(CMS - CAIMS Books in Mathematics 11) Arian Novruzi - A Short Introduction To Partial Differential Equations-Springer Nature Switzerland (2023)

CMS/CAIMS Books in Mathematics
Canadian
Mathematical Society
Société mathématique
du Canada
Arian Novruzi
A Short Introduction
to Partial Differential
Equations
Volume 11
Series Editors
Karl Dilcher
Department of Mathematics and Statistics, Dalhousie University, Halifax, NS, Canada
Frithjof Lutscher
Department of Mathematics, University of Ottawa, Ottawa, ON, Canada
Nilima Nigam
Department of Mathematics, Simon Fraser University, Burnaby, BC, Canada
Keith Taylor
Department of Mathematics and Statistics, Dalhousie University, Halifax, NS, Canada
Associate Editors
Ben Adcock
Department of Mathematics, Simon Fraser University, Burnaby, BC, Canada
Martin Barlow
University of British Columbia, Vancouver, BC, Canada
Heinz H. Bauschke
University of British Columbia, Kelowna, BC, Canada
Matt Davison
Department of Statistical and Actuarial Science, Western University, London, ON, Canada
Leah Keshet
Department of Mathematics, University of British Columbia, Vancouver, BC, Canada
Niky Kamran
Department of Mathematics and Statistics, McGill University, Montreal, QC, Canada
Mikhail Kotchetov
Memorial University of Newfoundland, St. John’s, Canada
Raymond J. Spiteri
Department of Computer Science, University of Saskatchewan, Saskatoon, SK, Canada
CMS/CAIMS Books in Mathematics is a collection of monographs and graduate-level
textbooks published in cooperation jointly with the Canadian Mathematical Society- Societé
mathématique du Canada and the Canadian Applied and Industrial Mathematics
Society-Societé Canadienne de Mathématiques Appliquées et Industrielles. This series offers
authors the joint advantage of publishing with two major mathematical societies and with a
leading academic publishing company. The series is edited by Karl Dilcher, Frithjof Lutscher,
Nilima Nigam, and Keith Taylor. The series publishes high-impact works across the breadth of
mathematics and its applications. Books in this series will appeal to all mathematicians,
students and established researchers. The series replaces the CMS Books in Mathematics
series that successfully published over 45 volumes in 20 years.
Arian Novruzi
A Short Introduction to Partial

Differential Equations
123
Arian Novruzi
Department of Mathematics
University of Ottawa
Ottawa, ON, Canada
ISSN 2730-650X ISSN 2730-6518 (electronic)

ISBN 978-3-031-39523-9 ISBN 978-3-031-39524-6 (eBook)
https://doi.org/10.1007/978-3-031-39524-6
Mathematics Subject Classification: 35FXX, 65M25, 35BXX, 46FXX, 46EXX, 35JXX, 35KXX, 35LXX
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or
part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and
retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and
regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book are believed
to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty,
expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been
made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Paper in this product is recyclable.

To my beloved parents, Xhiko and Milo
Preface
There are many publications on Partial Differential Equations (PDEs) in a multitude of forms, varying from
research-oriented papers and books to graduate/undergraduate textbooks. The number of topics and the depth of
analysis in these publications also varies. Some of them address a limited number of topics, while others cover a
broader spectrum of topics, with a level of difficulty varying from an introductory to an advanced or expert level.
This book provides a short introduction to PDEs. It is primarily addressed to graduate students and researchers,
who are new to PDEs. The book offers a user-friendly approach to the analysis of PDEs, by combining elementary
techniques with important concepts and fundamental modern methods.
This book focuses on the analysis of four prototypes of PDEs: first-order PDEs, second-order linear elliptic PDEs,
and two linear evolution PDEs—heat and wave equations, each with a second-order linear elliptic PDE principal
term. To facilitate a smooth introduction into the analysis for the reader, two approaches are presented for each of the
PDEs. The first approach consists of the method of analytical and classical solutions. It includes the method of
characteristics, the method of separation of variables, and Perron’s method. The simplicity of this approach, its
potential to provide classical solutions, and the impossibility of providing a classical solution in a general setting are
highlighted, and used to motivate the second approach.
The second approach is the method of weak (variational) solutions. While for the first-order PDEs we focus the
analysis on a short discussion of scalar conservation laws, we present a detailed analysis for the other PDEs. The main
ingredients we use are the Lax-Milgram lemma, Fredholm alternative and spectral decomposition theorems, and a
number of results from Sobolev spaces. As an application of the results from the second-order linear elliptic PDEs and
Sobolev spaces, we give a very short introduction to the solution to certain nonlinear PDEs by using a fixed point
approach.
In connection with the second approach, we give an introduction to distributions, Fourier transform, and Sobolev
spaces, which are fundamental for the study of PDEs. Though in a number of cases we deal with general fractional
Sobolev spaces Ws,p, we mainly analyze Hs Sobolev spaces by using Fourier transform. We give some fundamental
results related to the concepts of density, extension, embeddings, compactness, and boundary traces. We provide
proofs of some of these results in particular cases, with the intention to highlight the most relevant techniques and to
avoid the complexity of general cases. The book ends with an appendix chapter, which complements the previous
chapters with proofs, examples, and remarks.
Chapter 1 (introduction) provides some concepts and results from analysis, functional analysis, and topology,
which will be used as the book develops. At first reading, one only needs to get familiar with them. Chapter 2 provides
the formal definition of PDEs and a number of examples. Chapters 3 (first-order PDEs) and 4 (second-order PDEs
and maximum principle) can be read independently. Each of them depends on certain results from chapter 1.
Chapters 7 (second-order PDEs and weak solutions) and 8 (evolution PDEs) are almost independent, and each
of them depends on chapters 6 (Sobolev spaces) and 5 (distributions), and some results from chapter 1.
vii
viii Preface
This book can be used for an intense one-semester, or normal two-semester, PDE course. The reader is expected to
have knowledge of linear algebra and differential equations, a good background in real and complex calculus, and a
modest background in analysis and topology. The book has many examples, which help to explain the concepts,
highlight the key ideas, and emphasize the sharpness of results, as well as a section of problems at the end of each
chapter.
I am grateful to professor Michel PIERRE for a number of comments and suggestions, which have been helpful for me
to improve this book.
Ottawa, Canada Arian Novruzi

June 2023
Contents
1 Notations and review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Continuous differentiable functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Domains and Ck ð@XÞ spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.1 Partition of unity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.2.2 Domains and Ck ð@XÞ spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3 Review of some important results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Some results from Lp ðXÞ spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Some results from (Functional) Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.3 An application: Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Partial differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.1 Some prototypes of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 First-order PDEs: classical and weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.1 Method of characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Classical local solutions to first-order PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.1 Classical local solutions: flat boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1.1 Boundary and initial conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1.2 Classical local solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.2 Classical local solutions: non-flat boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Conservation laws and weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4 Second-order linear elliptic PDEs: maximum principle and classical solutions . . . . . . . . . . . . . . 49

4.1 Laplace equation and the method of separation of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Dirichlet problem in a ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.3 Maximum principle for Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.4 Solution to the Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.1 Sub(super) harmonic functions and sub(super) solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.4.2 Solution to the Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2.1 Test functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.2.2 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.2.3 Derivatives of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.3 Convolution of distributions and fundamental solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
ix
x Contents
5.4 Tempered distributions and Fourier transform . . . . . . . . . . . . . . . . . . . . . 81

5.4.1 Fourier transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.4.2 Tempered distributions and Fourier transform . . . . . . . . . . . . . . . 85
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1 Definitions and some first properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.1.1 Density of D in W k;p . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
6.1.2 Some applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.2 H s spaces and Fourier transform: W s;p and W0s;p spaces . . . . . . . . . . . . . . 101
6.3 Continuous, compact, and dense embedding theorems in H s ðXÞ . . . . . . . . 105
6.3.1 Case X ¼ RN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ....... 106
6.3.2 Case X ( RN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.4 Boundary traces in Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5 Poincaré inequality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
6.6 H s ðXÞ and W s;q ðXÞ spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
7 Second-order linear elliptic PDEs: weak solutions . . . . . . . . . . . . . . . . . 123

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
7.2 Existence and uniqueness of weak solutions . . . . . . . . . . . . . . . . . . . . . . . 125
7.2.1 Preliminary results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.2.2 Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
7.2.3 Neumann problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
7.3 Nonlinear second-order elliptic PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
8 Second-order parabolic and hyperbolic PDEs . . . . . . . . . . . . . . . . . . . . . 139

8.1 Heat and wave equations and the method of separation of variables . . . . 139
8.1.1 Heat equation and the method of separation of variables . . . . . . . 140
8.1.2 Wave equation and the method of separation of variables . . . . . . . 143
8.2 Some preliminary results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8.3 Weak solution to the heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.4 Weak solution to the wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
9 Annex . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.1 Notations and review (Ch. 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
9.1.1 Continuous differentiable functions . . . . . . . . . . . . . . . . . . . . . . . . 159
9.1.2 Some results from Lp ðXÞ spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
9.1.3 An application: Ordinary Differential Equations . . . . . . . . . . . . . . 161
9.2 First-order PDEs: classical and weak solutions (Ch. 3) . . . . . . . . . . . . . . 163
9.2.1 Classical local solutions to first-order PDEs . . . . . . . . . . . . . . . . . 163
9.2.2 Conservation laws and weak solutions . . . . . . . . . . . . . . . . . . . . . . 166
9.3 Second-order linear elliptic PDEs: maximum principle and classical
solutions (Ch. 4) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 169
9.3.1 Dirichlet problem in a ball . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 169
9.3.2 Maximum principle for second-order linear elliptic PDEs . . . .... 170
Contents xi
9.3.3 Solution to the Dirichlet problem . . . . . . . . . . . . . . . . . . . ...... 174

9.3.3.1 Sub(super) harmonic functions and sub(super)
solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
9.3.3.2 Some auxiliary results from Analysis . . . . . . . . . . . . . . . . 176
9.3.3.3 Proof of Theorem 4.4.7 . . . . . . . . . . . . . . . . . . . . . . . . . . 178
9.4 Distributions (Ch. 5) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.4.1 Some useful inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
9.4.2 More operations with distributions. Examples . . . . . . . . . . . . . . . . 181
9.4.3 Convergence of distributions. Distributions of finite order . . . . . . . 184
9.4.4 Convolution of distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
9.4.5 Tempered distributions and Fourier transform . . . . . . . . . . . . . . . 187
9.4.6 Tempered distributions and convolution . . . . . . . . . . . . . . . . . . . . 191
9.5 Sobolev spaces (Ch. 6) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.5.1 Continuous and compact embeddings . . . . . . . . . . . . . . . . . . . . . . 193
9.5.2 Extension and density results in Sobolev spaces . . . . . . . . . . . . . . 196
9.5.3 Boundary traces in Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . 200
9.6 Second-order linear elliptic PDEs: weak solutions . . . . . . . . . . . . . . . . . . 202
9.6.1 Regularity of weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9.6.1.1 Regularity in the interior . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.6.1.2 Regularity near the boundary . . . . . . . . . . . . . . . . . . . . . 206
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
1. Notations and review
The objective of this chapter is to help the reader become familiar with a number
of concepts and some important results that will be used throughout this book. It is
expected that the reader has a minimum background in analysis in RN and topology.
1.1 Continuous differentiable functions

In this book, N = {1, 2, . . .} is the set of positive integers, N0 = N ∪ {0}, R is
the set of real numbers, C the set of complex numbers, and for N ∈ N, RN = {x =
(x1 , . . . , xN ), x1 , . . . , xN ∈ R}, CN = {x = (x1 , . . . , xN ), x1 , . . . , xN ∈ C}.
Before we start with the definition of classical continuous differentiable functions in
Banach spaces, we are reminded the simplest of Banach spaces (R, | · |), with | · | the
absolute value norm. Another Banach space is (RN , | · |p ), where N ∈ N, p ∈ [1, ∞] and
| · |p is given by

N (|x1 |p + · · · + |xN |p )1/p , p ∈ [1, ∞),
∀x = (x1 , . . . , xN ) ∈ R , |x|p = (1.1.1)
max{|x1 |, . . . , |xN |}, p = ∞.
In general, a Banach space is defined as follows.

Definition 1.1.1 (Banach spaces) A Banach space is a couple (X, · X ), where X
is a vector space over the field of real numbers R (or of complex numbers C) and · X
is a norm, such that X is complete with respect to the norm · X , i.e.1 every Cauchy2
sequence (xn ) in X converges to an element x ∈ X in the norm · X , which is equiva-
lent to
lim xn − xX = 0.

n→∞
n
We note that if (Xi , · Xi ) are Banach spaces, Z = i=1 Xi = X1 × · · · × Xn and
zZ = max{x1 X1 , . . . , xn Xn }, (or zZ = x1 X1 + · · · + xn Xn ) (1.1.2)
1
In Latin id est, which means that is
2
(xn ) is a Cauchy sequence if for every > 0 there exists n ∈ N such that xn+m − xn X < for
all integers n > n and m ∈ N.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023 1
A. Novruzi, A Short Introduction to Partial Differential Equations, CMS/CAIMS Books in Mathematics 11,
https://doi.org/10.1007/978-3-031-39524-6 1
1.1. Continuous differentiable functions Chapter 1
for z = (x1 , . . . , xn ) ∈ Z, then it is easy to show that (Z, · Z ) is another Banach space.
If X1 = · · · = Xn = X we write X n instead of X × · · · × X .

n times
Definition 1.1.2 (C 0 (Ω; Y ) spaces) Let (X, · X ), (Y, · Y ) be two Banach spaces,
Ω ⊂ X be an open set, and u : Ω → Y a function. We say “u is continuous at x ∈ Ω” if
lim u(x + h) − u(x)Y = 0 (1.1.3a)

hX →0
∀ > 0, ∃δ,x > 0, ∀h ∈ Ω, hX < δ ⇒ u(x + h) − u(x)Y < . (1.1.3b)
If u is continuous at every x ∈ Ω, we say “u is continuous in Ω”. The set of such

functions u is denoted by C 0 (Ω; Y ), or simply C 0 (Ω) if Y = R.
A function u ∈ C 0 (Ω; Y ) is said to be “uniformly continuous in Ω” if there exists
δ > 0 such that δ,x in (1.1.3b) satisfies δ,x ≥ δ for all x ∈ Ω.
We will see below that the definition of the derivatives of functions involve the spaces
of linear continuous functions, which are simple examples of Banach spaces.
Definition 1.1.3 (L(X n ; Y ) spaces) Let (X, ·X ) and (Y, ·Y ) be two Banach spaces
and n ∈ N.
i) : X n → Y is said to be “n-linear” if
(x1 , . . . , xi−1 , ci xi + ci xi , xi+1 , . . . , xn ) = ci (x1 , . . . , xi−1 , xi , xi+1 , . . . , xn )
+ ci (x1 , . . . , xi−1 , xi , xi+1 , . . . , xn ),
for all i = 1, . . . , n, x1 , . . . , xi−1 , xi , xi , xi+1 , · · · , xn ∈ X and ci , ci ∈ R.
ii) L(X n ; Y ), or simply L(X; Y ) when n = 1, is defined as
L(X n ; Y ) = { ∈ C 0 (X n ; Y ), is n-linear}. (1.1.4)
Given ∈ L(X n ; Y ) and x ∈ X, sometimes instead of (x) we will write , xL(X n ;Y )×X n ,
and we will remove L(X n ; Y ) × X n from , x whenever there is no ambiguity. Also, we
will omit Y when Y = R, so we will write L(X n ) instead of L(X n ; R), and furthermore
we will write X , resp. X instead of L(X), resp. L(X ).
Remark 1.1.4 A n-linear map from X n in Y is continuous if and only if
(x)Y ≤ CxX n , ∀x ∈ X n ,
with a certain C ≥ 0 independent of x, see [26], and
(x)Y = inf{C ≥ 0, (x)Y ≤ CxX n , ∀x ∈ X n }. (1.1.5)
2
Chapter 1 1.1. Continuous differentiable functions
For this reason, a “continuous n-linear map” is equivalent to a “bounded n-linear map”.
Furthermore, one can show that
L(X n ;Y ) = sup{(x)Y , x ∈ X n , xX n = 1}, ∈ L(X n ; Y ), (1.1.6)
is a norm in L(X n ; Y ) and (L(X n ; Y ), · L(X n ;Y ) ) is a Banach space.
Example 1.1.5 In the case when X = RN and Y = R we have
N
N N
L(R ) = : R → R, (x) = Ai xi ,
i=1
A = (Ai ) ∈ RN , x = (xi ) ∈ RN ,
N
L(R2N ) = : RN × RN → R, (x1 , x2 ) = Ai,j x1i x2j ,
i,j=1
N2
A = (Ai,j ) ∈ R , x1 = (x1i ), x2 = (x2j ) ∈ RN ,
··· : ···
Now we define the C k spaces in general Banach spaces.
Definition 1.1.6 (C k (Ω; Y ) spaces) Let (X, · X ), (Y, · Y ) be two Banach spaces,
Ω ⊂ X be an open set, and u ∈ C 0 (Ω; Y ).
1) We say u is differentiable at x ∈ Ω if there exists u(1) (x) ∈ L(X; Y ), called
“derivative of u at x”, such that
u(x + h) − u(x) − u(1) (x; h)Y
lim = 0.
hX →0 hX
Here u(1) (x; h) denotes the value of u(1) (x) at h. If u is differentiable at every
x ∈ Ω, then u(1) : Ω → L(X; Y ) denotes “the embedded derivative of u”. Further-
more we set
C 1 (Ω; Y ) = {u ∈ C 0 (Ω; Y ) such that u(1) ∈ C 0 (Ω; L(X; Y ))}. (1.1.7)
2) For k = 2, 3, . . ., “the k-th order derivative of u at x” is defined by u(k) (x) :=

(u(k−1) )(1) (x) ∈ L(X; L(X; . . . , L(X; L(X; Y )))), and if it is continuous in Ω set

k times
C k (Ω; Y ) = {u ∈ C 0 (Ω; Y ) such that for alln = 1, . . . , k we have

u(n) ∈ C 0 (Ω; L(X; L(X; L(X; . . . , L(X; L(X; Y )))))}. (1.1.8)

n times
We also define C ∞ (Ω; Y ) = C k (Ω; Y ). (1.1.9)

k∈N
3
3) The space of Lipschitz functions is defined by
Lip(Ω; Y ) = C 0,1 (Ω; Y )

= {u ∈ C 0 (Ω; Y ), ∃L = L(u) ∈ [0, ∞) such that
∀x, y ∈ Ω, u(x) − u(y)Y ≤ Lx − yX }. (1.1.10)
4) When Y = R we will omit Y in the spaces of this definition, so we will write

C 1 (Ω), C k (Ω), C ∞ (Ω), Lip(Ω). Also, we will write u , resp. u , u , instead of
u(1) , resp. u(2) , u(3) .
Remark 1.1.7 Instead of the spaces L(X; L(X; . . . , L(X; L(X; Y )))) in Definition

k times
1.1.6 we can use L(X k ; Y ). This follows from the fact that there exists a linear bijective
isometry i : L(X n ; L(X m , Y )) → L(X n+m ; Y ) defined by for n,m ∈ L(X n ; L(X m , Y )),
i(n,m )(x1 , . . . , xn , xn+1 , . . . , xn+m ) = n,m (x1 , . . . , xn )(xn+1 , . . . , xn+m ). (1.1.11)
The proof is simple and we refer the reader to [8].
Example 1.1.8 Let a, b ∈ R, a < b, X = C 0 ([a, b]) be the set of continuous functions
from [a, b] to R endowed with the maximum norm, uX = max{|u(x)|, x ∈ [a, b]}. It
is a classical fact that (X, · X ) is a Banach space; see [7, 26]. Let f : X → X,
1
f (u) = . Clearly f (u) ∈ X and f ∈ C 0 (X; X) because
1 + u2
1 1
f (u + h) − f (u)X = 2
− ≤ (hX + 2uX )hX
1 + (u + h) 1 + u2 X
→ 0 as hX → 0.
uh
Furthermore, f (u) exists for all u ∈ X and f (u; h) = −2 . Indeed, clearly h →
(1 + u2 )2
f (u; h) ∈ L(X; X) because for h ∈ X we have f (u; h)X ≤ 2uX hX . Furthermore
1
f (u + h) − f (u) − f (u; h)X ≤ (1 + 2hX uX + 3u2X )hX
hX
→ 0 as hX → 0.
Finally, f ∈ C 0 (X; L(X; X)) because
f (u + v) − f (u)L(X;X) = sup{f (u + v; h) − f (u; h)X , h ≤ 1}

(u + v)h uh
= 2sup − , h ≤ 1
(1 + (u + v)2 )2 (1 + u2 )2 X
(u + v) u
≤ 2 2 2
−
(1 + (u + v) ) (1 + u2 )2 X
→ 0 as vX → 0.
Actually, one can show that f ∈ C ∞ (X; Y ).
4
In the following we consider X = RN and Y = R. In this case we have equivalent

definitions of C k (Ω) spaces, which are given in terms of partial derivatives with which
the reader is familiar. But first, let us introduce the following notations that will be used
throughout this book.
Notations 1.1.9 A α = (α1 , . . . , αN ) ∈ NN N
0 is called a “multi-index”. For α, β ∈ N0 ,
x = (x1 , . . . , xN ) ∈ RN , and m ∈ N, we define
⎧
⎪
⎪ |α| = α1 + · · · + αN , α ≤ β ⇐⇒ αi ≤ βi , ∀i,
⎨
α! = α1 ! · · · αN !, α ≤ m ⇐⇒ αi ≤ m, ∀i, (1.1.12)
⎪
⎪
⎩ x α = x α1 · · · x αN .
1 N
For α ∈ NN 0 and u : R
N
→ R having continuous partial derivatives of order |α| in the
usual sense, we write
∂ |α| u αN
Dα u = = ∂xα11 · · · ∂xαNN u = ∂1α1 · · · ∂N u = ux1 · · · x1 ······xN · · · xN , (1.1.13)
∂xα11 · · · ∂xαNN
α1 times αN times
α
D u = u, if |α| = 0.
We say Dα u is an “α derivative of u”, and a “Dα u is a derivative of order |α|”. Finally,
for m ∈ N0 we denote by Dm u = (u1 , u2 , . . . , uN m ) the vector of m-th order partial
derivatives of u, defined by recurrence as follows:
D0 u := u,
D1 u = (D11 , . . . , DN
1
) := (∂1 u, . . . , ∂N u),
D u = (D1 , . . . , DN 2 ) := (D1 D11 u, . . . , D1 DN
2 2 2 1
u),
···
D u = (D1m , . . . , DN
m m 1 m−1
m ) := (D D1
m−1
u, . . . , D1 DN m−1 u).
Often we write Du or ∇u instead of D1 u.
Proposition 1.1.10 Let Ω ⊂ RN be an open set. Then the spaces C k (Ω) of Definition
1.1.6 with X = RN , Y = R, are equivalently defined as
C k (Ω) = {u : Ω → R, Dα u ∈ C 0 (Ω), ∀α ∈ NN
0 , |α| ≤ k}, (1.1.15)
For the proof the reader can see [7, 8]. The idea of the proof, in the case k = 1 to avoid
the technicalities, is to point out that if u ∈ C 1 (Ω) in the sense of Definition 1.1.6, then
necessarily u ∈ C 0 (Ω) and all ∂xi u(x) := u (x; ei ), ei = (0, . . . , 0, 1, 0, . . . , 0) with 1 on
the i-th place, exist and are continuous in Ω. Conversely, if all ∂xi u ∈ C 0 (Ω) then one
n
defines u (x) ∈ L(RN ) by u (x; h) = i=1 ∂xi u(x)hi , h = (h1 , . . . , hn ), and proves that
u ∈ C 1 (Ω) in the sense of Definition 1.1.6 by using the mean value theorem and the
continuity of all ∂xi u.
5
In general, C k (Ω) is not a Banach space because it contains unbounded functions.

However, its subspace of bounded functions is a Banach space.
Proposition 1.1.11 Let Ω ⊂ RN be open, k ∈ N0 , λ ∈ (0, 1], and define
Cb0 (Ω) = {u ∈ C 0 (Ω), u is bounded in Ω}, (1.1.16)

k α N
Cb (Ω) = {u ∈ Cb (Ω), D u ∈ Cb (Ω), ∀α ∈ N0 with |α| ≤ k},
0 0
(1.1.17)

|Dα u(x) − Dα u(y)|
Cbk,λ (Ω) = u ∈ Cbk (Ω), |u|C k,λ (Ω) := sup < ∞ , (1.1.18)
b
|α|=k |x − y|λ
Ωx=y∈Ω
Cb∞ (Ω) = Cbk (Ω). (1.1.19)

k∈N0
The space Cbk (Ω), resp. Cbk,λ (Ω), k ∈ N0 , equipped with the norm
uCbk (Ω) = sup{Dα uCb0 (Ω) , |α| ≤ k}, resp. (1.1.20)

uC k,λ (Ω) = uCbk (Ω) + |u|C k,λ (Ω) , (1.1.21)
b b
is a Banach space.
Proof. See Proposition 9.1.1 in Annex.
The spaces C k,λ are called “Hölder spaces”. They are useful when describing addi-
tional regularity of functions in Sobolev spaces; see chapter 6. We note that in connection
with C k,λ spaces there is a complete theory associated with second-order linear partial
differential equations, called “Hölder theory”; see, for example, [20].
We also note that a function in Cbk (Ω) is not necessarily continuous up to the bound-
2x1 x2
ary. For example, u(x) = 2 ∈ Cb0 (Ω), Ω = (0, 1) × (0, 1), but u ∈/ Cb0 (Ω). The
x1 + x22
following subspaces of Cb0 (Ω) are useful when approximating, or describing additional
regularity, of Sobolev spaces functions, and are defined by
Cb0 (Ω) = {u ∈ Cb0 (Ω), u extends continuously in Ω}, (1.1.22)

k α N
Cb (Ω) = {u ∈ Cb (Ω), D u ∈ Cb (Ω), ∀α ∈ N0 with |α| ≤ k},
0 0
(1.1.23)

α α
|D u(x) − D u(y)|
Cbk,λ (Ω) = u ∈ Cbk (Ω), |u|C k,λ (Ω) := sup <∞ ,
|α|=k |x − y|λ
Ωx=y∈Ω
Cb∞ (Ω) = Cbk (Ω), (1.1.24)

k∈N0
6
where Ω is the closure of Ω.3,4 If Ω is bounded, we can remove the subscript “b” from
the spaces Cbk (Ω) and Cbk,λ (Ω), because every continuous function in a compact set5
is bounded. The space Cbk (Ω), resp. Cbk,λ (Ω), endowed with the norm (1.1.20), resp.
(1.1.21), is a Banach space.
When working with the distributions or the weak solution to PDEs, we are interested
in C k (Ω) functions which are zero in a neighborhood of the boundary ∂Ω of Ω.
Definition 1.1.12 (C0k (Ω) spaces) Let Ω ⊂ RN be an open set. Then we define
C00 (Ω) = {u ∈ C 0 (Ω),

supp(u) := {x ∈ Ω, u(x) = 0} is compact and included in Ω}, (1.1.25)
C0 (Ω) = {u ∈ C k (Ω), Dα u ∈ C00 (Ω), α ∈ NN , |α| ≤ k},
k
(1.1.26)
C0∞ (Ω) = C0k (Ω) =: D(Ω). (1.1.27)
k∈N∪{0}
Example 1.1.13 Let ρ : RN → R be defined by

1
ρ(x) =
1 e |x|2 −1 , |x| < 1, C=
1
e |x|2 −1 dx. (1.1.28)
C 0, |x| ≥ 1, {|x|<1}
It is easy to verify that

N
i) ρ ∈ D(R ), supp(ρ){|x| ≤ 1}, ii) ρ(x)dx = 1. (1.1.29)
RN
We set ρn (x) = nN ρ(nx), n ∈ N. The sequence (ρn ) satisfies

⎧ N
⎨ ρn ∈ D(R ), ρn ≥ 0,
⎪
supp(ρn ) ⊂ x ∈ RN , |x| ≤ 1
, (1.1.30)
⎪
⎩
n
ρ (x)dx = 1.
RN n
Every function satisfying (1.1.29) is called a “mollifier”. Given a mollifier ρ, the

sequence (ρn ), which necessarily satisfies (1.1.30), is called a “mollifier sequence”. The
mollifier sequences are useful when used with the convolution operator; see Theorem
1.3.5.
3
The closure Ω of Ω ⊂ X, X Banach, is defined as Ω = Ω ∪ ∂Ω, where ∂Ω is the boundary of Ω.
4
The boundary ∂Ω of Ω ⊂ X, X Banach, is defined as ∂Ω = {x ∈ X, B(x, r) ∩ Ω = ∅, B(x, r) ∩
Ωc = ∅, ∀r > 0}.
5
In general, K ⊂ X, X Banach, is compact if every sequence (xn ) in K has a convergent sequence
with limit in K.
7
1.2. Domains and C k (∂Ω) spaces Chapter 1
Usually, PDEs are equipped with boundary conditions, which describe the solution on
the boundary. It is then necessary to define some spaces of functions on the bound-
ary. First we define the space C 0 (∂Ω). The definition of C k (∂Ω) spaces requires more
regularity of ∂Ω and will be introduced in the next section.
Definition 1.1.14 (C 0 (∂Ω) spaces) Let Ω ⊂ RN . A function u : ∂Ω → R is said to be
“continuous at x ∈ ∂Ω” if lim |u(y) − u(x)| = 0, and it is said to be “continuous on
|y−x|→0
y∈∂Ω
∂Ω” if it is continuous at every x ∈ ∂Ω. We define C 0 (∂Ω) by
C 0 (∂Ω) = {u : ∂Ω → R, u continuous on ∂Ω}. (1.1.31)
When ∂Ω is unbounded one defines Cb0 (∂Ω) = C 0 (∂Ω) ∩ {u bounded}.
We note that when Ω is bounded, we have Cb0 (∂Ω) = C 0 (∂Ω). The space Cb0 (∂Ω) is a
Banach space when equipped with the norm
uC 0 (∂Ω) = max{|u(x)|, x ∈ ∂Ω}. (1.1.32)
1.2 Domains and C k (∂Ω) spaces

In this section we will introduce the concept of regularity of domains, which is neces-
sary for the definition of classical C k (∂Ω) spaces (and also boundary Sobolev spaces—we
will see this in chapter 6). Regularity is an important property of the domain and it
strongly affects the regularity to the solution to partial differential equations near the
boundary; see section 9.6.
1.2.1 Partition of unity

Many proofs of PDE results involve a reduction to local considerations around a
point, by using the so-called “partition of unity”.
Definition 1.2.1 (Partition of unity) Let Γ ⊂ RN be a compact m set and {G1 , . . . , Gm }
N
be a family of open bounded sets in R covering Γ, i.e. Γ ⊂ i=1 Gi . A partition of unity
subordinate to the covering {G1 , . . . , Gm } of Γ is {ϕ0 , ϕ1 , . . . , ϕm }, with ϕ0 ∈ C ∞ (RN )
and ϕ1 , . . . , ϕm ∈ D(RN ) satisfying
i) 0 ≤ ϕi ≤ 1, for all i = 0, 1, . . . , m,
m
ii) ϕi (x) = 1 for all x ∈ RN ,
i=0
iii) supp(ϕi ) ⊂ Gi for all i = 1, . . . , m, supp(ϕ0 ) ∩ Γ = ∅.
N
mwith Ω ⊂ R an open bounded set, we have θ0 = 0 in
Note that in the case Γ = ∂Ω
a neighborhood of ∂Ω, and so i=1 ϕi = 1 in the same neighborhood of ∂Ω.
8
Chapter 1 1.2. Domains and C k (∂Ω) spaces
Theorem 1.2.2 (cut-off function and partition of unity)

i) (cut-off function) Let K ⊂ RN be compact and G ⊂ RN open with K ⊂ G.
There exists η ∈ D(RN ) such that
0 ≤ η ≤ 1, η = 1 in K and supp(η) ⊂ G. (1.2.2)
Such a function η is called a “{K, G} cut-off function”.
ii) (partition of unity) Let Γ ⊂ RN and {G1 , . . . , Gm } be as in Definition 1.2.1.
Then
1) for every Gi there exist Ui⊂ RN open such that Ui Gi , i.e. “U i ⊂ Gi and
U i is compact”, and Γ ⊂ m i=1 Ui , and
2) there exists a partition of unity subordinate to the covering {G1 , . . . , Gm } of Γ.
Proof of i). There exists δ > 0 such that {B(x, 3δ), x ∈ K} ⊂ G. Indeed, if not, there
exist a positive sequence of numbers (δn ) with limn→∞ δn = 0, and sequences (xn ) in K
and (yn ) in Gc such that yn ∈ B(xn , 3δn ). As K is compact, there exists a subsequence
of (xn ), denoted again by (xn ), converging to a certain x0 ∈ K. We still denote by
(yn ), resp. (δn ), the corresponding subsequence of the original sequence (yn ), resp. (δn ),
associated with the subsequence (xn ). Then necessarily limn→∞ yn = x0 and so x0 ∈ Gc ,
because Gc is closed. So x0 ∈ K ∩ Gc , which is a contradiction.
Set Uδ = ∪{B(x, δ), x ∈ K} and Vδ = ∪{B(x, δ), x ∈ Uδ }. Clearly K ⊂ Uδ ⊂ Vδ ⊂
G. For n ∈ N define

ηn (x) = ρn (x − y)dy = ρn (x − y)dy,
Uδ Uδ ∩B(x,1/n)
where (ρn ) is the mollifier sequence in Example 1.1.13. Clearly ηn ∈ D(RN ), ηn ≥ 0 and
for n such that 1/n < δ we have
(i) ηn (x) = 1 for x ∈ K, because B(x, 1/n) ⊂ Uδ for all x ∈ K, and
(ii) ηn (x) = 0 for x ∈ Vδc , because B(x, 1/n) ∩ Uδ = ∅,
which proves i).
Proof of ii). For x ∈ Γ there exists i ∈ {1, . . . , m} and x > 0 such that B(x, x ) Gi .
As Γ is compact, from its open covering {B(x, x ), x ∈ Γ} there exists a finite covering
{B(xj, j ), j = 1, . . . , M } of Γ with B(xj , j ) Gij for a certain ij ∈ {1, . . . , m}. Then
Ui := {B(xj , j ), B(xj , j ) Gi }, i = 1, . . . , m, satisfies the claim ii.1).
Let ηi ∈ D(R n
), i = 1, . . . , m, be a {U i , Gi } cut-off function, and set η = m i=1 ηi .
Clearly η ≥ 1 in m i=1 i U . Set U 0 = {x ∈ RN
, η(x) > 3/4}, G 0 = {x ∈ RN
, η(x) > 1/2}
and let η0 be a {U 0 , G0 } cut-off function. Then
ηi
η0 in G0 ,
ϕ0 = 1 − η0 , ϕi = η i = 1, . . . , m,
0 in Gc0 ,
9
1.2. Domains and C k (∂Ω) spaces Chapter 1
is the desired partition.6
1.2.2 Domains and C k (∂Ω) spaces

To characterize the regularity of an open set Ω, we introduce the following definitions;
see, for example, [20].
Notations 1.2.3 For x ∈ RN , we write x = (x , xN ), with x = (x1 , . . . , xN −1 ), and set
|x| = (x21 + · · · + x2N −1 + x2N )1/2 , |x | = (x21 + · · · + x2N −1 )1/2 ,
RN
+ = {x ∈ RN , xN > 0},
Q = {x = (x , xN ) ∈ RN , |x | < 1, −1 < xN < 1},
Q+ = {x = (x , xN ) ∈ RN , |x | < 1, 0 < xN < 1},
Q0 = {x = (x , xN ) ∈ RN , |x | < 1, xN = 0}.
Definition 1.2.4 (C k domains) Let Ω ⊂ RN be an open and bounded set. We say Ω
is of class C k , k ∈ R, k ≥ 1, if for every x ∈ ∂Ω there exists an open set Gx ⊂ RN ,
x ∈ Gx , and a one-to-one map θx satisfying
i) θx ∈ C k (Q; Gx ), θx−1 ∈ C k (Gx ; Q),
x := Ω ∩ Gx ,
ii) θx (Q+ ) = G+
iii) θx (Q0 ) = G0x := ∂Ω ∩ Gx .
This definition states that a C k domain is more than i) and iii), which one could think
at a first attempt. The feature ii) ensures that Ω is on one side of ∂Ω.
Some properties of functions in Sobolev spaces hold for Lipschitz domains which
have a weaker regularity than C k domains.
Definition 1.2.5 (Lipschitz domains) Let Ω ⊂ RN be an open and bounded set. We
say Ω is Lipschitz if Ω satisfies Definition 1.2.4 except i), which is replaced by
i)L ∃Lx ≥ 0, ∀x1 , x2 ∈ Q L−1 x |x − x | ≤ |θx (x ) − θx (x )| ≤ Lx |x − x | (in this case
1 2 1 2 1 2
we say “θx is a bi-Lipschitz function with Lipschitz bound Lx ”).

Remark 1.2.6 An equivalent formulation of Definition 1.2.4, resp. 1.2.5, is as follows:
Ω, an open and bounded set in RN , is of class C k , resp. Lipschitz, if for every x ∈ ∂Ω
there exists an open set Gx containing x such that ∂Ω ∩ Gx is the graph of a C k , resp.
Lipschitz, function of N − 1 variables (in a local system of coordinates), and Ω ∩ Gx is
on one side of ∂Ω ∩ Gx .
Also, more generally, an open set Ω (not necessarily bounded) is said to be minimally
smooth, see [48], if there exists Gi , i ∈ I ⊆ N, open sets, r0 > 0, m ∈ N and L > 0 such
that
6
Of course, instead of 1/2, 3/4 one could choose any a, b with 0 < a < b < 1.
10
Chapter 1 1.3. Review of some important results
i) if x ∈ ∂Ω then B(x, r0 ) ⊂ Gi , for a certain i ∈ I,

ii) Gi1 ∩ · · · ∩ Gim+1 = ∅, for every arbitrary choice of distinct i1 , . . . , im+1 ∈ I,
iii) Gi ∩ Ω = Gi ∩ Ωi , for every i ∈ I, where Ωi is a domain that up to a rotation is
above the graph of a Lipschitz function ψi : RN −1 → R with Lipschitz bound L.
Definition 1.2.7 (C k (∂Ω) spaces) Assume Ω is an open bounded set of class C k , k ≥

1. Let {(Gi , θi ), i = 1, . . . , m} be as in Definition 1.2.4, corresponding to a certain {xi ∈
∂Ω, i = 1, . . . , m} such that {Gi , i = 1, . . . , m} forms on open covering of ∂Ω. Let also
{ϕ0 , ϕ1 , . . . , ϕm } be a partition of unity subordinateto the covering {Gi , i = 1, . . . , m}
of ∂Ω so that for every u : ∂Ω → R we have u = m i=1 (uϕi ) in a neighborhood of ∂Ω.
k
The space C (∂Ω) is defined by

C k (∂Ω) = u : ∂Ω → R, (uϕi ) ◦ θi (·, 0) ∈ C0k (Q0 ), ∀i = 1, . . . , m . (1.2.3)
Remark 1.2.8 The space C k (∂Ω) is a Banach space when equipped with the norm
uC k (∂Ω) = (uϕi ) ◦ θi (·, 0)C0k (Q0 ) . (1.2.4)

i=1,...,m
It can be shown that two C k (∂Ω) norms, n(·), resp. ñ(·), corresponding to the choice
{(Gi , θi , ϕi ), i = 1, . . . , m}, resp. {(G̃i , θ̃i , ϕ̃i ), i = 1, . . . , m̃}, are equivalent. Indeed
m̃
(uϕi ) ◦ θi (·, 0) = ((uϕ̃j ) ◦ θ̃j (·, 0)) ◦ hi,j (·, 0))(ϕi ◦ θi (·, 0)),
j=1
where hi,j (·, 0) = θ̃j−1 ◦ θi (·, 0), which implies n(u) ≤ C ñ(u), with C > 0 depending only
on {(Gi , θi , ϕi ), i = 1, . . . , m} and {(G̃i , θ̃i , ϕ̃i ), i = 1, . . . , m̃}. We proceed similarly to
prove the reverse inequality.
1.3 Review of some important results

The readers are not expected to know the following results, but they must get familiar
with them before we consider Sobolev spaces.
1.3.1 Some results from Lp (Ω) spaces

We will review very briefly a number of results related to Lp (Ω) spaces. For a detailed
analysis of Lp (Ω) spaces we refer the interested reader to, for example [5, 21, 28, 29, 34].
11
1.3. Review of some important results Chapter 1
Definition 1.3.1 (Lp (Ω) spaces) Let p ∈ [1, ∞], Ω ⊂ RN measurable for the
N -dimensional Lebesgue measure dx and define
⎧
⎪ p
⎪ L (Ω) = u : Ω → R, u dx-measurable and |u(x)|p dx < ∞ , p ∈ [1, ∞),
⎪
⎨ Ω
∞ (1.3.1)
⎪
⎪ L (Ω) = {u : Ω → R, u dx-measurable and ∃M > 0, |u(x)| ≤ M, a.a. x ∈ Ω},
⎪
⎩ p
Lloc (Ω) = {u : Ω → R, u ∈ Lp (ω), ∀ω ⊂ Ω, ω open, ω Ω}.
The space Lp (Ω) is equipped with the norm

⎧ p1
⎪
⎨ p
|u(x)| dx , p ∈ [1, ∞),
uLp (Ω) = (1.3.2)
⎪
⎩
Ω
inf{M ∈ R, |u(x)| ≤ M, a.a.x ∈ Ω}, p = ∞.
Whenever there is no confusion, instead of Lp (RN ) we will write Lp .

Now we list some important results in Lp spaces.
Theorem 1.3.2 (Fisher-Riesz) The spaces Lp (Ω), p ∈ [1, ∞], are vector spaces.
Equipped with the norm · Lp (Ω) , they are Banach spaces.
Note that a Banach space is said to be reflexive if E = E.7
Theorem 1.3.3 (Reflexivity of Lp spaces) For every p ∈ (1, ∞), the space Lp (Ω) is
reflexive. The spaces L1 (Ω) and L∞ (Ω) are not reflexive. The dual space of L1 (Ω) is
L∞ (Ω) and the dual space of L∞ (Ω) is a space, called “space of Radon measures” and
denoted by M(Ω), strictly including L1 (Ω).
In the context of reflexivity, we have the following theorem.
Theorem 1.3.4 (Riesz representation theorem in Lp ) Let p ∈ [1, ∞) and p ∈
1 1
(1, ∞], the conjugate8 of p, satisfying + = 1. Then we have
p p
⎧
⎪ 1 1
⎨ ∀ ∈ (Lp (Ω)) , ∃u ∈ Lp (Ω) with + = 1 such that
p p (1.3.3)
⎪
⎩ , ϕ(Lp (Ω)) ×Lp (Ω) = p
u(x)ϕ(x)dx, ∀ϕ ∈ L (Ω), (L (Ω)) = uLp (Ω) .
p
Ω
The following result shows that Lp functions are not very irregular, in the sense that
they can be approximated by D functions.
7
In the sense that there exists an isomorphism from E to E .
8
In general, the conjugate p of p is defined for all p ∈ [1, ∞].
12
Theorem 1.3.5 (separability and density of D in Lp ) We have

i) For every p ∈ [1, ∞), the space Lp (Ω) is separable.9
ii) The space D(Ω) is dense in Lp (Ω), p ∈ [1, ∞). In particular, let (ρn ) be a mollifier
sequence, u ∈ Lp (RN ), p ∈ [1, ∞), and set

un (x) = (u ∗ ρn )(x) := ρn (x − y)u(y)dy = ρn (y)u(x − y)dx.
RN RN
Then (un ) is in Lp (RN ) ∩ C ∞ (RN ) and lim un = u in Lp (RN ).

n→∞
iii) L∞ (Ω) is not separable and C 0 (Ω) is not dense in L∞ (Ω).

Theorem 1.3.6 (Lebesgue dominated convergence theorem, Lebesgue DCT)
Let (un ) be a sequence of functions in L1 (Ω), u : Ω → R and v ∈ L1 (Ω) such that
i) lim un (x) = u(x) for a.a. x ∈ Ω,
n→∞
ii) |un (x)| ≤ v(x) for all n and a.a. x ∈ Ω.

Then u ∈ L1 (Ω) and lim un − uL1 (Ω) = 0.
n→∞
The following is a sort of reciprocal of the theorem above.

Theorem 1.3.7 Let (un ) be a sequence in Lp (Ω), lim un = u in Lp (Ω), p ∈ [1, ∞].
n→∞
Then there exists a subsequence (unk ) of (un ) and v ∈ Lp (Ω) such that
i) lim unk (x) = u(x) for a.a. x ∈ Ω,
k→∞
ii) |unk (x)| ≤ v(x) for all k and a.a. x ∈ Ω.

We have also this weak compactness result in Lp (Ω) spaces; see [5, 29].
Theorem 1.3.8 (Weak compactness in Lp ) Let Ω ⊂ RN be an open set, p ∈ (1, ∞)
and (un ) a bounded sequence in Lp , i.e. there exists C > 0 such that un Lp (Ω) ≤ C for
all n. Then there exist u ∈ Lp (Ω) and a subsequence (unk ) of Un ) converging “weakly to
uin Lp (Ω)”, i.e.

1 1
lim unk ϕdx = uϕ, ∀ϕ ∈ Lp (Ω), + = 1.
k→∞ Ω Ω p p
The following result gives an integral criteria for an L1 function to be zero.
N
Lemma 1.3.9 (Fundamental lemma) Let Ω ⊂ R be open and f ∈ L (Ω) such that
1
Ω
f (x)ϕ(x)dx = 0 for every ϕ ∈ D(Ω). Then f = 0.
9
A Banach space E is said to be separable if it contains a dense countable set F , i.e. F ⊂ E and for
every u ∈ E there exists (ϕn ) in F such that limn→∞ ϕn − uE = 0.
13
1.3.2 Some results from (Functional) Analysis

For the proof of the following results, see, for example, [7, 8, 28, 34].
Theorem 1.3.10 (Contraction mapping theorem) Let (M, d) be a complete metric
space10 and T : M → M a contraction, i.e. there exists κ ∈ [0, 1) such that
∀x, y ∈ M, d(T x, T y) ≤ κd(x, y).
Then T has a unique fixed point in M , i.e. the equation T z = z has a unique solution
z ∈ M . Furthermore, for every x ∈ M , lim d(T n x, z) = 0.
n→∞
Theorem 1.3.11 (Inverse mapping theorem) Let X, Y be Banach spaces, U ⊂ X

open, and f ∈ C k (U ; Y ), k ∈ N, k ≥ 1. Furthermore, let x0 ∈ U and assume that
f (x0 ) ∈ L(X; Y ), the derivative of f at x0 , is an isomorphism from X to Y , i.e.
f (x0 ) : h ∈ X → f (x0 ; h) ∈ Y is continuous and invertible. Then f is a local C k
diffeomorphism at x0 from X to Y , i.e. there exists U0 ⊂ X open with x0 ∈ U0 , V0 ⊂ Y
open with f (x0 ) ∈ V0 such that f −1 : V0 → U0 , the inverse of f : U0 → V0 , exists and
f −1 ∈ C k (V0 ; X).
As a corollary of the inverse mapping theorem we have the implicit mapping theorem,
which is important in applications.
Theorem 1.3.12 (Implicit mapping theorem) Let X, Y, Z be Banach spaces, U ⊂
X, V ⊂ Y open sets, and f ∈ C k (U × V ; Z), k ∈ N , k ≥ 1. Assume furthermore
that f (x0 , y0 ) = 0 for some (x0 , y0 ) ∈ U × V , and ∂y f (x0 , y0 ) ∈ L(Y ; Z), the partial
derivative of f with respect to y at (x0 , y0 ), defines an isomorphism from Y to Z. Then
there exists an open set U0 ⊂ U , x0 ∈ U0 , and a unique g ∈ C k (U0 ; Y ) such that
y0 = g(x0 ) and f (x, g(x)) = 0, ∀x ∈ U0 .
1.3.3 An application: Ordinary Differential Equations

In this section we review some results about the existence, uniqueness, stability, and
regularity of solutions to ODEs. Some of these results will be used in chapter 3 when
studying first-order PDEs. For the proof of the following results, see section 9.1.3, or for
a detailed analysis see, for example, [28, Chapter XIV, §3].
Let I ⊂ R be an interval, 0 ∈ I, X be a Banach space11 , U ⊂ X an open set, and
f ∈ C 0 (U ; X). Let x ∈ U and look for α : I × U → U such that α(·, x) ∈ C 1 (I; X) and
α (t, x) = f (α(t, x)), t ∈ I, (1.3.4a)
α(0, x) = x. (1.3.4b)
10
A metric space (M, d) is said to be complete if every Cauchy sequence in M converges to an element
of M .
11
All along this section the reader may take X = RN
14
Here denotes the derivative with respect to the variable t and x is the initial condi-
tion. The function f is sometimes called an “(autonomous) vector field”. The function
α is called an “integral curve with initial condition x”. Regarding the existence and
uniqueness of (1.3.4), we have the following result.
Theorem 1.3.13 Let (X, · X ) be a Banach space, U ⊂ X open, and f : U → X
Lipschitz with a Lipschitz constant L. Let x0 ∈ U , ρ ∈ (0, 1) such that B(x0 , 2ρ) ⊂ U
and K > 0 such that f C 0 (B(x0 ,2ρ)) ≤ K. Finally let 0 < r < min{1/L, ρ/K} and
set I r = [−r, r]. Then there exists α : I r × B(x0 , ρ) → U , such that for every x ∈
B(x0 , ρ) there exists a unique solution α(·, x) ∈ C 1 (I r ; X) of (1.3.4). Furthermore, if
f ∈ C k (U ; X) then α(·, x) ∈ C k+1 (I r ; X).
Proof. See section Theorem 9.1.4 in Annex.
The following theorem shows that the solution α(t, x) to (1.3.4) is stable with respect
to x, in the sense of (1.3.5). See Theorem 9.1.6 for the proof of this result.
Theorem 1.3.14 Assume the conditions of Theorem 1.3.13 hold. Let x ∈ B(x0 , ρ) and
α(·, x) ∈ C 1 (I r ; U ) as given by Theorem 1.3.13. Then x → α(·, x) is uniformly Lipschitz
from B(x0 , ρ) to C 0 (I r ; X), i.e. for every x, y ∈ B(x0 , ρ) we have
α(·, x) − α(·, y)C 0 (I r ;X) ≤ Lx − yX . (1.3.5)
Theorem 1.3.13 shows that if f is of class C k then α(·, x) is of class C k+1 with respect
to t. The following theorem shows that α(·, ·) is also of class C k with respect to (t, x).
The reader can find the proof, for example, in [28, Chapter XIV, §3].
Theorem 1.3.15 Assume f ∈ C k (U ; X), k ∈ N ∪ {∞}. For x ∈ U let J(x) ⊂ R be
the interval such that a solution α(·, x) ∈ C k+1 (J(x); U ) exists, and let D = {(t, x) ∈
R × U, t ∈ J(x), x ∈ U }. Then α ∈ C k (D; X).
Remark 1.3.16 Note that often we are interested in a more general problem than
(1.3.4), namely
β (τ, y) = g(τ, β(τ, y)), τ ∈ J, (1.3.6a)

β(s, y) = y, (1.3.6b)
where y ∈ V with V ⊂ Y open and Y a Banach space, J ⊂ R an open interval, s ∈ J,

β(·, y) ∈ C 1 (J; Y ), and g ∈ C 0 (J × V ; Y ) (in this case g is called a “non-autonomous
vector field”). In fact, (1.3.6) can be written in the form (1.3.4), which is the reason for
considering only the problem (1.3.4). Indeed, let β(·, y) ∈ C 1 (J; Y ). Set X = R × Y , with
I = {τ − s, τ ∈ J}, U = I × V , and consider
x ∈ U, x = (s, y),
15
α : I × U → U, α(·, x) ∈ C 1 (I; X),

α(t, x) := (s + t, β(s + t, y)) ∈ U,
f ∈ C (U ; X), f (x) := (1, g(x)) = (1, g(s, y)), so that
0
f (α(t, x)) := (1, g(α(t, x)) = (1, g(s + t, β(s + t, y))).
Then α(t, x) solves
α (t, x) = (1, β (s + t, y)) = (1, g(s + t, β(s + t, y))) = f (α(t, x)), t ∈ I, (1.3.7a)
α(0, x) = (s, y) = x, (1.3.7b)
if and only if β(·, y) solves (1.3.6).

It follows that all the results proved for the solution α(t, x) to (1.3.4) hold for the
solution β(τ, y) to (1.3.6), with the assumptions for g in (1.3.6a) such that f in (1.3.7a)
satisfies the assumptions of problem (1.3.4). For example, we note that
a) “g is Lipschitz with respect to β with Lipschitz constant L” is equivalent to “f is

Lipschitz with constant Lipschitz L”, and
b) “f C 0 (B(x0 ,2ρ) ≤ 1 + K” is equivalent to “gC 0 (B(s0 ,y0 ),2ρ) ≤ K” where x0 =

(s0 , y0 ), s0 ∈ J, y0 ∈ V , provided B(x0 , 2ρ) = B((s0 , y0 ), 2ρ) ⊂ J × V .
Then (1.3.6)has a unique C 1 (J r ; Y ) solution β = β(·, y), with J r = [s0 − r, s0 + r],

1 ρ
0 < r < min , for every (s, y) ∈ B((s0 , y0 ), ρ).
L 1+K
We are interested in the following questions for the problem (1.3.4):

⎧
⎨ i) Does it have a solution? [existence]
ii) Is the solution unique? [uniqueness] (1.3.8)
⎩
iii) Is the solution stable, say like in (1.3.5)? [stability].
Similarly for PDEs, we are interested in questions i), ii), and iii). Given that usually the
solutions to PDEs are “weak”, meaning that they do not have the classical regularity
C k (we will see what it means precisely in the following chapters), for PDEs we consider
also the following question:
iv) How regular is the solution? [regularity] (1.3.9)
We will address i) and ii) for all the PDEs we consider in this book. For ease of reading,
in section 9.6 in Annex we have addressed question iv) for second-order linear elliptic
PDEs.
16
Problems
Problem 1.1 Let X, Y, and Z be three Banach spaces, f ∈ C 0 (X; Y ), g ∈ C 0 (Y ; Z)
with Im(f ) ⊂ Dom(g), where, in general, for a function f : X → Y we set
dom(f ) = {x ∈ X, f (x) is well-def ined}, Im(f ) = {f (x), x ∈ dom(f )}.
i) Prove that g ◦ f ∈ C 0 (X; Z), where g ◦ f : X → Z, (g ◦ f )(x) = g(f (x)).
ii) Assume f ∈ C 1 (X; Y ), g ∈ C 1 (Y ; Z). Prove12 that g ◦ f ∈ C 1 (X; Z) and for

h ∈ X we have (g ◦ f ) (x; h) = g (f (x); f (x; h)).
Problem 1.2 Find the norm of the elements of L(RnN); see Example 1.1.5, Remark 1.1.4.
Problem 1.3 Let Ω ⊂ RN be open, k ∈ N, and λ ∈ (0, 1]. Prove that · Cbk (Ω) , resp.
· C k,λ (Ω) , as given by (1.1.20), resp. (1.1.21), defines a norm in Cbk (Ω), resp. Cbk,λ (Ω).
b
Problem 1.4 Find a C ∞ domain in RN , and another one which is C 1 but not C 2 .
Problem 1.5 i) Show that if Ω ⊂ RN is a Lipschitz domain then it is minimally smooth.

ii) Find an unbounded domain Ω ⊂ RN which is minimally smooth.
Problem 1.6 Prove the equivalence in Remark 1.2.6.
Problem 1.7 Show that if p ∈ (0, 1) then · Lp (Ω) is not a norm by showing that the
triangle inequality is not satisfied.
Problem 1.8 Show that Theorem 1.3.6 does not hold if only one of i), ii) is satisfied.
12
One may use this fact: f differentiable at x is equivalent to f (x + h) = f (x) + f (x; h) + h hX
for all x ∈ X, where h ∈ Y is such that limhX →0 h Y = 0.
17
2. Partial differential equations
Infinitesimal calculus is one of the most important mathematical developments of the

seventeenth century. The need for better understanding and prediction of the behavior
of real-world phenomena, such as the motion of planets or the trajectory of a particle,
for example, motivated mathematicians to consider (systems of) Ordinary Differential
Equations (ODEs). The practical interest on more complex phenomena, such as the
distribution of heat, of an electric or magnetic field, or the motion of fluids, made the
understanding of PDEs and their rigorous analysis a central point of mathematical
analysis and the birthplace of the development of new areas of mathematics, such as
Calculus of Variations and Functional Analysis.
Very fundamental results about linear PDEs were achieved around the end of the
nineteenth century and during the twentieth century, for example results about the exis-
tence and uniqueness of weak and strong solutions, and the regularity of weak solutions
to linear and certain nonlinear PDEs. Due to the explosion of the number of applica-
tions, the interest in PDEs continues to grow and is more oriented toward nonlinear
PDEs.
It would be useful to classify PDEs into classes such that the PDEs of the same
class share the same qualitative properties. For example, a classification such that the
proof of the existence, uniqueness, regularity, and even numerical analysis is the same
(or similar) for all the PDEs of the same class. If restricted to linear second-order PDEs,
a complete classification can be done by using Fourier transform. We will not enter into
details and suggest the interested reader [12]; see also Definition 2.0.3 for the case of a
second-order linear PDE with constant coefficients.
Here we will simply define linear and nonlinear PDEs. This definition serves as a
basis for identifying types of PDEs, but in general the PDEs of the same type under
this definition do not share the same qualitative properties.
Definition 2.0.1 Let Ω ⊂ RN be an open set.
i) A (N dimensional) PDE of order m ∈ N in Ω is an equation of an unknown
function u = u(x), of its derivatives up to order m and of x of the form
E(x, u, D1 u, . . . Dm u) = 0, x ∈ Ω, (2.0.1)
N2 Nm
where E : RN × R × RN × R × ··· × R → R and Dk u given by (1.1.14).
https://doi.org/10.1007/978-3-031-39524-6 2
Chapter 2
ii) Usually a PDE is equipped with boundary conditions of the form
B(x, u, D1 u, . . . , Dmb u) = 0, x ∈ ∂Ω, (2.0.2)

2 mb
where mb ≤ m − 1 and B : RN × R × RN × RN × · · · × RN → R.
iii) A function u is called a “classical solution to the problem (2.0.1), (2.0.2)” if u ∈

C m (Ω) ∩ C mb (Ω) and satisfies (2.0.1), (2.0.2) pointwise.
iv) If (2.0.1), resp. (2.0.2), is linear with respect to u and all its derivatives then
(2.0.1), resp. (2.0.2), is called “linear”. Otherwise it is called “nonlinear”.
Example 2.0.2 Let A = (ai,j ) ∈ RN ·N be a symmetric matrix, b = (bi ) ∈ RN , c ∈ R,

u = u(x) a smooth real function, x = (x1 , . . . , xN ) ∈ RN and consider the PDE

N
n
Lu := − ai,j ∂x2i xj u + bi ∂xi u + cu = f, with f given. (2.0.3)
i,j=1 i=1
Let A = V · Λ · t V be the spectral decomposition of A, with Λ = diag(λi ) the diagonal

matrix of eigenvalues of A, V = (vi,j ) the matrix having for columns the eigenvectors of
A with1 t V = V −1 . Let x = V · ξ, ξ = t V · x and define v(ξ) := u(x). Then

N
N
∂x i u = ∂ξk v∂xi ξk = vi,k ∂ξk v, (we used ∂xi ξk = vi,k )
k=1 k=1
N
∂x2i xj u = vi,k vj,l ∂ξk ξl v; so
k,l=1

N
N
N
N
N
ai,j ∂x2i xj u = ai,j vi,k vj,l ∂ξk ξl v = (vi,k · ai,j · vj,l ) ∂ξ2k ξl v
i,j=1 i,j=1 k,l=1 k,l=1 i,j=1

N
N
= (t V · A · V )k,l ∂ξ2k ξl v = Λk,l ∂ξ2k ξl v
k,l=1 k,l=1

N
= λk ∂ξ2k ξk v,
k=1
N
N
b · ∇u = μk ∂xk v, μk = vi,k bi ; hence
k=1 i=1
N
N
Lu = − λk ∂ξ2k ξk v + μk ∂ξk v + cv.
k=1 k=1
1
For a matrix M ∈ RN ×N , t M will denote its transposed matrix and M −1 will denote its inverse
matrix.
20
Chapter 2
Note that if η = diag(|λk |−1/2 ) · ξ and w(η) = v(ξ), we get

N
N
Lu = − αk ∂η2k ηk w + βk ∂ηk w + cw, (2.0.4)
k=1 k=1
with αk ∈ {±1, 0}, βk = μk |λk |−1/2 .

