GRADUATE STUDIES

I N M AT H E M AT I C S 197

A Course on
Partial Differential
Equations
Walter Craig

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197

GRADUATE STUDIES
I N M AT H E M AT I C S 197

A Course on
Partial Differential
Equations

Walter Craig

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 EDITORIAL COMMITTEE
Daniel S. Freed (Chair)
Bjorn Poonen
Gigliola Staffilani
Jeff A. Viaclovsky

2010 Mathematics Subject Classification. Primary 35-XX, 42-XX.

For additional information and updates on this book, visit

www.ams.org/bookpages/gsm-197

Library of Congress Cataloging-in-Publication Data

Names: Craig, Walter, 1953- author.
Title: A course on partial differential equations / Walter Craig.
Description: Providence, Rhode Island : American Mathematical Society, [2018] | Series: Gradu-
ate studies in mathematics ; volume 197 | Includes bibliographical references and index.
Identifiers: LCCN 2018035511 | ISBN 9781470442927 (alk. paper)
Subjects: LCSH: Differential equations, Partial–Textbooks. | AMS: Partial differential equations.
msc | Harmonic analysis on Euclidean spaces. msc
Classification: LCC QA377 .C85 2018 | DDC 515/.353–dc23
LC record available at https://lccn.loc.gov/2018035511

Copying and reprinting. Individual readers of this publication, and nonprofit libraries acting
for them, are permitted to make fair use of the material, such as to copy select pages for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for permission
to reuse portions of AMS publication content are handled by the Copyright Clearance Center. For
more information, please visit www.ams.org/publications/pubpermissions.
Send requests for translation rights and licensed reprints to [email protected].

c 2018 by the American Mathematical Society. All rights reserved.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at https://www.ams.org/
10 9 8 7 6 5 4 3 2 1 23 22 21 20 19 18

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Contents

Preface vii

Chapter 1. Introduction 1
§1.1. Overview of the subject 1
§1.2. Examples 2
Exercises: Chapter 1 8

Chapter 2. Wave equations 9

§2.1. Transport equations: The Fourier transform 10
§2.2. Transport equations: The method of characteristics 12
§2.3. Conservation laws 15
§2.4. The d’Alembert formula 16
§2.5. Duhamel’s principle 19
§2.6. The method of images 21
§2.7. Separation of variables 24
Exercises: Chapter 2 34
Projects: Chapter 2 37

Chapter 3. The heat equation 43

§3.1. The heat kernel 44
§3.2. Convolution operators 46
§3.3. The maximum principle 48
§3.4. Initial and initial-boundary value problems 49
§3.5. Conservation laws and the evolution of moments 56

iii

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 iv Contents

§3.6. The heat equation in Rn 61

§3.7. Entropy 62
§3.8. Gradient flow 65
Exercises: Chapter 3 66
Projects: Chapter 3 69
Chapter 4. Laplace’s equation 71
§4.1. Dirichlet, Poisson, and Neumann boundary value problems 71
§4.2. Green’s identities 72
§4.3. The fundamental solution 74
§4.4. Maximum principle 77
§4.5. Green’s functions and Dirichlet–Neumann operators 80
§4.6. Poisson kernel on Rn+ 86
§4.7. Maximum principle again 90
§4.8. Oscillation and attenuation estimates 91
§4.9. Hadamard variational formula 93
Exercises: Chapter 4 95
Projects: Chapter 4 99
Chapter 5. Properties of the Fourier transform 103
§5.1. Hilbert spaces 104
§5.2. Schwartz class 107
§5.3. Fourier transform of L1 -integrable functions 113
Exercises: Chapter 5 119
Chapter 6. Wave equations on Rn 121
§6.1. Wave propagator by Fourier synthesis 122
§6.2. Lorentz transformations 123
§6.3. Method of spherical means 126
§6.4. Huygens’ principle 132
§6.5. Paley–Wiener theory 134
§6.6. Lagrangians and Hamiltonian PDEs 139
Exercises: Chapter 6 144
Projects: Chapter 6 147
Chapter 7. Dispersion 151
§7.1. Schrödinger’s equation 152
§7.2. Heisenberg uncertainty principle 156

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Contents v

§7.3. Phase and group velocities 159

§7.4. Stationary phase 162
Exercises: Chapter 7 166
Projects: Chapter 7 168
Chapter 8. Conservation laws and shocks 177
§8.1. First-order quasilinear equations 177
§8.2. The Riemann problem 183
§8.3. Lax–Olenik solutions 186
Exercises: Chapter 8 195
Projects: Chapter 8 196
Bibliography 201
Index 203

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Preface

Under the Wave off Kanagawa, K. Hokusai, 1829–33.

This textbook is intended to cover material for an introductory course in

partial differential equations (PDEs), aimed for fourth year undergraduate
or first year graduate mathematics students. The material focuses on the
three most important aspects of the subject: namely (i) theory, which is to
say the questions of existence, uniqueness, and continuous dependence on
given data or parameters; (ii) phenomenology, meaning the study of prop-
erties of solutions of the PDEs that we examine; and (iii) applications, by
which we mean that we will make an attempt to exercise good scientific
taste in choosing the topics in this text, often based on physical or geomet-
rical applications. In taking up this subject, it is invariable that there are

vii

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 viii Preface

many editorial choices that one must make. The area of PDEs is not one
unified topic; rather it is a union of numerous directions of research with
different sources of problems, motivations, and goals. My own point of view
in this area has been strongly influenced by my teachers and colleagues,
including Louis Nirenberg, Jürgen Moser, Henry McKean, Joe Keller, and
Constantine Dafermos, as well as my many friends and collaborators, in-
cluding Catherine Sulem, Claude Bardos, Jean-Claude Saut, Eugene Wayne,
Sergei Kuksin, and Håkan Eliasson. Members of my McMaster PDE class
in 2015–2016 helped with the preparation of the text and the figures for
this publication; for their input I would like to thank in particular Alexandr
Chernyavsky, Nikolay Hristov, Kyle MacDonald, and Adam Sliwiak. I also
want to extend a special thanks to Sarnia (Nia) Sulaiman for her rendering
of numerous figures, to Ramsha Khan for her work on the images, to Emma
Holmes for her proofreading, and especially to Deirdre Haskell who read
the text thoroughly and critically and suggested many improvements to the
presentation. However, the material itself and its presentation in this text
is based on my own judgement and experiences, both scientific and peda-
gogical, and responsibility for omissions and incompleteness is entirely my
own.
The present form of the course material follows from having given courses
on PDEs many times over the years; certainly I have taught similar courses
during my time on the faculty at McMaster University. But the kernel of
the course material also comes from previous experiences with the gradu-
ate Mathematical Physics course at Stanford University, the two-semester
sequence of courses on PDEs at Brown University, and a yearlong graduate
course in PDEs at the Fields Institute in Toronto, in conjunction with the
yearlong thematic program on PDEs.
Suggestions on how to use this book. The first chapters, plus a selection
of the more in-depth material of the book, in particular Chapters 5–7 and
sections 8.1 and 8.2, make quite a complete and interesting first semester
graduate course in PDEs. If one includes the full content of Chapters 4,
6, and 8, it provides a more challenging semester-long first year graduate
level course. Each chapter concludes with a set of exercises; these are on
topics that augment the text, and to a large extent they are selected for
their independent interest. In addition to the exercises there are a number
of more challenging in-depth and open-ended projects in most chapters.
Alternatively, I have taught a pleasant semester-long undergraduate course
on PDEs at McMaster University based on the material in Chapters 2–4
and sections 7.1 and 7.2, including some of the material from Chapter 5
for analytic rigor. The course started with an introductory lecture that is
appropriate for this level of class.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Preface ix

I am hopeful that the point of view adopted in this text, which brings
phenomenology and applications along with the theory and intends to in-
troduce the class to a spectrum of techniques, will be useful to students
entering the field.
This work was partially supported by the Canada Research Chairs Pro-
gram, the Fields Institute, and by NSERC with grant number 238452-16.

Walter Craig,
Hamilton, Ontario

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/01

Chapter 1

Introduction

Wente Torus: a genus-one surface of constant nonzero mean curvature

discovered by Henry Wente in 1986. Image made in 1988 by Jim Hoff-
man and Yi Fang from the Geometry, Analysis, Numerics & Graphics
research group at the University of Massachusetts and the Mathematical
Sciences Research Institute.

1.1. Overview of the subject

It is the irony of taking university courses that one doesn’t understand the
real reason for studying a subject until one knows it already and has been
steeped in its culture. With this paradox in mind, I will attempt to give an

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2 1. Introduction

introduction that will motivate the material we are going to address in this
course, so we can at least start with a sense of its content.
The first questions are possibly “Where do partial differential equations
(PDEs) arise, and why are they useful?”. In fact, the language of the sciences
is mathematics (the joke has it that the language of the sciences is English
with an accent). Many if not most statements in the physical sciences are
in the form of mathematical equations, and the vast majority of these are
differential equations, quantifying the change of one quantity in terms of
others. Indeed, equations describing physical, chemical, and sometimes bi-
ological phenomena are for the most part PDEs, and the same statement
holds for the engineering sciences.
Disciplines of mathematics such as geometry and dynamical systems also
often give rise to PDEs. Conditions such as (i) a surface is of minimal area,
(ii) a submanifold is invariant under a flow, (iii) a mapping is conformal,
or (iv) the curvature tensor satisfies a particular property, are PDEs. For
example, the statement that “the tensor (gμν ) is an Einstein metric for a
manifold M ” is a system of PDEs.
The course material will discuss the most commonly occurring PDEs
and the implications they have for a function to be a solution. We are
particularly interested in knowing whether solutions exist, whether they are
unique, and their properties of smoothness, positivity, etc.

1.2. Examples
Here is a set of examples of some common PDEs with a brief overview of
the settings where they appear, either as a characterization of mathematical
properties of a geometrical object or as a description of a physical phenom-
enon.

Conformal mappings. A mapping between two Euclidean spaces x ∈

Rn → u(x) ∈ RN can be explicitly locally represented as
x = (x1 , . . . , xn ) ∈ Rn → u(x) = (u1 (x), . . . , uN (x)) ∈ RN ,
u : x → u(x) .
A conformal mapping is one that preserves angles. The tangent space of
the image of the mapping u(x) is spanned by the vectors (∂x1 u, . . . , ∂xn u).
An image of a conformal mapping is given in Figure 1. Set n = N = 2 for
our example. In the x-variables the tangent vector fields ex1 = (1, 0) and
ex2 = (0, 1) are orthogonal and have the same length in the domain R2 ; a
conformal mapping u(x) must preserve this property for the two tangent
vectors ∂x1 u and ∂x2 u in the image:
(1.1) ∂x1 u · ∂x2 u = 0 , |∂x1 u|2 = |∂x2 u|2 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 1.2. Examples 3

xn uN

x1

u1

Figure 1. A conformal mapping between Rn and its image as a sub-

manifold of RN .

Equations (1.1) are equivalent to the Cauchy–Riemann equations for either

u = (u1 (x1 , x2 ), u2 (x1 , x2 )) or u = (u1 (x1 , x2 ), −u2 (x1 , x2 )). That is, either
u is holomorphic (analytic) in x1 + ix2 ,
(1.2) ∂x1 u1 = ∂x2 u2 , ∂x2 u1 = −∂x1 u2 ,
or else u is antiholomorphic:
(1.3) ∂x1 u1 = −∂x2 u2 , ∂x2 u1 = ∂x1 u2 .
The study of such mappings is thus a central topic of complex analysis.
Exercise 1.1 of this chapter is to show that (1.1) implies either (1.2) or
(1.3).

Laplace’s equation. Starting with the Cauchy–Riemann equations (1.2)

and differentiating again, we have
∂x21 u1 = ∂x1 (∂x2 u2 ) = ∂x2 (∂x1 u2 ) = −∂x22 u1 ;
thus
(1.4) (∂x21 + ∂x22 )u1 = 0 ,
that is, u1 is harmonic. The same goes for u2 of course. The operator
Δ = ∂x21 + ∂x22 is called the Laplacian; its higher-dimensional version is

n
(1.5) Δu := (∂x21 + ···+ ∂x2n )u(x) =( ∂x2j )u(x) .
j=1

In general Laplace’s equation is

(1.6) Δu = 0 ,
and solutions of (1.6) are called harmonic as well. This equation is a princi-
pal example of an elliptic equation, which will be discussed in more detail in
Chapter 4. The Laplacian operator appears very often in mathematics, par-
tially because it is the only second-order linear differential operator which

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4 1. Introduction

is invariant under the symmetries of Euclidean space, that is, translations

and rotations:
(1.7) x → x + c , x → Rx ,
with the rotation R represented by an orthogonal matrix RT = R−1 . This
is one elementary manifestation of the principle of relativity, which is that
equations which describe physical phenomena should be invariant under
symmetries of the underlying space.

The heat equation. Given coordinates of time and space as (t, x) =

(t, x1 , . . . , xn ) ∈ R1+n , the heat equation for a function u(t, x) is
1
(1.8) ∂t u = Δu ;
2
it is the principal example of a parabolic equation. It was originally derived
to study problems of heat conduction, and it occurs very often in probability
and in problems of gradient flow, among other places.

The wave equation. Again in coordinates (t, x) ∈ R1+n of space-time, the

wave equation for a function u(t, x) is
(1.9) ∂t2 u = Δu .
It is a principal example of a hyperbolic equation; we will study a number
of related equations in the next chapter, including the case of the wave
equation in one space dimension.

Maxwell’s equations. The propagation of electromagnetic radiation is

governed by Maxwell’s equations, the derivation of which is a towering
achievement of nineteenth-century science. This is a system of equations
coupling two vector quantities in R3 , the electric and the magnetic vector
fields:
(1.10) E(t, x) = (E1 (t, x), E2 (t, x), E3 (t, x)) ,
B(t, x) = (B1 (t, x), B2 (t, x), B3 (t, x)) .
The six components of (E, B) satisfy the coupled system of eight equations:
∂t E = ∇ × B − 4πj ,
(1.11) ∂t B = −∇ × E ,
∇ · E = 4πρ ,
∇·B =0 .
The vector calculus notation is that the divergence operator ∇ · E is

3
∇ · E := ∂x1 E1 + ∂x2 E2 + ∂x3 E3 = ∂xj Ej
j=1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 1.2. Examples 5

and the curl is given by

∇ × B := (∂x2 B3 − ∂x3 B2 , ∂x3 B1 − ∂x1 B3 , ∂x1 B2 − ∂x2 B1 ) .
The vector function j(t, x) = (j1 (t, x), j2 (t, x), j3 (t, x)) of the first line of sys-
tem (1.11) represents the electric current density, and function ρ(t, x) of the
third line represents the electric charge density. Physical experiments have
never discovered elementary particles that carry a magnetic charge analo-
gous to electrons, protons, and other particles that carry units of electric
charge. If there were such magnetic monopoles, then the other two equation
components of (1.11) would have the analogous magnetic charge and current
densities.
Proposition 1.1. In the case that ρ = 0 and j = 0 (the conditions for
electromagnetic wave propagation in a vacuum), Maxwell’s equations (1.11)
are equivalent to wave equations for the components of the electric and the
magnetic fields.

Proof. In the case that ρ = 0 and j = 0, one differentiates equations (1.11)

to find
∂t2 E = ∂t (∇ × B) = ∇ × (∂t B)
= −∇ × (∇ × E) = −(−ΔE + ∇(∇ · E))
= ΔE .
This is to say that each component of the electric field Ej , j = 1, 2, 3 indi-
vidually satisfies the wave equation (1.9). The second line of this calculation
uses a vector calculus identity and the fact that in a vacuum ρ = 0, and
therefore the third line of Maxwell’s equations is that ∇ · E = 0. There is a
similar calculation for the components of the magnetic field. 

The three equations above, namely the wave equation, the heat equa-
tion, and Laplace’s equation, are the standard second-order linear constant
coefficient PDEs. Most of the elementary theory of PDEs is developed to
describe properties of their solutions. However, this list is far from being
exhaustive; many other forms of equations, both linear and nonlinear, occur
frequently and are of great interest and importance.

Monge–Ampère equations. This equation arises when the gradient of a

potential function f (x) is used to construct a mapping:
u(x) = ∇f (x) : Rn → Rn .
If the mapping is constrained to be volume preserving, then f (x) satisfies
the Monge–Ampère equation
(1.12) det(∂xj ∂x f (x)) = 1 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6 1. Introduction

Similar equations play a role in the construction of Calabi–Yau metrics on

manifolds. The matrix of second partial derivatives of a function f (x) is
known as the Hessian of f , and (1.12) is a fully nonlinear equation for f
involving the determinant of the Hessian of f . If the function f is convex,
then this is an elliptic equation. We note in comparison that the Laplacian
is the trace of the Hessian

Δf = tr(∂xj ∂x f ) .

Schrödinger’s equation. The complex valued function ψ(t, x) = X(t, x)+

iY (t, x) is called the wave function of quantum mechanics when it solves the
equation
1
(1.13) i∂t ψ = − Δψ + V (x)ψ .
2
This is a linear equation. The lower-order coefficient V (x) is called the
potential term; it serves to represent the environment in which the quantum
particle evolves. A commonly occurring nonlinear Schrödinger equation is
1
(1.14) i∂t Ψ = − ΔΨ + c|Ψ|2 Ψ ,
2
where the local intensity |Ψ|2 of the wave function creates its own potential.
This nonlinearly self-interacting potential has the effect of focusing the wave
function when c < 0 and defocusing it when c > 0.

Einstein’s equations. In general relativity our space-time is a manifold

M 4 with a metric tensor g = (gμν (x)) with Lorentzian signature. The Rie-
mann curvature tensor (Rijk ) is a certain nonlinear function of the metric
coefficients and their derivatives, Rijk (g, ∂xj g, ∂xj x g), as is the Ricci cur-
vature tensor (Rμν ). Einstein’s equations in a vacuum consist of the system
of PDEs
1
Rμν − Rgμν = 0 ,
2
 μν is the scalar curvature tensor. The
where R = trRμν = μν Rμν g
Einstein summation convention allows us to omit the summation symbol
over repeated indices in the latter expression, so it can be represented as
R = Rμν g μν . When energy and/or matter are present, Einstein’s equations
become
1
(1.15) Rμν − Rgμν = Tμν ,
2
where T is the energy-momentum tensor.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 1.2. Examples 7

Minimal surfaces. Given a mapping u : D → RN defined on a domain

D, the surface area of the image is
 
A(u) := det(∂xj u · ∂x u) dx .
D

In the case that the mapping u is given as a graph of a function f (x) over
D ⊆ Rn (where N = n + 1), namely that u(x) = (x1 , . . . , xn , f (x)), then
 
A(f ) = 1 + (∂x f )2 dx .
D

A minimal surface is the result of a mapping for which the area functional
A(u) is minimal (or at least a critical point), generally with given additional
constraints such as boundary conditions. That is to say, the variations of A
about u all vanish; namely,
d 
0 = δA(u) · v :=  A(u + τ v)
dτ τ =0
for all variations v(x) such that u + τ v continues to satisfy the constraints.
For the case where n = 2 and for a graph, if f is such a critical point, then
it must satisfy the PDE

(1.16) (1 + (∂x2 f )2 )∂x21 f + (1 + (∂x1 f )2 )∂x22 f − 2∂x1 f ∂x2 f ∂x1 ∂x2 f = 0 .

x3

x1

Figure 2. A minimal surface that is a graph in R3 whose boundary lies

on a specified curve.

An example of a minimal surface is given in Figure 2, where n = 2 and

N = 3. It is a certain portion of a Scherk surface, given by the equation
x3 = log( cos(x2)
cos(x1 ) ).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8 1. Introduction

The Navier–Stokes equations. In Eulerian coordinates one describes a

fluid in motion in a region D ⊆ Rn by its velocity vector field u(t, x) =
(uj (t, x))nj=1 at each point of D and at each time t. An incompressible but
viscous fluid will satisfy the system of equations
(1.17) ∂t u + (u · ∇)u + ∇p = νΔu ,
∇·u=0 .
The pressure p(t, x) can be thought of as the extra degrees of freedom that
allow the flow determined by (1.17) to satisfy the constraint of incompress-
ibility ∇ · u = 0. The vorticity of the fluid is given by ω(t, x) := ∇ × u.
When n ≥ 3, it is an open question whether every solution of (1.17), even
if smooth initial data is prescribed, will remain smooth at all future times.

Exercises: Chapter 1
Exercise 1.1. Prove that statement (1.1) characterizing conformal map-
pings implies that either (1.2) or (1.3) holds.
Exercise 1.2. Determine the class of conformal mappings in the case that
n = N ≥ 3. The conditions (1.1) are much more rigid than in two dimen-
sions.
Exercise 1.3. Prove that the Laplacian (1.6) is invariant under Euclidean
symmetries (1.7).
Exercise 1.4. In dimension n = 2, some elementary harmonic functions
are polynomials. They include the polynomials of degrees one and two:
u(x) = xj , j = 1, 2 ,
u(x) = x21 − x22 , u(x) = x1 x2 .
How many independent harmonic polynomials are there of general degree
?
When the dimension n ≥ 3, how many independent harmonic polyno-
mials are there of degree ?
Exercise 1.5. Prove the vector calculus identity
∇ × (∇ × V ) = −ΔV + ∇(∇ · V )
for a vector field V (x), where in our notation the gradient of a function F is
∇F = (∂x1 F, . . . , ∂xn F ) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/02

Chapter 2

Wave equations

Windblown ocean waves at sunrise. Image credit: Lindsay imagery /

iStock / Getty Images Plus.

We will start the topic of PDEs and their solutions with a discussion of a
class of wave equations, initially with several transport equations, which will
lead us to the study of the standard second-order wave equation (1.9) in one
space dimension. These transport equations are often called first-order wave
equations for reasons that will become clear as we describe their solutions.
We will use this study of wave equations as motivation to introduce the
Fourier transform as well as to give examples of the concept of a field of
characteristic curves.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10 2. Wave equations

2.1. Transport equations: The Fourier transform

The transport equation is a first-order equation that describes the time
evolution of a quantity u, carried by a background flow field that possesses
a (possibly) variable velocity c(t, x), namely

(2.1) ∂t u + c(t, x) · ∂x u = 0 .

This typically gives rise to an initial value problem where one asks that

u(0, x) = f (x) .

The function f (x) is called the initial data. The problem is to describe how
the solution u(t, x) evolves given its initial “shape” f (x) at t = 0. Often
these equations are known as first-order wave equations for the unknown
function u(t, x). The coefficient c(t, x) is a vector known as the “speed of
propagation”; we will see the reason for this when we describe the solution
process.
A first example is given by the case in which the coefficient c is constant.
We will see in the solution why c is known as the speed of propagation. We
will first solve this equation using the Fourier transform. For convenience
we restrict ourselves to one space dimension. The process is elementary, but
nevertheless it serves to introduce a surprisingly powerful technique for the
study of PDEs.

Definition 2.1. The Fourier transform of a function f (x) is given by

 +∞
1
(2.2) f (ξ) = √
ˆ e−iξy f (y) dy := F f (ξ) .
2π −∞

One aspect of the power of the Fourier transform is its invertibility.

Theorem 2.2 (Fourier inversion formula). Under suitable hypotheses on

f (x), there is an inversion formula for the Fourier transform
 +∞
1
f (x) = √ eiξx fˆ(ξ) dξ .
2π −∞

At a later point we will discuss the hypotheses under which this theorem
holds in much more detail; for now a perfectly reasonable condition would
be that f ∈ L2 (R), that is,
 +∞ 1
2
|f (x)| dx := f L2 (R) < +∞ .
2
−∞

The Fourier transform is well adapted for representing derivatives.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.1. Transport equations: The Fourier transform 11

Proposition 2.3. Under further suitable hypotheses,

 +∞  +∞
1 1
∂x f (x) = √ ∂x eiξx fˆ(ξ) dξ = √ ∂x (eiξx )fˆ(ξ) dξ
2π −∞ 2π −∞
 +∞
1
=√ eiξx iξ fˆ(ξ) dξ .
2π −∞
That is,
(2.3) ∂x f (ξ) = iξ fˆ(ξ) .

The suitable hypotheses mentioned above will be needed in order to

justify the Fourier inversion formula and the exchange of the derivative ∂x
with the integral in the above calculation.
Let us assume that the propagation speed c in (2.1) is constant and that
the solution u(t, x) satisfies the required hypotheses of Proposition 2.3 at
each time t (to ultimately have a rigorous argument, this assumption must
be verified, but for the present we will not do so). Then in order to be a
solution of (2.1), the function
 +∞
1
u(t, x) = √ eiξx û(t, ξ) dξ
2π −∞
must satisfy
 +∞
1
(∂t + c∂x )u(t, x) = √ eiξx (∂t û + ciξ û)(t, ξ) dξ = 0 .
2π −∞
Since the Fourier transform is invertible, this means that for all ξ ∈ R1 ,
∂t û(t, ξ) + icξ û(t, ξ) = 0 ,
and this is a straightforward ordinary differential equation (ODE) in time
for each value of the Fourier transform variable ξ:
∂t v(t) + icξv(t) = 0
with solution
v(t) = e−icξt v(0) = e−icξt fˆ(ξ) .
Therefore an expression for our solution is
 +∞  +∞
1 1
u(t, x) = √ e û(t, ξ) dξ = √
iξx
eiξx e−icξt fˆ(ξ) dξ
2π −∞ 2π −∞
 +∞
1
=√ eiξ(x−ct) fˆ(ξ) dξ = f (x − ct) .
2π −∞
The Fourier inversion formula is used in the last line. A picture of the
evolution of a solution is given in Figure 1.
You could ask whether the identical procedure using the Fourier trans-
form works for the case of a variable speed of propagation c(t, x); it does

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 12 2. Wave equations

(x0+ct, t)

x0 x

Figure 1. This is an example of the evolution of a solution of a trans-

port equation.

not in general, nor for a time-independent speed c(x). However, when c(t) is
independent of x, the answer is “yes” and a similar solution representation
is possible. This is given as an exercise at the end of the section.

2.2. Transport equations: The method of characteristics

There is a second method for solving equations of the form (2.1) which
doesn’t involve the Fourier transform. Again for simplicity we will take
the case that x ∈ R1 . Let f ∈ C 1 (R1 ), the space of once continuously
differentiable functions, and examine the equation

(2.4) (∂t + c∂x )u = 0 ,

again with initial data given by

u(0, x) = f (x) .

This is the statement that the directional derivative of u(t, x) in the (1, c)
direction in the (t, x) plane vanishes:

(1, c) · (∂t , ∂x )u = 0 .

That is to say, u(t, x) is constant along lines with tangent direction (1, c).
Therefore for all lines defined by x, β such that (x − β) = ct, we have
u(t, x) = u(0, β) = f (β) = f (x − ct). Verifying via the chain rule that
this expression satisfies the transport equation (2.4) for given initial data
f ∈ C 1 (R1 ),

∂t u(t, x) = ∂t f (x − ct) = −cf  (x − ct) ,

c∂x u(t, x) = c∂x f (x − ct) = cf  (x − ct) ,

and the sum of the two terms cancels. The lines in (t, x) space defined by
(x − β) = ct, exhibited in Figure 2, are the characteristic curves for this
equation.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.2. Transport equations: The method of characteristics 13

c
(c, 1)

(x, t)

x
x = x0 (x-ct)

Figure 2. Geometric interpretation of a transport equation.

This observation extends to give a solution method for the general trans-
port equation, which uses ODEs and is known as the method of character-
istics. Consider the equation
(2.5) ∂t u + c(t, x)∂x u = 0 ,
u(0, x) = f (x) ,
where the speed of propagation c(t, x) is allowed to vary. We will assume
that c(t, x) ∈ C 1 . Interpret equation (2.5) in terms of directional derivatives
(1, c(t, x)) · (∂t , ∂x )u = 0 ,
which in turn is the statement that u(t, x) is constant along integral curves
of the vector field X = (1, c(t, x)) in the (t, x) plane:
(2.6) X u(t, x) = (∂t + c(t, x)∂x )u = 0 .
The characteristic equations for the transport equation (2.5) are
dt dx
(2.7) =1, = c(t, x) ,
ds ds
with initial conditions
(2.8) x(0) = β
whose solutions are the integral curves for X; these are the characteristic
curves for equation (2.5). This family of curves (t(s), x(s)) = (s, x(s, β)) is
parametrized by the points β ∈ R1 at which the curves intersect the initial
surface {t = 0}. A family of characteristic curves is illustrated in Figure 3.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 14 2. Wave equations

(c (x, t), 1)

(x (t, ), t)

Figure 3. Characteristic curves of the vector field (2.6).

The equation states that a solution is constant along characteristic curves.

Therefore, given (t, x), we define u(t, x) = f (β), where β = β(t, x) is the
point of intersection with the initial surface of the characteristic curve which
passes through (t, x). This involves the task of finding the point β such
that the characteristic curve starting at β will pass through (t, x), basically
following the characteristic curve through (t, x) backwards to the x-axis
{t = 0}. To verify that our construction satisfies PDE (2.5), take into
account that our proposed solution u = u(t(s), x(s)) is constant along a
characteristic curve, and calculate using the chain rule:
d
0= u(t(s), x(s))
ds
dt dx
= ∂t u + ∂x u = ∂t u + c(t, x) ∂x u .
ds ds
These equations are considered to be wave equations because they de-
scribe the propagation in time of information specified by the initial data
f (x) along the characteristics, at a speed given by the coefficient c(t, x). Two
examples of transport equations with differing global behavior are given by
the following:
(2.9) ∂t u + x∂x u = 0 ,
and
(2.10) ∂ t u + x2 ∂ x u = 0 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.3. Conservation laws 15

Exercise 2.3 asks to solve these two equations, giving a solution representa-
tion for the initial value problem u(0, x) = f (x) using explicit expressions
for the respective functions β(t, x).

2.3. Conservation laws

A conservation law is a first-order PDE similar to a transport equation. To
be precise we are looking at linear scalar conservation laws; these take the
form of an initial value problem
(2.11) ∂t v + ∂x c(t, x)v = 0 , v(0, x) = f (x) .
If we differentiate out the second term, the equation becomes
∂t v + c(t, x)∂x v + ∂x c(t, x)v = 0 ,
which is a transport equation with an extra term. A related variant of
a transport equation appears in Exercise 2.2. Equation (2.11) is called a
conservation law because if v(t, x) ∈ C 1 is a solution of the equation, then
the integral of v(t, x) between any two characteristic curves (t, x1 (t)) and
(t, x2 (t)) is a constant.
Proposition 2.4. If v(t, x) ∈ C 1 solves (2.11), and (t, x1 (t)) and (t, x2 (t))
are two characteristic curves, then the quantity
 x2 (t)
v(t, y) dy
x1 (t)
is independent of t. In particular it can be determined by the initial data
 x2 (t)  x2 (0)
v(t, y) dy = f (y) dy .
x1 (t) x1 (0)

Proof. Recall that the characteristic curves (t, x1 (t)) and (t, x2 (t)) satisfy
the characteristic equations (2.7) with initial conditions x1 (0) = β1 and
x2 (0) = β2 . The two curves do not intersect (as long as β1 = β2 ) because
solutions of ODEs such as the characteristic vector field are unique. Differ-
entiate the quantity in question with respect to time:
(2.12)
 x2 (t)  x2 (t)
dx2 dx1
∂t v(t, y) dy = ∂t v(t, y) dy + v(t, x2 (t)) − v(t, x1 (t)) .
x1 (t) x1 (t) dt dt
Now use the fact that x1 (t) and x2 (t) satisfy the characteristic equations:
dx2 dx1
v(t, x2 (t)) − v(t, x1 (t)) = v(t, x2 (t))c(t, x2 (t)) − v(t, x1 (t))c(t, x1 (t))
dt dt
 x2 (t)
= ∂x c(t, y)v(t, y) dy .
x1 (t)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 16 2. Wave equations

Using this calculation in equation (2.12), we have

 x2 (t)  x2 (t)
∂t v(t, y) dy = ∂t v(t, y) + ∂x c(t, y)v(t, y) dy = 0 .
x1 (t) x1 (t)


Conservation laws of the form (2.11) describe the advection of a quantity

v(t, x) in time, such as a pollutant or a tracer, as it is carried by a back-
ground flow field given by c(t, x). In this setting the function v(t, x) is called
the density of this quantity, while c(t, x)v(t, x) is called the quantity’s flux.
Proposition 2.4 is the statement that the amount of quantity is conserved.

2.4. The d’Alembert formula

We now take up the study of the second-order wave equation, starting with
the one space-dimensional case. In one dimension there is a particularly sim-
ple and explicit formula for solutions of the initial value problem, sometimes
called the Cauchy problem:
(2.13) ∂t2 u − ∂x2 u = 0 ,
u(0, x) = f (x) , ∂t u(0, x) = g(x) .
The initial data comprises the two functions f and g that are to be specified.
Introduce new coordinates
x + t := r , x − t := s ,
so that the differential operators are transformed as follows:
∂t − ∂x = 2∂r , ∂t + ∂x = −2∂s .
Then equation (2.13) is transformed to
(2.14) (∂t + ∂x )(∂t − ∂x )u = −4∂s ∂r u = 0 .
The meaning of equation (2.13) is now interpreted to be that ∂s (∂r u) = 0,
which is to say that ∂r u is independent of the variable s. That is,
∂r u = v(r)
for some function v(r). Therefore, using W = W (s) as an integration con-
stant in the r-integral,
 r
u(r, s) = v(r ) dr + W (s) := V (r) + W (s) .

Returning to the original variables (t, x), the solution takes the form
(2.15) u(t, x) = V (x + t) + W (x − t) ,
which has the interpretation of a decomposition of the solution u(t, x) into
a left-moving component V and a right-moving component W . That is, any

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.4. The d’Alembert formula 17

solution of the one-dimensional wave equation (2.13) can be decomposed into

the sum of a left-moving wave V (x + t) and a right-moving wave W (x − t).
The components V and W are to be expressed in terms of the initial
data u(0, x) = f (x) and ∂t u(0, x) = g(x). In order to do this, evaluate the
expression (2.15) of the decomposition on the surface t = 0:
u(0, x) = V (x) + W (x) = f (x) ,
∂t u(0, x) = V  (x) − W  (x) = g(x) .

Differentiating the first line with respect to x and solving for V  and W  , we
find
1 1 1 1
V  (x) = f  (x) + g(x) , W  (x) = f  (x) − g(x) ;
2 2 2 2
thus

1 1 x
V (x) = f (x) + g(x ) dx + c1 ,
2 2 0

1 1 x
W (x) = f (x) − g(x ) dx + c2 ,
2 2 0
where necessarily c1 + c2 = 0. This is to say that
u(t, x) = V (x + t) + W (x − t)
 x+t
1 1
(2.16) = f (x + t) + f (x − t) + g(x ) dx .
2 2 x−t

This is known as d’Alembert’s formula for solutions of the one-dimensional

wave equation. We note that the wave equation is a second-order PDE;
however, (2.16) is well defined for data f, g ∈ C(R1 ) (continuous functions)
and gives a certain sense of a solution even though the resulting expression
for u(t, x) is not necessarily C 2 . It gives a sense of a weak solution to the
wave equation in this case, a topic that will be explored more deeply at a
later point.
The solution formula (2.16) shows how the solution u at a given space-
time point depends upon the initial data. Indeed, at the point (t, x) the
solution u(t, x) depends upon data (f (x ), g(x )) only over the interval x−t ≤
x ≤ x + t, and in particular it depends upon f at the two points x ± t, and
on g(x ) for x − t ≤ x ≤ x + t. The set of space-time points (t, x) such
that {x + t = x } ∪ {x − t = x } is called the light cone over x . Given x,
the region {x − t = x} ∪ {x + t = x} is called the backwards light cone of
x. Following the d’Alembert formula, the solution u(t, x) of (2.13) depends
upon the initial data (f (x), g(x)) only for values of x where the interior
of the backwards light cone intersects the x-axis {t = 0}. This property is
illustrated in Figure 4.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 18 2. Wave equations

(x, t)

x
x -t x +t

Figure 4. Solution of (2.13) using the d’Alembert formula (2.16), ex-

hibiting the dependence on the initial data.

This is an instance of solutions satisfying the principle of finite propaga-

tion speed. Defining the support of a function to be

supp(h) := {x ∈ R1 : h(x) = 0} ,

and taking both (f, g) to have compact support, then if supp(f ), supp(g) ⊆
[−R, +R], then the support of the solution u(t, x) satisfies supp(u(t, x)) ⊆
[−R − |t|, R + |t|], as can be seen in Figure 5. At time t the solution vanishes
outside of the region {−R−|t| ≤ x ≤ R+|t|}, which is to say that information
given by the initial data in the interval [−R, +R] does not travel faster
than speed c = 1 and therefore at time t > 0 has not propagated farther
than [−(R + t), (R + t)]. This property of finite propagation speed is called
Huygens’ principle and is a common feature of solutions of wave equations.
Furthermore, it is evident from the d’Alembert formula and Figure 5 that
the solution u(t, x) is constant inside the union of light cones emanating from
the support of the initial data. Indeed, for t > R and then for |x| < t − R,
 
1 x+t 1 +R
u(t, x) = g(x ) dx = g(x ) dx
2 x−t 2 −R
which is a constant independent of both (t, x). This property is called the
strong Huygens’ principle; it holds for solutions of the wave equation (1.9) in
odd space dimensions, but not in even space dimensions nor for most other
wave equations.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.5. Duhamel’s principle 19

(x, t)

x
x-t -R R x+t

Figure 5. The d’Alembert formula (2.16) exhibits the principle of finite

propagation speed and the strong Huygens’ principle.

2.5. Duhamel’s principle

In addition to the initial value problem for a PDE, it is also common to be
given an inhomogeneous problem in the form
∂t2 u − ∂x2 u = h(t, x) ,
(2.17) u(0, x) = f (x) , ∂t u(0, x) = g(x) .
By a simple decomposition of the solution of this linear equation into two
components u(t, x) = u1 (t, x) + u2 (t, x), we may reduce problem (2.17) to
solving for u1 (t, x) given f (x) and g(x), with h(t, x) = 0, and then solving
for u2 (t, x) for which f = g = 0 and h(t, x) is nonzero. A solution for u1 (t, x)
is given by the d’Alembert formula (2.16).
Duhamel’s principle gives a method to solve for u2 (t, x) that is based on
a general principle. For the wave equation we first introduce the notion of a
standard form for the initial value problem. Namely, consider the problem
above,
(2.18) ∂t2 u − ∂x2 u = 0 ,
u(0, x) = 0 , ∂t u(0, x) = g(x) ,
that is, with f = h = 0. Using the d’Alembert formula, the solution is given
by
 x+t
W(t)(g)(x) := u(t, x) = g(x ) dx ,
x−t

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 20 2. Wave equations

which we are using to define the propagator, or solution operator W(t).

Using this operator notation, a formula for the solution to the general initial
value problem may be written

u(t, x) = ∂t W(t)f (x) + W(t)g(x) .

To check that this indeed solves the wave equation with the proper initial
data, one can either take the time derivative ∂t W(t)f (x) directly on the
integral expression for W(t), or check that for continuous f (x) and g(x),
the desired initial values are taken on:

∂t W(0)f (x) = f (x) , W(0)g(x) = 0 ,

∂t ∂t W(0)f (x) = ΔW(0)f (x) = 0 , ∂t W(0)g(x) = g(x) .

We can now express a solution to the inhomogeneous problem using the

propagator W(t). This is known as Duhamel’s principle:
 t
(2.19) u2 (t, x) = W(t − s)h(s, ·)(x) ds .
0

To verify that this is indeed the solution we seek, check the following:

u2 (0, x) = 0 ,
 t 

∂t u2 (t, x) = ∂t W(t − s)h(s, ·)(x) ds = W(0)h(t, ·)(x) = 0 ,
0 t=0
 t
2
∂t u2 (t, x) = ∂t2 W(t − s)h(s, ·)(x) ds + ∂t W(0)h(t, ·)(x)
0
= Δu2 (t, x) + h(t, x) .

In the case of the one-dimensional wave equation, using the explicit form
of the operator W(t), the solution formula for the inhomogeneous problem
(2.17) is similar to the d’Alembert formula itself. Indeed,

u(t, x) = u1 (t, x) + u2 (t, x)


1 1 x+t
= f (x + t) + f (x − t) + g(x ) dx
2 2 x−t
 
1 t x+(t−s)
+ h(s, x ) dx ds .
2 0 x−(t−s)

The term with the double integral accounts for the component u2 (t, x) of
the solution; it is the integral of h(t, x) over the interior of the backwards
light cone centered at the point (t, x).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.6. The method of images 21

Duhamel’s principle stems from a formula of ODEs known as the varia-

tion of constants. To illustrate this, suppose we are to solve the inhomoge-
neous problem
dv
(2.20) = Av + h(t) , v(0) = g .
dt
When h = 0, the solution is given by
v(t) = etA g ;
indeed, etA is the analog of the wave propagator W(t). A solution can be
decomposed as follows: v(t) = v1 (t) + v2 (t) where
v1 (t) = etA g
is the solution to the initial value problem with h = 0. The variation of
constants formula gives the form of a solution to take the inhomogeneous
term into account:  t
v2 (t) = e(t−s)A h(s) ds .
0
It is straightforward to verify the following:
dv2
= Av2 + h , v2 (0) = 0 ;
dt
therefore the full solution v(t) = v1 (t) + v2 (t) satisfies (2.20).

2.6. The method of images

We have been discussing the initial value problem for the one-dimensional
wave equation with initial data given for x ∈ R1 . However, in many situa-
tions the spatial domain has boundaries on which the solution is normally
required to satisfy additional conditions. Such problems are called initial-
boundary value problems. For the one-dimensional problem, let us take the
case of the spatial domain R1+ := {x > 0} with the boundary at x = 0. Two
typical boundary conditions are
(2.21) u(t, 0) = 0 ,
which go by the name of Dirichlet conditions, and secondly
(2.22) ∂x u(t, 0) = 0 ,
which are called Neumann conditions. Of course, if the support of the initial
data does not include x = 0, by the property of finite propagation speed the
solution will satisfy both of these boundary conditions for a certain time
interval. However, even if so, this situation will not generally last for all
time t ∈ R1 .
The method of images is a classical technique used to solve the two
cases of boundary conditions, (2.21) and (2.22). It also works for some

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 22 2. Wave equations

other boundary conditions, which we will not pursue here. The method
as applied to the problem of Dirichlet boundary conditions is based on the
simple fact that a function that is odd, h(−x) = −h(x), and continuous at
zero must necessarily satisfy h(0) = 0. Since the spatial domain is R1+ , the
initial data f (x), g(x) are only defined for x > 0. Assume for simplicity that
both f, g ∈ C(R1+ ). We require that both f (0), g(0) = 0 for compatibility
between the initial and boundary values of the solution. The method of
images consists of defining extensions f (x) = −f (−x) and g(x) = −g(−x)
for x < 0, the odd reflections of f (x) and g(x) through the origin; these are
defined and continuous on all of x ∈ R1 . The d’Alembert formula gives a
solution to the Cauchy problem, namely

1 1 x+t
u(t, x) = f (x + t) + f (x − t) + g(x ) dx .
2 2 x−t
This expression is clearly an odd function of x, as the extended values of f
and g are odd. For any x > t, this expression only involves the arguments
of f, g for positive x; however, for x < t, the expression involves negative
arguments even though x > 0. To remedy this we use the fact that the
extended f, g are odd to rewrite the expression as

1 1 x+t
(2.23) u(t, x) = f (t + x) − f (t − x) + g(x ) dx .
2 2 t−x
The effect is that the solution depends only upon values of the initial data
(f (x), g(x)) for x > 0. Examining this formula, it is as though the solution
is reflected off of the boundary with a negative mirror image of the incident
signal, as in Figure 6.
In the case of Neumann boundary conditions at x = 0, ask that the
boundary data f, g ∈ C 1 (R1+ ) and that it satisfy the compatibility conditions
that
∂x f (0) = 0 = ∂x g(0) .
Then reflect f, g to be even functions of x ∈ R1 , which is possible because
an even function h(x) whose first derivative is continuous at x = 0 must
satisfy ∂x h(0) = 0. Using the even property in the d’Alembert formula, for
x < t,

1 1 x+t
u(t, x) = f (x + t) + f (x − t) + g(x ) dx
2 2 x−t
  t−x
1 1 t+x  
= f (t + x) + f (t − x) + g(x ) dx + g(x ) dx .
2 2 t−x 0

The reflection of the solution off of the boundary is now a positive mirror

image of the incident signal, as illustrated in Figure 7. There is also a

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.6. The method of images 23

Figure 6. The method of images for Dirichlet boundary conditions.

component of the solution coming from the continuing reflection off the

boundary of the contribution of g.
One further case is where nonzero boundary conditions are imposed
along the t-axis; for example, one might specify the inhomogeneous data
(2.24) u(t, 0) = h(t)
along with the initial conditions
u(0, x) = f (x) , ∂t u(0, x) = g(x) .

The wave equation is linear so that a sum of two solutions is also a solution
of the PDEs. Therefore we may take the solution to be a sum of two compo-
nents: u1 = u1 (t, x) which solves (2.13) with zero boundary data (2.21), and
a second solution u2 (t, x) which has homogeneous initial data f = g = 0 but
satisfies the requested boundary conditions (2.24). The first component u1
can be expressed as above in (2.23). The second component can be taken to
vanish for t < x as the backwards light cone does not intersect the boundary
for positive times t. However, for t > x define
u2 (t, x) = h(t − x) ,
which satisfies both the wave equation (2.13) with f (x) = 0 = g(x) and the
boundary conditions (2.21). The full solution is then u(t, x) = u1 (t, x) +
u2 (t, x). The initial and boundary data for this solution have their own
compatibility conditions to satisfy. Namely, they must have h(0) = f (0)
to avoid discontinuities in u(t, x) being introduced at u(0, 0), then also

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 24 2. Wave equations

Figure 7. The method of images for Neumann boundary conditions.

∂t h(0) = g(0) = ∂t u(0, 0), and finally ∂t2 h(0) = ∂t2 u(0, 0) = ∂x2 u(0, 0) =
∂x2 f (0), all to avoid incompatibilities in the first and second derivatives of
the solution at (t, x) = (0, 0).

2.7. Separation of variables

The method of separation of variables is a time-honored technique that,
if one’s information comes from the standard undergraduate mathematics
textbook on PDEs, constitutes the entirety of the subject. In this section we
will attempt to bring the topic to life by presenting it in several compelling
contexts, working with the example of the wave equation. Other applications
follow with straightforward modifications.

Wave equation. As a first example consider the wave equation

(2.25) ∂t2 u − ∂x2 u = 0 , x ∈ (0, 2π)

as in (2.13), with Dirichlet boundary counditions

(2.26) u(t, 0) = 0 = u(t, 2π) .

The premise is to seek special solutions in the form of a product u(t, x) =

v(t)w(x) so that

(2.27) ∂t2 v w = v ∂x2 w ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.7. Separation of variables 25

which implies that there is a common constant λ (ignoring possible zeros of

v and w) such that
∂t2 v ∂ 2w
=λ= x .
v w
Here λ is necessarily a constant because the LHS depends only upon t while
the RHS depends only on x. Thus the solution form (2.27) decomposes
the problem into ODEs: an evolution problem for v(t) in order to satisfy
the initial conditions, and an eigenvalue problem for an ordinary differential
operator for w(x) so that the solution satisfies the boundary conditions.
The eigenvalue problem is for the pair (w(x), λ), where the function w(x)
satisfies Dirichlet boundary conditions on x ∈ (0, 2π):
(2.28) ∂x2 w = λw , w(0) = 0 = w(2π) .
Given an eigenfunction and an eigenvalue from this problem, the evolution
equation for v is
d2
v = λv ,
dt2
with initial data v(0) = a, v̇(0) = b to be specified. The eigenvalue problem
(2.28) has a countable number of solutions, which in the case above are
1 1 k2
wk (x) = √ sin( kx) , λk = − , k = 1, 2, . . . .
π 2 4
Thus we may choose countably many constants (ak , bk ), k = 1, 2, . . . , such
that
sin(ωk t)
vk (t) = ak cos(ωk t) + bk
ωk
√
with temporal frequencies ωk = −λk = 2 k. For each k ∈ N, this consti-
1

tutes a two-parameter family of such special solutions. Linear combinations

of these special solutions solve the initial value problem (2.13) with general
initial data and satisfy Dirichlet boundary conditions (2.26), as the following
theorem attests.
Theorem 2.5. The general solution of the initial-boundary value problem
(2.25) is a linear superposition of these special solutions:

+∞
sin(ωk t) 1 1
u(t, x) = ak cos(ωk t) + bk √ sin( kx) ,
ωk π 2
k=1
where initial data is specified by

+∞
1 1  +∞
(2.29) u(0, x) = f (x) = ak √ sin( kx) = ak wk (x) ,
π 2
k=1 k=1

+∞
1 1 
+∞
(2.30) ∂t u(0, x) = g(x) = bk √ sin( kx) = bk wk (x) .
π 2
k=1 k=1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 26 2. Wave equations

Furthermore, the sine series used to describe f (x), g(x) is complete as an

orthonormal basis of the Hilbert space L2 (0, 2π); thus any initial f (x), g(x) ∈
L2 can be represented by their Fourier sine series coefficients (ak , bk )k∈N ,
giving rise to a solution u(t, x).

