PDF

Universitext
DavidBorthwick
Introduction
to Partial
Differential
Equations
Universitext
Universitext
Series editors
Sheldon Axler
San Francisco State University, San Francisco, CA, USA
Vincenzo Capasso
Universit degli Studi di Milano, Milan, Italy
Carles Casacuberta
Universitat de Barcelona, Barcelona, Spain
Angus MacIntyre
Queen Mary University of London, London, UK
Kenneth Ribet
University of California, Berkeley, CA, USA
Claude Sabbah
CNRS cole Polytechnique Centre de mathmatiques, Palaiseau, France
Endre Sli
University of Oxford, Oxford, UK
Wojbor A. Woyczyski
Case Western Reserve University, Cleveland, OH, USA
Universitext is a series of textbooks that presents material from a wide variety of

mathematical disciplines at masters level and beyond. The books, often well
class-tested by their author, may have an informal, personal, even experimental
approach to their subject matter. Some of the most successful and established books
in the series have evolved through several editions, always following the evolution
of teaching curricula, into very polished texts.
Thus as research topics trickle down into graduate-level teaching, rst textbooks
written for new, cutting-edge courses may make their way into Universitext.
More information about this series at http://www.springer.com/series/223

David Borthwick
Introduction to Partial
Differential Equations
123
David Borthwick
Department of Mathematics
and Computer Science
Emory University
Atlanta, GA
USA
ISSN 0172-5939 ISSN 2191-6675 (electronic)

Universitext
ISBN 978-3-319-48934-6 ISBN 978-3-319-48936-0 (eBook)
DOI 10.1007/978-3-319-48936-0
Library of Congress Control Number: 2016955918
Springer International Publishing AG 2016

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microlms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specic statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained herein or
for any errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Dedicated to my parents
Preface
Partial differential equations (PDE) rst appeared over 300 years ago, and the vast
scope of the theory and applications that have since developed makes it challenging
to give a reasonable introduction in a single semester. The modern mathematical
approach to the subject requires considerable background in analysis, including
topics such as metric space topology, measure theory, and functional analysis.
This book is intended for an introductory course for students who do not nec-
essarily have this analysis background. Courses taught at this level traditionally
focus on some of the more elementary topics, such as Fourier series and simple
boundary value problems. This approach risks giving students a somewhat narrow
and outdated view of the subject.
My goal here is to give a balanced presentation that includes modern methods,
without requiring prerequisites beyond vector calculus and linear algebra. To allow
for some of the more advanced methods to be reached within a single semester, the
treatment is necessarily streamlined in certain ways. Concepts and denitions from
analysis are introduced only as they will be needed in the text, and the reader is
asked to accept certain fundamental results without justication. The emphasis is
not on the rigorous development of analysis in its own right, but rather on the role
that tools from analysis play in PDE applications.
The text generally focuses on the most important classical PDE, which are the
wave, heat, and Laplace equations. Nonlinear equations are discussed to some
extent, but this coverage is limited. (Even at a very introductory level, the nonlinear
theory merits a full course to itself.)
I have tried to stress the interplay between modeling and mathematical analysis
wherever possible. These connections are vital to the subject, both as a source of
problems and as an inspiration for the development of methods.
vii
viii Preface
I owe a great debt of gratitude to my colleague Alessandro Veneziani, with

whom I collaborated on this project originally. The philosophy of the book and
choice of topics were heavily influenced by our discussions, and I am grateful for
his support throughout the process. I would also like to thank former student Dallas
Albritton, for offering comments and suggestions on an early draft. Thanks also to
the series editors at Springer, for comments that helped improve the writing and
presentation.
Atlanta, USA David Borthwick

September 2016
Contents
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Partial Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Example: dAlemberts Wave Equation . . . . . . . . . . . . . . . . . . . 2
1.3 Types of Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Well Posed Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.1 Real Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Complex Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Domains in Rn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 Differentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Ordinary Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6 Vector Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.7 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3 Conservation Equations and Characteristics . . . . . . . . . . . . . . . . . . . 25
3.1 Model Problem: Oxygen in the Bloodstream . . . . . . . . . . . . . . . 25
3.2 Lagrangian Derivative and Characteristics . . . . . . . . . . . . . . . . . 27
3.3 Higher-Dimensional Equations . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 Quasilinear Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.5 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4 The Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 Model Problem: Vibrating String . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Boundary Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4 Forcing Terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.5 Model Problem: Acoustic Waves . . . . . . . . . . . . . . . . . . . . . . . . 59
4.6 Integral Solution Formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.7 Energy and Uniqueness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.8 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
ix
x Contents
5 Separation of Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.1 Model Problem: Overtones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2 Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.3 Circular Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.4 Spherical Symmetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.5 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6 The Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
6.1 Model Problem: Heat Flow in a Metal Rod . . . . . . . . . . . . . . . . 97
6.2 Scale-Invariant Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
6.3 Integral Solution Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.4 Inhomogeneous Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.5 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.1 Inner Products and Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
7.2 Lebesgue Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.3 Lp Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.4 Convergence and Completeness . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.5 Orthonormal Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.6 Self-adjointness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
7.7 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
8 Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.1 Series Solution of the Heat Equation . . . . . . . . . . . . . . . . . . . . . 131
8.2 Periodic Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
8.3 Pointwise Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
8.4 Uniform Convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
8.5 Convergence in L2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
8.6 Regularity and Fourier Coefcients . . . . . . . . . . . . . . . . . . . . . . . 145
8.7 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
9 Maximum Principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
9.1 Model Problem: The Laplace Equation . . . . . . . . . . . . . . . . . . . . 155
9.2 Mean Value Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9.3 Strong Principle for Subharmonic Functions . . . . . . . . . . . . . . . . 165
9.4 Weak Principle for Elliptic Equations . . . . . . . . . . . . . . . . . . . . . 167
9.5 Application to the Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . 170
9.6 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
10 Weak Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10.1 Test Functions and Weak Derivatives . . . . . . . . . . . . . . . . . . . . . 177
10.2 Weak Solutions of Continuity Equations . . . . . . . . . . . . . . . . . . 180
10.3 Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.4 Sobolev Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
Contents xi
10.5 Weak Formulation of Elliptic Equations . . . . . . . . . . . . . . . . . . . 194

10.6 Weak Formulation of Evolution Equations . . . . . . . . . . . . . . . . . 196
10.7 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
11 Variational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
11.1 Model Problem: The Poisson Equation . . . . . . . . . . . . . . . . . . . . 206
11.2 Dirichlets Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
11.3 Coercivity and Existence of a Minimum. . . . . . . . . . . . . . . . . . . 208
11.4 Elliptic Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
11.5 Eigenvalues by Minimization . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
11.6 Sequential Compactness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
11.7 Estimation of Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
11.8 Euler-Lagrange Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
11.9 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
12 Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
12.1 Model Problem: Coulombs Law . . . . . . . . . . . . . . . . . . . . . . . . 239
12.2 The Space of Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
12.3 Distributional Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
12.4 Fundamental Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
12.5 Greens Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
12.6 Time-Dependent Fundamental Solutions . . . . . . . . . . . . . . . . . . . 257
12.7 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
13 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
13.1 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
13.2 Tempered Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
13.3 The Wave Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
13.4 The Heat Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
13.5 Exercises. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Appendix A: Analysis Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Notations
: Equal by denition
Equal except on a set of measure zero
Asymptotic to (ratio approaches 1)
Comparable to (ratio bounded above and below)
k k Norm
k kp Lp norm
h; i Inner product (L2 by default)
; Distributional pairing
An Volume of the unit sphere Sn1 Rn
Bn Open unit ball fjxj\1g Rn
Bx0 ; R Open ball fjx x0 j\Rg Rn
ck Fourier coefcient
C Complex numbers
C m m-times continuously differentiable functions ! C
C m ; R Real valued C m functions
C m C m functions that admit extension across o
m
Ccpt C m functions with compact support in
A Characteristic (or indicator) function of a set A
D Unit disk in R2
D Multivariable derivative
x Dirac delta function
o Boundary of
D0 Distributions on
Df Dirichlet energy
dS Surface integral element
E Energy of a solution
F Fourier transform
Gamma function
r Gradient operator
H m Sobolev space, functions with weak derivatives in L2 to order m
xiii
xiv Notations
m
Hloc Local Sobolev functions
1
H01 Closure of Ccpt in H 1
Ht Heat kernel
Laplacian operator on Rn
Lp p-th power integrable functions ! C
L1loc Locally integrable functions ! C
p Discrete Lp space
N Natural numbers {1; 2; 3; . . .}
N0 Non-negative integers {0; 1; 2; 3; . . .}
Outward unit normal
Domain (open, connected set) in Rn

Closure of ( [ o)
Fundamental solution
r jxj in Rn
R Real numbers
R Rayleigh quotient
Sn1 Unit sphere fjxj 1g Rn
SRn Schwartz functions (smooth functions with rapid decay)
S 0 Rn Tempered distributions
Sn Partial sum of Fourier series
T R=2Z
Wt Wave kernel
Z Integers {. . .; 1; 0; 1; . . .}
Chapter 1
Introduction
1.1 Partial Differential Equations
Continuous phenomena, such as wave propagation or fluid flow, are generally mod-
eled with partial differential equations (PDE), which express relationships between
rates of change with respect to multiple independent variables. In contrast, phenom-
ena that can be described with a single independent variable, such as the motion
of a rigid body in classical physics, are modeled by ordinary differential equations
(ODE).
A general PDE for a function u has the form

u mu
F x, u(x), (x), . . . , (x) = 0. (1.1)
x j x j1 . . . x jm
The order of this equation is m, the order of the highest derivative appearing (which
is assumed to be finite). A classical solution u admits continuous partial derivatives
up to order m and satisfies (1.1) at all points x in its domain. In certain situations the
differentiability requirements can be relaxed, allowing us to define weak solutions
that do not solve the equation literally.
A somewhat subtle aspect of the definition (1.1) is the fact that the equation
is required to be local. This means that functions and derivatives appearing in the
equation are all evaluated at the same point.
Although classical physics provided the original impetus for the development
of PDE theory, PDE models have since played a crucial role in many other fields,
including engineering, chemistry, biology, ecology, medicine, and finance. Many
industrial applications of mathematics are based on the numerical analysis of PDE.
Most PDE are not solvable in the explicit sense that a simple calculus problem can
be solved. That is, we typically cannot obtain a exact formula for u(x). Therefore
much of the analysis of PDE is focused on drawing meaningful conclusions from an
equation without actually writing down a solution.
Springer International Publishing AG 2016 1

D. Borthwick, Introduction to Partial Differential Equations,
Universitext, DOI 10.1007/978-3-319-48936-0_1
2 1 Introduction
1.2 Example: dAlemberts Wave Equation
One of the earliest and most influential PDE models was the wave equation, devel-
oped by Jean dAlembert in 1746 to describe the motion of a vibrating string. With
physical constants normalized to 1, the equation reads
2u 2u
= 0, (1.2)
t 2 x2
where u(t, x) denotes the vertical displacement of the string at position x and time
t. If the string has length and is attached at both ends, then we also require that
u(t, 0) = u(t, ) = 0 for all t. We will discuss the formulation of this model in
Sect. 4.1.
DAlembert also found a general formula for the solution of (1.2), based on the
observation that (1.2) is solved by any function of the form f (x t), assuming f is
twice-differentiable. Given two such functions on R, we can write a general solution
u(t, x) := f 1 (x + t) + f 2 (x t). (1.3)
A similar formula applies in the case of a string with fixed ends. If f is 2-periodic
on R, meaning f (x + 2) = f (x) for all x, then it is easy to check that
1
u(t, x) := [ f (x + t) f (t x)] (1.4)
2
satisfies u(t, 0) = u(t, ) = 0 for any t.
One curious feature of this formula is that it appears to give a sensible solution
even in cases where f is not differentiable. For example, to model a plucked string
we might take the initial displacement to be a simple piecewise linear function in the
form of a triangle from the fixed endpoints, as shown in Fig. 1.1.
If we extend this to an odd, 2-periodic function on R, then the formula (1.4) yields
the result illustrated in Fig. 1.2. The initial kink splits into two kinks which travel in
opposite directions on the string and and appear to rebound from the fixed ends.
This is not a classical solution because u is not differentiable at the kinks. However,
u does satisfy the requirements for a weak solution, as we will see in Chap. 10.
Although a physical string could not exhibit sharp corners without breaking, the
piecewise linear solutions are nevertheless physically reasonable. Direct observations
of plucked and bowed strings were first made in the late 19th century by Hermann
von Helmholtz, who saw patterns of oscillation quite similar to what is shown in
Fig. 1.2. The appearance of kinks propagating along the string is striking, although
the corners are not exactly sharp.
Fig. 1.1 Initial state of a