Based on the form (2.0.4), the following definition gives a detailed classification of
second-order linear PDEs with constant coefficients.
Definition 2.0.3 Consider the PDE (2.0.3) and let λi , i = 1, . . . , N , be the eigenvalues
of A. We say the PDE (2.0.3) is
“elliptic”, if all λi have the same sign,
“hyperbolic”, if all λi have the same sign, except one which has the opposite sign,
“parabolic”, if all λi have the same sign except one which is zero.
It turns out that when (2.0.3) is elliptic (resp. hyperbolic, parabolic) then in (2.0.4) all
αi have the same sign (resp. all αi have the same sign except one which has the opposite
sign, all αi have the same sign except one which is zero).
Example 2.0.4 Let u = u(x, t), x, t ∈ R and consider the second-order PDE
autt + 2butx + cuxx = 0, (2.0.5)
with a, b, c ∈ R, b2 − ac > 0. Assume2 a2 + c2 = 0, for example c = 0. Then (2.0.5) is

a b
of the form (2.0.3) with A = , whose product of eigenvalues is ac − b2 < 0, so
b c
(2.0.5) is hyperbolic. Furthermore, the solution is given by

ct − bx ct − bx
u(x, t) = A x + √ +B x− √ , (2.0.6)
b2 − ac b2 − ac
with A and B arbitrary C 2 functions. In particular, if a = 1, b = 0, c = −ω 2 , so the
original equation is utt − ω 2 uxx = 0, then every solution is of the form3
u(x, t) = A(x − ωt) + B(x + ωt). (2.0.7)
Indeed, proceeding as in Example 2.0.2 it is easy to find the eigenvectors of A, which

leads this change of variables and function v:
⎧ ⎧
⎨ x = ξ, ⎨ ξ = x,
√
v(ξ, τ ) = u(x, t), ct − bx
⎩ t = ξ b + τ b − ac ,
2 so
⎩ τ = √ .
c c b2 − ac
2
If a2 + c2 = 0 then the equation is reduced to uxt = 0, so its solution is u(x, t) = A(x) + B(t), with
A and B arbitrary C 2 functions.
3
Or, equivalently u(x, t) = A(ωt − x) + B(ωt + x).
21
2.1. Some prototypes of PDEs Chapter 2
It follows by straightforward computations that vτ τ − vξξ = 0. Next, introducing the new

variables η = ξ + τ , s = ξ − τ and the new function w(η, s) = v(ξ, τ ) implies that w
satisfies wsη = 0. Therefore w(η, s) = A(η) + B(s), with A, B arbitrary C 2 functions,
v(ξ, τ ) = A(ξ + τ ) + B(ξ − τ ) and therefore (2.0.6) holds.
2.1 Some prototypes of PDEs

In this section we will present some PDEs that have important applications. Although
some of them may look similar at first, they have very different properties. Each of them
represents the simplest element of a distinct class of PDEs, and typically, the methods
we develop for each of them are also used for the PDEs of the same class.
In all the following examples Ω ⊂ RN is open, 0 < T ≤ ∞, f is a given function, and
u is the unknown function. Also, here x = (x1 , . . . , xN ) ∈ RN , (x, t) = (x1 , . . . , xN , t),
and the unknown function u = u(x, t) or u = u(x), depending on whether or not the
problem involves the variable t.
Transport equation
The transport equation, which sometimes is referred to as the continuity equation, models
the transport of a certain quantity, for example, the transport of a cloud of particles. In
its simplest form it is

N
ut + b · ∇u := ut + b i ∂x i u = f in Ω × (0, T ), b = (b1 , . . . , bN ) ∈ RN . (2.1.1)
i=1
In chapter 3, we will consider general first-order PDEs, which include the equation 2.1.1.
Laplace equation
The Laplace equation models, for example, the distribution of heat in a stationary regime,
i.e. the heat does not depend on time. This is one of the most studied PDEs and the
most representative of second-order elliptic PDEs. It has the form
−Δu := − ∂x21 x1 u + · · · + ∂x2N xN u = f in Ω. (2.1.2)
We will study the classical and weak solutions to (2.1.2) in chapters 4 and 7.
Heat equation
The heat equation models the distribution of heat in a time-dependent regime, i.e. the
heat depends on time. In its simplest form it is
ut − c2 Δu = f in Ω × (0, T ), c > 0. (2.1.3)
22
Chapter 2 2.1. Some prototypes of PDEs
Equation (2.1.3) is the most representative of second-order PDEs of parabolic type. We

will study (2.1.3) in chapter 8.
Wave equation
The wave equation models the oscillations of a certain quantity, for example the motion
of an elastic membrane. In it simplest form it has the form
utt − c2 Δu = f in Ω × (0, T ), c > 0. (2.1.4)
Equation (2.1.4) is the most significant representative of second-order PDEs of hyper-
bolic type. We will study (2.1.4) in chapter 8.
Nonlinear PDEs
There are many nonlinear PDEs, often originating from applications. For example,
minimal surface equation in R3 , see [20],
(1 + u2x )uyy − 2ux uy uxy + (1 + u2y )uxx = 0 in Ω, u = g on ∂Ω, (2.1.5)
where u = u(x), x = (x1 , x2 ) ∈ Ω ⊂ R2 , describes the surface with the least area among
all graphs of functions in Ω equaling g on ∂Ω.
Another important example of nonlinear PDEs are incompressible Navier-Stokes
equations, see, for example, [52], given by
ρ(∂t u + (u · ∇)u) − μΔu + ∇p = f in Ω × (0, T ), (2.1.6a)
∇ · u = 0 in Ω × (0, T ), (2.1.6b)
where Ω ⊂ RN , N ≥ 2, 0 < T ≤ ∞, μ, ρ > 0 are given physical parameters, f =
(f1 , . . . , fN ) is a given function, and u = u(x, t), u = (u1 , . . . , uN ) and p = p(x, t) ∈ R
are unknown. The variable u represents the velocity and p represents the pressure. Here
the nonlinearity is due to the term (u · ∇)u.
The analysis of nonlinear PDEs is more difficult than linear PDEs. However, typically
the question of the existence of solutions relies strongly on the existence of solutions
to certain associated linear PDEs, on compactness results, and on tools of Functional
Analysis such as fixed point or topological degree theorems; see, for example, [8, 20, 42].
We will discuss very briefly some nonlinear PDEs mostly in the context of examples; see
Section 7.3.
Problems
Problem 2.1 Find a C 1 solution u = u(x, y) of
a) ux = 3x2 y + y, uy = x3 + x, u(0, 0) = 0,
b) ux = y cos x + 1, uy = sin x, u(0, 0) = 0.
23
2.1. Some prototypes of PDEs Chapter 2
Problem 2.2 Find a C 2 solution u = u(x, t), (x, t) ∈ R × (0, ∞), of

a) utt − uxx = sin t + x100 ,
b) utt − uxx = tx,
c) utt + uxx = cos t − x100 .
Show that in fact each of these problems has an infinite number of solutions.
Problem 2.3 Find all classical solutions to

a) uttx − uxxx = 0 in R × (0, ∞), ux (x, 0) = 0, utx (x, 0) = sin x, x ∈ R,
b) uttt − uxxt = 0 in R × (0, ∞), ut (x, 0) = 0, utt (x, 0) = cos x, x ∈ R.
Problem 2.4 Let u = u(x, t), (x, t) ∈ R × (0, ∞) and consider the PDE
utt + 4utx + 4uxx + (ux − 2ut ) = 0.
Use the transformation in Example 2.0.2 to show that this equation is of parabolic type,
and in the new coordinates (ξ, τ ) the PDE is written in the form vτ − 5vξξ = 0.
24
3. First-order PDEs: classical and
weak solutions
First-order PDEs describe many real-world phenomena. Conservation laws, which

state that a certain property of a physical system is conserved, are first-order PDEs.
Hamilton-Jacobi equations, which describe the motion of a mechanical system as well
as the motion of a cloud of particles, are also first-order PDEs.
In this chapter, we will briefly discuss first-order PDEs with the intention of high-
lighting mainly the method of characteristics and the concepts of classical and weak
solutions. For a deeper analysis of the subject, see the pioneering works [6, 9, 30, 31],
or [16] for a modern overview.
Let E : RN × R × RN → R, E = E(x, z, p), p = (p1 , . . . , pN ), z ∈ R, x = (x1 , . . . , xN ),

u : Ω ⊂ RN → R, with Ω open, and consider the problem
E(x, u(x), ∇u(x)) = 0, x ∈ Ω, (3.0.1a)

u(x) = g(x), x ∈ Γ ⊂ ∂Ω, (3.0.1b)
where Γ is relatively open1 in ∂Ω. Hereafter, we assume that E, g, and Ω are sufficiently
smooth, unless otherwise specified.
First, we will look for a classical solution to (3.0.1), and next we will demonstrate the
idea of weak solutions restricted to scalar conservation laws. In both cases, we will use
the method of characteristics, which consists of writing (3.0.1) in the form of a system
of ODEs, called the “characteristic equations”. These characteristic equations describe
the solution u along some curves, called “characteristic curves”.
Motivated by physical phenomena, such as the motion of a cloud of particles, the
idea for solving (3.0.1) is as follows. For a given x ∈ Ω, there is a particle moving along
a trajectory, or a characteristic curve, passing through x; see Fig. 3.0.1. This trajectory
eventually connects x with a certain point y 0 ∈ ∂Ω. If y 0 ∈ Γ then u(y 0 ) = g(y 0 ) is
known and therefore the characteristic equations associated with y 0 will provide the
value of u(x) along the characteristic curve.
1
So Γ = G ∩ ∂Ω, where G is a certain open set in RN .
https://doi.org/10.1007/978-3-031-39524-6 3
Chapter 3
From the description of the method,

we have to deal with the following two
problems.
i) Characterize the characteristic

curves and the system of charac-
teristic equations associated with y0
(3.0.1). x
Γ
ii) Knowing the solution to charac-
teristic equations, find a solution
to the problem (3.0.1). Figure 3.0.1: Characteristic curves
As a first example of a first-order PDE that addresses these two questions, let us
consider the transport equation (2.1.1)
ut + b · ∇u = f in RN × (0, ∞), (3.0.2a)

u = g on R × {0}.
N
(3.0.2b)
We assume that (3.0.2) has one classical solution u ∈ C 0 (RN ×[0, ∞))∩C 1 (RN ×(0, ∞)).
For a fixed (x, t) ∈ RN × (0, ∞), we set z(s) = u(x + sb, t + s) and then we obtain
z (s) = ut (x + sb, t + s) + b · ∇u(x + sb, t + s) = f (x + sb, t + s), s ∈ (−t, ∞),

z(−t) = u(x − tb, 0) = g(x − tb).
So, z satisfies a differential equation. It implies

0 0

u(x, t) − g(x − tb) = z(0) − z(−t) = z (s)ds = f (x + sb, t + s)ds
−t −t
t
= f (x + (s − t)b, s)ds.
0
Therefore, if g ∈ C 1 (RN ) and f ∈ C 0 (RN ×[0, ∞)) then the classical solution u is given by
t
u(x, t) = g(x − tb) + f (x + (s − t)b, s)ds. (3.0.3)
0
The following is also true: if g ∈ C 1 (RN ) and f, ∂xi f ∈ C 0 (RN × [0, ∞)) for all i, then u
given by (3.0.3) is a classical solution to (3.0.2).
Note that we have solved (3.0.2) by converting it to an ordinary differential equation
of the form z (s) = f (x + sb, t + s). Such an equation describes the solution u along
the curve {(x + sb, t + s), s ∈ (−t, ∞)}, which is called a “characteristic curve”. This
method is a particular case of the so-called “method of characteristics”; see Section 3.1
for more details.
26
Chapter 3 3.1. Method of characteristics
In general, g ∈
/ C 1 (RN ) and ∂xi f ∈
/ C 1 (RN × [0, ∞)), and so (3.0.3) does not provide
1
a C “classical” solution to (3.0.2). However, u given by (3.0.3) is the only reasonable
candidate for a solution to (3.0.2). To give a meaning to the solution u in this case, we
introduce the so-called “weak solution to (3.0.2)” satisfying
∞ ∞
g(x)ϕ(x, 0)dx+ (uϕt +u(b·∇x ϕ))dxdt+ f (x, t)ϕ(x, t)dxdt = 0, (3.0.4)
RN 0 RN 0 RN

for all ϕ ∈ D(RN × R), where b · ∇x ϕ = N i=1 bi ∂xi ϕ. Note that (3.0.4) is obtained by
multiplying (3.0.2a) by ϕ, integrating by parts, and then using (3.0.2b). It is well-defined
for all f ∈ L1loc (RN × (0, ∞)) and g ∈ L1loc (RN ).
3.1 Method of characteristics

In this section and the next one, we will present the method of characteristics and
show that under some assumptions, (3.0.1) is equivalent to a system of differential (char-
acteristic) equations, which provides a local classical solution. The results are classical
and we follow the pioneering works [6, 9, 19], and [16] for a modern presentation of the
results.
Let {y = y(t) = (y1 (t), . . . , yN (t)), t ∈ I ⊂ R} ⊂ Ω, with I an open interval, be a
given smooth curve, u ∈ C 1 (Ω) and define
z(t) := u(y(t)), p(t) = (p1 (t), . . . , pN (t)) := ∇u(y(t)), t ∈ I. (3.1.1)
We have the following result.
Proposition 3.1.1 Let E be a C 1 function, u ∈ C 2 (Ω) be a solution to (3.0.1a), and
z, p defined by (3.1.1). If y solves (3.1.2a) then z = z(t) and p = p(t) solve (3.1.2b),
(3.1.2c), where
y = ∇p E(y, z, p), (3.1.2a)
z = p · ∇p E(y, z, p), (3.1.2b)
p = −∇x E(y, z, p) − p∂z E(y, z, p). (3.1.2c)
Proof. We can differentiate z(t) in (3.1.1), which in combination with p in (3.1.1) and
(3.1.2a) gives

N
N

z (t) = ∂xj u(y(t))yj (t) = pj (t)∂pj E(y(t), z(t), p(t)). (3.1.3)
j=1 j=1
Now, let us differentiate p(t) in (3.1.1), which after using (3.1.2a) gives

N
N
pi (t) = ∂xi xj u(y(t))yj (t) = ∂xi xj u(y(t))∂pj E(y(t), u(y(t)), ∇u(y(t)))
j=1 j=1
27
3.1. Method of characteristics Chapter 3

N
= ∂pj E(y, u, ∇u)∂xi xj u, on {y(t), t ∈ I}. (3.1.4)
j=1
In order to eliminate second-order derivative terms in (3.1.4), we differentiate (3.0.1a)

w.r.t. xi , i = 1, . . . , N , and get

N
∂xi E(x, u, ∇u) + ∂z E(x, u, ∇u)∂xi u + ∂pj E(x, u, ∇u)∂xi xj u = 0, ∀x ∈ Ω. (3.1.5)
j=1
Then (3.1.4) combined with (3.1.5) evaluated at x = y(t) (so, u(y(t)) = z(t) and
∇u(y(t)) = p(t)) gives
pi (t) = −∂xi E(y(t), z(t), p(t)) − pi (t)∂z E(y(t), z(t), p(t)), t ∈ I, (3.1.6)
which completes the proof.
Equations (3.1.2) are called “characteristic equations (associated with (3.0.1a)”, and
the curve {y(t), t ∈ I} is called a “characteristic curve”. Proposition 3.1.1 gives a nat-
ural method for solving (3.0.1). Namely, it suggests that if (y, z, p) solves (3.1.2) then
u(x) := z(y(t)), with x = y(t), is a good candidate for the solution to (3.0.1a).
Before we move to the analysis of the method to characteristics, let us see how it
works through a number of typical examples. Consider first an example of first-order
linear PDEs, which have the general form
E(x, u, ∇u) = a(x) · ∇u(x) + b(x)u + c(x) = 0, x ∈ Ω, (3.1.7)
with a, b, and c being smooth functions. So, here E(x, z, p) = a(x) · p + b(x)z + c(x) and
then (3.1.2) is equivalent to
y = a(y), (3.1.8a)
z = p · a(y) = −(c(y) + b(y)z), (3.1.8b)
p = −∇x E(y, z, p) − pb(y), (3.1.8c)
because on the characteristic curve we have E(y, z, p) = 0, so p · a(y) = −(c(y) + b(y)z).
Here we do not need to solve for p as the characteristic equations are closed for y and z.
Example 3.1.2 Consider

2x1 x2 ∂x1 u + ∂x2 u = u in Ω = {(x1 , x2 ), x2 > 0},
(3.1.9)
u(x1 , 0) = g(x1 ) on Γ = {(x1 , x2 ), x2 = 0}.
This equation is of the form (3.1.7) with a = (2x1 x2 , 1), b = −1, c = 0, and so
E(x, z, p) = 2x1 x2 p1 + p2 − z. According to (3.1.8), y and z solve
y1 = 2y1 y2 , y2 = 1, z = z, with y(0) = (y10 , 0) ∈ Γ, z(0) = g(y10 ).
28
Chapter 3 3.1. Method of characteristics
The initial conditions for y and z are chosen such that the characteristic curves start
on Γ, so y(0) ∈ Γ, and that z(0) = u(y(0)). We solve this system of ODEs and get
2
y1 (t) = y10 et , y2 (t) = t, z(t) = g(y10 )et .
We note that y and z depend on t and y10 , so y = y(t, y10 ), z = z(t, y10 ).
Now we “construct” the solution u at x = (x1 , x2 ) ∈ Ω by assigning to it the value
2
of z(t, y10 ) with y(t, y10 ) = x. Solving the last equation gives y10 = x1 e−x2 , t = x2 . Hence
2
u(x1 , x2 ) = u(y(t)) = z(t) = g(y10 )et = g(x1 e−x2 )ex2 .
It is easy to verify that u is a classical solution to (3.1.9) if g is C 1 .
Next we consider an example of so-called first-order quasi-linear PDEs, which have

the form
E(x, u, ∇u) = a(x, u(x)) · ∇u(x) + b(x, u(x)) = 0, x ∈ Ω. (3.1.10)
So, here E(x, z, p) = a(x, z) · p + b(x, z). Therefore, (3.1.2) is equivalent to
y = a(y, z), (3.1.11a)

z = p · a(y, z) = −b(y, z), (3.1.11b)
p = −∇x E(y, z, p) − p∂z E(y, z, p), (3.1.11c)
because on the characteristic curve we have E(y, z, p) = 0, so p · a(y, z) = −b(y, z).

u∂x1 u + ∂x2 u = 1, in Ω = {(x1 , x2 ), x1 < x2 < 0},
(3.1.12)
u(x1 , x1 ) = g(x1 ) = 2 x1 on Γ = {(x1 , x1 ), x1 < 0}.
1
This PDE system is of the form (3.1.10) with a = (z, 1), b = −1, and so E(x, z, p) =
zp1 + p2 − 1. Applying (3.1.11) and noting that ∇x E = 0 and ∂z E = p1 , we obtain

y1 = z, p1 = −p21 ,
z = 1,
y2 = 1, p2 = −p1 p2 .
We note that the system is closed in (y, z), so we do not need to solve for p. The initial
conditions for y and z are

y1 (0) = y10 , 1
0 z(0) = y10 .
y2 (0) = y1 , 2
Then we get

y1 (t) = 12 t2 + 12 y10 t + y10 , 1
z(t) = t + y10 , y10 ∈ R.
y2 (t) = t + y10 , 2
29
3.1. Method of characteristics Chapter 3
To find u(x) we first find the characteristic curve y(t, y01 ) passing through x ∈ Ω, so
y(t, y 0 ) = x or equivalently
⎧
⎪ x2 − x 1
1 2 1 0 ⎪
⎨ t = 2 ,
2
0
t + 2 y1 t + y1 = x1 , 2 − x2
which implies
t + y10 = x2 , ⎪
⎪ 2x1 − x22
⎩ 1
y 0
= .
2 − x2
This implies that
1 1 4x2 − x22 − 2x1
u(x1 , x2 ) = z(t, y10 ) = t + y10 = .
2 2 2 − x2
It is easy to check that u is a classical solution to (3.1.12).
Now we consider an example of a fully nonlinear first-order PDE, which are all first-
order PDEs that are neither linear nor quasi-linear. In this case we have to integrate
the full system of differential equations (3.1.2), which in general is not possible.

(∂x1 u)2 + (∂x2 u)2 = 2 in Ω = {(x1 , x2 ), x1 > 0},
(3.1.13)
u(0, x2 ) = x2 on Γ = {(0, x2 ), x2 ∈ R}.
Here E(x, z, p) = p21 + p22 − 2. Therefore (3.1.2a)–(3.1.2c) becomes

⎧
⎪
⎨ y1 = 2p1 ,
z = 4, p1 = 0,
y2 = 2p2 ,
⎪
⎩ y(0) = (0, y 0 ), y 0 ∈ R,
0
z(0) = y2 , p2 = 0.
2 2
The general solution (y, z, p) can be written as

y1 = 2p01 t, p1 = p01 ,
0 0
z = 4t + y20 ,
y2 = 2p2 t + y2 , p2 = p02 .
Here y20 and p0 = (p01 , p02 ) are arbitrary constants. Using (3.1.13), which we assume
holds in Ω, we can eliminate two of these three constants, for example p01 and p02 , so that
(y, z, p) depends only on y20 . Indeed, (∂x1 u)2 + (∂x2 u)2 = 2 in Ω and ∂x2 (u − x2 ) = 0 on
Γ give
0 2 0
(p1 ) + (p02 )2 = 2, p1 = ±1,
so
0
p2 = 1, p02 = 1.
Hence, there are two solutions (y, z, p) given by

y1 = ±2t, 0 p1 = ±1,
0
z = 4t + y2 ,
y2 = 2t + y2 , p2 = 1.
30
Chapter 3 3.2. Classical local solutions to first-order PDEs
1
Now, let x = (x1 , x2 ) ∈ Ω. We can find (t, y10 ) such that y(t, y10 ) = x. It follows t = ± x1 ,
2
y20 = x2 ∓ x1 . Therefore, two classical solutions to 3.1.13 are given by
1
u(x1 , x2 ) = u(y(t)) = z(t) = 4 ± x1 + (x2 ∓ x1 ) = x2 ± x1 .
2
3.2 Classical local solutions to first-order PDEs

In this section, we will construct local classical solutions to (3.0.1) by following the
method of characteristics as described in Section 3.1. Note that in general the boundary
Γ in (3.0.1b) is not flat, where, without loss of generality, we say that Γ is flat if
N −1
Γ ⊂ {y 0 = (y10 , . . . , yN
0
−1 , 0), (y1 , . . . , yN −1 ) ∈ R
0 0
}. (3.2.1)
The following lemma shows that modulo a diffeomorphism, (3.0.1) is equivalent to a
similar problem with a flat boundary. This motivates the study of (3.0.1) with Γ flat,
because due to the simplicity of the boundary the analysis is easier. Next we will return
to (3.0.1) with a non-flat boundary by using the following lemma again.
Lemma 3.2.1 Consider the problem (3.0.1). Let x0 ∈ Γ and assume that Γ is of class
C k+1 at x0 , i.e. there exist G ⊂ RN open with x0 ∈ G, and θ ∈ C k+1 (Q, G) invertible,
such that θ(Q+ ) = Ω ∩ G =: G+ , θ(Q0 ) = Γ ∩ G =: Γ0 ; see Definition 1.2.4.
i) Assume that E is of class C k , g ∈ C k (Γ0 ) and u ∈ C k (G+ ) solves (3.0.1) with
Ω = G+ and Γ = Γ0 . If ũ(x̃) = u ◦ θ−1 (x̃), x̃ ∈ Q+ , then ũ ∈ C k (Q+ ) and it solves
Ẽ(x̃, ṽ, ∇ṽ) = 0 in Q+ , (3.2.2a)

ũ(x̃) = g̃(x̃) on Q0 , (3.2.2b)
with Ẽ of class C k and g̃ ∈ C k (Q0 ).
ii) Conversely, assume Ẽ is of class C k , g̃ ∈ C k (Q0 ), and ũ ∈ C k (Q+ ) solves (3.2.2).
If u(x) = ũ ◦ θ−1 (x), x ∈ G+ , then u ∈ C k (G+ ) and it solves (3.0.1) with Ω = G+ ,
Γ = Γ0 , E of class C k , and g ∈ C k (Γ0 ).
Proof. By the composition rule, it is clear that ũ ∈ C k (Q+ ) and
∇u(x) = t [∇θ(x̃)]−1 · ∇u(x̃), x̃ ∈ Q+ , x = θ(x̃) ∈ G+ .
Therefore, (3.0.1) implies
0 = E(x, u(x), ∇u(x)) = E(θ−1 (x̃), ũ(x̃), t [∇θ(x̃)]−1 · ∇ũ(x̃))
=: Ẽ(x̃, ũ(x̃), ∇ũ(x̃)), x̃ ∈ Q+ , (3.2.3)
0 = u(x) − g(x) = ũ(x̃) − g(θ(x̃))
=: ũ(x) − g̃(x̃), x̃ ∈ Q0 , (3.2.4)
which proves i) with Ẽ(x̃, z̃, p̃) = E(θ−1 (x̃), z̃, t [∇θ(x̃)]−1 · p̃) and g̃(x̃) = g(θ(x̃)). For ii)
we proceed similarly.
31
3.2. Classical local solutions to first-order PDEs Chapter 3
3.2.1 Classical local solutions: flat boundary

Under the assumptions of Lemma 3.2.1, the problem (3.0.1) is equivalent to the
problem (3.2.2), which has a flat boundary. In this section, we assume that Γ in (3.0.1b)
is flat and prove the existence of a classical solution to (3.0.1). In the following section
we will show the existence of a classical solution to (3.0.1) in the case Γ is not flat, by
using again Lemma 3.2.1.
3.2.1.1 Boundary and initial conditions

The equations (3.1.2) describe the evolution of u from (3.0.1) along the characteristic
curves {y(t), t ∈ I}. We have to equip (3.1.2a)–(3.1.2c) with initial conditions; let’s say
(y(0), z(0), p(0)) = (y 0 , z0 , p0 ). (3.2.5)
It is worth emphasizing that z0 and p0 depend on y 0 , so z0 = z0 (y 0 ), p0 = p0 (y 0 ). This

implies that y, z, and p depend on (t, y 0 ), hence y = y(t, y 0 ), z = z(t, y 0 ), p = p(t, y 0 )
(we will drop the variable y 0 whenever there is no confusion).
Clearly, the initial conditions for y and z follow naturally from the boundary condi-
tion u = g on Γ, i.e.
y(0, y 0 ) = y 0 , z(0, y 0 ) = z0 (y 0 ) := g(y 0 ), y 0 ∈ Γ. (3.2.6)
Differentiating the equation for z in (3.2.6) with respect to yi0 , i = 1, . . . , N − 1, gives

∂i z0 = ∂i g on Γ, which leads to these conditions for p0i :
p0i (y 0 ) = ∂i g(y 0 ), i = 1, . . . , N − 1, y 0 ∈ Γ. (3.2.7)
Condition (3.2.7) is called a “compatibility condition”. Conditions (3.2.6) and (3.2.7)

uniquely define all the initial conditions except p0N . For p0N we assume that (3.0.1a)
holds in Ω ∪ Γ, which implies the following equation counting for the remaining p0N :
E(y 0 , z0 , p0 ) = 0, y 0 ∈ Γ. (3.2.8)
A triple (y 0 , z0 , p0 ) satisfying (3.2.6)–(3.2.7) is called the “admissible initial condition”.
Remember that we intend to find the solution u to (3.0.1) by starting at y 0 ∈ Γ, and

then solving (y, z, p) along the characteristic curves, which will provide u(y(t, y 0 )) =
z(y(t, y 0 )) along this curve. For this, the curves y(t, y 0 ), which start on Γ, must evolve
inside Ω. As the vector y (0, y 0 ) is tangent to the characteristic curve at y 0 , this condition

is satisfied, at least locally near Γ, if yN (0, y 0 ) > 0, which from (3.1.2a) is equivalent to
∂pN E(y 0 , z0 , p0 ) > 0, ∀y 0 ∈ Γ. (3.2.9)
Note that this condition is equivalent to ∂pN E(y 0 , z0 , p0 ) = 0 because we can consider
−E instead of E. The condition (3.2.9) is called a “transversality condition”. An initial
32
condition (y 0 , z0 , p0 ) satisfying (3.2.6)–(3.2.9) (so admissible and transversal) is called a

“non-characteristic”.2
The initial conditions z0 = z0 (y 0 ) and (p01 , . . . , p0N −1 ) = (∂1 g(y 0 ), . . . , ∂N −1 g(y 0 )) are
explicit and are given by (3.2.6) and (3.2.7), respectively. The remaining initial condition
p0N is given implicitly by (3.2.8). The following result shows that if (y 0 , z0 (y 0 ), p0 (y 0 )) is
non-characteristic at a certain x0 , then it can be extended uniquely as a non-characteristic
initial condition near x0 .
Lemma 3.2.2 Assume E and g are C k , k ≥ 2. Furthermore, assume that there exists
x0 ∈ Γ and p0 (x0 ) ∈ RN such that (3.2.6)–(3.2.9) hold at (x0 , z0 (x0 ), p0 (x0 )). Then
there exist ρ0 > 0 and a unique p0 ∈ C k−1 (Γ0 ; RN ) with Γ0 = B(x0 , ρ0 ) ∩ Γ, such that
(3.2.6)–(3.2.9) hold at (y 0 , z0 (y 0 ), p0 (y 0 )) for all y 0 ∈ Γ0 , i.e. (y 0 , z0 (y 0 ), p0 (y 0 )) is a
non-characteristic initial condition for all y 0 ∈ Γ0 .
Proof. Note that from (3.2.6) and (3.2.7) we have z0 (y 0 ) = g(y 0 ), and p0i (y 0 ) := ∂i g(y 0 ),
i = 1, . . . , N − 1, are well-defined and of class C k−1 in a neighborhood of x0 . Now we
will define p0N by using the (implicit function) Theorem 1.3.12. Consider
G : (RN −1 × {0}) × R →
R,
0 0
(y , pN ) → E(y 0 , z0 (y 0 ), (p01 (y 0 ), . . . , p0N −1 (y 0 ), p0N )).
Note G is C k−1 near x0 , G(x0 , p0N (x0 )) = 0, and
∂pN G(x0 , p0N (x0 )) = ∂pN E(x0 , z0 (x0 ), p0 (x0 )) > 0.
From Theorem 1.3.12 there exists a unique p0N ∈ C k−1 (Γ0 ), with Γ0 = B(x0 , ρ0 ) ∩ Γ
for a certain ρ0 > 0, and G(y 0 , p0N (y 0 )) = 0 on Γ0 . As ∂pN E(x0 , z0 (x0 ), p0 (x0 )) > 0, by
continuity it follows that (3.2.9) holds in Γ0 provided ρ0 is small, which proves the lemma.
3.2.1.2 Classical local solutions

We have already seen that in order to solve (3.0.1) by using the method of charac-
teristics, we are led to
⎧ ⎧
⎨ y = ∇p E(y, z, p), ⎨ y(0, y 0 ) = y 0 ,

z = p · ∇p E(y, z, p), z(0, y 0 ) = z0 (y 0 ), (3.2.10)
⎩ ⎩
p = −∇x E(y, z, p) − p∂z E(y, z, p), p(0, y 0 ) = p0 (y 0 ),
with y = y(t, y 0 ), z = z(t, y 0 ), p = p(t, y 0 ), and y 0 ∈ Γ, z0 = z0 (y 0 ), p0 = p0 (y 0 ) satisfy-

ing (3.2.6)–(3.2.9).
We have the following local existence theorem.

2
The name refers to the fact that the boundary Γ is not tangent to the characteristic curves.
33
Theorem 3.2.3 Assume that E and g are C k functions, k ≥ 2, and (x0 , z0 (x0 ), p0 (x0 ))
is non-characteristic for a certain x0 ∈ Γ. Then there exist ρ0 > 0 and r0 > 0 such that
the following hold.
(i) There exists y 0 → p0 (y 0 ) ∈ C k−1 (Γ0 ; RN ) such that (y 0 , z0 (y 0 ), p0 (y 0 )) is non-
characteristic for all y 0 ∈ Γ0 := Γ ∩ B(x0 , ρ0 ).
(ii) For every y 0 ∈ Γ0 , the initial value problem (3.2.10) has a unique C k ([−r0 , r0 ]; RN ×
R × RN ) solution. Furthermore, if T is defined by
T : [0, r0 ] × Γ0 =: U0 → Ω0 := T (U0 ) ⊂ Ω,
(t, y 0 ) → T (t, y 0 ) := y(t, y 0 ),
then T is C k−1 , invertible, and its inverse is C k−1 .
(iii) The function u : Ω0 → R, u(x) = z ◦ T −1 (x) = z(y(t, y 0 )), is a C k (Ω0 ) solution to
E(x, u, ∇u) = 0 in Ω0 , (3.2.11a)
u = g on Γ0 . (3.2.11b)
(iv) If ∂pN E(x0 , z0 (x0 ), (p01 (x0 ), . . . , p0N −1 (x0 ), q)) does not change sign (remains posi-
tive or negative) for all q ∈ R, then (3.2.11) has a unique C k solution.
Proof. See Theorem 9.2.1 in Annex.
3.2.2 Classical local solutions: non-flat boundary

When the boundary Γ in (3.0.1b) is not flat, we proceed as follows. Given x0 ∈ Γ, by
using Lemma 3.2.1 we transform the problem (3.0.1) near x0 to a similar problem with
flat boundary, for which Theorem 3.2.3 provides a classical solution. By using Lemma
3.2.1 again we obtain a classical local solution to the problem (3.0.1). More specifically
we have this result.
Theorem 3.2.4 Let x0 ∈ Γ and assume Γ is C k+1 at x0 , k ≥ 2, E and g are C k , and
the initial conditions z0 (x0 ), p0 (x0 ) satisfy

z0 (x0 ) = g(x0 ), E(x0 , z0 (x0 ), p0 (x0 )) = 0,
(3.2.12)
p0τ (x0 ) = ∇τ g(x0 ), ∇p E(x0 , z0 (x0 ), p0 (x0 )) · ν(x0 ) > 0.
Here ν(x0 ) is the unit normal vector to Γ at x0 oriented inward Ω, p0τ is the tangential
component of p0 , and ∇τ g is the tangential gradient of g defined by
p0τ = p0 − (p0 · ν)ν, ∇τ g = ∇g − (∇g · ν)ν, (3.2.13)
with g being any C k extension of g. Then there exists a solution u ∈ C k (Ω ∩ B(x0 , ρ0 ))
to (3.0.1), for a certain ρ0 > 0 small. Furthermore, if
∇p E(x0 , z0 (x0 ), p0τ (x0 ) + qν(x0 )) · ν(x0 ) does not change sign, q ∈ R, (3.2.14)
then the solution is unique at least locally near x0 .
34
Proof. The proof is based on Lemma 3.2.1 and Theorem 3.2.3. As Γ is C k+1 at x0 , there
exists θ ∈ C k+1 (Q; G) as in Lemma 3.2.1. We set x̃0 = θ−1 (x0 ), and consider the problem
(3.2.2) with Ẽ, g̃, ũ as in Lemma 3.2.1, which establishes a correspondence between the
local solutions to (3.0.1) with non-flat boundary and (3.2.2) with flat boundary.
Claim: Ẽ, g̃ and (x̃0 , z̃0 (x̃0 ), p̃0 (x̃0 )), where z̃0 (x̃0 ) = z0 (x0 ), p̃0 (x̃0 ) = t [∇θ(x̃0 )] · p0 (x0 ),
satisfy the conditions of Theorem 3.2.3 at x̃0 . Assuming the claim holds implies that (i)–
(iii) of Theorem 3.2.3 hold, and the existence of a C k solution u follows from Lemma 3.2.1.
Now we prove the claim. From (3.2.3) and (3.2.4), clearly Ẽ and g̃ are C k . It remains
to show that (x̃0 , z̃0 (x̃0 ), p̃0 (x̃0 )) is non-characteristic. First we note that the vectors
[∇θ(x̃0 )]·,j , j = 1, . . . , N − 1, are tangent to Γ at x0 and the vector [∇θ(x̃0 )]·,N is
oriented inside Ω. Combined with [∇θ(x̃0 )]−1 · [∇θ(x̃0 )] = Id,3 it implies that
(t [∇θ(x̃0 )] · ν(x0 ))i = 0, i = 1, . . . , N − 1, (3.2.15)

−1
[∇θ(x̃0 )]N,· = λν(x0 ), λ > 0. (3.2.16)
As every vector can be decomposed in its tangential and normal components, by using
(3.2.15), for i = 1, . . . , N − 1 we get
p̃0i (x̃0 ) = (t [∇θ(x̃0 )] · p0 (x0 ))i = (t [∇θ(x̃0 )] · (p0τ (x0 ) + (p0 (x0 ) · ν(x0 ))ν(x0 )))i
= (t [∇θ(x̃0 )] · (p0τ (x0 ))i = (t [∇θ(x̃0 )] · ∇τ g(x0 ))i
= (t [∇θ(x̃0 )] · ∇g(x0 ))i − (t [∇θ(x̃0 )] · ν(x0 ))i (∇g(x0 ) · ν(x0 ))
= (t [∇θ(x̃0 )] · ∇g(x0 ))i = (∇(g ◦ θ)(x̃0 ))i
= ∂i g̃(x̃0 ). (3.2.17)
Also, from E(x0 , z0 (x0 ), p0 (x0 )) = 0 and (3.2.3) we get
Ẽ(x̃0 , z̃0 (x̃0 ), p̃0 (x̃0 )) = E(θ−1 (x̃0 ), z̃0 (x̃0 ), t [∇θ(x̃0 )]−1 · p̃0 (x̃0 ))
= E(x0 , z0 (x0 ), p0 (x0 )) = 0. (3.2.18)
Furthermore, using (3.2.16) and the first line of (3.2.18) gives
∂p̃N Ẽ(x̃0 , z̃0 (x̃0 ), p̃0 (x̃0 )) = [∇θ(x̃0 )]−1

N,· · ∇p E(x , z0 (x ), p (x ))
0 0 0 0
(3.2.19)
= ([∇θ(x̃ 0
)]−1
N,· · ν(x ))(∇p E(x , z0 (x ), p (x )) · ν(x0 )) > 0.
0 0 0 0 0
From (3.2.17)–(3.2.19) it follows that (x̃0 , z̃0 (x̃0 ), p̃(x̃0 )) is non-characteristic.

Claim: If furthermore (3.2.14) holds, the solution is unique. In view of Lemma 3.2.1, it is
enough to show that the problem (3.2.2) has a unique solution. We show that Condition
(iv) of Theorem 3.2.3 holds, which proves the claim. Indeed, first note that by using
(3.2.16), for p̃N ∈ R we have
t
[∇θ(x̃0 )]−1 · (p̃01 (x̃0 ), . . . , p̃0N −1 (x̃0 ), p̃N )
3
Id : RN → RN is the identity map.
35
[∇θ(x̃0 )]−1 · (p̃01 (x̃0 ), . . . , p̃0N −1 (x̃0 ), p̃0N (x̃0 )) + t [∇θ(x̃0 )]−1 · (0, . . . , 0, p̃N − p̃0N (x̃0 ))
t
= p0 (x0 ) + (p̃N − p̃0N (x̃0 ))[∇θ(x̃0 )]−1 N,·

= p0 (x0 ) + λ(p̃N − p̃0N (x̃0 ))ν(x0 )
= p0τ (x0 ) + qν(x0 ),
for a certain q ∈ RN , which by using (3.2.16) again implies that
∂p̃N Ẽ(x̃0 , z̃0 (x̃0 ), (p̃01 (x̃0 ), . . . , p̃0N −1 (x̃0 ), p̃N ))

= [∇θ(x̃0 )]−1
N,· · ∇p E(x , z0 (x ), [∇θ(x̃ )]
0 0 t 0 −1
· (p̃01 (x̃0 ), . . . , p̃0N −1 (x̃0 ), p̃N ))
= λ(∇p E(x0 , z0 (x0 ), p0τ (x0 ) + qν(x0 )) · ν(x0 ))
does not change sign. Theorem 3.2.3 implies the uniqueness of the solution ũ, and Lemma
3.2.1 proves the uniqueness of the solution u.
The following example shows that, in general, if the condition (3.2.14) fails then the
problem (3.0.1) does not have a unique solution.
Example 3.2.5 Consider problem (3.1.13), where E(x, z, p) = p21 +p22 −2, g(x1 , x2 ) = x2 .
Hence for x0 = (0, x02 ) ∈ Γ we have ν(x0 ) = (1, 0), z0 (x0 ) = x02 , p0τ (x0 ) = ∇g(x0 ) −
(∇g(x0 ) · ν(x0 )ν(x0 ) = (0, 1), and
∇p E(x0 , z0 (x0 ), ∂τ p0 (x0 ) + qν(x0 )) · ν(x0 ) = ∂p1 E(x0 , x02 , (q, 1)) = 2q.
Hence, condition (3.2.14) is not satisfied. On the other hand, we have seen in Example
3.1.4 that problem (3.1.13) has at least two solutions.
The following example shows that if the transversality condition in (3.2.12) (or
(3.2.9)) is not satisfied then, depending on g, (3.0.1) may not have a solution, or may
have infinitely many solutions.
∂1 u + ∂2 u = 1 in Ω := {(x1 , x2 ), x1 , ∈ R, x2 > x1 }, (3.2.20a)

u(x1 , x1 ) = g(x1 ) on Γ := ∂Ω. (3.2.20b)
Here E = p1 + p2 − 1 and g(x1 ) is a smooth function. Note that for whatever choice of
p0 , the transversality condition is not satisfied, because
√
ν = (−1, 1)/ 2 and so ∇p E(y 0 , z0 , p0 ) · ν = 0.
So Theorem 3.2.3 does not apply. The system (3.2.10) associated with (3.2.20) is

y1 (t, y10 ) = 1, y2 (t, y10 ) = 1, z (t, y10 ) = 1,
y1 (0, y10 ) = y10 , y2 (0, y10 ) = y10 , z(0, y10 ) = g(y10 ),
36
and their solution is
y1 (t, y10 ) = t + y10 , y2 (t, y10 ) = t + y10 , z(t, y10 ) = t + g(y10 ).
The characteristic curves are straight lines parallel to Γ, so they do not intersect Γ unless
y10 = 0, in which case the characteristic curve coı̈ncides with Γ. Therefore, we cannot
find a solution to (3.2.20) by starting on Γ and following the characteristic curves.
i) Case of no solution. Note that if a classical solution u exists, if we set z(s) =
u(x1 + s, x2 + s) then z (s) = 1, and therefore by integrating z (s) in (−x1 , 0) we find that
u must satisfy u(x1 , x2 ) = u(0, x2 −x1 )+x1 . For x2 = x1 we get u(x1 , x1 ) = u(0, 0)+x1 =
g(0) + x1 . Hence, if g(x1 ) = g(0) + x1 for any x1 , then (3.2.20) has no solution.
ii) Case of infinitely many solutions. We assume that g(x1 ) = C+x1 for all x1 and
a certain C ∈ R. We can find infinitely many solutions to (3.2.20) in the following way.
We consider PDE (3.2.20a) but with a different boundary condition instead of (3.2.20b).
Namely, we choose a boundary Σ ⊂ R2 intersecting Γ at a certain x0 , such that the
transversality condition on Σ is satisfied. Next we look for a solution u to (3.2.20a)
equipped with a boundary condition u = h on Σ, where h is such that h = g on Σ ∩ Γ,
i.e. h(x0 ) = g(x0 ).
For example, we can take Σ = {(y10 , 0), y10 ∈ R} and h smooth such that h(0) = C,
because here Σ ∩ Γ = (0, 0). So we consider
∂1 u + ∂2 u = 1 in Ω, u(x1 , 0) = h(x1 ), x1 ∈ R. (3.2.21)
Problem (3.2.21) satisfies the conditions of Theorem 3.2.3. The solution to characteristic
equations of (3.2.21) is given by
y1 (t, y10 ) = t + y10 , y2 (t, y10 ) = t, z(t, y10 ) = t + h(y10 ).
For arbitrary x = (x1 , x2 ) ∈ Ω, we look for (t, y10 ) such that x = y(t, y 0 ). It follows
t = x2 , y10 = x1 − x2 , and then u(x1 , x2 ) = x2 + h(x1 − x2 ) is a classical solution to
(3.2.21) and to (3.2.20), provided h is C 1 and h(0) = C. So for g(x1 ) = C + x1 , (3.2.20)
has infinitely many solutions as we can take for example h(x1 ) = C + xn1 , n ∈ N.
Theorems 3.2.3 and 3.2.4 guarantee only local classical solutions, which are obtained
in the situation where the characteristic curves do not intersect. The following example
shows that depending on the behavior of characteristic curves, we may have a local or
global classical solution, or no classical solution at all.
1 2
∂2 u + u∂1 u = ∂2 u + ∂1 u = 0 in Ω = {(x1 , x2 ), x2 > 0}, (3.2.22a)
2
u(x1 , 0) = g(x1 ) on Γ = ∂Ω, (3.2.22b)
37
referred to as “Burgers’ equation”, which describes the dynamics of rarefied gases (vis-
cous effects are neglected). The system (3.2.10) associated with (3.2.22) has the form

y1 (t, y10 ) = z, y2 (t, y10 ) = 1, z (t, y10 ) = 0,
y1 (0, y10 ) = y10 , y2 (0, y10 ) = 0, z(0, y10 ) = g(y10 ),
which implies
y1 (t, y10 ) = tg(y10 ) + y10 , y2 (t, y10 ) = t, z(t, y10 ) = g(y10 ).
So the characteristic curves {y(t, y10 ) := (y1 (t, y10 ), y2 (t, y10 )), t ∈ R} are straight lines
starting at (y10 , 0) with slope g(y10 ) given equivalently by
y1 = y2 g(y10 ) + y10 , y10 ∈ R. (3.2.23)
Given x = (x1 , x2 ), x2 > 0, we look for (t, y10 ) such that y(t, y10 ) = x, which gives
t = x2 , y10 = x1 − x2 g(y10 ). (3.2.24)
As z is constant along characteristic curves we get u(x) = g(y10 ), or equivalently
u(x1 , x2 ) = g(x1 − x2 u(x1 , x2 )). (3.2.25)
The solution is implicit and its qualitative behavior depends on g.
x2 x2
u(x1 , 0) u(x1 , 0)
x1 x1
x2 x2
no char.
curves
region
x1 x1
Figure 3.2.1: Rarefaction waves. Top: an increasing continuous (left) and discontinuous
(right) initial condition. Bottom: the corresponding characteristic curves—they fill all
the domain {x2 > 0} in the case of continuous initial condition (left), while there is a
region where the characteristic curves do not cross (right).
38
Rarefaction waves. Assume g is an increasing function (see Fig. 3.2.1, top). In this
case the characteristic curves do not intersect. If g is continuous, see Fig. 3.2.1, top left,
(3.2.22) has a solution in the entire domain, because the equation y10 = x1 − x2 g(y10 ) has
a unique solution y10 for every (x1 , x2 ). If g is discontinuous, see Fig. 3.2.1, top right,
there is a region where there are no characteristic curves. In this region the method
of characteristics does not provide a solution. Such a solution is called a “rarefaction
wave”. Clearly, if g is only continuous then u is only continuous. Furthermore, in the
case g is discontinuous the method of characteristics does not provide a solution u defined
globally; see Fig. 3.2.1, bottom, right.
x2 x2
u(x1 , 0) u(x1 , 0)
x1 c x1
x2 x2
0
y1,l 0
y1,r x1 0
y1,l 0
y1,r x1
Figure 3.2.2: Shock waves Top: a decreasing continuous (left) and discontinuous (right)
initial condition. Bottom: the corresponding characteristic curves, which in both cases
intersect.
Shock waves. Assume g is a decreasing function (see Fig. 3.2.2, top). The only dif-
ference with the case when g is increasing is that for every x2 > 0, the equation
y10 = x1 − x2 g(y10 ) in (3.2.24), for certain (x1 , x2 ), has two solutions y1,l
0 0
, y1,r 0
, y1,l 0
< y1,r .
0 0
This implies that the two characteristic curves starting at (y1,l , 0) and (y1,r , 0) intersect.
0 0
As z is constant along characteristic curves, z will attain the values g(y1,l ) and g(y1,r )
at the intersecting point (x1 , x2 ). These values are in general different and therefore u
cannot be continuous at the intersection point. Such solutions are called “shock waves”.
They are interesting, even though they are not continuous.
For example, in the case g = 1(−∞,c) , for every (x1 , x2 ) with 0 < x1 − c < x2 we have
y1,l = x1 − x2 , y1,r
0 0
= x1 . So u cannot be continuous at such (x1 , x2 ) because z = 1 on the
0
characteristic curve emanating from y1,l and z = 0 on the characteristic curve emanating
0
from y1,r . In particular, there are no classical solutions near (c, 0). However, in the region
{(x1 , x2 ), x1 < c, x2 > max{0, x1 − c}, resp. {(x1 , x2 ), x1 > c, 0 < x2 < x1 − c}, the
problem has a unique classical solution u = 1, resp. u = 0.
39
3.3. Conservation laws and weak solutions Chapter 3
3.3 Conservation laws and weak solutions

The examples we have seen so far in this chapter show that the method of char-
acteristics, depending on the functions E, g, and on the boundary Γ, provides a local
or global classical solution, which may be unique or not, or no classical solution at all.
Of particular interest are the cases in which the classical local solutions interact and
develop discontinuities, like in the shock waves case of Example 3.2.7.
We note that the method of characteristics provides the most natural way of con-
structing the solutions of (3.0.1). In this section we will discuss the concept of the weak
solution to first-order PDEs, which extends the concept of classical solution obtained
with the method of characteristics. We will also present the concept of entropy solution,
which in the cases of non-uniqueness selects the solution that is physically admissi-
ble/reasonable.
We will restrict the discussion on the following class of PDEs:
∂2 u + ∂1 (f (u)) = 0, (x1 , x2 ) ∈ R × R+ := (−∞, +∞) × (0, +∞), (3.3.1a)

u(x1 , 0) = g(x1 ), x1 ∈ R, (3.3.1b)
where u = u(x1 , x2 ) is the unknown function, and g and f are given functions. The
equation (3.3.1a) is a (scalar) conservation law in one space dimension and (3.3.1b) is
the initial condition. Furthermore, u is the conserved quantity and f is its flux.
The name “conservation law” for (3.3.1) follows from this argument. Set x = x1 (the
space variable) and t = x2 (the time), so (3.3.1) becomes
∂t u(x, t) + ∂x (f (u(x, t)) = 0, (3.3.2a)

u(x, 0) = g(x), x ∈ R. (3.3.2b)
Assuming u is a smooth solution to (3.3.2), if (a, b) ⊂ R is a given fixed interval we get

b b
d b
u(x, t)dx = ∂t u(x, t)dx = − ∂x (f (u(x, t))dx
dt a a a
= f (u(a, t)) − f (u(b, t)).
Hence, the total amount of u inside any given interval [a, b] at any given time t remains
constant, if for example the fluxes at the endpoints f (u(a, t)) and f (u(b, t)) are zero, or
even if f (u(a, t)) − f (u(b, t)) = 0.
Remark 3.3.1 The wave equation in one dimension
∂22 w − ∂1 (h(∂1 w)) = 0, w = w(x1 , x2 )
can be written as follows. Let u = (u1 , u2 ) := (∂1 w, ∂2 w). Then u solves the first-order
vector PDEs
∂2 u + ∂1 f (u) = 0, f (u) = −(u2 , h(u1 )). (3.3.3)
40
Chapter 3 3.3. Conservation laws and weak solutions
The equation (3.3.1) with u = (u1 , . . . , uN ) and f (u) = (f1 (u), . . . , fN (u)) (the equation
(3.3.3) is an example with N = 2) is called “vector conservation laws in one space
dimension”. In order to avoid technicalities we will not consider vector conservation
laws. The reader can read more on this topic in [16, 30–32].
Note that equation (3.3.1) is of the form (3.0.1a) with E(x, z, p) = p2 + f (z)p1 .
About the classical solution to (3.3.1) we have this result, which is a straightforward
corollary of Theorem 3.2.3.
Proposition 3.3.2 Let x0 = (x01 , 0) ∈ R × {0} and assume f is C k+1 near x0 in R2 and
g is C k near x01 in R, k ≥ 2. Then
(x0 , z0 (x0 ), p0 (x0 )) = ((x01 , 0), g(x01 ), (g (x01 ), −f (g(x01 ))g (x01 )))
is a non-characteristic initial condition and ∂p2 E((x0 , z0 (x0 ), p0 (x0 ))) = 1, and therefore
there exists a unique C k solution to (3.3.1) near x0 .
Note that the solution to the characteristic equations of (3.3.1) is given by
y1 (t, y10 ) = f (g(y10 ))t + y10 , y2 (t, y10 ) = t, z(t, y10 ) = g(y01 ), y10 ∈ R. (3.3.4)
We have seen this solution in Example 3.2.7, which dealt with a conservation law with
f (u) = 12 u2 . As demonstrated in this example, if g is a decreasing function then the
solutions along the characteristic curves create a discontinuity at the intersection of
characteristic curves. The candidate solution obtained by the method of characteristic
is discontinuous. The following definition extends the concept of the classical solution
to a conservation law in order to allow a discontinuous solution.
Definition 3.3.3 Assume f ∈ L1loc (R) with |f (z)| ≤ C|z| for all z ∈ R and C ≥ 0, and
g ∈ L1loc (R). A function u ∈ L1loc (R × R+ ) is said to be a “weak solution to (3.3.1)” if for
all ϕ ∈ D(R2 ) we have

(u(x)∂2 ϕ(x) + f (u)∂1 ϕ(x)) dx1 dx2 + g(x1 )ϕ(x1 , 0)dx1 = 0. (3.3.5)
R+ R R
This definition is equivalent to say that (3.3.1) holds in the sense of distributions. It
is obtained by multiplying (3.3.1) with the test function ϕ, integrating in R × R+ and
finally using (3.3.1b). Note that the condition |f (z)| ≤ C|z| in Definition 3.3.3 is to
ensure that f (u) ∈ L1loc (R × R+ ) which implies that (3.3.5) is well-defined.
In general, a weak solution to (3.3.1) is not a (classical) solution—we will see this
in the following examples. The following proposition shows that if the weak solution is
smooth then the concepts of weak and classical solutions coincide. The proof is elemen-
tary and can be found in many textbooks; see for example [16, 30, 43]. For the reader’s
convenience, we have included the proof in Theorem 9.2.2 in Annex.
41
Proposition 3.3.4 Let f ∈ C 1 (R), g ∈ C 0 (R). If u ∈ C 1 (R × R+ ) ∩ C 0 (R × [0, ∞)) is

a classical solution to (3.3.1), then u is a weak solution to (3.3.1). Reciprocally, assume
u is a weak solution to (3.3.1) and u ∈ C 1 (Ω) ∩ C 0 (Ω), where Ω ⊂ R × (0, ∞) is open.
Then u is a classical solution to (3.3.1a) in Ω and satisfies u = g on ∂Ω ∩ {x2 = 0}.
The following result, so-called “Rankine-Hugoniot theorem”, essentially shows that
the slope of the curve x1 = χ(x2 ), where the solution to (3.3.1) is discontinuous, equals
the ratio of the jump of f (u) over the jump of u on the discontinuity curve. The proof
is simple and can be found, for example, in [16, 30, 43]. For the reader’s convenience,
we have given the proof in Theorem 9.2.3 in Annex.
Theorem 3.3.5 (Rankine-Hugoniot)
Let Ω ⊂ R × R+ be a simply connected x2
open set such that Ω = Ωl ∪ γ ∪ Ωr , where νl νr

γ is the graph of a function x1 = χ(x2 ), νl
x2 ∈ (α, β), γ ∈ C 1 (α, β), and Ωl , resp. Ωl γ Ωr
νr
Ωr , is the open set at the left, resp. at the
right, of γ; see Fig. 3.3.1. Assume that u is x1
a weak solution to (3.3.1) in Ω such that
u ∈ C 1 (Ωl ) ∩ C 0 (Ωl ) and u ∈ C 1 (Ωr ) ∩
C 0 (Ωr ), and set Figure 3.3.1: Ωl , Ωr and γ
[u] = lim (u(χ(x2 ) + h, x2 ) − u(χ(x2 ) − h, x2 )),

h→0+
[f (u)] = lim (f (u(χ(x2 ) + h, x2 )) − f (u(χ(x2 ) − h, x2 )) x2 ∈ (α, β).
h→0+
Then
χ [u] = [f (u)] on γ. (3.3.6)
Before we consider some examples, let us note that in physical situations x2 is the
time variable and x1 is the space variable. So physically, χ represent the speed of
propagation of the discontinuity curve (chock speed). In the particular case of Burgers’
equation (3.2.22), it gives
f (ur ) − f (ul ) 1 u2r − u2l 1

χ = = = (ur + ul ).
ur − u l 2 ur − u l 2
Note that the characteristic curves on the left of the discontinuity curve have equations
0 0 0 dy1
y1 = y2 g(y1,l ) + y1,l = y2 ul + y1,l ; see (3.2.23) and Example 3.2.7. Hence, = ul is the
dy2
propagation speed of the left characteristic curve (characteristic speed), and similarly
ur is the propagation speed of the right characteristic curve at a discontinuity point.
42
Therefore, the equation above for χ states that the chock speed of Burgers’ equation is
the average of the characteristic speeds.
Burgers’ equation (3.2.22) is a prototype of nonlinear first-order PDEs that, depend-
ing on the initial condition (3.2.22b), represents a number of features such as discontin-
uous weak solutions and non-uniqueness, as the following examples show.
Example 3.3.6 We look for a weak solution to Burgers’ equation (3.2.22) with
g(x1 ) = 1(−∞,0) (x1 ). (3.3.7)
Following Example 3.2.7, if a weak solution has a discontinuity curve γ, then from
Rankine-Hugoniot Theorem 3.3.5 we get γ = {x1 = 12 x2 , x2 > 0}, because [u] = −1,
[f (u)] = −1/2, and so χ = [f[u]
(u)]
= 1/2. It implies that
1, x1 < 12 x2 ,
u(x1 , x2 ) = (3.3.8)
0, x1 ≥ 12 x2
is a candidate for a weak solution to (3.2.22). Let us show that u is a weak solution. For
ϕ ∈ D(R2 ), let R > 0 such that supp(ϕ) ⊂ BR . Then we set Ωl = BR ∩ {(x1 , x2 ), x1 <
1
x , x2 > 0}, Ωr = BR ∩ {(x1 , x2 ), x1 > 12 x2 , x2 > 0}, and let ν l , resp. ν r , the exterior
2 2
unit normal vector on ∂Ωl , resp. Ωr . From the equation of γ, we find that ν l = (2, −1) √15 ,
ν r = −ν l on γ. Then using Gauss theorem4 we get

(u∂2 ϕ + f (u)∂1 ϕ)dx1 dx2 + g(x1 )ϕ(x1 , 0)dx1
R+ R {x2 =0}

= (ul ∂2 ϕ + f (ul )∂1 ϕ)dx + g(x1 )ϕ(x1 , 0)dx1
Ωl ∂Ωl ∩{x2 =0}

+ (ur ∂2 ϕ + f (u)∂1 ϕdx + g(x1 )ϕ(x1 , 0)dx1
Ωr ∂Ωr ∩{x2 =0}

l l
= (ul ν2 + f (ul )ν1 )ϕdx1 + g(x1 )ϕ(x1 , 0)dx1
∂Ωl ∩{x2 =0} ∂Ωl ∩{x2 =0}

r r
+ (ur ν2 + f (ur )ν1 )ϕdx1 + g(x1 )ϕ(x1 , 0)dx1
∂Ωr ∩{x2 =0} ∂Ωr ∩{x2 =0}

− (∂2 u + ∂1 (f (u)))ϕdx − ([u]ν2l + [f (u)]ν1l )ϕds
Ωl ∪Ωr γ
= 0.
In general, conservation law equations (3.3.1) do not have a unique solution, as the
following example shows.
4
If Ω ⊂ RN is open and Lipschitz, ν = (ν1 , . . . , νN ) is the exterior unitary normal vector on ∂Ω and
f, g ∈ C 1 (Ω), then Ω ∂i f (x)g(x)dx = ∂Ω f (x)g(x)νi ds − Ω f (x)∂i g(x)dx.
43
Example 3.3.7 Consider again Burgers’ equation (3.2.22) with

g(x1 ) = 1(0,∞) (x1 ), (3.3.9)
as in Example 3.2.7. Note that the method of characteristics gives the solution

0, x1 < 0,
u(x1 , x2 ) = (3.3.10)
1, x1 ≥ x2 ,
leaving the solution undefined in U = {(x1 , x2 ), 0 < x1 < x2 }; see Fig. 3.3.2. We can
obtain many weak solutions by defining the solution in U in many ways, for example
⎧
⎨ 0, x1 < 0,
u(x1 , x2 ) = x1 /x2 , 0 ≤ x1 < x2 , (3.3.11)
⎩
1, x1 ≥ x2 ,
which gives a continuous solution, and
0, x1 < 12 x2 ,
u(x1 , x2 ) = (3.3.12)
1, x1 ≥ 12 x2 ,
which gives a discontinuous solution; see Fig. 3.3.2.
There are several criteria to identify one among many solutions to a first-order conser-
vation law, which is physically admissible/reasonable. We will consider only Lax entropy
condition.
Definition 3.3.8 Let u be a weak solution to the conservation law (3.3.1) that more-
over is a piecewise classical solution, i.e. u is C 1 everywhere in R × R+ except on
a set of smooth curves, where the solution is discontinuous but having side limits.
We denote by ul , resp. ur , the restriction of u near a point on a discontinuity curve
γ = {(χ(x2 ), x2 ), x2 ∈ (a, b)} of u on the left, resp. right, of C. Such a solution u is
said to be “admissible” if it satisfies the condition
f (ul ) > χ > f (ur ), (3.3.13)
at every point on the discontinuity curve γ. This condition is called a “Lax entropy
condition”. A curve of discontinuity of u is called a “shock”, if Lax entropy condition
holds on the curve. A weak solution satisfying Lax entropy conditions is called a “Lax
entropy solution”.
The motivation for the condition (3.3.13) is the idea that “the information should
always flow from the boundary onto the shock curve”—so the values of u are transported
from R × {0}, where u is known, to the chock curve. This is the same as saying that the
characteristic curves start on R × {0}, where u is known, and end on the chock curve,
which based on (3.3.4) is equivalent to (3.3.13). If the condition (3.3.13) is not satisfied
then in general there are characteristic curves which start at the discontinuity curve and
do not intersect the boundary—hence the information is “created” on this discontinuity
curve, which is not physically admissible/reasonable.
44
Example 3.3.9 Consider again Burgers’ equation (3.2.22). If g is given by (3.3.7), see
Fig. 3.2.2 right, then the solution given by (3.3.8) is an entropy solution because it
satisfies Lax entropy condition on the discontinuity curve {x1 = x2 /2}:
1 x2
f (ul ) = 1 > χ = > 0 = f (ur ). u(x1 , 0)
2
If g is given by (3.3.9), see Fig. 3.3.2,
there are infinitely many solutions. The
x1
equation (3.3.11) gives a continuous
solution, so it is not an entropy solution. x2
The equation (3.3.12) gives a weak dis-

continuous solutions, however it is not
an entropy solution because along its
discontinuity curve γ = {(x1 , x2 ), x1 = x1
1
x , x2 > 0} it satisfies Rankine-
2 2
Hugoniot condition but not Lax entropy Figure 3.3.2: Some characteristic
condition: curves start at the discontinuity curves
1 [f (u)]
χ = = , and f (ul ) = 0 ≯ χ ≯ 1 = f (ur ).
2 [u]
Here, we presented a short introduction to the first-order PDEs. We developed the

concepts of characteristic equations and proved the existence of local classical solutions.
We showed that, in general, a first-order PDE does not have a global classical solution,
and developed the notion of the “weak solution”, which we demonstrated for single/scalar
conservation laws. A first-order PDE may have many weak solutions, some of which may
have discontinuities. A necessary condition for the weak solution along the discontinuity
curves is given by the Rankine-Hugoniot condition. By using Lax entropy condition,
we select one solution among many weak discontinuous solutions, which is physically
admissible. There are other conditions which are used to select a physical reasonable
solution among weak solutions, such as entropy condition or viscosity solutions.
Real physical problems involve systems of first-order multi-dimensional PDEs. The
research on these PDEs is very active and the issues of existence and uniqueness are not
completely solved. See, for example, [4, 10, 32] for more details on these topics.
Problems
Problem 3.1 Let b ∈ RN , c ∈ R and f , resp. h, be a smooth function in RN × (0, ∞),
resp. on RN , u = u(x, t), and consider

ut + b · ∇u + cu = f in RN × (0, ∞),
u = h on RN × {0}.
45
d
Find a classical solution u by considering ds
u(x + sb, t + s).
Problem 3.2 Find a non-trivial classical solution u = u(x1 , x2 ) to the following PDEs,
by using the method of characteristics (here g is C 1 ).
a) x1 ∂1 u − u∂2 u = x2 in {x1 > 1}, u(1, x2 ) = x2 , x2 ∈ R,

√
b) u2 ∂1 u + ∂2 u =0 in (0, ∞) ,
2
u(x1 , 0) = x1 , x1 ∈ (0, ∞),
c) x21 ∂1 u + x22 ∂2 u = u2 in {x2 > 2x1 }, u(x1 , 2x1 ) = x21 , x1 ∈ R,
d) ∂1 u − ∂2 u + (∂2 u)2 =0 in R,2
u(x1 , 0) = 0, x1 > 0,
e) x 1 ∂1 u + x 2 ∂2 u + u =0 in R2 \(0, 0), u(x1 , x2 ) = 1, {x21 + x22 = 1},
f) (∂1 u)2 + (∂2 u)2 = u2 in R2 , u(x1 , x2 ) = 1, {x21 + x22 = 1},
g) ∂1 u∂2 u =u in {x1 > 0}, u(0, x2 ) = x22 , x2 ∈ R,
h) ∂1 u + ∂ 2 u =1 in {x2 > 0}, u(x1 , 0) = g(x1 ), x1 ∈ R,
i) x1 ∂1 u + (x1 + x2 )∂2 u =1 in {x1 > 1}, u(1, x2 ) = g(x2 ), x2 ∈ R,
j) x1 ∂ 1 u + x 2 ∂ 2 u = 2u in R × (0, ∞), u(x1 , 1) = g(x1 ), x1 ∈ R,
k) (x1 + x2 )∂1 u + ∂2 u =0 in R × (0, ∞), u(x1 , 0) = g(x1 ), x1 ∈ R,
l) ∂1 u + ∂2 u = u2 in {x2 > 0}, u(x1 , 0) = g(x1 ), x1 ∈ R.
Discuss the largest domain where u is defined and the uniqueness of the solution.
Problem 3.3 Assume that g is an increasing C 1 function and consider

1
u∂1 u + ∂2 u + u = 0 in {x2 > 0}, u(x1 , 0) = g(x1 ).
2
i) Solve the characteristic equations to this PDE.
ii) Let y(t, y10 ) be the equations of the characteristic curve starting at (y10 , 0). Use the
implicit function theorem to show that the map (y10 , t) → (x1 , x2 ), with y(t, y10 ) = (x1 , x2 )
and x2 > 0, is invertible and C 1 .
iii) Conclude that there exists an implicit classical solution u.
Problem 3.4 Consider the PDE

(x22 + u)∂1 u + x2 ∂2 u = 0, u = 0 on Γ = (x1 , x2 ), 2x1 = x22 , x2 > 1 .
i) Show that every characteristic curve starting on Γ coı̈ncides with Γ.

ii) Find a classical solution u = 0 in {x2 > 1}.
Problem 3.5 Consider the PDE
x2 ∂1 u − x1 ∂2 u = 0 in R × (0, ∞).
Verify whether the method of characteristics provides a classical solution to this PDE
when equipped with one of the following boundary conditions:
i) u(x1 , 0) = x21 , x1 ∈ R, ii) u(x1 , 0) = x1 , x1 ∈ R, iii) u(x1 , 0) = x1 , x1 < 0.
46
Problem 3.6 Consider the problem (3.3.1) with f and g as below. For each of them
show that there exists an entropy solution.5
a) f (u) = 12 u2 , g(x1 ) = sign(−x1 ), c) f (u) = u3 , g(x1 ) = ex1 H(−x1 ),

3
b) f (u) = 23 u 2 , g(x1 ) = H(−x1 ), d) f (u) = 12 u2 , g(x1 ) = x1 1(−∞,0) − 1(0,∞) .
Problem 3.7 Consider the problem (3.3.1) with f ∈ C 1 (R) non-decreasing and f (0) =
0, and assume that u is a weak piecewise solution with discontinuity along the curve
γ = {(χ(x2 ), x2 ), x2 ≥ 0}. Show that if f is convex, resp. concave, then shocks travel
forward, resp. to the left, i.e. γ is an increasing, resp. decreasing, function.
Problem 3.8 Let u : R × R+ → R be a piecewise smooth monotone decreasing in x1

weak solution to (3.3.1) with a finite number of jumps. Show that if f ∈ C 1 (R) is a
convex function then u satisfies the Lax entropy condition at each discontinuity point.
Show also that if f is not convex then Lax entropy condition does not hold in general.
5
Hint: after solving the characteristic equations, you may “guess” the discontinuity curve γ by
solving the differential equation χ = [f[u]
(u)]
; next you find a solution candidate by using the method of
characteristics and show that it is indeed an entropy solution.
47
4. Second-order linear elliptic PDEs:
maximum principle and classical
solutions
In this chapter, we will study the problem
−Δu = 0 in Ω, (4.0.1a)
u = g on ∂Ω, (4.0.1b)
called “Dirichlet problem for the Laplacian”, as a prototype of second-order linear elliptic
PDEs. A classical solution to the problem (4.0.1) is any function u ∈ C 0 (Ω) ∩ C 2 (Ω)
satisfying (4.0.1) pointwise.1 We will prove the existence and uniqueness of a classical
solution to this problem by using Perron’s method. This method looks for the solution
at the intersection of subsolutions and supersolutions to (4.0.1), and strongly uses the
maximum principle, which roughly speaking states that the solution to a certain PDE in
a domain Ω attains its maximum value on the boundary ∂Ω. Note this method applies
to more general second-order linear PDEs; see, for example, [20].
There are other methods which provide classical solutions, such as Schauder’s theory,
which in fact provides C m,α (Ω) solutions; see, for example, [20]. We note that, in general,
(4.0.1) does not have a classical solution. This is a reason why the weak/variational
solutions are introduced; see Chapter 7.
4.1 Laplace equation and the method of separation

of variables
We start this chapter with the method of separation of variables (MSV) for (4.0.1).
This elementary method, valid in Ω of rectangular form, provides a candidate solution
in the form of a series. By using the mean value theorem and the maximum principle,
we will show that this series provides a classical solution to (4.0.1).
1
As we will see in this chapter, a classical solution u to (4.0.1) is C ∞ , i.e. u ∈ C 0 (Ω) ∩ C ∞ (Ω).
https://doi.org/10.1007/978-3-031-39524-6 4
4.1. Laplace equation and the method of separation of variables Chapter 4
We consider the problem

−Δu = 0 in Ω := (0, a) × (0, b), a, b > 0, (4.1.1a)
u = 0 on ∂Ω\Σ, Σ := [0, a] × {b}, (4.1.1b)
u = g on Σ, with g ∈ C00 (Σ). (4.1.1c)
Disregarding at this time the issues of existence and uniqueness of the solution (see
Remark 4.1.1), we are interested only in finding a formula for the solution u. We assume
that u = u(x, y) can be written as u(x, y) = V (x)W (y) (where the term method of
separation of variables comes from). Replacing u in (4.1.1a) gives (Figure 4.1.1)
−(V (x)W (y) + V (x)W (y)) = 0. (4.1.2)
Assuming V W = 0, from the last y

V (x) W (y) b
Σ
equation we obtain =− . g
V (x) W (y)
Then, necessarily there exists a con-
0 Ω 0
stant λ such that
0
V (x) W (y) a x
= λ, = −λ,
V (x) W (y) Figure 4.1.1: Domain, boundary,
or equivalently and boundary data
V − λV = 0, W + λW = 0. (4.1.3)
The differential equations (4.1.3) are of second order, linear, and homogeneous, so their
solution spaces are of dimension two and their general solutions are
V (x) = Ax + B, W (y) = Cy + D, if λ = 0, (4.1.4)
√ √ √ √
x λ −x λ y −λ −y −λ
V (x) = Ae + Be , W (y) = Ce + De , if λ = 0, (4.1.5)
√
with A, B, C, D being arbitrary constants. Here, ±λ can be a complex number, and
in this case A, B, C, and D must be such that the functions V and W are real. It follows
that
u(x, y) = (Ax + B)(Cy + D), λ = 0, (4.1.6)
√ √ √ √
x λ −x λ y −λ −y −λ
u(x, y) = (Ae + Be )(Ce + De ), λ = 0, (4.1.7)
is always a solution to (4.1.1a). The constants A, B, C, D are chosen such that u solves
the boundary conditions (4.1.1b), (4.1.1c). Condition (4.1.1b) is written as

u(0, y) = u(a, y) = u(x, 0) = 0,
⇐⇒ V (0) = V (a) = W (0) = 0.
∀x ∈ (0, a), ∀y ∈ (0, b),
In the case λ = 0, by using (4.1.6) we get
B = Aa + B = D = 0,
50
Chapter 4 4.1. Laplace equation and the method of separation of variables
which implies A = B = D = 0 and therefore u = 0. In the case λ = 0, by using (4.1.7)

we get
√ √
A + B = Aea λ
+ Be−a λ
= C + D = 0.
Assuming u = 0, necessarily we get
√
λ
√ π √ π
0 = B = −A, 0 = D = −C, e2a = 1, λ = ik , −λ = −k , k ∈ N.
a a
Therefore, u(x, y) given by (4.1.7) can be written as
uk (x, y) = (−AD)(eikπx/a − e−ikπx/a )(ekπy/a − e−kπy/a )
= (−4iAD) sin(kπx/a) sinh(kπy/a)
=: γk sin(kπx/a) sinh(kπy/a).
So, there are infinitely many uk (x, y) solving (4.1.1a) and (4.1.1b), and so formally

u(x, y) = uk (x, y) = γk sin(kπx/a) sinh(kπy/a) (4.1.8)
k∈N k∈N
solves (4.1.1a) and (4.1.1b). Now, we address (4.1.1c). Replacing (4.1.8) in (4.1.1c)
implies that the coefficients γk must satisfy

u(x, b) = γk sin(kπx/a) sinh(kπb/a) = g(x), x ∈ (0, a), (4.1.9)
k∈N
which is the sine Fourier expansion of g in [0, a]. Note that we have the following for-
mulas2
a π π a
sin k x sin l x dx = δkl , k, l ∈ N. (4.1.10)
0 a a 2
Then multiplying (4.1.9) by sin(lπx/a) and integrating in (0, a) gives
a
2 1
γk = sin(kπx/a)g(x)dx, k ∈ N.
a sinh(kπb/a) 0
Therefore, the solution u is given by

2 sin(kπx/a) sinh(kπy/a) a
u(x, y) = sin(kπx/a)g(x)dx. (4.1.11)
a k∈N sinh(kπb/a) 0
Finally, we note that we can use MSV for solving problems like
−Δu = 0 in Ω = (0, a) × (0, b), u = g on ∂Ω. (4.1.12)
Indeed, if Σi , i = 1, 2, 3, 4 are the sides of ∂Ω, then one can look for u = u1 +u2 +u3 +u4 ,
−Δui = 0 in Ω = (0, a) × (0, b), ui = g1Σi on ∂Ω, (4.1.13)
where 1Σi is the characteristic function of Σi . Each ui can then be solved using MSV.
a a
2
Similarly, we have 0
cos k πa x cos l πa x dx = δkl a2 , 0 sin k πa x cos l πa x dx = 0, k, l ∈ N.
51
4.1. Laplace equation and the method of separation of variables Chapter 4
Remark 4.1.1 Note that (4.1.11) is not found rigorously, and therefore we cannot say
that the series (4.1.11) converges, nor that u is a classical solution to (4.1.1). To empha-
size this, sometimes we will write ∼ instead of = in (4.1.11).
How can we verify whether the series in (4.1.11) provides a classical solution? The
answer relies on the convergence of the Fourier series and the maximum principle.
Indeed, if g is smooth then the sine Fourier series of g converges uniformly, which
implies that the (4.1.11) converges in C 0 (∂Ω). Assuming this convergence holds, in com-
bination with the maximum principle it implies the convergence of the series (4.1.11) in
C 0 (Ω), which in turn implies u ∈ C 0 (Ω) and u = g on ∂Ω. Furthermore, by using the
mean value theorem one proves that u ∈ C 2 (Ω) and Δu = 0 in Ω, which shows that u
is a classical solution to (4.1.1). See Corollary 4.3.5 and examples 4.3.5, 4.3.6 for more
details.
MSV, although useful, cannot be used when the domain is not a rectangle or a polar
sector (see below). Theorem 4.4.8 provides general criteria for the solution to (4.0.1),
which for the problem (4.1.1) implies that it has a unique classical solution if and only
if g ∈ C00 (Σ).
Remark 4.1.2 MSV can be used to solve (4.1.1) with
u = 0 on ∂Ω\Σ, Σ := (0, a) × {b}, (4.1.14a)

∂y u = g on Σ. (4.1.14b)
instead of (4.1.1b), (4.1.1c). One can proceed as above with the only difference that we
define γk in (4.1.8) by using ∂y u(x, b) = g(x) instead of u(x, b) = g(x).
Remark 4.1.3 One can use MSV in polar coordinates, which is useful when the domain
is a disk or a polar sector. For example, it is easy to show that if (r, θ) are polar coor-
dinates, x = r cos θ, y = r sin θ, then the problem
−Δu = 0 in Ω := B(0, R), u=g on ∂Ω,
can be written in polar coordinates as

1 1
− ∂r (r∂r u) + 2 ∂θ2 u = 0, (r, θ) ∈ (0, R) × [0, 2π), (4.1.15a)
r r
u(R, θ) = g(θ), θ ∈ [0, 2π). (4.1.15b)
Then one looks for u(r, θ) = R(r)Θ(θ), which after replacing in (4.1.15a) gives
1 1
(rR ) Θ + 2 RΘ = 0.
r r
Assuming RΘ = 0, dividing the equation above by RΘ gives
2
r R + rR − λR = 0, Θ + λΘ = 0, Θ(0) = Θ(2π) = 0,
(4.1.16)
r ∈ (0, R) θ ∈ (0, 2π), Θ (0) = Θ (2π),
52
Chapter 4 4.2. Dirichlet problem in a ball
with λ a constant. These equations have infinitely many solutions Rk , Θk , k = 0, 1, . . .

given by
R0 (r) = C0 + D0 ln r, Θ0 = A0 ,
Rk (r) = Ck rk + Dk r−k , k = 1, 2, . . . , Θk (θ) = Ak cos(kθ) + Bk sin(kθ), k ∈ N.
Therefore, a series candidate solution to −Δu = 0 in B(0, R) is given by

∞
u(r, θ) ∼ A0 (C0 + D0 ln r) + (Ck rk + Dk r−k )(Ak cos(kθ) + Bk sin(kθ)).
k=1
Assuming u is bounded implies D0 = D1 = · · · = 0, and then
α0 r k
∞
u(r, θ) ∼ + (αk cos(kθ) + βk sin(kθ)). (4.1.17)
2 k=1
R
Finally, the coefficients αk , βk are defined formally by replacing the series (4.1.17) in
(4.1.15b), multiplying by cos(lθ), sin(lθ), and integrating in [0, 2π]. It implies

1 2π
α0 = g(s)ds, (4.1.18)
π 0

1 2π 1 2π
αk = g(s) cos(ks)ds, βk = g(s) sin(ks)ds, k = 1, 2, . . . . (4.1.19)
π 0 π 0
We conclude this remark by pointing out that we can use MSV to solve problems like
−Δu = 0 in Ω, u = g on ∂Ω,
with Ω = {(r, θ), r ∈ [0, R), θ ∈ (0, σ), σ ∈ (0, 2π)}, or Ω = B(0, R)c .
4.2 Dirichlet problem in a ball

In this section, we will solve problem (4.0.1) in the case Ω is a ball. This problem is
a building block of Perron’s method that we will use for solving (4.0.1) with Ω general.
Theorem 4.2.1 (Dirichlet problem in a ball) Let R > 0, BR := {x ∈ RN , |x| < R},
VN the volume of B1 , g ∈ C 0 (∂BR ) and set

R2 − |x|2 g(y)
u(x) = dσ(y), x ∈ BR . (4.2.1)
N VN R ∂BR |x − y|N
Then u is the unique C 0 (B R ) ∩ C ∞ (BR ) solution to (4.0.1) in Ω = BR .

