(ro^(x)),

x e X,

(5.2.1)

x e X,

(5.2.2)

k^\

converges; and the identity

F(x) = WiToix))F(ro(x)),
holds if and only ifPx ({0}) > 0.

5.3 The general case

85

Proof. As in Chapter 2, set Q := {0, 1 , . . . , TV - 1}^, and let 0 = {0, 0, 0 , . . . } be

the point in Q which is given by an infinite string of zeroes. Let ^ ( 0 , 0 , . . . , 0) be
'
^
'
n times

the cylinder setsfi-om(2.3.5), and let Px,x e X,hQ the Radon measures constructed
in Lemma 2.4.1. By (2.4.2), we have
P, [ ^ ( 0 , 0 , . . . , 0 ) ) = n
n times

^ ( 4 (^))

(5-2.3)

'

Since
{0}=f|^(0,0,...,0)
n times

and
^(0,0,---,0) C ^(0,0, . . , 0 ) ,
+ l times

times

we get
P^({0})= lim

PAA((),(i,...,0)\.
n times

In view of (5.2.3), the desired conclusion (5.2.1) follows, and

F ( x ) = P^({0}).
The second conclusion (5.2.2) is also immediatefi*omthe convergence of the products in (5.2.3) as ^ cxD.
n

5.3 The general case

We begin with a remark concerning the embedding of natural numbers in Q.
Remark 5.3.1. Let ( / i , . . . , z, 0) be the point in Q which is the concatenation
of (z'l,..., /) with an infinite string of zeroes. For the measure of the singleton
{(/i,...,/,0)}wethenget
i'^({(/i,...,/,0)})
= W (r/jx) 1^ (T/ T/jx) F (T, . Ti^x),

Xe X

(5.3.1)

By the Euclidean algorithm, the natural numbers k e No = {0, 1,2,...} have a

unique representation
A: = /i + /2A^ + + /^""^
(5.3.2)

86

5 Infinite products

recall that No is naturally embedded in Q. If

with is EZN = {0,\, ...,N -\}.WQ
n is fixed, (5.3.2) is a unique representation of all A:, 0 < A: < A^'^ 1. It is convenient
to identify k with the singleton {co(k)} = { ( / i , . . . , z, 0)} in Q. So each A: 6 No
represents a unique singleton in Q, and the mapping k i-> (/'i,..., z, 0) is 1-1.
Hence
oo

or more precisely,
oo

Px (No) = X

Z / - ('!' ''"' 0)i)

(5.3.4)

Proposition 5.3.2. Let X, a, ro, . . . , rjv-i, ^ , ^^<^ {Px)x^x ^^ ^^ specified in

Lemma 5.2.1. Then the fidnction
h(x):=PAm,

^eX,

(5.3.5)

is harmonic for RwProof. Since Px is a Radon probability measure on Q for each x G X, by Lemma
2.4.1, it follows that the following infinite-sum representation for h (x) is convergent:
oo

With the Ruelle operator i? = i?^, we get

AA-l

(Rh) (x)=^W

(TSX) h (TSX)

5=0
s

iU-.Jn

= h{x),
which is the desired conclusion.

5.4 A uniqueness theorem

In our next result, we give a sufficient condition for the property
Px (No) = 1

(5.4.1)

to hold, hi [GunOO], the process {Px)x^x is said to be tight if (5.4.1) is satisfied.

5.4 A uniqueness theorem

87

By now the transfer operator is often called the Ruelle operator because of D.
Ruelle's use of it in the 1960s on phase-transition problems in statistical mechanics
[Rue69]. The probability aspect of this endeavor was stressed by R. Gundy [GunOO],
and others, e.g., [CoRa90]. Our presentation here relies in crucial ways on Gundy's
viewpoint, as it seems to unify the different threads in the subject that came before it.
Three early elements of the theory may be summarized in three different but closely
related aspect of filters (as filter response fijnctions) in the form of: (1) tilings (A.
Cohen), (2) the transfer operator (W. Lawton), and (3) cycles (I. Daubechies, among
others). For more citations, we refer to the comments at the end of this chapter, and
to the references in the papers and books cited above.
Mathematically, a filter is just a sequence, but a very special one, and filters
were used originally in operations on discrete time signals. In its simplest form,
this operation is merely the Cauchy product of sequences. If the filter sequence is
taken as the Fourier coefiicients of a periodic fijnction m, then we say that m is
thefi*equencyresponse fimction. The operation on fimctions is now just pointwise
multiplication. The numbers that occur as filter sequences are closely related to the
masking coefficients of computer graphics; see, e.g., [BrJo02b] and [Jor03].
The filters have been used in the form of subband filters, and quadrature-mirror
filters, in image/signal processing for half of a century, and it isfi*omsignal engineering of processing that the subject of wavelet theory has adopted such engineering
terms as low-pass/high-pass filters, down-sampling, up-sampling, and perfect signal
(or image) reconstruction.
Since wavelets ideally aim for orthogonal bases, initially in the Hilbert space
Z^(E), properties of filters which detect orthogonality of the basis fimctions were the
focus of attention for both the initial work on wavelet filters, and the subsequent more
probabilistic approach. The probabilistic approach in fact merges the three initial
criteria, i.e., tiling (A. Cohen), the transfer operator (W. Lawton), and cycles (graph
theory on trees, discrete paths, and random walk). Each of the three elements alone
misses some important features of orthogonality tests for filters, with filter response
fimctions that are singular, for example filter fimctions that are only measurable, and
not continuous. This generalization is not just idle abstraction, but is motivated by
uses on wavelet bases with localization in fi-equency bands; see, e.g., [BaMe99]. If
more smoothness in the time domain is imposed, then (by uncertainty) irregularities
in the fi-equency domain tend to pop up.
A fiirther advantage of the new meeting ground for analysis and probability is
its impact on other basis issues in harmonic analysis, such as those arising in the
joint work of the author with S. Pedersen and D. Dutkay, e.g., [Kat87], [JoPe96],
[DuJo06b], and related research by R. Strichartz, Y. Wang and others. See [JorOS],
[JoPe98], [DuJo06b], and the papers cited there, for additional discussion and references.
We aim to have the theorems which follow in this chapter combine the diverse
ideas and results that came out in the last two decades. While these results originated

88

5 Infinite products

with diverse mathematical problems or engineering applications in mind, they all

seem to share a fimdamental mathematical principle, which perhaps now with hindsight finally emerges as a sharper contour. So at this time, the subject seems to have
reached a level of maturity, and we hope it is by now ready for the classroom. But the
individual results (over the two decades) taken in isolation are likely to miss some of
the common threads.
Theorem 5.4.1. Let X, a, ro, . . . , TAT-I, W, cind {Px)xex ^^ ^^ above. Suppose for
someX e Xthere is anno eN^J^db e R_|_ such that
Y[ W {n^ . . . Ti,x) > b

for all (/I, / 2 , . . . ) .

(5.4.2)

p>no

Then it follows that

Px(m

= ^'

(5.4.3)

Remark 5.4.2. In probabilistic terms, condition (5.4.2) is an assertion about the tail
sequences, i.e., the estimate
PX(ZN

X'"

XZN^

X {co}) > b

for all co e Q..

(5.4.4)

o times

Note that each set in (5.4.4) is finite, of cardinality N^^, but there is an infinite number of sets.
Proof of Theorem 5.4.1. Suppose condition (5.4.2) above holds. Let no and b be chosen as stated. Since ^ < 1, we get
Y[ ^ i'^h '"'^h^)
p>n

^ n

^ {'^h ''' '^^i^) - *

^^^ ^^^ n>no.

(5.4.5)

p>no

Using the unique representation

k = i^-\-i2N-\

+ /7V''"^

for A: E No,

we set length (k) = n if in is the last non-zero term, and

Y[ W [tip ' . Tf^x)

if length (k) < n,

p=\

otherwise.

Then

X^i"nA:) = l,

(5.4.6)

and
lim F W {k) = P^ m)

(5.4.7)

5.4 A uniqueness theorem

89

Ir

Fig. 5.1. The 2-adic fractions.

For > ^0, we get the estimate

F^^"> (K) < b-^P^ ({k})

for k E No.

(5.4.8)

Since
A:No

the estimate (5.4.8) shows that the convergence (5.4.7) is dominated. Hence we may
exchange the limits. Condition (5.4.6) shows that, as claimed,

A:eNo

Example 5.4.3. W (x) := cos^ (3;rx), X = [0, 1 ], cr:x -^ 2x mod 1, r/ (x) =

(x + z) /2, / 6 {0, 1}. See Figure 5.1. The condition (5.4.2) is not satisfied for any
X e [0, 1 ], and
To see this, recall that if
A: = /i+2z2 + ---+2''-^z,
then

To show that (5.4.2) does not hold, we check that for all b e M+, all no e N, we can
find n > no and k e No, k < 2"~^, such that
2/37r(x+^)\
cos

2-

<b.

90

5 Infinite products
1|

a^

Fig. 5.2. Powers of a.

Since lim^oo (37rx/2") = 0, the conclusion follows: just approximate 1/6 with a
sequence of dyadic rationals k/2^, 0 < k < 2"~^ Then
lim cos I
{n,k)
V

I = lim cos^ I
I
(n,k)
V 2" /
= COS

As an application to the usual 7V-adic representation of fractions, we now read

off the following corollary. First set X = [ 0,1 ] = the unit interval, and let
a (x) = Nx mod 1,
T^s(x)

X -i-s

and
(5.4.9)
X E [0,1], s =0,1,...,7V-1.

If A^ = 2, these mappings are graphed as in Figure 5.1. Moreover, for this example,
we can compute
h (x) = P, (Z)
explicitly. In Section 6.2, we show that
.^.
/sin3;rx\^
P,(Z)=l
) .
\3sm;rx/
Corollary 5.4.4. Let X = [0, I], N e N, N > 2, be given, and let a, TQ, ..., TN-I
be the N-adic maps o/(5.4.9). LetW\[Q,\\^
[0,1 ] satisfy (2.4.1), i.e.,
N-\

(5.4.10)
Let (^:!c);ce[0,l ] ^^ ' ^ transition measures of Lemma 2.4.1. Then

5.5 Wavelets revisited

hix) = PAm
OO

91
(5.4.11)

OO

is convergent and represents an Rw-harmonic function o^ [0, 1 ].

Following Remark 1.4.2, we extend the measure Pjc on No to Z, making use of
two copies of Q = {0, 1 , . . . , TV 1}^. Identifying A: e No with co (A:), consider the
case N^ < k < 0. Then set
Px (W) := Px ({co [N^^' + A:))) .

(5.4.12)

5.5 Wavelets revisited

Then dulcet music swelled
Concordant with the life-strings of the soul;
It throbbed in sweet and languid beatings there,
Catching new life from transitory death;
Like the vague sighings of a wind at even
That wakes the wavelets of the slumbering sea...
^Percy Bysshe Shelley
We now use the martingale method of Remark 2.7.2 in the computation of the
cocycle V associated with the distinguished harmonic function h from (5.4.11).
Example 5.5.1. It may very well happen that the measure P^ assigns zero mass to
singletons in Q, i.e., that it is non-atomic.
(a) Let X = [ 0, 1 ], and let G be the endomorphism in Figure 5.1, i.e., a (x) =
2x mod 1. Let W be the constant function, i.e., W = 1/2, x G [0, 1 ]. Then it is
immediate that P^ is the product measure on Q = {0, 1}^ with weights pQ = pi =
1/2 for all X e [ 0,1 ], and therefore P^ ({co}) = 0 for all co e Q.
(b) Let (p e L'^ (R), and suppose q) satisfies the scaling identity (1.3.1) for TV = 2
and some sequence a^ of masking coefiicients with
^^akcik+ll = 2^0J '
keZ
Set

W(x) =

Zk akc

2
ilnkx
I
-i

Suppose
/ \(p{x)\^ dx = 1.

92

5 Infinite products

Then it follows from Corollary 5.4.4 that

P, {{co{k)}) = 1^ (x-}-k)\^,

keZ,

(5.5.1)

where co (k) = (co\, co2, ..) is given by Euclid and k = co\ -{- oyil H
0 < A: < 2^^ - 1.

+ co2"~^,

Proposition 5.5.2. There is a shift-invariant function t>: Q ^ [ 0? 1 ] ^^^h that

V(x,co) = h(v(co)),

CO en,

(5.5.2)

is the cocycle corresponding to the harmonic function h in (5.4.11). Moreover, if

a; = (coi, C02,...) e ^ = { 0 , . . . , A^ - 1}^, then
V (co) = lim I

r H

I.

(5.5.3)

In particular, v vanishes on the N-adic rational fractions.

Proof Since there is a 1-1 correspondence between harmonic ftinctions and cocycles, see Theorem 2.7.1, the cocycle corresponding to /z in (5.4.11) is unique. If we
show that it has the form (5.4.11), it follows from Corollary 5.4.4 that the fimction v
must be shift-invariant, i.e., that
t; (cwi, cy2,...) = t) {co2, C03,...) .

(5.5.4)

In view of Remark 2.7.2 (or the martingale convergence theorem), we know that, for
Px a.e. CO, we have
(5.5.5)
V (x, co)= \\m h (T(O^ Tfo^x),
but
X

CO\

CO2

COn

^^

^^

and it follows that the limit in (5.5.5) is independent of x. But the sequence of fractions in (5.5.3) or (5.5.6) is non-increasing, i.e.,
CO\

02

COn

CO\

C02

COn-\-\

The monotonicity follows from an easy induction argument. The fractions are clearly
in the compact interval [ 0, 1 ] c M.
In order to understand the fimction t>: Q > [ 0,1 ] from (5.5.3), note that
v{co\, C02, . . . , COn, 0, 0, 0, . ) = 0,

(5.5.7)

00 string of zeroes

i.e., V vanishes on the 7V-adic rationals.

Exercises

93

Exercises
5.1. Let
keZ
and set
(Sf)(z) = V2m(z)f(z^y

feL^T).

(By L^ (T), we mean the Hilbert space of all X^-functions on T where T is equipped
with the unique normalized Haar measure.)
Show that the following three conditions are equivalent.
(i) S defines an isometry of Z^ (T).
(ii) \m (z)p + \m (-z)\^ = 1 a.e. z e T.
(iii) ^akak+2i = -Soj, I e Z.
keZ
^
5.2. LQtN eN,N > 2, and let
m (z)
{ =

Y,^kZ^^L^(T).

Set
(Sf)(z) = VNm(z)f(z''y

feL\T).

Show that the following three conditions are equivalent,

(i) iS'defines an isometry of Z^ (T).
N-l . .
2
(ii) ^ \m (ze^'2^^/^) I = 1 a.e. z e T.
lu.ni. I ATI =
^ T7^0,/, / e Z.
(iii) y^akak^-Nl

N
5.3. Let iV e N, iV > 2, and let
OT(

As in Exercise 5.2, set

Sf(z) = ^m{z)f(fy

/l2(T).

The adjoint operator F S* \s, defined by the identity

94

5 Infinite products
(a) Show that
1
( F / ) ( z ) = - Y,
^/N

m(w)f(w),

feL\T),zeT.

weT

(b) Let M= M,hQ the multipHcation operator (Mf) (z) = zf (z), f e L^ (T),
z e T. Then show that M commutes with S*S, i.e., (S*S) M = M (S*S).
(c) Does M commute with SS*?
(d) Does M^ = M^N commute with SS""!
5.4. Let A^, m, and S be defined as in Exercise 5.3. Then show that S is 2i partial
isometry (i.e., S*S\S2i projection operator) if and only if there is a measurable subset
^ C T such that
S*S = M XE
where
{M^J)

(z) :=

XE

(^) / (^) >

f s L\T)

, z B

T.

5.5. Let (X, B) be a measure space, and let /^ be a probability measure defined on the
cr-algebra B. Let N e N, N >2, and let cr: X -> X and Zk'.X ^ X be measurable
endomorphisms of X such that
for 0 < A: < TV.

a oTk = idx

Consider functions m e Z'^ (X, S) =: Z ^ . Set

JV-l

SF(r):=

meL'

|m o TA;|^ = 1 /i-a.e. o n X

A:=0

Set
{Smf)ix)

=^

m{x) f {a {x)) ,

feL^{X,fji),

x e X.

Show that Sm'- L^ (X, /^) -^ L^ (X, //) is an isometry for all m e QF (r) if and only
if/^ satisfies
A^-l

-1

(5E.1)

^ = 77^=0S/ / o r ^
5.6. Let A^ e N, # > 2, be given, and set
a (x) = A/^x mod 1,
X + A:
7;^ ( ^ ) =
N

0 <k

<N.

(a) Then show that the restriction ^a to [0, 1) of Lebesgue measure on M satisfies
1

N-\N-\

A:=0

Exercises

95

(b) With the identification R / Z = [0, 1), and a function m e L^ (0, 1), set
(Sf) (x) = VNm (x) f (Nx mod 1) ,

(5E.2)

or
Sf = ^/Nmfoa,

feL^(0,l).

Then show that S is an isometry in Z^ (0, 1) if and only if

Z r ( ^ )

=^

a.e.on[0,l).

(c) Let S = SmhQ defined as in (5E.2) for some m e L^. Then show that S*S is
a multiphcation operator, i.e., that S*S = Mf for some F e L^.
(d) For F in (c), show that
A^-1 I

/^

7.\ |2

a.e. X.
I

A/

/ I

k=0

(e) What can you say about SS*"?

(f) Let mj, j = 1,2, be two 1-periodic fimctions in Z ^ , and let Sj be the corresponding subdivision operators defined as in (5E.2). Then show that S\S2 is a
multiplication operator; specifically, AS'**S2 is multiplication by
N-l

5.7. Let O: M -> C be a given function such that the sum

Av O (x)

:=y^(!)(x-\-n)
neZ

is pointwise convergent a.e. x E M. Let ^ be a 1-periodic fimction on R (i.e.,

W(x -\-l)=W (x), X e R), and suppose that for some TV e N, iV > 2,
W (x) (D (x) = O (Nx),

Xe

holds.
Set

<-/)<5V(T)/(^)
for periodic fimctions / on R. Then show that the function
/2(x):=AvO(x)
satisfies
i.e., that /? is a harmonic function for the Ruelle operator Rw-

96

5 Infinite products

5.8. Let W:R -^ C be a measurable fiinction, and suppose (i), (ii), and (ill) below
hold.
(i) W(x) = W(x-\-1) a.e. x eR, i.e., W is 1-periodic.
(ii) For some N eN, N >2, the infinite product
oo

is convergent a.e. x G R.
X G

Then prove that the fiinction O in (5E.3) is in Z ^ (M), and that

| 0 ( x ) | dx < 1.
)
5.9. Let W G L^ (R), and assume that W is 1-periodic, and fiirther that it satisfies
W(x)-j-w(x-}--\

=1

a.e. x G M.

(5E.4)

(a) Show^ that J^ W(x) dx = ^.

(b) Show thatft)rall ^ G N, the fi)llowing identity holds:

(c) Show thatft)rall k eN, the ft)llowing identity holds:

(d) Suppose, in addition to the above, that W > 0 a.e. on R, and fiuther that the
infinite product

is (pointwise) convergent on R a.e. x. Then show that

oo

^(x)dx<l.

(5E.5)

-OO

(e) Under the assumptions on W listed in (d), give examples of when the inequality (5E.5) is strict (i.e., <), and when it is "=".

References and remarks

97

Hints to 5.9: (a) Introduce the M / Z -Fourier coefficients {ck)kE:'L f^^ the function
W, and note that the identity (5E.4) is equivalent to
2c2w = ^o,,

n elj.

Also recall that CQ = JQ W {X) dx.

(b) Prove this by induction, noting that the case A: = 1 is equivalent to

But both sides of this last equation agree with the Fourier coefficient co (= \) from
(a).
(c) Again this may be proved by induction. While the case A: = 1 is clear, the
next case A: = 2 is instructive:

using that W is 1-periodic

= 41
Jo

E.4)
(5E.

A jyw{x)

+ w(x

W{2x) dx=2

+ ] \ \ W (2x) dx

I Wix)dx
Jo

= 1.
(a)

(d) For each A: e N, vs^e have

J-2^-'nJ[

^2^^^

by(c)

Letting A: ^ oc, we get the estimate (5E.5) by an application of Fatou's lemma.

Verify that Fatou applies in this case.

References and remarks

Some relevant references (books and papers) covering Riesz products are [Riel8,
Mey79], and a list covering related infinite-product constructs includes the following:
[Zyg32, May91, Kat87, Kak48, BeBKOS].
The idea of using the Ruelle operator and the Kolmogorov extension principle in
the construction of wavelets from Perron-Frobenius eigenfunctions was suggested
in Jorgensen's Memoir [JorOla], and in two papers by Dorin Dutkay, [Dut02] and
[Dut04a].
One of the advantages of this approach is that we get multiresolution wavelets in
other Hilbert spaces than the familiar L^ (M^): for example, in Hilbert spaces built
on fractals by the use of Hausdorff measure; see [DuJo06b]. See also [Dut04b].

98

5 Infinite products

In the prelude, we mentioned that random-walk models are realized on infinite discrete sets S such as the standard lattices Z^, d = 1 , 2 , . . . , and on various
combinatorial trees rooted at points in a given compact space. The first model is
widely studied in probability (see, e.g., Revuz [Rev84], Stroock [Stro05], and Spitzer
[Spi76]); and we have stressed here uses of the tree models in the study of scalesimilarity, as it is used in the analysis of wavelets and of fi-actals.
hi all the models, it makes sense to introduce spaces of paths, and path-space
measures, and there is a separate literature on this; see e.g., [Nel73] and [Sim79].
Somewhere between these two structures (lattice and combinatorial tree) there are
yet other uses of the idea and related probabilistic models. They are encountered for
example in thermodynamics, e.g., in percolation models fi-om mathematical physics
[FiEs61]. The infinite sets S we meet there could be Cay ley trees, Bethe lattices, or
bond graphs; see figures 1 and 4 in [FiEs61]. In these models the standard combinatorial tree is modified. Instead of a distinguished "root" or starting node, these models
allow more complicated cactus-like root configurations, and the resulting shapes are
called pseudolattices. While this theory does combine analysis and probability, it is
however beyond the scope of our book; and we refer the interested reader to [FiEs61]
and the papers cited there.
The relation between our tree in Figure 1.1 (p. 8) and the Bethe lattice is like the
relation between a lattice of N-points (non-negative integer coordinates) and a lattice
of Z-points (integer coordinates of either sign). (In a Bethe lattice there would be a
third subtree attached above the point labeled "x.") The "Bethe lattice" is uniform,
having no distinguished root, point, origin. The shapes in the figures fi-om [FiEs61]
are "cactus-like" in appearance.

The minimal eigenfunction

Anything that, in happening, causes itself to happen again, happens again.

^Douglas Adams

PREREQUISITES :

Eigenvalues, eigenvectors; eigenfunctions; a curiosity about wavelets as seen by a petroleum engineer.

Prelude
This brief chapter serves as a bridge between the iterated function systems (IFS) in
Chapter 4 and the more detailed wavelet analysis in Chapters 7 and 8. We wish to
examine a variation of scale, i.e., examine the effect on the Perron-Frobenius-Ruelle
theory induced by a change of scale in a wavelet basis. After stating a general result
(Theorem 6.1.1), we take a closer look at a single example: Recall that Haar's wavelet
is dyadic, i.e., it is a wavelet basis for L^ (M) which arises from the operations of
translation by the integers Z, and by scaling with all powers of two, i.e., scaling
by 2^, as j ranges over Z. But the process begins with the unit box fimction, say
supported in the interval from x = 0 to x = 1. The scaling by 3, i.e., f ^ f (x/3),
stretches the support to the interval [ 0, 3 ]. It is natural to ask what happens to an
ONB dyadic wavelet under scaling by 3, i.e., f -^ f (x/3). This is related to oversampling: The simplest instance of this is the following scaling of the low-pass filter
mo (x), i.e., mo (x) -^ mo (3x).
Using Perron-Frobenius-Ruelle theory, it can be shown that the ONB property
typically is lost, but still a number of essential wavelet features are preserved. Recall
that a consequence of the ONB property is a version of the Parseval identity. Even
though the ONB property is lost, it follows from our discussion below that the Parseval property is preserved under stretching, i.e., the stretched wavelet will still be a
Parseval frame, also called a normalized tight frame.

100

6 The minimal eigenfunction

We have chosen to restrict our discussion of this to one example, the "stretched
Haar wavelet." However, we emphasize that much more is known, but beyond the
scope of our book. While there are many ways to extend these developments, we
have restricted the discussion here to one example. Still a lot more can be done:
We know that if y/ is a frame wavelet {N = 2) and if /> is a fixed odd integer,
then the scaled fimction (l/p) y/ (x/p) is also a frame wavelet. And even in higher
dimensions, the theory takes a nice form.
From wavelets in one dimension to higher dimensions: Instead of the familiar
1-D translation by Z and scaling with all powers of two, in higher dimensions d the
translation will instead be by a mink-d lattice and we will be scaling by an integral
and expansive matrix A, i.e., scaling by AJ\ as j ranges over Z.
In the context of the path-space measures Px and the associated harmonic fiinctions from Chapter 2 a lot is known about change of scale. Specifically, in Chapter
2 we introduced the family of probability measures Px for the case of wavelets on
W^ (with an expansive matrix A over Z), and we showed that each Px is atomic. Is
the transformed Px family again atomic? Or in other words, are the transformed (or
stretched) Px measures still concentrated on a rank-J lattice, i.e., on some version
of Z*^, just as in the 1-dimensional case. Further it is important to consider the effect
on cycles (for the random walk) under "matrix-stretching," and the effect on other
structures such as the zeroes of the Pjc-harmonic fijnctions. And try the same for
self-affine tilings. Or develop a theory for an over-sampling in the general case of
higher dimensions and even for fractals.

6.1 A general construction of/xmin

The fiinction h (x) := Px (No) is special among the solutions h >0to the eigenvalue
problem
Rwh = h.
(6.1.1)
Our next result shows that x v^ Px (No) is minimal among the solutions to (6.1.1),
and we shall use the notation
hrnm{x):=Px{^o).

xeX.

(6.1.2)

In an indexed family of wavelet examples. Figure 6.2 (p. 107) contains plots of these
fiinctions/^min-

6.1 A general construction of /zmin

1^1

Theorem 6.1.1. Let X, B, N, W be as in Chapter 2, and a: X -> X an endomorphism such that #a~^ (M) = N, x e X, and assume in addition that X is a
compact Hausdorjf space. Suppose branches of o~^ may be chosen such that, for
some measure VQ on X,
}^oo

(^^"^ ^ ) (^- ^-1^) = ^0 {g) = f gdvo

(6.1.3)

for all g ^C (Q). The function W^^"^ is defined by

^() {y)=W{y)W{a{y))---W

(^""^ {y)) .

(6.1.4)

Let h e C (Q) be a solution ^o (6.1.1) such that h > 0. Then

^min (x) VQ (h) < h (x) ,

X E X.

(6.1.5)

Remark 6.1.2. With the normalization VQ (//) = 1,(6.1.5) reads/zmin < /z, and hence
the reference to hmin as the minimal eigenfunction. We saw that there are examples
where /zmin = 0.
Proof of Theorem 6.1.1. Let k e No, and consider the A^-adic representation k =
h inN'^'K Note that
i\ +i2N -\
co(k) = (iu

. . . , in, 0, 0, 0, ) >
(X) string of zeroes

and that
h(x) = R''j^Ph(x)

(6.1.6)

For n fixed, let p -^ oo. Then

^i^

^ i^m^) " ' ^ {'^o, " "^coix)' "

(6.1.7)

(co+l,...,co+p)
^

i'^COrr+p "'Tco^x)h

[Xcon^p ^ To,^x)

= P^ {[cO {K)}) VQ Ql)

for k = co\+ C02N H

We now exchange limits, and use Fatou's lemma in the form

V I lim i < lim V

h conN"^'^.
(6.1.8)

102

6 The minimal eigenfunction

Combining (6.1.6) and (6.1.7), we get

Y,PA{co{mvo{h)<h{x).

(6.1.9)

A:eNo

But the factor on the left-hand side in (6.1.9) is Px (No) = //min (^), so the desired
estimate (6.1.5) follows.
n
Remark 6.1.3. The wavelet representation is included in the framework of the theorem. Then X = [ 0, 1 ], and VQ = S{) = the Dirac mass at x = 0 ; see [BrJo02a], and
[BrJo02b, Proposition 5.4.11].

6.2 A closed expression for /zmin

Example 6.2.1. Haar wavelets. We now illustrate the minimal eigenfunction for
Rw with the following family of Haar wavelets; see also [BrJo02b] for more details.
First, note that the simplest version of the scaling identity (1.3.1) from Chapter 1 is
(p{t)=(p (20 + ^ (2^ - 1) ,

t E

(6.2.1)

t e

(6.2.2)

The corresponding wavelet fiinction y/ is given by

yj{t) = (p (20

-(p{2t-\).

Even without Fourier transformation (1.3.3), it is immediate by inspection that

the equations (6.2.1)-(6.2.2) are solved by the Haar father fimction,
1,
0

(p(t) =

0 < / < 1,
otherwise,

(6.2.3)

and the Haar mother frinction,

^(0 =

1,

0 < ^ < 1/2,

-1,

1/2 < ^ < 1,

otherwise;

(6.2.4)

see also Figure 6.1(a). Also consider the functions

t e

(6.2.5)

t e

(6.2.6)

and

wp(0 = ^w{^).

for /> = 3, 5, odd numbers. The pair ^3 and y/3 are sketched in Figure 6.1(b).

6.2 A closed expression for /^min

103

n
1 /

-1

2>
1

-1

Fig. 6.1. Haar's wavelets (compare Figure 1.2, p. 13): Haar's two functions cp (6.2.3) and y/
(6.2.4) are illustrated in the two columns: The system in the column on the left in the figure
illustrates Haar's orthonormal (ONB) wavelet basis, and the scaling identities are visually
apparent. The system of functions ^3 (6.2.5) and ^3 (6.2.6), for /? = 3, in the right-hand
column illustrates the two stretched Haar functions. The two scaling identities (6.2.1) and
(6.2.2) are visually apparent for both function systems, on the left and on the right. Moreover,
for both systems, the ^ function yields a double-indexed Parseval basis (see (1.3.16)), i.e., a
function system for which the Parseval identity (6.2.12) holds for every function / in L^(R).
\f p = \ (the left-hand-side case), then the inequality in (6.2.11) is an equality, i.e., is an =;
while for p = 3 (the right-hand-side column in the figure), the term on the left in (6.2.11)
is strictly smaller than ll/ll^2rro\ ^^^ non-zero / . The reason for this is the overlap for the
Z-translated functions, which is graphically illustrated too.

104

6 The minimal eigenfunction

As suggested from Figure 6.1(b), the wavelet ^^, for /> = 3, 5 , . . . , is called the
/?-stretched Haar wavelet.
We leave it to the reader to check that the low-pass filter fijnctions nip for the
wavelet pair cpp, xj/p are
\ _j_

^-i27tpx

nip (x) =

X G M,

and as a result,
I

|2

^P (^) = y^p (^)| = cos {npx) .

It is known, and easy to check, that the orthonormality relations hold for ^, y/,
but not for (pp, y/p, p = 3,5,
For (p, ^ , the orthonormality relations are
II^IIL2(]R) = lkllz2(M) = 1,
{(pi- -k)\k

eZ]

(6.2.7)

is orthonormal,

(6.2.8)

i.e.,
{(p \(p(- -k))^2(]^) = So^k.

keZ,

and
12"/^ v^ (2"^ - ^) I ^, A: e Z } is an orthonormal basis (ONB) for L^ (M). (6.2.9)
In contrast, (pp, y/p satisfy
lki^lL2(R) = lk;.llL2(M) = - ^ .

(6.2.10)

ZI(^/^(--^)i/>l'^ii/iiW)

(^-^-ii)

keZ
for all / in the closed subspace VQ spanned by the translates [cppi- k) \ k e Z],
and

^ X | ( 2 / V ; . (2^^ -k) I f)f = ||/|li2(M)

forall/ G L\R) . (6.2.12)

k,neZ
The properties (6.2.11)-(6.2.12) are called^awe properties. Basis fimctions satisfying such a generalized PARSEVAL identity are called Parsevalframes, or normalized tight frames. This frame requirement is evidently weaker than the requirements
of an ONB; see (6.2.8)-(6.2.9).
The discrepancy between them may be measured by the autocorrelation coefficients from (6.2.8): since (6.2.8) is not satisfied for cpp, /? = 3, 5 , . . . , the 1-periodic
fimction

6.2 A closed expression for /zmin

105

hp{x)'.= Y,{cpp\(pp{^ - ^ ) ) L 2 ( M ) ^ ^ ' " ^ '

(6.2.13)
A;eZ
is not constant. Of course /?i (x) = 1, x G M.
In many applications, it is too much to require that the wavelet produce a true
ONB, i.e., that (6.2.9) be satisfied, and the weaker requirement (Parseval fi-ame)
(6.2.12) is typically good enough. For these wavelets, the minimal eigenfimction
satisfies
0</\(Z)<l,
XG[0, 1],
as we shall now illustrate.
From the analysis in Chapter 1 above, we note that there is a wavelet transfer
An inspection of (6.2.5)-(6.2.6)
operator Rp = Rwp for each p = 1,3,5,
shows that
Wp (x) = cos^ (npx) .
(6.2.14)
Now take
v=do=S{x-0).

(6.2.15)

Then it is clearfi-om(6.2.14) that

vRp = v,

(6.2.16)

or equivalently,

J Rpfdv=J

fdv,

/GC([0,1]).

Using now Lemma 1.4.1 and the Fourier transform (1.3.3), we also conclude that

kk

kk\

^P(^-k)

p'KsmTcx

J
(6.2.17)
which is a closed expression for the minimal harmonic fimction; see also [BrJo02b,
page 332, Figure 6.3].
It follows fi-om (6.2.13)-(6.2.17) that
v{hp) =3o{hp) = 1,
and that
RWphp=hp,

(6.2.18)

and we conclude fi-om Theorem 6.1.1 that hp is the minimal harmonic fimction.
Specifically, ifh is any other continuous fimction satisfying
/z(0) = l,

Rwph = h,

0</z<l,

(6.2.19)

then
hp <h <\
See Figure 6.2.

pointwise.

(6.2.20)

106

6 The minimal eigenfunction

Exercises
6.1. (a) Set W (x) = cos^ (37vx), and show that

n-(?)

^^1

'

sin'^ (STTX)
(STUXY

(b) Verify by a direct argument that

while

L-Mri^)'-'I
^ ^ sin^ (STTX) ,

-oo

ax =

Jc

(STTX)^

6.2. Let y/ be the Haar wavelet fiinction from Figure 6.1 (p. 103), and set
y^j

,k(t):=2J^^y/(2Jt-ky

(a) Show that if y e No, and k = 0, 1 , . . . , 2^ 1, then the functions y/j^k are
supported in the unit interval / = [0, 1].
(b) Show that the fiinctions y/j^k indexed by ^ := {(/, k) \ j e No, k =
0, 1 , . . . , 2^ 1} form an orthonormal basis (ONB) in I ^ (/) when / is given the
Lebesgue measure, i.e., that { y/j^k \ (j^k) e A] is an ONB for L^ (/).
(c) Sketch the basis fiinctions in the cases 7 = 0,7 = 1, and j =2.
6.3. Can you find other "interesting" fiinctions ^ on / [0, 1] and subsets A c
Z X Z such that the family { y/j^k 1 0 ? ^ ) ^ ^ } forms an ONB for L^ (/), where
yfj,k(t):=2J^^y/{2Jt~k)l
6.4. Let y/3 be the stretched Haar wavelet function from Figure 6.1 (p. 103), and set

^)5(0:=2^/V3(2^V-^).
(a) Show that if 7 G No, and A: = 0, 1 , . . . , 3 (2^ - l), then the fimctions y/^^l
are supported in the stretched interval /s := [0, 3]; see Figure 6.1.
(b) Show that the functions y/^l indexed by As := {(j\k) \ j e No, k =
0, 1 , . . . , 3 (2> - 1)} form a Parseval frame for L^ (/s), i.e., | y/^^l \ (j\ k) e As]
is a normalized tight frame in Z^ (/s); specifically that

z
holds for all f e L^ (h).

(^s|/Uj'=/i/i^^^

107

References and remarks

h = hi, (3;c)

1
3^

/^^i^--^ ^;t^\ /L- ^X 924

Fig. 6.2. Solutions to the eigenvalue problem (6.2.19).

References and remarks

r and its
Some relevant references (books and papers) covering the Ruelle operato
, Jam64,
CoRa90
b,
BrJo02
BalOO,
,
PaPo90
,
[Kea72
are
Perron-Frobenius eigenspace
.
May91, Rue02, Rue69, Rue94]
tions not
The wavelet transfer operator is used in a variety of wavelet applica
s, regcovered here, or only touched upon tangentially: stability of refinable function
mining, to
ularity, approximation order, unitary matrix extension principles, and data
details
more
for
papers
g
followin
the
mention only a few. The reader is referred to

108

6 The minimal eigenfunction

on these subjects: [DaHRSOS, RoSh03, RSTOl, JiJSOl, RoShOO, RiShOO, JiSh99,

She98, RoSh98, LaLS98, RoSh97].
Several of these topics and papers invite the kind of probabilistic tools that we
stress in our book, but a more systematic discussion is outside the scope of our book.
We only hope to offer an introduction to a variety of more specialized topics.

Generalizations and applications

One cannot expect any serious understanding of what wavelet analysis

means without a deep knowledge of the corresponding operator theory.
^Yves Meyer

PREREQUISITES: Having seen a few approximation problems; spectrum; tensor

product of matrices; unitary matrices; Hilbert space; a rough idea about recursive
algorithms (visual representations); large matrices; a curiosity about some cute ideas
from engineering.

Prelude
In our earlier discussion of "wavelet-like" bases in Hilbert space H, we stressed
the geometric point of view, which begins with a subspace FQ and two operations. For standard wavelets in one variable, H will be the Hilbert space L^ (R),
and a suitable "resolution subspace" To will be chosen and assumed invariant under translation by the group of integers Z. In addition, it will be required that VQ
be invariant under some definite scaling operator, for example under "stretching"
/ -^ f {x/2). Thirdly, the traditional multiresolution (MRA) approach to L^ (M)wavelets demands that the chosen subspace FQ be singly generated, i.e., generated
by a single function cp, the father fiinction, i.e., the normalized Z^-fiinction which
solves the scaling identity (see (1.3.1) in Chapter 1). As is known, it turns out that
these demands for a subspace FQ are rather stringent.
Hence in this and the next chapter, we shall generalize the familiar MRA picture
in two ways. First, we shall allow Fo to be multiply generated: more than one father
fimction. When FQ is multiply generated, it is said to give a generalized multiresolution analysis (GMRA). Secondly, we shall extend the MRA idea to other Hilbert
spaces than L^ (M) or L^ '

110

7 Generalizations and applications

We saw in Chapter 4 that the idea of a "multiresolution" applies to fractals. An

inherent feature about fractals Jf is a scaling law. Lituitively a fractal JT is a "fernlike" geometric object which looks the same when a natural "scale" is varied, i.e.,
the same at small scales and at large scales. The notion of scaling can take many
forms: it can be arithmetic (e.g., dyadic scaling, 7V-adic scaling, or matrix scaling), or
geometric (e.g., iterated composition of rational functions of one complex variable),
or symbolic such as is employed in encoding of discrete automata-dynamics by shifts
or by substitution in a finite alphabet.
We will see that the geometric feature of a traditional wavelet "multiresolution" is very adaptable. If we think of a multiresolution as a selection of a distinguished subspace FQ in a "fractal" or symbolic Hilbert space Ji, then it turns out that
computational algorithms from wavelet theory carry over. But in these more nontraditional variants of "multiresolutions," perhaps the analogue of the familiar group
of Z-translations in the Hilbert space L^ (M) is not immediately transparent. Yet it
is well known, and recalled in (7.1.4) below, that the group of Z-translations in the
Hilbert space L^ (M) is Fourier equivalent to multiplication in L^ (R) by the Fourier
frequencies e^^^^^. (Note that the spectral transform in (7.1.4) below is more general. Specifically, the transform for a generalized multiresolution (GMRA) should be
thought of as an extension of the basic Fourier equivalence principle!) And it turns
out that all the three non-traditional variants of "multiresolutions" do have abelian
algebras of multiplication operators which leave invariant suitable resolution subspaces FQ.

7.1 Translations and the spectral theorem

Special to this chapter is a set of sequential figures and illustrations: Figures 7.1,
7.2, and 7.3 (decision-tree view, pp. 111-113), and Figures 7.4 and 7.5 (graphs of
two classes of wavelet packets, pp. 118-121); the sequence from Figure 7.6 to 7.15
(subspace analysis, pp. 123-128, 132-133); and Figure 7.16 (visual exhibit of the
sequential cascade algorithm, here applied to the two Daubechies wavelet fiinctions
(p and I//, p. 134.)
frideed, the central theme in the whole book is captured in the figures here in
Chapter 7. We resume specific and detailed citation of these figures by number later
in Chapter 7.

7.1 Translations and the spectral theorem

111

m 1 = high

mo = low

low/
^12

(P6
Miigh
^13

low/
^14

Miigh
^15

Fig. 7.1. The Fourier transform of the basis sequence (before scaling and selection) generated
from the scaling function g)Q, using the full combinatorial tree.

The first block of figures, i.e., Figures 7.1, 7.2, and 7.3, illustrate how the pyramid algorithmfi*omChapter 1 takes shape as a decision tree. While in our present
applications we emphasize the wavelet context for all of this, in fact it is the same
underlying idea which is used in our fi-actal constructions fi-om Chapters 3 and 4.
And the same idea may be used even more generally in a variety of data structures
which happen to come equipped with some inherent notion of scaling and similarity,
such as is known to be the case for digital data clouds, for graphs and for certain
manifolds.
For a number of applications, it is convenient to have a matrix variant of the
measures Px constructed in Chapter 2; see especially (2.4.6). For the wavelet applications, this is relevant when the scaling identity, or refinement equation, (1.3.1)
does not have solutions that are scalar-valued Z^ (E)-fimctions. For multiwavelets
this happens when the starting resolution subspace VQ (Z L-^ (M) has multiplicity. By
this, we mean that the representation of Z by translation in VQ is not cyclic. Equivalently, setting
Tkf(x)

= f(x-k),

feVo,xeR,keZ,

(7.1.1)

112

7 Generalizations and applications

{Ay/

\(5)

(^)/

(PA

95

\(5)

96

91

Fig. 7.2. The generation of the basis functions ^o? ^b ^2? t>y use of the algorithm (7.5.5)
initiated with cpo = cp such that ^(0 = 2 X / <^j9{^t j), the functional equation by which,
in thefirst(A) branch, (pQ "re-generates" itself

we get a spectral representation in the following form: there is an isometric isomorphism

J: Fo ^ L^ {Sx) e L^ (S2) 0
(7.1.2)
such that
^y+l C ^y C C ^2 C ^1 C [0, 1 ]

(7.1.3)

and
JTkf = e^^^^- J / ,

keZ,

fe

VQ.

(7.1.4)

To see this, note that, by the spectral theorem, there is a projection-valued measure
p:[0,l]-^

projections of 1^ (R)

(7.1.5)

such that

Tkf = I e'^^^' dp {X) /,

k^Z,

fe Vo.

(7.1.6)

JO

To get the sets Sj,J = 1,2,..., take

Sj:={Xe[OA]\dimp(X)>j},

(7.1.7)

The reader is referred to [BaMe99], [BaMM99], [BaCM02], and [BaJMP04].

These applications motivate two generalizations of the more traditional wavelet
setup, see, e.g., [Dau92]: first, the wavelet filters and the functions W (see Chapter 1
above) are not continuous, and second, they are matrix-valued.

7.1 Translations and the spectral theorem

113

(AL
nit)

= 2Y,aj(px{2t

- j)

(P2

<P3{t} =

2j:{-\yai_jn(2t-j}
J

VA

(B)
<P3

n
(AV
'16

v>9
UB)
m

'

(AV UB)
<PIS

/
(P6

(B)

<P19
(A)/

\(B)

UB)

^13
(AV UB)

cpsit)=2^{-\)j-ai_jcp2i2t-j)
(AY/
^24

(P25

(P16

(P5

(pi{t)=2Y.{-^V~a\-jni2t-j)

n\

^10
(A)/ U B )

(AV

^20

^22

^21

\
\(B)

^7

^23
(A)/

(PU
(AV \XB)
^28

^29

\(B)

(P\5
(AV \(B)
^30

^31

Fig. 7.3. The generation of the basis functions (p\, (p2, (P3, . . . by use of the algorithm (7.5.5)
initiated with cpi = y/.

114

7 Generalizations and applications

7.2 Multiwavelets and generalized multiresolution analysis

(GMRA)
We now outline the generalizations of the multiresolutions from Chapter 1 which
are entailed by the case of multiply generated resolution subspaces VQ C L^ (M).
The multiplicity ftinction ju is identically = 1 for the cases covered by (1.3.1). For
the general case, we consider ^:IR - ^ H for some Hilbert space H, and we are
interested in solutions cp e L^ (K). In this interpretation, the coefficients ak from
(1.3.1) are operators ak'. H ^ H, and the filter fimction (1.3.2) is operator-valued. Its
Fourier-transformed variant (1.3.4) must also be interpreted in this context: ^: R ^
n,m(x):n-^ K and
Hx) = m{^)^{^)
(7.2.1)
is an identity for vectors in H. Since

we need a product formula for ||^ (^) ||^.

Iteration of (7.2.1) yields

\\^(^nn =

{^{-^)\^n(x)^(-^)).

where
The function m is assumed to satisfy the normalization condition
2 ] m(y)m(yr^in,

x e X.

(7.2.3)

(T(y)=x

To state our result, we first provide the setting for our operator-valued variant of
Lemma 2.4.1.

7.3 Operator-coefficients
Definitions 7.3.1. (a) Let W be a complex Hilbert space. An operator T mHis said
to be positive (alias semidefinite) if {u \ Tu) > 0 for all w e H. The inner product
in 7i is denoted ( | ), and is assumed linear in the second variable. If (Q, Ba) is a
probability space, we say that a fimction

P:Ba-^Vos(n)
is an operator-valued measure if:

7.4 Operator-valued measures

115

(i) P (')is o--additive,

(ii) 0 < P (A) < tj-c^oY equivalently
0<{u\PiA)u)<

\\uf,

ueH,

A e BQ,

and
(iii) P (Q) = In.
(b) Let (X, B) be a measure space, and let W:X ^^ Pos (H) be a measurable
fimction, i.e., for every u eH,x \-^ {u\ W {x)u) is measurable, and
^<{u\W{x)u)<

\\uf.

(7.3.1)

Let a.X ^> X be an 7V-to-1 measurable mapping, with measurable inverse branches
TO,..., Tiv-i. Suppose
N-\

Y^WoTi = tn.

(7.3.2)

We then define a corresponding transfer operator

Rwf{x)=

^^W/W

(7.3.3)

yeX, o {y)=x

acting on L^ (X, 7Y), i.e., bounded measurable functions f'.X ^

H.

With these definitions, we now outline how the main results in Chapter 2 carry
over. We state them with only sketches of proof, as the modifications in Chapter 2
that are required are mainly technical. The main added issue is the non-commutativity, so when we write products like
W (^''" V ) ^ ( ^ " " M

-W {ay) W iy) ,

(7.3.4)

then we mean products of operators acting on H.

lA Operator-valued measures
Our next result. Lemma 7.4.1, is an operator-valued variant of the scalar case above,
in the form of Lemma 2.4.1. In the operator-valued context, we get again the existence and uniqueness of the measures P^.
Lemma 7.4.1. Let X, a, N, TQ, ..., TN-I, W, and W be as described above.
(a) Then, for every x e X, there is a unique positive operator-valued Radon probability measure / \ ow Q = {0, 1 , . . . , TV 1}^ such that
P, (A ( / i , . . . , /)) = W {zi.x) W {Ti,Xi,x) 'W
where A(i\,...,

(r, r,,x) ,

/) denotes a cylinder set as defined in (2.3.5).

(7.4.1)

116

7 Generalizations and applications

(b) Let the assumptions be as in (a), but consider instead the operator function
Wn(iu...,in;x)
= m (r, r/jx)* --m

{ri^x)* m (r/^x) m (T/ Zj^x),

with m satisfying (7.2.3). Then, for every x e X, there is a unique Radon measure Px on Q such that
Px (A O i , . . . , /)) = Wn ( / i , . . . , /; x).
Proof sketch. For each n e N, consider fiinctions f:Q
f (coi,..., con), and define
^i"^ [ / ] =

^ (^^1^) '"^{'con'"

H such that / (co) =

^co,x)f(cou

. . . , co) .