plucked string
1.3 Types of Equations 3
Fig. 1.2 Evolution of the

plucked string, starting from
t = 0 at the top
1.3 Types of Equations
There is no general theory of PDE that allows us to analyze all equations of the form
(1.1). To make progress it is necessary to restrict our attention to certain classes of
equations and develop methods appropriate to those.
The most fundamental distinction between PDE is the property of linearity. A
PDE is called linear if it can be written in the form
Lu = f, (1.5)
where f is some function independent of u, and L is a differential operator. Many of

the important classical PDE that we will discuss in this book are linear and of first
or second order. For such cases L has the general form

n
2
n

L= ai j + bj + c, (1.6)
i, j=1
xi x j j=1
xj
4 1 Introduction
where the coefficients ai j , b j , and c are functions of x. The second-order coefficients

are assumed to be symmetric, ai j = a ji , because the mixed partials derivatives of a
twice continuously differentiable function commute.
Linearity implies that a linear combination of solutions is still a solution, a fact that
is referred to as the superposition principle. Superposition often lets us decompose
problems into simpler components, which is the main reason that linear problems
are much easier to handle than nonlinear. It also makes it possible to work with
complex-valued solutions, which is sometimes more convenient, because the real
and imaginary parts of a complex solution will solve the equation independently.
Most linear PDE are derived as approximations to more realistic, nonlinear mod-
els. We will focus primarily on the linear case in this book. The main reason for this
is that nonlinear PDE are inherently more complicated, and for an introduction it
makes sense to start with the more basic theory. Furthermore, the analysis of nonlin-
ear problems frequently involves the study of associated linear approximations, so
that one must understand at least some of the linear theory first.
Linear equations are further classified by the properties of the terms with the
highest orders of derivatives, since this determines many qualitative properties of
solutions. Elliptic equations of second order are associated to an operator L of the
form (1.6), such that the eigenvalues of the symmetric matrix [ai j ] are strictly positive
at each point in the domain. The prototype of an elliptic operator is L = where
denotes the Laplacian,
2 2
:= 2 + + 2 , (1.7)
x1 xn
named after the mathematician and physicist Pierre-Simon Laplace.

Equations that include time as an independent variable are called evolution equa-
tions. The time variable usually plays a very different role from the spatial variables,
so in such cases we adapt the form (1.6) by separating out the time derivatives
explicitly.
The two classic types of second-order evolution equations are hyperbolic and
parabolic. Hyperbolic equations are exemplified by dAlemberts wave equation
(1.2). The general form is (1.5) with
2u n
2u
L= ai j (x) + (lower order terms), (1.8)
t 2 i, j=1 xi x 2j
where once again [ai j ] is assumed to be a strictly positive matrix. Hyperbolic equa-
tions are used to model oscillatory phenomena.
Parabolic evolution equations have the form (1.5) with
u n
2u
L= ai j (x) + (lower order terms), (1.9)
t i, j=1
xi x 2j
1.3 Types of Equations 5
where [ai j ] is a strictly positive matrix. The heat equation, whose derivation we will
discuss in detail in Sect. 6.1, is the prototype for this type of equation. Parabolic
equations are generally used to model phenomena of conduction and diffusion.
Note that hyperbolic and parabolic equations revert to elliptic equations in the
spatial variables if the solution is independent of time. Elliptic equations thus serve
to model the equilibrium states of evolution equations.
Because of their association with phenomenological properties of a system, the
terms elliptic, hyperbolic, and parabolic are frequently applied more broadly
than this simple classification would suggest. A nonlinear equation is typically
described by the category of its linear approximations, which can change depending
on the conditions.
For problems on a bounded domain, the application usually dictates some restric-
tion on the solutions at the boundary. Two very common types are Dirichlet bound-
ary conditions, specifying the values of u at the boundary, and Neumann conditions,
specifying the normal derivatives of u at the boundary. These conditions are named
for Gustave Lejeune Dirichlet and Carl Neumann, respectively. By default we will
use these terms in the homogeneous sense, meaning that the boundary values of the
function or derivative are set equal to zero. For evolution equations, we also impose
initial conditions, specifying the values of u and possibly its time derivatives at some
initial time.
1.4 Well Posed Problems
The set of functions used to formulate a PDE, which might include coefficients or
terms in the equation itself as well as boundary and initial conditions, is collectively
referred to as the input data. The most basic question for any PDE is whether a
solution exists for a given set of data. However, for most purposes we want to require
something more. A PDE problem is said to be well posed if, for a given set of data:
1. A solution exists.
2. The solution is uniquely determined by the data.
3. The solution depends continuously on the data.
These criteria were formulated by Jacques Hadamard in 1902. The first two properties
hold for ODE under rather general assumptions, but not necessarily for PDE. It is
easy to find nonlinear equations that admit no solutions, and even in the linear case
there is no guarantee.
The third condition, continuous dependence on the input data, is sometimes called
stability. One practical justification for this requirement is it is not possible to specify
input data with absolute accuracy. Stability implies that the effects of small variations
in the data can be controlled.
For certain PDE, especially the classical linear cases, we have a good under-
standing of the requirements for well-posedness. For other important problems, for
6 1 Introduction
Fig. 1.3 Numerical simulations of blood flow in the aorta. Courtesy of D. Gupta, Emory University
Hospital, and T. Passerini, M. Piccinelli and A. Veneziani, Emory Mathematics and Computer
Science
example in fluid mechanics, well-posedness remains a difficult unsolved conjecture.

Furthermore, many interesting problems are known not to be well posed. For exam-
ple, problems in image processing are frequently ill posed, because information is
lost due to noise or technological limitations.
1.5 Approaches
We can organize the methods for handling PDE problems according to three basic
goals:
1. Solving: finding explicit formulas for solutions.
2. Analysis: understanding general properties of solutions.
3. Approximation: calculating solutions numerically.
Solving PDE is certainly worth understanding in those special cases where it is
possible. The solution formulas available for certain classical PDE provide insight
that is important to the development of the theory.
The goals of theoretical analysis of PDE are extremely broad. We wish to learn
as much as we can about the qualitative and quantitative properties of solutions and
their relationship to the input data.
Finally, numerical computation is the primary means by which applications of
PDE are carried out. Computational methods rely on a foundation of theoretical
analysis, but also bring up new considerations such as efficiency of calculation.
Example 1.1 Figure 1.3 shows a set of numerical simulations modeling the insertion
in the aorta of a pipe-like device designed to improve blood flow. The leftmost frame
shows the aorta before surgery, and the three panes on the right model the insertion
1.5 Approaches 7
at different locations. The PDE model is a complex set of fluid equations called the
Navier-Stokes equations. These fluid equations are famously difficult to analyze and
an exact solution is almost never possible. However, the cylinder is one case that can
be handled explicitly. For the numerical simulations, exact solutions for a cylindrical
pipe were used to provide boundary data at the point where the pipe meets the aorta.
Theoretical analysis also plays an important role here, in that the regularity theory
for the fluid equations is used to predict the accuracy of the simulation. (The complete
well-posedness analysis of the Navier-Stokes equations remains a famously unsolved
problem, however.)
The simulated flows displayed in Fig. 1.3 were computed numerically by a tech-
nique called the finite element method. This involves discretizing the problem to
reduce the PDE to a system of linear algebraic equations. Modeling a single heart-
beat in this simulation require solving a linear system of about 500 million equations.
Chapter 2
Preliminaries
In this chapter we set the stage for the study of PDE with a review of some core
background material.
2.1 Real Numbers
The real number system R is constructed as the completion of the field of ratio-
nal numbers. This means that in addition to the algebraic axioms for addition and
multiplication, R satisfies an additional axiom related to the existence of limits. To
state this axiom we use the concept of the supremum (or least upper bound) of
subset A R. The supremum is a number sup(A) R such that (1) all elements of
A are less than or equal to sup(A); and (2) no number strictly less than sup(A) has
this property. The completeness axiom says that every nonempty subset of R that is
bounded above has a supremum. An equivalent statement is that a nonempty subset
that is bounded below has an infimum (greatest lower bound), which is denoted
inf(A).
It is convenient to extend these definitions to unbounded sets by defining sup(A) :=
when A is not bounded above, and inf(A) := when the set is not bounded
below. We also set sup() = and inf() := +. With these extensions, sup
and inf are defined for all subsets of R.
To illustrate the definition, we present a simple result that will prove useful in the
construction of approximating sequences for solutions of PDE.
Lemma 2.1 For a nonempty set A R, there exists a sequence of points xk A
such that
lim xk = sup A,
k
and similarly for inf A.

10 2 Preliminaries
Proof If A is not bounded above then there exists a sequence of xk A with xk .

Therefore the claim holds when sup A = .
Now suppose that sup A = R. By the definition of the supremum, 1/k
is not an upper bound of A for k N. Therefore, for each k there exists xk A such
that 1/k < xk . This yields a sequence such that xk .
There is an important distinction between supremum and infimum and the related
concepts of maximum and minimum. The latter are required to be elements of the set
and thus may not exist. For example, the interval (0, 1) has sup = 1 and inf = 0, but
has neither max nor min.
2.2 Complex Numbers
The complex number system C consists of numbers of the form z = x + iy, where
x, y R and i 2 := 1. The numbers x and y are called the real and imaginary parts
of z. The conjugate of z is
z := x iy,
so that
z + z z z
Re z := , Im z := .
2 2i
A nonzero complex number has a multiplicative inverse, given by
1 x iy
= 2 .
x + iy x + y2
The absolute value on C is the vector absolute value from Euclidean R2 ,

|z| := x 2 + y 2 .
This can be written in terms of conjugation,

|z| = z z ,
which shows in particular that the absolute value is multiplicative,
|zw| := |z||w|
for z, w C.
The basic theory of sequences and series carries over from R to C with only minor
changes. A sequence {z k } in C converges to z if
lim |z k z| = 0,
k
2.2 Complex Numbers 11

and a series ak converges if the sequence of partial sums nk=1 ak is convergent.
The series converges absolutely if

|ak | < .
k=1
It follows from the completeness axiom of R that absolute convergence of a series

in C implies convergence.
The exponential series,

zk
e z := , (2.1)
k=0
k!
converges absolutely for all z C. The special case where z is purely imaginary
gives an important relation called Eulers formula:

2 4 3
ei = 1 + + + +i +
2! 4! 1! 3!
= cos + i sin . (2.2)
Leonhard Euler, arguably the most influential mathematician of the 18th century,
published this identity in 1748. It yields a natural polar-coordinate representation of
complex numbers,
z = r ei ,
where r = |z| and is the angle between z and the positive real axis.
The product rule for complex exponentials,
e z ew = e z+w , (2.3)
follows from the power series definition just as in the real case. In combination with
(2.2) this allows for a very convenient manipulation of trigonometric functions. For
example, setting z = i and w = i in (2.3) and taking the real and imaginary parts
recovers the identities
cos( + ) = cos cos sin sin ,

sin( + ) = cos sin + sin cos .
The calculus rules for differentiating and integrating exponentials are derived
from the power series expansion, and thus extend to the complex case. In particular,
d ax
e = aeax
dx
for a C.
12 2 Preliminaries
2.3 Domains in Rn
For points in Rn we will use the vector notation x = (x1 , . . . , xn ). The Euclidean
dot product is denoted x y, and the Euclidean length of a vector is written

|x| := x x.
The Euclidean distance is used to define limits: limk x k = w means that
lim |x k w| = 0.
k
The ball of radius R > 0 centered at a point x 0 Rn is

B(x 0 ; R) := x Rn ; |x x 0 | < R .
A small ball centered at x 0 is called a neighborhood of x 0 . If x A has a neighbor-

hood contained in A then x is called an interior point.
A subset U Rn is open if all of its points are interior. This generalizes the notion
of an open interval in one dimension. The ball B(x 0 ; R) is open, for example, as is
Rn itself. The empty set is open by default.
A boundary point of A Rn is a point x Rn such that every neighborhood of x
intersects both A and its complement. The distinction between interior and boundary
points is illustrated in Fig. 2.1. Note that boundary points may or may not be included
in the set itself. The boundary of A is denoted
A := {boundary points of A} .
For example, the boundary of the ball B(x 0 ; R) is the sphere

B(x 0 ; R) = x Rn ; |x x 0 | = R .
A set is open if and only if it contains no boundary points.
Fig. 2.1 Open rectangle

in R2
boundary point
interior point
2.3 Domains in Rn 13
A subset of Rn is connected if any two points in the set can be joined by a

continuous path within the set. For an open set U this is equivalent to the condition
that U cannot be written as the disjoint union of two nonempty open sets.
We will refer to a connected open subset Rn as a domain, and reserve the
notation for this usage. For some problems we assume the domain is bounded,
meaning that B(0; R) for sufficiently large R.
The concept of a closed interval can also be generalized to higher dimension. A
subset F Rn is closed if it contains all of its boundary points, i.e.,
F F.
The union of a subset A Rn with its boundary is called the closure of A and
denoted
A := A A.
For example,
B(x 0 ; R) := x Rn ; |x x 0 | R .
It is potentially confusing that an overline is used for set closure and complex con-
jugation, but these notations are standard. Note that closure applies only to sets and
not to numbers or functions.
A closed set F Rn contains the limits of all sequences in F that converge in
R . This is because the limit of a sequence contained in a set must either be a point
n
in the set or a boundary point.