The proof of this theorem is classical, see, for example, [20], and follows from direct
calculations; see Theorem 9.3.1 in Annex for the details.
53
4.2. Dirichlet problem in a ball Chapter 4
Remark 4.2.2 Theorem 4.2.1 shows that if u ∈ C 0 (Ω) ∩ C 2 (Ω) is a classical solution
to (4.0.1) then u ∈ C 0 (Ω) ∩ C ∞ (Ω) because if B Ω is a ball and h solves −Δh = 0 in
B, h = u on ∂B, then u = h so u ∈ C ∞ (B), which implies u ∈ C ∞ (Ω) because B Ω
is arbitrary.
Remark 4.2.3 (Mean value theorem (MVT)) The formula (4.2.1) is called “Pois-
son’s formula”. In an arbitrary ball BR (z) := B(z, R), it has the form

R2 − |x − z|2 g(y)
u(x) = dσ(y), x ∈ BR (z), (4.2.2)
N VN R ∂BR (z) |x − y|N
which for x = z gives

R2 g(y) 1
u(z) = dσ(y) = g(y)dσ(y)
N VN R ∂BR (z) RN |∂BR | ∂BR (z)

=: − u(y)dσ(y), (4.2.3)
∂BR (z)
where |∂BR | = N VN RN −1 is the area of ∂BR .

Formula (4.2.3) is called a “mean value theorem in Ω”: if u ∈ C 2 (Ω) with −Δu = 0
in Ω, then u(x), x ∈ Ω, is equal to the mean value of u over the boundary of any ball
B(x, r), B(x, r) Ω.
In the case N = 2, one can easily find the formula for u(x) in (4.2.1) by using the
method of separation of variables in polar coordinates; see Remark 4.1.3. Indeed, assume
for simplicity R = 1. Then from (4.1.17), taking into account (4.1.19) and permuting
the sum with the integrals gives

1 k
2π ∞
1
u(r, θ) ∼ g(s) + r (cos(kθ) cos(ks) + sin(kθ) sin(ks)) ds
π 0 2 k=1

1 k
2π ∞
1
= g(s) + r cos(k(θ − s)) ds
π 0 2 k=1

∞
1 2π k 1
= g(s) Re rei(θ−s) − ds
π 0 k=0
2
2π
1 1 1
= g(s) Re − ds
π 0 1 − re i(θ−s) 2

1 2π 1 1
= g(s) Re − ds
π 0 (1 − r cos(θ − s)) − ir sin(θ − s) 2

1 2π 1 − r cos(θ − s) 1
= g(s) − ds
π 0 (1 − r cos(θ − s)) + (r sin(θ − s))
2 2 2
54
Chapter 4 4.3. Maximum principle for Laplacian
2π
1 2(1 − r cos(θ − s)) − 1 + 2 cos(θ − s) − r2
= g(s) ds
2π 0 1 − 2 cos(θ − s) + r2

1 − r2 2π g(s)
= ds
2π 1 − 2 cos(θ − s) + r2

0
1 − |x|2 g(s)
= ds.
∂B(0,1) |x − y|
2π 2
The following proposition shows that the property (4.2.3), even though involving
just an integral, which is well-defined for u ∈ C 0 (Ω), actually is very strong.
Proposition 4.2.4 Let Ω ⊂ RN be open and u ∈ C 0 (Ω). The followings are equivalent:
(i) u ∈ C 2 (Ω) and Δu = 0 in Ω.
(ii) u satisfies the mean value theorem.
Proof. Indeed, (i) implies (ii) follows from Remark 4.2.3. Now let us prove that (ii)
implies (i). Let B = B(x, R) Ω, h ∈ C 0 (B) ∩ C 2 (B) such that
h = u on ∂B and − Δh = 0 in B.
We will show that u = h in B. Indeed, set M = max{|u(y) − h(y)|, y ∈ B} and

KM = {y ∈ B, |u(y) − h(y)| = M }. Let us assume M > 0. As u − h = 0 on ∂B, there
exists xM ∈ ∂KM ∩ B and rM > 0 such that B(xM , rM ) B, |u(xM ) − h(xM )| = M ,
and u − h does not change sign in B. From property (4.2.3) for u and h in B, we get

M = ±(u(xM ) − h(xM )) = ± − (u(y) − h(y))dσ(y) < M,
∂B(xM ,rM )
with strict inequality holding because u−h is continuous and |u−h| M on ∂B(xM , rM ).
The contradiction proves M = 0 and hence u = h in B. From Theorem 4.2.1 and the
arbitrariness of B, we get u ∈ C 2 (Ω) and Δu = 0 in Ω.
4.3 Maximum principle for Laplacian

The proof of existence and uniqueness of solutions to the problem (4.0.1) for general
domain Ω requires the maximum principle, subharmonic and superharmonic functions,
and certain analysis results. We will prove the results we need for the operator −Δ.
Most of them hold for general second-order elliptic PDEs, and the related proofs are
presented in section 9.3.2 in Annex.
All along this chapter, u ∈ C 0 (Ω) ∩ C 2 (Ω), unless otherwise specified, and
Lu := −Δu. (4.3.1)
55
4.3. Maximum principle for Laplacian Chapter 4
For the weak maximum principle, the reader can think of a function u in [a, b] ⊂ R
with −u negative, resp. positive. Then u is concave up, resp. concave down, and there-
fore it reaches the maximum, resp. minimum, on the boundary of (a, b). The following
theorem generalizes these facts in dimension N .
Theorem 4.3.1 (Weak maximum principle) Let Ω ⊂ RN be open bounded and
assume Lu ≤ 0, resp. Lu ≥ 0, in Ω. Then the maximum, resp. minimum, of u in
Ω is achieved on ∂Ω, i.e.
max u = max u, resp. min u = min u,

Ω ∂Ω Ω ∂Ω
where min f = min{f (x), x ∈ A}, max f = max{f (x), x ∈ A} for every f ∈ C 0 (A).
A A
Proof. We deal first with the case Lu ≤ 0. Let x0 ∈ Ω such that u(x0 ) = max{u(x), x ∈
Ω}. We distinguish two sub-cases.
(i) Case Lu < 0 in Ω. If x0 ∈ ∂Ω the theorem is proved. So we assume x0 ∈ Ω. Hence
∇u(x0 ) = 0 and ∂ii u(x0 ) ≤ 0, for all i. Therefore Lu(x0 ) = −Δu(x0 ) ≥ 0, which is a
contradiction and proves that x0 ∈ Ω.
(ii) Case Lu ≤ 0 in Ω. Let > 0 and consider u (x) = u(x) + ex1 , x = (x1 , . . . , xN ).
Then
Lu = Lu + Lex1 = Lu − ex1 < 0.
From Case (i), it follows that there exists x = (x1 , . . . , xN ) ∈ ∂Ω such that

u (x) ≤ u (x ), so u(x) ≤ u(x ) + (ex1 − ex1 ), ∀x ∈ Ω.
From the compactness of ∂Ω there exists a subsequence of x , still denoted by x ,

and x0 ∈ ∂Ω, such that lim→0 x = x0 , which implies that for all x ∈ Ω we have
u(x) ≤ u(x0 ) = max{u(x), x ∈ Ω}. So the claim for the case Lu ≤ 0 is proved.
For the case Lu ≥ 0 we set v = −u, so Lv ≤ 0, and then the result follows from (ii).
Corollary 4.3.2 (Comparison principle) Let Ω be open bounded.
(i) If Lu ≤ Lv in Ω and u ≤ v on ∂Ω, then u ≤ v in Ω (monotonicity3 of L).
(ii) If Lu = Lv in Ω and u = v on ∂Ω, then u = v in Ω (uniqueness of Lu = 0).
Proof. Set w = u − v.
(i) We have Lw ≤ 0 in Ω, w ≤ 0 on ∂Ω. From Theorem 4.3.1 it follows max w ≤ max w ≤ 0.
Ω ∂Ω
So, w ≤ 0, or u ≤ v.
3
This is the reason for the negative sign “−” in front of Δ.
56
Chapter 4 4.3. Maximum principle for Laplacian
(ii) As Lw = 0 and w = 0, (i) implies 0 ≤ w ≤ 0, so u = v.
The result of Theorem 4.3.1 does not exclude that u achieves the maximum also at a
certain point inside the domain. The strong maximum principle excludes this possibility.
The following lemma and theorem are classical results; see, for example, [20].
Lemma 4.3.3 (Hopf lemma) Let Ω ⊂ RN be an open set with ∂Ω a C 2 boundary
near4 x0 ∈ ∂Ω. Assume furthermore Lu(x) ≤ 0, u(x) < u(x0 ) for all x ∈ Ω near x0 and
u is differentiable at x0 . Then ∂ν u(x0 ) := ∇u(x0 ) · ν(x0 ) > 0, where ν(x0 ) is the unit
outward normal vector to ∂Ω at x0 .
Proof. See Lemma 9.3.5 for the details of the proof.
In the case of a local minimum, we have the following result, which we still refer to as
“Hopf lemma”. Let Ω ⊂ RN be an open set with ∂Ω a C 2 boundary near x0 ∈ ∂Ω. Assume
furthermore Lu ≥ 0 in Ω, u(x) > u(x0 ) for all x ∈ Ω near x0 and u is differentiable at
x0 . Then ∂ν u(x0 ) < 0. The proof follows directly from Lemma 4.3.3 applied to −u.
We note that under the assumptions of lemma it follows easily that ∂ν u(x0 ) ≥ 0. The
strength of this lemma is that it rules out the case of equality. This result is strongly
used in the following theorem (Figure 4.3.1).
Theorem 4.3.4 (Strong maximum principle) Let Ω ⊂ RN be an open bounded set,

and assume Lu ≤ 0, resp. Lu ≥ 0. Then u is constant or, if not, u does not achieve its
maximum, resp. minimum, in Ω.
Proof. Let us consider first the case Lu ≤ 0. Let M = max{u(x), x ∈ Ω and set
K = {x ∈ Ω, u(x) = M }. Note that K is closed. Note also that K = ∅ because u
attains its maximum M in Ω.
If K = Ω then u = M and the theorem is proved.
Otherwise, we assume the theorem does not hold,
i.e. u achieves its maximum in Ω. Hence, K ∩Ω r
Ω. Then there exists a ball B0 Ω, B0 ∩ K = ∅, z
and x0 ∈ ∂B0 such that u achieves at x0 a strict x0 z0
maximum in B0 . Indeed, there exists z ∈ ∂K ∩ Ω
and r > 0 such that B = B(z, r) Ω. The set
B\K is open and nonempty. Let B0 = B(z0 , r0 ), K ∂K Ω\K
for a certain z0 ∈ B\K and r0 > 0, be the biggest
open ball included in B\K. From the maximality
of B0 , necessarily B0 ∩ ∂K = ∅, and let x0 ∈
∂B0 ∩ K. It follows that |∇u(x0 )| = 0. Figure 4.3.1: B(z0 , r0 ) and x0 .
4
Here, “∂Ω a C 2 boundary near x0 ” means that in Definition 1.2.4 we replace “for every x ∈ ∂Ω”
with “for x = x0 ”. Note that this implies that Ω ∩ B(x0 , r0 ) lies on one side of ∂Ω, for r0 > 0 small.
57
4.3. Maximum principle for Laplacian Chapter 4
Now we can apply Lemma 4.3.3 with B0 instead of Ω. It gives ∂ν u(x0 ) > 0, where
ν = (x0 − z0 )/|x0 − z0 |. But this contradicts the fact that |∇u(x0 )| = 0. Hence the
theorem’s claim holds.
The case Lu ≥ 0 is proved by setting v = −u and using the case Lu ≤ 0.
The following corollary is related to the solution to the problem (4.0.1) with the
MSV; see Section 4.1. It shows that if the series representing the boundary function g
convergences in C 0 (∂Ω) then the MSV provides a classical (unique) solution.
Corollary 4.3.5 Let Ω ⊂ RN open bounded and g, gn ∈ C 0 (∂Ω), n ∈ N, such that

limn→∞ gn = g in C 0 (∂Ω). Assume also that un ∈ C 0 (Ω) ∩ C ∞ (Ω), n ∈ N, satisfies
Δun = 0 in Ω, un = gn on ∂Ω. Then there exists u ∈ C 0 (Ω) ∩ C ∞ (Ω) such that
limn→∞ un = u in C 0 (Ω) and
Δu = 0 in Ω, u=g on ∂Ω.
Proof. We note that (un ) is a Cauchy sequence in C 0 (Ω) because from
Δ(un − um ) = 0 in Ω, un − um = gn − gm on ∂Ω,
and Theorem 4.3.1, we get un − um C 0 (Ω) ≤ gn − gm C 0 (∂Ω) . Therefore, limn→∞ un = u
in C 0 (Ω) and u = g on ∂Ω, for a certain u ∈ C (Ω).
0
As Δun = 0 in Ω, we have un (x) = −∂BR (x) un (y)dσ(y), for all BR (x) Ω. Passing to

the limit in this equality implies u(x) = −∂BR (x) u(y)dσ(y), which from Proposition 4.2.4,
gives u ∈ C 0 (Ω) ∩ C ∞ (Ω) and Δu = 0 in Ω. Then Theorem 4.2.1 implies u ∈ C ∞ (Ω).
Example 4.3.6 Let us consider the problem

Δu = 0 in Ω = (0, π) × (0, π),

0, (x, y) ∈ ∂Ω\Σ,
u(x, y) = g(x, y), with g(x, y) =
x(π − x), (x, π) ∈ Σ = (0, π) × {π}.
a) Find a series solution candidate u = u(x, y) of this problem.
b) Prove that the series converges to a solution u ∈ C 1 (Ω) ∩ C ∞ (Ω).
Solution. 1−(−1)k
i) We use the fact that the sine Fourier series of x(π − x) is ∞ k=1 π ·
4
k3
sin(kx),
0
and that the convergence 5 is in C ([0, π]).
Then using the MSV it is easy to find a series solution candidate
∞
4 1 − (−1)k sinh(ky) ∞
u(x, y) ∼ · sin(kx) =: vk (x, y).
k=1
π k3 sinh(kπ) k=1
n
ii.1) Set un (x, y) = k=1 vk (x, y) in Ω and gn = un on ∂Ω. Then limn→∞ gn −
gC 0 (∂Ω) = 0. Indeed, gn = g = 0 on ∂Ω\Σ, g ∈ C 1 (Σ) ∩ C00 (Σ) and from results
58
Chapter 4 4.4. Solution to the Dirichlet problem
related to Fourier series5 we have limn→∞ gn − gC 0 (Σ) = 0.

ii.2) As Δun = 0 in Ω, from Corollary 4.3.5 it follows that u ∈ C 0 (Ω) ∩ C ∞ (Ω) and
Δu = 0 in Ω.
ii.3) In fact we even have u ∈ C 1 (Ω). Indeed, let ∂i un = nk=1 ∂i vk , where i ∈ {x, y}.
Then ∂i un is a Cauchy sequence in C 0 (Ω) because
C
|∂i vk | ≤ , C > 0.
k2
Therefore, (un ) is a Cauchy sequence in C 1 (Ω) and so un converges6 to u in C 1 (Ω), so
u ∈ C 1 (Ω).
4.4 Solution to the Dirichlet problem

In this section, we will introduce the sub(super) harmonic functions and present some
related results. We will also present some auxiliary results from analysis and conclude
with the solution to the Dirichlet problem.
4.4.1 Sub(super) harmonic functions and sub(super) solutions

It turns out that the solution to (4.0.1) is at the intersection of two classes of func-
tions, namely the subsolutions and supersolutions to (4.0.1). These sub(super) solutions
have interesting properties listed below. We will show that the solution to (4.0.1) can be
found as the supremum of subsolutions, or similarly as the infimum of supersolutions,
to (4.0.1).
Definition 4.4.1 Let Ω ⊂ RN open. We say u is subharmonic (superharmonic) in Ω if
i) u ∈ C 0 (Ω) and
ii) for every ball B Ω, for every h ∈ C 0 (B) ∩ C 2 (B) with −Δh = 0 in B we have
u ≤ h (u ≥ h) on ∂B implies u ≤ h (u ≥ h) in B.
We say u is a subsolution (supersolution) to Dirichlet problem (4.0.1) if u is subharmonic

(superharmonic) in Ω and u ≤ g (u ≥ g) on ∂Ω.
The concept of subsolution (supersolution) is trivially extended for operators L like
(9.3.5), by just replacing −Δ by L in Definition 4.4.1. Below, we list a number of
properties of subharmonic (superharmonic) functions. For their proof, see section 9.3.3
in Annex, or [20].
5
If f ∈ C00 ([a, b]) ∩ Lip(a, b) then the sine Fourier series of f converges in C 0 ([a, b]) to f .
6
∞This is the so-called “Weierstrass M-test”, which says: given a set ∞Ω ⊂ RN and a seriesof functions
∞
k=1 uk (x), with |uk (x)| ≤ Mk for all x ∈ Ω, where Mk ≥ 0 and k=1 Mk < ∞, then k=1 |uk (x)|
0
converges in C (Ω).
59
4.4. Solution to the Dirichlet problem Chapter 4
Proposition 4.4.2 Let u ∈ C 0 (Ω) ∩ C 2 (Ω) such that −Δu = 0 in Ω.
i) u is a subharmonic and a superharmonic function.
ii) Assume u solves (4.0.1). Then u is a subsolution and supersolution to (4.0.1).
Proposition 4.4.3 Let u, v ∈ C 0 (Ω).
i) Assume u is subharmonic and v is superharmonic in Ω with u ≤ v on ∂Ω. Then

either u < v or u = v in Ω.
ii) Assume u is a subsolution and v is a supersolution to (4.0.1). Then either u < v

or u = v in Ω.
Proposition 4.4.4 Let u be subharmonic in Ω, B a ball with B Ω, and u ∈ C 0 (B) ∩

C 2 (B) such that −Δu = 0 in B and u = u on ∂B. Then U given by
U (x) = u(x), x ∈ B, U (x) = u(x), x ∈ Ω\B,
called “harmonic lifting of u with respect to B”, is also subharmonic in Ω.
Proposition 4.4.5 If u1 , . . . , um are subharmonic functions in Ω then max{u1 , . . . , um }

is also subharmonic in Ω.
It follows that the solution u to (4.0.1) belongs to the intersection of subsolutions

and supersolution to (4.0.1). Furthermore, one may look for it as the supremum of all
subsolutions to (4.0.1), which is the idea behind Perron’s method that we will present
in section 4.4.2.
The following result is key to showing that the supremum of subsolutions to (4.0.1)
is a harmonic function. For its proof, see Theorem 9.3.15 in Annex.
Theorem 4.4.6 Let B ⊂ RN be a ball and (un ) a sequence of uniformly bounded and
harmonic functions in B. Then (un ) has a subsequence converging pointwise in B to a
harmonic function u in B. Furthermore, the convergence is in C 0 (K) for every compact
K ⊂ B.
4.4.2 Solution to the Dirichlet problem

Now we state the main result of this section.
Theorem 4.4.7 (Perron) Let Ω ⊂ RN be an open bounded C 2 domain. Then for every
g ∈ C 0 (∂Ω) there exists a unique u ∈ C 0 (Ω) ∩ C 2 (Ω) solution to (4.0.1).
60
Comments on the proof of the theorem. The proof is classical and can be
found in several books; see, for example, [20, 43]. For the reader’s convenience, we have
included it in section 9.3.3.3 in Annex. Here we give some comments which highlight
the main steps and technique of the proof.
The proof considers separately the problem in the interior, −Δu = 0 in Ω, and the
problem on the boundary, u = g on ∂Ω.
For the problem in the interior, one looks for the solution u of the form
u = sup{v, v subsolution of (4.0.1)}.
By heavily using the maximum principle, it is proved that −Δu = 0 in Ω.

For the problem on the boundary, one uses the maximum principle again, and a
so-called “barrier function” w at ξ ∈ ∂Ω, defined by
|x − y|
w(x) = ln , for N = 2,
R
w(x) = R2−N − |x − y|2−N , for N ≥ 3,
with y and R such that B(y, R) ⊂ Ωc and ∂B(y, R) ∩ ∂Ω = {ξ}. Note that w satisfies
w ∈ C ∞ (RN \{y}), −Δw = 0 in Ω, w > 0 in Ω\{ξ}, w(ξ) = 0. (4.4.1)
Therefore w (resp. −w) is a supersolu-

tion (resp. subsolution) to the problem
y
R
−Δu = 0 in Ω, u = 0 on ∂Ω. ξ Ω
Any function w satisfying (4.4.1) is
called a “barrier function (for −Δ) at ξ
relative to Ω”. Comparing the function
u(x) with g(ξ)±( +kw(x)), with > 0
Figure 4.4.1: Ω, ξ, and B(y, R)
arbitrary, and k > 0 chosen appropri-
ately, one proves
−( + kw(x)) ≤ u(x) − g(ξ) ≤ ( + kw(x)), ∀x ∈ Ω,
which shows that u = g on ∂Ω.
One may wonder whether the barrier function w is just a tool used in the proof of
Theorem 4.4.7, or if it has a fundamental role in the existence of the solution to (4.0.1).
It is also important to know whether the C 2 regularity of Ω can be weakened. Note that
the C 2 regularity of Ω is used when constructing the barrier function w (namely when
choosing the ball B(y, R)) (Figure 4.4.1).
61
A careful look at the proof of Theorem 4.4.7 shows that instead of the definition
(4.4.1) for the barrier function, one can define a weak barrier function w (for −Δ) at ξ
relative to Ω as follows:

(i) w is superharmonic in Ω ∩ B(ξ, Rξ ),
(4.4.2)
(ii) w > 0 in Ω ∩ B(ξ, Rξ )\{ξ}, w(ξ) = 0,
for a certain Rξ > 0. Then the proof of Theorem 4.4.7 remains unchanged.
In this context, we say that a point ξ ∈ ∂Ω is regular with respect to the Laplacian
if there exists a (weak) barrier function (for −Δ) at ξ relative to Ω.
The following theorem explains the connection between barrier functions and the
problem (4.0.1); see [20].
Theorem 4.4.8 Let Ω ⊂ RN be open, bounded. The Dirichlet problem (4.0.1) has a
unique solution u ∈ C 0 (Ω) ∩ C 2 (Ω) for arbitrary g ∈ C 0 (∂Ω) if and only if all ξ ∈ ∂Ω
are regular w.r.t. the Laplacian (i.e. for every ξ ∈ ∂Ω there exists a weak barrier function
for −Δ at ξ relative to Ω).
Note that the only regularity for Ω needed in Theorem 4.4.8 is the existence of
a weak barrier function at every ξ ∈ ∂Ω. This is a much weaker condition than the
assumption “Ω is of class C 2 ”. However, characterizing the domains that have a weak
barrier function at each point of their boundary is not an easy problem, especially in
high dimension. In dimension N = 2 one can show that if ξ ∈ ∂Ω then
1 ln r
w(x) = w(r, θ) = −Re =− 2 ,
ln z θ + ln2 r
with (r, θ) the polar coordinates, x = ξ + (r cos θ, r sin θ), z = reiθ , are a weak barrier at
ξ for −Δ relative to Ω, provided ln z is well-defined near every ξ on ∂Ω. Hence, when
N = 2 the problem (4.0.1) is solvable in Ω for arbitrary g ∈ C 0 (∂Ω), if for example each
point ξ ∈ ∂Ω is the endpoint of a simple arc lying in the exterior of Ω because in this
case one can define a branch of ln z near every ξ ∈ ∂Ω.
In dimension N ≥ 3, the characterization of Ω such that (4.0.1) is solvable is more
difficult. An example given by Lebesgue, see [9], shows that in dimension N = 3 there
are domains with non-regular points, such that (4.0.1) is not solvable for all g ∈ C 0 (∂Ω).
We conclude this chapter by emphasizing that we proved the existence and unique-
ness of (4.0.1) using the method of subsolutions and supersolutions due to Perron. It
provides classical solutions, i.e. u ∈ C 0 (Ω) ∩ C 2 (Ω). This method separates the prob-
lem in the interior from that on the boundary, and can be applied to general linear
second-order elliptic PDEs. It is based heavily on the maximum principle.
62
Another approach to classical solutions is Schauder’s theory, which is a complete and

attractive PDE theory that provides C 2,α solutions to linear second-order elliptic PDEs,
if both Ω and boundary data are of class C 2,α ; see [20] for details.
At first look, one would expect that problem (4.0.1) has a classical solution for all
Ω bounded and open, and g ∈ C 0 (∂Ω). But as we mentioned above, this is not true in
general (Lebesgue’s example; see [9]). Furthermore, there are problems such as
−Δu = 0, in Ω = (−1, 1) × (0, 1), (4.4.3a)

u = ln |1 − ln |x1 ||, on Σ = (−1, 1) × {0}, (4.4.3b)
u = 0, on ∂Ω\Σ, (4.4.3c)
which do not have classical solutions but have “weak solutions” (we will review this
problem in Chapter 7). The method of weak solutions has been developed to pro-
vide solutions to (4.0.1) in the cases where the method of classical solutions fails; see
chapter 7.
Problems
Problem 4.1 Consider the problem
−Δu = f in Ω = B(0, 1) ⊂ R2 ,
u = g on ∂Ω.
i) Let (r, θ) ∈ (0, ∞) × [0, 2π) be the polar coordinates, x1 = r cos θ, x2 = r sin θ.
1 1
Show that Δu = ∂r (r∂r u) + 2 ∂θθ u.
r r
ii) Assume u, f , and g are smooth radially symmetric functions and solve for u.
Problem 4.2 Show that u(x) = ln |x| is a classical solution to Δu = 0 in R2 \{(0, 0)}.
Problem 4.3 For each of the following problems, find a series solution candidate and
analyze the existence of a classical solution as required.
i) uxx + uyy = 0 in Ω = (0, 1) × (0, 2),

u = 0 on ∂Ω\Σ,
u = x(1 − x) on Σ = (0, 1) × {2}.
Show the existence of a classical solution u ∈ C 1 (Ω) ∩ C ∞ (Ω).
ii) uxx + uyy = 0 in Ω = (0, 2) × (0, 1),

u = 0 on ∂Ω\Σ,
u = x1(0,1) + (1 − x)1(1,2) on Σ = (0, 2) × {1}.
63
Problem 4.4 For each of the following problems, find a series solution candidate and
analyze the existence of a classical solution as required.
i) uxx + uyy = 0 in Ω = (0, π) × (0, π),

∂ν u = 0 on ∂Ω\Σ,
∂ν u = y − a on Σ = {π} × (0, π), a = π2 .
ii) uxx + uyy = 0 in Ω = (0, π) × (0, π),

∂ν u = 0 on ∂Ω\Σ,
π2
∂ν u(2, y) = x2 − a on Σ = (0, π) × {π}, a = 3
.
In these problems, ν is the unitary exterior normal vector on ∂Ω, excluding the corners
of the domain.
Problem 4.5 Let α ∈ (0, π), Ω = {(r, θ), r ∈ (0, 1), θ ∈ (0, α)}, γ0 = {(r, 0), r ∈
(0, 1)}, γ = {(1, θ), θ ∈ (0, α)} and γα = {(r, α), r ∈ (0, 1)}. Consider the problem
−Δu = 0 in Ω, u = 0 on γ0 ∪ γα , u = g on γ,
with g(1, θ) = θ(α−θ). Use the method of separation of variables to find a series classical
solution u ∈ C 0 (Ω) ∩ C 2 (Ω).
Problem 4.6 Prove that the problem
−Δu = 0 in Ω := {x ∈ R2 , 0 < |x| < 1}, u = 1 on {|x| = 1}, u(0) = 0,
does not have a C 0 (Ω) ∩ C 2 (Ω) solution.
Problem 4.7 Prove that the problem
−Δu = 1 in Ω := {x ∈ R2 , 0 < |x| < 1}, u = 0 on ∂Ω,
does not have a C 0 (Ω) ∩ C 2 (Ω) solution.
Problem 4.8 Let Ω ⊂ RN be a C 2 bounded open connected set, ν the unit outward
normal vector to ∂Ω, and u ∈ C 1 (Ω) ∩ C 2 (Ω) satisfying
−Δu = 0 in Ω, ∂ν u = 0 on ∂Ω.
Prove that u is constant in Ω.
64
Problem 4.9 Let Ω ⊂ RN be an open set and u ∈ C 0 (Ω). Prove that u is subharmonic
in Ω if and only if

1
u(x) ≤ udσ =: − udσ,
N VN RN −1 ∂BR (x) ∂BR (x)
for all x ∈ Ω and R > 0 such that BR (x) Ω.
Problem 4.10 Let Ω ⊂ RN be an open set and u ∈ C 0 (Ω)∩C 2 (Ω). Prove that −Δu ≤ 0
in Ω if and only if

u(x) ≤ − udσ,
∂BR (x)
for all x ∈ Ω and R > 0 such that BR (x) Ω.
Problem 4.11 Let Ω ⊂ RN be an open set and u ∈ C 0 (Ω) ∩ C 2 (Ω). Prove that u is
subharmonic in Ω if and only if −Δu ≤ 0 in Ω.
Problem 4.12 Let Ω ⊂ RN be an open set. Prove that u ∈ C 0 (Ω)∩C 2 (Ω) and −Δu = 0
in Ω if and only if u ∈ C 0 (Ω) and satisfies the MVT in Ω.
Problem 4.13 Show that MVT in Ω ⊂ RN is equivalent to

1
u(x) = u(y)dy, ∀B(x, r) Ω.
|B(x, r)| B(x,r)
Problem 4.14 Prove the so-called “Liouville’s theorem”, i.e. if u ∈ C 2 (RN ) satisfies
−Δu = 0 in RN , |u| ≤ M < ∞,
for a certain M > 0, then u is constant.
Problem 4.15 Assume Ω ⊂ RN is an open bounded set and 0 = u ∈ C 0 (Ω) ∩ C 2 (Ω)

solving
−Δu = λu in Ω, u = 0 on ∂Ω (λ ∈ R).
Prove that λ > 0.
Problem 4.16 Let Ω = {(r, θ), r ∈ [0, 1), θ ∈ (0, α), α ∈ (0, π)} and u ∈ C 0 (Ω) ∩
C 2 (Ω) be the solution to
−Δu = 0 in Ω, u = g on ∂Ω,
where g ∈ C 0 (∂Ω) with g(r, 0) = g(r, α) = 0, r ∈ [0, 1].
65

i) Assume that |g(1, θ)| ≤ C sin απ θ , for all θ ∈ [0, α]. Use the comparison principle
to show that the u is differentiable at (0, 0) and ∇u(0, 0) = (0, 0).
ii) Prove that the claim in (a) holds even with the condition for g replaced by g(1, θ) ∈
C 1 ([0, α]).
Problem 4.17 Assume Ω+ ⊂ R2+ := {(x1 , x2 ), x2 > 0} is an open bounded set with
Γ0 = [0, a] × {0} = ∂Ω+ ∩ ∂R2+ and
u+ ∈ C 0 (Ω+ ) ∩ C 2 (Ω+ ), with − Δu+ = 0 in Ω+ , u+ = 0 on Γ0 .
Set Ω− = {(x1 , x2 ) s.th. (x1 , −x2 ) ∈ Ω+ } and
u− ∈ C 0 (Ω− ) ∩ C 2 (Ω− ), u− (x1 , x2 ) := −u+ (x1 , −x2 ).
Set furthermore Ω = Ω+ ∪ Γ0 ∪ Ω− and u : Ω → R with u = u+ in Ω+ and u = u− in

Ω− . Prove7,8 that u ∈ C 0 (Ω) ∩ C 2 (Ω) and −Δu = 0 in Ω.
Problem 4.18 Assume Ω ⊂ R2 is a polygon with nodes xn , n = 1, . . . , m, and consider
the problem
−Δu = 1 in Ω, u = 0 on ∂Ω.
i) Prove that this problem has a unique solution u ∈ C 0 (Ω) ∩ C 2 (Ω).

ii) Prove that u ∈ C ∞ (Ω\{x1 , . . . , xm }).
iii) Prove that if Ω is a convex polygon then u is differentiable at xn and |∇u(xn )| =
0, for all n.
Problem 4.19 Assume Ω ⊂ R2 is a regular polygon with nodes xn , n = 1, . . . , m, and
consider the problem
−Δu = 1 in Ω, u = 0 on ∂Ω.
Prove that u ∈ C ∞ (Ω\{x1 , . . . , xm }) and there exist9 M > 0 and points y1 , . . . , ym on

∂Ω such that |∂ν u(yi )| ≥ M , for all i.
Problem 4.20 Assume Ω ⊂ RN is a C 2 bounded open set and u ∈ C 1 (Ω) ∩ C 2 (Ω) solves
−Δu = 0 in Ω, u = g on ΓD ⊂ ∂Ω, ∂ν u + cu = 0 on ΓN = ∂Ω\ΓD ,
with ΓD = ∅ and c > 0. Prove that uC 0 (Ω) ≤ gC 0 (ΓD ) .
7
You may use Proposition 4.2.4.
8
This is the so-called “Schwarz reflection principle”.
9
You may compare u with v ∈ C 0 (B) ∩ C 2 (B) solving −Δv = 1 in B and v = 0 on ∂B, where B in
the ball inscribed to the polygon.
66
5. Distributions
Well before Laurent Schwartz developed the theory of distributions, see [47], physi-
cists and engineers used distributions in an informal way, for example when solving
ordinary differential equations with discontinuous right-hand sides.
Distributions generalize the concept of a function and as such they can represent
very irregular “functions”. At the same time, distributions are “very regular”, as they
do have derivatives of any order. It turns out that these features are powerful tools,
because they allow us to elegantly write and solve differential and partial differential
equations, which otherwise would not even have meaning.
The literature on distributions is rich, and we refer the reader to [18, 22, 23, 47, 50].
5.1 Motivation
Consider the initial value problem yh = fh in R, yh (−∞) := lim yh (t) = 0, where
t→∞
1
, t ∈ (a, a + h),
fh (t) = h with a ∈ R, h > 0.
0, t ∈/ (a, a + h),
t
We look for the limit of yh as h → 0. Clearly, as yh (t) = −∞ f (τ )dτ we get y(t) =
limh→0 yh (t) = H(t − a), t = a, where the limit exists pointwise a.e. and in L1loc (R), and
H(t) is the Heaviside function, H(t) = 1(0,∞) (t).
One would prefer to consider the limit from a functional viewpoint—first consider
the limit of fh and then by using the differential equation deduce the limit for yh . In this
case, we note that lim fh (t) = 0 for all t. If naively we would pass to the limit in yh = fh ,
h→0
we would get y = 0, so y = 0, which is incorrect. If one would accept that δa := lim fh
h→0
exists, let us say in L1 (R), then
necessarily δa = 0. However,
this is impossible because,
as for arbitrary h > 0 we have R fh (t)dt = 1, we obtain R δa (t)dt = 1, which contradicts
δa = 0. So, if the limit δa of (fh ) exists, it cannot be in L1 (R).
One notes that for ϕ ∈ C00 (R), we have

lim fh (t)ϕ(t)dt = ϕ(a).
h→0 R

https://doi.org/10.1007/978-3-031-39524-6 5
5.2. Distributions Chapter 5
This equality serves as a definition for the limit of fh as h → 0. By definition, δa is a

distribution, commonly called “Dirac measure (or Dirac function) at a”. It belongs to
(C00 (Ω)) and is characterized by
δa , ϕ(C00 (Ω)) ×C00 (Ω) = ϕ(a), ∀ϕ ∈ C00 (R).
Here and afterwards, T, ϕV ×V denotes a pairing/duality between T ∈ V and ϕ ∈ V ,

where V is a vector space and V its dual space. Whenever there is no ambiguity, we
will drop the symbol V × V from T, ϕV ×V .
So we expect the limit y of yh to satisfy y = δa , which is to be understood in the
distribution sense, i.e. y , ϕ = ϕ(0), for all ϕ ∈ D(R). This simple example motivates
not only the question of convergence of fh , but also of yh , yh , and the distributional
differential equation y = f . These questions will be answered in a more general context
with the results we will present in this chapter.
5.2 Distributions
In this section, we will introduce the space of test functions as well as the definition
of distributions and of their derivatives, and we will provide a number of examples.
5.2.1 Test functions

We will introduce the class of test functions together with an associated topology,
and some properties related to them.
Definition 5.2.1 Let Ω ⊂ RN be a nonempty open set. A function ϕ ∈ D(Ω), where
D(Ω)1 defined in (1.1.28) is called a “test function in Ω”, and D(Ω) is called “the space
of test functions in Ω”. Sometimes we will write D instead of D(RN ).
For ϕ ∈ D(Ω) we define supp(ϕ) := {x ∈ Ω, ϕ(x) = 0}, the closure of the set {x ∈
Ω, ϕ(x) = 0}, which is compact. We call it “support of ϕ”.
Given a compact set K ⊂ Ω we define DK (Ω) = {ϕ ∈ D(Ω), supp(ϕ) ⊂ K}.
Clearly, the set D(Ω) is a linear space. For the needs of this introduction, it is enough
to define in D(Ω) only the convergence of sequences.
Definition 5.2.2 Let ϕ, ϕn , n ∈ N be elements of D(Ω). We say “ϕn converges to ϕ in
D(Ω)” and write “ lim ϕn = ϕ in D(Ω)”, if
n→∞
(i) ∃K ⊂ Ω with K compact, supp(ϕ) ⊂ K, supp(ϕn ) ⊂ K for all n ∈ N, and

α α
(ii) ∀α ∈ NN
0 , lim D ϕn − D ϕ C 0 (K) = 0.
n→0
1
All along the remainder of this book, unless otherwise specified, the functions of D(Ω) are with
values in C.
68
Chapter 5 5.2. Distributions
We note that D(Ω) is not empty. For example, if ρ is the function defined in Example
1.1.13, then ρ(n(x − x0 )) belongs to D(Ω) provided x0 ∈ Ω and n ∈ N is big enough.
Remark 5.2.3 Let ϕ ∈ D(Ω) and (ϕn ) be a sequence in D(Ω) converging to ϕ in

D(Ω). Then from Definition 5.2.2 it follows that (Dα ϕn ) converges to Dα ϕ in D(Ω),
for arbitrary α ∈ NN
0 .
Using the sequence (ρn ) of Example 1.1.13 and the “convolution” operator, we can
construct infinitely many D(Ω) functions.
Theorem 5.2.4 Let p ∈ [1, ∞] and define the convolution operator ∗ by

∗ : L1 (RN ) × Lp (RN ) → Lp (RN ), (5.2.1)
(u, v) → u ∗ v, (u ∗ v)(x) = RN u(x − y)v(y)dy.
The operator ∗ is well-defined and satisfies
u∗v = v ∗ u ∈ Lp (RN ), u∗v Lp (RN ) ≤ u L1 (RN ) v Lp (RN ) , (5.2.2)

supp(u ∗ v) ⊂ supp(u) + supp(v). (5.2.3)
Theorem 5.2.5 Let (ρn ) be a mollifier sequence. We have
i) If v ∈ Lp (RN ) and p ∈ [1, ∞) then
ρn ∗ v ∈ C ∞ (RN ) ∩ Lp (RN ), and lim ρn ∗ v = v in Lp (RN ). (5.2.4)

n→∞
ii) If v ∈ C k (RN ) and K ⊂ RN is compact then

ρn ∗ v ∈ C ∞ (RN ), Dα (ρn ∗ v) = ρn ∗ Dα v, ∀α ∈ NN
0 , |α| ≤ k, and
(5.2.5)
limn→∞ ρn ∗ v = v in C k (K).
It follows from (5.2.3) and (5.2.5) that for ϕ ∈ D, limn→∞ ρn ∗ ϕ = ϕ in D.

For the proof of the results above and more on the convolution of functions, see, for
example, [5, 20, 34].
5.2.2 Distributions
As we will see, distributions generalize the notion of a function. In the context of
PDEs, they are fundamental to giving a precise meaning to PDEs, which otherwise
would not have one, and to solve them.
Roughly speaking distributions can be very irregular functions. However, they have
derivatives of any order.
Definition 5.2.6 Let Ω ⊂ RN be an open set.
69
1) A distribution in Ω is a map T : D(Ω) → C satisfying
(i) T is linear in the following sense:
∀ϕ1 , ϕ2 ∈ D(Ω), ∀α1 , α2 ∈ C : T, α1 ϕ1 +α2 ϕ2 = α1 T, ϕ1 +α2 T, ϕ2 ,
where T, ϕ denotes the value of T at ϕ,

(ii) T is continuous in the sense:
∀(ϕn ) in D(Ω), lim ϕn = 0 in D(Ω) implies lim T, ϕn = 0.

n→∞ n→∞
2) The set of distributions in Ω is denoted by D (Ω). Often, instead of D (RN ) we

will write D .
3) We say T ∈ D (Ω) is of finite order if

∃m ∈ N, ∀K ⊂ Ω, K compact, ∃C > 0, ∀ϕ ∈ DK (Ω),
(5.2.6)
|T, ϕ| ≤ C ϕ C m (Ω) .
The integer m is called an “order of T ”. Distributions of order zero are called

“measures”.
The interest of distributions of finite order is that they can be identified with the dual
space of C0m (Ω).
Example 5.2.7 Let f ∈ L1loc (Ω) and define

T : D(Ω) → C, T, ϕ = f (x)ϕ(x)dx. (5.2.7)
Ω
Then T ∈ D (Ω). Indeed, the linearity of T is clear. If lim ϕn = 0 in D(Ω) then

n→∞

|T, ϕn | = f (x)ϕn (x)dx ≤ |f (x)||ϕn |dx

Ω Ω
≤ f L1 (K) ϕn C 0 (K)
−−−→ 0,
n→∞
where K ⊂ Ω is a compact such that supp(ϕ) ⊂ K for all n. The

distribution T identifies
the function f uniquely. This follows from the fact that if f (x)ϕ(x)dx = 0, for all
Ω
ϕ ∈ D(Ω), then f = 0 a.e. (for Lebesgue measure); see Lemma 1.3.9.
Example 5.2.7 shows that Definition 5.2.6 extends the notion of L1loc (Ω) functions. Iden-
tifying f with Tf , we write L1loc (Ω) ⊂ D (Ω). The following example shows that the
inclusion is strict.
70
Example 5.2.8 Let δ0 : D(RN ) → C, δ0 , ϕ = ϕ(0), be the distribution called “Dirac
measure”. It is easy to show that δ0 is a distribution of order zero.
Note that for all p ∈ [1, ∞], we have δ0 ∈ / Lp , i.e. there is no δ0 ∈ Lp such that
δ (x)ϕ(x)dx = ϕ(0), for all ϕ ∈ D. Indeed, let us prove this for p = 2 (for the case
RN 0
p arbitrary, see [5]). Assume for the moment that δ0 ∈ L2 . Then there exists ϕn ∈ D,
lim ϕn = δ0 in L2 ; see Theorem 1.3.5. Consider un = ϕn (1 − ηn ), where ηn ∈ D is a
n→∞
{{0}, B(0, 1/n)} cut-off function. Assuming for the moment lim un = δ0 in L2 , we get
n→∞

0 < δ0 2L2 (RN ) = lim δ0 (x)un (x)dx = lim δ0 , un = un (0) = 0,
n→∞ RN n→∞
/ L2 (RN ).
which is a contradiction and proves δ0 ∈
Now we prove lim un = δ0 in L2 (RN ). We have
n→∞

lim ϕn − un 2L2 (RN ) = lim |ϕn ηn | ≤ lim
2
|ϕn |2
n→∞ n→∞ B(0,1/n) n→∞ B(0,1/n)

≤ C lim |ϕn − δ0 | + lim
2
|δ0 | = 0,
2
n→∞ B(0,1/n) n→∞ B(0,1/n)
where we have used the Lebesgue dominated convergence theorem. This implies lim un = δ0
n→∞
in L2 (RN ) because lim ϕn = δ0 in L2 .
n→∞
D (Ω) is a vector space, and has a topology, which for the needs of this introduction
will be characterized by the convergence of sequences as given by the following definition.
Definition 5.2.9 Let T , T1 , T2 in D (Ω), α1 , α2 ∈ C, and k ∈ C ∞ (Ω). We define
1) α1 T1 + α2 T2 ∈ D (Ω) by
α1 T1 + α2 T2 , ϕ = α1 T1 , ϕ + α2 T2 , ϕ.
2) kT ∈ D (Ω) by
kT, ϕ = T, kϕ.
3) Given T and a sequence (Tn ) of distributions in D (Ω), we say “Tn converges to

T in D (Ω)” if limn→∞ Tn , ϕ = T, ϕ, for all ϕ ∈ D(Ω).
It is easy to check that 1), 2) are consistent, while the consistency2 of 3) follows from
Theorem 9.4.5 in Annex.
2
Actually, we have the following stronger result: if limn→∞ Tn , ϕ exists for all ϕ ∈ D(Ω) then T
defined by T, ϕ := limn→∞ Tn , ϕ defines a distribution in D (Ω) and limn→∞ Tn = T in D (Ω).
71
Example 5.2.10 Let (ρn ) be a mollifier sequence; see Example 1.1.13. Then lim ρn = δ0
n→∞
in D because for ϕ ∈ D, by using (5.2.7) we get

ρn , ϕ = r i
1 ρn (x)ϕ(x)dx = Re(ϕ(θn )) + iIm(ϕ(θn )) = ϕ(θn ) −−−→ ϕ(0),
n→∞
B 0,
n

r i 1
where θn , θn ∈ B 0, .
n
Example 5.2.11 Let a ∈ R and Tn = n1(a,a+ 1 ) . Then lim Tn = δa in D (R) because

n n→∞
for ϕ ∈ D(R) we have
1
a+ n
Tn , ϕ = nϕ(x)dx −−−→ ϕ(a) =: δa , ϕ.
a n→∞
Example 5.2.12 The convergence in D (Ω) generalizes (or is consistent with) the con-
vergence in L1 (Ω). Indeed, let (fn ) in L1 (Ω), f ∈ L1 (Ω), lim fn = f in L1 (Ω). Then, if
n→∞
Tn = fn we have lim Tn = T in D (Ω) because for ϕ ∈ D(Ω) we have
n→∞

|Tn , ϕ − T, ϕ| = (f (x) − f (x))ϕ(x)dx ≤ fn − f L1 (Ω) ϕ
n L∞ (Ω)
Ω
−−−→ 0.
n→∞
Note that in general the distributions do not have meaning at any point as they are
defined only in D(Ω). However, it is useful to speak about the support of a distribution.
Definition 5.2.13 Let Ω ⊂ RN be an open set and T, S ∈ D (Ω).
1) We say “T = 0 in G”, with G ⊂ Ω, G open, if T, ϕ = 0 for all ϕ ∈ D(Ω) with

supp(ϕ) ⊂ G.
2) We set N (T ) to be the largest open set in Ω where T = 0, and call it “null set of
T ” (or “zero set of T ”).
The support of T ∈ D (Ω), denoted by supp(T ), is defined by supp(T ) = Ω\N (T ).
3) We say “T = S in D (Ω)” if N (T − S) = Ω, or equivalently supp(T − S) = ∅, or

even T, ϕ = S, ϕ, for all ϕ ∈ D(Ω).
Example 5.2.14 We have N (δ0 ) = R\{0} and supp(δ0 ) = {0}.
The following theorem shows that a distribution with compact support is of finite
order, i.e. it belongs to the dual space of C0m (Ω) (see 3) in Definition (5.2.6), for a cer-
tain m ∈ N0 .
72
Theorem 5.2.15 Let Ω ⊂ RN be open and T ∈ D (Ω) be with compact support. Then
T is of finite order, i.e. T satisfies (5.2.6).
Proof. See Theorem 9.4.6 in Annex.
This theorem shows that any T ∈ D (Ω) restricted in DK (Ω), K ⊂ Ω compact, is of

finite order.
5.2.3 Derivatives of distributions

In general, a distribution is not a function. At the same time, a distribution is very
regular as it has derivatives of any order as defined below.
Definition 5.2.16 Let T ∈ D (Ω), Ω ⊂ RN be open. The derivative of T with respect to
xi , i = 1, . . . , N , or the xi -partial derivative of T , denoted by ∂xi T , is defined by
∂xi T, ϕ = −T, ∂xi ϕ, ∀ϕ ∈ D(Ω). (5.2.8)
If ϕ = ϕ(x1 , x2 , . . . , xN ), resp. ϕ = ϕ(x, y, . . . , z), sometimes instead of ∂x1 T , ∂x2 T , . . .

we will write ∂1 T , ∂2 T , . . ., resp. ∂x T , ∂y T , . . ..
In general, if α ∈ NN α α
0 , D derivative of T , denoted by D T , is defined by
Dα T, ϕ = (−1)|α| T, Dα ϕ, ∀ϕ ∈ D(Ω). (5.2.9)
We note that Definition 5.2.16 is motivated by the usual derivative of smooth func-
tions. For example, if T = f ∈ C 1 (Ω) then

∂i T, ϕ = ∂i f (x)ϕ(x)dx = − f (x)∂i ϕ(x)dx = −T, ∂i ϕ,
Ω Ω
where we have integrated it by parts in Ω. Hence ∂i T = ∂i f .

Also, Dα T of (5.2.9) is well-defined in D (Ω). Indeed, Dα T is linear. It is also con-
tinuous because if (ϕn ) is a sequence in D(Ω) converging to 0 in D(Ω), then (Dα ϕn )
converges also to 0 in D(Ω), and therefore
Dα T, ϕn = (−1)|α| T, Dα ϕn −−−→ 0.

n→∞
Example 5.2.17 Let α ∈ NN |α| α

0 and for ϕ ∈ D set T, ϕ = (−1) D ϕ(0). Then T =
α
D δ0 , supp(T ) = {0}, and T is a distribution of order |α|.
Example 5.2.18 Let T = H(x) = 1(0,∞) (x), x ∈ R, be the Heaviside function. Then
∞

T , ϕ = −T, ϕ = − H(x)ϕ (x)dx = − ϕ (x)dx = ϕ(0)
R 0
= δ0 , ϕ.
So, H = δ0 (here denotes the derivative w.r.t. x).
73
Example 5.2.19 Let T = ln |x| = ln r, x ∈ R2 , r = |x|. Then

ΔT, ϕ := Txx + Tyy , ϕ = T, Δϕ

= ln |x|Δϕ(x)dx
RN

= lim+ ln |x|Δϕ(x)dx
→0 {|x|>}

= lim+ (ϕ∂r ln |x| − ln |x|∂r ϕ)dσ + ϕ(x)Δ ln |x|dx
→0 {|x|=} {|x|>}

1
= lim+ ϕ − ln ∂r ϕ dσ
→0 {|x|=}
= 2πϕ(0); hence

1
−Δ − ln |x| = δ0 in D (R2 ). (5.2.10)
2π
n
Example 5.2.20 Let T (x) = ln |x|, x ∈ R. As 1
ln rdr = 1 + n(ln n − 1) it follows
1
T ∈ L1loc (R) and so T ∈ D (R). Clearly T ∈ D (R) exists. However, (ln |x|) = and
x
1
/ D (R). So, necessarily T = (ln |x|) . Let us compute T . For ϕ ∈ D(R), we have
∈
x

T , ϕ = −T, ϕ = − ln |x|ϕ (x)dx = − lim+ ln |x|ϕ (x)dx
R →0 {|x|>}

ϕ(x)
= lim+ (ϕ( ) − ϕ(− )) ln + dx
→0 {|x|>} x

ϕ(x)
= lim+ dx.
→0 {|x|>} x
We define pv(x−1 ) ∈ D (R) by

−1 ϕ(x)
pv(x ), ϕ = lim+ dx. (5.2.11)
→0 {|x|>} x
So we have proved
(ln |x|) = pv(x−1 ), ϕ in D (R). (5.2.12)
The definition (5.2.11) defines pv(x−1 ) ∈ D (R) because for ϕ ∈ D(R) we have

−1 ϕ(x) − ϕ(0) ϕ(x) − ϕ(0)
pv(x ), ϕ = lim+ dx = dx, so
→0 {<|x|<M } x {|x|<M } x
|pv(x−1 ), ϕ| ≤ (2M ) ϕ C 1 (R) ,
74
where M = M (ϕ) > 0 is such that supp(ϕ) ⊂ [−M, M ]. It follows furthermore that
pv(x−1 ) is of order one (so, pv(x−1 ) is not a measure).
The distribution pv(x−1 ) is called a “principal value of x−1”. Its name is motivated
ϕ(x)
by the fact that it corresponds to the Cauchy principal value of dx. Finally, note
R x
that pv(x−1 ) is equivalently defined by

−1 ϕ(x) ϕ(x) − ϕ(0)
pv(x ), ϕ = dx + dx, ∀ > 0. (5.2.13)
{|x|>} x {|x|<} x
Example 5.2.21 Let Ω ⊂ RN be open, simply connected, and look for T ∈ D (Ω):
∇T := (∂i T ) = 0, i.e. T, ∂i ϕ = 0 for every ϕ ∈ D(Ω). (5.2.14)
We show that T is equal to a constant. For simplicity and without loss of generality, we
assume Ω = (a1 , b1 ) × · · · × (aN , bN ) =: I1 × · · · × IN .
We note that a function φ ∈ D(Ω) is the ∂i derivative of a certain function ϕ ∈ D(Ω)
if and only if
bi
φ(x1 , . . . , xi−1 , xi , xi+1 , . . . , xN )dxi = 0.
ai
One direction of this statement is clear and the other follows by considering
xi
ϕ(x1 , . . . , xi−1 , xi , xi+1 , . . . , xN ) = φ(x1 , . . . , xi−1 , t, xi+1 , . . . , xN )dt.
ai

Next we choose ηi ∈ D(Ii ), with Ii ηi (xi )dxi = 1, for all i = 1, . . . , N . Then for every
ϕ =: ϕ1 ∈ D(Ω) and x = (x1 , . . . , xN ) ∈ Ω define ϕ2 and φ1 by
b1
ϕ2 (x2 , . . . , xN ) := ϕ1 (x1 , x2 , . . . , xN )dx1 ,
a1
ϕ1 (x1 , . . . , xN ) = φ1 (x1 , . . . , xN ) + η1 (x1 )ϕ2 (x2 , . . . , xN ).
b
It is easy to show that η1 ϕ2 ∈ D(Ω), φ1 ∈ D(Ω), and a11 φ1 (x1 , x2 , . . . , xN )dx1 = 0. So
φ1 is equal to the ∂1 derivative of a test function in Ω. Therefore T, φ1 = 0 and
T, ϕ1 = T, η1 ϕ2 .
We proceed with ϕ2 as with ϕ1 . So we consider ϕ3 and φ2 defined by
b2
ϕ3 (x3 , . . . , xN ) = ϕ2 (x2 , x3 , . . . , xN )dx2 ,
a2
ϕ2 (x2 , . . . , xN ) = φ2 (x2 , . . . , xN ) + η2 (x2 )ϕ3 (x3 , . . . , xN ).
Therefore we have
η1 ϕ2 = η1 φ2 + η1 η2 ϕ3 .
75
5.3. Convolution of distributions and fundamental solutions Chapter 5
b
Clearly η1 η2 ϕ3 ∈ D(Ω), η1 φ2 ∈ D(Ω), and a22 η1 (x1 )φ2 (x2 , x3 , . . . , xN )dx2 = 0. There-
fore η1 φ2 is equal to the ∂2 derivative of a test function in Ω. So T, η1 φ2 = 0 and
T, η1 ϕ2 = T, η1 η2 ϕ3 .
Continuing in this way we get
T, ϕ = T, ϕ1
= T, η1 ϕ2
..
.
= T, η1 η2 · · · ηN ϕN +1
b
= T, η1 (x1 ) · · · ηN (xN ) aNN ϕN (xN )dxN
b b −1
= T, η1 (x1 ) · · · ηN (xN ) aNN aNN−1 ϕN −1 (xN −1 , tN )dtN −1 dxN
..
.
b b −1 b
= T, η1 (x1 ) · · · ηN (xN ) aNN aNN−1 · · · a11 ϕ1 (x1 , . . . , xN −1 , xN )dx1 · · · dxN −1 dxN
= T, η1 (x1 ) · · · ηN (xN )1, ϕ
= T, η1 (x1 ) · · · ηN (xN ) 1, ϕ
= T, η1 (x1 ) · · · ηN (xN ), ϕ ,
which proves T = T, η1 (x1 ) · · · ηN (xN ) ∈ C.
5.3 Convolution of distributions and fundamental

solutions
All throughout this section we consider the space of test functions, resp. distributions,
in RN , which is denoted by D, resp. D . We have seen the convolution operator ∗ is
L2 (RN ) × Lp (RN ); see (5.2.1). In this section we will extend it in D × D and use it to
solve certain PDEs. For more details, we refer the reader to [18, 22, 23].
Definition 5.3.1 Let T, S ∈ D with at least one of them, for example S, with compact
support. Then we define T ∗ S ∈ D , called “convolution of T with S”, by
T ∗ S, ϕ = T (x), S(y), ϕ(x + y), ∀ϕ ∈ D. (5.3.1)
We note that the definition of T ∗ S, ϕ is motivated by the convolution of functions.

For example, if T = f and S = g with f, g ∈ L1 , then (see (5.2.1))

T ∗ S, ϕ = (T ∗ S)(z)ϕ(z)dz = f (z − y)g(y)ϕ(z)dydz
RN RN RN
76
Chapter 5 5.3. Convolution of distributions and fundamental solutions

= f (x)g(y)ϕ(x + y)dydx = f (x), g(y), ϕ(x + y).
RN RN
Also, the definition (5.3.1) is consistent, i.e. T ∗ S ∈ D ; see Proposition 9.4.10 in Annex.
Example 5.3.2 Let T ∈ D . Then T ∗ δ0 = T because for every ϕ ∈ D we have
T ∗ δ0 , ϕ = T (x), δ0 (y), ϕ(x + y) = T (x), ϕ(x).
The following theorem shows the density of D in D and that the convolution is
commutative. The proof uses the lemmas 9.4.8 and 9.4.9 in section 9.4.4, which show
that the differentiation and the integration commute with distribution pairing.
Theorem 5.3.3 For every T ∈ D there exists (Tn ) in D converging to T in D , i.e.
D ⊂ D densely. (5.3.2)
Furthermore, if T, S ∈ D , with for example S with compact support, then

T ∗ S = S ∗ T, (5.3.3)
D (T ∗ S) = Dα T ∗ S = T ∗ Dα S,
α
∀α ∈ NN
0 . (5.3.4)
Proof. For (5.3.2), consider η ∈ D a {B(0, 1), B(0, 2)} cut-off function, and set ηn (x) =
η(x/n). It is clear that ηn T has compact support and lim (ηn T ) = T in D . Therefore,
n→∞
without loss of generality we may assume T has compact support. Set Tn = T ∗ ρn ,
where (ρn ) is a mollifier sequence. Then
Tn , ϕ = T ∗ ρn , ϕ = T (x), ρn (y), ϕ(x + y)

= T (x), ρ̌n ∗ ϕ
−−−→ T, ϕ,
n→∞
where we used the fact that if (ρn ) is a mollifier sequence then (ρ̌n ), with ρ̌n (x) = ρn (−x),
is also a mollifier sequence and lim ρ̌n ∗ ϕ = ϕ in D (see Theorem 5.2.5).
n→∞
For the equality (5.3.3) we note that from Remark 9.4.7 we have S = ηS, with a
certain η ∈ D. Then considering Tn = T ∗ ρn , using Lemma 9.4.8 and Lemma 9.4.9
implies
T ∗ S, ϕ = T (x), S(y), ϕ(x + y)

= T (x), S(y), η(y)ϕ(x + y)
= lim Tn (x), S(y), η(y)ϕ(x + y)
n→∞

= lim Tn (x)S(y), η(y)ϕ(x + y)dx
n→∞ RN

= lim S(y), Tn (x)η(y)ϕ(x + y)dx
n→∞ RN
77
= lim S(y), (T ∗ ρn )(x), η(y)ϕ(y + x)

n→∞
= lim S(y), T (x), ρn (z), η(y)ϕ(y + x + z).
n→∞
Now we use the fact limn→∞ ρn (z), ϕ(y + x + z) = ϕ(y + x) in D, see the comment after
Theorem 5.2.4, which by using Lemma 9.4.8 implies limn→∞ T (x), ρn (z), η(y)ϕ(y + x +
z) = T (x), η(y)ϕ(y + x) in D, to conclude that
T ∗ S, ϕ = S(y), T (x), η(y)ϕ(y + x)
= η(y)S(y), T (x), ϕ(y + x)
= S ∗ T, ϕ.
Finally (5.3.4) follows from Lemma 9.4.8 as follows:
Dα T ∗ S, ϕ = (−1)|α| T (x), Dxα S(y), ϕ(x + y) = (−1)|α| T (x), S(y), Dxα ϕ(x + y)
= (−1)|α| T (x), S(y), Dyα ϕ(x + y) = T (x), Dyα S(y), ϕ(x + y)
= T ∗ Dα S, ϕ,
By using the convolution and the so-called “fundamental solutions”, one can solve
linear PDEs with constant coefficients as follows.
Definition 5.3.4 Let L be a m-th order linear PDE operator with constant coefficients,
L= cα Dα , cα ∈ C.
|α|≤m
A fundamental solution to L is a distribution E ∈ D such that
LE := cα Dα E = δ0 . (5.3.5)
|α|≤m
As a corollary of Theorem 5.3.3, we have the following result.

Theorem 5.3.5 Let L be a m-th order linear PDE operator with constant coefficients,
and let E ∈ D be a fundamental solution to L. For f ∈ D , if at least one of E and f
has compact support, then
L(f ∗ E) = f ∗ LE = f. (5.3.6)
So u = f ∗ E solves Lu = f in D .
Proof. The proof follows directly from Theorem 5.3.3 and Example 5.3.2:
L(f ∗ E) = f ∗ LE = f ∗ δ0 = f.

78
Chapter 5 5.3. Convolution of distributions and fundamental solutions
Example 5.3.6 We have seen that if u ∈ C 2 (Ω) with Δu = 0 in Ω then u ∈ C ∞ (Ω).

We have a stronger result: if u ∈ L1 (RN ) and Δu = 0 in D (RN ) then u ∈ C ∞ (RN ).
Indeed, let un = ρn ∗ u, where (ρn ) is a mollifier sequence. Then from theorems 5.2.4
and 5.2.5 we have (un ) is uniformly bounded in L1 (RN ), un ∈ C ∞ (RN ), limn→∞ un = u
in L1 (RN ), and Δun = ρn ∗ Δu = 0. Also, from Problem 4.13 it follows that (un ) is
uniformly bounded. Then (un ) satisfies the conditions of Theorem 4.4.6 in every ball in
RN , which implies that u ∈ C ∞ (RN ).
As another application of the convolution and the fundamental solution to a PDE,

we will present two examples involving heat and wave equations in dimension one. Here,
calculus are without rigor, and for a rigorous proof we refer the reader to any specialized
monograph on distributions, for example [3, 22].
Example 5.3.7 Let G(·, t) ∈ C 0 ([0, ∞); D (R)) ∩ C 1 ((0, ∞); D (R)) (see footnote3 ) sat-
isfying
∂t G(·, t) − ΔG(·, t) = 0 for t > 0, G(·, 0) = δ0 (·) in D (R). (5.3.7)
The distribution G is called a “fundamental solution to the heat equation”. Let u be given
by
t
u(·, t) = G(·, t) ∗ u0 (·) + G(·, t − s) ∗ f (·, s)ds, in D (R). (5.3.8)
0
Then u(·, t) ∈ C 1 ((0, ∞); D (R))∩C 0 ([0, ∞); D (R)) (see footnote3 ), provided u0 ∈ D (R)
and f (·, t) ∈ C 0 ((0, ∞); D (R)) are with compact support. Furthermore, u solves4
∂t u(·, t) − Δu(·, t) = f (·, t) for t > 0, u(·, 0) = u0 (·) in D (R). (5.3.9)
Indeed, by direct computations we obtain

t
∂t u(·, t) − Δu(·, t) = ∂t G(·, t) ∗ u0 + G(·, 0) ∗ f (·, t) + ∂t G(·, t − s) ∗ f (·, s)ds
t 0
− ΔG(·, t) ∗ u0 − ΔG(·, t − s) ∗ f (·, s)ds

0
= (∂t G(·, t) − ΔG(·, t)) ∗ u0
+ δ0 ∗ f (·, t)
t
+ (∂t G(·, t − s) − ΔG(·, t − s)) ∗ f (·, s)ds
0
= f (·, t),
3
This means (i) t ∈ [0, ∞) → G(·, t) ∈ D (R) is continuous, which is limh→0 G(·, t + h) = G(·, t) in
D (R), and (ii) ∂t G(·, t) = limh→0 (G(·, t + h) − G(·, t))/h exists in D (R) and t ∈ (0, ∞) → ∂t G(·, t) ∈

D (R) is continuous at every t > 0.