(7.4.2)

(a>i,...,)6Z^

As in Chapter 2, we check that P^'"'^^^ [f] = /"i"^ [ / ] Using this, the conclusion
in Lemma 7.4.1 may be obtained as in Chapter 2, mutatis mutandis.
n
We now list some of the other conclusions which carry over.
Theorem 7.4.2. Let the setting be as described in the lemma, and let (Px)xeX ^^ ^^^
corresponding operator-valued process. Then (i)-(v) below hold.
(i) The infinite product

(7.4.3)

Y{W{T^{X))=:F(X)
n=\

exists if and only ifPx ({0}) > 0, where 0 = ( 0, 0, 0, . . . ) e Q. The condition

cxD String of zeroes

is that the operator Px ({0}) has zero kernel.

(ii) The function
/2(x) = P^(No),

xGjr,

(7.4.4)

solves the eigenvalue problem

Rwh = h.
(iii) As before, Pxi)

(7.4.5)

extends ft-om No to Z, according to formula (1.4.5).

(iv) Returning to the multiwavelet case of (7.1.1)-(7.1.7), let a, ro, . . . , TAT-I be

given as in (5.4.9). Choose (p\, (p2, ... in FQ such that
^
Then

(i>i (X + K) (pj {x + k) = Sijxs,'

(7-4-6)

7.4 Operator-valued measures

h(x) = PAZ)=\

117

Xs,---

solves (7.4.5). In this case the matrix function x \-^ W (x) is as described in
[BaJMP04].
(v) In the setting of (iv), the multiplicity function fi of (7.1.7) is
/

keZ

and the projection Q{x) onto the range subspace of

JVo = {{f{-+ky)kez\f^VQ}

(7.4.8)

is
Qk,i (^) = X ^ ' ( ^ + 0 (pi {x + k).
Specifically, the range ofQ(x)

(7.4.9)

is the subspace

{if(x + k))k^z^f^\f^

Vo],

and ju (x) is the dimension of this subspace.

Our next result is motivated by the theory of wavelet packets of Coifinan et al.;
see [CoMW95], [CoWi93], [Wic93], and the book [Wic94] by Wickerhauser.
Lemma 7.4.3. Let X, o, N, TQ, ..., rjq-x, H, and {Wi)Q<i^^ be given satisfying
the conditions in Definitions 7.3.1, and (7.4.10).
We continue with the setupfi'omDefinitions 7.3.1, but with thefollowing addition:
instead of a single operator function
W:X^

Pos (W),

we consider a family
Wi\ X -> Pos (W),

/ = 0, 1 , . . . , TV - 1,

each one satisfying (7.3.2), i.e., we assume that

X

^i(y) = ^n,

/=0,1,...,7V-1, xeX.

(7.4.10)

yeX, G {y)=x

Then for every x e X, there is a unique positive operator-valued Radon probability measure Px o No x Q such that
Px {{{cO, O e No X Q I ft>i = 71, . . . , COn= in; f l =JU

. . , in=

jn})

= Wi, {xj,x) Wi, {TJ,TJ,X) . Wi (ry, .. Tj,x).

As usual, we are identifying No with a subset ofQ.

(7.4.11)

118

7 Generalizations and applications

Fig. 7.4. The first thirty-two functions in the sequence (pn

7.4 Operator-valued measures

l-h
^4

l-||nnnnn

nnnniPD

(P20 I I |N 1

11 ill i

j | |!i j 1

oil 11
INI Iti 1 Jll I'l
iriH 1 nl

_i JIUilLlI

11 III

-1

in the case of the Haar wavelet construction.

119

120

7 Generalizations and applications

(Pl

^1
\

(5

T)

" 1 ~1
il

wl/^!

ill

1
0

-1

ImlH

1-2

11

Fig. 7.5. The first thirty-two functions in the sequence (pn

7.4 Operator-valued measures

121

1 1^

1
0

Tii

IML

y|[E|fl

|| jlj

-1
-2

'
0

in the case of the Daubechies wavelet construction.

>

122

7 Generalizations and applications

Proof sketch. For each e N, consider fimctions / : No x Q -^ H such that

/(cy, (f) = / ( c y i , . . . , co; (fl,..., (f). Define

(coi,...,fy)

(^i,...,4)

"^con (74 Ki^) fim....,

a;; ( f i , . . . , 4 ) ,

(7.4.12)

and check that

p^(+l)[y]^p()[/].

(7.4.13)

To see this, note that by taking n suflBciently large, we may assume that co has the
form co = co{k) for some A: G No, i.e., that
co{k) = (coi, . . . , con, 0, 0, 0, . . . ) ,
00 string of zeroes

where
k = m+ C02N + .. + conN''-^.

(7.4.14)

Using (7.4.12), we then get

((o\,...,o)) (i\,...,i)
W<o {ri ri^x)

But note that 5^4 ^ ^0 (^4+1 ' ") = ^n^y

f(co\,--',COn;^U...,^n)'

(7.4.10), so the desired consistency re-

lation (7.4.13) follows. Now the last step, extending the consistent family ( Px )
to a Radon measure on No x Q, follows the reasoning used earlier in Lemmas 7.4.1
and 2.4.1 above.
n
Remark 7.4.4. The measures Px in Lemma 7.4.3 are typically not probability measures. See Proposition 7.5.2 for details.

7.5 Wavelet packets

Remark 7.5.1. In the remaining chapters, we illustrate our presentation of wavelets
and wavelet packets with graphics series, both in Chapters 7 and 9. We begin by
recalling the wavelet filters and the corresponding pyramid algorithm which is used
in the construction of a sequence of wavelet packet fimctions g)o,g)i,(p2,
Starting
with the Daubechies wavelet filter. Figure 7.5 illustrates the first 32 fimctions in the

7.5 Wavelet packets

123

Vo
V\
Vi
F3

W^

W2

Wx

Fig. 7.6. A decreasing family of resolution subspaces VQ o Vi o V2 D ',- and corresponding detail spaces JVi, W2,
Dyadic wavelets. Vn = V^-^i + ^+i, n:0,1,....

sequence. As is expected from the theory, one observes the time-frequency behavior
which makes the wavelet packets especially adaptable to a given data set, or a given
image (in the case of two variables).
The next seven figures serve to illustrate the multifaceted features of pyramid
algorithms, and to stress the distinction between the algorithmic paths of the standard
wavelet construction and the choices going into the selection of a function in the
library of a wavelet packet construction.
The standard dyadic case begins with formulas (1.3.18) and (1.3.19), while the
N-adic case uses (1.3.1) instead of (1.3.18), but then (1.3.19) is replaced by TV 1
identities corresponding to the A'^ 1 higher-frequency subbands, and the associated
subband-filter functions mi, mi, . . . , ^ A ^ - I , as illustrated in Figures 7.9 and 7.10.
The distinction between the dyadic case, i.e., N = 2, versus the case of more than
two subbands in the encoding of subspaces, is illustrated in Figure 7.10 in the special
caseofA^ = 4.
Although subband filters were first used in signal processing, see Figure 7.7, they
have now been adapted to the wavelet algorithm as outlined in Figures 7.8, 7.9, and
7.10. The coefficients in the two filter fimctions mo and m\ of (7.5.1) and (7.5.2)
are at the same time the masking coefficients for the first two wavelet functions, the
father function cp = cpQ and the mother function y/ = cp^. While for the wavelet
algorithm, the same dyadic choice is made in each scaling/subdivision step, in contrast the wavelet packet algorithm is encoded by a string of separate dyadic choices,
as is illustrated in Figures 7.11 and 7.12, and in the theory part of this section; see
especially formulas (7.5.5)(A)-(B), as well as Figure 7.5 (pp. 120-121). The Mathematica program code for the graphics in Figure 7.5 is included in the "References
and remarks" section at the end of this chapter.

124

7 Generalizations and applications

dual

Fig. 7.7. Subbandfilteringfor dyadic wavelets: WQ = low-passfilter,m\ = high-pass filter,

Q down-sampling, (J) up-sampling. Thefigurerefers to input (from left) and output (to the
right) of signals from standard signal processing; for example, of speech signals. From the left,
we start with (signal) input. The signal is then split into two frequency bands with two filters
ntQ and wj. Thefirstfilter,rriQ, passes low frequencies, and the second, mi, passes high. The
split into frequency bands is called "analysis." We get two signal strands which are now each
followed by signal down-sampling. As we move right, the two bands are then up-sampled.
The processed signal bands then pass dual filters, and arefinallymerged again with + into
an output signal. "Perfect reconstruction" means that the output signal to the right is matched
up perfectly with the input from the left. From engineering, we know that it is possible to
findfiltersrriQ and wi that achieve perfect reconstruction; i.e., recovering the input signal by
synthesis of the bands. Magically, and by hindsight, thesefiltersystems serve at the same time
to give us orthogonal wavelets, subject to a technical condition discussed in Chapter 5. (There
can be more than two frequency bands, and we refer to Section 7.6 for the general case.)

The tiling aspect of the fundamental subdivision/wavelet packet issue is perhaps

especially clear for Haar's construction as illustrated in the corresponding series of
pictures in Figure 7.4 (pp. 118-119). Different geometric and algorithmic formulas are further illustrated in other forms in the pictures and formulas collected in
Figure 7.6 (subspaces), Figure 7.7 (the engineering approach to subbands), Figures
7.8-7.9 (subbands and matrix operations), Figures 7.10-7.11 (a particular path in the
wavelet-packet tree), and Figure 7.13 which spells out the relation between the subspaces and the associated matrix operations. The same issues are illustrated with the
representations from Section 7.6 below. Figure 7.16 (p. 134) outlines the subdivision
algorithm which goes into the graphics, creating a step-by-step design of both the
father function cp and the mother function y/ for the Daubechies wavelet. The pair
of formulas behind Figure 7.16 are special cases of (1.3.1), (d) in Example 1.3.3,
(1.3.19) from Chapter 1, as well as (7.5.1)-(7.5.2), and (7.6.6}-(7.6J) in the present

7.5 Wavelet packets

Vo

rEih

125

Wi

cHS)
rElHS)

fV^

43-ir
Fig. 7.8. Wavelet algorithm in the dyadic case: Vn = V^-^-i + ^w+i, = 0, 1,

chapter. The general use of the pyramid is discussed and illustrated further in Figures
8.1 and 8.2 in the next chapter (pp. 159-161).
The corresponding algorithms and figures are created by Brian Treadway in
Mathematica. The series of pictures to follow have been created so as to offer visual
illustrations of some of the basic wavelet algorithms, both the father function/mother
function algorithms, and the related pyramid algorithms used in the creation of wavelet packets. Following the progressions in the picture series, the reader will be able
to follow visually the corresponding algorithmic steps, see for example the (A) and
(B) parts of equations (7.5.5) and (7.5.6).
Caution: Readers comparing the present figures and definitions with other related ones in the literature, for example in [BrJo02b], will notice some differences
in normalization conventions: For example, here we normalize so that the low-pass
property reads | w ( ) | ^ = 1 at frequency zero. This is to stress the probabilistic
meaning of the term | w ( ) p for low-pass. The meaning of "low-pass filter" from
signal processing: The signals with low frequency will move through the eye of the
low-pass filter with high probability, while they will be essentially blocked by the
high-pass filters.
Besides the normalization convention, there is another difference between (2.5.25)
in [BrJo02b] and our present equation (7.5.5)(B): The index on a (or a) is 5 A: in
[BrJo02b], while it is 1 - A: (actually 1 - j) here.
This shifl;s the support interval for the corresponding mother function by an integer. To keep Daubechies' scaling fimction (po, and the wavelet packet series cpi, (p2,
. . . confined to the interval [ 0, 3 ], we need the index to be 3 A:.
This does not affect the computation, but it changes the tick marks on the horizontal axes of the final figure plots. When comparing the present pictures with others,

126

7 Generalizations and applications

(1)

i4w7|--(jjv)-^
Vo

r-p7[(jAf)

fF-(1)

fr:(2)

pip](I^

r-p7[-(^ r

f^i

(1)

fp:(2)

Lp^](IA^

fF!(1)
V2

W(2)

po](jA^

F3

po](liv)-

^^4

Fig. 7.9. Wavelet algorithm in the A^-adic case: F_i = F + fF, W = fF^'^ + ffi^^ +
Multiple subbands w i , . . . , wjv-i + W ~ ',n = 1,2,

VQ

Vi
Vl

ri*

nf

^'*

<i)

<>

fFf>

Fig. 7.10. A decreasing family of resolution subspaces FQ D Fj D F2 D , 7V-adic case,

TV = 4, F_i = F + Wn, Wn = WJi^^ + wP-\- W^, = 1, 2 , . . . .

the reader should allow for translation on the x-axis by integer amounts for the function (p\ (= y/); note that there is an integer translation in the frame setup. And for
the higher wavelet packets, the offset appears to be by non-integral amounts such as
1/2, 1/4, 1/8, etc.

7.5 Wavelet packets

EiHS)H
FEIHSM
Fi[vv~n

127

rH""^

i^^Kfvi

Fig. 7.11. A selection of a dyadic wavelet packet. (An example of a path in a pyramid!)

Vo
Vi
Vi

F3

r^

^3

W2

Fig. 7.12. Illustration of a path in the pyramid: /WQ -^ m\ > m\ -^ mQ> m\ ^ , or

Fo -^ Fi -> ^2 -^ ^3 -^ ^3 -^ ^5-

Similarly, the normalization of m mentioned above will not affect the computation of ^ provided that a suitable factor appears in (7.5.5): in (7.5.5)(A), for example, ^0 is given as a convolution of itself with the row of a's, multiplied by the initial
factor, so the magnitude of that initial factor is determined by a sum rule.
As we have our "normalization" in (7.5.5), the Haar case, i.e.. Example 1.3.3(a),
is ^ 0 = ^ 1 = 1/2, and the other ai 's zero. As noted above, our "normalization" here
is a little different from [BrJo02b].
Here we have ^ai = \\ see (7.6.6).
The Qi numbers for the Daubechies wavelet are given in Example 1.3.3(d), and
they are ao, . . . , as for the non-zero ones, starting with ao = (1 + \/3)/8, etc.; see
(7.6.6). Remember X a/ = 1 from our choice (1.3.14).

128

7 Generalizations and applications

mo\^-@-*

Fi

Vx + Wx = Vo
m\

v~~*

Wi

Matrix multiplications:

o ax

a2

0 ao ai
0 0 0
ai 3 0 0
bo bx bl 63
0 0 bo ^1
0 0 0 0
bl Z>3 0 0
0
0

0
0
2 3
ao ax ai
0 0 0
0 0 0
bl 63 0
bo ^1 bl
0 0 bo

0
0
3
<^1

0
0
b3
bl

Fig. 7.13. Combinedfilter/down-samplingin matrix form.

The algorithm in (7.5.5) is simple. Here are the steps. The two first cpt fimctions
are familiar: cpo = (p (father function), cpx = w (mother fimction), and we have them
in Fig. 1.2(a), the Haar case (p. 13), and Fig. 7.16, the Daubechies case (d) (p. 134).
The general case: List the numbers ao, ai, . . . , up to the last non-zero number.
Then do the first step (A) in (7.5.5) to getfi*om(px to cpi.
The second step (B) in (7.5.5) amounts to running the numbers a/ in reverse, and
alternating the signs; see (7.6.7). The second step gets us from cpx to ^3.
Now back to the first step in (7.5.5), and we get from cpi to ^4. Then return back
to the second step; this time applied to cpi, and we get ^5.
And so on: All the while, walking zig-zag through the algorithm; the first step
takes us from cpk to cpik, and the second step from cpk to (pik+\ (see Figures 7.2 and
7.3, pp. 112-113).
After having found the first two, cpo and ^ 1 , in the usual way, one procedure for
doing all the steps in succession is simply making a string of zig-zag steps as follows:

7.5 Wavelet packets

129

(A): 1 -> 2; (B): 1 -^ 3; (A): 2 ^ 4 ; (B): 2 -> 5;

(A): 3 ^ 6 ; (B): 3 -> 7; (A): 4 ^ 8 ; (B): 4 ^ 9 ; etc.
This zig-zag procedure is done "wholesale" in the Mathematica program: instead of
1 > 2, 1 > 3, 2 ^ 4, 2 ^ 5, etc., it is expressed (equivalently) in terms of lists:
1 ^ {2, 3} ^ {{4, 5}, {6, 7}} ^ {{{8, 9}, {10, 11}}, {{12, 13}, {14, 15}}},
etc., with each arrow representing one instance of the "map" operation. The nested
sublists are "flattened" to a single list only in the final step.
In fact, the implementation starting with (po gives the whole seriesfi*om0 at each
stage because cpo regenerates itself in step (A):
0 ^ {0, 1} ^ {{0, 1}, {2, 3}} -> {{{0,1}, {2, 3}}, {{4, 5}, {6, 7}}} -^ etc.
We now turn to one of the uses of Lemma 7.4.3 in the theory of wavelet packets.
To make the ideas more transparent, we state the result only in the special case when
N = 2, and where the Z-translation-invariant subspace VQ is generated by a single
fiinction q)o = <p in L^ (W) which is known to satisfy (1.3.1), or equivalently (1.3.4).
Set
mo(x)=:^a;^^-'2^^^

(7.5.1)

mi (x) = ^ ( - l ) ^ a l - K - ' ' " ^ ^

(7.5.2)

and
Mi = \mi\^,

i=0,l.

(7.5.3)

The orthogonality conditions state that

the matrix

mo (x) mo (x + ^ j
/
( I
m\ (x) mi (x H- jj

is unitary for x e M.

(7.5.4)

The two recursive formulas (7.5.5)(A)-(B) form the basis for the so-called/Tyramid algorithm, i.e., the building basis functions in algorithmic steps following paths
in a combinatorial tree, or algorithm. This is illustrated graphically both with the
series of figures. Figures 7.1-7.3 (pp. 111-113) and 7.14-7.15 (pp. 132-133), and
especially the pair Figures 7.2 and 7.3 (pp. 112-113). It is fiuther illustrated with the
figures which follow in the rest of this chapter, as well as the next.
Proposition 7.5.2. Let the functions (po, (pi, (pi, - -- be defined by

(A)

(p2k (0 = 2 2 ] aj(pk (2t - j) ,

(7.5.5)

(B)

(p2k+i it) = 2 ^ i-iy

j^^

ai-jn

(2t - J),

130

7 Generalizations and applications

or equivalently by
I (A)

^2k i?c) = mo y-J h ( 2 ) '

I (B)

nk-^x (x) = m\ ^ - j (pk ( - j .

(7.5.6)

Suppose k = i\ -\-i22-\

+ /2"~^ Then

h {x) = rrii, ( I ) mi, ( ^ ) mi (|^) ^0 (|^)

(7.5.7)

satisfies
00

/.

X Z l < ^ ^ ( ^ - ^ * ) l / > l ^ = ll/ll^= / 1 / ( 0 1 ' ^ ^

/eZ2(M).

(7.5.8)

Let (^;c)jce[0,i ] ^^ ^^^ measure on No x Qfrom Lemma 7A3. Then

P, ({CO (k) , a> (/)}) = \n (X + / ) | ^

(7.5.9)

where the correspondence co(k) <r^ k, co(l) <r^ I is determined as in (7.4.14), i.e.,
co(k) = (coi,... ,con, 0), and similarly for I e NQ. Moreover, {cpk ( /)}A;ENO, /GZ
is an orthonormal basis for L^ (M) if and only if
Px ({co{k)} X Z) = 1

a.e. x e [0, 1 ] and alike

NQ.

(7.5.10)

Proof. The detailed steps are quite analogous to those given in Chapter 2, the main
difference being that now Lemma 7.4.3 is used in place of Lemma 2.4.1 in the standard wavelet case.
n
Remark 7.5.3. Note that if A: = 0 in (7.5.5)(A), then we recover the scaling identity (1.3.1) in the special case N = 2. Depending on the conditions placed on the
coeflicients (aj), we get solutions q) = (poin various function spaces, or in spaces of
distributions. If (7.5.5)(A), or (1.3.1), is known to have a solution in I ^ ^ ) ? then an
integration on M yields X/eZ ^j ^^ ^^^ ^ recover the familiar low-pass property
for the wavelet filter
jeZ

in the form ofmo(0) = 1. This is an interesting (perhaps unexpected) link between

wavelets and signal processing.
The appearance of the graphs of the progression of functions cpo, (p\, ^2?
in Figure 7.5 (pp. 120-121) offers a visually convincing argument for the name
"wavelet packet" for this cascade construction. As we progress in the series of
functions fi-om the union of the two figures, we see that each of the fimctions (pn,
for n sufiiciently large, is made up of two separate and distinct visual parts: one is a

7.6 Representations of the Cuntz algebra O2

131

high-frequency wave, which appears packed or shaped inside a shadow envelope of

another form with low frequency (i.e., long wavelength).
The term "wave packet" appears to have been used first by Werner Heisenberg
in connection with particle-wave duality in quantum theory in the 1930s. The visual
situation in Heisenberg's theory was much the same as the one which comes out
from the much later wavelet constructions. In any case, Heisenberg's aim for his
construction in the early days of quantum theory was quite different from what came
later with wavelets, and Heisenberg most likely did not know about Haar's wavelet
from 1910, see [HaalO]. The name "wavelet packet" came much later than both
Haar and Heisenberg. But it was probably chosen by analogy to Heisenberg's wave
packets by the authors of [CoMW95] and [CoWi93].
Remark 7.5.4. hi (A)-(B) in Proposition 7.5.2 and their analogues (1.3.18)-{ 1.3.19),
we considered the crucial recursive identities only in the scalar case. But in fact there
are matrix versions. The main modification in the matrix case is that the coefficients
are then matrices, and the functions involved are vector-valued. However, both in the
scalar case, and in the matrix case, there is a pair of masking coefficients ak and bk.
The above-mentioned quadrature conditions (see, e.g.. Figure 7.7, p. 124) dictate a
certain relationship between the two: Both in the scalar and the matrix versions of
the pair ak, bk, the relationship is
^^ = ( - l ) ^ ^ ^ d d ^ .

(7.5.11)

Depending on the plan of the picture and the number of a^; 's, there are different usefiil
choices for the odd number in (7.5.11).
The pair of frequency response functions which correspond to the assignment
(7.5.11) made above will then be as follows:
[mo(z) =

^akz^,

7.6 Representations of the Cuntz algebra O2

The Cuntz algebra ON, see [Cun77], is a simple C*-algebra on the relations
N-l

S*Sj=SijI,

^ ^ , ^ ; = /.

(7.6.1)

When ON is realized by operators on a Hilbert space H, we take / to be the identity

operator on H, i.e., I:h \-^ h,h e H. These operators are fundamental in signal
processing, where they take the form of signal/wavelet diagrams as illustrated in the
series of Figures 7.7-7.15 (pp. 124-128,132-133), and especially in Figures 7.7 and
7.14 from signal processing.

132

7 Generalizations and applications

^0
'

ANALYSIS

'

'

SYNTHESIS

Q
low-pass

'

filter/

\dual low-pass filter

SIGNAL OUT

SIGNAL IN

down-sampling

up-sampling
'dual high-pass filter

high-pass filter

<Sy

<sy
J

L
Si

S*i

Fig. 7.14. Perfect reconstruction of signals.

The choice of the numbers (aj) . ^ used in mo, m\ from (7.5.1)-(7.5.4) guarantees that the corresponding two operators
{Si / ) (z) = V^ m, (z) / (z2) ,

/ = 0, 1, / i 2 (T),

(7.6.2)

satisfy (7.6.1) for W = 2 (T) and iV = 2.

If we use l-periodic functions on E instead of functions on T, then (7.6.2) takes
the equivalent form
iSif) ix) = -Jimi (x) / ( 2 x ) ,

/ = 0, 1, / i 2 (0,1).

(7.6.3)

The two functions WQ and m\ are called the low-pass/high-pass filters respectively,
and we have m^ (0) = 1, m\ (1/2) = 1, when the m/'s are viewed as l-periodic
functions.
Lemma 7.6.1. Let thefimctions mo, m\ satisfy the unitarity condition (7.5.4) and
let Si, i = 0, 1, be the corresponding operators. Then the operator relations (7.6.1)
ofCuntz are satisfied.
Proof. We refer to [BrJo02b] for the detailed verification that (7.6.1) is satisfied
when the unitarity property (7.5.4) is assumed. The essential step is the following
identity:
1

Zll^r/f
=
I
II2

fBL'(0,l).

(7.6.4)

7.6 Representations of the Cuntz algebra ^2

1^3

Fig. 7.15. Iteration of thefilter/down-samplingoperation as it arises in wavelet algorithms.

To verify (7.6.4), first check that

w/)=^|:".m/m.
And so
1

Zll^;/lr

/.I

(first the /-summation)

which is the desired identity (7.6.4).

/ = 0,1.

(7.6.5)

134

7 Generalizations and applications

(a): Father function

The father function (p is also called the scaling function and it is the solution to the scaling
identity (d) in Example 1.3.3. It is also the first of the sequence (po,(pi,(p2,..- of functions
generated by the algorithm (7.5.5), i.e., (p = g)Q.

(b): Mother fiinction

The mother function y/ is also called the wavelet function, and it is the second term in the
sequence of functions generated by the algorithm (7.5.5), i.e., ^ = ^ i . So this four-tap
Daubechies y/ is the same function as ^i in Figure 7.5 (p. 120).
Fig. 7.16. Daubechies wavelet functions and series of cascade approximants.

7.6 Representations oftheCuntz algebra ^2

135

We note, see [Jor03], that (7.6.4) is known in signal processing as the quadraturemirror identity for the filter operators defined by the two fiinctions m/, / = 0, 1, or
equivalently by the sequence (aj) . ^.
The case of four coeflScients ao, a\, ai, ^3 is called "four-tap", and it includes
(with our current normalization)
l+x/3
ao = - ,

3 + V3
a\ = - ,

3-V3
a2 = - ,

^
1 - V 3 ^^^^^
and as = - , (7.6.6)

the Daubechies wavelet; see [Dau92, p. 235]. If we take

1-V3
m (z) =

3 - V 3 . 3 + V3 2
z + -z^

1+V3 3
z^

(7.6.7)

for z = e~^^^^, then the two fimctions cp (father function) and y/ (mother function)
will be supported in the interval [0, 3 ], they are diflferentiable (see [Dau92]), and
they graph out as in Figure 7.16.
We now consider the probability space
n = {0,l}^,
and the finite spaces
Q(w) = {0, 1}^^'^' ''^^
Specifically, Q (n) consists of all fimctions co: { 1 , 2 , . . . , } -^ {0, 1}.
Lemma 7.6.2. For /? e N and co e Q (ri), set
So)-=Sco{l)"'Sco(n),

(7.6.8)

and
Eco := SayS^ = Sa)(i) *^cw(w)*^i()'' * S^^^y

(7.6.9)

Then {Eo)}a)eQ.(n) ^^ ^ commuting family of orthogonal projections satisfying the following matrix-unit identities:
Ea^Eco' = Say^co'Eco,

CO.oJ G Q () ,

(7.6.10)

and

X ^'o = lProof This is a direct verification; see also [BrJo02b, Jor05].

(7-6.11)
n

Theorem 7.6.3. There is a unique projection-valued measure P defined on the Borel

subsets of[0,1 ] such that

pd'^^^K

,^(")^w.

.'"W.M^ F
forco&Q. (n) andneN.

(7.6.12)

136

7 Generalizations and applications

Proof. The result follows from Lemmas 2.5.1 (Kolmogorov's construction), 2.6.2,
and 7.6.2 above; see also [Jor05, Jor04b].
n
Remark 7.6.4. In view of (7.6.11), the measure P from Theorem 7.6.3 satisfies

P(dx) = I,

(7.6.13)

where / is the identity operator in the Hilbert space H which carries the representation of the Cimtz algebra O2. Specifically, if / G W, and ||/|| = 1, then the measure
f^f(')
satisfies

In the case

= {f\P(')f)

= \\P(-)ff

(7.6.14)

djuf(x)

= fif([0,l])

= l.

(7.6.15)

n = L^ (T) ^ L^ (0, 1) ^ ^ (Z),

the case f = t (the constant fimction 1) was used in [CoMW95] and [CoWi93] in
the construction of libraries of orthonormal wavelet packets, and it was conjectured
that the measure
Mi(-)^\\Pi-)if
(7.6.16)
is absolutely continuous with respect to Lebesgue measure on [ 0, 1 ] for all the wavelet filter systems of (7.5.1)-(7.5.4).
We do not solve the conjecture here, but we illustrate the question with the use
of the random-walk model from Chapter 2.
We first take a closer look at the measure (7.6.16). Since e (z) = z " , a n d e o = 1?
we shall also use the notation JLLQ for the measure ju ^.
Lemma 7.6.5. For the case n = \ in Theorem 7.63, and the measure JUQ of Remark
1.6.A, we have

(a)M([o,^))=2;^|a2,|^
(b)M([^,l))=2XK+i|',
(c) //o ({x}) = 0 for all X e [0, 1 ].
Remark 7.6.6. When (a)-(b) are applied to the Daubechies wavelet filters (7.6.6){1.6.1\ we get

In the case of the dyadic Haar wavelet, it is easy to check that

7.6 Representations oftheCuntz algebra ^2

ncoil)

coin) CO (I)

^"'W^IXA

137

for CO eQ.(n) , n e N.
So for the dyadic Haar wavelet, the measure yuo is the Lebesgue measure restricted
to the unit interval [ 0 , 1 ] .
Proof of Lemma 7.6.5. Part (a):

VL

2 / 7 by(7.6.12) " "

" by(7.6.5)

by Parseval

Jo \2 ^

V 2

2y|a2yf
^
'
JE.Z

"^ '

Part (b):

"^ \\_2

//by(7.6.12) " 1

"

by(7.6.5)

by Parseval and (7.5.2)

Part (c): Here we refer the reader to [BrJo02b, Theorem 2.2.1, p. 92; Lemma
2.2.3, p. 95].
n
Theorem 7.6.7. Let (aj) . g. '"0. and m\ be as described above. Let N, and
G) e Q (). Then

. l ^ + - + ^ , ^ + - + ^2.w

{[-

-2"X

V"

+ 1))

/_i\cy(2)6+-+<w()4

' ^cw(l)-A:2+2<^2+-+2"-i4 * ^(2)-<^2 * * * ^co(n)-^

Proof For G N, co G Q (), set

m^"^ (x) = meo(i) (x) ma,{2) (2x) may(n) (2"-^x) .

(7.6.17)

Then

(5,/)(x) = /")(x)/(2"x)
and

(7.6.18)

138

7 Generalizations and applications

{Slf){x)^2-"'^

Y,

m^:^(y)f(y)

(7.6.19)

2^'y=x mod 1

by the argument from (7.6.5). Note that the summation in (7.6.19) is over the set of
all dyadic fractions
X -\-r

y^

r = 0 , l , 2 , . . . , 2 - i

(7.6.20)

where each r in (7.6.20) has a unique representation

(7.6.21)

M and (7.6.18), we get

Using now Lemma 7.6.5 on wa>
coin) coil)
2" ' 2

([^

coin)
2"

h))

= \SM
iy)
"^ I

dx

2-y=;( mod 1
(a>(2))

2Z

" A : . 2 " - ; 2^2

2^^-14 ^.
* "<^2

(coin))
^n

keZ ^2,...,4eZ

2^T

y^

(_l\G>(2)<^2+-+cy()4

* ^o)(l)-A:2+2^2+-+2"-i4 * ^>(2)-<^2 ' * * ^G>(W)-'^

which is the desired result.

7.7 Representations of the algebra of the canonical

anticommutation relations (CARs)
The purpose of the present section is twofold: (1) to illustrate the non-abelian framework for infinite-product constructions; and (2) to point out how a class of measures that are known as determinantal measures (see [LyStOS]) arises from a construction from operator algebra theory^Powers-St0rmer's construction of the quasifree states on the C*-algebra of the canonical anticommutation relations, the CARalgebra; see [PoSt70].
The construction for both the abelian and the non-abelian setup begins with a
fixed operator ^ in a Hilbert space, such that A is assumed selfadjoint with spectrum

7.7 Representations of the CAR algebra

139

in the closed unit interval [ 0, 1 ]. In the abelian setting, the measure JUA^S defined on
cylinder sets (i\,..., in) in Q, as the determinant of the submatrix obtained fi-om A
by using the subset of an orthonormal basis corresponding to the selection z'l,..., /.
As it turns out, ju A is the restriction of a state pA on the CAR-algebraa quasi-fi*ee
state, in the terminology of [PoStTO]^where pA is defined as the Hilbert norm of
fermion particles. This is a multistate computed fi-om / ' i , . . . , in, and computed in a
multiple-particle Hilbert space which is constructed directly fi*om A.
In both cases, the abelian and the non-abelian, the extensionfi*omthe cylinders
to the fiill measure space, or the fiill C*-algebra, is performed with a variant of the
Kolmogorov extension principle from Section 2.5 above.
We begin with some terminology and definitions.
Definitions 7.7.1. Let M2 denote the algebra of all 2 x 2 complex matrices, and set
2t = M2 (8) (8) M2 = M2.
^
'

(7.7.1)

n times

0\
j defines a natural embedding of

/ 2li C 2I2 C C 2l C 2l+i C

(7.7.2)

as a non-commutative analogue of (2.4.5) above.

Let /, j e {0,1}, let ejj be the 2 x 2 matrix etj (k, I) = Sf^k^jj, and set
^^l.,i;n ,-Jn = ^'1 '^ ^hJ2 >> eij.

(7.7.3)

For CO, ^ G Q (n) set

E^J!^ = ^coS*^^

(1.1 A)

where Sco is defined fi-om a representation of02 as in (7.6.8). As noted in [BrJo02b],

the assignment
identifies a subalgebra (up to C*-algebraic isomorphism) of O2 with the infinite
tensor product
00

21 = (9) M2 = ind lim 2l.

(7.7.6)

The C*-algebra 21 in (1.1,6) is the inductive limit of the system (7.7.2) of matrix algebras, and 21 is also called the C*-algebra of the canonical anticommutation relations
(CARs) for reasons which are spelled out in [PoSt70].
The purpose of this section is to demonstrate that a family of measures which
were studied independently in [PoSt70] and in [LyStOS] for very different reasons

140

7 Generalizations and applications

may be viewed under the heading of this book, i.e., measures defined from functions
Q -^ (matrices), or inductive limits of systems Q. (n) -^ (matrices), for n eN.
Let n = f^ (N) be the familiar Hilbert space of ^^-sequences, and let (^y) ,^|^
H
be the standard ONB in H, i.e., defined by Sj (k) := Sj^k. j \ k e N.LQI A:n-^
be a linear operator such that
0<{h\Ah)<

||/^||^

hen.

(7.7.7)

The matrix representation for A will be defined relative to the ONB (gy), i.e., by
A(iJ):={8i\A8j),

/,7eN.

(7.7.8)

Let Q = {0, 1}^, and define the matrix function PF on Q by

W (CO) (/, j) := (co (/) Si J + (-l)-(^> {Si J - A (/, y)))

(7.7.9)

For the C*-algebra 21 we shall need the following representation from [PoSt70]:
there is an antilinear mapping a:H -^ H such that
a(hra{k)

+ a(k)a(hr

a(h)a(k)-\-a(k)a(h)

{h\k)I,

=0

forallh^keH.

A state /? on 51 is a linear functional /?: 21 -> C such that />(/) = !, and p {T* T) >0
for all r e 2t.
Let ^ be a fixed operator as specified in (7.7.7). It is known [PoSt70] that there
is then a unique state p = pAonOi such that
PA {a {siy "-a {si,y a (sj,) a (sj^))
= Sn,mdQtiA(ir,js))\<r,s<n'

(7.7.11)

When the CARs (7.7.10) are used, we get the formulas (7.7.3) which are used in
the embeddings (7.7.2). As a result, we get the following formulas for the matrix
elements in the algebras in (7.7.2), with the superscript referring to the n'th tensor
slot (see [PoSt70]):
(7.7.12)
I ^ 3 = a {snY Vn-x,

e g = a {snY a {sn),

where Vo = I and
n

Vn:=Yl{l-2a(skra(sk)).
Using the diagonal matrices from (7.7.12), it is now clear that C (Q) is naturally
embedded in % and therefore in O2 as well; see (7.7.3)-(7.7.5).

Exercises

141

Theorem 7.7.2. Let A be a given operator satisfying {1.1.1), and let W = WA be

the corresponding matrix function (7.7.9) defined on Q. Ifn e N, and co e Q (w),
then W {co) is ann x n complex matrix, and we set
P^''\co) = AQXW {co) .

(7.7.13)

Then the family {P^^^)^fq is Kolmogorov consistent, see Lemma 2.5.1, and therefore extends to a Borel measure PA on Q. This measure is uniquely determined by
(7.7.13), and is the restriction to C (Q) of the C*-state pA of {1.1.11). Moreover the
state is given uniquely by the two formulas
PA {a {si,) a {si.y . . . a [st^) a {st^Y) = det (/ -

PEOAPE,)

(7.7.14)

and
PA {a {sj,y a {sj,)...

a {sj^

a {sj,))
= dtt[ln-PEAI-A)PE,)-

(7.7.15)

Proof We know that the state pA on 21 is well defined, and determined uniquely by
(7.7.11). Let C (Q) be embedded in 21 as described, and let pA |c(0) be the restriction.
Clearly the restriction is positive, and so determines a Borel measure on Q by Riesz's
theorem.
We claim that this measure is the same as the one coming fi*om the family
(^^''^)EN i^ C7.7.13); see also (7.7.9).
Let ^ e N, ando) e Q {n). Set E^ = co'^ (0) and Ei = CQ-^ (1). Then EQUEI =
An application of (7.7.11) and (7.7.10) yields the following: Let
Eo = { / i , . . . , />},

Ei=

{Ju

Js),

and let PEQ, resp. PEI, denote the corresponding orthogonal projections. Then the
two formulas (7.7.14)-(7.7.15) hold. Moreover, an inspection shows that (7.7.14)(7.7.15) determine the state uniquely. In view of the identification of Q as the matrix
diagonals in (7.7.12), the two formulas (7.7.14) and (7.7.15), taken together, prove
that the restriction ofpA satisfies the desired conclusion (7.7.13).
n

Exercises
Organization. This chapter is longer than the preceding one, and it has more exercises. They are organized as follows: The first seven exercises serve to link the

142

7 Generalizations and applications

operator theory from inside the text with matrix algorithms. As already stressed in
Chapter 1, wavelet bases exist abstractly in an ambient Hilbert space H (which could
be Z^ ( M ^ ) ) , but for computations, Hilbert spaces as such are not very useful: we
need to make a connection from these fimction spaces to some concrete sequence
spaces, i.e., ^^-spaces V, and to efficient matrix representations of the objects and
transformations in those spaces. The basic idea of a multiresolution wavelet basis is
nothing more nor less than an efficient way to make that connection.
A wavelet transform is an operator which relates the sequence space V to a
subspace Vo of the ambient Hilbert space H. With such a choice the subdivision
operators of wavelet geometry are turned into slanted matrices, which we saw are
(computationally) fast.
Exercises 7.4-7.7 are a guided tour of the matrix computations connecting the
discrete wavelet transform with "real" wavelet computations; see the material in
Section 7.6 and Chapters 5 and 8. Li a simple case we see that the distribution of
expansion coefficients for a given wavelet packet is determined by the measures
from (7.6.12) and Theorem 7.6.7. Recall that our starting point is a discrete wavelet
transform (7.6.1). As outlined in Lemma 7.6.1, it takes the form of a system of operators Si in the Hilbert space I ^ (T), or equivalently in ^^, while in contrast a wavelet
basis refers to functions in Z^ (E). Li computations, not all Hilbert spaces are equal.
The second block of exercises deals primarily with tensor products of Hilbert
space, and the associated geometry for operators and for matrices. One of these. Exercise 7.8, about image processing, serves to link two ideas, slanted matrices and
tensor products: For images, i.e., 2D, or other problems in higher dimensions, a popular approach is to build the models simply by taking tensor products of well chosen
ID wavelet algorithms.
The last one. Exercise 7.12, is about representations of the C*-algebra ON for
N fixed. It stresses that in the present context the representations of ON are more
important than ON itself.
Notational conventions for slanted matrices. The slanted matrices used in this
chapter, and later, are infinite-by-infinite, in fact doubly infinite, i.e., with rows and
columns infinity in both positive and negative directions. Specifically, both rows and
columns are indexed by the integers Z. That is, the counting of rows and columns
starts in the negative segment of Z, going up to 0 in the middle, and then continuing
the coimting with 1 , 2 , . . . , with three dots indicating recursion. Rows are counted
from top to bottom, and columns from left to right: So, in each case, the row index
increases as we move from top to bottom in a matrix display; and the column index
increases from left to right, consistent with the counting of columns. The matrices
have uses in computation, where their slanted feature is significant. The slanting of
these matrices reflects the subdivision (from scaling) built into the algorithm.
In computations, we rely on matrix multiplication but choose our index convention to be consistent with the associated operator formalism, i.e., with multiplication

Exercises

143

of the corresponding operators in Hilbert space. An operator and an orthonormal basis (ONB) will always be specified; typically the ONB is chosen to be the standard
canonical basis in ^^ (Z). But we will have occasion to use more general bases.
Our index convention means that the row/column position (0, 0), i.e., zero row
and zero column, is in the center position of each of the slanted matrix displays. We
have used shading to illustrate this point: The matrix entries in the center cross, with
crossing at (0, 0) is highlighted, meaning that the entries in both row 0 and column
0 are highlighted with a light shading. This is to help the reader see how matrix
multiplication is implemented; and it fiirther serves to call attention to the slant in
each of the slanted matrices, and to visualize how slanting changes as matrices are
multiplied, or as the scaling number changesfi*omone matrix to another.
7.1. Let W be a complex Hilbert space with inner product ( | ), and let A be an
index set. Two systems of vectors in W, {en)neA ^^^ (^n)neA^ ^^ said to form a dual
basis system, also called a bi-orthogonal basis in H, if the following three conditions
are satisfied:
(i) {en\em) = Sn,m foralln,m e A,
(ii) {en\f)=0\/neA=^f
= OmK
(iii) {f\en)=0\/neA=f>f
= Omn.
(a) Prove the following assertion: If ((en), (em)) is a dual basis system, then
every vector f eH has two representations,

/ = neA
X<^|/>^ = Z<^|^>^neA
(b) Make precise the notion of convergence used in (a), and supply a detailed
epsilon-delta argument.
(c) Prove that if (en)neA is a system of vectors in H, then (en)nGA is an orthonormal basis (ONB) in W if and only if the pair {(en)n^A -> (^n)neA) is a dual basis
system, i.e., if and only if we get a dual basis system by taking en = en for a\\n e A.
Definition. Let (en)neA ^^ ^ system of vectors in H. If there are two constants ci,
C2, 0 < ci < C2 < oo, such that
ci Wff

< J^\{en

I / > P < C2 Wff

f o r a l l / G H,

neA

then we say that (en)neA is ^ frame basis (or just 2Lframe)for the Hilbert space, and
c\ and C2 are CBXIQAframebounds.
(d) Suppose (en)neA is afi*ame.Then show that there is a system (en)neA such
that {(en)neA ? i^rdneA) is a dual basis system.
(e) If (en)neA is afi*amefor H, show that the operator T:H ^> ^ (A) defined
by Tf= ({en\f ))neA foT f e H is bounded.

144

7 Generalizations and applications

(f) Show that the adjoint operator T*\^ {A) -^ His given by the formula

neA

(g) Show that T*T:H^n\s

invertible, and that

T'Tf = Y,^en\f)en

forall/eH.

neA

(h) Setting e := {T* T)~^ en for n e A, verify that the two reproduction formulas
neA

neA

hold for a l l / e W .
7.2. Let H and /C be two complex Hilbert spaces, and let T: 7i^ -^ /C be a bounded
linear operator. Let (en)neA ^^^ (fm)meB ^^ ONBs for the respective Hilbert spaces
H and /C. Hence we have the two representations
h = ^^^XnCn

withXn = {Cn \ h) forh

EH,

neA

and
Th=^ymfm

withym =

{fm\Th).

meB

A matrix {T (n, k))^^^, keA is said to represent the operator T in the two bases if
yn = ^

T (n, k) Xk

forneB.

keA

Show that when the two ONBs are given, then the matrix representation is
unique, and
T(n,k) =
{fn\Tek).
7.3. Generalize the result in Exercise 7.2 for T:H -^ IC where {(en) , (en)) is a dual
basis system in H, and ((fm) , (fm)) is a dual basis system in /C. Specifically, show
that then the formula T (n,k) := {fn \ Tek) provides a matrix representation for T
relative to the two basis systems.
7.4. Consider the Hilbert space H = L^ (T) (= f^ (Z)), and let ek (z) = z^, k e Z,
be the usual Fourier basis.
Show that (ek)keZ is an ONB for H.
7.5. Let m e L^ (T) and consider the operator
(Sf)(z)

= V2m(z)f{z^y

fen,zeT,

Exercises

145

and its adjoint operator

1
{S*f) (z) = ^{rn iVz) fWz)^m

[-^z) f ( - ^ ) ) ,

/ e H, z E T.

Let m (z) = X^eZ ^k^k (^) be the Fourier series for m.

(a) Now show that the matrix representations for the two operators S and S*
relative to (ek)keZ ^^^ ^s follows:
S (n, k) = V2an-2k

and S* (n, k) = sflak-in

for n,k eX.

(b) Show that the two matrices in (a) have the following "slanted form":
k=0

S = V2

a_2 - 4 - 6

a,

a_l a_3 a_5

04

a_2 a_4

a,

a_i

a_3

^2

a_2

n=0

(7E.1)

and
k=o
/
^2

^3 ^4 ^5

^^6

^0

^1 ^2 ^3 ^4

F:=S* =V2

n=0

a_4 a_3 a_2 ^ - i | ^ o

a_6 a_5 5^4 a_3 a_2

/ .

(7E.2)

146

7 Generalizations and applications

7.6. In the case of only two non-zero terms, say QQ and a\, the matrix F from (7E.2)
takes the form
k=0

/l* * o o o o o o o o o o o o o o o o \
0
0
0
0
0
0
0
0

n=0

0
0
0
0
0
0
0
0

*
0
0
0
0
0
0
0

*
0
0
0
0
0
0
0

0
*
0
0
0
0
0
0

0
*
0
0
0
0
0
0

0
0
*
0
0
0
0
0

0
0
*
0
0
0
0
0

0
0
0
*
0
0
0
0

0 0 0 0
0 0 0 0
0 0 0 0
0 0 0
0 * * 0
0 0 0 *
0 0 0 0
0 0 0 0

0
0
0
0
0
*
0
0

0
0
0
0
0
0
*
0

0
0
0
0
0
0
*
0

0
0
0
0
0
0
0
*

0
0
0
0
0
0
0
*

(7E.3)

i;

and hence the term "slanted Toeplitz" matrix.