Closed and open sets are related in the sense that the complement of an open set
is closed, and vice versa. However, the terms are not mutually exclusive, and a set
might not have either property. The interval (a, b] R is neither open nor closed,
for example. The only subsets of Rn with both properties are Rn itself and .
2.4 Differentiability
The space of continuous, complex-valued functions on a domain Rn which

admit continuous partial derivatives up to order m is denoted by C m (). The assump-
tion of continuity for derivatives insures that mixed partials are independent of the
order of differentiation. A smooth function has continuous derivatives to all orders;
the corresponding space is written C ().
We use the notation C m (; R) to specify real-valued functions, and similarly
C (; Rn ) denotes the space of vector-valued functions. It is common to use C m as
m
an adjective, short for m-times continuously differentiable.

The definition of C m () makes no conditions on the behavior of functions as
the boundary is approached. To impose such restrictions, we use the notation C m ()
to denote the space of functions that admit C m extensions across the boundary. For
example, the function x C (0, 1) is an element of C 0 [0, 1], but not C 1 [0, 1].
14 2 Preliminaries
The support of f C 0 () is defined as
supp f := {x ; f (x)
= 0}. (2.4)
Note that the definition includes a closure. This means that the support does not
exclude points where the function merely passes through zero. For example, the
support of sin(x) is R rather than R\ Z.
A closed and bounded subset of Rn is said to be compact. We denote by Ccptm
()
the space of functions on that have compact support, meaning that supp f is a
compact subset of . Since is open and the support is closed, this requires in
particular that supp f be a strict subset of . For example, 1 x 2 vanishes at the
boundary of (1, 1), but does not have compact support in this domain because its
support is [1, 1].
Example 2.2 To demonstrate the existence of compactly supported smooth func-

tions, consider
e1/(1x ) , |x| < 1,
2
h(x) =
0, |x| 1,
which has support [1, 1]. As illustrated in Fig. 2.2, the function becomes extremely
flat as x 1.
To show that h is in fact smooth, we note that

qm (x) 1/(1x 2 )
(m) 2 me , |x| < 1,
h (x) = (1x )
0, |x| 1,
where qm denotes a polynomial of degree m. As x 1, the term (1 x 2 )m blows

up while the exponential term tends rapidly to zero. Using lHpitals rule one can
check that the exponential dominates this limit, so that all derivatives of h vanish as

x 1 from |x| 1. This shows that h Ccpt (R).
The function h can be integrated to produce a smooth function that is constant for
|x| 1. By translating and rescaling, this construction gives, for a < b, a function
C (R) satisfying
0, x a,
(x) =
1, x b.
Fig. 2.2 Compactly

supported smooth function
2.4 Differentiability 15
These smooth step functions are useful as building blocks for pasting together
smooth functions on different domains.
Certain problems require regularity of the boundary of . For example, many

theorems in vector calculus require the existence of a normal vector on the boundary,
which does not exist for a general domain. A standard hypothesis for such theorems
is that is piecewise C 1 . This means that consists of a finite number of compo-
nents which admit regular coordinate parametrization. A coordinate parametrization
is a map C 1 (U; Rn ) where U is a domain in Rn1 . To say the parametrization is
regular means that the tangent vectors defined by

1 n
:= ,..., ,
w j w j w j
j = 1, . . . , n 1, are linearly independent at each point of .

The piecewise C 1 boundary assumption guarantees that the points of each bound-
ary component have well-defined tangent spaces and normal directions. As an exam-
ple, the unit cube in R3 has piecewise C 1 boundary, consisting of 6 planar components
with normal directions parallel to the coordinate axes.
In this text we will focus on relatively simple domains with straightforward bound-
ary parametrizations.
2.5 Ordinary Differential Equations
Our development of PDE theory will not rely on any advanced techniques for the
solution of ODE, but it will be useful to recall some basic material.
First-order ODE can often be solved directly by methods from calculus. The
easiest cases are equations of the form
dy
= g(y)h(t),
dt
where the variables can be separated to yield an integral formula

dy dt
= .
g(y) h(t)
Integrating both sides yields a family of solutions with one undetermined constant.
16 2 Preliminaries
Example 2.3 Consider the equation
dy
= ay, y(0) = y0 .
dt
for a
= 0. This is called the growth or decay equation, depending on the sign of a.
Separating the variables gives

dy
= a dt,
y
which integrates to ln y = at + C. Solving for y and using the initial condition gives
y(t) = y0 eat .
Higher-order ODE are generally analyzed by reducing to a system of first-order

equations. To reduce the nth-order equation

y (n) (t) = F t, y, y , . . . , y (n1) , (2.5)
we define the vector-valued function w = (y, y , . . . , y (n1) ). This satisfies the first-
order system
w1 w2
d . .
.. ..
= .
dt wn1 wn
wn F(t, w)
First-order systems can be solved generally by the strategy of Picard iteration,

named for mathematician mile Picard. The first step is to write the vector equation,
dw
= F(t, w), w(t0 ) = w0 , (2.6)
dt
in an equivalent form as a recursive integral equation,

t
w(t) = w0 + F(s, w(s)) ds.
t0
For the construction, we set u0 (t) := w 0 and define a sequence of functions by

t
uk (t) = w 0 + F(s, uk1 (s)) ds (2.7)
t0
2.5 Ordinary Differential Equations 17
for k = 1, 2, . . . . It can be shown that the limit of this sequence exists and solves
(2.6) under some general assumptions on F, which leads to a proof of the following
result.
Theorem 2.4 (Picard iteration) Suppose that F is a continuous function on I
where I is an open interval containing t0 and is a domain in Rn containing w0 , and
that F is continuously differentiable with respect to w. Then (2.6) admits a unique
solution on some interval (t0 , t0 + ) with > 0.
Applying Theorem 2.4 to (2.5) shows that an nth order ODE satisfying the reg-
ularity assumptions has a unique local solution specified by the initial values of the
function and its first n 1 derivatives.
The C 1 hypothesis on F is stronger than necessary, but this version will suffice for
our purposes. The point we would like to stress here is the relative ease with which
ODE can be analyzed under very general conditions. This is very different from the
PDE theory, where no such general results are possible.
Example 2.5 The harmonic ODE is the equation
d 2y
= 2 y,
dt 2
for > 0. In view of the solution to the growth/decay equation in Example 2.3, it is
reasonable to start with an exponential solution as a guess. Substituting et into the
equation yields 2 = 2 . From = i we obtain the general solution,
y(t) = c1 eit + c2 eit .
To see how this relates to the Picard iteration method described above, consider
the corresponding system (2.6) for w = (y, y ):

dw 0 1
= w.
dt 2 0
With w0 = (a, b), the recursive formula (2.7) yields the sequence of functions
k j
tj 0 1 a
uk (t) =
j! 2 0 b
j=1
for k N. In the limit k this gives

(t)2 (t)4 a
w(t) = 1 +
2! 4! b

(t) 3
(t) 5
b
+ t + .
3! 5! 2 a
18 2 Preliminaries
Reading off y as the first component of w gives the familiar trigonometric solution,
y(t) = a cos(t) + b sin(t).
The trigonometric and complex exponential solutions are related by Eulers formula
(2.2).
2.6 Vector Calculus
The classical theorems of vector calculus were motivated by PDE problems arising
in physics. For our purposes the most important of these results is the divergence
theorem. We assume that the reader is familiar with the divergence theorem in the
context of R2 or R3 . In this section we will cover the basic definitions needed to state
the result in Rn and develop its corollaries.
As noted in Sect. 2.3, we always take a domain Rn to be connected and
open. The gradient of f C 1 () is the vector-valued function

f f
f := ,..., .
x1 xn
For a vector-valued function v C 1 (; Rn ) with components (v1 , . . . , vn ), the

divergence is
v1 vn
v := + + .
x1 xn
The Laplacian operator introduced in (1.7) is the divergence of the gradient
u := (u).
For this reason is sometimes written 2 .

If is bounded then the Riemannian integral of f C 0 () exists and is denoted
by
f (x) d n x,

where d n x is a shorthand for d x1 d xn . The integral can be extended to unbounded

domains if the appropriate limits exist. We will discuss a further generalization of
the Riemann definition in Chap. 7.
One issue we will come across frequently is differentiation under the integral.
If Rn is a bounded domain and u and u/t are continuous functions on
(a, b) , then the Leibniz integral rule says that
2.6 Vector Calculus 19

d u
u(t, x) d n x = (t, x) d n x.
dt t
Differentiation under the integral may still work when the integrals are improper, but
this requires greater care.
To set up boundary integrals for a domain with piecewise C 1 boundary, we need
to define the surface integral over a regular coordinate patch : U Rn1 .
Let : U Rn denote the unit normal vector pointing outwards from the domain.
The surface integral for such a patch is defined by

f d S :=
f ( (w)) det ,..., , d n1 w, (2.8)
(U ) U w w 1 n1
where det[. . . ] denotes the determinant of a matrix of column vectors. The full surface
integral over is defined by summing over the boundary coordinate patches. For
simplicity, we notate this as a single integral,

f d S.

In R2 , a boundary parametrization will be a curve (t) and (2.8) reduces to the

arclength integral

d
f d S := f ( (t)) dt.
(U ) dt U
In R3 , the unit normal for a surface patch can be computed from the cross product
of the tangent vectors. This leads to the surface integral formula

2
f d S := f ( (w)) d w.
(U ) U w1 w2
Even in low dimensions surface integrals can be rather complicated. We will make
explicit use of these formulas only in relatively simple cases, such as rectangular
regions and spheres.
We can use (2.8) to decompose integrals into radial and spherical components.
This is particularly useful when the domain is a ball. Let r := |x| be the radial
coordinate, and define the unit sphere
Sn1 := {r = 1} Rn .
A point x
= 0 can be written uniquely as r for Sn1 and r > 0. Let (y) be a
parametrization of Sn1 by coordinates y U Rn1 . For the change of variables
(r, y) x = r (y), the Jacobian formula gives
20 2 Preliminaries

x x x
d n x = det ,..., , dr dy1 dyn1 (2.9)
y1 yn1 r

= det ,..., , r n1 dr dy1 dyn1 .
y1 yn1
On the unit sphere, the outward unit normal is equal to . Thus (2.9) reduces to
d n x = r n1 dr d S(y).
For an integral over the ball this yields the radial integral formula,

R
f (x) d n x = f (r (y)) r n1 dr d S(y). (2.10)
B(0;R) Sn1 0
With these definitions in place, we turn to the divergence theorem, which relates
the flux of a vector field through a closed surface to the divergence of the field in the
interior. This result is generally attributed to Carl Friedrich Gauss, who published a
version in 1813 in conjunction with his work on electrostatics.
Theorem 2.6 (Divergence theorem) Suppose Rn is a bounded domain with
piecewise C 1 boundary. For a vector field F C 1 (; Rn ),

F dn x = F d S,

where is the outward unit normal to .

A full proof can be found in advanced calculus texts. To illustrate the idea, we
will show how the argument works for a spherical domain in R3 .
Example 2.7 Let B3 = {r < 1} R3 . Because a vector field can be decomposed

into components, it suffices to consider a field parallel to one of the coordinate axes,
say F = (0, 0, f ). The divergence is then
f
F= .
x3
In cylindrical coordinates, x = ( cos , sin , z), the volume element is
d 3 x = d d dz,
so the left side of the divergence formula becomes


1 2
2
f
Fd x = 3
r dz dr d
B3 0 1 2 z

2

1
= f cos , sin , 1 2 (2.11)
0 0

f cos , sin , 1 2 d d.