4
The formula for u in (5.3.8) holds also in D (RN ) and u solves (5.3.9) in D (RN ).
79
u(·, 0) = u0 (·).
One can solve G in (5.3.7) and find (see Example 5.4.18)

1 − 1 x2
G(x, t) = √ e 4t . (5.3.10)
4πt
Example 5.3.8 Let G = G(·, t) ∈ C 2 ((0, ∞); D (R)) ∩ C 1 ([0, ∞); D (R)) the solution to
∂tt G(·, t)−ΔG(·, t) = 0 for t > 0, G(·, 0) = 0, ∂t G(·, 0) = δ0 (·), in D (R). (5.3.11)
G be called a “fundamental solution to the wave equation”. Now we define

t
u(·, t) = ∂t G(·, t)∗ u0 (·) +G(·, t)∗u1 (·) + G(·, t−s)∗ f (·, s)ds, in D (R). (5.3.12)
0
If u0 , u1 ∈ D (R) and f (·, t) ∈ C 0 ((0, ∞); D (R)) have compact support then u(·, t) ∈
C 2 ((0, ∞); D (R)) ∩ C 1 ([0, ∞); D (R)), and solve
∂tt u(·, t) − Δu(·, t) = f (·, t) for t > 0, u(·, 0) = u0 , ∂t u(·, 0) = u1 . (5.3.13)
Indeed by direct computations we obtain
utt (·, t) − Δu(·, t) = ∂ttt G(·, t) ∗ u0 (·) + ∂tt G(·, t) ∗ u1 (·)

t
+ ∂t G(·, 0) ∗ f (·, t) + ∂tt G(·, t − s) ∗ f (·, s)ds
0
t
− Δ∂t G(·, t) ∗ u0 (·) − ΔG(·, t) ∗ u1 (·) − ΔG(·, t − s) ∗ f (·, s)ds
0
= ∂t (∂tt G(·, t) − ΔG(·, t)) ∗ u0 (·) + (∂tt G(·, t) − ΔG(·, t)) ∗ u1 (·)
t
+ ∂t G(·, 0) ∗ f (·, t) + (∂tt G(·, t − s) − ΔG(·, t − s)) ∗ f (·, s)ds
0
= f (·, t),
u(·, 0) = u0 , ∂t u(·, 0) = u1 .
The solution to (5.3.11) is given by (see Example 5.4.18)

1 1
G(x, t) = (δ0 (x + t) + δ0 (x − t))dt = (H(x + t) − H(x − t))
2 2
1
2
, for |x| < t,
= (5.3.14)
0, for |x| > t.
80
Chapter 5 5.4. Tempered distributions and Fourier transform
5.4 Tempered distributions and Fourier transform

The set S of tempered distributions consists of linear continuous maps in S, where
S is the set of infinitely differentiable functions with rapid decay (at infinity). As such,
a tempered distribution generalizes the concept of distribution. Fourier transform F,
which is initially defined in S, is invertible in S and S , and transforms a PDE with
constant coefficient to a polynomial. These properties make F a powerful tool for the
study of PDEs.
Definition 5.4.1 Let C ∞ denote the set of continuous infinitely differentiable functions
in RN with complex values.
(i) Define S, sometimes denoted by S(RN ) to avoid any ambiguity, by

⎧ ⎫
⎨ ⎬
∞ α β
S = ϕ ∈ C , ∀m ∈ N, sm (ϕ) := x D ϕ C 0 (RN ) < ∞ .
⎩ ⎭
|α|≤m, |β|≤m
The set S is called a “space of test functions with fast decay (at ∞)”.
(ii) If ϕn , ϕ ∈ S, n ∈ N, we say “ϕn converges to ϕ in S”, and write “ lim ϕn = ϕ in

n→∞
S”, if for every m ∈ N we have
lim sm (ϕn − ϕ) = 0.
n→∞
Note that S is a vector space because α1 ϕ1 + α2 ϕ2 ∈ S for arbitrary α1 , α2 complex

constants and ϕ1 , ϕ2 ∈ S.
2 2
Example 5.4.2 Let u(x) = e−x and v(x) = ex . Then u ∈ S and v ∈
/ S.
Definition 5.4.3 The space of tempered distributions and the convergence in it are
defined as follows.
(i) A tempered distribution is a map T : S → C satisfying
1) T is linear
∀ϕ1 , ϕ2 ∈ S, ∀α1 , α2 ∈ C : T, α1 ϕ1 + α2 ϕ2 = α1 T, ϕ1 + α2 T, ϕ2 ,
2) T is continuous in the sense
∀(ϕn ) ∈ S, lim ϕn = 0 in S implies lim T, ϕn = 0.

n→∞ n→∞
The set of tempered distributions in RN is denoted by S , or to avoid any ambiguity

by S (RN ).
81
5.4. Tempered distributions and Fourier transform Chapter 5
(ii) For T, S ∈ S we say “T = S” if T − S, ϕ = 0 for all ϕ ∈ S.
(iii) For T, Tn ∈ S , n ∈ N, we say “Tn converges to T in S ”, and write “ lim Tn = T

n→∞
in S ”, if 5
lim Tn − T, ϕ = 0, for all ϕ ∈ S.

n→∞
The following proposition shows that S is invariant with respect to the derivative of
any order and multiplication by any polynomial. Furthermore, the inclusions D ⊂ S ⊂
Lp ⊂ S ⊂ D , p ∈ [1, ∞), hold and are dense.6
Proposition 5.4.4 Let α, β ∈ NN
0 and p ∈ [1, ∞). Then
i) {xα Dβ u, u ∈ S} ⊂ S and, (5.4.1)

ii) D ⊂ S ⊂ Lp ⊂ S ⊂ D , (5.4.2)
with all the inclusions dense, the third inclusion being dense even for p = ∞.
Proof. See Proposition 9.4.11 in Annex.
5.4.1 Fourier transform

We now introduce the Fourier transform, which we will use for solving certain PDEs
and while studying Sobolev spaces.
Definition 5.4.5 Let u ∈ L1 (RN ). Fourier transform of u, denoted by F[u] or û, is
defined by

N 1
F : R → C, F[u](ξ) := e−i(ξ·x) u(x)dx, ∀ξ ∈ RN . (5.4.3)
(2π)N/2 RN
Here i2 = −1 and ξ · x = ξn xn .
n=1,N
The following properties in S are very important. In plain terms, they show that
after applying F, every operator Dα u becomes ξ α û in Fourier space. We will see that
this property holds for tempered distributions as well and leads to the conclusion that a
linear partial differential operator with constant coefficients becomes a linear equation
with a polynomial coefficient in Fourier space; see Remark 5.4.12.
5
This definition is consistent, because actually the following holds: Assume limn→∞ Tn , ϕ exists
for all ϕ ∈ S and define T : S → C, T, ϕ := limn→∞ Tn , ϕ. Then T ∈ S and Tn → T in S . See [22].
6
If A and B are sets and B is endowed with a convergence of sequences, we say “the inclusion A ⊂ B
is dense” if A ⊂ B and for every b ∈ B there exists a sequence (an ) in A converging to b in B.
82
Proposition 5.4.6 For all u ∈ S and α, β ∈ NN

0 we have F[u] ∈ S and
Dα F[u](ξ) = F[(−ix)α u](ξ), (5.4.4)

(iξ)β F[u](ξ) = F[Dβ u](ξ), (5.4.5)
(iξ)β Dα F[u](ξ) = F[Dβ ((−ix)α u)](ξ). (5.4.6)
Proof. Let u ∈ S. From (5.4.1) we have xα Dβ u ∈ L1 for all α, β ∈ NN 0 . Hence, all
the terms on the right-hand sides of (5.4.4), (5.4.5), and (5.4.6) are well-defined, and in
particular F[u] is well-defined.
To prove (5.4.4), for α = (1, 0, . . . , 0) for example, one can write

1 1
lim (F[u](ξ + he1 ) − F[u](ξ)) = lim e−i(ξ·x) (e−ih(e1 ·x) − 1)u(x)dx
h→0 h (2π)N/2 h→0 RN

1
−−→ e−i(ξ·x) (−ix1 )u(x)dx
h→0 (2π)N/2 RN
= F[(−ix)α u](ξ),
where for the limit we used Lebesgue DCT. The proof for arbitrary α is made by
recurrence. For the proof of (5.4.5) let (un ) be a sequence in D converging to u in S and
in L1 ; see Proposition 5.4.4. Then by passing in limit, integrating by parts, and passing
in limit again gives

β β 1
F[D u](ξ) = lim F[D un ](ξ) = N/2
lim e−i(ξ·x) Dβ un (x)dx
n→∞ (2π) n→∞ RN
|β|
(−1)
= lim Dβ e−i(ξ·x) un (x)dx
(2π)N/2 n→∞ RN x

1
= lim (iξ)β e−i(ξ·x) un (x)dx
(2π)N/2 n→∞ RN
= (iξ)β F[u](ξ).
Finally, (5.4.6) is proved by combining (5.4.4), (5.4.5). It implies F[u] ∈ S because the
right-hand side of (5.4.6) is bounded
1
F[Dβ ((−ix)α u)](ξ) ≤ Dβ ((−ix)α u) L1 (RN ) < ∞.
(2π)N/2
2
Example 5.4.7 Given a ∈ R, a = 0, we will find F[e−ax ], x ∈ RN . First consider
2
u(x) = e−x , x ∈ RN and let us find û. Note that u ∈ S, so û ∈ S. In the case N = 1,
by using Proposition 5.4.6 we get
2xu + u = 0, 2û + ξ û = 0.

− 14 ξ 2 1 2
Hence, û = Ae with A = û(0) = 1/2
e−x dx. Note that
(2π) RN
1/2 2π ∞ 1/2
−x2 −(x2 +y 2 ) −r 2
e dx = e dxdy = re drdθ = π 1/2 . (5.4.7)
R R2 0 0
83
1
Therefore, A = and
21/2
2 1 1 2
F[e−x ](ξ) = e− 4 ξ , ξ ∈ R. (5.4.8)
21/2
In the case N ∈ N, taking into account (5.4.8) we have
1
−x2 1 −i(ξk xk ) −x2k −x2k − 14 ξk2
F[e ](ξ) = e e dx k = F[e ](ξk ) = e .
k=1,N
(2π)1/2 R k=1,N k=1,N
21/2
Hence
2 1 1 2
F[e−x ](ξ) = e− 4 ξ , ξ ∈ RN . (5.4.9)
2N/2
√
Finally, by using (5.4.9) and the change of variable y = ax, a > 0, we get
2 1 1 2
F[e−ax ](ξ) = N/2
e− 4a ξ . (5.4.10)
(2a)
One may ask whether F[u] ∈ D for any u ∈ D. By using the analyticity of F[u],
one can prove that F[u] ∈ D if and only if u = 0. The introduction of S has been
motivated, among others, by the idea of finding a non-trivial space of test functions
which is invariant under F. It turns out that F[S] = S, as the following theorem shows.
Theorem 5.4.8 Fourier transform F is a linear continuous7 bijection from S to itself.

Its inverse, denoted by F−1 , is given by

−1 1
F [U ](x) = ei(x·ξ) U (ξ)dξ = F[U ](−x) = F[Ǔ ](x), ∀U ∈ S, (5.4.11)
(2π)N/2 RN
where w̌(ξ) = w(−ξ) for every function w. Furthermore, similar to (5.4.4), (5.4.5), for
all u ∈ S we have
Dα F−1 [u](x) = F−1 [(iξ)α u](x), (5.4.12)
(−ix)β F−1 [u](x) = F−1 [Dβ u](x). (5.4.13)
Proof. See Theorem 9.4.13. See Corollary 5.4.15 for another proof of (5.4.11).
The following theorem is important because it shows that Fourier transform preserves
the inner product in L2 . The action of F in Lp spaces is more complex; see Lemma 9.4.14
and Theorem 9.4.15 in Annex for more.
7
For the topology induced by the seminorms sm , in the sense limn→∞ un = u in S implies
limn→∞ F[un ] = F[u] in S; see Definition 5.4.1.
84
Theorem 5.4.9 (Parseval’s theorem) Let u, v ∈ S. Then (u, v) = (û, v̂), where (·, ·)
is the inner product in L2 . Furthermore, Fourier transform F is extended continuously
in L2 and it defines an isometry from L2 to itself, i.e. for all u ∈ L2 , u L2 = û L2 .
Proof. For u, v ∈ S we have

1
(u, v)L2 = u(x)v(x)dx = N/2
ei(x·ξ) û(ξ)v(x)dξdx
RN (2π) RN RN

1
= v̂(ξ) N/2
e−i(ξ·x) u(x)dxdξ
RN (2π) RN
= (û, v̂)L2 . (5.4.14)
As S ⊂ L2 with dense inclusion, it follows that F extends in L2 and (5.4.14) holds in L2 .
5.4.2 Tempered distributions and Fourier transform

The definition of Fourier transform of tempered distributions is motivated by the
simple case when T ∈ L1 . Indeed, from Lemma 9.4.14 we have F[T ] ∈ C 0 ∩ L∞ , and
with ϕ ∈ S we obtain

1
F[T ], ϕ = F[T ](ξ)ϕ(ξ)dξ = N/2
ϕ(ξ) e−i(ξ·x) T (x)dxdξ
N (2π) RN RN
R
1
= T (x) N/2
e−i(x·ξ) ϕ(ξ)dξdx = T, F[ϕ].
RN (2π) RN
Definition 5.4.10 For T ∈ S , the Fourier transform of T , denoted by F[T ], is the

tempered distribution defined by
F[T ], ϕ = T, F[ϕ], ∀ϕ ∈ S. (5.4.15)
Note that Theorem 5.4.8 ensures that F[T ] ∈ S . Note also that in some texts F[T ]
is defined by F[T ], ϕ = T, F−1 [ϕ], which is motivated by the multiplication by the
complex conjugate.
Remark 5.4.11 Similar to Theorem 5.4.8, F maps continuously8 S onto itself, and
formulas (5.4.4) and (5.4.5) hold for every T ∈ S . Also, F−1 : S → S , the inverse of
F in S , is continuous and invertible from S to itself, and is given by
F−1 [S], ϕ = S, F−1 [ϕ], ∀S ∈ S . (5.4.16)
Furthermore, (5.4.12) and (5.4.13) hold in S .

8
In the sense limn→∞ un = u in S implies limn→∞ F[un ] = F[u] in S ; see Definition 5.4.3.
85
Remark 5.4.12 The properties (5.4.4) and (5.4.5) are useful when solving partial dif-
ferential equations. For example, consider the following PDE:
L(D)u := cα D α u = f in S ,
|α|≤m
with cα ∈ C. By taking Fourier transform of both sides of this equation, we obtain
fˆ
û cα (iξ)α = fˆ, so û(ξ) = α
=: ĝ(ξ).
|α|≤m |α|≤m cα (iξ)
If ĝ ∈ S , so ĝ = F[g], for a certain g ∈ S , then u = F−1 [ĝ] = g. This formula for u

serves as a criteria for classifying PDEs with constant coefficients, because, given f , the
existence and the regularity of u depends uniquely on the operator L (or coefficients cα )
α
of the PDE, or equivalently on the polynomial |α|≤m cα (iξ) .
Example 5.4.13 Let a ∈ RN and δa ∈ S defined by δa , ϕ = ϕ(a), for all ϕ ∈ S. Then
1 1
F[δa ] = N/2
e−i(a·ξ) , F[δ0 ] = , (5.4.17)
(2π) (2π)N/2
because

1
F[δa ], ϕ = δa , F[ϕ] = F[ϕ](a) = e−i(a·x) ϕ(x)dx
(2π)N/2 RN
e−i(a·x)
= ,ϕ .
(2π)N/2
Example 5.4.14 We have
F[1] = (2π)N/2 δ0 . (5.4.18)
Indeed, from the second equality of (5.4.17) we get F[(2π)N/2 δ0 ] = 1, which implies
F[1], ϕ(x) = 1, F[ϕ(x)] = 1, F−1 [ϕ(−x)] = F−1 [1], ϕ(−x)
= (2π)N/2 δ0 , ϕ(−x) = (2π)N/2 ϕ(0).
By using Example 5.4.14 we get another (simple) proof of the inverse Fourier transform
F−1 in S.
Corollary 5.4.15 F−1 : S → S is given by (5.4.11).
86
Proof. Indeed, let u ∈ S and û = F[u]. For z ∈ RN we set

−1 1
F [û](z) := ei(z·ξ) û(ξ)dξ.
(2π)N/2 RN
If we prove that F−1 [û](z) = u(z) then we have proved the formula (5.4.11). Note that

i(ξ·z) 1 −i(ξ·(x−z)) 1
e û(ξ) = N/2
e u(x)dx = N/2
e−i(ξ·x) u(x + z)dx
(2π) RN (2π) RN
= F[u(· + z)](ξ).
Therefore from (5.4.18) we obtain

1 1
F−1 [û](z) = N/2
1, F[u(· + z)](ξ) = F[1], u(· + z)) = δ0 , u(· + z)
(2π) (2π)N/2
= u(z).
Hence, F−1 [F[u]] = u. In a similar way one can prove F[F−1 [U ]] = U , for all U ∈ S.
Example 5.4.16 Let x ∈ R. Clearly x ∈ S and from (5.4.4), (5.4.5) we get
1 d d √
F [x] = F[(−ix)1] = i F[1] = i (2π)1/2 δ0 = i 2πδ0 .
−i dξ dξ
Example 5.4.17 Let us find the solutions to
−Δu = δ0 in S , N ≥ 1. (5.4.19)
First we note that u ∈ C ∞ (RN \{0}); see Problem 5.17. Also, u must be radially sym-
1
metric in RN \{0}. Indeed, taking Fourier transform of (5.4.19) gives |ξ|2 û = .
(2π)N/2
So û is radially symmetric. Then for every rotation matrix R ∈ RN ·N , t R = R−1 , and
ϕ ∈ D(RN ) with supp(ϕ) ∩ {0} = ∅ we have
u, ϕ = û, F−1 [ϕ] = û ◦ R, F−1 [ϕ] = û, (F−1 [ϕ]) ◦ (t R)
= û, F−1 [ϕ ◦ (t R)]
= F−1 [û], ϕ ◦ (t R)
= u ◦ R, ϕ,
which implies u(x) = u(R · x), |x| = 0. So u = u(r), r = |x|. From the formula of Δ in
spherical coordinates, we get
1 N −1
Δu(x) = ∂r (r ∂r u(r)) = 0 in RN \{0}.
rN −1
87
Integrating this equation and neglecting the constant terms implies
N = 1 : u(x) = C|x|, C ∈ R,
N = 2 : u(x) = C ln |x|,
N ≥ 3 : u(x) = C|x|2−N .
The constant C is found by taking a particular test function in −Δu, ϕ = δ0 , ϕ, for
example ϕ(x) = e−|x| (which is allowed because δ0 is a distribution of order zero). In the
case N ≥ 3 we get
∞
1 = δ0 , ϕ = −Δu, ϕ = − Cr2−N rN −1 r1−N ∂r (rN −1 ∂r e−r )drdσ
S N −1
∞ 0
= C|SN | ((N − 1)e−r − re−r )dr

0
= C(N − 2)|SN |,
1
where |SN | is the area of the unit sphere in RN . Hence C = .
(N − 2)|SN |
In general, for arbitrary dimension N ≥ 1, the unique (up to a constant) solution in S
of (5.4.19) is given by
⎧
⎪ 1
⎪
⎪ − |x|, N = 1,
⎪
⎨ 2
1
E(x) = − ln |x|, N = 2, (5.4.20)
⎪
⎪ 2π
⎪
⎪ 1 1
⎩ , N ≥ 3.
(N − 2)|SN | |x|N −2
The distribution E is called a “fundamental solution to −Δ”. Using it one shows that
the solution to
−Δu = f, f ∈ D
is given by u = G ∗ f . For similar results for general linear operators with constant
coefficients, see Definition 5.3.4 and Theorem 5.3.6.
Example 5.4.18 Let us find a fundamental solution G to the heat equation solving
(5.3.7); see Example 5.3.7. By taking Fourier transform of the first equation of (5.3.7)
with respect to x and then using (5.4.10) with a = t gives
2
∂t Ĝ + ξ 2 Ĝ = 0, so Ĝ(ξ, t) = Ce−tξ and
C 1 x2
G(x, t) = √ e− 4 t , C ∈ R.
2t
88

The constant C is chosen such that limt→0 R G(x, t)ϕ(x)ds = ϕ(0). By changing the
variable and using (5.4.7), one finds easily C = √12π , which leads to (5.3.10).
Similarly, we can find the fundamental solution G to wave equation solving (5.3.11);
see Example 5.3.8. If T = ∂t G then T satisfies
∂tt T (·, t) − ΔT (·, t) = 0 for t > 0, T (·, 0) = δ0 .
By taking Fourier transform of the first equation above with respect to x and then using
(5.4.17), we get
T̂tt + ξ 2 T̂ = 0; so
T̂ (ξ, t) = Aeiξt + Be−iξt with A, B ∈ C and therefore
√
T (x, t) = 2π(Aδ0 (x + t) + Bδ0 (x − t)).
1 1 1
We choose A = B = √ so that T (·, 0) = δ0 and T (x, t) = (δ0 (x + t) + δ0 (x − t)).
2 2π 2
From ∂t G(x, t) = T (x, t) and G(·, 0) = 0 it follows that
1 1
G(x, t) = (H(x + t) − H(x − t)) = H(t − |x|).
2 2
Problems
Problem 5.1 Prove that ρ ∈ D(RN ), where
1
N
∀x ∈ R , ρ(x) = e |x|2 −1 , |x| < 1,
0, |x| ≥ 1.
Problem 5.2 Let Ω ⊂ RN open, u, v ∈ C 0 (Ω), u = v, and Tu , Tv be their corresponding

distributions in D (Ω). Show that Tu = Tv .
Problem 5.3 Find which of the following maps in D define a distribution in R:

∞ ∞
a) T, ϕ = ϕ(k), b) T, ϕ = ϕ2 (k),
k=1 k=1
∞ ∞
1
c) T, ϕ = ϕ(k) (1), d) T, ϕ = √ ϕ(k) (k).
k=1 k=1
k
Problem 5.4 Prove that if T ∈ D (Ω) then Dα T ∈ D (Ω).
Problem 5.5 Find a distribution in R which is not of finite order.
89
Problem 5.6 Let Ω ⊂ RN be open and T ∈ D (Ω). Assume T ≥ 0, i.e. T, ϕ ≥ 0 for all
real-valued ϕ, 0 ≤ ϕ ∈ D(Ω). Prove that T is a distribution of order zero (a measure).
Problem 5.7 Let n ∈ N0 , T ∈ D (R) and consider the equation xn+1 T = 0.
i) Assume n = 0. Prove that T = c0 δ0 , for arbitrary c0 ∈ C.

ii) Assume n ≥ 1. Prove that T = nk=0 ck δ0 , ck ∈ C.
(k)
Problem 5.8 Prove that if T ∈ D (R) and supp(T ) = {0} then there exists m ∈ N0 and
m
(k)
cα ∈ C, |α| ≤ m, such that9 T = ck δ0 .
k=0
Problem 5.9 Let Tn := nχ(0, 1 ) . Prove each of the following.

n
i) Tn has a limit in D (R).
ii) Tn2 does not have a limit in D (R).
iii) Tn2 − nδ0 has a limit in D (R).
sin(nx)
Problem 5.10 Let un = √ . Find the limit of un and un in D (R).
n
Problem 5.11 Find the limit in D (R) of the following distributions:

2 2
a) Tn = n1/2 e−nx , b) Tn = en δ0 (x − n),
sin(nx)
c) Tn = , d) Tn = cos(nx) + i(cos(nx))2 , i2 = −1.
nx
Problem 5.12 (Leibniz rule) Let k ∈ C ∞ (Ω) and T ∈ D (Ω), Ω ∈ RN open. Prove
that kT, ϕ = T, kϕ defines a distribution kT ∈ D (Ω) and
∂i (kT ) = k∂i T + (∂i k)T, i = 1, . . . , N, (5.4.21)
α!
Dα (kT ) = Cβα Dβ kDα−β T, Cβα = , ∀α ∈ NN
0 . (5.4.22)
0≤β≤α
β!(α − β)!
Problem 5.13 Let u = |x|, x ∈ R. Find u and u in D (R).
Problem 5.14 Let u ∈ C 1 (R\{xn , n ∈ Z}) with u having left and right limits at every
xn . Write the distribution u .
9
Note that the result holds in dimension N : if T ∈ D (RN ) and supp(T ) = {0} then there exist
m ∈ N0 and cα ∈ C, |α| ≤ m, such that T = cα Dα δ0 .
|α|≤m
90
Problem 5.15 Let u(x) = e−|x| , x ∈ R. Find −Δu + u in D (R).
Problem 5.16 Let T ∈ D (R), h ∈ R, δ > 0 and define Th by Th , ϕ = T, ϕ(· + h).
Show that Th ∈ D (R). Furthermore, show that if Th = T for all |h| < δ, then T is con-
stant.
Problem 5.17 Let u ∈ D (RN ), Δu = 0 in D (RN ). Prove that u ∈ C ∞ (RN ).
Problem 5.18 Let f ∈ D (R) and k ∈ C ∞ (R). Find the (general) solution T ∈ D (R)
to T = f and of T + kT = f .
Problem 5.19 Find F[u], where u = u(x), x ∈ R, is given by

1 sin x
a) u = sin(x), b) u = , c) u = ,
1 + x2 1 + x2
d) u = sin(x2 ), e) u = H(x), f ) u = |x|,
where H is the Heaviside function.
Problem 5.20 Find the general solution to y ± y = 0 in S (R).10

1 |x|2
− 41 t
Problem 5.21 Let u be given by (5.3.8) with G(x, t) = N/2
e , x ∈ RN , t > 0.
(4πt)
Show that u solves (5.3.9) and G solves (5.3.7) in dimension N .
Problem 5.22 Assume u ∈ D and F[u] ∈ D. Prove that u = 0.
Problem 5.23 Prove that Fourier transform in L1 (RN ) is injective.
10
n
In general, an equation k=0 ck y (k) = 0 is transformed to P (ξ)ŷ = 0, with P a polynomial. By
mi −1 (j)
using the roots ξi of P (ξ) = 0, each with multiplicity mi , it leads to ŷ = i j=0 ai,j δ0 (ξ − ξi ),
which after applying F−1 gives y.
91
6. Sobolev spaces
Sobolev spaces were introduced by Sergei Lvovich Sobolev in the 1930s, and have
since been widely embraced and developed by other mathematicians. Sobolev spaces
represent a natural functional framework to describe a rich variety of real-world prob-
lems, and they provide solutions to a large number of PDEs. Mathematically, they
provide elegant tools for studying PDEs because they have rich properties in terms of
approximation, compactness, and boundary values.
In order to avoid technicalities related to the analysis of general W s,p (Ω) spaces,
s ∈ R, p ∈ [1, ∞], we will focus mainly on H s (Ω) = W s,2 (Ω) spaces, s > 0. The reason
for this is that the analysis of H s (Ω) spaces is greatly simplified by the use of Fourier
transform, as well as the fact that H s (Ω) spaces are Hilbert spaces.
For more on W s,p (Ω) spaces in connection with the development of this chapter, see
section 9.5 in Annex, and for an extensive treatment of Sobolev spaces, see classical
books such as [1, 5, 14, 33, 40, 44, 51].
6.1 Definitions and some first properties

The reader is invited to review chapter 1.
Definition 6.1.1 Let Ω ⊂ RN be an open set, k ∈ N, p ∈ [1, ∞].
1) For p ∈ [1, ∞), we set
W k,p (Ω) = {u ∈ Lp (Ω), Dα u ∈ Lp (Ω), ∀α ∈ NN 0 , |α| ≤ k}, (6.1.1)

1/p 1/p
uW k,p (Ω) = Dα upLp (Ω) , |u|W k,p (Ω) = Dα upLp (Ω) , (6.1.2)
|α|≤k |α|=k
and for p = ∞, set
W k,∞ (Ω) = {u ∈ L∞ (Ω), Dα u ∈ L∞ (Ω), ∀α ∈ NN , |α| ≤ k}, (6.1.3)

0
uW k,∞ (Ω) = Dα uL∞ (Ω) , |u|W k,∞ (Ω) = Dα uL∞ (Ω) . (6.1.4)
|α|≤k |α|=k

https://doi.org/10.1007/978-3-031-39524-6 6
6.1. Definitions and some first properties Chapter 6
2) The space W k,p (Ω) is called a “W k,p Sobolev space in Ω”. The integer k is called
an “order of W k,p (Ω)”, and k − Np is called a “differential dimension of W k,p (Ω)”.
The quantity uW k,p (Ω) , resp. |u|W k,p (Ω) , is called a “norm, resp. seminorm of u
in W k,p (Ω)”.
3) For p = 2, we write H k (Ω) instead of W k,2 (Ω). In H k (Ω), we define an inner

product

∀u, v ∈ H (Ω) : (u, v)H k (Ω) :=
k
Dα uDα vdx. (6.1.5)
|α|≤k Ω
4) For p ∈ [1, ∞), sometimes we use the other equivalent norm and seminorm:

uW k,p (Ω) = Dα uLp (Ω) , |u|W k,p (Ω) = Dα uLp (Ω) . (6.1.6)
|α|≤k |α|=k
Also, when Ω = RN , we will write W k,p , H k , instead of W k,p (RN ), H k (RN ).
5) If E(Ω) is any of the spaces above, we define Eloc (Ω) = {u ∈ E(ω), ∀ω Ω}.
For u ∈ Lp (Ω), we have Dα u ∈ D (Ω). Definition 6.1.1 requires Dα u to be a Lp (Ω)

function. Furthermore, functions of W k,p (Ω) are defined almost everywhere in Ω. More
precisely, the elements of W k,p (Ω) are classes of equivalences defined by the relation
“u = v if and only if (iff ) u = v a.e. in Ω, u, v ∈ W k,p (Ω)”; see Problem 6.1. The ele-
ments of the same equivalent class have the same W k,p (Ω) norm. This is the reason why
throughout this text we will not make the distinction between a function of W k,p (Ω)
and its equivalent class.
The following result is classical. The proof of completeness is simple and based on the
completeness of Lp (Ω) spaces, while the separability and reflexivity require some results
from Functional Analysis. The proof can be found in many books treating Sobolev
spaces; see, for example, [1, 5, 14].
Theorem 6.1.2 Let k ∈ N and Ω ⊂ RN be open. W k,p (Ω) equipped with the norm
· W k,p (Ω) is a Banach space for all p ∈ [1, ∞]. W k,p (Ω) is separable for all p ∈ [1, ∞),
and reflexive for all p ∈ (1, ∞).
Example 6.1.3 Let u(x) = |x|q , Ω = B(0, 1) ⊂ RN , q = 0.
i) We have u ∈ W 1,p (B) iff p(q − 1) > −N . Indeed, considering spherical coordinates

(r, θ) and if SN = {x ∈ RN , |x| = 1} is the unit sphere in RN , we have
1 1
N −1 pq
|u| dx =
p
r r drdθ = |SN | rN −1+pq dr.
B SN 0 0
94
Chapter 6 6.1. Definitions and some first properties
Hence u ∈ Lp (B) iff pq > −N . For the derivatives, note that ∂i u = q|x|q−2 xi and
1 1
N −1 p(q−2)
|∂i u| dx = q
p p
r r |xi | drdθ ∼
p
rN −1+p(q−1) dr,
B SN 0 0
where a ∼ b means C1 a ≤ b ≤ C2 a, for certain C1 , C2 > 0. Hence ∂i u ∈ Lp (B) iff

p(q − 1) > −N , which proves the claim.
ii) In general, for k ∈ N, p ∈ [1, ∞], u ∈ W k,p (B), iff p(q − k) > −N , or equivalently
q > k − Np .
c
iii) Similar calculus show that k ∈ N, p ∈ [1, ∞], u ∈ W k,p (B ), iff q < k − N
p
.
Example 6.1.4 Let u(x) = ln |x|, x ∈ RN , Ω = B(0, 1) ⊂ RN . Then as in the example
above, we can prove that u ∈ W 1,p (Ω) iff p < N .
Example 6.1.5 Let u(x) = 1(0,1) (x), x ∈ (−1, 1). Then u ∈ / W 1,p (−1, 1), p ∈ [1, ∞].
Indeed, let u ∈ D (−1, 1) be the distributional derivative of u. Then u = δ0 and δ0 ∈

Lp (−1, 1) for all p ∈ [1, ∞]; see Example 5.2.8.

Example 6.1.6 Let Ω = A ∪ γ ∪ B, where A, B are open bounded sets, and assume
γ = ∂A ∩ ∂B = ∅ is a smooth boundary. Set u = α1A + β1B with α, β ∈ R. If α = β
/ W 1,p (Ω), p ∈ [1, ∞].
then u ∈
We prove the claim by contradiction, so assume u ∈ W 1,p (Ω), p ∈ [1, ∞]. Also, for
simplicity we consider A = (−1, 0) × (0, 1), B = (0, 1) × (0, 1), γ = {0} × (0, 1), and
p = 2. So u(x1 , x2 ) = (β − α)H(x1 ) + α. Then for every ϕ ∈ D(Ω), we have
∂2 u, ϕ = 0,
1
∂1 u, ϕ = − u, ∂1 ϕ = (β − α)ϕ(0, x2 )dx2 =: (β − α)δγ , ϕ,
0
with δγ ∈ L2 (Ω). Note also that δγ = 0 because if not we get ∇u = 0 and then u is
constant (see Example 5.2.21), which is impossible because α = β.
From the density of D(Ω) in L2 (Ω), there exists (ϕn ) in D(Ω) converging to δγ in
2
L (Ω). WLOG we may assume that the convergence of (ϕn ) is also pointwise in Ω and
|ϕn | ≤ v a.e. Ω for all n, for a certain function v ∈ L2 (Ω); see Theorem 1.3.7. Set
ψn (x1 , x2 ) = (1 − ηn (x1 ))ϕn (x1 , x2 ), where ηn (x1 ) = η(nx1 ) and η is a {[−1, 1], (−2, 2)}
cut-off function. It follows that limn→∞ ψn − δγ L2 (Ω) = 0, because

ψn − ϕn L2 (Ω) =
2
|ηn (x1 )ϕn (x)| dx =
2
|ηn (x1 )ϕn (x)|2 dx
Ω {|x1 |<2/n}
→ 0,
n→∞
where for the limit we used Lebesgue DCT. It implies
1
0 = δγ L2 (Ω) = lim
2
δγ ψn dx = lim δγ , ψn = lim ψn (0, x2 )dx2 = 0,
n→∞ Ω n→∞ n→∞ 0
because ψn (0, x2 ) = 0. The contradiction shows that u ∈
/ W 1,2 (Ω).
95
6.1.1 Density of D in W k,p

For k ∈ N and p ∈ [1, ∞), we have D ⊂ W k,p . In the following theorem we show that
this embedding is dense. The theorem involves two steps, regularization and truncation,
which are frequently used when working with Sobolev space functions. We emphasize
that the following theorems 6.1.7 and 6.1.8 do not hold for p = ∞.
Theorem 6.1.7 (Density of D and S in W k,p ) Let k ∈ N, p ∈ [1, ∞). Then the spaces
D and S are included densely in W k,p , p ∈ [1, ∞). We write D −→
d
W k,p , S −→
d
W k,p .
Proof. The proof is made in two steps.
Regularization. Let (ρn ) be a mollifier sequence, and for u ∈ W k,p set vn = u ∗ ρn . For
α ∈ NN0 , |α| ≤ k, from Theorem 5.3.3 we have
Dα v = Dα (u ∗ ρn ) = Dα u ∗ ρn , (6.1.7)
which in combination with i), Theorem 5.2.5, implies vn ∈ C ∞ ∩ Lp and lim vn = u in

n→∞
W k,p . In general, (vn ) is not in D and to generate a sequence in D we proceed as follows.
Truncation. Let η be a {B(0, 1), B(0, 2)} cut-off function and set ηn (x) = η(x/n), un =
ηn vn . Then un ∈ D and limn→∞ un = u in W k,p because
Dα un − Dα uLp ≤ ηn (Dα vn − Dα u)Lp + (1 − ηn )Dα uLp

+ Cαβ Dβ ηn Dα−β vn Lp ,
0<β≤α
where we used the result of Problem 5.12. The first integral tends to zero from the
construction of vn , and the second integral tends to zero from Lebesgue DCT. For
the third integral, one notices that Dβ ηn (x) = n−|β| Dβ η(x/n), which implies that this
integral also tends to zero.
The claim for S follows from the fact that given ϕ ∈ S, if ϕn = ηn ϕ then limn→∞ ϕn =
ϕ in W k,p , which is proven similarly as the convergence of un to u above.
The following result, even though it only provides the density of D(Ω) in Lp (Ω) and in
k,p
Wloc (Ω), is useful in many applications because it does not depend on the regularity of Ω.
Theorem 6.1.8 (Friedrich’s theorem) Let k ∈ N, p ∈ [1, ∞), and Ω RN . For

every u ∈ W k,p (Ω), there exists (un ) in D(Ω) such that limn→∞ un = u in Lp (Ω) and
k,p
limn→∞ un = u in Wloc (Ω), i.e.
lim un = u in Lp (Ω) and lim un = u in W k,p (ω), ∀ω Ω. (6.1.8)
n→∞ n→∞
Proof. The proof will be carried out in two steps.

i) Let ω ⊂ ω =: K ⊂ G Ω, with ω and G open, and let η ∈ D(RN ) be a {K, G}
96
cut-off function as given by Theorem 1.2.2. Note that η = 1 in K, supp(η) ⊂ G and

dist(supp(ϕ), Gc ) > 0.
Let U be the extension in RN of u by zero and set V = ηU . From Lemma 9.5.7, we
have V ∈ W k,p (RN ). Set Vn = ρn ∗ V . As Dα (V ∗ ρn ) = Dα V ∗ ρn , see (6.1.7), from
Theorem 5.2.5 it follows limn→∞ Vn = V in W k,p (RN ). Furthermore

Vn (x) = ρn (x − y)η(y)u(y)dy = ρn (x − y)η(y)u(y)dy.
supp(η) supp(η)∩B(x,1/n)
It follows that Vn ∈ C ∞ (RN ) and if 1/n < dist(supp(η), Gc ) then Vn (x) = 0 for x ∈ Gc .
Hence, if un = Vn |Ω we have un ∈ D(Ω). Also, as η = 1 in K, from limn→∞ Vn = V
in W k,p (RN ) it follows limn→∞ un = u in W k,p (ω). Therefore, we have proved that for
every ω G Ω and for every n ∈ N, there exists un ∈ D(Ω),
un − ηuLp (Ω) + un − uW k,p (ω) < 1/n.
ii) Now we will repeat i) for an increasing sequence of domains like ω and G. Namely, for
n ∈ N set On = {x ∈ Ω, d(x, ∂Ω) < 1/n, |x| < n}. Note that On is open, On is compact,
On ⊂ On+1 , and ∪n∈N On = Ω.
For every n ∈ N, from part i) with ω = On and G = On+1 , there exists un ∈ D(Ω)
and ηn ∈ D(Ω), a {On , On+1 } cut-off function, such that
un − ηn uLp (Ω) + un − uW k,p (On ) < 1/n.
Then (un ) is the desired sequence. Indeed, from Lebesgue’s dominated convergence the-
orem we have

un − uLp (Ω) ≤ C
p
|un − ηn u| + |1 − ηn | |u| → 0 as n → ∞.
p p p
Ω Ω
Finally, for a given ω Ω there exists nω such that ω ⊂ On for all n > nω . Therefore
for such n, we have
un − uW k,p (ω) ≤ un − uW k,p (Gn ) < 1/n,
Using a similar technique as above, one can prove the following result; see, for exam-
ple, [1, 5, 14].
Theorem 6.1.9 Let k ∈ N, p ∈ [1, ∞) and Ω ⊂ RN be open. Then C ∞ (Ω) ∩ W k,p (Ω) is
dense in W k,p (Ω).
97
6.1.2 Some applications

In this section, we present some applications of Theorem 6.1.8, which show that we
can make a number of operations with Sobolev space functions as if they were usual
classical functions.
We note that in general, if u, v ∈ W k,p (Ω) then uv ∈ / W k,p (Ω). For example, let
Ω = B(0, 1), and u = |x|q , q ∈ R. Then from Example 6.1.3, u ∈ W k,p (Ω) if and only if
q > k − Np . We can choose q = − 12 and p ∈ [1, ∞] such that k − Np = −1 (for example
N
p = k+1 ), so that u ∈ W k,p (Ω). However, if v = u then uv = |x|2q ∈
/ W k,p (Ω) because
2q = −1 ≯ −1 = k − Np . The following proposition gives a sufficient condition for the
product uv to be in W 1,p (Ω).
Proposition 6.1.10 Let Ω ⊂ RN be an open set and u, v ∈ W 1,p (Ω)∩L∞ (Ω), p ∈ [1, ∞].
Then uv ∈ W 1,p (Ω) and for all |α| = 1, we have
Dα (uv) = vDα u + uDα v. (6.1.9)
Proof. Clearly uv, vDα u, uDα v ∈ Lp (Ω). So it is enough to prove (6.1.9). We assume
first p ∈ [1, ∞). For ϕ ∈ D(Ω), there exists ω Ω such that supp(ϕ) ⊂ ω. Let (un ),
resp. (vn ), be a sequence in D(Ω) satisfying Theorem 6.1.8, associated with u, resp. v,
and such that furthermore un , resp. vn , converges a.e. and in W 1,p (ω) to u, resp. v (see
Theorem 1.3.7). Then for α ∈ NN 0 , |α| = 1, we get

Dα (uv), ϕ = − uvDα ϕdx
Ω

= − lim α
un vn D ϕ = lim ϕ(Dα un vn + un Dα vn )
n→∞ Ω n→∞ ω

= (Dα uv + uDα v)ϕ,
Ω
where when passing to the limit we used Lebesgue’s dominated convergence theorem.
In the case p = ∞, given ϕ ∈ D(Ω) we choose an open bounded set ω, supp(ϕ) ⊂
ω Ω. We can apply the case p ∈ [1, ∞) in ω, which proves (6.1.9) and completes the
proof.
We leave as an exercise the proof of the following proposition.

Proposition 6.1.11 Let u ∈ W k,p (Ω) and v ∈ W k,∞ (Ω). Then uv ∈ W k,p (Ω) and for
all |α| ≤ k, we have
α!
Dα (uv) = Cβα Dβ uDα−β v, Cβα = . (6.1.10)
β≤α
β!(α − β)!
The following two propositions show how the operation of composition holds in the
context of Sobolev spaces.
98
Proposition 6.1.12 Let Ω ⊂ RN be an open set, h ∈ C 1 (R), h ∈ L∞ (R), and h(0) = 0.

If u ∈ W 1,p (Ω), p ∈ [1, ∞] then h ◦ u ∈ W 1,p (Ω) and
h ◦ u ∈ W 1,p (Ω), Dα (h ◦ u) = h (u)Dα u, ∀α ∈ NN

0 , |α| = 1. (6.1.11)
Proof. As |h(t)| ≤ h L∞ (R) |t|, it follows that h ◦ u, h (u)Dα u ∈ Lp (Ω). It is enough
then to prove the equality in (6.1.11).
We consider first the case p ∈ [1, ∞). Let ω be an open set, ω Ω. From Theorem
6.1.8, there exists un ∈ D(Ω) such that limn→∞ un = u in Lp (Ω) and limn→∞ ∇un = ∇u
in Lp (ω). Also, without restriction we may assume that un → u pointwise in Ω, and
∇un → ∇u pointwise in ω; see Theorem 1.3.7.
From the assumption for h and the convergence of un , we have lim h(un ) = h(u)
n→∞
in Lp (Ω). Using Lebesgue DCT and the convergence a.e. of un and of ∇un yields
lim h (un ) = h (u) and lim h (un )Dα un = h (u)Dα u in Lp (ω). Then we get
n→∞ n→∞
Dα (h ◦ u), ϕ = − h ◦ u, Dα ϕ
= − lim h ◦ un , Dα ϕ = lim Dα (h ◦ un ), ϕ = lim h (un )Dα un , ϕ
n→∞ n→∞ n→∞
α
= h (u)D u, ϕ,
which completes the proof in the case p ∈ [1, ∞).
In the case p = ∞, we choose an arbitrary ω Ω. Then u ∈ W 1,p (ω), for all
p ∈ [1, ∞), and so the equality in (6.1.11) holds in ω, so in Ω.
We note that Proposition 6.1.12 √does not hold in general if h is not C 1 . For example,
let Ω = (0, 1), u = xα , α ∈ (1/2, 1/ 2), and h(x) = xα . From Example 6.1.3, we have
2
u, h ∈ H 1 (Ω) because p(α−1) > −N (here p = 2 and N = 1), but h◦u(x) = xα ∈ / H 1 (Ω)
because p(α2 − 1) < −N .
Proposition 6.1.13 Let Ω, G ⊂ RN be open sets and θ = (θ1 , . . . , θN ) ∈ W 1,∞ (G; Ω) be

invertible with θ−1 ∈ W 1,∞ (Ω; G). If u ∈ W 1,p (Ω), p ∈ [1, ∞], then u ◦ θ ∈ W 1,p (G) and

Dα (u ◦ θ) = Dα θβ Dβ u ◦ θ, ∀α ∈ NN 0 , |α| = 1, (6.1.12)
|β|=1
where θβ = θ1β1 · · · θN
βN
.
Proof. Let u ∈ W 1,p (Ω). From the classical change of variable theorem for Lp (Ω) func-
tions, we have

∀u ∈ L (Ω), u ◦ θ ∈ L (G) and
p p
|u| dx =
p
|u ◦ θ|p |Dθ|dy. (6.1.13)
Ω G
So u ◦ θ and all the right-hand side terms of (6.1.12) are in Lp (G). Then to prove
u ◦ θ ∈ W 1,p (G), it is enough to prove (6.1.12).
99
Consider first the case p ∈ [1, ∞). For ϕ ∈ D(Ω), there exists ω Ω such that
supp(ϕ) ⊂ ω. Let (un ) in D(Ω), limn→∞ un − uW 1,p (ω) = 0, as in Theorem 6.1.8. Note
that (6.1.13) implies that lim un ◦ θ = u ◦ θ and lim Dα un ◦ θ = Dα u ◦ θ in Lp (G), for
n→∞ n→∞
all |α| = 1. Therefore we get
Dα (u ◦ θ), ϕ = − u ◦ θ, Dα ϕ = − lim un ◦ θ, Dα ϕ
n→∞
= lim Dα (un ◦ θ), ϕ
n→∞

= lim Dα θβ Dβ un ◦ θ, ϕ
n→∞
|β|=1

= Dα θβ Dβ u ◦ θ, ϕ,
|β|=1
which proves (6.1.12) and u ◦ θ ∈ W 1,p (G).

In the case p = ∞, we can choose an open and bounded set ω as in Proposition
6.1.12. Then u ∈ W 1,p (ω), for all p ∈ [1, ∞), and (6.1.12) holds in θ−1 (ω), which from
the choice of ω proves (6.1.12).
Note that Proposition 6.1.13 extends in the W context. Namely, let Ω, G ⊂ R be
k,p N
two open sets and θ = (θ1 , . . . , θN ) ∈ W k,∞ (G; Ω) be invertible with θ−1 ∈ W k,∞ (Ω; G).
If u ∈ W k,p (Ω) then u ◦ θ ∈ W k,p (G), p ∈ [1, ∞], and, furthermore, Dα (u ◦ θ), |α| = k,
is a finite sum of terms of the form Dβ u ◦ θ, |β| ≤ k, multiplied by Dγ θσ terms, with
|γ| ≤ k, |σ| = 1.
The proof is carried out by recurrence on k. Indeed, for k = 1 the claim holds. Assume
it holds for k − 1, so Dα (u ◦ θ), |α| = k − 1, is a sum of terms Dβ u ◦ θ, |β| ≤ k − 1,
multiplied by Dγ θσ terms, with |γ| ≤ k − 1, |σ| = 1. Then applying Proposition 6.1.13
to Dα (u ◦ θ), |α| = k − 1, proves the claim.
Example 6.1.14 As a corollary of Proposition 6.1.12, let us prove the following. Let
Ω ⊂ RN be an open set, u ∈ W 1,p (Ω), p ∈ [1, ∞], and set
u+ = max{0, u(x), x ∈ Ω}, u− = min{0, u(x), x ∈ Ω}. (6.1.14)
Then
u± ∈ W 1,p (Ω), (6.1.15)
∇u± = (∇u)1{u≷0} , and (6.1.16)
∇u = 0 a.e. in {u = C} := {x ∈ Ω, u(x) = C}, C ∈ R. (6.1.17)
Indeed, for > 0 let f (t) = ((t2 + 2 )1/2 − )H(t), and u = f ◦ u. From Proposition
6.1.12, we have u ∈ W 1,p (Ω) and for ϕ ∈ D(Ω) we get

Du , ϕ = − u Dϕdx = − lim u Dϕdx
+ +
Ω →0 Ω
100
Chapter 6 6.2. H s spaces and Fourier transform: W s,p and W0s,p spaces

u∇u
= lim Du ϕdx = lim ϕ
→0 Ω →0 {u>0} (u2 + 2 )1/2

= 1{u>0} ∇uϕ,
Ω
which proves the lemma for u+ . The result for u− follows from the result for u+ by
noting that u− = −(−u)+ . For (6.1.17) set v = u − C. Since u = C + v + + v − , from
(6.1.16) it follows
(∇u)1{u=C} = (∇v)1{v=0} = (∇v + + ∇v − )1{v=0} = 0.
6.2 H s spaces and Fourier transform: W s,p and W0s,p

spaces
Sobolev spaces in RN are easily analyzed by using Fourier transform. The advantage
of working with these spaces is that they are homeomorphic to the space L2s defined
below, which makes it possible to “transfer” well-known properties of L2s spaces to H s
spaces.
Theorem 6.2.1 For u ∈ H k and α ∈ NN

0 , |α| ≤ k, we have
F[Dα u] = (iξ)α F[u], Dα uL2 = (iξ)α F[u]L2 . (6.2.1)
Furthermore, Fourier transform F is a homeomorphism1 from H k onto the weighted

space L2k defined by

2 12
Lk = v ∈ L , vL2 :=
2 2 2
ξ |v(ξ)| dξ < ∞,
2k 2
ξ = (1 + |ξ| ) . (6.2.2)
k
RN
Proof. Note that from Theorem 5.4.9 and Proposition 5.4.6, (6.2.1) holds in S.
i) F : H k → Lk is well-defined and continuous. Indeed, for u ∈ S, from (6.2.1) we get

F[u]L2 =
2
ξ |û| dξ ∼
2k 2
|(iξ)α |2 |û|2 dξ = F[Dα u]2L2 (RN )
k
RN RN |α|≤k |α|≤k
= u2H k . (6.2.3)
As S −→ d
H k , (6.2.3) implies that F extends to a continuous linear operator from H k
to Lk and (6.2.1) holds for all u ∈ H k .
2
ii) F : H k → L2k is invertible and its inverse F−1 is continuous. Indeed, let v ∈ L2k . Then
v ∈ L2 and from Theorem 5.4.9 there exists a unique u ∈ L2 such that F[u] = v. From
Remark 5.4.11 and v ∈ L2k , it follows F[Dα u] = (iξ)α v ∈ L2 for all |α| ≤ k, which in
1
A continuous linear invertible operator from a normed space X onto another normed space Y .
101
6.2. H s spaces and Fourier transform: W s,p and W0s,p spaces Chapter 6
combination with Theorem 5.4.9 implies Dα u ∈ L2 , so u ∈ H k which proves that F is

invertible. Finally, the boundedness of F−1 follows from (6.2.3).
The estimate (6.2.3) shows that ûL2k is a norm in H , equivalent to the usual
k

norm u2H k = |α|≤k D uL2 , which motivates the introduction of the spaces H ,
α 2 s
with arbitrary s > 0.
Definition 6.2.2 For s > 0, the “fractional order Sobolev space H s ” is defined by
H s = {u ∈ L2 , û := F[u] ∈ L2s }, equipped with (6.2.4)
(u, v)H s = (û, v̂)L2s , (6.2.5)
where

L2s = v ∈ L2 , v2L2s = (v, v)L2s < ∞ , and (6.2.6)

(v, w)L2s = ( ξs v)( ξs w)dξ. (6.2.7)
RN
Note that in some texts is used ξ2s = 1+|ξ|2s . The results are exactly the same because
(1 + |ξ|2 )s and 1 + |ξ|2s are equivalent for s > 0. This is the reason why we will use
ξ2s = 1 + |ξ|2s , whenever it is more appropriate.
Example 6.2.3 Let u(x) = 1(0,1) (x), x ∈ R. Clearly u ∈ L2 . As u(x) = H(x)−H(x−1),

we get u = δ0 − δ1 . So u ∈/ L2 and therefore u ∈/ H 1 . But, for what s > 0 do we have
u ∈ H s ? Note that we have
1
1 −i(ξx) 1 1 1 − e−iξ
û = √ e u(x)dx = √ e−i(ξx) dx = √ ,
2π
R 2π 0 2π iξ
|ξ|2s |û|2 dξ = C |ξ|2(s−1) |1 − cos(ξ)|2 dξ
R R
1
< ∞ if f 2(s − 1) < −1 or equivalently s < .
2
Hence, 1(0,1) ∈ H only for s ∈ (0, 1/2).
s
Example 6.2.4 Let u ∈ H s , s > 0, and α ∈ NN

0 , |α| ≤ s. Then D u ∈ H
α s−|α|
. Indeed,
F[Dα u] = (iξ)α F[u],

ξ 2(s−|α|)
|F[D u]| dξ =
α 2
ξ2(s−|α|) |(iξ)α |2 |F[u]|2 dξ
RN RN

≤C ξ2(s−|α|)+2|α| |F[u]|2 dξ
RN
= CF[u]2L2s
< ∞,
which proves the claim.
102
Chapter 6 6.2. H s spaces and Fourier transform: W s,p and W0s,p spaces
Remark 6.2.5 There is another norm for H s spaces which serves as motivation for
defining W s,p (Ω) spaces. Indeed, let s ∈ (0, 1) and u ∈ H s . Using the fact

1 −i(ξ·x) 1
F[u(· + y)](ξ) = N/2
e u(x + y)dx = e i(ξ·y)
N/2
e−i(ξ·z) u(z)dz
(2π) RN (2π) RN
= e i(ξ·y)
F[u](ξ),
and Plancherel theorem, one gets

|u(x + y) − u(x)|2 |F[u(· + y)](ξ) − F[u](ξ)|2
dxdy = dξdy
RN RN |y|N +2s RN RN |y|N +2s

|ei(y·ξ) − 1|2 1
= · N · |F[u](ξ)|2 dξdy
RN RN |y| 2s |y|

|ei(y·ξ)
− 1|2
dy
= · N |F[u](ξ)|2 dξ
N RN |y| 2s |y|
R
=: h(ξ)|F[u](ξ)|2 dξ.
RN
Note that h(ξ) is radial and homogeneous of degree 2s:

|ei(y·(tξ)) − 1|2 dy |ei(z·ξ) − 1|2 |t|−N dz
h(tξ) = · N = · · = |t|2s h(ξ),
RN |y|2s |y| RN |z|2s |t|−2s |t|−N |z|N
which implies2

ξ ξ |ei(y·ζ) − 1|2 dy
h(ξ) = h |ξ| = |ξ| h
2s
= Cs |ξ|2s , Cs = · N , |ζ| = 1.
|ξ| |ξ| RN |y|2s |y|
Therefore, for s ∈ (0, 1) and u ∈ H s we get

|u(x) − u(y)|2
Cs |ξ| |F[u]| dξ =
2s 2
dxdy. (6.2.8)
RN RN |x − y|
N +2s
RN
Note that for s ∈ R+ , we have the equivalence

ξ2s ∼ ξ2s + |ξ|2(s−s ) |ξ α |2 , (6.2.9)
|α|=s
where s is the largest integer smaller than s. Hence from (6.2.8), for 0 < s ∈
/ N we get

uH s ∼
2
ξ |F[u]| dξ +
2s 2
|ξ|2(s−s ) |(iξ)α F[u]|2 dξ
RN |α|=s RN
2
Here Cs < ∞ iff s ∈ (0, 1).
103
6.2. H s spaces and Fourier transform: W s,p and W0s,p spaces Chapter 6

= u2H s + |ξ|2(s−s ) |F[Dα u]|2 dξ
|α|=s RN

|Dα u(x) − Dα u(y)|2
= u2H s + dxdy.
RN RN |x − y|N +2(s−s )
|α|=s
This implies that for s ∈ R+ \N, an equivalent inner product in H s is given by

(u, v)H s = (u, v)H s
(Dα u(x) − Dα u(y))(Dα v(x) − Dα v(y))
+ dxdy, (6.2.10)
RN RN |x − y|N +2(s−s )
|α|=s
which motivates the following definition.
Definition 6.2.6 Let Ω ⊆ RN open.
i) For s ∈ R+ \N, p ∈ [1, ∞), define (W s,p (Ω), · W s,p (Ω) ) by

W s,p (Ω) = u ∈ Lp (Ω), upW s,p (Ω) := upW s,p (Ω) + |u|pW s,p (Ω) < ∞ , (6.2.11)
p
Dα u(x) − Dα u(y)
|u|pW s,p (Ω) = N ,
|α|=s |x − y| p +(s−s )
Lp (Ω×Ω)
where · W s (Ω) is defined in (6.1.2).
ii) For s > 0 and p ∈ [1, ∞), we denote by W0s,p (Ω) the closure of D(Ω) for the norm
(6.2.11) or (6.1.2).
iii) If p = 2 we write H s (Ω) = W s,2 (Ω), H0s (Ω) = W0s,2 (Ω), and equip them with the
inner product
(u, v)H s (Ω) := (u, v)H s,p (Ω)

(Dα u(x) − Dα u(y))(Dα v(x) − Dα v(y))
+ dxdy. (6.2.12)
Ω Ω |x − y|N +2(s−s )
|α|=s
iv) The number s − N

p
is called a “differential dimension of W s,p (Ω)”.
Note that the spaces W s,p (Ω) and W0s,p (Ω) are Banach spaces and for p = 2, they
are Hilbert spaces; see, for example, [1, 44, 51]. Also, it follows H s of (6.2.10) and is the
same as W s,2 of (6.2.11).
104
Chapter 6 6.3. Continuous, compact, and dense embedding theorems in H s (Ω)
Remark 6.2.7 The case s ∈ / N and p = ∞ is excluded from Definition 6.2.6. The
s ,s−s
reason is that W s,∞ (Ω) = Cb (Ω); see Proposition 1.1.11. Indeed, by formally
letting p → ∞ in Definition 6.2.6, one would define (see, for example, [41])
W s,∞ (Ω) = u ∈ L∞ (Ω) such that (6.2.13)
Dα u(x) − Dα u(y)
uW s,∞ (Ω) := uW s,∞ (Ω) + <∞ .
|x − y|s−s L∞ (Ω×Ω)
|α|=s
Example 6.2.8 Let Ω = (−1, 1) ⊂ R and u(x) = H(x), x ∈ Ω. We wonder for what
(s, p) we have u ∈ W s,p (Ω). We know that u ∈ / W 1,p (Ω), for all p ∈ [1, ∞]. Then we
restrict the discussion to the case s ∈ (0, 1), p ∈ [1, ∞). Clearly u ∈ Lp (Ω). We compute
1 1 0 1
|u(x) − u(y)|p 1
|u|W s,p (Ω) =
p
dxdy = 2 dxdy
−1 −1 |x − y| 1+sp
−1 0 (x − y)
1+sp

2 0 −sp
x=1 2 0
= − (x − y) dy = − (1 − y)−sp − (0 − y)−sp dy
sp −1 x=0 sp −1
2 1 0
= ((1 − y)−sp+1 − (0 − y)−sp+1
sp sp − 1 −1
< ∞, if f sp < 1.
So, u ∈ W s,p (Ω) for all s ∈ (0, 1), p ∈ [1, ∞) such that sp < 1.
6.3 Continuous, compact, and dense embedding

theorems in H s(Ω)
Continuous and compact embedding results hold in W s,p spaces. In order to simplify
the proofs, we will restrict our discussion to H s spaces. Let us start with a general
definition.
Definition 6.3.1 Let E and F be two Banach spaces.
i) We say “E is continuously embedded in F”, and write “E −→ F ”, if 3
E ⊂ F and ∃C > 0, ∀u ∈ E, uF ≤ CuE . (6.3.1)
ii) We say “E is compactly embedded in F”, and we write “E −→ c

F ”, if E −→ F
and the unit ball in E is precompact in F , i.e. every bounded sequence in E has a
convergent subsequence in F .
3
In other words, the map I : u ∈ E → I(u) = u ∈ F is continuous.
105
6.3. Continuous, compact, and dense embedding theorems in H s (Ω) Chapter 6
iii) We say “H is dense in E”, or “H is densely embedded in E”, and write “H −→ d

E”,
if H ⊂ E and for every element u ∈ E there exists a sequence (un ) in H such that
lim un = u in E.
n→∞
Continuous, resp. compact, embedding theorems in Sobolev spaces show that if the
differential dimension s − Np is large enough then the Sobolev space W s,p is continuously
embedded in, for example, more familiar spaces such as C 0 , resp. compactly embedded
in larger spaces, such as L2 . Compact embeddings are particularly important when
studying nonlinear PDEs.
Density theorems are significant when H is a set of regular functions. For example,
one can prove a certain property only for u ∈ H, which a priori is simpler than the proof
for u ∈ E. Next, by using the density of H in E, one can prove that the same property
holds for u ∈ E. This is done by writing the property for a sequence (un ) in H, with
limn→∞ un = u in E, and then passing to the limit.
6.3.1 Case Ω = RN
First, we develop the results in H s spaces. Here, the analysis is simplified based on
the fact that H s spaces are homeomorphic with L2s spaces. Next, the results are extended
to H s (Ω) spaces. Here, the regularity of Ω is important. It allows us to extend H s (Ω)
functions as H s functions and use the results developed for H s spaces.
Theorem 6.3.2 For all s > 0, the space H s equipped with the inner product (u, v)H s is
a Hilbert space. Furthermore, D −→
d
H s and S −→d
H s.
Proof. From Definition 6.2.2, F : H s → L2s is an isometry,4 which implies that

(H s , (u, v)H s ) is an Hilbert space because (L2s , · L2s ) is so. The embedding S −→d
Hs
follows from S −→ Ls , F[S] = S and the isometry of F : H → Ls .
d 2 s 2
For the embedding D −→ d

H s , given u ∈ H s , let (un ) in S converging to u in H s .
As D −→ d
S, see Proposition 5.4.4, for every un there exists (ϕm ) in D such that (ϕm )
converges to un in S. From the continuity of F : S → S in S, we get

un − ϕm H s (RN )
2
= ûn − ϕ̂m L2s (RN ) =
2
ξ2s |ûn − ϕ̂m |2 dξ
RN

1
≤ 2N
ξ 2(N +s)
|ûn − ϕ̂m |2 dξ
RN ξ
≤ CN |s(N +s ) (ûn − ϕ̂m )|2
−−−→ 0.
m→∞
We take ψn = ϕmn with mn such that un − ϕmn H s (RN ) < 1/n. So (ψn ) is in D and
converges to u in H s .
4
A homeomorphism that preserves the norm.
106
Theorem 6.3.3 Let s − N

2
> 0. Then5 H s −→ Cb0 ∩ {u, lim|x|→∞ u(x) = 0}.
Proof. Note that u(x) = F−1 [F[u](ξ)](x), so u(−x) = F[F[ξ]](x). Recall also Riemann-
Lebesgue Lemma 9.4.14, which says if v ∈ L1 (RN ) and v̂ = F[v] then
1
v̂ ∈ Cb0 ∩ { lim v̂(ξ) = 0} and v̂L∞ ≤ vL1 .
|ξ|→∞ (2π)N/2
If û ∈ L1 , applying this lemma with v = û gives the required result. So, we need to show
that û ∈ L1 . As u ∈ H s we have û ∈ L2s and
1/2
−s −s dξ
|û|dξ = ξ ξ ûL1 ≤ ξ L2 ξ ûL2 ≤ C
s s
2s
uH s .
RN RN ξ
N
Since s − > 0, we obtain
2
∞ ∞
dξ rN −1 N −1−2s
= dr ∼ 1 + r ds < ∞,
RN ξ2s 1 (1 + |r|2 )s 1
which proves ûL1 < ∞ and completes the proof.
For the proof of the following result, see Theorem 9.5.1 in Annex.
Theorem 6.3.4 Assume s − N2 ∈ (m, m + 1], m ∈ N0 and let σ ∈ (0, s − N2 − m}). Then
H s −→ Cbm,σ ∩ {u, lim Dα u = 0, ∀α ∈ NN

0 , |α| ≤ m}. (6.3.2)
|x|→∞
In particular, there exists C > 0 such that for all α ∈ NN

0 with |α| = m and x, y ∈ R
N
we have
|Dα u(x) − Dα u(y)| ≤ CDα uH s−m |x − y|σ . (6.3.3)
Remark 6.3.5 One may wonder whether the condition s − N2 > 0 in theorems 6.3.3 and
6.3.4 is due to the technique we have used. Actually, this condition is necessary, in the
sense that if s− N2 ≤ 0 the theorem does not hold. Indeed, we have seen in Example 6.2.3
that u = 1(0,1) ∈ H s (R) only for s−1/2 < 0 (so, here N = 1 and s−N/2 = s−1/2 < 0).
Clearly u ∈/ C 0 . In this context, we say that the condition s − N2 > 0 in theorems 6.3.3
and 6.3.4 is optimal.
Theorems 6.3.3 and 6.3.4 do not hold in general even in the critical case s−N/2 = 0.
For example, let u(x) = 1{|x|<1} ln |1 − ln |x||, x ∈ R2 . Note that u ∈ H 1 (R2 ), however
u∈/ C 0 (here, s − N/2 = 1 − 2/2 = 0). The proof of u ∈ H 1 (R2 ) goes as follows:
1
|u| dx =
2
|u| dx ≤ C
2
r| ln |1 − ln r|2 dr < ∞,
R2 {|x|<1} 0
5
We remind that if a space symbol is not followed by a domain, then it is understood that the
domain is RN .
107
and as ∂i u = 1{|x|<1} ∂i ln |1 − ln |x|| we get

1 1 |xi |2 1 dx
|∂i u| dx =
2
· 2· dx ≤ · 2
{|x|<1} |1 − ln |x|| |x| |x| {|x|<1} |1 − ln |x|| |x|
2 2 2
R2
1 1
1 dr
≤C · =C ((1 − ln r|)−1 ) dr
0 |1 − ln r|
2 r 0
< ∞, i = 1, 2.
6.3.2 Case Ω RN
We start with the definition of “extension” and “restriction” operators. They are
used to extend H s (Ω) in H s and then to restrict H s to H s (Ω), which makes it possible
to apply H s embedding results to H s (Ω) case.
Definition 6.3.6 Let Ω RN be an open set, s > 0 and p ∈ [1, ∞].
i) A linear continuous operator E (or EΩ ), E : W s,p (Ω) → W s,p , satisfying U = Eu,

U (x) = u(x) for all x ∈ Ω, is called a “W s,p (Ω) extension operator”.
We say that a domain Ω satisfies “W s,p extension property” if there exists a
W s,p (Ω) extension operator. If p = 2, we say that E is an “H s (Ω) extension
operator” and Ω satisfies the “H s extension property”.
ii) A linear continuous operator R (or RΩ ), R : W s,p → W s,p (Ω), satisfying u = RU ,

u(x) = U (x) for all x ∈ Ω, is called a “W s,p (Ω) restriction operator”.
The existence of the restriction operator R is straightforward from Definition 6.2.6,

which shows that R is well-defined, linear, and continuous. The existence of the extension
operator is a difficult problem that we will consider below. Note also that whenever E
exists, it is a right inverse of R.
Theorem 6.3.7 If Ω is a bounded Lipschitz domain, then it satisfies W s,p extension

property for all s > 0 and p ∈ [1, ∞].
For the proof of the theorem in the case s ∈ N and p ∈ [1, ∞], see [48]. For the proof in
the case 0 < s ∈/ N and p ∈ (1, ∞) and a proof by interpolation, see [54]. For a general
direct proof, see [45].
The theorem below gives a detailed instructive step-by-step proof in the case s = 1,
p ∈ [1, ∞], and Ω = RN + of the existence of an extension operator.
+ ) →
Theorem 6.3.8 There exists a linear continuous extension operator E : W 1,p (RN
W (R ), p ∈ [1, ∞].
1,p N
108
Proof. For u ∈ W 1,p (RN

+ ), let U = Eu be given by

u(x , xN ), xN ≥ 0,
U (x , xN ) = (6.3.4)
u(x , −xN ), xN < 0.
E is the required extension operator. Indeed, clearly E is linear. Let us prove that E is
well-defined and continuous.
(i) E is well-defined, so if u ∈ W 1,p (RN + ) then U ∈ W
1,p
(RN ).
(i.1) Let α = (α , αN ) ∈ NN0 , |α| = 1. Then for ϕ ∈ D(R ), we have
N

D U, ϕ = −
α
U (x)Dα ϕ(x)dx
N
R
= − u(x)Dα ϕ(x)dx + u(x , −xN )Dα ϕ(x , xN )dx
RN RN
+ −
= − u(x)Dα ψ(x)dx, (6.3.5)

RN
+
with
ψ(x , xN ) = ϕ(x , xN ) + (−1)αN ϕ(x , −xN ),
which we found after using the change of variable formulas (see Proposition 6.1.13)

u(x , −xN )Dx v(x , xN )dx = (−1)
α αN
u(y , yN )Dyα (v(y , −yN ))dy. (6.3.6)
RN
± RN
∓
(i.2) Note that in general ψ ∈ / D(RN α

+ ) and therefore, a priori, we cannot pass D on u in
N α
(6.3.5). However, ψ is “small” on ∂R+ and this is sufficient to pass D on u. We proceed
as follows. Let η ∈ C ∞ (R), η = 0 in (−∞, 1/2], η = 1 in [1, ∞) and set ηn (xN ) = η(nxN ).
Note that

α
uD ψdx = lim u(ηn Dα ψ)dx
RN n→∞ RN
+
+

= lim uDα (ηn ψ) − uψDα ηn dx. (6.3.7)
n→∞ RN
+
We claim that the limit of the second term in (6.3.7) is zero. Indeed, if αN = 0 then
Dα ηn = 0. If αN = 1 we get
Dα ηn (xN ) = nη (nxN ), ψ(x , 0) = 0, ψ(x , xN ) = xN Dα ψ(x , θxN ),
where θ = θ(x) ∈ (0, 1), which implies

α
uψD ηn dx = u(x)(xN Dα ψ(θx))(nη (nxN ))dx
RN
+ RN
+
109

≤ ηC 1 (R) ψC 1 (RN ) |u|dx
{|xN |≤1/n}
−−−→ 0, (6.3.8)
n→∞
where we used Lebesgue DCT and the fact that η (nxN ) = 0 for xn > n1 .
Therefore as ηn ψ ∈ D(RN
+ ), from (6.3.7) and (6.3.8), we get

α
uD ψdx = lim uD (ηn ψ)dx = − lim
α
(ηn ψ)Dα udx
RN n→∞ RN n→∞ RN
+
+ +
= − ψDα udx.
RN
+
(i.3) Finally, by combining (6.3.5) with the last equality and changing the variable again,
we obtain

α
D U, ϕ = ψDα udx
RN
+

= α
ϕD udx + (−1) αN
ϕ(x , −xN )Dα u(x , xN )dx
RN RN
+
+
= α
ϕD udx + ϕ(x , xN )Dα u(x , −xN )dx
RN
+ RN
−
= E(Dα u), ϕ.

