ZURICH LECTURES IN ADVANCED MATHEMATICS ZURICH LECTURES IN ADVANCED MATHEMATICS

Alessio Figalli
Alessio Figalli
Alessio Figalli

The Monge-Ampère Equation and its Applications

The Monge-Ampère Equation
and its Applications
The Monge–Ampère equation is one of the most important partial differential equations,
appearing in many problems in analysis and geometry. The Monge-Ampère
Equation and
This monograph is a comprehensive introduction to the existence and regularity theory of
the Monge–Ampère equation and some selected applications; the main goal is to provide
the reader with a wealth of results and techniques he or she can draw from to understand

Its Applications
current research related to this beautiful equation.

The presentation is essentially self-contained, with an appendix wherein one can find
precise statements of all the results used from different areas (linear algebra, convex
geometry, measure theory, nonlinear analysis, and PDEs).

This book is intended for graduate students and researchers interested in nonlinear PDEs:
explanatory figures, detailed proofs, and heuristic arguments make this book suitable for
self-study and also as a reference.

ISBN 978-3-03719-170-5

www.ems-ph.org

Figalli Cover (ZLAM) | Fonts: RotisSemiSans / DIN | Farben: 4c Pantone 116, Pantone 287, Cyan | RB 10.4 mm
Zurich Lectures in Advanced Mathematics

Edited by

Erwin Bolthausen (Managing Editor), Freddy Delbaen, Thomas Kappeler (Managing Editor), Christoph Schwab,
Michael Struwe, Gisbert Wüstholz

Mathematics in Zurich has a long and distinguished tradition, in which the writing of lecture notes volumes
and research monographs plays a prominent part. The Zurich Lectures in Advanced Mathematics series
aims to make some of these publications better known to a wider audience. The series has three main con-
stituents: lecture notes on advanced topics given by internationally renowned experts, in particular lecture
notes of “Nachdiplomvorlesungen”, organzied jointly by the Department of Mathematics and the Institute for
Research in Mathematics (FIM) at ETH, graduate text books designed for the joint graduate program in
Mathematics of the ETH and the University of Zürich, as well as contributions from researchers in residence.
Moderately priced, concise and lively in style, the volumes of this series will appeal to researchers and
students alike, who seek an informed introduction to important areas of current research.

Previously published in this series:

Yakov B. Pesin, Lectures on partial hyperbolicity and stable ergodicity

Sun-Yung Alice Chang, Non-linear Elliptic Equations in Conformal Geometry
Sergei B. Kuksin, Randomly forced nonlinear PDEs and statistical hydrodynamics in 2 space dimensions
Pavel Etingof, Calogero–Moser systems and representation theory
Guus Balkema and Paul Embrechts, High Risk Scenarios and Extremes – A geometric approach
Demetrios Christodoulou, Mathematical Problems of General Relativity I
Camillo De Lellis, Rectifiable Sets, Densities and Tangent Measures
Paul Seidel, Fukaya Categories and Picard–Lefschetz Theory
Alexander H.W. Schmitt, Geometric Invariant Theory and Decorated Principal Bundles
Michael Farber, Invitation to Topological Robotics
Alexander Barvinok, Integer Points in Polyhedra
Christian Lubich, From Quantum to Classical Molecular Dynamics: Reduced Models and Numerical Analysis
Shmuel Onn, Nonlinear Discrete Optimization – An Algorithmic Theory
Kenji Nakanishi and Wilhelm Schlag, Invariant Manifolds and Dispersive Hamiltonian Evolution Equations
Erwan Faou, Geometric Numerical Integration and Schrödinger Equations
Alain-Sol Sznitman, Topics in Occupation Times and Gaussian Free Fields
François Labourie, Lectures on Representations of Surface Groups
Isabelle Gallagher, Laure Saint-Raymond and Benjamin Texier, From Newton to Boltzmann: Hard Spheres and
Short-range Potentials
Robert J. Marsh, Lecture Notes on Cluster Algebras
Emmanuel Hebey, Compactness and Stability for Nonlinear Elliptic Equations
Sylvia Serfaty, Coulomb Gases and Ginzburg–Landau Vortices

Published with the support of the Huber-Kudlich-Stiftung, Zürich

Alessio Figalli

The Monge-Ampère
Equation and
Its Applications
Author:

Alessio Figalli
Department Mathematik
ETH Zürich
Ramistrasse 101
8092 Zürich
Switzerland
E-mail: [email protected]

2010 Mathematics Subject Classification: Primary: 35J96; secondary: 35B65, 35J60, 35J66, 35B45, 35B50,
35D05, 35D10, 35J65, 53A15, 53C45

Key words: Monge–Ampère equation, weak and strong solutions, existence, uniqueness, regularity

ISBN 978-3-03719-170-5

The Swiss National Library lists this publication in The Swiss Book, the Swiss national bibliography,
and the detailed bibliographic data are available on the Internet at http://www.helveticat.ch.

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permission
of the copyright owner must be obtained.

© 2017 European Mathematical Society

Contact address:

European Mathematical Society Publishing House

Seminar for Applied Mathematics
ETH-Zentrum SEW A27
CH-8092 Zürich
Switzerland

Phone: +41 (0)44 632 34 36

Email: [email protected]
Homepage: www.ems-ph.org

Typeset using the author’s TEX files: Alison Durham, Manchester, UK

Printing and binding: Beltz Bad Langensalza GmbH, Bad Langensalza, Germany
∞ Printed on acid free paper
987654321
To Mikaela
Preface

This book originates from a series of lectures given by the author at ETH Zürich
during the fall of 2014, in the framework of a Nachdiplomvorlesung, on the Monge–
Ampère equation and its applications.
The Monge–Ampère equation is a fully nonlinear, degenerate elliptic equation
arising in several problems in the areas of analysis and geometry, such as the pre-
scribed Gaussian curvature equation, affine geometry, and optimal transportation. In
its classical form, it consists of prescribing the determinant of the Hessian of a convex
function u inside some domain Ω, that is,

det(D2 u) = f in Ω.

This is in contrast with the “model” elliptic equation ∆v = g, which prescribes the
trace of the Hessian of a function v. There are several boundary conditions that one
may consider for u, and in this book we shall focus on Dirichlet boundary conditions
prescribing the value of u on ∂Ω.
Our goal is to give a comprehensive introduction to the existence and regularity
theory of the Monge–Ampère equation, and to show some selected applications.
Although some of the results contained here have already been discussed in the
classical book by Gutiérrez [61], recent developments in the theory have motivated
us to write a new book on the subject.
In the same spirit as the lectures given at ETH Zürich, the structure of this book
follows a “historical” path. More precisely, after a brief introduction in Chapter 1 to
the Monge–Ampère equation and its history, Chapter 2 is dedicated to the theory of
weak solutions introduced by Alexandrov in the 1940s. This notion of solutions is
powerful enough to allow one to obtain existence and uniqueness of weak solutions
with any nonnegative Borel measure as a right-hand side. Then in Chapter 3 we
address the issue of existence of global smooth solutions. This theory, developed
between the 1960s and 1980s, combines the continuity method and some interior a
priori estimates due to Pogorelov to show existence of smooth solutions when the
domain and the boundary data are smooth. The largest part of this book is devoted
to the interior regularity of weak solutions. Specifically, in Chapter 4 we study, in
detail, the geometry of solutions, mostly investigated by Caffarelli in the 1990s, and
we prove interior C 1,α , W 2, p , and C 2,α estimates. Finally, in Chapter 5 we describe
some extensions and generalizations of the results described in the previous chapters.
Since this theory needs some general knowledge from different areas (linear
algebra, convex geometry, measure theory, nonlinear analysis, and PDEs), we have
decided to write an appendix where the reader can find precise statements of all the
viii Preface

results that we have used. Whenever possible we have included the proofs of such
results, and otherwise we have given a precise reference.
By no means is this book intended to cover all the topics and recent developments
in the theory of the Monge–Ampère equation and its variants. Rather, our hope
and intent is that, after reading this book, the reader will be able to understand and
appreciate contemporary literature on the topic.
The reader may notice that every chapter is divided into several sections and
subsections. We believe this will facilitate comprehension and help the reader when
moving between different results. Also, for the same reason, long proofs are always
split into several steps. Finally, because of the geometric arguments presented, we
have included several supporting figures. We hope that the reader will benefit from
this presentation style.
This book would not exist without the support and help of many friends and
colleagues. First, I have been lucky enough to have Luis Caffarelli as a colleague and
department neighbor for many years. His beautiful results on Monge–Ampère have
been a constant source of inspiration. Second, I have been fortunate to have Guido
De Philippis as a long-time collaborator; investigating the Monge–Ampère equation
with him has been tremendously inspiring and enjoyable.
I owe a debt of gratitude to Tristan Rivière, Michael Struwe, and the entire staff
at ETH. I cannot begin to express my appreciation for their warm hospitality during
my semester at ETH Zürich. Also, I wish to thank the faculty and students who
attended my course; their interest, curiosity, and participation was a great source of
support and motivation. I am particularly grateful to Camillo De Lellis and Thomas
Kappeler for their encouragement and to Maria Colombo and Mikaela Iacobelli for
their carefully written notes.
This book has benefited from valuable comments and suggestions by Connor
Mooney and Neil Trudinger. Furthermore, I have to thank Yash Jhaveri for a detailed
and careful reading of the whole manuscript.
Finally, I am very grateful for my family’s continued encouragement and my wife
Mikaela’s constant support and love. Thank you for always being by my side!
Contents

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 On the degeneracy of the Monge–Ampère equation . . . . . . . . . 1
1.2 Some history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Alexandrov solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1 The Monge–Ampère measure . . . . . . . . . . . . . . . . . . . . . 7
2.2 Alexandrov solutions: Definition and basic properties . . . . . . . . 11
2.3 The Dirichlet problem: Uniqueness . . . . . . . . . . . . . . . . . 17
2.4 The Dirichlet problem: Existence . . . . . . . . . . . . . . . . . . 20
2.5 C 1 regularity in 2-D . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.6 Application 1: The Minkowski problem for curvature measures . . . 34

3 Smooth solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1 Existence of smooth solutions by the continuity method . . . . . . . 39
3.2 Pogorelov’s counterexample to interior regularity . . . . . . . . . . 53
3.3 Pogorelov’s interior estimates and regularity of weak solutions . . . 56

4 Interior regularity of weak solutions . . . . . . . . . . . . . . . . . . . . . 65

4.1 Sections and normalized solutions . . . . . . . . . . . . . . . . . . 65
4.2 On the strict convexity of solutions . . . . . . . . . . . . . . . . . . 74
4.3 A Liouville theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 86
4.4 Application 2: Petty’s theorem . . . . . . . . . . . . . . . . . . . . 90
4.5 Interior C 1,α estimates . . . . . . . . . . . . . . . . . . . . . . . . 93
4.6 Application 3: The optimal transport problem with quadratic cost . 95
4.7 Geometry of sections . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.8 Interior W 2, p estimates . . . . . . . . . . . . . . . . . . . . . . . . 111
4.9 Application 4: The semigeostrophic equations . . . . . . . . . . . . 127
4.10 Interior C 2,α regularity . . . . . . . . . . . . . . . . . . . . . . . . 130
4.11 Wang’s counterexamples . . . . . . . . . . . . . . . . . . . . . . . 137

5 Further results and extensions . . . . . . . . . . . . . . . . . . . . . . . . 141

5.1 Further results on the Monge–Ampère equation . . . . . . . . . . . 141
5.2 The linearized Monge–Ampère equation . . . . . . . . . . . . . . . 148
5.3 A general class of Monge–Ampère-type equations . . . . . . . . . . 150
x Contents

Appendix: Useful facts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

A.1 Some general facts from linear algebra . . . . . . . . . . . . . . . . 159
A.2 Hausdorff measure . . . . . . . . . . . . . . . . . . . . . . . . . . 162
A.3 Convex geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
A.4 Convex functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
A.5 Tools from measure theory, nonlinear analysis, and PDEs . . . . . . 185

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
1

Introduction

The Monge–Ampère equation is a fully nonlinear, degenerate elliptic equation that

draws its name from its initial formulation in two dimensions, by the French mathe-
maticians Monge [92] and Ampère [8], about two hundred years ago. The classical
form of this equation is
det D2 u = f (x, u, ∇u) in Ω, (1.1)
where Ω ⊂ Rn is an open set, u : Ω → R is a convex function, and f : Ω × R × Rn →
R+ is given. As we shall explain below, the convexity of u is a necessary condition
to make the equation elliptic and to hope for regularity results.
The prototypical place where the Monge–Ampère equation appears is the “pre-
scribed Gaussian curvature equation", also known as the “Minkowski problem”: if we
take f = K(x)(1 + |∇u| 2 )(n+2)/2 , then (1.1) corresponds to imposing that the Gaussian
curvature of the graph of u at the point (x, u(x)) is equal to K(x) (see Section 2.6).
Other classical appearances of the Monge–Ampère equation can be found in
affine geometry (for instance, in the “affine sphere problem” and the “affine maximal
surfaces” problem; see [25, 98, 27, 115, 116, 117] and the references in [118]) and in
convex geometry (see, for instance, Section 4.4). More recently, the Monge–Ampère
equation has found important applications in optimal transportation (see Section 4.6)
and in meteorology (see Section 4.9). The goal of this book is to develop the existence,
uniqueness, and regularity theory for (1.1), and to show how this equation appears in
the above-mentioned problems.

1.1 On the degeneracy of the Monge–Ampère equation

Before entering into the theory of Monge–Ampère, we wish to discuss the terms
“fully nonlinear” and “degenerate elliptic” that we have used above. Also, we want
to explain why we are considering this equation only on convex functions.

1.1.1 A classification of second-order elliptic PDEs. The model second-order

elliptic PDE is the Laplace equation
∆u = 0,
2 1 Introduction

or, more generally, the Poisson equation

∆u = f ,

where f : Ω → R is some given function. These equations are called linear as they
depend linearly on the unknown function u.
Given a family of coefficients ai j : Ω → R, bi : Ω → R, c : Ω → R, i, j =
1, . . . , n, the linear equation
Õ Õ
ai j ∂i j u + bi ∂i u + cu = f
ij i

is called uniformly elliptic provided the coefficients ai j define positive-definite

bounded matrices, that is,

λ|ξ | 2 ≤ ai j (x)ξ i ξ j ≤ Λ|ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω,

Õ
(1.2)
ij

for some constants 0 < λ ≤ Λ < ∞, and degenerate elliptic if λ can be equal to zero
or Λ can be equal to infinity. When (1.2) holds, we shall also write λ Id ≤ ai j ≤ Λ Id.
When PDEs are nonlinear in the unknown u, one can classify them depending on
the kind of nonlinear structure.
More precisely, if the leading-order term (the term involving the second deriva-
tives of u) is linear in u, one says that the equation is semilinear; the model example is

∆u = f (x, u, ∇u)

for some given function f : Ω × R × Rn → R.

Instead, if the leading-order term is linear in the Hessian of u but depends nonlin-
early on u via lower-order terms, the equation is called quasilinear: elliptic equations
of this form are Õ
ai j (x, u, ∇u)∂i j u = f (x, u, ∇u),
ij

where the coefficients ai j = ai j (x, z, p) satisfy (1.2) for all (x, z, p) ∈ Ω × R × Rn . A

classical example is the p-Laplace equation.
Finally, a PDE is fully nonlinear if it is nonlinear in the Hessian of u. Model
elliptic examples are the Bellman equation
Õ 
α α α α
Õ
0 = sup ai j ∂i j u + bi ∂i u + c u − f
α∈ A ij i
 1.1 On the degeneracy of the Monge–Ampère equation 3

and the Isaacs’ equation

Õ 
α,β α,β α,β α,β
Õ
0 = inf sup ai j ∂i j u + bi ∂i u +c u− f ,
β ∈B α∈ A
ij i

α,β
where the coefficients aiαj , ai j satisfy (1.2) with constants independent of α and β.
Since the Monge–Ampère equation depends nonlinearly on the Hessian of u, it
falls into the last category.

1.1.2 Degeneracy and convexity. To understand the degenerate elliptic structure

of Monge–Ampère, we consider u : Ω → R a smooth solution of (1.1) with f =
f (x) > 0 smooth.
A standard technique to deal with nonlinear equations is to differentiate the
equation solved by u to obtain a linear second-order equation for its derivatives.
More precisely, fix a direction e ∈ Sn−1 and differentiate (1.1) in the direction e.
To simplify notation, we shall use subscripts to denote partial derivatives, that is,
ue = ∂e u, fe = ∂e f , ui j = ∂i j u, etc.
Since D2 u(x + εe) = D2 u(x) + εD2 ue (x) + o(ε), we see that
d d
det D2 u(x + εe) = det D2 u(x) + εD2 ue (x) .



dε ε=0 dε ε=0

Then using (A.1), we get

d
det D2 u(x + εe) = det D2 u(x) tr (D2 u(x))−1 D2 ue (x) .




dε ε=0

Hence, if we use ui j to denote the inverse matrix of ui j = ∂i j u and we use Einstein’s

convention of summing over repeated indices (ai j bi j = i j ai j bi j ), we deduce that
Í

(det D2 u)ui j ∂i j ue = fe in Ω.

Recalling that det D2 u = f > 0, we can rewrite the above equation as

fe
ui j ∂i j ue = in Ω. (1.3)
f
Thus, setting
fe
ai j := ui j , v := ue, g := ,
f
we see that v solves the linear equation

ai j ∂i j v = g.
4 1 Introduction

Now, if we want this equation to be uniformly elliptic, we need ai j = ui j to be positive

definite as in (1.2), which can be written in terms of its inverse ui j = ∂i j u as follows:

1 2 1
|ξ | ≤ D2 u(x)[ξ, ξ] ≤ |ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω.
Λ λ
So, u must be uniformly convex and C 1,1 . In particular, in order for the coefficients
ai j to be at least nonnegative definite, we are forced to restrict our attention to convex
functions. However, since ai j may vanish or be unbounded at some points, the
equation is degenerate elliptic.
Notice that if
1
Id ≤ D2 u ≤ C Id inside Ω (1.4)
C
for some constant C > 0, then Id/C ≤ ui j ≤ C Id and the linearized equation (1.3)
becomes uniformly elliptic. For this reason, proving the bound Id/C ≤ D2 u ≤ C Id
is key for the regularity of solutions to (1.1).
We shall use a crucial observation in the sequel:
Remark 1.1. Let u solve (1.1) with f ≥ a0 > 0, and assume that

kD2 u(x)k := sup ∂ee u(x) ≤ A ∀ x ∈ Ω.

e ∈S n−1

Then (1.4) holds. Indeed, given x ∈ Ω, we can choose a system of coordinates so

that D2 u(x) is a diagonal matrix with eigenvalues (λ1, . . . , λn ). Since det D2 u(x) =
i=1 λi , it follows that
În

n
Ö
λi = f (x) ≥ a0 and max λk ≤ A,
1≤k ≤n
i=1

and we get that

λj
Î
j a0
λi = Î ≥ ∀ i = 1, . . . , n,
k,i λk An−1
which proves (1.4) with C := max{A, An−1 /a0 }.

1.2 Some history

The first notable results on the Monge–Ampère equation are due to Minkowski
[90, 91]. At the end of the 19th and the beginning of the 20th century, he proved
the existence of weak solutions to the “prescribed Gaussian curvature problem” (this
 1.2 Some history 5

is now called the “Minkowski problem”): Given a function K on the sphere, find
a convex surface whose Gaussian curvature in polar coordinates is equal to K (see
Section 2.6).
Using convex polyhedra with given generalized curvatures at the vertices, forty
years later, Alexandrov proved the existence of a weak solution in all dimensions, as
well as the C 1 smoothness of solutions in two dimensions [1, 2, 3, 4]. Then, based on
these results, Alexandrov [5] (and also Bakelman [9] in two dimensions) introduced
the notion of a generalized solution to the Monge–Ampère equation, and proved
existence and uniqueness of solutions to the Dirichlet problem (see Chapter 2). Their
treatment also led to the Alexandrov–Bakelman maximum principle which plays a
fundamental role in the study of non-divergence elliptic equations (see, for instance,
[58, Section 9.8]).
The notion of weak solutions introduced by Alexandrov (now called “Alexandrov
solutions”) has often been used in recent years, and a lot of attention has been given
to proving smoothness of Alexandrov solutions under suitable assumptions on the
right-hand side and the boundary data.
The regularity of Alexandrov solutions in higher dimensions is a very delicate
problem. In the 1960s, Pogorelov found a convex function which is not of class
C 2 but satisfies the Monge–Ampère equation (1.1) inside B1/2 with positive analytic
right-hand side (see Section 3.2). As we shall explain in detail in Chapter 4, the main
obstacle to regularity is the presence of a line segment in the graph of u (in other
words, u is not strictly convex). Indeed, Calabi [24] and Pogorelov [97] were able
to prove a priori interior second- and third-derivative estimates for strictly convex
solutions, in turn proving the smoothness of strictly convex Alexandrov solutions
(see Section 3.3, where instead of Calabi’s estimates we use the interior regularity
theory for fully nonlinear uniformly elliptic equations established by Evans [41] and
Krylov [75] in the 1980s).
Later on, using the continuity method, Ivochkina [65], Krylov [76], and Caffarelli–
Nirenberg–Spruck [23] were able to show the existence of globally smooth solutions
to the Dirichlet problem. In particular, Alexandrov solutions are smooth up to the
boundary provided all given data are smooth (cf. Section 3.1).
In all the situations mentioned above one assumes that the right-hand side f
is positive and sufficiently smooth, but the many applications of Monge–Ampère
motivated the development of a regularity theory under weaker assumptions on f .
In the 1990s, under only the hypothesis that f is bounded away from zero and
infinity, Caffarelli proved the C 1,α regularity of strictly convex solutions [15]. Then,
if in addition f is continuous (resp. C 0,α ), using a perturbation argument, Caffarelli
proved an interior W 2, p estimate for any p > 1 (resp. C 2,α interior estimates) [14].
More recently, the author and De Philippis proved interior L log L estimates on D2 u
6 1 Introduction

when f is merely bounded away from zero and infinity [30], and together with Savin
2,1+ε
they improved this result showing that u ∈ Wloc [36] (see also [106]). All these
results, as well as some selected applications, are described in Chapter 4.
We also mention that these interior regularity results have a natural counterpart
at the boundary, which is briefly described in Section 5.1.2.
In recent years, applications to optimal transportation and antenna design prob-
lems have motivated the study of a much more general class of Monge–Ampère-type
equations, as well as their boundary regularity and their linearization (cf. Chapter 5).
It is important to remark that, despite all these recent developments, several
important questions on the regularity of solutions to Monge–Ampère are still open,
and the Monge–Ampère equation and its applications remain very active areas of
research.
2

Alexandrov solutions

In this chapter, we introduce the concept of weak solutions “à la Alexandrov” to the

Monge–Ampère equation, we describe their main properties, and prove existence and
uniqueness for the Dirichlet problem. Then we prove C 1 regularity in two dimensions,
and we apply this result to show the C 1 regularity of convex surfaces with bounded
Gaussian curvature.

2.1 The Monge–Ampère measure

To motivate the definition of weak solutions, given an open set Ω ⊂ Rn , consider

u : Ω → R a convex function1 of class C 2 satisfying
det D2 u = f in Ω (2.1)
for some f : Ω → R+ . Then given any Borel set E ⊂ Ω, it follows by the area
formula (see Theorem A.31) that
∫ ∫
f dx = det D2 u dx = |∇u(E)|.
E E

Notice that while the above argument needs u to be of class C 2 , the identity
∫
f = |∇u(E)|
E

makes sense if u is only of class C 1 . To find a definition when u is merely con-

vex one could try to replace the gradient ∇u(x) with the subdifferential ∂u(x) (see
Appendix A.4) and ask for the above equality to hold for any Borel set E.
This motivates the following definition:
Definition 2.1. Given an open set Ω ⊂ Rn and a convex function u : Ω → R, we
define the Monge–Ampère measure associated to u by
Ø
µu (E) := ∂u(x) .
x ∈E
1As explained in Appendix A.4, the definition of a convex function inside a general open set requires some
attention. Hence, we suggest the reader looks at Definition A.17 before continuing.
8 2 Alexandrov solutions

As we shall see, the

∫ basic idea of Alexandrov was to say that u is a weak solution
of (2.1) if µu (E) = E f for all Borel sets E ⊂ Ω. Although this seems natural, to
justify such a definition one needs some preliminary results.
However, in order to get some familiarity with the notion of Monge–Ampère
measure, we first give some examples.
Example 2.2. (1) If∫u ∈ C 1,1 (Ω), then the area formula (see Theorem A.31) gives
us that µu (E) = E det D2 u dx for any Borel set E. Hence,

µu = det D2 u dx.

(2) If u(x) = M |x|, then

 BM (0) if x = 0,


∂u(x) =

x
 M |x|
 if x , 0.

This implies that |∂u(Rn \ {0})| = |∂BM (0)| = 0 and we see that

µu = |BM (0)|δ0 = M n |B1 (0)|δ0 .

(3) Given k ∈ {1, . . . , n − 1}, write x = (x 0, x 00) ∈ Rk × Rn−k , and consider

u(x) = |x 0 |. Then,

 0, 0) ∈ R × R
0 k n−k : |y 0 | ≤ 1} if x 0 = 0,
 {(y


∂u(x) =

x
 |x 0 | , 0
 if x 0 , 0.

In particular, |∂u(Rn )| = |B1k (0) × {0}| = 0. Thus

µu ≡ 0.

|x | 2
(4) Let u(x) = 2 + |x1 |. Then, denoting the canonical basis by {e1, . . . , en }, we
get


 {x + e1 } if x1 > 0,
∂u(x) = {x − e1 }


if x1 < 0,
 {(t, x2, . . . , xn ) : |t| ≤ 1} if x1 = 0,



from which one deduces that

µu = dx + 2Hn−1 {x1 = 0},

where Hn−1 denotes the (n − 1)-dimensional Hausdorff measure (see Ap-

pendix A.2).
 2.1 The Monge–Ampère measure 9

2.1.1 On the subdifferential of disjoint sets. In order to justify the word “measure”
in Definition 2.1, we need to prove that µu is actually a Borel measure. This is not
obvious since if E and F are disjoint sets, it is not true that ∂u(E) and ∂u(F) are
disjoint (think of the trivial case when u is constant).
Theorem 2.3. Let Ω ⊂ Rn be an open set, u : Ω → R a convex function, and define
µu as in Definition 2.1. Then µu is a nonnegative locally finite Borel measure in Ω.
Proof. We first show that all Borel sets are µu -measurables, and then we prove that
µu is a measure.
Step 1: Borel sets are µu -measurables. Let us define

F := {E ⊂ Ω : ∂u(E) is Lebesgue measurable}.

We want to show that F is a σ-algebra containing the Borel sets. To simplify

notation, we denote the σ-algebra of Lebesgue measurable sets by L. We shall
use the following two facts, which can be easily checked using the definition of a
subdifferential:
Ø
∂u ∪k Ek = ∂u(Ek ) ∀ Ek ⊂ Ω (2.2)
k
and

∂u Ω \ E = ∂u(Ω) \ ∂u(E) ∪ ∂u(Ω \ E) ∩ ∂u(E) (2.3)

∀ E ⊂ Ω.

Step 1-a: F contains all compact sets. Let K ⊂ Ω be compact. Then it follows by
Lemma A.22 that ∂u(K) is compact as well; hence, in particular, it belongs to L,
which gives us that K ∈ F .
Step 1-b: F is closed under countable union. Let {Ek }k ∈N ⊂ F , that is, ∂u(Ek ) ∈ L
for all k. Since L is a σ-algebra and using (2.2), we deduce that

Ø
∂u ∪k Ek = ∂u(Ek ) ∈ L.
k

Therefore, ∪k Ek ∈ F .
Step 1-c: Ω ∈ F . It suffices to write Ω as a countable union of a compact set and
apply Steps 1-a and 1-b.
Step 1-d: F is closed under complement. Let Z be the set of measure zero defined
in Lemma A.30, and notice that

∂u(Ω \ E) ∩ ∂u(E) ⊂ Z ∀ E ⊂ Ω;
10 2 Alexandrov solutions

this implies that

∂u(Ω \ E) ∩ ∂u(E) ∈ L
is a set of measure zero.
Also, ∂u(Ω) ∈ L by Step 1-c. Thus, if ∂u(E) ∈ L, it follows
by (2.3) that ∂u Ω \ E is obtained by set operations on elements in L. Hence it
belongs to L as well. This proves that

E∈F ⇒ Ω \ E ∈ F,

as desired.
We have shown that F is a σ-algebra containing all compact sets. Consequently
it contains the Borel sets, which concludes the first part of the proof.
Step 2: µu is a measure. It is obvious that µu is nonnegative and µu (∅) = 0. Also,
given a compact set K ⊂ Ω, it follows by Lemma A.22 that ∂u(K) is compact, thus
µu (K) = |∂u(K)| < ∞. So, we only need to show that µu is σ-additive.
Let {Ek }k ∈N ⊂ F be a disjoint family. Although the sets {∂u(Ek )}k ∈N may not
be disjoint, we notice that

E j ∩ Ek = ∅ ⇒ ∂u(E j ) ∩ ∂u(Ek ) ⊂ Z,

where Z is again the set defined in Lemma A.30. Therefore, in particular,

|∂u(E j ) ∩ ∂u(Ek )| = 0 ∀ j , k.

Combining this estimate with (2.2), we deduce that

N
Ø N
Õ N
Õ
µu ∪k=1
N
Ek = ∂u(∪k=1
N
Ek ) = ∂u(Ek ) = |∂u(Ek )| = µu (Ek )


k=1 k=1 k=1

for any N ≥ 1, and the result follows by monotone convergence letting N → ∞. 

Before concluding this section, we discuss the relation between the Monge–
Ampère measure and the Legendre transform (see (A.16) and Remark A.28). This
will be useful later for the proof of Theorem 2.19.
Lemma 2.4. Let Ω ⊂ Rn be a convex set, u : Ω → R a convex function, and
u∗ : Rn → R the Legendre transform of u defined as in (A.16). Assume that
1
µu = f dx for some f ∈ Lloc (Ω). (2.4)

Then, ∫
f dx = |F | ∀ F ⊂ Rn Borel s.t. ∂u∗ (F) ⊂ Ω.
∂u ∗ (F)
 2.2 Alexandrov solutions: Definition and basic properties 11

Proof. First, we extend u as in (A.18). Then we notice that by the locality of the
subdifferential (see Remark A.18), u still satisfies (2.4) inside Ω.
Given a Borel set F satisfying ∂u∗ (F) define := ∂u ∂u ∗ (F) \F. Notice


⊂ Ω, G
that F ∩G = ∅ and that ∂u∗ (G) ⊂ ∂u∗ ∂u ∂u∗ (F) = ∂u∗ (F) (since ∂u = (∂u∗ )−1



by (A.20)). Hence, since F and G are disjoint, it follows by Lemma A.30 applied to
u∗ that
|∂u∗ (G)| = |∂u∗ (F) ∩ ∂u∗ (G)| = 0.

Thus, noticing that ∂u ∂u∗ (G) ⊃ G and that ∂u∗ (G) ⊂ Ω, it follows by (2.4) that

∫
0= f dx = µu ∂u∗ (G) = ∂u ∂u∗ (G) ≥ |G| = ∂u ∂u∗ (F) \ F .




∂u ∗ (G)

Thanks to this fact and, again, since µu = f dx, we conclude that

∫
f dx = µu ∂u∗ (F) = ∂u ∂u∗ (F) = |F |,



∂u ∗ (F)

as desired. 

2.2 Alexandrov solutions: Definition and basic properties

In the previous section, we introduced the Monge–Ampère measure associated to a

convex function u. Motivated by Example 2.2(1), we can now define the notion of
an Alexandrov solution to the Monge–Ampère equation.
Notice that since µu is a Borel measure, we can actually give meaning to the
equation “det D2 u = ν” when ν is merely a Borel measure (and not necessarily
a function). The possibility of considering general measures as right-hand sides
will be particularly useful when proving the existence of solutions (see the proof of
Theorem 2.13).
Here and in the sequel, when we talk about Borel measures, we shall always
implicitly assume that they are nonnegative and locally finite.

Definition 2.5. Let ν be a Borel measure in Ω, let u : Ω → R be a convex function,

and define µu as in Definition 2.1. We say that u is an Alexandrov solution of

det D2 u = ν

if µu = ν.
12 2 Alexandrov solutions

2.2.1 Stability of solutions under uniform convergence. An important property

of the Monge–Ampère measure is that it is stable under uniform convergence:
Proposition 2.6. Let Ω ⊂ Rn be an open set, and let uk : Ω → R be a sequence of
convex functions converging locally uniformly to a convex function u : Ω → R. Then
the associated Monge–Ampère measures µuk weakly∗ converge to µu , that is,
∫ ∫
ϕ dµuk → ϕ dµu ∀ ϕ ∈ Cc (Ω).
Ω Ω

Proof. It is a classical fact in measure theory (see [42, Section 1.9, Theorem 1]) that
demonstrating weak∗ convergence in Ω is equivalent to showing that
(1) µu (K) ≥ lim supk→∞ µuk (K) for all K ⊂⊂ Ω compact;
(2) µu (A) ≤ lim inf k→∞ µuk (A) for all A ⊂⊂ Ω open.
We now prove both properties.
Proof of (1). Fix K ⊂⊂ Ω compact. We claim that
∞ Ø
Ù ∞
∂u(K) ⊃ lim sup ∂uk (K) := ∂uk (K). (2.5)
k→∞ i=1 k=i

To prove this, we consider p ∈ lim supk→∞ ∂uk (K). This implies that there exists a
subsequence k j → ∞ such that p ∈ ∂uk j (K), that is,

p ∈ ∂uk j (xk j ) for some xk j ∈ K.

Up to a further subsequence we can assume that xk j → x ∈ K. Thus, writing the

subdifferential inequality

uk j (z) ≥ uk j (xk j ) + hp, z − xk j i ∀z ∈ Ω

and letting j → ∞, we get that p ∈ ∂u(x) ⊂ ∂u(K), proving the claim.

Thanks to (2.5) and by dominated convergence, we get
∞ Ø
Ù ∞ ∞
Ø
|∂u(K)| ≥ ∂uk (K) = lim ∂uk (K)
i→∞
i=1 k=i k=i
≥ lim lim sup |∂uk (K)|
i→∞ k→∞

= lim sup |∂uk (K)|,

k→∞

which concludes the proof of (1).

 2.2 Alexandrov solutions: Definition and basic properties 13

Proof of (2). Fix A ⊂⊂ Ω open. Since

|∂u(A)| = sup |∂u(K)|,

K ⊂ A,K cpt

it suffices to prove that |∂u(K)| ≤ lim inf k→∞ |∂uk (A)| for any compact set K ⊂ A.
Let Z be the set defined in Lemma A.30. Since | Z | = 0, it suffices to prove that

|∂u(K) \ Z | ≤ lim inf |∂uk (A)|,

k→∞

and this will follow, as in the proof of (1), provided we can show that
∞ Ù
Ø ∞
∂u(K) \ Z ⊂ lim inf ∂uk (A) := ∂uk (A). (2.6)
k→∞
i=1 k=i

To prove this fact, fix p ∈ ∂u(K) \ Z . Then

(
p ∈ ∂u(x) for some x ∈ K,
p < ∂u(y) for any y ∈ Ω \ {x},

and Remark A.19 yields

u(z) > `x, p (z) ∀ z ∈ Ω \ {x}, where `x, p (z) := u(x) + hp, z − xi.

Since ∂ A is compact, we see that

0 < δ := min u(z) − `x, p (z).

z ∈∂A

Set
vk := uk − `x, p and v := u − `x, p .
Thus, as vk → v locally uniformly, we deduce that for k  1,
δ δ
vk ≥ on ∂ A and vk (x) = uk (x) − u(x) ≤ ,
2 4
(see Figure 2.1). It follows that vk attains its minimum on A at some point xk ∈ A.
This implies that xk is a local minimum for vk ; therefore, xk is a global minimum by
the convexity of vk . Thus,

vk (z) ≥ vk (xk ) ∀ z ∈ Ω,

or equivalently (by the definition of vk ),

uk (z) ≥ uk (xk ) + hp, z − xk i ∀ z ∈ Ω,

which proves that p ∈ ∂uk (A) for k  1. This concludes the proof of (2.6). 
14 2 Alexandrov solutions

Figure 2.1. Since vk → v = u − `x, p locally uniformly and v attains its minimum at x ∈ A, for k
large enough vk attains its minimum at some point xk in the interior of A.

2.2.2 Monotonicity of the subdifferential. Before proving an important mono-

tonicity property of the Monge–Ampère measure, we make an assumption that we
will implicitly use through the whole book: Given an open set Ω ⊂ Rn and convex
function u : Ω → R, whenever we write that

u=g on ∂Ω,

we assume that u extends continuously on Ω and that the continuous extension of u

coincides with g on ∂Ω.
Lemma 2.7. Let Ω ⊂ Rn be an open set, O ⊂ Ω an open bounded set, and
u, v : Ω → R be two convex functions satisfying
(
u = v on ∂ O,
u ≤ v in O .

Then,
∂u(O) ⊃ ∂v(O),
and, in particular, µu (O) ≥ µv (O).
Proof. Let p ∈ ∂v(x) for some x ∈ O. As observed in Appendix A.4.1, from a
geometric point of view, this means that the affine function `x, p (z) := v(x) + hp, z − xi
touches v from below at x.
Since u ≤ v, we see that the constant

a := max `x, p (z) − u(z)

z ∈O

is nonnegative and that `x, p − a touches u from below at some point x0 ∈ O (see
Figure 2.2). We now distinguish two cases.
 2.2 Alexandrov solutions: Definition and basic properties 15

Figure 2.2. Let `x, p be a supporting hyperplane for v at x. Either by moving `x, p down by an
amount a > 0 we obtain a supporting hyperplane for u at some point x0 ∈ O (left), or `x, p is already
a supporting hyperplane for u at x (right).

Case 1: x0 ∈ O. If x0 ∈ O, then w := u − (`x, p − a) attains a local minimum at x0 ,

which is also a global minimum by the convexity of w. In other words,

`x, p − a ≤ u in Ω, `x, p (x0 ) − a = u(x0 ),

and p ∈ ∂u(x0 ) ⊂ ∂u(O), as desired.

Case 2: x0 ∈ ∂ O. Since u = v on ∂ O and x0 ∈ ∂ O, we deduce that a = 0. Therefore,

`x, p ≤ u ≤ v in O .

As `x, p (x) = v(x), this gives `x, p (x) = u(x). So, `x, p touches u from below at x and
p ∈ ∂u(x), which concludes the proof. 

2.2.3 Alexandrov maximum principle. A consequence of the monotonicity result

proved in the previous section is the celebrated Alexandrov maximum principle.
Notice that if a convex function u vanishes on ∂Ω, then necessarily Ω is convex.
Theorem 2.8. Let Ω be an open bounded convex set, and let u : Ω → R be a convex
function such that u = 0 on ∂Ω. Then there exists a dimensional constant Cn > 0
such that
|u(x)| n ≤ Cn diam(Ω)n−1 dist(x, ∂Ω)|∂u(Ω)| ∀ x ∈ Ω.
Proof. Let (x, u(x)) be a point on the graph of u, and consider the convex “conical”
function y 7→ Ĉx (y) with vertex at (x, u(x)) that vanishes on ∂Ω (see Figure 2.3).
Since u ≤ Ĉx in Ω (by the convexity of u), Lemma 2.7 implies that2

|∂ Ĉx (x)| ≤ |∂ Ĉx (Ω)| ≤ |∂u(Ω)|; (2.7)

so, to conclude the proof, it suffices to bound |∂ Ĉx (x)| from below.
16 2 Alexandrov solutions

Figure 2.3. The function Ĉx touches u from above at the point (x, u(x)) and vanishes on ∂Ω.
If |p| < |u(x)|/diam(Ω), then `x, p touches Ĉx from below at x. Moreover, there exists q with
|q| = |u(x)|/dist(x, ∂Ω) such that `x,q touches Ĉx from below at x.

Step 1: ∂ Ĉx (x) contains the ball B% (0) with % := |u(x)|/diam(Ω). Take p with
|p| < |u(x)|/diam(Ω), and consider the affine function

`x, p (z) := u(x) + hp, z − xi.

Notice that `x, p (x) = u(x) = Ĉx (x) and

`x, p (z) ≤ u(x) + |p| |z − x| ≤ u(x) + |p| diam(Ω) ≤ 0 ∀ z ∈ ∂Ω,

so (by convexity) `x, p ≤ Ĉx inside Ω (see Figure 2.3). Hence, `x, p touches Ĉx from
below at x, which implies that p ∈ ∂ Ĉx (x). By the arbitrariness of p, this proves that
∂ Ĉx (x) ⊃ B% (0), as desired.
Step 2: ∂ Ĉx (x) contains a vector of norm |u(x)|/dist(x, ∂Ω). Consider now x̄ ∈ ∂Ω
such that |x − x̄| = dist(x, ∂Ω), and set
x̄ − x |u(x)|
q := .
| x̄ − x| dist(x, ∂Ω)
Then `x,q (z) := u(x) + hq, z − xi satisfies

`x,q ( x̄) = u(x) + hq, x̄ − xi = u(x) + |u(x)| = 0,

and the hyperplane {`x,q = 0} is tangent to Ω at x̄ (see Figure 2.4). This implies that
`x,q ≤ 0 = Ĉx on ∂Ω. Indeed, given any z ∈ ∂Ω, we have that dist(x, ∂Ω) ≤ |z − x|,
so  
|z − x|
`x, p (z) ≤ u(x) 1 − ≤ 0.
dist(x, ∂Ω)
Arguing as in Step 1, this proves that q ∈ ∂ Ĉx (x).
2 Actually, as a consequence of the argument in the proof, one can show that ∂Ĉ x (x) = ∂Ĉ x (Ω).
 2.3 The Dirichlet problem: Uniqueness 17

Figure 2.4. On the left, the hyperplane {`x,q = 0} is tangent to Ω at x̄. On the right, the subdifferential
of Ĉx at x contains the cone C generated by q and Σ % .

Step 3: Conclusion. Thanks to Steps 1 and 2, we know that

∂ Ĉx (x) ⊃ B% (0) ∪ {q}, (2.8)

where % = |u(x)|/diam(Ω) and |q| = |u(x)|/dist(x, ∂Ω). Now, let Σ % denote the
intersection of B% (0) with the hyperplane passing through the origin and orthogonal
to q, that is,
Σ % := B% (0) ∩ {z : hz, qi = 0}.
Since ∂ Ĉx (x) is convex, it follows by (2.8) that ∂ Ĉx (x) contains the cone C generated
by q and Σ % (see Figure 2.4). Hence,

|u(x)| n
|∂ Ĉx (x)| ≥ | C | = cn |q| %n−1 = cn ,
diam(Ω)n−1 dist(x, ∂Ω)
where cn > 0 is a dimensional constant. Recalling (2.7), this concludes the proof. 

2.3 The Dirichlet problem: Uniqueness

The aim of this section is to show a comparison principle between Alexandrov

solutions. This implies, in particular, uniqueness of Alexandrov solutions, and a
fortiori uniqueness of classical solutions (since any classical solution is a weak
solution in the sense of Alexandrov).
Before proving the comparison principle, we first need a preliminary result on the
behavior of the Monge–Ampère measure with respect to addition and multiplication.

Lemma 2.9. Let u, v : Ω → R be convex functions. Then,

µu+v ≥ µu + µv and µλu = λ n µu ∀ λ > 0.

18 2 Alexandrov solutions

Proof. We prove only the first statement; the proof of the second statement is
analogous.
We begin by noticing that when u, v ∈ C 2 (Ω) the statement becomes

det(D2 u + D2 v) ≥ det D2 u + det D2 v,

and this follows immediately from (A.7).

To prove the result in the general case, we argue by approximation: Extend u and
v to convex functions on all of Rn , and define uk := u ∗ ρk and vk := v ∗ ρk , where
ρk ∈ Cc∞ (Rn ) is a smooth sequence of convolution kernels. Since uk , vk ∈ C 2 (Ω),
we know that
µuk +vk ≥ µuk + µvk ,
or that
∫ ∫ ∫
ϕ dµuk +vk ≥ ϕ dµuk + ϕ dµvk ∀ ϕ ∈ Cc (Ω), ϕ ≥ 0. (2.9)
Ω Ω Ω

Then, since uk → u and vk → v locally uniformly inside Ω, it follows by Proposi-

tion 2.6 that

µuk *∗ µu, µvk *∗ µv, and µuk +vk *∗ µu+v,

and the result follows by letting k → ∞ in (2.9). 

2.3.1 A comparison principle. We now prove a comparison principle between

convex functions having ordered Monge–Ampère measures and boundary data.
Theorem 2.10. Let U ⊂ Ω be an open bounded set, and let u, v : Ω → R be two
convex functions satisfying (
µu ≤ µv in U ,
u≥v on ∂ U .
Then,
u≥v in U .
Proof. We begin by making the inequality between µu and µv strict. To do this, we
consider the function
|z| 2 − R2
vε := v + εw with w(z) := ,
2
where R  1 is chosen so that U ⊂ BR (0). By construction we have vε ≤ v ≤ u on
∂Ω. Also, it follows by Lemma 2.9 and Example 2.2(1) that

µvε ≥ µv + µεw = µv + det(εD2 w) dx = µv + ε n dx. (2.10)

 2.3 The Dirichlet problem: Uniqueness 19

Suppose, to the contrary, that the open set O := {u < v} ⊂⊂ U is nonempty. Then
Oε := {u < vε } ⊂⊂ U is also nonempty for ε  1, and it follows by the inequality
µv ≥ µu and (2.10) that
µvε (Oε ) ≥ µv (Oε ) + ε n | Oε | ≥ µu (Oε ) + ε n | Oε | > µu (Oε ). (2.11)
On the other hand, since (
u = vε on ∂ Oε,
u ≤ vε in Oε,
we can apply Lemma 2.7 to deduce that µu (Oε ) ≥ µvε (Oε ). This contradicts (2.11)
and concludes the proof. 

2.3.2 Uniqueness and stability of weak solutions. An immediate consequence of

Theorem 2.10 is the uniqueness of solutions to the Dirichlet problem.
Corollary 2.11. Let Ω be an open bounded set, g : ∂Ω → R a continuous function,
and ν a Borel measure in Ω. Then, there exists at most one convex function u : Ω → R
solving the Dirichlet problem
(
µu = ν in Ω,
u=g on ∂Ω.

Another consequence of Theorem 2.10 is the stability of solutions to the Dirichlet

problem. For simplicity, we first restrict our attention to the case g = 0. We shall
discuss the general case in Proposition 2.16.
Corollary 2.12. Let {Ωk }k ∈N be a sequence of open bounded convex sets, let νk be
a Borel measure in Ωk , and let uk : Ω → R solve the Dirichlet problem
(
µuk = νk in Ωk ,
uk = 0 on ∂Ωk .

Assume that
sup νk (Ωk ) < ∞
k
and that there exist an open bounded convex set Ω and a Borel measure ν in Ω such
that
dist(∂Ωk , ∂Ω) → 0 and νk *∗ ν as k → ∞. (2.12)
Then uk |Ω converges locally uniformly in Ω to the unique solution u of the Dirichlet
problem (
µu = ν in Ω,
(2.13)
u=0 on ∂Ω.
20 2 Alexandrov solutions

Proof. Since supk νk (Ωk ) < ∞ and diam(Ωk ) → diam(Ω) < ∞ (because of (2.12)),
it follows by Theorem 2.8 that

|uk (x)| ≤ C dist(x, ∂Ωk )1/n ∀ x ∈ Ωk , (2.14)

for some constant C independent of k. This implies, in particular, that the functions
uk are uniformly bounded in Ωk . Therefore, they are locally Lipschitz (uniformly in
k) by Corollary A.23.
Since Ωk → Ω (by (2.12)), the bound (2.14) and the Ascoli–Arzelà theorem
imply that there exists a subsequence uk j converging locally uniformly in Ω to some
convex function û : Ω → R satisfying | û(x)| ≤ C dist(x, ∂Ω)1/n . This implies, in
particular, that û|∂Ω = 0. In addition, it follows by Proposition 2.6 that µuk j |Ω *∗ µû ,
and since by assumption µuk = νk *∗ ν, we deduce that µû = ν.
Hence, we have proved that û solves the Dirichlet problem (2.13). As Corol-
lary 2.11 implies that such a solution is unique, we deduce that any unique cluster
point for the family {uk }k ∈N coincides with such a solution; thus the whole sequence
uk must converge to the unique solution of (2.13), as desired. 

2.4 The Dirichlet problem: Existence

The aim of this section is to prove existence of weak solutions on bounded convex
domains. We shall first focus on the case of zero boundary data, and later we will
address the general case.

2.4.1 Existence of weak solutions with zero boundary data.

Theorem 2.13. Let Ω be an open bounded convex set, and let ν be a Borel measure
on Ω with ν(Ω) < ∞. Then there exists a unique convex function u : Ω → R solving
the Dirichlet problem (
µu = ν in Ω,
(2.15)
u=0 on ∂Ω.

Proof. Since uniqueness follows by Corollary 2.11, we only need to show existence.
By the stability result proved in Corollary 2.12, since any finite measure can be
approximated in the weak∗ topology by a finite sum of Dirac deltas, we only need to
solve the Dirichlet problem (2.15) when ν = i=1 αi δxi with xi ∈ Ω and αi > 0.
ÍN
 2.4 The Dirichlet problem: Existence 21

Figure 2.5. The “conical” function Ĉxi vanishes on ∂Ω and takes its minimum value −1 at xi .

To prove existence of a solution, we use the so-called Perron method: we define

S [ν] := {v : Ω → R convex : v|∂Ω = 0, µv ≥ ν in Ω},

and we show that the largest element in S[ν] is the desired solution. We split the
argument into several steps.
Step 1: S[ν] , ∅. To construct an element of S[ν], we consider the “conical”
function Ĉxi that is 0 on ∂Ω and takes the value −1 at its vertex xi (see Figure 2.5).
The Monge–Ampère measure of this function is concentrated at xi and has mass equal
to some positive number βi corresponding to the measure of the set of supporting
hyperplanes at xi (compare with Example 2.2(2)).
Now, consider the convex function v̄ := i=1 λĈxi where λ > 0 has to be chosen.
ÍN
We notice that v̄|∂Ω = 0. In addition, provided λ > 0 is sufficiently large, Lemma 2.9
implies that

N
Õ N
Õ N
Õ N
Õ
µv̄ ≥ µλĈ x = λ n µĈ x = λ n βi δxi ≥ αi δxi .
i i
i=1 i=1 i=1 i=1

This proves that v̄ ∈ S [ν].

Step 2: v1, v2 ∈ S [ν] ⇒ max{v1, v2 } ∈ S [ν]. To simplify notation, set w :=
max{v1, v2 }, and define

Ω0 := {v1 = v2 }, Ω1 := {v1 > v2 }, and Ω2 := {v1 < v2 }.

Also, given a Borel set E ⊂ Ω, consider Ei := E ∩ Ωi for i = 0, 1, 2.

Since Ω1 and Ω2 are open sets, w|Ω1 = v1 , and w|Ω2 = v2 , it follows by Re-
mark A.18 (and see Figure 2.6) that

∂w(E1 ) = ∂v1 (E1 ) and ∂w(E2 ) = ∂v2 (E2 ).

22 2 Alexandrov solutions

Figure 2.6. The function w := max{v1, v2 } coincides with v1 (resp. v2 ) in Ω1 ∪ Ω0 (resp. Ω2 ∪ Ω0 ).

While in Ω1 (resp. Ω2 ) the subdifferentials of w and v1 (resp. w and v2 ) coincide, in Ω0 the
subdifferential of w contains both the subdifferential of v1 and the subdifferential of v2 (in the
picture, p1 ∈ ∂v1 (x), p2 ∈ ∂v2 (x), and we see that {p1, p2 } ⊂ ∂w(x)).

In addition, since w = v1 on Ω0 and w ≥ v1 everywhere else, using the definition of

a subdifferential, we see that

p1 ∈ ∂v1 (x), x ∈ Ω0 ⇒ p1 ∈ ∂w(x),

which proves that ∂w(E0 ) ⊃ ∂v1 (E0 ) (and analogously for v2 ). So, recalling that
v1, v2 ∈ S [ν], we get
2
Õ 2
Õ
µw (E) = µw (E0 ) ≥ µv1 (E0 ) + µv1 (E1 ) + µv2 (E2 ) ≥ ν(Ei ) = ν(E).
i=0 i=0

This proves the result.