Given the desired initial data f (x), g(x), the remaining task is to cal-
culate the coefficients ak of their eigenfunction expansion (2.29) for f and,
respectively, bk in the expansion (2.30) for g. It is here that the utility of
an orthonormal set of basis elements becomes clear. Using our example of
the sine series wk (x) = √1π sin( k2 x), one calculates that
 2π 
1 2π k
wk (x)w (x) dx = sin( x) sin( x) dx = 0 for k = ,
0 π 0 2 2
 2π  2π
1 k
wk2 (x) dx = sin2 ( x) dx = 1 .
0 π 0 2
Denote the inner product between two functions w(x) and z(x) (in the
Hilbert space L2 ([0, 2π])) by
 2π
w, z := w(x)z(x) dx .
0
Then the expressions above for the various integrals between the eigenfunc-
tions wk (x) are manifestations of the fact that they form an orthonormal set
in L2 ([0, 2π]), namely wk , w  = 0 for k = and wk , wk  := wk 2L2 = 1.
Theorem 2.5 also asserts that this set forms a complete set, or basis, for the
Hilbert space, an important result from the theory of Fourier series (again
a proof that we defer for the sake of continuity of the presentation).
To calculate the coefficients ak , bk of the sine series for f and g, take the
inner product of (2.29) and (2.30), respectively, with the basis element wk ,
and using the orthonormality relations,

+∞
f, wk  =  a w , wk  = ak ,
=1

+∞
g, wk  =  b w , wk  = bk .
=1

The coefficients {ak }∞

are known as the Fourier sine series for the function
k=1
f , and {bk }∞
k=1 the Fourier sine series for g.

Quantum mechanics. There are a number of fundamental problems in

mathematical physics, particularly in quantum mechanics, that have a cen-
tral role in the foundations of condensed matter theory and quantum chem-
istry. The Schrödinger equation (1.13) for the quantum mechanical wave

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.7. Separation of variables 27

function ψ(t, x) is one of our examples in Chapter 1. The connection be-

tween the evolution problem given in terms of the Schrödinger equation and
the spectral problem presented by the Schrödinger operator is made through
separation of variables. Indeed, writing
ψ(t, x) = e(t)ϕ(x) ,
the spatial dependence of the wave function is then described by a spectral
problem for an ordinary or partial differential operator for the complex val-
ued function ϕ(x), along with appropriate boundary conditions. The time
evolution of this special solution is given by
e(t) = e−iλt e(0)
for λ in the spectrum of this operator. Two settings that continue to play
central roles in the physical sciences are the quantum harmonic oscillator
and the hydrogen atom. The statement that the time evolution for arbitrary
initial data ψ(0, x) ∈ L2 is captured in terms of the spectral problem is a
basic fact that is related to the completeness of spectral representations.
Theorem 2.6. Let A be a self-adjoint operator on a Hilbert space H. Then
its spectrum is contained in R and its spectral resolution is complete. If for
some λ ∈ C the resolvant operator (λI − A) is compact, then completeness
means that there is a complete set of orthonormal eigenfunctions {ϕk ∈ H}
whose eigenvalues are real:
Aϕk = λk ϕk , λk ∈ R ,
ϕk 2H =1, ϕk , ϕ H = 0 for k= .
The key point to using this theorem is to show that the operator A
is self-adjoint. For operators that are bounded on H, this is equivalent to
being Hermitian symmetric:
ϕ1 , Aϕ2 H = Aϕ1 , ϕ2 H .
In the case of unbounded operators including differential operators such as
d2
A = dx 2 , self-adjointness depends upon the operator and boundary condi-
d2
tions. For example, for dx 2 , one requires boundary conditions so that A
is Hermitian symmetric, and one also needs to produce a dense subspace
DA ⊆ H of elements of H which satisfy the boundary conditions, with the
property that the graph (DA , A(DA )) ⊆ H × H is closed. We will not give
a proof of this theorem, as it lies outside the main focus of the course.
Quantum harmonic oscillator. For x ∈ R1 , the quantum harmonic oscil-
lator describes the quantum mechanical wave functions that correspond to
electron states that are confined by a quadratic potential well V (x) = 12 x2 :
1 1
i∂t ψ = − ∂x2 ψ + x2 ψ .
2 2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 28 2. Wave equations

By the method of separation of variables, the time evolution can be reduced

to be
ψ(t, x) = e−iλk t ϕk (x) ,
where the boundary value problem for ϕk (x) is the spectral problem
1 d2 1
− ϕk + x2 ϕk = λk ϕk , x ∈ R1 , ψk ∈ L2 (R1 ) = H .
2 dx2 2
The eigenfunctions of this Schrödinger operator are the classical Hermite
functions {hk (x)}k∈N , whose eigenvalues are given by λk = 12 (k + 1), k =
0, 1, . . . . The eigenfunction corresponding to the lowest eigenvalue is the
ground state, in this case
1
h0 (x) = c0 e− 2 x ,
1 2
λ0 = .
2

Hydrogen atom. In this case x ∈ R3 , and one solves for the orbitals of
electrons in the presence of the attractive Coulomb potential of the atomic
nucleus. The Schrödinger equation describing this situation is
1 Z
(2.31) i∂t ψ = − Δ2 ψ − ψ.
2 |x|
Again, to solve this problem we separate variables in time and space to
reduce the question to a spectral problem:
1 Z
(2.32) − Δ2 ϕ − ϕ = λϕ .
2 |x|

n = 1, l = 0 n = 2, l = 0 n = 3, l = 0

n = 2, l = 1 n = 3, l = 1

n = 3, l = 2

Figure 8. Cross section of probability densities |ϕnm (x)|2 of electron

orbitals of the hydrogen atom, for the first three electronic shells.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.7. Separation of variables 29

The eigenfunctions of this Schrödinger operator are the orbitals of electrons

bound to the atomic nucleus. Cross sections of the densities of the first
several orbitals are given in Figure 8. The spectral lines seen in atomic
spectroscopy correspond to the differences in energy between these eigen-
states. Besides the obvious scientific relevance of the spectrum and the
structure of the eigenstates, the interest in this spectral problem is that it
also is handled through separation of the spatial variables, for which one
encounters an interesting variety of classical special functions. Indeed, one
writes x = (r, φ, θ) ∈ R3 in polar coordinates, and then making the Ansatz
of a product decomposition
ϕk (x) = Rn (r)Φm (φ)Θm
 (θ) ,

one finds that Φm (φ) are given in terms of complex exponentials, while
Θm (θ) are associated Legendre functions of cos(θ), and the radial depen-
dence Rn (r) is given in terms of associated Laguerre functions. The op-
erator (2.32) has continuous spectrum for λ ≥ 0 and an infinite number of
Z2
discrete negative eigenvalues λn = − 2n 2 , n = 1, 2, . . . , corresponding to the
energies of bound electron orbitals.

Physics of music. In the hands of a good player, a musical instrument

is a machine that creates a sequence of vibrations in the air that we enjoy
as music. Different instruments produce these vibrations in different ways,
which can be studied by modeling them mathematically.
Strings. Stringed instruments, from violins and basses to guitars to pianos,
produce sounds from a bowed, plucked, or hammered string, which is a
taut elastic continuum that vibrates according to the wave equation (2.25)
(with units chosen so that the speed of propagation in the string is c =
1). The string is held fast at the bridge (x = 0) and at the nut or by a
musician’s finger on the fingerboard (at x = 2π); hence the solution satisfies
Dirichlet boundary conditions. According to our analysis of the initial-
boundary value problem for the wave equation above, the eigenmodes wk (x)
and eigenfrequencies ωk of the vibration of a string are given by
1 1  1
w1 (x) = √ sin( x) , ω1 = −λ1 = ,
π 2 2
1 
w2 (x) = √ sin(x) , ω2 = −λ2 = 1 . . .
π
1 1  1
wk (x) = √ sin( kx) , ωk = −λk = k .
π 2 2
The first five of these eigenmodes are graphed in Figure 9.
There is certainly significance in the fact that the kth frequency satisfies
ωk = kω1 , forming an arithmetic progression. When a string is played,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 30 2. Wave equations

0 2

Figure 9. The first five eigenmodes of a string. Blue = w1 (x), Orange

= w2 (x), Yellow = w3 (x), Purple = w4 (x), and Green = w5 (x).

the resulting note is a superposition of the eigenfrequencies ω1 , ω2 , . . . , with

some weights (ak , bk ) which generally decrease with increasing k. These are
the harmonics of the fundamental tone ω1 ; namely, ω2 = 2ω1 , etc. Part of
the skill of the musician is to control these weights, as our ears hear this
combination of these frequencies, giving the character of the sound of the
instrument. In fact, the diatonic scale of western music, as well as the
pentatonic scales of eastern music and the tone selections of many other
music traditions, are based very closely on these frequency relations; see
Table 1.

Table 1. Overtones of a string with their correspondences to the dia-

tonic scale.

overtone relation to fundamental common name

ω1 fundamental
ω2 2ω1 octave
ω3 3ω1 twelfth, or major fifth plus one octave
ω4 4ω1 second octave
ω5 5ω1 major third above second octave
ω6 6ω1 major fifth above second octave
ω7 7ω1 first anharmonic overtone
ω8 8ω1 third octave

One may check the fact that the hammers of a piano are designed to
strike the string at one-seventh the distance from the end, which tends to
suppress the presence of the anharmonic seventh overtone in the sound of
the note.
Woodwinds. The vibrations in an air column of a wind instrument obey the
Euler equations of fluid dynamics, which, in the case of the small air veloc-
ities and small pressure variations that hold in the body of the instrument,
can be assumed to satisfy the linearized equation for variations around the
state of constant ambient density ρ0 and zero velocity. This is the system

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.7. Separation of variables 31

of equations
(2.33) ∂t ρ + ρ0 ∂x v = 0 ,
ρ0 ∂t v + p (ρ0 )∂x ρ = 0
for small density variations ρ(t, x) and small air velocities v(t, x) in the air
column of the instrument. The pressure p(ρ) is determined as a function of
ρ by a constitutive law, given by the theory of thermodynamics. Taking an
additional time derivative of the first (or second) equation of (2.33) and then
using the second (or the first, respectively), we are led to the two essentially
equivalent wave equations
(2.34) ∂t2 v − p (ρ0 )∂x2 v = 0 ,
∂t2 ρ − p (ρ0 )∂x2 ρ = 0 ,
which differ through treatment at the boundaries x = 0, 2π. The wave
speed is c = p (ρ0 ), and we will take the liberty of choosing units so that,
conveniently, c = 1. Boundary conditions for ρ(t, x), respectively for v(t, x),
vary according to the design of the instrument.
The flute. This instrument has large holes at the mouthpiece and at the
end of the active part of the flute body, which effectively imposes that the
pressure of the air column in the flute coincides with atmospheric pressure
at these two boundary points; this gives boundary conditions
ρ(t, 0) = 0 = ρ(t, 2π) , equivalently ∂x v(t, 0) = 0 = ∂x v(t, 2π) .
Solving for the density ρ(t, x) using the method of separation of variables
gives us the eigenmodes wk (x) = π −1/2 sin( 12 kx), resulting in the expression
 sin(ωk t) 1 1
ρ(t, x) = ak cos(ωk t) + bk √ sin( kx) .
ωk π 2
k≥1

The frequencies are clearly ωk = 12 k in our chosen units, and evidently

ω2 = 2ω1 , ω3 = 3ω1 , etc. Typically only several overtones are present in a
well-played flute note (so that ak ∼ 0, bk ∼ 0 for k > 3 or so), reflecting
the purity and simplicity of the tone of the instrument. Overblowing the
flute to the second overtone to achieve the second register of the instrument
increases the tone by one octave.
The clarinet. The clarinet is roughly the same length as a flute and has
a similar cylindrical air column; however, the mouthpiece holds a reed in
position which essentially closes this end. This imposes Dirichlet boundary
conditions v(t, 0) = 0 for the velocity, while at the bell the boundary con-
ditions are ∂x v(t, 2π) = 0, the same as for the flute. Solving by separation
of variables for the velocity in this instance, we find that the fundamental
mode is w1 (x) = π −1/2 sin(x/4) with frequency ω1 = 1/4, while the higher

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 32 2. Wave equations

0 2

Figure 10. The first five eigenmodes of a clarinet. Blue = w1 (x),

Orange = w2 (x), Yellow = w3 (x), Purple = w4 (x), and Green =
w5 (x).

eigenmodes are wk (x) = π −1/2 sin((2k − 1)x/4) and the frequencies of the
overtones are ωk = 12 (k − 12 ). The first five eigenmodes are illustrated in
Figure 10. Evidently ω1 = 1/4, ω2 = 3/4 = 3ω1 , and ω3 = 5/4 = 5ω1 , etc.
We remark that all of the even multiples of the fundamental frequency are
absent and that the lowest note of a clarinet is about one octave lower than
that of a flute, while the instruments are about the same size. A clarinetist
overblowing to the second overtone to achieve the second register of the in-
strument increases the tone by one-twelfth, or one octave and a major fifth.
This absence of even overtones is an explanation for the composer’s rule
of thumb to write the clarinet parts above the string parts. Otherwise the
strings may be obliged to accord their notes with overtones that are absent
in the notes of the clarinet.
The oboe and the saxophone. Both the oboe and the saxophone are reed
instruments; the saxophone uses a single reed similar to the clarinet, while
the oboe has a famously difficult double reed arrangement. Both impose
boundary conditions at the mouthpiece and at the bell, similar to the clar-
inet. However, both overblow an octave rather than a twelfth. The answer
to this conundrum is that they are both conical bore instruments, unlike
the cylindrical bore clarinet, and must be analyzed in terms of a three-
dimensional rather than one-dimensional air column. Making the reason-
able assumption that the density variation in the air column (expressed in
polar coordinates) is ρ(t, r), namely independent of angles (θ, φ), the wave
equation (2.34) becomes
2
(2.35) ∂t2 ρ − p (ρ0 )(∂r2 + ∂r )ρ = 0
r
with boundary conditions ∂r ρ(t, 0) = 0 and ρ(t, 2π) = 0. Making the sub-
stitution R = rρ(t, r) (d’après Beals [3]) leads to the equation
∂t2 R − p (ρ0 )∂r2 R = 0 ,
satisfying boundary conditions R(t, 0) = 0 = R(t, 2π). This is now addressed
as before by separation of variables techniques.
 The fundamental is R1 (r) =
π −1/2 1 1 
sin( 2 r) with frequency ω1 = 2 p (ρ0 ) while, just as in the case of the

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 2.7. Separation of variables 33

0 2

Figure 11. The first five eigenmodes of a saxophone or oboe, expressed

in terms of the velocity v(r) (see Project 2.14). Blue = w1 (x), Orange
= w2 (x), Yellow = w3 (x), Purple = w4 (x), and Green = w5 (x).


stringed instruments or the flute, the higher eigenfrequencies are 12 p (ρ0 ) k.
Again we find that ω2 = 2ω1 , ω3 = 3ω1 , etc. Eigenmodes in terms of the
original variables are given by ρn (r) = Rnr(r) . Alternatively, the problem
can be phrased in terms of the velocity v(t, r); see Project 2.14 at the end
of this chapter. A similar analysis can be performed in terms of the velocity
variables v(r), giving eigenmodes wn (r) and the same sequence of eigenvalues
ωn ; this is the topic of Project 2.14.
The drum. The eigenmodes of a drum correspond to those of a two-
dimensional disk BA (0) = {x = (x1 , x2 ) : |x|2 < A2 }. By separation of vari-
ables one seeks solutions (wk (x), λk ) of the eigenvalue problem for Laplace’s
equation, which in polar coordinates is
1 1
(∂r2 + ∂r )wk + 2 ∂θ2 wk = λk wk , w(A, θ) = 0 .
r r
The angular dependence can be described in terms of complex exponentials
eiθ . However, the radial dependence involves Bessel functions whose√zeros
do not generally give rise to an arithmetic series of frequencies ωk = −λk ,
as is the case for the one-dimensional instruments described above. And as
we know, drums are quite noisy instruments.
Synthesizers have allowed musicians to create new instruments with the
sound of a note determined by a chosen overtone series. D. DeTurck has
used this idea to synthesize the sound of an imaginary piano-like instrument,
where the role of the strings is replaced by higher-dimensional geometrical
objects such as spheres, tori, or hyperbolic Riemann surfaces, and the sub-
sequent overtone series of each note is determined by the square roots of the

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 34 2. Wave equations

first number of Laplace–Beltrami eigenvalues. You can hear some of these

instruments in action on the website www.toroidalsnark.net/som.html.

Exercises: Chapter 2
Exercise 2.1. For the wave speed c(t) which is a general (locally integrable)
function of time, give a Fourier integral representation of the solution of the
transport equation (2.1).
Exercise 2.2. Find an explicit expression for the solution of the equation
∂t u + c∂x u + ru = 0 , u(0, x) = f (x) ,
using either (or both) of the above solution methods. When r > 0, what
happens to the solution u(t, x) asymptotically for t → +∞?
Exercise 2.3. Give a solution formula of the following equations using the
method of characteristics:
∂t u + x∂x u = 0 , u(0, x) = f (x) ;
∂ t u + x2 ∂ x u = 0 , u(0, x) = f (x) .
Then differentiate to check your answer. These two examples appear in
the text as (2.9) and (2.10). Explain the differences between the two cases
with regards to the behavior of solutions. Are the solutions u(t, x) unique
as functions defined over all of (t, x) ∈ R2 in both cases? What are the
asymptotic states
v(x) = lim u(t, x) ,
t→±∞
if the limit exists, that is?
Exercise 2.4. Express the solution to the following transport equation on
x ∈ R1 :
∂t u − xt∂x u = 0
for general initial data u(0, x) = f (x).
Suppose that the initial data f (x) has compact support, that is, it is
zero outside of an interval [−R, R]. For x = 0, what is the limit
lim u(t, x) ?
t→+∞

Exercise 2.5. Show that if the function u(t, x) solves the transport equation
(2.1), then the function v(t, x) = ∂x u(t, x) solves the conservation law (2.11).
Exercise 2.6. Give an explicit expression for the solution of the dispersive
equation
∂t u ± x∂x3 u = 0 , u(0, x) = f (x) ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 2 35

using the Fourier transform and then the method of characteristics [4]. Com-
ment on properties of the solution when the indicated choice of sign is +
and when it is −.

Exercise 2.7 (Conservation of energy). This problem concerns the wave

equation
∂t2 u − ∂x2 u = 0 .
Suppose that the initial data f (x) and g(x) for the wave equation satisfy
 +∞  +∞
|∂x f (x)|2 dx < ∞ , |g(x)|2 dx < ∞ .
−∞ −∞

Define the energy of a solution u(t, x) to be


1 +∞
E(u) = |∂t u(t, x)|2 + |∂x u(t, x)|2 dx .
2 −∞
Show that for u(t, x) ∈ C 2 ,
d
E(u(·, t)) = 0 ,
dt
namely that solutions conserve the energy functional E.

Exercise 2.8 (Equipartition of energy). Consider the wave equations as

in Exercise 2.7, and assume in addition that the functions f (x) and g(x)
are compactly supported. That is, there is an interval [−R, +R] such that
f (x) = 0 and g(x) = 0 for all |x| ≥ R.
Show that for t > R, the solution u(t, x) exhibits equipartition of energy.
Namely, for sufficiently large t,
 
1 +∞ 1 +∞
|∂t u(t, x)| dx =
2
|∂x u(t, x)|2 dx .
2 −∞ 2 −∞
Exercise 2.9 (Traveling waves). (i) Starting from the d’Alembert formula,
show that the solution of the wave equation

∂t2 u − ∂x2 u = 0 ,
u(0, x) = f (x) , ∂t u(0, x) = g(x)

can be decomposed into left-propagating and right-propagating components

u(t, x) = V (x + t) + W (x − t) .

(ii) Suppose that ∂x f (x) = −g(x). Show that the left-moving component of
the solution vanishes.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 36 2. Wave equations

Exercise 2.10 (Propagation of singularities). Use the d’Alembert formula

to express the solution of the wave equation, x ∈ R1 , with discontinuous
initial data as in the following two different cases of initial data.
Case (i):
f (x) = 1 for − 1 ≤ x ≤ +1
=0 for x < −1 and 1 < x ,
g(x) = 0 .
Where are the discontinuities of the solution u(t, x) in space-time (t, x) ∈ R2 ?
Case (ii):
f (x) = 0 ,
g(x) = 1 for − 1 ≤ x ≤ +1
=0 for x < −1 and 1 < x .
Is u(t, x) discontinuous anywhere? Is ∂x u(t, x) discontinuous, and where?
Exercise 2.11 (Periodic boundary conditions). Use the method of separa-
tion of variables to give an expression for the wave equation
∂t2 u − ∂x2 u = 0
on the circle T1 := {0 ≤ x < 2π}. That is, impose periodic boundary
conditions on your solution in the form
u(t, x + 2π) = u(t, x) .
You may assume that f (x), g(x) are 2π-periodic C 2 functions.
Exercise 2.12 (Impedence boundary conditions). Use the method of sep-
aration of variables to represent the solution of the wave equation
∂t2 u − ∂x2 u = 0
for 0 < x < π with the following boundary conditions:
u(t, 0) = 0 , α∂x u(t, π) − u(t, π) = 0 .
The parameter α can be any real number. The behavior of a solution can be
very different, depending upon α. In particular, it is different over the range
π < α than over α ≤ π. Describe this difference in terms of the stability of
solutions.
Exercise 2.13 (Bloch eigenfunction expansion). Consider the Schrödinger
equation (1.13) in one space dimension with x ∈ T1 = {0 ≤ x < 2π}, V = 0,
and Floquet boundary conditions for the wave function
ψ(t, x + 2π) = eiϑ ψ(t, x) ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 2 37

where ϑ is a parameter on the circle 0 ≤ ϑ < 2π. Use the method of sepa-
ration of variables to give an expression for the solution to the Schrödinger
equation in this setting.
In your analysis show that the eigenvalue problem
−∂x2 ϕk (x) = λk ϕk (x) , ϕk (x + 2π) = eiϑ ϕk (x)
gives rise to a symmetric operator A on L2 (T1 ) and that all of the eigenvalues
λk (ϑ) are real. Indexing the eigenvalues in increasing order, show that they
are 2π-periodic in ϑ. For which values of ϑ are the eigenvalues simple, and
for which values are they double? Comment on the properties of continuity
of these eigenvalues and of the associated eigenfunctions.

Projects: Chapter 2
Project 2.14 (Eigenmodes of the oboe). Formulated in terms of the density
ρ(t, r), we will assume that the air column in an oboe (or a saxophone) de-
pends only upon the radial coordinate r, where (r, θ, φ) are polar coordinates
in R3 .
(i) Making the substitution R(r) = rρ(r), show that the wave equation
(2.35) can indeed be rephrased for the radial function R(r) as
∂t2 R − p (ρ0 )∂r2 R = 0
with boundary conditions
R(t, 0) = 0 = R(t, 2π) .
See Beals [3] for a reference.
(ii) Using the method of separation of variables, the eigenmodes for the air
column are given by an eigenvalue problem for ρ(r), which in turn becomes
2
an eigenvalue problem for R(r), with evidently the eigenvalues λn = n4 and
eigenmodes
1 n
Rn (r) = sin( r) .
cn 2
Therefore rρn (r) = Rn (r) solves the eigenvalue problem (2.35). The L2
inner product for radial functions defined on B2π (0) is given by
 2π
f, g = f (r)g(r) r2 dr .
0

Show that Rn , Rm  = 0 for n = m. Compute the value of the constant cn

such that
ρn 2
L2 = ρn , ρn  = 1 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 38 2. Wave equations

(iii) The linearized Euler’s equations in three dimensions can be written for
(ρ, v), where ρ is the density as above and v is a velocity vector field:
∂t ρ + ρ0 ∇ · v = 0 ,
ρ0 ∂t v + p (ρ0 )∇ρ = 0 .
Show that the vector field v satisfies the wave equation
∂t2 v − p (ρ0 )∇(∇ · v) = 0
with boundary conditions v(t, 0) = 0 and ∇ · v(t, 2π) = 0.
(iv) The eigenvalue problem for the normal modes of the oboe can also be
phrased in terms of the velocity variables v, which is a vector field in the
three-dimensional problem. In polar coordinates (r, θ, φ) it can be written
v = (vr , vθ , vφ ) .
The gradient of a function f is given in polar coordinates by
1 1
∇f = ∂r f r̂ + ∂θ f θ̂ + ∂φ f φ̂
r r sin(θ)
and the divergence of a vector field u is given by
1 1 1
∇ · u = 2 ∂r (r2 ur ) + ∂θ (uθ sin(θ)) + ∂φ uφ .
r r sin(θ) r sin(θ)
Assuming that the motion of the air column in the instrument is radial, that
is,
v(r) = (vr (t, r), 0, 0) ,
show that the component vr (t, r) satisfies the scalar equation
1
∂t2 vr − p (ρ0 )∂r ∂r (r2 vr ) = 0
r2
with boundary conditions
1
vr (t, 0) = 0 , ∂r (r2 vr ) = 0 .
r2

(v) Setting vr = ∂r ( 1r W ), show that

∂t2 W − p (ρ0 )∂r2 W = 0
with boundary conditions W (t, 0) = 0 = W (t, 2π). Thus the eigenmodes
given in terms of the velocity field wn (r), corresponding to ρn (r), are
1 1 n
wn (r) = ∂r sin( r) .
dn r 2
Show that wn , wm  = 0 for n = m. Calculate the normalizing constants
dn such that wn , wn  = wn 2 = 1. The first five of these eigenmodes are
illustrated in Figure 11.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 2 39

Project 2.15 (Quantum harmonic oscillator). Solving the Schrödinger

equation for the quantum harmonic oscillator
1 1
i∂t ψ = − ∂x2 ψ + x2 ψ ,
2 2
by separation of variables one is led to the expression

ψ(t, x) = ak e−iλk t ϕk (x) .
k
The eigenfunctions and eigenvalue pairs (ϕk (x), λk ) are given by the problem
in spectral theory:
1 1
(2.36) − ∂x2 ϕk (x) + x2 ϕk (x) = λk ϕk (x) .
2 2
(i) The first eigenfunction ψ1 (x) is given above as a normalized Gaussian.
In the following steps, derive the orthonormalized higher eigenfunctions for
(2.36), and show that the corresponding eigenvalues λk are all real. Verify
that they are as stated in the text.
(a) For λ finite and for |x| large, the potential term V (x) = 12 x2 is dominant,
and the asymptotic behavior of a solution of (2.36) is given by
+ be−x
2 /2 2 /2
ϕ(x) ∼ aex .
Therefore in order to have a localized solution, a = 0. This motivates the
Ansatz that
ϕ(x) = e−x /2 H(x)
2

for functions H(x) to be specified.

(b) Substituting this Ansatz into equation (2.36), show that
d2 H dH
2
− 2x + (2λ − 1)H = 0 .
dx dx
Make a formal power series expansion for H(x) in the form
∞

H(x) = h k xk
k=0
so that
 ∞  ∞
dH d2 H
= khk xk−1 , = k(k − 1)hk xk−2 .
dx dx2
k=1 k=2
Show that the equation results in a recursion formula for the coefficients:
(2λ − 1) − 2k
hk+2 = hk .
(k + 1)(k + 2)
The first two coefficients h0 and h1 are not specified. Furthermore the
recursion for even index coefficients h2m is independent of the recursion for
odd index coefficients h2m+1 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 40 2. Wave equations

(c) For most values of λ, the coefficients of this series grow rapidly and
the resulting formal power series does not converge. However, the series
terminates when
1
λ = λk = k + .
2
The resulting polynomials H(x) = Hk (x) are the Hermite polynomials, and
the eigenfunctions and eigenvalues of the quantum harmonic oscillator are
given by
1
ϕk (x) = ck Hk (x)e−x /2 , λk = k + .
2

2
The normalizing constants ck for the Hermite functions hk (x) = Hk (x)e−x /2
2

are in order to have ϕk L2 = 1. In fact this gives an orthonormal basis for

L2 (R1 ). The additive constant 12 to each eigenvalue is a correction to the
original quantization postulate of Planck.
(d) Another approach to constructing the eigenfunction series ϕk (x) is to
apply a Gram–Schmidt procedure to the set of functions
f (x) = x e−x
2 /2
.
One finds, to start with,
1 −x2 /2 4 −x2 /2
ϕ0 (x) = √
4
4
e , ϕ1 (x) = xe .
π π
For a continuing exercise, calculate several more normalized Hermite func-
tions.
(ii) Show that under Fourier transform the eigenvalue problem (2.36) is
transformed to itself. Conclude that the eigenfunctions of the quantum
harmonic oscillator are also eigenfunctions of the Fourier transform, and
vice versa.
(iii) What are the corresponding eigenvalues of the Fourier transform?
Project 2.16 (Electron orbitals for the hydrogen atom). The Schrödinger
equation for the motion of an electron in the quantum mechanical potential
of an atomic nucleus is given in equation (2.31), which gives rise to a spectral
problem for the Schrödinger operator (2.32) after a separation of variables
involving time. Namely, for an eigenfunction-eigenvalue pair (ϕn (x), λn ) of
the spectral problem
1 Z
(2.37) − Δ2 ϕ − ϕ = λϕ ,
2 |x|
the time evolution of a solution that is initially in this eigenstate is given by
ψ(t, x) = e−iλn t ϕn (x) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 2 41

Further separation of variables allows us to give a complete picture of the

spectrum of the Schrödinger operator (2.32). In this process we are intro-
duced to several of the classical special functions of mathematical physics.
(i) Because the potential function
Z
V (x) = −
|x|
only depends upon the radial variable |x| = r, one transforms the problem
to spherical polar coordinates x = (r, φ, θ). In these coordinates the Laplace
operator is given by
1 1 1
Δϕ = ∂r r2 ∂r ϕ + 2 2 ∂φ2 ϕ + 2 ∂θ sin(θ)∂θ ϕ .
r2 r sin (θ) r sin(θ)
The principle of the method of separation of variables is to seek special
solutions of the form
ϕ(r, φ, θ) = R(r)Φ(φ)Θ(θ) .
Show that a solution of (2.37) in this special form satisfies the three separate
equations
d2 Φ
= −m2 Φ , m ∈ Z ,
dφ2
1 d d m2
− sin(θ) Θ + Θ = kΘ ,
sin(θ) dθ dθ sin2 (θ)
1 d 2d k
2
r R + λ − V (r) R = 2 R .
2r dr dr 2r
The first of these equations evidently has the solution Φ(φ) = eimφ for m ∈
Z. Solutions Θ(θ) for the second equation are polynomials in cos(θ), which,
after a transformation, are given by associated Legendre functions. Solutions
R(r) of the final equation of this list are associated Laguerre functions.
These classical special functions are explored at a later point in this book,
in Project 7.9 of Chapter 7.
Orbitals of electrons are the eigenfunctions of the Schrödinger operator
(2.37). The physical significance of the integer m is as the z component of
the angular momentum of the orbital. In the second equation, the allowed
values of k are k = ( + 1) for ∈ N, ≥ |m|. Combining the angular
behavior of the solution gives the classical spherical harmonics
Ym (φ, θ) = Φm (φ)Θm
 (θ) ,

which are themselves the restrictions to the unit sphere in R3 of harmonic

polynomials, a fact discussed in Exercise 1.4. The total angular momentum

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 42 2. Wave equations


of the eigenstate is L = ( + 1). Once given k = ( + 1), the third equa-
tion is a spectral problem for the function R(r), which has eigenfunctions
that are ordered by n ∈ N and whose eigenvalues
Z2
λn = −
2n2
are determined by the solution process of the third equation. These are the
energies of the eigenstates. The full orbital is given by the eigenfunction
ϕ(x) = Rn (r)Φm (φ)Θm
 (θ) .
The three numbers (n, , m) are the quantum numbers of the eigenstates;
they satisfy
n = 1, 2, 3 . . . , = 0, 1, 2, . . . , n − 1 , m = − , −( − 1) . . . ( − 0, ) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/03

Chapter 3

The heat equation

White hot steel on an anvil. Image credit: Fmajor / E+ / Getty Images.

The Fourier transform was originally introduced by Joseph Fourier in

an 1807 paper in order to construct a solution of the heat equation on an
interval 0 < x < 2π, and we will also use it to do something similar for the
equation
1
(3.1) ∂t u = ∂x2 u , t ∈ R1+ , x ∈ R1 ,
2
u(0, x) = f (x) .
The first thing to notice is that equation (3.1) is invariant under Brownian
scaling, which is the change of variables
√
t = εt , x = εx .

43

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 44 3. The heat equation

This rescaling transformation is closely related to certain principles of prob-

ability and the properties of Brownian motion. To see how the equation
behaves, the change of variables gives
√
∂t = ε∂t , ∂x = ε∂x ,
and therefore if u(t, x) solves the heat equation (3.1), then so does u :=
√
u (t , x ) = u(εt, εx):
1 1 √
ε ∂t − ∂x2 u (t , x ) = ∂t − ∂x2 u(εt, εx) = 0 .
2 2
√
We will see that the ratio x/ t, which is invariant under Brownian scaling,
appears in a central way in expressions for the solution, and indeed it plays
a related role in probability.

3.1. The heat kernel

A derivation of the solution of (3.1) by Fourier synthesis starts with the
assumption that the solution u(t, x) is sufficiently well behaved so that it
satisfies the hypotheses of the Fourier inversion formula. This will be verified
a posteriori in section 5.3. Writing
 +∞
1
u(t, x) = √ eiξx û(t, ξ) dξ ,
2π −∞
solutions of (3.1) must satisfy
 +∞
1 1 1
0 = ∂t − ∂x2 u(t, x) = √ eiξx ∂t − (iξ)2 û(t, ξ) dξ .
2 2π −∞ 2
We have used Proposition 2.3 to express the spatial derivatives in terms of
the Fourier transform. Reasoning as before, the Fourier transform of the
solution must satisfy the family of ODEs
d 1
û + ξ 2 û = 0 ,
dt 2
parametrized by ξ. The solution is
û(t, ξ) = e− 2 ξ t û(0, ξ) = e− 2 ξ t fˆ(ξ) ,
1 2 1 2

where fˆ(ξ) is the Fourier transform of the initial data. With this we can
express the solution in integral operator form:
 +∞
1
eiξx e− 2 ξ t fˆ(ξ) dξ
1 2
u(t, x) = √
2π −∞

1
eiξ(x−y) e− 2 ξ t f (y) dydξ
1 2
(3.2) =
2π
   
1
eiξ(x−y) e− 2 ξ t dξ f (y) dy := H(t, x − y)f (y) dy .
1 2
=
2π

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.1. The heat kernel 45

Justification of the exchanges of the order of integration comes from the

Fubini theorem, as long as the initial data f is absolutely integrable, meaning
that it satisfies 
f L1 (R1 ) := |f (x)| dx < +∞ .
R1
The linear space of such functions is known as L1 (R1 ). The function H(t, x−
y) is the heat kernel, that is, the integral kernel (or the first component of
the integrand) for the solution operator H(t) for the heat equation with the
initial data f (x):
u(t, x) = H(t)f (x) .
Complete the square to evaluate this Fourier integral definition of the heat
kernel:
 +∞
1
eiξx e− 2 ξ t dξ
1 2
H(t, x) =
2π −∞

1 − 1 x2 /t +∞ 1 x2 /t+iξx− 1 ξ2 t
= e 2 e2 2 dξ
2π −∞

1 − 1 x2 /t +∞ − 1 (ξ√t−ix/√t)2
= e 2 e 2 dξ .
2π −∞
√ √
Substitute ξ  = ξ t − ix/ t (which is actually an application of Cauchy’s
theorem of complex analysis) to find the expression
 
1 − 1 x2 /t 1  2 dξ 1 −x2 /2t
(3.3) H(t, x) = e 2 e− 2 (ξ ) √ = √ e .
2π t 2πt
Theorem 3.1. A solution of the heat equation for initial data f ∈ L1 (R1 )
is given by the convolution of the initial data with the heat kernel:
 +∞  +∞
1 −(x−y)2 /2t
(3.4) u(t, x) = H(t, x − y)f (y) dy = √ e f (y) dy .
−∞ −∞ 2πt
Take note of the invariance under Brownian scaling of the quantity
1 −x2 /2t
dPt (x) = √ e dx ,
2πt
which is a Gaussian measure that is relevant to heat flow. It is a probability
measure since  +∞
1 −(x−y)2 /2t
√ e dy = 1 .
−∞ 2πt
This procedure via the heat kernel gives a unique solution among the
class of bounded solutions, or among solutions which are in L1 (R1x ), or simply
among solutions which do not grow too rapidly as |x| → +∞. It is a
surprising fact, which will come up in a later discussion, that solutions of the
heat equation are not unique if we admit ones that are badly unbounded [20].
This topic is covered in Project 3.12 at the end of this chapter.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 46 3. The heat equation

Figure 1. Superposition of the graph of the heat kernel at several times

0 < t1 < t2 < t3 .

The heat kernel H(t, x − y) for t > 0 has its maximum at x √ = y, where
1
the maximum value is √2πt , and width, or standard deviation, is t (defined
as the distance between its center and inflection point). A graph of the heat
kernel taken at several snapshots in time is given in Figure 1.
It is reasonable to expect that we would obtain the same solution u(t, x)
at time t > 0 starting with data f (x) at time t = 0 if we alternatively solved
the heat equations up to time 0 < s < t, and then, starting again at s
with data u(s, x), continued the solution for time t − s. That is, in operator
notation
(3.5) H(t)f = H(t − s)H(s)f , 0<s<t.
This is the semigroup property, and it is related to the question of uniqueness.

3.2. Convolution operators

The evolution operator for the heat equation is an example of a convolution
operator, with convolution kernel the function H(t, x).
Definition 3.2. Let h(x) ∈ L1 (R1 ), and define the convolution product with
f (x) ∈ L1 (R1 ) to be
 +∞
h ∗ f (x) = h(x − y)f (y) dy .
−∞

In these terms the evolution operator for the heat equation can be writ-
ten as
 +∞
u(t, x) = H(t)f (x) = H(t, ·) ∗ f (x) = H(t, x − y)f (y) dy .
−∞

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.2. Convolution operators 47

The convolution product is well defined between integrable functions:


h L1 (R1 ) = |h(x)| dx < +∞ .
R1

In fact, whenever f (x), h(x) ∈ L1 (R1 ), then h ∗ f (x) ∈ L1 (R1 ), as is shown

in the following proposition.
Proposition 3.3. Given two functions h(x), f (x) ∈ L1 (R1 ), the convolution
product g(x) := h ∗ f (x) is well defined. Furthermore g(x) ∈ L1 (R1 ) and
g L1 (R1 ) ≤ h L1 (R1 ) f L1 (R1 ) .

Proof. These assertions will follow from the proof that g(x) is integrable.
To this end, estimate
  
 
|g(x)| dx =  h(x − y)f (y) dy  dx
R1 R1 R1
 
≤ |h(x − y)||f (y)| dydx = h L1 (R1 ) f L1 (R1 ) ,
R1 R1
the last step following after an exchange of order of integration. 

Convolution operators with kernels h ∈ L1 (R1 ) have a number of con-

venient algebraic features; the most elementary ones are covered in the fol-
lowing proposition.
Proposition 3.4. Let h(x), f (x), g(x) ∈ L1 (R1 ). Then
(i) The convolution product is commutative and associative:
h ∗ f (x) = f ∗ h (x) ,
h ∗ g ∗ f = h ∗ g ∗ f (x) .

(ii) If in addition ∂x f ∈ L1 (R1 ), then differentiation commutes with the

operation of convolution:
∂x h ∗ f (x) = h ∗ ∂x f (x) .

(iii) Under Fourier transform, convolution with h(x) becomes a multiplica-

tion operator
√
h∗ f (ξ) = 2π ĥ(ξ)fˆ(ξ) .

Proof. Address statement (iii) first:

 +∞  +∞
1 −iξx
√ e h(x − y)f (y) dy dx
2π −∞ −∞

1
=√ e−iξ(x−y) h(x − y)e−iξy f (y) dydx
2π

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 48 3. The heat equation

by using Fubini’s Theorem. Then this quantity can be rewritten as

√  1  −iξy  1 
= 2π √ e f (y) √ e−iξ(x−y) h(x − y) dx dy
2π 2π
√
= 2π fˆ(ξ)ĥ(ξ) ,
which is the statement of (iii). This can now be used to prove (i) and (ii).
Indeed, since
√
h ∗ f (ξ) = 2π ĥ(ξ)fˆ(ξ) ,
the order of multiplication does not play a role, and therefore the properties
of commutativity and associativity clearly hold. Furthermore,
√
∂x
(h ∗ f )(ξ) = iξ h
∗ f (ξ) = 2π(iξ)ĥ(ξ)fˆ(ξ) ,
and the derivative can be seen to be acting either on the convolution product
or on f alone. The fact is that the Fourier transform has simultaneously
diagonalized the operations of convolution with h(x) and differentiation,
from which the results of the proposition are corollaries. 

3.3. The maximum principle

This fundamental principle is a feature of solutions of parabolic equations
such as the heat equation; we will also encounter it when we take up the
topic of elliptic equations such as Laplace’s equation. The following section
concerns the solution u(t, x) of the heat equation that is given by the heat
operator u(t, x) = (H(t)f )(x) = (H(t, ·) ∗ f )(x).

Theorem 3.5 (Maximum principle). Suppose that the initial data f ∈

L1 (R1 ) satisfies f (x) ≥ 0 for all x ∈ R1 . Then either u(t, x) > 0 for all
t > 0, or else both u(t, x) ≡ 0 and f (x) ≡ 0.

The conclusion is powerful because of the fact that the existence of one
zero of the solution implies that both the whole solution and the initial data
must vanish identically.

e−x /2t > 0 for

1 2
Proof. Observe that the heat kernel satisfies H(t, x) = √2πt
all t > 0. Hence if f (x) ≥ 0, then clearly
 +∞
u(t, x) = H(t, ·) ∗ f (x) = H(t, x − y)f (y) dy ≥ 0 .
−∞

This nonstrict inequality is called the weak maximum principle. By a further

argument we can prove that u(t, x) > 0 unless f ≡ 0. Suppose that f (x) ≥ 0
and define sets Bδ = {x : f (x) > δ}. If f is not identically zero (excluding

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.4. Initial and initial-boundary value problems 49

sets of zero Lebesgue measure), then Bδ is nonempty for some δ and there
is a bounded set A ⊆ Bδ with positive measure. Then
 +∞
u(t, x) = H(t, x − y)f (y) dy
−∞

≥ H(t, x − y)f (y) dy
 A

≥ H(t, x − y)δ dy .
A
This last expression is surely positive by the positive character of the heat
kernel H(t, x). There is more precise information available on a lower bound
for u(t, x) if we consider the Gaussian nature of the heat kernel:
 1 (x−y)2
u(t, x) ≥ δ meas(A) inf √ e− 2t .
y∈A 2πt


The maximum principle provides a partial ordering of solutions of the

heat equation, at least for those given by the heat kernel, and thus gives a
uniqueness theorem for the heat equation.
Theorem 3.6. Suppose that f1 (x), f2 (x) ∈ L1 (R1 ) and that f1 (x) ≤ f2 (x)
for all x ∈ R1 . Then the respective solutions u1 (t, x), u2 (t, x) either satisfy
u1 (t, x) < u2 (t, x)
for all t > 0, or else u1 (t, x) ≡ u2 (t, x) and f1 (x) ≡ f2 (x).

Proof. The difference u2 − u1 also satisfies the heat equation, and it is

given by convolution of the heat kernel with the initial data f2 − f1 ≥ 0.
The maximum principle asserts that either (u2 − u1 )(t, x) > 0 for all t > 0,
or else u2 − u1 = 0 and f2 − f1 = 0. 

3.4. Initial and initial-boundary value problems

The expression in (3.4) gives a solution to the heat equation, as we have
derived. We now address the question as to whether this expression satisfies
the desired initial conditions, that is, the conditions for which, as time t
decreases to zero through positive values, the following limit holds:
 +∞
lim u(t, x) = lim H(t, x − y)f (y) dy = f (x) .
t→0+ t→0+ −∞

There is a satisfactory answer at points x0 ∈ R1 at which f (x) is continu-

ous. Exercise 3.5 addresses the case in which the initial data has a jump
discontinuity at x = x0 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 50 3. The heat equation

Theorem 3.7. Suppose that f ∈ L1 (R1 ) and additionally that f (x) is con-
tinuous at a point x = x0 . Then
 +∞
lim H(t, x0 − y)f (y) dy = f (x0 ) .
t→0+ −∞

This is an instance of a Dirac δ-function. The δ-function is expressed

formally as a distribution that satisfies the relation
 +∞
δ(x0 − y)f (y) dx = f (x0 )
−∞

for any continuous function f (x). Its mathematical sense is defined in terms
of a limit, as in the limit of the statement of the theorem.

Proof. Since we know f (x) to be continuous at x = x0 , for any ε > 0 there

is a δ > 0 such that
ε
(3.6) |f (x) − f (x0 )| <
3

for all |x − x0 | < δ. Secondly, we recall the fact that

 +∞
1 − (x0 −y)2
√ e 2t dy = 1
−∞ 2πt

for all t > 0 because of the normalization of the heat kernel. Furthermore,
for y = x0 the heat kernel satisfies

1 − (x0 −y)2
lim √ e 2t =0.
t→0+ 2πt

This limit is uniform for y bounded away from x0 , and in fact for δ > 0
fixed, for any ε > 0 there exists a time t(ε) > 0 such that for any smaller
0 < t < t(ε), both

1 − (x0 −y)2 ε
(3.7) √ e 2t f L1 < for any |y − x0 | ≥ δ ,
2πt 3

1 − (x0 −y)2 ε
(3.8) √ e 2t dy |f (x0 )| < .
|y−x0 |≥δ 2πt 3

Now split up the integral in expression (3.4) for the solution of the heat
equation in the following way. Given ε > 0, choose δ > 0 so that (3.6) holds

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.4. Initial and initial-boundary value problems 51

for |x − x0 | < δ, and consider

 +∞
1 − (x0 −y)2
f (x0 ) − √ e 2t f (y) dy
−∞ 2πt
 +∞
1 − (x0 −y)2
= √ e 2t [f (x0 ) − f (y)] dy
−∞ 2πt
 x0 −δ  +∞
1 − (x0 −y)2
= + √ e 2t [f (x0 ) − f (y)] dy
−∞ x0 +δ 2πt
 x0 +δ
1 − (x0 −y)2
+ √ e 2t [f (x0 ) − f (y)] dy .
x0 −δ 2πt

For 0 < t < t(ε), the first two integrals are bounded in terms of ε using, re-
spectively, (3.8) for the terms containing f (x0 ) and (3.7) for those containing
f (y):

 x0 −δ  +∞ 
 1 − (x0 −y)2 
 + √
e 2t [f (x0 ) − f (y)] dy 
−∞ x0 +δ 2πt
 x0 −δ  +∞ 1
 (x0 −y)2
≤ |f (x0 )| + √ e− 2t dy
−∞ x0 +δ 2πt
 +∞
1 − (x0 −y)2
+ |f (y)| dy sup √ e 2t
−∞ |x0 −y|≥δ 2πt
2ε
≤ .
3

Using (3.6), the remaining integral over the interval [x0 − δ, x0 + δ] is also
shown to be bounded in terms of ε:
 x0 +δ  ε  x0 +δ 1
 1 − (x0 −y)2  (x0 −y)2
 √ e 2t [f (x0 ) − f (y)] dy  ≤ √ e− 2t dy
x0 −δ 2πt 3 x0 −δ 2πt
 +∞ 2
ε 1 − (x0 −y) ε
≤ √ e 2t dy = .
3 −∞ 2πt 3

Since ε can be chosen arbitrarily small, this proves the statement of the
theorem. 

Method of images. As with the wave equation, there are methods for
solving the heat equation in a region with a boundary, where boundary
conditions are imposed. Not surprisingly, this involves the method of images.
The two basic boundary conditions are homogeneous Dirichlet conditions,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 52 3. The heat equation

which we will impose at x = 0 for the region {x > 0} := R1+ ,

1
∂t u = ∂x2 u , x > 0 , t > 0 ,
2
u(0, x) = f (x) , x > 0 ,
u(t, 0) = 0 ,

and homogeneous Neumann boundary conditions

∂x u(t, 0) = 0 .

In physical terms, homogeneous Dirichlet conditions correspond to the

boundary being held at a fixed temperature, such as the reference tem-
perature u(t, 0) = 0, as though it were connected to an infinite reservoir of
heat energy at this temperature. Homogeneous Neumann boundary condi-
tions correspond to the boundary being a perfect heat insulator, allowing
no heat flux through it.
Initial data f (x) is given for x > 0; the method of images corresponds to
extending the initial data to x < 0 in a way that the resulting solution will
obey the desired boundary conditions. In the case of Dirichlet boundary
conditions we reflect the data to be odd in x; namely, for x < 0 define
f (x) = −f (−x). Then the solution of the heat equation on the whole line
is given by

u(t, x) = H(t, x − y)f (y) dy
R
1

= H(t, x − y)f (y) dy + H(t, x − y)f (y) dy
y≥0 y<0

= [H(t, x − y) − H(t, x + y)]f (y) dy .
y≥0

The heat kernel for the Dirichlet problem on R1+ is therefore

1 (x−y)2 (x+y)2
HD (t, x, y) = [H(t, x − y) − H(t, x + y)] = √ [e 2t − e 2t ] ,
2πt
where both x > 0 and y > 0. We may denote the solution operator for heat
flow on R+1 with Dirichlet boundary conditions by

 +∞
HD (t)f (x) = HD (t, x, y)f (y) dy .
0

Similarly, initial data for the case of Neumann boundary conditions use
the method of images with an even reflection of the initial data, defining

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.4. Initial and initial-boundary value problems 53

f (x) = f (−x) for x < 0. In this case


u(t, x) = H(t, x − y)f (y) dy
R
1

= H(t, x − y)f (y) dy + H(t, x − y)f (y) dy
y≥0 y<0

= [H(t, x − y) + H(t, x + y)]f (y) dy .
y≥0

The resulting heat kernel for the Neumann problem on R1+ is

1 (x−y)2 (x+y)2
HN (t, x, y) = [H(t, x − y) + H(t, x + y)] = √ [e 2t + e 2t ] ,
2πt
where again both x > 0 and y > 0. One may similarly denote the solution
1 with Neumann boundary conditions by
operator for heat flow on R+
 +∞
HN (t)f (x) = HN (t, x, y)f (y) dy .
0

Duhamel’s principle. The inhomogeneous problem for the heat equation

is the following:
1
(3.9) ∂t u − ∂x2 u = h(t, x) ,
2
u(0, x) = f (x) ,

which we consider first for x ∈ R1 . The general principle given in section 2.5
indicates the solution process. As before, we decompose the solution as
u(t, x) = u1 (t, x) + u2 (t, x) where u1 is a solution of the initial value problem
with initial data f (x), for which we take the inhomogeneous term to be zero.
The component u2 (t, x) then accounts for the presence of a nonzero h(t, x),
which, following Duhamel’s principle, is given by the expression
 t
u2 (t, x) = H(t − s)h(s, ·)(x) ds
0
 t  +∞ (x−y)2
1
=  e 2(t−s) h(s, y) dyds .
0 −∞ 2π(t − s)

One can check that u2 (0, x) = 0, so that the component u1 (t, x) accounts
for the initial data f (x), and that
1
∂t u2 = ∂x2 u2 + h(t, x) .
2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 54 3. The heat equation

The inhomogeneous problems on R1+ have similar solution formulae.