(a) Identifying the oc x oo matrix with the corresponding transformation in ^ (Z)
relative to the standard basis in ^ (Z), find the structure of the invariant subspaces
for F. Make an infinite list of finite-dimensional subspaces which are invariant under
F.
(b) Answer the same questions as those listed in part (a), but now in the case
of only four non-zero terms (the four-tap case); say ao, a\, a2, and ^3 non-zero, but
a ; t = O i f A : e Z \ {0,1,2,3}.
(c) In the four-tap case from part (b), find some finite-dimensional subspaces in
i^ (Z) which are F-invariant awJ minimally F-invariant.
7.7. Suppose the matrix F has the form from (7E.3) in Exercise 7.6 corresponding
to the two-tap case. Then show that F^ has the following shape:

/ 0

^
I * * * * ^J
I****
C\
^r i
\J
f\
I* *
11

1....
0
'
I* *

Hint and discussion: From the theory inside the chapter, the reader will notice
that there are two numbers which determine the shape of the pair of matrices S and
F which occur in Exercises 7.5 to 7.7: First there is the number of taps T (two-tap,
four-tap, six-tap, etc.), and secondly there is the scaling number. The scaling number
is denoted N, and for the dyadic wavelets, N = 2. Each operator, alias matrix, S
and F has two lives: first it is represented as an operator in a fimction space, and

Exercises

147

secondly we have outlined how it is represented as a matrix, albeit an infinite-by-infinite matrix. IfS is represented as an operator acting on fimctions (see Exercise 7.5,
N = 2), it is composition with z^ followed by multipUcation with a fimction of z.
In each incarnation, we may talk about an adjoint F = S*:ln the operator world,
the adjoint refers to the iimer product of the Hilbert space of fimctions on the circle
T, and in the matrix picture, the adjoint is the usual matrix adjoint, i.e., there the
adjoint matrix arises simply as transposed and complex conjugate. We use the same
notation in either case, namely F = S*, with the * denoting adjoint. The amount
of slanting (i.e., shifting of the terms fi-om one column in iS* to the next) is equal to
the number A^. But the matrix transposed explains why the direction of the slanting
changes fi-om formula (7E.1) for S to (7E.2) for F in Exercise 7.5. The direction
of the slanting is significant for computation, as it turns out. As already noted, the
number of taps T is even, 2,4,6, etc., and it is simply the length of a minimal band in
a column ofS, or equivalently a row in F. All the terms outside a band are zero, and
minimal means "the shortest" such band. In the present discussion, we will arrange
the bands so the index runs from 0 to T 1, so four-tap corresponds to a minimal
band {ao, ai, ^2, as}. Using the fimction representation, it is easy to see that if/S' had
tap number T and scale number N, then S^ has tap number T + N(T 1), and scale
number N'^. For the present discussion (Exercise 7.7), N = T = 2. Hence, in this
case, both the tap number and the scale number for S'^ are 4.
Images. The next multipart exercise is about images. It involves two features:
First, it is a recursive matrix algorithm, and secondly it is a tensor-product construction. The environment of images is 2D, and the tensor product of two separate ID algorithms thus serves to create a new 2D algorithm. In fact this simple
tensor-product construction is the easiest way of producing the type of 2D wavelet
algorithms needed for images.
7.8. (a) Search the internet for your favorite visual sequence created with a digital
camera, and illustrating image processing for a digital portrait-photo; for example of
Lena [WWW5]. Look closely at subdivisions of pictures, and compare the cascading
successions. Identify how the pictures changefi*omone subdivision step to the next
in the cascade.
(b) Having understood the pictures on the web, then compare with the algorithmic resolution steps in Figures 7.9 and 7.10 in this chapter (pp. 126-126).
A portrait (such as Lena's) consists of numbers arranged into matrix forms: only
one matrix is needed for recording of grayscales, but several matrices are used for
digitizing color images. An exposure with a digital camera imprints numbers for
each pixel. Pixels form checkerboard arrangements which in turn correspond to a
chosen resolution. However a digital camera creates both scales of resolution and
intermediate differences', and everything is digitized.
So a picture L which is chosen and fixed at the outset will be represented by a
vector in the subspace FQ fi-om Figure 7.10. Since subdivision in the planar picture

148

7 Generalizations and applications

is made with horizontal and vertical lines, the first algorithmic analysis step applied
to L (in vector form) will split it up into four pieces, i.e., N = 4.
(c) In the image of Lena (Z) on the Web, identify the corresponding four subpictures. They are usually arranged inside a square, with the NW scaled "quarter-box"
showing the same image Z but in a slightly more blurred resolution. Identify how
the blurring corresponds to "local averages." The other three scaled pictures represent differences in grayscale (for black/white) in three directions, horizontal, vertical,
and diagonal.
In the Hilbert-space representation, each of the four subsquares of a picture is a
component in the corresponding four subspaces Vi (the coarser resolution), and the
three intermediate spaces Wl^\ wf\ and r f ^ fi-om Figure 7.10.
(d) Work out that the three ^-subspaces are registers for intermediate differences
in grayscales.
(e) Finally, fi-om the pictures on the web, identify how this is the start of a cascading process. Show how the process continues on the part of Z in the space V\. As
you follow the pictures on the web, note that the process is repeated, and at each step
the part of Z in the NW subsquare is fiirther subdivided. Check that this algorithm
corresponds to the pyramid which is illustrated with Figure 7.9, for N = A.
Note: The following citations throw light on the use of wavelet algorithms/filters
in 2D image processing: [SongOS] and [StHS+99]. Dr. Myung-Sin Song has created a
sample of images which are decomposed this way using the wavelet system of Haar,
plus the analogous image decompositions using Coiflets and Symlets. The processing
is posted at [Song05a] and the images themselves at [Song05b].
Tensor products. The next three exercises, 7.9, 7.10, and 7.11, highlight tensor
products of Hilbert space, both finite and infinite tensor products. Each is composed
in a multipart format, so that it is easier for the student to work through the concepts
step by step. The exercises reinforce fimdamental concepts already used in Chapter 7 above; and at the same time, they motivate and introduce some essential facts
(such as Schmidt's decomposition in its infinite-dimensional variant) for tensor products that we will need in the remaining chapters. The tensor-product idea for Hilbert
space is central in our present basis constructions and in our analysis in general in
that it lets us use Hilbert-space geometry to reduce or to factor "complicated" data
into its simpler forms or factors. The reader is encouraged to compare Schmidt's decomposition (Exercise 7.11 (e)) with the closely related formulas from Exercises 2.7
and 9.8 for the Karhunen-Loeve decomposition.
7.9. Let W/, z = 1,2, be Hilbert spaces, and consider tensors f\^

fi, ft ^ Hi. Set

( / l ^ / 2 Ul<8)g2> : = ( / l | ^ l ) { / 2 1^2).
(a) Show that this formula extends by sesquilinearity to define an inner product,
also denoted ( | ), on the space of all finite sums

Exercises

149

k
(b) Spell out the completion step which turns the tensor product Hx^Hi
into a
Hilbert space.
(c) Let A and B be index sets, and consider two indexed families of vectors
fa en\,gfi eHi^a e A,fi e B, such that {fa)a^A is an ONB in Hx and {gp)p^g
is an ONB in Hi^ Then show that (/ (g) gp)^^ s)eAxB ^^ ^^ ^^^ mHi <S> Hi(d) Conclude from (c) that the tensor-product formula
f {A) (g) ^ (B) = f(Ax

B)

holds for the respective ^^-sequence Hilbert spaces.

(e) Let {Xi, Bi, ///), z = 1,2, be two measure spaces, and let B be the cr-algebra
generated by all sets in X\ x X2 of the form E\ x E2 where Et G BJ. Verify that the
product measure ju := ju\ x jU2 is determined uniquely on B by completion and the
ansatz
ju(Ei X E2) = jui (Ei)jU2(E2).
(f) Let (Xi,Bi, jUi), I = 1,2, and /^ be as in (e). Then prove that there is a
"natural" Hilbert-space (isometric) isomorphism
L^ (Xu //i) 0 L^ (X2, M2) = L^ (Xi X X2, ju) .
(g) For / N, consider compact spaces Xj with Borel probability measvires juf.
Using Kolmogorov's construction (Lemma 2.5.1), carry out the determination of the
infinite-product measvire
/i :

mnt:fi{{{EiX
\i=\

00

[{

onX:=Y[Xi.

Xk] = {{fii{Ei),

k=n-]-\

X\EiX

= lljUi

W Xk\ =

EteBt.

i=l

X\^i{Ei),

(h) Extend the tensor-product formula from (f) above to infinite tensor products.
/=1
k=n^r\ J
i=\
7.10. Infinite tensor products
Let W be a Hilbert space, and let (A)eN be an infinite sequence of vectors such
that ||/z|| = l,and

lim \\hn h\\ =0 for somQh e7i,

CXD

^\\hn
=1

and

- hn-[.l\\ < OC,

150

7 Generalizations and applications

oo

^ { a r c c o s \ { h n \ h)\}^ = oo.
n=\

Consider infinite tensors (g)^ fk = A ^ fi^ " where all but finitely many of
the factors fk agree with the unit vectors hk, i.e., from a finite step onwards, fk = hk
holds. On such tensors, define the inner product

oo

I oo

oo

I 1

k=\

(a) Use a completion argument to show that this ansatz yields a Hilbert space
which depends on the chosen system {hk).
(b) Give a formula for comparing two Hilbert spaces H% (hk) and 7Y^ (/z^)
constructed this way from sequences (hk) and (/z^) of unit vectors, both satisfying
the three conditions stated before (a).
7.11. Let Tti,i = 1,2, be Hilbert spaces, and let C be a conjugation in Tii, i.e., C is
conjugate linear and satisfies C^ = / . Let y/ eH\<^ Hi, \W\\ = 1, be given.
(a) Show that there is a unique bounded linear operator K := Kyj'.Hi > H\
such that

{\Cfx^f2)n,^H2

= ifi \^h)n,^

foraliy;eW,-, z - 1 , 2 .

(b) Show that the operator T := K* K:H2 -^ H2 is of trace class with trace one,
i.e., that for every ONB (ba)aeA ^^ ^2? we have

In particular, justify this convergence.

(c) Use the spectral theorem to show that there is an ONB (ba)aeA ^^ ^ 2 and
numbers ! > 0 such that

aeA

(d) Let K =W (K*K)^^^ be the polar decomposition of the operator K with W

denoting the partial-isometry factor. Then prove that
Kba = A / I ^ Wba

for all

aeA.

(e) Schmidt's decompositioii (also called Schmidt's theorem or Schmidt's factorization). Let y/ be given, ^ e TYi 0 W2 as in (a), and let C, W, la, and ba be as
specified in (c)-(d). Then prove that
= ^y/KCWba
aeA

<S)ba.

References and remarks

151

This last exercise shows that Kolmogorov's extension principle from Lemma
2.5.1 generalizes naturally to the non-commutative setting that we encountered in
our use of the C*-algebra ON- The relations which define ON are essential for wavelet algorithms, and for our representation of corresponding processes from subband
filtering. The formulation below is adapted from [BrJKWOO].
7.12. First some notation: Let A'^ e {2, 3 , . . . } and let IAN be a set oiN elements. Let
Qfin be the set of finite sequences (/'i,..., im) where /> e Zjv and m e { 1 , 2 , . . . } .
We also include the empty sequence 0 in Qfin, and denote elements in Qfin by
/, J,
If / = O'l,..., im) ^ ^fin and / e Z^, we let / / denote the element
( / i , . . . , /;;,, /) in Qfin, andsi = Sf^Sf^ Sj^ e ON and s} = s*j*^_^ ^* e ONIn particular s^ = t = s^.
Let A'^ G {2, 3 , . . . } . Show that there is a canonical one-to-one correspondence
between the following objects.
(a) States P on ON(b) Functions C: Qfin x Qfin -^ C with the following properties:
(i) C ( 0 , 0 ) = 1,
(ii) for any fimction f: Qfin -> C with finite support we have
Z/,yeQfi?a)C(/,/)f(/)>0,
(iii) ZieZM C {lU Ji) - C (/, J) for all /, J G Qfin(c) Unitary equivalence classes of objects (/C, OQ, V\, ..., VN) where
(i) /C is a Hilbert space,
(ii) DO is a unit vector in /C,
(iii) F i , . . . , F i v G B ( / C ) ,
(iv) the linear span of vectors of the form F/uo, where / G Qfin, is dense in /C,
The correspondence is given by
(d) P {sjs}) =C(I,J)

= { V*vo I V}vo).

References and remarks

Images and higher dimension
The early 1990s brought rapid advances on two fronts: wavelets in mathematics
and image processing in engineering. The engineering applications include both the
projects going into compressing and converting fingerprints into digital files [Bri95],
and also the creation of processes used in chips in digital cameras.
On the mathematical side, we note in [Dau92] the use of tensor products in transferring ID wavelet algorithms into higher-dimensional analogues. This was suggested in Daubechies' classic book [Dau92]. In fact our discussion of this in the

152

7 Generalizations and applications

general context of Hilbert space (see Figure 7.10, p. 126, and Exercise 7.8 above) is
close to the ideas in [Dau92, Chapter 10].
Daubechies suggests the use of wavelets for image processing on page 313 in
[Dau92]. She first uses the tensor-product construction in passing fi-om problems
in ID to 2D, and then gives a geometric interpretation of the recursive splitting of
resolution subspaces of Z^ (M?) as an iterated subdivision of squares. With this construction, elements of subsquares simultaneously play two separate roles: On the one
hand, (1) vectors in cascades of subspaces of an ambient Hilbert space, as well as
(2) the digital components of an image represented in grayscale and processed by a
2D wavelet algorithm. She then reproduces a visual rendition in [Dau92, Fig. 10.3,
page 316], and she credits this image-cascade to M. Barlaud.
Other pioneering mathematical presentations involving 2D wavelet algorithms
include [GrMa92] and [LaRe91]. Several results fi-om [LaRe91] were later extended
and appeared in the next generation of papers such as [LaLS96] and [LaLS98]. (We
thank W.M. Lawton for explaining this to us.)
Representations of the Cuntz algebras
The idea of using the representations of the Cuntz algebras On in the study of iterated function systems (IFS) and more generalfi*actalsmay originate with the paper
[JoPe96] by Steen Pedersen and the author.
The figures in this chapter (and elsewhere) were created by Brian Treadway with
Mathematica, and the Mathematica program used for Figure 7.5 (pp. 120-121) has
been included below.
paO[\[Theta]_] \rcolon= (1/4) - (1/4) (Cos[\[Theta]])
- (1/4) Sin[\[Theta]]
pal[\[Theta]_] \rcolon= -(1/4) (-1 + Sin[\[Theta]])
+ (1/4) Cos[\[Theta]]
pa2[\[Theta]_] \rcolon= (1/2) + (1/2) Sin[\[Theta]]
pa3[\[Theta]_] := -(1/2) Cos[\[Theta]]
pa4[\[Theta]_] := (1/4) + (1/4) (Cos[\[Theta]])
- (1/4) Sin[\[Theta]]
pa5[\[Theta]_] := (1/4) (-1 + Sin[\[Theta]])
+ (1/4) Cos[\[Theta]]
aO[\[Theta]_] := paO[\[Theta] + (\[Pi]/2)]
- pa5[\[Theta] + (\[Pi]/2)]
al[\[Theta]_] := pal[\[Theta] + (\[Pi]/2)]
+ pa4[\[Theta] + (\[Pi]/2)]
a2[\[Theta]_] := pa2[\[Theta] + (\[Pi]/2)]
- pa3[\[Theta] + (\[Pi]/2)]
a3[\[Theta]_] := pa3[\[Theta] + (\[Pi]/2)]
+ pa2[\[Theta] + (\[Pi]/2)]
a4[\[Theta]_] := pa4[\[Theta] + (\[Pi]/2)]

References and remarks

153

- pal[\[Theta] + (\[Pi]/2)]
a5[\[Theta]_] := pa5[\[Theta] + (\[Pi]/2)]
+ paO[\[Theta] + (\[Pi]/2)]
loctwont[\[Theta]_] := N[Transpose[{{a2[\[Theta]],
aO[\[Theta]]}, {a3[\[Theta]], al[\[Theta]]}}]]
cascadestep[phitable_, \[Theta]_] := Flatten[Partition[
Flatten[{0, phitable, 0}], 2, 1] . loctwont[\[Theta]]]
correlatewavelet[phistart_, itercount_, \[Theta]_] :=
Flatten[Transpose[Partition[PadRight[
Nest[cascadestep[#, \[Theta]] &, phistart,
itercount], 6 (2"itercount)], (2"itercount)]]]
avec[\[Theta]_] := {aO[\[Theta]],
al[\[Theta]], a2[\[Theta]], a3[\[Theta]]}
signavec[\[Theta]_] := {-aO[\[Theta]],
al[\[Theta]], -a2[\[Theta]], a3[\[Theta]]}
ABstep[wltsc_,\[Theta]_] := {ListConvolve[avec[\[Theta]],
wltsc. If 0], ListCorrelate[signavec[\[Theta]],
wltsc, - 1 , 0]}
unscramble[wltsc_] := Flatten[Transpose[Partition[
wltsc, 6]]]
rescramble[wltsc_] := Flatten[Transpose[Partition[
Partition[Partition[PadRight[wltsc, 2 Length[wltsc]],
2], 3 ] , Length[wltsc] / 6 ] , {3, 2, 4, 1}]]
startit = 7

philevelcount = 5
c u r r e n t t h e t a = -5 \ [ P i ] / 6
f i r s t p h i l e v e l [ p h i s t a r t _ , i t e r c o u n t _ , \ [ T h e t a ] _ ] :=
Map[ABstep[#, \ [ T h e t a ] ] &, c o r r e l a t e w a v e l e t [
p h i s t a r t , i t e r c o u n t , \ [ T h e t a ] ] , {-2}]
n e x t p h i l e v e l [ w l t t r e e _ , \ [ T h e t a ] _ ] := Map[ABstep[#,
\ [ T h e t a ] ] &, Map[rescramble, w l t t r e e , {-2}], {-2}]
Table[Show[Graphics[Map[Point, T r a n s p o s e [ { T a b l e [ i (2^(
- ( s t a r t i t + p h i l e v e l c o u n t ) ) ) , { i , 1, 3 ( 2 " ( s t a r t i t
+ p h i l e v e l c o u n t ) ) } ] . Flatten[Map[unscramble. Nest[
n e x t p h i l e v e l [ # , c u r r e n t t h e t a ] &, f i r s t p h i l e v e l [ { 1 } ,
s t a r t i t , c u r r e n t t h e t a ] , p h i l e v e l c o u n t - 1 ] , {-2}],
p h i l e v e l c o u n t - l ] [ [ f ] ] } ] ] , {AspectRatio \ [ R u l e ]
Automatic, Axes \ [ R u l e ] {True, F a l s e } , Ticks \ [ R u l e ]
None, Frame \ [ R u l e ] True, ImageSize \ [ R u l e ] 74,
FrameTicks \ [ R u l e ] {{0, 1, 2, 3 } , {-2, - 1 , 0, 1, 2 } ,
None, None}, PlotRange \ [ R u l e ] {{-0.125, 3 . 1 2 5 } ,
{-2.5625 , 2 . 5 6 2 5 } } } ] ] , {f, 1, 3 2 } ] ;
We offer the following minimal list of relevant references (books and papers):
covering multiresolutions, [Jor03, BaCM02, BaMM99, BrJo02b]; multiwavelets,
[BaMe99]; and wavelet packets, [CoWi93, Wic93, Wic94]. In our discussion of

154

7 Generalizations and applications

families of filters, high-pass and low-pass, in Remark 7.4.4 above, we rely on an

operator-theoretic formulation which is only implicit in the formulation we offer
here. The operator-theoretic approach uses certain representations of C*-algebras,
the Cuntz algebras O^, A^ = 2, 3 , . . . ; see [Cun77]. However, this representationtheoretic viewpoint is stressed more, and made explicit, in our research papers
[BrJo02a, BrJo02b, Jor03, Jor05, Jor04b].
Combining formulas (7.7.9) with (7.7.14)-(7.7.15), we see that the measures
fi*om Theorem 7.7.2 are simply restrictions of certain quantum-mechanical states
to the canonical maximal abelian subalgebra V in the CAR-algebra 21, i.e., the C*algebra 21 on the canonical anticommutation relations, also called the fermion algebra. The subalgebra D in 21 is simply the C*-subalgebra in the CAR-algebra which
is generated by the diagonal matrix elements from the list (7.7.12), i.e., by the matrix
elements e^^^ when n runs over No and / = 0 , 1 .
While the formulas (7.7.14)-(7.7.15) appear to define pA only on V, i.e., appear to only define a measure, the paper [PoSt70] makes it clear that the measure
is a restriction of a unique state pA on the CAR-algebra 21, the so-called quasi-free
states. These states on the CAR-algebra are called quasi-free states in [PoSt70]. In
fact [PoSt70] initiated an important study (in operator-algebra theory) of a notion
of equivalence (called quasi-equivalence) of these states (generalizing the more familiar notion of equivalence of measures^two measures are said to be equivalent
if they are relatively absolutely continous). The results of Powers-Stormer [PoSt70]
are motivated in turn by seminal work of Kakutani [Kak48] on equivalence of infinite-product measures.
The definition of the C*-algebra 21 on the CARs simply dictates the known relations that arise in models of infinite systems of quantum particles called fermions.
In fact, the CARs axiomatize Pauli's exclusion principle for the fermion particles, or
fermion gases. At the same time, the CAR-algebra 21 models infinite lattice systems
in quantum statistical mechanics; see [Rue69]. The framework applies independently
of the lattice dimension, and it allows the development of a mathematical formulation of equilibrium states that does not rely on a chosen way of taking limits of finite
subsystems.
More generally, the quasi-free states on the CAR-algebra are important in physics.
They are widely used in the study of Fermi gases, and they have played a role in statistical mechanics and in axiomatic quantum field theory over the years; see, e.g.,
[Hal83] and [Rue69].
Independent of the long history of the study of states on the CAR-algebra, the
measures from the conclusion of Theorem 7.7.2 have emerged in statistics and combinatorial probability theory under the name "determinantal measures;" see, e.g., the
paper [LySt03] by Lyons and Steif, and the papers cited there. Lyons et al. and Diaconis et al. [DiFr99] used the determinantal measures in giving a new independent
proof of Szego's limit theorem, and in fiirther extending the scope of Szego's theorem to cover applications to infinite particle models in physics.

References and remarks

155

The determinantal measures have especially attractive ergodicity and monotonicity properties; see [LyStOS]. In fact, they have been used in other research, e.g., in
new proofs of rigorous results on Fredholm determinants in the theory of orthogonal
functions, and in the representation theory of infinite-dimensional unitary groups;
see the references in [DiFr99].
We saw in this chapter that there are specific non-abelian operator algebras whose
representations are of significance for wavelets,fi*actals,and for the kind of subband
filters which are used in both signal processing and in the study of wavelets. Specifically, we studied for each N the Cuntz algebra ON and two of its distinguished
subalgebras. In the special case of N = 2, one of the important subalgebras of O2
is known as the CAR-algebra. Here we are more interested in particular classes of
representations of these C*-algebras than in the C*-algebras themselves. The names
for the various C*-algebras are current lingo in the theory of operator algebras and
in physics. It turns out that the C*-algebras we encountered in this chapter have
been used for a long time in mathematical physics (especially in statistical mechanics [Rue69, Rue89]), where their representations play a role in the study of infinite
particle systems. Ruelle's paper [Rue89] was the one which introduced the transfer
operator, the operator which now goes under the name of the Perron-FrobeniusRuelle operator. As for the C*-algebras themselves, see, e.g., [Dav96] for a friendly
and current treatment.
Unfortunately technical jargon differs from one field to the other. To reduce this
linguistic confiision, we have collected a "translation guide," i.e., a separate glossary
(pp. xvii-xxv above). Readers not already familiar with operator algebras might find
Davidson's little book [Dav96] to be an agreeable introduction.
However, we wish to stress that the representations of the Cuntz algebras ON
play a role in a rich variety of applications, most of which are outside the scope
of this book. Even within electrical engineering itself, signal processing is not the
only place where the Cuntz relations show up. They play a crucial role in systems
theory as well. Indeed, it has been shown by Joe Ball and his coauthors (see, e.g.,
[BaCVOS] and [BaVi05]) that multidimensional linear input/output systems and their
scattering theory may be couched elegantly in terms of representations of the Cuntz
algebras ONThe central idea in this chapter may be encapsulated in Figure 7.7 (p. 124). While
this diagram is from engineering (signal processing), we have stressed that various
equivalent forms of it have come up quite independently in mathematics, and for
entirely different reasons. We believe that the interdisciplinary cormections stressed
here have great potential, but apparently that in the mathematics community they
have been largely overlooked so far.
For mathematics students looking for a friendly survey of filter banks, i.e., the
subdivision method for frequency subband filtering (in an engineering treatment),
we recommend Vaidyanathan's book [Vai93] (a classic!); and for an excellent review
of optimal subband and transform coders, see [VaAkOl]. Many uses of filter banks

156

7 Generalizations and applications

and the associated algorithms in wavelets (pure and applied) are covered in [Mal98],
[VeKo95], and [StNg96].
The next chapter continues the same theme, butfroma more geometric viewpoint
than is traditional.

8
Pyramids and operators

Mathematics is an experimental science, and definitions do not come first,

Oliver Heaviside
but later on.

Operators and their adjoints; Hilbert space; product measure;

set-theoretic partitions; multiindices; a curiosity about some appHcations outside of
mathematics.
PREREQUISITES:

Prelude
In Chapters 4 and 7, we stressed that the crucial feature of localization is shared by
a number of basis constructions, most notably by those of wavelets and of certain
classes of fractals. This includes basis constructions in Hilbert spaces built recursively on fractals and on state spaces in dynamics. The recursive approach to the
more general basis constructions is a special case of a refined tool from probability
which is based on martingales. (It should be contrasted to classical Fourier expansions, which are notoriously poorly localized.)
The localization can be fiirther refined with the use of pyramid algorithms (see,
e.g., the multipart images in Figures 7.4 and 7.5, pp. 118-121), and we will follow
up on this in the present chapter. What should emerge is that our constructions based
on scale-similarity have an intrinsic recursive nature which immediately suggests
numerical implementations of efiicient and iterative matrix algorithms.
We further saw that the use of iteration and scale similarity for bases in fimction
spaces is intrinsic and closely tied to localization. This in turn is part of active and
current research, and has been made precise in a number of brand new research
papers which go far beyond the scope of our book. Sufiice it to call attention to
work by Stephane Jaflfard (e.g., [Jaf05]) on oscillation spaces, including the so-called
Besov spaces (also beyond the scope of our book).

158

8 Pyramids and operators

In connection with localization, we mention at the conclusion of Chapter 5 that

there are further implications for approximation of functions. In particular we cited
papers by Terence Tao (for wavelets) and R.S. Strichartz (for Hilbert spaces on fractals) which show that localized basis expansions in fimction spaces yield much better
pointwise approximations than is possible for traditional Fourier bases; so contrasting again to classical Fourier expansions. In addition, Chapter 7 above emphasizes
that localization is crucial for computation!
In this chapter we shall take up an additional and separate issue related to localization, that of separation of variables. Our approach to this will be geometric, and
as far as possible, it will be based on the notion of tensor products in Hilbert space,
including infinite tensor products. Readers not familiar with Hilbert-space concepts
are encouraged to first do the block of multipart exercises in Chapter 7 (starting
with Exercise 7.9) which are entirely devoted to the kind of tensor products that are
needed.

1r
8.1 Why pyramids
In this chapter, we follow up on and generalize the ideas which form the basis for
the pyramid algorithm (7.5.5) of dyadic wavelet packets. To see how this ties into
more general models, based on branching, we begin with a closer examination of the
unitary operator U2'. L^ (M) -> L^ (R) given by

U2f{t):=^f{2t),

/EZ2(E),

te

(8.1.1)

In particular, we show that there are two representations, {Pi) and (iS/), of the Cuntz
algebra O2, see (7.6.1), such that

U2 =

Y.^i^Sl

(8.1.2)

i=0

We will also study more general pairs of representations (Pf) and (St) of the Cimtz
algebra ON such that the factorization
N-l

(8.1.3)
i=0

has attractive spectral properties.

8.2 Dyadic wavelet packets

(PA
TO

(P6

(P5

y4o T i \

Tl\

(P9

^10

/TO

^11

(P\2

159

(Pl

Tl\

/TO
(PU

(PU

T\

(P15

Fig. 8.1. The fundamental dyadic pyramid.

8.2 Dyadic wavelet packets

The simplest instance of the framework in Section 2.1, especially (2.1.2), is the following: Let X = No = {0, 1 , 2 , . . . } , and set
ro (n) = 2/7,

Ti (n) = 2n -\-1,

ne

NQ.

(8.2.1)

Defining
a (2n + /) := n,

/ = 0, 1, n e No,

(8.2.2)

we see that (2.1.2) is satisfied. Moreover the sequence (po, (p\, (p2, . . . of wavelet
packet functions from (7.5.5) in Proposition 7.5.2 is generated by the pyramid in
Fig. 8.1 below.
We say that the pyramid in Fig. 8.1 is singly generated.
Definition 8.2.1. Let X, c, r/, z = 0 , . . . , TV 1, be a branching system as defined in
(2.1.2). We say that the system has d generators if there is a subset E (Z X satisfying
#E = d

(8.2.3)

160

8 Pyramids and operators

and

|J(T-(^)=X,
,

(8.2.4)

neNo

where
,T- {E) = {x eX\c7''(x)eE}.
(8.2.5)
We now introduce the following modification in Example 2.2.1 fi-om Chapter 2.
We set
To(n)=2n,

Ti(n)=2n-\-3,

G (2n + 3/) = n,

neNo,
ne No, / = 0 , 1 .

Setting X = Z, we see that condition (2.1.2) is satisfied, so the system (Z, cr, T/)
is a dyadic branching system. Moreover the following diagram in Fig. 8.2 and an
induction show that the subset
^:={-3,-2,-l,0}

(8.2.7)

satisfies the condition in Definition 8.2.1. In contrast, (8.2.2) is singly generated.

Another feature which separates the system (8.2.6), with step size 3, is that it
contains a non-trivial two-cycle, i.e., C = {2, 1}. Note that a (2) = ~ 1 , and
cr (1) = 2. In addition, it has two distinct one-cycles, i.e., a (3) = 3, and
a (0) = 0.
Lemma 8.2.2. Let N e N, N > 2, and let a.X ^> X be an N-to-l mapping which
is onto X. Pick branches of the inverse,
T/:X->X,

/ = 0, 1 , . . . , A ^ - 1 ,

(8.2.8)

such that (2.1.2) holds.

On ^ (X), let \x),forx

e X, be the canonical basis vector, and set

Pi \x) := \Ti (x)>,

/ = 0 , 1 , . . . , iV - 1.

(8.2.9)

/ = 0, 1 , . . . , TV - 1.

(8.2.10)

The adjoint operator P* is

P; \x) = XziiX) (x) \a (X)),

Then this system of operators satisfies the Cuntz relations

^1

(8.2.11)

_ /=o
Proof The formulas (8.2.11) follow directly by an application of the expressions
(8.2.9) and (8.2.10) for the operators Pj and their adjoints P*. The computation of
the adjoint operator to Pi in (8.2.9) is straightforward.
n

8.2 Dyadic wavelet packets

-3
70/

TO/

. -.
E3]

-6
y
-12

-1
\Ti

-\
\ ^

0
Vl

-.
p. -n. .P21

-4

N^ '
- 9 -

Xy

y ^
8 - 5

'^ '^ ^
2

-24 -21 -18 -15 -16 -13 -10 - 7

161

fol

^1

'\
5

10

^y

^1
9

13

12

15

18

21

Fig. 8.2. Branching with step size 3.

Lemma 8.2.3. (a) Let (SJ).^Q

Hubert space /C, and set

be a representation of the Cuntz algebra

ON

on the

N-\

U:=Y,Pi^Sl

(8.2.12)
/=o
Then U: i^ {X) 0 /C ^ ^^ {X) ^Kis a unitary operator.
(b) Conversely, if {St) is a system of operators in /C such that U in (8.2.12) is
unitary, then the operators Sf satisfy the Cuntz relations

N-\

(8.2.13)

i=0

Proof We will prove (a), but leave (b) to the reader.

First,
UW

by (8.2.13) ^ - ^

'

''

'^

by (81.11) '^^'W'^'
and secondly.

u*u

X S ^'* ^J ^' ^j
'

by(8.2.11)-^-^--

'^ '^ '^'

by (8=2.13) "^^'W^-

'

162

8 Pyramids and operators

Lemma 8.2.4. Let ((Pn)neNo ^^ ^^^ sequence in L^ (R)from Proposition 7.5.2, and
let K be the Hilbert space L^ (T) with basis
Cj (z) :=zJ,

zeT,

j eZ.

(8.2.14)

Setting
W{\n)^ej):=(pn(t-j),

(8.2.15)

this assignment extends to define a co-isometry of^ (No) (8)/C into L^ (IR) with range
equal to a dense subspace in L^ (M). This co-isometry will also be denoted W.
Proof. The conclusion (7.5.8) from Proposition 7.5.2 is the assertion that the doubleindexed family
j eZ}
(8.2.16)
{(Pni- -j)\neno,
forms a tight frame (also called a PARSEVAL frame) for the Hilbert space Z^ (M). It
follows from this that W in (8.2.15) maps onto a dense subspace in L^ (R); and we
need only check that it is co-isometric.
For f ^ L^ (M), we have the following computation:

ll"-/irp^,ZI(^.i"/)i^
(3ri5)XK^(--i)i/)P

Remark 8.2.5. Let W be the operator from (8.2.15) in the lemma. Then ^ is a unitary isomorphism of ^^ (No)<S)L^ (T) onto L^ (M) if and only if the family (pn(' j)
in (8.2.16) forms an orthonormal basis in Z^ (M).
hi our next result, we show that the familiar dyadic scaling operator U2 of (8.1.1)
has a tensor representation with respect to the two representations of the Cuntz algebra O2 from above.
Theorem 8.2.6. Let Pi, i = 0,1, be the representation ofOi which is described in
Lemma 8.2.2 a<i(8.2.1)-(8.2.2), and let St, i = 0, 1, be one of the representations
of02 described in Lemma 7.6A. Let
W: f (No) 0 L^ (T) -> L^ (R)

(8.2.17)

be the co-isometry of Lemma 8.2.4.

Then
W^Pi^S*

= U2W,

where U2 is the dyadic scaling operator (S.l.l) in L^ (R).

(8.2.18)

8.2 Dyadic wavelet packets

163

Proof. First
^

Pi ^ S* {\n) (8) ej) = ^ ( X 1 ^ " + '^ ^ ^i^J 1

(^-^-l^)

1
?=0 A:eZ
1

= S S ^^y-2)t ^2+/ (t-k),

z=0 A:eZ

where
a f ^ = ak

and

4^^ = (-1)^ ^ i - ^

f^^ ^ ^ ^ '

(8.2.20)

see also (7.5.1H'7-5.2). Substituting this back into (8.2.19), we get

W ( X ^ ^ ^ ^i (l^> ^ ^i)) = ^ ^ (2^ -^') '

(^-^-^l)

which is the desired conclusion (8.2.18).

Corollary 8.2.7. Let the two representations ofOi be as described in the theorem,
and assume in addition that the functions in (S.2.16) form an orthonormal basis
(ONE) in L^ (M).
Then the unitary scaling operator Ui in L^ (M) is unitarily equivalent with
Proof Immediate from the theorem, and Remark 8.2.5.

Proposition 8.2.8. (a) Let the two representations ofO^ be as described in Theorem
8.2.6, and assume the functions in (8.2.16) are orthonormal.
Then
1
( ^m ( - ^ ) I ^l(Pn ( - 7 ) ) =

^^^m,lnM

( ^i^k

I ej )

'=^

for all m,ne

No.

(8.2.22)

(b) Let a e L^ (T) with Fourier series

a{z) = Y^akz\

ZGT,

(8.2.23)

and set
akcpn ( -k).

a*(pn = ^

(8.2.24)

A:eZ

Then
1

{a^(Pm\U2b^(Pn)

= ^^m,2+/ {Sia\b).
/=0

(8.2.25)

164

8 Pyramids and operators

Proof. It follows from sesquilinearity that the second conclusion is implied by the
first. To verify (8.2.22), we carry out a computation and use the results in Section
7.5.
Before starting the computation of the left-hand side in (8.2.22), write m G No
in the form m = p -\-21, p e {0, 1}, and / e NQ. (This is the familiar Euclidean
algorithm in its simplest form!) We now calculate the term form = /? + 2/; but since
p runs over {0, 1}, the result must be stated as a sum, as in (8.2.22).
We have
{(Pm(- -k)\

U2(Pn(' -J))
/

(by Plancherel)

(Pmit-k)s/2(pn{2t-j)dt

j ek (x) (pm (x) - ^ (pn ( ~ ) ej ( - ) dx

V2

ek (2x) ^m (2x) ^n (x) ej (x) dx

JR

by (7.5.8)

Vl
\fl

(by orthonormality)

by (T.e.S)

ek (2x) rrip (x) (^/ (x) ^n (x) ej (x) dx

JR

\ ek (2x) nip (x) ^

cpi (x + r) cpn (x + r) ej (x) dx

v 2 / ek (2x) nip (x) di n^j (x) dx

Jo
^/, V2 / ek (x) ^0

rnpiy) ej (y) Jx

^ 2>;=;c mod 1

which is the desired result. Recall that

1

^l,n = ^^w,2w+/,
and only one term on the right-hand side in (8.2.26) is non-zero.

(8.2.26)
n

Corollary 8.2.9. ([CoWi93], [Wic93]) Let the two representations ofOi be as described in Theorem 8.2.6, and let (p{), (p\, ... be the functions generated by the algorithm (7.5.7). Consider a subset
^ C No X No.
Then the following two conditions are equivalent:

(8.2.27)

8.2 Dyadic wavelet packets

165

(i) the segments

[2Pn, 2P {n + 1)) c No,

(/?, n) e A,

(8.2.28)

form a non-overlapping partition {or tiling) o/No;

and
(ii)
{ UP(pn {' -j)\

(p, n)eA,

jeZ}

(8.2.29)

is an ONB in L^ (E).
Proof. We give a proof which rehes on Theorem 8.2.6 above. Having the representation
1

U2 = Y^Pi^S*

(8.2.30)

i=0

makes it clear that

UP = ^Pi^S}

(8.2.31)

(where the summation is over all multiindices / of length p).

Using elementary properties of tensor product in Hilbert space, we similarly get
(U^y = U*P = ^ p ; (8) SI.

(8.2.32)

Recall, the formula for Pj is given in (8.2.10). With the multiindex notation
/ = (/i, / 2 , . . . , /p),

ij e {0,1}, 1 < 7 < p,

it is then easy to verify the formula (8.2.32) for {^2)*. The reader only has to insert
the expressions for the multiindexed operator monomials. Specifically, we have

SI = S i , S i ^ ,

(8.2.33)

s} = si...s*^s*^.
We fiirther refer to the formulas (7.6.18)-(7.6.19) for the explicit computation of
the multioperators Sj, and adjoints Sj.
With this, it is clearfi-omProposition 8.2.8 that
{(Pmi- -k)\UP(pn{'

-j))=

(Pm{t-k)2P'^cpn{2Pt-j)dt

m
I

The two transformation rules follow immediatelyfi*omthis. We have

166

8 Pyramids and operators

U^(Pn (- -J)=

(^^^^ \ej)(pm(' -k),

(8.2.34)

(ii,...,ip),m,
2Pn-\-ii-\-'-\-ip2P-^=m

Moreover, when A is chosen, we get the converse basis transformation:

(Pmi- -k)=

Y.

(^i I ^^^^) ^2^" ( -J)'

(8.2.35)

2Pn-\-ii-{-'Hp2P~^=m

The equivalence of (i) and (ii) is immediate from this.

The next section gives more details on the choice of the sets A which make the
segments [l^n, 2^ (n + 1)) tile No- We will show that the measures from Section 7.6
allow us to make "good" selections of sets v4 c No x No and therefore "good" ONBs.

8.3 Measures and decompositions

In this section we revisit the measures from Section 7.6. Indeed these measures were
introduced in [CoWi93] and [CoMW95] for the purpose of choosing between the
families of orthonormal bases (ONBs) which are listed in (8.2.28) from Corollary
8.2.9. The measures help in the use of entropy methods in the selection of the "best"
ONBs.
In (8.2.28) from Corollary 8.2.9, we consider segments in No, and when two
segments [2Pn, 2P (n + 1)) and \2P'n\ 2P' [W + l ) ) are disjoint then the corresponding basis fimctions t/^^^w ( j) and U2 (Pn' {' j') ^r^ orthogonal, and vice
versa. This follows from the two transformation rules (8.2.34)-(8.2.35) which make
it clear that a segment [2^w, 2P (n + 1)) consists precisely of the numbers m e No
which admit a dyadic representation
m

2^ 4- /i + 4- ip2P-^ = 2Pn + x.

(8.3.1)

Working in M / Z = [0, 1), we see from (8.3.1) that the partitions from (i) in Corollary 8.2.9 correspond to dyadic partitions of the unit interval. The correspondence
between (a) the "/?-subintervals," or segments of the integers No, and (b) the dyadic
fractional intervals in [0, 1) is
[2^, 2^ ( n + ! ) ) - >

'^+-2P

+ ^2'. 42P- ^ - - + ^ + ^ ) .

(8-3.2)

where the integer x in (8.3.1) is specified by the multiindex / , and the Euclidean
algorithm, i.e.,
X = zi + Z22 + + ip2P~^,

with ij e {0,1}.

(8.3.3)

8.3 Measures and decompositions

167

In the sequel, we will use this 1-1 correspondence m (natural number) <-^ / (/'i,...).
If m is given, and ip is the last non-zero term in (8.3.3), we say that p is the length
ofw, orof/.
Comparing with (8.3.1), the length is the smallest p for which there is a representation (8.3.1) with ^ = 0. The number p in (8.3.2) is this length, i.e., p := length (/).
As a result, p is determined by w in (8.3.2), and 2~P is then the length of the corresponding dyadic subinterval in [0, 1) on the right-hand side in (8.3.2).
Ifm is as in(8.3.1), then2"^m = + zi2-^H
[-ip2~^ = /i2~^ + - +/p2-^
+ ip2~^ is the left endpoint in the subinterval on
modulo Z; and indeed ii2~P -\
the right-hand side in the formula (8.3.2).
Left endpoints to left endpoints, and the same for rights! So 2Pn should correh ip2-\ and 2^ (n + 1) to ii2-P -\
h ip2~^ + 2~P. NOW
spond to i\2-P H
2P (ii2-P + .. + / ^ 2 - i ) = / ! + . . . + ip2P-^ = X,
and
2P (^h2-P + . . . + ip2-^ + 2-P^ = X + 1.
So when / varies over the /7-tuples (/'i,..., ip), then x covers [0,2^); and when n
is chosen as in (8.3.1), then the integers m = 2Pn + x cover \2Pn, 2P (n -f 1)).
Hence, a specific multiindex / = {h, - - -^ip) "selects" a unique subinterval in
[0, 1) of length 2~P, and with endpoints which are dyadic rationals.
Definition 8.3.1. If (Si)j^Q is a representation of 2 in Z^ (T), and
*J/ = *Jzi ^ip^

then
P (I) := SjS}
is the orthogonal projection of Z^ (T) onto the subspace Sjl-^ (T).
We saw in Section 7.6 (Chapter 7 above) that the correspondence
/ 1-^ P (/)

(8.3.4)

extends naturally to a projection-valued measure defined on the Borel subsets B of

[0, 1); see also [Jor05] and [Jor04b] for additional details.
Specifically, this measure extension (also denoted P ()) will be cr-additive, and
it is orthogonal. This means that the identity
P(BinB2)

= P(Bi)P(B2)

(8.3.5)

holds for all 51,^2 e B.

As a result, every h e L^ (T), \\h\\^ = 1, induces a probability measure juh
according to the formula
juh (B) := \\P iB)hf,

for ^ B.

(8.3.6)

168

8 Pyramids and operators

Proposition 8.3.2. Let {Si)]^^ be some representation ofOi on L^ (T) which corresponds to a quadrature-mirror filter (mo,m\) as described in Lemma 7.6.1. Let
p,n e No, and J e Z.
Then the measure jUej (-) on[0,1) prescribes the distribution of the wavepacket
coefficients ofthefiinction
{U^cpn) {t - j) = 2^/2^ {2Pt - j)

(8.3.7)

relative to the ONB

{(Pmi- -k)\me^Q,

keZ}.

(8.3.8)

Proof. The result follows from (8.2.34) in Corollary 8.2.9 above. This is the formula
which gives the expansion of U2(pn ( 7) in the ONB (8.3.8).
Hence, to prove the proposition, we only need to compute the ^^-norm of the
expansion coefficients from (8.2.34). The computation is just a simple application of
Parseval's formula to the standard Fourier basis
for A: e Z, and z

ek {z) := z^

T.

Indeed, summing the expansion coefficients over k, we get

keZ

keZ
(byParseval) " ^ -^ "

{ej\SiS*jej)

{ej\PiI)ej)

(see (8.3.4)) "

^ ^ -^ "

jUej (the dyadic interval labeled by / ) ,

which is the desired conclusion.

8.4 Multiresolutions and tensor products

In this section we show that the tensor factorization of the unitary scaling operator
U from Section 8.2 gives rise to a general class of multiresolutions. Multiresolutions were introduced into wavelet theory by Mallat [Mal89], and they are based on
ideas of Kohnogorov [Kol77] from dynamics, and on the notion of a martingale from
probability theory [Wil91], as we noticed in Chapters 2 and 3 above.
In the present context, we shall take "multiresolution" to be defined as follows.

8.4 Multiresolutions and tensor products

169

Definition 8.4.1. Let (/ be a unitary operator in a Hilbert space 7Y, and let Ho be
a closed subspace in H. We say that Ho generates a multiresolution for U if the
following two conditions (i)-(iO hold (if all three conditions (i)-(iiO hold, we say
that the multiresolution is pure):
(i) Ho c UHo,

(ii) V U'^Ho = K
neZ

(iii) / \ f/Wo = {0},

neZ

where the symbols V and / \ denote the lattice operations from Hilbert space, i.e.,
V applied to a family of closed subspaces in H means "the closed linear span", and
/\ means "intersection".
Remark 8.4.2. Let a system (U, H, Ho) be given as in Definition 8.4.1, but suppose
only (i)-(iO ^r satisfied. Then U and H may be modified such that the resulting
reduced system {U\ H\ HQ) satisfies all three conditions (i)-(iii). This follows from
a simple application of the Wold decomposition; see [BrJo02b].
We now sketch the details of proof: Suppose only (i)-(ii) hold. Then set
/C:= /\WHo.
H' :=HeK:

(8.4.1)

= {f ^H\{f\k)

=0 for 2i\\ke1C},

(8.4.2)

and
HQ

= Ho /C.

We leave to the reader the verification of the assertions; i.e., that the reduced
system Ho -^ HQ, H -^ H\ and U -^ U\ U' := U\w, satisfies all three conditions
(iHiii).

The significance of Definition 8.4.1 is reflected in the following general result on

unitary equivalence.
Theorem 8.4.3. Let (U, H, Ho) and (U\ H\ HQ) be two given systems as stated in
Definition 8.4.1, and assume that all three conditions (i)-(iii) holdfor both systems,
i.e., that the two multiresolutions are pure.
Then there is a unitary isomorphism
W'.H-^H'

(8.4.3)

which satisfies the following two conditions:

WHo = HQ

(8.4.4)

and
WU = U'W

{intertwining).

{We say that the systems are unitarily equivalent.)

(8.4.5)

170

8 Pyramids and operators

Proof. This is a standard result in operator theory, and we refer the reader to
[BrJo02b, Chapter 2]. We further add that the underlying ideas date back to Kolmogorov [Kol77] in the 1930s.
n
We now show that the tensor factorization U = X / ^i ^ ^t from Section 8.2
above gives rise to multiresolutions.
Theorem 8.4.4. Let N e N, N > 2, and let {X, a, r/) satisfy the conditions in
Lemma 8.2.2. Let (Pi)f^^ be the corresponding representation of ON in t^ {X). Let
{Si)^J^ he a representation of ON in a Hilbert space JC, and set
H:=^^(X)(g)/C,

(8.4.6)

A^-l

U:= ^ A ^ ^ ; .

(8.4.7)

Let a subset E (Z X satisfy (a) and (b) below,

(a)

CT{E) = E

and
[ja-''{E)^X,

(b)
eNo

and set
H:=^^((T-"())(8)/C.

(8.4.8)

Then
U^HQ

= Hn

(8.4.9)

and
{U, H, Ho) is a multiresolution system,

(8.4.10)

I.e.,

no:=^(E)^IC
generates a multiresolution for U.
Proof To prove that (i) in Definition 8.4.1 is satisfied note that for every subset
E o X,we have
(8.4.11)
E iz (J-^ ((7 (E)) .
Substituting (a) from the theorem, we get
N-l

Eza-\E)=
But if ^ e E and k e IC, then

U Ti{E).
/=o

(8.4.12)

8.4 Multiresolutions and tensor products

\e)^k

= UU*\e)k

171

(8.4.13)

= U{

The desired conclusion Ho O UHo follows.

Ifne No, then
^^^(1^)^^)=

|r/^r,v..r,^e)0 5*,..-^**.