Step 3: The function u := supv ∈S[ν] v belongs to S[ν]. Take {vk }k ∈N ⊂ S [ν] a
countable subset which is dense with respect to the local uniform convergence, and
observe that
u = sup vk = lim wm for wm := max vk .
k ∈N m→∞ 1≤k ≤m

It follows by Step 2 (applied finitely many times) that wm ∈ S [ν]; hence, µwm ≥ ν,
or equivalently,
∫ ∫
ϕdµwm ≥ ϕ dν ∀ ϕ ∈ Cc (Ω), ϕ ≥ 0. (2.16)
Rn Rn

Since wm → u locally uniformly, it follows by Proposition 2.6 that µwm *∗ µu .

Therefore, it suffices to let m → ∞ in (2.16) to obtain that µu ≥ ν. Also, we deduce
immediately that u|∂Ω = 0 by construction; hence, u ∈ S [ν].
 2.4 The Dirichlet problem: Existence 23

Figure 2.7. On the left, if the measure µu is not supported at the points {x1, . . . , x N }, then we can
construct a function u larger than u that still belongs to S[ν]. On the right, if the subdifferential of
u at x j is too large, we can construct a function ũ larger than u that still belongs to S[ν].

Step 4: µu = ν. This is the most delicate step. Notice that this step combined with
Step 3 proves that u solves (2.15).
Step 4-a: The measure µu is supported at the points {x1, . . . , x N }. Suppose that this
is not the case. Then there exists a set E ⊂ Ω such that

E ∩ {x1, . . . , x N } = ∅ and |∂u(E)| = µu (E) > 0. (2.17)

Because xi < E, the sets {∂u(E) ∩ ∂u({xi })}1≤i ≤ N are contained inside the set Z
defined in Lemma A.30, hence they are all of measure zero. Also, for the same
reason, |∂u(Ω) ∩ ∂u(∂Ω)| = 0, so it follows by (2.17) that
N
Ø 
∂u(E) \ ∂u({xi }) ∪ ∂u(∂Ω) > 0.
i=1

In particular, the above set is nonempty, and we can find x̄ ∈ Ω \ {x1, . . . , x N } and
p̄ ∈ Rn such that
N
Ø
p̄ ∈ ∂u( x̄) and p̄ < ∂u({xi }) ∪ ∂u(∂Ω).
i=1

Then, it follows by Remark A.19 that there exists δ > 0 such that

u ≥ `x̄, p̄ + 2δ on {x1, . . . , x N } ∪ ∂Ω, (2.18)

where `x̄, p̄ (z) := u( x̄) + h p̄, z − x̄i.

We now consider the function

u := max{u, `x̄, p̄ + δ}

(see Figure 2.7). Notice that u is convex, u u, and it follows by (2.18) that u = u in
a neighborhood of {x1, . . . , x N } ∪ ∂Ω. In particular, u|∂Ω = 0 and ∂u(xi ) = ∂u(xi )
24 2 Alexandrov solutions

for all i = 1, . . . , N (see Remark A.18), which implies that u ∈ S [ν]. Yet u was
defined as the largest element in S [ν], a contradiction.
Step 4-b: µu ({xi }) = ν({xi }) for all i = 1, . . . , N. By Step 3 we already know that
u ∈ S [ν]; hence µu ({xi }) ≥ ν({xi }) = αi for all i = 1, . . . , N. Assume, to the
contrary, that there is an index j ∈ {1, . . . , N } such that

β j := µu ({x j }) = |∂u(x j )| > α j .

Since ∂u(x j ) is a convex set of positive measure, pick a vector p ∈ Rn that belongs to
the interior of ∂u(x j ), define `x j , p (z) := u(z) + hp, z − x j i, and consider the function
U := u − `x j , p . Notice that ∂U(z) = ∂u(z) − p for all z, which implies, in particular,
that |∂U(x j )| = β j and that 0 belongs to the interior of ∂U(x j ).
Choose δ > 0 small enough so that
 
U ≤ U(x j ) + δ ∩ {x1, . . . , x N } ∪ ∂Ω = ∅,


and define the function

if U > U(x j ) + δ,

e := U(z)
U(z)
(1 − δ)U(z) + δ[U(x j ) + δ] if U ≤ U(x j ) + δ.
Taking δ > 0 even smaller if necessary, we observe that
e j )| = (1 − δ)n |∂U(x j )| = (1 − δ)n β j > α j .
|∂U(x

Notice that U
e U and U e = U in a neighborhood of {x1, . . . , x N } ∪ ∂Ω. Hence,
considering ũ := U
e + `x j , p (see Figure 2.7), we see that ũ u,

ũ|∂Ω = u|∂Ω = 0, ∂ ũ(xi ) = ∂u(xi ) ∀ i , j, and |∂ ũ(x j )| > α j .

Thus, ũ ∈ S [ν], but this contradicts the maximality of u and concludes the proof. 

2.4.2 Existence of weak solutions with continuous boundary data. In the previ-
ous section, we proved the solvability of the Dirichlet problem when the boundary
data is zero. Now we explain how to extend this result to general continuous bound-
ary data. For this case, we need to assume the strict convexity of Ω, that is, any
supporting hyperplane (see Definition A.8) touches Ω at only one point.
Theorem 2.14. Let Ω be an open bounded strictly convex set, let g : ∂Ω → R be a
continuous function, and let ν be a Borel measure on Ω with ν(Ω) < ∞. Then there
exists a unique convex function u : Ω → R solving the Dirichlet problem
(
µu = ν in Ω,
u=g on ∂Ω.
 2.4 The Dirichlet problem: Existence 25

Figure 2.8. Since Ω is strictly convex, there exists η > 0 such that the set ∂Ω \ Bδ (0) is contained
inside the half-space {z n ≤ −η}.

In order to prove this result, we need the following simple lemma:

Lemma 2.15. Let Ω be an open bounded strictly convex set, and let g : ∂Ω → R
be a continuous function. Then, for any ε > 0 and x ∈ ∂Ω, there exist two affine
± : Rn → R such that
functions `x,ε
+
`x,ε
−
≤ g ≤ `x,ε on ∂Ω (2.19)

and
+
g(x) − ε ≤ `x,ε
−
(x) ≤ `x,ε (x) ≤ g(x) + ε. (2.20)
Proof. Fix ε > 0 and assume, without loss of generality, that

x = 0 ∈ ∂Ω and Ω ⊂ {z = (z1, . . . , z n ) ∈ Rn : z n < 0}.

By the continuity of g, there exists δ > 0 such that

|g − g(0)| ≤ ε on ∂Ω ∩ Bδ (0) (2.21)

(see Figure 2.8). Also, by the strict convexity of Ω, there exists η > 0 such that

∂Ω \ Bδ (0) ⊂ {z n ≤ −η}. (2.22)

Now, consider the affine functions

kg − g(0)k L ∞ (∂Ω)
`0,ε
−
(z) := g(0) − ε + zn
η
and
+ kg − g(0)k L ∞ (∂Ω)
`0,ε (z) := g(0) + ε − z n,
η
and notice that with this definition, (2.20) is immediate.
26 2 Alexandrov solutions

Also, since Ω ⊂ {z n < 0}, it follows by (2.21) that

+
`0,ε
−
≤ g ≤ `0,ε on ∂Ω ∩ Bδ (0),

while (2.22) and the definition of `0,ε

± imply

+
`0,ε
−
≤ g ≤ `0,ε on ∂Ω \ Bδ (0),

proving (2.19). 
A first consequence of this result is the stability of solutions for continuous
boundary data. Like Corollary 2.12, one could state the following result with varying
domains (and varying boundary data if desired). However, since we do not need
such a general result, for simplicity we state and prove a stability result given a fixed
domain.
Proposition 2.16. Let Ω be an open bounded strictly convex set, let g : ∂Ω → R
be a continuous function, let {νk }k ∈N be a family of Borel measures in Ω, and let
uk : Ω → R solve the Dirichlet problem
(
µuk = νk in Ω,
uk = g on ∂Ω.

Assume that supk νk (Ω) < ∞ and that νk *∗ ν as k → ∞. Then uk converges

uniformly in Ω to the unique solution u of the Dirichlet problem
(
µu = ν in Ω,
(2.23)
u=g on ∂Ω.

Proof. The proof of this result is essentially the same as the proof of Corollary 2.12,
the only difference being that we cannot apply Theorem 2.8 to control uk near ∂Ω.
Instead, we simply notice that by the convexity of uk and Theorem 2.10,

vk − kgk L ∞ (∂Ω) ≤ uk ≤ kgk L ∞ (∂Ω),

where vk solves (
µvk = νk in Ω,
vk = 0 on ∂Ω.
Since
− C dist(z, ∂Ω)1/n ≤ vk (z) ≤ 0 ∀z ∈ Ω (2.24)
by Theorem 2.8, we deduce that the functions vk are uniformly bounded; hence, so
are the functions uk . With this in hand, we can argue as we have done in the proof
 2.4 The Dirichlet problem: Existence 27

of Corollary 2.12, to deduce that there exists a subsequence uk j converging locally

uniformly to a function u satisfying µu = ν.
We claim that u|∂Ω = g. To this aim, fix ε > 0 and x ∈ ∂Ω, and consider the
affine functions `x,ε
± provided by Lemma 2.15. Since µ
vk +`x,
±
ε
= µvk , (2.19) and
Theorem 2.10 imply that
+
vk + `x,ε
−
≤ uk ≤ vk + `x,ε in Ω.

Combining this with (2.24) gives

−C dist(·, ∂Ω)1/n + `x,ε

− +
≤ uk ≤ `x,ε in Ω.

Hence, choosing k = k j and taking the limit as j → ∞, we obtain

−C dist(·, ∂Ω)1/n + `x,ε

− +
≤ u ≤ `x,ε in Ω.

Recalling (2.20), this implies that

g(x) − ε ≤ lim −C dist(z, ∂Ω)1/n + `x,ε

−

(z)
z→x
≤ lim inf u(z)
z→x
≤ lim sup u(z)
z→x
≤ lim ` + (z) ≤ g(x) + ε,
z→x x,ε

and since ε > 0 and x were arbitrary, we deduce that u is continuous up to the
boundary and that u|∂Ω = g.
This proves that u solves the Dirichlet problem (2.23), and as such a solution is
unique (by Corollary 2.11), it follows that the whole sequence uk converge to u. 
We are now ready to prove Theorem 2.14.
Proof of Theorem 2.14. The proof is very similar to that of Theorem 2.13, so we only
provide a sketch. Again, thanks to Corollary 2.11, we only need to show existence.
Step 1: Construction of u. By the stability result in Proposition 2.16, it suffices to
solve the Dirichlet problem (2.15) when ν = i=1 αi δxi with xi ∈ Ω and αi > 0.
ÍN
Define
S [ν] := {v : Ω → R convex : v|∂Ω ≤ g, µv ≥ ν in Ω}.
Considering the “conical” functions Ĉxi that are 0 on ∂Ω and take the value −1 at
their vertices xi and setting
N
Õ
v̄ := λĈxi , (2.25)
i=1
28 2 Alexandrov solutions

we have that µv̄ ≥ ν if λ is large enough (see Step 1 in the proof of Theorem 2.13);
hence, v̄ − kgk L ∞ (∂Ω) belongs to S[ν], which proves that S[ν] is nonempty.
Then we define
u := sup v, (2.26)
v ∈S[ν]

and an identical argument to the one given when g = 0 shows that µu = ν. So, the
only thing left to prove is that u = g on the boundary of Ω.

Step 2: u is continuous up to the boundary and u|∂Ω = g. The argument here is

similar to the one used in the proof of Proposition 2.16.
Fix ε > 0 and x ∈ ∂Ω, and consider the affine functions `x,ε ± provided by

Lemma 2.15. Recalling that the function v̄ defined in (2.25) satisfies µv̄ ≥ ν, it
follows by (2.19) that v̄ + `x,ε
− ∈ S [ν]. Thus, from (2.26), we get

v̄ + `x,ε
−
≤u in Ω.

On the other hand, by the convexity of Ω, we see that any convex function that is
+ inside Ω. It follows that
below g on ∂Ω will be below `x,ε

+
u ≤ `x,ε in Ω.

In particular,

g(x) − ε = lim v̄(z) + `x,ε

−

(z)
z→x
≤ lim inf u(z)
z→x
≤ lim sup u(z)
z→x
+
≤ lim `x,ε (z) ≤ g(x) + ε,
z→x

and the arbitrariness of ε and x proves the result. 

2.5 C 1 regularity in 2-D

The aim of this section is to show the C 1 regularity of two-dimensional solutions

to the Monge–Ampère equation with bounded right-hand side. As we shall see, the
same argument shows that solutions with right-hand side bounded away from zero
are strictly convex.
 2.5 C 1 regularity in 2-D 29

Figure 2.9. If u < C 1 (Ω), then u(x) = 0 for some x ∈ Ω. Since u|∂Ω = 0 and u is convex, this
implies that u ≡ 0.

2.5.1 C 1 regularity in 2-D for locally bounded right-hand side. The following
result, due to Alexandrov [3], shows that either a solution is C 1 or any singularity
propagates along a segment crossing the whole domain. This dichotomy is a sign
of the degeneracy of the Monge–Ampère equation, which we already mentioned in
Section 1.1.2. Unless the boundary data prevent such behavior, the solution may be
singular. Notice that in view of Example 2.2(3), the following result is optimal.
Theorem 2.17. Let Ω ⊂ R2 be an open set, and let u : Ω → R be a convex function.
Assume that
µu = f dx for some f ∈ Lloc ∞
(Ω).
Then, either
(1) u ∈ C 1 (Ω);
(2) or there exists a segment Σ crossing Ω along which u is affine and nondiffer-
entiable. More precisely, there exist p0, p1 ∈ R2 , with p0 , p1 , such that

∂Σ ⊂ ∂Ω, u|Σ = affine, and ∂u(z) ⊃ [p0, p1 ] ∀ z ∈ Σ.

Before proving this result, we first show how it implies the C 1 regularity of
solutions with zero boundary data.
Corollary 2.18. Let Ω ⊂ R2 be an open bounded convex set, f : Ω → R+ a locally
bounded function, and let u : Ω → R be a convex function solving
(
µu = f dx in Ω,
u=0 on ∂Ω.

Then u ∈ C 1 (Ω).

Proof. Assume, to the contrary, that u < C 1 (Ω). Then it follows by Theorem 2.17
that there is a segment Σ crossing Ω along which u is affine (see Figure 2.9). Since
30 2 Alexandrov solutions

u|∂Ω = 0 we deduce that u|Σ ≡ 0. In particular, there exists a point x ∈ Ω such that
u(x) = 0, and the convexity of u implies that u ≡ 0 (recall u|∂Ω = 0). This contradicts
that u < C 1 (Ω). 
Proof of Theorem 2.17. Assume that u < C 1 (Ω). Then it follows by Lemmas A.24,
A.20, and A.21 that ∂u(x) is not a singleton at some point x̄ ∈ Ω.
Up to subtracting an affine function and a change of coordinates, we can assume
that x̄ = 0 and

u ≥ 0, u(0) = 0, and ∂u(0) ⊃ [−αe1, αe1 ],

where α > 0 and {e1, e2 } denotes the canonical basis of R2 . In particular,

u(x) ≥ α|x1 | ∀ x = (x1, x2 ) ∈ R2 . (2.27)

We can also assume that α ≤ 1 by replacing α with min{α, 1}.

Claim: The singularity propagates along the e2 -axis. More precisely, we set Σ :=
Ω ∩ {x1 = 0} and claim that
u|Σ = 0. (2.28)

We first show how (2.28) implies the desired result, and then we prove it.
Step 1: (2.28) concludes the proof. We need to show that

∂u(z) ⊃ [−αe1, αe1 ] ∀ z ∈ Σ,

and this is an immediate consequence of (2.28), (2.27), and the definition of subdif-
ferential.
Step 2: Proof of (2.28). Define

τ̂ := sup{τ ≥ 0 : u(0, τ) = 0} and τ̄ := inf{τ ≤ 0 : u(0, τ) = 0},

and assume, for the sake of contradiction, that { τ̂e2, τ̄e2 } 1 ∂Ω. Then either τ̂e2 or
τ̄e2 is in Ω, and, without loss of generality, we assume that τ̂e2 ∈ Ω.
Let Ω0 ⊂⊂ Ω be a neighborhood of [0, τ̂e2 ] and L ≥ 1 be a Lipschitz constant for
u|Ω0 (since u is locally Lipschitz in Ω by Corollary A.23, L is finite). Fix δ > 0 small
enough (to be chosen later), set h := u(0, τ̂ + δ) > 0, and consider the rectangle

R := [−h, h] × [τ̂ − δ, τ̂ + (1 + L)δ].

 2.5 C 1 regularity in 2-D 31

Figure 2.10. The rectangle R and its lateral and upper boundary G.

Notice that by choosing δ sufficiently small, we can ensure that R ⊂ Ω0. Moreover,
define
G := {±h} × [τ̂ − δ, τ̂ + (1 + L)δ] ∪ [−h, h] × { τ̂ + (1 + L)δ}
(see Figure 2.10).
Step 2-a: u|G ≥ αh. It follows immediately from (2.27) that

u ≥ αh on {±h} × [τ̂ − δ, τ̂ + (1 + L)δ].

Also, as u is L-Lipschitz inside R and u(0, τ̂) = 0, we deduce that

h
u(s, τ̂) ≤ h ∀ |s| ≤ . (2.29)
L
Thus, since u is convex,

u(z + t(z − x)) ≥ u(z) + t[u(z) − u(x)] ∀ t ≥ 0, ∀ x, z ∈ R.

Now, thanks to (2.29) and that u(0, τ̂ + δ) = h, applying this inequality with z =
(0, τ̂ + δ), x ∈ − Lh , Lh × { τ̂}, and t = L, we see that

u≥h on [−h, h] × { τ̂ + (1 + L)δ}

(see Figure 2.11). Recalling that α ≤ 1, this proves the desired estimate.
Step 2-b: µu (R) & h/δ. Given
α α αh
− ≤a≤ and 0<b≤ , (2.30)
4 4 4(1 + L)δ
32 2 Alexandrov solutions

h i
Figure 2.11. Let z = (0, τ̂ + δ) and x ∈ − Lh , Lh × { τ̂}. Since u(z) = h and u(x) ≤ h, the convexity
of u implies that u(z + L(z − x)) ≥ h.

consider the affine function

`a,b (z) := az1 + b(z2 − τ̂),

where z = (z1, z2 ). Notice that thanks to (2.30),

α αh α
`a,b |G ≤ h+ (1 + L)δ ≤ h < u|G .
4 4(1 + L)δ 2
Also, since b > 0 and recalling (2.27),
α
`a,b |∂R\G < |x1 | ≤ α|x1 | ≤ u.
4
Therefore, (
`a,b |∂R < u|∂R,
`a,b (0, τ̂) = u(0, τ̂).
This implies that we can choose ε > 0 (depending on a, b) sufficiently small so that

O := {u < `a,b + ε} ⊂⊂ R.
Then Lemma 2.7 yields that

∂u(R) ⊃ ∂u(O) ⊃ ∂`a,b (O) = {(a, b)}.

Since (a, b) ∈ R2 was an arbitrary point satisfying (2.30), we get

α α αh
   
∂u(R) ⊃ − , × 0, ;
4 4 4(1 + L)δ
hence
α2 h
µu (R) ≥ . (2.31)
8(1 + L)δ
 2.5 C 1 regularity in 2-D 33

Step 2-c: | R | ≈ hδ and conclusion. Since

| R | = 2(2 + L)hδ,

it follows by the Monge–Ampère equation that

∫
µu (R) = f dx ≤ | R |k f k L ∞ (Ω0 ) ≤ 2(2 + L)hδk f k L ∞ (Ω0 ) .
R

Combining this bound with (2.31), we conclude that

α2
δ2 ≥ ,
16(1 + L)(2 + L)k f k L ∞ (Ω0 )
which is impossible if δ is sufficiently small. This contradiction concludes the
proof. 

2.5.2 Strict convexity in 2-D for right-hand side bounded away from zero. When
µu = f dx with f bounded away from zero, the natural analogue of Theorem 2.17 is
that u is strictly convex. As shown in [18], one can give a direct proof of this result.
However, it is also possible to show that this is a consequence of Theorem 2.17
by exploiting the Legendre transform to exchange upper and lower bounds on the
Monge–Ampère measure. We follow the second strategy because it is independently
interesting.
Theorem 2.19. Let Ω ⊂ Rn be an open set and u : Ω → R be a convex function.
Assume that
µu = f dx with 1/ f ∈ Lloc
∞
(Ω). (2.32)
Then u is strictly convex inside Ω.
Proof. Assume, to the contrary, that u coincides with an affine function `(z) :=
h p̄, zi + ā on a nontrivial segment Σ ⊂ Ω. Let Ω0 ⊂ Ω be an open convex domain
such that Σ ∩ Ω0 , ∅, and consider u∗ the Legendre transform of u|Ω0 , that is,

u∗ (p) := sup hp, xi − u(x).

x ∈Ω0

Since p̄ ∈ ∂u(z) for all z ∈ Σ ∩ Ω0, (A.17) implies that

Σ ∩ Ω0 ⊂ ∂u∗ ( p̄);

hence, u∗ is not differentiable at p̄. In addition, (2.32) and Lemma 2.4 imply that

µu∗ (F) ≤ k1/ f k L ∞ (Ω0 ) |F | ∀ F Borel s.t. ∂u∗ (F) ⊂ Ω0.

34 2 Alexandrov solutions

Since Ω0 is convex, it follows by Lemma A.29 that ∂u∗ (R2 ) ⊂ Ω0, and as |∂Ω0 | = 0,
we deduce that for any Borel set F ⊂ R2 ,

µu∗ (F) = |∂u∗ (F)|

= |∂u∗ (F) ∩ Ω0 |
= ∂u∗ F ∩ (∂u∗ )−1 (Ω0)

≤ k1/ f k L ∞ (Ω0 ) F ∩ (∂u∗ )−1 (Ω0)

≤ k1/ f k L ∞ (Ω0 ) |F |,

which proves that

µ∗u ∈ L ∞ (R2 ).
Thus, it follows by Theorem 2.17 applied with Ω = Rn that u∗ is affine on an infinite
line containing p̄, and Lemma A.25 applied to u∗ implies that ∂u∗ (R2 ) is contained
in a hyperplane.
On the other hand, (A.20) implies that ∂u∗ (R2 ) ⊃ Ω0, impossible. 

2.6 Application 1: The Minkowski problem for curvature measures

A classical problem in convex geometry is to prescribe some geometric quantity

(surface area, Gaussian curvature, etc.) and find necessary and sufficient conditions
ensuring that such a quantity comes from a convex domain.
In this section, we briefly discuss the “prescribed Gaussian curvature” problem,
as it is strictly related to Monge–Ampère.

2.6.1 Setting of the problem. Let K ⊂ Rn be an open bounded convex domain

containing the origin, and parameterize ∂K in polar coordinates as follows:

∂K = ρ(x) x : x ∈ Sn−1, ρ : Sn−1 → R+ .



Then to any point z ∈ ∂K, we associated the normal mapping

NK (z) := y ∈ Sn−1 : K ⊂ {y : hy, w − zi ≤ 0} . (2.33)



Geometrically, the normal mapping finds the normals of all supporting hyperplanes
at z (see Figure 2.12), and we can think of NK as an analogue of the subdifferential
map.
Finally, we consider the (multivalued) Gauss map G K : Sn−1 → Sn−1 defined by

G K (x) := NK ρ(x)x ,


 2.6 Application 1: The Minkowski problem for curvature measures 35

Figure 2.12. Given a convex set K 3 0, we can parameterize its boundary in polar coordinates using
a map ρ : Sn−1 → R+ . Given a point z ∈ ∂K, the set NK (z) consists of all outer normals to K at z.
Such a set is a singleton when K has a unique tangent plane (as for the point z2 in the figure), while
it is multivalued at a “corner” point such as z1 .

and define the Gaussian curvature measure

µK (E) := Hn−1 G K (E) ∀ E ⊂ Sn−1 Borel,

where Hn−1 denotes the (n − 1)-dimensional Hausdorff measure on Sn−1 (see Ap-
pendix A.2).
Using the argument in Section 2.1, one can show that µK is a Borel measure. One
then asks the following question: Given a Borel measure ν on Sn−1 , can we find an
open bounded convex set K 3 0 such that ν = µK ?

2.6.2 Alexandrov result. In order to describe the existence and uniqueness theorem
of Alexandrov [1, 2], we first need to introduce some notation.

Definition 2.20. Let ω ⊂ Sn−1 .

• We say that ω is convex if the cone

R+ ω := {t x : t > 0, x ∈ ω}

is convex.

• We define the polar set to ω as

ω∗ := {y ∈ Sn−1 : hx, yi ≤ 0 ∀ x ∈ ω}.

We can now state the Alexandrov result (recall that all measures considered in
this book are nonnegative).
36 2 Alexandrov solutions

Theorem 2.21. Let ν be a Borel measure on Sn−1 . Then



 (a) ν(Sn−1 ) = Hn−1 (Sn−1 );


ν = µK for some K


⇔


 (b) ν(Sn−1 \ ω) > Hn−1 (ω∗ )
∀ ω ( Sn−1 compact and convex.



In addition, if K exists, then it is unique up to a dilation.
As for the existence of solutions to the Monge–Ampère equation (see the proof
of Theorem 2.13), the existence of K is first proved when ν is a finite sum of Dirac
deltas, and then one obtains the general case by approximation.
The original existence proof of Alexandrov when ν is discrete was based on a
topological argument relying on the invariance of domain theorem [1] (see also [4]).
More recently, in [95] Oliker gave a new proof based on a variational argument
inspired by optimal transportation theory.
Here, we just prove the easy implication in the theorem, and we refer to [95] for
more references and a detailed proof. Then, in the next section, we show how the
regularity theory for Monge–Ampère applies to this problem.
Proof of Theorem 2.21: the sufficiency part. Assume that ν = µK for some convex
domain K containing the origin. We prove that (a) and (b) hold.
Step 1: Proof of (a). We first notice that the map NK : ∂K → Sn−1 is surjective.
Indeed, given y ∈ Sn−1 , consider the family of hyperplanes

Hy,a := {z ∈ Rn : hy, zi = a} ∀ a ∈ R,

and notice that Hy,a ∩ K = ∅ if a  1. Let

â := inf{a > 0 : Hy,a ∩ K = ∅}

(see Figure 2.13), in other words, set â to be the largest positive number such that
Hy, â ∩ ∂K , ∅. Then y ∈ NK (z) for all z ∈ Hy, â ∩ ∂K. Since y was arbitrary, this
proves that
G K (Sn−1 ) = NK (∂K) = Sn−1,
and point (a) follows.
Step 2: Proof of (b). By the convexity of K and since 0 ∈ K, it is easy to check that

hy, zi > 0 ∀ z ∈ ∂K, y ∈ NK (z),

or equivalently,
ρ(x)hy, xi > 0 ∀ x ∈ Sn−1, y ∈ G K (x).
 2.6 Application 1: The Minkowski problem for curvature measures 37

Figure 2.13. Given a direction y ∈ Sn−1 , we can find a point z ∈ ∂K such that y ∈ NK (z).

Recalling the definition of ω∗ , this implies that

y < ω∗ ∀ y ∈ G K (ω);

thus,
G K (ω) ⊂ Sn−1 \ ω∗ .
Since ω is compact by assumption, it follows that G K (ω) is compact as well (this can
be proved following the ideas of the proof of Lemma A.22), and since Sn−1 \ ω∗ is
open, we obtain that H n−1 n−1 \ ω ) \ G K (ω) > 0. Hence,
∗


(S

ν(ω) = µK (ω) = Hn−1 G K (ω) < Hn−1 (Sn−1 \ ω∗ ).

Combining this estimate with (a), we get

ν(Sn−1 \ ω) = ν(Sn−1 ) − ν(ω) > Hn−1 (Sn−1 ) − Hn−1 (Sn−1 \ ω∗ ) = Hn−1 (ω∗ ),

as desired. 

2.6.3 C 1 regularity for convex surfaces with bounded Gaussian curvature. In

this section, we apply Theorem 2.17 to show that convex surfaces with bounded
Gaussian curvature are C 1 . A generalization of this fact to higher dimensions will
follow immediately from the regularity results in Chapter 4.
Theorem 2.22. Let K ⊂ R3 be an open bounded convex domain containing the
origin, and assume that µK = f d H2 for some f : S2 → R+ bounded. Then ∂K is
C1.
Proof. Assume that ∂K is not C 1 at some point z0 ∈ ∂K. Up to a change of
coordinates, we can write ∂K near z0 as the graph of a convex function

u : Ω ⊂ R2 → R,
38 2 Alexandrov solutions

so that z0 = (x0, u(x0 )) and u is not differentiable at x0 . The equation µK = f d H2

written in terms of u becomes3

2
µu = f 1 + |∇u| 2 , (2.34)

where ∇u exists at almost every point since u is locally Lipschitz (being convex) near
x0 . In particular, µu is locally bounded, and it follows by Theorem 2.17 that the
singularity propagates from x0 along a segment where u is affine, or equivalently ∂K
contains a segment Σ along which ∂K is not differentiable. Since ∂K can be written
as a graph in a neighborhood of Σ, applying Theorem 2.17 again, we see that the
singularity propagates indefinitely; thus, ∂K has to contain an infinite line. This is
impossible since K is bounded. 

3It is a classical fact that the Gaussian curvature of the graph of a C 2 function u : Ω ⊂ R n → R is given by
det D 2 u
.
(1 + | ∇u | 2 )(n+2)/2
Then an approximation argument based on Proposition 2.6 yields (2.34).
3

Smooth solutions

The aim of this chapter is to show the existence of smooth solutions to the Dirichlet
problem when the domain and the boundary data are smooth. In addition, we shall
discuss Pogorelov’s counterexample to the smoothness of solutions with smooth
right-hand side in dimensions greater than 2. Finally, we will prove Pogorelov’s
interior estimates for strictly convex solutions.

3.1 Existence of smooth solutions by the continuity method

As shown in Section 2.3, the uniqueness of solutions to the Dirichlet problem holds
even at the level of weak solutions. So, the main issue is existence. Existence
of smooth solutions is obtained through the so-called continuity method. This is
a classical tool to prove existence of solutions to nonlinear PDEs. For simplicity,
we first focus on the case of zero boundary data, and then we will discuss general
boundary data in Section 3.1.5.

3.1.1 The continuity method. Let ū : Ω → R be a smooth convex function

satisfying
det D2 ū = f¯
(
in Ω,
(3.1)
ū = 0 on ∂Ω,
for some f¯ > 0, and assume that we want to find a smooth solution to
(
det D2 u = f in Ω,
(3.2)
u=0 on ∂Ω,

for some other function f > 0. Define fσ := (1 − σ) f¯ + σ f , and for σ ∈ [0, 1],
consider the one-parameter family of problems
(
det D2 uσ = fσ in Ω,
(3.3)
uσ = 0 on ∂Ω.

The method of continuity consists of showing that the set of σ ∈ [0, 1] for which (3.3)
is solvable is both open and closed in [0, 1]. Since the problem is solvable for σ = 0
40 3 Smooth solutions

(the solution being ū), this will imply that the problem is solvable for all σ ∈ [0, 1],
which yields, in particular, the existence of a solution to our original problem. Let
us now discuss this construction in detail.

3.1.2 The setting. Fix α ∈ (0, 1), assume that Ω is a uniformly convex domain of
class C 4,α , and consider ū ∈ C 4,α (Ω) a uniformly convex function that vanishes on
∂Ω.1 Define f¯ := det D2 ū ∈ C 2,α (Ω), so that (ū, f¯) solves (3.1). Notice that since ū
is uniformly convex, f¯ ≥ c̄ > 0 in Ω.
Given f ∈ C 2,α (Ω) with f ≥ c0 > 0, we want to find a smooth solution to (3.2).
To this aim, we set

C := C04,α (Ω) ∩ {u : Ω → R : u convex},

we define
fσ := (1 − σ) f¯ + σ f ∈ C 2,α (Ω) ∀ σ ∈ R, (3.4)
and notice that, taking c0 smaller if necessary,

fσ ≥ c0 ∀ σ ∈ [0, 1]. (3.5)

Then we consider the nonlinear map

F : C × [0, 1] −→ C 2,α (Ω),

(u, σ) 7−→ det D2 u − fσ,

and the set

T := {σ ∈ [0, 1] : ∃ uσ ∈ C s.t. F (uσ, σ) = 0}.
Notice that 0 ∈ T (since F (ū, 0) = 0), so to prove existence of a C 4,α solution to
(3.2), it suffices to show that T is both open and closed in [0, 1].

3.1.3 T is open: The implicit function theorem. Openness follows from the
implicit function theorem in Banach spaces. More precisely, we set

X := C04,α (Ω), Σ := R, and Y := C 2,α (Ω),

and we consider the map G : X × Σ → Y defined by

G (u, σ) := det D2 u − [(1 − σ) f¯ + σ f ] ∀ (u, σ) ∈ X × Σ.

1A possible way to construct ū is first to set
ū := − dist(·, ∂Ω) + dist(·, ∂Ω)2 in a neighborhood of ∂Ω
(see [40, Section 5] for a detailed analysis of the distance function to a smooth hypersurface), and then to extend
ū to a smooth uniformly convex function inside Ω. For instance, if Ω = B1 , one may take ū(x) := 12 |x | 2 − 1 .


 3.1 Existence of smooth solutions by the continuity method 41

It is easy to check that G is of class C 1 and that the Frechèt differential of G with
respect to u is given by the linearized Monge–Ampère operator (see (1.3))
(
det(D2 u)ui j ∂i j h in Ω,
Du G (u, t)[h] = (3.6)
h=0 on ∂Ω

(recall that ui j is the inverse of ui j = ∂i j u and we are summing over repeated indices).
In particular, if σ̂ ∈ [0, 1] and û ∈ C , it follows by (3.5) and Remark 1.1 that the
coefficients ai j := det(D2 û)ûi j = fσ̂ ûi j are uniformly elliptic and of class C 2,α . So,
we can apply Theorem A.39 to deduce that the linear elliptic problem
(
fσ̂ ûi j ∂i j h = g in Ω,
u=0 on ∂Ω

has a unique solution h ∈ X for any g ∈ Y . Let σ̂ ∈ T and û ∈ C be such that

F(û, σ̂) = 0. Clearly, (û, σ̂) ∈ X × Σ and G (û, σ̂) = 0. Hence, we can apply
Theorem A.37 to conclude that the equation
G (uσ, σ) = 0
is solvable in X for all σ ∈ R close to σ̂, and kuσ − uσ̂ kC 4, α (Ω) ≤ C|σ − σ̂|.
Notice that, since û ∈ C , it follows by Remark 1.1 that û is uniformly convex;
hence, uσ is convex for |σ − σ̂|  1, which proves that the equation G (uσ, σ) = 0 is
solvable in C and not just in X. Since G = F on C × [0, 1], we have demonstrated the
local solvability of F (uσ, σ) = 0 for σ close to σ̂.

3.1.4 T is closed: Global a priori estimates. The task now is to prove closedness
of T . This is done through interior and boundary a priori estimates.
The main idea is to prove uniform bounds on the C 2 norm of a solution uσ of
(3.3) in Ω. Then, once this bound is established, we shall see that Theorem A.41
applies and gives us a universal C 4,α (Ω) bound on uσ . This uniform control on the
C 4,α norm of a solution uσ will allow us to conclude that T is closed.

3.1.4.1 Global C 2 estimates. The first step of the proof is to show global C 2
estimates for solutions to the Monge–Ampère equation when the right-hand side is
of class C 2 .
In order to simplify notation, we remove the parameter σ. Furthermore, as we
shall see, the dependence on the right-hand side is better expressed through the norm
of its logarithm.
Remark 3.1. Notice that if f ∈ C k,α (Ω) and f ≥ c0 > 0, then
 
k log f kC k, α (Ω) ≤ C = C k, k f kC k, α (Ω), c0 .
42 3 Smooth solutions

Theorem 3.2. Let Ω be a uniformly convex domain of class C 3 , f ∈ C 2 (Ω) with

f ≥ c0 > 0, and u ∈ C 2 (Ω) ∩ C 4 (Ω) solves (3.2). Then,
 
kukC 2 (Ω) ≤ C = C Ω, k log f kC 2 (Ω) .

Notice that the uniform convexity of Ω is necessary to obtain regularity up to the

boundary. Indeed, if D2 u is uniformly bounded, then u is uniformly convex on Ω
(see Remark 1.1); hence, ∂Ω = {u = 0} is uniformly convex as well. The assumption
that u is of class C 4 inside Ω is also important, since we will need to differentiate the
Monge–Ampère equation twice.
Proof. We begin by noticing that, since we do not yet have a uniform bound on the
eigenvalues of D2 u, the maximum principle is essentially the only tool at our disposal
when considering the linearized operator in (3.6).
Step 1: C 0 and C 1 estimates. Both estimates are obtained by a simple barrier con-
struction. More precisely, fix û ∈ C 2 (Ω) a uniformly convex function that coincides
with − dist(·, ∂Ω)+dist(·, ∂Ω)2 near ∂Ω, and set fˆ := det D2 û. Notice that fˆ ≥ ĉ0 > 0
on Ω (by the uniform convexity of û) and the pair (û, fˆ) solves (3.2).
Set
C0 := k f k L ∞ (Ω) and Ĉ0 := k fˆk L ∞ (Ω), (3.7)
and define
1
v + := M û and v − :=
û,
M
where M > 0 is a large constant that will be chosen later. Then, recalling that f ≥ c0
and fˆ ≥ ĉ0 , we see that

det D2 v + = M n fˆ ≥ f
(
in Ω, C0
+
provided M n ≥ ,
v =u=0 on ∂Ω, ĉ0

and
1
fˆ ≤ f
(
det D2 v − = in Ω, Ĉ0
Mn provided M n ≥ .
v− = u = 0 on ∂Ω, c0

Hence, if we choose M := max (C0 /ĉ0 )1/n, (Ĉ0 /c0 )1/n , we can apply Theorem 2.10


both to u, v + and to u, v − to deduce that

v+ ≤ u ≤ v− in Ω (3.8)

(see Figure 3.1), which implies, in particular, that

kuk L ∞ (Ω) ≤ kv + k L ∞ (Ω) = M k ûk L ∞ (Ω) .

 3.1 Existence of smooth solutions by the continuity method 43

Figure 3.1. Since v + ≤ u ≤ v − , we obtain a universal bound both on kuk L ∞ (Ω) and on k∂ν uk L ∞ (∂Ω) =
k∇uk L ∞ (Ω) .

To obtain a bound on |∇u|, we note that by the convexity of u, |∇u| attains its
maximum on the boundary, that is,

max |∇u| = max |∇u|.

Ω ∂Ω

So, it suffices to estimate |∇u| on the boundary. Since u = 0 on ∂Ω, any tangential
derivative of u is zero. Therefore, we only have to estimate the exterior normal
derivative of u. Notice that thanks to (3.8), we have

∂ν v − (x) ≤ ∂ν u(x) ≤ ∂ν v + (x) ∀ x ∈ ∂Ω

(see Figure 3.1). Moreover, ∂ν û = 1 on ∂Ω since û = − dist(·, ∂Ω) + dist(·, ∂Ω)2 near
∂Ω. (Recall that the gradient of the distance function on ∂Ω is precisely −ν, the inner
unit normal to ∂Ω.) Hence,
1
≤ ∂ν u ≤ M on ∂Ω. (3.9)
M
In particular,
k∇uk L ∞ (Ω) = k∇uk L ∞ (∂Ω) = k∂ν uk L ∞ (∂Ω) ≤ M,
which concludes the proof of the C 1 bound.
Step 2: C 2 estimates. To get C 2 bounds we first use that pure second derivatives of
u satisfy a differential inequality to reduce ourselves to a boundary estimate. Then
a direct computation gives us a bound on the second tangential derivatives, while
a barrier argument allows us to control the mixed derivatives. Finally, the second
normal derivatives are controlled using the Monge–Ampère equation.
Step 2-a: Reduction to a boundary estimate. Given a unit vector e ∈ Sn−1 , we
differentiate the equation
log det D2 u = log f (3.10)
44 3 Smooth solutions

once and twice in the e-direction. As done in Section 1.1.2, we shall use subscripts
to denote partial derivatives. However, we shall not limit this convention just to
ue = ∂e u, but we shall also write (ue )i j = ∂i j ue .2 Recall that ui j denotes the inverse
of ui j , and we use the convention of summation over repeated indices.
Differentiating (3.10) in the direction e, we get

ui j (ue )i j = (log f )e (3.11)

(see (1.3)). In order to differentiate again, using (A.5), we see that

d d
D2 u(x + εe) D2 u(x) + εD2 ue (x)

−1
−1
=
dε ε=0 dε ε=0
= −[D2 u(x)]−1 D2 ue (x)[D2 u(x)]−1,

that is,
∂e (ui j ) = −uil (ue )lk uk j .
Hence, the differentiation of (3.11) in the direction e gives

ui j (uee )i j − uil uk j (ue )i j (ue )lk = (log f )ee . (3.12)

We claim that
uil uk j (ue )i j (ue )lk ≥ 0. (3.13)
Indeed, for any point x ∈ Ω, we can fix a system of coordinates so that ui j (and also
ui j ) is diagonal, and since uii > 0 (by convexity), we get
2
uil uk j (ue )i j (ue )lk = uii u j j (ue )i j ≥ 0,


which proves (3.13). Combining (3.12) and (3.13) yields

ui j (uee )i j ≥ (log f )ee . (3.14)

Now, consider the linear elliptic operator

Lv := ui j ∂i j v = ui j vi j . (3.15)

It follows from (3.14) that

Luee ≥ −k(log f )ee k L ∞ (Ω) ≥ −k log f kC 2 (Ω),

and since Lu = ui j ui j = n ≥ 1, we deduce that

L(uee + Ku) > 0

2Although this convention may look a bit confusing at first sight, it is extremely convenient for computations.
This will become particularly evident in Section 3.3.
 3.1 Existence of smooth solutions by the continuity method 45

provided we choose K > k log f kC 2 (Ω) . In particular, it follows by Lemma A.38 that
uee + Ku cannot have an interior maximum. Therefore,

max(uee + Ku) = max(uee + Ku).

Ω ∂Ω

Recalling that uee ≥ 0 (by the convexity of u) and that u|∂Ω = 0, we obtain

max uee ≤ max(uee + Ku) + K kuk L ∞ (Ω)

Ω Ω
= max(uee + Ku) + K kuk L ∞ (Ω)
∂Ω
= max uee + K kuk L ∞ (Ω) .
∂Ω

Hence, since u is universally bounded (by Step 1), to estimate the L ∞ norm of D2 u
over Ω, it is enough to do so on ∂Ω.
To this aim, we fix a point x0 ∈ ∂Ω, and we look for uniform control on D2 u(x0 ).
Up to a translation and a rotation, we can assume that x0 = 0 and that
n−1
n κα 2 o
xα + O(|x 0 | 3 ) . (3.16)
Õ
Ω ∩ Bρ (0) = (x 0, xn ) = (x1, . . . , xn ) : xn ≥ ϕ(x 0) =
α=1
2

Here, ρ > 0 is some universal constant, and ϕ ∈ C 3 satisfies

κ 0 0
© 1 ..
ϕ(0) = 0, ∇ ϕ(0) = 0, and D2x0 x0 ϕ(0) =­ 0 (3.17)
ª
x0
­ . 0
®
®
« 0 0 κn−1 ¬
(see Figure 3.2). Notice that by the smoothness and uniform convexity of Ω,

κ− ≤ κα ≤ κ+ ∀ α = 1, . . . , n − 1, (3.18)

for some universal constants 0 < κ− ≤ κ+ .

Step 2-b: A uniform bound on the second tangential derivatives. Since u|∂Ω = 0, we
can differentiate the identity u(x 0, ϕ(x 0)) = 0 with respect to x 0 to get

∇x0 u + un ∇x0 ϕ = 0, (3.19)

and differentiating again, we find

D2x0 x0 u + ∇x0 un ⊗ ∇x0 ϕ + ∇x0 ϕ ⊗ ∇x0 un + unn ∇x0 ϕ ⊗ ∇x0 ϕ + un D2x0 x0 ϕ = 0,

where we have used the notation un = ∂xn u and unn = ∂xn xn u, and the derivatives of
u are evaluated at (x 0, ϕ(x 0)).
46 3 Smooth solutions

Figure 3.2. Given any point x0 ∈ ∂Ω, up to an isometry, we can assume that x0 = 0 and that
∂Ω ∩ Bρ (0) is given by the graph of a uniformly convex C 3 function ϕ : Rn−1 → R such that
ϕ(0) = 0 and ∇ x0 ϕ(0) = 0.

In particular, it follows from (3.17) that

κ 0 0
© 1 ..
D2x0 x0 u(0) = −un (0) ­ 0 ®. (3.20)
ª
­ . 0 ®
« 0 0 κn−1 ¬
Denoting the canonical basis in Rn by {e1, . . . , en }, we notice that the exterior outer
normal at 0 is given by −en . Thus, un (0) = −∂ν u(0). Therefore, combining (3.9),
(3.20), and (3.18), we obtain the following universal bound on the second tangential
derivatives:
κ−
Idn−1 ≤ D2x0 x0 u(0) ≤ M κ+ Idn−1 . (3.21)
M

Step 2-c: A uniform bound on the second mixed derivatives. For this, fix α ∈
{1, . . . , n − 1}, and use polar coordinates (ρα, θ α ) for points in the plane generated
by {eα, en }. Then, differentiating equation (3.10) with respect to θ α , and recalling
(3.15) and (3.11), we get

L(∂θα u) = ui j (∂θα u)i j = ∂θα (log f ).

Noticing that in Cartesian coordinates ∂θα = xα ∂n − xn ∂α , we deduce that

L(xα un − xn uα ) = xα (log f )n − xn (log f )α . (3.22)

In order to use a barrier argument, we have to combine pure and angular derivatives
of u to get a quantity which vanishes “enough” on ∂Ω.
More precisely, it follows by (3.17) that

∂xα ϕ = κα xα + O(|x 0 | 2 );
 3.1 Existence of smooth solutions by the continuity method 47

hence, recalling that uα = −∂xα ϕun (see (3.19)), we get

uα = −∂xα ϕun = −κα xα un + O(|x 0 | 2 ) on ∂Ω ∩ Bρ (0). (3.23)

This suggests that we should consider the quantity uα + κα xα un . However, this

quantity does not solve any equation, so it is not suitable for a barrier argument. On
the other hand, we know that κα xα un − κα xn uα solves an equation (see (3.22) and
notice that κα is a constant). Hence, it makes sense to consider the function

vα := uα + κα xα un − κα xn uα,

which satisfies (by (3.11) and (3.22))

Lvα = (log f )α + κα [xα (log f )n − xn (log f )α ] ≤ C,

e (3.24)

for a constant C
e that depends only on Ω and k log f k 1 . We now want to estimate
C (Ω)
vα on ∂Ω.
To this aim, we first notice that since |x| 2 = |x 0 | 2 + xn2 , as a consequence of (3.16)
and (3.18) there exist small constants c 0, δ > 0 such that
κ− 0 2 2
 
xn ≥ max |x | , xn ≥ c 0 |x| 2 for (x 0, xn ) ∈ ∂Ω, xn ≤ δ.
4
Also, since ∂Ω is bounded, we have
C0
|x| 2 ≤ C 0 ≤ xn for x = (x 0, xn ) ∈ ∂Ω, xn ≥ δ.
δ
Thus, combining these two estimates, we get

K̂ xn − |x| 2 ≥ 0 on ∂Ω, (3.25)

for some constant K̂ > 0 depending only on ∂Ω.

Analogously, since uα + κα xα un is globally bounded on Ω (recall that ∇u is
universally bounded by Step 1), the estimate uα + κα xα un = O(|x| 2 ) holds trivially
when |x| ≥ ρ; hence, (3.23) yields

uα = −κα xα un + O(|x| 2 ) on ∂Ω. (3.26)

We can now estimate vα on ∂Ω: using (3.26), (3.25), and that ∇u is universally
bounded (by Step 1), we see that
|vα | = |(uα + κα xα un ) − κα xn uα |
≤ C|x| 2 + |κα xn uα | (3.27)
2 2
≤ Ĉ(xn + |x| ) ≤ (Ĉ + K̂ A)xn − (A − Ĉ)|x| on ∂Ω, ∀ A > 0.
48 3 Smooth solutions

To conclude the barrier argument, we fix a point x ∈ Ω, and we choose a system

of coordinates so that ui j is diagonal. Since L |x| 2 = 2 i uii and f ≤ C0 (see (3.7)),
Í
using the arithmetic–geometric inequality and that
Ö  −1   −1
= det D2 u
Ö
u =
ii
uii = f −1,
i i

we see that for any a, b > 0,

 1/n
2nb 2nb
Ö
2 2
Õ
L(b|x| − axn ) = bL |x| = 2b u ≥ 2nb
ii
u ii
= 1/n
≥ 1/n .
i i
f C0

Hence, it follows from (3.24) and (3.27) that the functions

ψα± := (A − Ĉ)|x| 2 − (Ĉ + K̂ A)xn ± vα

satisfy
Lψα± > 0 in Ω and ψα± ≤ 0 on ∂Ω,
provided we choose A large enough so that

2n(A − Ĉ)
> C.
e
C01/n

Thanks to Lemma A.38, this implies that ψα± cannot have interior maxima; hence,

max ψα± = max ψα± ≤ 0.

Ω ∂Ω

Since ψα± (0) = 0, we deduce that ∂xn ψα± (0) ≤ 0. Noticing that

∂xn ψα± (0) = −(Ĉ + K̂ A) ± uαn (0),

we conclude that

|uαn (0)| ≤ Ĉ + K̂ A ∀ α = 1, . . . , n − 1,

as desired.
Step 2-d: A uniform bound on the second normal derivatives. To get a bound on
unn (0), we notice that

f (0) = det D2 u(0) = det D2x0 x0 u(0)unn (0) + Pol D2x0 x0 u(0), D2x0 xn u(0)

where Pol D2x0 x0 u(0), D2x0 xn u(0) is a polynomial of degree n depending only on


second tangential and mixed derivatives of u at 0.
 3.1 Existence of smooth solutions by the continuity method 49

Hence, since Pol D2x0 x0 u(0), D2x0 xn u(0) is universally bounded (by Steps 2-a and

2-b) and det D2x0 x0 u(0) ≥ 1/M n−1 (thanks to (3.21)), we conclude that


0 ≤ unn (0) ≤ M n−1 | f (0)| + Pol D2x0 x0 u(0), D2x0 xn u(0) ≤ C̄.

The result above demonstrates the existence of a uniform a priori C 2 bound for
solutions of the Monge–Ampère equation with a positive right-hand side of class
C 2 . Then it follows directly from Proposition A.43 and Theorem A.41 that if f is
smoother, the following higher a priori estimate holds.
Theorem 3.3. Let Ω be a uniformly convex domain, k ≥ 2, α ∈ (0, 1), and assume
that ∂Ω is of class C k+2,α . Let u ∈ C 2 (Ω) ∩ C 4 (Ω) solve (3.2), and assume that
f ∈ C k,α (Ω) and f ≥ c0 > 0. Then u ∈ C k+2,α (Ω) with
 
kukC k+2, α (Ω) ≤ C = C Ω, n, k, k log f kC k, α (Ω) .

3.1.4.2 T is closed. Let {σj } j ∈N ⊂ T be a sequence converging to σ̂. We need to

show that σ̂ ∈ T .
Since σj ∈ T , we know that there exist uσ j ∈ C 4,α (Ω) solving
(
det D2 uσ j = fσ j in Ω,
(3.28)
uσ j = 0 on ∂Ω,
where fσ j ≥ c0 satisfies
k fσ j kC 2, α (Ω) ≤ k f kC 2, α (Ω) + k f¯kC 2, α (Ω) ∀j ∈N
(see (3.4) and (3.5)). Hence, it follows from Theorems 3.2 and 3.3 (see also Re-
mark 3.1) that
kuσ j kC 4, α (Ω) ≤ C (3.29)
for some constant C independent of j. Since C 4,α (Ω) embeds compactly in C 4 (Ω) (by
the Ascoli–Arzelà theorem), we deduce that, up to a subsequence, uσ j → û ∈ C 4 (Ω)
where û is convex. Also, it follows by (3.4) that
k fσ j − fσ̂ kC 2, α (Ω) ≤ k f¯kC 2, α (Ω) + k f kC 2, α (Ω) |σj − σ̂| → 0.