This problem with Dirichlet boundary conditions imposed at x = 0 is
1
∂t u = ∂x2 u + h(t, x) , x>0,
2
u(0, x) = f (x) , x > 0 , u(t, 0) = 0 .
Decomposing the solution u(t, x) = u1 (t, x) + u2 (t, x) as above, and using
the solution operator u1 = HD (t)f for the first component, the second
component is given by Duhamel’s principle:
 t
u2 (t, x) = HD (t − s)h(s, ·) ds
0
 t
= HD (t − s, x, y)h(s, y) dyds .
0
Similarly, one may instead impose Neumann boundary conditions at the
boundary x = 0:
∂x u(t, 0) = 0 .
In this case the Neumann heat kernel is used so that u1 (t, x) = HN (t)f and
 t
u2 (t, x) = HN (t − s, x, y)h(s, y) dyds .
0

Green’s identities. The heat kernel H(t − s, x − y) that appears in the

formulae of Duhamel’s principle, considered as a function of (t, x), is a solu-
tion of the heat equation. Considered as a function of (s, y), it is a solution
of the heat equation with reversed time 0 < s < t:
1
(3.10) ∂s H = − ∂y2 H .
2
Given any functions G(s, y) and u(s, y) that are both smooth in the domain
x ∈ R1 and 0 < s < t, integrations by parts result in the following calculus
identities:
 t  +∞
1
G(s, y) ∂s u − ∂y2 u dyds
0 −∞ 2
 t  +∞  +∞ t
1 2 
= −∂s G − ∂y G u(s, y) dyds + G(s, y)u(s, y) dy  ,
0 −∞ 2 −∞ s=0

where we will assume that G and u decrease to zero as |y| → +∞ sufficiently

rapidly so that there are no remaining boundary terms. If G(s, y) also
satisfies (3.10), then the RHS of the integral expression simplifies. If we
take G(s, y) = H(t − s, x − y), we also know from Theorem 3.7 that
 +∞
lim H(t − s, x − y)u(s, y) dy = u(t, x) ,
s→t− −∞

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.4. Initial and initial-boundary value problems 55

and therefore
 t +∞
1
(3.11) H(t − s, x − y) ∂s u − ∂y2 u dyds
0 −∞ 2
 +∞
= u(t, x) − H(t, x − y)u(0, y) dy .
−∞

This is known as Green’s identity for the heat equation. From it we recover
the statement that if u(t, x) is a solution of the heat equation (3.1) which
decays at |x| → +∞ sufficiently well, then
 +∞
u(t, x) = H(t, x − y)f (y) dy .
−∞

Furthermore, if u(t, x) is a solution of the inhomogeneous heat equation

(3.9), then this gives an alternate derivation of the formula stemming from
Duhamel’s principle,
 +∞  t +∞
u(t, x) = H(t, x − y)f (y) dy + H(t − s, x − y)h(s, y) dyds .
−∞ 0 −∞

Inhomogeneous boundary conditions. Green’s identities are also useful

for the initial-boundary value problem for the heat equation. The analog
of the calculus identities for the initial-boundary value problem with given
Dirichlet boundary data is that
 t +∞
1
HD (t, x, y) ∂s u − ∂y2 u dyds
0 0 2
 t  +∞
1
= −∂s HD − ∂y2 HD u(s, y) dyds
0 0 2
 +∞ t

+ HD (t − s, x, y)u(s, y) dy 
0 s=0
 t +∞
1 
+ −HD (t − s, x − y)∂y u − ∂y HD (t − s, x − y)u ds .
2 0 y=0

The heat kernel HD (t − s, x, y) satisfies (3.10). We also know that the limit
as s → t− of the heat kernel HD (t, s, x, y) = δx (y). Using this as well as the
fact that HD (t − s, x, 0) = 0, and furthermore assuming that the solution
u(s, y) does not grow faster than the decay of the heat kernel as y → +∞,
the above formula gives us a second version of Green’s identities (3.11),

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 56 3. The heat equation

namely
 t +∞
1
HD (t, x, y) ∂s u − ∂y2 u dyds
0 0 2
 +∞
(3.12) = u(t, x) − HD (t, x, y)u(0, y)dy
0
 t
1
+ ∂y HD (t − s, x, 0)u(s, 0) ds .
0 2
Initial data u(0, x) = f (x), boundary data u(t, 0) = g(t), and inhomogeneous
data h(t, x) gives a full solution formula:
 +∞
u(t, x) = HD (t, x, y)f (y)dy
0
 t
1
+ ∂y HD (t − s, x, 0)g(s) ds
0 2
 t  +∞
+ HD (t, x, y)h(s, y) dyds .
0 0

A similar line of reasoning gives a formula for solutions of the heat equa-
tion on R1+ with prescribed inhomogeneous Neumann data on the bound-
ary x = 0. Namely, given initial data u(0, x) = f (x), boundary data
∂x u(t, 0) = g(t), and inhomogeneous data h(t, x), one finds the following:
 +∞
u(t, x) = HN (t, x, y)f (y)dy
0
 t
1
(3.13) − HN (t − s, x, 0)g(s) ds
0 2
 t  +∞
+ HN (t, x, y)h(s, y) dyds .
0 0

3.5. Conservation laws and the evolution of moments

The heat equation is often called the diffusion equation, and indeed the
physical interpretation of a solution is of a heat distribution or a particle
density distribution that is evolving in time according to equation (3.1).
That is, in probabilistic terms, if the distribution of outcomes of a random
event at time t is described by the function u(t, x), the quantity
 b
Pt [a, b) = u(t, x) dx
a
represents the probability of the outcome of a random event taking its value
in the interval [a, b) at that time. For instance Pt [a, b) might represent
the probability that a random particle lies in the interval at that time.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.5. Conservation laws and the evolution of moments 57

For this interpretation we should

 ask that the initial data f (x) satisfy f (x) ≥
0 and normalize it so that R1 f (x) dx = 1. In particular we require that f ∈
L1 (R1 ). We can derive the heat equation itself from this interpretation given
one more piece of information, namely Fourier’s law of heat conservation and
flux. Taking [a, b) to be an arbitrary interval, consider the change in u(t, x)
in the interval over time. Fourier’s law of heat conservation posits that this
change is given by the flux of heat across the boundary. Namely, the change
in heat over this interval in time is given by
 b
∂t Pt [a, b) = ∂t u(t, x) dx = F (a) − F (b) ,
a

where F (a) is interpreted to be the flux of heat across the boundary point
a into the interval, and F (b) is interpreted to be the flux of heat out of the
interval across the boundary point b. Fourier’s law for the heat flux gives
the form as
1
F (a) := − ∂x u(t, a) ,
2
and similarly for b so that
 b
∂t Pt [a, b) = ∂t u(t, x) dx = F (a) − F (b)
a
 b
1 1 1 2
= ∂x u(t, b) − ∂x u(t, a) = ∂ u(t, x) dx .
2 2 a 2 x
Since the interval [a, b) is arbitrary, the heat density u(t, x) must satisfy the
heat equation ∂t u = 12 ∂x2 u.
This interpretation must satisfy two conditions in order to be consis-
tent: (i) the total amount of heat must be conserved, and (ii) if the initial
heat distribution f (x) ≥ 0, then for all t ≥ 0 we should have u(t, x) ≥ 0.
The second condition is already satisfied for us by the maximum principle.
Regarding the first, there is the following result.
Theorem 3.8. If u(t, x) ∈ L1 (R1 ) is a solution of (3.1), then for all t ≥ 0
 +∞  +∞
u(t, x) dx = f (x) dx .
−∞ −∞

Proof.
 +∞  +∞
∂t u(t, x) dx = ∂t u(t, x) dx
−∞ −∞
 +∞
1 2 1
= ∂x u(t, x) dx = lim ∂x u(t, R) − ∂x u(t, −R) = 0 .
−∞ 2 2 R→+∞


Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 58 3. The heat equation

In fact the proof is incomplete, for we should also have shown that
when f (x) ∈ L1 (R1 ), then for t > 0 we have ∂x2 u(t, x) ∈ L1 (R1 ), and then
justified the exchanges of differentiation and integration and the integrations
by parts. These facts will be addressed in the later chapter on the properties
of the Fourier transform.
The interpretation of a solution u(t, x) of (3.1) as the density of a prob-
ability measure motivates a discussion of the time evolution of its moments.
For a given distribution f (x) dx, define the kth moment to be the quantity
 +∞
(3.14) mk (f ) = xk f (x) dx .
−∞

For an arbitrary f (x) ∈ L1 (R1 ) this may be infinite or not well defined, but
for sufficiently well-localized functions it makes perfectly good sense. To
give examples, take the case of the heat kernel itself, then
 +∞
1 − x2
m0 (H(t, ·)) = √ e 2t dx = 1 ,
−∞ 2πt
 +∞
1 x2
m1 (H(t, ·)) = √ xe− 2t dx = 0 ,
−∞ 2πt
 +∞
1 x2
m2 (H(t, ·)) = √ x2 e− 2t dx = t .
−∞ 2πt
Now consider the initial value problem for the heat equation; our solution
procedure yields u(t, x) = H(t, ·) ∗ f (x), and the resulting heat distribution
u(t, x) dx has moments which are given by
 +∞ 
1 − (x−y)2
mk (u(t, ·)) = xk u(t, x) dx = xk √ e 2t f (y) dydx .
−∞ 2πt
Proposition 3.9. The first two moments of the solution u(t, x) are con-
served:
m0 (u(t, ·)) = m0 (f ) , m1 (u(t, ·)) = m1 (f ) .

To deduce the evolution properties of the higher moments of u(t, x) dx,

we will derive further elementary properties of the Fourier transform.

Proposition 3.10. For sufficiently well-behaved functions g(x), the Fourier

transform has the property that

F (xk g(x)) = (i∂ξ )k ĝ(ξ) ,

 +∞ √
g(x) dx = 2πĝ(0) .
−∞

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.5. Conservation laws and the evolution of moments 59

Proof. This first statement is essentially dual to that of Proposition 2.3.

Observe by integrations by parts that
  
−iξx k k −iξx
e x g(x) dx = (i∂ξ ) e g(x) dx = (i∂ξ ) k
e−iξx g(x) dx ,

which is the first statement of the proposition. To show the second state-
ment,  
 √
g(x) dx = e−iξx g(x) dxξ=0 = 2πĝ(0) .

Hypotheses on the function g(x) are necessary in order to justify the inte-
grations by parts and the limits. 

We use this information in order to compute moments; since

√
ˆ
xk u(t, ξ) = (i∂ξ )k û(t, ξ) = 2π(i∂ξ )k Ĥ(t, ξ)f(ξ) ,
therefore
 √ 
mk (u(t, ·)) = xk u(t, x) dx = 2π (i∂ξ )k û(t, ξ) ξ=0

ˆ
= 2π (i∂ξ )k Ĥ(t, ξ)f(ξ)  .
ξ=0

In particular for k = 0,
√
m0 (u(t, ·)) = 2π Ĥ(t, 0)fˆ(0) = 2π fˆ(0) = m0 (f ) ,
which gives the first statement of Proposition 3.9. For k = 1,

ˆ
m1 (u(t, ·)) = 2πi∂ξ Ĥ(t, ξ)f(ξ) 
ξ=0

= 2πi ∂ξ Ĥ(t, ξ)f(ξ) + Ĥ(t, ξ)∂ξ fˆ(ξ) ξ=0
ˆ

= 2πiĤ(t, 0)∂ξ fˆ(ξ)ξ=0 = m1 (f ) .
This proves the second statement of Proposition 3.9, for initial data with
finite zeroth and first moments, at least. The general case is now clear:
ˆ
mk (u(t, ·)) = 2π(i∂ξ )k Ĥ(t, ξ)f(ξ) |ξ=0
k  
k
= 2πik ∂ξj Ĥ(t, ξ)∂ξk−j fˆ(ξ) |ξ=0
j
j=0
k  
 k
= mj (H(t, ·))mk−j (f ) .
j
j=0
j even

We used the Leibnitz product rule for differentiation in the second line, and
in the third line we used the fact that odd moments of the heat kernel are
zero, as the kernel itself is an even function of x.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 60 3. The heat equation

A property of solutions of the heat equation, related to that of conserva-

tion of heat as in Theorem 3.8, is that the evolution has a certain contraction
property on a number of common spaces of functions.
Proposition 3.11. Let u(t, x) solve the heat equation (3.1) with initial data
f ∈ L2 (R1 ). Then the L2 -norm of u(t, ·) is a decreasing function of t:
u(t, ·) L2 ≤ f L2 .

Proof. Compute the time derivative of the norm:


∂t u(t, ·) 2L2 = ∂t u2 (t, x) dx
 
1
= 2u(t, x)∂t u(t, x) dx = 2u(t, x) ∂x2 u(t, x) dx
2

= − |∇u(t, x)|2 dx ≤ 0 .

Thus we find that

 t
u(t, ·) 2
L2 − f 2
L2 = ∂t u(s, ·) 2
L2 ds ≤ 0 .
0
One needs to revisit this calculation in order to justify the integration by
parts. 

We have already seen that the setting of L1 (R1 ) functions is sometimes

more natural for the heat equation than that of L2 (R1 ). It turns out that
time evolution by heat flow is also a contraction in L1 (R1 ).
Proposition 3.12. Let f ∈ L1 (R1 ), and let u(t, x) solve the heat equation
with initial data f . Then u(t, ·) L1 is a nonincreasing function of t:
u(t, ·) L1 ≤ f L1 .

Proof. By nonincreasing we mean that the norm in question could stay the
same or it could decrease in time; in any case it cannot increase. The easy
case is for f ∈ L1 to be of one sign; say for convenience that f (x) ≥ 0. Then
the maximum principle implies that u(t, x) ≥ 0 as well, so that
 
∂t |u(t, x)| dx = ∂t u(t, x) dx
 
1 2
= ∂t u(t, x) dx = ∂ u(t, x) dx = 0 ,
2 x
and the L1 -norm is unchanged by the flow. However, when the initial data
f (x) changes sign, the L1 -norm will in fact be decreasing in time. To show
this, decompose the initial data f (x) = f+ (x) − f− (x) into, respectively, its
positive and negative parts. Both f± (x) ≥ 0, and furthermore they have

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.6. The heat equation in Rn 61

disjoint support; {f+ = 0} ∩ {f− = 0} = ∅. The L1 -norm of f is then the

sum of the L1 -norms; indeed,
   
|f (x)| dx = (f+ (x) + f− (x)) dx = f+ dx + f− dx .
supp(f+ ) supp(f− )
The previous result can now be applied to the two initial data, f+ and f− ,
respectively:
u+ (t, ·) L1 = f+ L1 , u− (t, ·) L1 = f− L1 .
The solution is obtained from the sum, u(t, x) = u+ (t, x)−u− (t, x); however,

Figure 2. Decomposition of initial data into positive (f+ (x)) and neg-
ative (−f− (x)) components, and their subsequent evolution under heat
flow, respectively u+ (t, x) and u− (t, x).

the maximum principle applies so that as long as neither of f± is zero almost

everywhere, supp(u+ ) = supp(u− ) = R1 . This decomposition of f (x) and
its time evolution is illustrated in Figure 2. We therefore have the estimate
u(t, ·) L1 = u+ − u− L1
≤ u+ L1 + u− L1 = f+ L1 + f− L1 = f L1 .
Since we are assuming that both u+ and u− are not zero, there must be some
cancellation in the inequality of the second line, and hence the inequality is
in fact strict. 

3.6. The heat equation in Rn

Most of the results of this chapter generally hold for the heat equation posed
in higher-dimensional space Rn for any dimension n. The heat equation is
1
(3.15) ∂t u = Δu , t ∈ R1+ , x ∈ Rn ,
2
u(0, x) = f (x) .
For initial data f (x), we follow the procedure of section 3.1 and use the
Fourier transform, this time in Rn :

1
ˆ
f (ξ) = √ n e−iξ·y f (y) dy := F f (ξ),
2π Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 62 3. The heat equation

ny are in R and where their Euclidean inner product is

where both ξ and n

given by ξ · y = j=1 ξj yj . The inverse transform is similarly given by


1
f (x) = √ n eiξ·x fˆ(ξ) dξ := F −1 fˆ(x) ,
2π Rn
and the Plancherel identity holds for f (x) ∈ L2 (Rn ):
(3.16) f 2
L2 (Rn
x)
= fˆ 2
L2 (Rn .
ξ)

A solution u(t, x) of (3.15) has Fourier transform that satisfies the

parametrized family of ODEs
d 1
û + |ξ|2 û = 0
dt 2
− 12 |ξ|2 t ˆ
which has solution û(t, ξ) = e f (ξ), analogous to our previous expres-
sion. As we did in (3.2), the heat kernel is given by a Gaussian integral

1 1 |x−y|2
eiξ·(x−y) e− 2 |ξ| t dξ = √ n e− 2t .
1 2
(3.17) H(t, x − y) = n
(2πt) Rn 2πt
Thus a solution to (3.15) is given by convolution with the heat kernel

(3.18) u(t, x) = H(t, x − y)f (y) dy .
Rn
The heat kernel gives rise to a probability measure that is important for
Brownian motion in Rn ,
1 |x|2
dPt (x) = √ n e− 2t dx ,
2πt

which is invariant under Brownian scaling and has total measure dPt (x) =
1, as one can check.

3.7. Entropy
The interpretation of nonnegative and normalized solutions of the heat equa-
tion as probability distributions has been discussed in section 3.5. Associ-
ated with such probability measures are a variety of concepts that concern
the entropy of a measure. Given a density f ∈ L1 (Rn ) with f (x) ≥ 0 and
f L1 = 1, the entropy is defined as

(3.19) S(f ) := f (x) log(f (x)) dx
Rn
with the entropy density given by
(3.20) s(x) = f (x) log(f (x)) .
The concept of entropy dates to the work of Boltzmann, who recognized
the quantity as a measure of the disorder of a distribution of particles in

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.7. Entropy 63

the context of the dynamics of gasses. The definition above is made as an

analogy to this idea when applied to heat flow. There are two fundamental
physical measurements when considering heat;1 one is the quantity of heat,
given by 
u(t, x) dx = m0 (u(t, ·)) ,
Rn
which is conserved as described in Proposition 3.9. The second is a measure
of the amount of disorder in the heat distribution, given by the entropy
S(u(t, ·)). Work can be performed by ordered distributions of heat, as in,
for example, a steam engine. Entropy decreases the amount of work that can
be performed for the same quantity of heat. If f (x) is taken as initial data for
the heat equation, the entropy of the solution u(t, x) evolves monotonically
in time; this is the conclusion of the following proposition.
Proposition 3.13. Suppose that f (x) ≥ 0 is taken as initial data for the
heat equation. Then the solution u(t, x) given by (3.18) satisfies
d
S(u(t, ·)) ≤ 0 .
dt
Proof. Unless f (x) = 0, we know by the maximum principle that u(t, x) >
0, and hence the entropy density s(t, x) is at least defined. The time evo-
lution of the total entropy density in a domain D ⊆ Rn can be expressed
by
 
d d
s(t, x) dx = u(t, x) log(u(t, x)) dx
dt D dt
 D 
∂t u
= ∂t u log(u) dx + u dx
u
 D D

1 1
= Δu log(u) dx + Δu dx .
D 2 D 2
From integrations by parts, this equals
 
1 1
= (log(u) + 1)∇u · N dS − ∇u · ∇ log(u) dx
∂D 2 2 D
 
1 1 |∇u|2
= (log(u) + 1)∇u · N dS − dx .
∂D 2 2 D u
Taking the limit of large domains D which exhaust all of Rn , for example
the limit of large balls D = BR (0) as R → +∞, the conclusion is that
 
d 1 |∇u|2
u(t, x) log(u(t, x)) dx = − dx ≤ 0 .
dt Rn 2 Rn u
Therefore S(u(t, ·)) is a nonincreasing quantity associated with heat flow.

1 Freeman Dyson, Scientific American, September 1954.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 64 3. The heat equation

In the proof one identifies the rate of change of entropy as being


1 |∇u|2
(3.21) − dx ,
2 Rn u
and the integration by parts also identifies the entropy flux to be
1
1 + log(u) ∇u .
F (u) :=
2
This shows that the quantity S(u(t, ·)) is decreasing; by mathematics tra-
dition we have evidently chosen the opposite sign in definition (3.19) from
that in practice in physics, where entropy is defined to be −S(u) and is
increasing. The entropy S(f ) is not guaranteed to be finite for any choice
of f ∈ L1 (Rn ), but if it is initially so, then it will remain finite for u(t, ·) for
all future t > 0.
There are other quantities that are studied in relation to the concept
of entropy. One is the relative entropy of one probability distribution with
respect to another. Suppose that u(x) and e(x) are densities of two prob-
ability measures; the relative entropy of dP1 (x) = u(x)dx with respect to
dP2 (x) = e(x)dx is defined to be
  u(x)
S(u|e) := u(x) log dx .
Rn e(x)
Note that there is an asymmetry between the roles of u and e. Suppose that
u(t, x) is a solution of the heat equation and that e(t, x) = H(t, x) is given
by the heat kernel (3.17). The change in time of the relative entropy is then
expressed as
 
d u
u log dx = ∂t u log(u) − u log(e) dx .
dt Rn e Rn

The first term of the RHS we know from (3.21). The second term evolves
as follows:
   1
d ∂t e 1 Δe
u log(e) dx = ∂t u log(e) + u dx = Δu log(e) + u 2 dx
dt Rn R n e R n 2 e
 
1 ∇e 1  u ∇e 1 |∇e|2
=− ∇u · + ∇ · ∇e dx = − ∇u · − u 2 dx .
Rn 2 e 2 e Rn e 2 e
Since e(t, x) = H(t, x), the heat kernel, the latter quantity is explicit:
∇e x
=− .
e t
We can therefore compute that
  
∇e x n
∇u · dx = − ∇u · dx = u dx ,
Rn e Rn t t Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 3.8. Gradient flow 65

and furthermore
  
1 |∇e|2 1 |x|2 1
− u 2 dx = − u 2 dx = − 2 |x|2 u(t, x) dx .
R n 2 e R n 2 t 2t R n

The latter is the variance of the probability distribution dP1 = u(t, x)dx,
which from section 3.5 we know satisfies
   
1 1
∂t |x|2 u dx = |x|2 Δu dx = Δ(|x|2 ) u dx = n u(t, x) dx .
Rn 2 Rn 2 Rn Rn
Since by Proposition 3.9 the moment m0 (u(t, ·)) = m0 (f ) is a constant, a
conserved quantity, we then have that

∇e 1 |∇e|2
− ∇u · − u 2 dx
Rn e 2 e
 n  
n 1
=− u dx + u dx + 2 |x|2 f (x) dx .
t Rn 2t Rn 2t Rn
Putting these together, we have a quite explicit form for the evolution of
the relative entropy S(u|e), namely

d 1 |∇u|2 n 1
(3.22) S(u|e) = − dx − m0 (f ) − 2 m2 (f ) .
dt 2 Rn u 2t t

3.8. Gradient flow

This section of the chapter has to do with an interpretation of the heat
equations as a gradient flow for the energy functional

1
(3.23) E(u) = |∇u|2 dx .
4
To illustrate gradient flow with a finite-dimensional example, let
e(v) : Rd → R
be a C 1 function of v ∈ Rd . Its gradient ∇e(v) (defined in terms of the
Euclidean inner product on Rd ) is given by the formula
d 
 e(v + σV ) = ∇e(v), V  .
dσ σ=0
The (negative) gradient flow of e(v) is the solution map of the ODE
v̇ = −∇e(v) .
A basic property of gradient flow is that the function e(v) is decreasing along
trajectories; indeed,
d
e(v(t)) = ∇e(v), v̇(t) = −∇e(v), ∇e(v) = − ∇e(v) 2 ≤ 0 .
dt
Such flows play a role in Morse theory, among many other things.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 66 3. The heat equation

The heat equation defines a flow in space of functions. Using a similar

calculation, we can identify this as a gradient flow for the energy functional
E(u):

d  d  1
 E(u + σw) =  |∇u + σw|2 dx
dσ σ=0 dσ σ=0 4
d  1 
=  |∇u|2 + 2σ∇u · ∇w + σ 2 |∇w|2 dx
dσ σ=0 4

1
= ∇u · ∇w dx .
2
In order to describe this quantity in terms of the L2 inner product, one
integrates by parts:

d  1
(3.24)  E(u + σw) = − Δu w dx .
dσ σ=0 2
This identifies gradE(u) = − 12 Δu. The (negative) gradient flow is therefore
1
∂t u = −gradE(u) = Δu ,
2
which is precisely the heat equation (3.15). A classical notation is to write
δE(u) for the Gâteau derivative of E(u); that is, for the directional derivative
of E(u) in the direction w, one writes that
d 
 E(u + σw) = δE(u), w .
dσ σ=0

Exercises: Chapter 3
Exercise 3.1. This problem concerns the semigroup property of the solution
process for the heat equation. Show that the heat kernel satisfies the identity
 +∞
H(t, x) = H(t − s, x − y)H(y, s) dy ∀0 < s < t .
−∞
Now show that the heat operator satisfies the property that for all 0 < s < t,
H(t) = H(t − s)H(s) .
Conclude that formula (3.5) holds for the one-parameter family of solution
operators {H(t)}t∈R1+ of the heat equation.
Exercise 3.2. Justify on a rigorous level of analysis the exchange of integra-
tions in the proof of Proposition 3.4(iii), therefore completing the rigorous
proof of the proposition’s three parts.
Exercise 3.3. Solve the following initial value problems for the heat equa-
tion in explicit terms.
(1) f (x) = x

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 3 67

(2) f (x) = x2
(3) f (x) = 0 for x < 0 and f (x) = 1 for x ≥ 0
(4) f (x) = eαx
(5) f (x) = sin(kx)
What is the asymptotic behavior of u(t, x) as t → +∞? What is their
asymptotic behavior as t → +∞? On what properties of the initial data
does this behavior depend? Does it affect the asymptotic behavior if f (x) ∈
L1 (R1 )?
Exercise 3.4 (Moments of the heat kernel). (i) Show that the second mo-
ment of the heat kernel
1 − x2
H(t, x) = √ e 2t = Pt (x)
2πt
is its variance and that
m2 (Pt ) = t .
Hint: Calculate the following integral:
 +∞
1 x2
G(α) = √ e−α 2 dx ,
2π −∞
and then use the fact that
 +∞
1 x2
√ e− 2 x2 dx = −2∂α G(1) .
2π −∞

(ii) Also calculate all the moments mj (Pt ) for j any odd integer.
(iii) Using the same technique as in (i), give an expression for all of the
moments mj (Pt ) for j even.
Exercise 3.5. Suppose that the initial data f (x) ∈ L1 (R1 ) for the heat
equation is piecewise continuous, meaning that it contains a possibly finite
number of jump discontinuities
lim f (x) = aj , lim f (x) = bj
x→x−
j x→x+
j

at points xj , j = 1, N . Show that the limit of the solution (3.4) of the heat
equation satisfies
1
lim u(t, xj ) = (aj + bj ) .
t→0 + 2
Exercise 3.6 (Method of images). Give an expression for the solution of
the heat equation over the bounded interval 0 ≤ x < 2π for the following
cases of boundary conditions.
(1) Dirichlet: u(t, 0) = 0 = u(t, 2π). Denote this solution by uD (t, x).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 68 3. The heat equation

(2) Neumann: −∂x u(t, 0) = 0 = ∂x u(t, 2π). Denote this solution by

uN (t, x).
(3) Periodic: u(t, x) = u(t, x + 2π). Denote this solution by up (t, x).
Suppose that the initial data satisfies 0 ≤ f (x) ≤ 1 for the heat equation on
[0, 2π] (and assume that f is not identically zero). Show that in each case
u = uD , uN , and up , and for all t > 0 the solution satisfies the inequality
0 < u(t, x) < 1
(unless f (x) ≡ 0 or f (x) ≡ 1, in which case u(t, x) ≡ 0 or u(t, x) ≡ 1,
respectively).
Exercise 3.7. (i) Using the method of separation of variables, give an ex-
pression for the solution u(t, x) for t > 0 to the heat equation with Dirichlet
boundary conditions
u(t, 0) = 0 = u(t, 2π) ,
given the initial conditions
u(0, x) = f (x) = x .

(ii) Clearly u(0, 2π) = 0. Show that for all t > 0 the solution nevertheless
satisfies u(t, 2π) = 0.
Exercise 3.8. Consider the three solutions with initial data 0 ≤ f (x) for
the three different boundary conditions in Exercise 3.6. Show that for all
0 ≤ x ≤ 2π and all t > 0, the three solutions satisfy
uD (t, x) < up (t, x) and uD (t, x) < uN (t, x) .
Is it also true that
up (t, x) < uN (t, x) ?

Exercise 3.9. Suppose that the initial data f (x) ∈ L1 (Rn ), and consider
the solution u(t, x) given by convolution with the heat kernel. Show that
lim u(t, x) = 0
t→+∞
uniformly in x ∈ Rn . Furthermore, give a decay rate for the solution by
showing that it is bounded by
1
|u(t, x)| ≤ √ n f L1 (Rn ) .
2πt
Exercise 3.10 (Entropy). (i) Define two distributions by their densities:
f1 (x) = χ−1/2,+1/2(x) , f2 (x) = 2χ−1/4,+1/4 .
They both have total mass m0 (fj ) = 1 for j = 1, 2. Compare their entropies.
Which is larger (less disordered)?

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 3 69

(ii) What is the entropy of the heat kernel (3.3)

1 − x2
H(t, x) = √ e 2t dx ?
2πt
How does it behave as t → 0+ and as t → +∞?
Exercise 3.11. Suppose the functions f1 (x) ≥ 0 and f2 (x) ≥ 0 have disjoint
support, meaning that {x : f1 (x) > 0} ∩ {x : f2 (x) > 0} = ∅.
(i) Show that the entropy of the function f (x) = f1 (x) + f2 (x),
 +∞
S(f ) = f (x) log(f (x)) dx ,
−∞
satisfies
S(f ) = S(f1 ) + S(f2 ) .
(ii) Now suppose that u1 (t, x) and u2 (t, x) are the solutions of the heat
equation with initial data f1 and f2 , respectively. Let u(t, x) = u1 (t, x) +
u2 (t, x). Show that
S(u(t, ·)) ≥ S(u1 (t, ·)) + S(u2 (t, ·)) .
Under what condition is this inequality an equality?

Projects: Chapter 3
Project 3.12 (Tikhonov-nonuniqueness of solutions of the heat equation).
It turns out that solutions of the heat equations are not unique, at least
if one admits solutions that grow very rapidly and oscillate as |x| → +∞.
This is a surprising result that appeared in Tikhonov [20]. To show this we
will solve the heat equation with data on the positive t-axis in the form
1
∂t u = ∂x2 u , x ∈ R1 ,
2
u(t, 0) = g(t) , ∂x u(t, 0) = 0 , t ∈ R+ .

(i) Derive a formal power series expansion in x for a solution u(t, x) in the
form 
u(t, x) = c j xj ,
j≥0
and use the heat equation to show that for all j ∈ N,
2j j
c2j = ∂ g(t) , c2j+1 = 0 .
(2j)! t

(ii) Give functions g(t) such that this formal series for u(t, x) converges for
all x ∈ R1 . For example, functions in the Gevrey class will do, such as
g(t) = e− t2 ,
1
t>0.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 70 3. The heat equation

Show that for this g(t),

g(0) = 0 , ∂tj g(0) = 0
and that for all j ≥ 0,
(j!)3/2 − 12
|∂tj g(t)| ≤ e t .
(Ct)j
Show that under these conditions the power series for u(t, x) converges for
t > 0. Furthermore show that
u(0, x) = 0 .

Such solutions are oscillatory and badly unbounded for large x ∈ R1 . It

turns out that if we impose growth bounds on u(t, x), such as
2
|u(t, x)| ≤ Cex ,
then solutions are unique in this class, meaning that u(0, x) = 0 implies
u(t, x) = 0 for t > 0. Alternatively, if solutions are required to be bounded
from below, then they are also unique.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/04

Chapter 4

Laplace’s equation

Soap bubbles: surfaces of constant mean curvature supported by a slight

pressure difference across the interface. Image credit: Godunova Tatiana
/ iStock / Getty Images Plus.

4.1. Dirichlet, Poisson, and Neumann boundary value

problems
The most commonly occurring form of problem that is associated with
Laplace’s equation is a boundary value problem, normally posed on a do-
main Ω ⊆ Rn . That is, Ω is an open set of Rn whose boundary is smooth

71

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 72 4. Laplace’s equation

enough so that integrations by parts may be performed; thus at the very

least the boundary is rectifiable. The most common boundary value problem
for Laplace’s equation (1.6) is the Dirichlet problem
(4.1) Δu(x) = 0 , x∈Ω,
u(x) = f (x) , x ∈ ∂Ω ,
where one specifies the values of the solution u(x) on the boundary ∂Ω,
analogous to specifying the value of a solution of the wave or heat equation
on the boundary of its domain. The function f (x) is known as the Dirichlet
data. In physics applications, f (x) corresponds to a density distribution
of dipoles of charged particles fixed on the boundary ∂Ω, whereupon the
solution u(x) corresponds to the resulting electrostatic potential. A function
satisfying Δu = 0 is called harmonic, as we have stated in Chapter 1.
Perhaps the second most common problem is called the Poisson problem,
(4.2) Δu(x) = h(x) , x∈Ω,
u(x) = 0 , x ∈ ∂Ω ,
for which the function h(x) represents a distribution of fixed charges in the
domain Ω, while the boundary ∂Ω is a perfect conductor. Again the solu-
tion u(x) represents the resulting electrostatic potential. There are several
other quite common boundary value problems that are similar in character
to (4.1), for example the Neumann problem
(4.3) Δu(x) = 0 , x∈Ω,
∂N u(x) = g(x) , x ∈ ∂Ω ,
where N (x) is the outward unit normal vector to Ω and ∂N u(x) = ∇u(x)·N .
The solution corresponds to the electrostatic potential in Ω due to a charge
density distribution on ∂Ω. The Robin problem, or boundary value problem
of the third kind, asks to find u(x) such that
(4.4) Δu(x) = 0 , x∈Ω,
∂N u(x) − βu(x) = g(x) , x ∈ ∂Ω ,
where β is a real constant or possibly a real function of x ∈ Ω. This problem
is often associated with an imposed impedence on the boundary.

4.2. Green’s identities

Consider two functions u(x) and v(x) defined on a domain Ω ⊆ Rn . Inte-
gration by parts gives the following formula, which is known as Green’s first
identity:
  
(4.5) vΔu dx = − ∇v · ∇u dx + v∂N u dSx .
Ω Ω ∂Ω

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.2. Green’s identities 73

The notation is that dSx is the differential of surface area in the integral
over the boundary. To ensure that the manipulations in this formula are
valid, we take Ω ⊆ Rn to be bounded and ask that u, v ∈ C 2 (Ω) ∩ C 1 (Ω);
that is, all derivatives of u and v up to second order are continuous in the
interior of Ω and at least their first derivatives have continuous limits on the
boundary ∂Ω.
Integrating again by parts (or alternatively using (4.5) in a symmetric
way with the roles of u and v reversed), we obtain Green’s second identity
  
(4.6) vΔu dx − Δv u dx = v∂N u − ∂N vu dSx
Ω Ω ∂Ω

for u, v ∈ ∩C 2 (Ω) C 1 (Ω).

The integrand of the integral over the boundary
∂Ω is the analog of the Wronskian in ODEs.
Considering the function v as a test function and substituting several
astute choices for it into Green’s identities, we obtain information about
solutions u of Laplace’s equation. First of all, let v(x) = 1; then (4.5) gives
 
Δu dx = ∂N u dSx .
Ω ∂Ω
In case u is harmonic, Δu = 0 and the LHS vanishes. This is a compatibility
condition for boundary data g(x) = ∂N u for the Neumann problem.
Proposition 4.1. In order for the Neumann problem (4.3) to have a solu-
tion, the Neumann data g(x) must satisfy
 
g(x) dSx = ∂N u(x) dSx = 0 .
∂Ω ∂Ω

For a second choice, let v(x) = u(x) itself. Then Green’s first identity
(4.5) is an identity involving the L2 -norm of ∇u, which is a form of an
“energy” identity
  
(4.7) |∇u(x)|2 dx + uΔu dx = u∂N u dSx .
Ω Ω ∂Ω
One consequence of this is a uniqueness theorem.
Theorem 4.2. Suppose that u ∈ C 2 (Ω) ∩ C 1 (Ω) satisfies the Dirichlet prob-
lem (4.1) with f (x) = 0, the Poisson problem (4.2) with h(x) = 0, or the
Neumann problem (4.3) with g(x) = 0. Then
 
|∇u(x)| dx =
2
u∂N u dSx = 0 .
Ω ∂Ω
Therefore u(x) = 0 in the cases of the Dirichlet problem (4.3) and the Pois-
son problem (4.2). In the case of the Neumann problem, the conclusion is
that u(x) is constant.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 74 4. Laplace’s equation

Proof. Identity (4.7) implies that


|∇u(x)|2 dx = 0
Ω

in each of the three cases, which in turn implies that ∇u = 0 almost ev-
erywhere in Ω. Therefore u(x) must be constant as we have assumed that
u ∈ C 2 (Ω). In cases (4.1) and (4.2), this constant must vanish in order to
satisfy the boundary conditions. In the case of the Neumann problem (4.3),
we only conclude that u(x) is constant. In this situation we conclude that
the constant functions u(x) ≡ C span the null space of the Laplace operator
with Neumann boundary conditions. 

The case of unbounded domains Ω is not too different, and a uniqueness

result is the goal. One asks that solutions u(x) are bounded and have the
limit u(x) = 0 as x becomes unbounded.

4.3. The fundamental solution

The Laplace operator is invariant under rotations, and one can imagine that
solutions which are also rotationally invariant are of special interest. In polar
coordinates (r, ϕ) in Rn , where r ∈ [0, +∞) and ϕ ∈ Sn−1 , the Laplacian is
expressed
n−1 1
(4.8) Δu = ∂r2 + ∂r u + 2 Δϕ u ,
r r
where Δϕ is the Laplace operator on the unit sphere Sn−1 ⊆ Rn . A rota-
tionally invariant solution Γ(r) of Laplace’s equation must satisfy
n−1
∂r2 Γ + ∂r Γ = 0 ,
r
C
that is, ∂r Γ = r n−1
, which in turn implies that

C 1
Γ(r) = , n≥3,
2−nr n−2

Γ(r) = C log(r) , n=2.

The function Γ(r) is harmonic for 0 < r < +∞ but singular for r = 0. It
is called a fundamental solution, which is part of the content of Theorem 4.3.
Its proof is the outcome of the following computation that uses Green’s
identities. Suppose that u ∈ C 2 (Ω) ∩ C 1 (∂Ω) and consider y ∈ Ω a point
inside the domain under consideration. Take ρ > 0 sufficiently small so that
the ball Bρ (y) ⊆ Ω, as in Figure 1. Using Green’s second identity (4.6) over

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.3. The fundamental solution 75

x2

y
B (y)

x1

Figure 1. Domain Ω with a ball Bρ (y) excised.

the region Ω\Bρ (y), we have


(4.9) Γ(|x − y|)Δu dx
Ω\Bρ (y)
 
= Γ∂N u − ∂N Γu dSx + Γ∂N u − ∂N Γu dSx
∂Ω Sρ (y)

+ ΔΓ u dx .
Ω\Bρ (y)

The fundamental solution Γ is harmonic in Ω\Bρ (y), the singularity at

x = y being inside Bρ (y); therefore that last term of the RHS is zero. The
first term of the RHS is our usual boundary integral from (4.6). We are to
calculate the second integral and its limit as ρ → 0. On the sphere Sρ (y) we
have Γ(|x − y|) = Γ(ρ) while ∂N Γ(|x − y|) = −∂r Γ(ρ), the latter minus sign
coming from the fact that the outward unit normal to Ω\Bρ (y) on Sρ (y) is
pointing inwards towards y. Therefore, using the divergence theorem,
  
Γ∂N u dSx = Γ(ρ) N · ∇u dSx = −Γ(ρ) Δu dx ,
Sρ (y) Sρ (y) Bρ (y)

and the limit of this quantity vanishes as ρ → 0. Indeed, in case n ≥ 3,

  
  C 1 n
(4.10) lim Γ(ρ) Δu dx ≤ lim ρ u C2
ρ→0 Bρ (y) ρ→0 n − 2 ρn−2

which vanishes like ρ2 as ρ → 0. The case n = 2 also vanishes in the limit,

which involves log(ρ) instead. On the other hand, the second term of (4.9) is
 
C(2 − n)
− ∂N Γ u dSx = u dSx ,
Sρ (y) ρn−1 Sρ (y)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 76 4. Laplace’s equation

and as ρ tends to zero, by the continuity of u(x) at x = y, this has the limit
(4.11) 
C(2 − n) C(2 − n)
lim u dSx = lim ωn ρn−1 u(y) = C(2 − n)ωn u(y) ,
ρ→0 ρn−1 Sρ (y) ρ→0 ρ n−1

where the quantity ωn is the surface area of the unit sphere Sn−1 ⊆ Rn .
Theorem 4.3. For u ∈ C 2 (Ω) ∩ C 1 (∂Ω) and for y ∈ Ω we have the identity
 
(4.12) Γ(|x −y|)Δu(x) dx = C(2−n)ωn u(y) + Γ∂N u−∂N Γ u dSx .
Ω ∂Ω

Setting C = ((2−n)ωn )−1 , we recover precisely the value of u(y) with this
identity. In fact this procedure need not be restricted to bounded
 domains
Ω. If we specify that u ∈ C (R ) is such that limR→+∞ SR (0) (Γ∂N u −
2 n

∂N Γ u) dSx = 0 for y ∈ Rn , then the statement of Theorem 4.3 can be

interpreted distributionally in terms of the Dirac δ-function:
Δx Γ(|x − y|) = δy (x) .

Formula (4.12) has further implications having to do with the smooth-

ness of solutions of Laplace’s equation away from the boundary of their
domain of definition, a topic known as the interior regularity of elliptic
equations. Away from its pole at x = y, the fundamental solution Γ is C ∞
smooth; in fact it is C ω , which is the notation for the class of real analytic
functions. The results of Theorem 4.3 can be interpreted to deduce that
a harmonic function is C ∞ , indeed even real analytic, in the interior of its
domain of definition; this is the conclusion of the following theorem.
Theorem 4.4. Let u ∈ C 2 (Ω) ∩ C 1 (∂Ω) be a harmonic function in Ω. Then
in fact u ∈ C ω .

Proof. For harmonic functions u(x), identity (4.12) reads


u(y) = u ∂N Γ − ∂N u Γ dSx .
∂Ω
Since y ∈ Ω in this expression, while the boundary integral only involves
x ∈ ∂Ω, then dist(x, y) is bounded from below and the RHS is clearly a C ω
function of y, hence the result. 

We may add any harmonic function w(x) to the fundamental solution

Γ(|x − y|) and retain the above properties, as
Δx (Γ + w) = Δx Γ + Δx w = δy (x) .
In particular consider any ball Bρ (y) ⊆ Ω within the domain of definition of a
harmonic function u(x). Use the quantity Γ(|x − y|) − Γ(ρ) as a fundamental

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.4. Maximum principle 77

solution in (4.12) to find that

  

u(y) = Γ(|x − y|) − Γ(ρ) ∂N u dSx − u ∂ r Γ dSx
Sρ (y) Sρ (y) r=ρ

1
= u(x) dSx .
Sρ (y) ωn ρn−1
This proves the following statement.
Theorem 4.5 (Gauss’ law of arithmetic mean). A harmonic function sat-
isfies the integral identity
 
1 1
(4.13) u(y) = u(x) dSx = u(y + ρv) dSv
ωn ρn−1 Sρ (y) ωn v∈S1 (0)
for all Bρ (y) contained in the domain of definition of u. An alternative
version of this statement is that

n
(4.14) u(y) = u(x) dx .
ωn ρn Bρ (y)

In words, this result states that a harmonic function u(y) is equal to its
average values over spheres Sρ (y) about y. This fact is consistent with dis-
cretizations of the Laplace operator in numerical simulations. For example,
the simplest version of a finite difference approximation of the Laplacian at
y ∈ Zn is
1 
Δh u(y) := 2 u(x) − 2nu(y) .
h n x∈Z :|x−y|=1
A discrete harmonic function Δh u = 0 explicitly satisfies the property that
u(y) is equal to the average of its values at the nearest neighbor points
|x − y| = 1.

4.4. Maximum principle

The maximum principle is a recurring theme in the theory of elliptic and
parabolic PDEs. It provides a strong and ubiquitous method to assure
the uniqueness of solutions and to provide estimates on them and their
derivatives.
Theorem 4.6 (Weak maximum principle). Suppose that Ω is a bounded
domain and that u(x) ∈ C 2 (Ω) ∩ C 0 (Ω) satisfies
(4.15) Δu ≥ 0
for x ∈ Ω. Then the maximum value of u(x) over the set Ω is achieved for
some x ∈ ∂Ω; namely, for all x ∈ Ω,
(4.16) u(x) ≤ max u(x) := M .
x∈∂Ω

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 78 4. Laplace’s equation

If u(x) ∈ C 2 (ω) ∩ C 0 (Ω) is harmonic, then both u and −u satisfy the

hypotheses of the theorem, and therefore there are upper and lower bounds
on u(x) that only depend upon its boundary values; namely, for x ∈ Ω,
min u(x) ≤ u(x) ≤ max u(x) .
x∈∂Ω x∈∂Ω

Proof. First of all, start with the stronger assumption of strict inequality,
Δu(x) > 0 in Ω. Suppose that at some point x0 ∈ Ω the function u(x)
achieves its maximum, u(x0 ) = maxx∈Ω (u(x)). Then ∇u(x0 ) = 0 and the
Hessian matrix of u satisfies
n
H(u) = ∂xj ∂x u(x0 ) j=1
≤0.

However, this contradicts the fact that Δu > 0; indeed,

Δu(x) = tr H(u) > 0 .
Hence no such maximum point can exist in Ω, and this argument even
rules out local maxima. For the general case where we assume only that
Δu ≥ 0 in Ω, consider the function v(x) = u(x) + ε|x|2 , which satisfies
Δv(x) = Δu(x) + 2εn > 0. Therefore v is covered by the hypotheses of the
first case, from which we conclude
max(u(x)) ≤ max(u(x) + ε|x|2 )
x∈Ω x∈Ω
= max (u(x) + ε|x|2 ) ≤ max (u(x)) + ε max (|x|2 ) .
x∈∂Ω x∈∂Ω x∈∂Ω

Since Ω is bounded, maxx∈∂Ω (|x|2 ) is finite, and since ε is arbitrary, we have

shown that
max(u(x)) ≤ max (u(x)) .
x∈Ω x∈∂Ω

The advantage of this proof is that it immediately generalizes to many

other elliptic equations, including

n 
n
aj (x)∂xj ∂x u + bj (x)∂xj u = 0 ,
j,=1 j=1

where the matrices (aj (x))nj,=1 are positive definite. The disadvantage is
that it relies upon the hypothesis that u(x) ∈ C 2 (Ω), which in the present
situation is stronger than necessary. The result on Gauss’ law of arithmetic
mean (4.13) in the previous section allows us to discuss maximum principles
for functions which are not C 2 (Ω) nor even C 1 (Ω), but which merely are
continuous.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.4. Maximum principle 79

Definition 4.7. A function u ∈ C(Ω) is subharmonic in Ω if for all y ∈ Ω

and all Bρ (y) ⊆ Ω, one has

1
u(y) ≤ u(x) dSx .
ωn ρn−1 Sρ (y)
A function u ∈ C(Ω) is superharmonic in Ω if for all y ∈ Ω and all Bρ (y) ⊆ Ω,
the opposite inequality holds:

1
u(y) ≥ u(x) dSx .
ωn ρn−1 Sρ (y)

This nomenclature is motivated by the fact that the graph of a subhar-

monic function lies beneath the graph of the harmonic function with the
same boundary values on ∂Ω, as we will show below. Similarly, superhar-
monic functions lie above harmonic functions sharing their boundary values.
Furthermore, if we also knew that u ∈ C 2 (Ω), then the inequality
Δu(x) ≥ 0 , respectively, Δu(x) ≤ 0 ,
implies that u(x) is subharmonic (respectively, superharmonic).
It turns out that subharmonic functions u ∈ C(Ω) satisfy the maximum
principle, a statement that substantially lessens the necessary hypotheses for
maximum principles and proves to be a very useful principle in nonlinear
problems and in other generalizations of Laplace’s equation.
Theorem 4.8 (Maximum principles for subharmonic functions). Let Ω be
a bounded domain and suppose that u ∈ C(Ω) is subharmonic. Then
(4.17) max(u(x)) = max (u(x)) .
x∈Ω x∈∂Ω

Furthermore, if Ω is connected, then either for all x ∈ Ω we have

(4.18) u(x) < max (u(x)) ,
x∈∂Ω

or else, if at some point x ∈ Ω equality holds, then u(x) is necessarily a

constant function u(x) ≡ M := maxx∈∂Ω (u(x)).