(8.4.14)

iu...,ip

Now, combining (8.4.14) and (8.4.12), we conclude that

However, the argument from (8.4.13) shows that this inclusion is in fact an identity.
The remaining property (ii) from Definition 8.4.1 clearly follows from this, and
(b) in the theorem.
n
Remark 8.4.5. In applications of Theorem 8.4.4, it is usefiil to reduce general multiresolutions to a system of minimal ones. If A^ and (X, a, r/) are as stated in Theorem
8.4.4, it is helpfiil to make a carefiil choice of the sets E to be used in (a) and (b) of
the theorem.
If ^ C X satisfies a (E) = , we say that E is minimal if the following holds:
0i^AizE,
\=^ A = E.
a{A) = A

(8.4.15)

In our example (8.2.6) above, we had E {3, 2, 1,0} satisfy G (E) = E,

but each of the sets
^i={0},
^2 = { - 2 , - 1 } ,
and
E3 = {-3}
satisfies a (Ej) = Ej, and each Ef is minimal.
Remark 8.4.6. The next result shows that such minimal sets Ef may be chosen in
general. However, it is only for rather special branching systems (X, cr, T/) that the
choices are natural.

172

8 Pyramids and operators

Lemma 8.4.7. Let N eN, N >2, and let (X, cr, xt) satisfy the conditions in Lemma
8.2.2. Suppose some subset E c X satisfies conditions (a) and (b) in Theorem 8.4.4.
Then there are subsets Ef c E such that
c7(Ei) = Ei,
EinEj=0

(8.4.16)
fori^j,

(8.4.17)

[JEi=E,

(8.4.18)

and
each Ei is a minimal solution to (8.4.16).

(8.4.19)

Proof. Zom's lemma.

Definition 8.4.8. If ^ is a subset of X, we set

A:= | J a - " ( ^ ) .

(8.4.20)

neNo

Proposition 8.4.9. Let N e N, N > 2, and let (X, a, T/) satisfy the conditions
in Lemma 8.2.2. Let [E, (Ei)j^j) be a system of subsets in X which satisfies the
conditions in Lemma 8.4.7, where I is a chosen index set.
Then
Ei HEj =0
for i + j
(8.4.21)
and

U Ei = X.

(8.4.22)

iel

Proof We leave the easy details to the reader.

Remark 8.4.10. The significance of conditions (8.4.21)-(8.4.22) in Proposition

8.4.9 is that they generate mutually orthogonal multiresolutions in the Hilbert space
^^ (X) (g) /C. Specifically, we get an orthogonal decomposition
^{X)^K=

^ f- {Ei) (g) K

(8.4.23)

which reduces the operators, and the representations on ^ (X) (8)/C considered above.
Returning to the three sets E\, E2, and E^ fi'om Remark 8.4.5 (and Figure 8.2,
p. 161,fi-om(8.2.6)), we get Z written as a disjoint union of the associated three sets
El, E2, and E3. The sets are equivalence classes for an equivalence relation studied
in[BrJo99a]:
El = {0, 3,6,9,12,15, 18,21,24,27, 30, 33, 3 6 , 3 9 , 4 2 , 4 5 , . . . } ,
E2 = {-2, - 1 , - 4 , 1 , - 8 , - 5 , 2 , 5, - 1 6 , - 1 3 , - 1 0 , - 7 , 4 , 7, 10, 1 3 , . . . } ,

Exercises

-2

-1

-4

-16

-13

173

- 1 - 2

-10

-7

10

13

Fig. 8.3. The pyramid on E2 = {2, 1} (may be seen as a subdiagram of Figure 8.2, p. 161).

and
E3 = {-3, - 6 , - 1 2 , - 9 , - 2 4 , - 2 1 , - 1 8 , - 1 5 , . . . } .
The two sets E\ and E3 are singly generated. The middle one can be understood
from the pyramid shown in Figure 8.3.

Exercises
8.1. Let H be a complex Hilbert space. By a projection, say P, we mean a linear
operator P:H -^ H which satisfies P = P^ = P*. An operator S'.H -^ His
called 2i partial isometry if *S'*iS' is a projection. Recall that the norm of a bounded
linear operator T:H ^ /C is defined as

lir||=sup{||r/^|||/z6H, \\h\\ = i}.

(a) Suppose that an operator P satisfies P = P-^. Show that P is a projection if
and only if IIPII = 1.
(b) Let S be an operator such that iS'**S' is a projection. Then show that SS* is also
a projection. The two projections are called the initial, resp., the final projection ifS
is given to be a partial isometry.
(c) Show that if AS is a non-zero partial isometry, then \\S\\ = 1.

174

8 Pyramids and operators

8.2. Let HbQs, Hilbert space, and let A^ e N. A set ^SQ, . . . , SN-\ ofN operators is
said to satisfy the Cuntz relations if (i) and (ii) hold:
(i) S*Sj = Si JI (where / denotes the identity operator), and
N-\

(ii) J^SiS; = L
i=0

(a) Suppose it is given that the individual operators Sf are isometries, and that (ii)
holds. Then show that (i) is automatic.
(b) Suppose that for some N, the Cuntz relations have a representation in H.
Then show that H must be infinite-dimensional.
8.3. Let n = f^ (Z), and let TV^ e N be fixed. Let be a subset of Z such that there
is a bijection between E and the cyclic group Z^ = Z/A^Z. Define
SrCk := er+Nk,

r e E,

keZ.

for all r,/

e E,

Then show that

S* Sr^ = dry I
L reE

8.4. Let n = L^ (T), and let A^ e N be given. Let mo, m i , . . . , MN-I e I ^ (T),
where we give T the representation R / Z . Set

Sjf(x)

= VNmj (x)fiNx)

for / G L^ (T),andO

<j<N.

(We identify fimctions on T with 1-periodic fimctions on E.)

(a) Show that the operators [Sj) .~Q yield a representation of the Cuntz relations
if and only if the matrix
/

mo (x)

mo (x + ;^ j

mi(x)

mi(x-\-j^j

\mN-\(x)

mN-i\x-{-j^j

mo (x-\- ^

mi(x +

mN-\{x-\-^j

(i) the fimctions (mj) .~Q satisfy the unitarity condition in (a),
'Wy I X + - I my I X + - I = Skj, and

^ j

is unitary for a.e. x e E.

(b) Show that the following three conditions are equivalent:

(ii) ^

Exercises

175

N-\

(iii) ^

rrik Ix -\r -j mi Ix -\- -j

= Skj.

7=0

8.5. Examples
For convenience, make the identification T = [0, 1), H = L^ (T) = L^ (0, 1), so
that Haar measure on T corresponds to the restricted Lebesgue measure on the unit
interval. Functions on [0,1) will be identified with 1-periodic fimctions on E via the
obvious extension. Set en (x) := e^-^^"^, n e Z, x e M.
(a) For N = 2, check that the following pairs of fimctions satisfy the unitarity
conditions which are listed in (a) (and (b))fi-omExercise 8.4.
Example 1: mo (x) = /[o,i/2) (x). mi (x) = /[i/2,i) (x).
1
e\ (x)
Example 2: mo (x) =, mi (x) = -^.
V2
V2
1
^3 (x)
Example 3: mo (x) ~ z=, mi (x) = ^ .
Example 4: mo (x) = cos (TUX), mi(x)

= sin (TVX).

Example 5: mo (x) =
, mi (x) =
, .
17
1 ic
/A
1 + ^3 (x)
Example 6: mo (x) =
, mi (x) =

.
l-esix)
.

(b) For N = 3, check that the following systems of three fimctions satisfy the unitarity conditions which are listed in (a) (and (b)) in Exercise 8.4.
Example 7: mo (x) = /[o,l/3) (x), mi (x) = /[i/3,2/3) (x),
m2ix) = X[2/3A)(X).
62 (x)
1
ei (x)
Example 8: mo (x) = T=, mi (x) = , m2 (x) = ^ .
y/j

v 3

s/j

Example 9: Set ^3 := e' 3 = - _ -j. j _ ^ and

mo (x) = - (1 + ^1 (x) + 62 (x)),
mi (x) = - (1 + C3^i (x) + C3^2 (x)),
^2 (X) = 3 ( 1 + C3^1 (X) + f3^2 (X)) .
17

1 1A

/^ ^

Example 10: /wo (x) =

m2(x)^-^.

^^2(x)

. .

j=, ^ i (x) =

1 - e2 (x)

=,

176

8 Pyramids and operators

8.6. For the function mo in the examples in Exercise 8.5, consider the corresponding
isometric operators in 7Y = I ^ (0, 1) given by
5o/(x) = ^ m o ( x ) / ( 2 x )

ifA^-2,

^o/(x) = V3mo(x)/(3x)

ifiV = 3.

and
Show that the condition
lim^o*"/-0,

feU,

(8E.1)

is satisfied in Examples 1,4, 5, 6, 7, 9, and 10, but not in Example 2, 3, or 8.

8.7. (a) For the representations ofOi in Examples 1,4, 5, and 6, verify irreducibility,
i.e., that there is no Hilbert space /C, 0 ^ /C ^ 7Y, which reduces the action of O2 on

n.
(b) For the representations of O3 in Examples 7, 9, and 10, verify irreducibility
on the Hilbert space H = L^ (0, 1).
(c) For the representations in Examples 2, 3, and 8, directly verify reducibility.
8.8. Show that the respective representations of O2 and O3 lift to wavelet bases (or
normalized tight fi-ames) for L^ (E) in the cases of Examples 1, 4, 5, and 6 (the
dyadic case), and for examples 7 and 9 (the triadic case).
8.9. Show that the respective representations of O2 and O3 do not lift to wavelet
bases (orfi-ames)in the Hilbert space L^ (M) in the cases of Examples 2, 3, 8, or 10.
8.10. Discuss the relevance of the representation of O3 from Example 10 (in Exercise 8.5) for the middle-third Cantor set X3, and its Hausdorflf measure hs of Haushi2
dorfif dimension 5 = logci2 = .
^^
In 3

References and remarks

We were [initially] entirely in Heisenberg's footsteps. He had the idea that
one should take matrices, although he did not know that his dynamical quantities were matrices
And when one had such a programme of formulating everything in matrix language, it takes some effort to get rid of matrices.
Though it seemed quite natural for me to represent perturbation theory in
the algebraic way, this was not a particularly new way.
^Max Bom
Wavelet packets were pioneered by Coifinan, Meyer, and Wickerhauser; see especially the seminal papers [CoMW95], [CoWi93], and [Wic93]. The idea and the
motivation come from the equivalence of the two properties (i) and (ii) in Corollary

References and remarks

177

8.2.9. By a carefiil choice of the set A, one is able to adjust the orthonormal basis
(8.2.29) in such a way that the frequency bands are adapted to the space (or time)
localization. Since the functions in (8.2.29) have the form

2P/^,p{2Pt-j),
it is clear that large choices ofp yield narrow frequency bands and vice versa.
Other helpfiil references on the subject may be found in Wickerhauser's book
[Wic94]; see also [CoMW95], [BeBe95], [Jor05], and [Jor04b].
We have outlined in this chapter a particular interplay between on the one hand,
certain selection, or optimal choice problems, in wavelet theory, and on the other,
the wider field which often goes under the label of "non-commutative probability
theory." While we are only scratching the surface of this very active research area,
we want to mention two more directions of non-commutative probability which are
not covered here: (1) Free probability, and (2) dilations of finite or infinite sets of
non-commuting operators in Hilbert space. Readers who wish to learn more about
(1) might wish to consult the monograph [VoDN92], [JoSW95], or [KoSp04]; and
for (2) we recommend [Pop89], and [DaKSOl].

Pairs of representations of the Cuntz algebras O^,

and their application to multiresolutions

If one finds a difficulty in a calculation which is otherwise quite convincing,

one should not push the difficulty away; one should rather try to make it the
^Werner Heisenberg
centre of the whole thing.

PREREQUISITES:

Permutations; trace; tensor products of matrices; Cantor.

Prelude
This chapter will resume our study of the geometric approach via Hilbert space to
separation of variables/tensor product stressed in the last two chapters; and we will
note a number of applications of this idea. The approach further serves to clarify a
number of themes involving combinatorics of the recursive bases studied throughout
the book. The separate themes are as follows.
(1) The tensor-product idea yields an explicit representation of the unitary operator U which we use to model scale-similarity in our constructions both for standard
wavelets and for fractals; see especially Lemma 9.3.2.
(2) Using tensor products we show that there are two representations of the Cuntz
relations involved in modeling basis constructions on scale-similarity (such as was
pioneered first for the standard wavelet bases in L^ (E^)). As already noted, our use
of multiresolutions naturally entails representation of the Cuntz relations; representations which come directly from subband filters. But in addition, we find a second
family of representations, one which reveals symmetry under certain permutations;
see Section 9.3. The resulting second class of representations of the Cuntz relations
has in fact already been studied earlier (for different purposes) under the name "permutative representations."
(3) The tensor-product idea fiuther shows that the representation-theoretic approach is as useful for new basis constructions in fimction spaces on fractals as it is

180

9 Representation duality and multiresolutions

for the standard wavelet bases in Z^ (R^). In Chapter 4 we introduced natural Hilbert
spaces on affine fractals, and in Theorems 9.6.1 and 9.6.2 below we show how these
spaces admit bases of generalized wavelets (i.e., "fractal wavelets") and of wavelet
packets.
(4) Our use of tensor products helps organize the possibilities for the wavelet
packets introduced in Chapter 7. In Theorem 9.4.3 we show that when a system of
(admissible) digital filters is given, then the feasible choices of wavelet packets are
in one-to-one correspondence with certain discrete pavings, or tilings.
(5) Finally, in Section 9.5 we outline how our use of tensor products clarifies the
discrete wavelet transform. Recall that a choice of a multiresolution (in the generalized sense outlined in Chapter 7) automatically entails a discrete wavelet transform.
When a resolution subspace is selected, this allows us to process data which is organized in an associated sequence space (i.e., in an ^^-space); and it is precisely in
these ^^-sequence spaces where our Cuntz relations are realized.

9.1 Factorization of unitary operators in Hilbert space

In this chapter we show that the examples from Sections 8.2 to 8.4 in the previous
chapter may be formulated and understood within an operator-theoretic framework.
This abstract formulation helps to unify the examples, and to bring to light an underlying representation-theoretic flavor to the subject. At the same time, this general
formulation may be of independent interest in operator theory. The examples from
Chapter 8, i.e., wavelet packets and permutative representations, have appeared in
various research papers in pure and applied mathematics; and it may now be an opportune time to try to give them a general and more axiomatic formulation.
This concluding chapter begins with a study of unitary operators U in Hilbert
space. Motivated by the examples from Sections 8.2, 8.3, and 8.4, we consider the
following general question: If [/ is a unitary operator in a Hilbert space H, when
does it admit factorizations

U = Y^Si^V,\

(9.1.1)

where {St) and ( ^ ) are representations of the Cuntz algebras O/l We outline a
number of results and applications regarding these factorizations.
We show that fractals and wavelet constructions have attractive computational
features. What the different examples have in common is a kind of scaling-similarity.
The scaling will be represented by the unitary operator U. As we move up and down

9.2 Generalized multiresolutions

181

a scale of powers of U, there will be intermediate bands corresponding to subdivisions. One representation {St) of On serves to encode these bands, and a second
representation is then used in recovering U from (Sf). This leads to a representation
(9.1.1) for the operator U.

9.2 Generalized multiresolutions

The relations known in operator theory as the Cuntz relations, i.e., S*Sj = Sjjt
and ^j SfS* = 1, and which define the Cuntz algebras On, n = 2, 3 , . . . , also
play a big role in signal processing and wavelet analysis. A common feature in these
applications is a detailed spectral analysis of some unitary operator U:H -^ 7i
which plays the role of scaling. This could be scaling between different resolutions
in a sequence of closed subspaces of the underlying Hilbert space H, or the scaling
could refer to a system of frequency bands.
Two structures are present: scaling from one band to the next, and operations
within each band. In this chapter, we show that these applications may be subsumed
in a certain tensor factorization of H.
Li the simplest case, H = L^ (M) and U is the dyadic scaling operator
Uf{t) = V 2 / ( 2 0

for / G L^ (M) and/ e R.

(9.2.1)

In the case of a multiresolution analysis (MRA), there is a subspace Vo C 7Y =

Z^(IR) such that
Vo C f/Vo,
(9.2.2)
/ \ ^ V o = {0},

(9.2.3)

Y ^Vo = H,

(9.2.4)

and

eZ

where / \ and V refer to the usual lattice operation on closed subspaces of 7i.
In the MRA case for wavelets, see [Dau92], there is a fimction cpQ e Vo such that
Vo is the closed span of the translates
{^o(- -k)\keZ}.

(9.2.5)

It is then possible (in favorable cases) to arrange that the translates in (9.2.5) are
mutually orthogonal vectors of unit norm, i.e., that
/ <Po ( 0 (po (.t -k)dt

= So,k.

(9.2.6)

182

9 Representation duality and multiresolutions

There is a pyramid algorithm (defined below) for generating a sequence cpQ.cpi,...

with
(Pn e WVo
(9.2.7)
and such that
{(Pn{- -k)

I ^ G N O , keZ]

(9.2.8)

is an orthonormal basis (ONB) for L^ (E), i.e., such that

{(Pn{'

-k)\(pn'{'

-k'))

Sn,n'h,k''

We show that this structure arisesfi-omtwo representations of the Cuntz algebra O2,
see [Cun77], {Si) and (FJ), such that the unitary operator U of (9.2.1) has the form
U = ^Si^v;,

(9.2.9)

The two representations are defined naturallyfi^omthe wavelet data: the isometrics
Si act in ^ (No) and the Vi 's in L^ (T). In the simplest case, we have
Si \n) = \2n + i)

for w G No and / = 0, 1,

(9.2.10)

where we used Dirac's terminology for the natural basis |w) in ^^ (No). The isometrics Vi,i = 0, 1, are defined by two functions rrii on T = R / Z by
{Vif)(x)

= ^/2mi(x)f(2x)

for/GZ^(T), X G E , z =0,1.

(9.2.11)

The Cuntz relations for (9.2.11) are equivalent to the requirement that the matrix
mo(x)

mo (x + ^) \
)
[
m\ (x) m\ yc -\- j \ J

(9.2.12)

is unitary a.e. x G T.
It turns out that there are much more general systems of representations of the
Cuntz algebras On which nonetheless involve the same geometry. The applications
include the theory of iterated fiinction systems (IPS), symbolic dynamics, tilings,
harmonic analysis, and fractals.

9.3 Permutation of bases in Hilbert space

Definition 9.3.1. Let 7i be a complex Hilbert space. We say that a finite set of isometrics Si'.T-L ^ HJ = 1 , . . . , w, forms a representation of the Cuntz algebra On if the
following system of identities holds:
(a)

S*Sj=dijt

and

(b)

Y.^iS*
i=\

where t denotes the identity operator.

= t,

(9.3.1)

9.3 Permutation of bases in Hilbert space

183

It is well known, see, e.g., [Cim77] and [BrJP96], that there is a one-to-one correspondence between representations of the C*-algebra On, defined by the relations
(9.3.1), and the set of all systems of isometrics (Sf) subject to (9.3.1). Moreover,
Cuntz [Cun77] showed that the C*-algebra On is simple.
As a result, we will identify the set of all representations of On on some Hilbert
space H with the systems (9.3.1): A system (9.3.1) will be said to be an element
in Rep (On, H). It is known that the set of equivalence classes of irreducible representations of On is "too large" to admit a meaningfiil classification, see [Gli60];
specifically, this set does not have a Borel cross section.
Lemma 9.3.2. Let Hi, i = 1,2, be complex Hilbert spaces, and let n e N, > 2,
be given. Let (Sf) e Rep (On, Hi), and let V\,... ,Vn be a system of operators in
the Hilbert space Hi- Then the following two conditions are equivalent:
(i) the operator
n

U:=^Si<^V;

(9.3.2)

i=\

is unitary inH'.= H\ 0 H2,

and
(ii) ^ E R e p ( a , W 2 ) .
Proof. Consider arbitrary pairs of vectors Xi.yt e Hj^i = 1,2. Using the relations
(9.3.1) for the iS-system, we find that the unitarity property (i) is equivalent to the
following two identities:

(^1 I yi ){x2 I X ^ ^ > 2 ) - (xi I yi){x2 I yi)

(9.3.3)

and
Z Z (^1 I ^'^jy^)

(^2 I v; Vjn ) = {xi\yx){x2\y2)

(9.3.4)

Setting e (/, j) := StS*, we see that (9.3.1) is a restatement of the famihar matrix
identities
e(i,jr=e(j,i)
(9.3.5)
and
e(i,j)e(kJ)

= Sj^ke(i,l)

(9.3.6)

for all /, j,kj e [l,.. .,n}.

If we form the n-hy-n block matrix
M(V) = {V; Vj)lj^^

e Mn (C) 0 B (Hi) = Mn (B (Hi)),

(9.3.7)

the two identities (9.3.3)-(9.3.4) simply state that M(V) is the identity matrix in
Mn (B (Hi)), i.e., that

184

9 Representation duality and multiresolutions

V;Vj=8ujIm-

(9.3.8)

Let trace be the normalized trace on M (C), and consider

trace 0 id: M (C) (g) B {Hi) -^ B {Hi) .
Applying this to the element (Vt V*Y

in M (C) (g) 5 (Hi), we get X / ^ ^*-

But identity (9.3.2) also yields (trace (g) id) ( Vt V* j = t^^. As a result, we get
^VtV;

= tn,.

(9.3.9)

Combining (9.3.8) and (9.3.9), we get the implication (i) => (ii). Since the converse
n
implication is immediate, the proof is completed.
Definition 9.3.3. Let X be a set, and let R\X -> X be an endomorphism. Let
R~^ (x) := {y e X \ R{y) =x} iox X e X. We say that R is an n-fold branch
mapping if R is onto and if
#7?-i (x) = n

for all X G X.

(9.3.10)

For a given n-fo\A branch mapping, we shall select n branches of the inverse, i.e.,
n distinct mappings
af.X-^ X,
/ = 1,...,,
(9.3.11)
such that
Roai=\dx

forall/= 1,...,^.

(9.3.12)

Definition 9.3.4. Let 7< be a complex Hilbert space, and let {Si) e Rep {On, H) for
some /7 G N, > 2. We say that {St) is apermutative representation, see [BrJo99a],
is each isometry *S'/ permutes the elements in some orthonormal basis (ONB) for H.
Lemma 9.3.5. [BrJo99a] Up to unitary equivalence of representations every permutative representation {Sf) of On in a Hilbert space H has the following form for
some set X and some n-fold branch mapping R:X -^ X.
Let H = ^ {X), and for x e X, let \x) be the corresponding basis vector in H.
IfcX-^
C is in l^ {X), then
c - ; ^ c ( x ) |x),

\\ct = ^

xeX

\c{x)\\

(9.3.13)

xeX

and
[x\y)

= S:c,y

forallx,yeX.

(9.3.14)

For a choice of distinct branches at, i = I,... ,n, set

Si\x):=\ai{x)),

(9.3.15)

9.4 Tilings

185

Proof. Let {Si) e Rep (0, Ji) be a permutative representation. Let X be an index
set for some ONB which is permuted by the respective isometrics St. As a result,
there are maps af.X ^ X such that
Si \x) = \(7i (x))

for / = 1 , . . . , w and x e X.

(9.3.16)

Using (a) in (9.3.1), we consider that each cr/ is one-to-one, and that
(7i (X) n cTj (X) = 0

if / / j .

(9.3.17)

Using (9.3.14) and (9.3.16), we derive the formula

S* \(7j (x)) = Si J \x)

for /, 7 = 1 , . . . , w and X 6 X,

(9.3.18)

for the adjoint operators S*, i = 1 , . . . , . When (9.3.18) is substituted into (b) of
(9.3.1) we conclude that
[Jc7i(X)=X.

(9.3.19)

/=i

As a result, we may define R:X -> X by setting

R ((7j (x))=x

for / = 1 , . . . , ^ and X G X,

(9.3.20)

and conclude that R, defined this way, is an n-fo\d branch mapping. Substituting
back into (9.3.18), we conclude that
S* \x) = XoiiX) 1^ (^))

for z = 1 , . . . , andx e X.

9.4 Tilings
Let X be a set, and let w e N, w > 2, be given. Let R:X ^> X be an -fold branch
mapping. For x e X and /> e N, set
R-P{x)\={y^X\RP{y)=x].

(9.4.1)

It is convenient to include the case /? = 0, and set

E(x,p)

R-P (x)

lip e N,

{x} (the singleton)

ifp = 0.

(9.4.2)

Definition 9.4.1. A subset .4 C X x No is said to define a tiling of X if

U

E(x,p)=X

(9.4.3)

(x,p)eA

and
E(x,p)nE

{x\ y ) =0

if (x, /7) # {x\ y ) in A,

i.e., the sets E (x, p), indexed by distinct points in A, are disjoint.

(9.4.4)

186

9 Representation duality and multiresolutions

Examples 9.4.2. In the next three examples we illustrate the use of tensor products
as outlined in Lemma 9.3.2. The scaling will be represented by the unitary operator
U. One representation {St) of 0 serves to encode subdivision bands, and a second
representation is then used in recovering U from {St). This leads to a representation
(9.3.2) for the operator U. The first representation {Si) of On will be a "permutative
representation," i.e., it will be defined from a permutation of a canonical basis in one
of the two tensor factors. Formula (9.2.10) is an example of such a representation,
but there are many more; see for example the memoir [BrJo99a]. The second representation {Vi) will be given by a quadrature-mirror filter as in (9.2.12), or more
generally filters corresponding to n subbands as given in Chapter 7. The reader is
encouraged to work out the two types of representations in Exercises 8.3 and 8.4.
The additional feature in this construction is tensor products of Hilbert spaces.
For this part the use of operator theory helps clarify the use of tensor products and
setting up a "wavelet transform." We are fiuther relying on Dirac's elegant bra-ket
notation: see the conclusion of Chapter 2 for details.
Example 9.4.2.1. Let X = No, and set
R (2n) = n

md

R (2n -f 1) =

forne

NQ.

(9.4.5)

Then for (n, p) e X x No, we have the identity

E (, p) == [n2P, (n + 1) 2^).

(9.4.6)

One easily checks that the mapping R in (9.4.5) is a 2-fold branch mapping with
branches
tro (n) = 2n, a\ (n) =2n-\-l
torn ENQ.
(9.4.7)
If / e Q (/?), i.e., / = ( z i , . . . , z^), then
aj (n) = h + Z22 + . . . + ip2P-^ + n2P,

(9.4.8)

and it follows that the sets E (n, p) are as described in (9.4.6). In conclusion, the
possible tilings of X = No associated with R in (9.4.5) are the partitions of No into
non-overlapping segments of the form (9.4.6). Here are three distinct types of such
tilings.
Case (a):

^ = {(zi, 0) | G No}, and

E (n, 0) = {n} (the singleton)

Case(b):

^ = {(0,0), (1, j^) |/> No}, and

E (1, p) = [2P, 2^+^ )

Case (c):

for w e No.
forpe

No.

^ = {(0,2), (1,2k) , (2, 2k) , (3,2A:) | A: G N } , where now

^ ( 0 , 2 ) = {0,1, 2, 3}, and
E0,2A:) = [722^ 0 + 1)2^^)

for 7 = 1,2, 3 and A: G N.

We stress, however, that there are many more types of examples.

9.4 Tilings

187

Example 9.4.2.2. It would be tempting to define R on No by the following rules,

R(2n) = n

and

R(2n-\-3)

= n,

(9.4.9)

by analogy to (9.4.5) in Example 9 A.2.1, but the equation I =2n-\-3 does not have
solutions in NQ. One checks that no subset X of No allows the rules (9.4.9) in the
definition of a 2-fold branch mapping. Nonetheless, if we take X = Z, then (9.4.9)
does define a 2-fold branch mapping.
So there are several differences betwen the two examples 9.4.2.1 and 9.4.2.2. For
Example 9.4.2.1, every point is attracted to {0} in the sense that ifne No, then there
is a /? such that R^n = 0. For Example 9.4.2.2, there is no singleton in X = Z
which serves as an attractor. Nonetheless, for every n e Z, there is a p such that
RPn e {3, 2, 1, 0}. The analogues of the tiling systems (a)-(c) in Example
9.4.2.1 carry over to Example 9.4.2.2 as follows.
Case (a):

A = {(n,0) \ n e Z] ,md
E in, 0) = {n} (the singleton)

Case (b):

Case (c):

for w G Z.

This tile system now contains a more varied system

of tiles for the set X = Z. The singleton tiles are
E (n, 0) = {n} for n = 3, 2, 1,0, and in addition there are the following five classes of (nonoverlapping) dyadic tiles:
E(-6,p)

for all;? e No,

E(-4,p)

forall/?GNo,

E(l,p)

foralljr? GNO,

E (3, p)

for all p e No, and

E(S,p)

forallj^GNo.

We leave this to the reader.

In conclusion, the two examples show how libraries of wavelet bases are
constructed fi-om two entirely different families of representations of the Cuntz
algebra (9.
The occurrence of one of the representations (Vf) is not surprising: it is determined by the kind of subband filters familiar from both signal processing and the
standard multiresolution approach to wavelet bases. It is the second representation
(Si) of On, the permutative representation, that completes formula (9.2.10) for the
scaling operator U, encodes our wavelet bases and accounts for the variety of tilings
as they arise in Proposition 8.3.2 and Theorem 8.4.4. As in Proposition 8.3.2, we start
with an ONB {cpn | G No} as in (8.3.8); we apply suitable powers of U to these
fimctions, creating the fimctions LfPcpn, and then ask which subsets of the double
indices yield new ONBs. The same principle applies to more general basis constructions, and Theorem 8.4.4 spells out the tilings that yield bases of wavelet packets.

188

9 Representation duality and multiresolutions

In the systematic study of permutative representations in [BrJo99a] it is shown

that each permutative representation can be understood from a single endomorphism
R of the index set. In the next example we introduce R: see formula (9.4.10).
Example 9.4.2.3. In this next example, we use an arbitrary but fixed scaling number
N eN,N >2, generalizing Example 9.4.2.1 above. We take X = No, and set
R{k + Nn) = n

for A: = 0, 1 , . . . , TV - 1, and for all n e NQ.

(9.4.10)

As in Example 9.4.2.1, one checks that this is a branch mapping, in this case an
A^-fold branch, i.e., there are A^ distinct branches of the inverse, cjk (x) := k -\- Nn,
k = 0,1,.. .,N I. For (n, p) e X x No, X = No, one checks the formula
Ein,p)

= [nNP, ( + 1)7V^).

(9.4.11)

The three types of tilings (a)-(c) from Example 9.4.2.1 then generalize as follows.
Case (a):

A = {(n,0) \ n ENQ} ,md

E (n, 0) = {n} (the singletons)

Case(b):

for n GNQ.

^ = {(0,0), (1,;?), ( 2 , p ) , . . . , (TV - 1,;?) |/7 e No}, and

E (k, p) = [kNP, (A: + 1) 7V^), A: = 0, 1 , . . . , TV - 1, p G No.

Case (c):

^ = {(0,2), (1, Nk) , (2,7VA:) , . . . , ((7V^ - 1) , 7VA:) | A: e N } ,

where now
E (0,2) = lo, 1,2, 3 , . . . , 7V^ - i j , and

E{j,2k) = {jN^\{j

+ ^)N^^)

for 7 = 0, 1 , . . . , A^^ - 1 and A: e N.

In the rest of the chapter, we study some more general features of the -fold
branching mapping R which occurred in a special case in Example 9.4.2.3. In our
general study of permutative representations in [BrJo99a], we refer to R as the
canonical endomorphism. What makes the next theorem different is that it yields
an axiomatic approach to libraries of bases in Hilbert space. It applies to the kind
of iterated-fimction-system fractals studied in Chapter 8 above and in more detail
in [BrJo99a], and at the same time it generalizes the pyramid approach to bases of
wavelet packets; see Proposition 8.3.2, Theorem 8.4.4, and Figure 7.5 (pp. 120-121).
An ONB {(pn I w e No} as in (8.3.8) yields one index set. The second arises
from the scale-similarity inherent in wavelet analysis, scaling by a power of 2 or iV,
or in several variables by a power of a fixed scaling matrix. The scaling is a unitary
operator U in the ambient Hilbert space, which may be Z^ (M^), and then we create
the fimctions UP(pn, and ask which subsets of the double indices yield new ONBs.
The setting in the next result. Theorem 9.4.3, is maximally flexible in that it allows a completely general ONB and index set X. Once this is given, we may consider

9.4 Tilings

189

the most general permutative representations as in [BrJo99a]. Depending on the application, e.g., an iterated-function-system fractal, a time-series problem from signal
processing, or an image-processing algorithm, we are then able to build a unitary
operator, a tensor-product representation, and associated tilings and bases of wavelet
packets, so generalized libraries of bases of wavelet packets.
A special class of iterated-ftinction-system fractal where this works well is the
aflBne iterated fiinction systems. In Section 9.6, we show how multiresolutions and
generalized libraries of bases of wavelet packets may be adapted to these systems.
Closing the circle from Chapter 4, we begin with the middle-third Cantor set, and
our associated wavelet families are given in Theorem 9.6.1, so-called space-filling
wavelet bases. In Section 9.7 we go on to study the wider class of these afiine fractals.
Theorem 9.4.3. Let R: X ^ X be an n-fold branch mapping, and let (Sf) be
the corresponding permutative representation on t^ {X). Let K be a Hilbert space
with orthonormal basis {ONE) 8 = {|^)}' ^^^ ^^^ ( ^ ) ^ Rep(0,/C). Set
U = X L l ^i ^ K' ^^^ letA(ZXxnobea
subset.
Then the following two conditions are equivalent:
(i) the vectors UP\x ^b) indexed by (x, p) e A and b e 6 form an ONE for
^ (X) 0IC; and
(ii) A defines a tiling ofX.
Proof Let n = { 1 , . . . , w}, and set Q (p) = n x n x

x n. Then

p times

UP\x^b)=

\Six)(^\Vj*b),

(9.4.12)

/eQ(p)

SLndSjx e E (x, p).

(ii) => (i). Assuming (ii), it is clear that the corresponding vectors in (i) form an
orthonormal family in H := i^ {X) 0 /C. Let j e X and V e B. Pick (x, />) G A
such that y e E{x,p). Then y = Sjx for a imique / = (z'l,..., ip) e Q (p), and
|j;> (g \b') = UP\x^

Vib'),

(9.4.13)

where we used the identity

V} Vi = djj

for J, / G Q (/?) .

(9.4.14)

A substitution of
Vib' = ^{b\

Vib')

(9.4.15)

I Vib')Up \x^b).

(9.4.16)

beB

into (9.4.13) yields

\y)^\b')

= ^{b
beB

190

9 Representation duality and multiresolutions

Since the vectors \y) 0 \b^) form an ONB for H, we conclude that the family (i) is
total in H: the vectors on the right-hand side in (9.4.16) are precisely the vectors
listed in (i).
(i) =^ (ii). We now assume that the vectors in (i) form an ONB in H.
Let (x,p), {x\p') e A. Then for every I e Q(p) and / ' e Q ( / ) , Sjx e
E(x,p) and Sfx' e E {x\ / ) . Since the vectors UP \x (g) b) and UP' \x' 0 b') from
(i) are orthogonal, we conclude from (9.4.12) that (9.4.4) holds, i.e., that the distinct
sets E (x, p) and E {x\ p') for (x, p) and (x^ p') in A are disjoint.
If (9.4.3) were not satisfied and y e X\\J^^
x^^ E (x, p), then it follows from
(9.4.13) that \y, b') is in the orthogonal complement of the whole family (i). Hence
(i) could not be total. Since we assume that (i) is an ONB, the proof is completed, n

9.5 Applications to wavelets

Let H be the space L^ (M) of all I^-fimctions on E, and let A^ G N, A^ > 2, be given.
Then the scaling operator
iUf) (0 = ^f(Nt),

feKteR,

(9.5.1)

is clearly unitary. The purpose of this section is to show that every multiresolution
wavelet decomposition of W, corresponding to scale number N, gives rise to a tensor
factorization of U of the form (9.3.2); specifically, U has the form
N-l

U=^Si(S^V;.

(9.5.2)

Here the tensor factorization in (9.5.2) refers to

n = ^ (No) (g) L^ (T)

where T = R / Z ,

(9.5.3)

and where
(Si) e Rep (ON, ^^ (No))

and (Vi) E Rep (ON, L ^ ( T ) ) .

(9.5.4)

Furthermore, the two representations in (9.5.4) are specified as follows. The representation (Si) in (9.5.4) is the permutative representation from Example 9.4.2.3
above. (ViN = 2, it is the special case in Example 9.4.2.1.)
The representation (Vf) from (9.5.4) is defined from a system of fimctions
mo,..., rriN-i in L^ (T) such that the TV x TV matrix
/

mo(x)

mo(x + ^ j

mo (x + ^

wi(x)

mi(x-\rjjj

mi\x-\-^j

ymN-\(x) mN-i(x-\-j^j mN-i(x-{-^j J

(9.5.5)

9.5 Applications to wavelets

191

is unitary a.e. x e T.
We recall the following two lemmas from [BrJo02b]; see also [Wic94].
Lemma 9.5.1. Let mo, m\,...,

rriN-i be in L^ (T), and set

(Vif)(x)

= ^mi(x)f(Nx)

(9.5.6)

for feL^
(T), x eR, andi =0,1,...
,N - I.
Then ( ^ ) G Rep {ON, L ^ ( T ) ) if and only if the matrix (9.5.5) is unitary a.e. x e
T.
The next lemma is also known, but we sketch it for reference.
Lemma 9.5.2. The Fourier transform
oo

e-'^''^'f it) dt
/

(9.5.7)

-OO

realizes a unitary isomorphism

(9.5.8)

W: L^ (M) = ^ (Z) (g) L^ (T)

as follows:

(Wf)(x):=[fix+k))^^^ef(^).

(9.5.9)

Proof Initially, Wf () is just a fimction from M mapping into ^^ (Z). But an application of Parseval's identity and Fubini's theorem shows that

ZK^^+^)

'^^=

V^^^l ^''=

\f(t)\^dt.

(9.5.10)

Hence the mapping W in (9.5.9) passes to a unitary isomorphism of L^ (R) onto

L^ (T, ^ (Z)) = ^2 (Z) (g) Z^ (T).
n
Lemma 9.5.3. Let m(),m\,..., m^-x be in L^ (T) and suppose the unitarity condition (9.5.5) is satisfied. Assume in addition that mo is Lipschitz near x = 0, and
that mo (0) = \fN.
(a) Then there is a sequence (po,(p\,... in L^ (E) such that
(pNn (Nx)

mo (x) (pn (x) ,

(PNn+\ (Nx)

m\{x)

(pn (x) ,

(9.5.11)
[ ^iV+iV-l (A^^)

m^-X (x) q>n (^) ,

and
(po(t)dt = \.

(9.5.12)

192

9 Representation duality and multiresolutions

(b) The following three additional conditions on the functions (po^cpi.cpi,... are
mutually equivalent:
(i) / \(pQ{t)\' dt = \,
(ii) ^

\(po (x + 7)| = 1
-^:

(iii) ^(pn{^+
jei

rr ^

a.e. x e R,
..

J)^n' (^ + J) = ^n,n'

a.e. X G

Remark 9.5.4. The clearest way to visualize the sequence {(pn | No} from
Lemma 9.5.3(b), see especially (9.5.11), is to picture the functions in the case of
TV = 2, for the Haar wavelet construction. Li that case,
(pin (0 = (Pn (20 + (PnQ't

-\),

(pm-^x (0 = (Pn (20 - (pn {2t - 1) ,

(9.5.13)

n = X[o,\)
The first 32 functions in this sequence are reproduced in Figure 7.4 (pp. 118-119)
where some features of their construction may be observed, such as the partial symmetry/antisymmetry of ^ (in the Haar case) at all dyadic scales:
(Pn(2~^ - 0 = ^ ( 0 .

= 0 , 1 , 2 , . . . , ^ = 0, 1 , 2 , . . . , a.e. t e (0,2~^).

Theorem 9.5.5. Let mo, mi,..., mjsf-x be functions in L^ (T) which satisfy the
conditions in Lemma 9.5.3(b), and let(po,(p\,(p2,... be the corresponding sequence
in L^ (M). Then the double-indexed family
{(Pni- -k)

I w e N o , keZ]

(9.5.14)

is an orthonormal basis for L^ (R). Set

^k (x) := exp(72;rA:x),

(9.5.15)

and let (^)y J^ be the representation of ON defined in Lemma 9.5.1; see (9.5.6).
Then
N-\

y/N(Pn{Nt-k)=Y,Y.i

^i^J I ^k)(PNnM {t - j) .

(9.5.16)

Proof The fact that (9.5.14) is an ONB in Z^ (R) follows from Lemma 9.5.3(b).
When the operators ^ , z = 0 , 1 , . . . , A/^ 1, are defined from the filter functions mj,
as in (9.5.6), it follows from Lemma 9.5.1 that (Ji) e Rep {ON, L^ ( T ) ) . Hence to
prove (9.5.16), it is enough to show that
/ (pNn+i(t-j)VN(Pn
JR

(Nt -k)dt

= { ViCj I Ck)

(9.5.17)

9.5 Applications to wavelets

193

for all n e No, and all j . A: e Z. On the right-hand side of (9.5.17), we use the inner
product in L^ (T).
The proof of (9.5.17) is again based on Parseval's identity, and a second application of Lemma 9.5.3(b). Starting with the left-hand side in (9.5.17), we set
h WjpiV^"^ where wi, 25 e {0, 1 , . . . , A'^ 1}. Then
n = n\ -\- niN H
(Pn {x) = rrin^ (x/N) "-mnp (x/NP) ^o {x/N^) ,
and
/ (PNn-}ri{t-J)^(Pn

(Nt - k) dt

m
= / ej (x) ^Nn+i {x) N~^l^^n () ek (^j dx
= ^

\ ej (Nx) ^Nn+i {Nx) (pn (x) Ck (x) dx

m
= \fN I Cj (Nx) Mi (x) \^n (x)\ ek (x) dx
JR

= VN / ej (Nx) jfii (x) ^ ^ |^ (x + / ) | ek (x) dx

= \fN I ej (Nx) mi (x) ek (x) dx
Jo

= / ^ T W {V;ek) (x) dx
Jo
= {ej I V;ek) = {Viej \ ek).

Since every system /WQ, . . . , rriN-i which satisfies the conditions in Lemma
9.5.3(b) generates functions (po^cpi,... in Z^ (R) such that
{(Pn(' -k)\.neNo,

keZ}

is an ONB in L^ (E), the assignment

(Pn(' -k)^

\n)^\k)

(9.5.18)

extends to a unitary isomorphism

L^ (R) ^ {^ (No) (g) {^ (Z) ^ ^ (No) 0 L^ (T),

(9.5.19)

where we use the familiar isomorphism ^^ (Z) = L^ (T) defined by the usual Fourier
basis {ek \ k eZ] introduced in (9.5.15).

194

9 Representation duality and multiresolutions

Corollary 9.5.6. Let the functions mo,m\,...

Lemma 9.5.3(h), and let

,mN-\

satisfy the conditions in

L^ (M) ^ ^ (No) (8) ^^ (Z)

be the unitary isomorphism defined by the corresponding ONB
{(Pn{' - A : ) | W E N O ,

A:GZ};

^^^ (9.5.19) aw^ (9.5.14).

Let (Vj) e Rep (On, ^^ (Z)) be definedfi^om the (mt) system as in Lemma 9.5.1,
and wet
Si \n) = \Nn + />

fi)ri

= 0,l,...,N

-I

andn e No

(9.5.20)

be the permutative representation of Example 9.4.2.3.

Then the unitary scaling operator
Uf(t):=VNf(Nt)

forfeL^iW),

t eR

(9.5.21)

has the tensor factorization

N-\

U=.Y,^i^K'

(9.5.22)

Proof Once we have identified the isomorphism (9.5.18), the result follows fi*om
(9.5.16) in Theorem 9.5.5. Specifically,
U {\n) ^\k))

= ^(pn(Nt

- k)

N-\

= ^^{^i^j

\^k) (PNnM it - j)

i=0 jeZ
N-\
i=0 JGZ

N-\

/=0

9.6 An application to fractals

In a paper with D. Dutkay [DuJo06b] we showed that a certain class offi*actalsadmits
a multiresolution analysis (MRA). It is the class of fi-actals which is generated by a

9.6 An application to fractals

195

finite system of contractive affine mappings in R^, and it includes the middle-third
Cantor set C3.
Rather than treating the most general case here, we will restrict attention to two
examples, C3 and C4. The Cantor set C3 is the unique compact subset of R (in fact
of [0, 1]) which satisfies
3C3=C3U(C3-f2),
(9.6.1)
and C4 is the unique compact subset of R which solves
4C4 = C4 U (C4 + 2) .

(9.6.2)

Both examples are generated by scaling and subdivision into similar "smaller" replicas, and we define the HausdorfF dimension of the fractals by
c :=

log (number of replicas)

log (magnification factor) *

(9.6.3)

Furthermore, we note that

c(C3) = log3(2) =

hi2

in3

(9.6.4)

and
c(C4) = log42 = ^

= l.

(9.6.5)

Using a standard theorem from geometric measure theory, see, e.g., [HutSl], we
observe that there are unique probability measures /^3 and /^4 with supports C3 and
C4 respectively such that

y / (0 dix^ (0 = \{j f ( 0 dii^ it) + Jf (4"^) djus (O) (9.6.6)

and

j f (0 dfi4 (0 = ^ ( / / (^) dfi4 (0 + Jf ( ^ ) d/i^ (O)

(9.6.7)

hold for all bounded continuous fiinctions / on R.

Set
7^3:=

t e

BneNo,

t = ^
and

U e {0, 1,2} such that

:^ and only a finite number of tj 's are 1 L

(9.6.8)

196

9 Representation duality and multiresolutions

7^4 :=

/ G

3 e No, ti e {0, 1, 2, 3} such that

t = ^

-J and only a finite number of// 's are 1 or 3 [.

(9.6.9)

Using the respective Hausdorff measures {dty on E we get the following separable
Hilbert spaces:
(9.6.10)
^(iog3 2) ^ ^2 /^^^ (^^yog3 2\
and
^(1/2) ^ ^2 /^^^ (^^1/2^

(9.6.11)

In general, we get unitary scaling operators U in these fi*actal Hilbert spaces

given by the formula
iUf) (0 = V'# (replicas)/(7V0

for t e E,

(9.6.12)

where N = the magnification factor. Specifically, for 7^3,

(^3/)(0 = V 2 / ( 3 0 ,

(9.6.13)

and for IZ4,

(UA/) (0 = V2 / (40

for t e E.

(9.6.14)

In each example, we have a system of filter fimctions (mj) satisfying condition

(9.5.5) before Lemma 9.5.1. (Recall that condition (9.5.5) is equivalent to the fact
that the operators Vt in (9.5.6) form a representation of the Cuntz algebra ON-) It follows that the two examples are associated to representations of the respective Cuntz
algebras O3 and O4 acting on Z^ (T).
For TZs, the three fimctions are
mo(z) =
mi(z)=z,

(l-\-z^y
(9.6.15)

m2 (
where we have set z = ^'^^^.
For TZ4, the four fimctions are
mo
mi (z) =z,

W3

(9.6.16)

9.6 An application to fractals

197

Theorem 9.6.1. (a) There is an orthonormal system (ON) ^o? ^i? ^2? ^^
L^ (TZS, (dtf^^^ ^) which solves the system of equations
(P3>n (0 = (Pn (30 + (Pn Ot " 2) ,
^3+l(0 =-/2(pn{2>t - 1),

(9.6.17)

^3+2 (0 = (Pn (30 - ^ (3^ - 2) ,

^0=/c3-

(b) The functions

{(Pn{' -k)\neno,

keZ}

(9.6.18)

form an ONBfor L^ {113, (dtf''^^ ^).

(c) If (Si) e Rep (O3, ^^ (No)) and(Vi) e Rep (O3, ^^ (T)) are the associated
representations, see Example 9.4.23, N = 3, and Lemma 9.5.1, then

U3 = X^i
i=0

(9.6.19)

v;.

Theorem 9.6.2. (a) There is an an orthonormal system (po,(p\,(p2, - -- in

L^ (TZ^, (dty^'^) which solves the system of equations
<P4n (0 = (Pn (40 + (Pn (4^ - 2) ,
^4+i(0 = V 2 ^ ( 4 ^ - l ) ,
^4+2(0 =V2q)n(4t

-3),

(9.6.20)

^4+3 (0 = ^n (40 - (Pn (4/ - 2) ,

(b) The functions

{(Prti- -k)\neNo,

keZ}

(9.6.21)

form an ONBfor L^ (7^4, (dty^^)(c) If (Si) e Rep {O4, ^ (No)) and(Vi) e Rep (O4, L^ (T)) are ^/ze associated
representations (see Example 9.4.2.3, N =4, and Lemma 9.5.1), ^Aew
^4 = ^ ^ , - 0 P ^ * .