Thus, letting j → ∞ in (3.28), we get

(
det D2 û = fσ̂ in Ω,
uσ̂ = 0 on ∂Ω.
Also, letting j → ∞ in (3.29), we deduce that k ûkC 4, α (Ω) ≤ C. This proves that û ∈ C
and F (û, σ̂) = 0; hence, σ̂ ∈ T , as desired.
50 3 Smooth solutions

3.1.5 Existence of smooth solutions and generalizations. To summarize, the con-

tinuity argument used in the previous sections proves the following:

Theorem 3.4. Let Ω be a uniformly convex domain, k ≥ 2, α ∈ (0, 1), and assume
that ∂Ω is of class C k+2,α . Let f ∈ C k,α (Ω) with f ≥ c0 > 0. Then there exists a
unique solution u ∈ C k+2,α (Ω) to the Dirichlet problem
(
det D2 u = f in Ω,
u=0 on ∂Ω.

In the previous theorem, we considered the case when u vanishes on ∂Ω. We now
briefly discuss how the argument above can be extended to general smooth boundary
data. Since this generalization is not needed in the sequel, we shall just explain the
main argument leaving some details to the interested reader.
Assume we want to solve
(
det D2 u = f in Ω,
u=g on ∂Ω,

where g : ∂Ω → R is a smooth function. In this case, the main difference arises in

proving closedness. In particular, one needs to modify the barrier argument in the
proof of Theorem 3.2 to establish global C 2 bounds up to the boundary.
Assume that g ∈ C 4 (∂Ω), extend it to a function ḡ : Ω → R of class C 4 up to the
boundary (see, for instance, [58, Lemma 6.38] for a possible construction of such an
extension), and consider the barrier

ũ := ḡ + M ū,

where ū ∈ C 2 (Ω) is a uniformly convex function which coincides with − dist(·, ∂Ω) +
dist(·, ∂Ω)2 near the boundary of Ω, and M > 1 is a large constant chosen in such a
way that ũ is uniformly convex in Ω and det D2 ũ > det D2 u.
We then notice the following:

• For the L ∞ bound on u, we simply use that u ≥ ũ (by comparison) and

u ≤ kgk L ∞ (∂Ω) (by convexity).

• To control ∇u, by convexity it suffices to control it on ∂Ω. Since u|∂Ω = g, the

tangential derivatives of u are controlled by k∇gk L ∞ (∂Ω) , and we only need to
control the normal derivative uν .
Then, again using u ≥ ũ, we get uν ≤ ũν ≤ C on ∂Ω, and since u is bounded and
convex inside Ω, one easily sees that uν ≥ −C. Therefore, k∂ν uk L ∞ (∂Ω) ≤ C.
 3.1 Existence of smooth solutions by the continuity method 51

• For the second derivatives, as in Step 2-a in the proof of Theorem 3.2, using
the maximum principle we reduce the problem to estimating D2 u on ∂Ω.
With the same notation as in the proof of Theorem 3.2, we assume that 0 ∈ ∂Ω
and that ∂Ω is as in Figure 3.2. Notice that since Ω is of class C 4,α , so is ϕ.
Now, if we use that u − ḡ ≡ 0 on ∂Ω we easily get that (see Step 2-b)
|D2x0 x0 u(0)| ≤ C,
while the argument in Step 2-c gives
|D2xα xn u(0)| ≤ C ∀ α = 1, . . . , n − 1.

Finally, as in Step 2-d, we would like to use the equation to control D2xn xn u(0),
but in order to do this, we need to know that the matrix Dx0 x0 u(0) is positive
definite with a universal bound. We now show this, following the argument in
[23, Section 2].
Fix a tangential direction τ = (τ1, . . . , τn−1, 0) ∈ Sn−1 and, up to a rotation of
coordinates, assume that τ = e1 . Notice that due to this additional rotation, we
cannot assume that D2x0 x0 ϕ(0) is diagonal as we did in (3.17) (τ = e1 may not
be an eigenvector of D2x0 x0 ϕ(0)); instead, we have (recall that ϕ ∈ C 4,α )
n−1 n−1
1 Õ
σαβγ xα xβ xγ + O(|x 0 | 4 ),
Õ
ϕ(x 0) = καβ xα xβ + (3.30)
2 α,β=1 α,β,γ=1

where the matrix (καβ )α,β=1

n−1 is positive definite (see (3.18)) and the coefficients

σαβγ ∈ R are universally bounded.

Up to subtracting an affine function from both u and g, we can also assume
that g(0) = ∇τ g(0) = 0. In this way, since g, ϕ ∈ C 4 , a Taylor expansion yields
u x 0, ϕ(x 0) = g x 0, ϕ(x 0)

n−1 n−1
1 Õ
Tαβγ xα xβ xγ + O(|x 0 | 4 )
Õ
= Sαβ xα xβ +
2 α,β=1 α,β,γ=1

for some universally bounded coefficients Sαβ, Tαβγ ∈ R. Hence, choosing

µ := S11 /κ11, since xn = ϕ(x 0) on ∂Ω, we deduce from (3.30) that the function
û := u − µxn satisfies

1 Õ
û x 0, ϕ(x 0) = [Sαβ − µκαβ ]xα xβ
2
(α,β),(1,1)

[Tαβγ − µσαβγ ]xα xβ xγ + O(|x 0 | 4 ).

Õ
+ (3.31)
α,β,γ
52 3 Smooth solutions

Also, as xn = ϕ(x 0) on ∂Ω, (3.30) allows us to write

2  1 Õ 
x12 = xn − καβ xα xβ + O(|x 0 | 3 ), (3.32)
κ11 2
(α,β),(1,1)

where κ− ≤ κ11 ≤ κ+ (see (3.18)). Hence, by the inequalities

x 2 + (x1 xβ )2 ≤ xα2 + |x 0 | 4 if α , 1,
(
|x1 xα xβ | ≤ α2
xβ + (x1 xα )2 ≤ xβ2 + |x 0 | 4 if β , 1,

we get

[T111 − µσ111 ]x13

2[T111 − µσ111 ]  1 
καβ x1 xα xβ + O(|x 0 | 4 )
Õ
= x1 x n −
κ11 2
(α,β),(1,1)
n−1
2[T111 − µσ111 ]
Õ 
≤ x1 x n + O xγ2 0 4
+ |x | .
κ11 γ=2

Substituting this expression into (3.31), we obtain the bound

n n−1
Õ 
xγ2 + |x 0 | 4

Õ
û x 0, ϕ(x 0) ≤ A1j x1 x j + Ĉ (3.33)
j=2 γ=2

for some universally bounded coefficients A1j ∈ R and Ĉ ∈ R.

Let ε, δ, B > 0 (to be chosen), and consider the barrier function
n
1 Õ 2
2
h(x) := −εxn + δ|x| + A1j x1 + Bx j .
2B j=2

Notice that thanks to (3.33), if we choose 0 < ε  δ so that δ|x| 2 ≥ εxn on

∂Ω,3 then
n n n
1 Õ
2 Õ
x 2j ,
Õ
h(x) ≥ A1j x1 + Bx j ≥ A1j x1 x j + B
2B j=2 j=2 j=2

and by choosing B sufficiently large, (3.33) and (3.25) imply that û|∂Ω ≤ h|∂Ω .
We also observe that the Hessian of h is universally bounded. In addition, if
2 h = 2δ|w| 2 (since the
we choose the vector w := (−B, A12, . . . , A1n ), then Dww
3Compare with (3.25), where the bound δ |x | 2 ≤ εx n is proved in detail in the opposite regime 0 < δ  ε.
 3.2 Pogorelov’s counterexample to interior regularity 53

last term in the definition of h is constant along w). This implies that the lowest
eigenvalue of D2 h is bounded from above by 2δ. Hence, since all the other
eigenvalues are bounded, det D2 h ≤ Cδ for some universal constant C > 0. In
particular, by choosing δ sufficiently small, we can ensure that
det D2 h < det D2 û inside Ω.
Then Theorem 2.10 implies that h ≥ û in Ω; thus, since h(0) = û(0),
∂xn û(0) ≤ ∂xn h(0) = −ε.
Therefore, differentiating (3.31) twice with respect to x1 , we get (compare with
(3.20) and recall (3.18))
D2x1 x1 u(0) = D2x1 x1 û(0) = −∂xn û(0)κ11 ≥ εκ−, (3.34)

which proves that D2x1 x1 u(0) is bounded away from zero, as desired.
Once uniform global C 2 estimates are established, one concludes as in the case
g = 0 by applying a suitable variant of the Evans–Krylov theorem with general
smooth boundary data (see, for instance, [20, Chapter 9] and [58, Chapter 17.8]).
This yields the following:
Theorem 3.5. Let Ω be a uniformly convex domain, k ≥ 2, α ∈ (0, 1), and assume
that ∂Ω is of class C k+2,α . Let f ∈ C k,α (Ω) with f ≥ c0 > 0. Then for any
g ∈ C k+2,α (∂Ω), there exists a unique solution u ∈ C k+2,α (Ω) to the Dirichlet
problem (
det D2 u = f in Ω,
u=g on ∂Ω.
Remark 3.6. Theorem 3.5 was proved independently by Ivochkina [65], Krylov [76],
and Caffarelli–Nirenberg–Spruck [23]. Then, the proof was simplified by Trudinger,
who introduced a trick allowing him to extend these global estimates to general
Hessian and curvature equations [110]. As shown later by Wang in [125], the above
assumptions on the data can be relaxed. More precisely, the existence of a solution
u ∈ C 2,α (Ω) is ensured provided Ω is uniformly convex and both ∂Ω and g are of
class C 3 . Also, this result is optimal: if either ∂Ω or g are only C 2,1 , then the solution
may fail to be C 2 up to the boundary.

3.2 Pogorelov’s counterexample to interior regularity

In the previous section, we proved the existence of smooth solutions provided the
boundary data are nice. To show that these assumptions on the boundary data are
54 3 Smooth solutions

crucial, we now describe Pogorelov’s counterexample to interior regularity. We

construct an Alexandrov solution u to the Monge–Ampère equation in Bρ (0) with
f ≥ 1 and f analytic, but u < C 2 .
As we shall see, this counterexample can exist only in dimension n ≥ 3. Indeed,
two-dimensional solutions with positive right-hand side are strictly convex (by Theo-
rem 2.19), and strictly convex solutions with smooth right-hand side are smooth (see
Theorem 3.10 below).
Let us write a point x ∈ Rn as (x 0, xn ) with x 0 ∈ Rn−1 . We look for solutions with
some special homogeneity. Given a function u, define

uλ (x 0, xn ) := λ−γ u(λx 0, xn ).

We want to find an exponent γ ∈ R that preserves the Monge–Ampère equation.

Recalling the notation un = ∂xn u, we have
 2 2
λ Dx0 x0 u λ∇x0 un

2
D uλ = λ −γ
,
λ∇x0 un unn

and a simple computation yields

det D2 uλ |(x0,xn ) = λ−γn λ2(n−1) det D2 u |(λx0, xn ) .

Hence, if we set
2
γ := 2 − , (3.35)
n
we get
det D2 uλ |(x0,xn ) = det D2 u |(λx0, xn ) . (3.36)

Motivated by this computation, we look for a convex function u so that uλ = u for all
λ > 0. Equivalently, we want u to be γ-homogeneous in the x 0-variable with γ as in
(3.35). So, we consider u of the form

u(x 0, xn ) := |x 0 | 2−2/n h(xn ) = |x 0 |γ h(xn )

for some function h : R → R to be chosen.

We now compute det D2 u. Notice that thanks to (3.36) and because u is invariant
under rotations in the x 0-variable, we only need to compute det D2 u at the point
(x 0, xn ) = (1, 0, . . . , 0, xn ). Since
 
D2x0 x0 u = γ|x 0 |γ−2 Idn−1 +γ(γ − 2)|x 0 |γ−4 x 0 ⊗ x 0 h(xn ),
∇x0 un = γ|x 0 |γ−2 h 0(xn )x 0,
 3.2 Pogorelov’s counterexample to interior regularity 55

and
unn = |x 0 |γ h 00(xn ),
we deduce that for x = (1, 0, . . . , 0, xn ),
γ(γ − 1)h 0 ··· 0 γh 0
0 γh 0 0 0
© ª
­ ®
2
D u = ­­
­ .. .. .. ®.
®
. 0 . 0 . ®
0 ··· 0 γh 0 ®
­ ®
­
« γh 0 0 ··· 0 h 00 ¬
Thus,

2
det D2 u = γ(γ − 1)γ n−2 h n−1 h 00 − γ n−2 h n−2 γ 2 h 0
h
2i
= γ n−1 h n−2 (γ − 1)hh 00 − γ h 0 .

In particular, choosing, for instance, h(xn ) := 1 + xn2 and recalling (3.35), one can
easily check that u is convex in a neighborhood of 0 and
2  n−1 2 4
      
det D2 u = g := 2 − 1 + xn2 − 6 xn2 .

n−2
2 1− +
n n n
Hence, there exists ρ > 0 such that
(
det D2 u = g in Bρ (0),
(3.37)
g > 0, g analytic in Bρ (0),

but u < C 2 (Bρ (0)). More precisely, u ∈ C 1,α (Bρ (0)) only for α ≤ 1 − 2/n.
Notice that, at the moment, we have computed the determinant of D2 u at any
point x 0 , 0, but we have not yet checked that u is an Alexandrov solution. To show
that u solves det D2 u = g in the Alexandrov sense, we notice that u ∈ C 1 (Rn ) ∩
C 2 (Rn \ {x 0 = 0}) and
∇u({x 0 = 0}) ⊂ {x 0 = 0}.
Therefore,
|∇u({x 0 = 0})| = 0,
so for any Borel set E ⊂ Bρ (0), we get
∫ ∫ ∫
|∇u(E)| = ∇u E \ {x 0 = 0} = det D2 u dx = g dx =


g dx,
E\{x 0 =0} E\{x 0 =0} E

which shows that u is an Alexandrov solution of (3.37).

Notice, however, that u is not strictly convex, as u = 0 on the line {x 0 = 0}.
56 3 Smooth solutions

Remark 3.7. If, instead of 1 + xn2 , one chooses h to be the solution of the ODE

2
γ n−1 h n−2 (γ − 1)hh 00 − γ h 0 = 1,
 




h(0) = 1,
 h 0(0) = 0



(which exists in a neighborhood of 0 by the Cauchy–Lipschitz theorem), one can even
construct a function u ∈ C 1,1−2/n (Bρ (0)) \ C 2 (Bρ (0)) that solves det D2 u = 1 in the
Alexandrov sense.
Remark 3.8. The example above shows that we can construct a non-strictly convex
solution of class C 1,1−2/n . As we shall see in Corollary 4.11, this is the best possible
regularity for a non-strictly convex solution of Monge–Ampère.

3.3 Pogorelov’s interior estimates and regularity of weak solutions

In the previous sections, we have shown the existence and the uniqueness of weak
solutions in general convex domains, and the existence of smooth solutions in smooth
uniformly convex domains. To obtain smoothness of weak solutions inside general
convex domains, we need the following interior estimate due to Pogorelov [97]. In
order to emphasize the dependence of our estimates on the geometry of the domain,
we shall prove that the bounds depend only on the radii of the smallest and largest
balls that, respectively, contain and are contained in Ω.
Theorem 3.9. Let Ω ⊂ Rn be an open bounded convex set satisfying

Br ( x̄) ⊂ Ω ⊂ BR ( x̄) (3.38)

for some x̄ ∈ Rn and 0 < r ≤ R, and let u ∈ C 4 (Ω) solve

(
det D2 u = f in Ω,
u=0 on ∂Ω,

with f ∈ C 2 (Ω) and f ≥ c0 > 0. Then,

 
dist(x, ∂Ω)3n+2 kD2 u(x)k ≤ C = C n, r, R, k log f kC 2 (Ω) ∀ x ∈ Ω. (3.39)
 3.3 Pogorelov’s interior estimates and regularity of weak solutions 57

Proof. For any h > 0 small, set

Ωh := {z ∈ Ω : dist(z, ∂Ω) ≥ h} ⊂⊂ Ω

and
Uh := {z ∈ Ω : u(z) ≤ −h} ⊂⊂ Ω.
We split the proof into several steps. In Step 1, we control the size of u and ∇u inside
Uh . Then in Step 2, we estimate the size of D2 u in Uh . In Step 3, we relate Ωh and
Uh . Finally, in Step 4, we combine these estimates. In the proof we will use C to
denote a generic constant that depends only on n, r, R, and k log f kC 2 (Ω) .
Step 1: |u| . 1 and |∇u| . h−n in Uh . To control u, we simply apply Theorem 2.8
and (3.38) to get
∫ 
|u(x)| ≤ Cn diam(Ω)n f dx ≤ C ∀ x ∈ Ω. (3.40)
Ω

To estimate the gradient, we notice that for any x ∈ Uh , Theorem 2.8 implies that

h ≤ |u(x)| ≤ C dist(x, ∂Ω)1/n ∀ x ∈ Uh ;

hence,
C dist(Uh, ∂Ω) ≥ h n,
and it follows by (3.40) and Corollary A.23 that
C
k∇uk L ∞ (Ωh ) ≤ .
hn

Step 2: A universal bound on D2 u inside Uh . Since second derivatives of u are

subsolutions to a linear elliptic equation (recall the argument used in Step 2-a in
the proof of Theorem 3.2), the basic idea is to look for a maximum point of kD2 uk
inside Uh and differentiate the Monge–Ampère equation twice at such a point to find
a universal bound. However, a priori this maximum point may be attained at the
boundary of Uh .
To avoid this issue, one may consider |u + h| a kD2 uk, where a > 0 has to be
chosen. This function has the advantage of vanishing on ∂ Uh , so its maximum is
attained in the interior (recall that, by assumption, u ∈ C 4 (Ω), so the maximum is
finite). However, if one tries to run the above strategy with such a function, one gets
into trouble as one cannot control some of the terms that arise from differentiating.
2
For this reason, Pogorelov chose to multiply |u + h| a kD2 uk by eb | ∇u | /2 for a
suitable b > 0, and consider the maximum of this new function. With this auxiliary
58 3 Smooth solutions

function, the terms coming from differentiating the additional exponential term are
positive enough to allow one to close the argument.
To simplify notation, we set v := u + h, and noticing that v ≤ 0 in Uh and that the
derivatives of u and v are equal, we define

a 2
w(x, ξ) := −v(x) vξ ξ (x)eb | ∇v(x)| /2, x ∈ Uh, ξ ∈ Sn−1,
where a, b > 0 are constants that we will choose later (recall that subscripts denote
partial derivatives, so vξ ξ = ∂ξ ξ v).
Let us consider the maximum of w over all x ∈ U h and ξ ∈ Sn−1 , and call this
maximum point x0, ξ0 . Notice that x0 ∈ Uh since v = 0 on ∂ Uh . Also, observe that ξ0
is the eigenvector corresponding to the maximal eigenvalue of D2 v(x0 ). In particular,
up to a rotation in the system of coordinates, we can assume that ξ0 = e1 and that the
matrix vi j is diagonal at x0 . From this point on, we consider w as a function of x
only, with ξ = e1 fixed.
Then we compute
vi v11i
(log w)i = a + + bvk vki
v v11
and
vi j vi v j v11i j v11i v11j
(log w)i j = a −a 2 + − + bvk vki j + bvki vk j .
v v v11 (v11 )2
As x0 is a maximum point for w (and, therefore, for log w) and the matrix v i j = (vi j )−1
is positive definite, we have
vi v11i
0 = (log w)i = a + + bvk vki (3.41)
v v11
and
0 ≥ v i j (log w)i j
v i j vi j v i j vi v j v i j v11i j v i j v11i v11j
=a −a + − + bvk v i j vki j + bv i j vki vk j
v v2 v11 (v11 )2
n v i j vi v j v i j (v11 )i j v i j v11i v11j
=a −a + − + bvk v i j (vk )i j + bv i j vki vk j ,
v v2 v11 (v11 )2
where all of the functions are evaluated at x0 , and we have also used that v i j vi j = n.
Since all derivatives of u and v coincide, we can apply (3.11) and (3.12) to see that
n v i j vi v j (log f )11 v il v k j v1i j v1kl
0 ≥ a −a + +
v v2 v11 v11
i j
v v11i v11j
− + bvk (log f )k + bv i j vki vk j . (3.42)
(v11 )2
 3.3 Pogorelov’s interior estimates and regularity of weak solutions 59

Now, using (3.41) and recalling that vi j and v i j are diagonal at x0 , we obtain

v il v k j v1i j v1kl v i j v11i v11j v i j vi v j

− − a
v11 (v11 )2 v2
v il v k j v1i j v1kl v i j v11i v11j 1 i j v11i
  
v11j
= − − v + bvk vki + bvk vk j
v11 (v11 )2 a v11 v11
2
v ii v kk v1ik v1ik v ii v11i v11i 1 ii v11i

= − − v + bvi vii .
v11 (v11 )2 a v11

In particular, choosing
a := 2 (3.43)
and because, for each i,
2  2  2
1 v11i

v11i
+ bvi vii ≤ + bvi vii ,
2 v11 v11

we get (in the following computation, we write out the indices over which we are
summing)

v il v k j v1i j v1kl v i j v11i v11j v i j vi v j

− − 2
v11 (v11 )2 v2
Õ v ii v kk v1ik v1ik Õ v ii v11i v11i Õ 1  v11i 2
= − − v ii
+ bv v
i ii
ik
v11 i
(v11 )2 i
2 v11
Õ v ii v kk v1ik v1ik Õ v ii v11i v11i
≥ −
ik
v11 i
(v11 )2
Õ h  v11i  2  2i 1  v 2
111
− v ii + bvi vii − v 11 + bv1 v11
i,1
v11 2 v11
Õ v ii v kk v1ik v1ik Õ  v11i  2 Õ  v11i  2
≥ − v ii − v ii
i=1 or k=1
v 11 i
v11 i,1
v11
2 1  v 2
111
Õ 
− v ii bvi vii − v 11 + bv1 v11
i,1
2 v 11
2 1  v 2
111
Õ 
=− v ii bvi vii − v 11 + bv1 v11
i,1
2 v11
(v1 )2
= −b2 vi2 vii − 2
Õ
.
i,1
v 2 v11
60 3 Smooth solutions

In the last equality, we used that v ii = vii−1 (recall, once again, that the matrix vi j is
diagonal at x0 ) and (3.41). Plugging the above equation into (3.42), again taking into
account that vi j is diagonal, and choosing

b := k∇vk L−2∞ (Uh ) = k∇uk L−2∞ (Uh ) (3.44)

so that b(vi )2 ≤ 1 for all i, we deduce that

2n (log f )11 (v1 )2

vii − b2 vi2 vii
Õ Õ
0≥ + + bvk (log f )k − 2 2 +b
v v11 v v11 i i,1
2n C |∇v| 2
≥ − − C|∇v| − 2 2 + bv11 .
v v11 v v11
2 2
Multiplying by v 4 v11 eb | ∇v | , noticing that 1 ≤ eb | ∇v | ≤ e (as a consequence of
(3.44)), and recalling (3.40), (3.43), and the definition of w, we find that

C  1
0 ≥ bw 2 − C 1 + |∇v| 2 − C|∇v|w ≥ bw 2 − √ w − C 1 + .


b b

Thus,
 1 
w ≤ C 1 + 3/2 .
b
From (3.44) and as v 2 vξ ξ ≤ w, we get
 
max v 2 kD2 vk ≤ C 1 + k∇uk L3 ∞ (Uh ) .
Uh

Step 3: Relating Ωh and Uh . In the previous steps, we obtained a priori C 2 bounds

for u inside Uh . We now want to relate Ωh and Uh .
Consider the function
 |x − x̄| 2 − r 2 
v(x) := c01/n ,
2
and notice that

v|∂Br (x̄) = 0 ≥ u|∂Br (x̄) and det D2 v = c0 ≤ det D2 u.

Hence, thanks to Theorem 2.10,

v≥u in Br ( x̄).
 3.3 Pogorelov’s interior estimates and regularity of weak solutions 61

Figure 3.3. Since Ω ⊂ BR ( x̄), given a point x 0 ∈ ∂Ω, we have that d := |x 0 − x̄| ≤ R and that Ah
contains all the points on the segment [x 0, x̄] that are at a distance larger than cd1 h from x 0 . As this
holds for any point x 0 ∈ ∂Ω, we see that Ah ⊃ z : dist(z, ∂Ω) ≥ cR1 h .


This implies that

r2
u( x̄) ≤ −c1 := −c01/n
.
2
Now, consider the “conical” function Ĉx̄ that is 0 on ∂Ω and takes the value −c1
at its vertex x̄. Since u ≤ Ĉx̄ (by the convexity of u), we deduce that

Uh ⊃ Ah := Ĉx̄ ≤ −h ,


and by a simple geometric argument (see Figure 3.3), we get

 
R
Ah ⊃ z : dist(z, ∂Ω) ≥ h ,
c1
or equivalently,
Uh ⊃ ΩRh/c1 . (3.45)

Step 4: Conclusion. Combining Steps 1 and 2, we deduce that for h > 0 small,
  C
(u + h)2 kD2 uk ≤ C 1 + k∇uk L3 ∞ (Ωh ) ≤ 3n in Uh .
h
In particular, since |u + h| ≥ h inside U2h , we get
C
kD2 uk ≤ in U2h .
h3n+2
Moreover, applying (3.45) to U2h , we obtain
C
kD2 uk ≤ in Ω2Rh/c1 ,
h3n+2
for any h ∈ (0, h0 ], with h0 > 0 a small universal constant.
62 3 Smooth solutions

Now, to conclude the proof of (3.39), given a point x ∈ Ω, we distinguish two

cases: If x ∈ Ω2Rh0 /c1 , then we use that dist(x, ∂Ω) ≤ diam(Ω) ≤ 2R to deduce that

C C(2R)3n+2
kD2 u(x)k ≤ ≤ .
h03n+2 h03n+2 dist(x, ∂Ω)3n+2

Otherwise, dist(x, ∂Ω) < 2Rh0 /c1 , and we simply notice that x ∈ Ω2Rh/c1 for
h := c1 dist(x, ∂Ω)/2R < h0 to get

C(2R)3n+2
kD2 u(x)k ≤ .
c13n+2 dist(x, ∂Ω)3n+2

We now show the smoothness of strictly convex solutions by combining the above
result with the stability of weak solutions and the existence of smooth solutions in
smooth domains.
Theorem 3.10. Let Ω ⊂ Rn be an open set, and let u : Ω → R be a strictly convex
function satisfying
µu = f dx in Ω
3,α
for some f ∈ C 2 (Ω) with f ≥ c0 > 0. Then u ∈ Cloc (Ω) for all α < 1.
k+2,β
In addition, if f ∈ C k,β (Ω) for some k ≥ 2 and β ∈ (0, 1), then u ∈ Cloc (Ω).
Proof. Fix x0 ∈ Ω and p ∈ ∂u(x0 ), and for t > 0, consider the open convex set

S(x0, p, t) := z ∈ Ω : u(z) < `x0, p,t (z) , (3.46)



where `x0, p,t (z) := u(x0 ) + hp, z − x0 i + t. Notice that by the strict convexity of u,
Ù
S(x0, p, t) = {x0 };
t>0

hence, we can choose t > 0 small enough so that S(x0, p, t) ⊂⊂ Ω.

We now consider Sε a sequence of C ∞ uniformly convex sets such that

dist ∂Sε, ∂S(x0, p, t) → 0 as ε → 0

and a sequence fε ∈ C ∞ (Ω) such that k fε − f kC 2 (Ω) → 0. Applying Theorem 3.4,

loc
we find a function vε ∈ C 4 (Sε ) solving
(
det D2 vε = fε in Sε,
vε = 0 on ∂Sε .
 3.3 Pogorelov’s interior estimates and regularity of weak solutions 63

Let S 0 ⊂⊂ S(x0, p, t) be a smooth open convex set containing x0 . Since Sε →

S(x0, p, t) and (u − `x0, p,t )|∂S(x0, p,t) = 0, it follows by Corollary 2.12 that

vε → u − `x0, p,t locally uniformly in S(x0, p, t) as ε → 0.

Hence, as
S 0 ⊂⊂ Sε for ε  1,
we can use Theorem 3.9 to conclude that

kvε kC 2 (S0 ) ≤ C for ε  1,

for some constant C independent of ε. Notice that the Monge–Ampère equation

satisfies the assumptions needed to apply the Evans–Krylov theorem (see Proposi-
tion A.43). Therefore, because k fε kC 1, α (S0 ) ≤ k fε kC 2 (S0 ) ≤ C for all ε > 0 and
α < 1, Theorem A.42 applied with Ω0 = S 0 implies that given an open set S 00 ⊂⊂ S 0
containing x0 ,
kvε kC 3, α (S00 ) ≤ C for ε  1,
where C is independent of ε. Letting ε → 0, we get

ku − `x0, p,t kC 3, α (S00 ) ≤ C,

which proves that u is of class C 3,α in a neighborhood of x0 . Then the arbitrariness

3,α
of x0 shows that u ∈ Cloc (Ω), as desired.
Notice that if we additionally knew that f ∈ C k,β (Ω), then we could have chosen
fε with k fε kC k, β (S0 ) ≤ C, and Theorem A.42 would provide a uniform bound on the
C k+2,β norm of vε in S 00, implying that u ∈ C k+2,β (S 00). 
4

Interior regularity of weak solutions

In this chapter, we focus on Alexandrov solutions of the Monge–Ampère equation

whose right-hand side is bounded away from zero and infinity, that is,

λ dx ≤ µu ≤ Λ dx in Ω

for some constants 0 < λ ≤ Λ < ∞. (We shall often emphasise only that λ is positive,
keeping implicit that Λ is finite.) We will investigate the geometry of solutions and
their strict convexity, and we will prove interior C 1,α , W 2, p , and C 2,α regularity.
Whenever we say that a constant is universal, we shall mean that it depends only
on n, λ, and Λ.

4.1 Sections and normalized solutions

“Sections” of solutions play an important role in the regularity theory for Monge–
Ampère. Our goal here is to introduce some basic definitions, describe the affine
invariance of the Monge–Ampère equation, motivate the concept of normalized
solutions, and describe their main properties.

4.1.1 Sections of convex functions and normalization. Let Ω ⊂ Rn be an open

set and u : Ω → R a convex function. Given x ∈ Ω, p ∈ ∂u(x), and t > 0, we define
the section centered at x with slope p and height t by

S(x, p, t) := z ∈ Ω : u(z) < `x, p,t (z) ,



where `x, p,t : Rn → R is the affine function defined by

`x, p,t (z) := u(x) + hp, z − xi + t ∀ z ∈ Rn

(see Figure 4.1). Moreover, given τ > 0, we denote by τS(x, p, t) the dilation of
S(x, p, t) by a factor τ with respect to x, that is,
n y−x o
τS(x, p, t) := y ∈ Rn : x + ∈ S(x, p, t) . (4.1)
τ
66 4 Interior regularity of weak solutions

Figure 4.1. Given a supporting hyperplane `x, p for u at x, we lift it by a factor t > 0 and we use this
hyperplane to cut the epigraph of u. The section S(x, p, t) is then obtained by projecting such a set
onto Ω.

Whenever S(x, p, t) ⊂⊂ Ω, it follows by the convexity of u that S := S(x, p, t) is

a bounded open convex set, and Lemma A.13 implies that there exists an ellipsoid
E = E A, x̂ (recall Definition A.11) such that

E ⊂ S ⊂ nE .

Consider the affine map T z := Az + x̂, and denote its inverse by L : Rn → Rn , that
is,
Lz := A−1 z − A−1 x̂.
Then since by definition E = T B1 (0) , we have L(E ) = B1 (0); therefore,

B1 (0) ⊂ L(S) ⊂ Bn (0).

This suggests the following definition:

Definition 4.1. An open convex set Z ⊂ Rn is normalized if

B1 (0) ⊂ Z ⊂ Bn (0).

With this definition, Lemma A.13 can be restated to say that, for any open bounded
convex set S, there exists a unique invertible affine transformation L that normalizes
it. In the sequel, we shall refer to L as the affine map that normalizes S.
 4.1 Sections and normalized solutions 67

Notice that given an affine transformation L, its linear part corresponds to its
gradient ∇L. Specifically,
Lz = (∇L)z + x̄ for some x̄ ∈ Rn .
In the sequel, to simplify the notation, we shall write
det L := det(∇L) and kL k := sup |(∇L)v|.
|v |=1

Also, we recall that

det L = det L ∗ and kL k = kL ∗ k,
where A∗ denotes the transpose of a matrix A.

4.1.2 Affine invariance and normalized solutions. Let u ∈ C 2 (Ω) satisfy

det D2 u = f in Ω,
consider S(x, p, t) a section of u, and assume that S := S(x, p, t) ⊂⊂ Ω. Notice that
since u ∈ C 2 , it follows by Lemma A.20 that ∂u(x) = {∇u(x)}; hence, p = ∇u(x).
Let L be the affine transformation that normalizes S, denote the normalized convex
set associated to S by S ∗ := L(S), and consider the function
v(z) := (det L)2/n u − `x, ∇u(x),t (L −1 z)
 
(4.2)
= (det L)2/n u(L −1 z) − u(x) − h∇u(x), L −1 z − xi − t .
 

Then it is immediate to check that

D2 v(z) = (det L)2/n [∇L ∗ ]−1 D2 u(L −1 z)[∇L]−1, (4.3)
so, in particular,
 
det D2 v(z) = (det L)2/n det [∇L ∗ ]−1 D2 u(L −1 z)[∇L]−1
 n

= (det L)2 (det L −1 )2 det D2 u(L −1 z) = f (L −1 z).

In addition, since u = `x, ∇u(x),t on ∂S (see Figure 4.1), we deduce that v = 0 on ∂S ∗ ;
thus, (
det D2 v = f ◦ L −1 in S ∗,
v=0 on ∂S ∗ .
In other words, v solves the Monge–Ampère equation with zero Dirichlet boundary
data inside a normalized convex set. We call v a normalized solution.
Using Proposition 2.6 and a simple approximation argument, these considerations
extend to Alexandrov solutions. We summarize this in the following proposition:
68 4 Interior regularity of weak solutions

Proposition 4.2. Let Ω ⊂ Rn be an open set, u : Ω → R a convex function, and

assume that
1
µu = f dx in Ω, for some f ∈ Lloc (Ω).
Assume that S := S(x, p, t) ⊂⊂ Ω, let L be the affine transformation that normalizes
S, set S ∗ := L(S), and define v as in (4.2). Then v solves
(
µv = f ◦ L −1 dx in S ∗,
(4.4)
v=0 on ∂S ∗ .

Remark 4.3. It is important to notice that if λ ≤ f ≤ Λ for some 0 ≤ λ ≤ Λ, then

λ ≤ f ◦ L −1 ≤ Λ. Hence, the upper and lower bounds on the right-hand side of
Monge–Ampère are invariant under the affine transformation described above. This
corresponds to what is called the affine invariance of Monge–Ampère.

4.1.3 Alexandrov estimates for normalized solutions. Theorem 2.8 provides con-
trol on the side of solutions depending only on dimension and the geometry of Ω.
We now show that if the right-hand side is bounded away from zero and infinity
and the domain is normalized, then we can find universal estimates on the size of a
solution. Actually, since it will be useful to apply these estimates in domains that
are not normalized, but still comparable to balls, we state and prove these results in
a slightly more general setting.
Proposition 4.4. Let Z be an open convex set satisfying
Br ( x̄) ⊂ Z ⊂ BR ( x̄)
for some 0 < r ≤ R and x̄ ∈ Rn , and let v : Z → R be a convex function satisfying
(
λ dx ≤ µv ≤ Λ dx in Z,
v=0 on ∂ Z,

for some 0 < λ ≤ Λ. Then there exist positive constants c0 = c0 (n, λ, r), C0 =
C0 (n, Λ, R), and Ĉ = Ĉ(n, Λ, R), such that

c0 ≤ min v ≤ C0 and |v(z)| ≤ Ĉ dist(z, ∂ Z)1/n ∀ z ∈ Z.

Z

Proof. Since µv (Z) ≤ Λ|Z | ≤ Λ|BR ( x̄)| and diam(Z) ≤ 2R, it follows from Theo-
rem 2.8 that
|v(z)| ≤ Ĉ dist(z, ∂ Z)1/n ∀ z ∈ Z,
for some constant Ĉ depending only on n, Λ, and R. In particular, this implies that

min v = max |v| ≤ Ĉ diam(Z)1/n ≤ Ĉ(2R)1/n =: C0 .

Z Z
 4.1 Sections and normalized solutions 69

To prove the lower bound, it suffices to consider the function

 r 2 − |z − x̄| 2 
w(z) := λ1/n .
2
Indeed, since

w|∂Br (x̄) = 0 ≥ v|∂Br (x̄) and µw = det D2 w dx = λ dx ≤ µv,

it follows from Theorem 2.10 that

v≤w in Br ( x̄);

hence,
λ1/n r 2
min v ≥ min v ≥ min w = =: c0 .
Z Br ( x̄) Br ( x̄) 2

An immediate consequence of the above result is that different level sets are
uniformly separated.
Corollary 4.5. Let Z be an open convex set satisfying

Br ( x̄) ⊂ Z ⊂ BR ( x̄)

for some 0 < r ≤ R and x̄ ∈ Rn , and let v : Z → R be a convex function satisfying

(
λ dx ≤ µv ≤ Λ dx in Z,
v=0 on ∂ Z,

for some 0 < λ ≤ Λ. Let x̂0 be a minimum point for v, set h := max Z |v| = |v( x̂0 )|,
and for t ∈ (0, 1], define the sections

Zt := S( x̂0, 0, th).

Then for any τ ∈ (0, 1) and any 0 < t1 ≤ t2 ≤ 1 with t2 − t1 ≥ τ, there exists a
constant ĉτ > 0, depending only on n, λ, Λ, r, R, and τ, such that

dist(Zt1 , ∂ Zt2 ) ≥ ĉτ .

In particular, choosing τ = 1/2, t2 = 1, and t1 ≤ 1/2, we see that

dist( x̂0, ∂ Z) ≥ ĉ,

where ĉ > 0 depends only on n, λ, Λ, r, and R.

70 4 Interior regularity of weak solutions

Proof. Fix 0 < t1 ≤ t2 ≤ 1 with t2 − t1 ≥ τ, and consider the function v̂ := v + t2 h.

Then, (
λ dx ≤ µv̂ ≤ Λ dx in Zt2 ,
v̂ = 0 on ∂ Zt2 ,
and by Proposition 4.4 we deduce that

| v̂(z)| ≤ Ĉ dist(z, ∂ Zt2 )1/n ∀ z ∈ Zt2 .

In particular, since | v̂(z)| ≥ (t2 − t1 )h ≥ τh for z ∈ Zt1 , we obtain

τh ≤ | v̂(z)| ≤ Ĉ dist(z, ∂ Zt2 )1/n ∀ z ∈ Zt1 ,

which implies n
τh

dist(Zt1 , ∂ Zt2 ) ≥ .
Ĉ
Recalling that h ≥ c0 (see Proposition 4.4), this proves the result. 

4.1.4 On the size of sections. In the previous section, we found universal upper
and lower bounds for normalized solutions of Monge–Ampère with right-hand side
bounded away from zero and infinity. We now use this information to show some
important facts on the size of the sections.
The first result states that the measure of a section of height t is comparable
to t n/2 .
Lemma 4.6. Let Ω be an open set and u : Ω → R a convex function such that

λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.

Assume that S(x, p, t) ⊂⊂ Ω. Then there exist constants 0 < c1 ≤ C1 , depending only
on n, λ, and Λ, such that

c1 t n/2 ≤ |S(x, p, t)| ≤ C1 t n/2 .

Proof. Let L be the affine transformation that normalizes S := S(x, p, t), set S ∗ :=
L(S), and define v as in (4.2). Then it follows from Propositions 4.2 and 4.4 that

c0 ≤ min
∗
v ≤ C0,
S

or equivalently, in terms of u,

c0 ≤ (det L)2/n min(u − `x, p,t ) ≤ C0 .

S
 4.1 Sections and normalized solutions 71

Noticing that
min(u − `x, p,t ) = −t
S

(see Figure 4.1), we find

c0 ≤ (det L)2/n t ≤ C0 . (4.5)
On the other hand, since L(S) is normalized and |L(S)| = | det L| |S|, we have

|B1 (0)| ≤ | det L| |S| ≤ |Bn (0)|. (4.6)

Combining (4.5) and (4.6), the result follows. 

By the above volume estimate and Proposition 4.4, we deduce that sections inside
bounded domains with height bounded away from zero contain nontrivial balls.
Corollary 4.7. Let Ω be a bounded open set and u : Ω → R a convex function such
that
λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.
Let x̂ ∈ Ω, τ > 0, and assume that S := S( x̂, p, t) ⊂⊂ Ω for some t ≥ τ. Let L be the
affine transformation that normalizes S. Then there exist K, % > 0, depending only
on τ, n, λ, Λ, and diam(Ω), such that

kL k ≤ K, kL −1 k ≤ K,

and
B% ( x̂) ⊂ S.
Proof. Since S ⊂ Ω ⊂ Bdiam(Ω) ( x̂) and |S| ≥ c1 t n/2 (by Lemma 4.6), we can apply
Lemma A.14(b) to obtain that there exists r > 0, depending only on τ, n, λ, Λ, and
diam(Ω), such that Br (x) ⊂ S for some x ∈ S.
Since
Br (x) ⊂ S ⊂ Bdiam(Ω) (x),
we deduce that

and



L Br (x) ⊂ Bn (0) L Bdiam(Ω) (x) ⊃ B1 (0),

or equivalently,

and L −1 B1 (0) ⊂ Bdiam(Ω) (x),

L Br (x) ⊂ Bn (0)

from which it follows that

n
kL k ≤ , kL −1 k ≤ diam(Ω). (4.7)
r
72 4 Interior regularity of weak solutions

If we now define S ∗ := L(S) and v as in (4.2), we see that ẑ := L x̂ is a minimum

point for v; hence, Corollary 4.5 implies that
dist( ẑ, ∂S ∗ ) ≥ ĉ > 0.
This combined with (4.7) yields
1 ĉr
dist( x̂, ∂S) ≥ dist( ẑ, ∂S ∗ ) ≥ =: %,
kL k n
proving the desired result. 
We now show that if an affine map normalizes a section, then all other sections
with comparable heights are still comparable to balls.
Lemma 4.8. Let Ω be an open set and u : Ω → R a convex function such that
λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.
Assume that S1 := S( x̂, p, t1 ) and S2 := S( x̂, p, t2 ) are both compactly contained in Ω,
let L be the affine transformation that normalizes S1 , and suppose that
t1 h 1 i
∈ ,T (4.8)
t2 T
for some T > 1. Then there exists a radius ρ ∈ (0, 1), depending only on T, n, λ, and
Λ, such that
Bρ (L x̂) ⊂ L(S2 ) ⊂ B1/ρ (L x̂).
Proof. Set S1∗ := L(S1 ), S2∗ := L(S2 ), and x̂ ∗ := L x̂. We distinguish two cases,
depending on whether t2 is larger or smaller than t1 .
Case 1: t1 ≤ t2 . Since S1∗ is normalized and S1∗ ⊂ S2∗ , we can apply Corollary 4.5
with Z = S1∗ , r = 1, R = n, x̄ = 0, and x̂0 = x̂ ∗ to get
dist( x̂ ∗, ∂S2∗ ) ≥ dist( x̂ ∗, ∂S1∗ ) ≥ ĉ
for some constant ĉ > 0 depending only on n, λ, and Λ. In other words,
Bĉ ( x̂ ∗ ) ⊂ S2∗ . (4.9)
For the other inclusion, we apply Lemma 4.6 both to S1 and to S2 , and (4.8) to get
|S2∗ | = | det L| |S2 | ≤ C1 | det L|t2n/2 ≤ C1T n/2 | det L|t1n/2
C1T n/2
≤ | det L| |S1 |
c1
C1T n/2 ∗
= |S1 |
c1
C1T n/2
≤ |Bn (0)|.
c1
 4.1 Sections and normalized solutions 73

Thus, |S2∗ | ≤ C2 for some constant depending only on T, n, λ, and Λ. This bound and
the inclusion (4.9) allow us to apply Lemma A.14(a) to obtain

S2∗ ⊂ B1/ρ ( x̂ ∗ )

for some small ρ > 0 depending only on T, n, λ, and Λ.

Case 2: t2 < t1 . In this case, we apply Corollary 4.7 with Ω = S1∗ , S = S2∗ , x̄ = x̂ ∗ ,
and τ = 1/T to get
B% ( x̂ ∗ ) ⊂ S2∗,
where % depends only on T, n, λ, and Λ. On the other hand, since x̂ ∗ ∈ S1∗ ⊂ Bn (0),
assuming without loss of generality that 1/ρ ≥ 2n, we have

S2∗ ⊂ Bn (0) ⊂ B1/ρ ( x̂),

concluding the proof. 

A direct consequence of the previous lemma and Proposition 4.4 is a universal

Lipschitz bound inside “strictly contained” round sections.

Corollary 4.9. Let Z be an open bounded convex set, v : Z → R be a convex function

satisfying (
λ dx ≤ µv ≤ Λ dx in Z,
v=0 on ∂ Z,
for some 0 < λ ≤ Λ. Let x̂0 be a minimum point for v, set h := max Z |v| = |v( x̂0 )|,
and consider the section
Z1/2 := S( x̂0, 0, h/2).
Assume that
Br ( x̄) ⊂ Z1/2 ⊂ BR ( x̄) (4.10)
for some 0 < r ≤ R and x̄ ∈ Rn . Then there exists a constant C, depending only on
n, λ, Λ, r, and R, such that
|∇v| ≤ C in Z1/2 .

Proof. Let L be the affine map that normalizes Z1/2 . Then it follows by (4.10) that

and L −1 B1 (0) ⊂ BR ( x̄).

L Br ( x̄) ⊂ Bn (0)

Therefore,
n
kL k ≤ and kL −1 k ≤ R. (4.11)
r
74 4 Interior regularity of weak solutions

Applying Lemma 4.8 with S1 = L(Z1/2 ), S2 = L(Z), and T = 2, we obtain that there
exists ρ ∈ (0, 1), depending only on n, λ, and Λ, such that

Bρ (L x̂0 ) ⊂ L(Z) ⊂ B1/ρ (L x̂0 ).

Hence, by (4.11), we see that

Br ρ/n ( x̂0 ) ⊂ Z ⊂ BR/ρ ( x̂0 ).

This allows us to apply Proposition 4.4 and Corollary 4.5 to deduce that

c0 ≤ h ≤ C0 and dist(Z1/2, ∂ Z) ≥ ĉ

for some constants c0 , C0 , ĉ > 0 depending only on n, λ, Λ, r, and R. So, it follows

by Corollary A.23 that
2C0
k∇vk L ∞ (Z1/2 ) ≤ ,
ĉ
as desired. 

4.2 On the strict convexity of solutions

In [13], Caffarelli was able to fully exploit the affine invariance of Monge–Ampère to
extend Theorem 2.17 to higher dimensions and show that solutions are either strictly
convex or affine on a segment crossing the domain. As we shall see, this implies the
strict convexity of solutions with zero boundary data, and actually the strict convexity
of solutions with boundary data of class C 1,α with α > 1 − 2/n. Also, whenever one
considers normalized solutions, simple compactness arguments show that the strict
convexity is universal.

4.2.1 Strict convexity of solutions with zero boundary data. The following result,
due to Caffarelli [13], extends Theorem 2.17 to higher dimensions. For the definition
of exposed points and their main properties, we refer to Appendix A.3.1.
Theorem 4.10. Let Ω be an open set and u : Ω → R a convex function such that

λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.

Fix x ∈ Ω and p ∈ ∂u(x), consider the supporting plane `(z) := u(x) + hp, z − xi, and
define the convex set Σ := {u = `}. Then one of the following two properties holds:
(1) Σ = {x} (i.e., u is strictly convex at x);
(2) Σ has no exposed points inside Ω.
 4.2 On the strict convexity of solutions 75

Notice that in view of Pogorelov’s counterexample (see Section 3.2), the above
result is optimal.
Before proving this theorem, we first state an important corollary.

Corollary 4.11. Let Ω be an open bounded convex set and u : Ω → R a convex

function satisfying (
λ dx ≤ µu ≤ Λ dx in Ω,
u=ϕ on ∂Ω,

for some 0 < λ ≤ Λ and ϕ ∈ C 1,α (∂Ω) with α > 1 − 2/n. Then u is strictly convex.

Proof. Assume, to the contrary, that u is not strictly convex. Then it follows from
Theorem 4.10 that u is affine on a convex set Σ ⊂ Ω, and Σ has no exposed points
inside Ω. Since any compact convex set is the closure of the convex hull of its
exposed points (see Theorem A.9), all the exposed points of Σ must be contained on
∂Ω. Furthermore, at least two exposed points of Σ must live on ∂Ω for otherwise Σ
would be just one point.
Let x̂, ŷ ∈ Σ ∩ ∂Ω be two such points. Since Σ is convex, we deduce that u is
affine along the segment Σ 0 := [ x̂, ŷ] connecting x̂ and ŷ. Notice that if ϕ is affine,
then one can conclude as in the proof of Corollary 2.18. We now explain how to
conclude in the general case.
Up to subtracting an affine function, we can assume that

u≥0 in Ω and u|Σ0 ≡ 0. (4.12)

Also, up to a change of coordinates, we can assume that x̂ = ae1 and ŷ = −ae1 for
some a > 0. Pick ρ > 0 small (to be fixed later) and define the set

Ωρ := {x = (x1, x 0) ∈ Ω ⊂ R × Rn−1 : |x 0 | < ρ}

(see Figure 4.2). Since ϕ ∈ C 1,α (∂Ω) coincides with u|∂Ω , it follows by (4.12) that
ϕ ≥ 0 and ϕ vanishes at ±ae1 ∈ ∂Ωρ ; hence,

ϕ ≤ C̃ ρ1+α on ∂Ω ∩ ∂Ωρ .

(Here, C̃ := kϕkC 1, α (∂Ω) .) Recalling that u|∂Ω = ϕ, the bound above combined with
the convexity of u yields
u ≤ C̃ ρ1+α on ∂Ωρ . (4.13)
Now, for h > 0 (to be chosen), we consider the function
 1 
v(x1, x 0) := λ1/n h n−1 x12 + |x 0 | 2 − ρ1+α .
h
76 4 Interior regularity of weak solutions

Figure 4.2. The set Ωρ .

Let C̃ > 0 be as in (4.13). If we choose h := ĉρ1−α with ĉ > 0 a small constant, as

|x 0 | = ρ on ∂Ωρ \ ∂Ω, we have

λ1/n 2−(1−α)  λ1/n 

v≥ ρ − ρ1+α = − 1 ρ1+α ≥ C̃ ρ1+α on ∂Ωρ \ ∂Ω,
ĉ ĉ
provided ĉ is sufficiently small (the smallness being independent of ρ). In addition,
since (n − 1)(1 − α) < 1 + α (because α > 1 − 2/n) and |x1 | ≥ a/2 on ∂Ω ∩ ∂Ωρ (see
Figure 4.2), it follows that

n−1 a2
v ≥ λ1/n ĉρ1−α − ρ1+α
4
λ1/n a2 (n−1)(1−α)
= ρ − ρ1+α
4ĉ
≥ C̃ ρ1+α on ∂Ωρ ∩ ∂Ω,

provided ρ is small enough. Hence, recalling (4.13), we have shown that

v≥u on ∂Ωρ,

for ρ small enough. Notice that

det D2 v dx = λ dx ≤ µu in Ωρ .

Thus, by Theorem 2.10, we see that v ≥ u inside Ωρ . However, this is impossible

since
v(0) = −ρ1+α < 0 = u(0)
(recall that 0 ∈ Σ 0), which proves that u has to be strictly convex. 
Proof of Theorem 4.10. Assume, to the contrary, that Σ is not a point, but instead
that there exists an exposed point for Σ inside Ω. Let x̄ ∈ Ω be this point.
 4.2 On the strict convexity of solutions 77

Figure 4.3. The set Σ = {u = 0}. In this figure, the set Σ is “thick”, but it could also be lower-
dimensional. For instance, Σ could be a segment contained in the line Re1 , with x̄ being one of its
endpoints.

Up to subtracting an affine function and a change of coordinates, we can assume

that

` ≡ 0, 0 ∈ Σ, Σ ∩ {z1 ≥ 0} ⊂ Ω, Σ ⊂ {z1 ≤ a}, and x̄ = ae1 (4.14)

for some a > 0 (see Figure 4.3).