The latter result is known as the strong maximum principle for subhar-
monic function. If u(x) is superharmonic, then −u(x) is subharmonic, and
therefore u satisfies a minimum principle. A corollary of Theorem 4.8 is that
harmonic functions, which are both sub- and superharmonic, satisfy upper
and lower estimates in the supremum norm. Namely, if u(x) is harmonic in
Ω and u(x) = f (x) for x ∈ ∂Ω, then the weak maximum principle implies
that
min (f (x)) ≤ u(x) ≤ max (f (x))
x∈∂Ω x∈∂Ω
for all x ∈ Ω.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 80 4. Laplace’s equation

Corollary 4.9 (Strong maximum principle). Let Ω be a bounded connected

domain, and suppose that u ∈ C(Ω) is harmonic and satisfies u(x) = f (x)
for x ∈ ∂Ω. That is,
Δu = 0 for x ∈ Ω .
Then for all x ∈ Ω,
min (f (x)) < u(x) < max (f (x)) .
x∈∂Ω x∈∂Ω

Or else, if equality holds at any point x ∈ Ω instead of either of these strict

inequalities, then
u(x) ≡ M := min (f (x)) = max (f (x)) .
x∈∂Ω x∈∂Ω

Proof of Theorem 4.8. The proof of this second version of the maximum
principle is more specific to Laplace’s equation, using the property of subhar-
monicity. Consider Ω a domain which is connected and u(x) a subharmonic
function, and set M := supx∈Ω (u(x)). Decompose the domain into two
disjoint subsets:
Ω = {x ∈ Ω : u(x) = M } ∪ {x ∈ Ω : u(x) < M } := Ω1 ∪ Ω2 .
Since u(x) ∈ C(Ω), then Ω2 is open as a subset of Ω because an inequality
is an open condition. The claim is that the set Ω1 is also open in Ω, which
we will prove using the property of subharmonicity. Therefore, because of
the fact of being connected, either Ω = Ω2 and Ω1 = ∅, whereupon strict
inequality holds throughout Ω, or else Ω = Ω1 and Ω2 = ∅, in which case
u(x) = M a constant.
To prove the claim above that Ω1 is open, consider x0 ∈ Ω1 and refer
to the property of subharmonicity on all sufficiently small spheres Sρ (x0 )
about x0 which lie in Ω. This property implies that

1
0≤ u(x) − u(x0 ) dSx .
ωn ρn−1 |x−x0 |=ρ
The integrand is nonpositive because u(x) ≤ M = u(x0 ) in Ω. But the
integrand cannot be negative anywhere near x0 either, for that would violate
the above inequality. Therefore we must have Sρ (x0 ) ⊆ Ω1 for all sufficiently
small ρ, which is the statement that Ω1 is open. 

4.5. Green’s functions and Dirichlet–Neumann operators

As we observed in section 4.3, we may add any harmonic function to the
fundamental solution Γ(|x − y|) and retain the property that Δ(Γ + w) = δy .
This is useful because the fundamental solution Γ(|x −y|) doesn’t satisfy the
proper Dirichlet boundary conditions on domains Ω, with the only exception

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.5. Green’s functions and Dirichlet–Neumann operators 81

being the case Ω = Rn . Thus we may add a function w(x, y) ∈ C 2 (Ω × Ω)

that satisfies
Δx w(x, y) = 0 for x∈Ω,
w(x, y) = −Γ(|x − y|) for x ∈ ∂Ω .
The function w(x, y) is harmonic in x ∈ Ω and depends parametrically on
y ∈ Ω. Then
G(x, y) := Γ(|x − y|) + w(x, y)
is still a fundamental solution in the distributional sense as in (4.12), mean-
ing that Δx G(x, y) = δy (x). Furthermore G(x, y) = 0 for x ∈ ∂Ω, and
therefore 
u(x) := G(x, y)h(y) dy
Ω
satisfies

Δx u(x) = Δx G(x, y)h(y) dy
 Ω

= Δx Γ(|x − y|)h(y) dy + Δx w(x, y)h(y) dy = h(x) .
Ω Ω
It also satisfies Dirichlet boundary conditions; indeed, for x ∈ ∂Ω,

u(x) = G(x, y)h(y) dy = 0 .
Ω
The function G(x, y) is called the Green’s function for the domain Ω; it is
the integral kernel of the solution operator P for the Poisson problem (4.2),

(4.19) u(x) = G(x, y)h(y) dy := Ph(x) .
Ω

The Green’s function for Rn+ . While it is usually not possible to have
explicit formulae for the Green’s function G(x, y) for a general domain, it
is straightforward for the domain consisting of the half space Rn+ . Let y =
(y  , yn ) ∈ Rn+ so that yn > 0, and denote the point that is its reflection
through the boundary {xn = 0} by y ∗ := (y  , −yn ) ∈ Rn− . The function
w(x, y) = −Γ(|x − y ∗ |) satisfies the property that it is harmonic in Rn+
(since its singularity at x = y ∗ is in Rn− ), and Γ(|x − y ∗ |) = Γ(|x − y|) when
x = (x , 0) ∈ ∂Rn+ . This gives rise to the following expression for the Green’s
function for the domain Rn+ :
(4.20) G(x, y) = Γ(|x − y|) − Γ(|x − y ∗ |)
1  1
=
(2 − n)ωn (|x − y | + (xn − yn )2 )(n−2)/2
  2

1
−  
(|x − y | + (xn + yn )2 )(n−2)/2
2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 82 4. Laplace’s equation

x3

x’

x2

x1

Figure 2. Conformal inversion of the point x ∈ BR (0) whose image is x∗ .

when n ≥ 3. When n = 2,
1  (x − y )2 + (x − y )2
1 1 2 2
(4.21) G(x, y) = log .
4π (x1 − y1 )2 + (x2 + y2 )2
This procedure is similar to the method of images, as one can see. Actually,
in two dimensions there is a connection with the theory of complex variables;
because of the Riemann mapping theorem there is an expression for the
Green’s function for arbitrary simply connected domains in terms of formula
(4.21) and the Riemann mapping of the domain to the upper half plane R2+ .

The Green’s function for BR (0). A second case of a domain Ω for which
the Green’s function has an explicit expression is the ball BR (0) ⊆ Rn .
Given x, y ∈ Ω and another point y ∗ ∈ Ω, the function constructed from the
fundamental solution Γ(|x − y|) given by
G(x, y) = Γ(|x − y|) − c(y)Γ(|x − y ∗ |)
is harmonic as a function of x ∈ Ω,
Δx G(x, y) = 0 , x=y ,
with the singularity of the fundamental solution at x = y. In the special
case of Ω = BR (0), one may choose the exterior point y ∗ to be the inversion
of y in the sphere Sn−1
R = ∂BR (0), given by
R2
y∗ = y.
|y|2
For x ∈ Sn−1
R , this choice satisfies the special geometrical property that
|x − y ∗ | R
(4.22) = ,
|x − y| |y|

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.5. Green’s functions and Dirichlet–Neumann operators 83

the RHS being an expression independent of x. This is illustrated in Figure

R n−2
2. Therefore if we take c(y) = |y| , then for x ∈ Sn−1
R we will have
1 1
c(y)Γ(|x − y ∗ |) = c(y)
(2 − n)ωn |x − y ∗ |n−2
1 1  |y| n−2 1 1
= c(y) = ,
(2 − n)ωn |x − y| n−2 R (2 − n)ωn |x − y|n−2
and therefore G(x, y) will have the properties that it is harmonic in the x-
variable and that G(x, y) = 0 for x ∈ Sn−1R . This gives the expression for
the Green’s function for the ball BR (0),
G(x, y) = Γ(|x − y|) − c(y)Γ(|x − y ∗ |)
1  1 Rn−2 1
(4.23) = − .
(2 − n)ωn |x − y| n−2 |y| n−2 |x − y ∗ |n−2
Although it is not so evident for the Green’s function for an arbitrary
domain Ω, or from the way we constructed the Green’s function for special
domains Ω, the Green’s function has the symmetry that G(x, y) = G(y, x).
This reflects the property of the Laplace operator with Dirichlet boundary
conditions being a self-adjoint operator.1 This property can be verified
explicitly for the Green’s functions for the upper half space and for BR (0)
upon inspecting the formulae in (4.20), (4.21), and (4.23).

The Poisson kernel. The Green’s function G(x, y) is the integral kernel
for the solution of the Poisson problem (4.2), but it also is relevant for the
Dirichlet problem (4.1). Let u ∈ C 2 (Ω) ∩ C 1 (Ω) be harmonic. Then by
Green’s identities

0= G(x, y)Δu(x) dx
Ω

= u(y) + G(x, y)∂N u(x) − ∂Nx G(x, y)u(x) dSx .
∂Ω
Using the fact that G(x, y) = 0 for x ∈ ∂Ω, we deduce that

(4.24) u(y) = u(x)∂Nx G(x, y) dSx
 ∂Ω

= f (x)∂Nx G(x, y) dSx := Df (y) ,

∂Ω
where Df (y) is the operator giving the solution to the Dirichlet problem
(4.1) for f (x) Dirichlet data for u(x) on the boundary ∂Ω. The quantity
D(x, y) := ∂Nx G(x, y) in the integrand is called the Poisson kernel for the
domain Ω.
1 Technically, the Laplace operator is a symmetric operator which is self-adjoint when re-

stricted to an appropriate subspace H01 (Ω) ⊆ L2 (Ω).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 84 4. Laplace’s equation

In the case that Ω = Rn+ , the explicit form of the Green’s function (4.20)
(and (4.21) in the case in which n = 2) gives an equally explicit expression
for the Poisson kernel. This is
 2  yn
(4.25) −∂xn G(x, y)xn =0 = D(x − y  , yn ) =  
.
ωn (|x − y |2 + yn2 )n/2
In the case of the ball Ω = BR (0), the Poisson kernel in (4.24) is given
by the expression
1  R2 − |y|2 
(4.26) D(x, y) = ,
Rωn |x − y|n
as one finds from (4.22) and (4.23) after a short calculation.

The Neumann problem. There is a related expression for the analog

of the Green’s function for the Neumann problem, which can be found by
choosing another function w(x, y) which is harmonic in x and forming the
fundamental solution
N (x, y) = Γ(|x − y|) + w(x, y) .
If w(x, y) is chosen so that ∂N Γ(|x − y|) = −∂N w(x, y) for all x ∈ ∂Ω, then
the function N (x, y) is the integral kernel for the solution operator of the
Neumann problem (4.3). That is, the analog of the formula  (4.24) holds.
On general domains, given Neumann data g(x) such that ∂Ω g(x) dSx = 0,

(4.27) u(y) = − N (x, y)∂N u(x) dSx
 ∂Ω

=− N (x, y)g(x) dSx := Ng(y) ,

∂Ω
analogous to (4.24), as one shows with a calculation using Green’s second
identity. In the case that Ω = Rn+ , the choice is that w(x, y) = Γ(|x − y ∗ |),
and
(4.28) N (x, y) = Γ(|x − y|) + Γ(|x − y ∗ |)
1  1
=
(2 − n)ωn (|x − y | + (xn − yn )2 )(n−2)/2
  2

1
+  
(|x − y | + (xn + yn )2 )(n−2)/2
2

when n ≥ 3. When n = 2,
(4.29)
1 
N (x, y) = log (x1 − y1 )2 + (x2 − y2 )2 (x1 − y1 )2 + (x2 + y2 )2 .
4π
Setting xn = 0 gives rise to the single layer potential
2  1
(4.30) S(x − y  , yn ) =
(2 − n)ωn (|x − y | + yN
  2 2 )(n−2)/2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.5. Green’s functions and Dirichlet–Neumann operators 85

with which one solves the Neumann boundary value problem with data g(x),
namely 
u(y) = S(x − y  , xn )g(x ) dx := Sg(y) .
x ∈Rn−1
The physical interpretation of this expression is that it represents the elec-
trostatic potential induced by a charge distribution of density g(x ) on the
boundary ∂Rn+ = Rn−1 of the domain Rn+ . Similarly, the double layer po-
tential is given by

1  yn 1
u(y) =  
f (x ) dx = Df (y) ,
x ∈Rn−1 ωn (|x − y | + yn )
2 2 n/2 2
and its interpretation is the resulting electrostatic potential from a distri-
bution of dipole charges with density f (x ) on the boundary of Rn+ .2 The
potential associated to a distribution ρ(x) of charge in Rn is given by a
convolution with the fundamental solution itself

u(y) = Γ(x − y)ρ(x) dx .
Rn

The Dirichlet–Neumann operator. Given Dirichlet data on the bound-

ary of a domain Ω, it is often the most important part of the solution process
for u(x) of the Dirichlet problem to recover the normal derivatives of the
solution ∂N u(x) on the boundary. For Ω a conducting body, this would be
the map from applied voltage u(x) = f (x) on ∂Ω to the resulting current
∂N u(x) = g(x) across the boundary. This map can be expressed in terms of
the Green’s function; in particular, using the Poisson kernel we can express
the solution to the Dirichlet problem,

u(y) = (Nx · ∇x )G(x, y)f (x) dSx .
∂Ω
Therefore we have an expression for its normal derivative on ∂Ω, namely
(4.31) 


(Ny · ∇y )u(y) = (Nx · ∇x )(Ny · ∇y )G(x, y)f (x) dSx := Gf (y) ,
y∈∂Ω ∂Ω

2 In more generality, the single layer potential on a codimension-one hypersurface Λ is given

by

1
u(x) = g(y) dSy ,
Λ |x − y|n−2
corresponding to the potential u(x) due to a charge distribution of density f (x) on Λ. The double
layer potential is given by

1
u(x) = g(y) N · ∂y dSy ,
Λ |x − y|n−2
which is interpreted physically as the potential due to a dipole distribution along Λ with density
g(x). In the case that Λ = Rn−1 ⊆ Rn , these two coincide with the kernel for the Neumann
problem and the Poisson kernel, respectively. However, this relationship does not hold in general.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 86 4. Laplace’s equation

where G is the Dirichlet–Neumann operator for the domain Ω. There is

symmetry in the exchange of x with y in the integrand of (4.31), from which
we deduce that the Dirichlet–Neumann operator is self-adjoint;3 GT = G.
Recall the energy identity for a harmonic function u(x),
  
(4.32) 0 ≤ |∇u(x)| dx =
2
u(x)∂N u(x) dSx = f (x)(Gf )(x) dSx .
Ω ∂Ω ∂Ω
This is to say that the operator G is nonnegative definite as well as self-
adjoint. Formula (4.32) also exhibits the relation between the Dirichlet
integral of a harmonic function over a domain Ω and the boundary integral
over ∂Ω involving the Dirichlet–Neumann operator.

4.6. Poisson kernel on Rn+

This section is dedicated to techniques based on the Fourier transform. For
Laplace’s equation we can do this directly only in particular cases, the most
straightforward being that the domain Ω = Rn+ := {x = (x1 , . . . , xn ) ∈ Rn :
xn > 0}, the half space, and this is the situation that we will discuss. It
would also be possible to directly use similar techniques based on Fourier
series to solve Laplace’s equation on the disk D2 := {x ∈ R2 : |x| < 1} or
the polydisc D2n := {z = (x1 , . . . , xn , y1 , . . . , yn ) ∈ R2n : (x2j + yj2 ) < 1 , j =
1, . . . , n}. We will, however, stick with the half space Rn+ .
The Dirichlet problem (4.1) on Rn+ is to solve
(4.33) Δu(x) = 0 , x = (x1 , . . . , xn ) ∈ Rn+ ,
u(x) = f (x ) , x = (x , 0) , x ∈ Rn−1 = ∂Rn+ .
We will assume that u(x) and ∇u(x) tend to zero as xn → +∞. For the
 
special boundary data f (x ) = eiξ ·x with ξ  ∈ Rn−1 , there are explicit
solutions
  
(4.34) u(x , xn ) = eiξ ·x e−|ξ |xn
since

n−1
     
Δu(x , xn ) = ∂x2j eiξ ·x e−|ξ |xn + ∂x2n eiξ ·x e−|ξ |xn
j=1


n−1
  
−ξ  j + |ξ  |2 eiξ ·x e−|ξ |xn = 0 .
2
=
j=1
  
The other possible solution is eiξ ·x e+|ξ |xn , but this is ruled out by its growth
as xn → +∞. The Fourier transform allows us to decompose a general
3 Analogous
to the case above, technically the operator G is symmetric on C 1 (∂Ω) and self-
adjoint on an appropriate subspace H 1/2 (∂Ω) ⊆ L2 (∂Ω).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.6. Poisson kernel on Rn+ 87

function f (x ) on the boundary (in L2 (Rn−1 ) or perhaps in L1 (Rn−1 )) into

a composite of complex exponentials

1  
f (x ) = √ n−1 eiξ ·x fˆ(ξ  ) dξ  .

2π
By using (4.34), the general solution u(x) can be expressed as a super-
position of the above special solutions with boundary data expressed via the
Fourier transform, namely

1   
u(x) = √ n−1 eiξ ·x e−|ξ |xn fˆ(ξ  ) dξ 
2π
 
1    
= n−1
eiξ ·(x −y ) e−|ξ |xn dξ  f (y  ) dy 
(2π)

(4.35) = D(x − y  , xn )f (y  ) dy  = Df (x) .

The function D(x , xn ) is an integral kernel for the solution operator, often
called the Poisson kernel for Rn+ or the double layer potential, and the solu-
tion u(x) is evidently given by convolution of the Dirichlet boundary data
with D(x , xn ). Evaluating the above Fourier integral expression (4.35), we
get an explicit expression:

 1   
D(x , xn ) = n−1
eiξ ·x e−|ξ |xn dξ 
(2π)
2  xn
(4.36) = 
.
ωn (|x | + x2n )n/2
2

The second line of (4.36) has already been derived as (4.25) in section 4.5
using the method of images. The constant ωn is the surface area of the unit
sphere in Rn .
Theorem 4.10. The solution of the Dirichlet problem on Rn+ with data
f (x ) ∈ L2 (Rn−1 ) is given by

(4.37) u(x) = D(x − y, xn )f (y  ) dy 
R n−1
 
1    
= eiξ ·(x −y ) e−|ξ |xn dξ  f (y  ) dy  .
(2π)n−1 Rn−1 Rn−1
For xn > 0, this function is C ∞ (differentiable an arbitrary number of
times).

Proof. For xn > 0, both of the above integrals in (4.37) converge absolutely,
as indeed we have
 
 iξ ·x −|ξ |x 
e e n ˆ  
f (ξ ) dξ  ≤ fˆ L2 (Rn−1 )

e−2|ξ |xn dξ 
1/2

Rn−1 Rn−1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 88 4. Laplace’s equation

(where we have used the Cauchy–Schwarz inequality and the Plancherel

identity (3.16)), and hence we also learn that the solution u(x) has an upper
bound in Rn+ which quantifies its decay rate as xn → +∞. Namely,

  C
e−2|ξ |xn dξ 
1/2 ˆ
|u(x , xn )| ≤ f L2 (Rn−1 ) ≤ f L2 (Rn−1 ) .
Rn−1 |xn | (n−1)/2

Further derivatives of expression (4.37) do not change the properties of ab-

solute convergence of the integral for xn > 0, and we verify that the function
we have produced is indeed harmonic:

1   
Δu(x) = √ n−1 Δ eiξ ·x e−|ξ |xn fˆ(ξ  ) dξ  = 0 .
2π Rn−1

The issue is whether the harmonic function u(x , xn ) that we have pro-
duced converges to f (x ) as xn → 0, and in what sense. In this paragraph
we will show that for every xn > 0, u(x , xn ) is an L2 (Rn−1 ) function of the
horizontal variables x ∈ Rn−1 , and that u(x , xn ) → f (x ) in the L2 (Rn−1 )
sense as xn → 0. By Plancherel,

u(x , xn ) − f (x ) L2 (Rn−1 ) = e−|ξ |xn fˆ(ξ  ) − fˆ(ξ  ) L2 (Rn−1 ) .

Because fˆ L2 < +∞, for any δ > 0 there is a (possibly large) R > 0 such
that 
|fˆ(ξ  )|2 dξ  < δ .
|ξ  |>R

We now estimate

u(x , xn ) − f (x ) 2L2 (Rn−1 ) = e−|ξ |xn − 1 fˆ(ξ  ) 2L2 (Rn−1 )
 
 −|ξ |x 2  −|ξ |x 2
=  e n
− 1 fˆ(ξ  ) dξ  +  e n
− 1 fˆ(ξ  ) dξ 
|ξ  |≤R |ξ  |>R

 
≤ e−Rxn − 1 |fˆ(ξ  )|2 dξ  + 2δ .
|ξ  |≤R

The first term of the RHS vanishes in the limit as δ → 0. Since δ is arbitrary,
we conclude that u(x , xn ) converges to f (x ) as xn → 0 in the L2 (Rn−1 )
sense, which is that the L2 -norm of their difference tends to zero. 

Dirichlet–Neumann operator. It is useful to work out in detail the form

of the Dirichlet–Neumann operator on the domain Rn+ . The solution to the
Dirichlet problem for Rn+ is given in (4.37) in terms of the Fourier transform
of the Poisson kernel:

1   
u(x) = √ n−1 eiξ ·x e−|ξ |xn fˆ(ξ  ) dξ  .
2π Rn−1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.6. Poisson kernel on Rn+ 89

Recall that 1
i ∂xj u(ξ) = ξj û(ξ), which motivates the notation for differential
operators that Dj = 1i ∂xj and the definition of a general Fourier multiplier
operator

1  
m(D  )f (x ) = √ n−1 eiξ ·x m(ξ  )fˆ(ξ  ) dξ 
2π Rn−1

1   
= n−1
eiξ ·(x −y ) m(ξ  )f (y  ) dξ  dy  .
(2π) Rn−1 ×Rn−1

In these terms the harmonic extension u(x) of the boundary conditions f (x )
can be written as

u(x) = e−xn |D | f (x ) := (Df )(x , xn ) .

The operator D extends the Dirichlet data f (x ) to a harmonic function

u(x) = Df (x , xn ) in the upper half space Rn+ . This is an interpretation
of formula (4.37) which uses a Fourier integral expression for the Poisson
kernel.
Differentiating (4.37) with respect to xn and evaluating the result on the
boundary {xn = 0}, the Dirichlet–Neumann operator has a related Fourier
integral expression, namely
 
 1
Gf (x ) = −∂xn xn =0 √ n−1
  
(4.38) eiξ ·x e−|ξ |xn fˆ(ξ  ) dξ 
2π Rn−1
 
= |D |f (x ) .

Finally, the Dirichlet integral can be expressed in terms of Fourier multipli-

ers, using that

1   
∂x u(x , xn ) = √ n−1 eiξ ·x iξ  e−|ξ |xn fˆ(ξ  ) dξ  ,
2π R n−1

1   
∂xn u(x , xn ) = √ n−1 eiξ ·x |ξ  |e−|ξ |xn fˆ(ξ  ) dξ  .
2π Rn−1

Therefore

|∇u(x)|2 dx
Rn
+
 +∞ 
 
= |iξ  fˆ(ξ  )|2 e−2|ξ |xn + |ξ  |2 |fˆ(ξ  )|2 e−2|ξ |xn dξ  dxn ,
0 Rn−1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 90 4. Laplace’s equation

where we have used the Plancherel identity on the hyperplanes {(x , xn ) :

xn = Constant}. Thus
   +∞
 2 ˆ  2 
|∇u(x)| dx =
2
2|ξ | |f (ξ )| e−2|ξ |xn dxn dξ
Rn Rn−1 0
+

= f (x ) |D  |f (x ) dx .
Rn−1

4.7. Maximum principle again

A property that is evident of the Poisson kernel in (4.25) is that for xn > 0,
2  xn
D(x − y  , xn ) = 
>0.
ωn (|x | + x2n )n/2
2

Therefore whenever a solution of the Dirichlet problem (4.33) has the prop-
erty that f (x ) ≥ 0,

u(x) = D(x − y  , xn )f (y  ) dy  > 0 ;
Rn−1
this follows from an argument that is very similar to the one we used for
Theorem 3.5 on the heat equation.
Theorem 4.11. Suppose that f (x ) ≥ 0; then for x ∈ Rn+ we have u(x) ≥ 0,
and in fact if at any point x ∈ Rn+ (meaning, with xn > 0) it happens that
u(x) = 0, then we conclude that u(x) ≡ 0 and f (x ) = 0.

From this result we have a comparison between solutions.

Corollary 4.12. Suppose that f1 (x ) ≤ f2 (x ) for all x ∈ Rn−1 . Then
either
u1 (x) = (D ∗ f1 )(x) < u2 (x) = (D ∗ f2 )(x)
on all of R+ , or else u1 (x) ≡ u2 (x) if equality holds at any point x ∈ Rn+ .
n

In particular if f1 = f2 , then both f1 ≤ f2 and f1 ≥ f2 so that u1 ≡ u2 .

This is our third encounter with the recurring theme in elliptic and
parabolic PDEs of comparison and maximum principles. In the case of
Laplace’s equation and other elliptic equations, analog results closely hold
on essentially arbitrary domains as well, although the Poisson kernel is not
generally so explicit and the proof is different. In our present setting, the
form of the Poisson kernel gives us a lower bound on the decay rates of
solutions u(x , xn ) for large xn .
Corollary 4.13. Suppose that f (x ) ≥ 0 and that A ⊆ {x : f (x ) ≥ δ} is
a bounded set of positive measure meas(A) > 0. Then
2  xn
u(x , xn ) ≥ δ meas(A) ,
ωn supy ∈A (|x − y  |2 + x2n )n/2


Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.8. Oscillation and attenuation estimates 91

and, in particular, solutions u(x , xn ) that are positive have a minimum

decay rate that is bounded below by C|xn |−n ; that is, they cannot decay too
rapidly as xn → +∞.

Proof. Express the solution in terms of the Poisson kernel


u(x) = D(x − y  , xn )f (y  ) dy 
R
n−1

≥ D(x − y  , xn )f (y  ) dy 
A

 
≥ inf

D(x − y , xn ) f (y  ) dy  ,
y ∈A A

and of course 
f (y  ) dy  ≥ δ meas(A)
A
while  xn
inf D(x − y  , xn ) ≥ .

y ∈A supy ∈A (|x − y  |2 + x2n )n/2


4.8. Oscillation and attenuation estimates

Another principle exhibited by solutions of Laplace’s equation is the prop-
erty of attenuation of oscillatory data. This is a feature that is related to
the interior elliptic regularity of solutions of elliptic equations. It also is very
relevant to applications such as imaging strategies in electrical impedence
tomography, which is a medical imaging technique where the idea is to use
electrostatic potentials to probe the interior of a patient’s body in real time.

Theorem 4.14. Suppose that f (x ) ∈ L2 (Rn−1 ) is the Dirichlet data for
(4.33) on Rn+ , and suppose in addition that

dist supp(fˆ(ξ  )), 0 > ρ .

Then the solution u(x , xn ) decays as xn → +∞ with the upper bounds
C
(4.39) |u(x , xn )| ≤ e−ρ|xn | .
|xn |(n−1)/2

This estimate (4.39) gives an effective penetration depth of the solution

u(x) into the interior of the domain, in the situation in which the Dirich-
let data f (x ) has no low frequency component, namely, that the Fourier
transform fˆ(ξ  ) vanishes in a neighborhood of ξ  = 0.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 92 4. Laplace’s equation

Proof. Since dist supp(fˆ(ξ  )), 0 > ρ, there is a δ > 0 such that inf ξ ∈supp(fˆ) |ξ|
> ρ(1 + δ). Using the Fourier representation for u(x),

1   
u(x) = √ n−1 eiξ ·x e−|ξ |xn fˆ(ξ  ) dξ  ,
2π supp(fˆ)

we conclude that

1  −|ξ |x 
|u(x , xn )| ≤ √ n−1 e n ˆ  
f (ξ ) dξ 
2π supp(fˆ)

1  
e−|ξ |xn ( 1+δ ) e−|ξ |xn ( 1+δ ) |fˆ(ξ  )| dξ 
1 δ
≤ √ n−1
2π supp(fˆ)
|ξ  |x

− inf ξ ∈supp(fˆ) ( 1+δn ) 1 
e−|ξ |xn ( 1+δ ) |fˆ(ξ  )| dξ 
δ
≤e √ n−1
2π supp(fˆ)
 1/2  
 1/2
e−|ξ |xn ( 1+δ ) dξ 
2δ
≤ e−ρ|xn | |fˆ|2 dξ  .

We have used the Cauchy–Schwarz inequality on the last line. We thus have
the estimate on the decay of u(x , xn ) for large xn , namely
C(δ)
|u(x)| ≤ e−ρ|xn | f L2 .
|xn |(n−1)/2


This result states that an electrostatic potential whose data on the

boundary do not have support of its Fourier transform near ξ  = 0 does
not penetrate the domain significantly, and therefore imaging techniques
based on such potentials do not give rise to accurate renditions of details
from deeper in a patient’s body.
A similar bound holds for derivatives of u(x) using the same lines of
argument as in the proof above. This is again related to the interior reg-
ularity of solutions of elliptic equations. We will give a bound on multiple
derivatives of u(x). Using multi-index notation

∂xα u(x) = ∂xα11 . . . ∂xαnn u(x) ,

and assuming the hypotheses on the support of the Dirichlet data f (x) as
in Theorem 4.14, one follows a similar argument as in the proof above to
show that


||ξ  |αn |fˆ(ξ  )|e−|ξ |xn ( 1+δ ) dξ 
δ
 −ρ αn−1
|∂x u(x , xn )| ≤ e
α
|ξ1α1 . . . ξn−1

C(δ, α)
≤ n−1 e−ρ|xn | f L2 .
|xn | 2
+|α|

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 4.9. Hadamard variational formula 93

4.9. Hadamard variational formula

It is rare to have such explicit formulae as (4.35) for the Dirichlet problem
or (4.38) for the Dirichlet–Neumann operator, as is the case for the domain
Rn+ . More often one considers general domains, with less explicit solution
procedures, so that it is reasonable to think to develop more general methods
in order to understand the solution operators. In particular, it is relevant
to ask whether the Green’s function or the Poisson kernel vary continuously
under perturbation of the domain. For the Dirichlet problem there is an
elegant idea, originally due to Hadamard, to describe the Taylor expansion
of the Green’s function with respect to variations of the domain itself. This
is described in the book of Garabedian [8] and is the topic of Project 4.15.
Here we will describe this idea in a specific case, namely in terms of the
Poisson kernel ∂N G(x , y) for domains which are perturbations of the upper
half space Rn+ .
Consider η(x ) ∈ C 1 (Rn−1 ), whose graph defines the boundary of a
domain Ω(η) = {(x , xn ) ∈ Rn : xn > η(x )}. In this notation, Ω(0) =
Rn+ . The domain Ω(η) has its Green’s function GΩ(η) (x , y), from which we
obtain the Poisson kernel ∂N GΩ(η) (x , y) := DΩ(η) (x , y). The solution to the
Dirichlet problem with Dirichlet data is thus given by the integral operator

u(x) = DΩ(η) f (x) = DΩ(η) (x , y)f (y) dSy ,
Rn−1
 
where dSy = 1 + |∇η|2 dy  . From (4.37) we see that DΩ(0) = e−xn |D | , a
Fourier multiplier operator. In general u(x) = DΩ(η) f (x) is the bounded
harmonic extension to the domain Ω(η) of the boundary data f (x) defined
on ∂Ω(η) = {(x , xn ) : xn = η(x )}. The solution operator DΩ(η) clearly
depends upon the domain given by η(x ) in a nonlinear, global, and possibly
complicated way. The Hadamard variational formula expresses the deriv-
ative of the operator DΩ(η) with respect to variations of η, giving a linear
approximation to changes of the Green’s function under perturbation of the
domain.
Definition 4.15. Let D(η) be a bounded linear operator defined on L2 that
depends upon functions η(x) ∈ C 1 . The bounded linear operator A(η) is
the Fréchet derivative of D(η) with respect to η(x) at the point η = 0 if it
satisfies
A(λη) = λA(η) ∀λ ∈ R ,
D(η)f − D(0)f − A(η)f L2 ≤ C|η|2C 1 f L2 .

The notation that is used

 for the variation, or linear approximation, to D(η)
is that A(η) = δD(η)η=0 . This definition can of course be adapted to the
more general situation of the operators D(η) mapping f in a Banach space

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 94 4. Laplace’s equation

X to a Banach space Y , for η varying over a neighborhood of a third Banach

space Z.
Without actually proving that D(η) is analytic with respect to η ∈ C 1
(which it is), we will derive a formula for the Fréchet derivative of D(η)
for the domain that is the upper half space, for small |η|C 1 . A standard
harmonic function on Rn+ , and indeed on any of the domains Ω(η), is of

course ϕk (x) = eik·x e−|k|xn for each parameter k ∈ Rn−1 fixed. Therefore
   
D(0) eik·x = eik·x e−|k|xn = e−xn |D | eik·x .

Given one of the domains Ω(η), the boundary values of ϕk (x) on ∂Ω(η) are
 
eik·x e−|k|η(x ) ; therefore for xn > η(x ), we know that
  
D(η) eik·x e−|k|η(x ) = eik·x e−|k|xn .

Thus taking any point (x , xn ) ∈ Ω(0) ∩ Ω(η),

  
0 = D(η) eik·x e−|k|η(x ) − D(0) eik·x
1  
= D(η) (1 − η(x )|k| + η 2 (x )|k|2 + . . . )eik·x − D(0) eik·x .
2
This is to say that
  
(4.40) D(η) eik·x − D(0) eik·x − D(η) η(x )|k|eik·x
1 
= D(η) ( η 2 (x )|k|2 + . . . )eik·x .
2
The facts are that the operator D(η) does have a Fréchet derivative A(η)
at η = 0 and that the RHS is bounded by C|η|2C 1 |k|2 for small |η|C 1 . This
allows us to compute A(η) from (4.40). Namely,
  
D(η)(eik·x ) − D(0)(eik·x ) = D(0) η(x )|D  |eik·x + O(|η|2C 1 ) ,

from which we read that

   
A(η)eik·x = D(0) η(x )|D  |eik·x = e−xn |D | η(x )|D  | eik·x .

In other words, ∂η D(η)η=0 f = A(η)f = e−xn |D | η(x )|D  |f (x).


Now consider the Dirichlet–Neumann operator G(η) on domains Ω(η)

which are perturbations of Ω(0) = Rn+ and its Fréchet derivative at η = 0.
Recall that

G(η)f (x ) = Nx · ∇u(x)  , xn =η(x )

where u(x) is the bounded harmonic extension of the Dirichlet data f (x )
to the domain Ω(η). Using again the family of harmonic functions ϕk (x) =

eik·x e−|k|xn , we compute its boundary values f (x ) and its normal derivative

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 4 95

Nx · ∇ϕk (x) on {xn = η(x )}:

  
f (x ) = eik·x e−|k|η(x ) = 1 − η(x )|k| + O(|η|2C 1 ) eik·x ,
(4.41) G(η)f (x ) = Nx · ∇ϕk (x)
1 
= (∂x η, −1) · (ik, −|k|)∇ϕk (x)xn =η(x ) .
1 + |∇η|2
The Dirichlet–Neumann operator is not bounded on L2 , but it is bounded
from H 1 (Rn−1 ) to L2 , as can be seen by its expression (4.38) as a Fourier
multiplier when Ω = Rn+ . We seek the linear approximation to it among
domain perturbations Ω(η) at the point η = 0. To calculate B(η) =
δG(η) η=0 := ∂η G(η)η=0 , compare the first two terms of the LHS with the

RHS of (4.41) in powers of η:
 
G(η)f (x ) = G(0) + B(η) eik·x + G(0) −η(x )|k|eik·x
  
= |D  |eik·x + B(η)eik·x − |D  | η(x )|D  |eik·x
which is the LHS, and where the RHS is
  
Nx · ∇ϕk (x) = |k|eik·x + ∂x η(x ) · ik eik·x − |k|2 eik·x + O(|η|2C 1 )
 
= |D  |eik·x + ∂x η(x )∂x eik·x + O(|η|2C 1 ) .
Equating these expressions, solving for B(η), and applying the operators

to a general f (x ) rather than the particular family of functions eik·x , we
obtain
(4.42) B(η)f (x ) = ∂x η(x )∂x f (x ) + |D  | η(x )|D  |f (x ) .
Notice the symmetry under adjoints that is evident in this expression for
B(η), reflecting the self-adjoint property of the Dirichlet–Neumann operator
itself.

Exercises: Chapter 4
Exercise 4.1. Derive expression (4.36) for the case n = 2 from the Fourier
integral, and show that ω2 = 2π. Hint: Complex variables techniques would
be useful.
Exercise 4.2. In the case n = 2, the fundamental solution has the prop-
erty that ΔΓ(|x − y|) = δy (x). Show this by proving that the limit of the
expression in (4.10) vanishes and that the limit in (4.11) holds.
Exercise 4.3 (Separation of variables). Consider the boundary value prob-
lem for the domain B1 (0) ⊆ R2 for Dirichlet boundary data u(x) = f (x)
when |x| = 1:
Δu(x) = 0 , u(x) = f (x) for x ∈ S1 (0) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 96 4. Laplace’s equation

In polar coordinates (r, ϑ) ∈ R2 , Laplace’s equation is written

1 1
Δu = (∂r2 + ∂r )u + 2 ∂ϑ2 u = 0 .
r r
Special solutions can be obtained in the form u(x) = v(r)w(ϑ) so that
1 λ
∂ϑ2 w(ϑ) = −λw(ϑ) , (∂r2 + ∂r )v(r) = 2 v(r) .
r r
(i) Show that this equation for w(ϑ) implies that

λ = k2 , w(ϑ) = ck eikϑ + c−k e−ikϑ for k ∈ Z .

(ii) Show that the resulting solutions v(r) are therefore of the form

v(r) = rk for k≥0.

This uses the information that u(x) is to be bounded at r = 0. Conclude

that

+∞
u(r, ϑ) = rk ck eikϑ + c−k e−ikϑ
k=0

is a harmonic function in B1 (0).

(iii) Use this formula to prove that the Dirichlet problem for Laplace’s equa-
tion on B1 (0) has a solution for general Dirichlet data f ∈ L2 (S1 (0)).

Exercise 4.4. Give the solution to Laplace’s equation on B1 (0) ⊆ R2 using

the method of separation of variables for the following boundary data.
(1) f (ϑ) = a1 sin(ϑ) + a3 sin(3ϑ). What are the values of u(0), ∂x1 u(0),
and ∂x2 u(0)?
(2) f (ϑ) = b2 cos(2ϑ). What are the values of u(0), ∂x u(0), and ∂x2 u(0)?
(3) f (ϑ) = 1 + 5 cos3 (ϑ). What is the value of u(0)?

Exercise 4.5. Use the method of separation of variables on the square

domain D := {0 < x1 < 2π , 0 < x2 < 2π} to solve the following boundary
value problems for Laplace’s equation

Δu = 0 .
(1) Dirichlet boundary data on the side {(x1 , 0) : 0 < x1 < 2π}

f (x) = u(x1 , 0) = x1 (2π − x1 ) ,

while on the remaining three sides set f (x) = 0.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 4 97

(2) Dirichlet boundary data on opposite sides {(x1 , 0) : 0 < x1 < 2π}
and {(x1 , 2π) : 0 < x1 < 2π}
f (x) = u(x1 , 0) = x1 (2π − x1 ) ,
f (x) = u(x1 , 2π) = −2x1 (2π − x1 ) ,
while on the remaining two sides set f (x) = 0.
(3) Dirichlet boundary data on adjacent sides {(x1 , 0) : 0 < x1 < 2π}
and {(0, x2 ) : 0 < x2 < 2π}
f (x) = u(x1 , 0) = 2π − x1 , f (x) = u(0, x2 ) = 2π − x2 ,
and f (x) = 0 on the remaining two sides.
Exercise 4.6 (Dirichlet Laplace operator on the square domain D). Eigen-
functions of the Laplace operator on a domain Ω with Dirichlet boundary
conditions satisfy the equation
Δu = λu , u(x) = 0 for x ∈ ∂D ,
and λ is the associated eigenvalue. In the case that Ω = D is the square,
eigenfunction-eigenvalue pairs can be expressed by separation of variables
methods.
(1) Use the Ansatz u(x) = v(x1 )w(x2 ) to solve for eigenfunctions and
eigenvalues of the Dirichlet Laplacian on the square D and their
associated eigenvalues.
(2) Show that all of the eigenvalues satisfy λ < 0 and that the eigen-
function u(x) associated with the eigenvalue λ that is smallest in
absolute value satisfies
u(x) = 0 for x ∈ D .
(3) Show that this set of eigenfunctions is complete, meaning that
linear combinations of them are dense in L2 (D). Enumerate the
eigenfunction-eigenvalue pairs (uk (x), λk ) as you prefer, and nor-
malize the eigenfunctions so that
 
|uk (x)| dx = 1 ,
2
uk (x)u (x) dx = 0
D D
for k = .
(4) Use an eigenfunction expansion to give an expression for the Green’s
function for D in the form
 1
G(x, y) = uk (x)uk (y) .
λk
k

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 98 4. Laplace’s equation

Now show that a solution to the Poisson problem

Δu(x) = h(x) , x∈D, u(x) = 0 , x ∈ ∂D
can be expressed by

u(x) = G(x, y)h(y) dy .
D

Exercise 4.7 (Mixed boundary conditions). Consider the domain Ω =

B1 (0) ∩ R2+ ; this is a half disk. On the boundary ∂Ω ∩ {x2 = 0}, give
the following boundary conditions for Laplace’s equation:
u(x1 , 0) = 0 for x1 > 0 , ∂x2 u(x1 , 0) = 0 for x1 < 0 .
On the boundary {|x| = 1, x2 > 0}, specify that u(x) = sin(ϑ/2). Show
√
that u(x) = r sin(ϑ/2) satisfies Laplace’s equation and these boundary
conditions. However, show that u(x) has a square root singularity on the
axis {x2 = 0} even though the boundary data is smooth.
Exercise 4.8. Prove the second version of Gauss’ law of arithmetic mean
(4.14) for harmonic functions,

n
u(y) = u(x) dx .
ωn ρn Bρ (y)

Exercise 4.9. Prove that a function u(x) ∈ C 2 (Ω) that satisfies

(4.43) Δu ≥ 0
is subharmonic in Ω in the sense of Definition 4.7. Prove that if u(x) ∈ C 2 (Ω)
is subharmonic, then it satisfies inequality (4.43).
Exercise 4.10. Show that the Green’s function G(x, y) on a domain Ω ⊆ Rn
satisfies the property of symmetry G(x, y) = G(y, x). Conclude that the
resulting operator Δ−1 with Dirichlet boundary conditions is self-adjoint on
L2 (Ω).
Exercise 4.11. Prove the geometric property expressed by (4.22) for points
x, y such that x ∈ Sn−1
R and y ∈ BR (0), with y ∗ = |y|
R2
2 y, the inversion of y

R . That is, for any y ∈ SR , the sphere SR

in the sphere Sn−1 n−1 n−1
is the locus
of points x such that the ratio
|x − y ∗ | R
=
|x − y| |y|
is constant.
Exercise 4.12. Calculate the normal derivative of the Green’s function
(4.23) for the domain BR (0) with respect to x on the boundary x ∈ Sn−1
R ,
verifying expression (4.26).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 4 99

Exercise 4.13. Show that in R2 , the Laplace operator is invariant under

conformal transformations. That is, if x = x1 + ix2 and y = y1 + iy2 satisfy
x = τ (y) for an analytic function τ , then under the change of variables
u(x) → w(y) = u(τ (y)) the Laplacian is invariant, namely
Δx u = Δy w .
Hint: One approach to this question is to use the fact that the Laplace
operator on a domain Ω1 ⊆ R2 is the gradient of the energy functional

1
L(u) = |∂x u(x)|2 dx .
2 Ω1
Namely, gradL(u) = −Δu in the sense that
 
d 
 L(u + σv) = grad L(u) v dx = −Δu v dx
dσ σ=0 Ω1 Ω1

for all compactly supported test functions v ∈ C 2 (Ω1 ). Under a change of

variables x = τ (y) with w(y) = u(τ (y)), one has
|∂y w(y)|2 = ∂x u, (∂y τ )T (∂y τ )∂x u , dx1 dx2 = det(∂y τ )dy1 dy2 .
It suffices to show that for analytic transformations x = τ (y),
(∂y τ )T (∂y τ ) = det(∂y τ )I .
These expressions appear when transforming the energy functional L(u) over
x ∈ Ω1 to L(w) over y ∈ Ω2 = τ −1 (Ω1 ).
Exercise 4.14. For simply connected domains Ω in R2 , describe the Green’s
function in terms of the Riemann map of Ω to R2+ .

Projects: Chapter 4
Project 4.15 (Hadamard variational formula). This project is an extension
of the discussion in section 4.9 on perturbations of harmonic functions with
respect to their domain of definition. In particular, it studies perturbations
of the Green’s function of the domain. The goal is to prove the formula
that under perturbation of a domain Ω by small displacements η(x) in the
normal direction, the Green’s function for the domain, G(x, y) for x, y ∈ Ω,
satisfies the variational formula

(4.44) δG(x, y) = − Nz · ∇z G(x, z) η(z) Nz · ∇z G(z, y) dSz .
∂Ω
This formula was derived by J. Hadamard in 1915; it is from one point of
view a version of the resolvant identity.
Consider a compact domain Ω ⊆ Rn with real analytic boundary, and
a nearby domain Ω1 ⊆ Rn which is a real analytic perturbation of Ω, in
the sense that the boundary ∂Ω1 can be parametrized by x1 ∈ ∂Ω1 , where

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 100 4. Laplace’s equation

x1 = x + η(x)N for x ∈ ∂Ω. The domain Ω and the perturbation η are

assumed to be real analytic, and considering η over a small complex neigh-
borhood of ∂Ω, the norm η L∞ is assumed to be sufficiently small so that
any harmonic function on Ω can be extended to a harmonic function in a
full neighborhood of Ω, and similarly for Ω1 . Actually, formula (4.44) ap-
plies in more generality than this; a derivation in the case of domains Ω
with boundaries that are C 2 was derived by Garabedian and Schiffer, and
is described in Garabedian [8].
(i) Let Ω and Ω1 be two domains, and let x, y ∈ Ω∩Ω1 be two points interior
to both domains. The Green’s functions G(x, y) for Ω and G1 (x, y) for Ω1
are both well defined. Furthermore, suppose that η(x) is sufficiently small
so that the harmonic extension of G(x, y) in y to a neighborhood of Ω, for
x fixed, is defined on Ω1 , and vice versa for G1 (x, y).
Use Green’s identity as in the proof of Theorem 4.3 to show that

∇z G(x, z) · ∇z G1 (z, y) dz
Ω

= G(x, z) Nz · G1 (z, y) dSz − G(x, y) = −G(x, y) ,
 ∂Ω

∇z G(x, z) · ∇z G1 (z, y) dz
Ω1

= Nz · G(x, z) G1 (z, y) dSz − G1 (x, y) = −G1 (x, y) .
∂Ω1
Conclude that
 
G1 (x, y)−G(x, y) = − ∇G(x, z) · ∇G1 (z, y) dz+ ∇G(x, z) · ∇G1(z, y) dz .
Ω1 Ω

(ii) Consider the above expression as an integral of ∇z G(x, z) · ∇z G1 (z, y)

over the difference between the two domains (using the sign convention that
one adds if z ∈ Ω1 \Ω and subtracts if z ∈ Ω\Ω1 ). Because ∇z G(x, z) · T = 0
for all tangent vectors to the boundary ∂Ω, show that

∇G(x, z) · ∇G(z, y) dz
Ω1 −Ω

= Nz · ∇z G(x, z)η(z)Nz · ∇z G1 (z, y) dz + O(|η|2L∞ ) .
∂Ω
Therefore

G1 (x, y) − G(x, y) = − ∇G(x, z) · ∇G1 (z, y) dz
Ω1 −Ω

(4.45) =− Nz · ∇z G(x, z)η(z)Nz · ∇z G1 (z, y) dz + O(|η|2L∞ ) .
∂Ω

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 4 101

(iii) Finally, use formula (4.45) again in order to show that

G1 (z, y) − G(z, y) ∼ O(|η|L∞ ) ,
therefore

G1 (x, y) − G(x, y) = − Nz · ∇z G(x, z)η(z)Nz · ∇z G(z, y) dz + O(|η|2L∞ ) ,
∂Ω
which is a statement of the variational formula of Hadamard.
(iv) For a more open-ended project, also derive the variational formula of
Hadamard for general C 2 smooth domains Ω and C 2 domain perturbations
η(x). A useful reference is Garabedian [8].

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/05

Chapter 5

Properties of the
Fourier transform

x2
Normalised Hermite function hn (x) = √
1
1 Hn (x)e− 2 , the graphs
(2n πn!) 2
of h1 , h2 , and h13 .

The purposes of this section are to raise our level of sophistication of

the analysis of the Fourier transform and to make up our “backlog” of ana-
lytic justification of our work in the previous several sections. The Fourier
transform plays a very important role in analysis, and for this reason it has
been thoroughly analyzed from many points of view and by many people in
many different settings. We will work through the content of three principal
settings in this chapter, the ones most important to analysis and PDE.