Note that z = 1 2 gives the restriction to the upper and lower hemispheres,
respectively.
We denote the two hemispheres S2 S2 and parametrize them as

(, ) = cos , sin , 1 2 .
The corresponding surface area elements are given by

d S = d d

= d d.
1 2
Thus,

2
1

f cos , sin , 1 d d =
2 f 1 2 d S.
0 0 S2

On S2 we have F = f 1 2 , so that

f 1 2 dS = F d S.
S2 S2
Applying this to (2.11) reduces the equation to

F d3 x = F d S,
B3 S2
verifying the divergence theorem in this special case.

Theorem 2.6 can be used to evaluate integrals of the Laplacian of a function by
substituting F = u for the vector field. Inside the volume integral this yields the
integrand
F = u.
On the surface side, the integrand becomes the directional derivative with respect to
the outward unit normal, which is denoted
22 2 Preliminaries
u
:= u .

Corollary 2.8 If Rn is a bounded domain with piecewise C 1 boundary, and

u C 2 (), then

u
u d n x = d S.

The application we will encounter most frequently is to the ball B(0; R) Rn .

The outward unit normal is parallel to the position vector, so that
x
= .
R
It follows from the chain rule that
u u
= . (2.12)
r
Example 2.9 Consider a radial function g(r ) where r := |x| for x Rn . For the
ball B(0; a), the radial integral formula (2.10) gives

a
g d x = An
n
g(r )r n1 dr,
B(0;a) 0
where
An := vol(Sn1 ). (2.13)
By (2.12),

g g
dS = (a) d S
B(0;a) B(0;a) r
g
= An a n1 (a).
r
The formula from Corollary 2.8 reduces in this case to

a
g
g(r )r n1 dr = a n1 (a). (2.14)
0 r
Differentiating (2.14) with respect to a gives, by the fundamental theorem of

calculus,
g
a n1 g(a) = a n1 (a) .
a r
This holds for all a > 0, so evidently the Laplacian of a radial function is given by

g = r 1n r n1 g. (2.15)
r r
In principle one could derive this formula directly from the chain rule, but the direct
computation is difficult in high dimensions.
There are two other direct corollaries of the divergence theorem that will be used
frequently. These are named for the mathematical physicist George Green, who used
them to develop solution formulas for some classical PDE.
The first result is a generalization of Corollary 2.8, obtained from Theorem 2.6
by the substitution F = vu for a pair of functions u, v. The product rule for
differentiation gives
(vu) = v u + vu, (2.16)
which can easily be checked by writing out the components of the gradient.
Theorem 2.10 (Greens first identity) If Rn is a bounded domain with piece-
wise C 1 boundary, then for u C 2 (), and v C 1 (),

u
[v u + vu] d x = n
v d S.

The second identity follows from the first by interchanging u with v and then
subtracting the result.
Theorem 2.11 (Greens second identity) If Rn is a bounded domain with
piecewise C 1 boundary, then for u, v C 2 (),

u v
[vu uv] d n x = v u d S.

2.7 Exercises
2.1 For r := |x| in Rn , and R, compute (r ) and (r ).

2.2 Polar coordinates (r, ) in R2 are related to Cartesian coordinates (x1 , x2 ) by
x1 = r cos , x2 = r sin .
(a) Use the chain rule to compute r and in terms of x1 and x2 .

(b) Find the expression for in the (r, ) coordinates. (The radial part should agree
with (2.15).)
2.3 In Rn let be the unit cube (0, 1)n . Define
w(x) = f (x)e j ,
24 2 Preliminaries
where f C () and e j is the jth coordinate vector, e j := (0, . . . , 1, . . . , 0).

Compute both sides of the formula from Theorem 2.6 in this case, and show that
the result reduces to an application of the fundamental theorem of calculus in the x j
variable.
2.4 For f C 0 (Rn ) set
h(t) := f (x) d n x.
B(0;t)
for t 0. Use the radial decomposition formula (2.10) to show that

dh
= f (w) d S(w).
dt B(0;t)
2.5 The gamma function is defined for z > 0 by

(z) := t z1 et dt. (2.17)
0
Note that (1) = 1 and integration by parts gives the recursion relation (z + 1) =
z(z). In this problem we will show that the volume of the unit sphere in Rn is given
by
n
2 2
An = . (2.18)
( n2 )
(a) Use the radial formula (2.10) and the substitution u := r 2 to compute that

1
er d n x =
2
An ( n2 ). (2.19)
Rn 2
(b) Observe that we can rewrite

er d n x = e(x1 ++xn ) d n x
2 2 2
Rn
R

n
n
ex d x .
2
= 2
0
Substitute t = x 2 to evaluate the one-dimensional integral in terms of ( 21 ).

(c) Compare (a) to (b) to obtain a formula for An .
(d) Use (c) and the fact that A2 = 2 to compute ( 21 ) and reduce the formula to
(2.18).
Chapter 3
Conservation Equations and Characteristics
A conservation law for a physical system states that a certain quantity (e.g., mass,
energy, or momentum) is independent of time. For continuous systems such as fluids
or gases, these global quantities can be defined as integrals of density functions. The
conservation law then translates into a local form, as a PDE for the density function.
In this section we will study some first-order PDE that arise from conservation
laws. We introduce a classic technique, called the method of characteristics, for
analyzing these equations.
3.1 Model Problem: Oxygen in the Bloodstream
To derive the conservation equation, we consider a simple model for the concentration
of oxygen carried by the bloodstream. For this discussion we ignore any external
effects that might break the conservation of mass, such as absorption of oxygen into
the walls of a blood vessel. (Some examples of external effects will be considered in
the exercises.)
Let us model an artery as a straight tube, as pictured in Fig. 3.1. We assume that the
concentration is constant on cross-sections of the tube, so that the problem reduces
to one spatial dimension. For the moment, suppose that the artery extends along the
real line and is parametrized by x R.
Let u(t, x) denote the oxygen concentration, expressed in units of mass per unit
length. Within a fixed interval [a, b], as highlighted in Fig. 3.1, the total mass at time
t is given by an integral,
b
m(t) := u(t, x) d x. (3.1)
a
The instantaneous flow rate at a given point x is called the flux q(t, x), expressed
as mass per unit time. The general relationship between flux and concentration is
flux = (concentration) (velocity).

26 3 Conservation Equations and Characteristics
x
a b
Fig. 3.1 One-dimensional model of an artery
For the bloodstream model we can reasonably assume that velocity is independent
of the oxygen concentration (because oxygen accounts for a relatively small portion
of the total density). This assumption implies that q has a linear dependence on u. In
other models the velocity might depend on the concentration, making q a nonlinear
function of u.
Conservation of mass implies that the total amount of oxygen within the segment
changes only as oxygen flows across the boundary points at x = a and x = b. Since
the flow across these points is given by the flux, the corresponding equation is
dm
(t) = q(t, a) q(t, b). (3.2)
dt
If q is continuously differentiable with respect to position, then the fundamental
theorem of calculus allows us to write the right-hand side of (3.2) as an integral,
b
q
q(t, a) q(t, b) = d x.
a x
We can also differentiate the integral in (3.1) to obtain

dm b
u
= d x,
dt a t
provided that u(t, x) is continuously differentiable with respect to time. These cal-
culations transform (3.2) into the integral equation
b
u q
+ d x = 0. (3.3)
a t x
Since the segment was arbitrary, (3.3) should hold for all values of a, b. This is
only possible if the integrand is identically zero, which gives the local form of the
law conservation of mass:
u q
+ = 0. (3.4)
t x
3.1 Model Problem: Oxygen in the Bloodstream 27
This relationship between concentration and flux is called the continuity equation
(or transport equation). The continuity equation applies generally to the physical
process of advection, which refers to the motion of particles in a bulk fluid flow.
To adapt (3.4) to a particular model, we need to specify the relationship between
q and u. As we remarked above, for the bloodstream model it is reasonable to assume
a linear relationship,
q = vu, (3.5)
where the velocity v(t, x) is part of the input data for the equation. Under this
assumption (3.4) reduces to
u u v
+v +u = 0, (3.6)
t x x
which is called the linear conservation equation.
3.2 Lagrangian Derivative and Characteristics
In this section we will discuss the strategy for solving a first-order PDE such as (3.6).
The basic idea is to adopt the perspective of an observer traveling with velocity v.
This is like taking measurements in a river from a raft drawn by the current. Once
we fix a starting point for the observer, the observed concentration depends only on
the time variable, thus reducing the equation to an ODE.
This principle applies to any first-order PDE of the form
u u
+v + w = 0, (3.7)
t x
where v = v(t, x) is independent of u. The zeroth-order term w could be a general

function w(t, x, u). A trajectory t x(t) is called a characteristic for the equation
(3.7) if
dx
(t) = v(t, x(t)). (3.8)
dt
For v and v/ x continuous, Theorem 2.4 shows that a unique solution exists in the
neighborhood of each starting point (t0 , x0 ).
Example 3.1 Suppose v(t, x) = at + b, with a and b constant. Integration over t

gives
a
x(t) = t 2 + bt + x0 .
2
The characteristics are a family of curves indexed by the parameter x0 , as illustrated
in Fig. 3.2.
Fig. 3.2 Sample x

characteristics for the
velocity v(t, x) = 1 + 2t
From the point of view of an observer carried by the flow, the measured concen-
tration is u(t, x(t)). The observed rate of change is the derivative of this quantity,
Du d
(t) := u(t, x(t)), (3.9)
Dt dt
called the Lagrangian derivative (or material derivative). This concept was devel-
oped by the 18th century mathematician and physicist Joseph-Louis Lagrange. Note
that Du/Dt depends also on the initial value (t0 , x0 ) that determines the character-
istic. For convenience we suppress the initial point from the notation.
Theorem 3.2 On each characteristic, (3.7) reduces to the ODE
Du
+ w = 0, (3.10)
Dt
where w is the restriction of w to the characteristic,

w(t)
:= w t, x(t), u(t, x(t) .
In particular, if w = 0 then u is constant on each characteristic.
Proof Applying the chain rule in (3.9) gives
Du u u d x
= + ,
Dt t x dt
with the understanding that the partial derivatives on the right are evaluated at the
point (t, x(t)). Because x(t) solves (3.8), this reduces to
Du u u
= +v , (3.11)
Dt t x
3.2 Lagrangian Derivative and Characteristics 29
If we restrict the variables in (3.7) to (t, x(t)), then the first two terms match the
right-hand side of (3.11), reducing the equation to (3.10).
If w = 0, then (3.10) becomes
Du
= 0.
Dt
This is equivalent to the statement that u(t, x(t)) is independent of t.
With Theorem 3.2 we can effectively reduce the PDE (3.7) to a pair of ODE,
namely the characteristic equation (3.8) and the Lagrangian derivative equation
(3.10). In many cases, solving these ODE will lead to an explicit formula for u(t, x).
This approach is referred to as the method of characteristics.
Example 3.3 For constants a, b R, assume that u(t, x) satisfies
u u
+ (at + b) = 0,
t x
with the initial condition
u(0, x) = g(x),
for some function g C 1 (R). The characteristics for this velocity, v(t, x) = at + b,
were computed in Example 3.1.
According to Theorem 3.2, u is constant along characteristics, implying that
a
u t, t 2 + bt + x0 = u(0, x0 ) = g(x0 ), (3.12)
2
for all t R. This is not yet a formula for u(t, x), but we can derive the solution
formula by identifying
a
x = t 2 + bt + x0 .
2
Solving for x0 in terms of x and substituting this into (3.12) gives
a
u(t, x) = g x t 2 bt .
2
Example 3.4 For steady flow through a pipe of changing diameter, the velocity would
vary with position rather than time. Let v(t, x) = a + bx for x 0, with a, b > 0.
The resulting characteristic equation (3.8) is
dx
= a + bx.
dt
This can be solved by the standard ODE technique of separating the t and x variables
to different sides of the equation:
dx
= dt.
a + bx
Integration of both sides gives the general solution
1
ln(a + bx) = t + C,
b
with C a constant of integration. (Note that a + bx > 0 by our assumptions.) Solving
for x gives
1 b(t+C)
x(t) = e a .
b
Given the assumption x 0, it is natural to index the characteristics by the start
time t0 such that x(t0 ) = 0. With this convention, the family of solutions is
a b(tt0 )
x(t) = e 1 . (3.13)
b
These characteristic curves are illustrated in Fig. 3.3.
With v = a + bx the linear conservation equation (3.6) becomes
u u
+ (a + bx) + bu = 0.
t x
Let us find the solution under the boundary condition
u(t, 0) = f (t). (3.14)
Since v/ x = b, (3.10) gives
Du
+ bu = 0.
Dt
This is a decay equation, with the family of exponential solutions
u(t, x(t)) = Aebt .