Now, given ε > 0 small, set

`ε (z) := εz1, uε := u − `ε, and Sε := {u < `ε },

and notice that

Sε → Σ ∩ {z1 ≥ 0} in the Hausdorff distance as ε → 0, (4.15)

and (
λ dx ≤ µuε ≤ Λ dx in Sε,
uε = 0 on ∂Sε .

Let Lε be the affine transformation that normalizes Sε , set Sε∗ := Lε (Sε ) and
xε∗ := Lε ( x̄), and let vε be the normalized solution associated to (uε, Lε ) (see Propo-
sition 4.2).
The basic idea is that the value of vε at xε∗ is close to the minimum of vε inside Sε∗ ,
but at the same time xε∗ is close to ∂Sε∗ , and this is in contradiction with Proposition 4.4
(see Figure 4.4).

Step 1: vε (xε∗ ) ' minSε∗ vε . Consider the ratio between vε (xε∗ ) and minSε∗ vε . Recall-
ing the definition of vε (see (4.2)) and that x̄ = ae1 , since u( x̄) = `( x̄) = 0 and u ≥ 0,
78 4 Interior regularity of weak solutions

Figure 4.4. We subtract the affine function `ε from u, and then we normalize. In this way, we get a
normalized solution vε almost attaining its minimum value at xε∗ . However, since xε∗ converges to
a point on ∂Sε∗ as ε → 0, this contradicts Alexandrov estimates for ε  1.

we get

vε (xε∗ ) (det Lε )2/n uε ( x̄) u( x̄) − ε x̄1

= =
minSε∗ vε (det Lε )2/n minSε∗ (uε ◦ Lε−1 ) minSε (u − `ε )
−εa
=
minSε (u − εx1 )
εa
=
maxSε (εx1 − u)
εa a
≥ = .
maxSε εx1 maxSε x1

Then it follows from (4.14) and (4.15) that the last term converges to
a
=1
maxΣ x1
as ε → 0 (see Figure 4.3). In particular,

vε (xε∗ ) 1
≥ for ε  1.
minSε∗ vε 2

Step 2: dist(xε∗ , ∂Sε∗ )  1. Consider the hyperplanes

Π0 := {z1 = 0}, Π1 := {z1 = a}, and Π2 := {z1 = aε },

where aε = a + o(1) is such that Π2 is a supporting hyperplane for Sε (see Figure 4.5),
and set Πi∗ := Lε (Πi ), i = 0, 1, 2. Since affine transformations preserve the ratio
between the distances of parallel hyperplanes, we get

dist(Π1∗, Π2∗ ) dist(Π1, Π2 ) aε − a

= = →0 as ε → 0.
dist(Π0∗, Π2∗ ) dist(Π0, Π2 ) a
 4.2 On the strict convexity of solutions 79

Figure 4.5. Since affine transformations preserve the ratio between the distances of parallel hyper-
planes, we see that dist(xε∗ , ∂Sε∗ ) → 0 as ε → 0.

On the other hand, since Π0∗ and Π2∗ are supporting hyperplanes for Sε∗ ⊃ B1 (0) and
xε∗ ∈ Π1∗ , we deduce that

dist(xε∗ , ∂Sε∗ ) ≤ dist(Π1∗, Π2∗ ) and dist(Π0∗, Π2∗ ) ≥ 2.

Combining these two estimates, we conclude that

dist(xε∗ , ∂Sε∗ ) → 0 as ε → 0.

Step 3: Conclusion. Combining Step 2 and Proposition 4.4, we get

|vε (xε∗ )| ≤ Ĉ dist(xε∗ , ∂Sε∗ )1/n → 0 as ε → 0,

while Step 1 and Proposition 4.4 yield

1 c0
|vε (xε∗ )| ≥ min vε ≥ for ε  1,
2 Sε
∗ 2

a contradiction. 

4.2.2 Universal strict convexity of normalized solutions. In this section, we focus

on normalized solutions (or solutions in “round” domains), and we show that their
strict convexity is “universal” in several ways. All the proofs rely on the compactness
of normalized solutions. Notice that as a consequence of Corollary 4.11, solutions
with zero boundary data are strictly convex; hence, they have a unique minimum
point.
Recall that given a section S = S(x, p, t) and τ > 0, the set τS is the convex set
obtained by dilating S by a factor τ with respect to x (see (4.1)).
80 4 Interior regularity of weak solutions

Figure 4.6. Since ṽ is convex, it follows that 21 Z ⊂ Z1/2 , or equivalently max 1 Z ṽ ≤ h/2. Proving
2
that max 1 Z ṽ ≤ (1 − δ)h/2 is a way of quantifying the strict convexity of ṽ (see also Figure 4.7).
2

Lemma 4.12. Let Z be an open convex set satisfying

Br ( x̄) ⊂ Z ⊂ BR ( x̄) (4.16)
for some 0 < r ≤ R and x̄ ∈ Rn , and let v : Z → R be a convex function satisfying
(
λ dx ≤ µv ≤ Λ dx in Z,
(4.17)
v=0 on ∂ Z,

for some 0 < λ ≤ Λ. Let x̂0 be the minimum point of v, set h := max Z |v| = |v( x̂0 )|,
and consider the section
Z1/2 := S x̂0, 0, h/2 .

Also, define ṽ := v + h, so that

ṽ ≥ 0, ṽ( x̂0 ) = 0, ṽ|∂Z = h, and ṽ|∂Z1/2 = h/2.
Then there exists δ > 0, depending only on n, λ, Λ, r, and R, such that
h
max ṽ ≤ (1 − δ) .
1
2Z
2

Notice that Z = S( x̂0, 0, h), so the convexity of ṽ implies that 12 Z ⊂ Z1/2 (see
Figure 4.6); thus, max 1 Z ṽ ≤ h/2. The factor 1 − δ in front of h/2 quantifies the
2
strict convexity of v.
Proof. Assume, to the contrary, that the result is false. Then we can find a sequence
of solutions vk of (4.17) in convex domains Zk satisfying (4.16) such that (with the
obvious notation)
 1  hk hk
1− ≤ max ṽk ≤ . (4.18)
k 2 1
2 Zk
2
 4.2 On the strict convexity of solutions 81

Figure 4.7. If the estimate max 1 Z ṽ ≤ (1 − δ)h/2 were false, by compactness, we could construct a
2
solution ṽ∞ such that ṽ∞ ( x̄) = h∞ /2 for some point x̄ ∈ ∂ Z∞,1/2 . This implies that ṽ∞ is affine on
the segment [ x̂∞, x̄], a contradiction to its strict convexity.

Up to a translation, we can assume that

Br (0) ⊂ Zk ⊂ BR (0) ∀ k ∈ N. (4.19)

In this way, the Blaschke selection theorem ensures that the family {Z k }k ∈N is com-
pact with respect to the Hausdorff convergence of sets. Also, the family of measures
{µvk }k ∈N is compact for the weak∗ topology. So, we can apply Corollary 2.12 to
deduce that, up to a subsequence, vk → v∞ locally uniformly, and v∞ solves (4.17)
in a convex set Z∞ satisfying (4.19).
On the other hand, letting k → ∞ in (4.18), we deduce that there exists a point
h∞
x̄ ∈ ∂ Z∞,1/2 such that ṽ∞ ( x̄) = .
2
By convexity of ṽ∞ , this implies that ṽ∞ is affine on the segment joining its minimum
point and x̄ (see Figure 4.7), which contradicts Corollary 4.11 (applied with ϕ ≡ 0).

A similar argument shows that sections centered at interior points cannot touch
the boundary if the height is sufficiently small.
Lemma 4.13. Let Z be an open convex set satisfying (4.16), and let v : Z → R be
a convex function satisfying (4.17) for some 0 < λ ≤ Λ. Fix ρ > 0, let x̂0 be the
minimum point of v, set h := max Z |v| = |v( x̂0 )|, and consider the section

Zρ := S x̂0, 0, (1 − ρ)h .

Then there exists τ > 0, depending only on n, λ, Λ, r, R, and ρ, such that

S(x, p, t) ⊂⊂ Z ∀ x ∈ Zρ, p ∈ ∂v(x), t ≤ τ.

82 4 Interior regularity of weak solutions

Proof. Again, assuming the result were false, we would have a sequence of functions
vk solving (4.17) in domains Zk satisfying (4.16) such that

S(xk , pk , tk ) ∩ ∂ Zk , ∅ for some xk ∈ Zk,ρ, pk ∈ ∂vk (xk ), tk > 0,

with tk → 0 as k → ∞. Up to a translation, taking a subsequence and in the limit, we

would get a solution v∞ of (4.17) in a domain Z∞ satisfying (4.16), a point x∞ ∈ Z∞,ρ ,
and a slope p∞ ∈ ∂v∞ (x∞ ) such that

S(x∞, p∞, t) ∩ ∂ Z∞ , ∅ ∀t > 0

(compare with the proof of Lemma 4.12). This contradicts the strict convexity of v∞
provided by Corollary 4.11, concluding the proof. 
Finally, we show that sections with small height cannot be close both to x̂0 , the
minimum of our solution, and ∂ Z1/2 .
Lemma 4.14. Let Ω ⊂ Rn be an open set, Z ⊂ Ω an open convex set satisfying
(4.16), and let v : Ω → R be a convex function satisfying (4.17) for some 0 < λ ≤ Λ.
Let x̂0 ∈ Z be the minimum point of v in Ω, set h := max Z |v| = |v( x̂0 )|, and consider
the section
Z1/2 := S x̂0, 0, h/2 .

Then there exists θ > 0 small, depending only on n, λ, Λ, r, and R, such that

x ∈ Ω, p ∈ ∂u(x), and S(x, p, θ) ∩ S( x̂0, 0, θ) , ∅ ⇒ x ∈ Z1/2 .

Proof. If the result were false, we would find a sequence of functions vk : Ωk → R

solving (4.17) inside convex sets Zk ⊂ Ωk satisfying (4.16) such that

S(xk , pk , θ k ) ∩ S( x̂k , 0, θ k ) , ∅ for some xk ∈ Zk \ Zk,1/2, pk ∈ ∂uk (xk ),

for a sequence θ k → 0, with x̂k the minimum of vk in Zk . In particular, if we set

`xk , pk (z) := vk (xk ) − hpk , z − xk i, since

0 ≤ vk − ` x k , p k ≤ θ k in S(xk , pk , θ k ),

we deduce that, for k large enough, there exists a segment Σk ⊂ S(xk , pk , θ k ) ∩ Zk

such that

0 ≤ (vk − `xk , pk )|Σk ≤ θ k , Σk ∩ ∂ Zk,1/2 , ∅, and Σk ∩ S( x̂k , 0, θ k ) , ∅.

Hence, letting k → ∞ we get a limit v∞ solving (4.17) inside a convex set Z∞

satisfying (4.16) and a segment Σ∞ ⊂ Z∞ such that

v∞ is affine on Σ∞, Σ∞ ∩∂ Z∞,1/2 , ∅, and Σ∞ ∩S( x̂∞, 0, t) , ∅ ∀ t > 0.

 4.2 On the strict convexity of solutions 83

(Again, compare with the proofs of the above two lemmas.) Since
Ù
S( x̂∞, 0, t) = { x̂∞ }
t>0

(by the strict convexity of v∞ ), we get that v∞ is affine on a segment joining ∂ Z∞,1/2
to x̂∞ , impossible. 

4.2.3 A quantitative strict convexity estimate. In the previous section, we obtained

several universal estimates on the sections of normalized solutions. We now show
that normalized solutions separate from their tangent plane at a point x at least as
|y − x| M for some (large) M > 0 (the case M = 2 would correspond to saying that u
is uniformly convex).
Notice that if u is a strictly convex solution in a domain Ω, we can pick any section
S := S(x, p, t) ⊂⊂ Ω, consider the affine transformation that normalizes S, and apply
Theorem 4.16 below to the function v defined as in (4.2) to obtain a quantified strict
convexity estimate for u with the same exponent M.
We begin with the following result, which is a simple consequence of strict
convexity and compactness (since the proof is completely analogous to the one for
Lemma 4.14, we skip it).
Lemma 4.15. Let Z be an open convex set satisfying (4.16), and let u : Z → R be
a convex function satisfying (4.17) for some 0 < λ ≤ Λ. Fix ρ > 0, let x̂0 be the
minimum point of u, set h := max Z |u| = |u( x̂0 )|, and consider the section

Z1/2 := S x̂0, 0, h/2 .

Then for any ρ > 0, there exists σ > 0, depending only on n, λ, Λ, r, R, and ρ, such
that

u(y) ≥ u(x) + hp, y − xi + σ ∀ x ∈ Z1/2, p ∈ ∂u(x), y ∈ Ω \ Bρ (x).

We can now prove the following:

Theorem 4.16. Let Z ⊂ Rn be a normalized convex set, and let u : Z → R be a
convex function satisfying (4.17) for some 0 < λ ≤ Λ. Let x̂0 be the minimum point
of u, set h := max Z |v| = |u( x̂0 )|, and consider the section

Z1/2 := S( x̂0, 0, h/2).

Then there exist constants c > 0 and M > 1, depending only on n, λ, and Λ, such
that
u(y) ≥ u(x) + hp, y − xi + c|x − y| M ∀ x, y ∈ Z1/2, p ∈ ∂u(x).
84 4 Interior regularity of weak solutions

Figure 4.8. Since dist(S1/2

∗ , ∂S ∗ ) ≥ ĉ > 0, we can choose γ ∈ (0, 1) sufficiently close to 1 so that

dist (1 − γ)z∗ + γz, ∂S ∗ ≤ ĉ for all z ∈ ∂S ∗ .

Proof. Fix x ∈ Z1/2 and p ∈ ∂u(x), and let τ > 0 be given by Lemma 4.13 with
ρ = 1/2. Up to subtracting an affine function from u, we can assume that

u(x) = 0 and p = 0. (4.20)

Step 1: There exists γ < 1 such that u(x + γy) ≥ u(x + y)/2 for all x + y ∈ S(x, 0, τ).
Fix t ≤ τ, let L be the affine transformation that normalizes S := S(x, 0, t), set
S ∗ := L(S), and define v as in (4.2). Also, let z∗ := L x ∈ S ∗ denote the minimum
point of v, set h∗ := |v(z ∗ )|, and define
∗
S1/2 := S(z ∗, 0, h∗ /2).

Then it follows by Corollary 4.5 that

dist(S1/2
∗
, ∂S ∗ ) ≥ ĉ

for some ĉ > 0 universal. Hence, choosing γ ∈ (0, 1) a universal constant close
enough to 1, we can ensure that

(1 − γ)z ∗ + γz ∈ S ∗ \ S1/2
∗
∀ z ∈ ∂S ∗

(see Figure 4.8), that is,

h∗
v (1 − γ)z ∗ + γz ≥ − ∀ z ∈ ∂S ∗ .


2
Recalling that v|∂S ∗ = 0 and that u(x) = 0 (see (4.20)), the above equation translated
back to u yields
u(x + y)
u(x + γy) ≥ ∀ x + y ∈ ∂S(x, 0, t).
2
 4.2 On the strict convexity of solutions 85

Since t is an arbitrary constant less than τ, we get

u(x + y)
u(x + γy) ≥ ∀ x + y ∈ S(x, 0, τ),
2
as desired.
Step 2: The iteration argument. Applying the estimate from Step 1 to the sequence
of points {γ k y}k ≥0 iteratively, we get

u(x + γ k y) ≥ 2−k u(x + y) ∀ x + y ∈ S(x, 0, τ), k ≥ 0.

We now recall that thanks to Corollary 4.7, there exists a universal radius % > 0 such
that B% (x) ⊂ S(x, 0, τ); therefore,

u(x + γ k y) ≥ 2−k u(x + y) ∀ y ∈ B% (0), k ≥ 0,

and choosing M > 1 such that γ M = 1/2, we obtain

u(x + γ k y) ≥ γ k M u(x + y) ∀ y ∈ B% (0), k ≥ 0.

Furthermore, it follows by Lemma 4.15 applied with ρ = γ % that u ≥ σ outside

Bγ% (x). Thus,

u(x + γ k y) ≥ γ k M σ ∀ y ∈ B% (0) \ Bγ% (0), k ≥ 0. (4.21)

Step 3: Conclusion. Given an arbitrary point y 0 ∈ B% (0), we can choose k = k(y 0) ≥

0 such that
γ k+1 % < |y 0 | ≤ γ k %.
Then apply (4.21) with y := y 0 γ −k ∈ B% (0) \ Bγ% (0) to get (notice that γ k ≥ |y 0 |/ρ)
σ 0M
u(x + y 0) ≥ γ k M σ ≥ |y | ∀ y 0 ∈ B% (0),
%M

which proves the desired bound when x + y 0 ∈ B% (x).

On the other hand, when x + y 0 < B% (x), it suffices to observe that u ≥ σ in
Z1/2 \ B% (x) and that y 0 ∈ B2n (0) (since x, x + y 0 ∈ Z1/2 ⊂ Bn (0)) to obtain
σ
u(x + y 0) ≥ |y 0 | M ∀ x + y 0 ∈ Z1/2 \ B% (x),
(2n) M

which concludes the proof. 

86 4 Interior regularity of weak solutions

4.3 A Liouville theorem

The classical Liouville theorem for elliptic equations allows one to classify global
solutions with some special growth. For instance, if u : Rn → R is a harmonic
function that grows no faster than |x| M at infinity, then u is a polynomial of degree
at most M.
We now want to find an analogous rigidity result for Monge–Ampère. We will
prove that the only global solutions of det D2 u = 1 are quadratic polynomials, and as
an application, we will show a rigidity result in convex geometry.

4.3.1 Global solutions of det D2 u = 1 are quadratic polynomials. Our goal here
is to prove that global Alexandrov solutions of det D2 u = 1 are quadratic polynomials.
This result was first proved by Jorgens in the two-dimensional setting [70], and then
extended by Calabi to the case n ≤ 5 [24]. Finally, Pogorelov proved this result in
any dimension [98].
An important step in the proof is to show that large sections remain comparable
to balls at all scales. This will be a consequence of the following:
Lemma 4.17. Fix M > 0, H ≥ 1, and assume that u ∈ C 2 (Z) is a convex function
solving (
det D2 u = 1 in Z,
u=h on ∂ Z,
1 
for some open convex set Z and h ∈ H , H . Assume that min Z u = 0 and that

|D2 u( x̄)| ≤ M for some x̄ ∈ {u ≤ 3h/4}. (4.22)

Then there exists ρ ∈ (0, 1), depending only on n, M, and H, such that
Bρ ( x̄) ⊂ Z ⊂ B1/ρ ( x̄).
Proof. The idea of the proof is that if we normalize Z, then the corresponding
normalized solution will have a bounded Hessian at the image of x̄ under the nor-
malizing map. However, this is compatible with (4.22) only if the normalizing map
is bi-Lipschitz, which means that the section Z was already round.
Step 1: A universal bound on the normalized solution. Let L be the affine function
that normalizes Z, set Z ∗ := L(Z), and define
v := (det L)2/n [u ◦ L −1 − h], (4.23)
so that (
det D2 v = 1 in Z ∗,
v=0 on ∂ Z ∗
 4.3 A Liouville theorem 87

(see Proposition 4.2). Set x̄ ∗ := L x̄, and h∗ := | min Z ∗ v|. Then (4.22) implies that

x̄ ∗ ∈ {v ≤ −h∗ /4}.

Also, it follows by Corollary 4.5 that

dist( x̄ ∗, ∂ Z ∗ ) ≥ ĉ > 0,

where ĉ depends only on the dimension. Hence, we can apply Theorem 3.9 to deduce
that
|D2 v( x̄ ∗ )| ≤ Ĉ.
So, since det D2 v( x̄ ∗ ) = 1, by Remark 1.1 we see that
1
Id ≤ D2 v( x̄ ∗ ) ≤ C̄ Id, (4.24)
C̄
where C̄ depends only on the dimension.
Step 2: A first bound on L. Recalling (4.23), we can rewrite (4.24) in terms of u to
get
1
Id ≤ (det L)2/n [∇L ∗ ]−1 D2 u( x̄)[∇L]−1 ≤ C̄ Id . (4.25)
C̄
We now observe that, since det D2 u( x̄) = 1, (4.22) and Remark 1.1 imply
1
Id ≤ D2 u( x̄) ≤ CM Id
CM
for some constant CM > 1 depending only on M and n. These bounds and (4.25)
yield
1
Id ≤ (det L)2/n [∇L ∗ ]−1 [∇L]−1 ≤ C̄CM Id .
C̄CM

Step 3: L is bi-Lipschitz. Let x0 be the minimum point of u. Since Z = S(x0, 0, h)

and h ∈ H1 , H , it follows by Lemma 4.6 that


c1
≤ |Z | ≤ C1 H n/2 .
H n/2
Recalling the estimate

|B1 (0)| ≤ |Z ∗ | = | det L| |Z | ≤ |Bn (0)|,

we see that
c̄1 ≤ | det L| ≤ C̄1,
88 4 Interior regularity of weak solutions

where c̄1, C̄1 > 0 depend only on n and H. Hence, combining this bound with Step
2, we obtain that
1
Id ≤ [∇L ∗ ]−1 [∇L]−1 ≤ C 0 Id
C0
for some C 0 = C 0(n, M, H) > 1. Thus, Lemma A.4 yields
√ √
kL k ≤ C 0 and kL −1 k ≤ C 0 . (4.26)

Step 4: Conclusion. Since

B1 (0) ⊂ L(Z) ⊂ Bn (0)

it follows by (4.26) that

B1/√C 0 (x0 ) ⊂ L −1 B1 (0) ⊂ Z ⊂ L −1 Bn (0) ⊂ B√C 0 n (x0 ), (4.27)

where x0 := L −1 0. This allows us to apply Corollary 4.5 with t1 = 3/4 and t2 = 1 to

deduce that
dist( x̄, ∂ Z) ≥ ĉ 0
√
for some ĉ 0 = ĉ 0(n, M, H) > 0. Now, as diam(Z) ≤ 2 C 0 n (by (4.27)), we conclude
that
Bĉ0 ( x̄) ⊂ Z ⊂ B2√C 0 n ( x̄),
n o
which proves the result with ρ := min ĉ 0, √1 0 . 
2 Cn

We can now classify global solutions:

Theorem 4.18. Let u : Rn → R be a convex function satisfying

µu = dx in Rn .

Then u is a quadratic polynomial, that is,

1
u(z) = hAz, zi + hb, zi + c ∀ z ∈ Rn,
2
where A ∈ Rn×n is a symmetric positive-definite matrix satisfying det A = 1.

Proof. The proof is divided into three steps. We first prove that u is strictly convex
and smooth. Then we use Lemma 4.17 to deduce that the sublevel sets of u are round.
Finally, we conclude using the Evans–Krylov interior C 2,α estimates.
 4.3 A Liouville theorem 89

Step 1: u is strictly convex and smooth. If u is not strictly convex, then it follows by
Theorem 4.10 that u coincides with an affine function on a line, and Lemma A.25
implies that ∂u(Rn ) is contained in a hyperplane. This gives

|E | = µu (E) = |∂u(E)| ≤ |∂u(Rn )| = 0 ∀ E ⊂ Rn Borel,

a contradiction.
This proves that u is strictly convex; hence, u ∈ C ∞ (Rn ) by Theorem 3.10.

Step 2: The sublevels of u are round. Up to subtracting an affine function from

u, we can assume that u(0) = 0 and u ≥ 0. Notice that since u is strictly convex,
u(z) → +∞ as |z| → +∞. Also, since u ∈ C ∞ and det D2 u(0) = 1, there exists a
constant M ≥ 1 such that
|D2 u(0)| ≤ M.

Consider the sublevel sets

UR := {u ≤ R2 } ∀ R > 0.

We want to show that UR is comparable to a ball of radius R. To this aim, we consider

the function
1
u R (x) := 2 u(Rx),
R
which solves
1
det D2 u R = 1 in  R, 
(
RU
uR = 1 on ∂ R1 UR ,

where R1 UR denotes the dilation of UR by a factor 1/R with respect to 0. Indeed,

since |D2 u R (0)| = |D2 u(0)| ≤ M, we can apply Lemma 4.17 with x̄ = 0 to deduce
that
1
Bρ (0) ⊂ R UR ⊂ B1/ρ (0) ∀ R > 0. (4.28)

Step 3: C 2,α interior estimates and conclusion. Thanks to (4.28), we can apply
Theorem 3.9, Proposition A.43, and Theorem A.42 inside Bρ/2 (0) to deduce that

kD2 u R kC 0, α (Bρ/4 (0)) ≤ C ∀ R > 0.

90 4 Interior regularity of weak solutions

Since1
[D2 u R ]C 0, α (Bρ/4 (0)) = Rα [D2 u]C 0, α (Bρ R/4 (0)),
we obtain
C
[D2 u]C 0, α (Bρ R/4 (0)) ≤ ,
Rα
and letting R → ∞, we get
[D2 u]C 0, α (Rn ) = 0.
This proves that D2 u is constant, and so u is a quadratic polynomial. 

4.4 Application 2: Petty’s theorem

Given an open bounded convex set K ⊂ Rn , one can associate to it the so-called
surface area measure. This is a small variant of the curvature measure defined in
Section 2.6.1, and it is defined as

νK (E) := Hn−1 NK−1 (E) ∀ E ⊂ Sn−1 Borel,

where NK is the normal mapping defined in (2.33) (see also Figure 2.12). If one
assumes that the measure νK is absolutely continuous with respect to the surface
measure on the sphere, then one can write

νk = fK Hn−1 Sn−1,

and fK : Sn−1 → R is called the curvature function of K.2 In [96], Petty studied
the following three inequalities: the affine isoperimetric inequality, the Blaschke–
Santalò inequality, and the geominimal surface area inequality. He proved that if K
is extremal for any of these inequalities, then there exists a positive constant cK such
that, up to a translation,
cK
fK = n+1 on Sn−1 . (4.29)
hK
1Recall that the Hölder norm of a function f in a domain Ω is defined as

k f kC 0, α (Ω) := k f k L ∞ (Ω) + [ f ]C 0, α (Ω),

where the Hölder seminorm [ f ]C 0, α (Ω) is given by

| f (x) − f (y)|
[ f ]C 0, α (Ω) := sup .
x,y∈Ω |x − y | α

1
2If K is a uniformly convex domain of class C 2 , then fK = −1 , where κ is the Gaussian curvature of ∂K.
κ◦N K
 4.4 Application 2: Petty’s theorem 91

Here
hK (y) := max hy, zi ∀ y ∈ Sn−1 (4.30)
z ∈K
is the support function of K (notice that (4.29) implies that hK > 0, from which
it follows that necessarily K has to contain the origin). Hence, the main question
becomes, how do we characterize the family of convex domains K that satisfy (4.29)?
As shown in [96], if K is of class C 2 , then K must be an ellipsoid. This regularity
assumption has recently been removed in [88] by an approximation argument.
In this section, we repeat Petty’s argument, and we show how the characterization
of convex domains satisfying (4.29) reduces to the classification of global convex
solutions of the equation det D2 u = 1. Then applying Theorem 4.18, we obtain
Petty’s result.
Theorem 4.19. Let K be bounded convex set K ⊂ Rn satisfying (4.29). Then K is
an ellipsoid centered at the origin.
Sketch of the proof. We first consider the case when K is a uniformly convex domain
of class C 2 , and then we discuss how to extend the result to general convex sets.
Step 1: K is a uniformly convex domain of class C 2 . Extend hK to a positively
1-homogeneous function on all of Rn by considering
 y 
H(y) := |y|hK ∀ y ∈ Rn, (4.31)
|y|
and define
1 2
F(y) := H (y). (4.32)
2
Notice that
D2 F(y) = ∇H(y) ⊗ ∇H(y) + H(y)D2 H(y). (4.33)
Also, since ∇H is 0-homogeneous, we see that
d
0= ∇H(t y) = D2 H(y) · y; (4.34)
dt t=1

that is, y belongs to the kernel of D2 H(y).

Now, fix y ∈ Sn−1 and choose a system of coordinates so that y is the first element
of an orthonormal basis for Rn . Consider the hyperplane y ⊥ := {z ∈ Rn : hz, yi =
0}, and denote by py ∈ Rn the projection of ∇H(y) onto y ⊥ . Thus,

∇H(y) = h∇H(y), yiy + py .

Then it follows by (4.33) and (4.34) that

h∇H(y), yi 2
 
2 h∇H(y), yipy
D F(y) = ,
h∇H(y), yipy H(y)D2 H(y) + py ⊗ py
92 4 Interior regularity of weak solutions

and a direct computation gives

det D2 F(y) = h∇H(y), yi 2 H(y)n−1 det D2 H|y ⊥ (y),

where det D2 H|y ⊥ denotes the determinant of D2 H(y) seen as a symmetric operator


on the (n − 1)-dimensional space y ⊥ .
We now observe that since H is 1-homogeneous and |y| = 1,
d d
h∇H(y), yi = H(t y) = tH(y) = H(y) = hK (y).


dt t=1 dt t=1

Also, it is well known that D2 H|y ⊥ is the inverse of the second fundamental form
of ∂K at the point ∇H(y) = NK−1 (y) (see, for instance, [107, Section 2.5]). Hence
(recall footnote 2 at the beginning of this section),
1
det D2 H|y ⊥ (y) = = fK (y).


κ ◦ NK−1 (y)

In summary, we have proved that

det D2 F(y) = hK
n+1
(y) fK (y),

and so, (4.29) implies that

det D2 F(y) = cK . (4.35)
Now, since y ∈ Sn−1 was arbitrary, (4.35) holds for all y ∈ Sn−1 .
Also, since F is
2-homogeneous, its Hessian is 0-homogenous and so (4.35) is actually valid for all
y ∈ Rn . In conclusion, we have proved that

det D2 F = cK in Rn .
1/n
So, Theorem 4.18 applied to the function F/cK implies that F is a quadratic
polynomial, and since F(0) = |∇F(0)| = 0 we get that F(y) = 12 hAy, yi for some
symmetric positive-definite matrix A.
Recalling (4.31) and (4.32), this implies that
p
hK (y) = hAy, yi, y ∈ Sn−1 .

Also, it follows by the definition of hK (see (4.30)) that

K = x ∈ Rn : hx, yi ≤ hK (y) ∀ y ∈ Sn−1 .



Therefore,
n p o 
K = x ∈ Rn : hx, yi ≤ hAy, yi ∀ y ∈ Sn−1 = x ∈ Rn : hA−1 x, xi ≤ 1 ,
 4.5 Interior C 1,α estimates 93

so, K is an ellipsoid. This proves the result under the assumption that K is C 2 and
uniformly convex.
Step 2: The general case. If K is just a convex set, we can approximate it with
a sequence Kε of uniformly convex sets of class C 2 . Now, following the notation
introduced in the previous step, the computations above prove that
n+1 y
   y 
det D2 Fε (y) = hK ε
fK ε ,
|y| |y|

where Fε (y) = 12 |y| 2 fKε (y/|y|)2 . Then it is not difficult to show that

hKε → hK and fKε → fK , uniformly on Sn−1 .

Also, it follows by Proposition 2.6 that the measures det D2 Fε dy converge weakly∗
to the Monge–Ampère measure of the convex function F associated to K. Hence,
using (4.29), we deduce that
µF = cK dy in Rn,
and Theorem 4.18 allows us to conclude as before. 

4.5 Interior C 1,α estimates

In [15], Caffarelli proved the interior C 1,α regularity of strictly convex solutions of
Monge–Ampère with right-hand side bounded away from zero and infinity. As we
shall see in Section 4.11, these assumptions are optimal.
This regularity result is obtained by suitably iterating Lemma 4.12 (see [57] for
a slightly different approach which avoids any compactness argument). As we shall
show below, it suffices to prove the result for normalized solutions.
Theorem 4.20. Let Z ⊂ Rn be a normalized convex set, and let u : Z → R be a
convex function satisfying
(
λ dx ≤ µu ≤ Λ dx in Z,
u=0 on ∂ Z,

for some 0 < λ ≤ Λ. Let x̂0 be the minimum point of v, set h := max Z |v| = |u( x̂0 )|,
and consider the section
Z1/2 := S( x̂0, 0, h/2).
Then there exist C, α > 0, depending only on n, λ, and Λ, such that
kukC 1, α (Z1/2 ) ≤ C.
94 4 Interior regularity of weak solutions

Before proving this result, we state and prove an immediate corollary.

Corollary 4.21. Let Ω ⊂ Rn be an open set and u : Ω → R be a strictly convex

function satisfying

λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.

1,α
Then u ∈ Cloc (Ω) for some α > 0 depending only on n, λ, and Λ.

Proof. Given x ∈ Ω and p ∈ ∂u(x), it follows by strict convexity of u that we can find
t > 0 small enough so that S := S(x, p, t) ⊂⊂ Ω. Let L be the affine transformation
that normalizes S. We set S ∗ := L(S) and define v as in (4.2).
Applying Theorem 4.20 to v we deduce that v ∈ C 1,α (S1/2
∗ ), which translated back

to u implies that u ∈ C 1,α inside S(x, p, t/2). By the arbitrariness of x, this proves
the result. 

Proof of Theorem 4.20. The argument is very similar to the one used in the proof of
Theorem 4.16.
Fix x ∈ Z1/2 and p ∈ ∂u(x), and let τ > 0 be given by Lemma 4.13 with ρ = 1/2.
Up to subtracting a function from u, we can assume that

u(x) = 0 and p = 0. (4.36)

Fix t ≤ τ, let L be the affine transformation that normalizes S := S(x, 0, t), set
S ∗ := L(S), and define v as in (4.2). Also, let z∗ := L x ∈ S ∗ denote the minimum
point of v, set h := |v(z∗ )|, and define ṽ := v + h. Then it follows by Lemma 4.12 that

h
ṽ(z/2) ≤ (1 − δ) ∀ z ∈ S∗ .
2
Since ṽ|∂S ∗ = h, the above equation yields

1−δ
ṽ(z/2) ≤ v(z) ∀ z ∈ ∂S ∗ .
2
In terms of u, we see that (recall (4.36))

1−δ
u(x + y/2) ≤ u(x + y) ∀ x + y ∈ ∂S(x, 0, t).
2
Noticing that t is an arbitrary constant less than τ, we have proved that

1−δ
u(x + y/2) ≤ u(x + y) ∀ x + y ∈ S(x, 0, τ).
2
 4.6 Application 3: The optimal transport problem with quadratic cost 95

Since S(x, 0, τ) ⊃ B% (x) for some universal radius % > 0 (see Corollary 4.7), itera-
tively applying the estimate above to the sequence of points {2−k y}k ≥0 , we deduce
that  1 − δ k
u(x + 2−k y) ≤ u(x + y) ∀ y ∈ B% (0), k ≥ 0.
2
Also, since u ≤ τ in B% (x), if we choose α > 0 such that 2−α = 1 − δ, we get

u(x + 2−k y) ≤ 2−k(1+α) τ ∀ y ∈ B% (0), k ≥ 0. (4.37)

We now pick y 0 ∈ B% (0), consider k = k(y 0) ≥ 0 such that

2−(k+1) % < |y 0 | ≤ 2−k %,

and apply (4.37) with y := 2k y 0 ∈ B% (0) to get

21+α τ −(k+1)
1+α 21+α τ 0 1+α

u(x + y 0) ≤ 2−k(1+α) τ = 2 % ≤ 1+α |y | ,
%1+α %
or equivalently (recall (4.36)),

21+α τ
u(y) − u(x) − hp, y − xi ≤ |y − x| 1+α ∀ y ∈ B% (x).
%1+α
By the arbitrariness of x and p ∈ ∂u(x), this proves that

u(y) − u(x) − hp, y − xi ≤ Ĉ|y − x| 1+α ∀ y ∈ B% (x) ∩ Z1/2, x ∈ Z1/2, p ∈ ∂u(x),

where Ĉ := 21+α τ/%1+α . Also, applying Lemma 4.8 to S1 = Z and S2 = Z1/2 with
T = 2, we deduce that
B1/R ( x̄) ⊂ Z1/2 ⊂ BR ( x̄)
for some R = R(n, λ, Λ). Hence, Corollary 4.9 yields that u| Z1/2 is uniformly Lips-
chitz. This allows us to apply Lemma A.32 and conclude that u| Z1/2 is universally
C 1,α , as desired. 

4.6 Application 3: The optimal transport problem with quadratic cost

Let µ and ν denote two probability measures on Rn . The optimal transport problem
(with quadratic cost) consists of finding the “optimal” way to transport µ onto ν given
that the transportation cost to move a point from x to y is |x − y| 2 . Hence, one is
naturally led to minimize ∫
|S(x) − x| 2 dµ(x)
Rn
96 4 Interior regularity of weak solutions

among all maps S that “transport µ onto ν”. Mathematically, this corresponds to
saying that S# µ = ν, that is, for any bounded Borel function ϕ : Rn → R,
∫ ∫
ϕ(y) dν(y) = ϕ S(x) dµ(x). (4.38)


Rn Rn
By a classical theorem of Brenier [12] (see also [28, 100], where the same result
was proved independently), existence and uniqueness of optimal maps always hold
provided that µ is absolutely continuous and both µ and ν have finite second mo-
ments. Moreover, such a map is given by the gradient of a convex function. This is
summarized in the next theorem:
Theorem 4.22. Let µ, ν be probability measures on Rn with µ = f dx and
∫ ∫
|x| 2 dµ(x) + |y| 2 dν(y) < ∞.
Rn Rn
Then,
• there exists a µ-a.e. unique optimal transport map T;
• there exists a lower semicontinuous convex function u : Rn → R ∪ {+∞}, with
u differentiable µ-a.e., such that T = ∇u µ-a.e.
We now show the connection between optimal transport and the Monge–Ampère
equation. Assume that both µ and ν are absolutely continuous, that is, µ = f dx
and ν = g dy, and suppose that T is a smooth diffeomorphism transporting µ onto ν.
Then it follows by (4.38) and the standard change of variable formula that
∫ ∫ ∫
ϕ(T(x)) f (x) dx = ϕ(y)g(y) dy = ϕ(T(x))g(T(x)) det∇T(x) dx
Rn Rn Rn
for any ϕ ∈ Cc∞ (Rn ). By the arbitrariness of ϕ, this implies the validity of the
Jacobian equation
f (x) = g(T(x)) det∇T(x)) . (4.39)
Then the condition T = ∇u implies that u solves the Monge–Ampère equation
f
detD2 u = . (4.40)
g ◦ ∇u
Concerning the “boundary conditions”, since T = ∇u was assumed to be a diffeo-
morphism and sends the mass f onto g, we deduce that
∇u({ f > 0}) = {g > 0}.
In particular, the map ∇u sends the support of f onto the support of g.
Notice that the above computations are formal since, in order to write the above
equation, we needed to assume a priori that T is smooth. Still, this fact is the starting
point behind the regularity theory of optimal transport maps.
 4.6 Application 3: The optimal transport problem with quadratic cost 97

4.6.1 Regularity theory: Brenier versus Alexandrov solutions. Let X and Y be

two bounded smooth open sets in Rn , let µ = f dx and ν = g dy be two probability
measures on Rn such that f |Rn \X = 0 and g|Rn \Y = 0. Assume that f and g are
bounded away from zero and infinity on X and Y , respectively, and denote the convex
function provided by Theorem 4.22 by u : Rn → R. As we have seen before,
formally, the transport condition

(∇u)# f dx = g dy (4.41)

can be seen as a weak form of the Monge–Ampère equation (4.40) coupled with the
“boundary condition”
∇u(X) = Y . (4.42)
For this reason, we shall say that if u satisfies (4.41), then u is a Brenier solution of
the Monge–Ampère equation.
In analogy with Definition 2.5, one can also define Alexandrov solutions as such:
we say that a convex function v : X → R is an Alexandrov solution of (4.40) if
∫
f (x)
|∂v(E)| = dx ∀ E ⊂ X (4.43)
E g(∇v(x))

(notice that v is differentiable a.e., so the right-hand side makes sense).

While for Alexandrov solutions one may apply the regularity theory developed in
the previous sections, as observed by Caffarelli in [16], even for smooth densities one
cannot expect any general regularity result for Brenier solutions (equivalently, for
optimal transport maps) without making some geometric assumptions on the support
of the target measure. Indeed, let n =
2 and suppose
that X = B1 is the unit ball
centered at the origin and Y = B1+ + 2e1 ∪ B1− − 2e1 is the union of two half-balls,
where
B1+ := B1 ∩ {x1 > 0} , B1− := B1 ∩ {x1 < 0} ,

1
and (e1, e2 ) denotes the canonical basis of R2 (see Figure 4.9). Then if f = |X | 1 X
and g = |Y1 | 1Y , the optimal transport map is given by
(
x + 2e1 if x1 > 0,
T(x) :=
x − 2e1 if x1 < 0,

which corresponds to the gradient of the convex function u(x) = |x| 2 /2 + 2|x1 |.
Thus, in order to hope for a regularity result for u, we need at least to assume the
connectedness of Y .3 However, starting from the above construction and considering
3This is actually obvious from a topological point of view since the image of a connected set by a continuous
map is still connected.
98 4 Interior regularity of weak solutions

Figure 4.9. On the left, the map T(x) = x ± 2e1 is the optimal transport map from the constant
density on X to the constant density on Y . On the right, if the strip connecting the upper and lower
half-balls has width ε > 0, the optimal transport Tε will still be discontinuous for ε  1.

a sequence of domains Yε where
one adds a small strip of width ε > 0 to glue
together B1+ + 2e1 ∪ B1− − 2e1 (see Figure 4.9), one can also show that for ε > 0


small enough, the optimal map will still be discontinuous (see [16, 43]). Hence,
connectedness is not enough to ensure regularity.
The reason for this lack of regularity is the following. For Alexandrov solutions,
whenever f and g are bounded away from zero and infinity on X and Y , respectively,
(4.43) implies that the multivalued map x 7→ ∂u(x) preserves the Lebesgue measure
up to multiplicative constants; that is, the volumes of E and ∂u(E) are comparable
for any Borel set E ⊂ X.
On the other hand, for Brenier solutions, the optimal map can only see the regions
where f and g live, and in fact, one can show that
∫ ∫
f (x) dx = g(y) dy ∀ E ⊂ X (4.44)
E ∂u(E)

(see, for instance, [122, Lemma 4.6]). Thus, if f and g are bounded away from zero
and infinity on X and Y , respectively, we only have
|E | ' |∂u(E) ∩ Y | ∀E ⊂ X
(and not |E | ' |∂u(E)| as in the Alexandrov case). This means that we do not have
full control on the Monge–Ampère measure of u.
In the counterexample above, with f = |X1 | 1X and g = |Y1 | 1Y , when x = (0, x2 )
we have that ∂u(x) = [−2, 2] × {x2 }. Thus ∂u maps the segment {0} × [−1, 1] onto
the rectangle [−2, 2] × [−1, 1] where g has no mass. Notice that in this case, the
determinant of D2 u is equal to 1 a.e. inside B1 (so u is a Brenier solution with right-
hand side 1), but the Monge–Ampère measure of u is equal to dx B1 + 4H1 ({x1 =
0} ∩ B1 ) (compare with Example 2.2(4)).
Hence, in order to avoid this kind of counterexample, one should ensure that the
target mass always covers the image of ∂u(X) (so that |∂u(E) ∩ Y | = |∂u(E)| for all
E ⊂ X), and this is the case when Y is convex. Indeed, as shown by Caffarelli [16],
if Y is convex, then ∂u(X) ⊂ Y and the following result holds:
 4.6 Application 3: The optimal transport problem with quadratic cost 99

Theorem 4.23. Let X, Y ⊂ Rn be two bounded open sets, let f , g : Rn → R+ be two

probability densities, that are zero outside X, Y and are bounded away from zero and
infinity on X, Y , respectively. Denote by T = ∇u : X → Y the optimal transport map
provided by Theorem 4.22, and assume that Y is convex. Then there exists α > 0
0,α
such that T ∈ Cloc (X).
Proof. It will be useful to consider

û(z) := sup u(x) + hp, z − xi ∀ z ∈ Rn . (4.45)

x ∈X, p ∈∂u(x)

Notice that û agrees with u inside X. (Actually, û is the smallest convex function that
does so.) Moreover, ∂ û = ∂u inside X (thanks to Remark A.18).
Our goal is to show that u (or equivalently û) is a strictly convex Alexandrov
solution of the Monge–Ampère equation inside X, so that we can apply Corollary 4.21
1,α
and deduce that u ∈ Cloc (X). We split the argument into a few steps.
Step 1: û is an Alexandrov solution inside X. As (∇û)# f dx = (∇u)# f dx = g dy and
f and g are supported and bounded away from zero inside X and Y , respectively, we
deduce that ∇û(x) ∈ Y for a.e. x ∈ X, and that for a.e. y ∈ Y there exists x ∈ X such
that ∇û is differentiable at x and y = ∇u(x). Since |∂Y | = 0 (since Y is convex), this
implies that
|Y \ ∂ û(X)| = 0. (4.46)
In addition, again, since Y is convex, it follows by the continuity of the subdifferential
(see the proof of Lemma A.22) and Corollary A.27 that

∂ û(X) = Y . (4.47)

Once again, using that f and g are bounded away from zero and infinity on X and Y ,
respectively, and combining (4.47) and (4.44), we deduce that

λ dx ≤ µû ≤ Λ dx in X, (4.48)

for some 0 < λ ≤ Λ; hence, û is an Alexandrov solution inside X.

Step 2: ∂ û(Rn ) = Y . In the previous step, we showed that û is an Alexandrov solution
of the Monge–Ampère equation inside X. As we shall see in Step 4, to be able to
prove the strict convexity of û inside X, we first need to understand what equations û
satisfies outside X, and for this we need to look at the subdifferential of û there.
We begin by noticing that since ∂u = ∂ û inside X, it follows by (4.47) that
∂u(X) ⊂ Y . In particular, since Y is bounded, there exists R > 0 such that Y ⊂ BR (0).
This implies that each function

z 7→ u(x) + hp, z − xi, x ∈ X, p ∈ ∂u(x),

100 4 Interior regularity of weak solutions

is R-Lipschitz; hence, so is û (being the supremum of a family of R-Lipschitz

functions; see (4.45)).4 Thus, since u = û inside X, we can rewrite (4.45) as
û(z) := max û(x) + hp, z − xi ∀ z ∈ Rn . (4.49)
x ∈X, p ∈∂û(x)

Now, fix ẑ ∈ Rn a differentiability point for û, and consider x̂ ∈ X and p̂ ∈ ∂ û( x̂) ⊂ Y
that attain the maximum in (4.49). Thus, the affine function z 7→ û( x̂) + h p̂, z − x̂i
touches the graph of u at ẑ from below, and so p̂ ∈ ∂ û( ẑ). Recalling that ẑ is a
differentiability point, it follows by Lemma A.20 that ∇u( ẑ) = p̂ ∈ Y . Hence, if we
denote the set of differentiability points of û in Rn by Dû , Corollary A.27 implies
that ∂ û(Rn ) ⊂ Y . Together with (4.47), this concludes the proof of Step 2.
Step 3: A Monge–Ampère equation for û in Rn . Thanks to (4.46), (4.47), and Step 2,
we know that there exists a set Y 0 ⊂ Y such that
∂ û(X) = Y 0, |Y \ Y 0 | = 0, and ∂ û(Rn ) = Y .
Therefore,
∂ û(Rn \ X) = ∂ û(Rn \ X) ∩ Y ⊂ ∂ û(Rn \ X) ∩ ∂ û(X) ∪ Y \ Y 0 .

Since the set ∂ û(Rn \ X) ∩ ∂ û(X) has zero measure (by Lemma A.30), we deduce
that µû (Rn \ X) = 0. Finally, using (4.48) yields
µû = h dx in Rn, where λ1X ≤ h ≤ Λ1X . (4.50)

Step 4: û is strictly convex inside X. Assume, to the contrary, that û is not strictly
convex inside X. Then there exists a supporting plane `(z) := û(x) + hp, z − xi for
some x ∈ X and p ∈ ∂u(x) such that the convex set Σ := {û = `} is not a singleton.
We now consider the following four cases, and we show that none of them can occur.
• Σ has no exposed points in Rn . If so, by Theorem A.10, we would have that û
is affine on an infinite line. Then Lemma A.25 yields that ∂ û(Rn ) is contained
in a hyperplane, contradicting Step 2.
• Σ has an exposed point inside X. This is excluded by Theorem 4.10.
• Σ has an exposed point inside Rn \ X. Let x̄ ∈ Rn \ X be an exposed point.
Then up to subtracting an affine function and a change of coordinates, we can
assume
` ≡ 0, Σ ∩ {z1 ≥ 0} ⊂ Rn \ X, Σ ⊂ {z : z1 ≤ a}, and x̄ = ae1
4 Notice that, a priori, if û is a supremum of R-Lipschitz functions, this does not exclude that û ≡ +∞.
However, this is excluded since û | X = u and u is differentiable a.e. inside X (by Theorem 4.22), therefore it
must be finite there.
 4.6 Application 3: The optimal transport problem with quadratic cost 101

for some a > 0 (see Figure 4.3). Set Sε := {û < εz1 }. Since Sε → Σ∩{x1 ≥ 0}
in the Hausdorff distance as ε → 0, we deduce that for ε > 0 small enough,
the sets Sε are bounded subsets of Rn \ X. So, (4.50) yields

µû (Sε ) = 0 for ε  1.

On the other hand, if we define ûε := û − εz1, we see that minSε ûε < 0.
Consequently, by Theorem 2.8, we see that

µû (Sε ) = |∂ û(Sε )| = |∂ ûε (Sε )| > 0,

a contradiction.

• Σ has an exposed point on ∂ X. Up to subtracting an affine function and a

change of coordinates, we can assume that the origin is our exposed point. In
particular,

` ≡ 0, Σ ⊂ {z1 ≤ 0}, Σ ∩ {z1 = 0} = 0, and 0 ∈ ∂ X.

In this case, we have an additional issue: the Monge–Ampère measure may

vanish near 0 (as h = 0 outside X), so we need to be careful in order to capture
some Monge–Ampère measure from the interior of X.
To this aim, let R0 > 0 be sufficiently large so that X ⊂ BR0 (0), and fix
x0 ∈ Σ ∩ X and δ > 0 such that Bδ (x0 ) ⊂ X (such a δ exists because X is
open). To make sure that we capture the Monge–Ampère measure around x0 ,
we consider

ûε (z) := û(z) − ε(z1 + 2R0 ) and Sε := {ûε < 0}

(notice that with this definition, x0 ∈ Sε ), and we use John’s map Lε to

normalize Sε into Sε∗ and get a normalized solution v̂ε inside Sε∗ .
We now observe that since Sε → Σ ∩ {z1 ≥ −2R0 } as ε → 0 and the latter set
is compact, for ε  1 the sets Sε are all contained inside a ball BR1 (0). Hence,
since Lε (Sε ) ⊃ B1 (0), it follows that Lε−1 (B1 (0)) ⊂ BR1 (0). This implies that
the maps Lε−1 are R1 -Lipschitz. Equivalently,

1
|Lε (x) − Lε (z)| ≥ |x − z| ∀ x, z ∈ Rn,
R1
which yields

X ε∗ := Lε (X) ⊃ Lε Bδ (x0 ) ⊃ Bδ/R1 (Lε x0 ).

102 4 Interior regularity of weak solutions

Consider now the convex set Sε,δ ∗ := {z ∈ Sε∗ : dist(z, ∂Sε∗ ) ≥ δ/(2R1 )}.
Since Sε is a normalized convex set containing Lε x0 and the Monge–Ampère
∗

measure of v̂ε is bounded from below by λ inside X ε∗ , we deduce that

 n
δ
∗
|∂ v̂ε (Sε,δ )| = µv̂ε (Sε,δ
∗
) ≥ λ|Sε,δ
∗
∩ Bδ/R1 (Lε x0 )| ≥ λ ĉn
R1

for some dimensional constant ĉn > 0. This implies that ∂ v̂ε (Sε,δ
∗ ) cannot be

contained inside the ball Bc0 δ/R1 (0) if c is too small; hence,
0

δ
sup |p| ≥ c 0 ,
∗ )
p ∈∂v̂ε (Sε, δ
R1

and Corollary A.23 implies that

2
δ

min v̂ε ≥ c 00
.
∗
Sε R1

On the other hand, since µv̂ε (Sε∗ ) ≤ Λ|Bn (0)| (recall that µv̂ε ≤ Λ dx and Sε∗ is
normalized), Theorem 2.8 implies that

| v̂ε (Lε 0)| ≤ C dist(Lε 0, ∂Sε∗ )1/n .