103

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 104 5. Properties of the Fourier transform

The cases are

(1) Hilbert spaces, covered in section 5.1,
(2) Schwartz class, covered in section 5.2, and
(3) L1 integrability, covered in section 5.3.

5.1. Hilbert spaces

Definition 5.1. The usual Hilbert space for us is L2 (Rn ), which consists of
the square-integrable measurable complex valued functions on Rn :

L (R ) = {f Lebesgue measurable in R ,
2 n n
|f (x)|2 dx < +∞}.
Rn

Proposition 5.2 (Elementary properties of L2 (Rn )). (1) The space of func-
tions L2 (Rn ) is a linear space; namely, if f, g ∈ L2 , then this implies that
af + bg ∈ L2 for all a, b ∈ C.
(2) The topology on L2 (Rn ) is given by a norm
 1/2
f L2 = |f (x)|2 dx , f L2 = 0 ⇔ f = 0 .
Rn
Norms are homogeneous of degree 1:
αf L2 = |α| f L2 .
(3) Norms satisfy the triangle (or Minkowski) inequality:
f +g L2 ≤ f L2 + g L2 .

(4) The linear space L2 (Rn ) is complete with respect to the norm f L2 .
That is, if {fn }∞
n=1 ⊆ L (R ) is a Cauchy sequence of functions, then there
2 n

exists a limit function f (x) ∈ L2 (Rn ) such that fn (x) →L2 f (x) (meaning
fn (x) − f (x) L2 → 0).
(5) Furthermore, the norm on L2 (Rn ) is given by an inner product:

(5.1) f 2L2 = f, f  where f, g = f (x)g(x) dx .

The inner product satisfies the Cauchy–Schwarz inequality:

|f, g| ≤ f L2 g L2 .

(6) The “dual space”, or space of bounded linear functionals on L2 (Rn ), is

L2 (Rn ) itself.
Definition 5.3 (Banach and Hilbert spaces). A linear space with a norm,
which is complete, is a Banach space. A Banach space whose norm is given
by an inner product is a Hilbert space.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.1. Hilbert spaces 105

Our favorite Hilbert space L2 (Rn ) has its inner product defined in (5.1).
One reason that L2 (Rn ) is a natural setting for the Fourier transform is
that it is preserved under the transform. In Rn , we define the transform as
follows. For ξ ∈ Rn ,

1
ˆ
f (ξ) = √ n e−iξ·x f (x)dx = (F f )(ξ) .
2π Rn

Theorem 5.4. The Fourier transform is an isometry of L2 (Rn ). That is,

f L2 (Rnx ) = fˆ L2 (Rnξ ) , or, stated in other words,
 
|f (x)| dx =
2
|fˆ(ξ)|2 dξ .
Rn
x Rn
ξ

This equality between the L2 -norms of a function and its Fourier transform
is known as the Plancherel identity; it is a general fact about the Fourier
transform that holds in many settings. The proof of Theorem 5.4 is deferred
until the end of our discussion of Schwartz class.
Other examples of Hilbert spaces and Banach spaces as tools of analysis
include the following.
(1) Sobolev spaces H s (Rn ) (Hilbert spaces based on L2 -norms): For s ∈ N,
  1
2
H (R ) = {f (x) :
s n
|∂xα f (x)|2 dx < +∞} .
0≤|α|≤s

The norm of a function f in this Sobolev space measures the size of deriva-
tives of f as well as the size of f itself. We are using our previously intro-
duced notation that a convenient device for multivariable objects involves
multi-indices α, β: α = (α1 , α2 , . . . , αn ), β = (β1 , β2 , . . . , βn ), with each
αi , βj ∈ N. The order of a multi-index is given by |α| = |α1 |+· · ·+|αn | so that
xα = xα1 1 xα2 2 . . . xαnn is a monomial of degree |α| and ∂xβ = ∂xβ11 ∂xβ22 . . . ∂xβnn =
β
Π∂xjj is a differential operator of order |β|. The family H s (Rn ) is a scale of
spaces which are nested in terms of decreasing index s: H s ⊆ H s−1 ⊆ · · · ⊆
L2 (Rn ).
(2) Lp spaces (Banach spaces): For 1 ≤ p < +∞,
 1
p
L (R ) = {f (x) :
p n
|f (x)|p dx < +∞} .
Rn
For p = ∞, the space of bounded measurable functions on Rn is denoted by
L∞ (Rn ) and has a norm given by
f L∞ := sup |f (x)| .
x∈Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 106 5. Properties of the Fourier transform

(3) Sobolev spaces W s,p (Rn ) (Banach spaces modeled on Lp -norms, which
again measure derivatives of f ):
  1
p
W s,p (Rn ) = {f : |∂xα f (x)|p dx < +∞} .
0≤|α|≤s

(4) Schauder spaces C s,γ (Ω) (Banach spaces modeled on C 0 -norms): For
Ω ⊆ Rn a domain, that is, an open set with smooth boundary,

C s,γ (Ω) = {f (x) : sup |∂xα f (x)| < +∞ ∀|α| ≤ s

x∈Ω
|∂xα f (x) − ∂xα f (y)|
and ∀|α| = s, sup < +∞} .
x,y∈Ω,x=y |x − y|γ

The space of Lipschitz functions is in this class; namely, f ∈ Lip(Rn ) =

C 0,1 (Rn ) when there is a constant C such that

|f (x) − f (y)| ≤ C|x − y| for all x, y ∈ Rn .

The many relationships of inclusion among these spaces, and their as-
sociated inequalities of norms, are important information for the analysis
of PDEs. A principal one that is very often used in PDEs is the Sobolev
Embedding Theorem. A version of this result is as follows.

Theorem 5.5 (Sobolev embedding). If f (x) ∈ H s (Rn ) for a Sobolev index

s > n2 , then
|f (x)|L∞ ≤ Cn f Hs .
In other words H s ⊆ L∞ for s > n2 , and there is a bound on the inclusion
operator ι : H s → L∞ given by the constant Cn .

Proof. Consider the value of f (x) at an arbitrary point of Rn ,


 1 
|f (x)| =  √ n eiξ·x fˆ(ξ) dξ 
2π Rn

 1 1 

= √ n eiξ·x 2 s ˆ 
s (1 + |ξ| ) 2 f (ξ) dξ .
2π Rn (1 + |ξ|2 ) 2

Using the Cauchy–Schwarz inequality on this last expression,

 1 
1  1 1

(1 + |ξ|2 )s |fˆ(ξ)|2 dξ
2 2
(5.2) |f (x)| ≤ √ n dξ .
2π (1 + |ξ|2 )s

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.2. Schwartz class 107

Because derivatives are transformed to multiplications by iξ under the Four-

ier transform, by the binomial theorem,
 1  s    1
2 s ˆ s
|ξ|2j |fˆ(ξ)|2 dξ
2 2
(1 + |ξ| ) |f (ξ)| dξ
2
=
j
j=0
s    1
s 2
= |∇j f (x)|2 dx ≤ Cn f H s ,
j
j=0

where we have used the Plancherel identity for f (x) and its derivatives in
the last line. This holds of course for any integer s. It remains to bound the
RHS of (5.2),
 1 1   +∞
1 1
2 n−1 2
dξ = r dr dSϕ ,
(1 + |ξ|2 )s S n−1 0 (1 + r 2 )s

and if s > n2 , then this quantity is finite, which finishes the proof. 

5.2. Schwartz class

This linear space of functions was introduced by Laurent Schwartz (École
Polytechnique, Paris) in order to embody a very convenient class of func-
tions with which to work, for which one can integrate by parts and inter-
change integrations with differentiations with impunity. Colloquially, f (x)
is a Schwartz class function if it is infinitely differentiable and if it decays
along with all of its derivatives as |x| → ∞ faster than any polynomial.
Recall our previously introduced notation that is a convenient device for
multivariable objects: the quantities α, β are multi-indices,
α = (α1 , α2 , . . . , αn ), and β = (β1 , β2 , . . . , βn ) .
Definition 5.6 (Schwartz class). The Schwartz class consists of those func-
tions f (x) defined over Rn such that for all α, β,
sup |xα ∂xβ f (x)| < +∞.
x∈Rn

It is clearly a linear space of functions and is denoted by S.

The multi-index notation is a device for describing many combinatorially

complex objects, which can be written very conveniently. For example,
Taylor series in n dimensions can be stated
 1
f (x) ∼ ∂xα f (y)(x − y)α
α
α!

with α! = α1 !α2 ! . . . αn !.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 108 5. Properties of the Fourier transform

Here are several examples of particular functions where we can compare

their properties of smoothness and decay in x ∈ Rn to those of Schwartz
class:
x2
f (x) = e− 2 ∈S ,
f (x) = (1 + |x|2 )− 2 ∈
b
/ S but f (x) ∈ L2 (Rn ) for b > n/2 ,
f (x) = e−|x| ∈
/ S for a different reason .

Every Schwartz class function f (x) is also in L2 (Rn ), as it is bounded

and decays rapidly. Furthermore, Schwartz class is a linear subspace of
L2 (Rn ) which is dense in the L2 topology. That is, for any f ∈ L2 (Rn ) there
is a sequence {fn }∞n=1 ⊆ S such that limn→∞ fn − f L2 = 0. The proof of
this fact appears in the exercises.
The space C0∞ of infinitely differentiable functions with compact support
is a subset of S as a subspace, C0∞ ⊆ S, and the space C ∞ of infinitely
differentiable functions contains Schwartz class, S ⊆ C ∞ . The space C ω (Rn )
of real analytic functions on Rn (the restriction to x ∈ Rn of functions that
are analytic in a neighborhood of Rn ⊆ Cn ) is a subset of C ∞ , but not every
C ∞ function is analytic.
The topology of S is defined using a countable family of seminorms
f α,β = sup |xα ∂xβ f (x)| .
x∈Rn

That is to say, a sequence {fn }∞n=1 converges to f if and only if for all
multi-indices α, β, the seminorms
fn − f α,β = sup |xα ∂xβ (fn (x) − f (x))| → 0
x

with n → ∞. No uniformity in α, β is implied, however. Some functional

analytic remarks are in order at this point. The space S is a linear space but
not a Banach space, as there is no way to describe this sense of convergence
by using a norm. It is, however, a metric space with the distance between
two points given by
 1 f − g α,β
ρ(f, g) = ,
2 n(|α|+|β|) 1 + f − g α,β
α,β

and it is complete with respect to this metric. Clearly {fn (x)}∞ n=1 converges
in S to f (x) if and only if ρ(fn , f ) → 0 if and only if all fn − f α,β → 0 as
n → ∞. The notation for limits of sequences of functions in this Schwartz
class sense is to write S − lim fn = f . Metric spaces such as this one whose
topology is given by a countable family of seminorms are called Fréchet
spaces.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.2. Schwartz class 109

Lemma 5.7 (Technical lemma). This result quantifies the properties of scal-
ing limits in two cases, and these limits give rise to well-defined functionals
on S.
(i) (Approximations of the identity). Let g ∈ S with g(0) = 1. Then for all
f ∈ S,
S − lim (g(εx)f (x)) = f (x) .
ε→0

 (Approximations of the Dirac δ-function). Let h(x) ∈ L (R ) with

(ii) 1 n

h(x)dx = 1, and suppose that f (x) is bounded, and continuous at x = 0.

Then 
1 x
lim f (x) n h( )dx = f (0) .
ε→0 ε ε

For h(x) ∈ L1 (Rn ) with h(x) dx = 1, the limit operation in (5.7) describes
the approximation of the Dirac δ-function by rescaling a suitably normalized
L1 function h(x).

Proof of Lemma 5.7(ii). By change of variables,

 
1 x
h( ) dx = h(x ) dx = 1
εn ε
with x = xε . Given δ > 0, choose r > 0 such that |f (x) − f (0)| < δ for all
|x| < r. Then write
  
1 x 1 x 1 x
f (x) n h( ) dx = f (0) n h( ) dx + (f (x) − f (0)) n h( ) dx .
ε ε ε ε ε ε
The first term is f (0). Split the second term in two,
 
 1 x   1 x 
 
(f (x) − f (0)) n h( ) dx ≤ δ  h( ) dx  ≤ δ h L1 ,
ε ε n
|x|<r |x|<r ε ε
 
 1 x  1 x
 (f (x) − f (0)) n h( ) dx ≤ 2 sup |f (x)| |h( )| dx
ε ε ε n ε
|x|≥r x |x|≥r

= 2 sup |f (x)| |h(x )|dx ,
x |x |≥ rε

and by the Lebesgue dominated convergence theorem, the RHS of the last
term with r fixed tends to zero with ε → 0. 

The proof of Lemma 5.7(i) appears in the exercises of this chapter.

Corollary 5.8. The space C0∞ ⊆ S is a dense subspace of S.

Proof. This uses Lemma 5.7: Take any g ∈ C0∞ with g(0) = 1. Then for
f ∈ S arbitrary,
g(εx)f (x) ∈ C0∞

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 110 5. Properties of the Fourier transform

and
S − lim (g(εx)f (x)) = f (x) .
ε→0


The space of distributions is the space of continuous linear functionals on

C0∞ , D = (C0∞ ) , otherwise known as its dual space. The space of tempered
distributions is the dual space of S. Also, we write E  for the dual space to
C ∞ , the space of infinitely differentiable functions. Since C0∞ ⊆ S ⊆ C ∞ ,
then E  ⊆ S  ⊆ D . An example of a distribution is the Dirac δ-function,
δ(f ) := f (0), which is a distribution in E  and hence in all of the above
spaces of continuous linear functionals.
The Fourier transform acts nicely on S; for f (x) ∈ S define the Fourier
transform as usual:

1
f (ξ) = √ n e−iξ·y f (y) dy = F (f )(ξ) .
ˆ
2π
One checks that this integral converges absolutely because f (x) ∈ S; indeed
1
|e−iξ·x f (x)| ≤ f N,0 N ,
(1 + |x|2 ) 2
and N > n + 1 will do for this. For Schwartz class f (x) we may exchange
integrations and differentiations as much as we wish; therefore the following
list of properties holds for fˆ(ξ).

Proposition 5.9 (Elementary properties of the Fourier transform on S).

ˆ
(1) ∂x f (ξ) = (iξ)f(ξ).
 (ξ) = i∂ξ fˆ(ξ).
(2) xf
Notice the vectorial notation; x and i∂ξ are vector operations.
Define two operations on functions: τh f (x) = f (x − h) (translations)
and (σλ f )(x) = f (λx) (dilations).
(3) τh f (ξ) = e−ihξ fˆ(ξ), a position boost.

(4) e
ihx f (ξ) = fˆ(ξ − h), a momentum boost.

(5) σλ f (ξ) = 1 ˆ 1 ˆ
|λ|n σ(1/λ) (f )(ξ) = |λ|n f (ξ/λ).

The principal reason why Schwartz class is well suited for Fourier anal-
ysis is that it is invariant under the Fourier transform.
Theorem 5.10. The Fourier transform maps S into itself; that is, whenever
f ∈ S, then fˆ = F (f ) ∈ S.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.2. Schwartz class 111

Proof. For f (x) ∈ S, at the very least fˆ(ξ) makes sense as a convergent
integral. In order to check that fˆ(ξ) ∈ S, we have to check that all of
the seminorms are finite, which we do using the properties described in
Proposition 5.9:
fˆ α,β = sup |ξ α ∂ β fˆ(ξ)| ξ
ξ
 1  
 1 1 
= sup  √ n e−iξ·x (( ∂x )α ( x)β f (x)) dx.
ξ 2π i i

This integrand is still bounded by Cβ+N,α (1 + |x|2 )− 2 , and hence the supre-
N

mum over ξ ∈ Rn is finite. Furthermore, the Fourier transform is continuous

on S. Suppose that {fn (x)}∞ n=1 , and fn → f . This implies that for each
S
α, β,
 1  
 1 1 
fˆn − fˆ α,β = sup  √ n e−iξ·x (( ∂x )α ( x)β (fn (x) − f (x))) dx → 0
ξ 2π i i
with n → ∞ as well. Hence fˆn (ξ) → fˆ(ξ), and therefore the mapping
S
fˆ : S → S is continuous. 

Finally we are prepared to prove the central theorem of Fourier inversion

on S.
Theorem 5.11 (Fourier inversion theorem on S). Suppose that f ∈ S; then

1
f (x) = √ n eiξ·x fˆ(ξ) dξ = F −1 (fˆ)(x) .
2π
Proof. Let’s first note that this works if f (x) is a Gaussian. For G(x) =
|x|2 |ξ|2
√ 1 n e− , Ĝ(ξ) = √2π
2
1
ne
− 2
by explicit calculation. One interpretation of
2π
this fact is that G(x) is an eigenfunction of the Fourier transform, F (G) = G,
with eigenvalue 1. Now for the case of general f (x) ∈ S, use properties given
in Lemma (5.7),
 
1 x 1
f (0) = lim f (x) n G( ) dx = lim f (x) n σ 1 (G) dx
ε→0 ε ε ε→0 ε ε
 
1
= lim f (x) n σ 1 (Ĝ) dx = lim f (x)σ ε (G) dx
ε→0 ε ε ε→0

= lim fˆ(ξ)σε (G)(ξ) dξ .
ε→0
We have to verify the last step of this sequence of calculations.
Lemma 5.12. For f, g ∈ S,
 
f (x)ĝ(−x)dx = fˆ(ξ)g(ξ)dξ .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 112 5. Properties of the Fourier transform

Proof of Lemma 5.12.

  
1 −iξ·x 1
f (x) √ n e g(ξ) dξ dx = √ n f (x)e−iξ·x dx g(ξ) dξ
2π 2π

= fˆ(ξ)g(ξ) dξ .



Now apply (5.7) to the result. Since G(ξ)ξ=0 = √2π 1
n,

 
1
f (0) = lim fˆ(ε)G(ε/ξ) dξ = √ n fˆ(ξ) dξ .
ε→0 2π

1
This recovers f (0) = √2π n fˆ(ξ) dξ. To generalize this procedure to cover all
x ∈ Rn is in fact simple, again using the list of properties in Proposition 5.9:
 
1 1
f (x) = (τ−x f )(0) = √ n F (τ−x f )(ξ) dξ = √ n eiξ·x fˆ(ξ) dξ .
2π 2π


We can finish this circle of ideas with a proof of the L2 (Rn ) Fourier
inversion theorem. We have already stated that Schwartz class S ⊆ L2 (Rn )
is a dense subspace. Lemma 5.12 states that for f, g ∈ S,
f, F −1 g = F f, g ,
where we also note that
(F −1 g)(x) = ĝ(−x) .
Therefore for every f ∈ S,
f 2
L2 = f, F −1 (F f )
= F (f ), F (f )
= fˆ 2
L2 ,
so the Fourier transform acts on the the subspace S ⊆ L2 (Rn ) isometrically
(in the sense of the L2 -norm). Since S is dense, given an arbitrary f ∈
L2 (Rn ), there is a sequence fn → f in the L2 (Rn ) sense of convergence,
with fn ∈ S. Then we define
F (1) (f ) := lim F (fn ) .
n→∞

Theorem 5.13 (Functional analysis). Suppose that an operator T is defined

on a dense subspace S ⊆ B, where B is a Banach space, and it is bounded;
Tf B ≤ C f B for all f ∈ S. Then there exists a unique extension T(1)
of T to all of B, which is bounded with the same bound.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.3. Fourier transform of L1 -integrable functions 113

Proof. For arbitrary f ∈ B, consider a sequence {fn }∞ ∞

n=1 → f with {fn }n=1 ⊆
S. Then {fn }∞ ∞
n=1 is Cauchy, and so is the sequence {Tfn }n=1 because
Tfn − Tfm B ≤ C fn − fm B →0.
Thus {Tfn }∞n=1 has a limit g ∈ B, which we define to be T f := g. The
(1)

extension is clearly linear. It is also well defined, for if {hn }∞ n=1 ⊆ S is

another sequence such that {hn }∞ n=1 → f , but T h n → g 1 , then
g − g1 B = lim Tfn − Thn B ≤ C lim fn − hn B =0.
n→∞ n→∞


The conclusion that is relevant to the Fourier transform is that F re-

stricted to S (such that it has norm F f L2 = f L2 on S) extends uniquely
to F (1) defined on all of L2 (Rn ). Furthermore, this extension is an isometry
of L2 (Rn ), as for {fn }∞
n=1 ⊆ S ⊆ L (R ), with fn → f :
2 n

f 2 2 = lim fn 2 2 = lim fˆn 2 2 = F (1) (f ) 2 2 .

L L L L
n→∞ n→∞

This completes the proof of Theorem 5.4. We will redefine the “extension
notation” F (1) := F at this time, dropping the extra superscript. The
statement is then that f L2 = F f L2 for all f ∈ L2 , which is the desired
result.
To finish our remarks we should note that complex isometries U have
the property that U∗ = U−1 ; namely, they are unitary operators. Let’s
verify this with the Fourier transform.
Definition 5.14. The adjoint T∗ of an operator T on a Hilbert space H
satisfies T∗ f, g = f, Tg for all f, g ∈ H.

Applying this definition of the adjoint to F , take any two f, g ∈ L2 .

Then
  
1
g(ξ)F (f )(ξ) dξ = g(ξ) √ n e−iξ·x f (x) dxdξ
Rn Rn 2π
 
1
= ( √ n eiξ·x g(ξ) dξ)f (x) dx
2π
R R
n n

= F −1 (g)(x)(f (x)) dx ,
Rn
which states that F∗ = F −1 , as desired.

5.3. Fourier transform of L1 -integrable functions

The standard space of integrable functions is L1 (R1 ), which has arisen in our
discussions of the heat equation and in particular its interpretation in terms
of probability. The space of functions f ∈ L1 (Rn ) is a Banach space but

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 114 5. Properties of the Fourier transform

not a Hilbert space, and in particular there is no way to consistently define

an inner product. The theory of the Fourier transform of L1 functions is
subsequently more delicate than its Hilbert space analog. Nonetheless the
Fourier transform of a L1 function is well defined; it has certain properties
that will be described below, and the inverse transform is well defined and
unique. More subtle details include the manner in which one recovers the
function f (x) from its Fourier transform and the characterization of the
space F L1 (Rn ) which is the range of the Fourier transform of L1 . We discuss
the first but not the second of these items below. An excellent reference is
Dym and McKean [5]. For reasons of convenience we will consider the case
of one space dimension, namely L1 (R1 ), although all the results below have
their direct analogs in L1 (Rn ).

Theorem 5.15. For f ∈ L1 (R1 ), define as usual the Fourier transform


1
fˆ(ξ) = √ e−iξx f (y) dy .
2π
The Fourier transform fˆ(ξ) of the function f (x) has the following properties.
(i) Boundedness: |fˆ(ξ)|L∞ ≤ √1
2π
f L1 .

(ii) Continuity: fˆ(ξ) ∈ C(R1 ).

(iii) Limits as |ξ| → +∞ (Riemann–Lebesgue Lemma):

lim fˆ(ξ) = 0 .
|ξ|→+∞

(iv) Uniqueness: fˆ(ξ) = 0 if and only if f (x) = 0.

(v) Convolution algebra:
√
f
∗ g(ξ) = 2π fˆĝ(ξ) .

Proof. The statement (i) of boundedness follows directly from the expres-
sion of the Fourier transform; indeed
 
1   1
|f (ξ)| = √
ˆ e −iξy
f (y) dy  ≤ √ |f (y)| dy .
2π 2π
To prove the property of continuity (ii), consider the difference

1
fˆ(ξ1 ) − fˆ(ξ2 ) = √ [e−iξ1 y − e−iξ2 y ]f (y) dy .
2π
Given f ∈ L1 , for any ε > 0 we will show that if |ξ1 − ξ2 | is sufficiently small,
then this difference between f (ξ1 ) and f (ξ2 ) is less than ε in absolute value.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.3. Fourier transform of L1 -integrable functions 115

We may write this difference as

 +R
1
ˆ ˆ
f (ξ1 ) − f (ξ2 ) = √ e−iξ1 y [1 − e−i(ξ2 −ξ1 )y ]f (y) dy
2π −R
 −R  +∞
+ + [e−iξ1 y − e−iξ2 y ]f (y) dy .
−∞ +R

Since f ∈ L1 , then given ε > 0, there is R > 0 such that

 −R  +∞ ε
+ |f (y)| dy ≤ .
−∞ +R 3
Therefore the second term of the RHS is bounded by 2ε/3 for R chosen
sufficiently large. The first term of the RHS is bounded by
 
1  +R −iξ1 y 
√  e [1−e−i(ξ2 −ξ1 )y ]f (y) dy  ≤ sup |[1−e−i(ξ2 −ξ1 )y ]| |f (y)| dy .
2π −R |y|≤R

As long as |ξ1 −ξ2 | < ε 6R

π
this term is also bounded by ε/3, and the continuity
result follows.
Statement (iii) of the theorem is a result from the theory of the Lebesgue
integral. Without going into a full proof, we just note that it is consistent
with the expression of fˆ(ξ) for f (x) = χ[a,b] (x), an indicator function of an
interval:
  b
1 −iξy e−iξb − e−iξa
√ e χ[a,b] (y) dy = e−iξy dy = ,
2π a −iξ
which has this decay property as |ξ| → +∞.
Statement (v) concerns the properties of convolution, which are that
 
1 −iξy
F (f ∗g)(ξ)= √ e f (y − y  )g(y  ) dy  dy
2π
 √
1  
=√ e−iξ(y−y ) f (y − y  ) dy e−iξy g(y  ) dy  = 2π fˆ(ξ)ĝ(ξ) .
2π
Fubini’s theorem is invoked to justify the exchange of integrations when
both f, g ∈ L1 .
Finally, for statement (iv) consider f ∈ L1 while g ∈ S is an arbitrary
Schwartz class test function. Then
  
1
f (x)g(x) dx = f (x) √ e−iξx ĝ(−ξ) dξdx
2π

= fˆ(ξ)ĝ(−ξ) dξ ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 116 5. Properties of the Fourier transform

again using Fubini’s theorem to interchange integrals. If we have fˆ(ξ) = 0,

then

f (x)g(x) dx = 0

for all g ∈ S, implying that f (x) itself must vanish identically. 

As remarked above, the convergence of the inverse Fourier transform of a

function in L1 is less straightforward than in the case of L2 . In this direction
we describe two results. It would seem that the simplest manner to invert
the Fourier transform of fˆ(ξ) would be as the limit of partial inverses
 +R
1
lim fR (x) := lim √ eiξx fˆ(ξ) dξ .
R→+∞ R→+∞ 2π −R
However, this does not converge well in general; in fact for the analogous
question for convergence of Fourier series, there are counterexamples to con-
vergence in what are called lacunary sequences. An alternate limiting pro-
cess involves taking an average of such partial inverses. In particular we
take limits of the T -averages of the above expression,
  +R
1 T 1
√ eiξx fˆ(ξ) dξ dR
T 0 2π −R
as T → +∞, which turns out to have substantially better convergence prop-
erties.

Theorem 5.16 (Cesaro integrability). Given f ∈ L1 (R1 ),

  +R
1 T 1
(5.3) lim √ eiξx fˆ(ξ) dξ dR = f (x) ,
T →+∞ T 0 2π −R
where the convergence of the functions on the LHS to the RHS is taken with
respect to the L1 -norm.

Proof. First of all we can describe the integrals over [−R, +R] in terms of
a convolution
√  +R iξx  +R  +∞
2π ˆ
e f (ξ) dξ = eiξ(x−y) f (y) dydξ
−R −R −∞
 +∞  +R
iξ(x−y)
= e dξ f (y) dy
−∞ −R
 +∞  iR(x−y)
e − e−iR(x−y)
= f (y) dy .
−∞ i(x − y)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 5.3. Fourier transform of L1 -integrable functions 117

Using this expression, we now average over 0 ≤ R ≤ T ,

  +R
1 T 1
√ eiξx fˆ(ξ) dξ dR
T 0 2π −R
 
1 +∞ T 1  eiR(x−y) − e−iR(x−y)
= √ dR f (y) dy
T −∞ 0 2π i(x − y)
 +∞  2
1 sin (T (x − y))
= √ f (y) dy ,
−∞ T 2π(x − y)2
which is also a convolution of f (x), with the kernel function
1  sin2 (T (x − y))
FT (x − y) = √ .
T 2π(x − y)2
The kernel FT (x) is the analog of the Féjer kernel in the proof of Cesaro
summability of Fourier series. It is nonnegative and its integral equals one.
Therefore the task of the proof of convergence is to show that for arbitrary
ε > 0, the difference
 
FT (x − y)f (y) dy − f (x) = FT (y)[f (x − y) − f (x)] dy
 +∞
1 sin2 (T y)
= [f (x − y) − f (y)] dy
−∞ T y2

can be made smaller than ε when measured in L1 . The L1 -norm of this

difference is
 +∞  +∞ 
 
 FT (y)[f (x − y) − f (x)] dy  dx
−∞ −∞
 +∞  +∞
1 sin2 (T y)  
 dx dy
≤ [f (x − y) − f (x)]
−∞ −∞ T y2
 +δ  +∞
1 sin2 (T y)  
 dx dy
≤ [f (x − y) − f (x)]
−δ −∞ T y2
 −δ  +∞  +∞ 1 sin2 (T y)  
(5.4) + + [f (x − y) − f (x)] dx dy .
2
−∞ +δ −∞ T y

Since translation by y ∈ R1 is continuous in the L1 -norm, for y restricted to

[−δ, +δ] and for δ > 0 sufficiently small,
 +δ  +∞
1 sin2 (T y)  
2
[f (x − y) − f (x)] dx dy
−δ −∞ T y
 +δ
ε
≤ f (· − y) − f (·) L1 FT (y) dy < .
−δ 2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 118 5. Properties of the Fourier transform

The remaining term of the RHS of (5.4) is also bounded by ε/2 for T suffi-
ciently large, namely
 −δ  +∞  +∞ 1 sin2 (T y)  
+ [f (x − y) − f (x)] dx dy
2
−∞ +δ −∞ T y
 −δ  +∞ 1 sin2 (T y)
≤ 2 f L1 + dy
−∞ +δ T y2
 −T δ  +∞ sin2 (y  )
≤ 2 f L1 +  )2
dy  .
−∞ +T δ (y
Since the function sin2 (y  )/(y  )2 is integrable, for δ > 0 fixed, one can take
T large so that this latter term is also bounded above by ε/2. 

A second setting is related to solutions of the heat equation, which also

can be considered as a smoothing or an averaging process.
Theorem 5.17. Suppose that f ∈ L1 (R1 ) and that u(t, x) is the solution
of the heat equation with initial data f (x), given by u(t, x) = H(t, ·) ∗ f (x).
Then 
1
eiξx e− 2 |ξ| t fˆ(ξ) dξ ,
1 2
f (x) = lim u(t, x) = lim √
t→0+ t→0+ 2π
where again the limit is in the sense of convergence in the L1 (R1 )-norm.

Proof. The heat kernel has Fourier transform H(t, ξ) = e− 21 |ξ|2 t of course,
and for t > 0, it is a member of Schwartz class. From Theorem 5.15(ii), we
know that fˆ(ξ) is bounded; thus the product satisfies e− 2 |ξ| t fˆ(ξ) ∈ L1 (R1 ).
1 2

Hence it has a well-defined inverse Fourier transform, which, again by the

Fubini theorem, satisfies
F (e− 2 |ξ| t fˆ)(x) = (H(t, ·) ∗ f )(x) .
1 2

It therefore suffices to study the limit as t → 0+ of the quantity

 
 
H(t, ·) ∗ f − f L1 =  H(t, x − y)[f (y) − f (x)] dy  dx
  +δ
 
≤ H(t, y)[f (x − y) − f (x)] dy dx
−δ
  
−δ +∞  
+ + H(t, y)[f (x − y) − f (x)] dy dx .
−∞ +δ

Consider an arbitrary ε > 0. Use the fact that translation by y ∈ R1 is

continuous in the L1 -norm so that for δ > 0 sufficiently small,
 +δ 
 
H(t, y)[f (x − y) − f (x)] dx dy ≤ sup f (· − y) − f (·) L1 ≤ ε .
−δ |y|≤δ 2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 5 119

The remaining term of the difference is bounded by

  −δ  +∞
 
+ H(t, y)[f (x − y) − f (x)] dy dx
−∞ +δ
 −δ  +∞
≤2 f L1 + H(t, y) dy .
−∞ +δ

As long as 0 < t is taken sufficiently small, this quantity can also be shown
to be less than ε/2, as in the estimate (3.8) of the heat kernel. 

The results of this subsection are related in spirit to Theorem 3.7 which
concerns the pointwise behavior of convolutions with the heat kernel as
t → 0+. It is the statement that at points x0 at which f (x) is continuous,
limt→0+ u(t, x0 ) = f (x0 ). A similar result holds for the Cesaro integrals,
with a similar proof. Namely, if f (x) is continuous at x0 , then

lim FT (x0 − y)f (y) dy = f (x0 ) .
T →+∞

This pointwise limit at points of continuity of f (x) is the subject of Exer-

cise 5.8.

Exercises: Chapter 5
Exercise 5.1. Show that the Fourier transform preserves angles between
vectors. For f, g ∈ L2 ,
Ref, g = f L2 g L2 cos θ
and
Refˆ, ĝ = fˆ L2 ĝ L2 cos ϕ .
Show that θ = ϕ.
Exercise 5.2. The Lp (Ω)-norm for functions on a domain Ω ⊆ Rn is defined
as
 1/p
f Lp = |f (x)|p dx
Ω
for 1 ≤ p < +∞. The Banach space L∞ (Ω) is defined for p = +∞ by
the norm
f L∞ = sup |f (x)| .
x∈Ω
Because of Lebesgue measure, this supremum is taken modulo sets of zero
measure (the essential supremum). The Hölder inequality states that

 
 f (x)g(x) dx ≤ f Lp g p ,
L
Ω

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 120 5. Properties of the Fourier transform

where p1 + p1 = 1 are dual indices. A special case for p = p = 2 is the

Cauchy–Schwarz inequality. Give a proof of the Hölder inequality for the
range of p, p given above.
Exercise 5.3. This problem concerns the cases of domains Ω = Rn and
bounded domains Ω ⊆ Rn .
(1) In the case of Ω = Rn , for which indices p and q does
Lp (Rn ) ⊆ Lq (Rn )
hold?
(2) In the case of bounded Ω, use the Hölder inequality to show that
Lp (Ω) ⊆ Lq (Ω)
for q ≤ p.
(3) In the case of bounded Ω, is it true that

L∞ (Ω) = Lp (Ω) ?
1≤p<+∞

Exercise 5.4. Prove part (i) of the technical Lemma 5.7 on Schwartz class
functions. Namely, show that for g ∈ S(Rn ) such that g(0) = 1, then for all
f ∈ S(Rn ),
S − lim g(εx)f (x) = f (x) .
ε→0

Exercise 5.5. Give examples of nonanalytic C ∞ functions on Rn . Hint:

See Project 3.12 of Chapter 3.
Exercise 5.6. Prove that Schwartz class S is a dense subspace of L2 (Rn ).
Hints: Show that the subspace of compactly supported functions is dense
in L2 (Rn ). Then show that every compactly supported function in L2 (Rn )
can be approximated in the L2 -norm by a sequence of functions in C0∞ ⊆ S.
Exercise 5.7. We have described the Fourier transform F as a unitary op-
erator on L2 (Rn ). What are the eigenvalues and eigenfunctions (λk , ψk (x))
of this operator? For simplicity you may consider only the case n = 1. Note:
this exercise is related to Exercise 2.6 of Chapter 2.
Exercise 5.8. Suppose that f ∈ L1 (R1 ) and that f (x) is continuous at the
point x = x0 . Prove that for large T , the limit satisfies
  +R
1 T 1
lim √ eiξx0 fˆ(ξ) dξ dR = f (x0 ) .
T →+∞ T 0 2π −R

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/06

Chapter 6

Wave equations on Rn

Numerical simulation of a merger of two black holes generating gravita-

tional waves, similar to those observed by the LIGO detectors. Im-
age credit: SXS, the Simulating eXtreme Spacetimes (SXS) project
(http://www.black-holes.org).

Solutions of the wave equation describe the propagation of light, sound

waves in a gas or fluid, gravitational waves in the interstellar vacuum, and
many other phenomena. It is one of my favorite equations. Posed in R1t ×Rnx ,
the initial value problem (or Cauchy problem) for the equation looks very
similar to (2.13) of Chapter 2:
(6.1) ∂t2 u − Δu = 0, x ∈ Rn , t ∈ R1 ,
with initial or Cauchy data for u(x, t) given by
u(0, x) = f (x) , ∂t u(0, x) = g(x) .

121

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 122 6. Wave equations on Rn

The given initial data u(0, x) = f (x) is often referred to as the initial position
or displacement of the field u(t, x), while the data for ∂t u(0, x) = g(x) is
called the initial velocity, or momentum, as it specifies the time derivative
of the field u(t, x) at the initial moment.

6.1. Wave propagator by Fourier synthesis

Assuming first that f (x), g(x) ∈ S, that is, it is in Schwartz class, we may
take the Fourier transform of the equation to obtain
(6.2) ∂t2 û(t, ξ) + |ξ|2 û(t, ξ) = 0 .
Solutions of this second-order ODE are composed of linear combinations
of e±i|ξ|t ; taking into account that û(0, ξ) = fˆ(ξ) and ∂t û(0, ξ) = ĝ(ξ), we
derive the expression

(6.3) ˆ + sin(|ξ|t) ĝ(ξ) .

û(t, ξ) = cos(|ξ|t)f(ξ)
|ξ|
The inverse Fourier transform gives the solution
  
1 ˆ sin(|ξ|t)
(6.4) u(t, x) = √ n e iξ·x
cos(|ξ|t)f(ξ) + ĝ(ξ) dξ .
2π |ξ|
Observe from (6.3) that û(t, ξ) ∈ S(Rnx ) for each time t, and therefore our
solution given in (6.4) is in Schwartz class as well. However, this high level
of smoothness is not at all necessary, and one notes that expression (6.4)
makes sense whenever fˆ(ξ), ĝ(ξ)
|ξ| ∈ L (Rx ), in which case it may not possess
2 n

derivatives in the classical sense.

Theorem 6.1. For f, g ∈ S, expression (6.4) gives a solution of the wave
equation u(t, x) ∈ S(Rnx ) for each time t ∈ R. For fˆ(ξ), ĝ(ξ)
|ξ| ∈ L (Rx ), (6.4)
2 n

gives a weak solution to the wave equation in the sense that (6.2) is satisfied.

One basic property satisfied by solutions of the wave equation is the

principle of “energy” conservation under time evolution. The energy of a
solution is defined to be the expression

1
(6.5) E(u) = (∂t u(t, x))2 + |∇x u(t, x)|2 dx .
2
Theorem 6.2. For (f, g) ∈ H 1 (Rnx ) × L2 (Rnx ), the energy (6.5) is conserved
for solutions of the wave equation (6.1).

Proof. Using expression (6.3) for the Fourier transform of the solution, we
give a similar formula for the vector (∂x u(t, x), ∂t u(t, x))T :
    iξ
 
∂x u(t, x) 1 cos(|ξ|t) |ξ| sin(|ξ|t) iξ fˆ(ξ)
=√ n e iξ·x
iξ dξ .
∂t u(t, x) 2π |ξ| sin(|ξ|t) cos(|ξ|t) ĝ(ξ)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.2. Lorentz transformations 123

By inspection this is a well-defined vector valued function in [L2 (Rn )]n+1 as

long as (∇x f (x), g(x))T ∈ [L2 (Rn )]n+1 , which it is by hypothesis. By the
Plancherel identity of Theorem 5.4,
 iξ
 
2 2
cos(|ξ|t) |ξ| sin(|ξ|t) iξ fˆ(ξ) 2
∂t u(x, t) L2 + ∂x u(x, t) L2 = iξ L2
|ξ| sin(|ξ|t) cos(|ξ|t) ĝ(ξ)
 
iξ fˆ(ξ) 2
L2 = ∇ x f L2 + g L2 .
2 2
=
ĝ(ξ)
The second to last equality holds because the 2 × 2 matrix featured in this
calculation is unitary. 

An alternative proof works in the case that we also have ∂t2 u, Δu ∈ L2 ;

then for solutions of (6.1),

d 1
E(u) = 2∂t u∂t2 u + 2∂x u∂t ∂x u dx
dt 2 Rn

= ∂t u ∂t2 u − Δu dx = 0 .
Rn

6.2. Lorentz transformations


Just as the Laplace operator Δ = nj=1 ∂x2j is invariant under translations
x = x + c and rotations x = Rx, where RT = R−1 , as we have seen in
Chapter 1, Exercise 1.3, the wave operator or d’Alembertian

n
= ∂t2 −Δ= ∂t2 − ∂x2j
j=1

is invariant under a group of transformations of the space-time R1t × Rnx

known as the Lorentz group. Elements of this transformation group are
generated by the same spatial rotations of Rnx as above and hyperbolic ro-
tations which involve time as well as space. Define a hyperbolic rotation in
the (t, x1 ) coordinate plane in R2 ⊆ Rn+1 by
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
t cosh(ψ) sinh(ψ) ... 0 t t
⎜x 1 ⎟ ⎜ sinh(ψ) cosh(ψ) ⎟ ⎜ x1 ⎟ ⎜ x1 ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟
⎜ .. ⎟ = ⎜ .. ⎟ ⎜ .. ⎟ = H(ψ) ⎜ .. ⎟ .
⎝ . ⎠ ⎝ . I ⎠⎝ . ⎠ ⎝ . ⎠
n−1×n−1
xn 0 ... xn xn
One calculates that det(H(ψ)) = cosh(ψ)2 −sinh(ψ)2 = 1 and that H −1 (ψ) =
H(−ψ). Vector fields in the (t, x1 ) coordinate plane transform as follows:
      
∂t cosh(ψ) sinh(ψ) ∂ t ∂ t
(6.6) = = H2×2 (ψ) ,
∂x1 sinh(ψ) cosh(ψ) ∂x 1 ∂x 1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 124 6. Wave equations on Rn

while all other ∂xj = ∂x j are invariant. The Lorentz group of transforma-
tions on the space-time R1t × Rnx is generated by all spatial rotations R and
by the hyperbolic rotations H(ψ). It is a Lie group known as SO(1, n). The
set of space-time translations (t , x ) = (t + b, x + c) leave the wave equation
invariant, and so do the elements of the Lorentz group.
Proposition 6.3. The Lorentz transformations leave the d’Alembertian op-
erator invariant.

Proof. Since the group generators involving translations obviously leave the
d’Alembertian invariant, and all spatial rotations R just involve the Lapla-
cian, it suffices to check invariance under the hyperbolic rotations H(ψ).
Write the d’Alembertian operator using matrix notation:
 T   
∂t 1 0 ∂t
u = (∂t − Δ)u =
2
u.
∂x 0 −In×n ∂x
Use the transformation rule (6.6) for vector fields to check the invariance of
the d’Alembertian:
 T  T    
∂ t H2×2 0 1 0 H2×2 0 ∂ t
.
∂x 0 I(n−1)×(n−1) 0 −In×n 0 I(n−1)×(n−1) ∂x
The only nontrivial part of this matrix calculation is the upper left hand
2 × 2 block,
     
T 1 0 cosh(ψ) sinh(ψ) 1 0 cosh(ψ) sinh(ψ)
H2×2 H2×2 =
0 −1 sinh(ψ) cosh(ψ) 0 −1 sinh(ψ) cosh(ψ)
   
cosh (ψ) − sinh (ψ)
2 2
0 1 0
= = ,
0 sinh2 (ψ) − cosh2 (ψ) 0 −1
which is precisely to say that the d’Alembertian  is invariant under H(ψ).
This is the stated result. 

The effects of the Lorentz transformation H(ψ) on the time axis {x = 0},
exhibited in Figure 1, are given by
x 1 = sinh(ψ)t , t = cosh(ψ)t ;
therefore their ratio
x 1
= tanh(ψ) := v
t
gives the velocity of the new frame of reference with respect to the old. We
note that |v| = | tanh(ψ)| < 1. In terms of v, the hyperbolic rotation can be
written in a more familiar form:
⎛ ⎞
√ 1 2 √ −v 2
H = ⎝ −v
1−|v| 1−|v| ⎠
.
√ 2
√ 1
2 1−|v| 1−|v|

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.2. Lorentz transformations 125

The coordinate plane {(0, 0, x2 , . . . , xn )} is invariant under H(ψ), but the

x1 -axis {(t, x2 , . . . , xn ) = 0} is moved,
x 1 = cosh(ψ)x1 , t = sinh(ψ)x1 ,
so that the {t = 0} coordinate plane is tilted in space-time as well, at slope
x 1
= coth(ψ) .
t
A diagram of the new space-time coordinates under a hyperbolic rotation is
as follows.

Figure 1. This is the image of the x1 - and t-axes and the light cone
under a Lorentz transformation consisting of a hyperbolic rotation of
the (t, x1 ) coordinates.

The set that remains invariant under the transformations of the Lorentz
group is the light cone LC = {t2 − |x|2 = 0}. This fact is checked as follows:
take (t, x) ∈ LC, then
 T   
t 1 0 t
=0.
x 0 −In×n x
As we have checked above,
   
T 1 0 1 0
L L= ;
0 −In×n 0 −In×n
therefore under a Lorentz transformation L : (t, x) → (t , x ), we have
  T       T    
t 1 0 t t T 1 0 t
  = L L =0.
x 0 −In×n x x 0 −In×n x

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 126 6. Wave equations on Rn

The geometric interpretation of these properties has to do with the matrix

 
1 0
g :=
0 −In×n
that defines the Minkowski metric

ds2 = (dt, dx1 , . . . , dxn )T g(dt, dx1 , . . . , dxn )

on the space-time R1t × Rnx . What we have just shown is that this metric is
invariant under the Lorentz group.
The invariance of the wave equation under the Lorentz group of trans-
formations was part of a paradox that physics faced towards the end of the
nineteenth century. As we have seen in (1.1), Maxwell’s equations are in-
timately tied to the wave equation. The equations of classical mechanics
are invariant under the Galilean transformation group, that is, the trans-
formation group consisting of rotations and translations of space along with
constant velocity translations. And the theory of electricity and magnetism
is invariant under the Lorentz group. The fact is that these are incom-
patible. Compatibility was restored in 1905 when Einstein introduced the
special theory of relativity. However, this was a revolutionary change in our
perception of the universe, for which many intuitive ideas about space-time
had to be modified. One of these we have seen above; the action of a hyper-
bolic rotation transforms the plane {t = 0} of spatial coordinates as well as
the time axis. Since we think of the spatial coordinate plane as being the
“present” state of the universe, it is a new idea that there is no universally
valid instant that can globally be considered to be the present, and that the
sense of the present is relative to the frame of reference of the observer. The
concept of simultaneity has to be sacrificed. In terms of the wave equation,
any one of the hypersurfaces {t = 0} can be used as a hypersurface on which
to pose initial data for the wave equation.

6.3. Method of spherical means

There is another method for representing the solution of the wave equation
in n space dimensions (6.1), based on the spherical means that we have
encountered in Chapter 4, formula (4.13). Recall the elementary solution
method in the case of spatial dimension n = 1 via the d’Alembert formula

1 1 x+t
u(t, x) = (f (x + t) + f (x − t)) + g(y) dy .
2 2 x−t
The goal of this section is to produce similar expressions for the solution of
the wave equation in the case of higher space dimensions.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.3. Method of spherical means 127

Definition 6.4. Given a function h(x) ∈ C(Rn ), its spherical mean centered
about the point x ∈ Rn is

1
(6.7) M (h)(x, r) := h(y) dSy ,
ωn rn−1 |x−y|=r

the average of h(x) over a sphere Sn−1 centered at x of radius r.

Recall that ωn rn−1 is the surface area of the sphere Sn−1 of radius r
in n dimensions. When n = 1, the spherical mean of a function f (x) is
M (f )(x, r) = 12 (f (x + r) + f (x − r)), reminding one of the first term of the
d’Alembert formula. Changing variables in the integral (6.7), y → x + rξ,
ξ ∈ S1 (0), there is another useful expression for the spherical mean:

1
(6.8) M (h)(x, r) = h(x + rξ) dSξ .
ωn |ξ|=1

Expression (6.8) is defined for r ≥ 0, but it is clear from this second expres-
sion that M (h)(x, r) is an even function of r:

M (h)(x, −r) = M (h)(x, r) .

Lemma 6.5 (Darboux equation). Given h ∈ C 2 (Rn ),

n−1
Δx M (h)(x, r) = ∂r2 + ∂r M (h)(x, r) ,
r

a formula that relates the Laplace operator of M (h) in the n-dimensional

x-variables to an operator involving only one-dimensional r derivatives.

Proof. The Darboux equation follows from a calculation using multivariate

calculus. First take one derivative:

1
∂r M (h)(x, r) = ∂r h(x + rξ) dSξ
ωn |ξ|=1
 
n
1
= ∂xj h(x + rξ)ξj dSξ .
ωn |ξ|=1 j=1


We notice that nj=1 ∂xj h(x + rξ)ξj = ∇h · N , the outwards normal de-
rivative of h on the unit sphere. Continue this line of calculation using

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 128 6. Wave equations on Rn

Green’s theorem:
 n 
1 1
∂xj h(x + rξ)ξj dSξ = rΔx h(x + rξ) dξ
ωn |ξ|=1 ωn |ξ|<1
j=1

1
= Δx h(y) dy
ωn rn−1 |x−y|<r
1  r 
= Δx h(y) dSy dρ
ωn rn−1 0 |x−y|=ρ
1  r
= n−1 Δx ρn−1 M (h)(x, ρ) dρ .
r 0

The second step is to take a second derivative after multiplying through by

rn−1 :
  r
n−1
∂r r ∂r M (h)(x, r) = ∂r Δx ρn−1 M (h)(x, ρ) dρ
0
n−1
= Δx r M (h)(x, r) .