3.2 Lagrangian Derivative and Characteristics 31
Fig. 3.3 Characteristic lines x

for the position-dependent
velocity function of
Example 3.4
To fix A, we substitute the starting point (t0 , 0) into the equation and obtain
u(t, x(t)) = f (t0 )eb(tt0 ) . (3.15)
Putting together (3.13) and (3.15) and applying the boundary condition (3.14)
gives a

u t, eb(tt0 ) 1 = f (t0 )eb(tt0 ) . (3.16)
b
To express this as a function of (t, x), we set

a b(tt0 )
x= e 1 ,
b
and solve for t0 to obtain

1 a
t0 = t + ln .
b a + bx
Substituting this expression into (3.16) gives the final form of the solution:

a 1 a
u(t, x) = f t + ln .
a + bx b a + bx
A sample solution is illustrated in Fig. 3.4 for a = 1, b = 21 . For this example

the boundary condition f (t) was taken to have support between t = 1 and t = 1,
with a maximum at t = 0. The plots of u(t, x) on the right show concentrations at
a succession of times. Mass conservation is reflected in the fact that the total area
under each of these curves is independent of t.
Fig. 3.4 Behavior of solutions for Example 3.4. In the contour plot on the left, darker regions
correspond to higher concentration. The change in colors corresponds to exponential decay along
the characteristics illustrated in Fig. 3.3
3.3 Higher-Dimensional Equations
For flow problems in more than one spatial dimension, we can develop a continuity
equation analogous to (3.4) by the same reasoning as in Sect. 3.1. Suppose u(t, x)
represents a concentration defined for t R and x Rn . Let R Rn be a bounded
region with C 1 boundary. The total mass within this region is given by the volume
integral
m(t) := u(t, x) d n x.
R
The flow of u is represented by a vector-valued flux density q(t, x). The

interpretation of the flux density is that the rate at which mass passes through an
(n 1)-dimensional surface is given by the surface integral of q over this surface.
In particular, the rate at which mass exits R through the boundary is the quantity

q d S,
R
where is the outward unit normal vector defined on R.

Conservation of mass dictates that the mass within R can change only as mass
enters or leaves through the boundary. In other words,

dm
= q d S. (3.17)
dt R
Assuming that q is C 1 with respect to x, the Divergence Theorem (Theorem 2.6)

allows us to rewrite the flux integral as

q dS = q d n x. (3.18)
R R
3.3 Higher-Dimensional Equations 33
Note that since q depends on both t and x, the notation q is slightly ambiguous.
We follow the standard convention that vector calculus operators such as and
act only on spatial variables.
If u is C 1 with respect to t, then we can also differentiate the integral for m to
obtain
dm u n
= d x.
dt R t
Combining this with (3.17) and (3.18) gives

u
+ q d n x = 0. (3.19)
R t
As in the one-dimensional case, we now observe that since (3.19) holds for an arbi-
trary region R, the integrand must vanish. This is the higher-dimensional continuity
equation:
u
+ q = 0. (3.20)
t
Suppose we make the linear assumption that q = vu for a velocity field v which
is independent of u. The product rule for the divergence of a vector field is
(vu) = ( v)u + v u.
Substituting this into (3.20) gives the higher-dimensional form of the linear conser-
vation equation
u
+ v u + ( v)u = 0. (3.21)
t
In the special case where v = 0 the velocity field is called solenoidal (or
divergence-free). This situation arises frequently in applications, because incom-
pressible fluids like blood or water have solenoidal velocity fields.
The method of characteristics from Sect. 3.2 can be adapted directly to (3.21).
Consider a somewhat more general first-order PDE in the form
u
+ v u + w = 0, (3.22)
t
with v = v(t, x) and w = w(t, x, u). The characteristics associated to this equation
are by definition the solutions of
dx
(t) = v(t, x(t)). (3.23)
dt
Theorem 2.4 guarantees that characteristics exist in the neighborhood of each start-
ing point (t0 , x 0 ) provided v(t, x) and its partial derivatives with respect to x are
continuous.
The Lagrangian derivative of u along x(t) is defined as before by
Du d
(t) := u(t, x(t)).
Dt dt
The higher-dimensional version of Theorem 3.2 is the following:
Theorem 3.5 On each characteristic curve, the PDE (3.22) reduces to the ODE
Du
+ w = 0, (3.24)
Dt
where w denotes the restriction of w to the characteristic. In particular, if w = 0

then u is constant on each characteristic.
Proof By the chain rule,
Du u dx
(t) = (t, x(t)) + u(t, x(t)) (t).
Dt t dt
Since x(t) satisfies (3.23), this gives
Du u
= + v u.
Dt t
Substituting this into (3.22) reduces the equation to (3.24).
If w = 0 the equation becomes
Du
= 0,
Dt
which means precisely that u is constant along the characteristic curves.
Example 3.6 Consider a two-dimensional channel modeled as = R [1, 1]
with coordinates x = (x1 , x2 ). The velocity field
v(t, x) := (1 x22 , 0). (3.25)
is solenoidal and vanishes on the boundary {x2 = 1}. The characteristic line orig-
inating from (a, b) at t = 0 is

x(t) = a + (1 b2 )t, b .
Let us consider the conservation equation (3.21) for (t, x) R , with v given
by (3.25), subject to the initial condition
3.3 Higher-Dimensional Equations 35
t=0 t=3 t=6
Fig. 3.5 Evolution of a circle according to the two-dimensional flow in Example 3.6
u(0, x) = g(x),
for g C 1 (). Since v is solenoidal, Theorem 3.5 implies that u is constant on

characteristics. This gives the relation

u t, a + (1 b2 )t, b := g(a, b).
Rewriting this as a function of (t, x, y) gives

u(t, x, y) = g x (1 b2 )t, y .
Figure 3.5 illustrates the evolution of a circular ink spot distribution under this
flow. Conservation of mass is reflected in the fact that the area of the spot is inde-
pendent of t.
For applications of Theorem 3.5 on a bounded domain Rn , the specification

of boundary conditions can be quite a complicated problem, especially if the velocity
is time-dependent. (We avoided this problem in Example 3.6 by taking v tangent to
.) We will illustrate this issue in the exercises.
3.4 Quasilinear Equations
The method of characteristics remains an important tool for analysis of first-order

PDE even in the nonlinear case. In this section we will illustrate the application of
this method to the continuity equation (3.20) in the case of a flux term q that depends
on the concentration u.
To simplify the analysis, we assume that q = q(u), with no explicit dependence
on t and x. By the chain rule, (3.20) then reduces to the form
u
+ a(u) u = 0, (3.26)
t
where a(u) := dq/du. This type of PDE is called quasilinear, which means that the
equation is linear in the highest-order derivatives (which are merely first order in this
case).
A comparison of (3.26) to the linear conservation equation (3.21) shows that a(u)
is now playing the role of velocity. This suggests a definition for the characteristics,
but we must keep in mind that a(u) depends on t and x implicitly through u.
Theorem 3.7 Suppose that u C 1 ([0, T ] ) is a solution of (3.26) for some

region Rn , with a C 1 (R; Rn ). Then for each x 0 , u is constant along the
characteristic line defined by
x(t) = x 0 + a(u(0, x 0 ))t.
Proof Suppose that a solution u exists. Let x(t) be the solution to the ODE
dx
(t) = a(u(t, x(t))), x(0) = x 0 ,
dt
for t [0, T ]. Existence of such a characteristic is guaranteed by Theorem 2.4, at
least for t near 0, because the composition a u is C 1 as a function of (t, x) by the
assumptions on a and u.
To establish the claim that u(t, x(t)) is independent of t, we use the chain rule to
differentiate
d u dx
u(t, x(t)) = (t, x(t)) + u(t, x(t)) (t)
dt t dt
u
= (t, x(t)) + a(u(t, x(t))) u(t, x(t)).
t
The right-hand side vanishes by (3.26), so that
d
u(t, x(t)) = 0.
dt
This implies that
u(t, x(t)) = u(0, x 0 ),
which means that a(u(t, x(t))) is also constant. The characteristic equation reduces
to
dx
(t) = a(u(0, x 0 )),
dt
and we can integrate over t to compute x(t).
In contrast to the characteristic equation (3.8) in the linear case, the equation for
x(t) here depends on the initial condition u(0, x 0 ). Furthermore, it is important to
keep in mind that Theorem 3.7 does not imply that a solution to (3.26) exists; this
3.4 Quasilinear Equations 37
is assumed as a hypothesis. As we will see below, it is possible that the conclusion

of the theorem will lead to a contradiction, in the form of multiple values for the
solution at the same point. The implication in such a case is that a classical solution
does not exist.
To illustrate the application of Theorem 3.7, let us consider a simple model for
traffic on a single-lane road of infinite length, parametrized by x R. Let u(t, x)
denote the linear density of cars at a given point and time. Cars are discrete objects,
of course, but for modeling purposes we can assume that u is a C 1 function that
describes the density in an aggregate sense.
In traffic flow, the density of cars affects the flow velocity, with traffic slowing
down and possibly stopping as the density increases. A standard way to model this
effect is to set a maximum value for the velocity vm (presumably the speed limit).
The velocity is assumed to take its maximum value at u = 0 and decrease linearly
as u increases, up to some maximum value u m for which v = 0. In other words, for
this model u [0, u m ] and

u
v(u) := vm 1 .
um
Since v 0, the model always assumes that traffic moves to the right.
To eliminate the constants and focus on the equation itself, let us set vm = 1 and
u m = 1, reducing the velocity equation to
v(u) = 1 u
for u [0, 1]. The corresponding flux is
q(u) = u u 2 .
Substituting these assumptions into (3.26), we obtain a quasilinear equation called

the traffic equation:
u u
+ (1 2u) = 0. (3.27)
t x
Suppose we impose a general initial condition of the form
u(0, x) = h(x),
for some h : R [0, 1]. Assuming a solution exists, Theorem 3.7 gives the family
of characteristics
x(t) = x0 + (1 2h(x0 ))t. (3.28)
Therefore, the solution u must satisfy

u t, x0 + (1 2h(x0 ))t = h(x0 ). (3.29)
Fig. 3.6 Initial traffic

density modeling a line of h (x)
cars stopped at a traffic light
As we will demonstrate in the examples below, (3.29) leads to a solution formula for
some choices of h, while for others it leads to a contradiction.
Example 3.8 Figure 3.6 shows a plot of the initial condition
1 1
h(x) = arctan(20x),
2
which could represent a line of cars stopped at a traffic light at the point x = 0. The
corresponding characteristic lines as given by (3.28) are plotted in Fig. 3.7.
To derive a formula for u(t, x) from (3.29), we need to invert the equation
x = x0 + (1 2h(x0 ))t,
to express x0 as a function of t and x. For the function h given above it is not possible
to do this explicitly. However, there is a unique solution for each (t, x), which can
easily be calculated numerically. The resulting solutions are shown in Fig. 3.8.