However, an identical argument to the one in the proof of Theorem 4.10 shows
that
v̂ε (Lε 0)
dist(Lε 0, ∂Sε∗ ) → 0 and → 1 as ε → 0,
minSε∗ v̂ε
a contradiction.
This proves that Σ is a singleton, and û is strictly convex inside X.
Step 5: Conclusion. Recalling that û|X = u, (4.48) and Step 4 allow us to apply
1,α 0,α
Corollary 4.21 and deduce that u ∈ Cloc (X), or equivalently T = ∇u ∈ Cloc (X), as
desired. 
A natural question concerns what happens when one removes the convexity
assumption on the target. As shown in [47] (see also [43] for a more precise
description of the singular set in two dimensions), in this case, one can prove that
the optimal transport map is smooth outside a closed set of measure zero. More
precisely, the following holds:
Theorem 4.24. Let X, Y ⊂ Rn be two bounded open sets, let f , g : Rn → R+ be two
probability densities, that are zero outside X, Y and are bounded away from zero and
infinity on X, Y , respectively. Denote by T = ∇u : X → Y the optimal transport map
 4.6 Application 3: The optimal transport problem with quadratic cost 103

provided by Theorem 4.22. Then there exist two relatively closed sets ΣX ⊂ X and
ΣY ⊂ Y with |ΣX | = |ΣY | = 0 such that T : X \ ΣX → Y \ ΣY is a homeomorphism of
0,α
class Cloc for some α > 0.

Sketch of the proof. As explained before, when Y is not convex we might find points
x ∈ X such that ∂u(x) * Y , and at these points we have no control on the Monge–
Ampère measure of u. For this reason, let us define

RegX := {x ∈ X : ∂u(x) ⊂ Y } and ΣX := X \ RegX .

By the continuity of the subdifferential, one can see that RegX is open. Moreover,
since g gives no mass to Rn \ Y , it follows by (4.47) that |ΣX | = 0. Then a
modification of the argument described in the proof of Theorem 4.23 allows one to
show that u is a strictly convex Alexandrov solution on RegX , and the result follows
from Corollary 4.21 applied with Ω = RegX . 

We conclude this section with the following observation that we shall use in the
next section.
Remark 4.25. Under the assumptions of the previous results, if one assumes in
k,α k,α
addition that f ∈ Cloc (X) and g ∈ Cloc (Y ), then Theorem 4.42 and Remark 4.44
k+1,α
imply that T ∈ Cloc . Furthermore, if both X and Y are smooth and uniformly
convex, one can prove that T : X → Y is a smooth diffeomorphism [17, 19, 121].

4.6.2 An application of optimal transport: The isoperimetric inequality. The

classical isoperimetric inequality in Rn is an important tool in analysis and geometry.
It relates the volumes and the perimeters of sets. More precisely, if X ⊂ Rn is a
smooth bounded set, then the perimeter Hn−1 (∂ X) (see Appendix A.2) controls the
volume |X | via the formula

Hn−1 (∂ X) ≥ n|B1 (0)| 1/n |X | (n−1)/n . (4.51)

Notice that for any λ > 0, the perimeter scales as Hn−1 ∂(λX) = λ n−1 Hn−1 (∂ X)


(see Proposition A.6(i)) while |λX | = λ n |X |; hence, (4.51) is both translation and
scale invariant.
Following an argument due to Knothe [74] and later rediscovered by Gromov
[89], we now give a short elementary proof of (4.51) using optimal transport. Also,
as a consequence of this proof, we will be able to characterize the equality cases.

Theorem 4.26. Let X ⊂ Rn be a smooth bounded set. Then (4.51) holds. In addition,
X attains the equality in (4.51) if and only if X is a ball.
104 4 Interior regularity of weak solutions

Figure 4.10. The optimal transport map sending the constant density inside X onto the constant
density inside B1 (0).

Proof. Let us define the probability densities

1 1
f := 1X and g := 1B (0), (4.52)
|X | |B1 (0)| 1
and let T = ∇u : X → B1 (0) be the optimal transport map provided by Theorem 4.22
(see Figure 4.10). Since B1 is convex and f and g are smooth and positive inside X
and B1 , respectively, it follows by Theorem 4.23 and Remark 4.25 that T is of class
C ∞ inside X.
We now observe that T satisfies the following three properties inside X:
(a) |T | ≤ 1;
(b) det ∇T = |B1 (0)|/|X |;
(c) div T ≥ n(det ∇T)1/n .
Indeed, (a) follows as T = ∇u takes values inside B1 (0).
For (b), we notice that since T = ∇u is smooth, (4.44) combined with (4.52),
Theorem A.31, and the fact that ∇u(E) ⊂ B1 (0) yields
1
∫
|E |
= det D2 u dx ∀ E ⊂ X,
|X | |B1 (0)| E
and the arbitrariness of E proves that det ∇T = det D2 u = |B1 (0)|/|X |.
Finally, to prove (c), given a point x ∈ X, we choose a system of coordinates so
that ∇T(x) = D2 u(x) is a diagonal matrix with eigenvalues λ1, . . . , λn . Since u is
convex, we notice that λi ≥ 0. Thus, the arithmetic–geometric inequality gives
n n  1/n
1
 Õ  Ö
div T(x) = ∆u(x) = n λi ≥ n λi
n i=1 i=1

1/n
1/n
= n det D2 u(x) = n det ∇T(x) , (4.53)
as desired.
 4.6 Application 3: The optimal transport problem with quadratic cost 105

We are now ready to prove (4.51), and characterize the equality cases.
Step 1: Proof of (4.51). Consider X ε ⊂⊂ X a family of smooth sets converging to X
as ε → 0 and satisfying

Hn−1 (∂ X ε ) → Hn−1 (∂ X) and |X ε | → |X | as ε → 0. (4.54)

Since T is smooth inside X, we have that T ∈ C ∞ (X ε ). Hence, if we denote the outer

unit normal to ∂ X ε by νXε , it follows by the divergence theorem and properties (a),
(b), and (c) above that
∫ ∫ ∫
Hn−1 (∂ X ε ) ≥ |T | d Hn−1 ≥ T · νXε d Hn−1 = div T dx
∂Xε ∂Xε Xε
∫   1/n
|B1 (0)|
≥n (det ∇T)1/n dx = n|X ε | ,
Xε |X |
(4.55)

and (4.51) follows by (4.54), letting ε → 0.

Step 2: Balls attain the equality case. Let X be a ball. Since (4.51) is translation and
scale invariant, up to a translation and a dilation, we can assume that X = B1 (0). In
this case, f = g and the optimal transport map is simply given by the identity map
T(x) = x (as the cheapest way to transport a density onto itself is by not moving
anything). In particular, since T ∈ C ∞ (Rn ), we can repeat the proof above directly
with X ε = X = B1 (0) and it is immediate to check that all inequalities in (4.55) are
equalities, proving the result.
Step 3: Characterization of the equality case. Assume that X is a set attaining
equality in (4.51).
We first show that X is connected. Indeed, if X = X1 ∪ X2 with X 1 ∩ X 2 = ∅,
then

Hn−1 (∂ X) = Hn−1 (∂ X1 ) + Hn−1 (∂ X2 ) and |X | = |X1 | + |X2 |,

so it follows by (4.51) applied both to X1 and X2 that

  (n−1)/n
n|B1 (0)| |X1 | + |X2 | = n|B1 (0)| |X | (n−1)/n
= Hn−1 (∂ X) = Hn−1 (∂ X1 ) + Hn−1 (∂ X2 )
 
≥ n|B1 (0)| |X1 | (n−1)/n + |X2 | (n−1)/n ,

which is possible only if either |X1 | = 0 or |X2 | = 0, proving that X must be

connected.
106 4 Interior regularity of weak solutions

We now go back to (4.55). Since X attains the equality in (4.51), recalling (4.55)
and (4.54), we get
∫  
lim div T − n(det ∇T)1/n dx = 0,
ε→0 Xε

so (c) implies that div T = n(det ∇T)1/n inside X. Since equality in the arithmetic–
geometric inequality holds only when all the numbers are equal, we deduce from
(4.53) that for any x ∈ X, the eigenvalues λi (x) of D2 u(x) satisfy

λ1 (x) = · · · = λn (x).

Combining this with the equality i=1 λi (x) = det ∇T(x) = |B1 (0)|/|X | (see (b)) and
În
1/n
setting η := (|B1 (0)|/|X |) , we deduce that

λi (x) = η ∀ i = 1, . . . , n, ∀ x ∈ X,

that is,
∇(T − ηx) = ∇T(x) − η Id = 0 ∀ x ∈ X.
Since X is connected, this implies that T − ηx is constant on X; hence, there exists
x0 ∈ Rn such that T(x) = η(x − x0 ). Recalling that T maps X onto B1 (0), this yields

X = T −1 B1 (0) = B1/η (x0 ),

and X is a ball, as desired. 

It is worth pointing out that this proof is very robust. Indeed, it can be used
to prove a much more general class of isoperimetric-type inequalities where one
replaces the classical perimeter by an anisotropic version. Second, this proof is the
starting point to show the following stability statement: If a set X almost attains the
equality in (4.51), then it is quantitatively close to a ball. For more details on these
results, we refer to [51].

4.7 Geometry of sections

We want to study the geometry of the sections of a solution u. Notice that thanks
to Corollary 4.21, strictly convex functions with Monge–Ampère measure bounded
away from zero and infinity are locally C 1,α . In particular, the subdifferential of u at
a point x reduces to ∇u(x); hence, we can use the compact notation

Sh (x) := S(x, ∇u(x), h) ∀ h > 0.

 4.7 Geometry of sections 107

If we think of Sh (x) as a “ball of radius h centered at x”, our goal is to show that
sections satisfy the basic properties that balls in “nice” metric spaces enjoy. Indeed,
as we shall see, they are suitable for covering lemmas. This nice geometric structure
of sections was first discovered in [21], and further investigated in [62].

4.7.1 Sections behave like balls. The next proposition summarizes some of the
basic geometric properties of sections. These will be crucial later to show interior
W 2,1+γ and W 2, p estimates (see Section 4.8).
Proposition 4.27. Let Ω ⊂ Rn be an open set, and u : Ω → R a strictly convex
function satisfying
λ dx ≤ µu ≤ Λ dx in Ω, for some 0 < λ ≤ Λ.
There exists η ∈ (0, 1) universal such that the following hold:
(a) Assume that Sh2 (x2 ) ⊂⊂ Ω. Then for any h1 ≤ h2 and x1 ∈ Ω,
Sηh1 (x1 ) ∩ Sηh2 (x2 ) , ∅ ⇒ Sηh1 (x1 ) ⊂ Sh2 (x2 ).

(b) Assume that Sh1 (x1 ) ⊂⊂ Ω. Then for any h2 ≥ h1 and x2 ∈ Ω such that
x1 ∈ Sh2 (x2 ), there exists a point z ∈ Ω such that
Sηh1 (z) ⊂ Sh1 (x1 ) ∩ Sh2 (x2 ).

(c) Assume that S2h (x2 ) ⊂⊂ Ω. Then

Sηh (x1 ) ⊂ S2h (x2 ) ∀ x1 ∈ Sh (x2 ).

Proof. The basic observation is that all these statements are invariant under affine
transformations. Hence, if we fix a section S and consider L the affine transformation
that normalizes it, we only need to prove the result with v in place of u (v as in (4.2))
and with all the sections replaced by their images under L. Thanks to this invariance,
when proving each of the three statements, we can assume that one of the sections is
normalized and that our solution vanishes on the boundary of that particular section.
Step 1: Proof of (a). As explained above, with no loss of generality, we can assume
that Sh2 (x2 ) is a normalized section and u|∂Sh2 (x2 ) = 0. Since x2 is the minimum point
of u, applying Lemma 4.14 with Z = Sh2 (x2 ), we deduce that
Sθ (x1 ) ∩ Sθ (x2 ) , ∅ ⇒ x1 ∈ Sh2 /2 (x2 ).
Also, Lemma 4.13 applied with ρ = 1/2 implies that there exists τ > 0 universal
such that
Sτ (x1 ) ⊂ Sh2 (x2 ) ∀ x1 ∈ Sh2 /2 (x2 ).
108 4 Interior regularity of weak solutions

Figure 4.11. A description of Proposition 4.27(a): If two sections intersect, dilating the “larger” one
(i.e., the one of larger height) by a factor 1/η we can encompass the other one.

Since h1 ≤ h2 ≤ C0 (see Proposition 4.4), taking

θ τ
 
η := min ,
C0 C0

proves the result.

Step 2: Proof of (b). In this case we assume that Sh1 (x1 ) is a normalized section and
u|∂Sh1 (x1 ) = 0. Noticing that x1 is the minimum point of u, we split the proof into
two cases, depending on whether or not x2 is close to x1 . To this aim, we let ρ > 0
small be chosen later.
Step 2-a: x2 ∈ Sρ (x1 ). Since Sρ (x1 ) → {x1 } in the Hausdorff distance as ρ → 0, a
simple compactness argument shows that there exists ρ > 0 universal such that

x2 ∈ Sρ (x1 ) ⇒ Sh1 (x2 ) ⊃ Sh1 /2 (x1 )

(compare with the arguments used in Section 4.2.2). Since h2 ≥ h1 , this implies (b)
with z = x1 and η = 1/2.
Step 2-b: x2 < Sρ (x1 ). Given ρ as in Step 2-a, it is easy to show by compactness that
there exists a universal constant σ > 0 such that

|∇u(y)| ≥ σ ∀ y ∈ Ω \ Sρ (x1 );

hence, in particular,
|∇u(x2 )| ≥ σ. (4.56)
Also, since by assumption x1 ∈ Sh2 (x2 ), we have

u(x1 ) ≤ u(x2 ) + h∇u(x2 ), x1 − x2 i + h2 . (4.57)

 4.7 Geometry of sections 109

Figure 4.12. A description of Proposition 4.27(b): If the point x1 belongs to Sh2 (x2 ), for any
h1 ≤ h2 , the intersection Sh1 (x1 ) ∩ Sh2 (x2 ) contains a section of height ηh1 .

Consider now the unit vector ŷ := ∇u(x2 )/|∇u(x2 )|. Then thanks to (4.56), we see
that
D E
− ∇u(x2 ), ŷ + 12 w ≤ −|∇u(x2 )| + 12 |∇u(x2 )| ≤ − σ2 ∀ |w| ≤ 1. (4.58)

In particular, since ∇u(x1 ) = 0 (recall that x1 is the minimum point of u), applying
Theorem 4.20, (4.57), and (4.58), for any r > 0, we get
 
u x1 + r ŷ + r2 w ≤ u(x1 ) + Cr 1+α
≤ u(x2 ) + h∇u(x2 ), x1 − x2 i + h2 + Cr 1+α
≤ u(x2 ) + ∇u(x2 ), x1 + r ŷ + r2 w − x2 + h2 + Cr 1+α − σ
2r

for all w with |w| ≤ 1. Thus, letting r0 be small enough so that Cr01+α < σ2 r0 , we
deduce that
u x1 + r0 ŷ + r2 w < u(x2 ) + ∇u(x2 ), x1 + r0 ŷ + r2 w − x2 + h2 ∀ |w| ≤ 1,

which proves that

Br0 /2 (z) ⊂ Sh2 (x2 ), z := x1 + r0 ŷ.
In addition, it follows by Theorem 4.16 that
u(y) ≥ u(z) + h∇u(z), y − zi + c(r0 /2) M ∀ y < Br0 /2 (z),
or equivalently,
Br/2 (z) ⊃ St (z) provided c(r0 /2) M ≥ t.
Hence, since h1 ≤ C0 (see Proposition 4.4), this proves that
cr0M
Sηh1 (z) ⊂ Sh2 (x2 ) for η := .
2 M C0
110 4 Interior regularity of weak solutions

Figure 4.13. A description of Proposition 4.27(c): Given any point x1 inside the section Sh (x2 ), the
section Sηh (x1 ) is contained inside S2h (x2 ).

Step 3: Proof of (c). Assume that S2h (x2 ) is a normalized section and u|∂S2h (x2 ) = 0.
Since x2 is the minimum point of u, we can apply Lemma 4.13 to Z = S2h (x2 ) and
ρ = 1/2 to deduce that
Sτ (x1 ) ⊂ S2h (x2 ),
and as 2h ≤ C0 (see Proposition 4.4), we obtain the result with η := 2τ/C0 . 

4.7.2 A Vitali covering lemma with sections. As a consequence of Proposi-

tion 4.27(a), we can show that sections are suitable for covering lemmas.

Lemma 4.28. Let Ω ⊂ Rn be an open set, and u : Ω → R a strictly convex function

satisfying
λ dx ≤ µu ≤ Λ dx in Ω for some0 < λ ≤ Λ.
Let D be a compact set in Ω and assume that to each x ∈ D we can associate a
section Sh x (x) such that S2tx ⊂⊂ Ω. Then we can find a finite number of these
sections {Sh x i (xi )}i=1,...,m such that

m
Ø
D⊂ Sh x i (xi ).
i=1

Furthermore, the sections {Sηh x i (xi )}i=1,...,m are disjoint, and η is as in Proposi-
tion 4.27.

Proof. The proof is identical to the proof of the classical Vitali covering lemma. For
completeness, we provide the details.
By compactness of D, we select a finite number of sections {Sηh x j (x j )} j ∈J1 that
cover D. Set
h̄1 := max h x j ,
j ∈J1
 4.8 Interior W 2, p estimates 111

and (up to renaming the points) assume that h x1 = h̄1 . Remove from the family
{Sηh x j (x j )} j ∈J1 all the sections that intersect Sηh x1 (x1 ), and call the set of indices
corresponding to this new family J2 ⊂ J1 . Then set

h̄2 := max h x j
j ∈J2

and (up to renaming the points) assume that h x2 = h̄2 . Again, remove from
{Sηh x j (x j )} j ∈J2 all the sections that intersect Sηh x2 (x2 ), and call this new set of
indices J3 . Iterating this procedure, in finitely many steps (say, m steps), the set Jm is
empty and the sections {Sηh x i (xi )}1≤i ≤m which we have constructed in this way are
disjoint. Hence, we only need to show that this family satisfies the desired covering
property.
For this we pick j ∈ J1 an arbitrary index and we notice that by construction,
there exists i ∈ {1, . . . , m} such that

Sηh x j (x j ) ∩ Sηh x i (xi ) , ∅

(indeed, either the section Sηh x j (x j ) belongs to the family {Sηh x i (xi )}1≤i ≤m , and we
pick xi = x j , or the section Sηh x j (x j ) was discarded along the procedure described
above, which means that it had to intersect one of the sections {Sηh x i (xi )}1≤i ≤m ).
Thanks to Proposition 4.27(a), this implies that

Sηh x j (x j ) ⊂ Sh x i (xi ),

which proves that

Ø m
Ø
Sηh x j (x j ) ⊂ Sh x i (xi ).
j ∈J1 i=1

Since the sections {Sηh x j (x j )} j ∈J1 covered D, so do the sections {Sh x i (xi )}1≤i ≤m ,
which concludes the proof. 

4.8 Interior W 2, p estimates

Our aim is to prove two different results. First, we show that if u is a strictly convex
solution of
λ dx ≤ µu ≤ Λ dx
for some 0 < λ ≤ Λ, then u ∈ W 2,1+γ for some γ = γ(n, λ, Λ) > 0. Second, we shall
prove that this result improves when λ and Λ are close: more precisely, if we assume
that
(1 − ζ) dx ≤ µu ≤ (1 + ζ) dx (4.59)
112 4 Interior regularity of weak solutions

for some small ζ > 0, then u ∈ W 2,σ with σ ≈ log(1/ζ) (in particular, σ → ∞ as
ζ → 0). The latter result was proved in [14], while the W 2,1+γ estimates have been
recently obtained in [30, 36, 106].
As already observed in [44], both results can be obtained with a single approach,
the only difference being that, to prove W 2,σ estimates, one needs an additional
lemma stating that solutions to (4.59) have bounded Hessian in a large set. Hence,
we shall first prove this technical lemma, and then we will show how to obtain both
results at once.
Since our solutions are C 1,α (see Corollary 4.21), as in the previous section, we
will use the notation
Sh (x) := S(x, ∇u(x), h).
Also, as we shall explain later, it suffices to prove a priori bounds on C 2 solutions.
So, we prove all our technical lemmas under the additional assumption that u is C 2 .

4.8.1 A preliminary estimate for solutions with pinched right-hand side. Here,
we prove that if the Monge–Ampère measure of u is close √to 1 up to a factor ζ, then
the set where D2 u can be large has measure controlled by ζ.
Lemma 4.29. Let Z be an open convex set satisfying

Br ( x̄) ⊂ Z ⊂ BR ( x̄) (4.60)

for some 0 < r ≤ R and x̄ ∈ Rn , and let u : Z → R be a convex function of class C 2

satisfying (
1 − ζ ≤ det D2 u ≤ 1 + ζ in Z,
(4.61)
u=0 on ∂ Z,
for some ζ ∈ (0, 1/4). Let x0 be the minimum point of u, set h := max Z |u| = |u(x0 )|,
and consider the section
Z1/2 := Sh/2 (x0 ).
Then there exist K, ζ0 > 0, depending only on n, r, and R, such that

Z1/2 ∩ {kD2 uk ≥ K } ≤ K ζ
p

provided ζ ≤ ζ0 .
Proof. Before starting the proof we notice that since Z1/2 ⊂ BR ( x̄) and |Z1/2 | ≥ c̄ > 0
(by Lemma 4.6), Lemma A.14 implies the existence of a radius ρ = ρ(R, n) > 0 such
that
Bρ ( x̂) ⊂ Z1/2 ⊂ BR ( x̄) ⊂ B2R ( x̂) for some x̂ ∈ Z1/2 . (4.62)
We split the proof into six steps.
 4.8 Interior W 2, p estimates 113

Step 1: Construction of a smooth comparison function. Let w solve

(
det D2 w = 1 in Z,
(4.63)
w=0 on ∂ Z.

It follows by Corollary 4.11 (with ϕ ≡ 0) and Theorem 3.10 that w ∈ C ∞ (Z). Also,
if we set Z3/4 := S3h/4 (x0 ) = {u < −h/4}, since

dist(Z3/4, ∂ Z) ≥ c̃ > 0

with c̃ = c̃(n, r, R) > 0 (see Corollary 4.5), we deduce from Theorem 3.9 and
Remark 1.1 that
1
Id ≤ D2 w ≤ C 0 Id in Z3/4, (4.64)
C0
where C 0 = C 0(n, r, R). Furthermore, since u ∈ C 2 (Z) and |∇u| , 0 on ∂ Z3/4 (since
by convexity, the gradient of u vanishes only at its minimum point), it follows by the
implicit function theorem that

∂ Z3/4 = {u = −h/4} is of class C 2 . (4.65)

Step 2: Closeness of u and w, and construction of a suitable convex envelope. Thanks

to (4.61) and (4.63) we have

det (1 + ζ)D2 w = (1 + ζ)n > det D2 u > (1 − ζ)n = det (1 − ζ)D2 w ;

hence, Theorem 2.10 implies that

(1 + ζ)w < u < (1 − ζ)w in Z,

or equivalently,
1
+ ζ w < u − 12 w < 1
−ζ w in Z.



2 2
Thus, since w is convex, we deduce that
1 1
2 +ζ w < Γ < −ζ w in Z3/4, (4.66)



2

where Γ denotes the convex envelope of u − 21 w in Z3/4 , that is,

Γ(x) := sup `(x) : ` : Rn → R is affine, ` ≤ u − 12 w in Z3/4 ∀ x ∈ Z3/4 .



Notice that since u, w ∈ C 2 (Z3/4 ) and ∂ Z3/4 is of class C 2 (see (4.65)), it follows by
1,1
Proposition A.34 that Γ ∈ Cloc (Z3/4 ).
114 4 Interior regularity of weak solutions

Figure 4.14. Since w is uniformly convex, we have the inclusions Uζ (x) ⊂ Wζ (x) ⊂ BĈ √ζ (x) for
some universal constant Ĉ.

Step 3: The images of ∇Γ and ∇w are close. Given x ∈ Z1/2 , consider the affine
function
`x,w (z) := 21 − ζ w(x) + h∇w(x), z − xi ,

 

and define the open set Uζ (x) := {Γ < `x,w } (see Figure 4.14). Notice that
Uζ (x) ⊂ Wζ (x) := 12 + ζ w ≤ `x,w ,


and because w is uniformly convex (see (4.64)) and recalling the definition of `x,w ,
we see that
Wζ (x) ∩ Z3/4 ⊂ z ∈ Z3/4 : 12 + ζ w(x) + h∇w(x), z − xi + 2c1 0 |z − x| 2

 

≤ 21 − ζ w(x) + h∇w(x), z − xi

 

⊂ z ∈ Z3/4 : 12 + ζ 2c1 0 |z − x| 2 ≤ 2ζ |w(x)| + |∇w(x)| |z − x| .


 

Observe now that (4.62) and Corollary 4.9 imply that both w(x) and |∇w(x)| are
bounded by a constant depending only on n, r, and R. Hence, since ζ ≥ 0, the set
above is contained inside
z ∈ Z3/4 : |z − x| 2 ≤ Ĉζ ⊂ BĈ √ζ (x)


for some constant Ĉ = Ĉ(n, r, R), which proves that

Uζ (x) ∩ Z3/4 ⊂ BĈ √ζ (x) ∀ x ∈ Z1/2 .
We now note that dist(Z1/2, ∂ Z3/4 ) ≥ c0 > 0 (by Corollary 4.5). Hence, the above
inclusion proves that
 2
c0
Uζ (x) ⊂ BĈ ζ (x) ∀ x ∈ Z1/2, provided ζ ≤ ζ0 :=
√ . (4.67)
Ĉ
 4.8 Interior W 2, p estimates 115

In addition, since `x,w ≥ Γ in Uζ (x) and Uζ (x) ⊂ Z3/4 ⊂ Z, we have that Γ = `x,w
on ∂ Uζ (x). Therefore, Lemma 2.7 implies that

∇ 12 − ζ w (x) = ∇`x,w (x) ∈ ∇Γ Uζ (x) . (4.68)




Thus, if we define
n p o
Z1/2,Ĉ √ζ := z ∈ Z1/2 : dist(z, ∂ Z1/2 ) ≥ Ĉ ζ , (4.69)

it follows by (4.67) and (4.68) that

1
− ζ w Z1/2,Ĉ √ζ (4.70)



∇Γ(Z1/2 ) ⊃ ∇ 2

provided ζ ≤ ζ0 .
Step 4: Estimating |∇Γ(Z1/2 )| from below. By (4.62) and the convexity of Z1/2 , we
deduce that for any x ∈ ∂ Z1/2 , the set Z1/2 contains the cone Cx generated by x
and Bρ ( x̂) (see Figure A.7), which implies that the boundary of Z1/2 is uniformly
Lipschitz. In particular, if we consider the set Z1/2,Ĉ √ζ defined in (4.69), it follows
that
|Z1/2,Ĉ √ζ | ≥ |Z1/2 | − C̄ Ĉ ζ,
p

where C̄ depends only on R and n. Thanks to this bound, combining (4.70), Theo-
rem A.31, and (4.63), we deduce that

|∇Γ(Z1/2 )| ≥ ∇ 12 − ζ w Z1/2,Ĉ √ζ



∫
= det 12 − ζ D2 w dx


Z1/2, Ĉ √ ζ
1

n
= 2 −ζ Z1/2,Ĉ √ζ
1
≥ n |Z1/2 | − C ζ .
p
2

Step 5: Estimating |∇Γ(Z1/2 )| from above. It follows by Proposition A.35 that

∫  
|∇Γ(Z1/2 )| ≤ det D2 u − 21 D2 w dx.
Z1/2 ∩{Γ=u− 12 w }

Since
D2 u − 12 D2 w ≥ 0 on Γ = u − 12 w (4.71)


(see Figure A.8), we can apply (A.6) with A = D2 u − 21 D2 w

and B = 1 2
2D w to deduce
that h
1/n
1/n i n
det(D2 Γ) ≤ det D2 u − det( 12 D2 w) 1 {Γ=u− 1 w } .
2
116 4 Interior regularity of weak solutions

Thus, recalling (4.61) and (4.63), we get

∫ h
1/n
1/n i n
|∇Γ(Z1/2 )| ≤ det D2 u − det( 21 D2 w) dx
Z1/2 ∩{Γ=u− 12 w }
∫  n
≤ (1 + ζ)1/n − 1
2 dx
Z1/2 ∩{Γ=u− 21 w }
 n
1
+ ζ Γ = u − 12 w .

≤ 2 Z1/2 ∩

Step 6: Conclusion. Combining Steps 4 and 5, we deduce that

Z1/2 ∩ {Γ = u − 12 w}
≥ 1 − K1 ζ (4.72)
p
|Z1/2 |
for some constant K1 = K1 (n, r, R) > 0. Also, thanks to (4.71) and (4.64) we see that
1 1 2 2 1
2C 0 Id ≤ 2 D w ≤ D u on Γ = u − 2 w .


1 1
This proves that the eigenvalues of D2 u are bigger than 2C 0 on {Γ = u − 2 w}, and
since their product is bounded by 1 + ζ ≤ 2, we deduce that

kD2 uk ≤ K2 on Γ = u − 12 w , (4.73)


for some constant K2 = K2 (n, r, R) > 0. Hence, it follows by (4.72), (4.73), and the
inclusion Z1/2 ⊂ BR ( x̄) that

Z1/2 ∩ {kD2 uk ≥ K2 } ≤ Z1/2 − Z1/2 ∩ {Γ = u − 12 w ≤ K1 BR (0) ζ,

p

which proves the result with K := max K1 |BR (0)|, K2 .




4.8.2 An estimate on the Hessian of solutions in renormalized sections. In the

next lemma, we show that the L 1 -size of kD2 uk in a normalized section Sh (x0 ) is
controlled by the measure of points with D2 u ' Id inside a smaller section Sηh (x0 ).
However, for technical reasons, instead of considering just the measure of such
points inside Sηh (x0 ), we need prove such an estimate when we replace Sηh (x0 ) with
its intersection with another section St (y) of height t ≥ h (see Figure 4.15). This
will be crucial in the proof of Proposition 4.34 as the presence of the section St (y)
will allow us to localize our estimates. Indeed, to prove a decay estimate on the
measure of the superlevel sets of kD2 uk in Proposition 4.34, we shall apply the next
lemma (actually, its rescaled version Corollary 4.32) to a suitable family of sections
{Sh x i (xi )}1≤i ≤m . There, the sections Sh x i (xi ) and Z1/2 will play the role of Sh (x0 )
and St (y) (see Figure 4.16).
 4.8 Interior W 2, p estimates 117

Figure 4.15. The L 1 -size of D2 u inside Sh (x0 ) is controlled by the measure of points in Sηh (x0 ) ∩
St (y), where D2 u ' Id.

Lemma 4.30. Let Ω ⊂ Rn be an open set, Z ⊂ Ω a normalized convex set, and

u : Ω → R a convex function of class C 2 satisfying

λ ≤ det D2 u ≤ Λ in Ω, for some 0 < λ ≤ Λ.

Assume, in addition, that u|∂Z = 0, let x0 be the minimum point of u, set h :=

max Z |u| = |u(x0 )| so that Z = Sh (x0 ), and suppose that S2h (x0 ) ⊂⊂ Ω. Also, let η
be as in Proposition 4.27. Then there exists a universal constant K̂ > 0 such that the
following holds for any y ∈ Ω and t ≥ h such that x0 ∈ St (y) ⊂⊂ Ω:
∫
(i) kD2 uk dx ≤ K̂ K̂1 Id ≤ D2 u ≤ K̂ Id ∩ Sηh (x0 ) ∩ St (y) .

Sh (x0 )

(ii) If, in addition, k det D2 u − 1k L ∞ (Ω) ≤ ζ, then

Sh (x0 ) ∩ kD2 uk ≥ K̂ ≤ K̂ ζ K̂1 Id ≤ D2 u ≤ K̂ Id ∩ Sηh (x0 ) ∩ St (y) .

 p 

Proof. Since u is convex, its Hessian is nonnegative definite. Hence, its norm is
controlled by its trace, that is,

kD2 u(x)k = sup ∂ee u(x) ≤ tr D2 u(x) = ∆u(x).

e∈S n−1

Then the divergence theorem and Corollary 4.9 (recall that by assumption, S2h (x0 ) ⊂
Ω) yield
∫ ∫ ∫
kD2 uk dx ≤ ∆u dx = ∂ν u
Sh (x0 ) Sh (x0 ) ∂Sh (x0 )
≤ k∇uk L ∞ (Sh (x0 )) Hn−1 (∂Sh (x0 )) ≤ C1 . (4.74)

Here, the last inequality follows from Corollary A.16 and that Sh (x0 ) is normalized.
118 4 Interior regularity of weak solutions

In particular, by Chebyshev’s inequality, we see that

C1
Sh (x0 ) ∩ kD2 uk > σ ∀ σ > 0. (4.75)

≤
σ
We now apply Proposition 4.27(b) to deduce the existence of a point z such that

Sηh (x0 ) ∩ St (y) ⊃ Sη 2 h (z),

which by Lemma 4.6 implies that

|Sηh (x0 ) ∩ St (y)| ≥ c1

for some c1 > 0 universal. Hence, by (4.75) with σ = 2C1 /c1 and the bound above,
we obtain
2C1
 
2 c1
kD uk ≤ ∩ Sηh (x0 ) ∩ St (y) ≥ .
c1 2
In addition, the lower bound det D2 u ≥ λ implies that there exists a universal constant
C2 > 0 such that
1 2C1 2C1
Id ≤ D2 u ≤ Id inside kD2 uk ≤

C2 c1 c1

(see Remark 1.1), from which we deduce that

2C1 c1
1
Id ≤ D2 u ≤ Id ∩ Sηh (x0 ) ∩ St (y) ≥ . (4.76)

C2 c1 2
• Case (i). By (4.74) and (4.76), we get
2C1  1
∫
2C1
kD2 uk dx ≤ 2
C2 Id ≤ D u ≤ c1 Id ∩ Sηh (x0 ) ∩ St (y) ,
Sh (x0 ) c1
 2C1
and the desired estimate follows by choosing K̂ := max c1 , C2 .
• Case (ii). Since Sh (x0 ) is normalized, it follows by Lemma 4.8 that S2h (x0 )
satisfies (4.60) for some universal radii r, R > 0. Hence, since S2h (x0 ) ⊂⊂ Ω,
we can apply Lemma 4.29 inside S2h (x0 ).
Let ζ0 be as in Lemma 4.29 applied inside S2h (x0 ). We distinguish between
two cases:
 4.8 Interior W 2, p estimates 119

(1) If ζ ≤ ζ0 , then Lemma 4.29 and (4.76) imply that

Sh (x0 ) ∩ kD2 uk ≥ K


2K p  1 2C1
≤ ζ C2 Id ≤ D2 u ≤ c1 Id ∩ Sηh (x0 ) ∩ St (y) ,
c1
2C1
which gives the desired estimate with K̂ := max K, 2K
c1 , C2, c1 .


2C1
(2) If ζ ≥ ζ0 , then (4.75) with σ = c1 and (4.76) yield

Sh (x0 ) ∩ kD2 uk ≥ 2C 1

c1
1 p 1 2C1
≤ √ ζ C2 Id ≤ D2 u ≤ c1 Id ∩ Sηh (x0 ) ∩ St (y) ,
ζ0
 2C1
and the result follows choosing K̂ := max c1 , √1ζ , C2 .
0

4.8.3 On the size of the Hessian in terms of the size of sections. In order to exploit
Lemma 4.30, we want to consider a general section, normalize it, apply that result
to the normalized solution, and transfer the information back to u. When doing this,
the following factor, depending on the map normalizing the section, will naturally
appear.

Definition 4.31. Let u : Ω → R be a convex function, consider a section S :=

S(x, p, t) ⊂⊂ Ω, and let L be the affine map normalizing S. Then we define the
normalized size of S as
kL k 2
α(S) := .
(det L)2/n
Corollary 4.32. Let Ω ⊂ Rn be an open set, and u : Ω → R a convex function of
class C 2 satisfying

λ ≤ det D2 u ≤ Λ in Ω, for some 0 < λ ≤ Λ.

Let S2h (x0 ) ⊂⊂ Ω, and assume x0 ∈ St (y) ⊂⊂ Ω for some y ∈ Ω and t ≥ h. Set
α := α(Sh (x0 )), let η be as in Proposition 4.27 and K̂ as in Lemma 4.30. Then,
∫
(i) kD2 uk dx ≤ K̂ α K̂1 α ≤ kD2 uk ≤ K̂ α ∩ Sηh (x0 ) ∩ St (y) .

Sh (x0 )
120 4 Interior regularity of weak solutions

(ii) If, in addition, k det D2 u − 1k L ∞ (Ω) ≤ ζ, then

Sh (x0 ) ∩ kD2 uk ≥ K̂ α


≤ K̂ ζ K̂1 α ≤ kD2 uk ≤ K̂ α ∩ Sηh (x0 ) ∩ St (y) .

p 

Proof. Let L be the affine map normalizing Sh
(x0 ), define Sh∗ (x0∗ ) := L Sh (x0 ) ,


∗ (x ∗ ) := L S (x ) , and S ∗ (y ∗ ) := L S (y) , and let v be as in (4.2). Then


S2h 0 2h 0 t t
applying Lemma 4.30 to v, we obtain
∫
kD2 vk dz ≤ K̂ K̂1 Id ≤ D2 v ≤ K̂ Id ∩ Sηh ∗
(x0∗ ) ∩ St∗ (y)

Sh∗ (x0∗ )
or
Sh∗ (x0∗ ) ∩ kD2 vk ≥ K̂ ≤ K̂ ζ K̂1 Id ≤ D2 v ≤ K̂ Id ∩ Sηh (x0∗ ) ∩ St∗ (y ∗ ) ,
 p  ∗

depending whether we are in case (i) or (ii).

Also, recalling (4.3) and the definition of α, we see that
kD2 u(x)k ≤ αkD2 v(L x)k,
1 1
Id ≤ D2 v(z) ≤ K̂ Id ⇒ α ≤ kD2 u(L −1 z)k ≤ K̂ α,
K̂ K̂
and
kD2 v(z)k ≥ K̂ ⇒ kD2 u(L −1 z)k ≥ K̂ α.
Hence, in case (i), we get
∫ ∫ ∫
2 2
kD u(x)k dx ≤ α kD v(L x)k dx = | det L| α
−1
kD2 v(z)k dz
Sh (x0 ) Sh (x0 ) Sh∗ (x0∗ )

≤ K̂ α| det L| −1 z : K̂1 Id ≤ D2 v(z) ≤ K̂ Id ∩ Sηh∗

(x0∗ ) ∩ St∗ (y)


≤ K̂ α| det L| −1 z : K̂1 α ≤ kD2 u(L −1 z)k ≤ K̂ α ∩ Sηh∗

(x0∗ ) ∩ St∗ (y)


= K̂ α x : K̂1 α ≤ kD2 u(x)k ≤ K̂ α ∩ Sηh (x0 ) ∩ St (y) ,



while in case (ii), we obtain

Sh (x0 ) ∩ x : kD2 u(x)k ≥ K̂ α ≤ Sh (x0 ) ∩ x : kD2 v(L x)k ≥ K̂
 

= | det L| −1 Sh∗ (x0∗ ) ∩ z : kD2 v(z)k ≥ K̂



≤ K̂ ζ | det L| −1 z : K̂1 Id ≤ D2 v(z) ≤ K̂ Id ∩ Sηh

p ∗
(x0∗ ) ∩ St∗ (y ∗ )


≤ K̂ ζ x : K̂1 α ≤ kD2 u(x)k ≤ K̂ α ∩ Sηh (x0 ) ∩ St (y) ,

p 

as desired. 
 4.8 Interior W 2, p estimates 121

We conclude this section with a useful observation.

Remark 4.33. Let Ω ⊂ Rn be an open set, and let u : Ω → R be a convex function of
class C 2 satisfying det D2 u > 0. Then,

kD2 u(x)k
lim α Sh (x) =



1/n ∀ x ∈ Ω.
h→0 det D2 u(x)

Indeed, without loss of generality, assume that x = 0 and u(0) = ∇u(0) = 0, and
set A := 21 D2 u(0). Noticing that A is positive definite (since det A > 0), a Taylor
expansion yields

Sh (0) = {u < h} = y : hAy, yi + o(|y| 2 ) < h = y : 1 + o(1) hAy, yi < h ,

 

that is, for any ε > 0 there exists hε > 0 such that

y : (1 + ε)hAy, yi < h ⊂ Sh (0) ⊂ y : (1 − ε)hAy, yi < h ∀ h ≤ hε .

 

This implies that the maps Lh normalizing Sh (0) satisfy

1 − ε  1/2 1 + ε  1/2
k Ak ≤ kLh k ≤ k Ak
h h
and  1 − ε  n/2  1 + ε  n/2
(det A)1/2 ≤ det Lh ≤ (det A)1/2 ;
h h
hence,
1 − ε k Ak 1 + ε k Ak
≤ α(Sh (0)) ≤ .
1 + ε (det A)1/n 1 − ε (det A)1/n
By the arbitrariness of ε, this proves

k Ak kD2 u(0)k
lim α Sh (0) = =
1/n .


h→0 (det A)1/n det D2 u(0)

4.8.4 A priori W 2, p interior estimates for smooth solutions. To obtain the desired
a priori estimates, we want to control the measure of the superlevel sets of kD2 uk.
This is done by combining Corollary 4.32 above with a covering argument.
Proposition 4.34. Let Z ⊂ Rn be a normalized convex set and u : Z → R be a
convex function of class C 2 satisfying
(
λ dx ≤ det D2 u ≤ Λ dx in Z,
u=0 on ∂ Z.
122 4 Interior regularity of weak solutions

Let x0 be the minimum point of u, set h := max Z |u| = |u(x0 )|, and define

Z1/2 := Sh/2 (x0 ).

Also, let K̂ be as in Lemma 4.30. Then there exists a universal constant M > 0 such
that the following holds: Let Dk denote the compact sets

Dk := x ∈ Z1/2 : kD2 u(x)k ≥ M k , k ≥ 1.



Then,
1
∫ ∫
2
(i) kD uk dx ≤ (1 − τ) kD2 uk dx, τ := .
D k+1 Dk 1 + K̂ 2

(ii) If, in addition, k det D2 u − 1k L ∞ (Z) ≤ ζ, then

|Dk+1 | ≤ K̂ ζ |Dk |.
p

Proof. Let M  K̂ (to be fixed later), and fix k ≥ 1.

Notice that for any point x ∈ Dk+1 , it follows by Remark 4.33 that

1 1
lim α St (x) ≥ kD2 u(x)k ≥ 1/n M k+1 .


t→0 Λ1/n Λ
On the other hand, since Sηh/2 (x) ⊂ Sh (x0 ) = Z (by Proposition 4.27(c)), it follows
by Corollary 4.7 that
α Sηh/2 (x) ≤ Ĉ

for some universal constant Ĉ. Choosing M large enough (independent of k) we can
ensure that
1
Ĉ < K̂ M k < 1/n M k+1 ;
Λ
so, by continuity, there exists h x ∈ (0, ηh/2) such that α Sh x (x) = K̂ M k .

We now apply Lemma 4.28 to the covering {Sh x (x)}x ∈Dk+1 to find a finite sub-
family of sections {Sh x i (xi )}i=1,...,m still covering Dk+1 and such that the sections
Sηh x i (xi ) are disjoint (see Figure 4.16). Then by Corollary 4.32 applied with y = x0
and t = h/2, for each i = 1, . . . , m, we get

• Case (i):
∫
kD2 uk dx ≤ K̂ 2 M k M k ≤ kD2 uk ≤ K̂ 2 M k ∩ Sηh x i (xi ) ∩ Z1/2 ;

Sh x (xi )
i
 4.8 Interior W 2, p estimates 123

Figure 4.16. Given a value M  1, we fix k ≥ 1 and for any x ∈ Z1/2 such that kD2 u(x)k ≥
M k+1 , we find a section Sh x (x) such that α(Sh x (xi )) ' M k . By Lemma 4.28, we can find a
i
subfamily of sections {Sh x (xi )}1≤i ≤m that cover the set Z1/2 ∩ {kD2 u(x)k ≥ M k } and such that
i
{Sηh x (xi )}1≤i ≤m are disjoint. Then applying Lemma 4.30 (more precisely, its rescaled version in
i
Corollary 4.32), we can control the L 1 -size of kD2 uk inside Sh x (xi ) by the measure of points in
i
Sηh x (xi ) ∩ Z1/2 , where kD2 uk ' M k . This allows us to obtain the desired decay on the measure
i
of the superlevel sets of kD2 uk.

• Case (ii):

Sh x i (xi ) ∩ kD2 uk ≥ K̂ 2 M k


≤ K̂ ζ M k ≤ kD2 uk ≤ K̂ 2 M k ∩ Sηh x i (xi ) ∩ Z1/2 .

p 

Adding these inequalities over i = 1, . . . , m, using that

m
Ø
Dk+1 ⊂ Sh x i (xi ), with Sηh x i (xi ) disjoint,
i=1

and assuming that M > K̂ 2 (so that K̂ 2 M k < M k+1 ), we see that
• Case (i):
∫
kD2 uk dx ≤ K̂ 2 M k M k ≤ kD2 uk ≤ K̂ 2 M k ∩ Z1/2

D k+1
∫
2
≤ K̂ kD2 uk dx
{ M k ≤ kD 2 u k ≤ K̂ 2 M k }∩Z1/2
∫
≤ K̂ 2 kD2 uk dx,
D k \D k+1
124 4 Interior regularity of weak solutions

and adding K̂ 2 kD2 uk to both sides of the above inequality, the conclusion
∫
D k+1
follows with τ = 1/(1 + K̂ 2 );
• Case (ii):
m
Sh x i (xi ) ∩ kD2 uk ≥ K̂ 2 M k
Õ 
|Dk+1 | ≤
i=1

≤ K̂ ζ M k ≤ kD2 uk ≤ K̂ 2 M k ∩ Z1/2
p 

≤ K̂ ζ |Dk |.
p


Thanks to this result, a priori interior W 2, p regularity follows easily:
Theorem 4.35. Let Z ⊂ Rn be a normalized convex set and u : Z → R be a convex
function of class C 2 satisfying
(
λ dx ≤ det D2 u ≤ Λ dx in Z,
u=0 on ∂ Z.

Let x0 be the minimum point of u, set h := max Z |u| = |u(x0 )|, and define

Z1/2 := Sh/2 (x0 ).

Then there exist universal constants γ, C > 0 such that the following holds:
∫
(i) kD2 uk 1+γ dx ≤ C.
Z1/2

(ii) For any σ ≥ 1, if k det D2 u − 1k L ∞ (Z) ≤ ζ with ζ ≤ e−Cσ , then

∫
kD2 uk σ dx ≤ C.
Z1/2

Proof. With the same notation as in Proposition 4.34, we recall that (by definition)
kD2 uk ≥ M k inside Dk . Hence, Proposition 4.34(i) and (4.74) (applied with Z1/2 in
place of Sh (x0 )) imply that

1 (1 − τ)k
∫ ∫
2
|Dk | ≤ k kD uk dx ≤ kD2 uk dx
M Dk Mk Z1/2

(1 − τ)k C1 | log(1 − τ)|

≤ C1 = k(1+2γ) , γ := ,
M k M 2 log(M)
 4.8 Interior W 2, p estimates 125

while Proposition 4.34(ii) gives (recall that Z1/2 ⊂ Bn (0))

 p k |Bn (0)| 1
|Dk | ≤ K̂ ζ |Z1/2 | ≤ k(σ+1) provided ζ ≤ .
M M 2σ+2 K̂ 2
Also, applying Lemma A.36 with g := kD2 uk, we see that
∫ ∫ ∞
kD2 uk p dx = p t p−1 |Z1/2 ∩ {kD2 uk ≥ t}| dt
Z1/2 0
∫ 1 Õ ∫ M k+1
≤ p|Z1/2 | t p−1
dt + p |Dk | t p−1 dt
0 k ≥0 Mk
∞
Õ
≤ |Bn (0)| + M (k+1)p |Dk | ∀ p ≥ 1.
k=1

Hence,
∫ ∞
2 1+γ 1+γ
Õ
• Case (i): kD uk dx ≤ |Bn (0)| + C1 M M −kγ ≤ C;
Z1/2 k=1
∫  ∞
kD2 uk σ dx ≤ |Bn (0)| 1 + M σ
Õ
• Case (ii): M −kσ
Z1/2 k=1 
1

= |Bn (0)| 1 + ≤ 2|Bn (0)|,
1 − M −σ
provided ζ ≤ 1/M 2σ+2 K̂ 2 and assuming, without loss of generality, that M ≥
2. Since e−Cσ ≤ 1/M 2σ+2 K̂ 2 if C is sufficiently large (the largeness being
universal), this concludes the proof. 

4.8.5 W 2, p regularity for Alexandrov solutions. In the previous sections, we

established interior W 2, p estimates for smooth solutions of Monge–Ampère. We first
extend this result to Alexandrov solutions by a simple approximation argument, and
then we deduce as a corollary the regularity of solutions with bounded right-hand
side.
Theorem 4.36. Let Z ⊂ Rn be a normalized convex set u : Z → R be a convex
function satisfying (
µu = f dx in Z,
(4.77)
u=0 on ∂ Z,
for some 0 < λ ≤ f ≤ Λ. Let x0 be the minimum point of u, set h := max Z |u| =
|u(x0 )|, and consider the section
Z1/2 := Sh/2 (x0 ).
126 4 Interior regularity of weak solutions

Then there exist C, γ > 0, depending only on n, λ, and Λ, such that the following
holds:
(i) kukW 2,1+γ (Z1/2 ) ≤ C.

(ii) For any σ ≥ 1, if | f − 1| ≤ e−Cσ , then kukW 2, σ (Z1/2 ) ≤ C.

Proof. Let fε ∈ C 2 (Z) be a sequence of functions such that λ ≤ fε ≤ Λ (respectively

| fε − 1| ≤ e−Cσ ) and fε → f a.e. in Z, and let uε : Z → R solve
(
µuε = fε dx in Z,
uε = 0 on ∂ Z.

Noticing that the functions uε are of class C 2 inside Z (by Theorem 3.10), the
desired result follows by applying Theorem 4.35 to uε and then letting ε → 0 (notice
that uε → u locally uniformly by Corollary 2.12, and the sections Z1/2 ε converge

to Z1/2 ). 
By the very same argument as in the proof of Corollary 4.21, we deduce interior
W 2, p regularity for strictly convex solutions.
Corollary 4.37. Let Ω ⊂ Rn be an open set and u : Ω → R be a strictly convex
function satisfying

µu = f dx in Ω, for some 0 < λ ≤ f ≤ Λ.

2,1+γ
Then u ∈ Wloc (Ω) for some γ > 0 depending only on n, λ, and Λ. In addition, there
2,σ
exists C = C(n, λ, Λ) such that for any σ ≥ 1, if | f − 1| ≤ e−Cσ , then u ∈ Wloc (Ω).
Next, we consider the case when the right-hand side is continuous, and we show
2,σ
that solutions belong to Wloc for all σ < ∞.
Corollary 4.38. Let Ω ⊂ Rn be an open set and u : Ω → R be a strictly convex
function satisfying

µu = f dx in Ω, for some f > 0 continuous.

2,σ
Then u ∈ Wloc (Ω) for all σ < ∞.
Proof. Fix x ∈ Ω and p ∈ ∂u(x), consider t > 0 small (to be fixed later) so that
St := S(x, p, t) ⊂⊂ Ω, and let Lt be the affine transformation that normalizes St . We
set St∗ := L(St ), define vt as in (4.2), and set

1 f (Lt−1 z)
ṽt (z) := vt (z) and f˜t (z) := ∀ z ∈ St∗,
f (x)1/n f (x)
 4.9 Application 4: The semigeostrophic equations 127

so that ṽt solves the equation

µṽt = f˜t dz
(
in St∗,
vt = 0 on ∂St∗ .