(ii) Hence Dα (Eu) = E(Dα u), which implies Dα (Eu)Lp (RN ) ≤ 21/p uW 1,p (RN+ ) . Hence
Eu ∈ W 1,p (RN ) and E is continuous from W 1,p (RN
+ ) to W
1,p
(RN ).
The following theorem is analogous of Theorem 6.3.4 in H s (Ω).

Theorem 6.3.9 Assume Ω ⊂ RN is an open set satisfying H s extension property, s −
N
2
∈ R+ \N, m = s − N2 and σ ∈ (0, s − m). Then H s (Ω) −→ Cbm,σ (Ω). In particular,
there exists C > 0 such that for all α ∈ NN
0 , |α| = m, and x, y ∈ Ω, we have
|Dα u(x) − Dα u(y)| ≤ CDα uH s−m (Ω) |x − y|σ . (6.3.9)
Furthermore, if Ω is bounded then H s (Ω) −→

c
Cbm (Ω).
Proof. Let E : H s (Ω) → H s (RN ) be an extension operator and R : H k (RN ) → H k (Ω)
be the restriction operator (see Definition 6.3.6). From Theorem 6.3.4 and the continuity
of operators E and R, we have the following continuous embeddings:
H s (Ω) −→
E
H s (RN ) −→ Cbm,σ (RN ) −→
R
Cbm,σ (Ω).
These embeddings in combination with estimate (6.3.3) prove the embedding H s (Ω) −→
c
Cb (Ω) and the estimate (6.3.9). Furthermore, if Ω is bounded then Cb (Ω) −→Cb (Ω)
m m,σ c m
follows from Arzela-Ascoli lemma.
110
Theorem 6.3.10 Let k ∈ N and assume Ω is an open bounded set that satisfies the
H k+1 extension property. Then H k+1 (Ω) −→
c
H k (Ω). In particular, H0k+1 (Ω) −→
c
H k (Ω)
with Ω only open bounded.
Proof. For the proof of the first part, see Theorem 9.5.2. The second part follows from
Lemma 9.5.6 and the first part of this theorem.
Remark 6.3.11 Embedding theorems hold in a more general setting; see, for example,
[1, 14, 44, 51]. Namely, we have the following embeddings, where s > 0, p ∈ [1, ∞),
m ∈ N:
Ω = RN , or Ω RN Ω Lipschitz bounded
s,p
satisf ies W extension property
a) s − Np < 0 W s,p (Ω) −→ Lq (Ω), W s,p (Ω) −→c

Lq (Ω),
1
p̂
= p1 − Ns f or all q ∈ [p, p̂] f or all q ∈ [1, p̂)
b) s − N
p
=0 W s,p (Ω) −→ Lq (Ω), W s,p (Ω) −→c
Lq (Ω),
f or all q ∈ [p, ∞) f or all q ∈ [1, ∞)
c) s − Np = m + σ̂ W s,p (Ω) −→ Cbm,σ (Ω), W s,p (Ω) −→c

Cbm,σ (Ω),
σ̂ ∈ (0, 1) f or all σ ∈ [0, σ̂] f or all σ ∈ [0, σ̂)
d) s − N
p
=m W s,p (Ω) −→ Cbm−1,σ (Ω), W s,p (Ω) −→c
Cbm−1,σ (Ω),
f or all σ ∈ [0, 1) f or all σ ∈ [0, 1)
e)
s − Np ≥ σ − N
q
W s,p (Ω) −→ W σ,q (Ω) W s+,p (Ω) −→c
W σ,q (Ω),
p≤q f or all > 0.
Furthermore
f ) the continuous embeddings hold in W0s,p (Ω) instead of W s,p (Ω), with Ω only open,
g) the compact embeddings hold with W0s,p (Ω) instead of W s,p (Ω), with Ω only open
bounded.
The following remark shows that Lipschitz regularity of Ω in Theorem 6.3.7 is essential.
111
6.4. Boundary traces in Sobolev spaces Chapter 6
Remark 6.3.12 Theorem 6.3.7 does not hold in general if Ω is not Lipschitz. For exam-
ple, let Ω = {(x1 , x2 ), x1 ∈ (0, 1), x2 < xα1 } with α > 1. Clearly, Ω is not Lipschitz
because it has a cusp at the origin.
− σp
Let u(x) = x1 , p ≥ 1, σ > 0. Given furthermore m ∈ N, we can choose α such that
α − σ − mp + 1 > 0. Then u ∈ W m,p (Ω) because for all integer k ≤ m, we have
1 xα1
− σ −k p
|∂x1 u| dx = C(k, p, σ)
k p
x1 p dx2 dx1
Ω 0 0
1
C(k, p, σ)
= C(k, p, σ) xα−σ−kp dx1 = < ∞.
0
1
α − σ − kp + 1
/ L∞ (Ω). This means that Theorem 6.3.7 does not hold because if it does,
However, u ∈
from the embedding H 2 (R2 ) −→ Cb0 (R2 ) (see Theorem 6.3.3 with s = 2, p = 2, N = 2
so s − N/p = 1 > 0), we should have Eu ∈ C 0 (R2 ).
Example 6.3.13 Let Ω = (0, 1) ⊂ R and u(x) = x. Then u ∈ / H01 (Ω). Indeed, if we
assume that u ∈ H0 (Ω), then there exists un ∈ D(Ω) such that un → u in H 1 (Ω). Note
1
that we have H01 (Ω) −→ C 0 (Ω) because here s − N/p = 1 − 1/2 = 1/2 > 0. Therefore,
u − un C 0 (Ω) ≤ Cu − un H 1 (Ω) so u − un C 0 (Ω) → 0 as n → ∞. But u(1) = 1
and un (1) = 0 for all n, so u − un C 0 (Ω) ≥ |u(1) − un (1)| = 1, which contradicts
u − un C 0 (Ω) → 0 and proves the claim.
Theorem 6.3.14 Assume Ω ⊂ RN is an open set that satisfies the H s extension prop-
erty, s > 0. Then Cb∞ (Ω) −→ d
H s (Ω). More generally, if Ω ⊂ RN is an open set that
satisfies the W s,p extension property, s > 0, p ∈ [1, ∞) then Cb∞ (Ω) −→
d
W s,p (Ω).
Proof. For the first part, let u ∈ H s (Ω) and U ∈ H s , U = Eu, where E is a H s (Ω)
extension operator. From Theorem 6.3.2, there exists a sequence (Un ) in D(RN ) con-
verging to U in H s (RN ). It follows that (un ), un = RUn , is in Cb∞ (Ω) and converges to
u in H s (Ω).
The proof for W s,p (Ω) spaces is similar: extension in W s,p , next using the embedding
D(RN ) −→ d
W s,p (RN ) and conclude with the restriction to W s,p (Ω); see [24].
6.4 Boundary traces in Sobolev spaces

Every u ∈ W k,p (Ω) is defined almost everywhere, and in general is not a regular
function in the classical sense, for example u ∈ / C 0 (Ω). Therefore, a priori, it does not
make sense to talk about the restriction of u on ∂Ω. Trace theorems explain in what sense
the boundary values of functions in Sobolev spaces are understood, and they provide
important qualitative and quantitative properties of boundary traces. These theorems
are important when studying PDEs equipped with boundary conditions.
112
Chapter 6 6.4. Boundary traces in Sobolev spaces
Definition 6.4.1 Let Ω ⊂ RN be an open and bounded set. The restriction of u ∈

S(Ω) := {U |Ω , U ∈ S} on ∂Ω is called “trace of u on ∂Ω” and will be denoted by γ(u).
The map u ∈ S(Ω) → γ(u) is called a “trace operator”.
Trace theorem in Sobolev spaces essentially extends the operator γ in H s (Ω) spaces
(or more generally in W s,p (Ω)). Typically, the boundary trace of H s (Ω) functions is
constructed as follows. First, one defines the trace of H s (RN ) functions on {xN = 0} by
using Fourier transform. Next, for Ω general, the boundary trace of H s (Ω) functions is
constructed by using a partition of unity of ∂Ω, see Definition 1.2.1, and the trace results
on {xN = 0}. Traces of W s,p (Ω) functions are constructed similarly, but the proofs are
substantially more difficult and technical; see, for example, [14, 33].
Theorem 6.4.2 Let s − 12 > 0. The trace operator γ : S(RN ) → S(RN −1 ), defined by
γ(u)(x ) = u(x , 0), x ∈ RN −1 , extends to a linear continuous operator denoted with the
1
same letter γ, from H s (RN ) onto H s− 2 (RN −1 ).
Proof. For x, ξ ∈ RN , we will write x = (x , xN ), ξ = (ξ , ξN ) with x , ξ ∈ RN −1 . Let
U ∈ S(RN ) and set u(x ) = U (x , 0), uN (xN ; x ) = U (x , xN ) (uN is considered as a
function of xN ). Note that u ∈ S(RN −1 ) and uN ∈ S(R). Furthermore, if
Û (ξ) = Fx [U ](ξ), resp. û(ξ ) = Fx [u](ξ ), ûN (ξN ; x ) = FxN [uN (·; x )](ξN )
is the Fourier transform of U in RN , resp. of u in RN −1 , of uN in R, then

1
û(ξ ) = Û (ξ , ξN )dξN .
(2π)1/2 R
Indeed, this follows from Fourier and Fourier inversion formulas (5.4.3), (5.4.11) and the
following equalities:

1
ûN (ξN ; x ) = e−i(ξN xN ) U (x , xN )dxN ,
(2π)1/2 R

1
u(x ) = uN (0; x ) = ei(0ξN ) ûN (ξN ; x )dξN ,
(2π)1/2 R

1
û(ξ ) = (N −1)/2
e−i(ξ ·x ) u(x )dx
(2π) N −1
R
1 −i(ξ ·x ) 1 i(0ξN ) 1
= N −1 e 1 e 1 e−i(ξN xN ) U (x , xN )dxN dξN dx
(2π) 2 RN −1 (2π) 2 R (2π) 2 R

1
= 1 Û (ξ , ξN )dξN .
(2π) 2 R
Using inequalities (9.4.3), (9.4.4), we can evaluate it as follows:

1
u s− 21 N −1 =
2
ξ 2(s− 2 ) |û(ξ )|2 dξ
H (R ) RN −1
113
2
2(s− 21 )
≤C ξ Û (ξ , ξN )dξN dξ
N −1
R R
s− 21
(using H ölder) ≤ C ξ 2( ) ξ2s |Û (ξ , ξN )|2 dξN
RN −1 R

ξ−2s dξN dξ . (6.4.1)
R
For s − 1
2
> 0, we have
2 −s
ξN
ξ−2s dξN = ( ξ 2 + |ξN |2 )−s dξN = ξ −2s 1+ dξN
R R R ξ

−2(s− 12 )
= ξ (1 + r2 )−s dr
R
1
= ξ −2(s− 2 ) Is , (6.4.2)

where Is := R
(1 + r2 )−s dr < ∞. Then (6.4.1) implies

u2 s− 21 N −1
≤C ξ2s |Û (ξ , ξN )|2 dξN dξ ≤ CÛ 2L2s (RN )
H (R ) RN −1 R
≤ CU 2H s (RN ) . (6.4.3)
By density of S(RN ) in H s (RN ), see Theorem 6.3.2, the inequality (6.4.3) is valid for all
U ∈ H s (RN ) and this proves the theorem.
Remark 6.4.3 The condition s − 12 > 0 in Theorem 6.4.2 is used to show that Is < ∞.
One can ask whether this condition is due to the technique we used or it is critical. Actu-
ally, this condition is indeed critical. One can prove that for 0 < s ≤ 12 , the trace operator
cannot be extended to a continuous operator from H s (RN ) to L2 (RN −1 ). Furthermore,
for such s, the closure of D(Ω) in H s (Ω) is H0s (Ω); see [35].
Now we address the question of defining the trace of H s (Ω) functions on ∂Ω. We
expect that the range of the trace operator will be in a certain space H s (∂Ω). The follow-
ing definition addresses a more general boundary space W s,p (∂Ω); see for example [17].
Definition 6.4.4 Let 1 ≤ m ∈ N, Ω ⊂ RN of class C m−1,1 be open and bounded. Let

{(Gi , θi ), i = 1, . . . , M } be as in Definition 1.2.4, with {Gi , i = 1, . . . , M } a finite open
covering of ∂Ω, with θi ∈ C m−1,1 (Q; Gi ) a diffeomorphism such that θi (Q+ ) = Ω ∩ Gi ,
θi (Q0 ) = ∂Ω ∩ Gi ; see Fig. 6.4.1. Furthermore, let {ϕi , i = 0, 1, . . . , M } be a partition of
unity subordinate to {Gi , i = 1, . . . , M }; see Theorem 1.2.2. For 0 < s ≤ m, p ∈ [1, ∞)
we define
W s,p (∂Ω) = {u : ∂Ω → R, (ϕi u) ◦ θi ∈ W s,p (∂RN

+ ), i = 1, . . . , M }, (6.4.4)
114
Chapter 6 6.4. Boundary traces in Sobolev spaces
θi
∂Ω θi−1
G+
i
Q+
Q0
−1
Figure 6.4.1: G+ +
i , Q , θi , and θi
and its norm by

uW s,p (∂Ω) = (ϕi u) ◦ θi W s,p (∂RN+ ) ,
i=1,M
where (ϕi u) ◦ θi is the extension of (ϕi u) ◦ θi in ∂RN

+ by zero.
In the same spirit as in Remark 1.2.8, one can show that · W s,p (∂Ω) norms corre-
sponding to two different partitions of unity of ∂Ω are equivalent; see also [17, 46]. We
have this general result.
Theorem 6.4.5 Let Ω ⊂ RN be an open bounded set of class C 0,1 , s > 0, and p ∈ (1, ∞).
1
i) Case s − ∈ (0, 1). There exists a linear continuous trace operator with continu-
p
ous right inverse,
1
γ : W s,p (Ω) → W s− p ,p (∂Ω). (6.4.5)
1
ii) Case s − ∈ [1, ∞). Assume furthermore that Ω is of class C m,1 , m ≥ 1. Then
p
1 1
for s − ∈ ( − 1, m + 1)\N, if p = 2, and s − ∈ ( − 1, m + 1), if p = 2, for a
p p
certain ∈ N, there exists a linear continuous trace operator with continuous right
inverse,

−1
1
γ = (γ0 , γ1 , . . . , γ −1 ) : W s,p
(Ω) → W s−i− p ,p (∂Ω), (6.4.6)
i=0
such that for U ∈ S(Ω), if (u0 , u1 , . . . , u −1 ) = γ(U ) then ui (x) = ∂νi U (x), for all
x ∈ ∂Ω, i = 0, . . . , − 1. Here, ∂νi U (x) means the i-th order normal derivative of
i!
u, i.e. ∂νi U (x) = Dα U ν α , where ν is the unit outward normal vector to ∂Ω.
α!
|α|=i
115
For the proof of this theorem in the case s ∈ N, see [40], and in the case s ∈
/ N see [25, 37].
1
We note that the condition for s − in ii) is due to these facts: a) Besov space
p
s
Bp,p (Ω) is equal to W s,p (Ω) only if (s, p) ∈ ((0, ∞)\N) × (1, ∞) or (s, p) ∈ N × {2} (see
1 −1 s−i− 1 ,p
[24, 44]) and b) Claim ii) holds for all s − ∈ ( − 1, m + 1) with i=0 W p (∂Ω)
p
−1 s−i− p1
replaced by i=0 Bp,p (∂Ω); see [25, 37]. For more details on the boundary traces, see
[24, 25, 33, 37, 39, 40, 46, 48].
The boundary trace results allow us to integrate by parts with Sobolev space func-
tions, as in the case of smooth functions.
Theorem 6.4.6 (Integration by parts) Let Ω ⊂ RN be an open set and α ∈ NN
0 ,
|α| = 1.
i) If u ∈ W01,p (Ω) and v ∈ W 1,q (Ω), where p ∈ [1, ∞), 1/p + 1/q = 1, then

uD vdx = − vDα udx, |α| = 1.
α
(6.4.7)
Ω Ω
ii) If Ω is furthermore a Lipschitz bounded set then for u ∈ W 1,p (Ω) and v ∈ W 1,q (Ω),
with p ∈ (1, ∞), 1/p + 1/q = 1, we have

α
uD vdx = γ(u)γ(v)ν dσ − vDα udx,
α
(6.4.8)
Ω ∂Ω Ω
where ν = (ν1 , . . . , νN ) is the unit outward normal vector on ∂Ω.
Proof. To prove (6.4.7), let (un ) in D(Ω) such that limn→∞ un = u in W 1,p (Ω). Then

α
uD vdx = lim un Dα vdx = lim un , Dα v
n→∞ n→∞
Ω Ω

= − lim v, D un = − lim
α
vD un dx = − vDα un dx.
α
n→∞ n→∞ Ω Ω
To prove (6.4.8) we consider (un ) and (vn ) in C ∞ (Ω) converging respectively to u in

W 1,p (Ω) and v in W 1,q (Ω); see Theorem 6.3.14. Then

α
uD vdx = lim α
D un vn dx = lim un vn ν dσ − lim
α
un Dα vn dx
n→∞ n→∞ n→∞
Ω
Ω
∂Ω Ω
= γ(u)γ(v)ν α dσ − uDα vdx,

∂Ω Ω
where for the limit in the boundary integral we used the continuity of the trace operator.
116
Chapter 6 6.5. Poincaré inequality
6.5 Poincaré inequality

Poincaré inequality, which exists in a variety of forms, essentially states that in a
certain subspace of W 1,p , the Lp norm of u is bounded uniformly by the Lp norm of
∇u. This makes the Lp seminorm an equivalent norm, which is useful in a number of
problems.
Theorem 6.5.1 (Poincaré inequality) Let Ω be an open bounded set in at least one
direction, for example, Ω ⊂ {(x1 , . . . , xN ), |xi | ≤ d < ∞}, for a certain d > 0 and,
i ∈ {1, . . . , N }. Then for p ∈ [1, ∞) there exists C = C(p, Ω) > 0 such that the following
inequality, called “Poincaré inequality”, holds
∀u ∈ W01,p (Ω), uLp (Ω) ≤ C|u|W 1,p (Ω) , C = pd. (6.5.1)
Proof. By the density of D(Ω) in W01,p (Ω), it is enough to prove (6.5.1) only for
u ∈ D(Ω). Integration by parts gives

uLp (Ω) =
p
∂i xi |u(x)| dx = lim ∂i xi (u2 + 2 )p/2 dx
p
→0 Ω
Ω
= − lim pxi (u2 + 2 )(p−2)/2 u∂i udx
→0 Ω

u
= − lim pxi (u2 + 2 )(p−1)/2 2 ∂i udx
→0 Ω (u + 2 )1/2

(use Lebesgue DCT ) = − pxi |u|p−1 sgn(u)∂i udx
Ω
(use H ölder inequality) ≤ pd|u|p−1 p ∂i uLp (Ω) ; hence
L p−1 (Ω)
uLp (Ω) ≤ (pd)∂i uLp (Ω) , (6.5.2)
where sgn(x) = 2H(x) − 1, because |u|p−1 p = up−1

Lp (Ω) , which proves (6.5.1).
L p−1 (Ω)
We note that Poincaré inequality (6.5.1) can be generalized. For example, if Ω is

bounded in one direction, then for k ∈ N, p ∈ [1, ∞) we have
∀u ∈ W0k,p (Ω), uLp (Ω) ≤ C|u|W k,p (Ω) , C = C(k, p, Ω). (6.5.3)
One can prove the inequality by applying (6.5.1) to all Dα u, α ∈ NN

0 , |α| ≤ k − 1.
Inequality (6.5.1) implies the following straightforward result.
Corollary 6.5.2 Let p ∈ [1, ∞) and Ω ⊂ RN be an open set bounded in at least one
direction. Then |u|W 1,p (Ω) is a norm in W01,p (Ω), equivalent to uW 1,p (Ω) . Furthermore,
when p = 2, (u, v)H 1 (Ω) = Ω ∇u · ∇vdx defines an inner product equivalent to the usual
inner product (6.1.5).
117
6.6. H −s (Ω) and W −s,q (Ω) spaces Chapter 6
Example 6.5.3 Assume Ω ⊂ RN is a Lipschitz open bounded connected set and let

M0 (Ω) = u ∈ W (Ω),
1,p 1,p
udx = 0 , p ∈ [1, ∞).
Ω
Then Poincaré inequality (6.5.1) holds in M01,p (Ω). Let us prove the claim for p = 2.
Assume the claim does not hold. So
∀n ∈ N, ∃un ∈ M01,2 (Ω), un L2 (Ω) ≥ n∇un L2 (Ω) . (6.5.4)
un
Dividing (6.5.4) by un L2 (Ω) and setting vn = we get for all n:
un L2 (Ω)
vn ∈ M01,2 (Ω), 1 = vn L2 (Ω) ≥ n∇vn L2 (Ω) . (6.5.5)
So, the sequence (vn ) is bounded in H 1 (Ω). As H 1 (Ω) −→ c

L2 (Ω), there exists a subse-
quence of (vn ) still denoted by (vn ), converging in L2 (Ω) to a certain v ∈ H 1 (Ω). From
(6.5.5), we get ∇vn → 0 in L2 (Ω), which implies that vn → v in H 1 (Ω) and ∇v = 0.
Necessarily we have
i) v is constant because ∇v = 0,

ii) v ∈ M01,2 (Ω) because from vn ∈ M01,2 (Ω) we have Ω vn dx = 0, which from the
convergence vn → v in L2 (Ω) implies v ∈ M01 (Ω),
iii) vL2 (Ω) = 1 because passing to the limit in vn 2L2 (Ω) = 1 implies vL2 (Ω) = 1.
But then we have a contradiction because from i) and ii) we get C = 0, and then iii)
becomes impossible. This proves that Poincaré inequality holds in M01,2 (Ω).
6.6 H −s(Ω) and W −s,q (Ω) spaces

When studying PDEs, often one has to consider elements of (H k (Ω)) , where in
general E is the topological dual space of the Banach space E. One would wonder
whether elements of (H k (Ω)) could be characterized by elements of D (Ω) which are
continuous for the H k (Ω) norm. The following lemma shows that D(Ω) is not dense in
H k (Ω). It implies that if ∈ (H k (Ω)) then cannot be characterized by any T ∈ D (Ω)
continuous for the H k (Ω) norm.
Roughly speaking, negative order Sobolev spaces H −s (Ω) are subspaces of (H s (Ω)) ,
elements of which are characterized by elements of D (Ω), continuous for the H s (Ω)
norm.
Lemma 6.6.1 Let Ω ⊂ RN be open, bounded. The set D(Ω) is not dense in H k (Ω),
k ∈ N.
Proof. Let 0 = u ∈ C 2k (Q) with Q = (−a, a)N such that Ω Q. Clearly u ∈ H k (Ω).
Define
L : H k (Ω) → D (Ω),
118
Chapter 6 6.6. H −s (Ω) and W −s,q (Ω) spaces

|α|
Lu, ϕ := (−1) D u, ϕ 2α
= Dα uDα ϕdx, ϕ ∈ D(Ω).
|α|≤k |α|≤k Ω
If u is chosen such that

(−1)|α| D2α u = 0 in Ω, (6.6.1)
|α|≤k
then Lu = 0. Note that (6.6.1) has nonzero solutions, for example u(x) = Re(eλx1 ), with
λ = 0 solution to 0≤α1 ≤k (−1)α1 λ2α1 = 0.
Now we can prove that D(Ω) is not dense in H k (Ω). If we assume D(Ω) is dense in
H k (Ω) then there exists ϕn ∈ D(Ω), lim ϕn = u in H k (Ω). Therefore
n→∞

0 = lim Lu, ϕn = lim α α
D uD ϕn dx = |Dα u|2 dx > 0,
n→∞ n→∞ Ω Ω
|α|≤k |α|≤k
which is a contradiction and proves the lemma.
This lemma with the comments preceding it motivates the following definition.
Definition 6.6.2 Let s > 0, p ∈ [1, ∞), q ∈ (1, ∞], 1/p + 1/q = 1, and Ω ⊂ RN . We
define6 W −s,q (Ω) as the dual space of W0s,p (Ω). If Ω is ∂RN
+ or a C
s +1
domain, then
−s,q s,p −s
W (∂Ω) denotes the dual space of W (∂Ω). When p = 2, we write H (·) instead of
W −s,2 (·).
Note that the spaces W −s,q (Ω) and W −s,q (∂Ω) equipped with their dual norms are
Banach spaces. Note also that p = ∞ is excluded from the previous definition, because
the closure of D(Ω) in W s,∞ (Ω) is a subspace of C s (Ω).
Furthermore, D(Ω) −→ d
W0s,p (Ω) implies that W −s,q (Ω) can be identified by a sub-
space of distributions in D (Ω) continuous with respect to · W s,p (Ω) norm, which is not
the case for the dual space of W s,p (Ω).
Remark 6.6.3 Let s > 0. In analogy to the characterization of H s (RN ) by L2s (RN ), we
have H −s (RN ) = (H s (RN )) , i.e.

−s −2s ˆ 22
H (R ) = ∈ S , H −s (RN ) =
N 2
ξ |F[]| dξ =:
2
L−s <∞ .
RN
For the proof see for example [12, 35].
6
See Remark 6.6.3 for the motivation of the term −s.
119
6.6. H −s (Ω) and W −s,q (Ω) spaces Chapter 6
Problems
Problem 6.1 Let Ω ⊂ RN be an open set, u ∈ W k,p (Ω), k ∈ N, p ∈ [1, ∞], and ũ a
function in Ω such that the N -dimensional Lebesgue measure of the set {x ∈ RN , ũ(x) =
u(x)} is zero. Prove that ũ ∈ W k,p (Ω) and ũW k,p (Ω) = uW k,p (Ω) .
Problem 6.2 Let H = 1(0,∞) be the Heaviside function. Prove that H ∈
/ W 1,p (Ω), where
p ∈ [1, ∞], Ω ⊂ R is an open set, 0 ∈ Ω.
Problem 6.3 Let B be the unit ball in RN . Prove that u(x) = ln(1 − ln |x|) ∈ W 1,p (B)
iff p < N .
Problem 6.4 Let h : R → R be a C 1 piecewise smooth function, i.e. h is C 0 in R,
h(0) = 0, h ∈ L∞ (R), and h is continuous in R except in a finite number of points
Z where side derivatives exist. Prove that h ◦ u ∈ H 1 (Ω), Ω ⊂ RN open, for every
u ∈ H 1 (Ω), and

∂i (h ◦ u) = 1{x∈Ω, / h (u)∂i u,
u(x)∈Z} i = 1, . . . , N.
Problem 6.5 Assume u ∈ H 1 (R) and set uh (x) = 1
h
(u(x + h) − u(x)). Prove that
lim uh = u in L2 (R).
h→0
Problem 6.6 Find where, and explain why, Ω in Lemma 6.6.1 is required to be bounded.
Prove that, in general, D(Ω) is not dense in H 1 (Ω) if Ω RN is unbounded.
Problem 6.7 Show that H01 (R\{0}) H 1 (R).
Problem 6.8 Let N ≥ 3, u ∈ H 1 (RN ) and f (t, x) = (fi (t, x)), t ∈ R, x ∈ RN , where
xi
fi (t, x) = ∂i u + tu 2 . Show that
|x|

u2
0≤ |f (t, x)| dx =
2
|∇u| dx + (t − (N − 2)t)
2 2
dx,
RN |x|
2
RN RN
and deduce that

u u 2
∈ H 1 (RN ) and ≤ ∇uL2 (RN ) .
|x| |x| L2 (RN ) N −2
As a corollary of this result, prove7 that H01 (RN \{0}) = H 1 (RN ).
Problem 6.9 Let Ω ⊂ RN be an open bounded Lipschitz set, s > 0, p ∈ [1, ∞), and
define

W̃ s,p (Ω) = u : uW̃ s,p (Ω) := inf{U W s,p (RN ) , U |Ω = u} .
Show that W s,p (Ω) = W̃ s,p (Ω) and the norms · W s,p (Ω) , · W̃ s,p (Ω) are equivalent. Show
furthermore that the statement does not hold if Ω is not bounded Lipschitz.
7
This statement is true even for N = 2, but the proof is more delicate; see [51].
120
Chapter 6 6.6. H −s (Ω) and W −s,q (Ω) spaces
Problem 6.10 Prove that there exist open bounds sets Ω ⊂ RN such that C 1 (Ω) is not
dense in H 1 (Ω).
Problem 6.11 Prove Theorem 6.1.9.
Problem 6.12 Show that the embedding H k+1 (Ω) −→ H k (Ω) is not compact, in gen-
eral, if Ω ⊂ RN is only an open bounded set.
Problem 6.13 Let u ∈ W 1,p (Ω), p ∈ [1, ∞], Ω = (a, b) ⊂ R. Prove that there exists
ũ ∈ W 1,p (Ω) ∩ C 0 (Ω), ũ = u a.e. in Ω and
x
∀x, y ∈ Ω, ũ(x) − ũ(y) = u (s)ds.
y
Problem 6.14 Prove that if u ∈ H 1 (R) then u cannot have jump discontinuities.
Problem 6.15 Prove that H 1 (RN ) C 0 (RN ) for N ≥ 2, i.e. there exists u ∈ H 1 (RN )
such that u ∈
/ C 0 (RN ).
Problem 6.16 Let Ω ⊂ RN be an open bounded Lipschitz set and s > r ≥ 0. Prove that
H s (Ω) −→
c
H r (Ω).
Problem 6.17 Let k ∈ N. Prove that the embedding H k+1 (RN ) −→ H k (RN ) is not
compact.
Problem 6.18 Let Ω ⊂ RN be an open bounded set. Prove that the embedding H01 (Ω) −→
H −1 (Ω) is compact.
Problem 6.19 Prove that Poincaré inequality does not hold in H 1 (RN ).
Problem 6.20 Let Ω ⊂ R2 be an open bounded Lipschitz connected set, γ ⊂ ∂Ω with
nonzero length, and
Wγ1,p (Ω) = {u ∈ W 1,p (Ω), u = 0 on γ}, p ∈ [1, ∞).
Prove that Poincaré inequality (6.5.3) holds in Wγ1,p (Ω).
Problem 6.21 Let Ω ⊂ RN be an open bounded Lipschitz connected set, ω ⊂ Ω with
nonzero measure, and
Wω1,p (Ω) = {u ∈ W 1,p (Ω), u = 0 on ω}, p ∈ [1, ∞).
Prove that Poincaré inequality (6.5.3) holds in Wω1,p (Ω).
N
Problem 6.22 Let k − > 0. Prove that δ0 ∈ H −k (RN ).
2
Problem 6.23 Let Ω ⊂ RN be of class C 1 , g ∈ H 1 (Ω), and ∈ L(H 1 (Ω)) defined
by (v) = g(x)v(x)dσ, v ∈ H 1 (Ω). Estimate the norm of seen as an element of
∂Ω
H −1 (Ω) or of L(H 1 (Ω)).
121
7. Second-order linear elliptic PDEs:
weak solutions
There exist several well-developed theories of elliptic linear PDEs of order two. We
have already seen Perron’s method for Laplace equation in Chapter 4, which provides
classical solutions. In this context, Schauder theory provides C 2,α classical solutions; see,
for example, [20].
Other methods, such as Hilbert spaces, variational, singular integral, or spectral
methods provide the so-called “weak solutions”, which belong to certain Sobolev spaces;
see, for example, [5, 16, 20, 35].
In this chapter, we will only deal with the existence and uniqueness of weak solutions
obtained by the variational method. By combining these results with some compactness
results in Sobolev spaces, we will also address the solution to certain nonlinear elliptic
PDEs.
7.1 Introduction
The objective of this chapter is to study the existence and uniqueness of a weak
solution u : Ω → R to the following problem:

N
N
Lu := − ∂j (aij ∂i u) + bi ∂i u + cu = f in Ω, (7.1.1a)
i,j=1 i=1
u = 0 on ∂Ω. (7.1.1b)
This problem is commonly referred to as the “Dirichlet problem”, which is in connection
with the boundary condition u = 0. One may consider the general Dirichlet boundary
condition u = g; see Remark 7.2.13.
Here and throughout this chapter, unless otherwise stated, Ω ⊂ RN is an open
bounded set, and aij , bi , and c satisfy
2
A := (aij ) = (aji ) ∈ W 1,∞ (Ω; RN ), b := (bi ) ∈ L∞ (Ω; RN ), c ∈ L∞ (Ω), (7.1.2a)

N
∃k, K > 0, ∀0 = ξ ∈ R , k|ξ| ≤ N 2
aij (·)ξi ξj ≤ K|ξ|2 , a.e. in Ω. (7.1.2b)
i,j=1
https://doi.org/10.1007/978-3-031-39524-6 7
7.1. Introduction Chapter 7
The assumptions for f will be provided later in the context.

We note that even in the case when L = −Δ, there are cases when the problem
Lu = 0 in Ω, u = g on ∂Ω does not have a classical solution; see problem (4.4.3). We
will revisit this problem in Example 7.2.14 and show that it has a unique weak solution.
There are three main questions relative to (7.1.1).
⎧
⎨ Q1. How does one understand the solution to (7.1.1)?
Q2. Does (7.1.1) have a solution? If yes, is the solution unique? (7.1.3)
⎩
Q3. How regular is the solution to (7.1.1)?
In this chapter we will address the first two questions. The third question is a classical
topic involving fine techniques and can be found in numerous PDE books, such as [5, 20].
For the ease of the reader, we have included the solution to Q3 in Section 9.6 in Annex.
Definition 7.1.1
i) Assume f ∈ C 0 (Ω). A strong, or classical, solution to (7.1.1) is every u ∈ C 0 (Ω) ∩
C 2 (Ω) satisfying (7.1.1) pointwise.
ii) Assume f ∈ H −1 (Ω). A weak solution to (7.1.1) is every u ∈ H01 (Ω) satisfying
N
N
aij ∂i u∂j v + bi ∂i uv + cuv = f v, ∀v ∈ H01 (Ω). (7.1.4)
Ω i,j=1 Ω i=1 Ω Ω
The equation (7.1.4) is called “the weak form equation of (7.1.1)”.

The motivation for (7.1.4) is as follows. The operator L defines a linear continuous
map L : H01 (Ω) → H −1 (Ω). Indeed, for u ∈ H01 (Ω) define Lu ∈ D (Ω) by

Lu, v := ∂j (−aij ∂i u), v + bi ∂i u, v + cu, v
i,j=1,N i=1,N

= aij ∂i u∂j v + bi ∂i uv + cuv dx
Ω i,j=1,N i=1,N
=: B(u, v). (7.1.5)
By density of D(Ω) in H01 (Ω), (7.1.5) is well-defined for all u, v ∈ H01 (Ω) and satisfies
Lu, v = B(u, v), | Lu, v | ≤ C u H 1 (Ω) v H 1 (Ω) , (7.1.6)
where C = C(A, b, c), which proves L ∈ L(H01 (Ω); H −1 (Ω)) and B ∈ L(H01 (Ω)×H01 (Ω)).
Actually, B defines a continuous bilinear form in H 1 (Ω), i.e. (7.1.6) holds for all u, v ∈
H 1 (Ω). Therefore, (7.1.4) is nothing but Lu, v = B(u, v) = f, v . It solves (7.1.1a) in
a weak sense, and satisfies (7.1.1b) in the sense of traces γ(u) = 0 because u ∈ H01 (Ω).
Finally, if u is a classical solution with u ∈ C 1 (Ω), (7.1.5) is obtained by multiplying
the left-hand side of (7.1.1a) by v, integrating by part, and canceling boundary integrals
by using u = 0 on the boundary. Hence, (7.1.4) gives a necessary condition for a strong
solution u.
124
Chapter 7 7.2. Existence and uniqueness of weak solutions
7.2 Existence and uniqueness of weak solutions

The existence of solutions to second-order linear elliptic PDEs relies on certain results
of Functional Analysis, such as Lax-Milgram lemma and the theory of (linear) compact
operators, as well as on a weak maximum principle in the context of Sobolev spaces,
which we present in the following section.
7.2.1 Preliminary results

We will give some results of Functional Analysis. For their proof, the interested
reader can see, for example, [20, 29]. We will also give the weak maximum principle in
Sobolev spaces context and some other auxiliary results.
Theorem 7.2.1 (Riesz representation theorem) Let H be an Hilbert space with
inner product (·, ·)H and ∈ H . Then there exists a unique u ∈ H such that
∀v ∈ H, , v H ×H = (u, v)H , u H = H . (7.2.1)
The following two theorems are the main ingredients for solving (7.1.1). While The-
orem 7.2.2 solves (7.1.1) in special cases, Theorem 7.2.3 gives a full characterization of
the weak solutions to (7.1.1).
Theorem 7.2.2 (Lax-Milgram lemma) Let H be a Hilbert space and B : H ×H → R
be a bilinear form satisfying
(continuity) |B(u, v)| ≤ K u H v H, ∀u, v ∈ H, (7.2.2a)

(ellipticity) k u 2H ≤ B(u, u), ∀u ∈ H, (7.2.2b)
with certain K, k > 0. Then, for every f ∈ H there exists a unique u ∈ H such that
B(u, v) = f, v H ×H , ∀v ∈ H, (7.2.3)
1
u H ≤ f H . (7.2.4)
k
Theorem 7.2.3 (Fredholm alternative) Let H be an Hilbert space and A : H → H,
a compact linear operator. Then all the following hold.
(i) N (I − A) = {v ∈ H, (I − A)v = 0}, the kernel of I − A, has finite dimension.
(ii) R(I − A) = (N (I − A∗ ))⊥ , where R(I − A) = {(I − A)v, v ∈ H} is the “range

of I − A”, A∗ is the “adjoint of A”, defined by (A∗ u, v)H = (u, Av)H , and (N (I −
A∗ ))⊥ = {w ∈ H, (w, v)H = 0, ∀v ∈ N (I − A∗ )}.
(iii) R(I − A) = H iff N (I − A) = {0}.
(iv) dim(N (I − A)) = dim(N (I − A∗ )).
125
7.2. Existence and uniqueness of weak solutions Chapter 7
The following definition gives a generalization of concepts of sub/super harmonic

functions related to the operator L and defines the concept of inequality on the boundary
in the context of Sobolev spaces.
Definition 7.2.4 Let u, w ∈ H 1 (Ω), v ∈ H01 (Ω), and Lu, v defined by (7.1.5).1
i) We say Lu ≤ 0 if Lu, v ≤ 0 for all 0 ≤ v ∈ H01 (Ω). Similarly, we say Lu ≥ 0 if

Lu, v ≥ 0 for all 0 ≤ v ∈ H01 (Ω).
ii) We say u ≤ w on ∂Ω if (u − w)+ := max{u − w, 0} ∈ H01 (Ω). Similarly, u ≥ w

on ∂Ω if (u − w)− := min{u − w, 0} ∈ H01 (Ω).
Now we give the “weak maximum principle” associated with the operator L in the
context of Sobolev spaces. It generalizes the classical weak maximum principle that we
have seen in Chapter 4.
Theorem 7.2.5 (Weak maximum principle) Let Ω ⊂ RN be an open bounded set,
c ≥ 0 and u ∈ H 1 (Ω) such that Lu ≤ 0 (Lu ≥ 0). Then
sup u(x) ≤ sup u+ (x), inf u(x) ≥ inf u− (x) ,

x∈Ω x∈∂Ω x∈Ω x∈∂Ω
where sup and inf are taken for x almost everywhere in Ω or on ∂Ω.
Proof. Consider first the case Lu ≤ 0. If M = supx∈∂Ω u+ (x) and m = supx∈Ω u(x) we
have to prove m ≤ M . Assume the claim does not hold, i.e. M < m. The proof continues
strongly using Lu, v ≤ 0 with some particular functions v, and Sobolev embeddings
in Remark 6.3.11.
Let r ∈ [M, m). From Example 6.1.14 we have w := (u − r)+ ∈ H 1 (Ω). Actually,
w ∈ H01 (Ω) for a certain r ∈ [M, m). Indeed, there exists r ∈ [M, m) such that
inf{|x − y|, x ∈ supp((u − r)+ ), y ∈ ∂Ω} > 0,
because if not we would get M = m, which by assumption is excluded. If we denote

still by w the extension of w in RN by zero, and consider wn = w ∗ ρn , with (ρn )
a mollifier sequence, then wn |Ω ∈ D(Ω), for n large, and converges to w in H 1 (Ω).
So (u − r)+ ∈ H01 (Ω). Furthermore, (u − s)+ ∈ H01 (Ω) for all s ∈ [r, m), because
supp((u − s)+ ) ⊂ supp((u − r)+ ).
Now we take v = (u − s)+ , s ∈ [r, m). As ∈ H01 (Ω) we have
N
N

aij ∂i u∂j v = Lu, v − bi ∂i uv − cuv ≤ C |v||∇u|, (7.2.5)
i,j=1 Ω i=1 Ω Ω Ω
1
Which is well-defined for all u, v ∈ H 1 (Ω).
126
because Lu, v ≤ 0. From (6.1.17) we have ∇v = (∇u)1{u>s} . Then (7.2.5) implies

N

aij ∂i v∂j v ≤ C v|∇v|, ωs = supp(∇v) ⊂ supp(v),
i,j=1 Ω ωs
which combined with (7.2.2b) yields k ∇v 2

L2 (Ω) ≤C v L2 (ω) ∇v L2 (Ω) , or equivalently
∇v L2 (Ω) ≤C v L2 (ωs ) , C = C(L, k, Ω). (7.2.6)
Now we use Sobolev embedding results, Remark 6.3.11, a) for N ≥ 3 and b) for N = 2.
For N ≥ 3 we obtain
v 2N ≤ v 2N (use P oincaré inequality,

L N −2 (ωs ) L N −2 (Ω)
≤ C1 ∇v L2 (Ω) (7.2.6) and

1/2
≤ C2 ωs
1v 2 H ölder inequality (9.4.1))
1/2
≤ C2 1 LN/2 (ωs ) v 2
N
L N −2 (ωs )
≤ C2 |ωs |1/N v 2N ,
L N −2 (ωs )
which gives C −N ≤ |ωs |. For N = 2 we obtain C −p ≤ |ω|, for any p > 2. In both cases
C does not depend on s. Setting C0 = min{C −N , C −p }, we get
0 < C0 ≤ |ωs | = |supp(∇(u − s)+ )| ⊂ supp((u − s)+ ), s ∈ [r, m). (7.2.7)
Letting s to m in (7.2.7) implies m < ∞, because, if not, from u ≥ s in ωs we get

u ∈
/ L2 (Ω). Therefore (7.2.7) holds for s = m, which implies that u attains its finite
supremum m in ωm , where at the same time necessarily ∇u = 0 and so |ωm | = 0,
because u = m in ωm (see (6.1.17), Example 6.1.14). This contradicts (7.2.7) for s = m
and proves the case Lu ≤ 0. The case Lu ≥ 0 is treated by considering L(−u) ≤ 0.
This theorem implies the following uniqueness result.

Corollary 7.2.6 Let c ≥ 0 and u ∈ H01 (Ω) be a weak solution to (7.1.1) with f = 0.
Then u = 0. So, (7.1.4) cannot have more than one solution.
We will see that in the case c ≥ 0, the existence of a weak solution to (7.1.1) is proved
easily by using Lax-Milgram lemma. The case c arbitrary is addressed by using Fredholm
alternative theorem and the following lemma.
Lemma 7.2.7 There exists k0 , K0 > 0, and λ0 ≥ 0 such that for all u, v ∈ H 1 (Ω) we have
|B(u, v)| ≤ K0 u H 1 (Ω) v H 1 (Ω) , (7.2.8)
127
k0 |u|2H 1 (Ω) ≤ B(u, u) + λ0 u 2

L2 (Ω) . (7.2.9)

If b = 0 and c ≥ 0 we can take k0 = k and λ0 = 0. Otherwise k0 = k − b L∞ (Ω) and

λ0 = (1/(4 )) b L∞ (Ω) + c− L∞ (Ω) , where b L∞ (Ω) = N i=1 bi L∞ (Ω) and > 0 small
so that k0 > 0.
Proof. Clearly, (7.2.8) follows by using Hölder inequality to all the terms of B(u, v).
For (7.2.9), clearly it holds with λ0 = 0 if b = 0 and c ≥ 0. Otherwise, from ellipticity
condition (7.1.2b), by using (9.4.2) for > 0 we have
N
N

k|u|2H 1 (Ω) ≤ aij ∂i u∂j u = B(u, u) − bi u∂i u − cu2
0
i,j=1 Ω i=1 Ω Ω
≤ B(u, u)
1
+ b L∞ (Ω) u 2L2 (Ω) + ∇u 2
L2 (Ω)
4
+ c− L∞ (Ω) u 2L2 (Ω) .
It implies

k− b L∞ (Ω) |u|2H 1 (Ω) ≤ B(u, u)
0
1
+ b L∞ (Ω) + c− L∞ (Ω) u 2
L2 (Ω) ,
4
which proves the lemma.
7.2.2 Dirichlet problem

We start with a simple existence result.
Theorem 7.2.8 Let Ω ⊂ RN be an open bounded set, f ∈ H −1 (Ω), b = 0, and c ≥ 0.
Then (7.1.1) has a unique weak solution u ∈ H01 (Ω) which satisfies
1
u H 1 (Ω) ≤ f H −1 (Ω) . (7.2.10)
k
Proof. Lemma 7.2.7 holds with λ0 = 0. Then it is clear that B(u, v) satisfies the
conditions of Lax-Milgram lemma, Theorem 7.2.2, which proves the theorem.
Example 7.2.9 Let Ω ⊂ RN be an open bounded set, c ∈ L∞ (Ω), and consider the
problem
−Δu + cu = f in Ω, u = 0 on ∂Ω. (7.2.11)
The weak form equation of this problem is given by

B(u, v) = (∇u · ∇v + cuv)dx, u, v ∈ H01 (Ω). (7.2.12)
Ω
128
Clearly, if c ≥ 0, Theorem 7.2.8 implies that (7.2.11) has a unique weak solution. In
fact, one can prove that there exists 0 > 0 such that if c− L∞ (Ω) < 0 then still (7.2.11)
has a unique weak solution. Indeed, using Poincaré inequality in H01 (Ω) we get

B(u, u) = (|∇u|2 + (c+ + c− )u2 )dx
Ω
≥ |∇u| dx +c− u2 dx
2
Ω Ω
−
≥ |∇u| − c L∞ (Ω) u2 dx
2
(use (6.5.1))
Ω Ω
≥ |∇u|2 − C 2 c− L∞ (Ω) |∇u|2

Ω
= (1 − C 2 c− 2
L∞ (Ω) )|∇u|H 1 (Ω) , (7.2.13)
0
where C is the (Poincaré) constant in (6.5.1). Inequality (7.2.13) shows that B is elliptic
in H01 (Ω) if c− L∞ (Ω) < C −2 . Then the existence and uniqueness of a weak solution
follows from Lax-Milgram lemma.
In the case when c 0 or b = 0, the issue of existence of solutions to (7.1.4) is more
complex. We will solve it as follows. First we will consider an auxiliary problem L0 u = f ,
f ∈ H −1 (Ω), L0 u = Lu + λ0 u, which, thanks to Lax-Milgram lemma, has a unique weak
solution. The operator L0 generates a map G0 : H01 (Ω) → H01 (Ω), which is compact.
Then, the equation (7.1.4) is written equivalently in the form (I − G0 )u = L−1 0 f , which
is analyzed by using Fredholm alternative Theorem 7.2.3.
More precisely, let L0 : H01 (Ω) → H −1 (Ω) and B0 ∈ L(H01 (Ω) × H01 (Ω)) be given by

N
N
L0 u := Lu + λ0 u = − ∂j (aij ∂i u) + bi ∂i u + (c + λ0 )u, (7.2.14)
i,j=1 i=1
B0 (u, v) = L0 u, v H −1 ×H01 (7.2.15)

= B(u, v) + λ0 (u, v)L2 (Ω)
N N

= aij ∂i u∂j v + bi ∂i uv + (c + λ0 )uv, ∀v ∈ H01 (Ω),
i,j=1 Ω i=1 Ω Ω
where λ0 is introduced in Lemma 7.2.7.

Before we address the solution to (7.1.1) in general, we present the following technical
lemmas.
Lemma 7.2.10 The equation L0 u = f , f ∈ H −1 (Ω), has a unique weak solution u ∈
H01 (Ω) satisfying
L0 u, v H −1 (Ω)×H 1 (Ω) = B0 (u, v) = f, v H −1 (Ω)×H01 (Ω) , ∀v ∈ H01 (Ω),
u H 1 (Ω) ≤ C f H −1 (Ω) . (7.2.16)
129
Hence, the maps L0 : H01 (Ω) → H −1 (Ω) and L−1

0 : H
−1
(Ω) → H01 (Ω) are linear, invert-
ible, and bounded.
Proof. Lemma 7.2.7 implies that B0 is continuous and elliptic in H01 (Ω), and therefore
the results follow from Lax-Milgram lemma.
The following lemma uses some properties of compact operators, and the reader can
consult [29] for more details about compact operators.
Lemma 7.2.11 Assume λ0 of Lemma 7.2.7 is positive, and let I : H01 (Ω) → H −1 (Ω)
be the embedding map, Iu = u, and G0 := λ0 L−1
0 ◦ I : H0 (Ω) → H0 (Ω). Then G0 is
1 1
compact.
Proof. Clearly G0 is linear. For the compactness, we write I = I0 ◦ I1 , where I1 :
H01 (Ω) → L2 (Ω), I0 : L2 (Ω) → H −1 (Ω), I1 u = I0 u = u. So G0 = λ0 L−1
0 ◦ I0 ◦ I1 , with
−1
L0 , I0 , I1 continuous, and I1 compact; see Theorem 6.3.10. Hence, G0 is compact as a
composition of a compact operator with a continuous operator. This proves (i).
Theorem 7.2.12 If c ≥ 0 then the problem (7.1.4) has a unique solution u ∈ H01 (Ω).
Otherwise we have if N (L) = {0} then (7.1.4) has a unique solution, and if N (L) = {0}
then (7.1.4) has a solution (and then a finite number of solutions) iff f ∈ R(L).
Proof. We note that (7.1.4), which we refer to as “Lu = f ”, is equivalent to (I −G0 )u =

f0 with f0 = L−10 f . In the following part of the proof, we will use Theorem 7.2.3 with
1
H = H0 (Ω) and A = G0 .
If c ≥ 0 and b = 0, Theorem 7.2.8 shows that Lu = f has a unique solution. If c ≥ 0
and b = 0, necessarily λ0 > 0. It implies N (I − G0 ) = {0}, because (I − G0 )u = 0
is equivalent to Lu = 0, which from Corollary 7.2.6 implies u = 0. Then from (iii),
Theorem 7.2.3, we get that (I − G0 )u = f0 has a unique solution, so Lu = f has a
unique solution.
Otherwise, in the case c 0 necessarily λ0 > 0. If N (I − G0 ) = {0}, like above,
Lu = f has a unique solution from (iii), Theorem 7.2.3. Finally, if N (I − G0 ) = {0},
(ii) Theorem 7.2.3 implies that (I − G0 )u = f0 has a solution (and in this case, a
finite number of weak solutions) if and only if f0 ∈ N (I − G∗0 )⊥ , which is equivalent to
f ∈ R(L) and completes the proof.
Remark 7.2.13 Consider the problem
f ind u ∈ H 1 (Ω), Lu = f in Ω, γ(u) = γ(g) on ∂Ω, (7.2.17)
where f ∈ H −1 (Ω), g ∈ H 1 (Ω), γ is the boundary trace operator, and Ω ⊂ RN an open

bounded Lipschitz set. A function u ∈ H 1 (Ω) is a weak solution to (7.2.17) if
Lu, ϕ = f, ϕ , ∀ϕ ∈ H01 (Ω).
130
If we look for u as u = v + g, then
v ∈ H01 (Ω), Lv, ϕ = f − Lg, ϕ H −1 (Ω)×H01 (Ω) , ∀ϕ ∈ H01 (Ω). (7.2.18)
We can apply Theorem 7.2.12 to (7.2.18), and the conclusions of this theorem hold for
(7.2.18) with f − Lg instead of f .
Example 7.2.14 Let us consider again the problem (4.4.3). It does not have a classical
solution because the boundary data is not continuous. However, it has a unique H 1 (Ω)
solution. Indeed, if G(x) = 1{|x|<1} ln |1 − ln |x|| then u = G on ∂Ω. As from Remark
6.3.5, ln |1 − ln |x|| ∈ H 1 (Ω), from Remark 7.2.13 the problem (4.4.3) has a unique
solution in H 1 (Ω).
7.2.3 Neumann problem

Using the method we developed in this chapter, we can solve the following “Neumann
problem”, which is similar to (7.1.1) but with the boundary condition (7.1.1b) replaced
by the normal derivative of u, commonly referred to as the “Neumann boundary condi-
tion”:

N
N
− ∂j (aij ∂i u) + bi ∂i u + cu = f ∈ (H 1 (Ω)) , (7.2.19a)
i,j=1 i=1

aij ∂i uνj = 0 on ∂Ω, (7.2.19b)
j=1,N
where Ω ⊂ RN is an open bounded Lipschitz set and ν = (ν1 , . . . , νN ) is the exterior

unit normal vector to ∂Ω. Note that by assuming u is smooth and (aij ) is the identity
matrix, the condition (7.2.19b) becomes ∇u · ν = 0.
The weak form equation of (7.2.19) is given by
B(u, v) = f, v (H 1 (Ω)) ×H 1 (Ω) , v ∈ H 1 (Ω), (7.2.20)
which formally is obtained by multiplying (7.2.19) by v ∈ H 1 (Ω), integrating by parts

in Ω, and canceling the boundary term by using the boundary condition (7.2.19b).
As a consequence of Lax-Milgram lemma, we have the following theorem.

Theorem 7.2.15 Let f ∈ (H 1 (Ω)) , b = 0, and c ≥ c0 > 0. Then the problem (7.2.20)
has a unique solution u ∈ H 1 (Ω). Furthermore, there exists a constant C > 0 indepen-
dent of f such that
u H 1 (Ω) ≤C f H −1 (Ω) .
131
7.3. Nonlinear second-order elliptic PDEs Chapter 7
Similar results to those of Theorem 7.2.12 for the Dirichlet problem (7.1.4) hold for the
problem (7.2.20). We leave the details as an exercise. For more on this topic, see, for
example, [5, 20].
Remark 7.2.16 Let us see in what sense a weak solution to (7.2.20) satisfies the bound-
ary condition (7.2.19b). For simplicity, we assume f ∈ L2 (Ω) (see NProblem 7.8 for

f ∈ (H (Ω)) ). Then taking v ∈ D(Ω) in (7.2.20) gives T :=
1
i,j=1 ∂j (aij ∂i u) =
b · ∇u + cu − f in D (Ω). As b · ∇u + cu − f ∈ L2 (Ω), it follows T = b · ∇u + cu in
L2 (Ω). Then from (7.2.20) with v ∈ H 1 (Ω), we get
N
T, v = b · ∇u + cu − f, v = (b · ∇u − cu)vdx − f, v = − aij ∂i u∂j vdx,
Ω Ω i,j=1
or equivalently
N N

∂j (aij ∂i u), v + aij ∂i u∂j vdx = 0, ∀v ∈ H 1 (Ω). (7.2.21)
i,j=1 Ω i,j=1

The equation (7.2.21) defines how N i,j=1 aij ∂i uνj = 0 is understood. This makes sense
because if u ∈ C (Ω) integrating by parts in (7.2.21) implies
2
N
0 = aij ∂i uνj vdx, ∀v ∈ H 1 (Ω),
∂Ω i,j=1
N
which implies i,j=1 aij ∂i uνj = 0.
7.3 Nonlinear second-order elliptic PDEs

Nonlinear PDEs describe a wide range of physical phenomena and are subject to
intense research. There are several methods for solving nonlinear PDEs, such as varia-
tional, continuity, and fixed point methods. For a deeper analysis of these methods see,
for example, [20].
In this section we give a short introduction to nonlinear second-order elliptic PDEs.
We will consider only the fixed point method. Typically, this method consists of three
steps: (i) write the solution to the nonlinear problem as a fixed point problem of a
certain linear operator A, (ii) solve the linear problem associated with the operator A
and establish appropriate estimates, and (iii) use the fixed point theorem.
In this short introduction to nonlinear PDEs, we will consider certain PDEs of the
form

N
−∂j (ai,j (x, u)∂i u) + b(x, u) = f (x). (7.3.1)
i,j=1
132
Chapter 7 7.3. Nonlinear second-order elliptic PDEs
Let us cite the following theorem which asserts the existence of fixed points of a compact
continuous operator; see, for example, [20].
Theorem 7.3.1 Let A : E → E be a continuous operator from the Banach space E to
itself. Assume that there exists K ⊂ E, convex closed set, such that A(K) ⊂ K and
A(K) is precompact2 in E. Then A has a fixed point, i.e. there exists u ∈ E such that
u = Au.
Now we give two examples, which demonstrate the use of the fixed point method to
solve various problems similar to (7.3.1).
Example 7.3.2 Consider the following nonlinear boundary value problem
−u + h(u)u = f in Ω = (0, 1), u = 0 on ∂Ω, (7.3.2)
with f ∈ H −1 (Ω), 0 ≤ h ∈ Cb0 (R). Assuming a smooth solution u exists, multiplying

(7.3.2) by ϕ ∈ H01 (Ω) and integrating in Ω yields

u ϕ + h(u)uϕ = f, ϕ , ∀ϕ ∈ H01 (Ω). (7.3.3)
Ω
This is the weak form equation of (7.3.2), and we look for a solution u ∈ H01 (Ω) to it.
For this we consider a fixed point approach as follows. Let A : C 0 (Ω) → C 0 (Ω), Av = u,
where u is the solution to

u ϕ + h(v)uϕ = f, ϕ , ∀ϕ ∈ H01 (Ω). (7.3.4)
Ω
If u is a fixed point of A, i.e. u = Au, then u ∈ H01 (Ω) and solves (7.3.3).
Claim: A is well-defined. Indeed, for every v ∈ C 0 (Ω) the equation (7.3.4) has a unique
solution u ∈ H01 (Ω) −→ C 0 (Ω); see Theorem 7.2.8 and Theorem 6.3.9. Furthermore,
the following estimate holds:
Av H 1 (Ω) ≤C f H −1 (Ω) . (7.3.5)
Claim: A is continuous in C 0 (Ω). Indeed, let v0 , v ∈ C 0 (Ω) with v tending to v0 in C 0 (Ω),

and set u = Av, u0 = Av0 . Then from (7.3.4) with ϕ = u − u0 , we get

|u − u0 |2 + h(v)(u − u0 )2 + (h(v) − h(v0 ))u0 (u − u0 ) = 0,
Ω
which implies

|Av − Av0 |2H 1 (Ω) ≤ |h(v) − h(v0 )||u0 ||Av − Av0 | (use H ölder)
Ω
2
A set K ⊂ E, E normed space, is precompact if its closure is compact.
133
≤ h(v) − h(v0 ) C 0 (Ω) u0 L2 (Ω) Av − Av0 L2 (Ω) .
This inequality in combination with Poincaré inequality yields
Av − Av0 H 1 (Ω) ≤ C h(v) − h(v0 ) C 0 (Ω) u0 L2 (Ω) .
From h(v) → h(v0 ) in C 0 (Ω) and H01 (Ω) → C 0 (Ω), we get Av − Av0 C 0 (Ω) → 0 as
v → v0 in C 0 (Ω), hence A is continuous from C 0 (Ω) to itself.
Claim: Let K = H01 (Ω). Then A(K) ⊂ K and A(K) is precompact in C 0 (Ω). This
follows from the estimate (7.3.5) and H 1 (Ω) −→
c
C 0 (Ω); see Theorem 6.3.9.
Claim: A has a fixed point. Conditions of Theorem 7.3.1 are satisfied, so A has a fixed
point u = Au ∈ H01 (Ω), which solves (7.3.3).
Example 7.3.3 Consider the following nonlinear PDE:
−Δu + h(u)u = f in Ω, u = 0 on ∂Ω, (7.3.6)
with Ω ⊂ R2 an open bounded Lipschitz set, f ∈ H −1 (Ω), 0 ≤ h ∈ Cb1 (R). Proceeding as

in example above, we get this weak form equation

u ∈ H0 (Ω) such that
1
∇u · ∇ϕ + h(u)uϕ = f, ϕ , ∀ϕ ∈ H01 (Ω). (7.3.7)
Ω
Similar to the method in Example 7.3.2, we consider a fixed point approach as follows.
Let A : L2 (Ω) → L2 (Ω), Av = u, where u is the solution to

∇u · ∇ϕ + h(v)uϕ = f, ϕ , ∀ϕ ∈ H01 (Ω). (7.3.8)
Ω
Claim: A is well-defined. Indeed, for v ∈ L2 (Ω) the equation (7.3.8) has a unique solu-
tion, see Theorem 7.2.8, and the following estimate holds:
Av H01 (Ω) ≤C f H −1 (Ω) . (7.3.9)
Claim: A is continuous in L2 (Ω). Indeed, let v0 , v ∈ L2 (Ω) with v tending to v0 in L2 (Ω),

and set u = Av, u0 = Av0 . Then, like in Example 7.3.2, from (7.3.8) with ϕ = u − u0
we get

|Av − Av0 |H 1 (Ω) ≤
2
|h(v) − h(v0 )||u0 ||Av − Av0 | (use H ölder)
0
Ω
≤ h(v) − h(v0 ) L2 (Ω) u0 L4 (Ω) Av − Av0 L4 (Ω)
≤ C h(v) − h(v0 ) L2 (Ω) u0 L4 (Ω) Av − Av0 H01 (Ω) ,
where we used the fact that H 1 (Ω) → Lq (Ω) for all q ∈ [1, ∞); see Remark 6.3.11.
Therefore, using Poincare inequality (6.5.1) and |h(v) − h(v0 )| ≤ h Cb1 (R) |v − v0 | implies
Av − Av0 H01 (Ω) ≤ C v − v0 L2 (Ω) u0 L4 (Ω) ,
134
which proves that A is continuous from L2 (Ω) to itself.

Claim: Let K = H 1 (Ω). Then we have A(K) ⊂ K and A(K) is precompact in L2 (Ω).
This follows from the estimate (7.3.9) and the compactness of the embedding H01 (Ω) →
L2 (Ω); see Theorem 6.3.10.
Claim: A has a fixed point. The conditions of Theorem 7.3.1 are satisfied. Then A has
a fixed point u = Au ∈ H01 (Ω), solution to (7.3.6).
Problems
Problem 7.1 Prove that for every ∈ H −k (Ω), k = 1, there exists a unique u ∈ H0k (Ω)
such that

(−1)|α| D2α u = in D (Ω).
|α|≤k
Extend the result in the case k ∈ N.
Problem 7.2 Let Ω = (0, 1), c ∈ L∞ (Ω), and f ∈ H −1 (Ω). Consider the problem
−u + c(x)u = f in Ω, u(0) = u(1) = 0.
Prove that it has a unique weak solution in H01 (Ω) if c− L∞ (Ω) < 2 (in fact, even if
c− L∞ (Ω) < π 2 ).
Problem 7.3 Let Ω ⊂ RN be an open bounded set and consider the problem

−Δu + bi ∂i u + cu = f in Ω, u = 0 on ∂Ω,
i=1,N

with bi ∈ W 1,∞ (Ω), c ∈ L∞ (Ω), and 2c − i=1,N ∂i bi ≥ 0. Prove that there exists a
unique weak solution u ∈ H01 (Ω).
Problem 7.4 Use these inequalities

u 2
≤ ∇u L2 (RN ) , ∀u ∈ H 1 (RN ), N ≥ 3,
|x| 2 N N − 2
L (R )

u
≤ C ∇u L2 (Ω) , ∀u ∈ H01 (Ω), Ω ⊂ B(0, R) ⊂ R2
|x| ln(|x|/R) 2
L (Ω)
(see [51]) to show that the weak solutions to the following problems
−Δu = 1, u ∈ H01 (B(0, R)\{0}); −Δu = 1, u ∈ H01 (B(0, R))
are the same.
135
Problem 7.5 Let Ω ⊂ RN be an open bounded Lipschitz set, ν the unitary exterior
normal vector on ∂Ω and consider the problem
−Δu + u = f ∈ (H 1 (Ω) in Ω, ∂ν u = g on ∂Ω,
where g ∈ H −1/2 (∂Ω). Write the weak form of this problem and prove that this problem
has a unique H 1 (Ω) weak solution.
Problem 7.6 Let Ω ⊂ RN be an open bounded Lipschitz set and ν the unitary
exterior
normal vector on ∂Ω. Furthermore let M0 (Ω) = u ∈ H (Ω), Ω udx = 0 and consider
1 1
the problem
−Δu = f in Ω, f ∈ (H 1 (Ω)) , f, 1 = 0,
∂ν u = 0 on ∂Ω.
i) Write the weak form of this problem.
ii) Prove that Poincaré inequality holds in M01 (Ω).
iii) Prove that this problem has a unique weak solution in M01 (Ω).
Problem 7.7 Let Ω ⊂ RN be an open bounded connected Lipschitz set and γ be a rela-
tively open set on ∂Ω with nonzero HN −1 measure. Furthermore let Hγ1 (Ω) be the closure

for the H 1 (Ω) norm of v ∈ C 1 (Ω), v|γ = 0 , and consider the problem: u ∈ Hγ1 (Ω)
solution to
−Δu = f in Ω, u = 0 on γ, ∂ν u = 0 on ∂Ω\γ, f ∈ L2 (Ω).
ii) Prove that Poincaré inequality holds in Hγ1 (Ω).
iii) Prove that this problem has a unique weak solution in Hγ1 (Ω).
Problem 7.8 Let Ω ⊂ RN be an open bounded Lipschitz set and let u be a solution
to (7.2.20), with b = 0 and c ≥ c0 > 0. Show that L2 (Ω) −→ d
(H 1 (Ω)) and, by
approximating f ∈ (H 1 (Ω)) by a sequence (fn ) in L2 (Ω), give a meaning to (7.2.21) in
the case f ∈ (H 1 (Ω)) .