(9.6.22)

z=0

Prao/ We have presented the details in Sections 9.4 and 9.5 so that the proofs of
the two theorems follow as an easy application. The main point is the observation
that the two recursive systems (9.6.17) and (9.6.20) admit solutions which satisfy the
respective orthogonality conditions. But the two Hilbert spaces L^ (TZs, (dt)^^^^-^)
and L^ (TZi, (dt)^^'^) have been defined so that this becomes clear. To see that the
respective orthonormal systems (9.6.18) and (9.6.21) really are total, one may use

198

9 Representation duality and multiresolutions

the MRAs which we introduced in [DuJo06b]. The multiresolutions in the respective

fractal Hilbert spaces share the same Hilbert-space geometry with L^ (M) which we
presented in Section 9.5 above. In particular, our tensor factorizations (9.6.19) and
(9.6.22) for f/3 and f/4 follow from the idea in the proof of Corollary 9.5.6 above. D

9.7 Phase modulation

The difference between the two examples of fractals in Section 9.6 above has to do
with phase modulation, i.e., with the operators of multiplication by
ex (0 := e-'^""^^

for t e E.

(9.7.1)

So far, our wavelet constructions have involved only scaling and translation. The
phases (9.7.1) have entered only because of their use in understanding translations
via Fourier duality.
Lemma 9.7.1. [JoPe98] Let C4 be the quarter Cantor set, and let jU4 be the corresponding normalized measure supported on C^ and satisfying (9.6.1). Set
finite sums, nj e {0,1}

A = A4 =

(9.7.2)

7>0

Then the fi^nctions {ex | 1 e A4} restricted to C4 form an orthonormal basis in

L^ (C4, //4).
Remark 9.7.2. It is known [JoPe98] that there are no more than two orthogonal
fimctions ex (for any A e M) in the Hilbert space L^ (C3, /^s).
Theorem 9.7.3. Let N = 4, and let {cpn | /? No } be the orthonormal sequence in
L^ {TZ4, (dt)^^-^) constructed in Theorem 9.6.2.
Then the family
{ex(t)(pj(t-k)\XeA4,

7 = 0 , 3 , k^Z]

(9.7.3)

and
{ ex it/A) (pj(t-k)\leA4,

j = h2, keZ}

(9.7.4)

is an orthonormal basis {ONE) in the Hilbert space L^ (TZ4, (dt)^^'^).

Proof The proof follows from a combination of (i) Lemma 9.7.1, (ii) the discussion
of the affine fractals in Section 9.6, and (iii) some basic observations of scaling and
the Hausdorff measure; see also [DuJo06b] and the following list.
Several observations are needed:
(1) A4 c 7^4.

Exercises

199

(2) Setting Xj (t) = (t-\- j) /4,7 = 0, 1,2, 3, we note that C4 = TQC^ U r2C4, as a
non-overlapping union, and

(9.7.5)
L^3

=XTOC4 - ^ T 2 C 4 -

(3) The Hausdorff measure h^^^ restricted to C4 is fx^.

(4) The integration {dt)^^^ in the definition of the Hilbert space refers ioh^^^.
(5) Borel subsets of M are /z^/^-measurable.
(6)
(9.7.6)
and
(9.7.7)
for all ft'/^-measurable sets B, all ? M, and all c R+.
(7)
/

/ (0 dh'l^ it) = \ [

/7^4

f {rj it)) dh"^ it)

(9.7.8)

^ J7^4

holds for all /^^/^-measurable fimctions on IZ^.

With this, we leave the remaining detailed verifications to the reader.

Exercises
9.1. Let (p e L^ (E) and suppose there is some sequence {ak)k&z such that
(pit) =2^Qkcp

(2t-k),

t e E.

keZ

(a) Show that if f^ cp (t) dt ^ 0, then

(b) Suppose there are constants c\,C2 such that 0 < ci < C2 < CXD, and

(9E.1)

200

9 Representation duality and multiresolutions

2

ciJ^m^<lY,^k(p(' -k)
WkeL

<C2Y,\ik\'.
Z2(E)

Let Vfp (=: Fo) denote the closed linear span of {^ ( A:) | A: 6 Z}.
Then show that W = W^ given by
(ik)keZ

\keZ

(9E.2)

defines a bounded linear operator fi-om VQ into ^^ (Z).

(c) Show that the operator W in (b) has zero kernel, i.e., that
{fe

Vo\Wf

= 0} = {0}.

(d) Define an operator SQ on fi (Z) by

A:eZ

and let U2 be the dyadic scaling operator in L^ (M), i.e..

Then show that the following intertwining relation holds:

WU2 = SoW

on To.

(9E.3)

9.2. We continue with the assumptions introduced in Exercise 9.1 above.

(a) Using Exercise 5.7, find a Ruelle operator R which has the fiinction
h(x) := Av|^| (x) := ^

|^ (x + ) |

neZ

as eigenfiinction, i.e., with Rh = h.

(b) Using the fimction X!eZ |^ (^ + '^)| ^ ^ ^ (^X ^^^ ^^ explicit formula for
the adjoint operator W*\^ (Z) -> FQ, when ^ ^ is the intertwining operator introduced in Exercise 9.1 (see (9E.2)).
9.3. Let the operators W, U2, and -So be as described in Exercise 9.1.
(a) Show that there is a unique isometry T:Vo -^ ^^ (Z) such that

(b) Show that the operator T in part (a) satisfies

Exercises

201

(c) Show that the operator T in part (a) satisfies

W*W

=T*{WW*)T.

(d) Show that the operator T in part (a) satisfies

TU2 = SoT

on

VQ.

Hint for part (d): T is the isometric part of the polar decomposition of W. First
use (9E.3) to show that (W* W) Ui = U2(W*W). Then use the spectral theorem on
W* W to show that {W* wf^'^ U2 = U2(W* W)^^^\ and the conclusion then follows
fi-om parts (b) and (c).
9.4. Let {ak) and {bk) be two sequences indexed by Z, and define operators SQ and

keZ

and
keZ

(a) Show that the following (i) and (ii) are equivalent.
(i) The Cuntz relations
1

S*Sj = SijI

and

^SfS*

= I

hold,
(ii) The three separate identities
2^

akak-\-2i = So J,

2^

bkbk+21 = SQJ,

akbk+21 = 0

hold.
(b) Suppose (ak)keZ satisfies
2y^akak-{-2i

=Soj.

Setting bk = {\)^a\-k, show that the two sequences then satisfy the conditions
in(ii).
(c) Given two sequences {uk) and {bk), define

keZ

\ b2k b2k-\-\ J

Now show that U (z) is unitary for all z e T if and only if the conditions in (ii) are
satisfied.

202

9 Representation duality and multiresolutions

9.5. Consider a dyadic wavelet basis in I ^ (M) indexed as follows:

and let W\ FQ -^ i^ (Z) be the intertwining operator from Exercise 9.1. Finally, let
{ek)kEl. be the canonical basis in ^^, i.e., e^: (/) = ^A:,^ Prove that
W {y/j^k) = sl'^Sxek

for all y = 1 , 2 , . . . , and A: e Z.

9.6. The discrete wavelet transform

Let (ak)keZ ^^^ (f^k)keZ be sequences satisfying the orthogonality conditions in
Exercise 9.4(a). Let F = SQ and G = S* he the corresponding slanted matrices in
(7E.2), F defined from (a^), and G from (bk). For x G ^^ (Z), we say that
oo

x = Y,Y.''j\kS{)^iek

(9E.4)

7=0 A:eZ

is the discrete wavelet transform.

(a) Show that the wavelet coefiicients cj^k in (9E.4) are given by the following
matrix multiplication:
(9E.5)
/eZ

(b) Carry out the matrix products in (9E.5) in the four-tap case, i.e., when the two
sequences have only four non-zero terms, say ao, a\, a2, as and bo, b\, bz, b^.
(c) In the four-tap case (part (b)) show that each of the finite-dimensional subspaces Sk := span{e_A;,..., e-\,eo, e\,.,., ek). A: > 2, is invariant under matrix
multiplication with F and with G.
(d) With assumptions as in part (c), show that multiplication with F and G maps
SkJ\-\ into 5^ for k >2.
(e) In the four-tap case, and representing vectors in 52 as
/X-2\

I x-i
X = \ XQ

I ^1
\:^2/

show that y = Fx has the following form, matrix multiplication in the four-tap case,
i.e., matrix times column vector:
y-2 = a2X-2-\-a3X-u
y-i = aox-2 + a\x-\ + a2Xo + ^3X1,
yo = aoxo H- aixi + ^2x2,
yi = aoX2,
y2 = o.

Exercises

203

9.7. Let 7Y be a complex Hilbert space, and let So, Si be bounded operators in H.
For y e {0, 1, 2 , . . . } , set 7} := SQSI. NOW show that the following two systems of
conditions (i) and (ii) are equivalent.
(i) So, Si satisfy the Oi-Cuntz relations, rndSX"" -> 0.
(ii) {7} }^^j^ satisfy the C^oo-Cuntz relations, i.e.,
T;Tk = Sj,kI

forall7,A:eNo,

yeNo
P.S.: The assertions in (i) and (ii) about convergence refer to the strong operator topology (SOT) on B (H) = the bounded operators on H. Recall that the
neighborhoods of 0 in 5 (H) are generated by finite sets {hf} of vectors in H and
0 (SOT) if and only if
s eR^:{T
eB(n)\
\\Thi \\ <s}. Specifically, S*"" ^
lim \\S*''h\\

=OforMhen.

9.8. Kolmogorov
(a) Let n eN, and let Sn be some set of cardinality n. Let (i? (x, y)) be a fixed
positive definite n x n matrix, i.e., satisfying

xeSnyeSn

for all (^x)xes i^ ^"- Show that there is an ^7-dimensional Gaussian probability
distribution v^^' '"^ which has Rn as its covariance matrix, and with mean vector
zero.
(b) Let S be an infinite set, and let 7?: iS x ^S ^ C be a positive definite fimction,
i.e., satisfying
(finite sum)
^ ^ </ 7? {xj, Xk) <^k>0
J k
for all (fi, (^2,... e C and all xi, X2,... e S.
Prove that there is a probability space (Q, B, v) and a fiinction X: iS > Z^ (Q, v)
such that
' R(x,y) = E, {X(x)X(y))
forallx,;; G S,
andv(X(x)) = 0.
Hint: Ifn e N and Sn = {xi, X 2 , . . . , x} is a finite subset of iS with # (Sn) = n,
then use (a) to pick a Gaussian probability distribution v(^i' '^) of dimension n
which has (i? (xy, x^;)) as covariance matrix, and with mean vector zero.
(c) If iS' c iS'^ are two finite subsets of S, and v*^, v^ are the corresponding finite-dimensional Gaussian probability distributions fi'om (b), then show that they are
consistent in the sense of Kolmogorov; see Lemma 2.5.1.

204

9 Representation duality and multiresolutions

(d) Set Q:= C^ = all functions from S into C. For x e S, set KX (CO) := co (x),
CO e Q,. Show that there is a smallest cr-algebra B on Q for which all the fimctions
[TUX I X G iS'} are measurable.
(e) Use Kolmogorov's theorem to conclude that there is a unique probability
measure v on (Q, 8) such that for all finite subsets S, the marginal distribution of v
is the measure v*^from(c).
(f) Set X(x, co) := TTX (CO) for x e S, and co e Q. Then show that X(x, ) e
L^ (Q, v) and that this random process satisfies the conditions (9E.6) in (b).
(g) State a special case of the conclusion in (b), S = [0, 1], which is the converse
implication to that of the Karhunen-Loeve theorem of Exercise 2.7 (pp. 55-56).
9.9. The following is a corollary to the conclusion in Exercise 9.8. Give a direct and
geometric proof independently of Kohnogorov's measure-theoretic construction in
Exercise 9.8(b).
Let iS be a set, and let R:S x S ^> Che a positive definite function. Show that
there is a Hilbert space (H, ( | )) and a fimction X:S ^> H such that
R(x,y)={X(x) \X(y))

forallx^y e S.

(9E.7)

Hint: Set X(x) = R(x, ), viewing R(x, ) as a fimction on ^S for x e S. Then

use linear span, quotient, norm-completion, and (9E.7) to finish the argument.

References and remarks

In an earlier textbook [BrJo02b], we consistently lived with the notational convention mo(0) \/N for the frequency response function mo\ but I noticed when giving
lectures about this material to a mixed audience of students in mathematics and in engineering that the alternative convention mo(0) = 1 seems to be less confiising. The
idea is that the entire signal is admitted through by the low-pass filter at frequencies
near zero, i.e., low frequencies. And the other frequency components of the signal
are blocked by the mo filter. So in this book we have broken with the convention
from [BrJo02b].
Add to that the confusion about a different notational dichotomy, multiplicative
vs. additive notation for the torus groups: Frequency 0 in additive notation corresponds to the point z = 1 on T in multiplicative notation. This dichotomy is dictated
Funcby the choice of the alternative meanings of the manifolds T^,d = 1,2,
tions on T^ may be viewed alternately as periodic functions on E^, or equivalently
as functions on the compact torus T^.
Although the material in the last exercise. Exercise 9.8 above, serves to walk the
student through some main ideas of Kolmogorov, this material is also available in
more detail in a number of books on probability theory. Here we wish to especially
recommend the presentation [PaSc72] by Parthasarathy and Schmidt. But see also
Nelson's paper [IS[el59].

Appendices: Polyphase matrices and the

operator algebra O^

Born wanted a theory which would generalize these matrices or grids of

numbers into something with a continuity comparable to that of the continuous part of the spectrum. The job was a highly technical one, and he counted
on me for aid.... I had the generalization of matrices already at hand in the
form of what is known as operators. Born had a good many qualms about
the soundness of my method and kept wondering ifHilbert would approve of
my mathematics. Hilbertdid, in fact, approve of it, and operators have since
remained an essential part of quantum theory.
^Norbert Wiener

PREREQUISITES :

Curiosity about functions mapping into the group of unitary ma-

trices.

Prelude
The dialog between engineers and mathematicians is hampered by a substantial difference in jargon in the two fields. Often the same idea makes its appearance in the
two fields but under a different name. A case in point is the representation of a class
of algebras which in the operator-algebraic community is known as the Cuntz algebras, see [Cun77], but which in signal processing is known as a dual system of
subband filters yielding a perfect reconstruction in the signal output. The breaking
up of a signal (say a speech signal) into a finite number N of filtered and downsampled subbands is called analysis. There is then a dual operation of up-sampling
and filtering. When the result is then merged (added), we call it synthesis. When
N = 2, the use of filters (high-pass/low-pass) with a perfect reconstruction is called
quadrature-mirror (referring to the duality) filters. Even when N is larger than 2, we
may on occasion abuse notation and still talk about quadrature-mirror filters.

206

Appendices: Polyphase matrices and the operator algebra Oj^

So even within mathematics, there are separate themes; and analogously for engineering! And it does not become any less confiising by the huge variety of terminology and occasionally inconsistent jargon. In order to help organize the presentation
and the mathematics around engineering themes, we have subdivided the discussion
into four separate appendices.

fe|

Appendix A: Signals and filters

Our use in this book of signal-processing diagrams such as Figures 7.7 and 7.14
(pp. 124 and 132) summarizes this combined process (analysis and synthesis) very
nicely. While the diagrammatic viewpoint in fact originated in the early days of signal processing, the diagrams are much more versatile, and they are now used widely
in modem applications such as image processing as well. In our present discussion,
for simplicity, we have stressed signal processing. So in the simplest case, a signal (also called a time series) is represented by a sequence x = (x) of numbers,
with the index n denoting time. For the purpose of devising algorithms, it is natural
to introduce generating functions of a complex variable z, i.e., X{z) := ^XnZ^.
We have illustrated some wavelet algorithms based on this approach in Figures
7.8-7.13 (pp. 125-128). One of the advantages of this approach is that a variety
of analytic steps which on the face of it appear not to be especially computational
get rewritten in a finite sequence of relatively elementary matrix steps. As emphasized in Figures 7.13 and 7.14, when the operations of down- and up-sampling
are introduced into filtering models, then the matrices become slanted. Indeed, the
slanted nature of the matrices in turn is responsible for a speed-up of the computations, as reflected in the well-known sparseness of the wavelet matrices; see, e.g.,
[BrRo91, Coh03, Dau92, PaSW99, StNg96]. As a result, the complexity of the steps
in a computation is reduced fi'om rp- to nlogn.
In signal processing, it is convenient to restrict the complex variable z to the circle
T, i.e., setting z = exp(/a>), so that z" = exp(inco), and we think of cw as fi*equency,
or more generally as the dual variable. Of course, this means that X(z) becomes a
Fourier series. Engineers refer to X(z) as thefi-equencyresponse. If numbers (a) are
given, the Cauchy product (yn) of the two sequences (an) and (x) is called a filter, or
a filtering of the signal x, and the fimction m(z) := X ^ ^ " is called the frequencyresponse filter. We will often refer to m{z) itself as the filter, or more precisely the

Appendix A: Signals and

filters

207

filter fimction. The reader will easily verify that the associated fi*equency response
Y(z) of the filtered signal is then the pointwise product Y(z) = m(z)X(z). But we
shall here use duality much more generally than for the classical duality of time
and fi-equency in time series analysis. The general duality we have in mind is thus a
special case of Fourier duality; see, e.g., [Mal98, Waln02].
A main point in this book is to organize the diverse themes (both mathematics
and engineering) within the fi-amework of Hilbert-space geometry. While mathematicians tend to create Hilbert spaces of fimctions, engineers are concerned with
computations, and hence with transformations of sequences. The sequences could
be indexed by the integers for time, i.e., time series; or they could be organized as
matrices with double index (e.g., pixels) as is used for the grayscale numbers for an
exposure in a digital camera.
As for the mathematics, the next step is then to break the transformations in
sequence spaces into smaller algorithmic steps. That is where matrix tricks become
handy. One such matrix trick centers on what engineers call "the polyphase matrix;"
see Figures C.l and C.2 (pp. 214-215). For sequence space we suggest the Hilbert
spaces f^ but we allow flexibility for the indexing.
As we saw in Chapter 7, the selection for a particular problem of a resolution subspace Vo (and a generating function, a scaling fimction (p) within an ambient Hilbert
space of fimctions allows us to set up an isometric isomorphism between a chosen
VQ and the Hilbert space fi of sequences. Hence we are faced with a correspondence
principle. This has been organized into Table C.l.
In the simplest case, we may take X^ (M^) to be the ambient Hilbert space, and
the subspace VQ to be the closed linear span of the translates [q) ( n) \ n e Z^],
where Z^ denotes the usual rank-J lattice in R^. If there are finite positive constants
(frame constants) A and B such that
A< ^\^(t-\-k)f

<B

for all t e M^,

we may then define an operator (a frame operator) W: FQ -^ i^ (Z) =: l^ as follows:

W\

x^ ( - ) I := {xn)n^id .

The reader can check that W is bounded, and invertible with bounded inverse.
Theorem A.I. IfU\=
example

U^"^^ denotes a chosen scaling operator on L^ (R^), fi)r

Ul^^^A (x) := Idet^r^/^ / (A~^X\ ,

X G E^,

fi)r some expansive integral d x d matrix A, then itfi)llows that the frame operator
W: Vo -^ t^ intertwines U with the low-pass synthesis operator SQ ^ on ^; i.e., that

208

Appendices: Polyphase matrices and the operator algebra Oj^

Proof. The reader is encourged to work out the details of the argument in Exercises
A.l and A.2 below; see especially part (b) of Exercise A.2. However, note that the
exercises only sketch the outline of proof in the special case of a dyadic scaling operator. Nonetheless, one easily sees that the general case follows by the same outline
with only minor modifications.
n
Recall that S^ ^ is the first operator in a whole Cuntz system of isometries,

[sf^ |0<y < |det^|-l}.

This system (see details in the next appendix) in turn depends on an associated system of subband filters { W2y () | 0 < 7 < |det^| 1}, which are now Z^-periodic,
or equivalently fimctions on the J-torus T^ := E ^ / Z ^ .
While the number N from O}^, the Cuntz algebra, is simply the scaling in the
wavelet analysis of fimctions on the real line E, i.e., in one dimension, it is a Httle
more subtle in higher dimension, i.e., in R^.
There, a scaling is specified by a choice of a d-hy-d matrix ^ . To be usefiil
in wavelet bases, and in the study of scale-similarity, the matrix A mustfialfiUtwo
conditions: The entries of ^ must be integers; and all its eigenvalues X must satisfy
\M > 1. Then N := \detA\ yields the A^ in ON- We saw in Chapter 7 that this
number N is also the number of frequency bands which is needed in a subband-filter
approach to wavelet bases.
Strictly speaking, this is true only if we do wavelet bases on W^ itself But if we
are working on fractals, then the number of frequency bands will typically be strictly
smaller than |det v41.

Exercises
A.I. Let (pel?- (E) and suppose that the fimction

A:eZ

is well defined, where cp denotes the Fourier transform of ^. Suppose in addition that
there are finite positive constants A and B such that

A < ^\9it
keZ

holds.

+ k)f < B,

ER,

Exercises

209

(a) Then show that the ansatz

W
yields a well-defined linear operator on the closed linear span D^ of the translatedfiinction family {(p (- n) \ n eZ] in L'^ (M), and that the following estimate
holds:
I|2

A^\Xn\^
neZ

< ly^Xn^i'
lUeZ

-n)
ll2(M)

neZ

(b) Using the usual ordering for hermitian operators in Hilbert space, show that
the estimates in (a) may be rewritten in the following equivalent form:

B-^I <W*W < A'^I,

where / denotes the identity operator in the closed subspace X>^.
A.2. Let W, (p, and X>^ be as in the previous exercise, with the operator W: V(p -^
e'^ (Z) =: e^ defined as before.
Suppose in addition that a sequence (ak)keZ is given and fixed such that an associated operator S: fi -> ^^ is bounded, where

(Sx)n := V2y^a2k-nXkkeZ

On L^ (E), set U = U2, where

(f/2/)(O = - ^ / ( 0 ,

felHrn.

teR.

(a) If the fimction (p satisfies the scaling identity

(p (t) = 2^ak(p

(2t - k) ,

keZ

then show that the space V(p is invariant under the dyadic scaling operator U.
(b) Conditions as in (a): Show that the operator-intertwining relation
WU2 = SW
holds as an identity on the subspace P ^ .
A.3. With assumptions as in the previous two exercises, find a formula for the adjoint
operator ^*:^^ ^ D^.
Hint: Up to the Fourier dualityfi*omTable C.l, the answer is multiplication by
S-^ followed by (x) 1-^ XeZ ^^ (* " ")

210

Appendices: Polyphase matrices and the operator algebra Oj^

Appendix B: Hilbert space and systems of operators

Our use of Hilbert space throughout this book facilitates the breaking up of a computation or an algorithm into a set of basic algorithmic steps. In fact, it gives a geometric
meaning to these matrix factorizations, as stressed for example in [Jor03]. Hence a
signal X = (x) will be viewed as a vector in the Hilbert space H = ^(Z). But by
Fourier duality (in this case, Parseval's identity), we will identify H with the Hilbert
space of Z^ fimctions on T, and it will be clear from the context which meaning H
has. This is the geometric viewpoint used throughout the book, and in particular in
Figures 7.8-7.13 (pp. 125-128). When some signal x is broken up into a finite number, say N, of frequency bands, the result is then a vector in the A^-fold direct (or
orthogonal) sum ofTi with itself, denoted HNI or equivalently the result is a vectorvalued function on T. The reader easily checks that the Hilbert norm on H^f agrees
with the natural Hilbert norm on the space of all square-integrable fimctions from T
into the A^-dimensional Hilbert space
As illustrated in Figures 7.14 and 7.9 (pp. 132 and 126), this operator-theoretic
viewpoint can be accomplished with a set of TV operators, or subband filters. To take
advantage of matrix algebra, we shall view this operator system as a column matrix
of operators. As is familiar for scalar matrices, we then note that the Hilbert-space
adjoint of a column matrix of operators will be a row matrix of operators, the entries
in the row being simply the adjoint operators from the column. And we note that a
row operator is conveniently viewed as a single operator from HN into H.
When composing several algorithmic steps, we are faced with operators from HN
into itself, or with a composition of such operators. The introduction ofN-hy-N matrices with operator entries lets us represent composition of operators on HN as matrix multiplication, again facilitating computations. Moreover, the requirement of stability from signal processing leads us to favor unitary operators from HN to HN- But
if frequency-response filters are involved, we will (as noted above) be studying matrices of operators where the 7V^ individual entry operators are multiplication operators
in H. In the literature of signal-processing engineering (see [StNg96], [GaNSOla],
[GaNSOlb], [GaNS02]), these matrices of fimctions are cdillcd polyphase matrices
(as in the heading of this group of appendices). So a polyphase matrix is a function
F from T into the algebra of all N-hy-N complex matrices. We say that it is unitary
if this fimction F maps T into the group of N-by-N unitary matrices. (Note that
F(z) will only be unitary for z in T, and not for z in the complement of T within the
complex plane C.)
It is the purpose of this appendix to point out that the operator-algebraic framework of Cuntz algebras from Chapter 9 (see also [Cun77] and [BrJo02b]) is especially usefiil in making the link between the theory of individual filters in a system
of frequency bands, and operators on the system of vector fimctions which makes up
the band under discussion. And we will fiuther be stressing the use of this theory in
the analysis of the wavelet and fractal algorithms throughout from inside the book.

Exercises

211

As noted, the two figures 7.7 and 7.14 (pp. 124 and 132) illustrate the concept
of subband filtering from signal processing, but in the more general context adapted
here to wavelets and to fractals. In this operator-theoretic approach, we work out
algorithmic diagrams starting with N fimctions, say mo, mi, . . . , m^-i on T; and
we build a system of operators
Sj = Smj,

jeZN

= Z/NZ

^ {0, 1 , . . . , TV - 1}

satisfying the following relations:

S*Sk = Sj^kL

(B.l)

X^y^; = /.

(B.2)

J
Definition B.l. If we think of (B.1)-(B.2) as axioms for operators in Hilbert space
H, then we arrive at the familiar Cuntz relations ON [Cun77] from operator-algebra
theory; and so subband filtering yields a particular family of representations of ON\
see Table C. 1 in the next appendix.

Exercises
B.I. Let e N, > 2, be given, and let {St)^^^ be a system of operators in a Hilbert
space H which satisfies the Cuntz relations.
(a) Let u'.H -^ TihQdi unitary operator, and set
S',:=uSi.
Show that the system (S^)^^^ also satisfies the Cuntz relations.
(b) Let (Si)^^i be as above. Let (a/,y) e U (C) = the group of all n xn xmitary
matrices, and set
7=1

Show that the system {S^)^^^ also satisfies the Cuntz relations.
(c) Formulate and prove the implications demonstrating that the conditions in (a)
and (b) are also necessary.
B.2. Using the formula in Exercise B.l(b), show that the group U (C) acts as a
transformation group on the C*-algebra On, i.e., as a group of ^-automorphisms.

212

Appendices: Polyphase matrices and the operator algebra Oj^

Appendix C: A tale of two Hilbert spaces

We now turn to Table C.l, a systematic table which will serve as a graphic dictionary translating between current terminology in mathematics and jargon from signalprocessing engineering.
Explanation of Table C.l. The two columns in the table represent the same operators, but in different guises, that of time series, and that of frequency functions. Each
is important, and each has its advantages and its drawbacks. The sequences on the
left represent time series or signals (where time is the discrete integer index), while
the QSSociditQAfiinctions on the right carry frequency information.
Organization of the table: Li the left-hand-side column, we have sequences
and operations/operators on sequences; and then the corresponding fimctions and
operations/operators on fimctions in the right-hand-side column. Thus, moving down
the table, each of the listed operators has a representation in each of the two columns;
i.e., each operator has two incarnations (time and frequency), but they will often
be denoted with the same symbol. In addition to time/frequency duality, there is a
second duality, operator vs. adjoint operator. Specifically, each operator, say S, has
an adjoint operator S*. When S is listed in one line, then the adjoint S* follows in the
next line. Each of the two is important as each has distinct significance in the context
of signal processing: The adjoint of up-sampling is down-sampling, the adjoint of
right-shift is left-shift, and so on.
Comments on the adjoints: The adjoint of each operator depends on the inner
product of the Hilbert space in question. On the left and the right, respectively, it is
{x \y) = ^xnyn

and

{f\g)=

neZ

/ /(z)g(z),
^^

where the integration is with respect to the Haar measure on T. Note that
lim Q^^=

/fJHaar.

(C.l)

In fact, we may take (C.l) as a definition of the Haar measure on T. Specifically,

one checks independently that the limit in (C.l) exists for all (f e C (T). Introducing
rotations on T, (f^^ (z) = <f (zw) for z, lo e T, one checks that
lim Q^(f = lim Q^^u,,
which is the invariance characterizing Haar measure.
The operation of adjoint (from operator theory) reflects the word "mirror" in
quadrature-mirror filter from signal processing. Recall, taking the adjoint ofS twice
gets you back S itself, i.e., we have

S**=S,
as is immediate from {Sx \ y) = {x \ S*y).

Appendix C: A tale of two Hilbert spaces

Table C.l. Operations on two Hilbert spaces: The correspondence principle.
xei^

^ G L^ (T)

(Z)
Vectors in each space

frequency functions

time signals (xn)neZ

weZ
Filters
filter by m (z) = ^ anz^
yn = {Fmx) (n) = Y.^kXn-k
k

multiplication Mm
(MmO(z) = m(z)^(z)
Shift operators

shift to the right

(Sx)(n)=Xn-\,

multiplication by z
{M,Oiz)=z^{z),
ZGT

neZ

The adjoint of a shift operator is also a shift:

multiplication by z ^ = z
forz G T
( M ; ^ ) ( Z ) = 714-1^(2) = z - i ^ ( z )

shift to the left

{S*x) (n) = Xyi^i, n eZ

Sampling operations (by N)

fn = kN

((D^)() = r^ !fN\n

(00(z)=^(z^)

Down-sampling is the adjoint of up-sampling:

{x){n) =XnN, rI G Z
weJ
Subband-filter operators
(Smx) (n) =

VN^akXn-kN
k

Sm= y/NMm
(SmO(z) =
VNm(z)i{z^)
The adjoint operators:

(^*f)(z) = - ^

^^^^(w)

213

214

Appendices: Polyphase matrices and the operator algebra O^r

input of bands

output signal

Fig. C.l. Representation of 5' = TU*. This and the next diagram illustrate the use of the
operator relations from Appendix C in synthesis of subbands of a signal, or of a subdivision
of an image. The two diagrams form a pair. The present first one represents a matrix action
transforming one band-configuration into another. This is followed by up-sampling applied to
each strand. There is then a system of translations before the strands are merged and added.
This diagram is a natural dual version of the next one, Figure C.2, i.e., the analysis step which
begins with subdivision, or breaking apart an input signal.

We now resume the discussion of representations of OAT.

For our reference point, we shall use the particular representation {TJ) .^^
ON given by

{Tj^){z) =

of

zJi(2^)

forz T, 7 e { 0 , 1 , . . . , AT - 1}, and.f e L^ (f) ;

(C.2)

and the reader readily checks that the C^f-relations are satisfied, i.e., that
N-\

T*Tk^dj,kI

and

(C.3)

^TjT*^!
7=0

In this appendix, we show that there is an alternative approach based instead on

certain matrix-functions U = {Uj,k), ^^^ from T into the group of ail N x N
unitary matrices =: UN ( C ) .
Introducing operator-valued matrices and adjoints

I S*o \
S = iSo,...,SN-i),

S*^\

(C.4)

Appendix C: A tale of two Hilbert spaces

input signal

215

output of bands

Fig. C.2. Representation of ^S* = UT*. The present twofigurestogether, Figures C.l and
C.2, represent equivalent formulations of the respective left-hand and right-hand sides in Figure 7.7 (p. 124). But the present figures in fact represent signal-processing algorithms (or
equivalently wavelet algorithms) that are more versatile than the more traditional polyphase
matrices: specifically, the matrix functions in our present diagrams may be arbitrary, and be
of arbitrary size. The size of a polyphase matrix equals the number of subbands which is
used. But we also allow non-unitary matrix functions. Of course, more general choices of
matrices in Figures C.1-C.2 will then correspond to more general choices of filters in Figure 7.7. These choices are highly relevant, for example for the algorithms based on lifting,
see [DaSw98]. Our present Figures C.1-C.2 are multiband versions of diagrammatic representations of subsampling/subband-filtering algorithms which appear frequently in the signalprocessing engineering literature; see for example [WWWl] with text, and [WWW2] and
[WWW3] with pictures.

and similarly for (7}) .^^ , we arrive at the following matrix formulas:
S = TU*

(C.5)

S* = UT*.

(C.6)

and

In signal processing, the representation of these two operator identities takes the
diagrammatic form shown in Figures C.l and C.2, and the operator matrix U is
called the polyphase matrix.
Specifically, in terms of processing diagrams the operator-theoretic formulas
(C.5) and (C.6) turn into Figures C.l and C.2.
Remark C.l. Our use here of the operator notation S and S* is reversed from the
convention in signal processing. (Conventions: engineers typically set F = S* and

216

Appendices: Polyphase matrices and the operator algebra Ojsf

F* = S.) Our choice of terminology is motivated by the traditions in operatoralgebra theory, and by the fact that the individual operators Sj will then be isometrics,
while the Sys are co-isometries. Specifically, S*Sj = I for J = 0 , . . . , TV 1; and
each operator Sj S* = Ej is a projection. Moreover, the projections [Ej). ^ satisfy
EjEk = Sj,kEj

(orthogonality),

(C.7)

and

EXPLANATION OF PROGRAMMING DIAGRAMS FROM SIGNAL PROCESSING

AND THE CORRESPONDING COMPOSITION OF OPERATORS IN HILBERT SPACE:
GUIDE TO FIGURES C.l AND C . 2 . Caution: in general note that we read figures
with input-output flow diagrams from left to right, and this corresponds to an action
of operators, but now reading the action of the respective operators from right to
left. So the first figure. Figure C.l, represents the following operator factorization:
S = TU*. The direction from the left to the right in a diagram corresponds to the
flow of the operations performed on signals. Specifically, the diagram in Figure C.l
begins with input of bands into a box for U* on the left in the figure. This is followed
by the diagram for the operator T which involves bands that are assembled into a single output signal. In contrast, the second picture. Figure C.2, is the result of taking
the adjoint of the operators from Figure C.l. So it corresponds to the operator factorization of the adjoint operator S* = UT*, and it begins with the diagram for the
operator T* on the left in Figure C.2. The diagram for T* is then followed by a box
placed to the right-hand side of T* representing the matrix U. The result of matrix
multiplication by C/ is a vector output representing a system of banded signals. These
output signals may then in turn be the input in a new operation representing a new
and different matrix. The various matrices are also called "gates" in programming.
The unitary matrices correspond to gates which preserve the energy of signals from
input to output.

The next appendix is largely concerned with Figures C.1-C.2, their use, their
interpretation, and their relationship to both the Cuntz formulas (B.1)-(B.2) and the
polyphase-matrix formalism from engineering.

Exercises

217

Exercises
C.l. Let n e N,n > 2,hQ given and let iSi,..., ^S^ be a system of bounded operators in a Hilbert space H. Then show that the following two a priori estimates are
equivalent:

I n

11

|/=i
\Y,Sim

II <^\\^^if
/=i

n
forallw/GW;

(CE.l)

and
^ l l ^ S / u f < \\vf

for all?; e W.

(CE.l)

/=1

Note: If the estimate (CE.l) holds, then we say that (S\,...,

contraction.

Sn) is a row-

C.2. Let e N, w > 2, be given, and consider a family of 1-periodic fimctions

m 1 , . . . , m on R which are assumed in Z ^ . On the Hilbert space W = Z,2 (0,1) we
introduce the system
(Sif) (x) := Vlrrn (x)f(2x)

x e [0, 1) ,

where x i-^ 2x is understood mod 1.

(a) Then write a set of necessary and sufficient conditions on the function system
(^/)/Li f^^ (*^i ? ? ^n) to be a row-contraction.
(b) Spell out how in the case n = 2 the conditions in (a) generalize the
quadrature-mirror filter condition, i.e., the assumption that the matrix
mi (x)

mi (x-\- 5J

m2(x)

m2\x-^^j

is in U2 (C) for a.a. x, i.e., is a unitary 2 x 2 complex matrix.

C.3. Verify the assertions made in Table C. 1 about adjoint operators, i.e., the formulas for S^, , , and the formula = *.
C.4. Using the identification of Hilbert spaces from Table C.l, ^^ (Z) = L^ (T) =
L^ (0, 1) = L^ (M/Z), return to the operator
W: Vcp -> f
from Exercises A. 1-A.2; same assumptions.
(a) Then show that when M := ^ ^ * is realized as an operator in Z^ ( R / Z ) ,
then it is a multiplication operator, i.e., multiplication by a 1-periodic fimction on M.
(b) Show that M = WW* from (a) is multiplication by the fimction
IR3/l-> ^ \ ^ (t-\-k)]^ .

218

Appendices: Polyphase matrices and the operator algebra Ojq

Appendix D: Signal processing, matrices, and programming

diagrams
We now turn to the Hilbert-space geometry of the polyphase matrices. If (Si)j^iN
and (Ti)j^i^ is SL pair of representations of ON in a Hilbert space H, i.e., both
satisfying (B.1)-(B.2), then we show that the operator matrix
U := {S*Tj)

(D.l)

is unitary, i.e., U isa. unitary operator in the Hilbert space

HN

:=ne"'n.
V

'

'

N times

The convention in (D. 1) is that UisanNxN

matrix with entries in the algebra B (H)
of all bounded linear operators in 7i, i.e., that Uij := S*Tj for (/, j) G ZAT X XN.
The formulas (C.5)-(C.6) are matrix products; specifically,
Sj=Y.^kUlk^

7E{0,...,7V-1};

(D.2)

7e{0,...,7V-l}.

(D.3)

and
^y^S^M^yT'
k

But if U is unitary, we also get T = SU\ or, in matrix form,

Tj=Y.^kUkj.

(D.4)

keeping in mind that the products on the right-hand side refer to products in the nonabelian algebra B (H) of all bounded linear operators on H. Note further that the
matrix representation of formula (D.l) now takes the form
U = S*T,

(D.5)

where S* is a column matrix of operators, while T is a row matrix.

Proposition D.l. Let (Si)j^^^ be a representation of ON in a Hilbert space TC, and
let U be a bounded operator in the Hilbert space HN = H 0 0 7i. Writing S,
N times

as before, and T = (To,..., TN-I) as row matrices of operators, we claim that the
following two conditions are equivalent:
(a) T SU is a representation of ON, and
(b) U = S*T is unitary in B (HNI

Appendix D: Signal processing, matrices, and programming diagrams

219

Proof, (a) => (b). Now the assumption is that both S and T are representations of
ON- TO verify (b), we calculate the (/, j) matrix entry of each of the matrix products
f/*^ and UU\ Specifically,

(^*^)/j = Y.T^:SkSiTj = T;TJ = s,ji

k

and
k

showing that U is unitary.

(b) ^ (a). If (b) holds, we now show that the matrix product T = SU is a
representation of ON- Specifically, we get the two ON relations:

^*^/ =mi^Ki
k

SIS, Uij =Y^Ul,Uk,j = SijI

V^

Ok,l

and

= X^ks;
k

= L
We now return to the setting of signal processing. In this application, we have
n = L^ (T) = ^ (Z), and the two operator systems S and T will be as outlined in
the formulas

Sj=Smj
or

=^/NMmj,

{Sj^)iz)^VNmjiz)^(z''),

in Table C.l, and in (C.2) for Tj. The representation (^Sj) will be called a subband
representation, and (Tj) will be called a base-point representation.
Corollary D.2. Let (mj) .^^ be a family of bounded measurable functions on T,
and let
z G T, (f e Z^ (T) , 7 = 0, 1 , . . . , TV - 1
(D.6)
be the corresponding operators in H = L^ (T).
Then {Sj) ^^ defines a representation of OM in H if and only if the N'^
functions
{Sji) (z) := VNmj (z) f ( z ^ ) ,

220

Appendices: Polyphase matrices and the operator algebra Oj^^

Uij (z): = - ^ 21 "^^^ "^''

^ ^ '^'

(D.7)

define a unitary matrix for all z e T, i.e., z \-^ {Utj (z)) defines afidnctionfi-omT
into the Lie group UAT ( C ) of all unitary N x N matrices.
Moreover, in that case,
mi (z) = = y]zJ

Ujj (z^) ,

z G T, / e ZTV.

(D.8)

Proof. The result is immediate from Proposition D. 1 once the two formulas (D.7)
and (D.8) have been established, subject to the stated assumptions.
Verification of (D.7): We have

{S*Tj) f (z) = - L 21 "^^(^ (^>^) (^)

= -7= X ^/(^)^'^^(^)
^

w^=z

= Uij{z)i{z)

forzGT,

proving (D.7).
Verification of (D.8): By Proposition D.l, we have T = SU, or
T^ (z) = (SUO (z) = U ( z ^ ) S^ (z),
for z G T, so (using unitarity) we get
(50(z) = C/(z^)*(rO(z),
or
J

= VNmi (z)^(z^)
proving the result.

forze T,

References and remarks: Systems theory

221

Exercises
D.l. Let W be a Hilbert space, and let e N, > 2, be given. Suppose a system of
H^-'-^H.
isometries Su ..., Sn satisfies the Cuntz relations in H. Set H^^^ :=
n times

(a) Then show that

B(J-L)3X\%

{S*XSj) e B {T&A

defines an isomorphism of the respective *-algebras of all bounded linear operators.

{Note: By *-isomorphism, we mean that a is bijective, and fiirther satisfies
a {XY) = a{X)a (Y) and a (X*) = a (X)* for all X, 7 G 0 (W).)
(b) Find the inverse
P = a-^:B{H^''^^ -^

B{n).

Hint: liA = {Afj) e B (W^")), show that

References and remarks: Systems theory

The reader will notice that our formulation in Proposition D. 1 above is "pure" Hilbert
space geometry: it is about linear operators in abstract Hilbert space, i.e., it is operator
theory. The proposition shows that every system of isometries {Si) in a Hilbert space
H subject to the Cuntz relations may be rewritten as a single unitary operator U in
a direct-sum Hilbert space, i.e., as a unitary operator matrix. By this we mean that
the matrix-entries in U are operators in H. Then in Corollary D.2 we specialize to
the case when the Cuntz system is of the kind that is used in subband filtering. In
this case we note that the corresponding unitaty operator matrix U then has entries
which are multiplication operators. So these are special operator matrices, so-called
polyphase matrices.
But it is worth stressing that the use of systems of isometries {Si) subject to
the Cuntz relations is ubiquitous in the analysis of problems involving some kind
of branching, and a feature involving scales, or a notion of similarity which can be
couched in terms of operator theory. A case in point is the formulation of scattering
theory due to Ralph Phillips and Peter Lax. hi that theory, there is a single unitary
operator, or a one-parameter group of unitary operators; the parameter is for time.
The isometries in the theory then arise by restriction to a chosen pair of subspaces
which are invariant forward in time and backward in time, respectively: so-called incoming and outgoing subspaces. As it turns out, the scattering then "takes place" in

222

Appendices: Polyphase matrices and the operator algebra Ojsi

the ortho-complement of these subspaces. This is "obstacle scattering:" the obstacle

is modeled on the ortho-complement of these subspaces. It turns out that the subspaces in the Lax-Phillips theory are examples of the kind of multiresolutions that
we have emphasized in the present book.
A number of authors (J.A. Ball, S. Cora, W. Helton, V. Vinnikov, and others;
see, e.g., [BaCV05] and [BaVi05]) have extended Lax-Phillips scattering to systems
theory, a branch of engineering. Our discussion of the dual pairs of Hilbert spaces in
Table C.l applies to this setting as well. Further, it is interesting to note that the family of isometrics used in systems theory may also be modeled on the Cuntz relations,
i.e., that they yield different classes of representations of the Cuntz algebra. Hence
the general framework of the present appendices applies to this branch of systems
theory as well.
It should be noted that our operator systems from Exercises C.l and C.2 arise
in a number of places in mathematics, and have already found many uses. Here we
mention only their use in operator theory as row-contractions (see, e.g., [Arv04],
[Pop89]) and in wavelet theory (see, e.g., [RoSh97], [She98], and the papers cited
there). In the context of [RoSh97] these authors have the functions and the operators
from Exercise C.2 as part of what they call the Unitary Extension Principle (UEP).
As demonstrated in [RoSh97], the types of generalized wavelets which may be generated via use of the UEP have a number of desirable properties.

Afterword

Comments on signal/image processing terminology

Introduction
A Google search on the word "wavelet" yields 2,650,000 entries, "fractal" 8,530,000,
and "signal processing" 45,000,000, as of September 2005. While this may seem
staggering, it is wise to keep in mind that by now general ideas in the subject which
do originate with wavelets are presently used in many sciences outside of mathematics. While the general subject may be said to have a mathematical core, the vast
number of applications and independent tracks followed in the last two decades have
lead to an amazing diversity, even when counting only the mathematical trends.
This diversity of ideas is both exciting, but at the same time, it often leaves students confiised. Ideas which started with wavelets turned out at the present time to be
significant elsewhere, for example in computational mathematics, in engineering and
in computer science; in part because of their algorithmic efiiciency. Another reason
is the prevalence of time-space scale similarity in both science (e.g., in physics of
large scale and physics at the quantum level) and in engineering. While the notion of
scale-similarity is not always made mathematically precise when it is used outside of
mathematics (for example in finance!), it is crystallized especially nicely in wavelet
theory through the notion of a "multiresolution," a concept from optics. This scale
similarity may explain why wavelets keep getting reinvented outside mathematics.
While the mathematical concept of wavelet dates back (in a special case) almost
one hundred years, to Alfred Haar (mathematics) and Oliver Heaviside (signal analysis), the potential of the interaction between mathematics and signal processing was
only realized much later. It is only the scientific and engineering developments since
the mid-1980s that have accounted for the impressive growth which is now reflected
in the gigantic totals we find in typical Google searches.
There are at least three reasons for the resurgence of activity since 1980: one
comes from signal processing in engineering (e.g., wireless communication and data

224

Afterword

mining); another from image processing (e.g., digital cameras, JPEG 2000, finger
prints) and medical imaging (vision, optics and more); and a third from statistics and
telecommunication (quantization, filtering of noisy signals). While the mathematics
of wavelets continues to be both very active and exciting, it is probably all the interdisciplinary connections and developments which account for the bulk of the activity
in the general area.
While interdisciplinary developments are wonderfiil and often take place along
parallel tracks, they have a tendency of making it difiicult for students, for users such
as engineers and computer scientists, and for researchers to get started in the subject,
and to get a good sense of the main trends. They are busy and look for friendly
introductions to a particular main trend, and each trend has its own scientific culture.
There are several reasons for why it may now be hard for a student to know where
to begin: The different disciplines have vastly differing lingo, probably reflecting the
fact that they have quite different aims.
For the mathematician who is used to seeing wavelets as a part of fijnction theory,
it may be hard to accept that many users in digital signal/image processing and its
applications are mainly interested in numbers and in algorithms. Because of the "cultural" differences between the diverse subjects, the connection between the fimction
theory on one side and the practical algorithms which process numbers on the other
is not always transparent.
Similarly, we noticed that colleagues from engineering (computer and electrical
engineering, industrial engineering, data mining, etc.) find the version of wavelet
theory in typical mathematics books on fimction theory rather forbidding.
While all the diversity of applications is truly impressive, the student from mathematics or from one of the other disciplines involved often has difiiculties in getting
started. This difiiculty is not because of a shortage of books, but rather because of an
apparent divergence of the various trends, developments and applications.
hi this book we have aimed at remedying some of this. To make the book usefiil
for diverse audiences, we have approached the task from two sides, and we have
added a discussion of some key terms below.
On the one hand, each chapter (including exercises) has elements of tutorial, and
a practical side as well. On the other hand, we have included separate sections which
serve to "translate" between the lingo which is used in the different, and sometimes
disparate disciplines, hi this endeavor we make an effort to both explain concepts
from engineering and statistics to mathematicians, and also to explain the various
areas of mathematics to engineers and scientists from the neighboring disciplines.
What follows are two small sections dealing with how ideas and terminology
are used differently in mathematics and in its applications: First we discuss technical
terms from inside the book and stress how they are used quite differently in engineering and in (relatively) pure mathematics. (While perhaps confiising, this dichotomy
is hardly surprising since the two conmiunities have quite different aims.) Mindfiil
that technical lingo is viewed and used differently by engineers and mathematicians.