Notice that since f is continuous and St → {x} in the Hausdorff distance as t → 0

(by strict convexity), we have that
f
k f˜t − 1k L ∞ (St∗ ) = −1 →0 as t → 0.
f (x) L ∞ (St )

In particular, given σ ∈ (1, ∞), we can choose t small enough so that

k f˜t − 1k L ∞ (St∗ ) ≤ e−Cσ,
2,σ ∗ 2,σ
and it follows by Corollary 4.37 that ṽt ∈ Wloc (St ), or equivalently u ∈ Wloc (St ).
By the arbitrariness of σ and x, this proves the result. 
Remark 4.39. A question that naturally arises in view of the previous results is the
following: Choose a sequence of functions fk such that λ ≤ fk ≤ Λ converging
strongly in L 1 (Ω) to f , and denote the solutions of (4.77) corresponding to fk and f
by uk and u, respectively. By convexity of uk , it is immediate to deduce that uk → u
uniformly, and ∇uk → ∇u in Lloc 1 (Ω). However, due to the highly nonlinear character

of the equation, it is not clear what can be said about the W 2,1 convergence of the uk .
In [32], the authors prove that if fk → f in Lloc1 (Ω), then u → u in W 2,1 (Ω).
k loc

4.9 Application 4: The semigeostrophic equations

The semigeostrophic equations are a simple model used in meteorology to describe

large-scale atmospheric flows. As explained, for instance in [10, Section 2.2] (see
also [29] for a more complete discussion), these equations can be derived from the
3-D Euler equations, with Boussinesq and hydrostatic approximations, subject to a
strong Coriolis force. Since for large-scale atmospheric flows, the Coriolis force
dominates the advection term, the flow is mostly bidimensional. For this reason, the
study of the semigeostrophic equations in 2-D and 3-D are similar, and in order to
simplify our presentation, we focus here on the two-dimensional periodic case.
The semigeostrophic system can be written as
 ∂ ∇p + (u t · ∇)∇pt + ∇⊥ pt + u t = 0 in [0, ∞) × R2,
 t t


in [0, ∞) × R2, (4.78)

div u t = 0
 p0 = p̄ on R2,



128 4 Interior regularity of weak solutions

where u t : R2 → R2 and pt : R2 → R are periodic functions corresponding to,

respectively, the velocity and the pressure.5 In the above system the notation ∇⊥ pt
denotes the rotation of the vector ∇pt by π/2, that is, ∇⊥ pt = (∂2 pt , −∂1 pt ), while
(u t · ∇) denotes the operator u 1t ∂1 + u 2t ∂2 (see [45] for more details).
As shown in [29], energetic considerations show that it is natural to assume that
the function Pt (x) := pt (x) + |x| 2 /2 is convex on all of R2 . Motivated by optimal
transport theory, where gradients of convex functions are optimal transport maps (see
Theorem 4.22), it is natural to look at the image of the Lebesgue measure under the
map ∇Pt : R2 → R2 and try to understand its behavior. Hence, letting dx denote the
Lebesgue measure on R2 , one defines the measure ρt as (∇Pt )] dx and computes its
time derivative to understand its evolution in time.
Using (4.78), one can show that ρt solves the following dual problem (see [45,
Section 1.2] for a detailed computation):


 ∂t ρt + div(Ut ρt ) = 0 in [0, ∞) × R2,
 U (y) = ∇P∗ (y) − y
⊥ in [0, ∞) × R2,

t
(4.79)

 t

 ρt = (∇Pt )] dx in [0, ∞) × R2,
 P = p̄ + |x| 2 /2 on R2,


 0
where Pt∗ is the convex conjugate of Pt , namely

Pt∗ (y) := sup y · x − Pt (x) .


x ∈R2

The dual problem (4.79) is nowadays pretty well understood. In particular, Benamou
and Brenier proved in [10] global existence of weak solutions to (4.79). On the
contrary, much less is known about the original system (4.78). Formally, given a
solution (ρt , Pt∗ ) of (4.79) and defining Pt through the relation

Pt (x) := sup {y · x − Pt∗ (y)}, (4.80)

y ∈R2

the pair (pt , u t ) given by

(
pt (x) := Pt (x) − |x| 2 /2,
(4.81)
u t (x) := [∂t ∇Pt∗ ](∇Pt (x)) − D2 Pt∗ (∇Pt (x)) ∇Pt (x) − x

⊥

solves (4.78).
However, a priori D2 Pt∗ is only a matrix-valued measure, since Pt∗ is a convex
function; thus, it is not clear what D2 Pt∗ ◦ ∇Pt means. However, since ρt solves
5We note that here the subscript t in u t and pt does not denote a time derivative, but only expresses fact that
the velocity and the pressure depend on the time variable t. In other words, u t (x) = u(t, x) and pt (x) = p(t, x).
 4.9 Application 4: The semigeostrophic equations 129

a continuity equation with a divergence-free vector field (because Ut is the rotated

gradient of the function Pt∗ (y) − |y| 2 /2; see (4.79)), one has that

0 < λ ≤ ρt ≤ Λ ∀ t > 0, (4.82)

provided this bound holds at t = 0. In addition, the relation ρt = (∇Pt )] dx implies

that (∇Pt∗ )] ρt = dx (since ∇Pt∗ is the inverse of ∇Pt ), from which it follows that Pt∗
is a Brenier solution of the Monge–Ampère equation

det(D2 Pt∗ ) = ρt on R2 (4.83)

(see Section 4.6), and since the image of ∇Pt∗ is convex (being all of R2 ), one can
show that Pt∗ is also an Alexandrov solution. Hence, if (4.82) holds, then we can
1+γ
apply Theorem 4.36 to obtain that D2 Pt∗ ∈ Lloc (R2 ), which allows us to prove that
1+γ
also D2 Pt∗ ◦ ∇Pt belongs to Lloc (R2 ).
To deal with the term ∂t ∇Pt∗ , one needs a different argument (see [82, Theorem
5.1]) to show that ∂t ∇Pt∗ ∈ Lloc 1+κ (R2 ) with κ := γ/(2 + γ). The basic idea is to

differentiate (4.83) with respect to t and use the first equation in (4.79) to find some
a priori estimates on ∂t ∇Pt∗ in terms of ρt and Ut (see [45, Theorem 4.5] for more
details).
These estimates allow one to give a meaning to the velocity field u t defined in
(4.81) and prove that (pt , u t ) solve (4.78). This is summarized in the following result
[6, Theorem 1.2]:
Theorem 4.40. Let p̄ : R2 → R be a periodic function such that p̄(x) + |x| 2 /2 is
convex, and assume that the measure ρ̄ := (Id +∇ p̄)] dx is absolutely continuous with
respect to the Lebesgue measure and satisfies 0 < λ ≤ ρ̄ ≤ Λ.
Let (ρt , Pt∗ ) be a solution of (4.79) starting from ρ̄, and let Pt be defined as in
(4.80). Then the couple (pt , u t ) defined in (4.81) is a distributional solution of (4.78).
In [7], the above result has been generalized to the case of three-dimensional
bounded convex domains. This situation presents several additional difficulties. First
of all, the three-dimensional version of (4.78) is much less symmetric than its 2-D
counterpart because the action of Coriolis force regards only the first and the second
space components. Moreover, even considering regular initial data and velocities,
in this case the proof requires a finer regularization scheme. Still, under suitable
assumptions on the initial data, global existence of distributional solutions hold (see
[7] for more details).
Let us conclude this section by mentioning that the assumption ρ̄ := (Id +∇ p̄)] dx
≥ λ > 0 is crucial in all the previous results, as it implies that ρt can never vanish.
However, from a physical point of view, it would be extremely important to remove
such a hypothesis. The main difficulty in this case concerns the nontrivial evolution
130 4 Interior regularity of weak solutions

of the support of the solution ρt . Indeed, since a priori the set {ρt > 0} may
not be open, one is not able to apply any regularity theory for the Monge–Ampère
equation, and completely new ideas need to be found in order to prove the existence
of distributional solutions.

4.10 Interior C 2,α regularity

In [14], in addition to proving W 2, p regularity for solutions with right-hand side close
to a constant, Caffarelli showed interior C 2,α regularity when the right-hand side is
Hölder continuous. Here, we give a proof of this result. First, we use the argument
in [66] to get C 1,1 estimates, and then, by applying the Evans–Krylov theorem, we
obtain the desired C 2,α regularity.
The idea of the proof is that the right-hand side is almost constant at small scales;
so, we can compare our solution to solutions with constant right-hand side for which
Pogorelov’s estimates apply. Then we use the closeness to these smooth functions to
show that our solution has bounded Hessian. An important part of the proof will be
to control the shape of small sections and prove that they remain comparable to balls
at all scales. Indeed, as observed in the next lemma, the roundedness of sections
gives a C 1,1 control.
Lemma 4.41. Let u ∈ C 2 (Ω) be a uniformly convex function satisfying det D2 u ≤ Λ,
let x ∈ Ω, and assume that there exists R > 0 such that
diam Sh (x)


lim inf √ ≤ R. (4.84)
h→0 h
Then there exists a constant C > 0, depending only on n, Λ, and R, such that

kD2 u(x)k ≤ C.

Proof. Without loss of generality, we assume that x = 0, u(0) = |∇u(0)| = 0. Let

A = 21 D2 u(0). Since u is uniformly convex, A is positive definite. Then by a Taylor
expansion,
Sh (0) = {u ≤ h} = y ∈ Ω : hAy, yi + o(|y| 2 ) ≤ h .


For each y ∈ Ω, observe that

√ √ √
hAy h, y hi + o(|y h| 2 )
hAy, yi ≤ 1 if and only if lim ≤ 1.
h→0 h
Thus, if we define the ellipsoid

E := y ∈ Ω : hAy, yi < 1 ,

 4.10 Interior C 2,α regularity 131

we see that
1
E = lim √ S h (0).
h→0 h

Hence, by (4.84), diam(E ) ≤ R, so E ⊂ BR/2 (0).

Since A is symmetric, there exists an orthonormal transformation diagonalizing
A. Hence,
 n 
2
Õ
y∈Ω: ai yi ≤ 1 ⊂ BR/2 (0),
i=1

where 0 < a1 ≤ a2 ≤ · · · ≤ an are the eigenvalues of A. Taking y = (a1−1/2, 0, . . . , 0),

we find that a1 ≥ 4/R2 . Recalling that i=1 ai = det A ≤ Λ, we see that
În

În   2(n−1)
i=1 ai Λ R
k Ak = an = Î ≤ n−1 ≤ Λ ,
i,n ia a1 2

as desired. 

We can now prove the interior C 2,α regularity of solutions with Hölder right-hand
side.

Theorem 4.42. Let Z ⊂ Rn be a normalized convex set, and let u : Z → R be a

convex function satisfying (
µu = f dx in Z,
u=0 on ∂ Z,

for some 0 < λ ≤ f ≤ Λ, with f ∈ C 0,α (Z) for some α ∈ (0, 1). Let x0 be the
minimum point of u, set h := max Z |u| = |u(x0 )|, and consider the section

Z1/2 := S(x0, 0, h/2).

Then there exist C > 0, depending only on n, λ, Λ, α, and k f kC 0, α (Z) , such that

kukC 2, α (Z1/2 ) ≤ C.

Before proving this result, we notice the following consequence, whose proof is
identical to that of Corollary 4.21.

Corollary 4.43. Let Ω ⊂ Rn be an open set, and u : Ω → R a strictly convex function

satisfying
µu = f dx in Ω,
2,α
for some 0 < λ ≤ f ≤ Λ, f ∈ C 0,α (Ω), and α ∈ (0, 1). Then u ∈ Cloc (Ω).
132 4 Interior regularity of weak solutions

Figure 4.17. For any k ≥ 0, we construct a function vk that solves det D2 v̂k = 1 and coincides with
û on ∂ Ŝk . Then by showing that kD2 vk ( x̂) − D2 vk+1 ( x̂)k . νk . 2−kα , we prove that kD2 û( x̂)k is
universally bounded.

Proof of Theorem 4.42. Up to approximating f with a sequence of C 2 functions and

arguing as in the proof of Theorem 4.36, it suffices to prove an interior C 2,α estimate
when u ∈ C 2 (Z), but with a bound that depends only on n, λ, Λ, and k f kC 0, α (Z) .
Fix x ∈ Z3/4 := S(x0, 0, 3h/4), and let τ > 0 be given by Lemma 4.13 with
ρ = 1/4. Up to subtracting an affine function from u, and replacing u by u/ f (x)1/n
and f by f / f (x), we can assume that

u(x) = 0, ∇u(x) = 0, and f (x) = 1.

Step 1: Construction of an approximating sequence of solutions with constant right-

hand side. Pick ĥ ∈ (0, τ] to be fixed later, and for k ≥ 0 define hk := ĥ/4k , and
Sk := S(x, 0, hk ).
Let L be the affine map normalizing S0 , and consider L̂ := (det L)−1/n L so that

det L̂ = 1, Bρ (0) ⊂ L̂(S0 ) ⊂ Bnρ (0)

with ρ := (det L)−1/n . Then we set û := u ◦ L̂ −1 , x̂ := L̂ x, and Ŝk := L̂(Sk ), and

notice that û solves (
det D2 û = f ◦ L̂ −1 in Ŝk ,
û = hk on ∂ Ŝk .
Define v̂k ∈ C ∞ (Ŝk ) to be the solution of
(
det D2 v̂k = 1 in Ŝk ,
v̂k = hk on ∂ Ŝk

(see Figure 4.17), set

νk := sup f (z) − 1 = sup f ( L̂ −1 z) − 1 , (4.85)

z ∈Sk z ∈ Ŝk
 4.10 Interior C 2,α regularity 133

and notice that by Theorem 2.10,

(1 + νk )1/n [û − hk ] ≤ v̂k − hk ≤ (1 − νk )1/n [û − hk ] in Ŝk . (4.86)
In particular, recalling that minŜk û = 0 for all k, we have that
− νk hk ≤ min v̂k ≤ νk hk . (4.87)
Ŝk

Since f is continuous and S(x, 0, ĥ) → { x̂} in the Hausdorff distance as ĥ → 0 (by
strict convexity of u), we can choose ĥ small enough so that νk ≤ 1/4 for all k ≥ 0.
With this choice, notice that
1
x̂ ∈ { v̂k ≤ hk /2} and k f − 1k L ∞ (Sk ) ≤ 4 ∀ k ≥ 0. (4.88)
Set
Mk := |D2 v̂k ( x̂)| ∀ k ≥ 0.
Our goal is to show that the numbers Mk are uniformly bounded by a dimensional
constant, which by Lemma 4.17 will imply that the sections Ŝk are round.
Step 2: Mk is universally bounded and Ŝk is round for all k ≥ 0. Thanks to
Lemma 4.6, we know that

c1 ≤ √ 1 ≤ C1 ∀ k ≥ 0,
h k Ŝk

for some dimensional constants 0 < c1 ≤ C1 . In particular, recalling that Bρ (0) ⊂

Ŝ0 ⊂ Bnρ (0) for some ρ > 0, we deduce that there exists Rn > 1, depending only on
dimension, such that
B1/Rn (0) ⊂ √h1 Ŝ ⊂ BRn (0).
0 0

Hence,
√ since
x̂ ∈ { v̂0 ≤ h0 /2}, it follows by1 Corollary 4.5 applied to the function
1
v̂
h0 0 h 0 that the distance between x̂ and √h Ŝ0 is bounded below by a dimensional
z
0
constant. Therefore, since k f − 1k L ∞ (S0 ) ≤ 14 (see (4.88)), Theorem 3.9 applied to
1
√

h0 v̂0 h0 z − 1 implies that M0 is bounded by a dimensional constant M̂.
Our goal is to show by induction on k that if ĥ was chosen sufficiently small (the
smallness being universal), then all numbers Mk are bounded by M̂ + 1.
To this aim, we notice that, since ν0 ≤ 14 , it follows by (4.87) and the definition of
v̂0 that
h0 h0
− ≤ min v̂0 ≤ and v̂0 = h0 on ∂S0 .
4 S0 4
 √

Hence, we can apply Lemma 4.17 to the function h10 v̂0 h0 z − minS0 v̂0 with
M := M̂ + 1 and H := 2 to deduce that
B1/R ( x̂) ⊂ √1 Ŝ0 ⊂ BR ( x̂) (4.89)
h0
134 4 Interior regularity of weak solutions

for some dimensional constant R, and we want to prove by induction that the above
inclusion still holds if we replace √1h Ŝ0 by any of the sets √1h Ŝk .
0 k
So, assume by the inductive hypothesis that the above inclusion holds up to k.
Consider the functions
1 p
1 p

v̄k (z) := v̂k hk z and ṽk (z) := v̂k+1 hk z ,
hk hk
and note that

 det D2 v̄k = 1 in √1 Ŝk ,  det D2 ṽk = 1 in √1 Ŝk+1,


 

hk hk
and
 
 
 v̄k = 1
 on ∂ √1h Ŝk ,  v̄k =
 1
4 on ∂ √1h Ŝk+1 .
 k  k

Since (by the inductive assumption) √1 Ŝk satisfies (4.89), Corollary 4.5 applied
hk
with t1 = 1/4 j+1 and t2 = 1/4 j
gives
   
dist ∂ √1h Ŝk+j , √1h Ŝk+j+1 ≥ ĉn > 0 ∀ j = 1, 2, 3, (4.90)
k k

for some dimensional constant ĉn . Therefore, as before, we can apply Theorem 3.9
to deduce that
kD2 v̄k k ≤ C and kD2 ṽk k ≤ C in √1 Ŝk+2 .
hk

This allows us to use Proposition A.43 and Theorem A.42 to get

k v̄k k   ≤C and k ṽk k   ≤ C. (4.91)
C 3 √1 Ŝk+3 C 3 √1 Ŝk+3
hk hk

We now observe that

0 = det D2 v̄k − det D̃2 vk
∫ 1
d  
= det (1 − t)D2 ṽk + tD2 v̄k dt
0 dt
= ai j ∂i j (v̄k − ṽk ),
where (recall (A.2))
∫ 1   
2 2
ai j (x) := cof (1 − t)D ṽk (x) + tD v̄k (x) dt .
0 ij

Furthermore, as a consequence of (4.89) applied to Ŝ0 = L̂(S0 ) and Ŝk = L̂(Sk ),

√
2 hk
2
Sk ⊂ hk L̂ BR ( x̂) = R hk L̂ B1/R ( x̂) ⊂ R √ S0 = R2 2−k S0
p p
−1
−1

h0
 4.10 Interior C 2,α regularity 135

(recall that hk = 4−k h0 ). Thus,

diam(Sk ) ≤ R2 2−k diam(S0 ). (4.92)

Also, thanks to (4.86), we have

| v̂k − v̂k+1 | ≤ | v̂k − û| + | û − v̂k+1 | ≤ νk hk + νk+1 hk+1 ≤ 2νk hk inside Ŝk+1,

therefore,
| v̄k − ṽk | ≤ 2νk inside √1 Ŝk+1 . (4.93)
hk
Hence, by (4.85), (4.92), and the Hölder regularity of f , we get

νk ≤ C diam(Sk )α ≤ C2−kα diam(S0 )α . (4.94)

Since the coefficients ai j are uniformly elliptic and of class C 1 inside √1h Ŝk+3 (thanks
k
to (4.91) and Remark 1.1), it follows by (4.90), (4.93), and Theorem A.40 that

k v̄k − ṽk k   ≤ Cνk ,

C 2 √1 Ŝk+4
hk

which in terms of the functions v̂k gives

k v̂k − v̂k+1 kC 2 (Ŝk+4 ) ≤ Cνk . (4.95)

In particular, since x̂ belongs to Ŝk for all k, recalling (4.94), we deduce that

kD2 v̂k ( x̂) − D2 v̂k+1 ( x̂)k ≤ Cνk ≤ C2−kα diam(S0 )α .

Noticing that by the inductive hypothesis, the above inequality must be true for
all functions { v̂i }0≤i ≤k , we obtain
k
Mk+1 = kD2 v̂k+1 ( x̂)k ≤ kD2 v̂0 ( x̂)k + kD2 v̂i ( x̂) − D2 v̂i+1 ( x̂)k
Õ

i=0
k
≤ M̂ + C diam(S0 )α
Õ
2−kα
i=0
diam(S0 )α
≤ M̂ + C .
1 − 2α
Hence, by choosing ĥ small enough so that diam(S0 ) is sufficiently small (notice that,
since the strict convexity of u is “universal”, the smallness of ĥ depends only on n,
λ, and Λ; see Theorem 4.16), we can ensure that
diam(S0 )α
C ≤ 1,
1 − 2α
136 4 Interior regularity of weak solutions

which proves that

Mk+1 = kD2 v̂k+1 ( x̂)k ≤ M̂ + 1.
Thus,
B1/R ( x̂) ⊂ √ 1 Ŝk+1 ⊂ BR ( x̂)
h k+1

by Lemma 4.17. This shows that √1 Ŝk satisfies (4.89) for all k ≥ 0, and conse-
hk
quently, the induction argument yields that Mk ≤ M̂ + 1 holds for all k, as desired.
2,α
Step 3: u ∈ Cloc . As shown in the previous step, (4.89) holds for all sections √1 Ŝk .
hk
So, in particular,
diam Ŝk


√ ≤ 2R ∀ k ≥ 1.
hk
Thanks to Lemma 4.41, this implies that

kD2 û( x̂)k ≤ C(n, Λ, R);

therefore, recalling that û = u ◦ L̂ and x̂ = L̂ x,

kD2 u(x)k ≤ C(n, Λ, R)k L̂ k 2 .

Since L̂ is the affine transformation that normalizes S(x, 0, ĥ) and ĥ has been fixed
universally, it follows by Corollary 4.7 that k L̂ k is also universally bounded. Hence,

kD2 u(x)k ≤ C.

Since the point x ∈ Z3/4 was arbitrary, this proves that the solution u is uniformly
C 1,1 inside Z3/4 , and because

dist Z1/2, ∂ Z3/4 ≥ ĉ > 0

(thanks to Corollary 4.5), we can apply Proposition A.43 and Theorem A.42 to deduce
that u ∈ C 2,α (Z1/2 ), as desired. 
Remark 4.44. Notice that if f ∈ C m,α (Z) for some m ≥ 0, then Theorem A.42
actually implies that u ∈ C m+2,α (Z1/2 ) (see Step 3 in the proof above).
Remark 4.45. In the proof of the previous theorem, we have decided to use the
Evans–Krylov theorem because it shortens the final step. It is worth noticing that one
could give a direct proof of C 2,α regularity using only the properties of the functions
vk constructed in Steps 1–2 (see [66, Proof of Theorem 1]).
Although this may seem conceptually better (as it would give a “self-contained”
proof of C 2,α regularity), at least with respect to how things have been presented in
this book, this would not really be the case: indeed, our proof of interior estimates for
 4.11 Wang’s counterexamples 137

solutions of Monge–Ampère with constant right-hand side (see Theorem 3.9) used
the Evans–Krylov theorem. So, in this way, our proof already relies on it. For this
reason, we did not see any conceptual improvement in avoiding the Evans–Krylov
theorem in Step 3 above.
Remark 4.46. By a careful inspection of the proof of Theorem 4.42, one can check
that the C 2,α norm of u is controlled by a polynomial function of k f kC 0, α : more
precisely, under the assumptions in Theorem 4.42, for any α ∈ (0, 1), there exist
constants C, γ > 0, depending only on n, λ, Λ, and α, such that
  1+γ
kukC 2, α (Z1/2 ) ≤ C 1 + k f kC 0, α (Z) .

As shown in [46], this nonlinear dependence is sharp (i.e., the above estimate is false
with γ = 0).

4.11 Wang’s counterexamples

1,α
In [124], Wang showed, by a series of counterexamples, that both the Cloc and the
2, p
Wloc estimates obtained in the previous sections are sharp. We illustrate just some
of these counterexamples, referring to the original paper for more detail.
Wang’s basic idea is to consider two-dimensional homogenous solutions defined
in a neighborhood of the origin. More precisely, let us first assume that u satisfies
u(λx, µy) = λα µβ u(x, y).
Then,
λ2 µ2 det D2 u |(λx,µy) = λ2α µ2β det D2 u |(x,y) .

Hence, if we are looking for a determinant of size 1 at all points, we should take
α = β = 1 (compare with Section 3.2).
We now rewrite λ = γ 1/a and µ = γ 1/b for some γ, a, b > 0, so that the condition
on u becomes
1 1
u(γ 1/a x, γ 1/b y) = γu(x, y) with + = 1,
a b
and we look for special solutions that depend homogeneously on one variable and on
some ratio of the two variables. We basically have two choices:
   
x y
u(x, y) = y b f b/a or u(x, y) = x a g a/b
y x
for some f , g : R → R (here and in what follows we use the convention x a = |x| a
and similarly for y).
138 4 Interior regularity of weak solutions

Since we do not want the arguments of f and g to go to infinity, the most natural
choice is to construct u by gluing two functions, but use each of the two choices
only in the region where its argument is bounded by 1. This suggests the following
definition:  
 y b f x/y b/a if |x| ≤ |y| b/a,


u(x, y) :=

 
 x a g y/x a/b
 if |x| ≥ |y| b/a .

Then we impose some compatibility conditions. First, we want the function u to be
continuous and C 1 across the interphase {|x| = |y| b/a } to make sure that we do not
create any Monge–Ampère measure there (compare with Section 3.2). In terms of f
and g, these conditions become (recall that b = a/(a − 1))

f (1) = g(1) and f 0(1) = ag(1) + (1 − a)g 0(1). (4.96)

Second, we compute the determinant of D2 u and we obtain, depending on the two

regions,
(
2
(b − 1) f 00(t) b f (t) − (b − 2)t f 0(t) − f 0(t) , t = x/y b/a,


2

det D u (x, y) =
2
(a − 1)g (t) ag(t) − (a − 2)tg (t) − g (t) ,
00 0 0 t = y/x a/b .

As in Section 3.2, we now try to look for very special functions, the easiest choice
being polynomials of degree 2. Hence, we consider

f (t) := A + Bt 2 and g(t) := 1 + Ct 2,

so that (4.96) becomes

A+B = 1+C and 2B = a + (2 − a)C.

Solving this linear system in terms of B, we get

2B − a 2 − 2a + aB
C= and A= ,
2−a 2−a
and one has to play with the choices of B and a in order to make sure that det D2 u is
bounded away from zero and infinity.
For instance, if we take a  1 and B ≈ 1/a (so that aB is bounded), then
a 2 + o(1)
b= = 1 + o(1), A = 2 + o(1), and C =1+
a−1 a
(here, o(1) is a quantity which goes to 0 as a → ∞). Therefore,
4B
(
2
− 4B2 t 2 + o(1)
a2

, t = x/y b/a,
det D u (x, y) = a
2a a(1 − t 2 ) + 2t 2 + o(1), t = y/x a/b .
 4.11 Wang’s counterexamples 139

1
In particular, choosing B = 2a , we obtain that the C 1 function u defined by

a2 −1 2−a 2
(
xa + a(a−2) x y if {|x| a−1 ≥ |y|},
u(x, y) := 4a−5 a/(a−1) 1 (a−2)/(a−1) 2
2(a−2) y + 2a y x if {|x| a−1 ≤ |y|}

is strictly convex and satisfies 0 < C1 (a) ≤ det D2 u ≤ C2 (a) in the Alexandrov sense.
However, for any ρ > 0,
∫ ρ ∫ ρ  ∫ ρ ∫ ρ 
a
p p ( a−1 −2)
|∂yy u| dy dx & y dy dx
0 x a−1 0 x a−1
∫ ρ
x (a−1)[ p( a−1 −2)+1] dx
a
≈
∫0 ρ
a
= x p(2−a)+a−1 dx = +∞ ∀ p > ,
0 a − 2
2, p 4a−5 a/(a−1)
which proves that u < Wloc for p > a−2 . Also,
a
since u = 2(a−2) y on the axis
1,α
{x = 0}, we see that u < for α
Cloc > a−1
a
.
This family of examples shows the optimality of the results obtained in Sec-
tions 4.5 and 4.8.
5

Further results and extensions

In this chapter, we briefly describe without proofs some further results concerning
the Monge–Ampère equation, as well as the extension of some of the results in the
previous chapters to a general class of Monge–Ampère-type equations.

5.1 Further results on the Monge–Ampère equation

5.1.1 Existence of smooth solutions with general right-hand side. We want to

find smooth global solutions to the equation
(
det D2 u = f (x, u, ∇u) in Ω,
u=g on ∂Ω,

where ∂Ω is a uniformly convex domain of class C 4,α and g : ∂Ω → R is of class C 4,α .

Following the argument used in Section 3.1, we assume that f ∈ C 2,α (Ω × R × Rn )
and that f ≥ 0.
Given a uniformly convex function ū ∈ C04,α (Ω) such that ū = g on ∂Ω (compare
with Section 3.1.5), set f¯ := det D2 ū, and define

fσ (x, z, p) := (1 − σ) f¯(x) + σ f (x, z, p) ∀ σ ∈ R, (x, z, p) ∈ Ω × R × Rn .

To apply the continuity method, we need to show that the set of σ ∈ [0, 1] such that
the equation (
det D2 uσ = fσ (x, uσ, ∇uσ ) in Ω,
(5.1)
uσ = g on ∂Ω

is solvable in C04,α (Ω) ∩ {u : Ω → R : u convex}, is both open and closed.

For openness, the linearization of (5.1) is given by

d  
det D2 uσ + εh) − fσ (x, uσ + εh, ∇uσ + ε∇h)
dε ε=0
= det(D2 uσ )uσ ∂i j h − ∂z fσ (x, uσ, ∇uσ )h − ∇ p fσ (x, uσ, ∇uσ ) · ∇h;
ij
142 5 Further results and extensions

hence, we need the unique solvability in C 4,α (Ω) of the linear equation
(
ai j,σ ∂i j h + bi,σ ∂i h + cσ h = H in Ω,
(5.2)
h=0 on ∂Ω,

for any function H ∈ C 2,α (Ω), where

ai j,σ := fσ uσ ∈ C 2,α (Ω) are uniformly elliptic,

ij

bi,σ := −∂pi fσ (x, uσ, ∇uσ ) = −σ∂pi f (x, uσ, ∇uσ ),

and
cσ := −∂z fσ (x, uσ, ∇uσ ) = −σ∂z f (x, uσ, ∇uσ ).

While the presence of the gradient term bi,σ ∂i h does not play any substantial role,
the zero-order term cσ h influences the validity of the maximum principle (hence, the
uniqueness of solutions; see [58, Theorem 3.7]). Indeed, as shown, for instance, in
[58, Theorems 6.8 and 6.19], the unique solvability of (5.2) in C 4,α (Ω) holds provided
ai j,σ, bi,σ, cσ ∈ C 2,α (Ω) and cσ ≤ 0. Thus, in terms of f , we need to assume that
∇ p f , ∂z f ∈ C 2,α and ∂z f ≥ 0 to ensure that openness holds.
For closedness, one needs to establish global a priori C 2 estimates. The proof of
C 0 and C 1 bounds are analogous to the case f = f (x). Concerning the C 2 bound,
by a barrier argument similar to the one used in Section 3.1.5, one first establishes a
universal bound kD2 uk L ∞ (∂Ω) ≤ K. Then to estimate D2 u in the interior, one defines
2 /2
M := sup eµ | ∇u(x)| ∂ξ ξ u(x),
x ∈Ω,ξ ∈S n−1

where µ > 0 is a positive constant to be suitably chosen. Because of the universal

bound at the boundary, if M > K, then the sup is attained at some interior point, and
the same computation as in Pogorelov’s argument (in this case, it is actually easier,
since we do not have the extra term (−u)a as in Step 3 in the proof of Theorem 3.9)
shows that M is universally bounded.
So, noticing that Theorem A.41 holds also when the equation depends on u and
∇u (see, for instance, [58, Chapter 17.8]) one obtains the following result:

Theorem 5.1. Fix α ∈ (0, 1), assume that Ω is a uniformly convex domain of class
C 4,α , and consider f ∈ C 2,α (Ω × R × Rn ) with

∇ p f ∈ C 2,α (Ω × R × Rn ; Rn ), ∂z f ∈ C 2,α (Ω × R × Rn ),

f (x, z, p) ≥ c0 > 0, and ∂z f (x, z, p) ≥ 0 ∀ (x, z, p) ∈ Ω × R × Rn .

 5.1 Further results on the Monge–Ampère equation 143

Then for any g ∈ C 4,α (∂Ω), there exists a unique solution u ∈ C 4,α (Ω) to the Dirichlet
problem (
det D2 u = f (x, u, ∇u) in Ω,
u=g on ∂Ω.

5.1.2 Boundary regularity. As we saw in Chapter 4, the interior regularity theory

for Alexandrov solutions relies on several geometric properties of sections of u that are
strictly contained inside the domain Ω where u solves the Monge–Ampère equation.
In particular, a consequence of Lemmas A.13 and 4.6 is that any section Sh (x) =
S(x, ∇u(x), h) contained inside Ω is comparable to an ellipsoid of volume h n/2 .
In order to develop a boundary regularity theory, it is crucial to understand the
geometry of sections Sh (x) when x ∈ ∂Ω. This has been done in [102]. Let us
describe the main results.
Assume that Ω ⊂ Rn is a bounded open convex set satisfying

Bρ (ρen ) ⊂ Ω ⊂ B1/ρ (0) ∩ {xn > 0} (5.3)

for some ρ > 0 and that u : Ω → R satisfies

µu = f dx in Ω, (5.4)

for some 0 < λ ≤ f ≤ Λ. Extend u by letting it equal +∞ in Rn \ Ω, and up to

subtracting a linear function, assume that `(x) ≡ 0 is the tangent plane to u at 0,
that is,
u ≥ 0, u(0) = 0, and u(x)  εxn ∀ ε > 0.
The main result in [102] shows that if u ≈ |x| 2 along ∂Ω∩ {xn ≤ ρ}, then the sections
Sh (0) = {x ∈ Ω : u(x) < h} are comparable to half-ellipsoids for h small. More
precisely, the following holds:
Theorem 5.2. Let Ω ⊂ Rn be a bounded open convex set satisfying (5.3), let
u : Ω → R be a convex function satisfying (5.4) for some 0 < λ ≤ f ≤ Λ, and
assume that
1
β|x| 2 ≤ u(x) ≤ |x| 2 on ∂Ω ∩ {xn ≤ ρ},
β
for some β ∈ (0, 1). Then there exists an ellipsoid Eh of volume h n/2 such that
1 

Eh ∩ Ω ⊂ Sh (0) ⊂ K Eh ∩ Ω,
K
where K > 1 depends only on√ n, λ, Λ, ρ, and β. In addition, the ellipsoid Eh is
comparable to a ball of radius h, up to a possible translation along the xn -direction
144 5 Further results and extensions

of size | log h|. Specifically, there exists a linear transformation Ah : Rn → Rn of the

form

Ah (x) = x − γxn, γ = (γ1, . . . , γn−1, 0) ∈ Rn, and |γ| ≤ K | log h|,

such that Eh = Ah B√h (0) .

The last part of the above result provides information about the behavior of the
second derivatives near the origin. Indeed, heuristically, this result states that inside
Sh (0), the tangential second derivatives are uniformly bounded both from above
and below, while the mixed second derivatives are bounded by | log h|. This is very
interesting given that µu is only bounded from above and below, and that the boundary
data as well as the boundary are only C 1,1 .
As a consequence of Theorem 5.2 and the interior estimates proved in Chapter 4,
in [103, 104], Savin obtained the following global C 2,α -W 2, p estimates:
Theorem 5.3. Let Ω ⊂ Rn be a bounded open uniformly convex set, let u : Ω → R
be a convex function satisfying (5.4) for some 0 < λ ≤ f ≤ Λ, and assume that both
u|∂Ω and ∂Ω are of class C 1,1 . Suppose also that u separates quadratically on ∂Ω
from its tangent plane, that is,

u(z) − u(x) ≥ h∇u(x), z − xi + β|z − x| 2 ∀ x, z ∈ ∂Ω.

Then,
• there exists γ > 0 such that u ∈ W 2,1+γ (Ω);
• for any σ > 1, if k f − 1k L ∞ (Ω) ≤ e−Cσ , then u ∈ W 2,σ (Ω);

• assume that f ∈ C 0,α (Ω) and that both u|∂Ω and ∂Ω are of class C 2,α ; then
u ∈ C 2,α (Ω).
As observed in [102], the assumption that u separates quadratically on ∂Ω from
its tangent plane is verified, for instance, whenever ∂Ω and u|∂Ω are of class C 3 with
Ω uniformly convex.

5.1.3 Singular solutions. As already discussed, the strict convexity of solutions is

crucial to prove regularity results. In particular, Pogorelov’s example (see Section 3.2)
shows that one can find a function u solving det D2 u = f with f positive and smooth,
but only C 1,1−2/n and belonging to W 2, p only for p < n(n − 1)/2. Another example
showing the importance of strict convexity is the following:

u(x 0, xn ) = |x 0 | + |x 0 | n/2 (1 + xn2 ), n ≥ 3.

 5.1 Further results on the Monge–Ampère equation 145

It solves µu = f dx with 0 < λ ≤ f ≤ Λ. Yet, u is merely Lipschitz. In [18],

Caffarelli generalized these examples to solutions that degenerate along subspaces,
and he proved that solutions can degenerate only on subspaces of dimension less
than n/2.
Since one cannot hope for C 1 regularity for non-strictly convex solutions, it is
natural to ask whether one can obtain some integrability estimates for the second
derivatives. In Section 4.8, we showed that strictly convex solutions of λ dx ≤ µu ≤
2,1+γ
Λ dx are Wloc for some γ = γ(n, λ, Λ) > 0. If one denotes by Σ the “singular set”
of points where u is not strictly convex, that is,

Σ := {x ∈ Ω : ∃ z ∈ Ω and p ∈ ∂u(x) s.t. u(z) = u(x) + hp, z − xi},

then one may wonder whether the second derivatives of u can concentrate on Σ.
This fact has been recently ruled out by Mooney [93] who showed that the (n − 1)-
dimensional Hausdorff measure of Σ vanishes. From this, he deduced the W 2,1
regularity of solutions without any strict convexity assumptions. Actually, in a sub-
sequent paper [94], he was able to improve this result by showing a small logarithmic
integrality improvement and proving that such a result is optimal.

Theorem 5.4. Let Ω ⊂ Rn be an open set, and u : Ω → R a convex function

2,1
satisfying (5.4) for some 0 < λ ≤ f ≤ Λ. Then Hn−1 (Σ) = 0 and u ∈ Wloc (Ω). In
addition, there exists ε = ε(n) > 0 such that
∫

ε
kD2 uk log 2 + kD2 uk dx < ∞ ∀ Ω0 ⊂⊂ Ω.
Ω0

On the other hand, if M > 0 is sufficiently large, one can construct a solution u such
that ∫
kD2 uk log 2 + kD2 uk dx = +∞ for some Ω0 ⊂⊂ Ω.

M
Ω0

5.1.4 Partial Legendre transform. As observed in Remark A.28, the Legendre

transform allows one to exchange upper and lower bounds on the determinant of the
Hessian, a fact used in the proof of Theorem 2.19 to show the strict convexity of
two-dimensional solutions with positive right-hand side. We now briefly describe
another transformation, the so-called partial Legendre transform, that was introduced
by Darboux at the beginning of the 19th century. It has been used by many authors
to convert the two-dimensional Monge–Ampère equation into a quasilinear elliptic
system.
To describe this, consider u : R2 → R a smooth uniformly convex function,
denote a point in R2 by (x, y) ∈ R × R, and use subscripts to denote partial derivatives
146 5 Further results and extensions

(u x = ∂x u etc.). For any fixed y ∈ R, we perform a Legendre transform in the

x-variable. More precisely, for any p ∈ R, we set

u?(p, y) := sup px − u(x, y).

x ∈R

Since the function u(·, y) is uniformly convex, it is possible to prove that there is a
unique x = X(p, y) where the above supremum is attained. This is characterized by
being the unique point x where u x (x, y) = p, in other words,

u x X(p, y), y = p. (5.5)

Moreover, one can prove that the map X : R2 → R depends smoothly on both p
and y.
To find an equation for u?, we notice that

u?(p, y) = pX(p, y) − u X(p, y), y ;

hence, differentiating this identity with respect to both p and y and using (5.5), we
get
u?y (p, y) = −uy X(p, y), y and u?p (p, y) = X(p, y). (5.6)

To compute the second derivatives of u?, we first differentiate (5.5) with respect to
both p and y to get
u xx X(p, y), y X p (p, y) = 1 (5.7)

and
u xx X(p, y), y X y (p, y) + u xy X(p, y), y = 0. (5.8)

Then, differentiating the first equation in (5.6) with respect to y and the second
equation with respect to p, it follows by (5.7) and (5.8) that

u?yy (p, y) = −u xy X(p, y), y X y (p, y) − uyy X(p, y), y

 2
u xy

=− + uyy X(p, y), y (5.9)


u xx
and
1
u?pp (p, y) = X p (p, y) =
. (5.10)
u xx X(p, y), y

Therefore, assuming that det D2 u = u xx uyy − u2xy is equal to some function f ≥ 0,

using (5.9), (5.10), and the second equation in (5.6), we deduce that u? solves the
quasilinear elliptic equation

f u?p (p, y), y u?pp (p, y) + u?yy (p, y) = 0. (5.11)

 5.1 Further results on the Monge–Ampère equation 147

Because of the simple form of (5.11), the partial Legendre transform is a very useful
tool to obtain a priori estimates for two-dimensional solutions of Monge–Ampère,
and we refer to [108] for more details and references.
We conclude this section by showing how the partial Legendre transform can
be used to prove Liouville-type theorems in two dimensions. Assume one wants to
classify all solutions of

det D2 u = 1 in R × (0, ∞) (5.12)

with boundary conditions

x2
u(x, 0) = on R. (5.13)
2
In this case, it follows by [102, 105] (together with Theorem 2.19 and Theorem 3.10
for interior regularity) that u is smooth and uniformly convex.
Taking the partial Legendre transform in the x-variable, it follows from (5.11)
that u? solves
p2
∆u? = 0 in R × (0, ∞) and u?(p, 0) = on R, (5.14)
2
where we used that the Legendre transform of x 7→ x 2 /2 is p 7→ p2 /2.
To classify all the solutions of (5.14), we differentiate u? twice with respect to p
to get
∆u?pp = 0 in R × (0, ∞) and u?pp (p, 0) = 1 on R.
Noticing that u?(·, y) is convex (since the Legendre transform preserves convexity),
we see that u?pp ≥ 0 on R × (0, ∞); thus, v := u?pp − 1 satisfies

∆v = 0 in R × (0, ∞), v=0 on R, and v ≥ −1 in R × (0, ∞).

Consider now the harmonic function w(p, y) := − log x 2 + (y + 1)2 , and observe
p

that
w(p, 0) ≤ 0 = v(p, 0) on R
and
w ≤ − log ε 2 ≤ log ε 2 v on ∂B1/ε (0) ∩ R × (0, ∞) ,

for any ε > 0. Hence, it follows from the maximum principle that
1
inside B1/ε (0) ∩ R × (0, ∞) ,


v≥ w
log ε 2
and letting ε → 0, we deduce that v ≥ 0 inside R × (0, ∞). So, v solves

∆v = 0 in R × (0, ∞), v = 0 on R, and v ≥ 0 in R × (0, ∞). (5.15)

148 5 Further results and extensions

It is now a classical fact that the only solutions to this equation are of the form
v(p, y) = Ay for some A ≥ 0 (see, for instance, [11, Theorem 1]). Therefore,

u?pp (p, y) = 1 + Ay,

and it follows that u? is of the form

p2
u?(p, y) = h1 (y) + ph2 (y) + (1 + Ay)
2
for some functions h1, h2 : [0, ∞) → R. Recalling (5.14), we see that

h1 (0) = h2 (0) = 0 and h100(y) + ph200(y) + (1 + Ay) = 0 on R × (0, ∞).

This implies that h100(y) + (1 + Ay) = 0 and h200(y) = 0, and integrating these ODEs
we obtain
y 2 Ay 3 p2
u?(p, y) = By − − + Cpy + (1 + Ay)
2 6 2
for some constants B, C ∈ R. Recalling that the Legendre transform is an involution
on convex functions (see Appendix A.4.2), we recover u by taking the partial Legendre
transform of u?:
y 2 Ay 3 (x − C y)2
u(x, y) = sup px − u?(p, y) = + − By + .
p ∈R 2 6 2(1 + Ay)

This gives us a complete classification of all solutions to (5.12)–(5.13).

It is interesting to observe that while the only global solutions of det D2 u = 1 are
quadratic polynomials (see Theorem 4.18), in the half-space there are also solutions
with cubic growth at infinity. However, if one assumes that u grows at most quadrat-
ically at infinity, then A = 0 and u is again a quadratic polynomial. We refer to [105]
for analogous Liouville-type results in higher dimensions.

5.2 The linearized Monge–Ampère equation

In Section 4.9, we described how the regularity theory for the Monge–Ampère
equation allows one to prove global existence of weak solutions to the semigeostrophic
equations. Basically, in that case, one has a time-dependent family of Monge–Ampère
equations
det D2 ut = ft
(see (4.83), and recall that the subscript t does not denote a time derivative but only
expresses dependence on the variable t). The goal is to understand the regularity of
the solutions ut in both space and time.
 5.2 The linearized Monge–Ampère equation 149

If one differentiates the above equations with respect to time, using (A.2), one
deduces that vt := ∂t ut is a solution of

cof(D2 ut )i j ∂i j vt = ∂t ft ,

that is, vt solves a linear elliptic equation whose coefficients ai j := cof(D2 ut )i j may
not be uniformly elliptic (indeed, the uniform ellipticity of ai j would correspond to
the fact that ut ∈ C 1,1 ; see (A.4) and Remark 1.1).
This motivates the following problem: let u solve det D2 u = f and let v solve the
linear equation
Lu (v) := cof(D2 u)i j ∂i j v = g.
What can we say about the regularity of v in terms of u, f , and g? (Here we assume
that both u and v are of class C 2 , and we are looking for a priori estimates.)
To understand how to get regularity estimates for v, one needs to think that the
geometry of the space is now dictated by the function u through its sections. To
explain this point, consider first a linear equation of the form

L(v) := ai j ∂i j v = g in Ω,

where the coefficients ai j are uniformly elliptic, that is,

γ1 |ξ | 2 ≤ ai j (x)ξ i ξ j ≤ γ2 |ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω,

for some 0 < γ1 ≤ γ2 . Then for any quadratic polynomial of the form
1
P(x) := a + hb, xi + |x| 2,
2
we have that L(P) ≈ 1, and the sublevel sets of P correspond (as a and b vary) to all
the balls in Rn . One of the key steps in the regularity theory for elliptic equations
with uniformly elliptic coefficients is the so-called Harnack inequality. It states that
if L(v) = 0 and v ≥ 0 inside some ball B2r (x) ⊂⊂ Ω, then

sup v ≤ C inf v,
Br (x) Br (x)

where C = C(n, γ1, γ2 ) > 1. This is a very strong property which can be seen as a
quantitative version of the strong maximum principle: if v vanishes at some point
inside Br (x), then the Harnack inequality implies that v is identically zero. Also, a
classical consequence of the Harnack inequality is the Hölder continuity of solutions
(see, for instance, [63, Corollary 4.18]).
Now, motivated by what we just said, consider a function v solving Lu (v) = 0,
where u is a solution of the Monge–Ampère equation with right-hand side bounded
150 5 Further results and extensions

away from zero and infinity. Since Lu (u − `) = Lu (u) = det D2 u = f (see (A.4)) for
any affine function `, if we assume that 0 < λ ≤ f ≤ Λ, we see that Lu (u − `) ≈ 1.
Hence, the functions u − ` play the role of the quadratic polynomials P defined above.
This suggests that the natural analogue of balls, in this context, is the sections of u
(or equivalently the sublevel sets of u − ` as ` varies), and the natural result to expect
is that, if Lu (v) = 0 and v ≥ 0 inside some section S2h (x) ⊂⊂ Ω, then
sup v ≤ C(n, λ, Λ) inf v.
Sh (x) Sh (x)

This result has indeed been proved in [22], where the authors developed a general
Calderon–Zygmund theory for sections of solutions to Monge–Ampère (see also
[21]), and it has been recently extended in [87] to the case when lower-order terms
are present. Other results in this direction include the global Hölder and W 2, p
regularity for solutions [78, 77], as well as the validity of a Liouville theorem
(i.e., affine functions are the only global solutions having linear growth) in two
dimensions [101].

5.3 A general class of Monge–Ampère-type equations

In Section 5.1.1, we described the regularity of solutions to Monge–Ampère equations

where the right-hand sides depend also on the lower-order terms u and ∇u. However,
lower-order terms may appear not only in the right-hand side but also inside the
determinant, and this gives rise to a general class of Monge–Ampère-type equations.

5.3.1 A motivation: Optimal transport maps with general costs. Let c : Rn ×

Rn → R be a smooth function representing the transportation cost, that is, c(x, y)
corresponds to the cost of transporting a unit mass from x to y. The optimal transport
problem with cost c consists of finding a transport map T that minimizes the total
transportation cost. Precisely, find a map T : Rn → Rn such that
∫ ∫ 
c(x, T(x)) dµ(x) = inf c(x, S(x)) dµ(x) .
Rn S# µ=ν Rn

As we saw in Section 4.6, when the cost is quadratic, then the optimal transport map
T is given by the gradient of a convex function u, and it follows from the Jacobian
equation (4.39) that u solves the Monge–Ampère equation (4.40).
A useful remark to understand what happens with general costs is that the cost
|x − y| 2 is equivalent to the cost −hx, yi. Indeed, by (4.38), for any transport map S,
∫ ∫
|S(x)| 2 dµ(x) = |y| 2 dν(y);
Rn Rn
 5.3 A general class of Monge–Ampère-type equations 151

hence,
∫
|S(x) − x| 2 dµ(x)
Rn
∫ ∫ ∫
2 2
= |S(x)| dµ(x) + |x| dµ(x) + 2 −hx, S(x)i dµ(x)
n n n
∫R ∫ R ∫ R
= |y| 2 dν(y) + |x| 2 dµ(x) + 2 −hx, S(x)i dµ(x).
Rn Rn Rn

Thus, since the first two integrals in the right-hand side are independent of S and,
of course, the multiplicative factor 2 does not play any role, we see that the two
minimization problems
∫ ∫
2
min |S(x) − x| dµ(x) and min −hx, S(x)i dµ(x)
S# µ=ν Rn S# µ=ν Rn

are equivalent, as they have the same set of minimizers.

As observed in Appendix A.4.2, a characterization of convex functions is that
the Legendre transform acts on them by involution, that is, a (lower semicontinuous)
function u is convex if it can be written as

u(x) = sup hx, yi − u∗ (y),

y ∈R n

where u∗ (y) is defined as in (A.16). In addition, the set of y ∈ Rn where the above
supremum is attained coincides with the subdifferential of u at x (see (A.21)).
Hence, regarding −hx, yi as a cost, the following definition comes quite naturally.
Definition 5.5. Given an open set Ω ⊂ Rn , a function u : Ω → R is c-convex if

u(x) = sup −c(x, y) − uc (y),

y ∈R n

where the function uc is defined as

uc (y) := sup −c(x, y) − u(x).

x ∈Ω

If u is c-convex, we define its c-subdifferential as the set of points y where the

supremum above is attained, that is,

∂ c u(x) := {y ∈ Rn : u(x) + uc (y) + c(x, y) = 0}.

Observe now that whenever y ∈ ∂ c u(x), the function

z 7→ u(z) + uc (y) + c(z, y)

152 5 Further results and extensions

attains a minimum at z = x. In particular, if u is differentiable at x (recall that by

assumption, c is smooth), then by differentiating at z = x, we get

0 = ∇u(x) + ∇x c(x, y) ∀ y ∈ ∂ c u(x). (5.16)

Notice that when c(x, y) = −hx, yi, the above relation reduces to y = ∇u(x), which
proves that the subdifferential is a singleton at differentiability points (compare with
Lemma A.20), and this is a key fact behind the proof of Theorem 4.22.
For this reason, in order to show the existence and uniqueness of optimal transport
maps for general cost, it is important to assume that the c-subdifferential is a singleton
at differentiability points, at least when y belongs to the support of the target measure
ν. So, we consider X, Y ⊂ Rn two open sets such that µ and ν are supported inside
X and Y respectively, and we assume that

the map Y 3 y 7→ ∇x c(x, y) ∈ Rn is injective ∀ x ∈ X. (5.17)

This assumption ensures that (5.16) uniquely identifies y in terms of x and ∇u(x),
and this allows one to prove that optimal transport maps exist and are unique (see,
for instance, [33, Theorem 3.6] for an idea of the proof and [123, Chapter 10] for a
complete and much more general result):

Theorem 5.6. Let X, Y ⊂ Rn be two bounded open sets, c : X ×Y → R be a function

of class C 1 , and let µ = f dx and ν = g dy be two probability densities supported on
X and Y respectively. Assume that (5.17) holds. Then,

• there exists a unique optimal transport map T;

• there exists a c-convex function u : Rn → R such that T(x) ∈ ∂ c u(x) and

∇u(x) + ∇x c(x, T(x)) = 0 (5.18)

for µ-a.e. x.