−Δ2 u + c(x)u = f in Ω ⊂ RN , where Δ2 u = Δ(Δ(u)),
u = ∂ν u = 0 on ∂Ω,
with Ω an open bounded Lipschitz set, ν the exterior unitary normal on ∂Ω, and 0 <
c ∈ C 0 (Ω). Write the weak form of this problem and use Corollary 9.5.4 to prove that
it has a unique weak solution u ∈ H02 (Ω) for every f ∈ H −2 (Ω).
136
Problem 7.10 Let Ω ⊂ R be an open bounded interval, a ∈ C 0 (R), m ≤ a ≤ M , with

m, M ∈ (0, ∞), 0 ≤ c ∈ L∞ (Ω).
i) Write the weak form of this problem
−(a(u)u ) + c(x)u = f in Ω, u ∈ H01 (Ω).
ii) Prove that this problem has a weak solution for every f ∈ H −1 (Ω) (you may
consider the map A : C 0 (Ω) → C 0 (Ω) with u = Av the solution to −(a(v)u ) +
c(x)u = f in H01 (Ω) and use Theorem 7.3.1).
Problem 7.11 Let Ω ⊂ RN be an open bounded Lipschitz set and consider the problem
2
−Δu = e−u in Ω, u = 0 on ∂Ω.
ii) Prove that this problem has a unique weak solution (you may apply Theorem 7.3.1)
2
to the map A : Lq (Ω) → Lq (Ω), for a certain q, with u = A(v) solving −Δu = e−v
in H01 (Ω)).
Problem 7.12 Let Ω ⊂ RN be an open bounded Lipschitz set, N ≤ 5, and consider the
problem:
u ∈ H01 (Ω), −Δu + h(u)u = f in Ω,
where f ∈ H −1 (Ω), 0 ≤ h ∈ Cb1 (R).
i) Write the weak form equation of this problem.
ii) Prove that this problem has a unique weak solution (you may apply Theorem 7.3.1)
to the map A : Lq (Ω) → Lq (Ω), for a certain q, with u = A(v) solving −Δu +
h(v)u = f in H01 (Ω)).
137
8. Second-order parabolic and
hyperbolic PDEs
In this chapter we consider the evolution equation of the form

u (t) + Lu(t) = f (t), t ∈ (0, T ), (8.0.1a)
u(0) = u0 , (8.0.1b)
du
where T > 0, u : (0, T ) → V is a map, V is a Banach space, u := is the time
dt
derivative of u, L : V → V is an operator, and u0 and f are given. With the same
notations as above, we will also consider the second-order evolution equation of the form
u (t) + Lu(t) = f (t), t ∈ (0, T ), (8.0.2a)

u(0) = u0 , u (0) = u1 , (8.0.2b)
d2 u
where u := is the second-order time derivative of u, and u0 , u1 are given. We will
dt2
define in sections 8.3 and 8.4 in which sense (8.0.1) and (8.0.2) are understood.
Often, the equations (8.0.1) and (8.0.2) are referred to as heat equation and wave
equation, respectively. In the simplest form they are given by

u (t) − Δu(t) = f (t), u (t) − Δu(t) = f (t), t ∈ (0, T ),
resp.
u(0) = u0 , u(0) = u0 , u (0) = u1 .
There are different methods to analyze the equations of type (8.0.1) and (8.0.2), such
as Hilbert spaces, variational (or weak formulation), semigroup, Laplace transform, and
spectral decomposition methods; see, for example, [5, 11, 27, 35, 53]. We will only present
in detail the method of spectral decomposition.
8.1 Heat and wave equations and the method

of separation of variables
Before we develop the analysis of (8.0.1) and (8.0.2), we will present the method of
separation of variables for heat and wave equations in dimension one with L = −Δ. It
https://doi.org/10.1007/978-3-031-39524-6 8
8.1. Heat and wave equations and the method of separation of variables Chapter 8
serves as a motivation for the method of separation of variables in the abstract case that
we will consider in this chapter.
8.1.1 Heat equation and the method of separation of variables

Here we consider the heat equation in one space dimension
u − Δu = 0 in Ω × (0, T ) := (0, π) × (0, T ), (8.1.1a)

u = 0 on ∂Ω × (0, T ), (8.1.1b)
u = g on Ω × {0}. (8.1.1c)
Here u = u(x, t), u(·, t) : Ω → R, u is the derivative of u w.r.t. t, and Δu = ∂xx

2
u.
Without developing a functional framework, at this point it is not clear what kind of
solution to look for (8.1.1). However, we will follow the well-known method of “separation
of variables” for non-rigorously solving (8.1.1). At the beginning of the next section, we
will provide more comments about this method before we formalize it for more general
problems.
We look for u = u(x, t) in the form u = V (t)W (x). Replacing u in (8.1.1a) gives
V (t)W (x) − V (t)W (x) = 0.

V (t) W (x)
Assuming V W = 0 for all (x, t), from the last equation we obtain V (t)
= W (x)
, which
V (t) W (x)
implies V (t)
= λ, W (x)
= λ, for a certain constantλ, or equivalently
V − λV = 0, W − λW = 0. (8.1.2)
The equations (8.1.2) are respectively of first and second order, linear, and homogeneous,
so their solution spaces are respectively of dimension one and two. Their general solutions
are

αx + β, λ = 0,
λt
V (t) = γe , W (x) = √ √
−x λ α, β, γ ∈ R.
αe x λ
+ βe , λ = 0,
Therefore

eλt (Ax + B), λ = 0,
u(x, t) = √ √
−x λ A, B ∈ R. (8.1.3)
λt
e (Ae x λ
+ Be 0,
), λ =
Such a u solves (8.1.1a) for arbitrary A, B, and λ. The constants A, B, and λ are chosen
such that u solves (8.1.1b) and (8.1.1c). Let us first deal with (8.1.1b). As we must have
u(0, t) = u(π, t) = 0 for all t ∈ (0, T ), in the case λ = 0 we get
√
B = Aπ√ + B = 0, so
A = B = 0 and u = 0. In the case λ = 0 it follows A + B = Ae π λ
+ Be−π λ = 0. Hence
√ √
B = −A, e2π λ = 1, λ = ik, λ = −k 2 , k ∈ N.
140
Chapter 8 8.1. Heat and wave equations and the method of separation of variables
Therefore, all non-trivial u(x, t) given by (8.1.3) have the form

2 2
uk (x, t) = Ae−tk (eikx − e−ikx ) = Ck e−tk sin (kx) , k ∈ N, Ck ∈ R.
From the linearity of (8.1.1a) and (8.1.1b), it follows (not rigorously) that
2
u(x, t) = uk (x, t) = Ck e−tk sin (kx) (8.1.4)
k∈N k∈N
solves (8.1.1a) and (8.1.1b). The coefficients Ck are unknown, but we can define them
by imposing the condition (8.1.1c). Replacing u of (8.1.4) in (8.1.1c) implies

u(x, 0) = Ck sin (kx) = g(x), x ∈ (0, ). (8.1.5)
k∈N

2 π
Using the relations (4.1.10) gives Ck = sin (ks) g(s)ds, which implies
π 0
π
2 −tk2
u(x, t) = e sin (kx) sin (ks) g(s)ds . (8.1.6)
π k∈N 0
We note that (8.1.6) gives a candidate for a solution to (8.1.1). Ideally, one looks for
a “classical solution” to (8.1.1), which is u ∈ C 2 (Ω × (0, T )) ∩ C 0 (Ω × [0, T )) satisfying
(8.1.1) pointwise. Classical solutions require strong assumptions on the data f , u0 , which
narrow the class of f and u0 . In this chapter we look for weak solutions, which exist for
a large range of data; see section 8.3.
Remark 8.1.1 Using the same method as above, we can also solve the problem
u − c2 Δu = 0 in Ω × (0, T ) := (0, ) × (0, T ), (8.1.7a)
u = 0 on ∂Ω × (0, T ), (8.1.7b)
u = p on Ω × {0}. (8.1.7c)
Indeed, we can change the variables as follows:
π π 2
y = x, τ = c t, v(y, τ ) = u(x, t).

If furthermore g(y) = p(x) then v(x, τ ) solves

π 2
v − Δv = 0 in (0, π) × 0, c T , (8.1.8a)

π 2
v = 0 on {0, π} × 0, c T , (8.1.8b)

v = g on (0, π) × {0}, (8.1.8c)
and is given by (8.1.6). Then
2 −t(ck π )2 π π
u(x, t) = e sin k x sin k s p(s)ds . (8.1.9)
k∈N 0
141
The method of separation of variables can be used to solve (8.1.1) with different
boundary conditions than (8.1.1b), for example ux (0, t) = 0 and u(, t) + ux (, t) = 0.
The method is the same as above, except that after applying these conditions we get
different formulas for the coefficients A, B, λ, and Ck .
Example 8.1.2 Let us find a classical solution to the problem
ut − 16uxx = 0 in (0, π) × (0, ∞), (8.1.10a)

u(0, t) = u(π, t) = 0 for t > 0, (8.1.10b)
u(x, 0) = x(π − x) for x ∈ (0, π), (8.1.10c)
by using the method of separation of variables. We can use (8.1.9) with = π, g(x) =
x(π − x), and c = 4. Integrating by parts
π π

1 − (−1)k
sin k s g(s)ds = sin (ks) s(π − s)ds = 2
0 0 k3
gives
4 1 − (−1)k −t(4k)2 4
u(x, t) = e sin (kx) = uk (x, t). (8.1.11)
π k∈N k3 π k∈N
We show that u is a classical solution to (8.1.10). Indeed, note first that the series
(8.1.11) converges in C 0 ([0, π] × [0, T ]), for arbitrary T > 0 because |uk (x, t)| ≤ k23 for
all (x, t) ∈ [0, π] × [0, ∞). Then the claim follows from Weierstrass M-test. Therefore,
u ∈ C 0 ([0, π] × [0, ∞)) and u satisfies (8.1.10b), (8.1.10c). As |∂x uk | + |∂t uk | ≤ kC2 for all
(x, t) ∈ [0, π] × [0, ∞)), by using again Weierstrass M-test we get u ∈ C 1 ([0, π] × [0, ∞)).
Furthermore, u ∈ C 2 ((0, π) × (0, ∞)) and satisfies (8.1.10a). Indeed, all uk satisfy
(8.1.10a). Next we point out that given > 0 and i, j ∈ {x, t} we have s
e−(4k)
2
e−(4k)2
|∂i,j uk | ≤ C in [0, π] × [ , T ], and < ∞.
k k∈N
k
Then from Weierstrass M-test the series (8.1.11) converges in C 2 ([0, π] × [ , T ]) and the
claim follows from the arbitrariness of and T . Finally, we point out that limt→∞ u(x, t) =
0 because
1
lim |u(x, t)| ≤ C lim e−16t = 0.
t→∞
k∈N
k3 t→∞
142
8.1.2 Wave equation and the method of separation of variables

Now we use the method of separation of variables again, this time to solve the wave
equation in dimension one, namely
utt − uxx = 0 in Ω × (0, T ), Ω = (0, π), T > 0 (8.1.12a)

u = 0 on ∂Ω × (0, T ), (8.1.12b)
u = g, ut = h on Ω × {0}. (8.1.12c)
Disregarding the issues of existence and uniqueness, we are interested only in finding
a formula for the solution u. We assume that u = u(x, t) can be written as u(x, t) =
V (t)W (x). Replacing u in (8.1.12a) gives
V (t)W (x) − V (t)W (x) = 0.

V (t) W (x)
Assuming V W = 0, from the last equation we obtain V (t)
= W (x)
. Then, necessarily
V (t) W (x)
there exists a constant λ such that V (t)
= λ, W (x)
= λ, or equivalently
V − λV = 0, W − λW = 0. (8.1.13)
The equations (8.1.13) are of second order, linear, and homogeneous, so their solution
spaces are of dimension two. The general solutions are of the form

At +
√
B, √
λ = 0, Cx +√
D, √
λ = 0,
V (t) = −t λ W (x) = −x λ
Aet λ
+ Be , λ, = 0, Ce x λ
+ De , λ = 0,
with A, B, C, D being arbitrary constants. Therefore

(At +
√
B)(Cx +√ D), √ √
λ = 0,
u(x, t) = −t λ −x λ (8.1.14)
(Aet λ
+ Be )(Cex λ
+ De 0,
), λ =
is always a solution to (8.1.12a). The constants A, B, C, D, and λ are chosen such that
u solves (8.1.12b), (8.1.12c). Let us first deal with (8.1.12b). The conditions u(0, t) =
u(π, t) = 0, t ∈ (0, T ) are equivalent to
D = Cπ + D √
= 0, √
λ = 0,
−π λ
C + D = Ceπ λ
+ De 0.
= 0, λ =
Then necessarily
C = D = 0, √ λ = 0,
√
D = −C, e2π λ = 1, λ = ik, λ = −k , 2
k ∈ N, λ = 0.
Therefore, u(x, t) given by (8.1.14) looks like

uk (x, t) = C(Aeikt + Be−ikt )(eikx − e−ikx )
143
= 2iC((A + B) cos(kt) + i(A − B) sin(kt)) sin(kx)

=: (Ak cos(kt) + Bk sin(kt)) sin(kx).
It follows that

u(x, t) = uk (x, t) = (Ak cos(kt) + Bk sin(kt)) sin(kx) (8.1.15)
k∈N k∈N
solves (8.1.12a) and (8.1.12b). Now, it remains to find Ak and Bk . They are chosen such
that u given by (8.1.15) satisfies (8.1.12c), which implies

u(x, 0) = Ak sin(kx) = g(x), ut (x, 0) = Bk k sin(kx) = h(x). (8.1.16)
k∈N k∈N
If the functions g and h have the following sine Fourier expansions

g(x) = gk sin(kx), h(x) = hk sin(kx), x ∈ (0, π),
k∈N k∈N
gk
then from (8.1.16) we obtain Ak = hk , Bk = and therefore
k
hk

u(x, t) = gk cos(kt) + sin(kt) sin(kx) (8.1.17)
k∈N
k
solves (8.1.12).
Like for (8.1.6), (8.1.17) gives a candidate for the solution to (8.1.12). On the first
attempt, we look for a “classical solution” to (8.1.12), which is u ∈ C 2 (Ω × (0, T )) ∩
C 1 (Ω × [0, T )) ∩ C 0 (Ω × [0, T )) satisfying (8.1.12) pointwise. Classical solutions require
strong assumptions on the data f , u0 , u1 . We will study the existence of weak solutions,
which exist for a large range of data f , u0 , and u1 ; see section 8.4.
Remark 8.1.3 Using the same method we can solve the problem
utt − c2 uxx = 0 in Ω × (0, T ) := (0, ) × (0, T ), (8.1.18)
u = 0 on ∂Ω × (0, T ), (8.1.19)
u = p, ut = q on Ω × {0}. (8.1.20)
Indeed, we can change the variables as follows:
π π
y= x, τ = c t, v(y, τ ) = u(x, t).

Note that we have
π 2 π 2
utt (x, t) = c vτ τ (y, τ ), c2 uxx (x, t) = c vyy (y, t).

144
Therefore, if g(y) = p(x), h(y) = q(x) then v(y, t) solves

π
vτ τ − vyy = 0 in (0, π) × 0, c T , (8.1.21)
π
v = 0 on {0, π} × 0, c T , (8.1.22)

π
v = g, vτ = c h on (0, π) × {0}. (8.1.23)

Note that the sine Fourier coefficients gk , resp. hk , of g, resp. h, are related to the sine
Fourier coefficients pk , resp. qk ,
of p, resp. q, by resp. gk = pk , hk = qk . Then from
π π
(8.1.17) and u(x, t) = v x, c t , we get

π π π
qk
u(x, t) = pk cos ck t + sin ck t sin k x ,
k∈N
cπ k
where pk and qk are given by

π π
2 2
pk = sin k x p(x)dx, qk = sin k x q(x)dx.
0 0
Finally we point out that following the same method, we can consider more general
boundary conditions than (8.1.12b), such as
a0 u(0, t) + b0 ux (0, t) = 0, aπ u(π, t) + bπ ux (π, t) = 0, (8.1.24)
with a0 , b0 , aπ , bπ given constants, a20 + b20 > 0, a2π + b2π > 0.
D’Alembert solution to the wave equation

When the space domain of the wave equation is R, we can get a formula for the
solution. Indeed, consider
utt − c2 uxx = 0 in R × (0, ∞), (8.1.25a)

u = g, ut = h on R × {0}. (8.1.25b)
As we have seen in Example 2.0.4, the general solution to (8.1.25a) is given by
u(x, t) = F (x − ct) + G(x + ct), (8.1.26)
with F and G being two C 2 (R) functions. The functions F (x − ct) and G(x + ct) are
called “traveling wave solutions”; more precisely F (x − ct) is called “forward wave”,
while G(x + ct) is called “backward wave”. The curves x ± ct = x0 , x0 ∈ R are called
“characteristic curves” of the wave equation.
To identify g and h, we replace (8.1.26) in (8.1.25b) and get
F (x) + G(x) = g(x), c(−F (x) + G (x)) = h(x).
145
8.2. Some preliminary results Chapter 8
It follows that
x
F (x) + G(x) = g(x), c(−F (x) + G(x)) = c(−F (0) + G(0)) + h(s)ds.
0
Hence

1 1 1 x
F (x) = g(x) + (F (0) − G(0)) − h(s)ds,
2 2 2c 0

1 1 1 x
G(x) = g(x) + (−F (0) + G(0)) + h(s)ds, and
2 2 2c 0
u(x, t) = F (x − ct) + G(x + ct)

1 1 x+ct
= (g(x − ct) + g(x + ct)) + h(s)ds. (8.1.27)
2 2c x−ct
The solution given by the formula (8.1.27) is called the “d’Alembert solution to the wave
equation (8.1.25)”.
8.2 Some preliminary results

In general, the formulas (8.1.6) and (8.1.17) do not provide classical solutions. How-
ever, the heat and wave equations describe natural phenomena and expanding their
solutions in series (8.1.6) and (8.1.17) is meaningful. This means that the concept of
classical solutions is restrictive. Instead, the concept of weak solutions is more general
and includes the classical solutions. Before we develop the analysis of the existence of
weak solutions to evolution equations, we will present some preliminary results.
A careful view of the method of separation of variables for heat and wave equations
shows that in all cases the solution is given in the form of a series like

∞
u(t) := u(·, t) = cn (t) sin (n(·)) , (·) ∈ Ω = (0, π),
n=1
with appropriate functions cn . Hence, u(·, t) belongs to span {un = sin(nx)), n ∈ N}.
Note that un and λn = n2 , n = 1, 2, . . ., are the only eigenfunction/eigenvalue pairs of
the operator −Δ : u ∈ H01 (Ω) ∩ H 2 (Ω) → L2 (Ω). Note also that if T denotes the inverse
of −Δ, T = (−Δ)−1 , then T is a compact self-adjoint operator, and un , n−2 are the only
eigenfunction/eigenvalue pairs of the operator T .
The property of writing the solution as a series of eigenfunctions un of T , often
called “spectral decomposition”, holds in general for any self-adjoint (symmetric) com-
pact operator T in a separable Hilbert space. It allows one to look for the solution to a
certain PDE involving such an operator T as a series of eigenfunctions of T . We have
this general result; see, for example, [5, 13, 29].
146
Chapter 8 8.2. Some preliminary results
Theorem 8.2.1 (Spectral decomposition) Let (H, (·, ·)) be a separable Hilbert space
and T : H → H be a self-adjoint compact operator, i.e.
[self-adjointness] : (T u, v) = (u, T v) for every u, v ∈ H, and (8.2.1)
[compactness] : for every bounded sequence (un ) in H, (T (un )) has
a convergent subsequence in H. (8.2.2)
Then there exists a countably infinite orthonormal basis (ϕn ) of H consisting of eigen-
vectors of T , with corresponding eigenvalues {μn } ⊂ R and limn→∞ μn = 0.
We will restrict the analysis of (8.0.1) and (8.0.2) to the cases where L is given by
⎧
⎪ L : H01 (Ω) → H −1 (Ω),
⎪
⎪
⎪
⎪ N
⎪
⎨ u → Lu = − ∂j (aij ∂i u) + cu,
i,j=1 (8.2.3)
⎪
⎪
⎪
⎪ N
⎪
⎪ Lu, ϕ = aij ∂i u∂j ϕ + cuϕ dx, ∀ϕ ∈ H01 (Ω).
⎩
Ω i,j=1
Here and throughout this chapter, unless otherwise stated, Ω ⊂ RN is an open bounded
set, and aij and c satisfy
A := (aij ) = (aji ) ∈ W 1,∞ (Ω; RN ×N ), c ∈ L∞ (Ω), (8.2.4)
N
∃k, K > 0, ∀0 = ξ ∈ RN , k|ξ|2 ≤ aij (·)ξi ξj ≤ K|ξ|2 , a.e. in Ω. (8.2.5)
i,j=1
In what follows, we will use these notations:

(u, v) := (u, v)L2 (Ω) , |u| := (u, u)1/2 ,
((u, v)) := (u, v)H01 (Ω) = Lu, v , u := ((u, u))1/2 ,
where for the equivalence of the norm in H01 (Ω) we have used the Poincaré inequality
in H01 (Ω) and (8.2.5). The following theorem characterizes the spaces L2 (Ω) and H01 (Ω)
in terms of series, involving the eigenvalues and eigenfunctions of L.
Theorem 8.2.2 Set T : L2 (Ω) → L2 (Ω), T f = u, where Lu = f , u ∈ H01 (Ω).
a) L is positive and invertible from H01 (Ω) onto H −1 (Ω), and T is a well-defined
self-adjoint and compact operator.
b) There exists a countably infinite orthonormal basis {ϕn } of L2 (Ω) for the inner
product (u, v), consisting of eigenvectors {ϕn } of L, Lϕn , ϕ = λn (ϕn , ϕ) for all
ϕ ∈ H01 (Ω), with moreover {λn } ⊂ (0, ∞), limn→∞ λn = ∞, and

∞ ∞
L2 (Ω) = u= (u, ϕn )ϕn , |u|2 = (u, ϕn )2 < ∞ . (8.2.6)
n=1 n=1
147
8.2. Some preliminary results Chapter 8
√
c) Similarly, {ψn } := {ϕn / λn } is an orthonormal basis of H01 (Ω) for the inner
product ((u, v)) and

∞
∞
H01 (Ω) = u = (u, ϕn )ϕn , u2 = λn (u, ϕn )2 < ∞ . (8.2.7)
n=1 n=1
d) Let

∞
∞
dom(L) = u= (u, ϕn )ϕn , |u|2dom(L) := λ2n (u, ϕn )2 < ∞ .
n=1 n=1
∞
Then dom(L) ⊂ L2 (Ω), Lu = n=1 λn (u, ϕn )ϕn ∈ L (Ω) and LuL2 (Ω) =
2
|u|dom(L) for every u ∈ dom(L). Furthermore, L ∈ L(dom(L); L2 (Ω)) and dom(L)

equipped with the norm udom(L) = (u2L2 (Ω) + |u|2dom(L) )1/2 is a Hilbert space.
Proof. The claim a) follows from Theorem 7.2.8 and the compact embedding H01 (Ω)→ c
L2 (Ω). The claim b) follows from the classical theorem of spectral decomposition of
compact self-adjoint operators; see Theorem 8.2.1. For c), clearly {ψn } is orthonormal
in H01 (Ω) because

λm
((ψm , ψn )) = Lψm , ψn = λm (ψm , ψn ) = (ϕm , ϕn ) .
λn
Furthermore {ψn , n ∈ N} is dense in H01 (Ω), because if u ∈ H01 (Ω) and ((u, ψn )) = 0 for
all n ∈ N then

0 = ((u, ψn )) = Lψn , u = λn (ψn , u) = λn (ϕn , u)
which implies u = 0 because

{ϕn , n ∈ N} is dense in L2 (Ω). For (8.2.7) we note that if
u ∈ H01 (Ω) then u = ∞n=1 ((u, ψn ))ψn and

∞
∞
∞
∞
u2 = ((u, ψn ))2 = Lψn , u 2
= λ2n (ψn , u)2 = λn (u, ϕn )2 .
n=1 n=1 n=1 n=1
∞ ∞
For d), let u ∈ dom(L), so u = n=1 (u, ϕn )ϕn and n=1 λn (u, ϕn ) < ∞. Clearly
2 2
u ∈ L2 (Ω) because λn ≥ λ1 > 0. For arbitrary ϕ ∈ D(Ω) we have
Lu, ϕ = u, Lϕ
∞

∞
= (u, ϕn )ϕn , Lϕ = (u, ϕn ) Lϕn , ϕ
n=1
n=1

∞ ∞
= λn (u, ϕn )(ϕn , ϕ) = λn (u, ϕn )ϕn , ϕ ,
n=1 n=1
148
Chapter 8 8.2. Some preliminary results

which shows that Lu = ∞ n=1 λn (u, ϕn )ϕn and LuL2 (Ω) = |u|dom(L) .
Finally, to show that dom(L) is a Hilbert space it is enough to show that it is
complete. Let (un ) be a Cauchy sequence in dom(L). Then (un ) and (Lun ) are Cauchy
sequences in L2 (Ω), so they converge in L2 (Ω) to u and v, respectively. Necessarily
v = Lu because for ϕ ∈ D(Ω), we have
v, ϕ = lim Lun , ϕ = lim un , Lϕ = u, Lϕ = Lu, ϕ ,

n→∞ n→∞
Similar to the proof of (8.2.7), one can show that

∞
∞
−1 −1
H (Ω) = f = f, ϕn ϕn ∈ H (Ω), f H −1 (Ω) =
2
λ−1
n f, ϕn
2
< ∞ , (8.2.8)
i=1 n=1
where f, v means the duality H −1 (Ω) × H01 (Ω) and the convergence of the series is in
H −1 (Ω). We leave the proof as an exercise; see Problem 8.9.
Definition 8.2.3 Let F = F (s), s ∈ R, be a function. Then we define

F (L) : dom(F (L)) → L2 (Ω),
∞
F (L)(u) = F (λn )(u, ϕn )ϕn , where (8.2.9)

n=1

∞
dom(F (L)) = u ∈ L2 (Ω), |F (λn )|2 |(u, ϕn )|2 < ∞ .
n=1
Example 8.2.4 For t ≥ 0 consider F (s) = e−ts , s ∈ R, and define

∞
−tL
e (u) = e−tλn (u, ϕn )ϕn , with (8.2.10)

n=1

∞
dom(e−tL ) = u ∈ L2 (Ω), e−2tλn |(u, ϕn )|2 < ∞ . (8.2.11)
n=1
We close this section by introducing the following concepts and associated results.
For the proof of these results and more details, the reader can see, for example, [15, 36].
Let (Y, · ) be a Banach space, a, b ∈ R, a < b, k ∈ N0 , p ∈ [1, ∞]. We have introduced
the spaces C k ((a, b); Y ); see Definition 1.1.6 with Ω = (a, b) and X = R. The spaces
C k ([a, b]; Y ) and Lp (a, b; Y ) are defined by

C k ([a, b]; Y ) = u : [a, b] → Y, u(i) extends as a C 0 ([a, b]; Y ) f or all i = 0, . . . , k,
149
8.3. Weak solution to the heat equation Chapter 8

k
equipped with uC k ([a,b];Y ) = sup u(i) (t)Y , (8.2.12)
i=0 t∈[a,b]

Lp (a, b; Y ) = t ∈ (a, b) → u(t) ∈ Y measurable and uLp (a,b;Y ) < ∞, where
b
uLp (a,b;Y ) =
p
u(t)pY dt, if p ∈ [1, ∞),
a

uL∞ (a,b;Y ) = u(t)Y L∞ (a,b) if p = ∞ . (8.2.13)
These spaces equipped with the respective norms are Banach spaces. If Y is a separable
Hilbert space with {ϕn , n ∈ N} an orthonormal basis, then for every u ∈ L2 (a, b; Y ) we
have

u(t) = (u(t), ϕn )Y ϕn =: un (t)ϕn , (8.2.14)
n∈N n∈N
b b
u2L2 (a,b;Y ) = u(t)2Y dt = |un (t)|2 dt, un ∈ L2 (a, b). (8.2.15)
a n∈N a
8.3 Weak solution to the heat equation

We introduce the concept of a weak solution to the heat equation.
Definition 8.3.1 Let u0 ∈ L2 (Ω) and f ∈ L2 (0, T ; H −1 (Ω)). A function u is called
“weak, or variational, solution to (8.0.1)” if
i) u ∈ C 0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)), (8.3.1a)
ii) u(0) = u0 , (8.3.1b)
d
iii) (u(t), v) + Lu(t), v = f (t), v , ∀v ∈ H01 (Ω), in D (0, T ). (8.3.1c)
dt
The problem i)–iii) is called “weak, or variational, form of (8.0.1)”.
The decompositions (8.2.6), (8.2.7),
∞ and (8.2.8) motivate
∞the following formal solu-
tion to (8.0.1). We assume u0 = n=1 u0,n ϕn and f (t) = n=1 fn (t)ϕn are ∞the decom-
2 2 −1
positions of u0 in L (Ω) and of f in L (0, T ; H (Ω)), and look for u = n=1 un (t)ϕn .
Replacing u in (8.0.1) gives
∞
∞
(un (t) + λn un (t))ϕn (x) = fn (t)ϕn (x), t > 0, x ∈ Ω,
n=1 n=1

∞
∞
un (0)ϕn = u0,n ϕn ,
n=1 n=1
which imply
un (t) + λn un (t) = fn (t), un (0) = u0,n , ∀n ∈ N. (8.3.2)
150
Chapter 8 8.3. Weak solution to the heat equation
This equation can be written equivalently as (eλn t un ) = eλn t fn (t), which has a unique
solution
un ∈ H 1 (0, T ) ∩ C 0 ([0, T ]), (8.3.3a)

t
−λn t
un (t) = e u0,n + eλn (s−t) fn (s)ds, (8.3.3b)
0
because fn ∈ L2 (0, T ). We have this result.
Theorem 8.3.2 Let L be as in (8.2.3) and assume (8.2.4), (8.2.5) hold, u0 ∈ L2 (Ω),
and f ∈ L2 (0, T ; H −1 (Ω)). Then the problem (8.0.1) has a unique weak solution u ∈
C 0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)) given by
∞
t
−λn t λn (s−t)
u(t)(x) = e u0,n + e fn (s)ds ϕn (x) (8.3.4)
n=1 0
t
=: e−tL u0 + e(s−t)L f (s)ds,
0
with un given by (8.3.3).

Proof. Set Suk (t) = kn=1 un (t)ϕn , S0k = kn=1 u0,n , and Sfk (t) = kn=1 fn (t)ϕn . It is
clear that (S0k ) converges to u0 in L2 (Ω) and (Sfk ) converges to f in L2 (0, T ; H −1 (Ω)).
We note also that from (8.2.15) and (8.2.7), for all u ∈ C 0 ([0, T ]; L2 (Ω)), resp. u ∈
L2 (0, T ; H01 (Ω)), we have

k
Suk 2C 0 ([0,T ];L2 (Ω)) = sup |un (t)|2 , (8.3.5)
t∈[0,T ] n=1
k T
Suk 2L2 (0,T ;H 1 (Ω)) = λn |un (t)|2 dt. (8.3.6)
0
n=1 0
Claim: u satisfies (8.3.1a) and (8.3.1b). Indeed, multiplying (8.3.2) by 2un and integrat-
ing in (0, t) ⊂ (0, T ) gives
t t t

2
(un (s)) ds + 2λn |un (s)| ds = 2
2
fn (s)un (s)ds, which implies
0
t
0 0
t t
1
2
un (t) + 2λn |un (s)| ds ≤ u0,n +
2 2
|fn (s)| ds +
2
λn |un (s)|2 ds.
0 λn 0 0
Therefore, for all t ∈ [0, T ] we have

T T
1
2
un (t) + λn |un (s)| ds ≤ u0,n +
2 2
|fn (s)|2 ds.
0 λn 0
151
8.3. Weak solution to the heat equation Chapter 8
Summing this result for n from k + 1 to k + l gives
Suk+l (t) − Suk (t)2L2 (Ω) + Suk+l − Suk 2L2 ((0,T );H 1 (Ω))
0
≤ S0k+l − S0k 2L2 (Ω) + Sfk+l − Sfk 2L2 ((0,T );H −1 (Ω)) , (8.3.7)
which shows that (Suk ) is a Cauchy sequence in C 0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)),
because (S0k ) converges to u0 in L2 (Ω) and (Sfk ) converges to f in L2 (0, T ; H −1 (Ω)).
As C 0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)) is Banach, the sequence (Suk ) converges to u in
C 0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)). Also from Suk (0) = S0k and (Suk ) converging to u in
C 0 ([0, T ]; L2 (Ω)), we get u(0) = u0 , which proves the claim.
Claim: u satisfies (8.3.1c). Indeed, the equation (8.3.2) is equivalent to
d k
(Su , ϕn ) + LSuk , ϕn = Sfk , ϕn , ∀n ∈ {1, 2, . . . , k}.
dt
Multiplying this equation by ψ = ψ(t) ∈ D(0, T ) and integrating by parts in (0, T ) gives
T T T
− (Suk , ϕn )ψ dt + LSuk , ϕn ψdt = Sfk , ϕn ψdt. (8.3.8)
0 0 0
Since limk→∞ Suk − uL2 (0,T ;H01 (Ω)) = 0 and limk→∞ Sfk − f L2 (0,T ;H −1 (Ω)) = 0, we get
T T
lim (Suk , ϕn )ψ dt
(u, ϕn )ψ dt,= (8.3.9)
k→∞ 0 0
T T T
k k
lim LSu , ϕn ψdt, = lim ((Su , Lϕn ))ψdt = ((u, Lϕn ))ψdt
k→∞ 0 k→∞ 0 0
T
= Lu, ϕn ψdt, and (8.3.10)
0
T T
k
lim Sf , ϕn ψdt = f, ϕn ψdt. (8.3.11)
k→∞ 0 0
Therefore, letting k → ∞ in (8.3.8) combined with (8.3.9), (8.3.11) gives

T T T

− (u, ϕn )ψ dt + Lu, ϕn ψdt = f, ϕn ψdt, ∀n ∈ N. (8.3.12)
0 0 0
This equality holds with ϕn replaced by any finite combination of ϕ1 , ϕ2 , . . ., which by

the density of {ϕn } in H01 (Ω) implies (8.3.1c).
Claim: The solution is unique. It is enough to show that if u solves (8.3.1) with f = 0
∞
and u0 = 0, then u = 0. Indeed, we have u = n=1 un ϕn because u ∈ L2 ((0, T ), H 1 (Ω)).
Taking ϕ = ϕn in (8.3.1c) gives un + λn un = 0 and un (0) = 0, which implies un = 0 and
so u = 0.
152
Chapter 8 8.4. Weak solution to the wave equation
8.4 Weak solution to the wave equation

We will solve the problem (8.0.2) with L given by (8.2.3) by using the same approach
we used for the heat equation. Let us first introduce the concept of weak solutions to
the wave equation.
Definition 8.4.1 Given u0 ∈ H01 (Ω), u1 ∈ L2 (Ω), and f ∈ L2 (0, T ; L2 (Ω)), a function
u is called a “weak, or variational, solution to (8.0.2)”, if
i) u ∈ C 1 ([0, T ]; L2 (Ω)) ∩ C 0 ([0, T ]; H01 (Ω)), (8.4.1a)

ii) u(0) = u0 , u (0) = u1 , (8.4.1b)
d
iii) (u (t), v) + Lu(t), v = (f (t), v), ∀v ∈ H01 (Ω), in D (0, T ). (8.4.1c)
dt
The problem (8.4.1) is called “weak, or variational, form of (8.0.2)”.
Assume the decomposition

∞ of u0 in H01 (Ω), 2
∞of u1 in L (Ω), and of f in L2 (0, T ; L2 (Ω))
∞
∞ given by u0 =
are n=1 u0,n ϕn , u1 = n=1 u1,n ϕn , f (t) = n=1 fn (t)ϕn . If u =
n=1 un (t)ϕn , then replacing it in (8.4.1c) yields

∞
(un (t) + λn un (t) − fn (t))ϕn (x) = 0, t > 0, x ∈ Ω,
n=1
which implies
un (t) + λn un (t) = fn (t), un (0) = u0,n , un (0) = u1,n . (8.4.2)
This equation has a unique solution satisfying
un ∈ H 2 (0, T ) ∩ C 1 ([0, T ]), (8.4.3a)

1
un (t) = cos t λn u0,n + √ sin t λn u1,n
λn
t
1
+ √ sin (t − s) λn fn (s)ds. (8.4.3b)
0 λn
The formula (8.4.3b) follows from the method of variations of constants, and it implies
theregularity (8.4.3a) after using the fact: for all g ∈ L1 (0, T ) and ϕ ∈ D(0, T ) we have
d t
dt 0
g(s)ds, ϕ = g, ϕ .
Theorem 8.4.2 Let L be as in (8.2.3) with (8.2.4), (8.2.5) satisfied, u0 ∈ H01 (Ω), u1 ∈
L2 (Ω), and f ∈ L2 (0, T ; L2 (Ω)). Then the problem (8.4.1) has a unique solution given
by
∞
1
u(t) = cos(t λn )u0,n + √ sin(t λn )u1,n ϕn
n=1
λn
153
8.4. Weak solution to the wave equation Chapter 8
∞
t
1
+ √ sin((t − s) λn )fn (s)ds ϕn (8.4.4)
n=1 0 λn
t
1 1
=: cos(t λn )u0 + √ sin(t λn )u1 + √ sin((t − s) λn )f (s)ds, (8.4.5)
λn 0 λn
where un are given by (8.4.3b).

Proof. Like in Theorem 8.3.2 the proof goes in several steps. Set Suk (t) = kn=1 un (t)ϕn ,
k k k
S0k = k
n=1 u0,n , S1 =
k
n=1 u1,n , and Sf (t) =
k
n=1 fn (t)ϕn . It is clear that (S0 )
converges to u0 in H01 (Ω), (S1k ) converges to u1 in L2 (Ω), and (Sfk ) converges to f in
L2 (0, T ; L2 (Ω)). We note also that from (8.2.6) and (8.2.7), we have

k
k
Suk 2C 1 ([0,T ];L2 (Ω)) = sup |un (t)|2 + sup |un (t)|2 , (8.4.6)
t∈[0,T ] n=1 t∈[0,T ] n=1
k
Suk 2C 0 ([0,T ];H 1 (Ω)) = sup λn |un (t)|2 . (8.4.7)
0
t∈[0,T ] n=1
Claim: u satisfies (8.4.1a) and (8.4.1b). Indeed, multiplying (8.3.2) by 2un and inte-
grating in (0, t) ⊂ (0, T ) gives
t t t
2 2
(|un (s)| ) ds + λn (|un (s)| ) ds = 2 fn (s)un (s)ds.
0 0 0
It implies
t t
|un (t)|2 + λn |un (t)| 2
≤ |u1,n | + λn |u0,n | +
2 2
|un (s)|2 ds + |fn (s)|2 ds. (8.4.8)
0 0
To proceed, we estimate un in L2 (0, T ). From (8.4.3b) we get

t

un (t) = − λn sin(t λn )u0,n + cos(t λn )u1,n + cos((t − s) λn )fn (s)ds.
0
This implies
T T

|un (t)| dt ≤ C λn u0,n + u1,n +
2 2 2
|fn (t)| dt , 2
0 0
which combined with (8.4.8) gives

T

|un (t)| + λn |un (t)| ≤ C u1,n + λn u0,n +
2 2 2 2
|fn (s)| ds . 2
0
Summing this result for n from k + 1 to k + l yields

(Suk+l ) (t) − (Suk ) (t)2L2 (Ω) + Suk+l (t) − Suk (t)2H 1 (Ω))
0
154
≤ S1k+l − S1k 2L2 (Ω) + S0k+l − S0k 2H 1 (Ω)

0
+ Sfk+l − Sfk 2L2 ((0,T );L2 (Ω)) , (8.4.9)
which shows that (Suk ) is a Cauchy in C 1 ([0, T ]; L2 (Ω)) ∩ C 0 ([0, T ]; H01 (Ω)). As this space
is Banach, the sequence (Suk ) converges to u in C 1 ([0, T ]; L2 (Ω))∩C 0 ([0, T ]; H01 (Ω)). Also
from Suk (0) = S0k and (Suk ) (0) = S1k we get u(0) = u0 , u (0) = u1 , which proves the claim.
Claim: u satisfies (8.4.1c). The equation (8.4.2) is equivalent to
((Suk ) , ϕn ) + LSuk , ϕn = (Sfk , ϕn ), ∀n ∈ {1, 2, . . . , k}.
Multiplying this equation by ψ = ψ(t) ∈ D(0, T ) and integrating by parts in (0, T ) gives
T T T
k
− ((Su ) , ϕn )ψ dt + k
LSu , ϕn ψdt = (Sfk , ϕn )ψdt. (8.4.10)
0 0 0
From limk→∞ (Suk ) − u C 0 ([0,T ];L2 (Ω)) = 0, limk→∞ Suk − uC 0 ([0,T ];H01 (Ω)) = 0, and
limk→∞ Sfk − f L2 (0,T ;L2 (Ω)) = 0, we get
T T
k
lim ((Su ) , ϕn )ψ dt = (u , ϕn )ψ dt, (8.4.11)
k→∞ 0 0
T T T
k k
lim LSu , ϕn ψdt = lim Su , Lϕn ψdt = u, Lϕn ψdt
k→∞ 0 k→∞ 0 0
T
= Lu, ϕn ψdt and (8.4.12)
0
T T
k
lim (Sf , ϕn )ψdt = (f, ϕn )ψdt. (8.4.13)
k→∞ 0 0
Therefore, letting k → ∞ in (8.4.10) combined with (8.4.11), (8.4.12), and (8.4.13) gives
T T T

− (u , ϕn )ψ dt + Lu, ϕn ψdt = (f, ϕn )ψdt, ∀n ∈ N, (8.4.14)
0 0 0
which like in Theorem 8.3.2 proves (8.4.1c).

Claim: The solution is unique. It is enough to show that if u solves ∞(8.4.1) with f = 0,
u0 = 0, and u1 = 0 then u = 0. Indeed, as necessarily u(t) = n=1 un (t)ϕn , taking
ϕ = ϕn in (8.4.1c) gives un + λn un = 0 and un (0) = 0, un (0) = 0, which implies un = 0
and so u = 0.
In this chapter, we considered linear heat and wave equations (8.0.1), (8.0.2), with
the main part given by the linear operator L in (8.2.3). The main ingredient of the
analysis was the spectral decomposition of the spaces L2 (Ω) and H01 (Ω) and H −1 (Ω).
There are theories associated with (8.0.1), (8.0.2) in a more abstract context. For more
details on this topic we refer the reader to, for example, [11, 27, 30, 53].
155
Problems
Problem 8.1 Using the method of the separation of variables, solve the following PDEs.
For each of them, consider whether the solution is classical or not, and find limt→∞ u(x, t).
(a) ut − uxx = 0 in (0, 1) × (0, ∞),

u=0 on {0, 1} × (0, ∞),
u(x, 0) = x(x − 1) on (0, 1).
(b) ut − uxx = u in (0, 1) × (0, ∞),

u(0, t) = ux (1, t) = 0 for t > 0,
u(x, 0) = x on (0, 1).
(c) ut − uxx = 0 in (0, π) × (0, ∞),

ux (0, t) = ux (π, t) for t > 0,
u(x, 0) = π − x on (0, π).
(d) ut − c2 uxx = 0 in (0, 2π) × (0, ∞),

u(0, t) = u(2π, t) for t > 0,
ux (0, t) = ux (2π, t) for t > 0,
u(x, 0) = g(x) on (0, 2π), g ∈ C 2 ([0, 2π]) and periodic.

ut − uxx = αu in (0, π) × (0, ∞),
u(0, t) = u(π, t) = 0 on {0, π} × (0, ∞),
u(x, 0) = x(π − x) on (0, π),
with α being a real parameter. Show that the method of separation of variables gives a
classical series solution and find limt→∞ u(x, t) in the case α < 1. Consider the solution
and limt→∞ u(x, t) for other values of α.
Problem 8.3 Using the method of separation of variables solve the following PDEs.
(a) utt − uxx = 0 in (0, π) × (0, ∞),

u(0, t) = u(π, t) = 0 for t > 0,
3
u(x, 0) = sin x, ut (x, t) = sin(2x) on (0, π).
(b) utt − uxx = 0 in (0, 1) × (0, ∞),

u(0, t) = u(1, t) = 0 for t > 0,
u(x, 0) = x(1 − x), ut (x, 0) = 0 on (0, 1).
In both cases, discuss whether the series solution is a classical solution or not.
Problem 8.4 Let u0 (x), u1 (x), f (x, t) be given smooth functions. Show that

1 1 x+ct
u(x, t) = (u0 (x − ct) + u0 (x + ct)) + u1 (s)ds
2 2c x−ct
156
t x+c(t−s)
1
+ f (r, s)drds
2c 0 x−c(t−s)
is a classical solution to
utt − c2 uxx = f in R × (0, ∞), u(·, 0) = u0 (·), ut (·, 0) = u1 (·) on R.
Problem 8.5 Find d’Alembert formula in Problem 8.4 by using the formula (3.0.3) for
vt − vx = f and ut + ux = v.
Problem 8.6 Consider the wave equation

utt − c2 uxx = f in (0, ∞) × (0, ∞),
u = u0 , ut = u1 on (0, ∞) × {0},
u = 0 on {0} × (0, ∞),
with u0 (x), u1 (x), f (x, t) given smooth functions, u0 (0) = u1 (0) = 0 and u0 (0+ ), u1 (0+ )
exist. Extend u0 , u1 , and f (·, t) in R as odd functions and then use the formula of
Problem 8.4 above to find a classical solution u(x, t).
Problem 8.7 Consider the wave equation

utt − c2 uxx = f in (0, ) × (0, ∞),
u = u0 , ut = u1 on (0, ) × {0},
u = 0 on {0, } × (0, ∞),
with u0 (x), u1 (x), f (x, t) given smooth functions with u0 (0) = u1 (0) = 0 and u0 (0+ ),
u1 (0+ ) well-defined. Extend u0 , u1 , and f (·, t) in (−, ) as odd functions and next in
R as 2 periodic functions, and then use the formula of Problem 8.4 above to find a
classical solution u.
Problem 8.8 Use d’Alembert formula to solve u(x, t) and t0 .
(a) utt − 4uxx = 0 in R × (0, ∞), with u(·, 0) = 1(−1,1) , ut (·, 0) = 1(0,1) , and t0 solution
to u(2, t0 ) = max{u(2, t), t ≥ 0}.
(b) utt − 9uxx = 0 in R × (0, ∞), with u(·, 0) = 5 · 1[−1,1] , ut (·, 0) = 1[−1,1] . Here, for a
given x0 , t0 solves u(x0 , t0 ) = max{u(x0 , t), t ≥ 0}.1
Problem 8.9 Prove (8.2.8). As an application, take Lu = −u in Ω = (0, 2π) and write
δπ (x) := δ0 (x − π) as a series in terms of eigenfunctions of L.
1
This problem has an application: if u is the pressure after a shock wave generated by an explosion
and a building located at the position x0 supports a pressure up to 4 (u is a dimensionless variable),
then this building will collapse or not depend on the sign of δ := 4 − max{u(x0 , t), t > 0}—it collapses
if δ < 0 and it does not if δ ≥ 0.
157

u − uxx = f in Ω × (0, T ), Ω = (0, 2π),
u(·, t) = 0 on ∂Ω × (0, T ),
u(x, 0) = u0 (x) on Ω.
with u0 ∈ L2 (Ω), f (x, t) = δπ (x). Prove that the series solution (8.3.4) associated with
this problem converges in C 0 ([0, T ]; L2 (Ω))∩L2 (0, T ; H01 (Ω)) and satisfies (8.3.1c). Prove
moreover that u ∈ L2 (0, T ; H −1 (Ω)).
Problem 8.11 Let u be the solution to the problem (8.0.1) as given by Theorem 8.3.2.
(a) Prove that u ∈ H 1 (0, T ; H −1 (Ω)) and therefore (8.0.1a) holds in L2 (0, T ; H −1 (Ω)).
(b) Prove that if u0 ∈ H01 (Ω) then u ∈ C 0 ([0, T ]; H01 (Ω)).
Problem 8.12 Let u be the solution to the problem (8.0.1) as given by Theorem 8.3.2.
By using the density of {ζ(x)ψ(t), ζ ∈ D(Ω), ψ ∈ D(0, T )} in D(Ω × (0, T )) show that
u satisfies
u + Lu = f in D (Ω × (0, T )).
Problem 8.13 Show that the problem (8.0.1) with (8.0.1b) replaced by u(0) = u(T )
has a unique solution in CT0 ([0, T ]; L2 (Ω)) ∩ L2 (0, T ; H01 (Ω)) in the sense of Definition
8.3.1 (with u(0) = u0 replaced by u(0) = u(T )) for every f ∈ CT0 ([0, T ]; L2 (Ω)) := {g ∈
C 0 ([0, T ]; L2 (Ω)), g(0) = g(T )}.
Problem 8.14 Let u be the solution to (8.0.2) as given by Theorem 8.4.2. Prove that

E(t) := |u (x, t)|2 dx + Lu(·, t), u(·, t)
Ω t
= |u1 (x)| dx + Lu0 , u0 + 2
2
f (x, s)u (x, s)dxds, t ∈ (0, T ). (8.4.15)
Ω 0 Ω
Hence, if f = 0 in a certain interval (a, b) ⊂ (0, T ) then E(t) remains constant in (a, b).
158
9. Annex
In this chapter, we will collect different results, with or without proof, which com-
plement the material considered in the previous chapters.
9.1 Notations and review (Ch. 1)

9.1.1 Continuous differentiable functions
Proposition 9.1.1 Let Ω ⊂ RN be open, k ∈ N0 , λ ∈ (0, 1]. Define Cbk (Ω) and Cbk,λ (Ω)
spaces as in Proposition 1.1.11, and endow them with the norms · Cbk (Ω) , · Cbk (Ω) ,
respectively. These spaces are Banach spaces.
Proof. It is easy to show that uCbk (Ω) and uC k,λ (Ω) are norms, which we leave as an
b
exercise. Now we show that the spaces are Banach.
i) Claim: Cb0 (Ω) is a Banach space. Let (un ) be a Cauchy sequence in Cb0 (Ω), so for given
> 0 there exists n such that
|un+m (x) − un (x)| < , ∀n, m ∈ N, n > n , ∀x ∈ Ω. (9.1.1)
We have to show that (un ) converges to a certain u ∈ Cb0 (Ω) in Cb0 (Ω) norm. From
(9.1.1), it follows that (un (x)) is uniformly bounded, Cauchy in R and converges to a
certain u(x), with necessarily {u(x), x ∈ Ω} bounded. For n > n and x ∈ Ω, we obtain
|un (x) − u(x)| = lim |un (x) − un+m (x)| ≤ , (9.1.2)
m→∞
which shows that (un ) converges uniformly1 to u in Ω. To complete the proof of the
claim, it is enough to show u ∈ C 0 (Ω). For x, x + h ∈ Ω, we have
|u(x + h) − u(x)| ≤ |u(x + h) − un (x + h)| + |u(x) − un (x)| + |un (x + h) − un (x)|,
which combined with the uniform convergence of (un ), of the uniform continuity of un
near2 x and of (9.1.1) implies u ∈ C 0 (Ω) and proves the claim.
1
A sequence (un ) in C 0 (Ω; Y ), Ω ⊂ X, “converges uniformly to u : Ω → Y in Ω” if for every > 0
there exists n ∈ N such that for all n > n and all x ∈ Ω we have un (x) − u(x)Y < .
2
It is easy to prove that if v ∈ C 0 (Ω) then v is uniformly continuous in B(x, rx ) = {y ∈ Ω, |y − x| ≤
rx } ⊂ Ω.
https://doi.org/10.1007/978-3-031-39524-6 9
9.1. Annex: Chapter 1 Chapter 9
ii) Claim: Cbk (Ω) is a Banach space. Proceeding as in i) above, we show that given a
Cauchysequence(un )inCb1 (Ω),thereexistu, v i ∈ Cb0 (Ω),i = 1, . . . , N ,withlimn→∞ un = u,
limn→∞ ∂i un = v i in Cb0 (Ω). Necessarily, ∂i u = v i because for x ∈ Ω and |t| small we
have
1 1
(u(x + tei ) − u(x)) = lim (un (x + tei ) − un (x)|
t t n→∞
1 1
= lim ∂i un (x + stei )ds = v i (x + stei )ds
n→∞ 0 0
−−→ v i (x),
t→0
which completes the claim for Cb1 (Ω) spaces. The proof for Cbk (Ω) spaces follows similarly
by recurrence.
iii) Claim: Cbk,λ (Ω) is a Banach space. Given a Cauchy sequence (un ) in Cbk,λ (Ω), from
cases above we get the convergence of (un ) to a certain u in Cbk (Ω). Furthermore, as
(un C k,λ (Ω) ) is necessarily bounded, for x, y ∈ Ω with x = y, α ∈ NN
0 with |α| = k, we
b
have
|Dα u(x) − Dα u(y)| |Dα un (x) − Dα un (y)|

= lim ≤ lim un C k,λ (Ω) < ∞,
|x − y|λ n→∞ |x − y|λ n→∞ b
and therefore |u|C k,λ (Ω) < ∞, so u ∈ C k,λ (Ω). Finally for n, m ∈ N, we have
|Dα (u − un )(x) − Dα (u − un )(y)| |Dα (un+m − un )(x) − Dα (un+m − un )(y)|

= lim
|x − y|λ m→∞ |x − y|λ
≤ lim un+m − un C k,λ (Ω) ,
m→∞
which from the Cauchy property of (un ) in Cbk,λ (Ω) implies limn→∞ un = u in Cbk,λ (Ω),
and completes the proof.
9.1.2 Some results from Lp (Ω) spaces

Theorem 9.1.2 (Tonelli) Let N, M ∈ N and A ⊂ RN , B ⊂ RM be two open sets, and
u : A × B → R be a measurable function such that

|u(x, y)|dy < ∞ for a.a. x ∈ A, and dx |u(x, y)|dy < ∞.
B A B
Then u ∈ L1 (A × B).
The following is a sort of the reciprocal from the theorem above.
160
Chapter 9 9.1. Annex: Chapter 1
Theorem 9.1.3 (Fubini) Let u ∈ L1 (A × B) with A, B as in Theorem 9.1.2. Then

i) for a.a. x ∈ A, u(x, ·) ∈ L (B), and x →
1
u(x, y)dy ∈ L1 (A),
B
ii) for a.a. y ∈ B, u(·, y) ∈ L (A), and y →

1
u(x, y)dx ∈ L1 (B),
A
iii) u(x, y)dydx = u(x, y)dxdy = u(x, y)dxdy.
A B B A A×B
9.1.3 An application: Ordinary Differential Equations

Theorem 9.1.4 Let (X, · X ) be a Banach space, U ⊂ X open and f : U → X a
Lipschitz function with Lipschitz constant L. Let x0 ∈ U , ρ ∈ (0, 1) such that B(x0 , 2ρ) ⊂ U
and K > 0 such that f C 0 (B(x0 ,2ρ)) ≤ K. Finally, let 0 < r < min{1/L, ρ/K} and
set I r = [−r, r]. Then there exists α : I r × B(x0 , ρ) → U , such that t ∈ I r → α(t, x) ∈ X
is the unique solution to (1.3.4) in C 1 (I r ; X).
Proof. The proof of this theorem is an application of Theorem 1.3.10. It is simple to

show that (1.3.4) in I r is equivalent to
t
find α(·, x) ∈ C (I r ; X), α(t, x) = x +
1
f (α(s, x))ds, t ∈ I r . (9.1.3)
0
Now consider the set
M = {z ∈ C 0 (I r ; B(x0 , 2ρ)), z(0) ∈ B(x0 , ρ)}.
Equipped with the distance d(y, z) := y(·)−z(·)C 0 (I r ;X) , M is a complete metric space.
Next, for x ∈ B(x0 , ρ) consider the map
t
Tx : M → M, (Tx z)(t) = x + f (z(s))ds, t ∈ I r , x ∈ B(x0 , ρ).
0
Clearly Tx is well-defined, because Tx z ∈ C 0 (I r ; X), (Tx z)(0) = x ∈ B(x0 , ρ), and for
t ∈ I r we have (Tx z)(t) ∈ B(x0 , 2ρ), because
t
(Tx z)(t) − x0 X ≤ x − x0 X + f (z(s))X ds ≤ ρ + rK ≤ 2ρ.
0
Then every solution to (9.1.3) is a fixed point of Tx . Note Tx is a contraction because

t
Tx z1 − Tx z2 C 0 (I r ;X) ≤ sup f (z1 (s)) − f (z2 (s))X ds, t ∈ I r
0
≤ κz1 − z2 C 0 (I r ;X) , 0 ≤ κ = rL < 1. (9.1.4)
161
Therefore, from Theorem 1.3.10 there exists a unique fixed point αx ∈ Z of Tx , and
α(t, x) = αx (t) is a solution to (9.1.3).
As f is Lipschitz from U in X, it follows that f ∈ C 0 (U ; X) and therefore from (9.1.3)
we obtain α(·, x) ∈ C 1 (I r ; U ) and α (t, x) = f (α(t, x)), t ∈ I r . Finally, if f ∈ C k (U ; X),
from α (t, x) = f (α(t, x)) we get α(·, x) ∈ C k+1 (I r ; X).
The uniqueness of the solution to (1.3.4) holds as follows.

Theorem 9.1.5 Assume the conditions of Theorem 9.1.4 hold. Let x ∈ U and assume
α1 (·, x) ∈ C 1 (J1 ; U ) and α2 (·, x) ∈ C 1 (J2 ; U ) solve (1.3.4), where J1 , I2 are closed
intervals containing 0. Then α1 (·, x) = α2 (·, x) on J1 ∩ J2 .
Proof. See [28, Chapter XIV, §3].
When f is Lipschitz, we can prove by direct calculations that the solution α is Lipschitz
with respect to x.
Theorem 9.1.6 Let x ∈ B(x0 , ρ) and α(·, x) ∈ C 1 (I r , U ) as given by Theorem 9.1.4.
Then x → α(·, x) is uniformly Lipschitz from B(x0 , ρ) to C 0 (I r ; X).
Proof. Let x, y ∈ B(x0 , ρ) and α(·, x), α(·, y) be the corresponding C 1 (I r ; X) solutions
of (1.3.4a) as given by Theorem 9.1.4. We have to prove
α(·, x) − α(·, y)C 0 (I r ;X) ≤ Cx − yX ,
for certain C > 0 independent of x and y. Note that from Theorem 1.3.10, limn→∞
Txn α(·, y) − α(·, x)C 0 (I r ;X) = 0, where Tx : Z → Z is defined in Theorem 9.1.4. As
Ty α(·, y) = α(·, y) and Tx z1 − Tx z2 C 0 (I r ;X) ≤ κz1 − z2 X , 0 ≤ κ = rL < 1 (see
(9.1.4)), we have
Tx α(·, y) − α(·, y)C 0 (I r ;X) = Tx α(·, y) − Ty α(·, y)C 0 (I r ;X) = x − yX ,
Txk α(·, y) − Txk−1 α(·, y)C 0 (I r ;X) ≤ κTxk−1 α(·, y) − Txk−2 α(·, y)C 0 (I r ;X) x − yX
···
≤ κk−1 Tx α(·, y) − α(·, y)C 0 (I r ;X) x − yX
= κk−1 x − yX ; hence

Txn α(·, y) − α(·, y)C 0 (I r ;X) ≤ Txk α(·, y) − Txk−1 α(·, y))C 0 (I r ;X)
k=1,...,n
≤ (1 + κ + · + κn−1 )x − yX

1
≤ x − yX ,
1−κ
which after letting n to infinity gives α(·, x) − α(·, y)C 0 (I r ;X) ≤ Cx − yX with
C = 1/(1 − κ).
162
9.2 First-order PDEs: classical and weak solutions

(Ch. 3)
9.2.1 Classical local solutions to first-order PDEs
Theorem 9.2.1 Assume that E and g are C k functions, k ≥ 2, and (x0 , z0 (x0 ), p0 (x0 ))
is non-characteristic, x0 ∈ Γ. Then there exist ρ0 > 0, r0 > 0 such that
(i) There exists y 0 → p0 (y 0 ) ∈ C k−1 (Γ0 ; RN ) such that (y 0 , z0 (y 0 ), p0 (y 0 )) is non-
characteristic for all y 0 ∈ Γ0 := Γ ∩ B(x0 , ρ0 ).
(ii) For every y 0 ∈ Γ0 , the initial value problem (3.2.10) has a unique C k ([−r0 , r0 ]; RN ×
R × RN ) solution. Furthermore, if T is defined by
T : [0, r0 ] × Γ0 =: U0 → Ω0 := T (U0 ) ⊂ Ω,
(t, y 0 ) → T (t, y 0 ) := y(t, y 0 ),
then T is C k−1 , invertible, and its inverse is C k−1 .
(iii) The function u : Ω0 → R, u(x) = z ◦ T −1 (x) = z(y(t, y 0 )), is a C k solution to
(3.2.11).
(iv) If ∂pN E(x0 , z0 (x0 ), (p01 (x0 ), . . . , p0N −1 (x0 ), q)) does not change sign (remains posi-
tive or negative) for all q ∈ R, then (3.2.11) has a unique C k solution.
Proof. The proof of this theorem is not difficult, but long and technical, and will be
achieved in several steps. The proof we present follows [6, 9, 16].
Proof of (i). The existence of the map p0 follows from Lemma 3.2.2.
Proof of (ii). As E is C k , Theorem 1.3.13 implies that the problem (3.2.10) has a unique
C k solution (y(t, y 0 ), z(t, y 0 ), p(t, y 0 )), t ∈ [−r0 , r0 ], for arbitrary y 0 ∈ Γ0 . Note that, a
priori, r0 depends on y 0 . However, if ρ0 > 0 is small enough the constants K and L of
Theorem 1.3.13 can be chosen independent of y 0 ∈ Γ0 , and then the existence of r0 > 0
independent of y 0 follows.
Furthermore, as E and g are C k , Theorem 1.3.15 implies z ∈ C k−1 ([−r0 , r0 ] × Γ0 )
and y, p ∈ C k−1 ([−r0 , r0 ] × Γ0 ; RN ), provided r0 and ρ0 are small enough.
Next, we consider
T : R × RN −1 × {0} →
RN ,
t
(t, y 0 ) → T (t, y 0 ) := y(t, y 0 ) = y 0 + 0 ∇p E(y(s, y 0 ), z(s, y 0 ), p(s, y 0 ))ds.
The map T is well-defined and C k−1 near (0, x0 ) in R × RN −1 × {0}. Furthermore, from
y(0, x0 ) = x0 and y = ∇p E(y, z, p) we get
⎡ ⎤
∂p1 E(x0 , z0 , p0 ) 1 ··· 0
⎢ .. .. . . .. ⎥
∂T (0, x0 ) ⎢ . . . . ⎥.
= ⎢ ⎥
0
∂(t, y ) ⎣ ∂pN −1 E(x , z0 (x ), p (x )) 0 · · · 1 ⎦
0 0 0 0
∂pN E(x0 , z0 (x0 ), p0 (x0 )) 0 · · · 0
163

∂T (0, x0 )
Clearly, det = ±∂pN E(x0 , z0 (x0 ), p0 (x0 )) = 0. Then from (inverse mapping)
∂(t, y 0 )
Theorem 1.3.11, it follows that T is C k−1 and invertible near (0, x0 ), with T −1 of class
C k−1 , which completes the proof of (ii).
Proof of (iii). Clearly u ∈ C k−1 (Ω0 ) and u(y 0 ) = g(y 0 ) on Γ0 . We will complete the
proof of (iii) in several steps.
Claim: For all (t, y 0 ) ∈ U0 and x = T (t, y 0 ), we have
E(x, u(x), p(t, y 0 )) = 0. (9.2.1)
Indeed, for fixed y 0 set (t) = E(y(t, y 0 ), z(t, y 0 ), p(t, y 0 )), t ∈ [0, r0 ]. Then we have
(0) = E(y(0, y 0 ), z(0, y 0 ), p(0, y 0 )) = E(y 0 , z0 (y 0 ), p0 (y 0 )) (f rom (3.2.8))
= 0,

(t) = ∇p E(y, z, p) · p + ∂z E(y, z, p)z + ∇x E(y, z, p) · y (f rom (3.2.10))
= 0.
Then Theorem 1.3.13 implies = 0 and proves (9.2.1).
Claim: For all (t, y 0 ) ∈ U0 and x = T (t, y 0 ), we have
∇u(x) = p(t, y 0 ). (9.2.2)

∂z ∂x
Indeed, from x = y(t, y 0 ) and (3.2.10) we have =p· , which gives
∂t ∂t
∂u ∂z ∂t
N −1
∂z ∂yj0 ∂xk ∂t
N
N −1
∂z ∂yj0
= + = p k + . (9.2.3)
∂xi ∂t ∂xi j=1
∂yj0 ∂xi k=1
∂t ∂x i j=1
∂y 0
j ∂x i
If we admit for a moment that
∂z ∂xk
N
= pk 0 , j = 1, . . . , N − 1, (9.2.4)
∂yj0 k=1
∂yj
then (9.2.3) implies

∂u N
∂xk ∂t
N −1
∂xk ∂yj0
N
∂xk
= pk + = pk = pi ,
∂xi k=1
∂t ∂xi j=1
∂yj0 ∂xi k=1
∂xi
which proves (9.2.2). Now we focus on the proof of (9.2.4), which is quite technical and
∂z(t, y 0 )
N 0
0 ∂xk (t, y )
relies on (3.2.10). Set (t) = − p k (t, y ) . Then
∂yj0 k=1
∂yj0
∂z(0, y 0 )
N 0
0 ∂xk (0, y ) ∂z0
(0) = 0
− pk (0, y ) 0
= 0 − p0j = 0, (9.2.5)
∂yj k=1
∂yj ∂yj
164
∂ N
N
∂2z ∂xk ∂pk ∂xk ∂ 2 xk
and using = pk = + pk gives
∂t∂yj0 k=1
∂yj0 ∂t k=1
∂yj0 ∂t ∂t∂yj0
N

∂ 2z ∂pk ∂xk ∂ 2 xk
(t) = − + pk
∂t∂yj0 k=1 ∂t ∂yj0 ∂t∂yj0
N
∂pk ∂xk ∂ 2 xk ∂pk ∂xk ∂ 2 xk
= 0
+ pk 0
− 0
+ pk
k=1
∂y j ∂t ∂t∂y j ∂t ∂y j ∂t∂yj0
N
∂pk ∂xk ∂pk ∂xk
= − (use (3.2.10))
k=1
∂yj0 ∂t ∂t ∂yj0
N
∂pk ∂E ∂E ∂E ∂xk
= − − − pk . (9.2.6)
k=1
∂yj0 ∂pk ∂xk ∂z ∂yj0
Differentiating E(y(t, y 0 ), z(t, y 0 ), p(t, y 0 )) = 0 (which is (9.2.1)) with respect to yj0 gives
N
N
∂E ∂pk ∂E ∂xk ∂E ∂z
0
=− 0
+ 0
.
k=1
∂p k ∂yj
k=1
∂x k ∂yj ∂z ∂y j
Replacing this equality in (9.2.6) yields

∂E ∂z ∂xk
N
∂E
(t) = − − pk =− · (t). (9.2.7)
∂z ∂yj0 k=1 ∂yj0 ∂z
From (9.2.5), (9.2.7), and Theorem 1.3.13, we get = 0, which proves (9.2.4) and
completes the proof of (9.2.2).
Claim: u ∈ C k (Ω0 ) and solve (3.2.11). Indeed, from (9.2.2) we have ∇u(x) = p(T −1 (x)),
which implies ∇u ∈ C k−1 (Ω0 ; RN ), so u in C k (Ω0 ). The equations (9.2.1), (9.2.2), and
u(y 0 ) = z(0, y 0 ) = g(y 0 ), y 0 ∈ Γ0 prove that u solves (3.2.11).
Proof of (iv): Proposition 3.1.1 and (iii) of this theorem set a correspondence between
local C 2 solutions u to (3.0.1) and the solutions (y, z, p) of (3.2.10). Therefore, from the
uniqueness of solutions to ODEs (from point (ii) of this theorem), it is enough to prove
that near x0 , the non-characteristic initial conditions are unique.
The compatibility condition (3.2.7) near x0 implies that the initial conditions y 0 ,
z0 (y 0 ), and (p01 (y 0 ), . . . , p0N −1 (y 0 )) = (∂1 g(y 0 ), . . . , ∂N −1 g(y 0 )) are uniquely defined. The
condition E(x0 , z0 (x0 ), (p01 (x0 ), . . . , p0N −1 (x0 ), p0N )) = 0 implies that if p0N is not unique
then necessarily ∂pN E(x0 , z0 (x0 ), (p01 (x0 ), . . . , p0N −1 (x0 ), q)) = 0 for a certain q, which
contradicts the assumption and proves that the initial condition (x0 , z0 (x0 ), p0 (x0 )) is
uniquely defined. Lemma 3.2.2 shows that (y 0 , z0 (y 0 ), p0 (y 0 )) are uniquely defined near
x0 , which completes the proof.
165
9.2.2 Conservation laws and weak solutions

Proposition 9.2.2 Let f ∈ C 1 (R), g ∈ C 0 (R). If u ∈ C 1 (R × R+ ) ∩ C 0 (R × [0, ∞)) is
a classical solution to (3.3.1), then u is a weak solution to (3.3.1). Reciprocally, assume
u is a weak solution to (3.3.1) and u ∈ C 1 (Ω) ∩ C 0 (Ω), where Ω ⊂ R × (0, ∞) is open.
Then u is a classical solution to (3.3.1a) in Ω and satisfies u = g on ∂Ω ∩ {x2 = 0}.
Proof.
x2
Assume u ∈ C 1 (R × R+ ) ∩ C 0 (R × [0, ∞))
is a classical solution to (3.3.1). So u sat-
isfies
0 = ∂2 u + ∂1 (f (u)). supp(ϕ)
x1
For ϕ ∈ D(R2 ), let BR+ = {(x1 , x2 ), x21 + -R R
x22 < R2 , x2 > 0} with R large enough so
that supp(ϕ) ∩ {x2 > 0} ⊂ BR+ ; see Fig. Figure 9.2.1: BR+ and supp(ϕ)
9.2.1. If ν + is the outward unit normal on
∂BR+ , using Gauss theorem in BR+ implies

0 = (∂2 u + ∂1 (f (u))) ϕdx
+
BR

= ϕ(uν2+ + f (u)ν1+ )ds −
(u∂2 ϕ + f (u)∂1 ϕ)dx
+ +
∂BR BR

+ +
= ϕ(uν2 + f (u)ν1 )dx1 + ϕ(uν2+ + f (u)ν1+ )ds
+ +
∂BR ∩{x2 =0} ∂BR ∩{x2 >0}

− (f (u)∂1 ϕ + u∂2 ϕ) dx1 dx2
R+ R

= − ϕ(x1 , 0)g(x1 )dx1 − (f (u)∂1 ϕ + u∂2 ϕ) dx1 dx2 ,
R R+ R
which proves the first part of the claim.