Example 3.9 In order to solve the traffic equation explicitly, let us simplify the initial
condition to the piecewise linear function

1, x 0,
h(x) = 1 x, 0 < x < 1,

0, x 1.
This is not C 1 , but the resulting solution could be interpreted as a weak solution in
the sense described in Sect. 1.2. We will discuss the precise definition in Chap. 10.
By the formula from Theorem 3.7, the characteristic lines are

x0 t, x0 0,
x(t) = x0 + (2x0 1)t, 0 < x0 < 1, (3.30)

x0 + t, x 1.
Solving these equations for x0 gives

Fig. 3.7 Characteristic lines for the initial density shown in Fig. 3.6
t=0
t=2
u(t, x)
Fig. 3.8 Solutions for the traffic light problem

x + t, x t,
x0 = 1+2t
x+t
, t < x < 1 + t,

x t. x 1 + t.
Therefore, by the solution formula (3.29), the solution is

1, x t,
u(t, x) = 1 x+t
, t < x < 1 + t, (3.31)

1+2t
0. x 1 + t.
This is a continuous function, but differentiability fails on the lines x = t and

x = 1 + t. Away from these lines it is easy to check that u solves (3.27).
Despite the lack of smoothness, this solution is quite reasonable. To illustrate
this, let us trace the motion of a particular car starting from the position x0 0. The
velocity of the car is given by the flow rate v(u) = 1 u. The initial density at x0 is
u = 1, so the car is stationary for a time. According to (3.31), at t = x0 the value
of (t, x) enters the region where t < x < 1 + t and so at this time the density starts
to decrease and the car starts to move. For (t, x) in this range, (3.31) gives
Fig. 3.9 Trajectories of x

individual cars according to
the model of Example 3.9
x +t
v(t, x) = 1 u(t, x) = . (3.32)
1 + 2t
Let s(t) denote the position of the car at time t. For t x0 the velocity formula
(3.32) gives the equation
ds s+t
= , (3.33)
dt 1 + 2t
The initial condition at t = x0 is the original starting point s(x0 ) = x0 . The

standard ODE method of integrating factors can be used to solve (3.33), yielding

x0 , 0 t x0 ,
s(t) =
1 + t (1 2x0 )(1 + 2t), t x0 .
These trajectories are illustrated in Fig. 3.9. As we might expect, the cars further back
in the line wait longer before moving, but each car eventually moves forward and
gradually accelerates.
Example 3.10 Consider the initial condition
1 1
h(x) = + arctan(20x),
2
as shown in Fig. 3.10. This is the reverse of the initial condition of Example 3.8.
The characteristics specified in Theorem 3.7 now cross each other, as illustrated in
Fig. 3.11. The existence of crossings implies that a classical solution with this initial
condition cannot exist beyond the time of the first crossing.
If we were to trace the trajectories of individual cars, as we did in Example 3.9,
we would see that these also intersect each other at the points where characteristics
cross. In effect, the model predicts the formation of a traffic jam.
Fig. 3.10 Initial traffic

density with a h(x)
near-maximum density of
cars to the right
x
Fig. 3.11 Conflicting x

characteristic lines for the
initial density shown in
Fig. 3.10
A crossing of characteristics as observed in Example 3.10 is called a shock. After

the shock, the solution is forced to have discontinuities. The proper interpretation of
this situation requires weak solutions, for which discontinuities are allowed. We will
return to this issue in Chap. 10.
3.5 Exercises
3.1 Consider the conservation equation with a constant velocity c > 0,
u u
+c = 0,
t x
on the quadrant t 0, x 0. Suppose the boundary and initial conditions are

u(0, x) = g(x), x 0,
u(t, 0) = h(t), t 0,
for g, h C 1 [0, ).
(a) Find a formula for the solution u(t, x) in terms of g and h.

(b) Find a matching condition for g and h that will ensure that u(t, x) is a C 1
function.
3.2 In the continuity equation (3.4), external factors that break the conservation of
mass are accounted for by adding terms to the right-hand side.
(a) A forcing term f (t, x) is independent of the existing concentration. (In the
bloodstream model of Sect. 3.1, this could represent intravenous injection, for
example.) Assume that c is constant, f C 1 (R2 ), and g C 1 (R). Solve the
equation
u u
+c = f, u(0, x) = g(x),
t x
to find an explicit formula for u(t, x) in terms of f and g.

(b) A reaction term depends on the concentration u. The simplest case is a linear
term u where the coefficient is some function (t, x). (This could represent
absorption of oxygen into the walls of the artery, for example.) Assume that c is
constant, C 1 (R2 ), and g C 1 (R). Solve the equation
u u
+c = u, u(0, x) = g(x),
t x
to find an explicit formula for u(t, x) in terms of and g.
3.3 Assume that u satisfies the linear conservation equation
u u
+ 2t = 0,
t x
for t R and x [0, 1]. Suppose the boundary conditions are given by
u(t, 0) = h 0 (t), u(t, 1) = h 1 (t).
Find a relation between h 0 and h 1 . (This shows that we can only impose a boundary
condition at one side of the interval [0, 1].)
3.4 If the spatial domain in the linear conservation equation (3.21) is a bounded
region Rn , then for a given velocity field v, the inflow boundary in
is defined as the set of boundary points where v points into . Fixing boundary
conditions on the inflow boundary will generally determine the solution in the interior.
Suppose = (1, 1) (1, 1) R2 with coordinates (x1 , x2 ). For the velocity
fields below, determine the characteristics and specify the inflow boundary. Draw a
sketch of for each case, indicating these features.
3.5 Exercises 43
(a) v(x1 , x2 ) = (x2 , 1).

(b) v(x1 , x2 ) = (1, x2 ).
3.5 Suppose that a section of of a river is modeled as a rectangle = (0, )

(0, 1) R2 , parametrized by (x1 , x2 ). Assume the flow is parallel to the x1 -axis,
with velocity
v(x1 , x2 ) = ( f (x2 ), 0),
for some positive function f on (0, 1). Assume also that the concentration on the
left boundary {x1 = 0} is given by
u(t, 0, x2 ) = h(t, x2 ).
Find a formula for u(t, x1 , x2 ) in terms of the functions h and f .
3.6 Burgers equation is a simple quasilinear equation that appears in models of gas
dynamics,
u u
+u = 0.
t x
(a) Use the method of characteristics as described in Sect. 3.4 to find a formula for
the solution u(t, x) given the initial condition

0, x 0,
u(0, x) = ax , 0 < x < a,

1, x a.
(b) Suppose a > b and

a, x 0,
u(0, x) = a(1 x) + bx, 0 < x < 1,

b, x 1.
Show that all of the characteristics originating from x0 [0, 1] meet at the same
point (thus creating a shock).
3.7 In the mid-19th century, William Hamilton and Carl Jacobi developed a formu-
lation of classical mechanics based on ideas from geometric optics. In this approach
the dynamics of a free particle in R are described by a generating function u(t, x)
satisfying the Hamilton-Jacobi equation:
2
u 1 u
+ = 0. (3.34)
t 2 x
Assume that u C 1 ([0, ) Rn ) is a solution of (3.34). By analogy with Theo-

rem 3.7, a characteristic of (3.34) is defined as a solution of
dx u
(t) = (t, x(t)), x(0) = x0 . (3.35)
dt x
(a) Assuming that x(t) solves (3.35), use the chain rule to compute d 2 x/dt 2 .
(b) Differentiate (3.34) with respect to x and then restrict the result to (t, x(t)),
where x(t) solves (3.35). Conclude from (a) that to
d2x
= 0.
dt 2
Hence, for some constant v0 (which depends on the characteristic),
x(t) = x0 + v0 t.
(c) Show that the Lagrangian derivative of u along x(t) satisfies
Du 1
= v02 ,
Dt 2
implying that
1
u(t, x0 + v0 t) = u(0, x0 ) + v02 t.
2
(d) Use this approach to find the solution u(t, x) under the initial condition
u(0, x) = x 2 .
(For the characteristic starting at (0, x0 ), note that you can compute v0 by eval-
uating (3.35) at t = 0.)
Chapter 4
The Wave Equation
As we noted in Sect. 1.2, dAlemberts derivation of the wave equation in the 18th
century was an early milestone in the development of PDE theory. In this chapter
we will develop this equation as a model for the vibrating string problem, and derive
dAlemberts explicit solution in one dimension using the method of characteristics
introduced in Chap. 3.
In higher dimensions the wave equation is used to model electromagnetic or
acoustic waves. We will discussion the derivation of the acoustic model later in
Sect. 4.5. A clever reduction trick allows the solution formula for Rn to be deduced
from the one-dimensional case. The resulting integral formula yields insight into the
propagation of waves in different dimensions.
The chapter concludes with a discussion of the energy of a solution, based on the
physical principles of kinetic and potential energy.
4.1 Model Problem: Vibrating String
Consider a flexible string that is stretched tight between two points, like the strings
on a violin or guitar. The stretching of the string creates a tension force T that pulls
in both directions at each point along its length. For simplicity, let us assume that
any other forces acting on the string, including gravity, are negligible compared to
the tension. The linear density of mass is taken to be constant along the string.
For a violin string it is also reasonable to assume that the displacement of the
string is extremely small relative to its length. This assumption justifies taking T to
be a fixed constant, ignoring the additional stretching that occurs when the string
is displaced. It also allows us to treat horizontal and vertical components of the
displacement independently, so we can restrict our attention to the vertical.
Let the string be parametrized by x [0, ]. The vertical displacement as a
function of time is denoted by u(t, x). To develop an equation for u, we first discretize
the model by subdividing the total length into segments of length x = /n for

46 4 The Wave Equation
some large n. Each segment has a mass x and is subject to the tension forces
pulling in the direction of its neighbors on either side.
For j = 0, . . . n, let x j := jx be the position of the jth segment along the string.
The segments j = 0 and j = n represent the fixed endpoints, with j = 1, . . . , n 1
in the interior. Let u(t, x j ) denote the vertical displacement of the jth segment as a
function of time. Figure 4.1 illustrates this discretization (with displacements greatly
exaggerated).
To develop an equation for the string, we apply Newtons laws of motion to the
segments of the discretization, as if they were single particles. The jth particle is
being pulled by its neighbors with a force T on each side. Unless the string is straight,
these forces are not quite aligned.
In terms of the angles labeled in Fig. 4.2, the net vertical force on a single segment
is
F(t, x j ) = T sin j + T sin j .
We have assumed that the relative displacements are extremely small, so the angles
j , j will be very small also. To leading order, we can replace the sines by tangents,
which are linear in u,
u(t, x j1 ) u(t, x j ) u(t, x j ) u(t, x j+1 )

sin j , sin j .
x x
With this linear approximation, the net vertical force at the point x j becomes
T
F(t, x j ) = u(t, x j+1 ) + u(t, x j1 ) 2u(t, x j ) . (4.1)
x
Fig. 4.1 Discrete model for the displacement of the string
Fig. 4.2 Discrete model for the displacement of the string

4.1 Model Problem: Vibrating String 47
The equation of motion for the jth segment now comes from Newtons law: mass
times acceleration equals force. At the point x j this translates to
2u
x (t, x j ) = F(t, x j ). (4.2)
t 2
Using (4.1) on the right then gives
2u T u(t, x j+1 ) + u(t, x j1 ) 2u(t, x j )

(t, x j ) = . (4.3)
t 2 (x)2
The final step is to take the continuum limit n and x 0. Assuming

that u is twice continuously differentiable as a function of x, we can deduce from
the quadratic Taylor approximation of u(t, x) that
u(t, x + x) + u(t, x x) 2u(t, x) 2u

lim = (t, x).
x0 (x)2 x 2
Hence, taking x 0 in (4.3) gives
2u T 2u
= 0. (4.4)
t 2 x 2
This is the one-dimensional wave equation. The fixed ends of the string correspond
to Dirichlet boundary conditions,
u(t, 0) = u(t, ) = 0.
4.2 Characteristics
For convenience, set c2 := T / in (4.4), assuming c > 0, and rewrite the equation
as
2u 2 u
2
c = 0. (4.5)
t 2 x 2
The constant c is called the propagation speed, for reasons that will become apparent
as we analyze the equation.
Let the physical domain be x R for the moment; we will discuss boundary
conditions later. The key to applying the method of characteristics to (4.5) is that the
differential operator appearing in the equation factors as a product of two first-order
operators, i.e.,
2 2
2

c = +c c . (4.6)
t 2 x 2 t x t x
Individually, these operators have characteristic lines t x0 ct. Both sets of

characteristics will play an important role here.
Theorem 4.1 Under the initial conditions

u
u(0, x) = g(x), (0, x) = h(x), (4.7)
t
for g C 2 (R) and h C 1 (R), the wave equation (4.5) admits a unique solution

1 1 x+ct
u(t, x) = g(x + ct) + g(x ct) + h( ) d . (4.8)
2 2c xct
Proof Consider the auxiliary function w(t, x) defined by
u u
w := c . (4.9)
t x
By (4.6), w satisfies the linear conservation equation
w w
+c = 0.
t x
The characteristics for this equation are given by x+ (t) = x0 + ct. By Theorem 3.2
the unique solution with an initial condition w(0, x) = w0 (x) is
w(t, x) = w0 (x ct). (4.10)
We will relate w0 back to the initial conditions g and h in a moment.