5.3.2 The MTW condition and the regularity of optimal transport maps. To
understand the regularity of optimal maps with general cost, we start (as for the
quadratic case) from the Jacobian equation

f (x)
| det ∇T(x)| =
g(T(x))

(see (4.39)). Since T(x) ∈ ∂ c u(x) (see Theorem 5.6), the map

z 7→ u(z) + uc (T(x)) + c(z, T(x))

 5.3 A general class of Monge–Ampère-type equations 153

attains a minimum at z = x. Thus, whenever u is twice differentiable at x, we get

D2 u(x) + D2xx c(x, T(x)) ≥ 0. (5.19)

Also, differentiating the relation (5.18) with respect to x, we obtain

D2 u(x) + D2xx c(x, T(x)) = −D2xy c(x, T(x)) · ∇T(x).

Then, taking the determinant of both sides and using the Jacobian equation for T, we
get (recall (A.3))
f (x)
det D2 u(x) + D2xx c(x, T(x)) = det D2xy c(x, T(x)) , (5.20)



g(T(x))
where we have used (5.19) to remove the modulus in the left-hand side.
Now, assuming that (5.17) holds, it follows from (5.18) that T can be uniquely
written in terms of x and ∇u, that is, there exists a map

X × Rn 3 (x, p) 7→ T (x, p) ∈ Y

such that T(x) = T (x, ∇u(x)). Hence, substituting this relation into (5.20), we get

det D2 u(x) + A(x, ∇u(x)) = h(x, ∇u(x)), (5.21)

where
A(x, p) := D2xx c x, T (x, p)

and
f (x)
h(x, p) := det D2xy c x, T (x, p) .



g(T (x, p))
Notice that if f and g are bounded away from zero and infinity on their respective
supports, the right-hand side h may still vanish because of the presence of the term
det(D2xy c). So, it makes sense to assume that c satisfies

det D2xy c(x, y) , 0 (5.22)

∀ (x, y) ∈ X × Y,

so that h > 0.
This discussion naturally leads to studying regularity theory for general equations
of the form (5.21) on the class of c-convex solutions. In order to understand the
regularity of solutions, a natural strategy is to try to mimic Pogorelov’s approach for
the classical Monge–Ampère equation (see Theorem 3.9).
Thus, given a c-convex function u solving (5.21), one considers a maximum point
of the second derivatives of u (or a variant of this quantity), and one differentiates
(5.21) once and twice at this maximum point in order to find a universal a priori bound.
154 5 Further results and extensions

Now, when considering second derivatives of (5.21), one of the terms appearing will
involve the Hessian of the matrix A(x, p) in the p-direction, and it is unclear whether
or not this additional term can be controlled (notice that this term was not present in
the classical Monge–Ampère equation, as in that case A ≡ 0).
For this reason, in [86], Ma, Trudinger, and Wang introduced a condition on A
(or equivalently on the cost c) that allowed them to show that the potentially bad term
had the right sign. This condition (now called the MTW condition after the three
authors) reads as follows:

D2pη pη A(x, p)[ξ, ξ] ≥ 0 ∀ x ∈ X, p ∈ −∇x c(x, Y ), ξ, η ∈ Rn × Rn with ξ ⊥ η

(5.23)
(notice that in our application, p = ∇u(x) = −∇x c(x, T(x)), and since T(x) ∈ Y ,
we need the condition above to hold only when p ∈ −∇x c(x, Y )). In addition to
this condition and a geometric assumption on the supports of the measures (which
is the analogue of the convexity assumption of Caffarelli in Theorem 4.23), it is
also important to assume that (5.17) holds when exchanging the role of x and y. In
other words,

the map X 3 x 7→ ∇y c(x, y) ∈ Rn is injective ∀ y ∈ Y . (5.24)

Then, under these assumptions, Ma, Trudinger, and Wang were able to prove the
following result [86, 119, 120] (see also [112]):
Theorem 5.7. Let X, Y ⊂ Rn be two bounded open sets, c : X ×Y → R be a function
of class C ∞ , and µ = f dx and ν = g dy be two probability densities supported on
X and Y , respectively. Assume that (5.17), (5.24), (5.22), and (5.23) hold. Also,
suppose that
(a) X and Y are smooth;
(b) ∇x c(x, Y ) is uniformly convex for all x ∈ X;
(c) ∇y c(X, y) is uniformly convex for all y ∈ Y ;
(d) f and g are smooth and bounded away from zero and infinity on X and Y ,
respectively.
Then u ∈ C ∞ (X), and the optimal map T is a global smooth diffeomorphism of X
onto Y .
Remark 5.8. Apart from being interesting on its own, this general theory of Monge–
Ampère-type equations satisfying the MTW condition turns out to be useful also in
the classical case.
 5.3 A general class of Monge–Ampère-type equations 155

Indeed, as shown in [86, 83, 72], the MTW condition is coordinate invariant. This
implies that if u solves an equation of the form (5.21) with A satisfying the MTW
condition, and Φ is a smooth diffeomorphism, then u ◦ Φ satisfies an equation of the
same form with a new matrix à that still satisfies the MTW condition. In particular,
in proving boundary regularity, one can reduce to a nicer situation (for instance, a flat
boundary) by applying a smooth diffeomorphism to the domain. In particular, this
applies to the classical Monge–Ampère equation (which trivially satisfies the MTW
condition since A ≡ 0): if u solves the Monge–Ampère equation, then u ◦ Φ satisfies
(5.21) for some matrix A satisfying (5.23).
This observation is also extremely natural in relation to the optimal transport
problem. Consider, for instance, the optimal transport problem with cost −x · y
between two smooth densities supported on smooth uniformly convex domains. By
the results described in Section 4.6.1, we know that the optimal map T is smooth. If
one performs smooth changes of coordinates Φ : X → Φ(X) and Ψ : Y → Ψ(Y ), the
cost function becomes −Φ−1 (x) · Ψ−1 (y), and the optimal transport map is given by
T̃ := Ψ ◦ T ◦ Φ−1 , which of course is still smooth. However, if one wants to directly
obtain the smoothness of T̃ (i.e., without using that it comes from T through a change
of variables) one can no longer apply the results in Section 4.6.1, and Theorem 5.7 is
needed.
Although the MTW condition seems to be the right assumption to obtain regularity
of optimal maps, it was only after Loeper’s work [83] that people started to have a
good understanding of and geometric insight into this condition.
Loeper realized that for the classical Monge–Ampère equation, a key property
behind the regularity of convex solutions is the convexity (and so, in particular, the
connectedness) of the subdifferential of a convex function. Indeed, roughly speaking,
this has the following consequence: whenever a convex function u is not C 1 at a point
x0 , there is at least a whole segment contained in the subdifferential of u at x0 , and one
can prove that this is incompatible with the subdifferential preserving the Lebesgue
measure (recall the discussions in Section 4.6.1, and see the proof of Theorem 2.17
for a detailed argument in two dimensions). Hence, Loeper wanted to understand
whether or not the c-subdifferential of a c-convex function is connected, believing
that this has a link with regularity. In [83], he was able to prove that the connectedness
of the c-subdifferential of a c-convex function is equivalent to the MTW condition,
and he showed that the MTW condition is necessary for the regularity of optimal
maps: if the MTW condition fails at one point x ∈ X for some p ∈ −∇x c(x, X) and
some ξ ⊥ η, then one can construct smooth positive probability densities f and g
such that the optimal map is discontinuous (see [33, Section 4.1] for more details).
The geometric insight provided by Loeper’s result has been important not only
in showing that the MTW condition is the right structural assumption on the cost
function to prove regularity of optimal transport maps, but also in obtaining Hölder
156 5 Further results and extensions

and Sobolev regularity estimates for transport maps under weak regularity assump-
tions on the densities (see [83, 120, 79, 50, 80, 48, 81, 31, 59, 68]). For instance, the
following result holds true (see [48]):
Theorem 5.9. Let X, Y ⊂ Rn be two bounded open sets, c : X ×Y → R be a function
of class C ∞ , and µ = f dx and ν = g dy be two probability densities supported on
X and Y , respectively. Assume that (5.17), (5.24), (5.22), and (5.23) hold. Also,
suppose that
(a) ∇x c(x, Y ) is uniformly convex for all x ∈ X;
(b) ∇y c(X, y) is uniformly convex for all y ∈ Y ;
(c) f and g are smooth and bounded away from zero and infinity on X and Y ,
respectively.
1,α
Then u ∈ Cloc (X), and the optimal map T is a homeomorphism of X onto Y .
The study of the regularity of optimal maps with general costs is useful also to
understand the case when the cost c(x, y) = d(x, y)2 /2 is the squared Riemannian
distance on X = Y = (M, g), a compact Riemannian manifold without boundary. In
this case, in addition to the validity of the MTW condition, one needs to face the
additional difficulty that the function d 2 /2 is never smooth on all of M × M due to
the presence of the cut-locus. This creates several additional difficulties which we
shall not discuss here, and we refer to [39, 83, 71, 84, 85, 52, 72, 53, 38, 55, 54, 56,
73, 49, 33] for more details and references.
We conclude this section by mentioning that the partial regularity result described
at the end of Section 4.6.1 can be generalized to general cost functions and one can
show that even if the MTW condition fails, optimal transport maps are always smooth
outside a closed singular set of measure zero [35, 26].

5.3.3 General prescribed Jacobian equations. Equations (2.1) and (5.21) can be
seen as particular cases of prescribed Jacobian equations of the following form:
det ∇[T (x, u, ∇u)] = ψ(x, u, ∇u), (5.25)

where T = T (x, z, p) : Ω × R × Rn → Rn is a smooth map. Assuming that

det(∇ p T ) , 0 and arguing as in Section 5.3.2, one can rewrite (5.25) in the form
det D2 u + A(x, u, ∇u) = f (x, u, ∇u). (5.26)

Note that the classical Monge–Ampère equation corresponds to the case T (x, z, p)=p,
and, more generally, the optimal transport problem described in the previous section
corresponds to the case
T (x, z, p) = T (x, p) and − ∇x c(x, T (x, p)) = p.
 5.3 A general class of Monge–Ampère-type equations 157

Motivated by problems arising in geometric optics, Trudinger began, in [111, 113],

a systematic study of equations of the form (5.25)–(5.26) in the particular case when
T is obtained through a “generating function” G : Rn × Rn × R → R. Let us briefly
present this theory, referring to [113] for more details.
Given a function G : Rn × Rn × R → R, a function u : Ω → R is said to be
G -convex if for every x0 ∈ Ω, there exist y0 and z0 such that

u(x0 ) = −G (x0, y0, z0 ) and u(x) ≥ −G (x, y0, z0 ) ∀ x ∈ Ω. (5.27)

Under suitable invertibility assumptions on G , one can define the maps T = T (x, z, p)
and Z = Z (x, z, p) through the relations

∇x G x, T (x, u, p), Z (x, u, p) = −p and − G x, T (x, u, p), Z (x, u, p) = u,

and with this choice, one gets from (5.27) that y0 = T (x0, u(x0 ), Du(x0 )) whenever
u is differentiable at x0 (compare with (5.16)). Note that the optimal transportation
case corresponds to the choice G (x, y, z) := c(x, y) − z. Then under some necessary
structural conditions on G (which are analogous to the MTW condition in this con-
text), in [111, 113, 67, 60] the authors started to develop a theory parallel to the one
described in Section 5.3.2, although a full regularity (and a partial regularity) theory
in this context is still missing.
Notice that the assumption of the existence of a generating function G ensures that
the matrix A(x, u, ∇u) appearing in (5.26) is symmetric. However, for the general
class of equations (5.26) arising from the prescribed Jacobian equations (5.25), there
is no reason why the matrix A should be symmetric. It would be very interesting
to understand under which structural assumptions on A one can develop a regularity
theory (see [111, Corollary 1.2] for a result in two dimensions).
A

Appendix: Useful facts

A.1 Some general facts from linear algebra

In this section, we show formulas for the derivative of the determinant and of the
inverse of a matrix, and then we gather some useful estimates.
Given a matrix A ∈ Rn×n , we denote its transpose by A∗ and its cofactor matrix
by cof(A).
Lemma A.1. Let A, B ∈ Rn×n , and assume that A is invertible. Then,
d
det(A + εB) = det(A) tr(A−1 B), (A.1)
dε ε=0

or equivalently,
d
det(A + εB) = tr cof(A)∗ B . (A.2)


dε ε=0
In addition, the latter formula holds also when A is not invertible.
Proof. Recall that for any A, B ∈ Rn×n ,

det(AB) = det(A) det(B). (A.3)

Hence, whenever A is invertible, we can write

det(A + εB) = det(A) det(Id +ε A−1 B),

and a direct computation shows that

det Id +ε A−1 B = 1 + ε tr(A−1 B) + O(ε 2 ).

Using this formula, we immediately obtain that

det(A + εB) − det(A) ε tr(A−1 B) + O(ε 2 )

lim = det(A) lim = det(A) tr(A−1 B),
ε→0 ε ε→0 ε
which proves (A.1).
It is interesting to have a formula also when A is not invertible. To this aim, we
recall the identity
det(A)A−1 = cof(A)∗ (A.4)
160 Appendix: Useful facts

(see, for instance, [64, Equation (0.8.2.2)]). Inserting (A.4) in (A.1), we get (A.2).
Although a priori (A.2) is true only when A is invertible, a simple approximation
argument shows that it holds also when A is not invertible. 

Lemma A.2. Let A, B ∈ Rn×n , and assume that A is invertible. Then,

d
(A + εB)−1 = −A−1 BA−1 . (A.5)
dε ε=0

Proof. Differentiating the identity

(A + εB)−1 (A + εB) = Id

with respect to ε, we get

 
d
(A + εB)−1 A + A−1 B = 0,
dε ε=0

proving (A.5). 

The following result relates sum and determinant.

Lemma A.3. Let A, B ∈ Rn×n be symmetric non-negative-definite matrices. Then,

det(A + B)1/n ≥ det(A)1/n + det(B)1/n, (A.6)

so, in particular,
det(A + B) ≥ det(A) + det(B). (A.7)
In addition, if both A and B are positive definite, then equality holds in (A.6) if and
only if A = λB for some λ > 0.

Proof. Notice that by approximation, it suffices to prove the result when A and B are
positive definite, and in this case it is a classical fact in linear algebra that A and B
can be simultaneously diagonalized, that is, there exists an invertible matrix X such
that
A = X ∗ M X, B = X ∗ N X,
M = diag(α1, . . . , αn ), and N = diag(β1, . . . , βn )
where αi, βi > 0 (see, for instance, [64, Theorem 7.6.4]).
We now notice that as a consequence of the arithmetic–geometric inequality,
n  1/n n  1/n n n
αi βi 1 Õ αi 1 Õ βi
Ö Ö
+ ≤ + = 1,
i=1
αi + βi i=1
αi + βi n i=1 αi + βi n i=1 αi + βi
 A.1 Some general facts from linear algebra 161

or equivalently,
n
Ö  1/n n
Ö  1/n n
Ö  1/n
αi + βi αi + βi ,


≥
i=1 i=1 i=1

with equality if and only if there exists λ > 0 such that αi = λβi for all i = 1, . . . , n.
Therefore, using (A.3), we get
  1/n
det(A + B)1/n = det X ∗ (M + N)X
= det(X)2/n det(M + N)1/n
Ön  1/n
= det(X)2/n (αi + βi )
i=1
n
Ö  1/n n
Ö  1/n 
≥ det(X)2/n αi + βi
i=1 i=1
 
= det(X)2/n det(M)1/n + det(N)1/n
= det(A)1/n + det(B)1/n,

which proves (A.6).

Noticing that
(a + b)n ≥ a n + bn ∀ a, b ≥ 0,
(A.7) follows immediately from (A.6) by applying the above inequality with a =
det(A)1/n and b = det(B)1/n . 

Given A ∈ Rn×n , we denote its operator norm by k Ak, that is,

k Ak := sup | Av|.
|v |=1

We conclude this section by showing how bounds from above and below on the
eigenvalues of the symmetric matrix A∗ A give control on the norm of A and A−1 .

Lemma A.4. Let A ∈ Rn×n , and assume that there exists a constant K > 1 such that
1
Id ≤ A∗ A ≤ K Id . (A.8)
K
Then, √ √
k Ak ≤ K and k A−1 k ≤ K. (A.9)
162 Appendix: Useful facts

Proof. Let us recall that by the polar decomposition of matrices, there exist O ∈ Rn×n
orthogonal and S ∈ Rn×n symmetric nonnegative definite such that A = OS (see, for
instance, [64, Theorem 7.3.1]).
Since O ∗ O = Id, substituting the identity A = OS in (A.8), we deduce that

1
Id ≤ S ∗ S = S 2 ≤ K Id .
K
√ √
Hence, all the eigenvalues of S are bounded between 1/ K and K. This yields
√ √
kSk ≤ K and kS −1 k ≤ K,

and noticing that A−1 = S −1 O −1 and that orthogonal transformations preserve the
operator norm, we obtain (A.9). 

A.2 Hausdorff measure

The aim of this section is to define the Hausdorff measure in Rn and collect its basic
properties. Before giving the precise definition of the Hausdorff measure, we first
give a brief informal presentation.
Given s ∈ [0, ∞), the s-dimensional Hausdorff measure of a set E ⊂ Rn is de-
fined as follows: First, fix δ > 0. Then, cover E with sets {E j } j ∈N of diameter
bounded by δ and compute the “s-dimensional volume” of this covering by consid-
ering j diam(E j )s . Finally, find the “best” possible covering. This gives rise to the
Í
definition of Hδs .
However, this preliminary definition has two downsides, one geometric and one
technical. First of all, if a set E has diameter smaller than δ, then a possible covering
is given by taking E1 = E and E j = ∅ for j ≥ 2, which implies that Hδs (E) is always
bounded by a multiple of diam(E)s . This means that at scales smaller than δ, Hδs (E)
will completely miss the geometry of E (for instance, E may be a very long curve
contained inside Bδ/2 (0), and for s = 1 we would like to recover the length of the
curve). Second, Hδs is not additive (still, one can show that Hδs is additive on sets
that are at mutual distance strictly larger than δ).
The solution to both these issues is to let δ → 0 and define Hs := limδ→0 Hδs .
Indeed, by forcing the covering to be as refined as possible, we capture the geometry
of smooth sets (see Proposition A.6(v) below), and one can show that Hs is a Borel
regular measure (see, for instance, [42, Section 2.1, Theorem 1]). As it will be
clear from the definition of Hδs , the existence of a limit as δ → 0 is an immediate
consequence of the fact that the map δ 7→ Hδs (E) is monotone nonincreasing.
 A.2 Hausdorff measure 163

To ensure that, for s integer and E a smooth s-dimensional surface, Hs (E)

coincides with the usual notion of s-dimensional area (see Proposition A.6(v) below),
it is convenient to add a multiplicative constant in front of j diam(E j )s .
Í

Definition A.5. Given E ∈ Rn and 0 ≤ s < ∞, one defines

(i) for any 0 < δ ≤ ∞, the s-dimensional Hausdorff δ-premeasure of E:1

∞ ∞
diam(E j )
Õ  s Ø 
Hδs (E) := inf ωs : E⊂ E j , diam(E j ) ≤ δ ,
j=1
2 j=1

where

π s/2 ∞
∫
ωs := and Γ(t) := e−x x t−1 dx ∀ t > 0;
Γ 2s + 1


0

(ii) the s-dimensional Hausdorff measure of E:

Hs (E) := sup Hδs (E) = lim Hδs (E).

δ>0 δ→0

We remark that the normalizing constant ωs coincides with the Lebesgue measure
of the k-dimensional unit ball in Rk when s = k is an integer. This is particularly
convenient if one wants to ensure that Hn coincides with the Lebesgue measure
on Rn .
In the definition of Hs , the number s plays the role of a dimension. This is
justified by the following facts:

Proposition A.6. The following properties hold:

(i) Hs (λE) = λ s Hs (E) for all λ > 0.

(ii) H0 is the counting measure.

(iii) If s > n, then Hs ≡ 0.

(iv) Hn (E) = |E |, that is, Hn coincides with the Lebesgue measure on Rn .

(v) If s ∈ {1, . . . , n − 1} is an integer and E ⊂ Rn is an s-dimensional Lipschitz

surface, then Hs (E) coincides with the classical s-dimensional area of E.
1 In the definition of Hδs , one is allowed to choose E j = ∅, in which case, diam(E j ) s = 0 for all s ≥ 0.
On the other hand, by convention, diam(E)0 ≡ 1 for any nonempty set E (in particular, if E is a point, then
diam(E)0 = 1).
164 Appendix: Useful facts

Sketch of the proof. We prove just the first three statements (as their proofs are ele-
mentary), and we give precise references for the last two.
Step 1: Proof of (i). This follows easily since {E j } is a covering of E if and only if
{λE j } is a covering of λE, noticing that diam(λE j ) = λ diam(E j ).
Step 2: Proof of (ii). It is easy to see that the best covering of a point x ∈ Rn is
obtained by choosing E1 = x and E j = ∅ for any j ≥ 2 (recall footnote 1). Since
ω0 = 1, this implies that H0 ({x}) = 1 for any point x ∈ Rn , which proves the result.
Step 3: Proof of (iii). Consider the unit cube Q = [0, 1]n , fix N  1, and split Q into
N n of side length 1/N. Since each of these small cubes has
N n disjoint cubes {Q j } j=1
√
diameter n/N, using this covering in the definition of H√s , we get
n/N

Nn √ s
Õ n/N ωs ns/2 1
H√s (Q) ≤ ωs = ,
n/N
j=1
2 2s N s−n

and we discover that Hs (Q) = 0 by letting N → ∞. Since Hs is translation invariant,

we can cover Rn with {Q + ξ}ξ ∈Zn to deduce that Hs (Rn ) = 0, as desired.
Step 4: Proofs of (iv) and (v). The proof of (iv) is rather delicate and relies on the
isodiametric inequality. We refer to [42, Section 2.2, Theorem 2]. Finally, (v) follows
from the area formula for Lipschitz functions (see, for instance, [42, Sections 3.3.2
and 3.3.4.C]). 
We conclude this brief introduction to the Hausdorff measure with a simple lemma
describing its behavior under Lipschitz maps.
Lemma A.7. Let E ⊂ Rn and f : Rn → Rm be a Lipschitz map with Lipschitz
constant L. Then for any s ≥ 0,

Hs f (E) ≤ L s Hs (E).

Proof. Fix δ > 0, and consider an arbitrary covering {E j }∞ j=1 of E with
diam(E j ) ≤ δ.
Then we see that the sets { f (E j )} j=1 cover f (E) and that diam f (E j ) ≤ L diam(E j )
∞

≤ Lδ; hence,
∞ ∞
diam f (E j ) s diam(E j ) s

  

Õ Õ
s
H Lδ f (E) ≤ ωs ≤L s
ωs .
j=1
2 j=1
2

Since the covering {E j }∞

j=1 is arbitrary, this implies that

HsLδ f (E) ≤ L s Hδs (E),

and the result follows letting δ → 0. 

 A.3 Convex geometry 165

Figure A.1. Both H1 and H2 are supporting hyperplanes for Σ. Also, while x1 is an exposed point,
x2 is not.

A.3 Convex geometry

In this section, we recall some general facts from convex geometry.

A.3.1 Exposed points.

Definition A.8. Let Σ ⊂ Rn be a closed convex set. A hyperplane H ⊂ Rn is
supporting if
• Σ is contained in one of the two closed half-spaces bounded by H;

• Σ ∩ H , ∅.
A point x ∈ Σ is exposed if there exists a supporting hyperplane H ⊂ Rn such that

Σ ∩ H = {x}

(see Figure A.1).

The following result shows that for compact convex sets, the collection of exposed
points is large enough to generate Σ (see [99, Theorem 18.7] for a proof):
Theorem A.9. Every compact convex set is the closure of the convex hull of its
exposed points.
While unbounded convex sets may have no exposed points (for instance, think of
a half-space), in such a case they have to contain a whole line (see, for instance, [99,
Corollary 18.5.3 and Theorem 18.6]):
Theorem A.10. Let Σ be a closed convex set, and assume that it has no exposed
points. Then Σ contains a line.
166 Appendix: Useful facts

A.3.2 John’s lemma. John’s lemma is a classical result in convex geometry. It

states that every open bounded convex set is essentially equivalent to an ellipsoid
[69]. Before stating and proving this result, we recall some basic definitions.
Definition A.11. An ellipsoid is the image of the unit ball B1 (0) under an affine map
of the form
Rn 3 z 7→ Az + x ∈ Rn,
where A ∈ Rn×n is invertible and x ∈ Rn . Such an ellipsoid will be denoted by E A, x ,
and it can be equivalently described as

E A,x = z ∈ Rn : A−1 (z − x)| < 1 .



The point x is called the center of E A,x , and the formula

| E A,x | = det(A)|B1 (0)| (A.10)

holds.
Given r > 0, we denote by r E the dilation of an ellipsoid E by a factor r with
respect to its center, or equivalently,

r E A, x = Er A, x .

Definition A.12. Given two Borel sets E, F, one defines their Minkowski semisum as
E+F z1 + z2
 
:= z = : z1 ∈ E, z2 ∈ F .
2 2
Lemma A.13. Let Z ⊂ Rn be an open bounded convex set. Then there exists a unique
ellipsoid E of maximal volume contained in Z. Moreover, this ellipsoid satisfies

E ⊂ Z ⊂ nE . (A.11)

Proof. Let K denote the set of all ellipsoids contained in Z, and set

a := sup | E |.
E ∈K

We split the proof into three steps.

Step 1: There exists an ellipsoid Ê ∈ K of maximal volume. Consider {Ek =
E Ak ,xk }k ∈N ⊂ K, a sequence such that | Ek | → a. Since each ellipsoid Ek is contained
inside Z, it is easy to check that the set of matrices Ak and centers xk defining Ek are
bounded. So, up to subsequences, they converge to a matrix  and a point x̂. Then
by continuity, the ellipsoid Ê := E Â, x̂ associated to them satisfies

Ê ∈ K and | Ê | = a.
 A.3 Convex geometry 167

Figure A.2. If Z contains two ellipsoids E1 and E2 = E1 + (x2 − x1 ) (in grey in the picture), it
suffices to “stretch” E1 in the direction of x2 − x1 to construct a new ellipsoid with larger volume
still contained inside Z.

Step 2: There is a unique ellipsoid in K of maximal volume. Assume that E1 =

E A1,x1 ∈ K and E2 = E A2,x2 ∈ K both satisfy | E1 | = | E2 | = a, and consider their
Minkowski semisum
E1 + E2
E1/2 := = E 1 (A1 +A2 ), 1 (x1 +x2 ) .
2 2 2

Since E1, E2 ⊂ Z, and Z is convex, we have

Z+Z
E1/2 ⊂ = Z;
2
thus E1/2 ∈ K. Hence, it follows from the definition of a, (A.10), and (A.6) that
1 1 
a1/n ≥ | E1/2 | 1/n = | E1 + E2 | 1/n ≥ | E1 | 1/n + | E2 | 1/n = a1/n .
2 2
The equality case in (A.6) implies that A1 and A2 are multiples of each other. Since

a = det(A1 )|B1 (0)| = det(A2 )|B1 (0)|,

it follows that A1 = A2 , and therefore E2 = E1 + (x2 − x1 ). To conclude, it suffices

to observe that because Z contains the convex hull of E1 and E2 (by convexity),
unless x1 = x2 , one can find an ellipsoid contained in Z of strictly larger volume (see
Figure A.2), a contradiction.
Step 3: Z ⊂ nÊ . Up to an affine transformation, we can assume that Ê = B1 (0), and
we want to show that Z ⊂ Bn (0). To prove this, assume, to the contrary, that there
exists x̄ ∈ Z with ρ := | x̄| > n. Up to a rotation, we can assume that x̄ = ρe1 and
consider C the convex hull of B1 (0) and x̄. Notice that C ⊂ Z (since Z is convex),
and
∂B1 (0) ∩ ∂ C = ∂B1 (0) ∩ {x1 ≤ 1/ρ} (A.12)
168 Appendix: Useful facts

Figure A.3. If ρ > n, using a suitable vector field V λ , we can deform B1 (0) into an ellipsoid Etλ that
is still contained inside C and has larger volume than |B1 (0)|.

(see Figure A.3). We claim that there exists an ellipsoid E ⊂ C with

| E | > |B1 (0)|,

which will contradict the maximality of B1 (0) and conclude the proof.
To construct such an ellipsoid, given λ ∈ (0, 1) (to be chosen) we consider the
vector field

V λ (x1, x 0) := (x1 + 1), −λx 0 with (x1, x 0) ∈ R × Rn−1,

and the family of affine maps generated by it: more precisely, solving the ODE
(
∂t Φλt = V λ Φλt ,

Φλ0 = Id,

we obtain  
Φλt (x1, x 0) = et x1 + (et − 1), e−λt x 0 .

Set Etλ := Φλt (B1 (0)). Since

Etλ = E At ,xt , At = diag(et , e−λt , . . . , e−λt ), xt = (et − 1, 0, . . . , 0),

we have

| Etλ | = det(At )|B1 (0)| = et e−(n−1)λt |B1 (0)| > |B1 (0)| ∀ t > 0,
1
provided λ < n−1 . Hence, to conclude the argument, it suffices to show that we can
1
choose λ < n−1 in such a way that Etλ ⊂ C for t > 0 sufficiently small.
For this, we show that on the common boundary ∂B1 (0) ∩ ∂ C , the vector-field V λ
points inside the interior of C (see Figure A.3). So, recalling (A.12), we consider the
 A.3 Convex geometry 169

scalar product of V λ with the outer normal ν∂B to ∂B1 (0): since ν∂B (x1, x 0) = (x1, x 0)
and |x 0 | 2 = 1 − x12 on ∂B1 (0), we get

V λ (x1, x 0) · ν∂B (x1, x 0) = x12 + x1 − λ(1 − x12 ) ∀ (x1, x 0) ∈ ∂B1 (0) ∩ x1 ≤ 1

.

ρ

As the function
x1 7→ ϕλ (x1 ) := x12 + x1 − λ(1 − x12 )
λ
satisfies ϕλ < 0 on −1, 1−λ , if ρ1 < 1−λ
λ
(or equivalently λ > 1
ρ−1 ), we deduce that

V λ · ν∂B < 0 on ∂B1 (0) ∩ − 1 < x1 ≤ 1

,

ρ

which implies that Etλ ⊂ C for t  1 (see Figure A.3). Hence, since ρ > n, it suffices
to choose
1 1
λ ∈ ρ−1 , n−1

to conclude the proof of the claim and find the desired contradiction. 

A simple consequence of the above result is the following estimate on the size of
convex sets.

Lemma A.14. Let Z ⊂ Rn be an open bounded convex set.

(a) Assume there exist r, K > 0 such that

Br ( x̄) ⊂ Z and |Z | ≤ K

for some x̄ ∈ Rn . Then there exists R > 0, depending only on n, r, and K, such
that
Z ⊂ BR ( x̄).

(b) Assume there exist R, κ > 0 such that

Z ⊂ BR ( x̄) and |Z | ≥ κ

for some x̄ ∈ Rn . Then there exists r > 0, depending only on n, R, and κ, such
that
Z ⊃ Br ( x̂)
for some x̂ ∈ Z.

Proof. Without loss of generality we can assume that x̄ = 0.

Step 1: Proof of (a). Fix an arbitrary point x ∈ Z, and consider the set Σx obtained
by intersecting the ball Br (0) with the hyperplane x ⊥ := {z ∈ Rn : hz, xi = 0}.
170 Appendix: Useful facts

Since Z is convex, it contains the cone Cx generated by Σx and {x} (compare with
Figure 2.4); therefore,
K ≥ |Z | ≥ | Cx | = cn |x|r n−1
for some dimensional constant cn > 0. This shows that
K
|x| ≤ =: R ∀ x ∈ Z,
cn r n−1

proving that Z ⊂ BR (0).

Step 2: Proof of (b). Let E = E A, x̂ denote John’s ellipsoid associated to Z (see
Definition A.11 and Lemma A.13), and assume up to a rotation that

(zi − x̂i )2
n Õn o
E = z = (z1, . . . , z n ) ∈ Rn : < 1 .
i=1
ai2

Since diam(E ) ≤ diam(Z) ≤ 2R, we deduce that all the semiaxes a1, . . . , an are
bounded by R. On the other hand, since S ⊂ nE , we get
n
Ö
κ ≤ |Z | ≤ |nE | = nn |B1 (0)| ai .
i=1

Hence,
κ 1 κ
ai ≥ ≥ n =: r ∀ i = 1, . . . , n,
nn |B1 (0)| j,i a j
Î
n |B1 (0)|Rn−1

which proves that Br ( x̂) ⊂ Z. 

We conclude this section with a classical fact on projections onto convex sets and
an elementary estimate on the surface measure of convex sets.

Lemma A.15. Let Z ⊂ Rn be an open convex set, and let π : Rn → Z denote the
projection onto Z. Then π is 1-Lipschitz.

Proof. First, we show that the map π is well defined.

Given a point x ∈ Rn , by definition π(x) is the closest point to x in Z, that is, π(x)
is a minimizer of
min |x − z| 2 : z ∈ Z .


The existence of a minimizer follows from compactness: if {zk }k ∈N ⊂ Z is a

minimizing sequence, then supk |zk | < ∞ (as otherwise |x − zk | 2 would tend to
infinity). Hence, one can extract a convergent subsequence to find a minimizer.
 A.3 Convex geometry 171

Figure A.4. The scalar product between x − π(x) and π(x) − z is nonnegative for all z ∈ Z.

We also notice that the minimizer is unique. Indeed, if z1, z2 ∈ Z are minimizers,
then by the convexity of the squared norm, we get
2
z1 + z2 1 
x− ≤ |x − z1 | 2 + |x − z2 | 2
2 2

with strict inequality unless z1 = z2 . Thus, since z1 +z2

2
∈ Z (as Z is convex), we
deduce that z1 = z2 .
This proves that π is well defined. To show that π is 1-Lipschitz, we consider
z ∈ Z and for t ∈ [0, 1], we define zt := (1 − t)π(x) + tz ∈ Z. Then by the minimality
of π(x), we get

|x − zt | 2 − |x − π(x)| 2
0 ≤ lim+ = hx − π(x), π(x) − zi,
t→0 2t
which proves that
hx − π(x), π(x) − zi ≥ 0 ∀z ∈ Z (A.13)
(see Figure A.4). In particular, if we pick y ∈ Rn and we apply (A.13) with z = π(y),
we obtain
hπ(x), π(x) − π(y)i ≤ hx, π(x) − π(y)i.
Also, exchanging the roles of x and y,

−hπ(y), π(x) − π(y)i ≤ −hy, π(x) − π(y)i.

Adding these two inequalities, we obtain

|π(x) − π(y)| 2 = hπ(x) − π(y), π(x) − π(y)i

≤ hx − y, π(x) − π(y)i
≤ |x − y| |π(x) − π(y)|;

hence, |π(x) − π(y)| ≤ |x − y|, as desired. 

As a consequence of this result, we get the following:
172 Appendix: Useful facts

Figure A.5. The projection π maps ∂BR ( x̄) onto ∂ Z.

Corollary A.16. Let Z ⊂ Rn be an open bounded convex set with

Z ⊂ BR ( x̄)
for some R > 0 and x̄ ∈ Rn . Then
Hn−1 (∂ Z) ≤ Rn−1 Hn−1 (∂B1 (0)).
Proof. Let π : Rn → Z denote the projection onto Z, and notice that π(∂BR ( x̄)) = ∂ Z
(see Figure A.5). Hence, since π is 1-Lipschitz (by Lemma A.15), we can apply
Lemma A.7 and Proposition A.6(i) to get
Hn−1 (∂ Z) ≤ Hn−1 (∂BR ( x̄)) = Hn−1 (∂BR (0)) = Rn−1 Hn−1 (∂B1 (0)),
as desired. 

A.4 Convex functions

Convex functions are usually defined by requiring that given two points x and y, the
value of the function along the segment connecting x to y is below the value of the
affine function that coincides with our function at x and y. In other words, a function
u is convex if
u(t x + (1 − t)y) ≤ tu(x) + (1 − t)u(y) ∀ x, y, ∀ t ∈ [0, 1].
Notice that this definition has a problem if the domain of definition of our function
is not convex: indeed, even if u is defined at x and y, u may not be defined at the
point t x + (1 − t)y for some t ∈ (0, 1). For this reason, we introduce the following
definition of a convex function inside a general open set.
Definition A.17. Given an open set Ω ⊂ Rn and a function u : Ω → R, we say that
u is convex if it can be extended to a convex function (possibility taking values in
R ∪ {+∞}) in all of Rn .
 A.4 Convex functions 173

Figure A.6. A convex function u and some of its supporting planes.

Notice that when Ω is convex, the above definition coincides with the usual
definition of convexity. Indeed, given a convex function u : Ω → R, it suffices to set
u ≡ +∞ in Rn \ Ω to obtain a convex function on Rn .
We recall that convex functions are continuous in the interior of the set where
they are finite (see, for instance, [99, Theorem 10.1]). In particular, since we al-
ways assume that u < ∞ inside Ω, u is continuous inside Ω. As we shall see in
Corollary A.23 below, u is actually locally Lipschitz inside Ω.

A.4.1 Subdifferential. Given an open set Ω ⊂ Rn and u : Ω → R a convex

function, for x ∈ Ω, we define the subdifferential of u at x as

∂u(x) := {p ∈ Rn : u(z) ≥ u(x) + hp, z − xi ∀ z ∈ Ω}.

It is immediate to check that ∂u(x) is a closed convex set. Geometrically, p ∈ ∂u(x)

if the affine function `x, p (z) := u(x) + hp, z − xi “touches u from below at x”, that is,

`x, p ≤ u in Ω and `x, p (x) = u(x),

and we shall say that `x, p is a supporting plane to u at x (see Figure A.6).

Remark A.18. A priori the definition of subdifferential given above may depend on
the domain Ω. To see that this is not the case, assume that `x, p ≤ u in a neighborhood
of x. This means that the convex function u − `x, p attains a local minimum at x;
hence, by convexity, this minimum is global. Thus, `x, p ≤ u inside Ω. This shows
that p ∈ ∂u(x) if and only if `x, p ≤ u in a neighborhood of x, proving that the notion
of subdifferential is local (and, in particular, independent of the domain Ω).
In the same respect, if x ∈ Ω and u, v : Ω → R are convex functions that coincide
in a neighborhood of x, then

p ∈ ∂u(x) ⇔ p ∈ ∂v(x).
174 Appendix: Useful facts

A convex function u is said to be strictly convex in Ω if for any x ∈ Ω and

p ∈ ∂u(x),
u(z) > u(x) + p · (z − x) ∀ z ∈ Ω \ {x},
that is, any supporting plane to u touches the graph of u at only one point (in Figure A.6
the function u is strictly convex at x1 and x4 , but not at x2 and x3 ).
Remark A.19. It follows immediately from the definition of subdifferential that if
p ∈ ∂u(x) \ ∂u(z) for some x , z ∈ Ω, then

u(z) > u(x) + hp, z − xi.

Indeed, if equality holds, then the affine function `x, p would touch u from below at
z; hence, p ∈ ∂u(z), a contradiction.

It is a standard fact that ∂u(x) is nonempty for any x ∈ Ω (see [99, Theorem
23.4]). The following result shows that ∂u(x) generalizes the concept of a gradient.
Lemma A.20. Let Ω ⊂ Rn be an open set and u : Ω → R be a convex function. If u
is differentiable at x ∈ Ω, then ∂u(x) = {∇u(x)}, so that, in particular,

u(z) ≥ u(x) + h∇u(x), z − xi ∀ z ∈ Ω.

Proof. Pick p ∈ ∂u(x), fix v ∈ Rn with |v| = 1, and consider ε > 0 small enough so
that x + εv ∈ Ω. Combining the definition of ∂u(x) (applied with z = x + εv) with
the differentiability of u at x, we get

u(x) + εhp, vi ≤ u(x + εv) = u(x) + εh∇u(x), vi + o(ε),

that is,
o(ε)
hp − ∇u(x), vi ≤ → 0 as ε → 0+ .
ε
This shows that hp − ∇u(x), vi ≤ 0, and exchanging v with −v, we deduce that
hp − ∇u(x), vi = 0 for all |v| = 1; hence, p = ∇u(x). 
There is actually a converse of this proposition (we refer to [99, Theorem 25.1]
for a proof):
Lemma A.21. Let Ω ⊂ Rn be an open set, u : Ω → R be a convex function, and
assume that ∂u(x) is a singleton. Then u is differentiable at x.
Another useful fact is that ∂u maps compact sets on compact sets (in particular,
∂u(x) is compact for any x ∈ Ω).
Lemma A.22. Let Ω ⊂ Rn be an open set and u : Ω → R be a convex function. If
K ⊂ Ω is compact, then ∂u(K) is compact.
 A.4 Convex functions 175

1
Proof. Let δ := 2 dist(K, ∂Ω) > 0, fix x ∈ K, and let z = x + δv for some |v| = 1.
Also, set
Kδ := {z ∈ Ω : dist(z, K) ≤ δ} ⊂⊂ Ω,
and notice that since u is continuous (hence, locally bounded) inside Ω, it is bounded
on Kδ . Thus, since x, z ∈ Kδ , we deduce that for any p ∈ ∂u(x),

δhp, vi ≤ u(z) − u(x) ≤ 2kuk L ∞ (K δ ) ;

therefore,
2kuk L ∞ (K δ )
|p| = sup hp, vi ≤ ∀ p ∈ ∂u(x), x ∈ K, (A.14)
|v |=1 δ
which proves that ∂u(K) is a bounded set.
To show that it is closed, consider a sequence {pk }k ∈N with pk ∈ ∂u(xk ) for some
xk ∈ K. Since K is compact and ∂u(K) is bounded, up to a subsequence, we have
that xk → x ∈ K and pk → p. Since pk ∈ ∂u(xk ), for any z ∈ Ω, we have that

u(z) ≥ u(xk ) + hpk , z − xk i,

and letting k → ∞, we obtain

u(z) ≥ u(x) + hp, z − xi,

that is, p ∈ ∂u(x) ⊂ ∂u(K) as desired. 

Combining Lemmas A.20 and A.22, we recover the classical fact that convex
functions are locally Lipschitz.
Corollary A.23. Let Ω ⊂ Rn be an open set and u : Ω → R be a convex function.
Then u is locally Lipschitz inside Ω and
2kuk L ∞ (Ω0 )
k∇uk L ∞ (K) ≤
dist(K, ∂Ω0)
for any K ⊂⊂ Ω0 ⊂ Ω.
Proof. We first show that u is locally Lipschitz. Let Br (x) ⊂ Ω and consider the set
∂u Br/2 (x) . Since this set is compact (by Lemma A.22), it is, in particular, bounded;

hence, there exists R > 0 such that ∂u Br/2 (x) ⊂ BR (0).

Now pick y, z ∈ Br/2 (x), and consider p ∈ ∂u(y) ⊂ BR (0). Then

u(z) ≥ u(y) + hp, z − yi ≥ u(y) − R|z − y|,

that is,
u(y) − u(z) ≤ R|z − y|.
176 Appendix: Useful facts

Reversing the roles of y and z, we get

|u(y) − u(z)| ≤ R|z − y|;

thus, u is R-Lipschitz on Br/2 (x). This proves that u is locally Lipschitz inside Ω,
which implies that it is differentiable a.e. Now consider K ⊂⊂ Ω0 ⊂ Ω. Then the
very same argument as in the proof of Lemma A.22 shows that
2kuk L ∞ (Ω0 )
|p| ≤ ∀ p ∈ ∂u(x), x ∈ K
dist(K, ∂Ω0)
(compare with (A.14)). In particular, since p = ∇u(x) at any differentiability point
of u (see Lemma A.20), we get
2kuk L ∞ (Ω0 )
|∇u(x)| ≤ for a.e. x ∈ K.
dist(K, ∂Ω0)

The argument in the proof of Lemma A.22 also shows that differentiable convex
functions are actually C 1 .
Lemma A.24. Let Ω ⊂ Rn be an open set, u : Ω → R be a convex function, and
assume that u is differentiable at every point of Ω. Then u ∈ C 1 (Ω).
Proof. As shown in the proof of Lemma A.22,

pk ∈ ∂u(xk ), pk → p, xk → x ⇒ p ∈ ∂u(x). (A.15)

Since u is differentiable inside Ω, it follows by Lemma A.20 that ∂u(z) = {∇u(z)}

for any z ∈ Ω; hence, (A.15) implies that the only possible cluster point of the
family {∇u(xk )}k ∈N is ∇u(x). This proves that ∇u(xk ) → ∇u(x). Thus, ∇u is
continuous. 
Another useful fact is contained in the following elementary lemma, stating that
if a convex function is affine on an infinite line, then its subdifferential is contained
in a hyperplane.
Lemma A.25. Let u : Rn → R be a convex function, and assume that u is affine on
ˆ Then ∂u(Rn ) is contained inside a hyperplane orthogonal to `.
a line `. ˆ

Proof. Assume that

u(z) = h p̄, zi + a on the line `ˆ := {tσ + x̄ : t ∈ R},

where σ ∈ Sn−1 .
 A.4 Convex functions 177

Then given x ∈ Rn and p ∈ ∂u(x), we have

th p̄, σi + h p̄, x̄i + a = u(tσ + x̄)

≥ u(x) + hp, tσ + x̄ − xi
= u(x) + hp, x̄ − xi + thp, σi ∀ t ∈ R,

or equivalently,

h p̄, x̄i + a − u(x) − hp, x̄ − xi ≥ thp − p̄, σi ∀ t ∈ R.

Since the left-hand side is bounded, letting t → ±∞, we deduce that hp − p̄, σi = 0.
Hence, we have shown that for any x ∈ Rn and p ∈ ∂u(x), the vector p belongs
to the hyperplane
Π := {q ∈ Rn : hq − p̄, σi = 0},
that is, ∂u(Rn ) ⊂ Π. 

The following result relates the theory of exposed points from Appendix A.3.1 to
the subdifferential of convex functions (see, for instance, the proof of [99, Theorem
25.6]).
Proposition A.26. Let Ω ⊂ Rn be an open set, u : Ω → R be a convex function,
let x ∈ Ω, and consider p an exposed point of ∂u(x). Then there exists a sequence
xk → x such that u is differentiable at xk and ∇u(xk ) → p.
Combining Theorem A.9 and Proposition A.26, we obtain the following:
Corollary A.27. Let Ω ⊂ Rn be an open set, u : Ω → R a convex function, and let
Du ⊂ Ω denote the set of differentiability points of u. Then ∂u(Ω) is contained in the
closure of the convex hull of ∇u(Du ).

A.4.2 Legendre transform. Given an open set Ω ⊂ Rn and u : Ω → R a convex

function, one defines the Legendre transform u∗ : Rn → R ∪ {+∞} of u as

u∗ (p) := sup hp, xi − u(x). (A.16)

x ∈Ω

Being the supremum of a family of linear functions, u∗ is convex.

We now show a simple implication between the subdifferential of u and that of u∗ :

p ∈ ∂u(x) ⇒ u(z) ≥ u(x) + hp, z − xi ∀ z ∈ Ω

⇒ hp, xi − u(x) ≥ hp, zi − u(z) ∀ z ∈ Ω
⇒ u∗ (p) = hp, xi − u(x).
178 Appendix: Useful facts

Since u∗ (q) ≥ hq, xi − u(x) (by the definition of u∗ ), we have shown that
p ∈ ∂u(x) ⇒ u∗ (q) ≥ u∗ (p) + hx, q − pi ∀ q ∈ Rn,
which proves the following useful implication:
x ∈ Ω, p ∈ ∂u(x) ⇒ x ∈ ∂u∗ (p). (A.17)
In order to reverse the above implication, we first need to extend u to a function
on all of Rn in a “suitable” way, and for this, we have to restrict to the case when Ω
is convex.
Let us define u in Rn \ Ω as
lim inf Ω3z→x u(z) if x ∈ ∂Ω,

u(x) := (A.18)
+∞ if x ∈ Rn \ Ω.
Since Ω is convex, it follows that u : Rn → R ∪ {+∞} is a lower semicontinuous
convex function and
u∗ (p) = sup hp, xi − u(x) = sup hp, xi − u(x). (A.19)
x ∈R n x ∈Ω

With this definition, it is a classical fact that the subdifferentials of u and u∗ are
inverse to each other, that is,
p ∈ ∂u(x) ⇔ x ∈ ∂u∗ (p) (A.20)
(see [99, Corollary 23.5.1]). Furthermore, the Legendre transform becomes an invo-
lution, that is, (u∗ )∗ = u (see [99, Theorem 12.2 and Corollary 12.2.1]). In addition,
the points x ∈ Ω where the supremum is attained are related to the subdifferential of
u in the following way:
given p ∈ Rn , the supremum in (A.19) is attained at x ∈ Ω ⇔ x ∈ ∂u∗ (p)
(A.21)
(see [99, Theorem 23.5]).
Remark A.28. If both u and u∗ are of class C 2 , one can rewrite (A.20) as
∇u ∇u∗ (p) = p,

and by differentiating this relation, one obtains

D2 u ∇u∗ (p) D2 u∗ (p) = Id .

In particular, taking the determinant on both sides, if det D2 u = f , then

1
det D2 u∗ = .
f ◦ ∇u∗
 A.4 Convex functions 179

It follows that the Legendre transform allows one to convert upper (resp. lower)
bounds on the determinant of the Hessian into lower (resp. upper) bounds, a fact that
is used, for instance, in the proof of Theorem 2.19. Note that although the above
argument requires the smoothness of both u and u∗ , an analogous statement holds
also at the level of Monge–Ampère measures (see Lemma 2.4).
The previous results allow us to show the following useful fact.
Lemma A.29. Let Ω ⊂ Rn be an open bounded convex set, u : Ω → R be a convex
function, and u∗ : Rn → R its Legendre transform. Then ∂u∗ (Rn ) ⊂ Ω.
Proof. Since Ω is convex, we can extend u as in (A.18) so that (A.19) holds, and
because Ω is bounded and u is lower semicontinuous, we deduce that
u∗ (p) = max hp, xi − u(x), (A.22)
x ∈Ω

that is, the supremum in (A.19) is attained at some point x ∈ Ω.

Let Du∗ ⊂ Rn denote the set of differentiability points of u∗ , and given p ∈ Du∗ ,
let x ∈ Ω be a point where the maximum in (A.22) is attained. Then it follows by
Lemma A.20 and (A.21) that ∇u∗ (p) = x ∈ Ω, which proves that ∇u∗ (Du∗ ) ⊂ Ω.
Since Ω is convex, Corollary A.27 implies that ∂u∗ (Rn ) ⊂ Ω, as desired. 

A.4.3 On the subdifferential of disjoint sets. If E and F are disjoint sets, it is not
true in general that ∂u(E) and ∂u(F) are disjoint (see, for instance, Figure A.6 where
x2 , x3 but ∂u(x2 ) = ∂u(x3 )). However, the following lemma shows that the set of
slopes for which this can happen has measure zero.
Lemma A.30. Given a convex function u : Ω → R, define
Z := {p ∈ Rn : p ∈ ∂u(x) ∩ ∂u(y) for some x , y ∈ Ω}.
Then | Z | = 0.
Proof. We first prove the result when Ω is bounded, and then we treat the general
case.
Step 1: Ω is bounded. Let u∗ be the Legendre transform of u (see (A.16)). Since Ω
is bounded (say, Ω ⊂ BR (0) for some R > 0), we see that for each x ∈ Ω, the map
p 7→ hp, xi − u(x)
is R-Lipschitz; hence, so is u∗ (being the supremum of a family of R-Lipschitz
functions). In particular, by Rademacher’s theorem, u∗ is differentiable a.e. Let
p ∈ Z . Then p ∈ ∂u(x) ∩ ∂u(y) for some x , y ∈ Ω, and it follows from (A.17) that
x, y ∈ ∂u∗ (p).
180 Appendix: Useful facts

So, Lemma A.20 implies that u∗ is not differentiable at p. Hence,

Z ⊂ {p ∈ Rn : u∗ is not differentiable at p},

and since the latter set has measure zero, we deduce that |S| = 0, as desired.
Step 2: The general case. For any k ∈ N, we consider the set Ωk := Ω ∩ Bk (0) and
define

Zk := p ∈ Rn : p ∈ ∂ u|Ωk (x) ∩ ∂ u|Ωk (y) for some x , y ∈ Ωk .