Now we assume u ∈ C 1 (Ω) ∩ C 0 (Ω) is a solution to (3.3.5). For ϕ ∈ D(Ω), we get

0 = (u∂2 ϕ + f (u)∂1 ϕ) dx1 dx2 + g(x1 )ϕ(x1 , 0)dx1
R+ R R

= (u∂2 ϕ + f (u)∂1 ϕ) dx (use Gauss theorem)
Ω

= − (∂2 u + ∂1 (f (u))) ϕdx.
Ω
From the arbitrariness of ϕ and Lemma 1.3.9, it follows ∂2 u + ∂1 (f (u)) = 0 in Ω.
166
Now, let x0 ∈ ∂Ω∩{x2 = 0} and B(x0 , r) be a ball such that B 0,+ := B(x0 , r)∩{x2 >
0} ⊂ Ω, and denote by ν 0,+ the exterior unit normal vector to ∂B 0,+ . Taking ϕ ∈ D(R2 )
with supp(ϕ) ⊂ B(x0 , r) and using Gauss theorem as above gives

0 = (u∂2 ϕ + f (u)∂1 ϕ)) dx + g(x1 )ϕ(x1 , 0)dx1
B 0,+ ∂B 0,+ ∩{x2 =0}

= − (∂2 u + ∂1 (f (u))) ϕdx
B 0,+

+ (uν20,+ + f (u)ν10,+ )ϕds
∂B 0,+ ∩{x2 >0}

+ (g(x1 ) + u(x1 , 0)ν20,+ + f (u)ν10,+ )ϕ(x1 , 0)dx1
0,+
∂B ∩{x2 =0}
= (g(x1 ) − u(x1 , 0))ϕ(x1 , 0)dx1
∂B 0,+ ∩{x2 =0}
which from the arbitrariness of x0 and ϕ proves u(·, 0) = g(·) on ∂Ω ∩ {x2 = 0} and
completes the proof.
Theorem 9.2.3 (Rankine-Hugoniot) Let Ω ⊂ R × R+ be a simply connected open set

such that Ω = Ωl ∪ γ ∪ Ωr , where γ is the graph of a function x1 = χ(x2 ), x2 ∈ (α, β),
γ ∈ C 1 (α, β), and Ωl , resp. Ωr , is the open set at the left, resp. at the right, of γ; see Fig.
3.3.1. Assume that u is a weak solution to (3.3.1) in Ω such that u ∈ C 1 (Ωl ) ∩ C 0 (Ωl )
and u ∈ C 1 (Ωr ) ∩ C 0 (Ωr ), and set
[u] = lim (u(χ(x2 ) + h, x2 ) − u(χ(x2 ) − h, x2 )),
h→0+
[f (u)] = lim (f (u(χ(x2 ) + h, x2 )) − f (u(χ(x2 ) − h, x2 )) x2 ∈ (α, β).
h→0+
Then
χ [u] = [f (u)] on γ. (9.2.8)
Proof. WLOG we may assume that both Ωl and Ωr are smooth, for example Lipschitz,
because otherwise we consider Ωl ∩ B(x, ρ), resp. Ωr ∩ B(x, ρ), instead of Ωl , resp. Ωr ,
with x ∈ γ and ρ small enough. We will use Gauss theorem in Ωl and Ωr , and we denote
by ν l , resp. ν r , the exterior unit normal vector to ∂Ωl , resp. ∂Ωr . As x1 − χ(x2 ) = 0 is
(1, −χ )
the equation of γ, it follows that ν l = = −ν r on γ.
1 + (χ )2
Next we note that as u ∈ C 1 (Ωl ∪ Ωr ) is a weak solution to (3.3.1) in Ωl ∪ Ωr , from
Proposition 3.3.4 it follows that u is a classical solution to (3.3.1) in Ωl and Ωr . Hence
∂2 u + ∂1 (f (u)) = 0 in Ωl ∪ Ωr .
167
As u is a weak solution in Ω, for ϕ ∈ D(Ω) we get

0 = (u∂2 ϕ + f (u)∂1 ϕ)dx + (u∂2 ϕ + f (u)∂1 ϕ)dx.
Ωl Ωr
Now we use Gauss theorem for each of these integrals. By taking into account the fact
that the integrals on ∂Ωl ∩ ∂Ω and on ∂Ωr ∩ ∂Ω vanish, we get

0 = l l
ϕ(uν2 + f (u)ν1 )ds + ϕ(uν2 + f (u)ν1 )ds −
r r
ϕ(∂2 u + ∂1 (f (u)))dx
Ωl ∪Ωr
∂Ωl
∂Ωr
= ϕ(uν2 + f (u)ν1 )ds + ϕ(uν2r + f (u)ν1r )ds

l l
γ γ
= ϕ([u]ν2r + [f (u)]ν1r )ds

γ
ϕ
= (−χ [u] + [f (u)])ds.
γ 1 + |χ | 2
From the arbitrariness of ϕ, the last equality implies (9.2.8).
168
9.3 Second-order linear elliptic PDEs: maximum

principle and classical solutions (Ch. 4)
Throughout this chapter, Ω ⊂ RN is open bounded and u ∈ C 0 (Ω) ∩ C 2 (Ω), unless
otherwise specified.
9.3.1 Dirichlet problem in a ball

Theorem 9.3.1 Let R > 0, BR := {x ∈ RN , |x| < R}, VN the volume of B1 , g ∈
C 0 (∂BR ) and set

R2 − |x|2 g(y)
u(x) = dσ(y), x ∈ BR . (9.3.1)
Then u is the unique C 0 (B R ) ∩ C ∞ (BR ) solution to (4.0.1) in Ω = BR .
Proof. Let K : ∂BR × BR → R be defined by

R2 − |x|2 1
K(x, y) = . (9.3.2)
N VN R |x − y|N

Then u(x) = ∂BR K(x, y)g(y)dσ(y). From direct calculus, one finds that u ∈ C ∞ (BR );
see, for example, [28]. Furthermore, Δu = 0 in BR because

Δu(x) = ∂ii u(x)
i=1,N

2xi 1 R2 − |x|2 (x − y)i
= ∂i − g(y)dσ − g(y)dσ
N VN R ∂BR |x − y|N VN R ∂BR |x − y|
N +2

2 1 4 x · (x − y)
= − g(y)dσ + g(y)dσ
VN R ∂BR |x − y| N VN R ∂BR |x − y|N +2

R2 − |x|2 N R2 − |x|2 N +2
− g(y)dσ + g(y)dσ
∂BR |x − y| ∂BR |x − y|
VN R N +2 VN R N +2

2 1 2 2x · (x − y) + R2 − |x|2
= − g(y)dσ + g(y)dσ
VN R ∂BR |x − y|N VN R ∂BR |x − y|N +2
= 0,
because y · y = R2 and so 2x · (x − y) + R2 − |x|2 = x · x − 2x · y + y · y) = |x − y|2 .

To prove u ∈ C 0 (B R ), we introduce the auxiliary function

u1 (x) = K(x, y)dσ, x ∈ BR , (9.3.3)
∂BR
169
which is obtained from (9.3.1) for g = 1. Thus u1 ∈ C ∞ (BR ) and Δu1 = 0 in BR .

Furthermore, it is easy to check that u1 is radially symmetric, and, then, from the formula
1
of Δu in spherical coordinates3 , we have Δu1 = N −1 ∂r rN −1 ∂r u1 = 0. Therefore, u1
r
R2 1
is constant in BR and as u1 (0) = N
dσ(y) = 1 it follows that u1 = 1 in BR .
∂BR N VN R R
Now, let us prove u ∈ C 0 (B R ). Let x0 ∈ ∂BR and > 0. We have to find δ >
0 such that
∀x ∈ B(x0 , δ) ∩ BR , |u(x) − g(x0 )| < . (9.3.4)
From g ∈ C 0 (∂BR ), there exists δ1 > 0 such that for all y, z ∈ ∂BR with |y − z| < 2δ1

we have |g(y) − g(z)| < . It follows that for x ∈ BR ∩ B(x0 , δ1 ), we have
2
|u(x) − g(x0 )| = |u(x) − g(x0 )u1 (x)|

= K(x, y)(g(y) − g(x0 ))dσ(y)
∂BR
≤ K(x, y)|g(y) − g(x0 )|dσ(y) + K(x, y)|g(y) − g(x0 )|dσ(y)
∂BR ∩{|y−x0 |<2δ1 } ∂BR ∩{|y−x0 |≥2δ1 }
2
R − |x| 2
≤ K(x, y)dσ(y) + 2gC 0 (∂BR ) dσ(y)
2 ∂BR N VN Rδ1N ∂BR
|∂BR |
≤ + 2gC 0 (∂BR ) (R2 − |x|2 ).
2 N VN Rδ1N
As 0 ≤ R2 − |x|2 ≤ 2R|x − x0 |, there exists δ2 > 0 such that for all x ∈ BR ∩ B(x0 , δ2 )
we have
|∂BR |
2gC 0 (∂BR ) N
(R2 − |x|2 ) < .
N VN Rδ1 2
Then (9.3.4) is satisfied with δ = min{δ1 , δ2 } and so u ∈ C 0 (B R ). Finally, the uniqueness

of the solution to (4.0.1) with Ω = BR follows from (ii), Corollary 4.3.2.
9.3.2 Maximum principle for second-order linear elliptic PDEs

Throughout this section, we will consider

N
N
Lu = − aij ∂ij u + bi ∂i u + cu, A = (aij ), b = (bi ), (9.3.5)
i,j=1 i=1
3
If (r, θ) ∈ (0, ∞) × SN −1 are the spherical coordinates connected to the Cartesian coordinates
x by
x 1 N −1 x
x = rθ, θ = , then the classical formula holds: Δx u(x) = N −1 ∂r (r ∂r u(rθ)) + Δx u r .
|x| r |x|
170
and we will assume the following:4
Ω ⊂ RN is an open set, u ∈ C 0 (Ω) ∩ C 2 (Ω), (9.3.6a)

aij , bi , c ∈ C (Ω) for all i, j,
0
(9.3.6b)
A(x) is symmetric for all x ∈ Ω and A(x) > 0 in Ω. (9.3.6c)
Theorem 9.3.2 (weak maximum principle) Assume Ω ⊂ RN is bounded and c = 0.

If Lu ≤ 0, resp. Lu ≥ 0, then
max u = max u, resp. min u = min u,

Ω ∂Ω Ω ∂Ω
where min f := min{f (x), x ∈ A}, max f := max{f (x), x ∈ A}.

A A
Proof. We deal first with the case Lu ≤ 0. Let x0 ∈ Ω such that u(x0 ) = max{u(x), x ∈
Ω}. We distinguish two sub-cases.
(i) Case Lu < 0 in Ω. If x0 ∈ ∂Ω the theorem is proved. So we assume x0 ∈ Ω. We have
|∇u(x0 )| = 0 and D2 u(x0 ) = [∂ij u(x0 )] ≤ 0. Therefore

N

Lu(x ) = −
0
aij (x0 )∂ij u(x0 ) = −tr A(x0 ) · D2 u(x0 ) ≥ 0,
i,j=1
because from A(x0 ) > 0 and D2 u(x0 ) ≤ 0 it follows tr[A(x0 ) · D2 u(x0 )] ≤ 0,5 which is a
contradiction and proves that x0 ∈ ∂Ω.
(ii) Case Lu ≤ 0 in Ω. Let , γ > 0 and consider u = u + eγx1 . We have
Lu = Lu + Leγx1 = Lu + (−γ 2 a11 + γb1 )eγx1 .
We can choose γ large enough such that (−γ 2 a11 + γb1 )eγx1 < 0 in Ω. This is possible
because Ω is bounded, a11 > 0 (this follows from A > 0), and (9.3.6b). So Lu < 0 in
Ω. From case (i), it follows that there exists x ∈ ∂Ω such that

u (x) ≤ u (x ), so u(x) ≤ u(x ) + (eγx1 − eγx1 ), x ∈ Ω.
There exists a subsequence of (x ), still denoted by (x ), and x0 ∈ ∂Ω such that
lim→0 x = x0 , which implies that for all x ∈ Ω we have u(x) ≤ u(x0 ).
4
A matrix M ∈ RN ×N is said to be positive, resp. strictly positive, and we write M ≥ 0, resp.
M > 0, if ξ · M · ξ ≥ 0, resp. ξ · M · ξ > 0, for every 0 = ξ ∈ RN .
5
Indeed, let A(x0 ) = U · Λ · t U be the spectral decomposition of A(x0 ) with U −1 = t U and Λ
the diagonal matrix of eigenvalues of A(x0 ). From the cyclic property of the trace, we get tr[A(x0 ) ·
D2 u(x0 )] = tr[Λ · t U · D2 u(x0 ) · U )] = tr[Λ · B], with B= t U · D2 u(x0 ) · U ; clearly B ≤ 0, because
D2 u(x0 ) ≤ 0, and then tr[A(x0 ) · D2 u(x0 )] = tr[Λ · B] = i λi Bi,i ≤ 0 because λi > 0 and Bi,i ≤ 0 for
all i.
171
For the case Lu ≥ 0 we set v = −u, so Lv ≤ 0, and then the result follows from (ii).
Note that this theorem does not hold if c = 0. For example, if Ω = B(0, 1) ⊂ R2 and
Lu = −Δu − u (so c = −1) then for u = 5 − |x|2 , we have
Lu = −1 + |x|2 ≤ 0 in Ω, and max u = 5 > 4 = max u.

Ω ∂Ω
However, if c ≥ 0 we have the following result.

Corollary 9.3.3 Assume Ω bounded and c ≥ 0. If Lu ≤ 0, resp. Lu ≥ 0, in Ω then
max u ≤ max u+ , resp. min u− ≤ min u, (9.3.7)

Ω ∂Ω ∂Ω Ω
where u+ = max{u, 0} and u− = min{u, 0}.

Proof. Let us consider first the case Lu ≤ 0. Set Ω+ := {x ∈ Ω, u(x) > 0}. As
u ∈ C 0 (Ω) it follows that Ω+ is open. There are two possibilities.
(i) Ω+ = ∅. So u ≤ 0 in Ω and then the claim is proved. N
(ii) Ω+ = ∅. Then cu ≥ 0 in Ω+ . If we set L+ u := − N i,j=1 aij ∂ij u + i=1 bi ∂i u + 0u,
we have L+ u ≤ 0 in Ω+ . Applying Theorem 9.3.2 with L+ in Ω+ (instead of L in Ω)
implies max u+ = max +
u+ .
Ω+ ∂(Ω )
Note that necessarily ∂(Ω+ ) ∩ ∂Ω = ∅, because otherwise ∂Ω+ ⊂ Ω and then
max u+ = max
+
u+ = 0, which is impossible because Ω+ = ∅. Then we get
Ω+ ∂(Ω )
max u = max u = max u+ = max

+
u+ = max u+ ,
Ω Ω+ Ω+ ∂(Ω ) ∂Ω
which proves the corollary in the case Lu ≤ 0.

(iii) The case Lu ≥ 0 is proved by setting v = −u and using Lv ≤ 0.
Note that if max u < 0 or min u < 0, the result of the previous corollary is not optimal.
Ω Ω
Corollary 9.3.4 (comparison principle) Let Ω be open bounded and c ≥ 0 in Ω.

(i) If Lu ≤ Lv in Ω and u ≤ v on ∂Ω, then u ≤ v in Ω (monotonicity of L).
(ii) If Lu = Lv in Ω and u = v on ∂Ω, then u = v in Ω (uniqueness of Lu = 0).

Proof. Set w = u − v.
(i) We have
Lw ≤ 0, c ≥ 0 in Ω, w ≤ 0 on ∂Ω.
From Corollary 9.3.3, it follows that max w ≤ max w+ ≤ 0. So, w ≤ 0, or u ≤ v.

Ω ∂Ω
(ii) As Lw = 0 and w = 0, (i) implies w = 0, so u = v.
172
Lemma 9.3.5 (Hopf lemma) Let Ω ⊂ RN be an open set with ∂Ω a C 2 boundary

near6 x0 ∈ ∂Ω, Lu(x) ≤ 0, u(x) < u(x0 ) for all x ∈ Ω near x0 and u differentiable at
x0 . Furthermore, assume that one of the following holds:
(i) c = 0 near x0 ,
(ii) u(x0 ) ≥ 0 and c ≥ 0 near x0 ,
(iii) u(x0 ) = 0.
Then ∂ν u(x0 ) := ∇u(x0 ) · ν(x0 ) > 0, where ν(x0 ) is the unit outward normal vector to
∂Ω at x0 .
Proof. From u(x0 ) − u(x) ≥ 0 near x0 in Ω, we obtain ∂ν u(x0 ) ≥ 0. The lemma would
be proven if we find a function v ∈ C 1 (Ω) such that u(x0 ) − u(x) ≥ v(x0 ) − v(x) near
x0 and ∂ν v(x0 ) > 0.
Construction of v. Let z ∈ Ω and r > 0 such that if Br (z) is the ball with center z and
radius r, then Br (z) ⊂ Ω, ∂Br (z)∩∂Ω = {x0 }; see Figure 9.3.1. Set V = Br (z)∩Br0 (x0 ),
2 2
with r0 small, and v(x) = (e−k|x0 −z| − e−k|x−z| ), with x ∈ V , , k > 0. Note that
v(x0 ) − v ≥ 0 in V,
2
L(v(x0 ) − v) = − 4k 2 (x − z) · A(x) · (x − z) + 2k(tr(A(x)) − (x − z) · b e−k|x−z|
2 2

−c (e−k|x0 −z| − e−k|x−z| ) ,
L(u(x0 ) − u) ≥ cu(x0 ).
With k large, small, and with one of (i)–(iii) hold-
ing, we obtain r0
x0
v(x0 ) − v ≤ u(x0 ) − u on ∂V, V
L(v(x0 ) − v) ≤ 0 ≤ L(u(x0 ) − u) in V,
z
because v(x0 ) − v = 0 on V ∩ ∂Br (z) and u(x0 ) − Ω

u > 0 on V ∩ ∂Br0 (x0 ), which is compact.
Cases (i) and (ii). As c ≥ 0 in V , by applying
Corollary 9.3.4 we get v(x0 ) − v(x) ≤ u(x0 ) − u(x)
in V . Therefore, if we take x = x0 − tν(x0 ), where Figure 9.3.1: V and x0
t > 0 small, we get v(x0 )−v(x0 −tν(x0 )) ≤ u(x0 )−
u(x0 −tν(x0 )), which after dividing by t and letting
2
t to zero gives 0 < 2kre−kr ≤ ∂ν u(x0 ).
Case (iii). As u(x0 ) = 0, we have u < 0 in Ω near x0 . Set c− = min{c, 0}. It follows
N
N
−
−
that for L defined by L u := − aij ∂ij u + bi ∂i u + (c − c− )u, we have L− u =
i,j=1 i=1
Lu − c u ≤ Lu ≤ 0 in Ω. Then we can apply Case (ii) above with L− u ≤ 0, u(x0 ) = 0
−
and (c − c− ) ≥ 0, which completes the proof.

6
Note that this implies Ω lies on one side of ∂Ω.
173
Remark 9.3.6 In the case of a local minimum, Hopf lemma is stated as follows. Let
Ω ⊂ RN be open with ∂Ω a C 2 boundary near x0 ∈ ∂Ω, Lu(x) ≥ 0, u(x) > u(x0 ) for all
x ∈ Ω near x0 , and u differentiable at x0 . Assume also that one of the following holds:
(i) c = 0 near x0 ,
(ii) u(x0 ) ≤ 0 and c ≥ 0 near x0 ,
(iii) u(x0 ) = 0.
Then ∂ν u(x0 ) < 0. The proof follows from Lemma 9.3.5 applied to −u.
As a consequence of Lemma 9.3.5, we have the following theorem.

Theorem 9.3.7 (Strong maximum principle) Let Ω ⊂ RN be an open bounded set
and assume Lu ≤ 0, resp. Lu ≥ 0, in Ω. Then u is constant or, if not, we have
(i) if c = 0 then u does not achieve its maximum, resp. minimum, in Ω,
(ii) if c ≥ 0 then u does not achieve a non-negative maximum, resp. non-positive mini-
mum, in Ω,
(iii) regardless of the sign of c, u does not achieve a maximum, resp. minimum, in Ω
equal to zero.
Proof. Let us consider first the case Lu ≤ 0. Let M = max{u(x), x ∈ Ω} and set
K = {x ∈ Ω, u(x) = M }. Note that K is closed. If K = Ω then u = M in Ω and so the
theorem is proved. Otherwise, we assume that one of (i)–(iii) fails. Each case implies
that u achieves its maximum value M in Ω. So, K ∩ Ω Ω.
As in Theorem 4.3.4, there exists x0 ∈ ∂K, B0 := B(z0 , r0 ) Ω\K (see Fig. 4.3.1),
such that u achieves at x0 a strict maximum in B 0 , and so |∇u(x0 )| = 0. Note that u
satisfies the conditions of Lemma 9.3.5 at x0 , with B0 instead of Ω. Applying Lemma
9.3.5 implies |∇u(x0 )| > 0. The contradiction shows that one of (i)–(iii) holds.
The result for the case Lu ≥ 0 is obtained by setting v = −u and using the case
Lu ≤ 0.
9.3.3 Solution to the Dirichlet problem

9.3.3.1 Sub(super) harmonic functions and sub(super) solutions
Proposition 9.3.8 Let u ∈ C 0 (Ω) ∩ C 2 (Ω) such that −Δu = 0 in Ω.
i) u is a subharmonic and superharmonic function.
ii) Assume u solves (4.0.1). Then u is a subsolution and supersolution to (4.0.1).
Proof. Let B be a ball, B Ω, h ∈ C 0 (B) ∩ C 2 (B), −Δh = 0 in B, and u ≤ h on

∂B. Up to a translation, we may assume that B = BR , for a certain R > 0. Then from
(4.2.1), for x ∈ BR we have

R2 − |x|2 u(y) − h(y)
(u − h)(x) = dσ(y) ≤ 0.
174
So u is a subharmonic function in Ω. Similarly, we prove that u is a superharmonic. The

Claim ii) is a corollary of i).
Proposition 9.3.9 Let u, v ∈ C 0 (Ω).
i) Assume u is subharmonic and v is superharmonic in Ω with u ≤ v on ∂Ω. Then

either u < v or u = v in Ω.
ii) Assume u is a subsolution and v is a supersolution to (4.0.1). Then either u < v

or u = v in Ω.
Proof. Let us prove first Claim i). Let M = max{u(x) − v(x), x ∈ Ω}, which exists as
u, v ∈ C 0 (Ω). If M < 0 then Claim i) holds. So it remains the case M ≥ 0. If u − v ≡ M
in Ω, the case M = 0 gives u = v in Ω, which again proves i), while the case M > 0 is
impossible because it gives u = M + v > v on ∂Ω, which is a contradiction.
So it remains the case that M ≥ 0 and u−v M in Ω. Set K = {x ∈ Ω, (u−v)(x) =
M }. The set K is nonempty and closed and Ω\K is nonempty and open; see Figure
9.3.2. Consequently,
∃x0 ∈ ∂K ∩ Ω, ∃r0 > 0 such that B0 := B(x0 , r0 ) Ω.
Now consider u, v ∈ C 0 (B 0 ) defined by

Δu = 0 in B0 , Δv = 0 in B0 ,
u = u on ∂B0 , v = v on ∂B0 .
Then
M = u(x0 ) − v(x0 )

1
≤ (u − v)(x0 ) = (u − v)(y)dσ(y) r0
|∂B0 | ∂B0 x0
< M,
Ω\K K
because u − v M on ∂B0 . The contradiction
∂K
shows that this case is impossible and proves i).
The Claim ii) follows from i).
Figure 9.3.2: K, Ω\K, and B(x0 , r0)
Proposition 9.3.10 Let u be subharmonic in Ω, B a ball with B Ω, and u ∈ C 0 (B) ∩

C 2 (B) such that −Δu = 0 in B and u = u on ∂B. Then U , given by
U (x) = u(x), x ∈ B, U (x) = u(x), x ∈ Ω\B,
called “harmonic lifting of u with respect to B”, is also subharmonic in Ω.
175
Proof. Let C be a ball, C Ω, and h ∈ C 0 (C) ∩ C 2 (C) satisfying −Δh = 0 in C,

U ≤ h on ∂C. To prove is U ≤ h in C.
If C ∩ B = ∅ then U ≤ h in C is equivalent to u ≤ h in C, which is true because
u is subharmonic in Ω. Similarly, if C ⊂ B the proposition is also true because U ≤ h
in C is equivalent to u ≤ h in C, which again is true because u is harmonic in C (see
Proposition 4.4.2 with Ω = B).
It remains the case that C ∩B = ∅, C ∩(Ω\B) = ∅;
see Figure 9.3.3. Note that u ≤ U in Ω and U ≤ h
in ∂C, so it follows u ≤ h in ∂C. Therefore, as u is
a subsolution, we have u ≤ h in C. From u = U in
C\B, it follows C
B∩C
U ≤ h in C\B.
B
Furthermore, we have
−ΔU = 0 = −Δh in C ∩ B,
U ≤ h on ∂(C ∩ B) ∩ B, Figure 9.3.3: Harmonic lifting
(from the assumption for h)
U ≤ h on ∂(C ∩ B) ∩ C.
(from U ≤ h in C\B)
Then from (ii), Corollary 4.3.2, it follows U ≤ h in C ∩ B, so U ≤ h in C and this
completes the proof.
Proposition 9.3.11 If u1 , . . . , um are subharmonic functions in Ω, then max{u1 , . . . , um }

is also subharmonic in Ω.
Proof. Define u by u(x) = max{u1 (x), . . . , um (x)}, x ∈ Ω. It is easy to show that

u ∈ C 0 (Ω). Let B be a ball with B Ω and h ∈ C 0 (B) ∩ C 2 (Ω) satisfying −Δh = 0
and u ≤ h on ∂B. We have ui ≤ u ≤ h on ∂B for all i. Therefore, ui ≤ h in B, for all i,
which implies u ≤ h in B.
9.3.3.2 Some auxiliary results from Analysis

To conclude the existence of solutions to (4.0.1), we need a compactness result in
C 0 (B), with B ⊂ RN a ball, for bounded harmonic functions. This result relies on the
concept of uniform equicontinuity of sequences given by this definition.
Definition 9.3.12 Let K ⊂ RN and fn : K → R, n ∈ N, be a sequence of functions.
i) The sequence (fn ) is called “equicontinuous at x ∈ K” if
∀ > 0, ∃δ = δx > 0 s.th. ∀y ∈ K with |x − y| < δ and ∀n ∈ N :

|fn (x) − fn (y)| < .
176
ii) We say (fn ) is equicontinuous in K if (fn ) is equicontinuous at every x ∈ K.
iii) We say (fn ) is uniformly equicontinuous in K if (fn ) is equicontinuous in K with

δ independent of x.
iv) Also (an independent concept from equicontinuity), we say (fn ) is bounded in K
if there exists M ≥ 0 such that |fn (x)| ≤ M for all x ∈ K and n.
Example 9.3.13 Let (fn ) be a uniformly Lipschitz sequence in C 0 (K), i.e. fn ∈ C 0 (K),
for all n, and
∃L > 0, |fn (x) − fn (y)| ≤ L|x − y|, ∀x, y ∈ K, ∀n ∈ N.
Then (fn ) is uniformly equicontinuous in K, because for |x − y| < δ := /L we have

|fn (x) − fn (y)| < for all n.
More generally, if (fn ) is equicontinuous in K and K is compact, then (fn ) is uni-
formly equicontinuous. Indeed, for given > 0 let δx be as in i), Definition 9.3.12.
From the open covering {B(x, δx /2), x ∈ K} of K, there exists a finite covering
{B(xi , δi /2), i = 1, . . . , m} of K. If δ = min{δi /2, i = 1, . . . , m} and x, y ∈ K with
|x − y| < δ, then necessarily x, y ∈ B(xj , δj ), for a certain j, and then i) in Definition
9.3.12 is satisfied with this δ, and for all x, y and n.
The following result states that an equicontinuous uniformly bounded sequence is

compact in C 0 . For its proof see any Real Analysis classic book, for example [28].
Theorem 9.3.14 (Arzela-Ascoli lemma) Let K ⊂ RN be a compact set and (fn )
be a equicontinuous and uniformly bounded sequence in K. Then (fn ) has a convergent
subsequence in C 0 (K).
The following result shows that a bounded sequence of harmonic functions is compact
in a certain sense. It is used in the proof of Theorem 4.4.7 in a substantial way to show
that the maximum of subsolutions to (4.0.1) is a harmonic function; see the comment
after Proposition 4.4.5.
Theorem 9.3.15 Let Ω ⊂ RN be an open set and (un ) be a sequence of functions,
harmonic and uniformly bounded in Ω by a constant M . Then (un ) has a subsequence
(unk ) converging pointwise in Ω to a harmonic function u in Ω. The convergence is in
C 0 (K) for every compact K ⊂ Ω.

2
Proof. For m ∈ N, set Km = x ∈ Ω, d(x, ∂Ω) ≥ , |x| ≤ m , where d(x, ∂Ω) =
m
inf{|x−y|, y ∈ ∂Ω} is the distance of x to ∂Ω. Then Km is compact and Ω = m∈N Km .
For x ∈ Km we have Bm := B(x, m1 ) ⊂ B m ⊂ Ω. Poisson’s formula (4.2.2) for un in
Bm gives

2
Rm − |z − x|2 un (y) 1
un (z) = dσ(y), z ∈ Bm , R m := .
∂Bm |z − y|
N VN Rm N m
177
As (un )n∈N is uniformly bounded, this formula gives a uniform bound for ∇un in Km .
Indeed, differentiating un (z) at z = x gives

2|x − x| M Rm2
M
|∇un (x)| ≤ N
dσ(y) + N +1
dσ(y) =: C(m), (9.3.8)
N VN Rm ∂Bm Rm VN Rm ∂Bm Rm
with a certain C(m) > 0.

It follows that (un )n∈N satisfies the conditions of Theorem 9.3.14 in Km (because the
bound (9.3.8) implies that (un ) is a uniformly Lipschitz sequence; see Example 9.3.13).
Therefore, there exists a subsequence (um n )n∈N of (un )n∈N and u ∈ C (Km ) such that
m 0
lim um m 0
n = u in C (Km ).
n→∞
We set (u0n ) = (un ). From the reasoning above, we obtain (u1n )n∈N , a subsequence
of (u0n )n∈N , and v 1 ∈ C 0 (K1 ) such that lim u1n = v 1 in C 0 (K1 ). Next we proceed with
n→∞
(um
n )n∈N as with (un
m−1
)n∈N , m = 1, 2, . . .. As a result of this process, we obtain the
sequences (un )n∈N , and v m ∈ C 0 (Km ), m = 1, 2, . . ., such that lim um
m m 0
n = v in C (Km ).
n→∞
Set un = unn for all n ∈ N, and v(x) = limn→∞ v n (x). Note that v(x) is well-defined
in Ω, because x ∈ Km , for a certain m, and from the construction, we have that up to
a finite number of terms, (un )n∈N is a subsequence of (um
n )n∈N .
The sequence (u ) and v satisfies the theorem. Indeed, from the definition of (un )
n
we have lim un = v in C 0 (Km ), because (un ) is (up to a finite number of terms) a

n→∞
subsequence of (umn )n∈N , for every m.
It remains to prove that u is harmonic in Ω. Let B be a ball with radius R, B Ω.
R2 − |z − x|2 un (y)
Then un (z) = dσ, z(y) ∈ B. Passing to the limit n → ∞
∂B |z − y|
N VN R N
inside the integral is allowed, because the convergence of un is uniform in B (B ⊂

R2 − |z − x|2 u(y)
Km for m large). This implies u(z) = dσ(y) and then from
∂B |z − y|
N VN R N
Proposition 4.2.4 we obtain −Δu = 0 in Ω.
9.3.3.3 Proof of Theorem 4.4.7

Set Sg = {v : Ω → R, v subsolution to problem (4.0.1)} and u(x) = sup{v(x), v ∈
Sg }. The function u is well-defined. Indeed, there exist subsolutions and supersolutions
to (4.0.1), for example v := −g||C 0 (∂Ω) is a subsolution and v := g||C 0 (∂Ω) is a super-
solution. Then, Sg is not empty and from Corollary 4.4.3 for all v ∈ Sg we have v ≤ v,
which proves that u is well-defined. Note also that, without restriction, we may assume
that (vn ) is uniformly bounded in Ω.
Claim: The function u is harmonic in Ω.
Proof. Let x ∈ Ω and (vn ) in Sg with lim vn (x) = u(x). Let B = B(x, R) with R > 0
n→∞
and B Ω, and consider Vn , the harmonic lifting of vn w.r.t. B. As vn ≤ Vn and Vn is a
subsolution, we get u(x) = lim Vn (x). From Theorem 4.4.6 there exists a subsequence
n→∞
178
of (Vn ), still denoted by (Vn ), converging pointwise in B to a harmonic function V in

B, and converging uniformly in any compact subset of B.
Of course V ≤ u in B. In fact we have V = u in B. Indeed, assume for the moment
that V (y) < u(y) for a certain y ∈ B. Then there exists w ∈ Sg such that V (y) < w(y).
Set wn = max{w, Vn } and let Wn be the harmonic lifting of wn w.r.t. B. As for (Vn ),
a certain subsequence of (Wn ), still denoted by (Wn ), converges pointwise in B and
uniformly in any compact subset of B to a certain harmonic function W in B. As
max{w, Vn } ≤ Wn in B, it follows
V ≤ W in B, V (y) < w(y) ≤ W (y) and V (x) = W (x) = u(x).
From the strong maximum principle, Theorem 4.3.4, V − W must reach the maximum
on ∂B (and not in B) or, otherwise it must be constant. Therefore, as (V − W )(x) = 0
and V − W ≤ 0 in B, it follows V − W = 0 in B, which contradicts V (y) − W (y) < 0. So
V = u in B and therefore u is harmonic in Ω. Theorem 4.2.1 implies that u ∈ C ∞ (Ω).
Claim: u ∈ C 0 (Ω) and u = g on ∂Ω.
Proof. It is enough to prove that u is continuous at every ξ ∈ ∂Ω and u = g on ∂Ω. Let
c
ξ ∈ ∂Ω, y ∈ Ω , R > 0, and B = B(y, R) such that B ∩ Ω = ∅, ∂B ∩ ∂Ω = ξ; see Figure
4.4.1. The choice of ξ and B is possible because ∂Ω is C 2 . Now consider the function w
given by
|x − y|
w(x) = ln , for N = 2,
R
w(x) = R2−N − |x − y|2−N , for N ≥ 3,
which satisfies
w ∈ C ∞ (RN \{y}), −Δw = 0 in Ω, w > 0 in Ω\{ξ}, w(ξ) = 0. (9.3.9)
Set v − (x) = g(ξ) − ( + kw(x)), resp. v + (x) = g(ξ) + ( + kw(x)), with > 0 given.
Clearly,
−Δv − = −Δv + = 0 in Ω.
Assume for a moment that v − ≤ g ≤ v + on ∂Ω, i.e. v − (x), resp. v + (x), is a subsolution,
resp. supersolution, to (4.0.1). Then from Corollary 4.4.3 and the fact that u(x) =
sup{v(x), v ∈ Sg }, we get
g(ξ) − ( + kw(x)) = v − (x) ≤ u(x) ≤ v + (x) = g(ξ) + ( + kw(x)).
This proves |u(x) − g(ξ)| ≤ + kw(x) for all x ∈ Ω, which implies lim u(x) = g(x)
x→ξ
because w is continuous and w(ξ) = 0.
Now we show that for an appropriate choice of , δ, the function v − , resp. v + , is a
subsolution, resp. supersolution. Let M := gC 0 (∂Ω) and choose , δ, k positive such that
|g(x) − g(ξ)| < , if x ∈ ∂Ω ∩ B(x, δ),
179
kw(x) ≥ 2M, if x ∈ Ω\B(x, δ).
Then
−Δv − = 0 in Ω,
v − (x) ≤ g(x) + |g(ξ) − g(x)| − − kw(x)

g(x) + − − kw(x), for |x − ξ| < δ
≤
g(x) + 2M − − 2M, for |x − ξ| ≥ δ
< g(x), ∀x ∈ ∂Ω.
We prove in a similar way g ≤ v + , which proves the claim.

Claim: The solution is unique.
Proof. This is a direct consequence of the comparison principle, Corollary 4.3.2.
180
9.4 Distributions (Ch. 5)

9.4.1 Some useful inequalities
The following inequalities are useful when working with distributions, Fourier trans-
form, Lp , and Sobolev spaces.
1 1
Hölder’s inequality: ∀p, p ∈ [1, ∞], + = 1, ∀(u, v) ∈ Lp (Ω) × Lp (Ω) :
p p

|u(x)v(x)|dx ≤ uLp (Ω) vLp (Ω) , (9.4.1)
Ω
1 1
Young’s inequality: ∀a, b ≥ 0, ∀p, p ∈ [1, ∞], + =1:
p p
1 p 1
ab ≤ a + bp , (9.4.2)
p p
Useful growth equiv. inequal: ∃C1 , C2 > 0, ∀ξ ∈ RN , ∀k ∈ N, ∀α ∈ NN0 , ∀s > 0 :

2k α 2 2k
C1 (1 + |ξ| ) ≤ |ξ | ≤ C2 (1 + |ξ| ), (9.4.3)
|α|≤k
C1 (1 + |ξ|2s ) ≤ (1 + |ξ|2 )s ≤ C2 (1 + |ξ|2s ). (9.4.4)
9.4.2 More operations with distributions. Examples

Product of distributions
In general, the product T1 T2 of two distributions is not well-defined. For example, if
1
T1 = T2 = √ then T1 , T2 ∈ D but T1 T2 ∈ D .
x
Even if T ∈ D (Ω) and k ∈ / C ∞ (Ω) then kT is not a distribution. For example,
ϕ(x) 1
T ∈ D (R), T, ϕ = dx, is a well-defined distribution. However, if k(x) =
R |x| |x|
ϕ(x)
then kT is not a distribution because the integral T, kϕ = dx, is not well-
R |x|
defined for all ϕ ∈ D(R), for example for ϕ such that ϕ(0) = 0. However, if k ∈ C ∞ (Ω)
then kT ∈ CD (Ω).
Change of variable
Let θ ∈ C ∞ (RN ; RN ) with inverse θ−1 ∈ C ∞ (RN ; RN ). If T ∈ D (RN ), we can define
the distribution T ◦ θ−1 ∈ D (RN ) by
T ◦ θ−1 , ϕ = T, ϕ ◦ θ| det[∇θ]|. (9.4.5)
181
The motivation for this definition is the following. Assume T = f ∈ L1 (RN ). Then

−1 −1 −1
T ◦ θ , ϕ = f ◦ θ , ϕ = f (θ (x))ϕ(x)dx = f (x)(ϕ ◦ θ)(x)|det[∇θ]|dx.
RN RN
Example 9.4.1 If θ(x) = Ax+b with A = (aij )i,j=1,N a N ×N invertible matrix, b ∈ RN

then θ−1 (x) = A−1 (x − b) and T ◦ θ−1 is defined by
T ◦ θ−1 , ϕ = T, ϕ(Ax + b)|det(A)|.
In particular, if θ(x) = x + a, a ∈ RN , then θ−1 (x) = x − a, so T ◦ (x − a), ϕ =

T, ϕ ◦ (x + a). For T = δ0 , we have
δ0 (x − a), ϕ = δ0 , ϕ(x + a) = ϕ(a).
We denote δa = δ0 ◦ (· − a). Hence δa , ϕ = ϕ(a).
Example 9.4.2 Let n ∈ N, f ∈ D (R) and let us look for the solutions T ∈ D (R) to
xn T = 0. (9.4.6)
First we consider n = 1. Then the solution is T = Cδ0 , with C constant. Indeed, clearly
ϕ(x) − ϕ(0)η(x)
T = Cδ0 is a solution. Conversely, for ϕ ∈ D set ψ(x) = , where η is a
x
ϕ(x)(1 − η(x)) 1
{[−1, 1], (−2, 2)} cut-off function. As ψ(x) = + η(x) ϕ (tx)dt, we get
x 0
ψ ∈ D. Then
0 = xT, ψ = T, ϕ − ϕ(0)T, η; so T, ϕ = T, ηδ0 , ϕ,
which proves the claim.

1
n−1 k
x
For the general case, we take ψ(x) = n ϕ(x) − η(x) ϕ(k) (0) . Then ψ ∈
x k!
k=0
ϕ(x)(1 − η(x)) η(x) 1
D(R), because ψ(x) = + ϕ(n) (tx)(1 − t)n−1 dt. Then
x n (n − 1)! 0
0 = x T, ψ = T, x ψ
n n
ϕ (0)
n−1 (k)

= T, ϕ − T, ηxk ; hence
k=0
k!

n−1
(k)
T = Ck δ0 , Ck constants.
k=0
182
Example 9.4.3 Let f ∈ D (R) and let us look for the solutions T ∈ D (R) to
xT = f. (9.4.7)
Note that we have already seen the cases f = 0 and f = 1. If we know one particular
solution S to (9.4.7), then all the solutions are given by T = Cδ0 + S, C constant.
Now let us find an S. Inspired by (5.2.13), we define pv(x−1 f ) by
pv(x−1 f ), ϕ = f, x−1 ϕ(1 − η) + f, x−1 (ϕ − ϕ(0))η, (9.4.8)
where η is a {{0}, (−, )} cut-off function, > 0. It is easy to show that T = pv(x−1 f ) ∈
D and it solves (9.4.7). So, S = pv(x−1 f ) is a particular solution to (9.4.7), and all
solutions to (9.4.7) are of the form T = Cδ0 + pv(x−1 f ).
Note that the solution does not depend on or η because if ηi , i = 1, 2, is a
{{0}, (−i , i )} cut-off function and Si = pv(x−1 T ) is defined by (9.4.8) with η = ηi ,
then S1 − S2 = Cδ0 .
Example 9.4.4 Let f ∈ D (R) and look for T ∈ D , solution to
T = f. (9.4.9)
x x
For ϕ ∈ D set ψ(x) = −∞ ϕ(t)dt− −∞ ϕ1 (t)dt1, ϕ, where ϕ1 ∈ D with R ϕ1 (t)dt = 1.
Then ψ ∈ D. Assuming T exists implies
−f, ψ = T, ψ = T, ϕ − T, ϕ1 1, ϕ.
Hence, if C = T, ϕ1 we get
T, ϕ = C, ϕ − f, ψ. (9.4.10)
Every T given by (9.4.9) solves (9.4.9), so (9.4.10) gives all the solutions to (9.4.9).
As an application, let us find the solutions to
xT = 0, x ∈ R, T ∈ D (R). (9.4.11)
From Example 9.4.2 we have T = Dδ0 , D constant. From (9.4.10) with f = Dδ0 we get
0 0
T, ϕ = C, ϕ − Dδ0 , ψ = C, ϕ − D ϕ(t)dt + D ϕ1 (t)dt1, ϕ
−∞ −∞
= C, ϕ + DH, ϕ,
for certain constants C, D, D = 0. So T = C + DH, where H is the Heaviside function.
183
9.4.3 Convergence of distributions. Distributions of finite order

Definition 3), 5.2.9, is consistent, thanks to the following theorem.
Theorem 9.4.5 Let Tn ∈ D (Ω) such that lim Tn , ϕ exists for all ϕ ∈ D(Ω). Set
n→∞
T : D(Ω) → R, T, ϕ = lim Tn , ϕ. Then T ∈ D (Ω) and lim Tn = T in D (Ω).
n→∞ n→∞
Proof. We refer the reader to [23] or [22] for a proof, which uses the uniform bound-
edness principle (Banach-Steinhaus theorem). For a proof by contradiction, the reader
can see [43].
Any distribution T ∈ D (Ω) restricted in any compact K is of finite order, with the
order eventually tending to infinity as K tends to Ω; see [22, 23]. The following theorem
shows that the distributions with compact support are of finite order.
Theorem 9.4.6 Let Ω ⊂ RN be an open set and T ∈ D (Ω) with compact support. Then
T is of finite order; see (5.2.6).
Proof. Let G Ω be open with supp(T ) ⊂ G, and η ∈ D a {supp(T ), G} cut-off
function. Then
T, ϕ = T, ηϕ + T, (1 − η)ϕ = T, ηϕ, ∀ϕ ∈ D(Ω).
Now, assume T is not of finite order. Then
∀m ∈ N, ∃ϕm ∈ D(Ω), |T, ϕm | = |T, ηϕm | > mϕm C m (Ω) .

1 ηϕm
Set ψm = . Then |T, ψm | > 1. But limm→∞ ψm = 0 in D(Ω), because
m ϕm C m (Ω)
Ck
supp(ψm ) ⊂ G and for given k ∈ N we have ψm C k (Ω) ≤ . Passing to the limit in
m
|T, ψm | > 1 implies 0 ≥ 1, which is a contradiction and proves the theorem.
Remark 9.4.7 Let T ∈ D (Ω) with supp(T ) compact. Then T can be extended to a
linear “continuous” map T : C ∞ (Ω) → R as follows.
Let G Ω with supp(T ) ⊂ G, and η ∈ D be a {supp(T ), G} cut-off function. Every
ϕ ∈ C ∞ (Ω) is written as ϕ = ϕη + ϕ(1 − η), with supp(ϕ(1 − η)) ⊂ N (T ), which implies
T, ϕ = T, ηϕ. Then T , the extension of T , is defined by
T , ϕ = T, ηϕ, ∀ϕ ∈ C ∞ (Ω).
Clearly T , ϕ = T, ϕ, for every ϕ ∈ D(Ω). Note also that T does not depend on the
choice of η because if ηi , i = 1, 2, are two cut-off functions as above then T, (η1 −η2 )ϕ =
0 because supp(η1 − η2 ) ⊂ N (T ). Finally, it is easy to check that T is continuous in the
sense:
lim T , ϕn = 0, ∀ϕn ∈ C ∞ (Ω) such that ∀α ∈ NN α 0

0 , lim D ϕn = 0 in Cloc (Ω).
n→∞ n→∞
184
9.4.4 Convolution of distributions

When dealing with the convolution of distributions, we have to consider functions
of two variables; let us say ϕ = ϕ(x, y) ∈ D(RN × RN ). In order to identify which
variable in the distribution T ∈ D is active, we will write T (x), or T (y). For example,
T (x), ϕ(x, y) := T, η, where η(x) = ϕ(x, y), y fixed. Similarly, T (y), ϕ(x, y) :=
T, ζ, where ζ(y) = ϕ(x, y), x fixed.
The following two auxiliary results show that the derivative and the integration
commute with the distribution pairing. They are needed in order to prove that the
convolution of two distributions is well-defined, as well as the density of D in D .
Lemma 9.4.8 (Derivative and distribution paring) Let T ∈ D (RN ), ϕ ∈ C ∞ (U ×
RN ), with U ⊂ RN open.
a) If ϕ satisfies: ∀x ∈ U , ∃r > 0 and K ⊂ RN compact such that
i) ∀x̃ ∈ B(x, r), supp(ϕ(x̃, ·)) ⊂ K, and

ii) ∀α, β ∈ NN 0 , Dx Dy ϕC 0 (B(x,r)×RN ) < ∞,
α β

T (y), ϕ(·, y) ∈ C ∞ (U ),
then (9.4.12)
Dα T (y), ϕ(x, y) = T (y), Dxα ϕ(x, y), ∀x ∈ U, ∀α ∈ NN
0 .
b) Furthermore, if T has compact support and ϕ(x, y) := ϕ(x+y), where ϕ ∈ D(RN ),

then (9.4.12) holds, T (y), ϕ(· + y) ∈ D and supp(T (y), ϕ(· + y)) ⊂ supp(ϕ) −
supp(T ).
Proof. Let us prove (9.4.12) for |α| = 1, for example α = (1, 0, . . . , 0). For x ∈ U , let r
and K be as in i) and ii) above. Let e = (1, 0, . . . , 0), 0 = h ∈ R, and consider
T (y), ϕ(x + he, y) − T (y), ϕ(x, y)
z(h) = − T (y), Dxα ϕ(x, y)
h
= T (y), h (y), with
1
h (y) := (ϕ(x + he, y) − ϕ(x, y) − hDxα ϕ(x, y)) .
h
Note that supp(h ) ⊂ K for |h| < r. Next, for every β ∈ NN
0 we have

1
|D h (y)|
h Dy ϕ(x + he, y) − Dy ϕ(x, y) − hDx Dy ϕ(x, y)
β β β α β
=

1
= hDx Dy ϕ(x + the, y) ≤ 1 hDx2α Dyβ ϕC 0 (B(x,r)×RN )
2α β
(here t ∈ (0, 1))
2 2
h→0
−−−→ 0,
185
which proves lim h = 0 in D. Hence, lim z(h) = 0, which proves (9.4.12) for |α| = 1.
h→0 h→0
One can prove (9.4.12) for arbitrary |α| = 1 by induction. This completes the proof of
Claim a).
For Claim b), we note that (9.4.12) holds because ϕ(x, y) satisfies the condition of
i) and ii). It remains to show that T (y), ϕ(· + y) has compact support. Let G ⊂ RN
be open with supp(T ) ⊂ G and η a {supp(T ), G} cut-off function. Then
ψ(y) := T (y), ϕ(x + y) = T (y), η(y)ϕ(x + y).
Clearly, if x is such that η(·)ϕ(x + ·) = 0 then ψ(x) = 0. This implies supp(ψ) ⊂
{x, η(·)ϕ(x + ·) = 0} ⊂ {x ∈ supp(ϕ) − supp(η)}, where in general A − B =
∪{p − q, p ∈ A, q ∈ B} for any A, B ⊂ RN . As supp(η) ⊂ G, and G is an
arbitrary open set including supp(T ), we obtain that x ∈ supp(ϕ) − supp(T ). So
supp(T (y), ϕ(· + y)) ⊂ supp(ϕ) − supp(T ).
The following two results are corollaries of this lemma.

Lemma 9.4.9 (Integration and distributing paring) Let T ∈ D , φ ∈ D(RN ×RN ).
Then
!
T (y), φ(x, y)dx = T (y), φ(x, y)dx . (9.4.13)
RN RN
Proof. Note that from Fubini’s theorem 9.1.3, we have

φ(x, y)dx = · · · φ(z1 , . . . , . . . , zN , y)dzN . . . dz1 . (9.4.14)
RN R R
x1 xN
Set ϕ(x, y) = ··· φ(z1 , . . . , zN , y)dzN · · · dz1 . The function ϕ satisfies the con-
−∞ −∞
ditions of Lemma 9.4.8 with U = RN and K = {y ∈ RN , φ(·, y) = 0} regardless of x,
which implies that T (y), ϕ(·, y) is a C ∞ (RN ) function and
T (y), φ(x, y) = T (y), ∂x1 · · · ∂xN ϕ(x, y) = ∂x1 · · · ∂xN T (y), ϕ(x, y).
Integrating this equality in RN gives (9.4.13).
Proposition 9.4.10 T ∗ S as defined by (5.3.1) is an element of D .
Proof. Indeed, let ϕ ∈ D. From Lemma 9.4.8, T (y), ϕ(·+y) ∈ D and supp(T (y), ϕ(·+
y)) ⊂ supp(ϕ) − supp(T ), so T ∗ S, ϕ is well-defined for every ϕ.
Next, we prove that if (ϕn ) is a sequence in D with limn→∞ ϕn = 0 in D, then
limn→∞ T ∗ S, ϕn = 0. For this, we note that if ψn (x) = S(y), ϕn (x + y)), clearly
we have ψn ∈ D with supp(ψn ) ⊂ supp(ϕn ) − supp(T ) ⊂ K − supp(T ), where K is a
compact including the support of all ϕn . Furthermore, for α ∈ NN α
0 we have D ψn (x) =
S(y), D ϕn (x + y), and as S is of finite order (see Theorem 5.2.15), say m, we have
α
|Dα ψn (x)| ≤ |S(y), Dα ϕn (x + y)| ≤ Cϕn C m+|α| (RN ) ,

which proves limn→∞ ψn = 0 in D. Hence limn→∞ T ∗ S, ϕn = limn→∞ T, ψn = 0.
186
9.4.5 Tempered distributions and Fourier transform

Proposition 9.4.11 We have
i) {xα Dβ u, u ∈ S} ⊂ S, for all α, β ∈ NN

0 .
ii) D ⊂ S, with dense inclusion, i.e. for every u ∈ S there exists (un ) in D such that
limn→∞ un = u in S.
iii) S ⊂ Lp , p ∈ [1, ∞], with dense inclusion for p ∈ [1, ∞), i.e. for every u ∈ Lp there
exists (un ) in S such that limn→∞ un = u in Lp .
iv) Lp ⊂ S , p ∈ [1, ∞], with dense inclusion, i.e. for every u ∈ S there exists (un )
in Lp such that limn→∞ un = u in S .
v) S ⊂ D with dense inclusion, i.e. for every u ∈ D there exists (un ) in D such
that limn→∞ un = u in D .
Proof of i). Set v = xα Dβ u. For arbitrary a, b ∈ Nn0 , we have

|xa Db v| ≤ C(b)|xa | Dc xα Dβ+b−c u| ≤ C(α, β, a, b)sm (u),
|c|≤|b|
with m = |α| + |a| + |β| + |b|, which proves i).

Proof of ii). Clearly D ⊂ S. Let us prove that the inclusion is dense. Let u ∈ S and
η ∈ D be a {B(0, 1), B(0, 2)} cut-off function. If un (x) = η(x/n)u(x) then un ∈ D.
Furthermore, for α, β ∈ NN0 we have
|xα Dβ (un − u)| = |xα Dβ ((1 − η(x/n))u)|

≤ C(β)|xα | |1 − η(x/n)||Dβ u| + |Dγ (η(x/n))Dβ−γ u|
0<γ≤β
|x|
≤ C(β)|x||α| ηC 1 (RN ) |Dβ u| +
n
1
ηC |β| (RN ) |x||γ| |Dβ−γ u|
n 0<γ≤β
C(β)
≤ ηC 1+|β| (RN ) sm (u)
n
n→∞
−−−→ 0,
with m = 1 + |α| + |β|, which proves lim un = u in S.

n→∞
Proof of iii). Note that ζ(x) = (1 + |x|2 )−N ∈ Lp , p ∈ [1, ∞], and
|u(x)| = |ζ(x)(1 + |x|2 )N u| ≤ C|ξ(x)|s2N (u),
187
which proves the inclusion S ⊂ Lp , p ∈ [1, ∞]. The density of the inclusion for p ∈ [1, ∞)
follows from ii) and Theorem 1.3.5.

Proof of iv). We have L ⊂ S in the sense that if u ∈ L then Tu defined by Tu , ϕ =
p p
RN
uϕdx, ϕ ∈ S, defines an element of S . It is easy to verify that, indeed, Tu ∈ S .
For the density of Lp ⊂ S , one can prove that S ⊂ S , with dense inclusion in the
sense that for every T ∈ S there exists (Tn ) in S such that limn→∞ Tn = T in S .
The proof of this result is technical and we will omit the details. It follows the same
steps as the proof of the dense embedding D ⊂ D ; see Theorem 5.3.3. Namely, take
Tn (x) = T (y), ρn (x − y), where (ρn ) is a regularizing sequence with ρn (x) = ρn (−x).
Next, like in Lemma 9.4.8 one can show that Tn ∈ S. Furthermore, like in Lemma 9.4.9
one can prove that Tn , ϕ = T, ρn ∗ ϕ, which shows that Tn → T in S .
Proof of v). We have S ⊂ D in the sense that if T ∈ S , then T restricted in D is an
element of D . For the density of the inclusion, let ηn be a {B(0, n), B(0, 2n)} cut-off
function, n ∈ N. For T ∈ D consider Tn = ηn T . Then Tn ∈ S and limn→∞ Tn = T in D .
Proposition 9.4.12 Let T ∈ S . Then T satisfies
∃C > 0, ∃m ∈ N0 , ∀ϕ ∈ S, |T, ϕ| ≤ Csm (ϕ). (9.4.15)
For every T ∈ S satisfying (9.4.15) we say “T has finite order”, and m is the order of T .
Proof. The proof is similar to the proof of Theorem 5.2.15 (Theorem 9.4.6). Indeed,
assume (9.4.15) does not hold. Then
∀n ∈ N, ∃ϕn ∈ S, |T, ϕn | > nsn (ϕn ).
1 ϕn
Set ψn = . Then |T, ψn | > 1 and limn→∞ ψn = 0 in S because sm (ψn ) ≤ n−1
n sn (ϕn )
for n > m. Passing to the limit in |T, ψn | > 1 implies 0 ≥ 1, which is a contradiction
and proves the proposition.
Theorem 9.4.13 Fourier transform F is a linear continuous bijection from S to itself.

Its inverse, denoted by F−1 , is continuous from S to itself, and is given by

−1 1
F [U ](x) = ei(x·ξ) U (ξ)dξ, ∀U ∈ S. (9.4.16)
(2π)N/2 RN
Furthermore, similar to (5.4.4), (5.4.5), for all u ∈ S and α, β ∈ NN

0 we have
Dα u(x) = F−1 [(iξ)α û](x), (9.4.17)

(−ix)β u(x) = F−1 [Dβ û](x). (9.4.18)
188
Proof.
i) Let us prove first F[S] ⊂ S and the continuity of F. For u ∈ S let û = F[u]. Note
that û ∈ C ∞ . From (5.4.6), we get
|ξ α Dβ û(ξ)| ≤ CDα (xβ u)L1 (RN )

≤ C(α) Dγ xβ Dα−γ uL1 (RN )
|γ|≤|α|
" "
" 1 " " "
≤ C(α) " " "(1 + |x|2 )N Dγ xβ Dα−γ u" ∞ N
" (1 + |x|2 )N " L (R )
|γ|≤|α| L1 (RN )
≤ C(α, N )sm (u), m = |α| + |β| + 2N , (9.4.19)
which proves û ∈ S and the continuity of F.

ii) Now let us prove (9.4.16).7 Let u ∈ S, û = F[u] and for z ∈ RN set

−1 1 1
v(z) = F [û](z) = ei(ξ·z)
û(ξ)dξ = ei(z−x)·ξ u(x)dxdξ.
(2π)N/2 RN (2π)N RN RN
We want to prove v(z) = u(z). For this, we want to use Fubini’s theorem in the previous
formula, which is not allowed because ei(z−x)·ξ ∈ L1 . However, we can proceed as follows.
For > 0 set

1 2
v (z) = N/2
ei(z·ξ) e−|ξ| û(ξ)dξ
(2π) N
R
1 1 2
= N/2
u(x) N/2
e−i(x−z)·ξ e−|ξ| dxdξ.
(2π) RN (2π) RN
Then from Lebesgue’s DCT, we have lim v (z) = v(z). Now, for v we can use Fubini’s
→0
theorem, which together with (5.4.10) gives

1 1 2
v (z) = N/2
u(x) N/2
e−i(x−z)·ξ e−|ξ| dξdx
(2π) N (2π) RN
R
1 2
= N/2
u(x)F[e−ξ ](x − z)dx
(2π) N
R
1 2
= N/2
u(z + x)F[e−|ξ| ](x)dx
(2π) RN

1 1 2
= N N/2 N/2
u(z + x)e− 4 x dx (using (5.4.10))
2 π RN

1 √ −x2
= u(z + 2x )e dx
π N/2 RN

1 2
e−x dx
→0
−−→ u(z) N/2 (using (5.4.7))
π RN
7
See Corollary 5.4.15 for another proof.
189
= u(z).
Hence F−1 [û] = u. It follows F[S] = S, because for every u ∈ S we have
u(x) = F−1 [F[u](ξ)](x) = F−1 [û(ξ)](x) = F[û(−ξ)](x)
and û(−ξ) ∈ S, which proves that F is invertible from S to itself, and its inverse F−1
is given by (9.4.16).
iii) Finally, (9.4.17) and (9.4.18) follow by simple calculations.
Theorem 5.4.9 shows that F is an isometry in L2 . The following two results deal
with the action of F in Lp spaces.
Lemma 9.4.14 (Riemann-Lebesgue lemma) Let u ∈ L1 . Then
ûL∞ ≤ (2π)−N/2 uL1 , û ∈ C 0 ∩ L∞ , lim û(ξ) = 0. (9.4.20)

|ξ|→∞
Proof. The L∞ bound follows from

−N/2
|û(ξ)| ≤ (2π) |e−i(ξ·x) u(x)|dx ≤ (2π)−N/2 uL1 (RN ) .
RN
The continuity of û follows from the Lebesgue DCT, because
lim (e−i(ξ·x) − e−i(η·x) )u(x) = 0, ∀x ∈ RN , |e−i(ξ·x) u(x)| ≤ |u(x)| ∈ L1 .

η→ξ
For the limit at infinity, assume first that u ∈ D. Note that Dα u ∈ D for all α ∈ NN
0 .
Therefore, from (5.4.5) we have F[Δu] = −ξ û. Then from (9.4.20) we get
2
F[Δu]L∞ ΔuL1 1
|û(ξ)| ≤ 2
≤ · , |ξ| = 0, (9.4.21)
ξ (2π)N/2 ξ 2
which proves the limit in (9.4.20) for u ∈ D.

Now let u ∈ L1 . From the density of D in L1 , for > 0 there exists ϕ ∈ D such that

u − ϕL1 (RN ) < (2π)N/2 .
2
From (9.4.21) and the bound (9.4.20), it follows
ΔϕL1 (RN ) 1
|û(ξ)| ≤ |(û − ϕ̂)(ξ)| + |ϕ̂(ξ)| ≤ (2π)−N/2 u − ϕL1 + < ,
(2π)N/2 ξ 2
2 ΔϕL1
for |ξ|2 > , which proves the limit in (9.4.20).
(2π)N/2
The action of F in Lp spaces is more complex. For p ∈ [1, 2] we have this result.
190
Theorem 9.4.15 For p ∈ [1, 2], F maps continuously Lp in Lq , where 1/p + 1/q = 1.
Proof. For p = 1 and q = ∞ the theorem is proven in Lemma 9.4.14, and for p = 2
and q = 2 in Theorem 5.4.9. For p ∈ (1, 2) and q ∈ (2, ∞), the theorem follows from
Riesz-Thorin interpolation theorem 9.4.16 with p0 = 1, q0 = ∞, p1 = 2, q1 = 2,
1/p = (1 − θ)/p0 + θ/p1 = (2 − θ)/2, 1/q = (1 − θ)/q0 + θ/q1 = θ/2, and T = F.
Note that for u ∈ Lp , p ∈ (2, ∞), in general F[u] is not a function; see [49].
Theorem 9.4.16 (Riesz-Thorin interpolation theorem) Let 1 ≤ p0 , q0 , p1 , q1 ≤

∞ and T : Lp0 + Lp1 → Lq0 + Lq1 be a bounded linear operator from Lp0 into Lq0 and Lp1
1 1−θ θ 1 1−θ θ
into Lq1 . For θ ∈ (0, 1), let pθ and qθ be defined by = + , = + .
pθ p0 p 1 qθ q0 q1
Then T is linear and bounded from Lpθ into Lqθ and satisfies
T L(Lpθ ;Lqθ ) ≤ T 1−θ
L(Lp0 ;Lq0 ) T L(Lp1 ;Lq1 ) .
θ
(9.4.22)
See [49] for the proof.
9.4.6 Tempered distributions and convolution

In this section, we will consider the convolution of tempered distributions.
Theorem 9.4.17 The convolution operator ∗ is continuous from S × S onto S. More-
over, for u, v ∈ S we have
u ∗ v = v ∗ u ∈ S, (9.4.23)
F[u ∗ v] = (2π)N/2 F[u]F[v], (9.4.24)
F−1 [uv] = (2π)−N/2 F−1 [u] ∗ F−1 [v]. (9.4.25)
Proof. From Theorem 5.2.5 we have u ∗ v ∈ L1 ∩ C ∞ , and for α, β ∈ NN0 we get

|x D (u ∗ v)(x)| =
α β
x D u(x − y)v(y)dy
α β
N
R

= (x − y + y) D u(x − y)v(y)dy
α β
RN

≤ C(α) |(x − y)γ Dβ u(x − y)||y σ v(y)|dy
|σ|,|γ|≤|α| RN
(as in Thm. 5.4.8) ≤ C(α, N )s|α|+|β| (u)s|α|+2N (v),
which proves (9.4.23) and the continuity of ∗ in S × S.

For (9.4.24), we use Fubini’s theorem as follows:

1 −i(ξ·x)
F[u ∗ v](ξ) = e u(x − y)v(y)dydx
(2π)N/2 RN RN
191

1
= v(y) e−i(ξ·x) u(x − y)dxdy
(2π)N/2 RN R N

1 −i(ξ·y)
= e v(y) e−i(ξ·(x−y)) u(x − y)dxdy
(2π)N/2 RN RN
= (2π)N/2 F[u](ξ)F[v](ξ),
which proves (9.4.24). Applying F−1 to (9.4.24) gives
F−1 [ûv̂] = (2π)−N/2 u ∗ v,
which together with the fact F[S] = S proves (9.4.25).
For the convolution in S , we avoid the technical details and will only give the formal
definition of the convolution and some properties. For the details of the proofs, the reader
can consult [12, 18, 22, 23].
Definition 9.4.18 Let T ∈ S and S ∈ E , where E is the set of linear maps from
C ∞ (RN ) to C continuous for the convergence in C ∞ (RN ).8 We define the convolution
T ∗ S : S → C by
T ∗ S, ϕ = T (x), S(y), ϕ(x + y), ∀ϕ ∈ S. (9.4.26)
Theorem 9.4.19 Let T ∈ S and S ∈ E . Then

T ∗ S = S ∗ T ∈ S, (9.4.27)
Dα [T ∗ S] = Dα T ∗ S = T ∗ Dα S, (9.4.28)
F[T ∗ S] = F[S ∗ T ] = (2π)N/2 F[T ]F[S]. (9.4.29)
8
A sequence (ϕn ) in C ∞ (RN ) converges to 0 in C ∞ (RN ) if for every compact K ⊂ RN and for every
α ∈ NN α
0 , limn→∞ D ϕn C 0 (K) = 0.
192
9.5 Sobolev spaces (Ch. 6)

9.5.1 Continuous and compact embeddings

Theorem 9.5.1 Assume s − N2 ∈ (m, m + 1], m ∈ N0 and let σ ∈ 0, s − N2 − m . Then
H s −→ Cbm,σ ∩ {u, lim Dα u = 0, ∀α ∈ NN

0 , |α| ≤ m}. (9.5.1)
|x|→∞
In particular, there exists C > 0 such that for all α ∈ NN

0 with |α| = m and x, y ∈ R
N
we have
|Dα u(x) − Dα u(y)| ≤ CDα uH s−m |x − y|σ . (9.5.2)
Proof. Let us consider first the case m = 0. From Theorem 6.3.3 we have H s −→ Cb0 .
To complete the proof in this case, it is enough to prove (9.5.2). It is easy to point out
that for σ as in theorem and arbitrary x, y, ξ, η ∈ RN , we have9
1
d iξ·(y+t(x−y))
|ei(ξ·x)
−ei(ξ·y)
|= e dt ≤ 2|ξ|σ |x − y|σ . (9.5.3)
0 dt
It follows that

1
|u(x) − u(y)| ≤ |ei(x·ξ) − ei(y·ξ) ||û|dξ
(2π)N/2 RN

2
≤ |x − y| σ
|ξ|σ |û|dξ (9.5.4)
(2π)N/2 N
R
2
≤ |x − y|σ |ξ|σ ξ−s ξs |û|dξ
(2π)N/2 RN
2
≤ |x − y|σ |ξ|σ ξ−s L2 ξs ûL2
(2π)N/2
= C|x − y|σ uH s ,
with C = 2
(2π)N/2
|ξ|σ ξ−s L2 . Note that we have (as in Theorem 6.3.3)
∞
|ξ| σ
ξ−s 2L2 ∼ 1+ r N +2(σ−s)−1
dr < ∞,
1
which proves the theorem in the case m = 0.