With w given by (4.10), the definition (4.9) can be regarded as a linear conservation
equation for u,
u u
c = w, (4.11)
t x
where w acts as a forcing term as described in Exercise 3.2. The characteristics of
(4.11) are x (t) = x0 ct. By Theorem 3.2, we can thus reduce the equation to the
form
d
u(t, x0 ct) = w(t, x0 ct). (4.12)
dt
The unique solution to (4.12) under the initial condition u(0, x) = g(x) is given
by direct integration with respect to time:
t
u(t, x0 ct) = g(x0 ) + w(s, x0 cs) ds.
0
4.2 Characteristics 49
Setting x = x0 ct then gives

t
u(t, x) = g(x + ct) + w(s, x c(s t)) ds.
0
Using the formula (4.10) for the solution w, we obtain

t
u(t, x) = g(x + ct) + w0 (x 2cs + ct) ds.
0
As a final step, the substitution := x + ct 2cs gives

x+ct
1
u(t, x) = g(x + ct) + w0 ( ) d . (4.13)
2c xct
The function w0 can be computed from the initial conditions (4.7),
u u
w0 (x) := (0, x) c (0, x)
t x
g
= h(x) c (x).
x
The w0 contribution to (4.13) is then given by
x+ct
1 x+ct
1 1 x+ct g
w0 ( ) d = h( ) d ( ) d
2c xct 2c xct 2 xct x
x+ct
1 1
= h( ) d g(x + ct) g(x ct) .
2c xct 2
Substituting back into (4.13) now gives the formula (4.8).
To highlight the role played by the characteristic lines in the solution of Theo-
rem 4.1, consider the functions
x
1 1
u (x) := g(x) h( ) d .
2 2c 0
In terms of u , the solution (4.8) simplifies to
u(x, t) = u + (x ct) + u (x + ct), (4.14)
matching the form of the solution stated in (1.3). The subscripts in u indicate the
propagation direction, i.e., u + propagates to the right and u to the left. In either
direction the speed of propagation is the parameter c.
Fig. 4.3 Evolution of a

solution to the wave equation
Example 4.2 Consider the wave equation (4.5) with the initial conditions h(x) = 0
and
(1 x 2 )2 , |x| 1,
g(x) =
0, |x| > 1.
By (4.8) the solution is the superposition of two localized bumps, which propagate
in opposite directions as illustrated in Fig. 4.3.
In Example 4.2 the initial condition was supported in [1, 1], and we can see
in Fig. 4.3 that the resulting solution has support in a V-shaped region. This region
could be identified as the span of the characteristic lines emerging from the initial
support interval.
This restriction of the support of a solution is closely related to Huygens principle,
an empirical law for propagation of light waves published by Christiaan Huygens
in 1678. The one-dimensional wave equation exhibits a special, strict form of this
principle:
Theorem 4.3 (Strict Huygens principle in dimension one) Suppose u solves the
wave equation (4.5) for t 0, x R, with initial data given by (4.7). If the functions
g, h are supported in a bounded interval [a, b], then

supp u (t, x) R+ R; x [a ct, b ct] [a + ct, b + ct] .
Proof Consider the components of the solution (4.8). The g term will vanish unless
x ct [a, b]. In the h term, the integral over will vanish unless the interval
[x ct, x + ct] intersects [a, b]. This again requires that x ct [a, b].
The restriction of support described in Theorem 4.3 is illustrated in Fig. 4.4. At a
fixed observation point, the solution passes through in some finite time interval and
then vanishes forever.
Example 4.4 Suppose the initial data from Example 4.2 are altered to include a
singularity at x = 0. For example,
4.2 Characteristics 51
Fig. 4.4 Support of a wave

solution with initial data in a
bounded interval
Fig. 4.5 Propagation of

singularities of the wave
equation along characteristic
lines

(1 |x|)2 , |x| 1,
g(x) =
0, |x| 1.
Then (4.8) still gives a formula for the solution even though g is not differentiable.
(This is a weak solution in the sense we will describe in Chap. 10). A set of solutions
at different points in time is plotted in Fig. 4.5. Observe that the original singularity
splits into two singularities, which propagate outward along the two characteristic
lines emanating from x = 0.
4.3 Boundary Problems
In the string model of Sect. 4.1 the domain of the wave equation (4.5) was restricted
to x [0, ], with Dirichlet boundary conditions
u(t, 0) = u(t, ) = 0, for all t 0. (4.15)
Suppose the initial data are given for x [0, ] by
u
u(0, x) = g(x), (0, x) = h(x), (4.16)
t
with g C 2 [0, ], h C 1 [0, ]. Both g and h are assumed to vanish at the endpoints
of [0, ].
The solution of the wave equation on R provided in Theorem 4.1 can be adapted
to the boundary conditions (4.15). The idea is to extend g, h to R in such a way that
the formula (4.8) gives a solution satisfying the boundary conditions for all t.
Theorem 4.5 The unique solution of the wave equation (4.5) on [0, ], with Dirichlet
boundary conditions and satisfying the initial conditions (4.16), is given by (4.8),
provided g and h are extended to R as odd, 2-periodic functions.
Proof By linearity we can consider the g and h terms independently. Assume that
the g term,
1
g(x + ct) + g(x ct) , (4.17)
2
is defined for all t and x and satisfies the boundary conditions on [0, ] for all values
of t. At x = 0 the condition u(t, 0) = 0 will be satisfied if and only if
g(ct) + g(ct) = 0, for all t 0.
In other words, u(t, 0) = 0 if and only if g is odd. At x = the condition is
g( + ct) + g( ct) = 0, for all t 0.
This is equivalent to the condition that g is odd with respect to reflection at the point
x = .
The composition of the reflections about 0 and gives translation by 2. Hence
the expression (4.17) satisfies the boundary conditions if and only if g is odd and
2-periodic.
A similar argument works for the h term,
x+ct
1
u(t, x) = h( ) d . (4.18)
2c xct
The requirement at x = 0 is
ct
h( ) d = 0, for all t 0. (4.19)
ct
Differentiation with respect to t, using the fundamental theorem of calculus, shows

that (4.19) is satisfied if and only if h is odd with respect to reflection at 0. Similarly,
the condition +ct
h( ) d = 0, for all t 0
ct
requires odd symmetry with respect to reflection at x = .

4.3 Boundary Problems 53
Example 4.6 Consider the vibrating string problem with c = 1 and = 1. Suppose
that the solution initially has the form (4.14) with u + = 0 and the left-propagating
solution given by the function u shown in Fig. 4.6. For small t > 0 the solution is
u(t, x) = u (x + t), (4.20)
but eventually the bump hits the boundary at x = 0, and we would like to understand
what happens then.
To apply Theorem 4.5, we must first solve for g and h in terms of u + . By (4.20)
we set g(x) = u + (x) and
du +
h(x) = u + (x + t)t=0 = (x).
t dx
The resulting functions g and h, extended to odd functions on R, are shown in Fig. 4.7.
According to Theorem 4.5 we can compute the solution from (4.8) using these
odd periodic extensions of g and h. The results are shown in Fig. 4.8. The bump
temporarily disappears at t = 0.3 and then reemerges as an inverted bump traveling
in the opposite direction.
4.4 Forcing Terms
The derivation of the string model in Sect. 4.1 assumed that no external forces act
on the string. Additional forces could be incorporated by adding extra terms to the
expression (4.1) for the force on a segment. In the continuum limit this yields a
Fig. 4.6 The initial waveform u +
Fig. 4.7 The odd extensions of the initial conditions g and h

Fig. 4.8 Reflection of a propagating bump at the endpoint of the string
forcing term on the right-hand side of (4.5):
2u 2 u
2
c = f, (4.21)
t 2 x 2
where f = f (t, x). The forcing term could be used to model plucking or bowing of
the string, for example.
In this section we introduce a technique, called Duhamels method, that allows us
to adapt solution methods for evolution equations to include a forcing term. The idea,
which is closely related to a standard ODE technique called variation of parameters,
is to reformulate the forcing term as an initial condition. This technique is named
for the 19th century French mathematician and physicist Jean-Marie Duhamel, who
developed the idea in a study of the heat equation.
4.4 Forcing Terms 55
Fig. 4.9 Domain of

dependence for the point
(t, x)
To focus our attention on the driving term, let us consider (4.21) on the domain
x R with the initial conditions set to zero. For a given c, define the domain of
dependence of a point (t, x) with t > 0 and x R by

Dt,x := (s, x ) R+ R : x c(t s) x x + c(t s) .
This is a triangular region, as pictured in Fig. 4.9. The terminology refers to the fact
that the solution u(t, x) is influenced only by the values of f within Dt,x , as the next
result shows.
Theorem 4.7 For f C 1 (R), the unique solution of (4.21) satisfying the initial
conditions
u
u(0, x) = 0, (0, x) = 0,
t
is given by
1
u(t, x) = f (s, x ) d x ds. (4.22)
2c Dt,x
Proof For each s 0, let s (t, x) be the solution of the homogeneous wave equation
(4.5) for t s, subject to the initial conditions
s
s (t, x)t=s = 0, (t, x)t=s = f (s, x). (4.23)
t
This function can be written explicitly by shifting t to t s in (4.8),
x+c(ts)
1
s (t, x) = f (s, x ) d x . (4.24)
2c xc(ts)
We claim that the solution of (4.21) is given by the integral

t
u(t, x) := s (t, x) ds. (4.25)
0
Note that the integration variable here is s rather than t.

We will first check that this definition of u satisfies the initial conditions. For
t = 0 the integral in (4.25) clearly vanishes, so that u(0, x) = 0 is satisfied. By the
fundamental theorem of calculus, differentiating (4.25) with respect to t gives

u t
s
(t, x) = s (t, x)s=t + (t, x) ds.
t 0 t
The first term vanishes for all t by the initial condition (4.23), leaving

u t
s
(t, x) = (t, x) ds, (4.26)
t 0 t
for all t 0. Setting t = 0 gives
u
(0, x) = 0.
t
Now let us check that the u defined in (4.25) solves (4.21). Differentiating (4.26)
once more gives
t 2
2u s s
(t, x) = (t, x) + (t, x) ds. (4.27)
t 2 t s=t 0 t
2
By (4.23) the first term on the right is equal to f (t, x). To simplify the second term,
we use the fact that s solves (4.5) and the definition of u to compute

t
2 s t
2 s
(t, x) ds = c2 (t, x) ds
0 t 2 0 x 2
2u
= c2 2 .
x
Therefore, (4.27) reduces to
2u 2 u
2
= f + c ,
t 2 x 2
proving that u solves (4.21).
Combining (4.24) and (4.25), we can write the formula for u as
t x+c(ts)
1
u(t, x) = f (s, x ) d x ds,
2c 0 xc(ts)
which is equivalent to (4.22).

4.4 Forcing Terms 57
Fig. 4.10 Range of

influence of the point (t0 , x0 )
To prove uniqueness, suppose u 1 and u 2 are solutions of (4.21). Then u 2 u 1 is

a solution of (4.5). Since u 2 u 1 also has vanishing initial conditions, Theorem 4.1
implies that u 2 u 1 = 0. Hence the solution is unique.
By the superposition principle, Theorem 4.7 is easily extended to the case of

nonzero initial conditions, by setting u = v + w where v is a solution of the form
(4.22) and w is a solution of the form (4.8).
The concept of domain of dependence still applies when the initial conditions are
nonzero. In the solution formula (4.8), u(t, x) depends only on the values of g and
h at the base of the triangle, Dt,x {t = 0}. Thus it is still the case that the solution
u(t, x) depends only on the data within Dt,x .
The existence of the domain of dependence is a limitation imposed by the prop-
agation speed c. For systems governed by the wave equations (4.5) or (4.21), no
information can travel at a speed faster than c.
The region of the space-time plane in which solutions can be affected by the data
at a particular point (t0 , x0 ) is called the range of influence of this point. By the
definition of the domain of dependence, the range of influence consists of the points
(t, x) such that (t0 , x0 ) Dt,x . This region is a triangle with vertex (t0 , x0 ) and sides
given by the characteristics (t, x0 c(t t0 )), as shown in Fig. 4.10.
Duhamels method applies also to the case of a vibrating string with fixed ends.
Assuming that f (t, x) satisfies the boundary conditions at x = 0 and x = , we
extend f to an odd 2-periodic function on R, just as in Theorem 4.5. This exten-
sion guarantees that the intermediate solution s defined by (4.24) will satisfy the
boundary conditions also. And then so will the solution u(t, x) defined by (4.25).
Example 4.8 Consider a string of length with propagation speed c = 1. Suppose

the forcing term is given by
f (t, x) = cos(t) sin(0 x), (4.28)
where 0 := / and > 0 is the driving frequency. Since sin(0 x) is odd and
2-periodic, the extension required by Theorem 4.5 is automatic. As in Theorem 4.7,
let us set the initial conditions g = h = 0 to focus on the forcing term.
Substituting (4.28) into (4.22) gives
t x+ts
1
u(t, x) = cos(s) sin(0 x ) d x ds
2 0 xt+s
t
1
= [cos(0 (x t + s)) cos(0 (x + t s))] cos(s) ds
20 0
A trigonometric identity reduces this to