Comments on signal/image processing terminology

225

we have tried to present the concepts and ideas evenhandedly from the perspectives
of the two communities.
Secondly, we briefly hint at the use of some central themes from the text in
"Computational mathematics." This expands on our initial presentation of related
ideas in the Glossary (a systematic discussion of terminology in the front matter
above, pp. xvii-xxv). Both this "Computational mathematics" section and the Glossary serve to motivate central ideas used in the text.
The Glossary is arranged as a table. In it each underlying idea has up to four
incarnations, hence the different terms! We outline item by item how each concept
is used in up to four related but different ways; in four contexts: in mathematics,
in probability, in engineering, and in physics. We hope to clarify (a little) how the
distinct aims of the four subjects are reflected in a variety of ways, for example in
the fact that there is often a multitude of names for what is essentially the same term.
JPEG 2000 vs. GIF
We begin with two engineering terms of a more recent vintage; both are from imageprocessing. They help us to strike a contrast between two themes in our book, wavelets vs. Fourier. Specifically, there are some practical features unique to A-to-D quantization (analog-to-digital) for wavelets (e.g., localization, algorithms, computability,
and resolutions) which are absent in analogous Fourier tools. For much more detailed
discussions, see, e.g., [JPEGOO], [JaMROl] and [MeyOS]. There is a huge engineering literature on this, and we only scratch the surface.
JPEG 2000
(Joint Photographies Experts Group)^A new image coding which is largely wavelet
based. Applications include digital cameras, remote sensing, image archives, scanning, and medical imaging.
It comes in different parts: Part 1 offers both lossless and lossy compression and
provides much better image quality at smaller file-sizes. Part 6 is aimed at compressing scarmed color documents containing both bi-tonal elements as well as images.
Fundamentally, JPEG 2000 is built much like a certain AJ computer-vision
model, following closely the biology of "human vision." See [Mar82]: Eye focus,
zoom, and scales of detail organized as visual resolutions.
hi engineering terms, image-input onto pixels contains information-detail at all
scales; the detail levels separate the resolution scales, and this hierarchical structure
can be digitized. Analogously, the human eye (as well as digital cameras) processes
intensity changes at a variety of resolution scales, much like in the pyramid algorithm
used in wavelets (where "intensity changes" are wavelet coefficients). Hence, images
are digitized taking advantage of scale-similarity in recursive processing, and using
iterated matrix powers, i.e., integral powers of the slanted wavelet matrices outlined

226

Afterword

in Chapter 7. In this form, the (wavelet) algorithm is thus a recursive matrix algorithm, and it takes advantage of a key feature of wavelet-scales: each degree of scale
is a level in the corresponding combinatorial pyramid. Further, it is localized, so it
does not make the redundant infinite additions that are notorious in Fourier approximations.
JPEG is commonly known as a format for compressing, digitizing and storing
color images. It supports a number of colors, in the millions: and it can reduce file
sizes to about 5% of their original size with only small amount of lost data, i.e.,
"small" relative to detection by the human eye. The latest JPEG2000 promises to
compress images 200 times with better resulting quality than earlier versions based
on Fourier waves. The key technology enabling such improvement involves switching to wavelets, awayfi*omthe earlier Fourier waves that were used in discrete cosine
transform (DCT).
From [JPEGOO]:
The coder is essentially a bit-plane coder, using the same Layered Zero Coding (LZC) techniques which have been employed in a nimiber of embedded
Wavelet coders. Key additions are:
The use offi*actionalbit-planes, in which the quantization symbols for
any given quantization layer (or bit-plane) are coded in a succession of
separate passes, rather than just one pass.
A simple embedded quad-tree algorithm is used to identify whether or
not each of a collection of "subblocks" contains any non-zero (significant) samples at each quantization layer, so that the encoding and decoding algorithms need only visit those samples which lie within subblocks
which are known to have significant samples.
GIF
(Graphics Interchange Format)^Largely "Fourier based."
A compressed file format used for storing and transmitting color graphics. The
GIF model is older, and it is more widely used (up to now!), but at the same time it
is more restricted in capabilities than JPEG 2000; more limited with respect to the
number of colors (a few hundred) and compression degree.
With the latest advances with wavelets in digital image processing I imagine new
titles in the next generation of horror films, such as:

He Came at Breakfast Time and Left at 8:30 p. m.

Marooned in Space with a Digital Camera Fanatic
Digital Photographers from Deneb
Revolt of the Memory Sticks
Countdown to jpeg
Digital Downloads/Uploads from Hell

Comments on signal/image processing terminology

227

Grayscale
This is a notion from digital image analysis. In image processing, for a given image,
a value is assigned to each subdivision square (called "pixel"). It is a certain sample
number: A single number for each pixel!
In general, the displayed images themselves are typically composed of shades
of gray, varying from black at the weakest intensity to white at the strongest. For
color images, each of the three primary colors is assigned pixel intensities in the
same way. This amounts to an encoding of various colors with different intensities.
A mixing then becomes part of the digital processing. Grayscale images are distinct
from black-and-white images, which in the context of computer imaging are images
with only two colors; grayscale images have many shades of gray in between.
Quadrature-mirror filter
"Quadrature-mirror filter" is engineering jargon, and signal-processing engineers
aren't exactly poets. [From the online Oxford English Dictionary [OED]: "quadrature, (a). Math- The action or process of squaring; spec, the expression of an area
bounded by a curve, esp. a circle, by means of an equivalent square. More widely, the
calculation of the area bounded by, or lying under, a curve, (b). 1942 H. M. BACON
-Differential & Integral Calculus: The desire was to find a square equal in area to the
area bounded by the given curve (in this case, a circle). For this reason the problem
has been called the problem of quadrature."]
Signal-processing engineers noticed early in the subject that the conditions imposed on each of the two filter fimctions, say mo and mi that are used in the analysis/synthesis of speech signals into two subbands (see Figure 7.14, p. 132) involve
sums of squares. Similarly the fimctions themselves mo and mi satisfy a sum-square
rule; or you could say, a circle-rule.
These quadratic conditions are summarized in the text around (9.2.12). Of course
this is also what allows us to get the unitary matrix fimctions (of the Appendices) into
the game. Or equivalently, the representations of the Cuntz relations from Chapter 9
serve to create a geometric framework for the popular pyramid algorithms of wavelet/fractal subdivision algorithms.
The essential conclusion in the Appendices above is that the so-called "quadrature conditions" from signal processing, or equivalently from wavelet filters, take a
simple equivalent form in terms of unitary matrices. See also [Jor03]. Each instance
of a sum-of-squares rule suggests a circle law, and that is perhaps the root of this engineering terminology. Actually for the engineers, it is the finite set of numbers (also
called masking, or wavelet coefficients) that go into the filters that are of more immediate interest, and the sum-of-squares rule for the filter fimctions merely reflects
a quadratic system of equations for the wavelet coefficients, see (1.3.20). They are
the frequency response to the system (1.3.20). The related condition (1.3.21) refers

228

Afterword

to the low-pass condition which characterizes the first signal-processing filter in a

subband system. For example the four numbers (alias, the four taps) that define the
Daubechies wavelet (see Chapter 1, Example 1.3.3, and (7.6.6)) result as one of the
solutions to this quadratic system of equations.
Additional discussion of the engineering term "quadrature-mirror filter" and its
relation to "mirror" can be found on the website [WWW4]. It explains the quadrature
part differently; i.e., in the following sense of "quadrature:" It is thefi-equencypoint
;r/2, the "quadrature point" that divides the circle into four parts.
What is a frame?
To the mathematics student:
It is a generalization of an orthonormal basis (ONB)! Let (Z?/), / in some index set,
be an orthonormal basis for a Hilbert space. To understand the generalization, begin
with Parseval's familiar identity for an ONB:

Wxf = Xi<*^i^>i'-

More generally, if a system of vectors (Z?/) satisfies (*), it is said to form a tight
frame, or a Parsevalframe. But it need not be an ONB: You might have \\bi\\ < 1,
i.e., the norm is "too small." And redundancy in the system will then smear out
orthogonality.
Or it could be even worse: If instead of an identity in (*), you only have estimates,
with two positive constants A, and B\ The term A \\x |p as a lower bound, and B \\x |p
as an upper bound for X / K ^^ I ^ )Pj then you talk about a general^rawe with A
and B as frame bounds.
It is easy to rewrite all of this in terms of dilations from operator theory: Frames
{bi) in a Hilbert space H are all of the form hi = T (c/) where (c/) is an ONB in an
expanded Hilbert space /C, and T:JC -^ H a bounded invertible linear operator.
The Parseval frames, for example, correspond to the case when this operator T
is the projection of/C onto H.
So the notion of a "frame" arises from relaxing the standard requirement on a
"basis." While at first this particular term "frame" might appear somewhat as a mathematical technicality or a mere curiosity, in fact we have seen in the past decade that
the subject "frames" emerging in leaps and bounds as a substantial mathematical discipline in its own right: one with a real presence both in the book literature (see, e.g.,
[Chr03] and the papers cited there) and in research journals. Mathematically there are
good reasons to relax the more stringent axioms that have previously been used for
bases in infinite-dimensional fimction spaces; and much of this has been motivated
by needs dictated by wavelets. Other motivations are drawn from engineering.

Comments on signal/image processing terminology

229

It turns out that it is much easier to generate "wavelet bases" which are only
frames, as opposed to orthonormal bases (ONBs). Moreover, a number of wavelet
problems in numerical analysis rely on easy and readily available libraries of wavelet bases. For this, the "library" of ONB wavelets is too small! Some of the frame
wavelets have been known in applied mathematics for a generation or more under
the name of "splines;" see [Chr03] for details. Add to this the fact that now, both
in wavelet mathematics and in telecommunications engineering, it has proved practical to give up the stringent and restrictive axiom of ONBs. What is needed is a
notion of "basis" which accommodates some degree of redundancy in the process
of synthesis. (By synthesis we mean the expansion of a vector in a (frame) "basis;"
i.e., reconstructing the vector from its base coeflScients.) Our Chapter 6 above illustrates this point by example, and for the simplest wavelet of them all. The issue in its
general form is closely tied into the role played by scale-similarity.
To an engineer:
Frames are used by signal-processing engineers, and they are motivated by practical
concerns regarding transmission and measurement in telecommunication. Many of
the non-trivial engineering applications and uses take place in finite dimensions.
The mmQ frame has to do with measurements within a visual "frame," referring
to an instrument. For mathematical reasons, we know that an "honest" basis, or an
ONB, would escape "out o f the frame!
Alias (aliasing)
Engineering:
(1) hi signal processing, the effect that causes different continuous signals to create multiplicities or to become indistinguishable (or aliases of one another) when
sampled.
(2) In computing, the indirect (usually unexpected) effect on other data when a
variable or reference is changed; typically referring to multiplicities of the original
data.
Mathematics:
This notion of "aliasing" from signal processing is also current in such areas of mathematics as operator theory and harmonic analysis. It is used in coimection with problems in a Hilbert space, say TL, when a super-structure is constructed in the form
of an ambient dilated Hilbert space /C, and an isometry embedding 7i into /C. Then
aliasing is seen as the effect on an initial signal resulting from oversampling. The
signal could be a vector in 7Y, or a frame basis for L^ i

230

Afterword

We introduced the term sampling in Chapter 1 above. The sampling could be

done with a selection of points on a lattice in M^ (i.e., a r?ivk-d discrete subgroup in
E^), and oversampling then refers to sampling points in a bigger lattice.
When oversampling is applied to a continuous signal (realized in Ti)^ it creates
one or more different signals (aliases, or multiplicities) indistinguishable from the
original signal; and it is then possible to realize the aliases in the dilated Hilbert space
/C. Hence some of our results in Chapter 6 and in [DuJo06c] make aliasing precise
via a super-structure. This is a dilated Hilbert space, and it arises, for example, in the
context of afiine wavelet frames in Z^ (R^).
Li this setting, there is a procedure in which sampling in an initial frame for
L^ (M^) yields a dilated Hilbert space /C and a super-frame in /C. Thus, in a sense
"nothing is lost" in oversampling: The oversampled frame is a comer of a third frame,
a super-structure, or an associated super-frame. The sampling setting in [DuJo06c]
admits a concrete representation of the Hilbert space /C as a direct sum of a finite
number of copies of L?- (M^), the number of copies depending on the amount of
oversampling.

Computational mathematics
On a personal note, after teaching some of the present material in courses, this author
has come to appreciate the usefiilness of Hilbert-space geometry and operator theory
in addressing even such practical problems of calculating wavelet coefiicients with
iterative and fast matrix multiplication algorithms.
In fact, I have become much more optimistic than I used to be about the practical
and the computational usefiilness of relying on even singular wavelet filters of one
kind or the other, in computation.
One of the things coming out from Chapter 7, the Appendices above, from my
teaching of wavelets, and from my work with engineers is the usefiilness of Hilbertspace geometry for iterative algorithms. Thus, once we agree to stay in a fixed resolution subspace in Z^ (R"), then we can do wavelet expansions without ever having
to calculate wavelet coefiicients the slow way, i.e., by integrating over R", or over
a subset of it. The trick is to compute wavelet coefiicients instead with the use of
suitable matrix iterations, and then using slanted matrices (see Figure 7.13, p. 128)
which are computationally much faster. For more detail, we refer the reader to the
series of figures in Chapter 7, see especially Figures 7.7, 7.8, and 7.9 (pp. 124-126),
and Figure 7.13. At first, I was not sure that it is possible to avoid integration, but
it is!
In concrete cases of image processing, we note that all the pictures you come
across on the web showing iterative image processing of Lena or of some other digital
image (see, e.g., [WWW3]) are done precisely with the standard matrix iteration
algorithm for wavelets. "Computers can't integrate!"

Computational mathematics

231

What further emerges by a closer scrutiny of these iterative algorithms is that

Cuntz (or Cuntz-like) relations are just perfect for this; see the discussion in the
Appendices above, especially (B.1)-{B.2) and Table C.l. Of course, with hindsight,
we now know that these relations have their root in old ideas from signal processing;
see [Jor03].
The only technical modification needed for all of these iterative schemes to work
for such singular filters that are used in fractals of one kind or the other (Chapters
7-8) is that the slanted matrices might now become infinite, but their slanted nature
makes this infinity only a mild nuisance. Engineers are good at devising threshold
schemes for truncating the infinite, as long as it is under control; and as we saw, it is
for this purpose.

Epigraphs

Quotation from John von Neumann on page vii above the Preface: This quote is popular on
web pages about von Neumann, and about computing and mathematics generally. It is
apparently not from a published work of von Neumann's, but Franz L. Alt recalls it as
a remark made from the podium by von Neumaim as keynote speaker at the first national meeting of the Association for Computing Machinery in 1947. The exchange at
that meeting is described at the end of Alt's brief article Archaeology of computers:
Reminiscences, 1945-1947, Communications of the ACM, vol. 15, issue 7, July 1972,
special issue: Twenty-fifth anniversary of the Association for Computing Machinery, p.
694. Alt recalls that von Neumann "mentioned the 'new programming method' for ENIAC
and explained that its seemingly small vocabulary was in fact ample: that future computers, then in the design stage, would get along on a dozen instruction types, and this was
known to be adequate for expressing all of mathematics.... Von Neumann went on to say
that one need not be surprised at this small number, since about 1,000 words were known
to be adequate for most situations of real life, and mathematics was only a small part of
life, and a very simple part at that. This caused some hilarity in the audience, which provoked von Neumann to say: Tf people do not believe that mathematics is simple, it is only
because they do not realize how complicated life is.'"
Quotation from David Mumford on page xv: David Mumford, The dawning of the age
of stochasticity, Atti della Accademia Nazionale dei Lincei, Classe di Scienze Fisiche,
Matematiche e Naturali, Rendiconti Lincei, Serie IX, Matematica e Applicazioni, vol. XI,
2000, special issue: Mathematics Towards the Third Millennium (Papers from the International Conference held in Rome, May 27-29, 1999), p. 107. This is quoted from the
beginning of Mumford's article, which offers a fascinating and refreshing view on thinking as Bayesian inference, and the use of probability spaces in image processing.
In the "Getting started" section of the front matter, on page xv, we "sort of quote" from
G.H. Hardy, A Mathematician's Apology, Canto, Cambridge University Press, Cambridge,
1992, with a foreword by C.P. Snow (first edition 1940, and latest edition 1992), in which
sixty-six years ago Hardy so eloquently apologized to the World for mathematics. Back
then Hardy, the Platonic puritan he was, had in mind pure mathematics: at the time, some
parts of applied mathematics had been used in an unpopular war. In my opening paragraph

234

Epigraphs
of the present "Getting started" section, I could not help wondering if in the meantime
the winds could have changed; wondering whether perhaps now a mathematics author
who trespasses into engineering topics and other applied domains might not be expected
to apologize; at least if he/she has in mind mathematics students as his primary audience. Aside from this, Hardy's lovely little book has become a paradigm for mathematical
apologies, and any apologetic mathematician ought to at least mention Hardy's A Mathematician 's Apology in her credits.

Quotation from Lewis Carroll on page xvii: Lewis Carroll, Through the Looking-Glass, The
Complete Illustrated Lewis Carroll, Wordsworth Editions, 1991, page 196.
Quotation from Soren Kierkegaard on page xxxvii in the "Getting started" section of the front
matter: Soren Kierkegaard, Journalen JJ:167 (1843), Soren Kierkegaards Skrifter, Seren
Kierkegaard Research Center, Copenhagen, 1997-, vol. 18, p. 306. Thanks to Karsten
Kynde, Soren Kierkegaard Forskningscenteret, web page http://www.sk.ku.dk/citater/, for
the Danish text of the quotation and directions to its location in print. The English translation of the long quote is my own. The Danish short form is due to Julia Watkin; see the
web page http://www.utas.edu.au/docs/humsoc/kierkegaard/resources/Kierkquotes.html.
Quotation from Stephen Hawking on page xliii in the Acknowledgments: The quote here is
from page ix in Stephen Hawking's book On the Shoulders of Giants: The Great Works
of Physics and Astronomy, Running Press, Philadelphia, 2002. This book in turn is a collection of reprints of original classics in the sciences. The book title "On the shoulders of
giants" is a quote from Isaac Newton.
Quotation from Edward B. Burger and Michael Starbird on page 1: Edward B. Burger and
Michael Starbird, Coincidences, Chaos, and All That Math Jazz: Making Light of Weighty
Ideas, W.W. Norton & Company, 2005. Quoted from the Front matter: Opening thoughts.
Quotation from Lewis Carroll on page 9: Lewis Carroll, Through the Looking-Glass, Chapter
3, The Complete Illustrated Lewis Carroll, Wordsworth Editions, 1991.
Quotation from CM. Brislawn on page 22: CM. Brislawn, Fingerprints go digital. Notices
of the American Mathematical Society, vol. 42, 1995, p. 1278.
Quotation from Piet Hein on page 27: Piet Hein, "Problems," Grooks, Borgens Forlag, Copenhagen, Denmark, The MIT Press, Cambridge, MA, 1966; Grooks 1, Doubleday & Company Inc., Garden City, NY, General Publishing Company Limited, Toronto, 1969; Collected Grooks I, Borgens Forlag, Copenhagen, Denmark, 2002.
Quotation from Lewis Carroll on page 39: Lewis Carroll, Alice's Adventures in Wonderland,
Chapter VI, The Complete Illustrated Lewis Carroll, Wordsworth Editions, 1991.
Quotation from P. A.M. Dirac on page 58: The Development of Quantum Theory (J. Robert
Oppenheimer Memorial Prize Acceptance Speech), Gordon and Breach Publishers, New
York, 1971, pp. 20-24. The quote used an an epigraph on p. 2 in Operator Commutation
Relations by Palle E.T. Jorgensen and Robert T. Moore, D. Reidel, Dordrecht, Boston,
1984, is an abridgement of the above passage. A longer excerpt (with the curious substitution of "commutation" for "noncommutation" where it first appears) is presented as an epigraph in the announcement of a "Program on Noncommutative Algebra" at the Mathematical Sciences Research Institute, http://www.msri.org/activities/programs/9900/noncomm/.

Epigraphs

235

Back when I thought about what to put into Operator Commutation Relations, I relied a
lot on The Historical Development of Quantum Theory by Jagdish Mehra and Helmut
Rechenberg, Springer-Verlag, New York, 1982- . I was also impressed by how well researched this lovely book-set is. The two authors, Mehra and Rechenberg, did long interviews over the span of time when they worked on their book set. They had known Bohr,
and they had many meetings with Heisenberg, Pauli, Dirac, and many more, and as I remember, much material in the book set results directly from these interviews. It is good
that the two coauthors had the interviews, as these giants in quantum theory passed on
shortly after the book set was completed. I think the book set is a treasure of information
on quantum theory (especially the mathematical part of it) and on the architects of the
theory. Late in life, Dirac would always tell the physicists at conferences to look to the
math for clues to the deep questions in physics, and he liked to use his (Dirac) equation for
the electron as an example, stressing that he was led to it by paying attention to the beauty
of the math, more than to the physics experiments. He was alive when I was working on
Operator Commutation Relations, and I talked to him a few times. He told me that he was
happy to be quoted. He used to visit his son Gabriel Dirac (graph theory) who was my
colleague in Aarhus.
Quotation from A.N. Kolmogorov on page 59: A.N. Kolmogorov, Foundations of the Theory
ofProbability, 2d English ed., translation edited by Nathan Morrison, Chelsea, New York,
1956, p. 1; for the German original see [Kol77].
Quotation from Benoit B. Mandelbrot on page 69: Benoit B. Mandelbrot, Multifractals and
\/f noise: Wild self affinity in physics (1963-1976), Selecta Volume N, Selected Works
of Benoit B. Mandelbrot, Springer- Verlag, New York, 1999, p. 9.
Quotation from Albert Einstein on page 83: These two quotes by A, Einstein are from pages
228-9 in The New Quotable Einstein, collected from Einstein's archives and edited by
Alice Calaprice, Princeton University Press, 2005. In their original, they are from conversations, the second one with Einstein and Valentine Bargmann, meaning that God makes
us believe we have understood something that in reality we are far from understanding.
Quotation from Percy Bysshe Shelley on page 91: Percy Bysshe Shelley, "Queen Mab: A
Philosophical Poem, with Notes," published by the author, London, 1813.
Quotation from Douglas Adams on page 99: Quoted from the front matter in Douglas Adams,
Hitchhiker's Guide to the Galaxy (Mostly Harmless), Del Rey, Ballantine Books, New
York, 2005.
Quotation from Yves Meyer on page 109: Yves Meyer, Wavelets and functions with bounded
variation from image processing to pure mathematics, Atti della Accademia Nazionale dei
Lincei, Classe di Scienze Fisiche, Matematiche e Naturali, Rendiconti Lincei, Serie IX,
Matematica e Applicazioni, vol. XI, 2000, special issue: Mathematics Towards the Third
Millennium (Papers from the International Conference held in Rome, May 27-29, 1999),
p. 95.
Quotation from Oliver Heaviside on page 157: Oliver Heaviside, On operators in physical
mathematics, part II, Proceedings of the Royal Society of London, vol. 54, 1893, p. 121.

236

Epigraphs

Quotation from Max Bom on page 176: Jagdish Mehra and Helmut Rechenberg, The Historical Development of Quantum Theory, vol. 3: The Formulation of Matrix Mechanics and
Its Modifications, 1925-1926, Springer-Verlag, New York, 1982, p. 129, footnote 146.
Quotation from Werner Heisenberg on page 179: Jagdish Mehra and Hehnut Rechenberg, The
Historical Development of Quantum Theory, vol. 3: The Formulation ofMatrix Mechanics
and Its Modifications, 1925-1926, Springer-Verlag, New York, 1982, p. 94.
Quotation from Norbert Wiener on page 205: Norbert Wiener, I Am a Mathematician, Doubleday, Garden City, NY, 1956; Victor Gollancz, London, 1956; The MIT Press, Cambridge,
MA, 1964, pp. 108-109.

References

Explanation: The list below includes both books and journal articles, listed alphabetically by author. The type style and punctuation are generated by BibTeX, and we
follow a well established AMS prescription. One general rule applies to all different
types of sources cited (books, journals, monographs, memoirs, edited volumes, web
sources): The title of the work cited is in italic type. This makes it easy for readers to
identify at a glance what is being cited. Thus "longer works" such as books are not
given the special status of being in italic. This BibTeX styling makes it clear which
item is a book and which a paper: entries for journal articles end with the volume
number in bold, the year in parentheses, and the page range, while entries for books
end with the publisher's name and address, and the year not in parentheses. Articles
published in books such as proceedings volumes have the same book information
near the end of the entry, but it is followed by the abbreviation "pp." and the page
range of the article within the book.
Comments on our citations of a variety of websites: Li general, websites might not
have well-defined authors, and well-defined year of "publication," so we cite them as
[WWWl], [WWW2], etc. Readers should keep in mind that we cite these URLs only
as supplements to issues covered inside the book; and we hope that they will inspire
readers to follow up on the themes covered in the book and in the various URLs.
[AgKu04a]

[AgKu04b]

[ACM04]

B. Agard and A. Kusiak, A data-mining based methodology for the design

ofproduct families. International Journal of Production Research 42 (2004),
no. 15, 2955-2969.
B. Agard and A. Kusiak, Data mining for subassembly selection, ASME
Transactions: Joumal of Manufacturing Science and Engineering 126 (2004),
no. 3,627-631.
A. Aldroubi, C.A. Cabrelli, and U.M. Molter, How to construct waveletframes
on irregular grids and arbitrary dilations in W^, Wavelets, Frames and Operator Theory (College Park, MD, 2003) (C. Heil, PE.T. Jorgensen, and D.R. Larson, eds.), Contemp. Math., vol. 345, American Mathematical Society, Providence, 2004, pp. 1-9.

238

References

[AlGrOl]
[Arv04]
[Ash90]
[AyTa03]
[BaCM02]

[BaJMP04]

[BaJMP05]

[BaJMP06]

[BaMM99]

[BaMe99]

[BalOO]
[BaCV05]

[BaVi05]

[BaNi70]
[Bas95]
[Bas03]

A. Aldroubi and K. Grochenig, Nonuniform sampling and reconstruction in

shift-invariant spaces, SIAM Rev. 43 (2001), 585-620.
W. Arveson, The free cover of a row contraction. Doc. Math. 9 (2004), 137161.
R.B. Ash, Information Theory, Dover, New York, 1990, corrected reprint of
the original 1965 Interscience/Wiley edition.
A. Ay ache and M.S. Taqqu, Rate optimality of wavelet series approximations
of fractional Brownian motion, J. Fourier Anal. Appl. 9 (2003), 451-471.
L.W. Baggett, J.E. Courter, and K.D. Merrill, The construction of wavelets
from generalized conjugate mirrorfilters in L^(R"), Appl. Comput. Harmon.
Anal. 13 (2002), 201-223.
L.W. Baggett, RE.T. Jorgensen, K.D. Merrill, and J.A. Packer, An analogue
of Bratteli'Jorgensen loop group actions for GMRA 's. Wavelets, Frames, and
Operator Theory (College Park, MD, 2003) (C. Heil, RE.T. Jorgensen, and
D.R. Larson, eds.), Contemp. Math., vol. 345, American Mathematical Society, Providence, 2004, pp. 11-25.
L.W. Baggett, RE.T. Jorgensen, K.D. Merrill, and J.A. Packer, Construction
ofParseval wavelets from redundant filter systems, J. Math. Phys. 46 (2005),
no. 8, 083502, 28 pp., doi: 10.1063/1.1982768.
L.W. Baggett, RE.T. Jorgensen, K.D. Merrill, and J.A. Packer, A non-MRA
C frame wavelet with rapid decay. Acta Appl. Math. 89 (2006), 251-270,
doi:10.1007/sl0440-005-9011-4.
L.W Baggett, H.A. Medina, and K.D. Merrill, Generalized multi-resolution
analyses and a construction procedure for all wavelet sets in R", J. Fourier
Anal. Appl. 5 (1999), 563-573.
L.W. Baggett and K.D. Merrill, Abstract harmonic analysis and wavelets in
R", The Functional and Harmonic Analysis of Wavelets and Frames (San Antonio, 1999) (L.W. Baggett and D.R. Larson, eds.), Contemp. Math., vol. 247,
American Mathematical Society, Providence, 1999, pp. 17-27.
V. Baladi, Positive Transfer Operators and Decay of Correlations, World Scientific, River Edge, NJ, Singapore, 2000.
J.A. Ball, C. Sadosky, and V. Vinnikov, Conservative input-state-output systems with evolution on a multidimensional integer lattice, Multidimens. Syst.
Signal Process. 16 (2005), no. 2, 133-198.
J.A. Ball and V. Vinnikov, Lax-Phillips Scattering and Conservative Linear
Systems: A Cuntz-Algebra Multidimensional Setting, Mem. Amer. Math. Soc.
178 (2005), no. 837.
M.N. Barber and B.W Ninham, Random and Restricted Walks: Theory and
Applications, Gordon and Breach, New York, 1970.
R.F. Bass, Probabilistic Techniques in Analysis, Probability and Its Applications, Springer-Verlag, New York, 1995.
R.F. Bass, Stochastic differential equations driven by symmetric stable processes, Seminaire de Probabilites, XXXVI (J. Azema, M. Emery, M. Ledoux,
and M. Yor, eds.). Lecture Notes in Math., vol. 1801, Springer-Verlag, Berlin,
2003, pp. 302-313.

References
[Bat87]
[Bea91]

[BeBe95]

[BeBK05]
[Bil99]
[BoNe03]
[BrEJOO]

[BrJo97]

[BrJo99a]

[BrJo99b]

[BrJo02a]

[BrJo02b]

[BrJKWOO]
[BrJO04]

[BrJP96]

239

G. Battle, A block spin construction of ondelettes, I: Lemarie functions,

Comm. Math. Phys. 110 (1987), 601^15.
A.F. Beardon, Iteration of Rational Functions: Complex Analytic Dynamical Systems, Graduate Texts in Mathematics, vol. 132, Springer-Verlag, New
York, 1991.
J.J. Benedetto and E.G. Bernstein, Pyramidal Riesz products associated with
subband coding and self similarity. Wavelet Applications II (H.H. Szu, ed.).
Proceedings of SPIE, vol. 2491, SPIE, Bellingham, WA, 1995, pp. 212-221.
J.J. Benedetto, E. Bernstein, and I. Konstantinidis, Multiscale Riesz products
and their support properties. Acta Appl. Math. 88 (2005), no. 2, 201-227.
P. Billingsley, Convergence of Probability Measures, second ed., Wiley Series
in Probability and Statistics, Wiley-Interscience, New York, 1999.
E. Bondarenko and V.V. Nekrashevych, Post-critically finite self-similar
groups. Algebra Discrete Math. 2003, 21-32.
O. Bratteli, D.E. Evans, and P.E.T. Jorgensen, Compactly supported wavelets
and representations of the Cuntz relations, Appl. Comput. Harmon. Anal. 8
(2000), 166-196.
O. Bratteli and P.E.T. Jorgensen, Isometrics, shifts, Cuntz algebras and multiresolution wavelet analysis of scale N, Integral Equations Operator Theory
28 (1997), 382-443.
0. Bratteli and P.E.T. Jorgensen, Iterated function systems and permutation
representations of the Cuntz algebra, Mem. Amer. Math. Soc. 139 (1999),
no. 663.
O. Bratteli and P.E.T. Jorgensen, Convergence of the cascade algorithm at
irregular scaling functions. The Functional and Harmonic Analysis of Wavelets and Frames (San Antonio, 1999) (L.W. Baggett and D.R. Larson, eds.),
Contemp. Math., vol. 247, American Mathematical Society, Providence, 1999,
pp. 93-130.
O. Bratteli and P.E.T. Jorgensen, Wavelet filters and infinite-dimensional unitary groups. Wavelet Analysis and Applications (Guangzhou, China, 1999)
(D. Deng, D. Huang, R.-Q. Jia, W Lin, and J. Wang, eds.), AMS/IP Studies in
Advanced Mathematics, vol. 25, American Mathematical Society, Providence,
International Press, Boston, 2002, pp. 35-65.
O. Bratteli and P.E.T. Jorgensen, Wavelets through a Looking Glass:
The World of the Spectrum, Applied and Numerical Harmonic Analysis,
Birkhauser, Boston, 2002.
O. Bratteli, P.E.T. Jorgensen, A. Kishimoto, and R. Werner, Pure states on O^,
J. Operator Theory 43 (2000), 97-143.
O. Bratteli, P.E.T. Jorgensen, and V. Ostrovs'kyi, Representation Theory and
Numerical AF-invariants: The Representations and Centralizers of Certain
States on O^, Mem. Amer. Math. Soc. 168 (2004), no. 797.
O. Bratteli, P.E.T. Jorgensen, and G.L. Price, Endomorphisms of B(H),
Quantization, Nonlinear Partial Differential Equations, and Operator Algebra (W. Arveson, T. Branson, and I. Segal, eds.), Proc. Sympos. Pure Math.,
vol. 59, American Mathematical Society, Providence, 1996, pp. 93-138.

240

References

[Bri95]
[BrRo91]

[BrMo75]

[CaHM04]
[Chr03]
[Coh90]
[Coh03]
[CoMW95]

[CoWi93]

[CoHR97]

[CoRa90]
[Cun77]
[DaSw98]
[CvSS85]
[Dau92]
[DaHRS03]
[Dav96]
[DaKSOl]

C. Brislawn, Fingerprints go digital. Notices Amer. Math. Soc. 42 (1995),

1278-1283.
C. Brislawn and I.G. Rosen, Wavelet based approximation in the optimal control of distributed parameter systems, Numer. Funct. Anal. Optim. 12 (1991),
33-77.
G. Brown and W. Moran, Products of random variables and Kakutani's criterion for orthogonality of product measures, J. London Math. Soc. (2) 10
(1975), no. part 4, 401-405.
C.A. Cabrelli, C. Heil, and U.M. Molter, Self-similarity and multiwavelets in
higher dimensions, Mem. Amer. Math. Soc. 170 (2004), no. 807.
O. Christensen, An Introduction to Frames and Riesz Bases, Applied and Numerical Harmonic Analysis, Birkhauser, Boston, 2003.
A. Cohen, Ondelettes, analyses multiresolutions etfiltres miroirs en quadrature, Ann. Inst. H. Poincare Anal. Non Lineaire 7 (1990), 439-459.
A. Cohen, Numerical Analysis of Wavelet Methods, Studies in Mathematics
and Its Applications, vol. 32, North-Holland, Amsterdam, 2003.
R.R. Coifman, Y. Meyer, and V. Wickerhauser, Numerical harmonic analysis. Essays on Fourier Analysis in Honor of Elias M. Stein (Princeton, 1991)
(C. Fefferman, R. Fefferman, and S. Wainger, eds.), Princeton Mathematical
Series, vol. 42, Princeton University Press, Princeton, NJ, 1995, pp. 162-174.
R.R. Coifinan and M. V. Wickerhauser, Wavelets and adapted waveform analysis: A toolkit for signal processing and numerical analysis. Different Perspectives on Wavelets (San Antonio, TX, 1993) (I. Daubechies, ed.), Proc.
Sympos. Appl. Math., vol. 47, American Mathematical Society, Providence,
1993, pp. 119-153.
J.-P. Conze, L. Herve, and A. Raugi, Pavages auto-affines, operateurs de
transfert et entires de reseau dans R^, Bol. Soc. Brasil. Mat. (N.S.) 28 (1997),
1^2.
J.-P. Conze and A. Raugi, Fonctions harmoniques pour un operateur de transition et applications. Bull. Soc. Math. France 118 (1990), 273-310.
J. Cuntz, Simple C*-algebras generated by isometrics, Comm. Math. Phys. 57
(1977), 173-185.
I. Daubechies and W. Sweldens, Factoring wavelet transforms into lifting
steps, J. Fourier Anal. Appl. 4 (1998), 247-269.
P. Cvitanovic, B. Shraiman, and B. Soderberg, Scaling laws for mode lockings
in circle maps, Phys. Scripta 32 (1985), 263-270.
I. Daubechies, Ten Lectures on Wavelets, CBMS-NSF Regional Conf. Ser. in
Appl. Math., vol. 61, SIAM, Philadelphia, 1992.
I. Daubechies, B. Han, A. Ron, and Z. Shen, Framelets: MRA-based constructions of wavelet frames, Appl. Comput. Harmon. Anal. 14 (2003), 1-46.
K.R. Davidson, C*-algebras by Example, Fields Institute Monographs, vol. 6,
American Mathematical Society, Providence, 1996.
K.R. Davidson, D.W. Kribs, and M.E. Shpigel, Isometric dilations of
non-commuting finite rank n-tuples, Canad. J. Math. 53 (2001), 506-545,
http://arxiv.org/abs/math.OA/0411521.

References
[DeSh04]
[Dev92]

[DiFr99]
[DoGHOO]

[DoJo94]

[DoMSS03]

[Doo94]
[DooO 1 ]
[Dut02]
[Dut04a]

[Dut04b]
[DuJo05a]

[DuJo05b]

[DuJo06a]

[DuJo06b]
[DuJo06c]
[EdwOl]

241

V. Deaconu and R Shultz, C*-algebras associated with interval maps. Trans.

Amer. Math. Soc, to appear, math.OA/0405469.
R.L. Devaney, A First Course in Chaotic Dynamical Systems: Theory and
Experiment, Addison-Wesley Studies in Nonlinearity, Addison-Wesley Advanced Book Program, Reading, MA, 1992, with a separately available computer disk.
R Diaconis and D. Freedman, Iterated random functions, SIAM Rev. 41
(1999), 45-76, graphical supplement 77-82.
V. Dobric, R. Gundy, and R Hitczenko, Characterizations of orthonormal
scale functions: A probabilistic approach, J. Geom. Anal. 10 (2000), 417434.
D.L. Donoho and I.M. Johnstone, Ideal denoising in an orthonormal basis
chosen from a library of bases, C. R. Acad. Sci. Paris Ser. I Math. 319 (1994),
1317-1322.
D.L. Donoho, S. Mallat, R. von Sachs, and Y. Samuelides, Locally stationary
covariance and signal estimation with macrotiles, IEEE Trans. Signal Process.
51 (2003), 614^27.
J.L. Doob, Measure Theory, Graduate Texts in Mathematics, vol. 143,
Springer-Verlag, New York, 1994.
J.L. Doob, Classical Potential Theory and Its Probabilistic Counterpart, Classics in Mathematics, Springer-Verlag, Berlin, 2001, reprint of the 1984 edition.
D.E. Dutkay, Harmonic analysis of signed Ruelle transfer operators, J. Math.
Anal. Appl. 273 (2002), 590-617.
D.E. Dutkay, The local trace function for super-wavelets. Wavelets, Frames,
and Operator Theory (Focused Research Group Workshop, College Park,
Maryland, January 15-21, 2003) (C. Heil, P.E.T. Jorgensen, and D. Larson,
eds.), Contemp. Math., vol. 345, American Mathematical Society, Providence,
2004, pp. 115-136.
D.E. Dutkay, Positive definite maps, representations and frames. Rev. Math.
Phys. 16 (2004), no. 4, 451-477.
D.E. Dutkay and P.E.T. Jorgensen, Wavelet constructions in non-linear
dynamics. Electron. Res. Announc. Amer. Math. Soc. 11 (2005), 21-33,
http://vmw.ams.org/era/2005-ll-03/S1079-6762-05-00143-5/home.html.
D.E. Dutkay and P.E.T. Jorgensen, Hilbert spaces of martingales supporting certain substitution-dynamical systems. Conform. Geom. Dyn. 9
(2005), 24-45, http://www.ams.org/ecgd/2005-09-02/S1088-4173-05-001359/home.html.
D.E. Dutkay and P.E.T. Jorgensen, Martingales, endomorphisms, and covariant systems of operators in Hilbert space, J. Operator Theory, to appear (in
galley proof 2006, pp. 101-145), http://arXiv.org/abs/math.CA/0407330.
D.E. Dutkay and P.E.T. Jorgensen, Wavelets on fractals. Rev. Mat. Iberoamericana22 (2006), 131-180.
D.E. Dutkay and P.E.T. Jorgensen, Oversampling generates super-wavelets,
Proc. Amer. Math. Soc, to appear, http://arxiv.org/abs/math.FA/0511399.
H.M. Edwards, Riemann 's Zeta Function, Dover, Mineola, NY, 2001, reprint
of the 1974 original [Academic Press, New York].

242

References

[Fal85]
[Fal90]
[FaLa99]
[Fel71]
[FeLa02]
[FiEs61]
[FrLM83]
[Fug74]
[GaNa98b]
[GaYu05]

[GaNSOla]

[GaNSOlb]
[GaNS02]

[GlZu80]
[Gli60]
[GrMa92]
[GroOl]
[GrMo84]

[GunOO]
[Gun66]

K.J. Falconer, The Geometry of Fractal Sets, Cambridge Tracts in Mathematics, vol. 85, Cambridge University Press, Cambridge, 1985.
K. Falconer, Fractal Geometry: Mathematical Foundations and Applications,
Wiley, Chichester, 1990.
A.H. Fan and K.-S. Lau, Iterated function system andRuelle operator, J. Math.
Anal. Appl. 231 (1999), 319-344.
W. Feller, An Introduction to Probability Theory and Its Applications, 3rd ed.,
vol. 2, Wiley, New York, 1971.
D.-J. Feng and K.-S. Lau, The pressure function for products of non-negative
matrices. Math. Res. Lett. 9 (2002), 363-378.
M.E. Fisher and J.W. Essam, Some cluster size and percolation problems, J.
Math. Phys. 2 (1961), 609-619.
A. Freire, A. Lopes, and R. Maiie, An invariant measure for rational maps,
Bol. Soc. Brasil. Mat. 14 (1983), 45-62.
B. Fuglede, Commuting self-adjoint partial differential operators and a group
theoretic problem, J. Funct. Anal. 16 (1974), 101-121.
J.-P. Gabardo and M.Z. Nashed, Nonuniform multiresolution analyses and
spectral pairs, J. Funct. Anal. 158 (1998), no. 1, 209-241.
J.-P. Gabardo and X. Yu, Wavelets associated with nonuniform multiresolution analyses and one-dimensional spectral pairs, preprint 2005, McMaster
University.
X.Q. Gao, T.Q. Nguyen, and G. Strang, Theory and lattice structure
of complex paraunitary filterbanks with filters of (Hermitian-) symmetry/antisymmetry properties, IEEE Trans. Signal Process. 49 (2001), 10281043.
X.Q. Gao, T.Q. Nguyen, and G. Strang, On factorization of M-channel paraunUary filterbanks, IEEE Trans. Signal Process. 49 (2001), 1433-1446.
X.Q. Gao, T.Q. Nguyen, and G. Strang, A study of two-channel complexvalued filterbanks and wavelets with orthogonality and symmetry properties,
IEEE Trans. Signal Process. 50 (2002), 824^833.
M.L. Glasser and I.J. Zucker, Lattice sums, Theor. Chem. Adv. Perspect. 5
(1980), 67-139.
J. Glimm, On a certain class of operator algebras. Trans. Amer. Math. Soc.
95 (1960), 318-340.
K. Grochenig and WR. Madych, Multiresolution analysis, Haar bases, and
self-similar tilings o/R", IEEE Trans. Inform. Theory 38 (1992), 556-568.
K. Grochenig, Foundations of Time-Frequency Analysis, Applied and Numerical Harmonic Analysis, Birkhauser, Boston, 2001.
A. Grossmann and J. Morlet, Decomposition of Hardy functions into square
integrable wavelets of constant shape, SIAM J. Math. Anal. 15 (1984), 723736.
R.F. Gundy, Low-pass filters, martingales, and multiresolution analyses, Appl.
Comput. Harmon. Anal. 9 (2000), 204-219.
R.F. Gundy, Martingale theory and pointwise convergence of certain orthogonal series. Trans. Amer. Math. Soc. 124 (1966), 228-248.

References
[Gun99]

[Gun04]

[GuKaOO]
[HaalO]
[Hal83]
[Hal67]

[HeJL04]

[HeWe96]
[HuTK05]

[Hug95]

[HuMS82]
[HuSM81]
[Hut81 ]
[IoMa50]

[Jaf05]
[JaMe96]
[JaMROl]

243

R.F. Gundy, Two remarks concerning wavelets: Cohen's criterion for low-pass
filters and Meyer's theorem on linear independence. The Functional and Harmonic Analysis of Wavelets and Frames (San Antonio, 1999) (L.W. Baggett
and D.R. Larson, eds.), Contemp. Math., vol. 247, American Mathematical
Society, Providence, 1999, pp. 249-258.
R.F. Gundy, Wavelets and probability, preprint, Rutgers University, material presented during the author's lecture at the workshop "Wavelets and Applications", Barcelona, Spain, July 1-6, 2002,
http ://www. imub. ub. es/wavelets/Gundy.pdf.
R.F. Gundy and K. Kazarian, Stopping times and local convergence for spline
wavelet expansions, SIAM J. Math. Anal. 31 (2000), 561-573.
A. Haar, Zur Theorie der orthogonalen Funktionensysteme, Math. Ann. 69
(1910), 331-371.
F.D.M. Haldane, Fractional quantization of the Hall effect: A hierarchy of
incompressible quantum fluid states, Phys. Rev. Lett. 51 (1983), 605-608.
P.R. Halmos, A Hilbert Space Problem Book, D. Van Nostrand, Princeton,
Toronto, London, 1967, Springer, 2nd edition, 1982, Graduate Texts in Mathematics.
C. Heil, P.E.T. Jorgensen, and D.R. Larson (eds.). Wavelets, Frames, and Operator Theory: Papers from the Focused Research Group Workshop held at the
University of Maryland, College Park, MD, January 15-21, 2003, Contemp.
Math., vol. 345, American Mathematical Society, Providence, 2004.
E. Hernandez and G. Weiss, A First Course on Wavelets, Studies in Advanced
Mathematics, CRC Press, Boca Raton, Florida, 1996.
C.-C. Huang, T.-L. Tseng, and A. Kusiak, XML-based modeling of corporate
memory, IEEE Transactions on Systems, Man, and Cybernetics, Part A 35
(2005), no. 5, 629-640.
B.D. Hughes, Random Walks and Random Environments, Vol. 1: Random
Walks, Oxford Science Publications, Clarendon Press, Oxford University
Press, Oxford, New York, 1995.
B.D. Hughes, E.W. Montroll, and M.F. Shlesinger, Fractal random walks, J.
Statist. Phys. 28 (1982), 111-126.
B.D. Hughes, M.F. Shlesinger, and E.W Montroll, Random walks with selfsimilar clusters, Proc. Nat. Acad. Sci. U.S.A. 78 (1981), 3287-3291.
J.E. Hutchinson, Fractals and self similarity, Indiana Univ. Math. J. 30 (1981),
713-747.
C.T. lonescu Tulcea and G. Marinescu, Theorie ergodique pour des classes
d'operations non completement continues, Ann. of Math. (2) 52 (1950), 140147.
S. Jaffard, Beyond Besov spaces, II: Oscillation spaces, Constr. Approx. 21
(2005), 29-61.
S. Jaffard and Y. Meyer, Wavelet Methods for Pointwise Regularity and Local
Oscillations of Functions, Mem. Amer. Math. Soc. 123 (1996), no. 587.
S. Jaffard, Y Meyer, and R.D. Ryan, Wavelets: Tools for Science & Technology, revised ed., SIAM, Philadelphia, 2001.