Therefore
1
(6.9) Δx M (h)(x, r) = ∂r rn−1 ∂r M (h)(x, r)
rn−1
n−1
= ∂r2 + ∂r M (h)(x, r) ,
r
which is the result of the lemma. 

Proposition 6.6. (i) For h(x) ∈ C(Rn ), the value of h(x) at any x ∈ Rn
can be recovered from its spherical mean:

h(x) = lim M (h)(x, r) = M (h)(x, 0) .

r→0

(ii) Additionally, for h(x) ∈ C 2 (Rn ),


r
∂r M (h)(x, 0) = lim h(x + rξ) dξ = 0 .
r→0 ωn |ξ|<1

Formulae involving spherical means can be used to give an expression for

the solution of the wave equation. Indeed, suppose that u(t, x) is a solution
of the Cauchy problem for the wave equation (6.1). Then its spherical mean
M (u)(t, x, r) can be defined from (6.7), and it satisfies an auxiliary equation
in the reduced space-time variables (t, r) ∈ R2 . Specifically, define

1
M (u)(t, x, r) = u(t, x + rξ) dSξ .
ωn |ξ|=1

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.3. Method of spherical means 129

Taking time derivatives and using the fact that u(t, x) is a solution of (6.1),
one obtains

2 1
∂t M (u)(t, x, r) = ∂ 2 u(t, x + rξ) dSξ
ωn |ξ|=1 t

1
= Δx u(t, x + rξ) dSξ = Δx M (u)(t, x, r) .
ωn |ξ|=1
Now use the Darboux equation of Lemma 6.5,
n−1
(6.10) ∂t2 M (u)(t, x, r) = ∂r2 +
∂r M (u)(t, x, r) ,
r
a PDE in the two variables (t, r) ∈ R2 . This is known as the Euler–Poisson–
Darboux equation.

Spherical means in R3 . Equation (6.10) is normally posed as an initial

value problem
n−1
(6.11) ∂t2 M (u)(t, x, r) = ∂r2 + ∂r M (u)(t, x, r) ,
r
M (u)(0, x, r) = M (f )(x, r) , ∂t M (u)(0, x, r) = M (g)(x, r) .
The precise methods and the character of the solution depend quite a bit on
the spatial dimension under consideration. The most straightforward case
is in dimension n = 3. In the Euler–Poisson–Darboux equation (6.11), use
the substitution M (u)(t, x, r) → rM (u)(t, x, r), giving
2
∂t2 rM (u)(t, x, r) = r ∂r2 + ∂r M (u)(t, x, r) = ∂r2 rM (u)(t, x, r) .
r
Therefore the function v(t, r) = rM (u)(t, x, r) is a solution of the wave
equation in one dimension, and the variable x is relegated to the role of a
parameter. The solution is given by the d’Alembert formula
v(t, r) = rM (u)(t, x, r)
1
= (r + t)M (f )(x, r + t) + (r − t)M (f )(x, r − t)
2

1 r+t
+ ρM (g)(x, ρ) dρ .
2 r−t
Since M (f )(x, r) and M (g)(x, r) are even functions under r → −r, then
rM (f )(x, r) and rM (g)(x, r) are odd; therefore we may rewrite the above
expression, dividing through by r:
1
M (u)(t, x, r) = (t + r)M (f )(x, t + r) − (t − r)M (f )(x, t − r)
2r
 t+r
1
+ ρM (g)(x, ρ) dρ .
2r t−r

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 130 6. Wave equations on Rn

 r−t
We have used the fact that t−r ρM (g)(x, ρ) dρ = 0, which holds because
rM (g)(x, r) is odd. Taking the limit as r → 0 and using Proposition 6.6, we
recover a representation for the solution u(t, x).

Theorem 6.7 (Kirchhoff’s formula). When n = 3, the solution to the wave

equation (6.1) is given by the expression

(6.12) u(t, x) = ∂t tM (f )(x, t) + tM (g)(x, t) ,

which is well defined as long as f (x) ∈ C 1 (R3 ) and g(x) ∈ C(R3 ).

It is quite explicit to see that in terms of pointwise regularity, the solu-

tion is generally less smooth than the initial data, for the Kirchhoff formula
depends upon the derivative of f (x). Specifically, carrying out the differen-
tiation in (6.12), we obtain

1 x−y
u(t, x) = tg(y) + f (y) + t∇f (y) · dSy .
2
4πt |x−y|=t |x − y|

Corollary 6.8. Given initial data g(x) ∈ C 2 (R3 ) and f (x) ∈ C 3 (R3 ), the
spherical means solution given in (6.12) is a solution u(t, x) ∈ C 2 (R1t × R3x ).

The corollary points to the phenomenon of derivative loss in the solution

process for the wave equation, where three derivatives of the initial data
are required in order to ensure that the solution is classical, namely twice
differentiable. This loss of differentiability for n ≥ 2, due to focusing effects,
is the topic of Exercise 6.2. In contrast, when solutions are viewed in the
sense of the Sobolev spaces H s through the energy, there is no loss visible:
 
1 1 1 2 1
E(u)(t) = (∂t u) + |∇x u| dx =
2 2
g + |∇x f |2 dx .
2 2 2 2
The Sobolev regularity of the solution u(t, x) is the same as the Sobolev
regularity of the initial data.

Spherical means in Rn for odd n. There are similar expressions for the
solution of the wave equation for x ∈ Rn , for n ≥ 3 and odd. Returning
to the Euler–Poisson–Darboux equation (6.10) for the n-dimensional wave
equation
n−1
∂t2 M (u) = ∂r2 + ∂r M (u) ,
r
we seek an algebraic reduction to the wave equation in two dimensions in
the variables (t, r). This is able to be carried out in the odd-dimensional
case.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.3. Method of spherical means 131

Proposition 6.9. Suppose that k ≥ 1 is an integer and that h = h(r) ∈

C k+1 (R1+ ). Then
1 k−1  1 k
∂r2 ∂r r2k−1 h = ∂r r2k ∂r h ,
r r
1 k−1  
k−1
∂r r2k−1 h = βjk rj+1 ∂rj h
r
j=0

with the combinatorial coefficients being β0k = 1 · 3 · . . . (2k − 1) := (2k − 1)!!.

Proof. An argument by induction will work. 

The result is useful because we may set n = 2k + 1 and take

1 k−1 
v(t, x, r) := ∂r r2k−1 M (u) (t, x, r)
r
for a given solution u(t, x) of (6.1). Then using the identities from Proposi-
tion 6.9,
1 k−1 
∂r2 v = ∂r2 ∂r r2k−1 M (u)
r
1 k
= ∂r r2k ∂r M (u)
r
1 k−1 
= ∂r r2k−1 ∂r2 M (u) + 2kr2k−2 ∂r M (u)
r
1 k−1  2k
= ∂r r2k−1 ∂r2 M (u) + ∂r M (u)
r r
1 k−1 
= ∂r r2k−1 ∂t2 M (u) = ∂t2 v(t, r) .
r
Therefore v(t, x, r) is the quantity that satisfies the one-dimensional wave
equation in the variables (t, r), for which we can apply the d’Alembert for-
mula:

1 1 r+t
v(t, x, r) = v(0, r + t) + v(0, r − t) + ∂t v(0, x, ρ) dρ
2 2 r−t

with initial data

1 k−1
v(0, x, r) = ∂r r2k−1 M (f ) (x, r) ,
r
1 k−1
∂t v(0, x, r) = ∂r r2k−1 M (g) (x, r) .
r

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 132 6. Wave equations on Rn

One recovers the solution as a limit

v(t, x, r)
(6.13) u(t, x) = lim M (u)(t, x, r) = lim
r→0 r→0 β0k r
1  1 (n−3)/2
= ∂t ∂t tn−2 M (f )(t, x)
(n − 2)!! t
1 (n−3)/2 
+ ∂t tn−2 M (g)(t, x) .
t
Theorem 6.10. For n odd, f ∈ C (n−1)/2 (Rn ), and g ∈ C (n−3)/2 (Rn ), the
solution formula (6.13) gives a classical solution to the wave equation in Rn .

It is interesting to quantify the possible loss of smoothness of the solution

over the initial data, made clear by formula (6.13). Indeed, the reduction
process from M (u) to v involves k − 1 = (n − 3)/2 derivatives, and the limit
involves one derivative; therefore in general the solution will be less regular
than the initial data by k = (n − 1)/2 many derivatives.

6.4. Huygens’ principle

Huygens’ principle is the expression of the principle of finite propagation
speed, analogous to the case of the one-dimensional wave equation in Chap-
ter 2. This general form of Huygens’ principle is valid for the wave equation
in any space dimension and for hyperbolic equations in general. The strong
form of Huygens’ principle says more than this; it is the property that so-
lutions are supported precisely on the union of light cones which have their
vertex on the support of the initial data. There is a difference in dimen-
sion concerning this strong form of the Huygens’ principle; it holds for odd
space-dimensional problems, but not in even dimensions. We state it here
for the case of three space dimensions.
Theorem 6.11. Consider solutions to the wave equation for x ∈ R3 , and
assume that the Cauchy data f (x) and g(x) are compactly supported, such
that supp(f ) ∪ supp(g) ⊆ BR (0). Then
(1) (Huygens’ principle). The solution u(t, x) has its support within the
bounded region BR+|t| (0):
supp(u(t, ·)) ⊆ BR+|t| (0) .
(2) (Strong Huygens’ principle). Additionally, for |t| > R, for any space-
time point (t, x) in the region inside the light cones given by {(t, x) : |x| ≤
|t| − R}, again the solution vanishes; u(t, x) = 0.

Proof. (1) of Huygens’ principle. The result follows from the form of the
Kirchhoff formula. Express the solution u(t, x) in terms of spherical means

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.4. Huygens’ principle 133

over spheres of radius |t| in the initial hyperplane {t = 0} that are of the form
{y : |x − y| = |t|} (the intersection of the backwards light cone emanating
from (t, x) and the Cauchy hypersurface). If |x| > R+|t|, then the associated
sphere does not intersect BR (0), which contains the support of the initial
data, and hence the solution at that space-time point vanishes.
(2) of the strong Huygens’ principle. When |t| > R and |x| < |t| − R,
again the backwards light cone emanating from the point (t, x) does not
intersect the support of the initial data, in this case because the sphere
{y : |x − y| = |t|} is too big and has passed outside of BR (0).
One notes that this proof is geometric, only depending upon the fact
that the solution is represented as a spherical mean. Therefore the result
holds for solutions of the wave equation in arbitrary odd dimensions, as
exhibited in (6.13). 

In fact a more precise statement is true; at times t > 0, the solution

u(t, x) is supported within the region consisting of the union of light cones
LC+ (t, x) = {(t, y) : t2 − |y − x|2 = 0, t > 0}
whose vertices x lie in the set supp(f ) ∪ supp(g). A space-time picture of
the support of the solution is depicted in Figure 2.

xn

x1

Figure 2. Space-time picture of the support of a solution. The vertical

green line is the world line of a stationary observer.

The finite propagation speed property is sometimes known as the weak

Huygens’ principle; it is a central feature of hyperbolic equations. It implies
that in particular a signal will travel with finite velocity. That is, for an
observer standing still at point x0 ∈ Rn , |x0 | > R, the solution satisfies
u(x0 , t) = 0 until |t| + R > |x0 |.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 134 6. Wave equations on Rn

Corollary 6.12. If the initial data f (x), g(x) vanish on BR (0), then for 0 <
|t| < R, the solution u(t, x) must vanish on the cone {(t, x) : |x| < R − |t|}.
In particular if f, g are identically zero, then u(t, x) will also be zero.

Proof. This is a local uniqueness theorem. Suppose that supp(f ) ∪ supp(g)

⊆ Ha,v = {x ∈ Rn : a ≤ v · x}, a half space; then supp(u(t, x)) ⊆ Ha+|t|,v .
Therefore no data in any of the enveloping half space HR−|t|,v , |v| = 1 to
BR−|t| (0) can propagate into the cone pictured in Figure 3. 

HR - |t|, v

xn

x1

Figure 3. The cone of uniqueness.

Definition 6.13. (i) The domain of influence of a set A ⊆ Rn is the

region in Rn+1 in which a solution u(x, t) of the wave equation can
be affected by data in A.
(ii) The domain of dependence of a set B ⊆ Rn+1 is the region A ⊆ Rn
in which data can influence the solution in B.

The results of Theorem 6.17 and Corollary 6.12 imply the following
pictures for the domains of influence and dependence of solutions of the
wave equation, given respectively in Figures 4 and 5.

6.5. Paley–Wiener theory

It is evident that the complex exponentials e−iξ·x , with ξ ∈ Rn and x ∈ Rn ,
extend as holomorphic functions e−iζ·x to ζ ∈ Cn . More specifically to its
character, the extension is defined over all of Cn , meaning that it is entire,
and furthermore the extension has bounds on its growth at infinity of the
form
|e−iζ·x | ≤ e|x||im(ζ)| .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.5. Paley–Wiener theory 135

Figure 4. Domain of influence of the set A ⊆ Rn .

Figure 5. Domain of dependence of the point B.

Definition 6.14. An entire function g(ζ), ζ ∈ Cn , is of exponential type R

if it satisfies the estimates
(6.14) (1 + |ζ|2 )N/2 |g(ζ)| ≤ CN eR|im(ζ)|
for all N ∈ N.

Examples of the behavior are given by the Fourier transform of C0∞

functions, as we will describe presently. Indeed, the factor (1 + |ζ|2 )N/2 is
present to ensure that all such functions have smooth, well-behaved Fourier
transforms. The complex exponential above is not strictly speaking of ex-
ponential type because it only satisfies (6.14) for the case N = 0. Another
example of an entire function, which, however, is not of exponential type, is
g(ζ) = e−ζ = e− 2 (ξ
2 /2 1 2 −η 2 )−iξη

for ζ = ξ + iη ∈ C. Given any function f ∈ S, its Fourier transform can be

viewed as a superposition of complex exponentials

1
ˆ
f (ξ) = √ n e−iξ·x f (x) dx ,
2π Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 136 6. Wave equations on Rn

where each complex exponential extends to an entire function on Cn but the

function fˆ(ξ) itself need not necessarily have a holomorphic extension off of
the real subspace Rn ⊆ Cn at all. However, if a function g(ξ) is such that
g(ξ) = fˆ(ξ) with f (x) ∈ C0∞ (Rn ), in particular if f (x) has compact support,
then g(ξ) will have an entire holomorphic extension.
Proposition 6.15. Suppose that f ∈ C0∞ (Rn ), with supp(f ) ⊆ {x ∈
Rn : |x| ≤ R}. Then fˆ(ξ) = g(ξ) extends to an entire function, which is
of exponential type R.

Proof. For each ζ ∈ Cn , the integral


1
g(ζ) = √ n e−iζ·x f (x)dx
2π BR (0)
converges absolutely, uniformly over bounded sets of ζ. To check that g(ζ)
is holomorphic, we simply test the Cauchy–Riemann equations:

1 1
∂ζ̄ g(ζ) = √ (∂ξ + i∂η )g(ζ) = √ n ∂ζ̄ (e−iζ·x )f (x) dx = 0 .
2 2π
Lastly, for α a multi-index with |α| = N , we need to estimate the integrals
     α 

 α   1 
ζ e −iζ·x 
f (x)dx =  e −iζ·x
∂x f (x)dx
  i 
R n BR (0)

≤ e|x||im(ζ)| |∂xα f (x)|dx ≤ C|α| eR|im(ζ)| .
BR (0)

The content of Paley–Wiener theory is to say that the converse also holds.
Theorem 6.16 (Paley–Wiener). Suppose that g(ζ) is an entire function of
exponential type R. Then there is a function f (x) ∈ C0∞ (Rn ) with supp(f ) ⊆
BR (0) such that

1
g(ζ) = f (ζ) = √ n e−iζ·x f (x) dx .
ˆ
2π
Proof. It is natural to restrict g to ξ ∈ Rn , and define

1
f (x) = √ n eiξ·x g(ξ)dξ.
2π R n

Since |g(ξ)| ≤ CN (1 + |ξ|2 )−N/2 as in (6.14), this integral is absolutely con-

vergent. We may also take an arbitrary number of derivatives of f (x),

1
∂x f (x) = √ n
β
eiξ·x (−iξ)β g(ξ) dξ ,
2π Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.5. Paley–Wiener theory 137

and the integral remains absolutely convergent. Thus f ∈ C ∞ (Rn ) and the
question that remains has to do with its support.
Using the fact that g(ζ) is holomorphic, consider deformations of the
region of integration off of the real axis in the complex space Cn ,

1  
√ n ei((ξ1 +iη1 )x1 +ξ ·x ) g(ξ1 + iη1 , ξ  ) dξ1 dξ 
2π
for η1 ∈ R. This is independent of η1 , as can be shown by Cauchy’s theorem,
taking the limit of an integral in ζ1 = ξ1 + iη1 over the contour given by
Figure 6 and letting T → +∞. The decay condition (6.14) implies that there

ξ1
-T T

Figure 6. The contour.

are no contributions from the boundaries ξ1 = ±T in the limit. Repeating

this argument in all variables, we show that for any desired η ∈ Rn ,

1
f (x) = √ n ei(ξ+iη)·x g(ξ + iη) dξ .
2π Rn
Now fix x = 0 in Rn , and choose the particular η = λ |x| x
with a real parameter
λ > 0. Then
   
1  x 
|f (x)| ≤ √ n  e iξ·x−λ|x|
g ξ + iλ dξ 
2π Rn |x|
   
1 
−λ|x|  x 
≤√ n e g ξ + iλ |x|  dξ
2π Rn

≤ CN (1 + |ξ|2 )−N/2 e−λ|x|+λR dξ .
Rn
Now suppose that |x| > R; as λ → +∞, the RHS tends to zero. Thus we
have shown that f (x) = 0. This proves that indeed supp(f ) ⊆ BR (0). 

Concerning the class of distributions mentioned in Chapter 5, there is

a similar theory for distributions of compact support, and their Fourier
transforms are entire functions of exponential type. Conversely if g(ζ) is an
entire function which for some N satisfies the estimate
(6.15) |g(ζ)| ≤ C(1 + |ζ|2 )N/2 eR|im(ζ)| ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 138 6. Wave equations on Rn

then there is a distribution f ∈ D with supp(f ) ⊆ BR (0) such that the

(generalized) Fourier transform of f is g.

Huygens’ principle revisited. Returning to the discussion of solutions

of the wave equation, our Fourier integral expression is that
  
1 ˆ + sin(|ξ|t)
u(x, t) = √ n e iξ·x
cos(|ξ|t)f(ξ) ĝ(ξ) dξ .
2π |ξ|
Let us suppose that the initial data is of compact support:

supp(f ) ∪ supp(g) ⊆ BR (0) = {x ∈ Rn : |x| < R} .

Then by Proposition 6.15, both fˆ(ξ) and ĝ(ξ) extend to entire functions
on Cn of exponential type R. Let us examine the Fourier transform of the
solution
ˆ + sin(|ξ|t) ĝ(ξ) .
û(ξ, t) = cos(|ξ|t)f(ξ)
|ξ|
 √
sin( ζ 2 t)
The individual functions cos( ζ 2 t)fˆ(ζ) and √ 2 ĝ(ζ) are entire func-
ζ
tions of ζ ∈ Cn , where we are using the notation that ζ 2 = ζ12 + ζ22 + · · · + ζn2
for ζj ∈ C. Furthermore,
 
    sin(ζ 2 t) 
  |im(ζ)||t|  
cos( ζ 2 t) ≤ e ,    ≤ e|im(ζ)||t| ;
 ζ 2 
   
ˆ
therefore the two products cos( ζ t)f(ζ) and sin( ζ 2 t)/ ζ 2 ĝ(ζ) are of
2

exponential type (R + |t|). We can conclude that the solution u(x, t) of the
wave equation is the Fourier transform of an entire function of exponential
type (R + |t|), and thus by the Paley–Wiener theorem, the solution u(x, t)
has its support in the ball BR+|t| (0) of radius R + |t|. This is a second proof
of Huygens’ principle for the wave equation.

Theorem 6.17 (Huygens’ principle again). A solution of the wave equa-

tion (6.1) with initial data with support such that supp(f )∪supp(g) ⊆ BR (0)
satisfies at nonzero times t ∈ R

supp(u(t, x)) ⊆ BR+|t| (0) .

This is a proof that is valid for all spatial dimensions n ≥ 1, independent

of the parity of the dimension. Because of this independence, techniques
based on the Fourier transform are, however, not so well adapted for a proof
of the strong Huygens’ principle that holds in odd space dimensions, as
stated in Theorem 6.11(2) for n = 3, but not in even dimensions.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.6. Lagrangians and Hamiltonian PDEs 139

6.6. Lagrangians and Hamiltonian PDEs

Many PDEs that arise in physics are derived from a Lagrangian functional,
following a formal but nevertheless compelling analogy of the principle of
least action of classical mechanics. The wave equation is a typical example.
The Lagrangian function for the wave equation (6.1) is defined by

1 1
(6.16) L := (∂t u)2 − |∇u|2 dx .
Rn 2 2
This Lagrangian has an associated action integral given by
 T  T
1 1
(6.17) S := L(u, ∂t u) dt = (∂t u)2 − |∇u|2 dxdt .
0 0 R n 2 2
From this action integral S, the wave equation arises from the principle of
least action, which dictates that the motion of a Lagrangian system is a
stationary point of the action integral. A stationary point of the action over
the time interval 0 ≤ t ≤ T is a function u(t, x) which satisfies δS = 0 for all
admissible variations v(t, x) = δu(t, x), where δu(t, x) denotes a small but
arbitrary variation of the function u(t, x). Admissible variations are smooth
and are such that v(0, x) = v(T, x) = 0, so that u(t, x) and u(t, x) + v(t, x)
have the same initial and final states over the time interval [0, T ]. In the
case of the Lagrangian L for the wave equation, a stationary point of the
action satisfies
 T
δS = ∂t u∂t v − ∇x u · ∇x v dxdt
0 Rn
 T   T

=− (∂t2 u − Δu)v dxdt + ∂t uv dx .
0 Rn Rn t=0

From the assumption that v(0, x) = v(T, x) = 0, the last term of the RHS
vanishes. The conclusion is that because v(t, x) is otherwise arbitrary, we
must have that
∂t2 u − Δu = 0 .
These are the Euler–Lagrange equations for the action (6.17). Notice that,
ironically, the formal principle of stationary action allows us to give an initial
position u(0, x) = f (x) but does not allow for setting the initial momentum
∂t u(0, x) = g(x). Furthermore this formal principle asks us to specify the
final position u(T, x), so it is not in fact compatible with an initial value
problem. Nonetheless, the principle of least action, or more generally of
stationary action, remains a guiding principle for many equations in physics,
while on a rigorous mathematical level the principle remains a formal one
in this and other cases.
A Lagrangian L and subsequently an action integral S can be defined
for more general systems; indeed, this is how wave equations are derived in

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 140 6. Wave equations on Rn

most problems in physics. Consider the more general Lagrangian


1
(6.18) L= (∂t u)2 − G(u, ∇x u) dx
Rn 2

whose associated action is given by

 T  T
1
S= L dt = (∂t u)2 − G(u, ∇x u) dxdt .
0 0 Rn 2

Suppose that the field u(t, x) is a stationary point of the action:

 T
d  1
0 = δS =  (∂t u + ε∂t v)2 − G(u + εv, ∇x (u + εv)) dxdt
dε ε=0 0 Rn 2
 T
= ∂t u∂t v − ∂∇u G(u, ∇x u) · ∇x v − ∂u G(u, ∇x u)v dxdt
0 Rn
 
T  2 
= −∂t u + ∇x · ∂∇u G(u, ∇x u) − ∂u G(u, ∇x u) v dxdt
Rn

0
T

+ ∂t uv dx .
Rn t=0

The notation is that the nonlinear function G = G(u, V ) depends upon

the variables u as well as the n components of V = ∇u. The notation for
its partial derivatives is that ∂∇u G(u, ∇x u) = ∂V G(u, V )|V =∇x u . The final
term vanishes because v(t, x) is an admissible variation. Since v is otherwise
arbitrary, the field u(t, x) must satisfy the Euler–Lagrange equations

∂t2 u − ∇x · ∂∇u G(u, ∇x u) + ∂u G(u, ∇x u) = 0 .

This is a hyperbolic equation if the matrix of partial derivatives of G with

respect to the variables V is positive definite. In the example of the wave
equation, G(u, ∇u) = 12 |∇u|2 , and ∂∇u
2 G = I.

In this formal treatment of the derivation of field theories, there is an-

other useful analogy with classical mechanics. Given a Lagrangian L, there
is a transformation of the Euler–Lagrange equations to a Hamiltonian sys-
tem, in our case to a system of Hamiltonian PDEs. This is illustrated in the
example Lagrangian

1 2
(6.19) L(u, u̇) = u̇ − G(∇x u) dx ,
Rn 2

where for clarity we have simplified the Lagrangian (6.18) above, and we are
using the notation that ∂t u = u̇. The action integral is as before
 T
S= L dt,
0

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.6. Lagrangians and Hamiltonian PDEs 141

and as above the principle of stationary action gives the Euler–Lagrange

equations
(6.20) −∂t δu̇ L + δu L = 0 ,
where δu L = ∇x · ∂∇u G(∇x u) and δu̇ L = u̇. In general a Lagrangian may
depend explicitly on time L(u̇, u, t), but in many cases, such as the one
at hand, it describes a physical process whose properties do not change
with time, and L is independent of t. In this situation, the Euler–Lagrange
equations exhibit a conservation law. This can be seen from the following
computation:
 
d
L= δu̇ L ü + δu L u̇ dx = δu̇ L ü + ∂t (δu̇ L) u̇ dx
dt

d
= δu̇ L u̇ dx .
dt
We used the Euler–Lagrange equations (6.20) in the last equality of the first
line. Therefore the conservation law is evident, namely

d
(6.21) δu̇ L u̇ dx − L = 0 .
dt
In example (6.19), this conservation law is

d 1
(∂t u)2 + G(∇x u) dx = 0 .
dt Rn 2
Definition 6.18. (i) Define the Hamiltonian of system (6.20) by

H := δu̇ L u̇ dx − L .

(ii) The conjugate momentum of the field u is defined to be

p := δu̇ L ,
giving p = p(u̇, u).
(iii) If the relationship between u̇ and p can be inverted to obtain u̇ = u̇(p, u),
at least locally, then the mapping u̇ → p is called the Legendre transform.
Using this mapping, we may rewrite the Hamiltonian

H = u̇ p dx − L = H(u, p)

in terms of the new Hamiltonian variables (u, p).

The Legendre transform offers an elegant way to transform a second-

order equation (in time) to a first-order system of equations.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 142 6. Wave equations on Rn

Theorem 6.19. The Euler–Lagrange equations (6.20) for u = u(t, x) are

equivalent in the new variables (u(t, x), p(t, x)) to the system
(6.22) ∂ t u = δp H ,
∂t p = −δu H .

The system of equations (6.22) is known as Hamilton’s canonical equations

for the evolution equations described by H.

Proof. The formal equivalence of (6.20) and (6.22) is a general fact. Firstly,

H = u̇ p dx − L

so that u̇ = δp H. Secondly, we note that δu H = −δu L so that

d
ṗ = δu̇ L = δu L = −δu H .
dt


Exhibiting this transformation in the setting of the wave equation, we

have 
1 2 1
L= u̇ − |∇x u|2 dx ,
R n 2 2
from which the Legendre transform gives p = δu̇ L = u̇. Then
  
1 2 1
H = u̇ p dx − L(u, u̇) = u̇ dx − L =
2
p + |∇x u|2 dx .
2 2
Hamilton’s canonical equations are then
(6.23) ∂ t u = δp H = p ,
∂t p = −δu H = Δu ,
which is of course the wave equation presented as a first-order system of
equations. The energy functional E(u) for the wave equation is the Hamil-
tonian H(u, p) for the system.

Hamiltonian PDEs. The above formal exposition is a lead-in to study

other PDEs that can be posed in the form of Hamiltonian systems. Consider
systems of equations of the form
(6.24) ∂t z = Jgradz H ,
where z is a vector valued function in a Hilbert space H, and the matrix
J = −J T is skew symmetric on H. The gradient gradz H of a functional is
defined using the inner product of H in this way:
d 
δz H · v =  H(z + sv) := gradz H, vH .
ds s=0

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 6.6. Lagrangians and Hamiltonian PDEs 143

Rewriting the wave equation (6.23) as

      
u p 0 I −Δu
(6.25) ∂t = = := Jgrad(u,p) H ,
p Δu −I 0 p
the wave equation takes this form, with H = L2u ⊕ L2p and
 
0 I
J := .
−I 0
We note that this matrix J satisfies J = −J T on H.
Proposition 6.20 (Conservation of energy). The Hamiltonian H(u, p) is a
conserved quantity for solutions of (6.25).

Proof. This can reasonably be called the law of conservation of energy as

the Hamiltonian function is often, not always, the energy of the system being
considered. The following classical calculation using the chain rule makes
the assumption that solutions to (6.24) exist. Indeed, from (6.24),
d
H(z(t)) = gradz H, żH = gradz H, Jgradz HH = 0 ,
dt
where we have used the skew symmetry of the matrix J. 

Pursuing our analogy with ODEs, this is a treatment of initial value

problems for PDEs that are evolution equations as dynamical systems in a
function space, naturally a space of infinite dimensions. A number of other
PDEs of interest can be viewed as Hamiltonian systems in infinitely many
variables. Several examples follow.

1. Nonlinear wave equations. Semilinear hyperbolic equations of this

form can be treated as Hamiltonian PDEs. Define the Hamiltonian to be

1 2 1
H(u, p) = p + |∇u|2 + G(x, u) dx .
R n 2 2
Then the gradient of H is given by
 
−Δu + ∂u G(x, u)
gradH(u, p) = ,
p
and therefore the equations of motion are posed as
 
u
∂t = JgradH ,
p
which is a first-order system equivalent to the nonlinear wave equation
∂t2 u = Δu − ∂u G(x, u) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 144 6. Wave equations on Rn

2. Nonlinear Schrödinger equations. The Hamiltonian for this set of

equations is in the form

1
H(ψ, ψ̄) = |∇ψ|2 + Q(x, ψ, ψ̄) dx ,
2
where Q(x, ψ, ξ) : Rx × Cψ × Cξ → C has the property that Q(x, ψ, ψ̄) is
real valued. Then δψ̄ H(ψ, ψ̄) = − 12 Δx ψ + ∂ψ̄ Q, and setting J = iI (which
is a nondegenerate, skew symmetric operator),
∂t ψ = Jδψ̄ H
1
= i(− Δψ + ∂ψ̄ Q(x, ψ, ψ̄)) .
2
When Q = ±|ψ| , this equation is the well-known cubic nonlinear Schrö-
4

dinger equation, where the + sign is the defocusing case and the − sign is
the focusing case.

3. Korteweg–de Vries equation (KdV). This famous dispersive equa-

tion first arose in the nineteenth century as a model of waves in the free
surface of water in a canal. Currently it is used in modeling numerous phe-
nomena including tsunami propagation. It is also well known as a PDE that
is a completely integrable Hamiltonian system, where the study of the phase
space of solutions has led to discoveries as far ranging as algebraic geometry
and inverse spectral theory. The Hamiltonian is
 ∞
1
H(q) = (∂x q)2 + G(q) dx .
−∞ 12
As above, the gradient of H(q) is given by
1
δq H = − ∂x2 q + ∂q G(q) .
6
Setting J = ∂x , which again is a skew symmetric operator, we arrive at
Hamilton’s canonical equations in the form
1
∂t q = ∂x (− ∂x2 q + ∂q G(q))
6
1 3
= − ∂x q + G (q)∂x q .
6
The most well-known versions of the KdV equation are when G(q) = 13 q 3
and when G(q) = 14 q 4 .

Exercises: Chapter 6
Exercise 6.1. This problem concerns the Klein–Gordon equation
∂t2 u − Δu + m2 u = 0 ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 6 145

where in physics applications the parameter m usually represents mass. This

equation, like the wave equation, is usually posed as an initial value problem
with initial data
u(0, x) = f (x) , ∂t u(0, x) = g(x) .
Show that the following energy function is a conserved quantity:

1
EKG (u) = (∂t u)2 + |∂x u|2 + m2 u2 dx .
2 Rn
Can you find other conserved quantities for this equation?
Exercise 6.2 (Focusing singularity of solutions of the wave equation in R3
(d’après John [9])). (i) Suppose that the initial data for the wave equation
in three dimensions has spherically symmetric data:
f (x) = f (r) , g(x) = g(r) , r2 = x21 + x22 + x23 .
Show that the general solution can be expressed as
1
u(t, r) = F (r + t) + G(r − t) ;
r
that is, it consists of an incoming wave and an outgoing wave.
(ii) With the special initial data u(0, r) = 0, ∂t u(r) = g(r) with g(r) an even
function of r, then
 r+t
1
u(t, r) = ρg(ρ) dρ .
2r r−t

(iii) Set the initial data to be

g(r) = 1 , 0≤r<1, g(r) = 0 , 1≤r.
Show that u(t, r) is continuous for |t| < 1 but at time t = 1 exhibits a jump
discontinuity. This is due to the focusing of the singularity in ∂t u(0, r) given
at t = 0.
Exercise 6.3. This problem addresses the decay rate of solutions of the
wave equation in R3 . Suppose that the initial data (f (x), g(x)) ∈ C 1 (R3 ) ×
C(R3 ) and that it is supported in the bounded set B1 (0). By inspecting the
Kirchhoff formula for the solution u(t, x), show that
C
|u(t, x)| ≤
|t|
for some constant C, which can be quantified using f C 1 (R3 ) and g C(R3 ) .

Exercise 6.4 (Global existence with small initial data for certain nonlin-
ear wave equations). This question is to show that certain nonlinear wave
equations possess smooth solutions for all t ∈ R; see [12]. This contrasts

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 146 6. Wave equations on Rn

with other cases where solutions form singularities in finite time. Consider
the equation
(6.26) ∂t2 v − Δv + (∂t v)2 − |∇v|2 = 0 ,
v(0, x) = f (x) , ∂t v(0, x) = g(x) ,
with (f (x), g(x)) ∈ C 1 (R3 ) × C(R3 ) supported in a compact set.
(i) Setting u = ev − 1, show that u(t, x) satisfies the wave equation
∂t2 u − Δu = 0 ,
u(0, x) = ef (x) − 1 := F (x) , ∂t u(0, x) = g(x)ef (x) := G(x) ,
(F (x), G(x)) ∈ C 1 (R3 ) × C(R3 ) .
Explain why (F, G) has compact support.
(ii) Restricting our attention to the case x ∈ R3 , show that for sufficiently
small F C 1 (R3 ) , G C(R3 ) , the solution u(t, x) is bounded by
|u(t, x)| < 1 ,
using the result of Exercise 6.3.
In this case the transformation v → u is invertible for all (t, x) ∈ R1t ×R3x ,
giving rise to a global solution v(t, x) of equation (6.26).
Exercise 6.5 (Method of descent for the wave equation for x ∈ R2 ).
(i) Show that if x ∈ R3 but the Cauchy data for the wave equation only
depends upon (x1 , x2 ), namely
(6.27) f = f (x1 , x2 ) , g = g(x1 , x2 ) ,
then the solution of the wave equation in R1t × R3x is also independent of x3 ,
u = u(t, x1 , x2 ) ,
and therefore u(t, x1 , x2 ) satisfies the wave equation in two space dimensions:
∂t2 u − (∂x21 + ∂x22 )u = 0 .

(ii) Use the Kirchhoff formula to express the solution to the wave equation
in R3 for data satisfying (6.27).
(iii) In the expression in (ii), reparametrize the spherical integrals by their
projection onto the (x1 , x2 ) plane, e.g.,

f (y) dSy
S2 :|x−y|=t
 
= f (y1 , y2 ) t2 − ((x1 − y1 )2 + (x2 − y2 )2 ) dy1 dy2 ,
|(x1 −y1 ,x2 −y2 )|<t

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 6 147

which gives a general formula in R2 for the solution of the wave equation
u = 0 .
(iv) Describe the nature of this solution in the case that the support of f
and g as functions on R2 is compact, say, supported in the ball BR (0) =
{|(x1 , x2 )| < R}. In particular comment on Huygens’ principle. Does the
solution violate the strong form of Huygens’ principle, and why? Describe
what an observer sees as time progresses when they are situated farther than
R from the origin.
Exercise 6.6 (Equipartition of energy). The energy function for the wave
equation is

1
E(u(t, ·)) = |∂t u(t, x)|2 + |∇u(t, x)|2 dx .
2 Rn
By reason of physical analogy, the first term

1
EK (u(t, ·)) = |∂t u(t, x)|2 dx
2 Rn
is called the kinetic energy, and the second term

1
EP (u(t, ·)) = |∇u(t, x)|2 dx
2 Rn
is called the potential energy. The total energy is a conserved quantity for
solutions of the wave equation.
Show that, on average, the kinetic energy and the potential energy for
a solution of the wave equation are equal; this is in the sense that
 
1 T 1 T
lim EK (u(t, ·)) dt = lim EP (u(t, ·)) dt .
T →∞ T 0 T →∞ T 0

Hint: You may assume that the initial data (f, g) are in Schwartz class. And
you may consider using the solution by Fourier synthesis for this problem
and apply the Plancherel identity, that is, the statement that
f 2 2 = fˆ 2 2 . L L

Projects: Chapter 6
Project 6.7 (Method of descent in Rn for even dimensions n = 2m). The
solution of the wave equation in even spatial dimensions is different in char-
acter than in odd dimensions. This is described in the case of R2 in Ex-
ercise 6.5 above. A similar procedure of descent gives an explicit spherical
means solution formula for solutions to the wave equation in higher even-
dimensional Euclidean space Rn , n = 2m. Derive a similar version of the
spherical means formulae to express the solution. You will note that, simi-
lar to the case of dimension n = 2, the solution does not satisfy the strong

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 148 6. Wave equations on Rn

Huygens’ principle. Assume that the initial data is smooth and compactly
supported, derive a decay rate for the solution (in the L∞ -norm), and show
that this rate is achieved on the light cones supported over the initial data.
Show that inside the light cones the solution decays faster, and derive the
rate of decay.
Project 6.8 (The Cauchy problem for the wave equation in the Fried-
man–Robertson–Walker space-time) [1],[13]). The metric for a Friedman–
Robertson–Walker (FRW) space-time is given in terms of the line element
in the form
ds2 = −dt2 + S 2 (t)dσ 2 , S(0) = 0 ,
defined on the half space-time R1+ × R3x , where dσ 2 = dx21 + dx22 + dx23 is the
Euclidean metric of each space-like hypersurface {(t, x) : t = Const.}  R3 .
Changing time variable
dt
= S(τ ) = τ 2 ,
dτ
the metric becomes
ds2 = S 2 (τ )(−dτ 2 + dσ 2 ) .
This metric describes an emerging space-time from a Big Bang at t = τ = 0.
Consider the wave equation on R1+ × R3x in this metric,
1 2 2Ṡ 1
∂ u + 3 ∂τ u − 2 Δσ u = 0 .
u :=
S2 τ S S
Initial data is given on the Cauchy hypersurface {(τ0 , x)}  R3 ,
u(τ0 , x) = g(x) , ∂τ u(τ0 , x) = h(x) .

(a) Making the change of variables

1
∂τ (τ 3 u) ,
v(τ, x) =
τ
show that v(τ, x) satisfies the usual wave equation in Minkowski space:
∂τ2 v = Δσ v , τ >0.
Show that the initial data for v at τ = τ0 > 0 is given by
v(τ0 , x) = 3τ0 g(x) + τ02 h(x) := φ(x) ,
∂τ v(τ0 , x) = 3g(x) + τ02 Δg(x) + τ0 h(x) := ψ(x) .
Thus the solution can be given in terms of spherical means; state this ex-
pression for the solution.
(b) The inverse of the transformation is given by
 τ −τ0
τ 3 u(τ, x) = (r + τ0 )v(r + τ0 , x) dr + τ03 g(x) .
0

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 6 149

Assume (for simplicity) that h(x) = 0. Give an expression for u(τ, x) in

terms of g(x) using the spherical means expression for v and the above
inverse.
(c) The above expression is for fixed τ0 > 0 defining the Cauchy surface, and
it gives the solution at time τ > 0. Now consider the solution expression at
a fixed time τ > τ0 , and take the limit as τ0 → 0. What do you get? Can
waves pass through a Big Bang singularity?
(d) Make a sketch of the light cone structure of this problem in the original
variables (t, x).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/07

Chapter 7

Dispersion

Schrödinger’s equation: this image needs no caption.

The phenomenon of dispersion for a PDE refers to the tendency of so-

lutions to spread out in space as time evolves. The most important cases of
this phenomenon occur when components of a solution with different spatial
frequencies travel at different velocities. Light passing through a medium
such as a raindrop or the glass of a prism is an example of this, with results
that give rise to a rainbow (from multiple internal reflections in a cluster of
raindrops) or the spectral decomposition of a light beam through a prism.

151

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 152 7. Dispersion

Other dispersive phenomena in physics include wave propagation in wave-

guides, ocean waves, and solutions of the Schrödinger equations. We will
start this chapter with the latter.

7.1. Schrödinger’s equation

The most common case of an evolution equation for which dispersion plays
a role is the Schrödinger equation
1
(7.1) i∂t ψ = − Δψ , ψ : R1t × Rnx → C ,
2
with prescribed initial data
ψ(0, x) = ψ0 (x) ∈ L2 (Rn ) .
In the quantum mechanical context, the solution ψ(t, x) is known as the
wave function or the complex probability amplitude. The solution of (7.1) is
straightforward via the Fourier transform,

1
ψ̂(t, ξ) = (F ψ)(t, ξ) = √ n e−iξ·x ψ(t, x) dx ,
2π
which then must satisfy the parameter family of ODEs
1
i∂t ψ̂ = |ξ|2 ψ̂ , ψ̂(0, ξ) = ψ̂0 (ξ) .
2
Thus
ψ̂(t, ξ) = e− 2 |ξ| t ψ̂0 (ξ) .
i 2

Reconstituting the wave function with the inverse Fourier transform gives
the following expression of the solution:

1
ψ(t, x) = √ n eiξ·x e− 2 |ξ| t ψ̂0 (ξ) dξ
i 2

2π

1
eiξ·(x−y) e− 2 |ξ| t ψ0 (y) dydξ .
i 2
(7.2) = n
(2π)
Proposition 7.1. Evaluation of the complex Gaussian integral in (7.2) gives
the expression

1 1 |x−y|2
iξ·(x−y) − 2i |ξ|2 t
e e dξ = √ n e i 2t
.
(2π)n 2πit
Thus the Fourier representation of the solution leads to the expression for
the solution involving the Schrödinger kernel

1 |x−y|2
(7.3) ψ(t, x) = √ ne
i 2t
ψ0 (y) dy
Rn 2πit

:= S(t, x − y)ψ0 (y) dy = S(t)ψ0 (x) .
Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.1. Schrödinger’s equation 153

One notices formal similarities of the Schrödinger kernel with the heat
kernel and its analogous representation of solutions of the heat equation.
On the other hand, the presence of the complex unit i in the exponent
radically changes the character of the Schrödinger kernel and the behavior
of solutions. Indeed, the heat kernel satisfies
1 |x|2
−
H(t, x) = √ ne 2t ∈S ∀t > 0 ,
2πt
and in particular for t > 0, the heat kernel H(t, x) decays very rapidly as
|x| → +∞ (faster than exponential decay). On the other hand,
1 i|x|2
S(t, x) = √ ne 2t ∈ C∞ ∀t ∈ R ,
2πit
but it is not Schwartz class as it does not decay as |x| → +∞, rather it
oscillates rapidly.
Theorem 7.2. (i) For initial data ψ0 (x) ∈ S, the solution expression (7.3)
gives rise to an absolutely convergent integral, and for all t ∈ R1 , ψ(t, x) ∈ S.
(ii) For initial data ψ0 (x) ∈ L2 (Rn ), the expression (7.3) for the solution
converges in the L2 sense, and for all t ∈ R1 , ψ(t, x) ∈ L2 (Rn ).

Proof. (i) The solution formula (7.3) is that


S(t, x − y)ψ0 (y) dy = F −1 e− 2 |ξ| t ψ̂0 (ξ) ,
i 2
ψ(t, x) =
Rn
which is the inverse Fourier transform of a Schwartz class function.
(ii) Again the solution formula gives us that
 
ψ(t, ·) L2 = |ψ̂(t, ξ)| dξ = |e− 2 |ξ| t ψ̂0 (ξ)|2 dξ
i 2
2 2
(7.4)

= |ψ̂0 (ξ)|2 dξ = ψ0 2L2 .

This last fact, that the L2 -norm is preserved by the Schrödinger flow,
is central to the interpretation of the wave function in quantum mechanics.
Consider initial data ψ0 (x) ∈ L2 (Rn ), normalized to satisfy ψ0 L2 = 1; this
can be viewed as a rescaling of ψ0 to lie on the unit sphere S∞1 ⊆ L . Then
2

the quantity
|ψ(t, x)|2 dx := dPt (x)
is a time-dependent distribution on x ∈ Rn with initial distribution
|ψ0 (x)|2 dx = dP0 (x) .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 154 7. Dispersion

The normalization of the wave function implies that

 
n
Pt (R ) = |ψ(t, x)| dx =
2
|ψ0 (x)|2 dx = 1 .
Rn Rn
Theorem 7.2(ii) is the fact that the Schrödinger flow preserves probability,
and therefore both dP0 (x) and dPt (x) are probability distributions. The
interpretation of the wave function is that the probability of a quantum
particle being in a (measurable) set A at time t is given by
 
Pt (A) = dPt (x) = |ψ(t, x)|2 dx .
A A
Comparing this with the heat equation, where the probability measure in
question is dPt (x) = u(t, x) dx (which is a measure when u0 (x) ≥ 0), under
heat flow this sign condition and the L1 -norm of the solution are preserved.

Conservation laws. There are other conservation laws for solutions of

Schrödinger’s equations. The kinetic energy represented by the wave func-
tion is given by 
1
E(ψ) := |∇ψ|2 dx .
Rn 2
Expressing this in Fourier transform variables and using the Plancherel iden-
tity, it is straightforward to verify that solutions conserve (kinetic) energy
(noting that no potential energy term is included in equation (7.1)):
 
1 1
E(ψ(t, ·)) = |∇ψ(t, x)|2 dx = |ξ ψ̂(t, ξ)|2 dξ
2 2
R R
n n

1  − i |ξ|2 t 2
= ξe 2 ψ̂0 (ξ) dξ = E(ψ0 ) .
Rn 2
The probability distribution dPt (x) is called the position distribution. One
can also study the momentum distribution
dP̂t (ξ) := |ψ̂(t, ξ)|2 dξ ,
which has initial distribution dP̂0 (ξ) = |ψ̂0 (ξ)|2 dξ. Again from the Plancherel
identity,
   
dP̂t (ξ) = |ψ̂(t, ξ)|2 dξ = |ψ(t, x)|2 dx = dP0 (x) ;
Rn Rn Rn Rn

thus the total momentum probability, or the zeroth moment of dP̂t (ξ), is
preserved under Schrödinger evolution. Furthermore, all moments of the
momentum distribution are preserved.
Theorem 7.3. Define a Fourier multiplier operator m(D) by the formula

1
m(D)ψ(x) = √ n eiξ·x m(ξ)ψ̂(ξ) dξ .
2π Rn

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.1. Schrödinger’s equation 155

All such operators acting on a solution of Schrödinger’s equation (7.1) sat-

isfy
 
|m(D)ψ(t, x)| dx =
2
|m(D)ψ0 (x)|2 dx .
Rn Rn

Moments of the momentum distribution dP̂t (ξ) are of this form. There-
fore all moments of the momentum distribution are preserved under free
Schrödinger flow.

Proof. The result follows from a calculation of the integrals in the statement
of the theorem,
 
|m(D)ψ(t, x)|2 dx = |m(ξ)ψ̂(t, ξ)|2 dξ
 
= |m(ξ)ψ̂0 (ξ)| dξ = |m(D)ψ0 (x)|2 dx .
2

All moments of the measure dP̂t (x) are of this form:

  
m̂k (ψ(t, ·)) = ξ dP̂t (ξ) = ξ |ψ̂(ξ)| dξ = ξ k dP̂0 (ξ) = m̂k (ψ0 (·)) .
k k 2

In particular, it follows that for k = 2s, these are Sobolev estimates of the
solution
 
m̂2s (ψ(t, ·)) = |ξ ψ(t, ξ)| dξ = |∂xs ψ(t, x)|2 dx
s 2


= |∂xs ψ0 (x)|2 dx = m̂2s (ψ0 ) .