Then,

p∈Z ⇒ p ∈ ∂u(x) ∩ ∂u(y) for some x , y ∈ Ω

⇒ p ∈ ∂u(x) ∩ ∂u(y) for some x , y ∈ Ωk , for k  1
⇒ p ∈ Zk for k  1,

which proves that Z ⊂ ∪k ∈N Zk . Since each set Zk has measure zero by Step 1, Z
also has measure zero. 

A.4.4 Area formula for the gradient of convex functions. In this section, we
discuss a change of variable formula for C 1,1 convex functions.
Theorem A.31. Let Ω ⊂ Rn be an open bounded set, and let u : Ω → R be a convex
1,1
function of class Cloc . Then,
∫
|∇u(E)| = det D2 u dx ∀ E ⊂ Ω Borel.
E
 A.4 Convex functions 181

Proof. Applying the area formula for Lipschitz functions (see, for instance, [42,
Section 3.3.2, Theorem 1]) to the map ∇u : Ω → Rn, it follows that
∫ ∫
2
det D u dx = #{x ∈ E : ∇u(x) = p} dp,
E ∇u(E)

where #A denotes the cardinality of a set A.

Thanks to Lemma A.30, we see that

#{x ∈ E : ∇u(x) = p} = 1 for a.e. p ∈ ∇u(E);

therefore, ∫
#{x ∈ E : ∇u(x) = p} dp = |∇u(E)|,
∇u(E)

as desired. 

A.4.5 A criterion for C 1,α regularity. In this section, we show that if a con-
vex function separates its supporting planes in a C 1,α fashion, then it is indeed of
class C 1,α .

Lemma A.32. Let Z be an open convex set satisfying

Br ( x̄) ⊂ Z ⊂ BR ( x̄) (A.23)

for some 0 < r ≤ R and x̄ ∈ Rn , let u : Z → R be a convex function, and assume

that there exist constants K, C, %̄ > 0 and α ∈ (0, 1] such that the following holds: u
is K-Lipschitz in Z, and for every x ∈ Z there exists px ∈ ∂u(x) satisfying

u(z) − u(x) − hpx, z − xi ≤ C|z − x| 1+α ∀ z ∈ Z ∩ B(x, %̄). (A.24)

Then u ∈ C 1,α (Z) with

 
k∇ukC 0, α (Z) ≤ C̄ = C̄ r, R, K, C, %̄ .

Proof. We first show that ∂u is a singleton at every point. For this, given x ∈ Z and
p ∈ ∂u(x), we apply the definition of subdifferential and (A.24) with z = x + εv,
where |v| = 1 and ε < %, to get

u(x) + εhp, vi ≤ u(x + εv) ≤ u(x) + εhpx, vi + Cε 1+α,

that is,
hp − px, vi ≤ Cε α ∀ |v| = 1.
182 Appendix: Useful facts

Figure A.7. Let x ∈ Z, and assume, with no loss of generality, that B3r ( x̄) ⊂ Z (otherwise, we simply
replace r by r/3). If x ∈ B2r ( x̄), then (A.27) holds with % ≤ r and z := x + %e/|e| ∈ B3r ( x̄) ⊂ Z.
If x < B2r (x), then we consider the cone C x generated by x and Br ( x̄), and given any vector e ∈ Rn
and % ≤ r, we can choose z ∈ ∂B% (x) ∩ C x such that (A.27) holds with some η > 0 small (the
smallness depending only on the opening of C x , which in turns depends only on r and R).

Letting ε → 0 and since v is arbitrary, we obtain that p = px , which proves that ∂u(x)
is a singleton. In particular, it follows by Lemmas A.21 and A.24 that u ∈ C 1 (Z), so
(A.24) can be rewritten as

u(z) − u(x) − h∇u(x), z − xi ≤ C|z − x| 1+α ∀ x ∈ Z, z ∈ Z ∩ B(x, %̄). (A.25)

Hence, we have to prove that there exists a constant M = M (r, R, K, C, %̄) > 0 such
that
|∇u(x) − ∇u(x 0)| ≤ M |x − x 0 | α ∀ x, x 0 ∈ Z. (A.26)
Let us observe that since Z is convex and satisfies (A.23), ∂ Z is uniformly Lipschitz:
more precisely, there exist %0 ∈ (0, %̄/2) and η > 0, depending only on r and R,
such that, for any x ∈ Z, for any vector e ∈ Rn , and for any % ≤ %0 , there is a point
z ∈ B(x, %0 ) ∩ Z satisfying

|he, z − xi| ≥ η|z − x| |e| and |z − x| = % (A.27)

(see Figure A.7).

We now prove (A.26). Up to subtracting a linear function, without loss of
generality, we can assume that ∇u(x) = 0. Let us observe that if |x − x 0 | ≥ %0 , then
we can use that k∇uk L ∞ (Z) ≤ K to deduce that

2K
|∇u(x 0)| ≤ |x − x 0 | α,
%α0

which proves (A.26) with M := 2K/%α0 . Hence, we can assume that |x − x 0 | < %0 .
 A.4 Convex functions 183

Fix x 0 ∈ B(x, %0 ) ∩ Z, and apply (A.27) to find a point z ∈ B(y, %0 ) ∩ Z such that

|h∇u(x 0), z − x 0i| ≥ η|z − x 0 | |∇u(x 0)| and |z − x 0 | = |x 0 − x|.

Then, since ∇u(x) = 0, it follows by (A.25) applied both at x 0 and x that

η|z − x 0 | |∇u(x 0)| ≤ |h∇u(x 0), z − x 0i|

≤ |u(z) − u(x 0) − h∇u(x 0), z − x 0i|
+ |u(z) − u(x)| + |u(x 0) − u(x)|
 
≤ C |z − x 0 | 1+α + |z − x| 1+α + |x 0 − x| 1+α

1+α 
≤ C |z − x 0 | 1+α + |z − x 0 | + |x 0 − x| + |x 0 − x| 1+α ,

and recalling that |z − x 0 | = |x 0 − x|, this proves (A.26) with M := (2 + 21+α )C. 

A.4.6 On the regularity of the convex envelope. Given an open set Ω ⊂ Rn and a
function v : Ω → R, we define its convex envelope in Ω as

Γv (x) := sup{`(x) : ` ≤ v in Ω, ` affine} ∀ x ∈ Ω. (A.28)

A possible characterization of the convex envelope is the following:

Lemma A.33. Given Ω ⊂ Rn open and v : Ω → R, the convex envelope of v in Ω is
the largest convex function below v in Ω.
Proof. Since Γv is the supremum of affine functions (see (A.28)), it is convex. Hence,
we need to show that any convex function below v in Ω is also below Γv .
Let w : Ω → R be a convex function with w ≤ v in Ω, and fix a point x̄ ∈ Ω.
Consider a vector p ∈ ∂w( x̄), and define the affine function

`(x) := w( x̄) + hp, x − x̄i.

Since ` ≤ w and w ≤ v, we deduce that ` ≤ v inside Ω, so ` ≤ Γv by the definition

of Γv . Recalling that w( x̄) = `( x̄), this implies that w( x̄) ≤ Γv ( x̄), and since x̄ is
arbitrary, we obtain that w ≤ Γv , as desired. 
We now want to understand the regularity of Γv in terms of the regularity of v
and Ω. This problem has been recently addressed in full generality in [34]. For our
purposes, we shall only need the following classical result, which can be seen as a
corollary of [34, Theorem 1.2] (see also [114, Lemma 2.3] for a direct proof).
Proposition A.34. Let Ω ⊂ Rn be an open bounded convex set of class C 1,1 , and let
1,1
v ∈ C 1,1 (Ω). Then Γv ∈ Cloc (Ω).
184 Appendix: Useful facts

Figure A.8. If x1 ∈ {v = Γv } is a point where Γv and v are twice differentiable, then D2 v(x1 ) ≥
D2 Γv (x1 ) ≥ 0 in the sense of symmetric matrices. If x2 ∈ {v > Γv } is a point where Γv is twice
differentiable, then D2 Γv (x2 ) must have a zero eigenvalue (in particular, the picture above is not
correct, as Γv should always be affine where it does not touch v). Indeed, if D2 Γv (x2 ) was positive
definite, then we could lift the linear function z 7→ Γ(x2 ) + h∇Γv (x2 ), z − x2 i by a small amount
δ > 0 to construct a convex function strictly larger than Γv but still smaller than v, contradicting the
fact that Γv is largest convex function below v in Ω.

1,1
Since Cloc functions are twice differentiable at almost every point, one can con-
sider the determinant of D2 Γv , which is defined almost everywhere.
The following result gives an estimate on the image of ∇Γv in terms of det D2 v
and the contact set between v and Γv .
Proposition A.35. Let Ω ⊂ Rn be an open bounded convex set of class C 1,1 , and let
v ∈ C 1,1 (Ω). Then,
∫ ∫
2
|∇Γv (E)| = det D Γv dx ≤ det D2 v dx ∀ E ⊂ Ω Borel.
E∩{v=Γv } E∩{v=Γv }
1,1
Proof. Thanks to Proposition A.34, we know that Γv ∈ Cloc (Ω), so we can apply
Theorem A.31 to get
∫
|∇Γv (E)| = det D2 Γv dx ∀ E ⊂ Ω Borel. (A.29)
E
1,1
We now notice that since both v and Γv are of class Cloc , they are twice differentiable
a.e. Let x be a point where both v and Γv are twice differentiable. If x ∈ {v = Γv },
then since v ≥ Γv , we see that
D2 v(x) ≥ D2 Γv (x) ≥ 0.
So, in particular,
det D2 v ≥ det D2 Γv a.e. in {v = Γv }. (A.30)
On the other hand, if x < {v = Γv } is a point where Γv is twice differentiable, then
D2 Γv (x) must have a zero eigenvalue. Indeed, if not, then the convex function
Γ̂v (z) := max{Γv (z), Γv (x) + h∇Γv (x), z − xi + δ}, δ > 0,
 A.5 Tools from measure theory, nonlinear analysis, and PDEs 185

would be smaller than v for δ sufficiently small, contradicting Lemma A.33 (see Fig-
ure A.8). Hence, since D2 Γv (x) has a zero eigenvalue, we deduce that det D2 Γv (x) =
0, which proves that

det D2 Γv = 0 a.e. in Ω \ {v = Γv }. (A.31)

Combining (A.29) with (A.30) and (A.31), the result follows. 

A.5 Tools from measure theory, nonlinear analysis, and PDEs

A.5.1 The layer-cake formula. The layer-cake formula is a very elementary but
useful tool to express the L p norm of a function in terms of the measure of its
superlevel sets.
Lemma A.36. Let E ⊂ Rn be a Borel set and g : E → R be a measurable function.
Then for any p ≥ 1,
∫ ∫ ∞
|g| dx = p
p
t p−1 E ∩ {|g| > t} dt.
E 0

Proof. Set h := |g| p ,

and notice that the following elementary identity holds:
∫ h(x) ∫ ∞
h(x) = dτ = χ{h>τ } (x) dτ ∀ x ∈ E.
0 0

Hence, integrating the above equality with respect to x and applying Fubini’s theorem,
we get ∫ ∫ ∫ ∞
|g| p dx = h dx = E ∩ {h > τ} dτ.
E E 0
Performing the change of variable τ = t p and noticing that {h > τ} = {|g| > t}, the
result follows. 

A.5.2 The implicit function theorem in Banach spaces. The implicit function
theorem in Banach spaces is a useful tool in nonlinear PDEs when one wants to
prove existence of solutions near a given one by studying the linearized operator (see
Section 3.1.3). This is the general statement (see, for instance, [37, Theorem 15.1
and Corollary 15.1] for a proof):
Theorem A.37. Let X, Σ, Y be Banach spaces, let G : X × Σ → Y , and assume that
there exists (û, σ̂) ∈ X × Σ such that
(a) G (û, σ̂) = 0;
186 Appendix: Useful facts

(b) G is of class C 1 in a neighborhood of (û, σ̂);

(c) the Fréchet differential of G at (û, σ̂) with respect to u, that is, the map
Dx G (û, σ̂) : X → Y given by

d
h 7→ Dx G (û, σ̂)[h] = G (û + εh, σ̂),
dε ε=0

is a linear isomorphism between X and Y .

Then there exists N , a neighborhood of σ̂ in Y , such that the following holds: For
all σ ∈ N , there is uσ ∈ X such that

N 3 σ 7→ uσ ∈ X is of class C 1 and G (uσ, σ) = 0.

A.5.3 The maximum principle. A useful tool in elliptic PDEs is the maximum
principle. It states that subsolutions to a uniformly elliptic equation cannot attain
their maximum in the interior. Here, we shall state and prove a variant of this
statement, concerning strict subsolutions to (possibly degenerate) elliptic equations.

Lemma A.38. Let Ω ⊂ Rn be an open bounded domain, let ai j : Ω → R be

measurable functions for all i, j = 1, . . . , n, and assume that ai j are elliptic, that is,

0 ≤ ai j (x)ξ i ξ j ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω. (A.32)

Let u ∈ C 2 (Ω) ∩ C 0 (Ω) satisfy

ai j ∂i j u > 0 in Ω. (A.33)

Then,
max u = max u.
Ω ∂Ω

Proof. If u attains its maximum at some point x0 ∈ Ω, then D2 u(x0 ) ≤ 0, and it

follows by (A.32) that
0 ≥ ai j (x)∂i j u(x0 ),
which contradicts (A.33). This proves that u cannot attain any interior maximum;
hence, its maximum over Ω must be attained at the boundary. 

A.5.4 Schauder’s theory. Schauder’s theory concerns the regularity of linear uni-
formly elliptic equations with Hölder coefficients. We state the results that we shall
need in our situation, referring to [58, Chapter 6] for a proof.
 A.5 Tools from measure theory, nonlinear analysis, and PDEs 187

Theorem A.39. Let Ω ⊂ Rn be an open bounded set satisfying

Br ( x̄) ⊂ Ω ⊂ BR ( x̄)

for some 0 < r ≤ R and x̄ ∈ Rn , let k ≥ 0, α ∈ (0, 1), ai j ∈ C k,α (Ω) for all
i, j = 1, . . . , n, g ∈ C k,α (Ω), and assume that ∂Ω is of class C k+2,α and that

λ|ξ | 2 ≤ ai j (x)ξ i ξ j ≤ Λ|ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω,

for some 0 < λ ≤ Λ. Then there exists a unique solution u ∈ C k+2,α (Ω) to
(
ai j ∂i j u = g in Ω,
u=0 on ∂Ω.

In addition, u satisfies
kukC k+2, α (Ω) ≤ C,
where C depends only on n, λ, Λ, kai j kC k, α (Ω) , kgkC k, α (Ω) , r, R, and k∂ΩkC k+2, α .
Theorem A.40. Let Ω ⊂ Rn be an open bounded set satisfying

Br ( x̄) ⊂ Ω ⊂ BR ( x̄)

for some 0 < r ≤ R and x̄ ∈ Rn , let k ≥ 0, α ∈ (0, 1), ai j ∈ C k,α (Ω) for all
i, j = 1, . . . , n, and assume that

λ|ξ | 2 ≤ ai j (x)ξ i ξ j ≤ Λ|ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ) ∈ Rn, x ∈ Ω,

for some 0 < λ ≤ Λ and that u ∈ C 2 (Ω) solves

ai j ∂i j u = 0 in Ω.
k+2,α
Then u ∈ Cloc (Ω), and for any Ω0 ⊂⊂ Ω, we have that

kukC k+2, α (Ω0 ) ≤ Ckuk L ∞ (Ω),

where C depends only on n, λ, Λ, kai j kC k, α (Ω) , dist(Ω0, ∂Ω), r, and R.

A.5.5 Evans–Krylov theorem. The Evans–Krylov theorem is a classical result in

fully nonlinear PDEs stating that C 1,1 solutions to concave uniformly elliptic fully
nonlinear equations are locally C 2,α [41, 75]. This result was later generalized to
prove regularity up to the boundary in [76, 23]. Note that if the operator is smooth,
one can differentiate the equation and apply Schauder’s theory to show that solutions
188 Appendix: Useful facts

are as smooth as linear elliptic theory allows them to be (see, for instance, [20,
Proposition 9.1]).
We begin by stating global regularity in the case of C ∞ operators, as this is enough
for our purposes, referring to [58, Chapter 17.8] or [20, Chapter 9]2 for a proof.
Theorem A.41. Let Ω ⊂ Rn be an open bounded set satisfying

Br ( x̄) ⊂ Ω ⊂ BR ( x̄) (A.34)

for some 0 < r ≤ R and x̄ ∈ Rn , let k ≥ 0 and α ∈ (0, 1), and assume that ∂Ω is
of class C k+2,α . Also, let F : Symn → R be a C ∞ function and g ∈ C k,α (Ω), and
assume that u ∈ C 1,1 (Ω) solves
(
F(D2 u) = g a.e. in Ω,
u=0 on ∂Ω.

Suppose that
(a) F is uniformly elliptic with respect to u, that is,

λ|ξ | 2 ≤ Fi j D2 u(x) ξ i ξ j ≤ Λ|ξ | 2 ∀ ξ = (ξ 1, . . . , ξ n ), for a.e. x ∈ Ω,

for some 0 < λ ≤ Λ;

(b) F is concave with respect to u, that is,

d2
F D2 u(x) + ε A ≤ 0 ∀ A ∈ Symn, for a.e. x ∈ Ω.


dε 2 ε=0

Then u ∈ C k+2,α (Ω) with

kukC k+2, α (Ω) ≤ C
for some constant C depending only on n, λ, Λ, kukC 1,1 (Ω) , kF kC k+2 (Symn ) , kgkC k, α (Ω),
r, R, and k∂ΩkC k+2, α .
We now state the interior version of the above result (see [58, Chapter 17.4] or
[20, Chapters 6 and 8] for a proof):
Theorem A.42. Let Ω ⊂ Rn be an open set, k ≥ 0, α ∈ (0, 1), F : Symn → R a C ∞
k,α
function, g ∈ Cloc (Ω), and u ∈ C 1,1 (Ω) a solution of

F(D2 u) = g a.e. in Ω.
2In [20, Chapter 9], the authors prove the result only in the case of zero right-hand side. However, once the
result is proved in this case, the extension to general right-hand sides can be obtained using the same techniques
as those used in [20, Chapter 8] for the interior case (see also [109, Section 4], where this strategy is used to
obtain more general boundary regularity results).
 A.5 Tools from measure theory, nonlinear analysis, and PDEs 189

Assume that F satisfies both assumptions (a) and (b) of Theorem A.41 inside some
open set Ω0 ⊂ Ω and that Ω0 satisfies (A.34) for some 0 < r ≤ R and x̄ ∈ Rn . Then,
0
u ∈ C k+2,α (Ω0 ) for any Ω00 ⊂⊂ Ω0 with

kukC k+2, α (Ω00 ) ≤ C

for some constant C depending only on n, λ, Λ, kukC 1,1 (Ω0 ) , kF kC k+2 (Symn ) , kgkC k, α (Ω0 ) ,
dist(Ω00, ∂Ω0), r, and R.
In order to be able to apply the above result to the Monge–Ampère equation, we
need to check that assumptions (a) and (b) of Theorem A.41 are satisfied.
Proposition A.43. Let Ω ⊂ Rn be an open set, and let u : Ω → R be a convex
function of class C 1,1 satisfying

det D2 u = f a.e. in Ω,

for some f ≥ c0 > 0. Then for any Ω0 ⊂ Ω, assumptions (a) and (b) of Theorem A.41
hold inside Ω0 with constants λ and Λ depending only on kukC 1,1 (Ω0 ) , and c0 .
Proof. Rewrite the equation as

log det D2 u = log f a.e. in Ω,

and notice that as a consequence of Remark 1.1 and the uniform bound on kukC 1,1 (Ω0 ) ,

1
Id ≤ D2 u(x) ≤ K Id for a.e. x ∈ Ω,
K
for some constant K > 1 depending only on kukC 1,1 (Ω0 ) , and c0 .
Consider a C ∞ function F : Symn → R that coincides with A 7→ log det A in a


neighborhood of
n 1 o
GK := A ∈ Symn : Id ≤ A ≤ K Id .
K
We now prove that F satisfies assumptions (a) and (b) of Theorem A.41.
For (a), we notice that since D2 u(x) ∈ GK for a.e. x ∈ Ω0,
d
Fi j (D2 u)ξ i ξ j = log det(D2 u + εξ ⊗ ξ = tr (D2 u)−1 ξ ⊗ ξ = ui j ξ i ξ j

 
dε ε=0

(see (3.11)); hence,

1 2
|ξ | ≤ Fi j (D2 u)ξ i ξ j ≤ K |ξ | 2 a.e. in Ω0 .
K
190 Appendix: Useful facts

For (b), arguing as in (3.12), we get

d2 d
F D2 u(x) + ε A = tr (D2 u + ε A)−1 A

 
dε 2 ε=0 dε ε=0
= − tr (D2 u)−1 A(D2 u)−1 A ≤ 0,
 

where the last inequality follows as in (3.13). 

Bibliography

[1] Alexandrov, A.D. An application of the theorem on the invariance of domains to existence
proofs. Izv. AN USSR 3 (1939), 243–256. Zbl 65.0828.02.
[2] Alexandrov, A.D. – Existence and uniqueness of a convex surface with a given integral
curvature. C. R. (Doklady) Acad. Sci. URSS (N. S.) 35 (1942), 131–134. Zbl 0061.37604.
[3] Alexandrov, A.D. – Smoothness of the convex surface of bounded Gaussian curvature. C.
R. (Doklady) Acad. Sci. URSS (N. S.) 36 (1942), 195–199. Zbl 0061.37605.
[4] Alexandrov, A.D. – Convex polyhedra. Translated from the 1950 Russian edition by N. S.
Dairbekov, S. S. Kutateladze, and A. B. Sossinsky. With comments and bibliography by
V. A. Zalgaller and appendices by L. A. Shor and Yu. A. Volkov. Springer Monographs in
Mathematics. Springer, Berlin, 2005. Zbl 1067.52011.
[5] Alexandrov, A.D. – Dirichlet’s problem for the equation det kzi j k = ϕ(z1, . . . , z n, z,
x1, . . . , xn ) I. (Russian). Vestnik Leningrad. Univ. Ser. Mat. Meh. Astr. 13 (1958), no. 1,
5–24. Zbl 0114.30202.
[6] Ambrosio, L., Colombo, M., De Philippis, G., and Figalli, A. – Existence of Eulerian
solutions to the semigeostrophic equations in physical space: The 2-dimensional periodic
case. Comm. Partial Differential Equations, 37 (2012), no. 12, 2209–2227. Zbl 1258.35164.
[7] Ambrosio, L., Colombo, M., De Philippis, G., and Figalli, A. – A global existence result for
the semigeostrophic equations in three dimensional convex domains. Discrete Contin. Dyn.
Syst., 34 (2014), no. 4, 1251–1268. Zbl 1287.35064.
[8] Ampère, A.M. – Mémoire contenant l’application de la théorie. Journal de l’École Poly-
technique (1820).
[9] Bakelman, I.Ya. – A generalization of the solution of Monge–Ampère equations. (Russian).
Dokl. Akad. Nauk SSSR (N. S.) 114 (1957), 1143–1145. Zbl 0114.29602.
[10] Benamou, J.-D. and Brenier, Y. – Weak existence for the semigeostrophic equation for-
mulated as a coupled Monge–Ampère/transport problem. SIAM J. Appl. Math., 58 (1998),
1450–1461. Zbl 0915.35024.
[11] Boas, H.P. and Boas, R.P. – Short proofs of three theorems on harmonic functions. Proc.
Amer. Math. Soc. 102 (1988), no. 4, 906–908. Zbl 0665.31004.
[12] Brenier, Y. – Polar factorization and monotone rearrangement of vector-valued functions.
Comm. Pure Appl. Math. 44 (1991), no. 4, 375–417. Zbl 0738.46011.
[13] Caffarelli, L.A. – A localization property of viscosity solutions to the Monge–Ampère equa-
tion and their strict convexity. Ann. of Math. (2) 131 (1990), no. 1, 129–134. Zbl 0704.35045.
[14] Caffarelli, L.A. – Interior W 2, p estimates for solutions of the Monge–Ampère equation. Ann.
of Math. (2), 131 (1990), no. 1, 135–150. Zbl 0704.35044.
[15] Caffarelli, L.A. – Some regularity properties of solutions of Monge–Ampère equation.
Comm. Pure Appl. Math. 44 (1991), no. 8-9, 965–969. Zbl 0761.35028.
192 Bibliography

[16] Caffarelli, L.A. – The regularity of mappings with a convex potential. J. Amer. Math. Soc.
5 (1992), no. 1, 99–104. Zbl 0753.35031.
[17] Caffarelli, L.A. – Boundary regularity of maps with convex potentials. Comm. Pure Appl.
Math. 45 (1992), no. 9, 1141–1151. Zbl 0778.35015.
[18] Caffarelli, L.A. – A note on the degeneracy of convex solutions to Monge–Ampère equation.
Comm. Partial Differential Equations 18 (1993), no. 7-8, 1213–1217. Zbl 0785.35028.
[19] Caffarelli, L.A. – Boundary regularity of maps with convex potentials. II. Ann. of Math. (2)
144 (1996), no. 3, 453–496. Zbl 0916.35016.
[20] Caffarelli, L.A. and Cabré, X. – Fully nonlinear elliptic equations. American Mathematical
Society Colloquium Publications, 43. American Mathematical Society, Providence, RI,
1995. vi+104 pp. Zbl 0834.35002.
[21] Caffarelli, L.A. and Gutiérrez, C.E. – Real analysis related to the Monge–Ampère equation.
Trans. Amer. Math. Soc. 348 (1996), no. 3, 1075–1092. Zbl 0858.35034.
[22] Caffarelli, L.A. and Gutiérrez, C.E. – Properties of the solutions of the linearized Monge–
Ampère equation. Amer. J. Math. 119 (1997), no. 2, 423–465. Zbl 0878.35039.
[23] Caffarelli, L.A., Nirenberg, L., and Spruck, J. – Dirichlet problem for nonlinear second
order elliptic equations I, Monge–Ampère equations. Comm. Pure Appl. Math. 37 (1984),
369–402. Zbl 0598.35047.
[24] Calabi, E. – Improper affine hyperspheres of convex type and a generalization of a theorem
by K. Jörgens. Michigan Math. J. 5 (1958), 105–126. Zbl 0113.30104.
[25] Calabi, E. – Complete affine hyperspheres. I. Symposia Mathematica, Vol. X (Convegno
di Geometria Differenziale, INDAM, Rome, 1971), pp. 19–38. Academic Press, London,
1972. Zbl 0252.53008.
[26] Chen, S. and Figalli, A. – Boundary ε-regularity in optimal transportation. Adv. Math., 273
(2015), 540–567. Zbl 1307.49045.
[27] Cheng, S.Y. and Yau, S.-T. – Complete affine hypersurfaces. I. The completeness of affine
metrics. Comm. Pure Appl. Math. 39 (1986), no. 6, 839–866. Zbl 0623.53002.
[28] Cuesta-Albertos, J.A. and Matrán, C. – Notes on the Wasserstein metric in Hilbert spaces.
Ann. Prob. 17 (1989), no. 3, 1264–1276. Zbl 0688.60011.
[29] Cullen, M. – A mathematical theory of large-scale atmosphere/ocean flow. Imperial College
Press (2006).
[30] De Philippis, G. and Figalli, A. – W 2,1 regularity for solutions of the Monge–Ampère
equation. Invent. Math., 192 (2013), no. 1, 55–69. Zbl 1286.35107.
[31] De Philippis, G. and Figalli, A. – Sobolev regularity for Monge–Ampère type equations.
SIAM J. Math. Anal., 45 (2013), no. 3, 1812–1824. Zbl 1271.49031.
[32] De Philippis, G. and Figalli, A. – Second order stability for the Monge–Ampère equation
and strong Sobolev convergence of optimal transport maps. Anal. PDE, 6 (2013), no. 4,
993–1000. Zbl 1278.35090.
 Bibliography 193

[33] De Philippis, G. and Figalli, A. – The Monge–Ampère equation and its link to optimal
transportation. Bull. Amer. Math. Soc. (N.S.), 51 (2014), no. 4, 527–580. Zbl 06377770.
[34] De Philippis, G. and Figalli, A. – Optimal regularity of the convex envelope. Trans. Amer.
Math. Soc., 367 (2015), no. 6, 4407–4422. Zbl 1316.49019.
[35] De Philippis, G. and Figalli, A. – Partial regularity for optimal transport maps. Publ. Math.
Inst. Hautes Études Sci., 121 (2015), 81–112. Zbl 1325.49051.
[36] De Philippis, G., Figalli, A., and Savin, O. – A note on interior W 2,1+ estimates for the
Monge–Ampère equation. Math. Ann., 357 (2013), no. 1, 11–22. Zbl 1280.35153.
[37] Deimling, K. – Nonlinear functional analysis. Springer, Berlin, 1985. xiv+450 pp.
Zbl 0559.47040.
[38] Delanoë, P. and Ge, Y. – Regularity of optimal transport on compact, locally nearly spherical,
manifolds. J. reine angew. Math., 646 (2010), 65–115. Zbl 1200.58025.
[39] Delanoë, P. and Loeper, G. – Gradient estimates for potentials of invertible gradient-
mappings on the sphere. Calc. Var. Partial Differential Equations 26 (2006), no. 3, 297–311.
Zbl 1136.35358.
[40] Delfour, M.C. and Zolésio, J.-P. – Shape analysis via oriented distance functions. J. Funct.
Anal. 123 (1994), no. 1, 129–201. Zbl 0814.49032.
[41] Evans, L.C. – Classical solutions of fully nonlinear, convex, second order elliptic equations.
Comm. Pure Appl. Math. 25 (1982), 333–362. Zbl 0469.35022.
[42] Evans, L.C. and Gariepy, R.F. – Measure theory and fine properties of functions. Studies in
Advanced Mathematics. CRC Press, Boca Raton, FL, 1992. Zbl 0804.28001.
[43] Figalli, A. – Regularity properties of optimal maps between nonconvex domains in the plane.
Comm. Partial Differential Equations 35 (2010), no. 3, 465–479. Zbl 1193.35086.
[44] Figalli, A. – Sobolev regularity for the Monge–Ampère equation, with application to
the semigeostrophic equations. Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov.
(POMI) 411 (2013), Teoriya Predstavleniĭ, Dinamicheskie Sistemy, Kombinatornye Metody.
XXII, 103–118, 242; translation in J. Math. Sci. (N.Y.) 196 (2014), no. 2, 175–183.
Zbl 1305.35086.
[45] Figalli, A. – Global existence for the semigeostrophic equations via Sobolev estimates for
Monge–Ampère. CIME Lecture notes, to appear.
[46] Figalli, A., Jhaveri, Y., and Mooney, C. – Nonlinear bounds in Hölder spaces for the Monge–
Ampère equation. J. Funct. Anal. 270 (2016), no. 10, 3808–3827. Zbl 1338.35196.
[47] Figalli, A. and Kim, Y.-H. – Partial regularity of Brenier solutions of the Monge–Ampère
equation. Discrete Contin. Dyn. Syst., 28 (2010), no. 2, 559–565. Zbl 1193.35087.
[48] Figalli, A., Kim, Y.-H., and McCann, R.J. – Hölder continuity and injectivity of optimal
maps. Arch. Ration. Mech. Anal., 209 (2013), no. 3, 747–795. Zbl 1281.49037.
[49] Figalli, A., Kim, Y.-H., and McCann, R.J. – Regularity of optimal transport maps on multiple
products of spheres. J. Eur. Math. Soc. (JEMS) 15 (2013), no. 4, 1131–1166. Zbl 1268.49053.
194 Bibliography

[50] Figalli, A. and Loeper, G. – C 1 regularity of solutions of the Monge–Ampère equation for
optimal transport in dimension two. Calc. Var. Partial Differential Equations 35 (2009),
no. 4, 537–550. Zbl 1170.35400.
[51] Figalli, A., Maggi, F., and Pratelli, A. – A mass transportation approach to quantitative
isoperimetric inequalities. Invent. Math. 182 (2010), no. 1, 167–211. Zbl 1196.49033.
[52] Figalli, A. and Rifford, L. – Continuity of optimal transport maps and convexity of injectivity
domains on small deformations of S2 . Comm. Pure Appl. Math., 62 (2009), no. 12, 1670–
1706. Zbl 1175.49040.
[53] Figalli, A., Rifford, L., and Villani, C. – On the Ma–Trudinger–Wang curvature on surfaces.
Calc. Var. Partial Differential Equations, 39 (2010), no. 3-4, 307–332. Zbl 1203.53034.
[54] Figalli, A., Rifford, L., and Villani, C. – Nearly round spheres look convex. Amer. J. Math.,
134 (2012), no. 1, 109–139. Zbl 1241.53031.
[55] Figalli, A., Rifford, L., and Villani, C. – Necessary and sufficient conditions for continuity
of optimal transport maps on Riemannian manifolds. Tohoku Math. J. (2), 63 (2011), no. 4,
855–876. Zbl 1262.58013.
[56] Figalli, A. and Villani, C. – Optimal transport and curvature. Nonlinear PDE’s and applica-
tions, 171–217, Lecture Notes in Math., 2028, Springer, Heidelberg, 2011. Zbl 1247.53003.
[57] Forzani, L. and Maldonado, D.– Properties of the solutions to the Monge–Ampère equation.
Nonlinear Anal. 57 (2004), 815–829. Zbl 137.35361.
[58] Gilbarg, D. and Trudinger, N.S. – Elliptic partial differential equations of second order.
Springer, New York, 1983. Zbl 0562.35001.
[59] Guillen, N. and Kitagawa, J. – The local geometry of maps with c-convex potentials. Calc.
Var. Partial Differential Equations 52 (2015), no. 1-2, 345–387. Zbl 1309.35038.
[60] Guillen, N. and Kitagawa, J. – Pointwise estimates and regularity in geometric optics and
other generated Jacobian equations. Preprint, 2015.
[61] Gutiérrez, C.E. – The Monge–Ampère equation, Progress in Nonlinear Differential Equations
and their Applications, 44. Birkhäuser Boston, Boston, MA, 2001. Zbl 0989.35052.
[62] Gutiérrez, C.E. and Huang, Q. – Geometric properties of the sections of solutions to
the Monge–Ampère equation. Trans. Amer. Math. Soc. 352 (2000), no. 9, 4381–4396.
Zbl 0958.35043.
[63] Han, Q. and Lin, F. – Elliptic partial differential equations. Courant Lecture Notes in
Mathematics, 1. New York University, Courant Institute of Mathematical Sciences, New
York. American Mathematical Society, Providence, RI, 1997. x+144 pp. Zbl 1051.35031.
[64] Horn, R.A. and Johnson, C.R. Matrix analysis. Second edition. Cambridge University Press,
Cambridge, 2013. xviii+643 pp. Zbl 1267.15001.
[65] Ivochkina, N.M. – Classical solvability of the Dirichlet problem for the Monge–
Ampère equation. (Russian). Questions in quantum field theory and statistical physics,
4. Zap. Nauchn. Sem. Leningrad. Otdel. Mat. Inst. Steklov. (LOMI) 131 (1983), 72–79.
Zbl 0522.35028. (Zbl 0569.35014 English.)
 Bibliography 195

[66] Jian, H.Y. and Wang, X.-J. – Continuity estimates for the Monge–Ampère equation. SIAM
J. Math. Anal. 39 (2007), 608–626. Zbl 1138.35326.
[67] Jiang, F. and Trudinger, N.S. – On Pogorelov estimates in optimal transportation and geo-
metric optics. Bull. Math. Sci. 4 (2014), no. 3, 407–431. Zbl 1307.90023.
[68] Jiang, F., Trudinger, N.S., and Yang, X.-P. – On the Dirichlet problem for Monge–Ampère
type equations. Calc. Var. Partial Differential Equations 49 (2014), no. 3-4, 1223–1236.
Zbl 1292.35131.
[69] John, F. – Extremum problems with inequalities as subsidiary conditions. In Studies and
Essays Presented to R. Courant on his 60th Birthday, January 8, 1948, pages 187–204.
Interscience, New York, 1948. Zbl 0034.10503.
[70] Jörgens, K. – Uber die Lösungen der Differentialgleichung rt − s2 = 1. (German). Math.
Ann. 127 (1954), 130–134. Zbl 0055.08404.
[71] Kim, Y.-H. – Counterexamples to continuity of optimal transport maps on positively
curved Riemannian manifolds. Int. Math. Res. Not. IMRN 2008, Art. ID rnn120, 15 pp.
Zbl 1160.49047.
[72] Kim, Y.-H. and McCann, R.J. – Continuity, curvature, and the general covariance of optimal
transportation. J. Eur. Math. Soc., 12 (2010), no. 4, 1009–1040. Zbl 1191.49046.
[73] Kim, Y.-H. and McCann, R.J. – Towards the smoothness of optimal maps on Riemannian
submersions and Riemannian products (of round spheres in particular). J. reine angew.
Math., 664 (2012), 1–27. Zbl 1239.53049.
[74] Knothe, H. – Contributions to the theory of convex bodies. Michigan Math. J., 4 (1957),
39–52. Zbl 0077.35803.
[75] Krylov, N.V. – Boundedly nonhomogeneous elliptic and parabolic equations. (Russian). Izv.
Akad. Nauk SSSR Ser. Mat. 46 (1982), 487– 523. English translation in Math. USSR. Izv. 20
(1983), 459–492. Zbl 0529.35026.
[76] Krylov, N.V. – Boundedly nonhomogeneous elliptic and parabolic equations in a domain.
(Russian). Izv. Akad. Nauk SSSR Ser. Mat. 47 (1983), 75–108. English translation in Math.
USSR. Izv. 20 (1983), 459–492. Zbl 0529.35026.
[77] Le, N.Q. and Nguyen, T. – Global W 2, p estimates for solutions to the linearized Monge–
Ampère equations. Math. Ann. 358 (2014), no. 3-4, 629–700. Zbl 1291.35093.
[78] Le, N.Q. and Savin, O. – Boundary regularity for solutions to the linearized Monge–Ampère
equations. Arch. Ration. Mech. Anal. 210 (2013), no. 3, 813–836. Zbl 1283.35137.
[79] Liu, J. – Hölder regularity of optimal mappings in optimal transportation. Calc. Var. Partial
Differential Equations 34 (2009), no. 4, 435–451. Zbl 1166.35331.
[80] Liu, J., Trudinger, N.S., and Wang, X.-J. – Interior C 2,α regularity for potential functions
in optimal transportation. Comm. Partial Differential Equations 35 (2010), no. 1, 165–184.
Zbl 1189.35142.
[81] Liu, J., Trudinger, N.S., and Wang, X.-J. – On asymptotic behaviour and W 2, p regularity of
potentials in optimal transportation. Arch. Ration. Mech. Anal. 215 (2015), no. 3, 867–905.
Zbl 1309.35183.
196 Bibliography

[82] Loeper, G. – A fully nonlinear version of the incompressible Euler equations: The semi-
geostrophic system. SIAM J. Math. Anal., 38 (2006), 795–823. Zbl 1134.35046.
[83] Loeper, G. – On the regularity of solutions of optimal transportation problems. Acta Math.,
202 (2009), no. 2, 241–283. Zbl 1219.49038.
[84] Loeper, G. – Regularity of optimal maps on the sphere: The quadratic cost and the reflector
antenna. Arch. Ration. Mech. Anal., 199 (2011), no. 1, 269–289. Zbl 1231.35280.
[85] Loeper, G. and Villani, C. – Regularity of optimal transport in curved geometry: The
nonfocal case. Duke Math. J., 151 (2010), no. 3, 431–485. Zbl 1192.53041.
[86] Ma, X.-N., Trudinger, N.S., and Wang, X.-J. – Regularity of potential functions of the
optimal transportation problem. Arch. Ration. Mech. Anal. 177 (2005), no. 2, 151–183.
Zbl 1072.49035.
[87] Maldonado, D. – Harnack’s inequality for solutions to the linearized Monge–Ampère op-
erator with lower-order terms. J. Differential Equations 256 (2014), no. 6, 1987–2022.
Zbl 1287.35046.
[88] Marini, M. and De Philippis, G. – A note on Petty’s theorem. Kodai Math. J. 37 (2014),
no. 3, 586–594. Zbl 1314.52003.
[89] Milman, V.D. and Schechtman, G. – Asymptotic theory of finite-dimensional normed spaces.
With an appendix by M. Gromov. Lecture Notes in Mathematics, 1200. Springer, Berlin,
1986. viii+156 pp. Zbl 0606.46013.
[90] Minkowski, H. – Allgemeine Lehrsätze über die convexen Polyeder. Nachr. Ges. Wiss.
Göttingen (1897), 198–219. Zbl 28.0427.01.
[91] Minkowski, H. – Volumen und Oberfläche. Math. Ann. 57 (1903), 447–495. Zbl 34.0649.01.
[92] Monge, G. – Sur le calcul intégral des équations aux differences partielles. Mémoires de
l’Académie des Sciences (1784).
[93] Mooney, C. – Partial regularity for singular solutions to the Monge–Ampère equation.
Commun. Pure Appl. Math. 68 (2015), no. 6, 1066–1084. Zbl 1317.35079.
[94] Mooney, C. – W 2,1 estimate for singular solutions to the Monge–Ampère equation. Ann. Sc.
Norm. Super. Pisa Cl. Sci., (5) 14 (2015), no. 4, 1283–1303. Zbl 06552095.
[95] Oliker, V. – Embedding Sn into Rn+1 with given integral Gauss curvature and optimal mass
transport on Sn . Adv. Math. 213 (2007), no. 2, 600–620. Zbl 1233.49024.
[96] Petty, C.M. – Affine isoperimetric problems. Discrete geometry and convexity (New York,
1982), 113–127, Ann. New York Acad. Sci., 440, New York Acad. Sci., New York, 1985.
Zbl 0576.52003.
[97] Pogorelov, A.V. – The regularity of the generalized solutions of the equation det(∂ 2 u
/∂ x i ∂ x j ) = ϕ(x1, x2, . . . , xn ) > 0. (Russian). Dokl. Akad. Nauk SSSR 200 (1971), 534–
537. Zbl 0246.35014.
[98] Pogorelov, A.V. – On the improper convex affine hyperspheres. Geometriae Dedicata 1
(1972), no. 1, 33–46. Zbl 0251.53005.
 Bibliography 197

[99] Rockafellar, R.T. – Convex analysis. Reprint of the 1970 original. Princeton Landmarks
in Mathematics. Princeton Paperbacks. Princeton University Press, Princeton, NJ, 1997.
xviii+451 pp. Zbl 0932.90001.

[100] Rüschendorf, L. and Rachev, S.T. – A characterization of random variables with minimum
L 2 -distance. J. Multivariate Anal. 32, 1 (1990), 48–54. Corrigendum in J. Multivariate Anal.
34, 1 (1990), p. 156. Zbl 0688.62034.

[101] Savin, O. – A Liouville theorem for solutions to the linearized Monge–Ampère equation.
Discrete Contin. Dyn. Syst. 28 (2010), no. 3, 865–873. Zbl 1193.35089.

[102] Savin, O. – A localization property at the boundary for Monge–Ampère equation. Advances
in geometric analysis, 45–68, Adv. Lect. Math. (ALM), 21, Int. Press, Somerville, MA, 2012.
Zbl 1317.35100.

[103] Savin, O. – Pointwise C 2,α estimates at the boundary for the Monge–Ampère equation. J.
Amer. Math. Soc. 26 (2013), no. 1, 63–99. Zbl 1275.35115.

[104] Savin, O. – Global W 2, p estimates for the Monge–Ampère equation. Proc. Amer. Math. Soc.
141 (2013), no. 10, 3573–3578. Zbl 1277.35192.

[105] Savin, O. – A localization theorem and boundary regularity for a class of degenerate Monge–
Ampère equations. J. Differential Equations 256 (2014), no. 2, 327–388. Zbl 1326.35170.

[106] Schmidt, T. – W 2,1+ε estimates for the Monge–Ampère equation. Adv. Math., 240 (2013),
672–689. Zbl 1290.35136.

[107] Schneider, O. – Convex bodies: The Brunn–Minkowski theory. Encyclopedia of Mathematics

and its Applications, 44. Cambridge University Press, Cambridge, 1993. Zbl 0798.52001.

[108] Schulz, F. – Regularity theory for quasilinear elliptic systems and Monge–Ampère equations
in two dimensions. Lecture Notes in Mathematics, 1445. Springer, Berlin, 1990, xviii+130
p. Zbl 0709.35038.

[109] Silvestre, L. and Sirakov, B. – Boundary regularity for viscosity solutions of fully nonlinear
elliptic equations. Comm. Partial Differential Equations 39 (2014), no. 9, 1694–1717.
Zbl 1308.35043.

[110] Trudinger, N.S. – On the Dirichlet problem for Hessian equations. Acta Math., 175 (1995),
151–164. Zbl 0887.35061.

[111] Trudinger, N.S. – On the prescribed Jacobian equation. Proceedings of International Con-
ference for the 25th Anniversary of Viscosity Solutions, 243–255, GAKUTO Internat. Ser.
Math. Sci. Appl. 30, Gakkotosho, Tokyo, 2008.

[112] Trudinger, N.S. – A note on global regularity in optimal transportion. Bull. Math. Sci. 3
(2013), no. 3, 551–557. Zbl 1283.35051.

[113] Trudinger, N.S. – On the local theory of prescribed Jacobian equations. Discrete Contin.
Dyn. Syst. 34 (2014), no. 4, 1663–1681. Zbl 1279.35052.

[114] Trudinger, N. and Urbas, J. – On second derivative estimates for equations of Monge–Ampère
type. Bull. Austral. Math. Soc. 30 (1984), no. 3, 321–334. Zbl 0557.35054.
198 Appendix: Useful facts

[115] Trudinger, N.S. and Wang, X.-J. – The Bernstein problem for affine maximal hypersurfaces.
Invent. Math. 140 (2000), no. 2, 399–422. Zbl 0978.53021.
[116] Trudinger, N.S. and Wang, X.-J. – Affine complete locally convex hypersurfaces. Invent.
Math. 150 (2002), no. 1, 45–60. Zbl 1071.53008.
[117] Trudinger, N.S. and Wang, X.-J. – The affine Plateau problem. J. Amer. Math. Soc. 18 (2005),
no. 2, 253–289. Zbl 1229.53049.
[118] Trudinger, N.S. and Wang, X.-J. – The Monge–Ampère equation and its applications.
Handbook of geometric analysis. No. 1, 467–524, Adv. Lect. Math. (ALM), 7, Int. Press,
Somerville, MA, 2008. Zbl 1156.35033.
[119] Trudinger, N.S. and Wang, X.-J. – On the second boundary value problem for Monge–
Ampère type equations and optimal transportation. Ann. Sc. Norm. Super. Pisa Cl. Sci. (5)
8 (2009), no. 1, 143–174. Zbl 1182.35134.
[120] Trudinger, N.S. and Wang, X.-J. – On strict convexity and continuous differentiability of
potential functions in optimal transportation. Arch. Ration. Mech. Anal. 192 (2009), no. 3,
403–418. Zbl 1214.49038.
[121] Urbas, J. – On the second boundary value problem for equations of Monge–Ampère type.
J. reine angew. Math. 487 (1997), 115–124. Zbl 0880.35031.
[122] Villani, C. – Topics in optimal transportation. Graduate Studies in Mathematics, 58. Amer-
ican Mathematical Society, Providence, RI, 2003. Zbl 1106.90001.
[123] Villani, C. – Optimal transport, old and new. Grundlehren des mathematischen Wis-
senschaften [Fundamental Principles of Mathematical Sciences], Vol. 338, Springer, Berlin-
New York, 2009. Zbl 1156.53003.
[124] Wang, X.-J. – Some counterexamples to the regularity of Monge–Ampère equations. Proc.
Amer. Math. Soc., 123 (1995), no. 3, 841–845. Zbl 0822.35054.
[125] Wang, X.-J. – Regularity for the Monge–Ampère equation near the boundary. Analysis, 16
(1996), no. 1, 101–107. Zbl 0844.35015.
Index

affine map area formula, 180

normalizing, see convex set definition, 172
Alexandrov maximum principle, 15 sections, 65
Alexandrov solution strictly, 174
boundary regularity, 143 subdifferential
comparison, 18 definition, 173
definition, 11 locality, 173
Dirichlet problem supporting plane, 173
existence, 20, 24 convex set
stability, 19 curvature function, 90
uniqueness, 19 exposed point, 165
interior regularity Gauss map, 34
C 1 regularity in 2-D, 29 Gaussian curvature measure, 35
C 1,α estimates, 93 normal mapping, 34
C 2,α estimates, 131 normalized, 66
W 2, p estimates, 125 normalized size, 119
counterexamples, 137 normalizing affine map, 66
higher regularity, 62 projection onto, 170
sections support function, 91
covering lemma, 110 supporting hyperplane, 165
geometry of, 107
singular solutions ellipsoid, 166
W 2,1 regularity, 145 center of, 166
stability, 12 elliptic equation
strict convexity, 74 fully nonlinear, 2
quantitative estimates, 83 Evans–Krylov theorem, 187
strict convexity in 2-D, 33 linear, 2
area formula, see convex function maximum principle, 186
Schauder’s theory, 186
Brenier solution, see optimal transport
quasilinear, 2
c-convex function, 151 semilinear, 2
c-subdifferential, 151 ellipticity
c-subdifferential, see c-convex function degenerate, 2
continuity method, 39 uniform, 2
convex envelope, 183 Evans–Krylov theorem, see elliptic
convex function equation
C 1,α regularity, 181 exposed point, see convex set
200 Index

Gauss map, see convex set normalized solutions, 67

Gaussian curvature measure, see convex Alexandrov estimates, 68
set sections, 70
Hausdorff measure, 163 strict convexity, 79

implicit function theorem in Banach optimal transport

spaces, 185 general cost, 150
isoperimetric inequality, 103 C 1,α regularity, 156
higher regularity, 154
John’s lemma, 166 MTW condition, 154
layer-cake formula, 185 partial regularity, 156
Legendre transform, 177 quadratic cost, 95
linearized Monge–Ampère, see Brenier solution, 97
Monge–Ampère partial regularity, 102
Liouville theorem, see Monge–Ampère regularity, 98
matrix partial Legendre transform, 145
determinant Petty’s theorem, 91
derivative of, 159 Pogorelov’s counterexample, 54
inverse prescribed Gaussian curvature, see
derivative of, 160 Minkowski problem
sum and determinant, 160 prescribed Jacobian equations, 156
maximum principle, see elliptic equation
Schauder’s theory, see elliptic equation
Minkowski problem, 1, 34
sections, see convex functions, see also
existence, 35
Alexandrov solution
regularity, 37
semigeostrophic equations, 127
Minkowski semisum, 166
smooth solutions
Monge–Ampère
a priori global estimates
affine invariance, 68
C 2 regularity, 42
degeneracy, 3
higher regularity, 49
history, 4
a priori interior estimates
linearized, 149
C 2 regularity, 56
Liouville theorem in Rn , 88
W 2, p regularity, 124
Liouville theorem in the half-plane,
existence, 50, 141
147
strictly convex function, see convex
Monge–Ampère measure
function
definition, 7
support function, see convex set
monotonicity, 14
supporting hyperplane, see convex set
Monge–Ampère-type equations, 150
supporting plane, see convex function
MTW condition, see optimal transport
universal constant, 65
normal mapping, see convex set
normalized size, see convex set
 ZURICH LECTURES IN ADVANCED MATHEMATICS ZURICH LECTURES IN ADVANCED MATHEMATICS

Alessio Figalli
Alessio Figalli
Alessio Figalli

The Monge-Ampère Equation and its Applications

The Monge-Ampère Equation
and its Applications
The Monge–Ampère equation is one of the most important partial differential equations,
appearing in many problems in analysis and geometry. The Monge-Ampère
Equation and
This monograph is a comprehensive introduction to the existence and regularity theory of
the Monge–Ampère equation and some selected applications; the main goal is to provide
the reader with a wealth of results and techniques he or she can draw from to understand

Its Applications
current research related to this beautiful equation.

The presentation is essentially self-contained, with an appendix wherein one can find
precise statements of all the results used from different areas (linear algebra, convex
geometry, measure theory, nonlinear analysis, and PDEs).

This book is intended for graduate students and researchers interested in nonlinear PDEs:
explanatory figures, detailed proofs, and heuristic arguments make this book suitable for
self-study and also as a reference.

ISBN 978-3-03719-170-5

www.ems-ph.org

Figalli Cover (ZLAM) | Fonts: RotisSemiSans / DIN | Farben: 4c Pantone 116, Pantone 287, Cyan | RB 10.4 mm