244

References

[Jam64]
[JiJSOl]

[JiSh99]
[Jor99]
[JorOla]
[JorOlb]
[JorOlc]

[Jor03]

[Jor04a]
[Jor04b]

[Jor05]
[JoKr03]
[JoPe92]
[JoPe93]
[JoPe96]
[JoPe98]
[JoSW95]

[JPEGOO]
[Kak48]

B. Jamison, Asymptotic behavior of successive iterates of continuous functions

under a Markov operator, J. Math. Anal. Appl. 9 (1964), 203-214.
R.-Q. Jia, Q. Jiang, and Z. Shen, Convergence of cascade algorithms associated with nonhomogeneous refinement equations, Proc. Amer. Math. Soc. 129
(2001), 415-427.
Q. Jiang and Z. Shen, On existence and weak stability of matrix refinable
functions, Constr. Approx. 15 (1999), 337-353.
RE.T. Jorgensen, A geometric approach to the cascade approximation operator for wavelets. Integral Equations Operator Theory 35 (1999), 125-171.
RE.T. Jorgensen, Ruelle Operators: Functions Which Are Harmonic with Respect to a Transfer Operator, Mem. Amer. Math. Soc. 152 (2001), no. 720.
RE.T. Jorgensen, Minimality of the data in wavelet filters. Adv. Math. 159
(2001), 143-228.
RE.T. Jorgensen, Representations ofCuntz algebras, loop groups and wavelets, Xlllth International Congress on Mathematical Physics (London, 2000)
(A. Fokas, A. Grigoryan, T. Kibble, and B. Zegarlinski, eds.). International
Press, Boston, 2001, pp. 327-332.
RE.T. Jorgensen, Matrix factorizations, algorithms, wavelets. Notices Amer.
Math. Soc. 50 (2003), 880-894, http://www.ams.org/notices/200308/200308toc.html.
RE.T. Jorgensen, Iterated function systems, representations, and Hilbert
space, Intemat. J. Math. 15 (2004), 813-832.
RE.T. Jorgensen, Use of operator algebras in the analysis of measures from
wavelets and iteratedfunction systems, accepted for Operator Theory, Operator Algebras, and Applications (Deguang Han, Palle Jorgensen, and David R.
Larson, eds.), Contemp. Math., American Mathematical Society, Providence,
to appear 2006. http://arxiv.org/abs/math.OA/0509360.
RE.T. Jorgensen, Measures in wavelet decompositions. Adv. Appl. Math. 34
(2005), 561-590.
RE.T. Jorgensen and D. W. Kribs, Wavelet representations and Fock space on
positive matrices, J. Funct. Anal. 197 (2003), 526-559.
RE.T. Jorgensen and S. Pedersen, Spectral theory for Borel sets in R" of finite
measure, J. Funct. Anal. 107 (1992), 72-104.
RE.T. Jorgensen and S. Pedersen, Group-theoretic and geometric properties
ofmultivariable Fourier series. Exposition. Math. 11 (1993), 309-329.
RE.T. Jorgensen and S. Pedersen, Harmonic analysis of fractal measures,
Constr. Approx. 12 (1996), 1-30.
RE.T. Jorgensen and S. Pedersen, Dense analytic subspaces in fractal L?-spaces, J. Analyse Math. 75 (1998), 185-228.
RE.T. Jorgensen, L.M. Schmitt, and R.F. Werner, Positive representations of
general commutation relations allowing Wick ordering, J. Funct. Anal. 134
(1995), 33-99.
The JPEG2000 resource webpage,
http://stargate.ecn.purdue.edurips/tutorials/j2k/#What-is.
S. Kakutani, On equivalence of infinite product measures, Ann. of Math. (2)
49 (1948), 214-224.

References
[Kat87]
[Kea72]
[KigOl]
[KnuSl]

[Knu84]
[Kol77]

[KoSp04]
[KuDS05]

[KusO 1]
[Kus02]

[Kus05]

[KuBu05]

[KuLDOl]

[LaWa02]
[LaNg98]
[LaNROl]
[LaRe91]
[Law91a]

245

Y. Katznelson, Integrales de produits de Riesz, C. R. Acad. Sci. Paris Ser. I

Math. 305 (1987), 67-69.
M. Keane, Strongly mixing g-measures. Invent. Math. 16 (1972), 309-324.
J. Kigami, Analysis on Fractals, Cambridge Tracts in Mathematics, vol. 143,
Cambridge University Press, Cambridge, 2001.
D.E. Knuth, The Art of Computer Programming: Vol. 2: Seminumerical Algorithms, 2nd ed., Addison-Wesley Series in Computer Science and Information
Processing, Addison-Wesley, Reading, Mass., 1981.
D.E. Knuth, An algorithm for Brownian zeroes, Computing 33 (1984), 89-94.
A.N. Kolmogorov, Grundbegriffe der Wahrscheinlichkeitsrechnung, SpringerVerlag, Berlin-New York, 1977, reprint of the 1933 original; English translation: Foundations of the Theory of Probability, Chelsea, 1950.
C. Kostler and R. Speicher, On the structure of non-commutative white noises,
http://arxiv.org/abs/math.OA/0411519.
A. Kusiak, B. Dixon, and S. Shah, Predicting survival time for kidney dialysis
patients: A data mining approach. Computers in Biology and Medicine 35
(2005),no. 4, 311-327.
A. Kusiak, Feature transformation methods in data mining, IEEE Transactions
on Electronics Packaging Manufacturing 24 (2001), no. 3, 214221.
A. Kusiak, A data mining approach for generation of control signatures,
ASME Transactions: Journal of Manufacturing Science and Engineering 124
(2002), no. 4, 923-926.
A. Kusiak, Selection of invariant objects with a data mining approach, IEEE
Transactions on Electronics Packaging Manufacturing 28 (2005), no. 2, 187196.
A. Kusiak and A. Bums, Mining temporal data: A coal-fired boiler case study,
Knowledge-Based Intelligent Information and Engineering Systems, Proceedings of the 9th International Conference, KES 2005, Melbourne, Australia,
September 14-16,2005, Vol. Ill (R. Khosla, R.J. Howlett, and L.C. Jain, eds.),
Lecture Notes in Artificial Intelligence, vol. 3683, Springer, Heidelberg, 2005,
pp. 953-958.
A. Kusiak, I.H. Law, and M.D. Dick, The G-algorithm for extraction of robust
decision rules: Children's postoperative intra-atrial arrhythmia case study,
IEEE Transactions on Information Technology in Biomedicine 5 (2001), no. 3,
225-235.
I. Laba and Y. Wang, On spectral Cantor measures, J. Funct. Anal. 193 (2002),
409-^20.
K.-S. Lau and S.-M. Ngai, L^ -spectrum of the Bernoulli convolution associated with the golden ratio, Studia Math. 131 (1998), 225-251.
K.-S. Lau, S.-M. Ngai, and H. Rao, Iterated function systems with overlaps
and self-similar measures, J. London Math. Soc. (2) 63 (2001), 99-116.
W.M. Lawton and H.L. Resnikoff, Multidimensional wavelet bases. Report
AD910130, AWARE, Inc., Cambridge, MA, 1991.
W.M. Lawton, Necessary and sufficient conditions for constructing orthonormal wavelet bases, J. Math. Phys. 32 (1991), 57-61.

246

References

[Law91b]
[LaLS96]
[LaLS98]
[LySt03]

[Mal89]
[Mal98]
[Mar82]

[May91]

[Mey79]

[Mey89]

[Mey97]

[Mey05]
[MeCo97]

[Mon64]

[Nek04]
[Nel59]
[Nel64]

W.M. Lawton, Multire solution properties of the wavelet Galerkin operator, J.

Math. Phys. 32 (1991), 1440-1443.
W.M. Lawton, S.L. Lee, and Z. Shen, An algorithm for matrix extension and
wavelet construction. Math. Comp. 65 (1996), no. 214, 723-737.
W.M. Lawton, S.L. Lee, and Z. Shen, Convergence of multidimensional cascade algorithm, Numer. Math. 78 (1998), 427-438.
R. Lyons and J.E. Steif, Stationary determinantal processes: phase multiplicity, Bernoullicity, entropy, and domination, Duke Math. J. 120 (2003), 515575.
S.G. Mallat, Multiresolution approximations and wavelet orthonormal bases
ofl?(K), Trans. Amer. Math. Soc. 315 (1989), 69-87.
S.G. Mallat, A Wavelet Tour of Signal Processing, Academic Press, San
Diego, 1998.
D. Marr, Vision: A Computational Investigation into the Human Representation and Processing of Visual Information, W.H. Freeman, San Francisco,
1982.
D.H. Mayer, Continued fractions and related transformations, Ergodic Theory, Symbolic Dynamics, and Hyperbolic Spaces (Trieste, 1989) (T. Bedford,
M. Keane, and C. Series, eds.), Oxford Science Publications, The Clarendon
Press, Oxford University Press, New York, 1991, pp. 175-222.
Y. Meyer, Produits de Riesz generalises, Seminaire d'Analyse Harmonique,
1978-1979, Publications Mathematiques d'Orsay 79, vol. 7, Universite de
Paris-Sud, Faculte des Sciences-Mathematiques, Orsay, 1979, pp. 38-48.
Y Meyer, Wavelets and operators. Analysis at Urbana, Vol. I: Analysis in
Function Spaces (Urbana, IL, 1986-1987) (E. Berkson and T. Peck, eds.),
London Math. Soc. Lecture Note Ser., vol. 137, Cambridge University Press,
Cambridge, 1989, pp. 256-365.
Y Meyer, Wavelets and fast numerical algorithms. Handbook of Numerical
Analysis, Vol. V: Techniques of Scientific Computing, Part 2 (P.G. Ciarlet and
J.L. Lions, eds.), North-Holland, Amsterdam, 1997, pp. 639-713.
Y Meyer, Compression des images fixes, Gaz. Math. (2005), no. 103, 9-23.
Y Meyer and R. Coifman, Wavelets: Calderon-Zygmund and Multilinear Operators, Cambridge Studies in Advanced Mathematics, vol. 48, Cambridge
University Press, Cambridge, 1997, translated from the 1990 and 1991 French
originals by David Salinger.
E.W. MontroU, Random walks on lattices. Stochastic Processes in Mathematical Physics and Engineering (R. Bellman, ed.), Proceedings of Symposia in
Applied Mathematics, vol. 16, American Mathematical Society, Providence,
1964, reprinted 1980, pp. 193-220.
VV. Nekrashevych, Cuntz-pimsner algebras of group actions, J. Operator
Theory 52 (2004), 223-249.
E. Nelson, Regular probability measures on function space, Ann. of Math. (2)
69 (1959), 630-643.
E. Nelson, Feynman integrals and the Schrodinger equation, J. Math. Phys. 5
(1964), 332-343.

References
pSFel67]
[Nel69]
pSfel73]
[Nev65]

[Nev75]

[OED]
[PaSW99]

[PaPo90]

[PaSc72]

[Per07]
[Pop89]
[PoSt70]
[Rad99]
[Rev84]
[RiShOO]
[Riel8]
[Rit79]
[RoWiOO]

[RoSh97]
[RoSh98]

247

E. Nelson, Dynamical Theories of Brownian Motion, Princeton University

Press, Princeton, NJ, 1967.
E. Nelson, Topics in Dynamics, I: Flows, Mathematical Notes, Princeton University Press, Princeton, NJ, 1969.
E. Nelson, Construction of quantumfieIds from Markofffields, J. Funct. Anal.
12 (1973), 97-112.
J. Neveu, Mathematical Foundations of the Calculus of Probability, HoldenDay Series in Probability and Statistics, Holden-Day, San Francisco, 1965,
translated from the French by A. Feinstein.
J. Neveu, Discrete-Parameter Martingales, revised ed., North-Holland Mathematical Library, vol. 10, North-Holland, Amsterdam, 1975, translated from
the French by T. P. Speed.
Oxford English Dictionary, second ed., Oxford University Press, Oxford,
1989.
M. Papadakis, H. Sikic, and G. Weiss, The characterization of low pass filters
and some basic properties of wavelets, scalingfunctions and related concepts,
J. Fourier Anal. Appl. 5 (1999), 495-521.
W. Parry and M. PoUicott, Zeta functions and the periodic orbit structure
of hyperbolic dynamics, Asterisque, vol. 187-188, Societe Mathematique de
France, Paris, 1990.
K. R. Parthasarathy and K. Schmidt, Positive Definite Kernels, Continuous
Tensor Products, and Central Limit Theorems of Probability Theory, Lecture
Notes in Mathematics, vol. 272, Springer-Verlag, Berlin-New York, 1972.
O. Perron, Zur Theorie der Matrizen, Math. Ann. 64 (1907), 248-263.
G. Popescu, Isometric dilations for infinite sequences of non-commuting operators. Trans. Amer. Math. Soc. 316 (1989), 523-536.
R.T. Powers and E. Stormer, Free states of the canonical anti-commutation
relations, Comm. Math. Phys. 16 (1970), 1-33.
C. Radin, Miles of tiles. Student Mathematical Library, vol. 1, American Mathematical Society, Providence, 1999.
D. Revuz, Markov Chains, second ed., North-Holland Mathematical Library,
vol. 11, North-Holland, Amsterdam, 1984.
S.D. Riemenschneider and Z. Shen, Interpolatory wavelet packets, Appl.
Comput. Harmon. Anal. 8 (2000), 320-324.
F. Riesz, Uber die Fourierkoefiizienten einer stetigen Funktion von
beschrdnkter Schwankung, Math. Z. 2 (1918), 312-315.
G. Ritter, On Kakutani 's dichotomy theorem for infinite products of not necessarily independent functions. Math. Ann. 239 (1979), 35-53.
L.C.G. Rogers and D. Williams, Diffusions, Markov processes, and martingales. Vol. 1: Foundations, Cambridge Mathematical Library, Cambridge University Press, Cambridge, 2000, reprint of the second ed., Wiley, Chichester,
1994.
A. Ron and Z. Shen, Affine systems in Z2(R'^): the analysis of the analysis
operator, J. Funct. Anal. 148 (1997), 408-447.
A. Ron and Z. Shen, Compactly supported tight affine spline frames in
Z2(R^), Math. Comp. 67 (1998), 191-207.

248

References

[RoShOO]
[RoSh03]
[RSTOl]

[Rud87]
[Rue69]
[Rue89]
[Rue94]

[Rue02]
[Sat99]

[SchCC]
[Schl871]
[SeCh67]
[ShKu04]

[Sha49]
[She98]
[StHS+99]

[Shu04]
[Shu05]
[Sim79]
[Sko61]

[Sko65]

A. Ron and Z. Shen, The Sobolev regularity of refinable functions, J. Approx.

Theory 106 (2000), 185-225.
A. Ron and Z. Shen, The wavelet dimension function is the trace function of a
shift-invariant system, Proc. Amer. Math. Soc. 131 (2003), 1385-1398.
A. Ron, Z. Shen, and K.-C. Toh, Computing the Sobolev regularity ofrefinable
functions by the Arnoldi method, SIAM J. Matrix Anal. Appl. 23 (2001), 5776.
W. Rudin, Real and Complex Analysis, thu"d ed., McGraw-Hill, New York,
1987.
D. Ruelle, Statistical Mechanics: Rigorous Results, W. A. Benjamin, Inc.,
New York-Amsterdam, 1969.
D. Ruelle, The thermodynamic formalism for expanding maps, Comm. Math.
Phys. 125 (1989), no. 2, 239-262.
D. Ruelle, Dynamical Zeta Functions for Piecewise Monotone Maps of the
Interval, CRM Monograph Series, vol. 4, American Mathematical Society,
Providence, 1994.
D. Ruelle, Dynamical zeta functions and transfer operators. Notices Amer.
Math. Soc. 49 (2002), 887-895.
K.-I. Sato, Levy Processes and Infinitely Divisible Distributions, Cambridge
Studies in Advanced Mathematics, vol. 68, Cambridge University Press, Cambridge, 1999, translated from the 1990 Japanese original, revised by the author.
Phil Schniter, Two-branch quadvalue mirrorfilterbank (QMF), see [WWW4].
E. Schroder, Uber iterirte Functionen, Math. Ann. 3 (1871), 296-322.
A. Selberg and S. Chowla, On Epstein's zeta-function, J. Reine Angew. Math.
227(1967), 86-110.
S.C. Shah and A. Kusiak, Data mining and genetic programming based
gene/SNP selection. Artificial Intelligence in Medicine 31 (2004), no. 3, 183196.
C.E. Shannon, Communication in the presence of noise, Proc. Inst. Radio Engineers 37 (1949), 10-21.
Z. Shen, Refinable function vectors, SIAM J. Math. Anal. 29 (1998), 235-250.
V. Strela, RN. Heller, G. Strang, R Topiwala, and C. Heil, The application
ofmultiwaveletfilterbanks to image processing, IEEE Transactions on Image
Processing 8 (1999), 548-563.
F. Shultz, Dimension groups for interval maps II: The transitive case, preprint,
math.DS/0405467.
F. Shultz, Dimension groups for interval maps. New York J. Math. 11 (2005),
477-517.
B. Simon, Functional Integration and Quantum Physics, Pure and Applied
Mathematics, vol. 86, Academic Press, New York, 1979.
A.V. Skorokhod, Issledovaniya po teorii sluchainykh protsessov (Stokhasticheskie differentsialnye uravneniya i predelnye teoremy dlya protsessov
Markova), Izdat. Kiev. Univ., Kiev, 1961.
A.V. Skorokhod, Studies in the Theory of Random Processes, AddisonWesley, Reading, Mass., 1965, Translated from the Russian by Scripta Technica.

References
[Song05]
[Song05a]
[Song05b]
[Spi76]

[StNg96]
[Str98]

[StrOO]
[Str05]
[Stro96]
[StroOO]

[Stro05]
[Tao96]
[Vai93]
[VaAkO 1 ]

[VeKo95]
[ViMu95]

[VoDN92]

[Waln02]
[WalOl]

249

M.-S. Song, Wavelet image compression, Ph.D. thesis, The University of Iowa,
2005.
M.-S. Song, Wavelet image decomposition,
http://www.biblethumper.org/~msong/Images/waveletdec.html.
M.-S. Song, http://www.biblethumper.org/~msong/Images/.
F. Spitzer, Principles of Random Walk, second ed.. Graduate Texts in Mathematics, vol. 34, Springer-Verlag, New York, 1976, first softcover printing,
2001.
G. Strang and T. Nguyen, Wavelets and Filter Banks, Wellesley-Cambridge
Press, Wellesley, Massachusetts, 1996.
R.S. Strichartz, Remarks on: "Dense analytic subspaces in fractal L^-spaces "
[J. Analyse Math. 75 (1998), 185-228] by P.E. T. Jorgensen and S. Pedersen,
J. Analyse Math. 75 (1998), 229-231.
R.S. Strichartz, Mock Fourier series and transforms associated with certain
Cantor measures, J. Anal. Math. 81 (2000), 209-238.
R.S. Strichartz, Convergence of Mock Fourier series, preprint, Cornell University.
D.W. Stroock, Gaussian measures in traditional and not so traditional settings. Bull. Amer. Math. Soc. (N.S.) 33 (1996), 135-155.
D.W. Stroock, An Introduction to the Analysis ofPaths on aRiemannian Manifold, Mathematical Surveys and Monographs, vol. 74, American Mathematical Society, Providence, RI, 2000.
D.W. Stroock, An Introduction to Markov Processes, Graduate Texts in Mathematics, vol. 230, Springer, Berlin, 2005.
T. Tao, On the almost everywhere convergence of wavelet summation methods,
Appl. Comput. Harmon. Anal. 3 (1996), no. 4, 384-387.
P.P. Vaidyanathan, Multirate Systems and Filter Banks, Prentice-Hall Signal
Processing Series, Prentice Hall, Englewood Cliffs, NJ, 1993.
P.P. Vaidyanathan and S. Akkarakaran, A review of the theory and applications
of optimal subband and transform coders, Appl. Comput. Harmon. Anal. 10
(2001), no. 3, 254-289.
M. Vetterli and J. Kovacevic, Wavelets and Subband Coding, Prentice-Hall
Signal Processing Series, Prentice Hall PTR, Englewood Cliffs, NJ, 1995.
B. Vidakovic and P. Mueller, Wavelets for kids: A tutorial introduction,
preprint, 1995, Duke University, Institute of Statistics and Decision Sciences,
and [WWW/wiki].
D.V. Voiculescu, K.J. Dykema, and A. Nica, Free Random Variables: A Noncommutative Probability Approach to Free Products with Applications to Random Mmatrices, Operator Algebras and Harmonic Analysis on Free Groups,
American Mathematical Society, Providence, 1992.
D.R Walnut, An Introduction to Wavelet Analysis, Applied and Numerical
Harmonic Analysis, Birkhauser, Boston, 2002.
P. Walters, Convergence of the Ruelle operator for a function satisfying
Bowen 's condition. Trans. Amer. Math. Soc. 353 (2001), 327-347.

250
[Wic93]

References

M.V. Wickerhauser, Best-adapted wavelet packet bases. Different Perspectives on Wavelets (San Antonio, TX, 1993) (I. Daubechies, ed.), Proc. Sympos. Appl. Math., vol. 47, American Mathematical Society, Providence, 1993,
pp. 155-171.
[Wic94]
M.V. Wickerhauser, Adapted Wavelet Analysis from Theory to Software, IEEE
Press, New York, A.K. Peters, Wellesley, MA, 1994.
[Wil91]
D. Williams, Probability with Martingales, Cambridge Mathematical Textbooks, Cambridge University Press, Cambridge, 1991.
[WWW 1 ]
Lifting-scheme-based wavelet transform,
http://wvs^w.electronicsletters.com/papers/example/paper.html; [WWW2] and
[WWW3] are Figures 5 and 6, respectively.
[WWW2]
http://www.electronicsletters.com/papers/example/Psubsampletransform.gif;
Figure 5 in [WWWl].
[WWW3]
http://www.electronicsletters.com/papers/example/Pinversetransform.gif;
Figure 6 in [WWWl].
[WWW4]
Phil Schniter, Two-branch quadvalue mirrorfilterbank (QMF),
http://cnx.rice.edu/content/ml0426/latest/. Connexions module.
[WWW5]
Ma Yi, Basic image processing demos (for EECS20),
http://robotics.eecs.berkeley.edurmayi/imgproc/.
[WWW/wiki] http://en.wikipedia.org/wiki/Wavelet.
[YaHK97]
M. Yamaguti, M. Hata, and J. Kigami, Mathematics of Fractals, Translations of Mathematical Monographs, vol. 167, American Mathematical Society, Providence, 1997, translated from the 1993 Japanese original by Kiki
Hudson.
[YiCC]
Ma Yi, Basic image processing demos (for EECS20), see [WWW5].
[Zyg32]
A. Zygmund, On lacunary trigonometric series. Trans. Amer. Math. Soc. 34
(1932), 435^46.

Symbols

Reminder: In the symbol list and in the chapters, fimction spaces are defined with
respect to various integrability conditions. For a fimction / on a space X, absolute integrability refers to | / | , i.e., to the absolute value o f / , and to a prescribed
(standard) measure on X. This measure on X is often implicitly understood, as is its
cr-algebra of measurable sets. Examples: If the space X is M^, the measure will be
the standard (i-dimensional Lebesgue measure; for the one-torus T (i.e., the circle
group), it will be normalized Haar measure, and similarly for the J-torus T^; for
X = Z, the measure will simply be counting measure; and for X3 (the middle-third
Cantor set), the measure will be the corresponding Hausdorff measure hs of fractal
dimension s = log3(2). In each case, we introduce Hilbert spaces of Z^^fimctions,
and the measure will be understood to be the standard one. Same convention for the
other Z^-spaces!
A{i\,...,
in) : the cylinder set
{COG Q I coi =/i,...,ft> = / } ,
i.e., the set of infinite strings
CO = (coi,...) specified by
coi = /i, ...,ft) = in
43,47,85
21 : the C*-algebra of the canonical
anticommutation relations
139, 140, 141, 154
2l : family of algebras increasing in
the index n,{f eC (Q)
I f(co)

44,45,139

f(cOuCD2,...,COn)}

B (H) : bounded linear operators on a

Hilbert space H
183,218
B : Borelcr-algebra
6,40, 53, 204
BQ : Borel cr-algebra on Q
115
C (Q) : continuous fimctions on Q
7, 27, 44,46
CAR : canonical anticommutation
relations
138-140, 154

252

Symbols

C : the complex numbers

25,43, 46, 48-52, 57, 61, 140, 184,
210,214,220

^min, hp minimal eigenfimction for

Rw
100-102, 105-107

C^ : A:-dimensional complex vector

space
31

/?3 : minimal eigenfimction corresponding to the scale-3 stretched

Haar wavelet
107

C3, C4 : Cantor sets

195, 197-199
V : maximal abelian subalgebra
154
Vy : smallest cr-algebra with respect
to which Y is measurable
xxiv
V(p : closed linear span
209,217
ex (t) := e^^^^\ ek {z) = z^ : Fourier
basis functions
61,71-79,130,192,198

special matrix element generators

135, 139, 183
T : (7-algebra
37
J^n '- system of cr-algebras
37
GMRA : generalized multiresolution
analysis
114
h : special (harmonic) fimction, a
Perron-Frobenius eigenfimction
for 7?pF, a measurable function on
X such that Rwh = h
xxxiv, 11, 19, 49, 55, 92, 101, 105,
116

hs : Hausdorff measure
14,17
H : some (complex) Hilbert space
14, 17, 114-117, 131, 136, 140,
169-170, 180-184, 189, 190,
196,210,218-219
/ : identity operator or identity matrix
(see also ty)
115,131,135,136, 139-141, 184,
211,214-219
/ : index set
172, 186, 189-190
/ : multiindex
165-168
IFS : iterated fimction system
XXXV, xliv, 5, 14, 15, 34, 35, 67, 70,
80,84,99,152,182
ind lim 2l : inductive limit of an
ascending family of algebras
139
/C : some Hilbert space
161, 169, 170, 172, 189
^ : all absolutely summable sequences
66

Symbols
^^(N), ^^(No) : all square-summable
sequences indexed by N, or by No
31,140,162,182,190,193,197

253

i?- {X, B, ju), L^ (ju) : all squareintegrable functions on the

cr-finite measure space (X, B, fi)
31,72

i?- (Z) : all square-summable

sequences indexed by Z
30,32, 117, 136, 143, 191, 193,
200-202,213,219

Z ^ (T) : all essentially bounded and

measurable functions on T
95, 163, 190, 191

i^ (JQj ^^ all square-summable

sequences indexed by a set X or
other index set
31,66, 143, 160, 161, 168, 170,
172, 184, 189

L^ {X) : all essentially bounded and

measurable functions on X with
respect to the standard measure
and cr-algebra of measurable
subsets
9,43,44,49,115

L ^ (M) : all absolutely integrable

functions on M
130

MRA : multiresolution analysis

6, 181, 194, 198

X^ (E) : all square-integrable functions

onR
xxxii, 4, 5, 10, 12-16, 29, 33, 65,
71,87,91, 103-105, 109, 112,
114,129, 130,158, 162-163, 165,
181,190-194, 198

m : function on T representing a digital

filter
4,10,114

L^ (R^) : all square-integrable

functions on R^
4,22,97,109,142,229,230

mo, mo : low-pass filter

111,124-129

L^ (T) : all square-integrable functions

onT
66, 132, 136, 162, 167, 182, 190,
192, 193, 196, 197,210,213,214,
219
I ^ () : all square-integrable fimctions
on some specified set with its
standard measure
14, 17,72,77,79, 112, 132, 136,
191, 196-198

rrii ' multiband filter functions

123,126,190,191,194,211

mi. mi : high-pass filter

111,124-129
M : multiplication operator
213
Mn = Mn (C)
matrices
139

n X n complex

M2 := M2 0 0 M2
139

254

Symbols

N : the positive integers or natural

numbers
6,11,135
No : = { 0 , 1 , 2 , . . . } = {0}UN
5, 11,59,66,85, 116, 117, 159,
160, 164,166, 182, 186, 188
ONB : orthonormal basis in a Hilbert
space
13,15, 16,56,71,72,76,77, 103,
104, 140-19S passim
On,ON,02
: Cuntz algebra
131, 136, 139-152, 154, 158,
161-164, 167,170, 176, 179-184,
189, 190, 192, 194, 196, 203, 205,
211,214,218,219
Px : transition probability initialized
at x; measure on Q such that
^ . [ / ] = ^i"^[/]forall/E2l
5-11, 19,26,37,43,44,62, 100
Px ( \ ') : conditional probability
initialized at x
51
Px (No) : path-space measure of the
natural numbers No as subset of Q
'= Z Px({co(k)}lwhQrQ
keNo
Px({co(k)})

n^(i

=
COp

' ' '

''COl

W)-

p=\
m=l

11,18,60,71,78,86,88-91,100,
102,116
Px C^) : path-space measure of the
integers Z as subset of Q
11,18,60,64,71,90,116

Px [f] transition probability

initialized at x and conditioned by
n coordinates
(Q)I,...,CO)P=1

' f(cou...,COri),

f^^n

21,44,45,63,64,116,122
Pos {H) : operator with spectrum
contained in [0, oo)
114,115,117
Rw, P Perron-Frobenius-Ruelle
transfer operator

(Rrrf)(x)=

W(y)f(y)

(7(y)=x

xxxiv, 9, 11, 19,26,43,45,49,

51-57,61,64,66,76,86,91,95,
100,101,105,115,116,200
R : the real numbers
33, 10, 14, 195, 199
71 : envelope of a fractal
195-199
s : Hausdorff dimension
14,17,71,72,77
S := F* : adjoint operator
67
Si, S*,Ti, T* : the operators (isometrics) and their adjoints (with
stars) in a representation of the
Cuntz relations (i.e., of the Cuntz
algebra)
131, 132, 135, 161, 181, 182, 184,
201,211,213,214,219
T:={zeC\\z\
= l} :
circle group, or one-torus
^ E / Z ^ [0, 1)
25,32,60,61,190,204

Symbols
Ui : dyadic scaling operator
200

Znix,co) : canonical martingale

50,51

V : cocycle, i.e., a measurable fiinction

on X X Q such that
V {TCO^X\ (a;2,cy3,...)) = V {x\(o)
43, 49, 92

Z : the integers
5, 19, 22, 59, 66

Vo,V\,Vn : resolution subspaces

22,33,104,111,123-128
F; : representation of Cuntz algebra
n^-\91 passim
W : a measurable function
X->[0,1]
xxxiv, 7-12, 17-21, 36, 41-45, 48,
49, 51-57, 61-66, 69, 71, 76, 77,
84-91,101,104, 105,112-115,
117,140,141,162
Wfi^ Wn detail subspaces
33,123-128
X : a fractal
110
X : a measurable space
xxxiv,6,7,39,47, 115, 117
X, X3,X4, X4 : Cantor sets
14,21,71-80,176
Xk {co) = cok : coordinate fiinctions on
a probability space
50

255

%2 := {0,1} : cyclic group of order 2

27
Zjv : cyclicgroupof order TV
: = Z / 7 V Z ^ { 0 , 1,...,7V-1}
41, 43, 44, 49, 52, 60, 86, 88, 116,
211,214-219
d : Kronecker delta fiinction
15,46,47,104,116,131,139,163,
164,183,192,211,216,219
^0 Dirac mass at x = 0
102, 105
X : Fourier frequency
71-78,112,198
A : index set for a Fourier orthonormal
basis
71-79, 198
yi : the Haar measure, or other measure
specified in the text
14, 41, 43, 61, 72, 77, 79, 136-139,
167, 168, 195, 198
// : multiplicity fimction
114,117
IJL oa~^ : is the measure given by

{^ioa-'){B):=^i{a-'{B))
{X,B) : a set X with a cr-algebra
B of measurable subsets
6,40,84,114,115
^.^.^iiTTt . Fourier variable
32

52,72
V : Perron-Frobenius-Ruelle measure,
or other measure specified in the
text
xxxiv, 52-54, 101, 105

256

Symbols

p : representation or state
47,48,139-141,154
a : one-sided shift, an onto map
(actually endomorphism) X ^ X
such that #cr~^ ({x}) is constant
xxxiv, 6-8, 12, 17, 41, 45, 47,
51-54,62,64,71,74-76,84,
89-91,101, 114, 115, 159, 160,
170-172, 184, 186, 188
G^ : shift on Q
47,51,52

co{k) : representation in Q of
A: e No: If A: =
co\ + (02N H
+ conN'"'^
is the Euclid A/'-adic representation, co{k) :=
(cou . . . , con, 0, 0, 0, . . . )
00 string of zeroes

11, 18,77,79,92, 101, 122, 130,

135, 137, 138, 140
Q : probability space
:={0,1,...,7V-1}^
= n{o,i,...,A^-i}
N

a~^ {B) : pre-image under the mapping a \= {x e X \ G (x) e B}

6,41,72,52,84
(cr^)

: pre-image under the mapping

a"
52
ro,...,T7v-i : branches of cr~^, maps
X > X such that a o n = idx
7,41,47,52,72,89,115,159
r p : branches of (cr^)~
47,48, 52
(p : scaling fiinction
3,10,12,13,15,23,102,103,114,
134

= all fimctions:
N^{0,1,...,7V-1}
=

{{C0UC02,...)

|a;,G{0,l,...,7V-l}}
5, 7, 11, 18, 20-37, 43, 46, 47, 49,
69, 85, 135
(Q, B, v) : probability space
56, 203
0 : one-sided infinite string of zeroes
= ( 0, 0, 0, . . . ) G Q
00 string of zeroes

12,85,116,130
{0} : the set with the one element 0
12,85,116

^0, ^ h ^2, . wavelet packet system

112,113,118-122,168,191

UT^ : identity operator (see

also / )
114, 115, 117, 122, 160, 161, 181,
182,184

X : characteristic fimction
14, 16,47

11 : constant fimction equal to 1

46,61,64,136

y/ : wavelet fimction
13,16,23,102,103,134

*-algebra, *-isomorphism
221

Wn,k '- wavelets

15

*-automorphism
211

Symbols
V : lattice operation applied to closed
subspaces in a Hilbert space: the
lim sup lattice operation
169, 181
A : lattice operation applied to closed
subspaces in a Hilbert space: the
lim inf lattice operation
169, 181
0 : empty set
171, 172, 185,

257

<g) : tensor product

139, 158, 161-163, 165, 170-172,
180-183, 189, 190, 193, 194, 197
0 : relative orthogonal complement
169
X : Cartesian product
11,43,49,52,88, 117, 130, 164,
166,185,188,218
# : counting fimction
6,41,72,101,159,184,196

E : closureof aset
44, 172

( I ) : inner product
16,75,77,79,104,114,140

^ : Fourier transform (of the scaling

function (p)
10,114,111

1 ) : Dirac vector
160-163, 171, 182, 184, 185, 189,
193, 194

< : relatively absolutely continuous

(relation between measures)
50, 53

[ , ) : interval closed to the left and

open to the right
41, 61, 63, 65, 136-138, 165, 166,
167, 192

: up-sampling
124,132,213,214
: down-sampling
124-128, 132, 133,212,213,215
0 : direct (orthogonal) sum
112,172,218

[, ) : segment of No
165-167, 186,188
[ , ] : interval closed at both ends
7, 11, 13, 16-18,47,62-66,71,77,
84,89-92, 102, 105, 112, 125,
130, 135-139, 195

Index

Comments on the use of the index: Some terms in the index may appear in the
text in a sHght variant, or variation of the actual index-term itself. For example, we
will have terms in the index referring to "theorem so and so." But when we use the
Stone-WeierstraB theorem, I just say Stone-WeierstraB. The word "theorem" will be
suppressed. It is implicitly understood.
Similarly, I often just say, "by domination" (or some variant thereof), when I mean,
"by an application of the dominated convergence theorem," or more fiilly: "By
Lebesgue's dominated convergence theorem." It will be the same theorem whether
the name is abbreviated or not.
For Fubini, the word "theorem" may be implicitly understood. Guido Weiss has made
a verb out of it: "Fubinate" means "to exchange the order of two integrals."
Similarly, the name Fatou often is used to mean "Fatou's lermna" (the one about
lim inf). For some reason poor Fatou only got credit for a lemma. But I do not mind
upgrading him to a theorem, although "Fatou's theorem" usually refers to the one
about existence a.e. of boundary values of bounded harmonic functions. I usually
call that one "the Fatou-Primalov theorem."
^-random variable, see random variable. Aabelian, see algebra, abelian; group, abelian;
maximal abelian subalgebra
absolutely continuous, see measure,
absolutely continuous
j..^
^. j . . ^ ^ j.-x
adjomt, see matrix, adjomt; operator, adjomt
a.e. convergence, see convergence, a.e.
affine
fractal, xxiii, 3, 5, 15, 22, 26, 70-72,
77, 80, 180, 194, 198

iterated function system, 5, 14, 15, 25,

" ' ' ^0-72, 80, 81
iteration, 21
map, xxui, 1, 3, 5, 81, 195
self tiling, see tiling, self-affine
wavelet frame, see frame, afifine
wavelet
^^^^^^^ ^^..^ ^ ^^ 3^ 27, 43, 44, 47, 54,
140, 176, 251, 252
abelian, 110, 138, 139
C*-, see C*-algebra

260

Index

algebra
CAR-, see CAR-algebra
Cuntz, see Cuntz algebra
Cuntz-Krieger, see Cuntz-Krieger algebra
dense sub-, 46
fermion, see fermion algebra
matrix, 139, 210
maximal abelian sub-, see maximal
abelian subalgebra
non-abelian, xxviii, 138, 139, 155,218
operator, see operator algebra
(J-, see o--algebra
*-, 221
sub-, 44, 139, 154, 155
algebraic structures
representations of, see representation of
algebraic structure
algorithm, xvi, xix, xx, xxv, xxvi, xxxiii,
xxxiv, xxxvi, xliii, 3, 4, 10, 33, 35,
124, 125, 128, 147, 148, 164, 206,
210,211,215,223-226,230,231
cascade, see cascade
Euclid's, XX, 11, 40, 63, 69, 85, 92, 164,
166
Gram-Schmidt, xv, xxvi
matrix, xxvi, xxviii, xxxii, 142, 147,
157, 226, 230, see also matrix step in
algorithm
pyramid, vi, xx, xxxiv, xxxv, xlv, 33,
111-113, 122, 123, 125, 128, 129,
134, 148, 157-159, 182, 225, 227, see
also pyramid
recursive, xxvii, xxxii, xxxiv, xxxv, xlv, 6,
109, 147, 157, 226
subdivision, xxix, xxxi, xxxii, xxxv, 124,
142, 227
wavelet, xiii, xvi, xix, xxvii-xxix,
xxxi-xxxiii, 25, 33, 110, 123, 125,
133, 142, 147, 148, 151, 152, 156,
206, 210, 215, 226, 227, 230
7V-adic, 125, 126
wavelet-like, xxv
wavelet packet, xxxiv, xxxv, 3, 34, 123,
125, 126
alias, 229, 230
ambient

.bases
Euclidean space, see space, Euclidean,
ambient
function space, see space, function,
ambient
Hausdorff measure, see measure,
Hausdorff, ambient
Hilbert space, see space, Hilbert,
ambient
analysis
data, xxv
(engineering), xvi, xxii, xxvi, 124, 132,
148, 205-207, 214, 227, see also
frequency band; perfect reconstruction; synthesis; signal analysis; signal
processing
Fourier, see Fourier analysis
fractal, see fractal analysis
harmonic, see harmonic analysis
(mathematics), xvii, xxvii-xxxi, xxxiii,
xliv, 3, 6, 9, 22, 26, 33-35, 37, 39,
59, 80, 81, 84, 87, 98, 206, see also
Fourier analysis; harmonic analysis;
spectral analysis
multiresolution, see multiresolution
analysis
numerical, xxvii, xxxii, 229
spectral, see spectral analysis
stochastic, 35
wavelet, see wavelet analysis
approximation, xxvi, 4, 6, 79, 90, 107, 109,
158,226
cascade, see cascade approximation
theory, xxix
atomic, see measure, atomic
attractor, 34, 187
automorphism
*-,211
jB-measurable, see measurable, BB-measure, see measure, Bband-limited wavelet, see wavelet,
band-limited
base-point representation, see representation,
base-point
bases, see basis

Index
basis
basis, XV, xvi, xx, xxvi, xxxi, xxxvi, xliv, 2,
5, 9, 22, 30, 59, 67, 70, 71, 87, H I ,
143, 146, 148, 157, 162, 179, 182,
184,228,229
bi-orthogonal, 143
canonical, 143, 160, 202
dual, 143, 144
Fourier, xvi, 67, 144, 158, 168, 193, 252,
255
fractal,21,26, 70-72, 79
frame, 28, 29, 143, 229
function, xv, xxxvi, 104, 106, 112, 113,
129, 166
localized, 67, 80, 157, 158, see also
localization property of vs^avelet bases
orthogonal, xxvi, 36, 69, 70, 87
orthonormal, xxxi, 13, 15, 16, 22, 26, 28,
29, 32, 36, 55, 56, 65, 71, 72, 74, 76,
77, 79, 99, 103-106, 130, 139, 140,
143, 144, 149, 150, 162, 163, 165,
166, 168, 177, 182, 184, 185, 189,
190, 192-194, 197, 198, 228, 229,
254, 255
Parseval, 103
permutation of, 182
recursive, 179
transformation, 166
wavelet, xvi, xix, xxvi, xxxiii, xliii, xliv,
13, 15, 22, 23, 67, 72, 80, 87, 99, 103,
142, 176, 179, 180, 187-189, 208,
229, see also localization
dyadic, 99, 202
fractal, 180
wavelet-like, 109
Bernoulli product measure, see measure,
/>-Bemoulli-product
Bethe lattice, see lattice, Bethe
bi-orthogonal, see basis, bi-orthogonal
black box, xx, xxi
Borel
cross section, 183
measure, see measure, Borel
fj-algebra, see (7-algebra, Borel
subsets, 135, 167, 199

261

Carleson
M.Bom, 176,205,236
boundary
for harmonic function, see harmonic
function, boundary for
representation, 19, 48
value, 21, 43, 48
branch mapping, 256
measurable, 41
-fold, 184, 185, 188, 189
2-fold, 186, 187
branching, 5, 158, 161, see also random
walk on branches
dyadic system, 160
system, 159, 171
C. Brislawn, xxxiii, xliv, 22, 234
Brownian motion, xxvi, xxvii, 56, 57
fractal, xxiii
fractional, xxvii, 57
5-fractal, xxiii
C*-algebra, xxix, 5, 6, 131, 138-140, 142,
151, 154, 155, 183,211,251
canonical anticommutation relations,
5, 138-140, 154, 251, see also
CAR-algebra
canonical basis vector, see basis, canonical
Cantor, 1, 69, 179
construction, 69, 70, 74
group, see group. Cantor
measure, see measure. Cantor
's example, see measure. Cantor
scaling identity, see scaling identity,
Cantor
set, 2, 5, 15, 71-73, 252, 255
conjugate, 73, 75-77
duality for s, 69
middle-third, 2, 5, 14, 15, 21, 25, 27,
69-71, 73, 74, 80, 176, 189, 195, 251
quarter, 5, 26, 72-77, 79, 198
scale-4, see Cantor set, quarter
CAR-algebra, 5, 138, 139, 154, 155
representations of, see representation of
CAR-algebra
L. Carleson, 32

262

Index

cascade
cascade, 9, 130, see also closed subspaces,
nested family of
approximant, 134
approximation, 4
Cauchy product, 87, 206
closed linear span, 15, 169, 200, 207, 209,
252
closed subspaces
nested family of, xviii-xx, xxii, xxix, 9,
33
cocycle, 43, 48-52, 91, 92, 255
identity, 19,20
property, 49
coefficient, 168
autocorrelation, 104
filter, 4
Fourier, xvi, xxii, 87, 97
masking, 4, 10, 16, 23, 87, 91, 114, 123,
130, 131,135,227
matrix, 131
operator, 114
response, 10
wavelet, xxvi, 23, 25, 202, 225, 227, 230
wavepacket, 168
A. Cohen, xxxii, xxxiii, 33, 87
co-isometry, 162, 216
combinatorial
probability theory, 5, 154
tree, 6, 111, 129
commute, 94, 135, 234, 235, see also
non-commutative setting
compact, 1, 8, 14, 25, 27, 35, 43, 69, 71, 83,
92, 98, 149, 195, 204
abelian group, see group, abelian,
compact
Hausdorff space, see space, compact
Hausdorff
operator, see operator, compact
support, see wavelet, compactly
supported
conditional expectation, 10, 57
conjugate Cantor set, see Cantor set,
conjugate
conjugation, 150

. decision tree
consistency, xx, 80, 122, see also
Kolmogorov consistency
continued fractions, 41
convergence, xxviii, xxx, 4, 6, 9, 10, 18, 80
a.e., 32, 50, 56, 92, 95, 96
dominated, 78, 89
dominated theorem, see theorem,
dominated convergence
martingale theorem, see theorem,
martingale convergence
of infinite product, 5, 8, 11, 17-19, 21,
60, 85
pointwise, 4, 5, 17-19
countable family of o -algebras, see
cr-algebras, countable family of
J. Cuntz, xxii, 183
Cuntz
algebra, xxix, 5, 6, 22, 131, 136, 152,
154, 155, 158, 161, 162, 179-183,
187, 196, 205, 208, 210, 222, 254,
255, see also representation of Cuntz
algebra
^Krieger algebra, xxix
relations, xxii, 6, 132, 155, 160, 161,
174, 179-182, 201, 203, 211, 214,
216, 219, 221, 222, 227, 231, 254, see
also representation of Cuntz algebra
representation, see representation of
Cuntz algebra
system, 208, 221
cycle, 26, 87
cyclic group, see group, cyclic
cylinder set, 43, 47, 78, 85, 115, 139, 251

data mining, xvii, xix, xxiv, xxv, xxviii, 107,

224
I. Daubechies, xxxii, xxxiii, 10, 33, 87
Daubechies
scaling function, see scaling function,
Daubechies
wavelet, see wavelet, Daubechies
decision tree, xxxv

Index
decomposition.
decomposition, 166
Karhunen-Loeve, see theorem,
Karhunen-Loeve
orthogonal, 172
Schmidt's, see theorem, Schmidt's
wavelet, xxxii, 190
Wold, 169
derivative
Radon-Nikodym, 50, 53
detail space, see space, detail
differentiability, 6
differentiable, xxxii-xxxiv, 14, 135
dimension, 4, 33, 117, 154, 210
fractal, 14, 195,251
Hausdorff, xxiii, 2, 71, 72, 74, 77, 176,
195,251,254
scaling, 72
Dini regularity, see regularity, Dini
G. Dirac, 235
P.A.M. Dirac, 58, 234, 235
Dirac
mass, 26, 102,255
notation, 55, 58, 182, 186, 257
discrete wavelet transform, see wavelet
transform, discrete
distribution, xxiii, 130, 142, 168, 204
Gaussian, 56, 203, see also random
variable, Gaussian
D. Donoho, xxxiii
J. Doob, xxiv, xxvi
down-sampling, see sampling, downdual
basis, see basis, dual
filter, see filter, dual
Fourier, see Fourier dual
high-pass filter, see filter, dual
high-pass
lattice, see lattice, dual
low-pass filter, see filter, dual low-pass
variable, xxi, 206, 212
wavelet, see wavelet, dual
duality, 69, 205, 207, 212
for Cantor sets, see Cantor set, duality
for

263

expansion
Fourier, 35, 60, 81, 207, 209, 210
particle-wave, 131
time-frequency, 207, 212
D. Dutkay, xliii, 57, 72, 87, 97, 194
dyadic
branching system, see branching,
dyadic system
fractional subinterval, 166-168
Haar wavelet, see wavelet, Haar,
dyadic
pyramid, see pyramid, dyadic
rationals, 66, 90, 167
representation, 166
scaling, see scaling, dyadic
subdivision, see subdivision, dyadic
tiles, see tiling, dyadic
wavelet, see wavelet, dyadic
wavelet packet, see wavelet packet,
dyadic
dynamics, xxix, xxx, xxxvi, xliv, 9, 34, 37,
42, 168
complex, 25, 34
symbolic, xxxv, 34, 182
eigenfunction, 19
minimal, 15, 19, 99-102, 105, 252
Perron-Frobenius, 26, 97, 252
eigenspace, 19
Perron-Frobenius, 107
eigenvalue, 9, 11, 19, 77, 100, 107, 116
endomorphism, xxxiv, 4, 52, 91, 101, 184,
256
engineering, xiii, xxviii, xxxi, xxxii,
xxxiv-xxxvi, xliv, 17, 87, 88, 124,
204, 210, 212, 215, 227, 228, 230
equivalence
class, 172, 183
relation, 172
ergodic theory, xliv, 84
ergodicity, 155
Euclidean algorithm, see algorithm, Euclid's
Euclid's algorithm, see algorithm, Euclid's
expansion, xxxi, 168
Fourier, see Fourier expansion