This leads to a natural question as to how the spatial moments of solutions

of the Schrödinger equation behave. These are
 
mk (ψ(t, x)) = xk dPt (x) = xk |ψ(t, x)|2 dx .
Rn Rn

It turns out that for ψ0 ∈ L2 (Rn ), the quantities mk (ψ(t, ·)) are not neces-
sarily bounded; see Projects 7.7 and 7.8 in this chapter. This is in contrast
to the recursive control of the moments of the analogous probability measure
given by the flow of the heat equation.
A second proof of the conservation law (7.4) for ψ(t, ·) 2L2 goes as fol-
lows. Take a domain A ⊆ Rn , that is, an open subset with sufficiently

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 156 7. Dispersion

smooth boundary. Then

  
d 1 1
ψ̄ψ dx = ∂t ψψ + ψ∂t ψ dx = Δψψ + ψ Δψ dx
dt A 2i 2i
 A
 A
1 1
=− Δψψ − ψΔψ dx = ∇ψ · ∇ψ − ∇ψ · ∇ψ dx
2i A 2i A

1
+ −ψ∇ψ · N + ψ∇ψ · N dSA .
2i ∂A
This calculation identifies the quantum mechanical flux F (ψ) across the
boundary ∂A as
1
F ·N = ψ∇ψ − ψ∇ψ · N .
2i
Now let A = R . We find that
n

d
ψ(t, ·) 2L2 = 0 ;
dt
hence ψ(t, ·) 2L2 = ψ0 2L2 for all t ∈ R1t . The advantage to this proof is that
it is very general, and in particular it applies to the Schrödinger equation
with potential
1
(7.5) i∂t ψ = − Δψ + V (x)ψ ,
2
where V (x) represents the interaction of the quantum particle in question
with a background potential energy. Two examples that we have already dis-
cussed in Chapter 2 are the quantum harmonic oscillator and the hydrogen
atom. The potential for the first is given by
1
V (x) = |x|2 ,
2
for which solutions of the associated Schrödinger equation (7.5) correspond
to the dynamics of a quantum particle trapped in a confining potential. For
the second example, let x ∈ R3 . The potential energy of the nucleus of the
hydrogen atom is given by
Z
V (x) = − ,
|x|
where the associated wave function ψ(t, x) describes the dynamics of the
electron of the hydrogen atom, which is a linear superposition of electron
orbitals and extended, or radiating, states.

7.2. Heisenberg uncertainty principle

Returning to the discussion of moments of a function ψ(x) ∈ L2 , we have
defined the kth spatial moments and momentum space moments as
 
mk (ψ0 ) = xk |ψ0 (x)|2 dx , m̂k (ψ0 ) = ξ k |ψ̂0 (ξ)|2 dξ .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.2. Heisenberg uncertainty principle 157

We have furthermore shown that the moments m0 (ψ(t, ·)) and m̂k (ψ(t, ·))
are constants of Schrödinger evolution. Introducing the Dirac “bra” := 
and “ket” :=  notation for inner products, the first moments of position
and momentum of an arbitrary wave function ψ(x) can be written as

m1 (ψ) = x|ψ(x)|2 dx = ψ|x|ψ ,
R
n

1 1
m̂1 (ψ) = ξ|ψ̂(ξ)| dξ = ψ(x) ∂x ψ(x) dx = ψ| ∂x |ψ .
2
i i
Since we are working in the setting x ∈ Rn , the operators 1i ∂x and x are
vectorial, and therefore so are the moments m1 (ψ) and m̂1 (ψ).

Proposition 7.4. Assume that the wave function ψ(x) ≡ 0 and that
m1 (ψ), m̂1 (ψ) < +∞. Under translation and change of phase, ψ(x) may
be adjusted so that m1 (ψ) = 0 = m̂1 (ψ).

Proof. Addressing first m1 (ψ), change variables by translation x → x −x0 ,

and use the notation that τx0 (ψ) := ψ(x − x0 ). Then
 
m1 (ψ) = x|ψ(x)| dx = (x − x0 )|ψ(x − x0 )|2 dx
2

= m1 (τx0 ψ) − x0 m0 (τx0 ψ) .

Since ψ ≡ 0, then m0 (ψ) = m0 (τx0 ψ) = 0, and thus we may choose

m1 (ψ)
x0 = −
m0 (ψ)
so that m1 (τx0 ψ) = 0. Similarly, consider
 
2 1
m̂1 (ψ) = ξ|ψ̂| dξ = ψ(x) ∂x ψ(x) dx .
i
Then with a change of phase,

iξ0 ·x 1
m̂1 (e ψ) = eiξ0 ·x ψ(x) ∂x eiξ0 ·x ψ(x) dx
i
 
1
= ψ(x) ∂x ψ(x) dx + ξ0 ψψ dx
i
= m̂1 (ψ) + ξ0 m̂0 (ψ) .

Again, one may choose ξ0 so that m̂1 (eiξ0 ·x ψ) = 0. 

The variance Δx (ψ) of a wave function quantifies its deviation from its
mean value, which we can write as follows. Assume that m1 (ψ) = 0 = m̂1 (ψ)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 158 7. Dispersion

without loss of generality due to Proposition 7.4. Then we may write


m2 (ψ) = xj x |ψ(x)|2 dx = ψ|xj x |ψ ,
 
1 1
 2 (ψ) =
m ∂x ψ ∂x ψ dx = ξj ξ |ψ̂(ξ)|2 dξ = ψ̂|ξj ξ |ψ̂ ,
i j i 
Δx (ψ) = tr (m2 (ψ)) ,  2 (ψ)) .
Δξ (ψ) = tr (m
Theorem 7.5 (Heisenberg uncertainty principle). Consider a normalized
ψ(x) ∈ B1 (0) ⊆ L2 such that
m0 (ψ) = 1 , m1 (ψ) = 0 ,  1 (ψ) = 0 .
m
Then there is a lower bound on the product of the variances; namely,
4
(7.6) 1 = m0 (ψ)2 ≤ 4m2 (ψ)m 2 (ψ) = 2 Δx (ψ)Δξ (ψ) .
n
In somewhat more generality, the above inequality follows from the state-
ment that for any ψ ∈ L2 and any 1 ≤ j ≤ n,
 
1 1 2
(7.7) ψ L2 ≤ xj |ψ(x)| dx  ∂xj ψ  dx .
4 2 2
4 i
Proof. We will prove (7.7). Assume that ψ ∈ S so that we may freely take
derivatives and moments. Calculate the commutator
1  1 1 1
∂xj , xj ψ = ∂xj (xj ψ) − xj ∂xj ψ = ψ .
i i i i
Using the commutator, there is an identity
  
1 1  1 1
ψ 2L2 = ψ̄ ∂xj , xj ψ dx = ∂xj ψ xj ψ dx − xj ψ ∂xj ψ dx .
i i i i
Therefore  
1  1
ψ ∂xj , xj ψ dx = 2 im ∂x ψ xj ψ dx ,
i i j
and using the Cauchy–Schwarz inequality, we have
  
     
 ψ̄ 1 ∂x , xj ψ dx ≤ 2  1 ∂x ψ 2 dx 1/2 x2j |ψ(x)|2 dx
1/2
.
i j i j
Therefore we obtain the statement in (7.7) for this ψ,
ψ 2
L2 ≤ 2 ∂xj ψ L2 xj ψ L2 .
To obtain the statement of Theorem 7.5 concerning the variances, square
this inequality and sum over 1 ≤ j ≤ n. The LHS gives n2 ψ 4L2 while the
RHS gives
2
4 Σnj=1 ∂xj ψ L2 xj ψ L2 ≤ 4 ∇ψ 2
L2 xψ 2
L2 .
This is the desired inequality, for ψ ∈ S at least. The general case follows
from the fact that S ⊆ L2 is a dense set. We may thus approach a general

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.3. Phase and group velocities 159

ψ ∈ L2 by a sequence of Schwartz class approximates in such a way that the

LHS of (7.7) converges in the sense of L2 (the RHS may well diverge if, for
example, our target L2 function is not also H 1 ). 

7.3. Phase and group velocities

The majority of the problems that we have addressed so far in this course
have been for constant coefficient PDEs, which is to say that they are of the
form 
P (D)u = pα D α u = 0
|α|≤m
!
n
α !
n
with D = 1i ∂x and D α = Dj j = (−i)α ∂xj . In particular, the evolu-
j=1 j=1
tion equations have been posed as
(7.8) ∂t u = P (D)u ,
u(x, 0) = u0 (x) .
Our examples have been the Schrödinger equation
i i
(7.9) P (D)u = (∂x21 + · · · + ∂x2n ) = − |D|2 ,
2 2
the heat equation
1 1
(7.10) P (D)u = Δ = − |D|2 ,
2 2
and the wave equation, for which we must write (7.8) as a system of PDEs,
    
u 0 I u
(7.11) ∂t = ,
p Δ 0 p
so that  
0 1
P (D) = .
−|D|2 0
The method that we have used in each case is to take the Fourier trans-
form of equation (7.8),
(7.12) ∂t v̂ = P (ξ)v̂,
and solve the resulting ODE which depends upon the parameter ξ ∈ Rn :
(7.13) v̂(ξ, t) = eP (ξ)t v̂0 (ξ).
Modulo questions of convergence, the solution should then be represented
as a Fourier integral

1
(7.14) v(x, t) = √ n eiξ·x eP (ξ)t û(ξ)dξ .
2π

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 160 7. Dispersion

The wave equation gives a good example of this method; in matrix form,
   1 
0 1 P (ξ)t cos(|ξ|t) |ξ| sin(|ξ|t)
P (ξ) = and e = .
−|ξ|2 0 −|ξ| sin(|ξ|t) cos(|ξ|t)
Since the above sequence of calculation can be reproduced for any constant
coefficient operator P (D), one should guess that the Fourier kernel eP (ξ)t of
the solution operator cannot give rise to a well-defined distribution kernel
in all cases. There is, however, a simple criterion in the scalar case for
guaranteeing that we can proceed in this fashion.
Definition 7.6. Suppose that the polynomial iP (ξ) is real valued for all
ξ ∈ Rn . Then P (ξ) is said to be of real type, and
ω(ξ) = iP (ξ)
is known as the dispersion relation.

The idea of a linear evolution equation with constant coefficients can be

extended to more general operators which are defined in terms of Fourier
multipliers which are symbols.
Definition 7.7. A Fourier multiplier is of real type if ω(ξ) is real for all
ξ ∈ R. A Fourier multiplier ω(D) ∈ C∞ (Rn \ {0}) is called a symbol if for
some m, for all multi-indices β,
|∂ξβ ω(ξ)| ≤ Cβ (1 + |ξ|2 )(m−|β|)/2 .
The least value m for which this holds is the order of the symbol.
Proposition 7.8. In the case that ω(D) is a continuous Fourier multiplier
of real type, the initial value problem
(7.15) ∂t u = −iω(D)u
is well-posed in L2 (Rn ).

Proof. Well-posedness will mean in this context that if the initial data
u0 (x) ∈ L2 (Rn ), then there exists a solution u(x, t) of (7.15) with the prop-
erties that for each t ∈ R, u(x, t) ∈ L2 (Rn ), and that limt→0 u(x, t) −
u0 (x) L2 (Rn ) = 0, which is to say that the solution is continuous in L2 . To
show this, we will use the Fourier integral expression for a solution

1
(7.16) u(x, t) = √ n eiξ·x e−iω(ξ)t û0 (ξ) dξ .
2π
Using again the Plancherel identity, u(x, t) L2 (Rn
x)
= e−iω(ξ)t û0 (ξ) L2 (Rnξ ) =
û0 (ξ) L2 (Rnξ ) = u0 L2 (Rnx ) . The continuity in L2 follows arguments that we
have also seen before. 

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.3. Phase and group velocities 161

I want to talk in more detail about the character of solutions of equa-

tion (7.15), and in particular their long time asymptotic behavior, using
an analysis of the Fourier integral representation (7.16) and the dispersion
relation ω(ξ). For this analysis, we ask that ω(ξ) be a symbol of real type.
The Fourier integral represents the solution as a superposition of complex
exponentials
 
1
u(x, t) = ei(ξ·(x−y)−ω(ξ)t) u0 (y) dydξ
(2π)n
 
1 (x−y)
= n
ei(ξ· t −ω(ξ))t u0 (y) dydξ .
(2π)
Definition 7.9 (Kelvin). The phase velocity of the solution given in equa-
tion (7.15) is defined to be
ξ
cp = ω(ξ) 2 .
|ξ|
The group velocity is given by
cg = ∂ξ ω(ξ) .

It is useful to describe the significance of these terms. Consider the

complex exponential for some fixed wave vector k ∈ Rn ,
k
ik·(x−ω(k) t)
u(x, t) = ei(k·x−ω(k)t) = e |k|2 = eiΘ(t,x;k) ,
which is a plane wave solution to (7.15). Inspecting the exponent of this
expression, the phase function Θ(t, x; k) takes a constant value θ on loci
of points in space-time which lie on hyperplanes k · (x − ω(k) k
|k| |k| t) = θ.
k
That is, the points of constant phase are moving in direction |k| with speed
ω(k)
|k| = |cp (k)|. For example, taking the real (or imaginary) part of the solu-
tion, this shows that the velocity of the peak or the trough of an individual
oscillation is given by cp (k).
Now consider a more distributed set of frequencies in the initial data in
the form of a wave packet. This is to say that we take data which decomposes
into two features, a complex exponential multiplying a localized function:
(7.17) u0 (x) = eik·x a0 (x) .
This localized function a0 (x) ∈ S represents an envelope which is slowly
varying with respect to the oscillation of the complex exponential factor.
We will show that this wave packet will propagate with group velocity cg (k).
To illustrate the fact that for a particular equation, the phase and group
velocities are not generally the same, here are several examples.
(1) The Schrödinger equation. ω(ξ) = 12 |ξ|2 , cp = 12 ξ , cg = ∂ξ ω = ξ.
The group velocity is twice the phase velocity.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 162 7. Dispersion

(2) The wave equation. The matrix

 
0 1
P (ξ) =
−|ξ|2 0
has eigenvalues ω(ξ) = ±|ξ|. The phase and group velocities coin-
cide for each of the two eigenvectors of P (ξ):
ξ
cp = ± = cg .
|ξ|
(3) The Airy equation (related to the KdV equation) for x ∈ R1 is
∂t u = ∂x3 u , P (D) = −iD 3 .
The dispersion relation ω(ξ) = ξ 3 has
cp = ξ 2 ≥ 0 , cg = 3ξ 2 = 3cp .

(4) Free surface water waves. The dispersion relation is ω(ξ) = g|ξ|
for surface waves over an infinitely deep body of water. The pa-
rameter g is related to the acceleration of gravity and the density
of water. Then
g ξ 1 g ξ
cp = , cg = ,
|ξ| |ξ| 2 |ξ| |ξ|
from which we see that cg is slower than cp . An astute observer
can see this relationship in the behavior of ripples in disturbances
in a pond.

7.4. Stationary phase

Starting with wave packet initial data (7.17), a solution of equation (7.8) is
expressed in Fourier integrals, giving the familiar expressions

1
(7.18) u(t, x) = √ n ei(ξ·(x−y)−ω(ξ)t) eik·y a0 (y) dydξ
2π

1  
= eik·x √ n ei(ξ ·x−ω(k+ξ )t) aˆ0 (ξ  ) dξ  .
2π
The latter integral follows from the change of variables ξ = k + ξ  . The
following argument is a common one in formal asympotic analysis, and al-
though it has many aspects which are difficult to make rigorous in general,
it provides very useful intuition for the behavior of solutions of dispersive
equations. On a formal level, consider ξ  as a perturbation of the wave
number k. The Taylor expansion of the phase function is
1
ω(k + ξ  ) = ω(k) + ∂ξ ω(k) · ξ  + ξ  , ∂ξ2 ω(k)ξ   + O(|ξ  |3 ) .
2

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.4. Stationary phase 163

In these terms the integral in (7.18) is

u(t, x)

1      3 ))
ei(ξ ·x−∂ξ ω(k)·ξ t) e( 2 ξ ,∂ξ ω(k)ξ +iO(|ξ |
i 2
=e i(k·x−ω(k)t)
√ n aˆ0 (ξ  ) dξ  .
2π
The phase function in this integral is
1
Θ(t, x, ξ; k) = ξ  · [x − ∂ξ ω(k)t] + ξ  , ∂ξ2 ω(k)ξ   + 1O(|ξ  |3 ) .
2
Consider the integral for a point (t, x) in space-time; its contribution to
the solution is negligible for large |k| or for long time |t| if [x − ∂ξ ω(k)t] = 0
(at least it is so under further hypotheses on the integrand). Therefore we
consider only (t, x) which satisfy [x − ∂ξ ω(k)t] = 0. For such space-time
points, the integral representing the solution u(t, x) is reduced to

1  2   3
√ n e 2 ξ ,∂ξ ω(k)ξ +iO(|ξ | ) aˆ0 (ξ  ) dξ 
i
(7.19)
2π
π 2
ei 4 sgn ∂ξ ω(k)
= â0 (0) + o(1) ,
| det ∂ξ2 ω(k)|tn

where sgn ∂ξ2 ω(k) is the signature of the Hessian matrix (the signature of
a symmetric matrix Λ, denoted by sgn Λ, is the number of positive eigen-
values of Λ minus the number of negative ones). Expression (7.19) follows
from contour deformation of ξ into the complex plane; it gives a valid as-
ymptotic formula under a number of further assumptions, including the
hypothesis that det ∂ξ2 ω(k) = 0. This asymptotic calculation is discussed
in Exercise 7.6. If the hypothesis on the nondegeneracy of the Hessian is
not satisfied, the contribution to the asymptotic behavior of the solution
can be more prominent as the resulting asymptotic expression decays less
rapidly; this plays a role in the Kelvin ship wake problem which is the topic
of Project 7.10. In any case, the conclusion is that the principal contribution
to the solution u(t, x) is for the space-time points such that
x
= ∂ξ ω(k) ,
t
which is to say that the wave packet propagates at the group velocity cg (k).
The second part of this section on the method of stationary phase is
dedicated to a more robust analysis of the Fourier integral expression for
the solution

1 x·ξ
(7.20) u(t, x) = √ n eit( t −ω(ξ)) uˆ0 (ξ) dξ .
2π
Consider the data uˆ0 (ξ) supported in Br (ξ0 ), and take any point (x, t) ∈
Rn+1 such that for ξ ∈ supp(uˆ0 (·)), | xt − ∂ξ ω(ξ)| > R. Make use of the

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 164 7. Dispersion

identity
−i( xt − ∂ξ ω(ξ))
x·ξ
−ω(ξ)) x·ξ
eit( t = · ∂ξ (eit( t −ω(ξ)) )
t| t − ∂ξ ω(ξ)|
x 2

in expression (7.20) to rewrite it as


1 −i( xt − ∂ξ ω(ξ)) x·ξ
(7.21) u(x, t) = √ n · ∂ξ (eit( t −ω(ξ)) )uˆ0 (ξ) dξ
2π t| t − ∂ξ ω(ξ)|
x 2

1 x·ξ i( x − ∂ξ ω(ξ))
= √ n eit( t −ω(ξ)) ∂ξ · ( xt uˆ0 (ξ)) dξ .
2π t| t − ∂ξ ω(ξ)|2
The critical fact is that the denominator does not vanish on the support of
the integrand, by the choice of (x, t) ∈ Rn+1 . Measuring the absolute value
of the solution,

1 1  i( x − ∂ξ ω(ξ)) 

|u(x, t)| ≤ √ n ∂ξ · ( xt u
ˆ
2 0
(ξ))  dξ
2π |t| t| t − ∂ ξ ω(ξ)|

C  1 ∂ξ2 ω 

≤  x ∂ξ uˆ0 (ξ) + x u
ˆ0 (ξ)  dξ
|t| | t − ∂ξ ω(ξ)| | t − ∂ξ ω(ξ)| 2

C1
≤ ,
R|t|
for R > 1 at least. Thus the solution is shown to decay in time along the ray
x
t = ∂ξ ω(ξ). Repeating the integration by parts that we performed above,
 1  i( x − ∂ξ ω(ξ)) N 
 x·ξ 
|u(x, t)| =  √ n eit( t −ω(ξ)) (i∂ξ · xt ) u
ˆ0 (ξ) dξ 
2π t| t − ∂ξ ω(ξ)|2
CN
≤ N N
R |t|
for large |t|. The picture of this is as follows.
Theorem 7.10. For initial data u0 (x) ∈ S such that supp(uˆ0 (ξ)) ⊆ Br (ξ0 ),
the solution u(x, t) is also in S for every t. Furthermore, on the set B =
{(x, t) : inf ξ∈Br (ξ0 ) | xt − ∂ξ ω(ξ)| > R}, we have for all N
CN
|u(x, t)| ≤ .
R |t|N
N

This is the statement that away from the rays of stationary phase { xt −
∂ξ ω(ξ) = 0; ξ ∈ supp(uˆ0 (·))}, the solution decays rapidly in space and time.
It may seem as though the above result only applies to quite special
initial functions u0 (x), but in fact it is rather generally applicable. For
example, the regions describing stationary phase asymptotics for two wave
packet solutions are illustrated in Figure 1. Indeed, we may furthermore
decompose any initial data u0 (x) into a sum of wave packets through a
partition of unity.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 7.4. Stationary phase 165

Figure 1. The channels travelled by wave packets with carrier oscil-

lation k and k . The black lines have slopes xt = ∂ξ (k) and ∂ξ (k ),
representing propagation of wave packets at their respective group ve-
locities. The surrounding green conical channels are the regions outside
of which the wave packets decay at rate RNC|t|
N
N .

Lemma 7.11. There exists {χj (y)}j∈Zn , a partition of unity of Rn , with

each χj (y) ∈ C∞ n
0 (R ) such that
3 
supp(χj ) ⊆ j + Q , χj (y) = 1 ,
2 n j∈Z
1
j+ Q,
χj (y) = 1 on
2
where Q is the unit cube {y ∈ Rn ; −1 ≤ yj ≤ 1, j = 1, . . . , n}.

Proof. This can be constructed from a test function χ̃0 (y) ∈ C∞ 0 which
satisfies supp(χ̃0 ) ⊆ 32 Q, χ̃0 ≥ 0, and χ̃0 (y) = 0 for y ∈ Q. Set χ̃j (y) =
χ̃0 (y − j) for all j ∈ Zn ; it is supported in the cube j + 32 Q. We finally
have to adjust the magnitude. But this is easy as we notice that for every
y ∈ m + Q,
 
χ̃m (y) = χ̃k (y) > 0 .
m |k−m|<2

We can simply divide to get

χ̃j (y)
χj (y) =  .
k χ̃k (y)

Notice that χ̃k (y) = 0 for |y −m| < 12 ; hence χm (y) = 1 identically
0<|k−m|<2
on m + 12 Q. 

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 166 7. Dispersion

Theorem 7.12. For initial data u0 ∈ L2 (Rn ), the solution u(x, t) of (7.8)
is in L2 (Rn ). Furthermore the initial data can bedecomposed into a su-
perposition of compactly supported L2 functions j∈Zn (u0 )j where each
(u0 )j (x) ∈ L (j + 2 Q). The resulting solution is a superposition of wave
2 3

packets ujk (x, t) ∈ S, each of which is asymptotically localized by Theo-

rem 7.10 around the ray { xt = ∂ξ ω(k)}.

Proof. The initial data u0 (x) can be decomposed into a sum of compactly
supported L2 functions using a partition of unity from Lemma 7.11:
 
u0 (x) = (u0 )j (x) = (χj u0 )(x) .
j∈Zn j∈Zn

Secondly, we may decompose the Fourier transform (u0 )j , using again a

partition of unity from Lemma 7.11, with the result that (u0 )j χk (ξ) ∈ C0∞ .
Therefore F −1 ((u0 )j χk (ξ)) ∈ S with Fourier support localized in a cube
2 Q + k ⊆ Rξ . Theorem 7.10 applies to each component, which behaves as a
3 n

wave packet, asymptotically being very well localized near rays determined
by the relevant group velocity { xt = ∂ξ ω(j)}. 

Exercises: Chapter 7
Exercise 7.1. The Schrödinger equation with a potential is given as
1
i∂t ψ = − Δψ + V (x)ψ .
2
(i) Show that for a potential V (x) that is bounded, the L2 (Rn )-norm of the
solution is conserved by the equation, namely

d
|ψ(t, x)|2 dx = 0 .
dt Rn
(ii) What is the quantum mechanical flux F in this case?
Exercise 7.2 (Conservation laws). There is an analogy between conserva-
tion laws and the Schrödinger equation, with or without a potential term.
The density function for the position distribution probability is given by
u(t, x) = |ψ(t, x)|2 .
Derive an evolution equation for the function u(t, x) of the form of a con-
servation law, namely
∂t u + ∇ · F = 0 .
What is the function F in terms of the wave function ψ(t, x)? In this equa-
tion, the vector function F plays the role of the quantum mechanical flux.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 7 167

Exercise 7.3. Consider the initial value problem for Schrödinger’s equation
for x ∈ [0, 2π], namely
1
i∂t ψ = − ∂x2 ψ ,
2
ψ(0, x) = ψ0 (x) .

(i) With Dirichlet boundary conditions imposed,

ψ(t, 0) = 0 = ψ(t, 2π) ,
and using the method of separation of variables, give an expression for the
general solution of the initial value problem. What are the eigenvalues λDk
of the Schrödinger operator in this case?
(ii) For periodic boundary conditions
ψ(t, x) = ψ(t, x + 2π) ,
give an expression via separation of variables for the general solution of
the initial value problem. What are the eigenvalues λpk of the Schrödinger
operator in this case?
Exercise 7.4 (Separation of variables). For x ∈ R2 , suppose that the po-
tential V (x) is a square-well, namely
1 1 1
V (x) = 0 , V (x) = −E for − < x < .
for |x| ≥
2 2 2
Writing ψ(x) = v(t)w(x), find all energies λ such that w(x) is bounded in
x ∈ R1 .
If w(x) → 0 as x → ±∞ in such a way that w L2 (R) < +∞, then it is
called a bound state and the associated λ is an eigenvalue of the Schrödinger
operator with this potential. How many eigenvalues and bound states does
this Schrödinger operator have, and how does this number depend upon E?
Express the time evolution of the solution whose initial data is this bound
state.
If w(x) is bounded but is not in L2 (R1 ), then it is called an extended
state. Discuss the set of extended states of the Schrödinger operator with
this potential.
Exercise 7.5. In one space dimension, the Heisenberg uncertainty principle
states that
1    1 
ψ L2 ≤
4
x |ψ(x)| dx
2 2  ∂x ψ(x)2 dx .
4 R1 R1 i
Is this inequality an equality for some function or functions ψ(x)? If so,
which function(s) achieves equality?

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 168 7. Dispersion

Exercise 7.6 (Method of stationary phase). In case â(ζ) is entire of expo-

nential type, reduce the one-dimensional integral
 +∞ 

1 1 ω  2
i ω2 ξ 2 t
√ e â(ξ) dξ = √ ei 2 ζ t â(ζ) dζ ,
2π −∞ 2π Γ
where ω  is real and Γ is a contour in the complex ζ plane consisting of two
rays Γ1 ∪ Γ2 = {reiϕ } ∪ {rei(π+ϕ) }. The angle ϕ = sgn (ω  )π/4 is positive
or negative depending upon the sign of ω  . The sign is chosen so that the
exponential factor is decaying along the rays in question. Furthermore, show
that the latter integral is the sum of two terms of similar form, the first being
  
ei 4 sgn (ω ) +∞ − ρ2  ρeiϕ
π
1 
− ω2 r 2 t
√ e iϕ iϕ
â(re ) e dr =  e 2 â  dρ .
2π Γ1 2π|ω  |t 0 |ω  |t
The second term is similar.
Now show that if â(ζ) is entire of exponential type, for ζ ∈ Cn , the
multidimensional complex Gaussian integral asymptotically takes the form
(7.19) for large time |t|.

Projects: Chapter 7
Project 7.7 (Gaussian wave packets for the Schrödinger equation). Give
an explicit solution to the Schrödinger equation
1
i∂t ψ = − Δψ
2
for all t ∈ R , with initial data given by
1

ψ0 (x) = a(0)eik0 ·x e− 2 x,Ax ,

1

where A = AT > 0 is a Hermitian matrix. What are the moments m1 (ψ0 )

and m1 (ψ(t, ·)), as well as m2 (ψ0 ) and m2 (ψ(t, ·))? From these quantities
you can calculate Δx (ψ0 ) and Δx (ψ(t, ·)). What are Δξ (ψ0 ) and Δξ (ψ(t, ·))?
Project 7.8 (Bounds on moments of the Schrödinger equation). Recall that
we have set
 
mj (ψ) = xj |ψ(x)|2 dx , m
  (ψ) = ξ  |ψ̂(ξ)|2 dξ .
R1 R1
For solutions ψ(t, x) of the Schrödinger equation (7.1) with initial data
ψ0 (x), we have shown that m0 (ψ(t, ·)) = m0 (ψ0 ) = m
 0 (ψ(t, ·)); this is stated
in (7.4).
(i) Show that for j ≥ 1, there do not exist constants Cj such that
|mj (ψ(t, ·))| ≤ Cj |mj (ψ0 )| .
Hint: Use the solutions derived in Project 7.7.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 7 169

(ii) On the other hand, show that for all indices ,

  (ψ(t, ·)) = m
m   (ψ0 ) .

(iii) However, for j = 2p even, show that there is a constant C2p such that

m2p (ψ(t, ·)) ≤ C2p m2p (ψ0 ) + m
 2p (ψ0 ) + m0 (ψ0 ) .

Compare this situation with what is true for t > 0 in the case of higher
moments of the heat equation.
Project 7.9 (Electron orbitals for the hydrogen atom). In Project 2.16 of
Chapter 2 we studied the orbitals of electrons in the quantum mechanical
potential of the hydrogen atom nucleus. In that project we derived the
equations for the components of the orbitals using the method of separation
of variables. In this project we will complete the picture by solving these
equations for the special functions that arise from the method. Recall that
the dynamics of the wave function for an electron in the potential of a point-
like nucleus are given by the Schrödinger equation
1 Z
i∂t ψ = − ∂x2 ψ − ψ.
2 |x|
The coefficient Z is related to the atomic number of the nucleus and the
charge of the electron. The Ansatz for a wave function from the method of
separation of variables is that
ψ(t, x) = e−iλt ϕ(x) = e−iλt R(r)Φ(φ)Θ(θ) ,
where (r, φ, θ) are spherical polar coordinates. The three functions Φ(φ),
Θ(θ), and R(r) satisfy the following three equations. Firstly, the dependence
of the wave function upon the longitudinal angle φ ∈ T1 is given by
d2 Φ
= −m2 Φ , m∈Z,
dφ2
for which the solution is evidently Φ(φ) = eimφ for some m ∈ Z. Secondly,
dependence on the azimuthal angle is given by solutions of the associated
Legendre equation
1 d d m2
(7.22) − sin(θ) Θ + Θ = kΘ .
sin(θ) dθ dθ sin2 (θ)
Finally, the radial dependence of the wave function is given through the
associated Laguerre equation,
1 d 2d Z k
(7.23) r R + λ+ R = 2R .
2
2r dr dr |x| 2r
The goal of this project is to describe the solutions Θ(θ) and R(r) of the
latter two equations.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 170 7. Dispersion

(i) (Associated Legendre functions). Equation (7.22) is defined on the do-

main −π/2 < θ < +π/2.
(a) Change variables by ξ = cos(θ) ∈ [−1, 1] so that (7.22) is written
d d m2
− (1 − ξ 2 ) Θ(ξ) + Θ(ξ) = kΘ(ξ) .
dξ dξ (1 − ξ 2 )
The points ξ = ±1 are singular points of the first kind, for which the second
term of the LHS is a Weyl limit point case. This is the Euler–Lagrange
equation for the Lagrangian functional
  dΘ 2
1 +1 m2
L(Θ) = (1 − ξ 2 ) + Θ2 (ξ) dξ
2 −1 dξ (1 − ξ 2 )
with the additional constraint that I(Θ) is held to a constant, where

1 +1 2
I(Θ) = Θ (ξ) dξ .
2 −1
Namely, the equations can be expressed as
δL = kδI
with Lagrange multiplier k.
(b) Make the substitution Θ(ξ) = (1 − ξ 2 )|m|/2 F (ξ) in the Lagrangian L and
in I, with the result that
  dF 2
1 +1
L= (1 − ξ 2 )|m|+1 + |m|(|m| + 1)(1 − ξ 2 )|m| F 2 dξ ,
2 −1 dξ

1 +1
I= (1 − ξ 2 )|m| F 2 dξ .
2 −1
Now derive the new Euler–Lagrange equations for the functions F (ξ).
(c) Consider power series expressions for F ,

F (ξ) = fj ξ j .
j≤0

In order to satisfy the Euler–Lagrange equations, the coefficients must satisfy

a recursion relation
(j + 2)(j + 1)fj+2 = k − (|m| + j + 1)(|m| + j) fj .
While in general the series cannot be expected to converge, if k = ( + 1)
for some > |m|, the recursion terminates. The resulting polynomial Fm (ξ)
is an associated Legendre polynomial, and we recover that
|m|
Θm
 (θ) = sin (θ)Fm (cos(θ)) .
The Fm (ξ) are the orthogonal polynomials over the interval ξ ∈ (−1, 1) for
the weight function (1 − ξ 2 )|m| .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 7 171

(ii) (Associated Laguerre functions). Equation (7.23) is defined for r ∈ R1+ .

(a) Rescale the independent variable in (7.23) by
ρ = 2βr .
Take k = ( + 1) to follow part (i), defining β as follows and defining γ in
terms of β:
λ 1 Z
2
=− , γ= .
2β 4 β
The resulting equation is now expressed as
1 d  2 dR  1 ( + 1) γ
2
ρ + − − + R=0.
ρ dρ dρ 4 ρ2 ρ

(b) For large ρ, bounded solutions behave as R(ρ) ∼ e−ρ/2 which motivates
the Ansatz
R(ρ) = e−ρ/2 ρ G(ρ) .
The Lagrangian for this problem is
 +∞   +∞ 
1 dR 2 2 1 1 ( + 1) γ
L= ρ dρ + − − + R2 dρ .
0 2 dρ 0 2 4 ρ2 ρ
Substitute the Ansatz into the Lagrangian and derive the Euler–Lagrange
equations δL = 0:
d2 G  2 dG  γ − 2
+ − 1 + − 2 G=0.
dρ2 ρ dρ ρ ρ

(c) As with the case of the Legendre polynomials, substitute a formal power
series for the solution 
G(ρ) = gj ρj .
j≥0
In order to satisfy the Euler–Lagrange equations, the coefficients must satisfy
the recursion relation
[(j + 2)(j + 1) + 2 (j + 1)]gj+2 = (j + 1 + − γ)gj+1 .
The series terminates when n = γ ≥ ( +1) is an integer, giving an associated
Laguerre polynomial Gn (ρ) as a solution.
Following our rescaling of the problem, γ = n; therefore
Z 2β 2 Z2
, λ = λn = −
β= =− 2 .
n 4 2n
2 3
These λn are the L (R ) eigenvalues of the hydrogen atom, while the radial
and angular dependence of the eigenfunctions are determined by the above
products of associated Laguerre functions Rn (r), associated Legendre func-
tions Θm
 (θ), and finally the complex exponentials Φm (φ). The possible

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 172 7. Dispersion

quantum numbers of this eigenstate are determined by the constraints on

the indices described above, namely
+1≤n , |m| ≤ .
The lowest energy electron state λ1 is for n = 1, where only , m = 0 are
permitted. The next energy level is λ2 , which corresponds to excited states
of electrons in the next energy shell. There are three excited states that
correspond to orbitals whose angular dependence is given by the spherical
harmonics Y1m (φ, θ) = Φm (φ)Θm 1 (θ) for m = −1, 0, 1 and whose radial de-
pendence is given by R21 (r), along with a fourth independent eigenstate
R20 (r)Y00 which is purely radial. The well-known fact that the electron
shells contain double these amounts of electrons is due to the extra spin
degree of freedom, which doubles the number of possible electrons in a shell
to 2, 8 · · · . When the spectral parameter λ is positive, this problem has con-
tinuous spectrum and extended states, solutions which are not in L2 (R3 ). It
corresponds to the presence of electrons that are not bound to the nucleus
of the hydrogen atom.
Project 7.10 (Kelvin ship wake problem). This problem concerns the shape
of the waves in the wake of a ship. It is a famous calculation of Thomson
(Lord Kelvin) [11] that explains that when moving in a straight line, the
waves in a ship’s wake form a vee behind the ship of universal opening
angle that is independent of the speed at which the ship is moving. This
calculation is at the heart of the origin of the method of stationary phase.
The equations of water waves. The Euler equations of fluid mechanics are a
system of equations articulated in terms of the velocity u(t, x) of the fluid.
In the case of water waves, it also involves the dynamics of the upper free
surface which defines the moving fluid domain. We consider a fluid which
is infinitely deep, and we use the linear equations of motion, which is a
valid hypothesis for small deformations of the free surface and small fluid
velocities. We therefore assume that the fluid domain is R3− = {(x1 , x2 , y) ∈
R3 , y < 0}. In this domain, the fluid vector field is incompressible and
irrotational, namely
∇·u=0 , ∇×u=0 ,
which implies that there exists a velocity potential function ϕ(t, x, y) defined
in R3− and satisfying
u = ∇ϕ , Δϕ = 0 .
For boundary conditions for ϕ, we assume that ∇ϕ → 0 as y → −∞, while
on the free surface we have
(7.24) ∂t η = ∂y ϕ ,
∂t ϕ = −gη + F .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 7 173

The first equation expresses the fact that η(t, x) moves vertically accord-
ing to the vertical component of the fluid velocity. The second equation
expresses Newton’s law of forces; here g is the acceleration of the restoring
force of gravity, and in the absence of other forces the value of the pressure
on the free surface is taken to be zero. The function F describes any other
additional forces that impinge upon the upper free surface.

(i) Equations (7.24) describe the evolution of the free surface in terms of the
variables η(t, x) and ϕ(t, x, y), for which the only reference to the motion of
the body of the fluid is through the term ∂y ϕ(t, x, 0), where ϕ(t, x, y) is a
harmonic function in the fluid domain. Conveniently, this is the quantity
expressed by the Dirichlet–Neumann operator for the fluid domain, which
in the case of Rn+ is given in formula (4.38). Denoting the surface bound-
ary values of ϕ(x, y) by ξ(x) = ϕ(x, 0), show that the Dirichlet–Neumann
operator for R3− is given by


∂y ϕ(x, y) = |Dx |ξ(x) .
y=0

Therefore the equations of motion (7.24) for the linear water wave problem
are given by
      
η 0 |Dx | η 0
(7.25) ∂t = + .
ξ −g 0 ξ F

(ii) Give a solution to equations (7.24) using Fourier synthesis.

Suggestions: To put ourselves in the spirit of applied mathematics, we will

denote the Fourier transform variable by k = (k1 , k2 ) ∈ R2 , known as the
wavenumber. Take the Fourier transform of equations (7.25) in the horizon-
tal spatial variables,
      
η̂ 0 |k| η̂ 0
(7.26) ∂t ˆ = + .
ξ −g 0 ξˆ F̂

Eigenvalues of the matrix of the RHS are ±iω(k), giving the dispersion
relation, which in 
our case of the water wave equations in deep water is
given by ω(k) = g|k|. This solution can be conveniently expressed in
complex coordinates, writing
"
1 4 g |k| ˆ
ẑ(t, k) = √ η̂ + i 4 ξ .
2 |k| g

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 174 7. Dispersion

In these variables, the problem is reduced to solving the resulting complex

scalar equation for ẑ(t, k),
"
i 4 |k|
(7.27) ∂t ẑ(t, k) = iω(k)ẑ(t, k) + √ F̂ (t, k) .
2 g
When F̂ is nonzero, this can be solved using Duhamel’s principle.
(iii) The perturbation of the free surface due to the presence of a ship is
modeled by a pressure source F . A ship moving at constant speed U is taken
to be the source of a uniformly moving Dirac δ-function, pδ(x − U t). Show
that after Fourier transform, the auxiliary force on the surface is expressed
by
p i(k·U t)
F̂ (t, k) = e .
2π
Then show that the solution of (7.27) is given by
"
ip |k|  eik·U t − eiω(k)t
ẑ(t, k) = eiω(k)t ẑ(0, k) + √ 4 .
2 2π g i(k · U − ω(k))

(iv) The inverse Fourier transform of this expression gives the solution in
complex coordinates in the spatial variables. Transforming to coordinates
in which the ship is stationary, X = x + U t, show that the solution is in the
form

1
z(t, X) = eik·(X−U t) eiω(k)t ẑ(0, k) dk
2π
 "  ik·X
ip 4 |k| e − ei(k·(X−U t)+ω(k)t)
+ √ dk .
4 2π 2 g i(k · U − ω(k))

(v) Instead of solving the initial value problem from t = 0, suppose that
motion started at some asymptotically large past time −T , from which we
assume that all time-dependent transient waves are now negligible. In this
situation the stationary wave pattern of the ship wake is given by
 " 
ip 4 |k| eik·X
z(X) = √ dk .
4 2π 2 g i(k · U − ω(k))
The integrand has a denominator that vanishes along two curves. Assume
that U = (U1 , 0), parametrize these curves in polar coordinates in the Fourier
transform variables as (k1 (α), k2 (α)) = (κ(α) cos(α), κ(α) sin(α)) where
g 1
κ(α) = 2 ,
U1 cos(α)
and use the calculus of residues to express this two-dimensional integral as
a line integral.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 7 175

(vi) The line integrals from part (v) take the form
 +π/2
z(X) = eiΘ(α,X) A(α) dα ,
−π/2
where the phase function is given by
Θ(α, X) = κ(α) cos(α)X1 + κ(α) sin(α)X2 .
In addition, the spatial variable can be represented in polar coordinates,
X = (X1 , X2 ) = (r cos(β), r sin(β)). For fixed X, show that the point of
stationary phase occurs when
tan(α) + (1 + 2 tan2 (α)) tan(β) = 0 .
Observe that this equation has two real roots when
1
| tan(β)| ≤ √ ,
2 2
which are distinct when the inequality is strict. These X = (r cos(β), r sin(β))
are in the region of the wake behind the ship. In the interior of the wake
region, two distinct wavenumbers are dominant at each point X.
(vii) As mentioned in section 7.4, points at which the derivative of the
phase degenerates have the potential to be more prominent features in the
solution. This is indeed the case with the wake of a ship. Show that the
second derivative of the phase function Θ(α, X) with respect to α vanishes
for
(1 + tan2 (α))(1 + 4 tan(α) tan(β)) = 0 .
Therefore both ∂α Θ(α, X) = 0 and ∂α2 Θ(α, X) = 0 when
1
| tan(β)| = √ ∼ 19.47◦ ,
2 2
that is to say, at the outer boundaries of the wake region described above.
Conclusion: The fact that people find remarkable is that this angle of the
most prominent features of a ship wake is independent of the speed |U | of the
ship. In addition, the wave number k(α) √associated with the waves produced
at this angle is such that α = arctan(1/ 2) ∼ 35.26◦ is independent of the
ship speed. What does change is the magnitude |k(α)|, as well as other
details of the wake. A continuation of this project is to discuss a similar
analysis of the wake of a ship which is in motion in a body of water over a
bottom of constant finite depth.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 10.1090/gsm/197/08

Chapter 8

Conservation laws and

shocks

Shock waves in the interstellar medium in the Vela Constellation. Im-

age credit and copyright: Angus Lau, Y. Van, SS Tong (Jade Scope
Observatory).

8.1. First-order quasilinear equations

This section is devoted to the study of nonlinear first-order scalar equa-
tions and their solutions. We will focus on the case of one space and one
time dimension. The first example of an equation of this form is Burger’s

177

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 178 8. Conservation laws and shocks

equation

(8.1) ∂t u + u ∂x u = 0 , x ∈ R1 ,
u(0, x) = h(x) .

Nonlinear equations where the derivatives of the unknown function enter

only linearly, as they do in (8.1) above, are called quasilinear. This equation
is reminiscent of the PDEs that were studied in Chapter 2 of the form

∂t u + c ∂x u = 0 ,

where c represents the speed of propagation. In quasilinear equations such as

(8.1), the speed of propagation depends upon the unknown function u(t, x),
that is to say, on the solution itself.
Equations of the general form of (8.1) are called conservation laws, which
is explained as follows. In physical settings, the function u(t, x) represents
the density of some quantity (such as a gas or a concentration of a component
of a fluid). A second function f = f (u) represents the flux of this quantity,
which is assumed to depend upon the density u itself. Then the law of
conservation of quantity dictates that over any interval [a, b) ⊆ R1 ,
 b
(8.2) ∂t u(t, x) dx + f (u(t, b)) − f (u(t, a)) = 0 .
a

Namely, the total quantity contained in the interval [a, b) changes in time t
only through flux out of the interval through the point x = b and flux into
the interval through the point x = a.
If f (u(t, x)) is differentiable at all x ∈ [a, b), (8.2) can be rewritten as
 b  b
0 = ∂t u(t, x) dx + ∂x f (u) dx
a a
 b
= ∂t u(t, x) + ∂x f (u) dx .
a

Of course if u(t, ·) ∈ C 1 (R1 ), then this holds for any interval [a, b); therefore
u(t, x) satisfies the PDE

∂t u + ∂x f (u) = 0 ,

which is to say

(8.3) ∂t u + f  (u)∂x u = 0 .
1 2
When the flux function is f (u) = 2u , this equation is precisely Burger’s
equation (8.1).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.1. First-order quasilinear equations 179

Figure 1. A solution to (8.5) as a graph (t, x, u(t, x)) ⊆ (t, x, z).

General first-order quasilinear equations. When the initial data is

smooth enough, u(0, x) = h(x) ∈ C 1 (R1 ), one solves these equations locally
in time by first-order methods consisting of integrating along characteristics.
The characteristic equations for Burger’s equation are
dt dx dz
=1, =z , =0.
ds ds ds
These characteristic ODEs can be solved explicitly:
(8.4) s=t, z(s) = z(0) , x(s) = x(0) + sz .

This merits an explanation; we will give an interpretation in geometric

terms of first-order quasilinear equations in general and their associated
characteristic ODEs. The general quasilinear equation when (t, x) ∈ R2 is
(8.5) a(t, x, u)∂t u + b(t, x, u)∂x u = c(t, x, u) .
A solution u(t, x) is presented as a graph (t, x, u(t, x)) in (t, x, z) ⊆ R3 ,
whose normal is
N = (−∂t u, −∂x u, 1) .
The coefficients of equation (8.5) are interpreted as a vector field T =
(a(t, x, z), b(t, x, z), c(t, x, z)) in (t, x, z) ∈ R3 (the characteristic vector field),
and equation (8.5) is the geometrical statement that
N ·T =0 .
The situation is illustrated in Figure 1. This implies that the integral curves
(t(s), x(s), z(s)) (the characteristic curves) of the characteristic vector field
dt dx dz
(8.6) = a(t, x, z) , = b(t, x, z) , = c(t, x, z)
ds ds ds

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 180 8. Conservation laws and shocks

lie on the graph z = u(t, x). A solution to (8.5) is given as the union
of a parameter family of such characteristic curves. This family is locally
specified by a condition that the characteristic curves pass through another
given curve Γ = (T (β), X(β), Z(β)). In the above case of an initial value
problem (8.1), this is the curve Γ = (0, β, h(β)) for β ∈ R1 . To ensure that
the solution surface is well defined, one requires that the tangent vectors
to the curve Γ and the tangent vectors to the characteristic curves (8.6)
at points on Γ are transversal, and that they locally span a tangent space
which itself is a graph over the coordinate plane (t, x). This is the condition
on the curve Γ that it be noncharacteristic. That is,
 
a b
det =0.
∂β T ∂β X
In the case (8.1) and for the initial value problem, this is to say that
 
1 z
det =0,
0 1
a condition which, in the case of Burger’s equation, holds independently of
the initial data z = h(β).
Applying this construction to Burger’s equation (8.4), we have derived
the solution in the form
(8.7) u(t, x) = z(s, β) ,
where t=s, x(s, β) = β + sz , z(s, β) = h(β) .
Describing this solution process in words, the values of the solution u(t, x)
are given by h(β) where (β, h(β)) is the point where the characteristic curve
(t(s, β), x(s, β), z(s, β)) intersects the {t = 0} coordinate plane. Let us as-
sume that the initial data satisfies h ∈ C 1 (R1 ). To present the solution as a
graph, we need to solve the above relations (8.7) for u(t, x) = z(s, β). That
is, one considers the Jacobian of the mapping (s, β) → (t, x), which is given
by
   
t s xs 1 z(0, β)
(8.8) det = det .
t β xβ 0 1 + s∂β z(s, β)
For small times t = s, the Jacobian will be invertible, but the fact is that
for larger times t, the Jacobian determinant has the possibility to vanish.
Indeed, this is the case whenever 1 + t∂β h(β) = 0, in other words when
−1
t= .
∂β h(β)
Therefore, for h ∈ C 1 (R1 ), ∂β h is bounded and there is an interval of time
for which the above construction gives a C 1 solution to (8.1). However, at

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.1. First-order quasilinear equations 181

the first time t for which the Jacobian vanishes, the graph z = u(t, x) will
fold, developing a vertical tangent, and a discontinuity, or shock, will form.

Theorem 8.1. If the initial data h(x) for Burger’s equation is decreasing at
some point x = β, then the solution will become singular at a time T ∗ > 0,
where
−1
T∗ = .
inf β ∂x h(β)

Proof. We will give two proofs of this result. The first follows from the
Jacobian calculation that appears above. Let β0 be a point which minimizes
−1
∂x h(x). As t → T ∗ = − ∂x h(β0 ) , the Jacobian determinant (8.8) tends
to zero as the tangent plane of the graph z = u(t, x) becomes vertical. In the
case where there is no point which achieves the minimum, nonetheless the
Jacobian will tend to zero on a minimizing sequence (tn , xn ) where |xn | →
+∞ and tn → T ∗ as n → +∞, implying that the solution u(t, x) has an
unbounded C 1 -norm in the limit.
A second proof of this theorem on the inevitable formation of shocks
comes from an implicit relation satisfied by the solution. Namely, since
u = z(s, β) = z(0, β) = h(β), t = s, and β = x − sz(s, β),
u(t, x) = h(β) = h(x − tu) .
Taking a derivative of this relation with respect to x,
∂x u(t, x) = h (x − ut)(1 − t∂x u) ,
and solving for ∂x u, we get
h (β)
∂x u(t, x) = .
1 + th (β)
If h (β0 ) < 0 for some β0 , this expression becomes singular for t = −(h (β0 ))−1 .
If β0 is the infimum of h (β), then the first time t > 0 at which this expression
becomes singular is precisely T ∗ . 