The general case m > 0 follows from the case m = 0 by applying it to Dα u, with
|α| ≤ m, and noting that Dα u ∈ H s−|α| (see Example 6.2.4) and (s − |α|) − N2 > 0.
Theorem 9.5.2 Let k ∈ N and assume Ω is an open bounded set satisfying the H k+1 (Ω)-
extension property. Then H k+1 (Ω) −→
c
H k (Ω).
9
Actually, this inequality holds for every σ ∈ (0, 1).
193
Proof. Let (un ) be a bounded sequence in H k+1 (Ω), i.e. un H k+1 (Ω) ≤ C1 for all n. From
Theorem 6.3.7, there exists a sequence Un = Eun ∈ H k+1 (RN ), with Un H k+1 (RN ) ≤
C2 := C1 E and E an H k+1 extension operator. Without loss of generality we may
assume that supp(Un ) ⊂ K, K compact, because we can multiply (Un ) with η(Rx), η a
{B(0, 1), B(0, 2)} cut-off function, and R >> 1.
The sequence (Ûn ), Ûn = F[Un ], is uniformly bounded and uniformly Lipschitz in
RN . Indeed, as Un are with compact support we have Un ∈ L1 (RN ), and from Lemma
N
9.4.14 obtain Ûn ∈ C 0 (R ) and

1 1
|Ûn (ξ)| ≤ N/2
|Un |dx ≤ C |K|1/2 =: C3 .
N/2 2
(9.5.5)
(2π) K (2π)
Furthermore, Ûn satisfies

1
|Ûn (ξ) − Ûn (η)| = (e−i(ξ·x) − e−i(η·x) )Un (x)dx
(2π)N/2 K

|ξ − η|
≤ |x||e−i(η+θ(ξ−η))·x ||Un (x)|dx (θ ∈ (0, 1))
(2π)N/2 K
≤ C4 |ξ − η|, (9.5.6)
1
with C4 = C2 xL2 (K) .
(2π)N/2
Therefore, (Ûn ) satisfies the conditions of Arzela-Ascoli in every ball B R = B(0, R),
R > 0. Then there exists a subsequence of (Ûn ), still denoted by (Ûn ) converging to a
certain ÛR in C 0 (B R ), for every R > 0. It follows that (Un ) is a Cauchy sequence in
H k (RN ). Indeed, first note that
un − un+m 2H k (Ω) ≤ Un − Un+m 2H k (RN ) (9.5.7)

= ξ |Ûn − Ûn+m | dξ +
2k 2
ξ2k |Ûn − Ûn+m |2 dξ
BR c
BR
≤ (1 + R2 )k Ûn − Ûn+m 2C 0 (B R )

ξ2k
+ ξ2(k+1) |Ûn − Ûn+m |2 dξ
c ξ
2(k+1)
BR
C5
≤ (1 + R2 )k Ûn − Ûn+m 2C 0 (B R ) + 2
Ûn − Ûn+m 2H k+1 (RN )
1 + R
1
≤ C6 (1 + R ) Ûn − Ûn+m C 0 (B R ) +
2 k 2
. (9.5.8)
1 + R2
C6
Next, given > 0 we choose R such that 2
< . As (Ûn ) is Cauchy in C 0 (B R ), there
1+R 2

exists n such that for all N > n and m ∈ N we have C6 (1 + R2 )k Ûn − Ûn+m 2C 0 (B R ) < ,
2
which shows that (un ) is a Cauchy sequence in H k (Ω), so (un ) converges in H k (Ω).
194
We also have the following embedding, which provides an example of how Fourier
transform can be successfully used for proving results in a context different from L2 or
H s spaces.
N N
Theorem 9.5.3 Let s > − with p > 2. Then H s −→ Lp .
2 p
Proof. Let u ∈ S, û = F[u], and q ∈ (1, 2) the conjugate number of p, p1 + 1q = 1.
From the continuity of F from Lq to Lp , see Theorem 9.4.15, and the fact that u =
it follows uLp = û(−ξ)
F[û(−ξ)] = û(−ξ), Lp ≤ CûLq .
If ûLq ≤ CuH s then uLp ≤ CuH s , and from S −→ d
H s it follows H s → Lp .
So, to prove, the theorem reduces to proving ûLq ≤ CuH s . From Hölder inequality
with α = 2/q and β = 1/(1 − q/2), we get

|û| dξ =
q
ξsq |û|q · ξ−sq dξ
RN RN
q/2 1−q/2
−sq
≤ ξ |û| dξ
2s 2
ξ 1−q/2 dξ
RN RN
12 p−1
p−2
2sp
− p−2
= uqH s ξ dξ .
RN
2sp
But RN ξ− p−2 dξ < ∞ iff N −2 p−2
sp
< 0, which is equivalent to s
N
> 12 − p1 and completes
the proof.
As an application of the compactness result in Theorem 9.5.2 and of Lemma 9.5.5

(which is a Functional Analysis result), we have the following result which is useful when
studying PDEs of high order.
Corollary 9.5.4 Let Ω ⊂ RN be an open bounded set satisfying H k (Ω) →
c
H k−1 (Ω) and
a, b > 0. Then the following norms in H k (Ω) are equivalent

uH k (Ω) = Dα uL2 (Ω) , uH k (Ω) = auL1 (Ω) + b|u|H k (Ω) . (9.5.9)
|α|≤k
Proof. As Ω is bounded, by using Hölder inequality it is easy to prove that uH k (Ω) ≤
C1 uH k (Ω) , with a certain C1 > 0. It remains to prove the inverse inequality, i.e.
uH k (Ω) ≤ C2 uH k (Ω) , with C2 > 0.
Set X = H k (Ω), Y = H k−1 (Ω), and Z = L1 (Ω). Then X, Y , and Z fulfill the
conditions of Lemma 9.5.5. From (9.5.11), for every ∈ (0, 1) there exists C() > 0 such
that
uH k−1 (Ω) ≤ uH k (Ω) + C()uL1 (Ω) .
195
This implies
1 C()
uH k (Ω) ≤ |u|H k (Ω) + uL1 (Ω) ≤ C(, a, b)uH k (Ω) ,
1− 1−
Lemma 9.5.5 (Ehrling) Let X, Y , and Z be Banach spaces and assume

c
X → Y → Z. (9.5.10)
Then for every > 0 there exists C() > 0 such that
∀x ∈ X, xY ≤ xX + C()xZ . (9.5.11)
Proof. Assume the lemma fails for a certain 0 > 0. Then, there exists a sequence xn
in X such that xn Y > 0 xn X + nxn Z . We can assume, without restriction, that
xn X = 1 (otherwise take xn /xn X instead of xn ). Then
xn Y > 0 + nxn Z , xn X = 1. (9.5.12)

c
Since X → Y , after passing to a subsequence we may assume that limn→∞ xn = y0
in Y and so xn Y is bounded. Therefore from (9.5.12), we obtain limn→∞ xn = 0 in
Z. From (9.5.10) we have xn − y0 Z ≤ Cxn − y0 Y , so limn→∞ xn = y0 in Z. The
uniqueness of the limit (of xn in Z) implies y0 = 0, which contradicts (9.5.12) for n
large, and so proves the lemma.
9.5.2 Extension and density results in Sobolev spaces

Density and extension results hold in W s,p (Ω) spaces, but for the sake of simplicity
we will focus our analysis only on W k,p (Ω) spaces. Let us start with two simple results,
which are practical in many situations.
Lemma 9.5.6 (Extension by zero) Let u ∈ W0k,p (Ω), p ∈ [1, ∞), and E0 u = U :
RN → R, the extension of u in RN by zero, i.e. U (x) = u(x) if x ∈ Ω, U (x) = 0 if
x∈/ Ω. Then E0 ∈ L(W0k,p (Ω); W k,p (RN )).
Proof. By the definition of W0k,p (Ω), there exists (un ) in D(Ω) with limn→∞ un = u
in W k,p (Ω). Set Un = E0 un . Then Un ∈ D(RN ) and limn→∞ Un = U in W k,p (RN ), so
U ∈ W k,p (RN ) and U W k,p (RN ) = uW k,p (Ω) .
0
Lemma 9.5.7 (Truncation and extension by zero) Let u ∈ W k,p (Ω), p ∈ [1, ∞],
η ∈ D(Ω). Then V = E0 (ηu) ∈ W k,p (RN ).
196
Proof. Indeed, for ϕ ∈ D(RN ) and for any α ∈ NN 0 , |α| ≤ k, by using (6.1.10) we have

|α| |α|
D V, ϕ = (−1)
α α
V D ϕ = (−1) α
uηD ϕ = Dα (uη)ϕ
RN Ω Ω

= ϕ Cβα Dβ uDα−β η
Ω 0≤β≤α

= ϕE0 Cβα Dβ uDα−β η .
RN 0≤β≤α

α
So D V = E0 α β
0≤β≤α Cβ D uD
α−β
η ∈ Lp (RN ), which proves V ∈ W k,p (RN ).
In the case when Ω is smooth, a function in W k,p (Ω) can be approximated by func-
tions smooth up to the boundary ∂Ω. The following theorem is instructive, because it
highlights the key steps of the method, in particular the step of regularization by convo-
lution, which furthermore shifts the support of the function. The general case is treated
by using a partition of unity; see Theorem 9.5.10.
N N
Theorem 9.5.8 Let k ∈ N, p ∈ [1, ∞). Then D(R+ ) −→
d
+ ), where D(R+ ) :=
W k,p (RN
{U |RN+ , U ∈ D(R )}.
N
Proof. Let u ∈ W k,p (RN + ). Without loss of generality, we may assume that supp(u) is
bounded. If not, we can always consider un = ηn u, where ηn (x) = η(x/n) and η is a
{B(0, 1), B(0, 2)} cut-off function, and show easily that un converges to u in W k,p (RN +)
similarly as in Theorem 6.1.7.
Now, let (ρn ) be a mollifier sequence with supp(ρn ) ⊂ RN − := {xN < 0}, and consider
the sequence (ρ̌n ), where in general ϕ̌(x) = ϕ(−x). Set U = E0 (u), where E0 is as in
Lemma 9.5.6, Un = U ∗ ρn .
We will show that (un ) = (Un |RN+ ) is the required sequence. Indeed, as u has bounded
support the U also has bounded support, Un ∈ D and furthermore limn→∞ Un = U in
Lp (RN ), so limn→∞ un = u in Lp (RN + ). Next we compute D un , |α| ≤ k.
α
For ϕ ∈ D(R ) ∩ {supp(ϕ) ⊂ R+ }, we have

N N

|α|
(−1) D un , ϕD (RN+ )×D(RN+ ) =
α
Un (x)Dα ϕ(x)dx
RN

= U (y)ρn (x − y)Dα ϕ(x)dydx
N N
R R
= U (y)(ρ̌n ∗ Dα ϕ)(y)dy
N
R
= U (y)Dα (ρ̌n ∗ ϕ)(y)dy
N
R
= u(y)Dα (ρ̌n ∗ ϕ)(y)dy, (9.5.13)
RN
+
197
where the last equality holds for n large, because for such n we have
supp(ρ̌n ) ⊂ RN
+, supp(ρ̌n ∗ ϕ) ⊂ sup(ρ̌n ) + supp(ϕ) ⊂ RN
+.
So Dα (ρ̌n ∗ϕ) ∈ D(RN + ) and then we can integrate by parts in (9.5.13) with no boundary
terms, which gives

D un , ϕD (RN+ )×D(RN+ ) =
α
Dα u(y)(ρ̌n ∗ ϕ)(y)dy
RN
+
= E0 (Dα u)(y)(ρ̌n ∗ ϕ)(y)dy

RN

= E0 (Dα u)(y)ρn (x − y)ϕ(x)dxdy
N N
R R
= E0 (Dα u)(y)ρn (x − y)ϕ(x)dydx
N RN
R
= (ρn ∗ E0 (Dα u))(x)ϕ(x)dx
RN
= ρn ∗ E0 (Dα u), ϕD (RN+ )×D(RN+ ) . (9.5.14)
From (9.5.14), we get Dα un = ρn ∗ E0 (Dα u) and therefore limn→∞ Dα un = Dα u in

Lp (RN
+ ), which completes the proof.
The classical construction of W k,p extension operator for p ∈ [1, ∞] is quite long and
technical. We will use Theorem 9.5.8 to construct a W k,p extension operator, p ≥ 1,
with a simple proof.
Theorem 9.5.9 Let k ∈ N, p ∈ [1, ∞). Then there exists a linear continuous extension
+ ) → W
operator10 E : W k,p (RN k,p
(RN ).
N
Proof. Let E : D(R+ ) → C0k (RN ) be defined by
⎧

⎪
⎨ u(x , xN ), xN ≥ 0,
x
(Eu)(x , xN ) := U (x , xN ) = c j u x , −
N
, xN < 0, (9.5.15)
⎪
⎩ j+1
j=0,...,k
where x = (x , xN ), x = (x1 , . . . , xN −1 ), and u ∈ D(R+ ). The coefficients cj are chosen

N
so that U ∈ C0k (RN ). For this, it is sufficient to ensure the continuity across ∂RN + of the

derivatives ∂xN U , 0 ≤ i ≤ k. Evaluating these derivatives at (x , 0) leads to
i
1
i
− cj = 1, i = 0, 1, . . . , k. (9.5.16)
j=0,...,k
j+1
10
See Theorem 6.3.7 for the general extension result.
198
The matrix of this system is the so-called “Vandermonde matrix”, which is nonsingular
and so the coefficients ci are uniquely defined.
E provides the required extension operator. Indeed, clearly E is linear. Furthermore,
there exists C > 0 such that
N
EuW k,p (RN ) ≤ CuW k,p (RN+ ) , ∀u ∈ D(R+ ). (9.5.17)
N
This estimate is proven easily by using simple classical calculus. Therefore as D(R+ ) is
dense in W k,p (RN
+ ), (9.5.17) implies that E extends to a linear continuous operator from
W k,p (RN
+ ) to W k,p
(RN ), still denoted by E, which satisfies (9.5.17) in W k,p (RN+ ). Finally,
Eu = u a.e. in R+ for every u ∈ W (R+ ), because we can choose a sequence (un ) in
N k,p N
N
D(R+ ) with Eun = un in RN + , converging to u in W
k,p
+ ) and pointwise in R+ .
(RN N
Theorem 9.5.10 Assume Ω ⊂ RN is an open bounded set of class C k , k ∈ N, p ∈ [1, ∞).

Then there exists an extension operator11 E = EΩ : W k,p (Ω) → W k,p (RN ).
Proof. Let {(Gi , θi ), i = 1, . . . , M } be as in Definition 1.2.4, corresponding to some

points {x1 , . . . , xM } on ∂Ω such that {Gi , i = 1, . . . , M } forms a finite covering of ∂Ω.
Let furthermore {ηi , i = 0, 1, . . . , M } be a partition of unity subordinate to {Gi , i =
1, . . . , M }; see Fig. 6.4.1 and Theorem 1.2.2.
For u ∈ W k,p (Ω) we have u = u0 + u1 + · · · + uM , with ui = uηi , i = 0, 1, . . . , M .
Then ui ∈ W k,p (Ω), i = 0, 1, . . . , M ; see Proposition 6.1.11. We can extend u0 by zero,
namely let W0 = E0 u0 . As supp(u0 ) Ω, from Lemma 9.5.7 we get W0 ∈ W k,p (RN ).
We can extend ui , i = 1, . . . , M , to a W k,p (RN ) function as follows. First define vi
by vi = ui ◦ θi in Q+ . From Proposition 6.1.13 and the remark following it, we have
vi ∈ W k,p (Q+ ). Note that from supp(ηi ) Gi ∩ Ω, we get supp(vi ) ⊂ Qi ∩ Q+ , with a
certain Qi Q. It is easy to prove that if ṽi is the extension by zero of vi in RN+ , then
ṽi ∈ W k,p (RN+ ). Next, let Ṽi ∈ W k,p (RN ) be the extension of ṽi in W k,p (RN ) as given by
Theorem 9.5.9, and set wi = Ṽi |Q ◦ θi−1 and Wi = E0 wi . Clearly wi ∈ W k,p (RN + ), see
Proposition 6.1.13, and as supp(wi ) Gi , we have Wi ∈ W (R ); see Lemma 9.5.7.
k,p N
Finally, set Eu := W0 + W1 + · · · + WM . Clearly we have
Eu ∈ W k,p (RN ),
(Ewi )(x) = wi (x) = u(x)ηi (x), x ∈ Ω, i = 1, . . . , M,

(Eu)(x) = u(x) ηi (x) = u(x), x ∈ Ω,
i=0,...,M
so Eu is a W k,p (RN ) extension of u.

Furthermore, E is linear and bounded in W k,p (Ω), because the maps u → ui → vi →

ṽi → Ṽi → wi → Wi are linear and continuous in the appropriate W k,p spaces.
11
See Theorem 6.3.7 for the general extension result.
199
Theorem 9.5.11 Assume Ω ⊂ RN satisfies the W k,p extension property, p ∈ [1, ∞).
Then D(Ω) −→
d
W k,p (Ω), where D(Ω) := {U |Ω , U ∈ D(RN )}.
Proof. Let u ∈ W k,p (Ω) and U = Eu its W k,p (RN ) extension. From Theorem 6.1.7,
there exists a sequence Un ∈ D(RN ) converging to U in W k,p (RN ). Set un = U |Ω . Then
(un ) is in D(Ω) and converges to u in W k,p (Ω).
9.5.3 Boundary traces in Sobolev spaces

The following result shows that the trace operator defined in Theorem 6.4.2 is onto.
We give an explicit formula for the right inverse γ −1 of γ.
1
Theorem 9.5.12 Let s > 12 . There exists a bounded linear map γ −1 : H s− 2 (RN −1 ) →
1
H s (RN ) such that γ ◦ γ −1 is the identity map in H s− 2 (RN −1 ).
Proof. We use the same notations as in Theorem 6.4.2. Let u ∈ S(RN −1 ) and set

−1 (2π)1/2 −1 ξ 2(s−1/2)
U (x) = (γ (u))(x) := F û(ξ ) , ξ = (ξ , ξN ),
Is ξ2s
where Is is defined in Theorem 6.4.2. Note that from the definition of U , we have
ξ 2(s−1/2) (2π)1/2
Û (ξ) = C1 û(ξ ), C1 = .
ξ2s Is
The map γ −1 is well-defined and continuous from H s (RN −1 ) to H s (RN ) because

ξ 4(s−1/2)
U H s (RN ) =
2
ξ |Û | dξ = C1
2s 2 2
ξ2s |û(ξ )|2 dξ
RN RN ξ 4s

1
(use (6.4.2)) = C12 ξ 4(s−1/2) |û(ξ )|2 dξN dξ
R ξ
2s
RN −1
≤ C2 uH s−1/2 (RN −1 ) .
2
Finally, we have U (x , 0) = u(x ) because from (6.4.2) we get

1 i(x ·ξ ) 2(s−1/2) 1
U (x , 0) = e û(ξ )ξ dξN dξ
R ξ
(2π) (N −1)/2 Is RN −1 2s

1
= (N −1)/2
ei(x ·ξ) û(ξ )dξ = u(x ),
(2π) RN −1
which together with the inclusion S(RN −1 ) → H k−1 (RN −1 ) completes the proof of the-
orem.
The following theorem shows that in the case when Ω is a bounded Lipschitz domain,
the W01,p (Ω) functions are characterized as W 1,p (Ω) functions with zero boundary trace.
The proof demonstrates typical techniques when working with Sobolev spaces, such
as truncation, extension, regularization, and shifting/translation of the support of the
function.
200
Corollary 9.5.13 Let Ω ⊂ RN be an open bounded Lipschitz set and p ∈ [1, ∞). Then
W01,p (Ω) = N 1,p (γ), where N 1,p (γ) = {u ∈ W 1,p (Ω), γ(u) = 0}. (9.5.18)
Proof. We will prove the corollary only in the case Ω = RN + . The proof of the general
case is made by transforming the problem to a finite number of problems in RN + by using
a partition of unity of ∂Ω.
It is easy to prove W01,p (Ω) ⊂ N 1,p (γ). Indeed, if u ∈ D(Ω) then clearly u ∈ N 1,p (γ).
As D(Ω) is dense in W01,p (Ω) and γ is continuous, it follows that W01,p (Ω) ⊂ N 1,p (γ).
Now, let u ∈ N 1,p (γ) and prove that u ∈ W01,p (Ω). The proof contains several steps:
first a truncation, then an extension by zero and finally an appropriate regularization
(which regularizes the function and translates/shifts its support). WLOG we may assume
that U has bounded support, as by “truncation” (see Theorem 6.1.7) we can always
consider a sequence (un ) with bounded support converging to u in W 1,p (RN + ). Next,
we denote by U the extension of u in RN by zero. Then U ∈ W 1,p (RN ). Indeed, for
ϕ ∈ D(RN ) and |α| = 1, we have

D U, ϕ = −
α
U D ϕdx = −
α
uDα ϕdx.
RN RN
+
There exists an open set ω ∈ RN + of class C such that supp(ϕ) ∩ R+ ⊂ ω. Then, from
1 N
(6.4.8), for |α| = 1 we get

α α
uD ϕdx = uD ϕdx = γ(u)ϕν dσ + Dα uϕdx
α
RN ω ∂ω ω
+

= Dα uϕdx.
RN
+
Hence, Dα U = E0 (Dα u) ∈ Lp (RN ), so U ∈ W 1,p (RN ).

Next, let (ρn ) be a mollifier sequence with supp(ρn ) ⊂ B (0, . . . , 0, n1 ), 2n
1
, and
∞
consider Un = U ∗ρn . Clearly Un ∈ C (R )∩W (R ) and limn→∞ Un =
N 1,p N 1,p
' U in W1 ( (R ).
N
Furthermore, from (5.2.3) we have supp(Un ) ⊂ supp(U ) + supp(ρn ) ⊂ xN ≥ 2n . This

proves that un = Un |RN+ ∈ D(RN+ ) and limn→∞ un = u in W
1,p
+ ), so u ∈ W0 (Ω).
(RN 1,p
201
9.6 Second-order linear elliptic PDEs: weak

solutions
9.6.1 Regularity of weak solutions
The method of finding the weak solutions as described in section 7.2 is an elegant
way to prove the existence of weak solutions. For example, the problem
−Δu + u = f in Ω, u ∈ H01 (Ω), f ∈ H −1 (Ω) (9.6.1)
has always a unique weak solution u ∈ H01 (Ω).
If f is more regular than just in H −1 (Ω) and Ω is more regular than just an open set,
then one expects that (9.6.1) (or even (7.1.1)) has a weak solution more regular than
H01 (Ω). For example, if Ω = RN and f ∈ L2 (RN ), by taking Fourier transform of both
sides of (9.6.1) we obtain
) *
fˆ fˆ
(1 + ξ 2 )û = fˆ, û = , u = F−1 .
1 + ξ2 1 + ξ2
fˆ
Then u ∈ H 2 (RN ) because ∈ L22 (RN ). By analogy, if f ∈ L2 (Ω) then one expects
1 + ξ2
that u ∈ H01 (Ω) ∩ H 2 (Ω). We will see that this is true provided Ω is regular.
In this section, we will address the “L2 -regularity theory” of the weak solutions to
(9.6.1). The results remain the same even for more general second-order elliptic linear
PDEs like (7.1.1). As the method of proof is the same, we will focus the analysis on
(9.6.1), so as to avoid the technicalities of the general case and highlighting the main
ideas.
There is also a “Lp -regularity theory” of the weak solutions to second-order elliptic
linear PDEs, but the analysis is more difficult. We refer the interested reader to [2, 20, 38]
for more in this topic. The result that we will prove in this section is the following.
Theorem 9.6.1 Let Ω ⊂ RN be an open bounded set and u ∈ W01,p (Ω), p ∈ (1, ∞), be a
weak solution to (9.6.1). If Ω is of class C m+2 and f ∈ W m,p (Ω), m = 0, 1, . . ., then
u ∈ W m+2,p (Ω) and uW m+2,p (Ω) ≤ Cf W m,p (Ω) , (9.6.2)
with C independent of f (C depends only on Ω).
The proof of this theorem in the case p = 2 is more difficult, and we refer the interested
reader to the pioneering works on this subject [2, 20, 38]. We will prove the theorem in
the case p = 2; see, for example, [5].
Note that with f ∈ L2 (Ω), the weak solution u ∈ H01 (Ω) to (9.6.1) satisfies
uH 1 (Ω) ≤ f L2 (Ω) . (9.6.3)
202
The regularity analysis of u will be divided in two parts. First is the regularity in the
interior of Ω, which depends only on the data f and not Ω. Next is the regularity near
the boundary ∂Ω, which depends on the regularity of the boundary as well. To this end,
let G1 , . . . , Gn , n ∈ N, be a finite covering with open sets of ∂Ω. From Theorem 1.2.2,
there exist η0 , η1 , . . . , ηn in D(RN ) and with values in [0, 1] such that12
n
ηi = 1 in RN ,
i=0
supp(ηi ) ⊂ Gi for all i = 1, . . . , n,
supp(η0 ) ∩ Σ = ∅ where Σ := {x ∈ RN , dist(x, ∂Ω) < },
for a certain > 0. Then we write

n
n
u = (η0 u) + (ηi u) =: u0 + ui . (9.6.4)
i=1 i=1
First, we will prove the regularity of u in the interior of Ω, namely for u0 . The idea is that
as supp(u0 ) Ω, we can use in (7.2.12) as test functions the so-called “difference quo-
tients” Dih u0 , which upon sufficient regularity of f allows one to estimate Dih u0 H m (Ω)
norms, m = 1, 2, . . .. These estimates imply H m+1 (Ω) regularity of u0 . The regularity of
u near the boundary is obtained by first making a change of variable which “flattens”
the boundary. This allows us to use difference quotients again, which provide similar
estimates for ui functions as for u0 , and ultimately the regularity of u in Ω.
9.6.1.1 Regularity in the interior

We start with some preliminary results. Let Ω ⊂ RN be an open set and {e1 , . . . , eN }
be the orthogonal standard basis of RN . For i = 1, . . . , N and given h ∈ R we define
Dih u, the difference quotient of u in direction ei , by
u(x + hei ) − u(x)

Dih u(x) = , x ∈ Ω ∩ {dist(x, ∂Ω) > |h|}. (9.6.5)
h
Note that by extending u in RN by zero allows to define Dih u in RN .
Lemma 9.6.2 Assume ω Ω, with ω open, and |h| < dist(ω, ∂Ω). Then for every
u ∈ Lp (Ω), p ∈ [1, ∞), and every v ∈ C 0 (Ω) bounded with supp(v) ⊂ ω, we have
Dih (uv)(x) = u(x)Dih v(x) + v(x + hei )Dih u(x), x ∈ ω, (9.6.6)

h
Di u(x)v(x) = u(x)Di−h v(x)dx. (9.6.7)
Ω Ω
12
Given A, B ⊂ RN , dist(A, B) = inf{|a − b|, a ∈ A, b ∈ B}.
203
Proof. Indeed, we get easily
Dih (uv)(x) = u(x)Dih v(x) + v(x + hei )Dih u(x),
and by changing the variable we obtain

h u(x + hei ) − u(x) v(x − hei ) − v(x)
Di u(x)v(x)dx = v(x)dx = u(x) dx,
Ω Ω h Ω h
Lemma 9.6.3 Let u ∈ W 1,p (Ω), p ∈ [1, ∞). Then
Dih uLp (ω) ≤ ∂i uLp (Ω) , ∀ω Ω, ω open, |h| < dist(ω, ∂Ω). (9.6.8)
Proof. Let u ∈ C ∞ (Ω) ∩ W 1,p (Ω). For t ∈ R set v(t) = u(x + thei ). Then v ∈ C ∞ (R),
v (t) = hei · ∇u(x + thei ), and

v(1) − v(0) 1 1

1
Dih u(x) = = v (t)dt = ei · ∇u(x + thei )dt.
h h 0 0
Therefore, it follows
1 p

Dih upLp (ω) = |e · ∇u(x + the )dt dx
i i
ω 0
1 1
≤ |∇u(x + thei )| dxdt ≤
p
|∇u(y)|p dydt
0 ω 0 Ω
= ∇upLp (Ω) ,
because from ω Ω, |h| < d(ω, ∂Ω) we get y = x + thei ∈ Ω. This inequality remains
true for all u ∈ W 1,p (Ω); see Theorem 6.1.8.
Lemma 9.6.4 Let p ∈ (1, ∞) and u ∈ Lp (Ω). Assume that there exists C > 0 such that
Dih uLp (ω) ≤ C, ∀ω Ω, ω open, |h| < dist(ω, ∂Ω), i = 1, . . . , N. (9.6.9)
Then u ∈ W 1,p (Ω) and
∂i uLp (Ω) ≤ C. (9.6.10)
Proof. Let ϕ ∈ D(Ω). Then

∂i u, ϕ = − u(x)∂i ϕ(x)dx = − lim u(x)Dih ϕ(x)dx = lim Dih u, ϕ, (9.6.11)
Ω h→0 Ω h→0
204
because ϕ ∈ D(Ω). As Dih uLp (ω) is bounded, there exists a sequence (h) tending to
zero and v ∈ Lp (ω), such that limh→0 Dih u = v weakly in Lp (ω), see Theorem 1.3.8, i.e.

lim Di u, ϕ =
h
vϕ, ∀ϕ ∈ D(ω). (9.6.12)
h→0 Ω
From the arbitrariness of ω, we have v ∈ Lp (Ω). Moreover, from (9.6.11) and (9.6.12) it
follows

∂i u, ϕ = vϕdx, ∀ϕ ∈ D(Ω),
Ω
which proves that ∂i u ∈ Lp (Ω). Finally, from (9.6.11) we obtain

∂i uϕdx ≤ Cϕ p , 1 + 1 = 1.
L (Ω)
p p
Ω

As D(Ω) is dense in Lp (Ω), we can take ϕ = ∂i u|∂i u|p−2 ∈ Lp (Ω) and then it follows
1/p p/p
Ω
|∂ i u|p
dx ≤ C Ω
|∂ i u|p
dx , so ∂i upLp (Ω) ≤ C∂uLp (Ω) and this proves (9.6.10).
We note that the advantage of using Dih u, i = 1, . . . , N , is that they represent a

constructive method for proving further regularity of u. For example, Lemma 9.6.4
asserts H 1 regularity for u only from a condition for Dih u. We will use this technique in
the following two theorems.
Theorem 9.6.5 Let Ω ⊂ RN be an open bounded set and assume u ∈ H01 (Ω) is a weak
solution to (9.6.1). If f ∈ H m (Ω) then u0 defined in (9.6.4) satisfies
u0 ∈ H m+2 (Ω), u0 H m+2 (Ω) ≤ Cf H m (Ω) , (9.6.13)
with C independent of f (C depends only on Ω).
Proof. Let us first deal with the case m = 0, so f ∈ L2 (Ω). Denote still by u0 the
extension of u0 in RN by zero. So u0 ∈ H 1 (RN ), see Lemma 9.5.7, and from (9.6.4) and
(9.6.1) it follows that u0 is a weak solution in RN to
−Δu0 + u0 = g := η0 f − 2(∇η0 · ∇u) − (Δη0 )u. (9.6.14)
Note that g ∈ L2 (RN ). For ϕ ∈ H 1 (RN ), we have

∇u0 · ∇ϕ + u0 ϕ = gϕ. (9.6.15)
RN RN
Let i ∈ {1, . . . , N }, h > 0, and take ϕ = D−h Dh u0 := Di−h Dih u0 . Note that ϕ ∈ H 1 (RN )
as a linear combination of H01 (RN ) functions. From (9.6.7), the fact that ∂j and Dh
commute, and (9.6.15), we obtain

|∇D u0 | +
h 2
|D u0 | =
h 2
gD−h Dh u0 .
RN RN RN
205
This gives Dh u0 2H 1 (RN ) ≤ gL2 (RN ) D−h Dh u0 L2 (RN ) . But from Lemma 9.6.3, we have
D−h (Dh u0 )L2 (RN ) ≤ Dh u0 H 1 (RN ) . Combining the two last inequalities gives
Dh u0 H 1 (RN ) ≤ gL2 (RN ) . (9.6.16)
Thanks to Lemma 9.6.4, this implies u0 ∈ H 2 (RN ), so u0 ∈ H 2 (Ω). Furthermore, from

(9.6.16), (9.6.14), and (9.6.3), we obtain
∂i u0 H 1 (Ω) ≤ gL2 (RN ) ≤ f L2 (Ω) + 2η0 H 2 (Ω) uH 1 (Ω)
≤ (1 + 2η0 H 2 (Ω) )f L2 (Ω) ,
which proves the theorem in the case m = 0.

For m ≥ 1, we proceed by recurrence w.r.t. m. So, assume that u satisfies Theorem
9.6.5 up to m − 1 and let us prove it for m. This means that u0 ∈ H m+1 (RN ), g ∈
H m (Ω), and to prove is u0 ∈ H m+2 (RN ). Let α ∈ NN
0 , |α| = m. Then D u0 ∈ H (R ),
α 1 N
Dα g ∈ L2 (RN ), and
−Δ(Dα u0 ) + Dα u0 = Dα g, weakly, i.e.

∇(D u0 ) · ∇ϕ + (D u0 )ϕ =
α α
Dα gϕ, ∀ϕ ∈ H 1 (RN ).
RN RN
This is obtained from (9.6.15) by taking Dα ϕ as a test function and integrating by

parts. Then as in the case m = 0, we obtain Dα u0 ∈ H 2 (RN ), so u0 ∈ H m+2 (Ω), and
the estimation (9.6.13) follows.
9.6.1.2 Regularity near the boundary

Now we will prove that ui := ηi u ∈ H m+2 (Ω), for all i = 1, . . . , N . The proof strongly
uses the regularity of Ω and of f . Note that ui ∈ H01 (G+
i ), where Gi = Gi ∩ Ω. Indeed, if
+
ϕk ∈ D(Ω) such that limk→∞ ϕk = ui in H 1 (Ω), then ηi ϕk ∈ D(G+ i ) and limk→∞ (ϕk ηi ) =
ui in H 1 (G+
i ). Furthermore, u i satisfies
−Δui = gi := ηi f − ηi u − 2(∇ηi · ∇u) − (Δηi )u in G+

i . (9.6.17)
From (9.6.17) and (9.6.3), we have gi ∈ L2 (G+

i ) and

(∇ui · ∇ψ)dx = gi ψdx, ∀ψ ∈ H01 (G+i ), (9.6.18)
G+
i G+
i
ui H 1 (G+i ) ≤ Cf L2 (Ω) , (9.6.19)
with C depending only on Ω.

We will prove the regularity of ui by using the method of translations. However, as
−h h
the equation (9.6.18) is in G+i , we cannot take ϕ = Di Di ui , because this ϕ may not
be in H01 (G+ +
i ). For this purpose, we will change the variable and write (9.6.18) in Q ,
see section 1.2, which allows the use of the method of translations.
206
More precisely, as Ω is of class C m+2 there exists θi ∈ C m+2 (Q; Gi ), invertible with
inverse ζi ∈ C m+2 (Gi ; Q) such that (see Fig 6.4.1)
θi (Q+ ) = G+
i , θi (Q0 ) = G0i , i ∩ ∂Ω.
G0i := G+
We change the variable and write (9.6.18) in terms of vi ∈ H01 (Q+ ). Namely, we set
x = θi (y) ∈ G, y = ζi (x) ∈ Q, vi (y) = ui (θi (y)), ui (x) = vi (ζi (x)). (9.6.20)
The strategy for m = 0 is as follows. After changing the variable, it turns out that vi
satisfies a certain PDE in Q+ (see (9.6.18)). In Q+ , we can make translations Dh := Djh ,
j = 1, . . . , N − 1, and by using the technique of the regularity in the interior, we obtain
∂ij2 vi ∈ L2 (Q+ ), j = 1, . . . , N − 1, i = 1, . . . , N . The regularity ∂N2
N vi ∈ L (Q ) is
2 +
obtained by using the weak form equation for vi , which concludes that vi ∈ H 2 (Q+ ),
and therefore from Proposition 6.1.13 we obtain ui ∈ H 2 (G+ i ). For m > 0 we will proceed
by recurrence.
The following lemma is important. It shows that under the change of variables
(9.6.20), the equation (9.6.18) is transformed to other elliptic PDEs.
Lemma 9.6.6 Under the conditions of Theorem 9.6.1, vi ∈ H01 (Q+ ) and it solves

N
akl (y)(∂k vi (y)∂l ϕ(y))dy = bi (y)ϕ(y)dy, ∀ϕ ∈ H01 (Q+ ), (9.6.21)
Q+ k,l=1 Q+
where akl ∈ C m+1 (Q+ ), bi ∈ H m (Q+ ) and satisfy

n
−1 −1
akl (y) = |Jac(θi )|([∇θi ] · [∇θi ] )kl ,
t
a0 |ξ| ≤
2
akl ξk ξl ≤ A0 |ξ|2 , (9.6.22)
k,l=1
bi (y) = |Jac(θi )|gi (θi (y)), (9.6.23)
with A0 , a0 > 0.
Proof. For arbitrary ϕ ∈ H01 (Q+ ), set ψ(y) = ϕ(x), x = θi (y). Then ψ ∈ H01 (G+
i ) and
∇ui (x) · ∇ψ(x) = t (t [∇θi (y)]−1 · ∇vi (y)]) · (t [∇θi (y)]−1 · ∇ϕ(y))
= ∇vi (y) · [∇θi (y)]−1 · t [∇(y)]−1 · ∇ϕ(y).
Therefore, changing the variable in (9.6.18) gives

bi (y)ϕ(y)dy = gi (x)ψ(x)dx
Q+ G+
i
= (∇ui (x) · ∇ψ(x))dx

G+
i
207

= ∇vi (y) · [∇θi (y)]−1 · t [∇θi (y)]−1 · ∇ϕ(y)|Jac(θi (y))|dy
G+
i

N
= akl (y)(∂k vi (y)∂l ϕ(y))dy.
Q+ k,l=1
For arbitrary ξ = (ξ1 , · · · , ξN ) and y ∈ Q+ , we have

N
A0 |ξ| ≥
2
akl ξk ξl = |t [∇θi (y)]−1 · ξ|2 |Jac(θi (y))| ≥ a0 |ξ|2 ,
k,l=1
for certain A0 , a0 > 0, because akl ∈ C m+1 (Q), [∇θi ] is not singular in Q+ and |t [∇θi (y)]−1 ·
ξ| is a norm in RN equivalent to |ξ| (independently of y).
Now we can prove this theorem.

Theorem 9.6.7 Let Ω ⊂ RN be a bounded open set of class C m+2 and assume u ∈ H01 (Ω)
is a weak solution to (9.6.1). If f ∈ H m (Ω) then ui ∈ H m+2 (Ω), i = 1, . . . , n, and the
following estimation holds:
ui H m+2 (Ω) ≤ Cf H m (Ω) , (9.6.24)
with C independent of f (C depends only on Ω)

Proof. Let us first deal with the case m = 0, so f ∈ L2 (Ω). From Proposition 6.1.13, it
is enough to prove that vi solving (9.6.21) satisfies vi ∈ H 2 (Q+ ). For j = 1, . . . , N − 1,
we consider |h| small enough such that ϕ = Dj−h Djh vi ∈ H01 (Q+ ) (this is possible because
sup(ui ) ⊂ G+
i ). Replacing ϕ in (9.6.21) and using (9.6.7) gives

N
Djh (akl (y)∂k vi (y)) Djh (∂l vi (y))dy = bi (y)Dj−h Djh vi (y)dy,
Q+ k,l=1 Q+
which in combination with (9.6.6) implies

N
akl (y + hej )(Djh ∂k vi )(Djh ∂l vi )dy = bi Dj−h Djh vi dy
Q+ k,l=1 Q+

N
− (Djh akl )(∂k vi (Djh ∂l vi ))dy.
Q+ k,l=1
Using the fact that akl are C 1 functions, (9.6.22) and Lemma 9.6.3, from the last equation
we obtain

a0 Djh ∇vi 2L2 (Q+ ) ≤ bi L2 (Q+ ) + akl C 1 (Q+ ) ∇vi L2 (Q+ ) Djh ∇vi L2 (Q+ ) .
208
Note that from bi in (9.6.23), gi in (9.6.17), and the estimations (9.6.19), (9.6.3), we
obtain
bi L2 (Q+ ) + akl C 1 (Q+) ∇vi L2 (Q+ ) ≤ Cf L2 (Ω) .
Therefore, the last two inequalities and Lemma 9.6.3 yield
∂jk
2
vi L2 (Q+ ) ≤ Cf L2 (Ω) , j = 1, . . . , N − 1, k = 1, . . . , N, (9.6.25)
with C depending only on Ω. Note that this inequality implies that for all ϕ ∈ H01 (Q+ ),
we have

∂ v ∂ ϕdy ≤ Cf L2 (Ω) ϕL2 (Q+ ) , i = 1, . . . , N − 1, j = 1, . . . , N. (9.6.26)
+ j i k
Q
From (9.6.25), to prove vi ∈ H 2 (Q+ ) is enough to show that ∂N N vi ∈ L2 (Q+ ). From

∂N N vi , ϕD (Q+ )×D(Q+ ) = − Q+ ∂N vi ∂N ϕdy and from Riesz representation Theorem
7.2.1 in L2 (Ω), if

∂N vi ∂N ϕdy ≤ CϕL2 (Q+ ) , ∀ϕ ∈ D(Q+ ), (9.6.27)

Q+

then ∂N vi ∂N ϕdy = (z, ϕ)L2 (Ω) with a certain z ∈ L2 (Ω), and so ∂N N vi = z ∈ L2 (Q+ )
Q+
ϕ
and zL2 (Ω) ≤ C. Taking instead of ϕ in (9.6.21) (note that aN N ≥ a0 ) gives
aN N

ϕ ϕ
∂N vi ∂N ϕdy = bi + ∂N aN N ∂N v i (9.6.28)
Q+ Q+ aN N Q+ aN N
ϕ
akl

+ ∂k vi ∂l akl − ∂l ϕ ∂k v i .
Q+ aN N aN N
(k,l)=(N,N )
Combining (9.6.28) and (9.6.26) gives (9.6.27) with Cf L2 (Ω) instead of C, and this
proves vi ∈ H 2 (Q+ ).
For m ≥ 1, we proceed by recurrence as in Theorem 9.6.5. So let us assume that the
theorem holds for up to m − 1 and prove it for m. As in Theorem 9.6.5, Dα vi , |α| = m
solves
N
N

α
akl (∂k D vi )(∂l ϕ)dy = (D bi )ϕ −
α α
(D akl )(∂k vi ∂l ϕ)dy dy, (9.6.29)
Q+ k,l=1 Q+ k,l=1
for all ϕ ∈ H01 (Q+ ). This equation is obtained from (9.6.21) by taking Dα ϕ instead of
ϕ, with ϕ ∈ D(Q+ ), integrating by parts, and then proceeding by density. Note that
Dα bi ∈ L2 (Q+ ), Dα akl ∈ C 1 (Q+ ). Then we proceed with (9.6.29) as with (9.6.21) and
then Dα vi ∈ H 2 (Q+ ), so vi ∈ H m+2 (Q+ ).
Finally, the estimation (9.6.24) follows (9.6.25) and (9.6.27), which completes the
proof of the theorem.
209

Proof of Theorem 9.6.1 We consider only the case p = 2. We have u = u0 + ni=1 ui .
From Theorem 9.6.5, we have u0 ∈ H m+2 (Ω), and from Theorem 9.6.7 we have ui ∈
H m+2 (Ω), so u ∈ H m+2 (Ω). The estimation (9.6.2) follows from (9.6.13) and (9.6.24).
210
Bibliography
[1] R. Adams and J. Fournier. Sobolev Spaces. Second edition. Elsevier, 2003.
[2] S. Agmon, A. Douglis, and L. Nirenberg. “Estimates near the boundary for solutions
of elliptic partial differential equations satisfying general boundary conditions”. In:
Comm. Pure Appl. Math. I, II (1959).
[3] H. Amann. Vector-valued distributions of Fourier transform multipliers. University
of Zürich. 2003. url: https://www.math.uzh.ch/amann/files/distributions.ps.
[4] S. Benzoni-Gavage and D. Serre. Multi-dimensional Hyperbolic Partial Differen-
tial Equations: First-order Systems and Applications. Oxford Mathematical Mono-
graphs. Oxford University Press, 2007.
[5] H. Brezis. Functional Analysis, Sobolev Spaces and Partial Differential Equations.
Springer, 2011.
[6] C. Caratheodory. Calculus of Variations and Partial Differential Equations. Part
I: Partial Differential Equations of the First Order. San Francisco, London, Ams-
terdam: Holden-Day, Inc., 1965.
[7] H. Cartan. Differential calculus. Paris: Hermann, 1971.
[8] Ph. Ciarlet. Linear and Nonlinear Functional Analysis with Applications. SIAM,
2013.
[9] R. Courant and D. Hilbert. Methods of Mathematical Physics. Vol. 2. Wiley-
Interscience, 1989.
[10] C. M. Dafermos. Hyperbolic Conservation Laws in Continuum Physics. Springer,
2010.
[11] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for
Science and Technology. Evolution Problems. Vol. V. Berlin: Springer-Verlag, 2000.
[12] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for Sci-
ence and Technology. Functional and Variational Methods. Vol. II. Berlin: Springer-
Verlag, 2000.
© The Editor(s) (if applicable) and The Author(s), under exclusive license 211
to Springer Nature Switzerland AG 2023
https://doi.org/10.1007/978-3-031-39524-6
Bibliography
[13] R. Dautray and J.-L. Lions. Mathematical Analysis and Numerical Methods for Sci-
ence and Technology. Spectral Theory and Applications. Vol. III. Berlin: Springer-
Verlag, 2000.
[14] F. Demengel and G. Demengel. Functional Spaces for the Theory of Elliptic Partial
Differential Equations. Fractional Sobolev Spaces. Springer, 2012.
[15] J. Diestel and Jr. J. J. Uhl. Vector Measures. Mathematical Surveys and Mono-
graphs. Vol. 15. AMS, 1977.
[16] L. C. Evans. Partial Differential Equations. Vol. 19. Graduate Studies in Mathe-
matics. AMS, 1998.
[17] E. Fabes, O. Mendez, and M. Mitrea. “Boundary Layers on Sobolev-Besov Spaces
and Poisson’s Equation for the Laplacian in Lipschitz Domains”. In: Journal of
Functional Analysis 159 (1998), pp. 323–368.
[18] G. G. Friedlander and M. Joshi. Introduction to the theory of distributions. Cam-
bridge University Press, 1999.
[19] P. Garabedian. Partial Differential Equations. Wiley, 1964.
[20] D. Gilbarg and N. S. Trudinger. Elliptic Partial Differential Equations of Secon-
dorder. Vol. 224. Springer, 2001.
[21] Ch. Heil. Introduction to Real Analysis. Graduate Texts in Mathematics. Springer,
2019.
[22] L. Hörmander. The analysis of linear partial differential operators. Distribution
Theory and Fourier Analysis. Vol. I. Springer-Verlag, 1983.
[23] G. Hörmann and R. Steinbauer. Lecture notes on Theory of dis-
tributions. Fakultät für Mathematik, Universität Wien, 2009. url:
https://www.mat.univie.ac.at/ stein/lehre/SoSem09/distrvo.pdf.
[24] A. Jonsson and H. Wallin. Function Spaces on Subsets of Rn . New York: Harwood
Academic, 1984.
[25] D. Kim. “Trace theorems for Sobolev-Slobodeckij spaces with or without weights”.
In: J. Funct. Spaces Appl. 5.3 (2007), pp. 243–268.
[26] E. Kreyszig. Introductory functional analysis with applications. New York: Wiley,
1978.
[27] O. A. Ladyzhenskaja, V. A. Solonnikov, and Ural’ceva N. N. Linear and quasilinear
equations of parabolic type. Translations of Mathematical Monographs. Providence,
RI: AMS, 1968.
[28] S. Lang. Real and Functional Analysis. Graduate Texts in Mathematics, Third
edition. Springer, 1991.
212
Bibliography
[29] P. Lax. Functional Analysis. Pure and Applied Mathematics, A Wiley-Interscience

Series of Texts, Monographs and Tracts. Wiley, 2002.
[30] P. Lax. Hyberbolic Systems of Conservation Laws and the Mathematical Theory of
Shock Waves. SIAM, 1973.
[31] P. Lax. “Nonlinear hyperbolic equations”. In: Communications on Pure and Applied
Mathematics VI (1953), pp. 231-258.
[32] P. G. Leoch. Hyperbolic systems of conservation laws. The theory of classical and
nonclassical shock waves. Lectures in Mathematics ETH Zürich. Basel: Birkhäuser
Verlag, 2002.
[33] G. Lefloni. A First Course in Sobolev Spaces. Graduate Studies in Mathematics.
AMS, 2009.
[34] E. H. Lieb and M. Loss. Analysis. AMS, 2001.
[35] J.-L. Lions and E. Magenes. Non-homogeneous boundary value problems and appli-
cations. Vol. 1. Springer-Verlag, 1972.
[36] J. Málek et al. Weak and Measure-valued Solutions to Evolutionary PDEs. CRC
Press, 1996.
[37] J. Marschall. “The trace of Sobolev-Slobodeckij spaces on Lipschitz domains”. In:
Manuscripta Math. 58 (1987), pp. 47–65.
[38] N. G. Meyers. “An Lp estimate for the gradient of solutions of second-order elliptic
divergence equations”. In: Annali della Scuola Normale Superiore di Pisa. Classe
di Scienze 3e série 17.3 (1963), pp. 189–206.
[39] D. Mitrea, M. Mitrea, and L. Yan. “Boundary value problems for the Laplacian in
convex and semiconvex domains”. In: Journal of Functional Analysis 258 (2010),
pp. 2507–2585.
[40] J. Neças. Direct Methods in the Theory of Elliptic Equations. Springer, 2012.
[41] E. Di Nezza, G. Palatucci, and E. Valdinoci. “Hitchhiker’s guide to the fractional
Sobolev spaces”. In: Bulletin des Sciences Mathématiques 136 (5 2012), pp. 521-573.
[42] L. Nirenberg. Topics in Nonlinear Functional Analysis. Vol. 6. AMS and Courant
Institute of Mathematical Sciences, 2001.
[43] M. Renardy and R. Rogers. An Introduction to Partial Differential Equations, Sec-
ond Edition. Springer, 2003.
[44] Th. Runst and W. Sickel. Sobolev Spaces of Fractional Order, Nemytskij Operators,
and Nonlinear Partial Differential Equations. Vol. 3. De Gruyter Series in Nonlinear
Analysis and Applications. De Gruyter, 1996.
213
Bibliography
[45] V. S. Rychkov. “On restrictions and extensions of the Besov and Triebel-Lizorkin
spaces with respect to Lipschitz domains”. In: J. Lond. Math. Soc. 60 (1 1999), pp.
237–257.
[46] C. Schneider. “Traces of Besov and Triebel-Lizorkin spaces on domains”. In: Math.
Nachr. 284.5-6 (2011), pp. 572–586.
[47] L. Schwartz. Théorie des distributions. Hermann, 1978.
[48] E. M. Stein. Singular Integrals and Differentiability of Functions. Vol. 30. Princeton
Mathematical Series. Princeton University Press, 1971.
[49] E. M. Stein and G. Weiss. Introduction to Fourier Analysis on Euclidean Spaces.
Vol. 32. Princeton Mathematical Series. Princeton University Press, 1990.
[50] R. S. Strichartz. A Guide to Distribution Theory and Fourier Transforms. World
Scientific Publishing, 2003.
[51] L. Tartar. An introduction to Sobolev spaces and interpolation spaces. Springer,
2007.
[52] R. Temam. Navier-Stokes Equations: Theory and Numerical Analysis. Vol. 343.
AMS Chelsea Publishing, 1984.
[53] F. Treves. Basic Partial Differential Equations. Academic Press, 1975.
[54] H. Triebel. Interpolation Theory, Function Spaces and Differential Operators. Ams-
terdam, New York, Oxford: North-Holland, 1978. 214
214
Index
abbreviations of Cb∞ (Ω) in W s,p (Ω), 112

a.a., almost all, 12 of D in W k,p , 96
a.e., almost everywhere, 67 derivative
i.e., in Latin id est, which means that Dα u, Dm u, 5
is, 1 ∂νn u, n-th order normal derivative,
iff, if and only if, 94 115
Lebesgue DCT, 13 ∂ν u, normal derivative of u, 57
MSV, method of separation of vari- difference quotient, 203
ables, 49 differential dimension
WLOG, without loss of generality, of W s,p (Ω), 104
31 of W k,p (Ω), 94
distributions
characteristic pv(x−1 ), principal value of x−1 , 74
curve, 25, 26
addition, subtraction, multiplication,
equations, 25 71
condition
composition with a C ∞ diffeomor-
admissible initial, 32
phism, 181
compatibility, 32
convergence, 71
ellipticity, 125
convolution, 76, 192
Lax entropy, 44
derivative, 73
transversality, 32
Dirac measure δ0 , 71
convergence
Dirac measure δa , 86
in D (Ω), 71
finite order of a tempered, 188
in D(Ω), 68
fundamental solution to heat equa-
in S, 81
tion, 79
in S , 82
k,p fundamental solution to Laplacian,
in Wloc (Ω), 96
88
uniform, 159
fundamental solution to wave equa-
density tion, 80
of D and S in H s , 106 measure, 70
k,p
of D in Wloc , 96 of finite order, 73
of Cb (Ω) in W k,p (Ω), 200
∞ order, 70
principal value of x−1 f , 183
© The Editor(s) (if applicable) and The Author(s), under exclusive license 215
to Springer Nature Switzerland AG 2023
https://doi.org/10.1007/978-3-031-39524-6
Index
support of, 72 Poincaré, 117

tempered, 81 Young, 181
zero set, 72
domain Laplacian
C k , 10 in polar coordinates, 63
Lipschitz, 10 in spherical coordinates, 170
minimally smooth, 10 lemma
Arzela-Ascoli, 177
embedding Ehrling, 196
compact, 105, 110 fundamental, 13
continuous, 105, 107 Hopf, 57, 173
dense, 106 Lax-Milgram, 125
Riemann-Lebesgue, 190
formula
d’Alembert solution to the wave equa- method
tion, 146 of characteristics, 26
Poisson’s, 54 of sep. of variables for heat equation,
function 140
Ck, 3 of sep. of variables for Laplace eq.,
{K, G} cut-off, 9 50
sgn(x) = 2H(x) − 1, 117 of sep. of variables for wave eq., 143
barrier, 61 Perron’s, 62
contraction, 14 mollifier
convolution, 69 function, 7
diffeomorphism, 14 sequence, 7
Dirac δ0 , 71
norm
Dirac δa , 86
||·||C k (∂Ω) , 11
domain/image of, 17
||·||X n , 2
harmonic lifting, 60
||·||L(X n ;Y ) , 3
Heaviside, H(x), 67
| · |p , 1
homeomorphism, 101
||·||Cbk (Ω) , ||·||C k,λ (Ω) , 6
isometry, 106 b
isomorphism, 14 ||·||Cbk (Ω) , ||·||C k,λ (Ω) , 7

b
Lipschitz, 4 ||·||Lp (Ω) , 12

positive part u+ , negative part u− , ||·||W k,p (Ω) , ||·||W k,p (Ω) , 94
100 ||·||W s,p (Ω) , ||·||W s,p (Ω) , 104
subharmonic, superharmonic, 59 notation
support of, 7, 68 A − B = ∪{a − b, a ∈ A, b ∈ B} for
uniformly continuous, 2 any A, B ⊂ RN , 186
weak barrier, 62 B(x, r) = {y ∈ RN , |y − x| < r}, 7
BR (x) = {y ∈ RN , |x − y| < R}, 54
inequality BR = {x ∈ RN , |x| < R}, 53
Hölder, 181 Dα u, 5
216
Index
M ≥ 0 or M > 0 for M ∈ RN ×N , kernel, 125

171 range, 125
M , the inverse of M ∈ RN ×N , 20
−1
order
M·,j , the j-th column vector of M , of W k,p (Ω) space, 94
35 of a distribution, 73
Mi,· , the i-th row vector of M , 35
pairing/duality V × V , 68
U G: U compact, G open in RN
partition of unity, 8
and U ⊂ G, 9
PDE
VN , the volume of the unit ball, 53
admissible condition, 32
[u], jump of u, 42
boundary conditions, 20
Ωc = {x ∈ X, x ∈ / Ω} = X\Ω, with
Burgers’, 37, 38
X Banach, Ω ⊂ X, 7
classical solution, 20
s, 103
compatibility condition, 32
(x), , x L(X n ;Y )×X n , 2
definition, 19
ξ = (1 + |ξ|2 )1/2 , 101 elliptic (with constant coefficients),
∂Ω, boundary of Ω, 7 21
∂ν u, normal derivative of u, 57 fully nonlinear first-order, 30
w̌, 84 heat equation, 22
B(x, r) = {y ∈ RN , |y − x| ≤ r}, 7 hyperbolic (with constant coefficients),
Ω = Ω ∪ ∂Ω, 7 21
|Sn |, 88 Laplace equation, 22
|α|, α ≤ β, α ≤ m, α21 , xα , 5 linear first-order, 28
a ∼ b, 95 linear/nonlinear, 20
dist(A, B) = inf{|a − b|, a ∈, b ∈ linear/nonlinear boundary conditions,
B}, 203 20
dx, N -dimensional Lebesgue measure, order, 19
12 parabolic (with constant coefficients),
p , the conjugate of p: p1 + p1 = 1, 12 21
u+ , u− , 100 prototypes, 22
N, N0 , 1 quasi-linear first-order, 29
R, RN , C, CN , 1 transport equation, 22
RN+ , 10 transversality condition, 32
supp(u), 7 wave equation, 23
t
M the transposed of M ∈ RN ×N , 20 weak form, 124
multi-index α ∈ NN 0 , 5 weak solution, 124
weak solution to a conservation law,
operator 41
E, W s,p (Ω) extension, 108 point
E0 , 196 regular w.r.t. Laplacian, 62
(boundary) trace, 112 principe
adjoint, 125 strong maximum, 57
convolution, 69 weak maximum, 56, 126
217
Index
regularization, 96 weak solution to transport equation,

27
seminorm weak solution to wave PDEs, 153
| · |C k,λ (Ω) , 6 space
b
| · |C k,λ (Ω) , 7 D(Ω), 68
b N
| · |W k,p (Ω) , 94 D(R+ ), 197
| · |W s,p (Ω) , 104 D(Ω), 200
sm , 81 DK (Ω), 68
sequence S, 81
Cauchy, 1 S(Ω), 112
equicontinuous, 177 C 0 (Ω; Y ), 2
uniformly bounded, 177 C 0 (∂Ω), Cb0 (∂Ω), 8
uniformly equicontinuous, 177 C00 (Ω), C0k (Ω), C0∞ (Ω), D(Ω), 7
set Cb0 (Ω), Cbk (Ω), Cbk,λ (Ω), Cb∞ (Ω), 7
compact, 7 Cb0 (Ω), Cbk (Ω), Cbk,λ (Ω), Cb∞ (Ω), 6
orthogonal, 125 C 1 (Ω; Y ), 3
precompact, 105, 133 C k (Ω; Y ), C ∞ (Ω; Y ), 4
solution C k (∂Ω), 11
classical solution to a PDE, 20 Eloc (Ω), 94
classical solution to hyperbolic (wave) H k (Ω), 94
PDE, 144 H s (RN ), 102
classical solution to parabolic (heat) L2k (RN ), 101
PDEs, 141 L2s (RN ), 102
classical solution to the Dirichlet prob- Lp (Ω), 12
lem for the Laplacian, 49 Lp (Ω), reflexive, 12
d’Alembert solution to the wave equa- Lp (Ω), separable, 13
tion, 146 W k,p (Ω), 94
fundamental solution to heat equa- W k,p (Ω), reflexive, 94
tion, 79 W k,p (Ω), separable, 94
fundamental solution to Laplacian, W s,∞ (Ω), 105
88 W s,p (∂Ω), 114
fundamental solution to wave equa- X , X , 2
tion, 80 Lip(Ω; Y ), C 0,1 (Ω; Y ), 4
Lax entropy, 44 L(X; Y ), L(X n ; Y ), 2
stable, 15 M(Ω), 12
subsolution, supersolution, 59 complete metric, 14
traveling backward wave, 145 separable, 13
traveling forward wave, 145
weak solution to conservation laws, theorem
41 Arzela-Ascoli, 177
weak solution to elliptic PDEs, 124 Friedrichs, 96
weak solution to heat PDEs, 150 Fubini, 161
218
Index
Gauss, 43 Riesz representation, 125

Lebesgue dominated convergence, 13 Tonelli, 160
mean value, 54 transform
Parseval, 85 Fourier, 82
Rankine-Hugoniot, 42, 167 inverse Fourier, 84, 188
219

(CMS - CAIMS Books in Mathematics 11) Arian Novruzi - A Short Introduction To Partial Differential Equations-Springer Nature Switzerland (2023)

Uploaded by

Copyright:

Available Formats

You might also like

(CMS - CAIMS Books in Mathematics 11) Arian Novruzi - A Short Introduction To Partial Differential Equations-Springer Nature Switzerland (2023)

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

(CMS - CAIMS Books in Mathematics 11) Arian Novruzi - A Short Introduction To Partial Differential Equations-Springer Nature Switzerland (2023)

Uploaded by

Copyright:

Available Formats

CMS/CAIMS Books in Mathematics

A Short Introduction to Partial

ISSN 2730-650X ISSN 2730-6518 (electronic)

Paper in this product is recyclable.

Ottawa, Canada Arian Novruzi

1 Notations and review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Partial differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 First-order PDEs: classical and weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4 Second-order linear elliptic PDEs: maximum principle and classical solutions . . . . . . . . . . . . . . 49

5.4 Tempered distributions and Fourier transform . . . . . . . . . . . . . . . . . . . . . 81

7 Second-order linear elliptic PDEs: weak solutions . . . . . . . . . . . . . . . . . 123

8 Second-order parabolic and hyperbolic PDEs . . . . . . . . . . . . . . . . . . . . . 139

9.3.3 Solution to the Dirichlet problem . . . . . . . . . . . . . . . . . . . ...... 174

1.1 Continuous diﬀerentiable functions

In general, a Banach space is deﬁned as follows.

lim xn − xX = 0.

lim u(x + h) − u(x)Y = 0 (1.1.3a)

∀ > 0, ∃δ,x > 0, ∀h ∈ Ω, hX < δ ⇒ u(x + h) − u(x)Y < . (1.1.3b)

If u is continuous at every x ∈ Ω, we say “u is continuous in Ω”. The set of such

for all i = 1, . . . , n, x1 , . . . , xi−1 , xi , xi , xi+1 , · · · , xn ∈ X and ci , ci ∈ R.

ii) L(X n ; Y ), or simply L(X; Y ) when n = 1, is deﬁned as

L(X n ; Y ) = { ∈ C 0 (X n ; Y ), is n-linear}. (1.1.4)

with a certain C ≥ 0 independent of x, see [26], and

(x)Y = inf{C ≥ 0, (x)Y ≤ CxX n , ∀x ∈ X n }. (1.1.5)

2) For k = 2, 3, . . ., “the k-th order derivative of u at x” is deﬁned by u(k) (x) :=

C k (Ω; Y ) = {u ∈ C 0 (Ω; Y ) such that for alln = 1, . . . , k we have

We also deﬁne C ∞ (Ω; Y ) = C k (Ω; Y ). (1.1.9)

3) The space of Lipschitz functions is deﬁned by

Lip(Ω; Y ) = C 0,1 (Ω; Y )

4) When Y = R we will omit Y in the spaces of this deﬁnition, so we will write

In the following we consider X = RN and Y = R. In this case we have equivalent

Often we write Du or ∇u instead of D1 u.

In general, C k (Ω) is not a Banach space because it contains unbounded functions.

Proposition 1.1.11 Let Ω ⊂ RN be open, k ∈ N0 , λ ∈ (0, 1], and deﬁne

Cb0 (Ω) = {u ∈ C 0 (Ω), u is bounded in Ω}, (1.1.16)

Cb∞ (Ω) = Cbk (Ω). (1.1.19)

uCbk (Ω) = sup{Dα uCb0 (Ω) , |α| ≤ k}, resp. (1.1.20)

Proof. See Proposition 9.1.1 in Annex.

Cb0 (Ω) = {u ∈ Cb0 (Ω), u extends continuously in Ω}, (1.1.22)

Cb∞ (Ω) = Cbk (Ω), (1.1.24)

C00 (Ω) = {u ∈ C 0 (Ω),

Example 1.1.13 Let ρ : RN → R be deﬁned by

It is easy to verify that

We set ρn (x) = nN ρ(nx), n ∈ N. The sequence (ρn ) satisﬁes

Every function satisfying (1.1.29) is called a “molliﬁer”. Given a molliﬁer ρ, the

uC 0 (∂Ω) = max{|u(x)|, x ∈ ∂Ω}. (1.1.32)

1.2 Domains and C k (∂Ω) spaces

1.2.1 Partition of unity

Theorem 1.2.2 (cut-oﬀ function and partition of unity)

is the desired partition.6

1.2.2 Domains and C k (∂Ω) spaces

we say “θx is a bi-Lipschitz function with Lipschitz bound Lx ”).

i) if x ∈ ∂Ω then B(x, r0 ) ⊂ Gi , for a certain i ∈ I,

Deﬁnition 1.2.7 (C k (∂Ω) spaces) Assume Ω is an open bounded set of class C k , k ≥

uC k (∂Ω) = (uϕi ) ◦ θi (·, 0)C0k (Q0 ) . (1.2.4)

1.3 Review of some important results

1.3.1 Some results from Lp (Ω) spaces

The space Lp (Ω) is equipped with the norm

lim xn − xX = 0.

lim u(x + h) − u(x)Y = 0 (1.1.3a)

∀ > 0, ∃δ,x > 0, ∀h ∈ Ω, hX < δ ⇒ u(x + h) − u(x)Y < . (1.1.3b)

for all i = 1, . . . , n, x1 , . . . , xi−1 , xi , xi , xi+1 , · · · , xn ∈ X and ci , ci ∈ R.

(x)Y = inf{C ≥ 0, (x)Y ≤ CxX n , ∀x ∈ X n }. (1.1.5)

uCbk (Ω) = sup{Dα uCb0 (Ω) , |α| ≤ k}, resp. (1.1.20)

Example 1.1.13 Let ρ : RN → R be deﬁned by

uC 0 (∂Ω) = max{|u(x)|, x ∈ ∂Ω}. (1.1.32)

uC k (∂Ω) = (uϕi ) ◦ θi (·, 0)C0k (Q0 ) . (1.2.4)

α(·, x) − α(·, y)C 0 (I r ;X) ≤ Lx − yX . (1.3.5)

β (τ, y) = g(τ, β(τ, y)), τ ∈ J, (1.3.6a)

α : I × U → U, α(·, x) ∈ C 1 (I; X),

b) “f C 0 (B(x0 ,2ρ) ≤ 1 + K” is equivalent to “gC 0 (B(s0 ,y0 ),2ρ) ≤ K” where x0 =

i) Prove that g ◦ f ∈ C 0 (X; Z), where g ◦ f : X → Z, (g ◦ f )(x) = g(f (x)).

with a, b, c ∈ R, b2 − ac > 0. Assume2 a2 + c2 = 0, for example c = 0. Then (2.0.5) is