264

Index

expansion
7V-adic, 11
orthogonal, xxix
wavelet, 23, 25
extension
unitary matrix, 107
unitary principle, see unitary extension
principle
factorization, 34
matrix, 210
of unitary operators, 158, 180
operator, 216
Schmidt's, see theorem, Schmidt's
tensor, 168, 170, 181, 186, 190, 194, 198
Fareytree, 41, 42
father function, see function, father
Fatou
^Primalov theorem, see theorem,
Fatou-Primalov
's lemma, see theorem, "Fatou's
lemma"
set, 25
fermion, 139, 154
algebra, 154
filter, xxxiii, 4, 87, 124, 133, 205, 206, 210,
213,215,227,231,253
dual, 124, 205
dual high-pass, 124, 132
dual low-pass, 124, 132
high-pass, 23, 87, 111, 124, 125, 132, 154,
205, 253
low-pass, xxxiv, 4, 23, 37, 87, 104, 111,
124, 125, 132, 154, 204, 205, 253
orthogonality, see orthogonality, filter
quadrature-mirror, 3, 4, 10, 23, 40, 87,
135, 168, 186, 205, 212, 217, 227, 228
subband, xxix, xxxiv, 23, 26, 87, 123, 124,
126,205,210,211,213,215,228
wavelet, see wavelet filter
fixed-point problem, 10
four-tap, vi, 22, 134, 135, 146, 147, 202, 228
Jean Baptiste Joseph Fourier, xxxi
Fourier
analysis, xv, xxviii, 2, 30, 225, 226

frequency
basis, see basis, Fourier
coefficient, see coefficient, Fourier
correspondence, 67
dual, xxi, XXX, 35, 60, 81, 198, 207,
209, 210
pair, xxi, 36
equivalence, 110
expansion, xxii, 61, 157, 158
frequency, 21, 26, 69, 70, 110, 255
Mock series, 80
series, xxi, xxii, xxvii, xxxi, 10, 32, 59,
67, 69, 80, 145, 163, 206
transform, xxi, 4, 10, 35, 102, 105,
111, 114, 191,208,257
inverse, xxii
variable, 255, see also dual variable
fractal, xxviii, xxix, xxxvi, 7, 15, 21, 22, 34,
35, 37, 67, 72, 74, 77, 80, 97, 152, 182,
194, 198, 210, 211, 227, 231, 254, 255
affine, see affine fractal
analysis, xxxv, xliv, 6, 36, 60, 98, 210
dimension, see dimension, fractal
Hilbert space, see space, Hilbert,
fractal
measure, see measure, fractal
theory, xxix, 25
wavelet, see wavelet, fractal
fractions
iV-adic, 90, 92
2-adic, 89, 138
frame, xliii, 104, 126, 143, 176, 228-230
affine wavelet, 230
bound, 29, 143,207,228
constant, see frame bound
normalized tight, 16, 99, 104, 106, 176,
see also frame, Parseval
operator, 207
Parseval, 13, 16, 99, 104-106, 162, 228
super-, 230
tight, 162,228
wavelet, 100,229
frequency, xxxi, xxxv, 23, 123, 125, 131,
204, 206, 207, 212, 213, 228
band, xx, 87, 124, 177, 181, 210

Index
frequency
domain, 4, 87
localized wavelet, see wavelet,
frequency-localized
response, 4, 17, 87, 131, 204, 206, 207,
210,227
subband, 123
Frobenius, xliii, see also eigenfunction, Perron-Frobenius; eigenspace, PerronFrobenius; matrix, Perron-Frobenius;
operator, Perron-Frobenius-Ruelle;
Perron-Frobenius-Ruelle theory;
theorem, Perron-Frobenius
Fubini's theorem, see theorem, Fubini's
function
basis, see basis function
bounded continuous, 14, 195
bounded measurable, 43, 47, 49, 51, 53,
61, 115,219,253
constant, 15, 46, 61, 91, 136, 256
continuous, 7, 43, 105, 251
eigen, see eigenfunction
father, xxx, 13, 102, XlZ-Yl'b, 128, 134,
135
filter, 4, 87, 104, 114, 123, 132, 192, 196,
207, 227, 253
filter response, 87
frequency response, see frequency
response
generating, 206
harmonic, see harmonic function
indicator, 14, 52
iterated system, see iterated function
system
L^-, 13, 14, 111, 190,210,251
limit, 50, 54
Lipschitz, see Lipschitz function
matrix, 4, 22, 112, 117, 140,141,214, 215
unitary, 227
measurable, xxxiv, 7, 47, 54, 60-62, 65,
70,84,87, 115, 199,252,255
mother, xxx, 13, 102, 123-125, 128, 134,
135
multiplicity, see multiplicity fiinction

265
. Haar

1-periodic, 18, 60-62, 66, 95-97, 104,

132, 174, 175, 217
operator, 114, 116, 117
unitary, xxix
periodic, 4, 69, 87, 204, 208, see also
function, 1-periodic
positive definite, 57, 203, 204
rational, 34
refinable, 107
scaling, 3, see scaling function
space, see space, function
square-integrable, 210, 253
step, xxxi
theory, xxviii, xliv, 46
time-localized, xxx, 177
vector-valued, 131,210
W-, xxxiv, 7, 10, 19, 36, 37, 54, 101, 112
wavelet, see wavelet function
zeta, 34
D. Gabor, xxxi
gap-filling, 14, see also wavelet, gap-filling
generalized multiresolution, xxv, 22, 110,
180,181,252
analysis, 109, 114
GMRA, see generalized multiresolution
grayscale, xviii, 22, 147, 148, 152, 207, 227
A. Grossman, xxxii
group, 1, 5, 28, 36, 70, 109, 110, 205, 210,
211,214,221
abehan, 27, 69
compact, 27
Cantor, 28
circle, 251, 254
cyclic, 60, 174, 255
infinite-dimensional unitary, 155
Lie, 220
non-abelian, xxvi
renormalization, xxix
sub-, 230
torus, 204
transformation, 211
R. Gundy, xxxiii, xliii, 6, 33, 87
A. Haar, xvi, xxvi, xxxi, 131, 223

266

Index

Haar .
Haar
dyadic wavelet, see wavelet, Haar,
dyadic
measure, see measure, Haar
wavelet, see wavelet, Haar
harmonic
analysis, xix, xxviii-xxx, xliii, 2, 5, 19,
22, 25, 26, 33, 60, 80, 87, 182, 229
discrete, 35
of iterated function systems, xxx, 14,
67, 80
function, 9, 18, 21, 22, 43, 48-52, 55,
76, 86, 91, 92, 95, 100, 252
boundary for, 21, 43, 48, 50
bounded, 43, 48-50
closed expression for, 15, 102, 105
construction of, 100
integral formula for, 50
minimal, 11,22, 105
Px-, 100
R-, xxxiv, 43, 50, 79, 91
Hausdorff
dimension, see dimension, Hausdorff
measure, see measure, Hausdorff
O. Heaviside, xvi, 157, 223, 235
W. Heisenberg, 58, 131, 176, 179, 235, 236
hermitian operator, see operator, hermitian
high-pass filter, see filter, high-pass
D. Hilbert, xxxi, 205
Hilbert space, see space, Hilbert
Hutchinson measure, see measure,
Hutchinson

image processmg, vi-vni, xv, xix, xx, xxn,

xxvi, xxviii, xxxii-xxxvii, xliv, 6, 10,
23, 33, 40, 87, 142, 147, 148, 151,
152, 189, 206, 223-227, 230, 233,
235, see also signal processing
infinite-dimensional unitary group, see
group, infinite-dimensional unitary
infinite product, xxviii-xxx, 4, 5, 7, 8,
10-12, 17-19, 21, 27, 33-35, 60, 67,
83-85, 96, 97, 116, 138, 154, see also

Kolmogorov
measure, infinite-product; Tychonofif
infinite-product topology
convergence of, see convergence of
infinite product
matrix, 4, 22
random, 22, 34
tensor, 139, 148, 149, 158
integers, 5, 21, 22, 37, 60, 71, 166, 254, 255
integral translates, 18, 104, 181
intermediate differences, 147, 148
intertwining, 169, 200, 202, 209
interval, unit, see unit interval
invariant, 51, 52, 66
measure, see measure, invariant
R-,53
shift-, 51, 52,92
cr-, 52, 53
subspace, 109, 110, 146, 202, 209, 221
translation-, 129
C.T. lonescu Tulcea, 57
irreducible representation, see representation, irreducible
isometry, xix, 32, 67, 93-95, 174, 176,
182-185, 200, 201, 208, 216, 221,
222, 229, 254
partial, 94, 150, 173
isomorphism, 47, 193, 194, 221
C*-algebraic, 139
isometric, 30, 112, 149,207
order-, 51
*-, 221
unitary, 162, 169, 191, 193, 194
iterated function system, xxx, xxxv, xliv, 34,
35, 47, 57, 67, 70, 84, 99, 152, 182,
252, see also afiine iterated function
system
JPEG 2000, xxxiii
Karhunen-Loeve decomposition theorem,
see theorem, Karhunen-Loeve
A.N. Kolmogorov, xxvi, xxxi, 7, 8, 39, 43,
46, 59, 84, 168, 170, 203, 204, 235

Index
Kolmogorov
Kolmogorov
consistency, xxxi, 7, 45, 46, 48, 49,
141, 203
extension, xxix, 21, 46, 48, 57, 97, 136,
139, 151
's lemma, see theorem, "Kolmogorov's
lemma"
's 0-1 law, xxxiii, 37
Krieger
Cuntz
algebra, see Cuntz-Krieger
algebra
Z^-normalization, 12
Z^-normalization, 12
^^-sequence, 140
lacunary trigonometric series, 84
lattice, 4, 154
Bethe, 98
dual, 28, 36, 69
operation, 169, 181,257
system, see statistical mechanics,
quantum
W. Lawton, xxxiii, xliii, xliv, 21, 33, 57, 87
Lebesgue
measure, see measure, Lebesgue
's dominated convergence theorem, see
theorem, dominated convergence
limit, 20, 21, 50, 75, 76, 92, 154, 212
exchange ofs, 89, 101
function, see function, limit
inductive, 139, 140, 252
martingale, 49
non-tangential, 50
Szego's theorem, see theorem, Szego's
limit
Lipschitz
continuous, 77
function, xxxiv, 57, 191
regularity, see regularity, Lipschitz
localization, 22, 158, see also basis,
localized; function, time-localized;
wavelet, frequency-localized
of Mock Fourier series, 80

267
matrix

property of wavelet bases, 80, 87, 157,

225, 226
low-pass
filter, see filter, low-pass
low-pass
condition, 17,228
property, 125, 130
S. Mallat, xxxii, 33, 168
Mallat subdivision, xxxi
Markov
chain, 50
process, 21
transition measure, see measure,
transition
martingale, xxix, xxx, xxxiii, 10, 21, 35, 36,
50,51,57,91, 168,255
convergence theorem, see theorem,
martingale convergence
limit, see limit, martingale
masking coefficient, see coefficient, masking
Mathematica, 35
graphics produced using, xxxiv, xliii, 2,
123, 125, 129, 152
matrix, 129, 139, 141, 182, 183, 190, 191,
210,215,216,230,252,253
adjoint, 147, 210, 214
algebra, see algebra, matrix
algorithm, see algorithm, matrix
coefficient, see coefficient, matrix
diagonal, 140
diagonal, 141
element, 140, 154
entry, 219
factorization, see factorization, matrix
function, see function, matrix
identity, 183, 252
infinite, xxvii, 142, 231
infinite product, see infinite product,
matrix
integral, 3
multiplication, 25, 128, 142, 143, 202,
210,216,230
operation, 124

268

Index

matrix
operator, 210, 214,215,218
Perron-Frobenius, 34
polyphase, 205, 210, 215, 218
positive, 34
positive definite, 203
positive semidefinite, 4
product, xxvi, 25, 218, 219, see also
infinite product, matrix
propagator, 34
representation, see representation,
matrix
slanted, 23-25, 142, 143, 145, 146, 202,
206,225,230,231
sparse, 23, 206
step in algorithm, xxxv, 206, 207, 210
sub-, 139
theory, 34
Toeplitz, 23, 146
unit, 135
unitary, 109, 129, 174, 182, 191, 201,
205, 210, 211, 214, 216-218, 220,
221, 227, see also extension, unitary
matrix; function, matrix, unitary
valued
function, see function, matrix
measure, see measure, matrix-valued
wavelet filter, see wavelet filter,
matrix-valued
wavelet, 206, 225
maximal abelian subalgebra, 154, 252
measurable, xxiv, xxxiv, 5, 6, 40, 41, 47,
115, 199,252,255
B-, 41, 53
branch, see branch mapping, measurable
space, xxxiv
measure, xxiv, xxvii, 1, 5, 26, 32, 36, 39, 72,
78, 85, 101, 136, 139-141, 154, 166,
168,204,251,253-255,257
absolutely continuous, 50, 136, 154, 257
atomic, 100
B-,53
Bernoulli-product, see measure,
/7-Bemoulli-product
Borel, xxxiv, 9, 46, 80, 141, 149, 154

. measure
Cantor, 12, 14, 74, 198
determinantal, 5, 138, 154, 155
Dirac, see Dirac mass
equivalent, 154
extension, 18, 21, 91, 116, 139, 141,
167, see also Kolmogorov extension
Feynman, 34
fractal, xxiii, 2, 3, 14, 70, 74, 77
full, 5, 22, 71, 79
Haar, xxvi, 28, 61, 66, 70, 93, 175, 212,
251,255
HausdorfiF, xxviii, 2, 14, 17, 72, 74, 97,
176,196,198, 199,251,252
ambient, 72
Hutchinson, 48
infinite-product, 149, 154
invariant, 52, 53, 70
Lebesgue, xxviii, 1, 3, 14, 15, 22, 32, 36,
55,94, 106, 136, 137, 175,251
matrix-valued. 111
non-atomic, 41, 91
operator-valued, 114, 115, 117
orthogonal, 167
p-BemouUi-product, 27
path-space, xxviii, xxx, xxxi, 1, 4-9, 11,
18, 19, 21, 26, 34, 35, 37, 43, 45, 51,
57, 59, 60, 65, 70, 71, 77, 79, 84, 91,
98, 100,111,115, 122,130,254
Perron-Frobenius, see measure,
Perron-Frobenius-Ruelle
Perron-Frobenius-Ruelle, 26, 255
Poisson, 50
positive, xxxiv, 4, 36, 44, 46, 51, 115, 141
probability, xviii, xxxiv, 14, 37, 41, 44,
46, 53, 54, 59, 70, 80, 86, 94, 100,
115,117,122, 149,167, 195,204
Borel, see measure, Borel
Radon, see measure. Radon
5-fractal, xxiii
product, xxvii, 91, 149, 157
projection-valued, 112, 135, 136, 142,
166, 167
Radon, 43, 44, 46, 49, 85, 86, 115-117,
122

Index
measure .
Ruelle, see measure, Perron-FrobeniusRuelle
^-additive, 115, 167
space, 6, 40, 84, 94, 115, 139, 149
a-finite, 31,253
spectral, 112
theory, viii, 5, 6, 20, 22, 70, 195, 204
transition, xxxiv, 21, 90
W-, xxxiv, 9, 26
Y. Meyer, xxxii, 33, 176
middle-third Cantor set, see Cantor set,
middle-third
minimal, 11, 171, 172
eigenfunction, see eigenfunction,
minimal
mirror, 33, 212, 228
quadrature, see filter, quadrature-mirror
monotonicity, 92, 155
J. Morlet, xxxii
mother function, see function, mother
MRA, see multiresolution analysis
multiindex, 165-167, 252
multiplicity. 111
function, 114, 117
multiplicity function, 255
multiresolution, xviii-xx, xxvi, xxxi,
xxxii, 9, 10, 35, 36, 59, 110, 114, 153,
168-171, 179,187,189,198,222,223,
see also generalized multiresolution;
wavelet, multiresolution
analysis, 6, 16, 36, 109, 181, 194, 198,
252, 253
orthogonal, 172
wavelet, see wavelet, multiresolution
multiwavelet, 5, 7, 111, 114, 116, 153
7V-adic, 126
map, 90
rationals, 92
subinterval, 47
-fold branch mapping, see branch mapping,
-fold
natural numbers, xviii, 5, 21, 40, 52, 66, 71,
85, 167, 254

269
operator

Nikodym
Radon
derivative, see derivative,
Radon-Nikodym
non-abelian, see algebra, non-abelian; group,
non-abelian
non-atomic, see measure, non-atomic
non-commutative setting, xv, xxvi, xxix,
5, 7,37,58, 115, 139, 151, 177,
234, see also canonical anticommutation relations; probability,
non-commutative
non-overlapping, 187, 199
partition, see partition, non-overlapping
norm, xviii, 12, 14, 17, 31, 44, 48, 139, 168,
173, 181,204,210,228
normalization, 12, 14, 15, 37, 53, 54, 61, 64,
70, 101, 114, 125, 127, 135
normalized solution, 12, 14
notational convention, 17, 204
ONB, see basis, orthonormal
one-torus, 60, 251,254
operator, 132, 140, 141, 158, 160, 162, 172,
192, 196, 210-212, 215, 216, 219, 254
adjoint, 93, 144, 145, 147, 157, 160, 165,
185, 200, 209, 210, 212-214, 216,
217, 254
algebra, xxviii-xxx, xxxvi, 6, 22, 37,
138, 154, 155, 205, 210, 211, 216, 218
bounded, 218
bounded linear, 218, 251
coefficient, see coefficient, operator
compact, 55, 58
composition ofs, 210, 216
conjugation, see conjugation
factorization, see factorization,
operator
filter, 135
subband-, 213
frame, seefi*ameoperator
function, see function, operator
hermitian, 209
Hilbert space, see space, Hilbert, operators

270

Index

operator
identity, 131, 136, 174, 182, 209, 252, 256
linear, 140
matrix, see matrix, operator
monomial, 165
multi-, 165
multiplication, 94, 95, 110, 142, 147, 198,
209, 210, 213, 217, 221, 253
non-commuting, see non-commutative
setting
Perron-Frobenius-Ruelle, xxxiv, xliv, 4,
8, 9, 19, 21, 26, 33, 43, 48, 49, 57, 61,
86, 87, 95, 97, 107, 155, 200, 254
positive, 117,254
positive semidefinite, 114
process, 116
product, 115,218
projection, see projection
row, 210
scaling, xx, 2, 3, 10, 109, 162, 180, 181,
187, 188, 190, 200, 207-209, 255
unitary, 163, 168, 194, 196
selfadjoint, 138
semidefinite, 114
shift, 213, 256
theory, xix, xxvi, xxix, 10, 37, 58, 109,
138, 142, 154, 170, 180, 181, 186,
210-212, 216, 221, 222, 229, 230
transfer, xxxiii, xHii, xliv, 4,19,26, 33-35,
57, 87, 115, 254, see also operator,
Perron-Frobenius-Ruelle
wavelet, xxxiii, 33, 105, 107
transition, xxxiv, 8, 9, 21, see also
operator, Perron-Frobenius-Ruelle
wavelet, xxxiv, 21
unitary, xix, 158, 161, 169, 179-183, 186,
188-190, 210, 211, 218, 219, 221, see
also factorization of unitary operators;
function, operator, unitary
valued measure, see measure,
operator-valued
zero-kernel, 116
ordering, 51, 209
orthogonal, 198
basis, see basis, orthogonal

. Perron-Frobenius-Ruelle theory
complement, 190, 257
decomposition, see decomposition,
orthogonal
expansion, see expansion, orthogonal
function theory, 155
measure, see measure, orthogonal
multiresolution, see multiresolution,
orthogonal
projection, see projection, orthogonal
sum, 210, 257
vectors, 181, 190
wavelet, see wavelet, orthogonal
orthogonality, 17, 129, 197
filter, 87
relations, 13, 57
wavelet, xxxii-xxxiv, 5
orthonormal basis, see basis, orthonormal
/?-Bemoulli product measure, see measure,
/?-Bemoulli product
p-subinterval, 166
Parseval
basis, see basis, Parseval
frame, see frame, Parseval
identity, 13, 32, 66, 99, 103, 104, 137,
162, 168, 191, 193, 210, 228
system, 13
wavelet, see wavelet, Parseval
partial isometry, see isometry, partial
partition, 166
non-overlapping, 165, 186
path, 7, 87, 123, 124, 127, 129
space, see space, path
perfect reconstruction, 87, 124, 132, 205
periodic function, see function, periodic
permutation of bases, see basis, permutation
of
permutative representation, see representation, permutative
Perron, xliii
Perron-Frobenius-Ruelle theory, 99, see
also eigenftinction, Perron-Frobenius;
eigenspace, Perron-Frobenius;
matrix, Perron-Frobenius; operator.

Index
Perron-Frobenius-Ruelle theory.
Perron-Frobenius-Ruelle; theorem,
Perron-Frobenius
phase modulation, 198
phase transition, 33, 34, 87
physics, xxviii, xxxi, xxxiii, xxxv, 33, 154,
see also quantum physics
mathematical, 6
pixel, 33
Plancherel formula, 164
pointwise convergence, see convergence,
pointwise
Poisson
integral, 50
measure, see measure, Poisson
polar decomposition, 150, 201
positional number system, xx, 33, 40, 69
Powers-Stormer, 138, 154
probability, xxviii-xxxi, xxxiii, xliv, 6, 17,
18, 33-37, 84, 87, 88, 108, 125, 168
combinatorial theory, see combinatorial
probability theory
conditional, 45, 254
distribution, Gaussian, see distribution,
Gaussian
free, 177
measure, see measure, probability
non-commutative, 37, 177
space, see space, probability
transition, see transition probability
process, xxviii, xxx, 48, 49, 86
branching, 5, see also branching
Markov, see Markov process
operator-valued, see operator process
processing, see signal processing; image
processing
product, 76, 114, 115, 207, see also infinite
product
Cartesian, 43, 257
Cauchy, see Cauchy product
infinite, see infinite product
infinite measure, see measure,
infinite-product
inner,9,75, 114, 193,212,257
matrix, see matrix product

271
Radon

measure, see measure, product

operator, see operator product
random, 34, 84
Riesz, see Riesz product
tensor, xxvii, 6, 56, 58, 109, 142, 147-149,
151,152, 158, 165,168,179,180, 186,
189, 257, see also infinite product,
tensor
projection, 10, 55, 94, 112, 117, 173, 216,
228
final, 173
initial, 173
orthogonal, 10, 135, 141, 167, 216
valued measure, see measure,
proj ection-valued
pure, 169
pyramid, 125, 127, 158, 159, 173, 188, 226
algorithm, see algorithm, pyramid
dyadic, 159
singly generated, 159
quadrature, xxix, 131, 227, 228
mirror, see filter, quadrature-mirror
quantization, xxxi, 224226
quantum
field theory, xxix, 154
mechanical state, see state, quantummechanical
particle, 154
physics, xxix, xxxii, 37
statistical mechanics, see statistical
mechanics, quantum
theory, 131
quarter
Cantor set, see Cantor set, quarter
division, 21
quasi-free state, see state, quasi-free
i?-harmonic function, see harmonic function,
R/?-invariant, see invariant, RRadon
measure, see measure, Radon
Nikodym derivative, see derivative,
Radon-Nikodym

272

Index

random
random, 18, 59, 83
process, 36, 55, 204
product, see product, random
variable, xviii, 55-57
A', xxiv
Gaussian, 56, 57
walk, vi, xviii, xxiv, xxviii-xxx, xxxiv,
xliii, 4-7, 16,21,34,39,40,42,
48, 57, 77, 83, 84, 87, 100, see also
process
model, viii, 2, 8, 9, 12, 21, 26, 40, 41,
83, 98, 136
on branches, xxviii, xxx, 70
on fractal, 21
range subspace, 117, 162
reconstruction, see perfect reconstruction
recursive, 129, 131
algorithm, see algorithm, recursive
system, 197
redundancy, 228, 229
refinement, 4
equation. 111
regularity, xxxiv, 5, 80, 87, 107
Dini, 4
Lipschitz, 4
renormalization group, see group,
renormalization
renormalize, xxix
representation, xxix, 124, 140, 165, 172,
190, 197, 214, 254-256
base-point, 219
boundary, see boundary representation
irreducible, 183
matrix, 140, 218
A^-adic, 90, 101, 256
of algebraic structure, xxix, 37
of CAR algebra, 138
of Cuntz algebra, 5, 22, 131, 136, 139,
152, 154, 155, 158, 161-164, 167,
168, 170, 174, 179, 180, 182, 183,
187, 192, 196, 205, 211, 214, 218,
219, 227
of Z by translation. 111

scaling
permutative, 6, 180, 184, 185, 189, 190,
194
spectral, 112
subband, 219
theory, 154, 155, 180
unitary equivalence of, 184
wavelet, 102, 182
reproduction formula, 144
resolution, xix, xxii, xxvi, xxix, xxx, 9, 10,
22, 33, 111, 147, 148, 181, 225, 230,
see also multiresolution
subspace, xxvi, 23, 109, 110, 123, 126,
152, 180, 207, 255
multiply generated, 114
visual, xvi, xviii, xx, xxii, 40, 225
Riesz
product, 84, 97
's theorem, see theorem, Riesz's
row-contraction, 217, 222
D. Ruelle, xliii, 33, 34, 87, 155
Ruelle, see also eigenfunction, PerronFrobenius; eigenspace, PerronFrobenius; matrix, Perron-Frobenius;
operator, Perron-Frobenius-Ruelle;
Perron-Frobenius-Ruelle theory;
theorem, Perron-Frobenius
measure, see measure, Ruelle
operator, see operator, Perron-Frobenius-Ruelle
's theorem, see theorem, Ruelle's
sampling, 17, 18, 21, 36, 37, 213, 215
down-, 87, 124, 128, 132, 133, 205, 206,
212, 213, 257
Shannon, xxxi, 18
theory, 5, 18
up-, 87, 124, 132, 205, 206, 212, 213, 257
scale-A/^
wavelet, see wavelet, scale-A^
scale number, 19, 40, 70, 147, 190, see also
scaling number
scaling, xxii, xxiii, 3, 22, 23, 25, 69, 80, 99,
100, 110, 111, 123, 142, 181, 186,
188, 195, 198, 208

Index
scaling
dimension, see dimension, scaling
dyadic, 2, 3, 110, 162, 181, 192, 200, 208,
209, 255
equation
wavelet, 25
function, xx, xxx, 3, 4, 12-14, 16, 23,
25,64, 111,134,207,256
Daubechies, xxxv, 12, 13, 125
Haar, xxxv, 13, 14, 103
stretched Haar, 12, 13, 103
identity, xx, 3, 10, 14, 15, 17, 25, 30,
91, 102, 103, 109, 111, 123, 130, 131,
134, 199, 209
Cantor, 14
in the large, 14
number, 143, 146, 147, 188, 208
operator, see operator, scaling
relation, 3, 75
similarity, 9, 33, 180
transformation, 9, 66, 79, 80
fixed, 3, 9, 10, 22, 66, 188
Schmidt
Gram-, see algorithm, Gram-Schmidt
's decomposition theorem, see theorem,
Schmidt's
segment, 165, 166, 186, 257
self-similarity, xxix, xxxv, 9, 48
separation of variables, 158, 179
C.E. Shannon, xxxi, 18
Shannon sampling, see sampling. Shannon
shift-invariant, see invariant, shifter-additive, see measure, cr-additive
^-algebra, xviii, xxiv, 27, 28, 37, 40, 47, 94,
149,204,251-253,255
Borel,xxiv, 43, 251
s, countable family of, 37
sub-, xxiv, 37
tail-, 37
cr-invariant, see invariant, asignal analysis, xvi, xxvi, 40, 223
signal processing, vi-viii, xv, xvi, xix-xxii,
xxvi, xxviii, xxxi-xxxvii, xliv, 4, 6, 10,
18, 23, 25, 26, 33, 37, 39, 87,123-125,
130, 131, 135, 155, 181, 187, 189,

273
spectral

205, 206, 210-212, 215, 216, 218,

219, 223, 224, 227-229, 231, see also
image processing; speech signal
singly generated, xxx, 159, 160, 173
six-tap, 22, 146
slanted, see matrix, slanted
space
compact Hausdorfif, 101
detail, 33, 123, 255
Euclidean, 22
ambient, 81
function, xxii, xxvi, 9, 22, 59, 109, 130,
142, 146, 147, 157, 158, 179, 190,
207,210,228,251
ambient, 3
Hilbert, xviii-xxi, xxiii, xxvii, xxxvi, 3, 5,
6, 10, 14, 15, 22, 28-31, 33, 36, 37,
58, 71, 79, 87, 93, 97, 109, 110, 114,
136, 139, 140, 142-144, 147-152,
157, 158, 161, 162, 165, 169, 170,
172, 174, 176, 180-184, 186, 188,
189, 196-199, 204, 207, 210, 212,
213, 217, 218, 221, 222, 228-230,
251,252,254,257
ambient, 25, 26, 142, 152, 188, 207, 229
complex, 58, 114, 143, 144, 173,
182-184, 203, 252
dilated, 229, 230
fractal, 14, 17, 72, 97, 110, 158, 196,
198
geometry, xxvi, 109, 142, 148, 179,
198, 207, 221, 230
operators in, xxxii, 6, 37, 131, 138, 142,
143, 161, 177, 180, 183, 209-211,
216,218,221,251
symbolic, 110
path, xxix, 5, 6, 34, see also measure,
path-space
probability, xviii, xxiv, 5, 7, 9, 21, 22, 37,
47, 50, 55, 56, 60, 70, 114, 135, 203,
233, 255, 256
sparse matrix, see matrix, sparse
spectral
analysis, 158, 181

274

Index

spectral
joint radius, xxxiii
measure, see measure, spectral
pair, 36, 71, 74
representation, see representation,
spectral
theorem, see theorem, spectral
theory, xliv, 37, 57
transform, 110
spectrum, 34, 138, 254
peripheral, 57
speech signal, xxxi, 124, 205, 227
state, xviii, xxiii, 6, 26, 37, 139-141, 151,
154, 157,256
equilibrium, 154
multi-, 139
on graph configuration, 35
quantum-mechanical, xviii, xxxii, 154
quasi-equivalent, 154
quasi-free, 138, 139, 154
statistical mechanics, 33, 34, 42, 87, 154
quantum, 34, 154
statistics, xxxi, xxxiii, 154
Stone-WeierstraB theorem, see theorem,
Stone-WeierstraB
stretched Haar wavelet, see wavelet, Haar,
stretched
R. Strichartz, 26, 76, 80, 87
subband, 23, 124, 205, 215, 227
coding, xxxi, 123
filter, see filter, subband
frequency, see frequency subband
representation, see representation,
subband
subdivision, xxxii, xxxv, 4, 123, 124, 195
algorithm, see algorithm, subdivision
dyadic, xxxv
Mallat, see Mallat subdivision
subinterval, 167
dyadic fractional, see dyadic fractional
subinterval
A/^-adic, see A^-adic subinterval
p-, see p-subinterval
subspace
invariant, see invariant subspace

transfer operator
resolution, see resolution subspace
substitution, iterated, 34
symbolic dynamics, see dynamics, symbolic
synthesis, 124, 132, 205, 206, 227, see also
analysis (engineering)
tap number, 147, see also two-tap; four-tap;
six-tap
tensor, 140, 148, 150, 162
factorization, see factorization, tensor
product, see product, tensor
theorem
dominated convergence, 65, 79
Fatou-Primalov, 48, 50
"Fatou's lemma", 101
Fubini's, 191
Karhunen-Loeve, 55, 58, 148, 204
Kohnogorov's, 204
"Kolmogorov's lemma", 48, 149
martingale convergence, 21, 48, 51, 92
Perron-Frobenius, 34, see also PerronFrobenius-Ruelle theory
Riesz's, 37, 45, 46, 141
Ruelle's, xxxiv, see also PerronFrobenius-Ruelle theory
Schmidt's, 58, 148, 150
"Schwarz's inequality", 31
spectral, 39, 55, 58, 110, 112, 150, 201
Stone-WeierstraB, 27, 39, 4 3 ^ 6 , 79
Szego's limit, 154
uniqueness, 86
"Zom's lemma", 172
tiling, 87, 124, 165, 166, 182, 185-189
dyadic, 187
self-afifine, 100
time-localized function, see function,
time-localized
torus, 25, 204, 251, see also one-torus
trace
formula, 34
normalized, 184
traditional wavelet setup, xxx, 4, 6, 10, 25,
26,71, 110,112
transfer operator, see operator, transfer

Index
transformation.
transformation
group, see group, transformation
rule, 165, 166, see also basis
transformation
transition
measure, see measure, transition
operator, see operator, transition
probability, xxxiv, 5, 9, 12, 17, 19, 21,
34, 37, 39-^1, 48, 50, 59, 62, 63, 70,
254
translation, 110, 111, 198
integer, 126, see also integral translates
invariant, see invariant, translationtree, xxxiv, 5, 42, 87, 124, see also combinatorial tree; decision tree; Farey
tree
two-cycle, 160
2-fold branch mapping, see branch mapping,
2-fold
two-tap, 146
Tychonofif infinite-product topology, 7, 43
uniqueness theorem, see theorem,
uniqueness
unit interval, 62, 90, 137, 139, 166
unitarity, 132, 174, 175, 183, 191, 220
unitary, xix
equivalence, 151, 163, 169, 184
extension principle, 107, 222
infinite-dimensional group, see group,
infinite-dimensional unitary
isomorphism, see isomorphism,
unitary
matrix, see matrix, unitary
operator, see operator, unitary
scaling operator, see operator, scaling,
unitary
up-sampling, see sampling, upvariable
dual, see dual variable
Fourier-dual, see Fourier dual
random, see random variable

275
wavelet

walk, 6, 41, 128, see also transition; random

walk
wavelet, vi, vii, xv-xvii, xix, xx, xxv-xxx,
xxxii-xxxvi, xliii, xliv, 1-7, 9, 14, 16,
17, 22, 23, 25, 33-37, 39, 40, 57, 58,
64, 71, 79-81, 83, 87, 91, 99, 100,
104, 105, 107, 109, 111, 122, 130,
131, 142, 151, 152, 155, 158, 179,
182, 189, 190, 211, 222-230, 235,
256, see also traditional wavelet setup
algorithm, see algorithm, wavelet
analysis, xv, xx, xxviii, xxx, xxxii,
xxxiii, xliii, 4, 6, 10, 34, 57, 60, 84, 98,
99, 109, 181, 188, 208, 210, 225, 226
band-limited, 4
basis, see basis, wavelet
coefficient, see coefficient, wavelet
compactly supported, xxx, 14
construction, xix, xxix-xxxi, xxxiii, 1,
4, 5, 7-9, 13, 14, 36, 57, 83, 97, 119,
121-124, 130, 131, 136, 157, 179,
180, 192, 198
Daubechies, 5, 14, 121, 122, 124, 127,
128, 134-136, 228
decomposition, see decomposition,
wavelet
dual, xxxii
dyadic, 15, 39, 99, 123-125, 146
expansion, 230
filter, xxxii, xxxiii, 5, 23, 33, 37, 83,
87, 112,122,130, 136,227,230
matrix-valued, 22
fractal, 15, 71, 180
frame, seefi*amewavelet
frequency-localized, xxx, 7
function, vi, xvi, xx, xxvii, xxx, 1,16,
102, 103, 123, 134, 256
Daubechies, 110
Haar, 103
gap-filling, 22, 72
Haar, xvi, xxxi, 5, 12-14, 99, 102, 103,
106, 119, 124, 127, 128, 131, 148,
192, 229
dyadic, 136, 137

276

Index

wavelet
stretched, 12, 13, 15, 16, 99, 100, 104,
106, 252
matrix, see matrix, wavelet
multi-, see multiwavelet
multiresolution, xxx, 10, 16, 97, 110, 142,
181, 190
7V-adic, 64, 123
orthogonal, 124
orthogonality, see orthogonality,
wavelet
packet, vi, xxxiv, xxxv, 4, 6, 22, 35,
110, 117, 122-126, 129-131, 136,
142, 153, 159, 176, 180, 187-189, 256
algorithm, see algorithm, wavelet packet
dyadic, 127, 158, 159
Parseval, 13, 16, 103, 105
representation, see representation,
wavelet
scale-A^, 64, 65, 190, 252
theory, xxix, 9, 33, 87, 110, 168, 177,
222-224
time-frequency, xxxi

. Zom's lemma
time-scale, xxxi
traditional setup, see traditional
wavelet setup
transfer operator, see operator, transfer,
wavelet
transform, 142, 186
discrete, 142, 180, 202
transition operator, see operator,
transition, wavelet
wavepacket coefficient, see coefficient,
wavepacket
WeierstraB
Stone
theorem, see theorem,
Stone-WeierstraB
weight function, see function, WM.V. Wickerhauser, 117, 176, 177
N. Wiener, 34
Wold decomposition, see decomposition,
Wold
zeta function, see function, zeta
Zom's lemma, see theorem, "Zom's lemma"

Graduate Texts in Mathematics

(con tinnedfrom p. ii)
64
65
66
67
68
69
70
71
72

EDWARDS. Fourier Series. Vol. I. 2nd ed.

WELLS. Differential Analysis on
Complex Manifolds. 2nd ed.
WATERHOUSE. Introduction to Affine
Group Schemes.
SERRE. Local Fields.
WEIDMANN. Linear Operators in
Hilbert Spaces.
LANG. Cyclotomic Fields IL
MASSEY. Singular Homology Theory.
FARKAS/KRA. Riemann Surfaces. 2nd
ed.
STILLWELL. Classical Topology and
Combinatorial Group Theory. 2nd ed.

95

SHIRYAEV. Probability 2nd ed.

96

CONWAY. A Course in Functional

Analysis. 2nd ed.
KoBLiTZ. Introduction to Elliptic
Curves and Modular Forms. 2nd ed.

97
98

BROCKER/TOM DIECK. Representations

99

of Compact Lie Groups.

GROVE/BENSON. Finite Reflection
Groups. 2nd ed.

100

BERG/CHRISTENSEN/RESSEL. Harmonic

Analysis on Semigroups: Theory of

Positive Definite and Related
Functions.
101 EDWARDS . Galois Theory

73

HUNGERFORD. Algebra.

102

74

DAVENPORT. Multiplicative Number

Theory. 3rd ed.
HOCHSCHILD. Basic Theory of
Algebraic Groups and Lie Algebras.
IiTAKA. Algebraic Geometry.
HECKE. Lectures on the Theory of
Algebraic Numbers.

Algebras and Their Representations.

103 LANG. Complex Analysis. 3rd ed.

75
76
77
78

79
80
81
82

WASHINGTON. Introduction to

84

IRELAND/ROSEN. A Classical

Cyclotomic Fields. 2nd ed.

86
87
88
89
90

Introduction to Modern Number

Theory. 2nd ed.
EDWARDS. Fourier Series. Vol. II. 2nd
ed.
VAN LINT. Introduction to Coding
Theory. 2nd ed.
BROWN. Cohomology of Groups.
PIERCE. Associative Algebras.
LANG. Introduction to Algebraic and
Abelian Functions. 2nd ed.
BR0NDSTED. An Introduction to
Convex Polytopes.

91

BEARDON. On the Geometry of

92

Discrete Groups.
DiESTEL. Sequences and Series in
Banach Spaces.

93

DUBROVIN /FOMENKO/NOVIKOV.

Modem GeometryMethods and

Applications. Part 1.2nd ed.
94

WARNER. Foundations of

Differentiable Manifolds and Lie

Groups.

DUBROVIN/FOMENKO/NOVIKOV.

Modern GeometryMethods and

Applications. Part II.
105

LANG.

106

SILVERMAN. The Arithmetic of Elliptic

Curves.
OLVER. Applications of Lie Groups to
Differential Equations. 2nd ed.
RANGE. Holomorphic Functions and
Integral Representations in Several
Complex Variables.
LEHTO. Univalent Functions and
Teichmiiller Spaces.
LANG. Algebraic Number Theory.
HUSEMOLLER. Elliptic Curves. 2nd ed.
LANG. Elliptic Functions.
KARATZAS/SHREVE. Brownian Motion
and Stochastic Calculus. 2nd ed.
KOBLITZ. A Course in Number Theory
and Cryptography. 2nd ed.

BURRIS/SANKAPPANAVAR. A Course in

Universal Algebra.
WALTERS. An Introduction to Ergodic
Theory.
ROBINSON. A Course in the Theory of
Groups. 2nd ed.
FoRSTER. Lectures on Riemann
Surfaces.
BOTT/TU. Differential Forms in
Algebraic Topology.

83

85

104

VARADARAJAN. Lie Groups, Lie

107
108
109
110
111
112
113
114
115

SL^(R).

BERGER/GOSTIAUX. Differential

Geometry: Manifolds, Curves, and

Surfaces.
116

KELLEY/SRINIVASAN. Measure and

Integral. Vol. I.
117 J. -P. SERRE . Algebraic Groups and
Class Fields.
118

PEDERSEN. Analysis Now.

119

ROTMAN. An Introduction to Algebraic

Topology.
120 ZIEMER. Weakly Differentiable
Functions: Sobolev Spaces and
Functions of Bounded Variation.
121 LANG. Cyclotomic Fields I and II.
Combined 2nd ed.
122 REMMERT . Theory of Complex
Functions.
Readings in Mathematics
123

EBBINGHAUS/HERMES et al.

Numbers.

124

125
126
127
128
129

130
131
132
133
134
135
136

Readings in Mathematics

154

DUBROVIN/FOMENKO/NOVIKOV.

Analysis.
155 KASSEL. Quantum Groups.
156 KECHRIS. Classical Descriptive Set
Theory.

Modem GeometryMethods and

Applications Part III.
BERENSTEIN/GAY. Complex Variables:
An Introduction.
BoREL. Linear Algebraic Groups.
2nd ed.
MASSEY. A Basic Course in Algebraic
Topology.
RAUCH. Partial DifFerential Equations.
FULTON/HARRIS . Representation

Theory: A First Course.

Readings in Mathematics
DODSON/POSTON. Tensor Geometry.
LAM . A First Course in
Noncommutative Rings. 2nd ed.
BEARDON. Iteration of Rational
Functions.
HARRIS. Algebraic Geometry: A First
Course.
ROMAN. Coding and Information
Theory.
ROMAN. Advanced Linear Algebra.
2nd ed.
ADKINSAVEINTRAUB. Algebra: An

Approach via Module Theory.

137

AXLER/BOURDON/RAMEY. Harmonic

Function Theory. 2nd ed.

138 COHEN. A Course in Computational
Algebraic Number Theory.
139 BREDON. Topology and Geometry.
140 AUBIN. Optima and Equilibria. An
Introduction to Nonlinear Analysis.
141

BECKER/ WEISPFENNING/KREDEL.

Grobner Bases. A Computational

Approach to Commutative Algebra.
142 LANG. Real and Functional Analysis.
3rded.
143 DooB. Measure Theory.
144

DENNIS/FARB. Noncommutative

Algebra.
145 ViCK. Homology Theory. An
Introduction to Algebraic Topology.
2nd ed.
146

147
148
149
150
151
152
153

BRIDGES. Computability: A

Mathematical Sketchbook.
ROSENBERG. Algebraic AT-Theory and
Its Applications.
RoTMAN. An Introduction to the
Theory of Groups. 4th ed.
RATCLIFFE. Foundations of Hyperbolic
Manifolds.
EISENBUD . Commutative Algebra with
a View Toward Algebraic Geometry.
SILVERMAN. Advanced Topics in the
Arithmetic of Elliptic Curves.
ZiEGLER. Lectures on Polytopes.
FULTON. Algebraic Topology: A First
Course.

157

BROWN/PEARCY. An Introduction to

MALLIAVIN. Integration and

Probability.
158 ROMAN. Field Theory
159 CONWAY. Functions of One Complex
Variable II.
160 LANG. Differential and Riemannian
Manifolds.
161

BORWEIN/ERDELYI. Polynomials and

Polynomial Inequalities.
162

ALPERIN/BELL. Groups and

163

DIXON/MORTIMER. Permutation

Representations.
Groups.
164 NATHANSON. Additive Number
Theory: The Classical Bases.
165 NATHANSON. Additive Number
Theory: Inverse Problems and the
Geometry of Sumsets.
166 SHARPE. Differential Geometry:
Cartan's Generalization of Klein's
Erlangen Program.
167 MORANDI. Field and Galois
Theory.
168 EwALD. Combinatorial Convexity and
Algebraic Geometry.
169 BHATIA. Matrix Analysis.
170 BREDON. Sheaf Theory 2nd ed.
171 PETERSEN. Riemannian Geometry.
172 REMMERT. Classical Topics in Complex
Function Theory.
173 DiESTEL. Graph Theory 2nd ed.
174 BRIDGES. Foundations of Real and
Abstract Analysis.
175 LiCKORiSH. An Introduction to Knot
Theory.
176 LEE. Riemannian Manifolds.
177 NEWMAN. Analytic Number Theory.
178

179
180
181
182
183

CLARKE/ LEDYAEV/STERNAVOLENSKI.

Nonsmooth Analysis and Control

Theory.
DOUGLAS . Banach Algebra Techniques
in Operator Theory. 2nd ed.
SRIVASTAVA. A Course on Borel Sets.
KRESS . Numerical Analysis.
WALTER. Ordinary Differential
Equations.
MEGGINSON. An Introduction to

Banach Space Theory.

184 BoLLOBAS. Modem Graph Theory.
185 COX/LITTLE/O'SHEA. Using Algebraic
Geometry. 2nd ed.
186

RAMAKRISHNANA^ALENZA. Fourier

Analysis on Number Fields.

187

HARRIS/MORRISON. ModuU of

Curves.

188

GoLDBLATi. Lectures on the

Hyperreals: An Introduction to
Nonstandard Analysis.
189 LAM. Lectures on Modules and Rings.

190

ESMONDE/MURTY. Problems in

191

Algebraic Number Theory. 2nd ed.

LANG. Fundamentals of Differential
Geometry.

192

HIRSCH/LACOMBE. Elements of

193

Functional Analysis.
COHEN. Advanced Topics in
Computational Number Theory.

194

ENGEL/NAGEL. One-Parameter

Semigroups for Linear Evolution

Equations.
195 NATHANSON. Elementary Methods in
Number Theory.
196 OSBORNE. Basic Homological Algebra.
197

EISENBUD/HARRIS. The Geometry of

198

Schemes.
ROBERT. A Course in /?-adic Analysis.

199

HEDENMALM/KORENBLUM/ZHU.

Theory of Bergman Spaces.

200

BAO/CHERN/SHEN. An Introduction to

201

HINDRY/SILVERMAN. Diophantine

Riemann-Finsler Geometry.

204

Geometry: An Introduction.
LEE. Introduction to Topological
Manifolds.
SAGAN. The Symmetric Group:
Representations, Combinatorial
Algorithms, and Symmetric Functions.
EscoFiER. Galois Theory.

205

FELIX/HALPERIN/THOMAS. Rational

202
203

Homotopy Theory, 2nd ed.

206 MuRTY. Problems in Analytic Number
Theory.
Readings in Mathematics
207 GODSIL/ROYLE. Algebraic Graph
Theory.
208 CHENEY. Analysis for Applied
Mathematics.
209 ARVESON. A Short Course on Spectral
Theory.
210 ROSEN. Number Theory in Function
Fields.

211
212
213

214
215
216
217
218
219

LANG. Algebra. Revised 3rd ed.

MATOUSEK. Lectures on Discrete
Geometry.
FRITZSCHE/GRAUERT. From

Holomorphic Functions to Complex

Manifolds.
JosT. Partial Differential Equations.
GoLDSCHMiDT. Algebraic Functions
and Projective Curves.
D. SERRE. Matrices: Theory and
Applications.
MARKER. Model Theory: An
Introduction.
LEE. Introduction to Smooth
Manifolds.
MACLACHLAN/REID. The Arithmetic of

Hyperbolic 3-Manifolds.
220 NESTRUEV. Smooth Manifolds and
Observables.
221 GRUNBAUM. Convex Polytopes.
2nd ed.
222 HALL. Lie Groups, Lie Algebras, and
Representations: An Elementary
Introduction.
223 VRETBLAD. Fourier Analysis and Its
Applications.
224 WALSCHAP. Metric Structures in
Differential Geometry.
225

BUMP: Lie Groups.

226

ZHU. Spaces of Holomorphic

Functions in the Unit Ball.

227

228
229
230
231

MILLER/STURMFELS. Combinatorial

Commutative Algebra.
DIAMOND/SHURMAN. A First Course in
Modular Forms.
EISENBUD. The Geometry of Syzygies.
STROOCK. An Introduction to Markov
Processes.
BJORNER/BRENTL Combinatorics of

Coxeter Groups.
232

EVERESTAVARD. An Introduction to

Number Theory.
233 ALBIAC/KALTON. Topics in Banach
Space Theory.
234 JoRGENSEN. Analysis and Probability:
Wavelets, Signals, Fractals.

Printed in the United States of America.