The outcome of these considerations is that we are forced to take into

account a sense of a solution of the differential equations which is not nec-
essarily even continuous.

Definition 8.2. The function u(t, x) is a weak solution of Burger’s equation

(8.1) if it satisfies the original conservation law (8.2) for all intervals [a, b) ⊆
R1 .

At points (t, x) at which u ∈ C 1 , the notion of weak solution coincides

with our usual sense of interpretation of the PDE. However, suppose that

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 182 8. Conservation laws and shocks

u(t, x) were discontinuous across a C 1 curve x = γ(t). Then on every interval

[a, b) which contains a point on the curve, a weak solution must satisfy
 γ(t)  b
0 = ∂t u dx + u dx + f (u(t, b)) − f (u(t, a))
a γ(t)
 γ(t)  b
= ∂t u dx + γ̇(t)u− (t, γ(t)) + ∂t u dx − γ̇(t)u+ (t, γ(t))
a γ(t)
+ f (u(t, b)) − f (u(t, a))
 γ(t)
= ∂t u dx + f (u− (t, γ(t))) − f (u(t, a))
a
 b
+ ∂t u dx + f (u(t, b)) − f (u+ (t, γ(t)))
γ(t)
+ γ̇(t)[u− (t, γ(t)) − u+ (t, γ(t))] − [f (u− (t, γ(t))) − f (u+ (t, γ(t)))] .
We have used the notation of the limit from the left limx→γ(t)− u(t, x) =
u− (t, x) and similarly for the right limit. Using the definition of a solution
in the two regions of regularity [a, γ(t)) and (γ(t), b], we obtain
(8.9) 0 = γ̇(t)[u− (t, γ(t)) − u+ (t, γ(t))] − [f (u− (t, γ(t))) − f (u+ (t, γ(t)))] .
This is the jump condition, matching the speed of propagation of the dis-
continuity with the magnitude of the jump, that is to be satisfied in order
to be a weak solution at a discontinuity.
Theorem 8.3. A piecewise C 1 (R2tx ) solution u(t, x) of (8.3) which is dis-
continuous across a curve x = γ(t) ∈ C(R1 ) is a weak solution if:
(1) it satisfies the PDE at a C 1 point, and
(2) across discontinuities it satisfies the jump condition (8.9), namely
(8.10) γ̇[u− − u+ ] = [f (u− ) − f (u+ )] .

One could ask about the variety of scalar conservation laws for different
flux functions f (u). Consider the general case as in (8.3),
∂t u + f  (u)∂x u = 0 .
Suppose that u(t, x) ∈ C 1 in some neighborhood Ω ∈ R2tx , and assume that
f (u) is strictly convex, which is to say that f  (u) is a strictly increasing func-
tion in u. Then g(v) = (f  )(−1) (v) is well defined, and the transformation
u = g(v) is such that in Ω, the function v(t, x) satisfies
∂t u = g  (v)∂t v , f  (u)∂x u = f  (g(v))g  (v)∂x v .
Since f  (g(v)) = v, v(t, x) satisfies
∂t v + v∂x v = 0 ,

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.2. The Riemann problem 183

which is of course Burger’s equation. That is, for hyperbolic conservation

laws with a convex flux function, for smooth solutions, all such equations are
identical up to a change of dependent variable. This is not necessarily the
case for weak solutions, however; Exercise 8.4 discusses the effect of changes
of variables on the jump condition.

8.2. The Riemann problem

Under dilations t = αt, x = αx, the equation
∂t u + c(u)∂x u = 0
remains invariant. The class of solutions that are also invariant under di-
lations is particularly simple, and it plays an important role in the theory
and applications of such conservation laws. In particular, initial data which
is dilation invariant is given by
u(0, x) = h(x) , where h(x) = h− for x < 0 , h(x) = h+ for x ≥ 0 ,
and where h± are two constants. This is called the Riemann problem. For
Burger’s equation
(8.11) ∂t u + u∂x u = 0 ,
solutions are simple and explicit.

Case (1). Given h− = 0 and h+ = 1, for t > 0 a solution is

u(t, x) = 0 for x ≤ 0 ,
x
u(t, x) = for 0 ≤ x ≤ t ,
t
u(t, x) = 1 for x ≥ 0 .
Such solutions are rarefaction waves, as they propagate a decrease in the
density of the quantity u. A solution with a rarefaction wave is illustrated
by its characteristic curves in Figure 2. Notice that the solution u(t, x) is
not C 1 (R2 ) for t > 0 but u ∈ Lip(R2 ).

Case (2). Given h− = 1 and h+ = 0, for t > 0 a solution is necessarily

discontinuous; indeed,
t
,
u(t, x) = 1 for x <
2
t
u(t, x) = 0 for x ≥ .
2
A solution with a shock wave is illustrated in Figure 3. The line of disconti-
nuity given by x = 2t is a shock curve. Its speed is determined by the jump

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 184 8. Conservation laws and shocks

x
h = -1 h = +1

Figure 2. A rarefaction wave for Burger’s equation (8.11).

condition (8.10); indeed, we check this:

u− = 1 , u+ = 0 ,
1 1 1 1
u− − u+ = 1 , f (u− ) − f (u+ ) = u2− − u2+ = ; thus γ̇(t) = .
2 2 2 2
Case (3). There is another case to consider. Namely, starting with the
initial data of Case (1), h− = 0 and h+ = 1, there is a discontinuous
solution that satisfies all of the conditions that we have so far imposed on
solutions. Namely,
t
,
u(t, x) = 0 for x<
2
t
u(t, x) = 1 for x ≥ .
2
Checking the jump condition (8.10), we have
u− = 0 , u+ = 1 ,
1 1 1 1
u− − u+ = −1 , f (u− ) − f (u+ ) = u2− − u2+ = − ; thus γ̇(t) = .
2 2 2 2
This is a primary example of nonuniqueness of weak solutions. In relaxing
the definition of a solution in order to accommodate discontinuities, we
have increased the class of allowable solutions, and in doing so have lost
the property of uniqueness of solutions for given initial data. In order to
regain it, one must make a selection of the appropriate solution by imposing
a further principle, in addition to the definition of a weak solution through
the law of conservation.
Definition 8.4 (Entropy condition). If u(t, x) is a weak solution of (8.3)
which is piecewise C 1 , and x = γ(t) is a curve along which it has a jump

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.2. The Riemann problem 185

Figure 3. A shock wave for Burger’s equation (8.11).

Figure 4. A shock wave for Burger’s equation (8.11) that does not
satisfy the entropy condition.

discontinuity, we define u(t, x) to be an entropy solution if

(8.12) f  (u− ) := c(u− ) > γ̇ and f  (u+ ) := c(u+ ) < γ̇ .

A solution that does not satisfy the entropy condition is illustrated in

Figure 4. This condition brings back uniqueness to the initial value problem,
as we will see in the following section. It is important to understand that
this is an additional condition that has been added to the evolution problem
for (8.3), above and beyond the concept of a weak solution to a conservation
law. In particular, this choice of acceptable solution has introduced a dis-
tinction between future and past, a so-called arrow of time. It means that

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 186 8. Conservation laws and shocks

as time increases towards the future, characteristic curves which neighbor a

shock curve must run into the shock. The solution of the Riemann problem
for Burger’s equation in Case (2) satisfies the entropy condition, while the
solution of the Riemann in Case (3) does not satisfy it, and characteristics
can and do originate on the shock curve and propagate away from it. In
terms of entropy, for a shock in a solution that satisfies the entropy condi-
tion, information carried along characteristics of a solution will pass into the
shock and disappear. However, in Case (3), information could be generated
in the shock curve itself and propagate into the future; indeed, another pos-
sible weak solution for the initial data given in Case (3) is a modification of
the shock curve γ(t) = t/2, for which at each second after t > 0 (and for
some small value of β), the shock curve satisfies γ̇ = 1/2 ± β. Choosing the
sign of ±β, each subsequent second gives a coding in binary, and in principle
we may encode all of the Oxford English Dictionary in a weak solution of
Burger’s equation, which itself is one of an infinite number of nonunique so-
lutions. Thus it makes sense to impose the entropy condition as a selection
principle for weak solutions.

8.3. Lax–Olenik solutions

So far, for scalar hyperbolic conservation laws we have constructed solutions
locally in time using the method of characteristics, and for special Riemann
initial data we have constructed quite explicit solutions that exist globally
in time and that exhibit shocks and rarefaction waves. There is a beautiful
extension of these considerations that gives global weak solutions for general
L1 (R1 ) initial data, solutions which satisfy the entropy condition of Defini-
tion 8.4. Furthermore, uniqueness holds for this class of solutions. A nice
reference for this material is by Lax [14].
Consider a general scalar conservation law for f ∈ C 2

∂t u + ∂x f (u) = 0 ,

which we rewrite by the chain rule as

(8.13) ∂t u + a(u)∂x u = 0 ,

where evidently a(u) = ∂u f (u). Without loss of generality, we may assume

that f (0) = 0. We take up the study of the initial value problem for initial
data h(x) = u(0, x) ∈ L1 (R1 ). The principal hypothesis is that the flux
function f (u) is strictly convex; this assures that the coefficient a(u) is
monotone increasing in u. This hypothesis is satisfied for example in the
case of Burger’s equation, where f (u) = 12 u2 .

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.3. Lax–Olenik solutions 187

f3

f
2

f’ (v) (u - v)
f1
u-v
u
v u

Figure 5. Property of convexity, as described in inequality (8.14).

Convexity and convex duality. The properties of f (u) and a(u) are used
in several ways. Firstly, the following inequality holds for arbitrary u and v:
(8.14) f (v) + a(v)(u − v) ≤ f (u) .
This inequality is a consequence of convexity, as illustrated in Figure 5.
Because of strict convexity, equality holds only for u = v. Being convex,
f (v) has a convex dual defined by
g(z) := a(v)v − f (v) , where z = a(v) .
The dual has the alternative definition
g(z) = min zv − f (v) ,
v

where the minimum is taken on when ∂v zv − f (v) = 0, namely for z −

∂u f (v) = 0. Figure 6 illustrates this relationship. Since a(v) is increasing in
v, we may define its inverse function, setting v = b(z) whenever a(v) = z.
Incidentally, when v = b(z), z is the minimizer of the function
f (v) = min zv − g(v) ,
z

namely v = ∂z g(z) = b(z). It follows from a short calculation that g(z) is

convex.
This is similar in the setting of a Lagrangian L(u̇, u) to the Legendre
transform, which is globally defined when L is convex in the variables u̇.
Examples of convex functions and their convex duals are given by
A 2 1 2
f (v) = v , g(z) = z ,
2 2A

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 188 8. Conservation laws and shocks

f(v)

u
v

g (z)

Figure 6. Convex duality.

and
1 1  1 1
f (v) = v p , g(z) =  z p , where +  =1,
p p p p
all of which are straightforward to check.

Hamilton–Jacobi equations. It is useful to change the problem to the

setting of a Hamilton–Jacobi equation, namely to study
 x
U (t, x) = u(t, x ) dx ,
−∞
so that ∂x U = u. Then U satisfies
(8.15) ∂t U + f (∂x U ) = 0 .
Initial data is given by
 x
H(x) = h(x ) dx ,
−∞

which is Lipschitz continuous since we have asked that h ∈ L1 . Furthermore,

for almost all x1 , we have1
H(x2 ) − H(x1 ) = h(x1 )(x2 − x1 ) + o(|x2 − x1 |) .
Whenever u(t, x) ∈ C 1 (R1 ), U (t, x) is a classical solution to (8.15). By
convexity, given a solution U of (8.15), for any v, (8.14) implies that
(8.16) ∂t U + ∂u f (v)∂x U = ∂u f (v)∂x U − f (∂x U ) ≤ ∂u f (v)v − f (v) .
Because of strict convexity, equality holds only for v = ∂x U .
1 We use big O(·) and little o(·) notation.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.3. Lax–Olenik solutions 189

For purposes of constructing a solution to equation (8.15), given a space-

time point (t, x) with t > 0, consider the line in space-time through it with
slope a(v) = ∂u f (v); the intersection of this line with the x-axis defines y
such that
x−y
= a(v) = z .
t
The LHS of inequality (8.16) has the interpretation as the directional deriv-
ative of the function U (t, x) along this line, while the RHS depends only on
the parameter v. Integrating (8.16) along this line gives that
(8.17) U (t, x) ≤ U (0, y) + t a(v)v − f (v) = H(y) + t a(v)v − f (v) .
Interpret this inquality in terms of convex duality; take z = a(v) and g(z) =
a(v)v − f (v) as above. That is, b(z) = v, the inverse function, so that when
x−y x−y
z= = a(v) , then b(z) = b =v .
t t
This gives a reinterpretation of inequality (8.17), articulated in the following
proposition.
Proposition 8.5. If u(t, x) is a solution of (8.13) which is Lipschitz at
x ∈ R1 at time t > 0, then
x−y
(8.18) u(t, x) = b = h(y(t, x)) ,
t
where y = y(t, x) minimizes the relation involving the initial data H(y):
x−y
(8.19) H(y) + tg := G(x, y; t) .
t
The ray (x − y(t, x))/t is the backwards characteristic between (t, x) and
(0, y(t, x)) over the time interval [0, t].

Lax–Olenik weak solutions. Now suppose that u(t, x) is a weak solution

with possible jump discontinuities. Then nonetheless
 x
U (t, x) = u(t, x ) dx ,
−∞
which is Lipschitz (as long as u(·, x) ∈ L1 ), and the inequality still holds
that
∂t U + ∂u f (v)∂x U ≤ ∂u f (v)v − f (v) .
If u(t, x) is a weak solution that in addition satisfies the entropy condition
(8.12) at each discontinuity,
dγ
u+ ≤ ≤ u− ,
dt
then every point (t, x) ∈ R+t × Rx is connected to the initial data at t = 0 by
at least one ray (x − y)/t = a(v) since backward characteristics do not inter-
sect shock curves. For each minimizing value of y, equality holds in (8.16).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 190 8. Conservation laws and shocks

The elegant aspect of the theory is that the converse holds. Namely, the
variational formula given by (8.18) and (8.19) defines a weak solution to
(8.13) that in addition satisfies the entropy condition at every discontinuity.
Theorem 8.6. Formulae (8.18) and (8.19) define a possibly discontinuous
function u(t, x) for arbitrary initial data h(x) ∈ L1 (R1 ), which is a weak so-
lution with at most countably many jump discontinuities, and which satisfies
the entropy condition at such discontinuities.

The function u(t, x) defined by (8.18) and (8.19) is called the Lax–Olenik
solution.

Proof. First of all, because h(x) ∈ L1 , H(x) is bounded, and the convex
dual g(z) = a(v)v − f (v) for v = b(z) taken over the range of H(y) has
its minimum at some z = c where ∂z g(z) = 0. Therefore the function
G(x, y; t) defined in (8.19) is minimized and takes on its minimum at some
finite y. This minimizer is possibly nonunique. If y is the unique minimizer
of G(x, y; t), then ∂y G(x, y; t) = 0, at least for a.e. y at which H  exists,
which implies that
x−y x−y
0 = H  (y) − g  = h(y) − b .
t t
In fact this y = y(t, x) is the minimizer for all points on the ray between
(0, y) and (t, x); indeed, for 0 < t1 < t and for x1 such that
x1 − y x−y
= ,
t1 t
∂y G(x1 , y; t1 ) = 0 so that h(y) = b( x1t−y
1
) as well. Thus the PDE (8.13) is
satisfied along this ray, and u(t, x) = h(y) = u(t1 , x1 ).
However, it could be that for given (t, x), G(x, y; t) has more than one
minimizer y(t, x); this situation is illustrated in Figure 8. In this case, the
following property of y(t, x) is useful.
Lemma 8.7. Fix t > 0 and denote y = y(t, x) the minimizers of G(x, y; t).
Then y(t, x) is a nondecreasing function of x.

Proof of Lemma 8.7. The result will follow from a fact that we shall prove
geometrically. Take x2 > x1 , then G(x2 , y1 ) < G(x2 , y) for all y < y1 . This
inequality depends upon the convexity of the function g:
x − y x − y x − y x − y
2 1 1 1 1 2 1
(8.20) g +g <g +g .
t t t t
The intervals [(x2 − y1 )/t, (x1 − y)/t] and [(x1 − y1 )/t, (x2 − y)/t] share
the same midpoint, and by convexity, the sketch in Figure 7 proves the
inequality.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.3. Lax–Olenik solutions 191

u
x1-y1 x1-y x2-y1 x2-y

Figure 7. Geometric proof of inequality (8.20).

Recall that G(x, y; t) = H(y) + tg((x − y)/t); multiply inequality (8.20)

by t and respectively add G(x1 , y1 ; t) ≤ G(x1 , y; t) to each side, obtaining
 x −y x1 − y
2 1
t g +g + G(x1 , y1 ; t)
t t
 x −y x2 − y
1 1
≤t g +g + G(x1 , y; t) .
t t
Using definition (8.19) for G(x, y; t), this is the statement that
x2 − y1 x2 − y
tg + H(y1 ) ≤ tg + H(y) ,
t t
which is to say that for y < y1 ,
G(x2 , y1 ; t) ≤ G(x2 , y; t) .
Thus if y1 is a minimizer for x = x1 , then any minimizer for x2 > x1 must
satisfy y2 > y1 . 

The first conclusion we draw from this monotonicity is that for fixed
t > 0, the mapping x → y(t, x) is increasing in x with only countably many
jump discontinuities. Suppose that G(x, y; t) has more than one minimizer,
and let y − be the least minimizer of G(x, y; t) and y + > y − the greatest.
Because of the property of monotonicity, for fixed t > 0 if x1 < x, then any
y1 that is a minimizer of G(x1 , y; t) must satisfy y1 < y − , while if x1 > x,
then y1 > y + . Therefore the interval [y − , y + ] is a jump discontinuity of
y(t, x), and it follows that there can only be countably many points x for
which the minimizer y = y(t, x) is not unique.
Now consider t1 > t; for all x1 ∈ R1 , a minimizer y(t1 , x1 ) either sat-
isfies y(t1 , x1 ) < y − (t, x) or else y + (t, x) < y(t1 , x1 ). This is because for
any y − (t, x) < y < y + (t, x), the ray from (t1 , x1 ) to (0, y) will intersect one

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 192 8. Conservation laws and shocks

(t,x1) (t,x2)

x
y-(t,x1) y+(t,x1) y(t,x2)

Figure 8. The characteristics in cases of nonunique (t, x1 ) and unique

(t, x2 ) minimizers of the Lax–Olenik variational principle.

of the rays (t, x) to (0, y ± (t, x)) and therefore cannot have the minimizing
slope of the latter. Hence when t1 > t, for each interval [y − (t, x), y + (t, x)]
of jump discontinuity of y(t, ·), there is a point x1 with multiple minimizers
of G(x1 , y; t1 ) and a corresponding interval [y − (t1 , x1 ), y + (t1 , x1 )] represent-
ing a jump in y(t1 , x). In particular, if {xk } and {xj1 } are the points of
discontinuity of y(t, x) and y(t1 , x) at times t and t1 > t, respectively, then
# #
[y − (t, xk ), y + (t, xk )] ⊆ [y − (t1 , xj1 ), y + (t1 , xj1 )] .
k j

The union of these points with nonunique minimizers form a set of shock
curves, which by the above property may merge and come into existence,
but may not split as t increases. The presence of such space-time points
for which G(x, y; t) has nonunique minimizers implies that a solution given
through the Lax–Olenik variational principle is not a classical solution.
We now show that it is a weak solution in the sense that it satisfies the
jump condition (8.9) at points on shock curves.
Consider t1 > t and points (t, x) and (t1 , x1 ) on a shock curve, where
each has nonunique minimizers y ± and y1± , respectively, and such that
[y − (t, x), y + (t, x)] ⊆ [y1− (t1 , x1 ), y1+ (t1 , x1 )] .
Since they are nonunique minimizers, from the Lax–Olenik variational prin-
ciple we know that
(8.21) G(x, y − ; t) = G(x, y + ; t) , G(x1 , y1− ; t1 ) = G(x1 , y1+ ; t1 ) ,

where of course G(x, y; t) = H(y) + tg( x−y

t ). Interpreting (8.21) in terms
± ± ± x1 −y1±
of f (v), and denoting v ± = b(z ± ) = b( x−y
t ), v1 = b(z1 ) = b( t1 ),

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 8.3. Lax–Olenik solutions 193

equalities (8.21) become

(8.22) t[f (v + ) − f (v − )] − [v + (x − y + ) − v − (x − y − )] = H(y + ) − H(y − )
and
(8.23) t1 [f (v1+ ) − f (v1− )] − [v1+ (x1 − y1+ ) − v1− (x1 − y1− )] = H(y1+ ) − H(y1− ) .
Taking the difference of (8.23) and (8.22) and using that
H(y1± ) − H(y ± ) = h(y ± )(y1± − y ± ) + o(y1± − y ± ) ,
we find
(8.24) (t1 − t)[f (v + ) − f (v − )] − (x1 − x)[v + − v − ]
= −t1 [f (v1+ ) − f (v + ) − f (v1− ) + f (v − )]
+ [(v1+ − v + )(x1 − y + ) − (v1− − v − )(x1 − y − )] + o(y1± − y ± ) .
Now
f (v1± ) − f (v ± ) = a(v ± )(v1± − v ± ) + o(v1± − v ± ) ;
therefore the RHS of (8.24) is given by
− t1 [a(v + )(v1+ − v + ) − a(v − )(v1− − v − )]
− [(v1+ − v + )(x1 − y + ) − (v1− − v − )(x1 − y − )]
+ o(v1± − v ± ) + o(y1± − y ± )
= −t1 [a(v + )(v1+ − v + ) − a(v − )(v1− − v − )
x − y+ x − y−
− (v1+ − v + ) + (v1− − v − ) ]
t t
+ o(v1± − v ± ) + o(y1± − y ± ) + o(x1 − x) + o(t1 − t) .
±
Observing that a(v ± ) = x−yt and that o(v1± − v ± ) and o(y1± − y ± ) are
o(x1 −x)+o(t1 −t), we have shown that the LHS is o(x1 −x)+o(t1 −t). This
is tantamount to the jump condition (8.9), and hence the function u(t, x)
defined through the Lax–Olenik variational principle ((8.18) and (8.19)) is
indeed a weak solution of (8.13).
It remains to verify that the solution given by (8.18) and (8.19) in Propo-
sition 8.5 satisfies the entropy condition for the conservation law (8.13). Use
again that f ∈ C 2 is convex so b(z) ∈ C 1 is increasing, and that y = y(t, x)
is increasing in x. Therefore for x1 < x2 ,
x1 − y1 x1 − y2
(8.25) u(t, x1 ) = b ≥b
t t
x2 − y2 x2 − x1
≥b − Cb
t t
x2 − x1
= u(t, x2 ) − Cb ,
t

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 194 8. Conservation laws and shocks

where one uses in the second line that b(z) has Lipschitz constant Cb :
b(z2 ) − b(z1 ) ≤ Cb (z2 − z1 ) for z1 ≤ z2 .
Therefore approaching a point x of discontinuity of u(t, x) = b((x − y)/t)
from the left by x1 and from the right by x2 , our solution satisfies
u− (t, x) ≥ u+ (t, x) ,
which is the entropy condition for a jump discontinuity. 

Uniqueness. The solution satisfying the entropy condition constructed by

the Lax–Olenik variational principle is unique, as we show with an argument
involving contractions in L1 . Consider two initial conditions h1 (x), h2 (x) ∈
L1 and their resulting Lax–Olenik solutions u1 (t, x) and u2 (t, x). We will
show that the difference in L1 -norm of the two solutions is a decreasing
function of time. The L1 -norm of their difference is
 +∞
u1 (t, ·) − u2 (t, ·) L1 = |u1 (t, x) − u2 (t, x)| dx
−∞
  yn+1 (t)
= (−1) n
(u1 (t, x) − u2 (t, x)) dx ,
n yn (t)

where yn (t) are the points of sign changes of the difference u1 (t, x)−u2 (t, x),
with sgn(u1 (t, x) − u2 (t, x)) = (−1)n over the interval (yn (t), yn+1 (t)). We
will analyze the behavior of the individual integrals of this sum.
Case (1). Suppose that yn (t) is a point of continuity for both u1 (t, x) and
u2 (t, x). In this case, both these solutions are constant along characteristics,
which in turn are linear. Therefore yn (t) is linear in t, and u1 (t, yn (t)) −
u2 (t, yn (t)) = 0.
Case (2). When yn (t) is a point of discontinuity of either u1 (t, x) or u2 (t, x),
or both, yn (t) is a shock curve.
Using the nature of conservation laws and their weak solutions, the time
derivative of the difference satisfies
 
d d yn+1 (t)
u1 (t, x) − u2 (t, x) L1 = (−1)n (u1 (t, x) − u2 (t, x)) dx
dt n
dt yn (t)
  yn+1 (t)
+ (−1)n ∂t (u1 (t, x) − u2 (t, x)) dx
n yn (t)
  dyn+1
+ (−1)n (u1 (t, yn+1 (t)) − u2 (t, yn+1 (t)))
n
dt
dyn (t) 
− u1 (t, yn (t)) − u2 (t, yn (t)) .
dt

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Exercises: Chapter 8 195

Each term of the sum after integration takes the form


d yn+1 (t)
(−1)n (u1 (t, x) − u2 (t, x)) dx
dt yn (t)
 yn+1 (t) dy yn+1 (t)

= (−1)n (f (u1 (t, ·))−f (u2 (t, ·))) + u1 (t, y)−u2 (t, y)  .
y=yn (t) dt y=yn (t)
For endpoints yn (t) in Case (1), u1 (t, yn ) = u2 (t, yn ) and there is no contri-
bution to the sum. For endpoints in Case (2), there are subcases of which
we will do one example. Suppose that it is u1 (x, t) that has an entropy
−
condition satisfying jump at y = yn+1 (t), at which u+ 1 ≤ u2 ≤ u1 , while
u2 is continuous. Then over the left-hand interval (yn , yn+1 ), we know that
u1 (t, x) − u2 (t, x) > 0, so n is even. The jump condition at yn+1 (t) is that
dy f (u−
1 ) − f (u1 )
+
= ,
dt u−
1 − u1
+

and by the convexity of f (u), we have

 f (u− ) − f (u+ )
f (u2 ) − f (u− −
1 ) + (u1 − u2 )
1 1
u−
1 − u1
+

 u − u+ u−
2 − 1 − u2
≤ f (u2 ) − −
1
+ f (u1 ) + − + f (u1 ) ≤ 0 .
+
u1 − u1 u1 − u1
Hence this boundary term gives a negative contribution:
dy 
(−1)n (u1 − u2 )(y)  ≤0.
dt y=yn+1 (t)
Inspecting the other cases, they also only give negative contributions. Hence
the L1 -norm of the difference is decreasing. In the case that the initial data
satisfies h1 = h2 , this implies the uniqueness of entropy condition-satisfying
solutions and thus uniqueness for the class of Lax–Olenik solutions.

Exercises: Chapter 8
Exercise 8.1. Suppose that f (x, u) = b(x)u is the flux function for a linear
scalar conservation law, and consider a solution u(t, x) ∈ C 1 (R2 ):
∂t u + ∂x (b(x)u) = 0 .
Suppose that x = X1 (t) and x = X2 (t) are two characteristic curves such
that X1 (t) < X2 (t). Show that for all t ∈ R1 ,
 X2 (t)  X2 (0)
u(t, x) dx = u(0, x) dx .
X1 (t) X1 (0)

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 196 8. Conservation laws and shocks

Exercise 8.2. Solve the linear scalar conservation law with flux function
f (x, u) = x3 u:

∂t u + ∂x (x3 u) = 0 , u(0, x) = h(x) .

Is the solution unique? Does h(x) ∈ C ∞ imply that u(t, x) ∈ C ∞ ?

Exercise 8.3. (i) Give the general (entropy condition-satisfying) solution

to Burger’s equation with the Riemann data

u(0, x) = h(x) ;
h(x) = h− , x<0, h(x) = h+ , x≥0.

(ii) Work out the more complicated but still explicit (entropy condition-
satisfying) solution to the family of problems with piecewise constant initial
data u(0, x) = h(x), where

h(x) = h− , x < −1 , h(x) = h0 , −1 ≤ x < 1 , h(x) = h+ , x ≥ 1 .

Note: These solutions are global in (t, x) ∈ R1+ × R1 , and they involve shock
interactions and interactions between shocks and rarefaction waves. There
are eight cases depending on the relative inequalities between h− , h0 , h+ .

Exercise 8.4. On a domain Ω ⊆ R2 in which u(t, x) ∈ C 1 solves a general

scalar conservation law with convex flux f (u), we showed above that (8.3)
is equivalent to Burger’s equation, up to transformation of dependent vari-
ables. Suppose now that u(t, x) is only a piecewise C 1 weak solution of this
conservation law with C 1 shock curves. Under transformation, does it sat-
isfy the jump condition (8.9) for Burger’s equation? If not, what condition
does it satisfy?

Projects: Chapter 8
Project 8.5 (The Euler equations and shear flows in fluid dynamics). The
motion of a fluid is often represented by its velocity field

u(t, x) = (u1 (t, x), . . . , ud (t, x)) ,

where x ∈ Ω ⊆ Rd , the domain in which the fluid is confined. For ob-

vious physical reasons, one usually considers d = 2 or 3. Fluid particles
which move according to this velocity field are described by their coordi-
nates X(t, x) which satisfy
dXj
= uj (t, X(t, x)) , Xj (0, x) = xj .
dt

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 8 197

The acceleration experienced by this fluid particle is given by

d2 Xj  d
dXj
2
= ∂t uj (t, X(t, x)) + ∂x uj (t, X(t, x))
dt dt
=1
= ∂t u + (u · ∇)u j
.
An incompressible fluid satisfies
∇·u=0 ,
and it is the most common setting to take the density ρ = 1 constant.
Newton’s laws state that this quantity is proportional to the forces acting
on the fluid particle, ρ(∂t u + (u · ∇)u) = F . For the simplest incompressible
flows with no viscous forces (inviscid flows), an intrinsic component of the
force F is given by the pressure
F (t, x) = −∇p(t, x) .
When no other forces are acting on the fluid, the equations of motion are
called the Euler equations
(8.26) ∂t u + (u · ∇)u = −∇p ;
this is a system of nonlinear equations that involves the components of the
d-dimensional vector field u(t, x) and the pressure p(t, x) which is a scalar
function. On the boundary ∂Ω of the fluid domain, the boundary conditions
are that u · N = 0 where N is the normal to the boundary. The vorticity of
this fluid vector field is given by
ω(t, x) := ∇ × u(t, x) .

(i) Shear flows are special solutions of the Euler equations in the domain
Ω = Rd+ . Considering the case d = 2, a shear flow is a velocity vector field
given by
u(t, x1 , x2 ) = (u1 (x2 ), 0) ,
where the function u1 (x2 ) is the shear profile. Remark that this flow is
time independent, so that if initially u(t = 0, x) is Hölder continuous with
exponent 0 ≤ α ≤ 1,
|u(t, x) − u(t, x )| ≤ C|x − x |α ∀x, x ∈ R2+ ,
then it remains so for all time t ∈ R1 .
(a) Show that this vector field is incompressible and that it satisfies the
Euler equations (8.26).
(b) Calculate the vorticity of the velocity vector field.
(c) Show that the velocity vector field satisfies the boundary conditions on
{x2 = 0}.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 198 8. Conservation laws and shocks

(ii) Now consider the case of shear flows in three-dimensional settings. In

the case that d = 3, a shear flow takes the form
u(t, x) = (u1 (t, x2 , x3 ), u2 (x3 ), 0) .
(a) Show that ∇ · u = 0.

(b) Show that if u(t, x) satisfies the Euler equations, then

∂t u1 + u2 ∂x2 u1 = 0 .
Therefore u1 (t, x) satisfies a transport equation whose speed is given by
u2 (x3 ). Show that the solution for u1 (t, x) takes the form
u1 (t, x) = u1 (t, x2 , x3 ) = u1 (x2 + tu2 (x3 )) .
Show that the velocity vector field satisfies the boundary conditions on {x3 =
0}. Calculate the vorticity of the flow.
(c) (Hölder modulus of continuity). Suppose that v(y) is a Hölder continuous
function of y with Hölder exponent β < 1,
|v(y) − v(y  )| ≤ Cβ |y − y  |β ,
and it is composed with a function u(v) which is Hölder continuous with
exponent γ < 1. Then show that the result is in general only Hölder con-
tinuous with exponent α = βγ < min{β, γ}.
(d) (Loss of Hölder continuity by solutions of the Euler equations). This
question pursues the fact that in three-dimensional flows, the solution at
later time t = 0 can be less smooth than the initial data for the solution.
This phenomenon was originally described in Lions [15] and Bardos and
Titi [2]. Show that if the initial conditions u(0, x) for a three-dimensional
shear flow are such that u1 (0, x2 , x3 ) is γ-Hölder continuous and u2 (x3 ) is
β-Hölder continuous, then the solution u(t, x) for t = 0 need not be any
smoother than α = βγ < min{β, γ}.
Project 8.6 (Moduli of continuity of the solution map of Burger’s equa-
tion). This project studies several aspects of the dependence of solutions of
evolution equations on their initial data. The concept of a well-posed prob-
lem is due to Hadamard, who articulated that an initial value problem is
well posed on a class of functions C if (i) solutions u(t) exist for data u0 ∈ C,
(ii) the solution u(t) is uniquely determined by u0 , and (iii) the solution
map ϕt : u0 → u(t) is continuous in the topology given for C.
(i) Consider the solution map for Burger’s equation
∂t u + u∂x u = 0 , x ∈ R1 .
Show that the solution map ϕt (u0 ) is not continuous for u0 in the space
Cpiecewise of piecewise continuous functions on R1 . To show this, you might

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Projects: Chapter 8 199

start with a solution that forms a shock, and consider the behavior of nearby
initial data and the resulting solutions. If instead the function class one con-
siders is that of continuous functions with piecewise continuous derivatives,
is the solution map continuous in this case?
(ii) Show that the solution map ϕt (u0 ) for Burger’s equation is continuous,
locally in time, when C is the Sobolev space H s (R1 ) for s > 1. That is, for
u0 ∈ K ⊆ H s (R1 ) a bounded open set, there exists a time T = T (K) > 0
such that the initial value problem satisfies the well posedness criterion of
Hadamard up to times |t| < T .
(iii) However, the solution map on the class H s (R1 ) is not smooth, nor even
Hölder continuous; this is a theorem of Kato [10]. His counterexamples to
any possible Hölder modulus of continuity are constructed as follows.
(a) Define a family of initial data for Burger’s equation to be

uλ0 (x) = (λ + xα+1

+ )ϕ(x) ,

where ϕ(x) is a C0∞ cutoff function such that ϕ(x) ≡ 1 for −1 < x < +1.
The notation is that

x+ = 0 (x < 0) , x+ = x (x > 0) ,

and the exponent α > 0 will be taken large, according to the Sobolev regular-
ity exponent s. Show that for α > s−1/2, both uλ0 (x) and u00 (x) are H s (R1 ).
(b) The characteristics for this problem for λ = 0 and λ > 0 are, respectively,

x = y 0 + u00 t = y 0 + (y+
0 α+1
) t, x = y 0 + uλ0 t = y λ + (λ + (y+
λ α+1
) )t ,

where we are restricting our attention to the region where |x|, |y| < 1. Show
that the respective solutions u0 (t, x) and uλ (t, x) satisfy

u0 (t, x) = 0 , x<0; u0 (t, x) = (y+

0 α+1
) , 0≤x;
uλ (t, x) = λ , x < λt ; uλ (t, x) = λ + (y+
λ α+1
) , λt ≤ x .

0 = y 0 (t, x) and y λ = y λ (t, x) in terms of (t, x).

(c) Describe the quantities y+ + + +
In particular, show that
0 α
x = [1 + (y+ ) t]y 0
and therefore for a sufficiently short time interval,
xα+
(y+0 )α = + O((x+ t)2 ) .
1 + αxα+ t

Therefore one concludes that ∂xs uλ (t, x) ∈ L2 (R1 ) and the same for ∂xs u0 (t, x).

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 200 8. Conservation laws and shocks

(d) In the space-time region 0 ≤ x ≤ λt, the solution uλ (t, x) is constant

while u0 (t, x) is not. Show that
 λt
|∂xs u0 (t, x)|2 dx ∼ (λt)2(α−s+1)+1 ,
0
and therefore
uλ (t, ·) − u0 (t, ·) 2
Hs ≥ ∂xs (uλ (t, ·) − u0 (t, ·)) 2
L2 ≥ C(λt)2(α−s+1)+1 .
Furthermore, the difference in the initial data is given by
uλ0 − u00 Hs =λ ϕ Hs .
We may take (α − s + 1) + 1
2 = α−s+ 3
2 arbitrarily close to zero. The
conclusion is that
3
uλ (t, ·) − u0 (t, ·) Hs ≥ Cλα−s+ 2
α−s+ 23 α−s+ 32
≥C λ ϕ Hs =C uλ0 − u00 Hs .
Thus the solution map for Burger’s equation posed on H s (R1 ) does not sat-
isfy the criterion for Hölder continuity with exponent γ = α − s + 32 . Since
γ can be made arbitrarily small, such examples exclude the possibility that
the solution map is Hölder continuous for any γ. As an extended project,
one could ask further whether the solution map satisfies a uniform modulus
of continuity that is weaker than Hölder continuity.
The well posedness criterion of Hadamard for evolution equations can
be compared with the concept of a flow for a dynamical system. Gener-
ally for the latter, one asks that a flow satisfy (i) existence of solutions, (ii)
uniqueness, and (iii) smoothness of the solution map. The above examples
illustrate that a partial differential evolution equation may have a well-posed
solution map which may not be a flow by this definition.

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Bibliography

[1] B. Abbasi and W. Craig, On the initial value problem for the wave equation in Friedmann-
Robertson-Walker space-times, Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci. 470 (2014),
no. 2169, 20140361, 13, DOI 10.1098/rspa.2014.0361. MR3238168
[2] C. Bardos and E. S. Titi, Loss of smoothness and energy conserving rough weak solutions
for the 3d Euler equations, Discrete Contin. Dyn. Syst. Ser. S 3 (2010), no. 2, 185–197, DOI
10.3934/dcdss.2010.3.185. MR2610558
[3] R. Beals, Analysis: An introduction, Cambridge University Press, Cambridge, 2004.
MR2098699
[4] W. Craig and J. Goodman, Linear dispersive equations of Airy type, J. Differential Equations
87 (1990), no. 1, 38–61, DOI 10.1016/0022-0396(90)90014-G. MR1070026
[5] H. Dym and H. P. McKean, Fourier series and integrals, Probability and Mathematical
Statistics, vol. 14, Academic Press, New York-London, 1972. MR0442564
[6] L. C. Evans, Partial differential equations, 2nd ed., Graduate Studies in Mathematics, vol. 19,
American Mathematical Society, Providence, RI, 2010. MR2597943
[7] G. B. Folland, Introduction to partial differential equations, 2nd ed., Princeton University
Press, Princeton, NJ, 1995. MR1357411
[8] P. R. Garabedian, Partial differential equations, John Wiley & Sons, Inc., New York-London-
Sydney, 1964. MR0162045
[9] F. John, Partial differential equations, 4th ed., Applied Mathematical Sciences, vol. 1,
Springer-Verlag, New York, 1982. MR831655
[10] T. Kato, The Cauchy problem for quasi-linear symmetric hyperbolic systems, Arch. Rational
Mech. Anal. 58 (1975), no. 3, 181–205, DOI 10.1007/BF00280740. MR0390516
[11] W. Thomson, On ship waves, Institution of Mechanical Engineers, Proceedings 38 (1887),
409–434.
[12] S. Klainerman, Global existence for nonlinear wave equations, Comm. Pure Appl. Math. 33
(1980), no. 1, 43–101, DOI 10.1002/cpa.3160330104. MR544044
[13] S. Klainerman and P. Sarnak, Explicit solutions of cmu = 0 on the Friedmann-Robertson-
Walker space-times, Ann. Inst. H. Poincaré Sect. A (N.S.) 35 (1981), no. 4, 253–257 (1982).
MR650132
[14] P. Lax, Hyperbolic Systems of Conservation Laws and the Mathematical Theory of Shock
Waves, CBMS-NSF Regional Conference Series in Applied Mathematics, Society for Indus-
trial and Applied Mathematics (1973). ISBN: 978-0-89871-177-6.

201

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 202 Bibliography

[15] P.-L. Lions, Mathematical topics in fluid mechanics. Vol. 2. Compressible Models, Oxford
Lecture Series in Mathematics and its Applications, vol. 10, Oxford Science Publications, The
Clarendon Press, Oxford University Press, New York, 1998. MR1637634
[16] R. McOwen, Partial Differential Equations: Methods and Applications, Second Edition, Pear-
son (2002). ISBN-10: 0130093351.
[17] L. Nirenberg, Lectures on linear partial differential equations, Conference Board of the Mathe-
matical Sciences Regional Conference Series in Mathematics, vol. 17, American Mathematical
Society, Providence, R.I., 1973. MR0450755
[18] J. Rauch, Partial differential equations, Graduate Texts in Mathematics, vol. 128, Springer-
Verlag, New York, 1991. MR1223093
[19] M. E. Taylor, Pseudodifferential operators, Princeton Mathematical Series, vol. 34, Princeton
University Press, Princeton, N.J., 1981. MR618463
[20] A. N. Tikhonov and V. Y. Arsenin, Solutions of ill-posed problems, Scripta Series in Math-
ematics, V. H. Winston & Sons, Washington, D.C.: John Wiley & Sons, New York-Toronto,
Ont.-London, 1977. Translated from the Russian. MR0455365
[21] A. Vasy, Partial differential equations: An accessible route through theory and applications,
Graduate Studies in Mathematics, vol. 169, American Mathematical Society, Providence, RI,
2015. MR3410751

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Index

associated Laguerre functions, 29, 171 electrical impedence tomography, 91

associated Legendre functions, 29, 170 entire functions, 134
entire functions of exponential type, 135
Banach space, 105 entropy, 62
Bloch decomposition, 36 entropy, relative, 64
Brownian scaling, 43 Euler equations, 196
Burger’s equation solution map, 198 Euler–Poisson–Darboux equation, 129

Cauchy–Riemann equations, 3
Féjer kernel, 117
Cesaro integrability, 116
first-order quasilinear equation, 178
complete orthonormal basis, 26
arrow of time, 185
conformal mapping, 2, 8, 99
Burger’s equation, 178
conservation law, 15, 34
characteristic curves, 179
constituitive law, 31
characteristic equations, 179
convex dual, 187
convex duality, 187 conservation laws, 178
convolution, 46 entropy condition, 185
convolution algebra, 114 Hamilton–Jacobi equation, 188
jump condition, 182
d’Alembert’s formula, 17 Lax–Olenik solutions, 186
Darboux equation, 127 Lax–Olenik variational principle, 189
diffusion equation, 56 Lax–Olenik weak solutions, 190
Dirac δ-function, 49, 109, 110 noncharacteristic Cauchy surface, 180
Dirichlet–Neumann operator, 86, 88, 94 Riemann problem, 183
dispersion, 151 shock curves, 192
dispersion relation, 160 shocks, 181
dispersive equation, 34 uniqueness, 194
distributions, 110 weak solutions, 181
double layer potential, 85, 87 flow of a dynamical system, 200
Duhamel’s principle, 19 fluid dynamics, 196
heat equation, 53 Fourier multiplier, 154
wave equation, 19 Fourier transform, 10, 105
Fourier inversion formula, 10
Einstein’s equations, 6 isometry, 105

203

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 204 Index

Schwartz class, 110 integrable functions, 113

Fréchet derivative, 93
Fréchet spaces, 108 John, F., 145
joke, 2
Gâteau derivative, 66
Klein–Gordon equation, 145
group velocity, 161
Lp spaces, 105
Hadamard variational formula, 93, 99 Laplace’s equation, 3, 71
Hadamard well posed criterion, 198 attenuation of oscillatory data, 91
Hamiltonian PDE, 139, 140, 141 conformal transformations, 99
action integral, 139 Dirichlet problem, 72
Euler–Lagrange equations, 139 eigenfunctions, 97
Hamilton’s canonical equations, 142 fundamental solution, 74, 95
Kortweg–de Vries equation, 144 Gauss’ law of arithmetic mean, 77, 98
Lagrangian, 139 Green’s function, 81, 98
nonlinear Schrödinger equation, 144 Green’s function for Rn + , 81
nonlinear wave equation, 143 Green’s function for BR (0), 82
principle of least action, 139 Green’s identities, 72, 73
harmonic function, 3, 72 interior regularity, 76
heat equation, 4, 43 Laplacian operator, 3
conservation laws and moments, 56 maximum principle for subharmonic
contraction on L1 , 60 functions, 79
contraction on L2 , 60 maximum principle, strong, 79, 90
Dirichlet boundary conditions, 52 maximum principle, weak, 77
energy functional, 65 minimum principle for superharmonic
entropy, 68 functions, 79
Fourier’s law, 57 mixed boundary conditions, 98
Green’s identities, 55 Neumann problem, 72, 84
heat kernel, 45 Poisson kernel, 83, 90
initial value problem, 49 Poisson kernel for Rn + , 84, 87
initial-boundary value problem, 55 Poisson kernel for BR (0), 84
maximum principle, 48 Poisson problem, 72
method of images, 51, 67 Robin problem, 72
Neumann boundary conditions, 52 separation of variables, 95, 96
nonuniqueness of solutions, 69 subharmonic function, 79
semigroup property, 46, 66 superharmonic function, 79
separation of variables, 68 Lax–Olenik solutions, 186
weak maximum principle, 48 Legendre transform, 187
Heisenberg uncertainty principle, 158, Lipschitz continuity, 106
167
Maxwell’s equations, 4
Hermite functions, 28
method of characteristics, 12
Hilbert space, 104, 105
method of stationary phase, 162, 163
Hölder continuity, 106
metric space, 108
Hölder continuity, lack of for solution
minimal surfaces, 7
map of PDE, 200
Monge–Ampère equations, 5
Hölder inequality, 119
Hölder spaces, 106 Navier–Stokes equations, 8
Huygens’ principle, 18, 132, 138
hydrogen atom, 28, 40, 156 Paley–Wiener theory, 134, 136
orbitals, 169 phase velocity, 161
hyperbolic rotation, 123 physics of music, 29

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Index 205

Beals, R., 32 variation, 93

clarinet, 31 variation of constants, 21
DeTurck, D., 33
drum, 33 water waves, 162
oboe, 32, 37 dispersion relation, 173
saxophone, 32 Kelvin ship wake, 172
strings, 29 wave equation, 4, 16, 121
synthesizers, 33 Cauchy problem, 16, 121
woodwinds, 30 Cauchy problem in
Plancherel identity, 105 Friedmann–Robertson–Walker
space-time, 148
quantum harmonic oscillator, 27, 39, conservation of energy, 35, 122
156 Dirichlet boundary conditions, 21
quantum mechanics, 26 domain of dependence, 134
domain of influence, 134
real analytic functions, 108 equipartition of energy, 35, 147
Riemann–Lebesgue Lemma, 114 finite propagation speed, 18, 133
focusing singularity, 145
Schauder spaces, 106 Fourier synthesis, 122
Schrödinger’s equation, 6, 152 global existence, 145
bounds on moments, 168 Huygens’ principle, 132, 138
complex probability amplitude, 152 Huygens’ principle, strong, 132
conservation laws, 154, 166 inhomogeneous problem, 19
Dirichlet boundary conditions, 167 initial value problem, 121
Gaussian wave packets, 168 John, F., 145
kinetic energy, 154 Kirchhoff’s formula, 130
momentum distribution, 154 light cone, 17, 125
periodic boundary conditions, 167 Lorentz transformations, 123
position distribution, 154 Minkowski space, 126, 148
potential, 166 Neumann boundary conditions, 21
quantum mechanical flux, 156 propagation of singularities, 36
Schrödinger kernel, 152 spherical means, 127
separation of variables, 167, 169 spherical means in R2 , 146
Sobolev estimates, 155 spherical means in R3 , 129
wave function, 152 spherical means in Rn for n odd, 130
Schwartz class, 107 spherical means in R2m , 147
self-adjoint operator, 27 traveling waves, 35
separation of variables, 24 weak solutions, 17
heat equation, 68
Laplace’s equation, 95
Schrödinger’s equation, 167
wave equation, 24
shear flow, 196
single layer potential, 84
Sobolev embedding theorem, 106
Sobolev space, 105
special theory of relativity, 126
strong Huygens’ principle, 18
symbol, 160

tempered distributions, 110

transport equations, 10

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.
 Does entropy really increase no matter what we do? Can light
pass through a Big Bang? What is certain about the Heisenberg
uncertainty principle? Many laws of physics are formulated in
terms of differential equations, and the questions above are about

Photo by Deborah Craig

the nature of their solutions. This book puts together the three
main aspects of the topic of partial differential equations, namely
theory, phenomenology, and applications, from a contemporary
point of view. In addition to the three principal examples of the
wave equation, the heat equation, and Laplace’s equation, the book has chapters on
dispersion and the Schrödinger equation, nonlinear hyperbolic conservation laws, and
shock waves.
The book covers material for an introductory course that is aimed at beginning
graduate or advanced undergraduate level students. Readers should be conversant
with multivariate calculus and linear algebra. They are also expected to have taken
an introductory level course in analysis. Each chapter includes a comprehensive set
of exercises, and most chapters have additional projects, which are intended to give
students opportunities for more in-depth and open-ended study of solutions of partial
differential equations and their properties.

For additional information

and updates on this book, visit
www.ams.org/bookpages/gsm-197

GSM/197
www.ams.org

Licensed to Univ Nac Autonoma de Mexico. Prepared on Sun Jul 16 20:20:10 EDT 2023for download from IP 132.248.9.34.