t
sin(0 x)
u(t, x) = sin(0 (t s)) cos(s) ds. (4.29)
0 0
For = 0 we obtain
sin(0 x)
u(t, x) = [cos(t) cos(0 t)] .
02 2
Note that the x dependence of the solution matches that of the forcing term. The
interesting part of this solution is the oscillation, which includes both frequencies
and 0 . Figure 4.11 illustrates the behavior of the amplitude as a function of time, in
a case where 0 . The large-scale oscillation has a period 1/, corresponding to
the low driving frequency. The solution also exhibits fast oscillations at the frequency
0 which depends only on .
For = 0 the formula (4.29) gives the solution
t
u(t, x) = sin(0 x) sin(0 t).
20
The resulting amplitude grows linearly, as shown in Fig. 4.12.
The physical phenomenon illustrated by Example 4.8 is called resonance. If the

string is driven at its natural frequency 0 then it will continually absorb energy from
the driving force. Of course, there is a limit to how much energy a physical string
could absorb before it breaks. Once the displacement amplitude becomes sufficiently
large, the linear wave equation (4.5) no longer serves as an appropriate model.
Fig. 4.11 Oscillation pattern

with a driving frequency
= 0 /10
4.5 Model Problem: Acoustic Waves 59
Fig. 4.12 Growth of the

amplitude at the resonance
frequency = 0
4.5 Model Problem: Acoustic Waves
The vibration of a drumhead can be modeled on a bounded domain R2 , with

a function u(t, x) representing the vertical displacement of the membrane at time t
and position x . With arguments similar to those in Sect. 4.1, one can derive the
equation
2u
c2 u = 0, (4.30)
t 2
where is the Laplacian operator (1.7). The wave equation (4.30) appears in many
other contexts as well, including the propagation of light and all other forms of
electromagnetic radiation. In all these cases the constant c represents the speed of
propagation.
In this section we will derive the three-dimensional wave equation as a model
for acoustic waves traveling through the air. Acoustic waves consist of fluctuations
of pressure which propagate through a gas. To analyze them, we must consider the
relationships between the pressure P, the velocity field v, and the density . For a
gas in motion these are all functions of both time and position.
Because acoustic waves involve minute pressure fluctuations with very little heat
transfer, the relationship between pressure and density is given by the adiabatic gas
law
P = C , (4.31)
where C and are physical constants. We will fix background atmospheric values
of the pressure P0 and density 0 and focus on the deviations
u := P P0 , := 0 .
Applying (4.31) to P/P0 gives the equation

u
1+ = 1+ .
P0 0
Since /0 is assumed to be very small, we can linearize by taking a first-order

Taylor approximation on the right side. This yields
P0
u= . (4.32)
0
The dynamics of the gas are modeled with two conservation laws. The first is
conservation of mass (3.20), which yields

+ (v) = 0.
t
Since and v are both assumed to be very small, for the leading approximation we
can replace by 0 to obtain

+ 0 v = 0. (4.33)
t
The second dynamical law is conservation of momentum. This is encapsulated in
a fluid equation derived by Euler in 1757, called Eulers force equation:

P = + v v.
t
Eulers equation is an aggregate form of Newtons second law (force equals mass
times acceleration). Note that the acceleration term on the right is the Lagrangian
derivative of the velocity field v. As above, we substitute P = P0 + u and = 0 +
and keep only the first order terms to derive the linearization
v
u = 0 . (4.34)
t
The final step is to eliminate the velocity field from the equation. Substituting
(4.32) into (4.33) and differentiating with respect to time gives
2u
= P0 ( v). (4.35)
t 2 t
On the other hand, by (4.34),
v
( v) =
t t
u
=
0
1
= u.
0
4.5 Model Problem: Acoustic Waves 61
Substituting this in (4.35) yields the acoustic wave equation,
2u P0
u = 0.
t 2 0
As with our previous derivations, many approximations are required to produce a

linear equation. These simplifications are well justified for sound waves at ordinary
volume levels, but more dramatic pressure fluctuations would require a nonlinear
equation.
4.6 Integral Solution Formulas
Let us consider the wave equation (4.30) on R3 with c = 1,
2u
u = 0.
t 2
This problem can be reduced to the one-dimensional case by a clever averaging trick.
For f C 0 (R3 ), define

1
f(x; ) := f (w) d S(w), (4.36)
4 B(x;)
where x R3 and > 0. The surface area of B(x; ) is 42 , so f is times the

spherical average of f . By continuity the spherical average approaches the value of
the function at the center point as 0, so that
f(x; )
lim = f (x). (4.37)
0
The dimensional reduction of the wave equation is based on the following formula
of Jean-Gaston Darboux.
Lemma 4.9 (Darbouxs formula) For f C 2 (R),
2
f (x; ) = x f(x; ).
2
Proof To compute the radial derivative of the spherical average, it is helpful to change
coordinates by setting w = x + y, so that the domain of y is the unit sphere S2 R3 ,

1 1
f (w) d S(w) = f (x + y) d S( y).
42 B(x;) 4 S2
Differentiation under the integral gives,

1
f (x + y) d S( y) = f (x + y) y d S( y).
S2 4 S2
In the original coordinates this implies

1
f (w) d S(w)
42 B(x;)
(4.38)
1 wx
= f (w) d S(w).
42 B(x;)
Since (w x)/ is the outward unit normal to B(x; ),

wx f
f (w) = (w).

Furthermore, by Corollary 2.8,

f
(w) d S(w) = f (w) d 3 w.
B(x;) B(x;)
Applying this to the right-hand side of (4.38) gives

1 1
f (w) d S(w) = f (w) d 3 w. (4.39)
42 B(x;) 42 B(x;)
Substituting the definition of f in (4.39) yields

1 1
f (x; ) = f (w) d S(w) + f (w) d 3 w
42 B(x;) 4 B(x;)
A further differentiation using (4.39) and the radial derivative formula from Exer-
cise 2.4 then gives

2 1
f (x; ) = f (w) d 3 w
2 4 B(x;)
(4.40)
1
= f (w) d S(w).
4 B(x;)
On the other hand

4.6 Integral Solution Formulas 63

1
x f(x; ) = x f (x + y) d S( y)
4 S2

1
= f (x + y) d S( y) (4.41)
4 S2

1
= f (w) d S(w).
4 B(x;)
The claim thus follows from (4.40).
Lemma 4.9 allows us to relate the three-dimensional wave equation in variables

(t, x) to a one-dimensional equation in variables (t, ). The result is a solution
formula first derived in 1883 by the physicist Gustav Kirchhoff.
Theorem 4.10 (Kirchhoffs integral formula) For u C 2 ([0, ) R3 ), suppose

that
2u
u = 0
t 2
under the initial conditions
u
u|t=0 = g, = h.
t t=0
Then

u(t, x) =
g(x; t) + h(x; t),
t
with g and h defined as in (4.36).
Proof Define
1
x; ) :=
u(t, u(t, w) d S(w).
4 B(x;)
Since u satisfies the wave equation, differentiating under the integral gives

2 1
x; ) =
u(t, u(t, w) d S(w).
t 2 4 B(x;)
By the calculation (4.41) this is equivalent to
2
x; ) = x u(t,
u(t, x; ).
t 2
Lemma 4.9 then shows that

2 2
2 x; ) = 0.
u(t, (4.42)
t 2
The initial conditions for u follow from the initial conditions for u,

u(0, x; ) = g(x;
),
u(0, x; ) = h(x; ).
t
By (4.37) we also have a boundary condition at = 0,
x; 0) = 0.
u(t,
Using Theorem 4.1 and the reflection argument from Theorem 4.5, we conclude

that the unique solution of (4.42) under these conditions is given by extending g(x; )

and h(x; ) to R with odd symmetry and then using the dAlembert formula,
+t
1 1

x; ) =
u(t,
g(x; + t) + g(x;
t) + h(x; ) d . (4.43)
2 2 t
By (4.37), we can recover u from this formula by setting
x; )
u(t,
u(t, x) = lim . (4.44)
0
To evaluate this limit, first note that for 0 t the odd symmetry of g and h with
respect to can be used to rewrite (4.43) as

1 1 t+

x; ) =
u(t,
g(x; t + ) g(x;
t ) + h(x; ) d .
2 2 t
The computations are now straightforward:
1
lim
g(x; t + ) g(x;
t ) = g(x;
t),
0 2 t
and t+
1
lim h(x; ) d = h(x; t).
0 2 t
The claimed solution formula thus follows from (4.44).

One interesting consequence of the Kirchhoff formula is the fact that three-
dimensional wave propagation exhibits a strict form of the Huygens principle. The-
orem 4.10 shows that the range of influence of the point (t0 , x 0 ) is the forward light
cone,
+ (t0 , x 0 ) := {(t, x); t > t0 , |x x 0 | = t t0 } .
This matches the result of Theorem 4.3 for the one-dimensional wave equation. The
strict Huygens phenomenon is readily observable for acoustic waves, in the fact that
Fig. 4.13 The plot on the right shows the observed waveform at a distance 3 from the origin, caused
by the initial radial impulse shown on the left
a sudden sound like a clap propagates as a sharp wavefront that is heard as a single
discrete event, without aftereffects unless there are reflective surfaces to cause an
echo. Figure 4.13 illustrates this effect; an observer located away from the origin
experiences a waveform of duration equal to the diameter of the initial impulse. The
strict Huygens principle holds in every odd dimension, but fails in even dimensions,
as we will illustrate below.
The spherical averaging trick used for Theorem 4.10 also works in higher odd
dimensions, although the solution formulas become more complicated. For even
dimensions, solution formulas can be derived from the odd-dimensional case by a
technique called the method of descent.
We will work this out for the two-dimensional case. Suppose u C 2 ([0, )R2 )
solves the wave equation with initial conditions
u
u|t=0 = g, = h,
t t=0
with g, h functions on R2 . If we extend g and h to R3 as functions that are independent

of x3 , then Kirchhoffs formula gives a solution to the three dimensional problem.
Since this solution is also independent of x3 , it descends to a solution in R2 . The
resulting formula was first worked out by Simon Poisson in the early 19th century
(well before Kirchhoffs three-dimensional formula).
Corollary 4.11 (Poissons integral formula) For u C 2 ([0, )R2 ), suppose that
2u
u = 0
t 2
under the initial conditions
u
u|t=0 = g, = h.
t t=0
Then

t g(x t y) 2 t h(x t y) 2
u(t, x) = d y + d y.
t 2 D 1 | y|2 2 D 1 | y|2
Proof Following the procedure described above, we extend g and h to functions on

R3 independent of x3 . In this case the integral (4.36) becomes

g(x; ) = g(x1 + y1 , x2 + y2 ) d S( y) (4.45)
4 S2
for x R2 . By symmetry we can restrict our attention to the upper hemisphere,

parametrized in polar coordinates by

y = r cos , r sin , 1 r 2 .
The surface area element is

r
dS = dr d,
1 r2
so that (4.45) becomes

2 r
g(x1 + r cos , x2 + r sin )

g(x; ) = r dr d
2 0 0 1 r2

g(x + y) 2
= d y.
2 D 1 | y|2
The claimed two-dimensional solution follows by substituting this formula for g and
the corresponding result for h in the Kirchhoff formula from Theorem 4.10.
Corollary 4.11 shows that the range of influence of (t0 , x 0 ) for the two-dimensional
wave equation is the solid region bounded by the forward light cone + (t0 , x 0 ), not
just the surface. Thus, in R2 the wave caused by a sudden disturbance has a lingering
tail after the initial wavefront has passed, as illustrated in Fig. 4.14.
Fig. 4.14 Two-dimensional

waveform observed at a
distance 3 from the origin,
corresponding to the radial
impulse shown on the left in
Fig. 4.13
In all dimensions, solutions of the wave equation exhibit a phenomenon known

as finite propagation speed. If the constant c is reinstated as in (4.30), then the range
of influence is restricted to spacetime points that are reachable at a speed less than
or equal to c.
4.7 Energy and Uniqueness
In Theorem 4.1, uniqueness of solutions was a consequence of the method of char-

acteristics. In this section we will present an alternative approach, which allows us
to deduce uniqueness directly from the equation without requiring any knowledge
of the solution other than differentiability. This argument is based on the concept of
energy of a solution, which proves to be a powerful tool for analyzing many different
types of PDE.
To motivate the definition, let us specialize again to the case of a string of length
with fixed ends. Assume that u C 2 ([0, ] [0, ]) satisfies the string wave
equation (4.4) with Dirichlet boundary conditions. In the discrete model of the string
used for the derivation of the equation, the segment of length x located at x j had
mass x and velocity u t
(x j ). By the standard expression for the kinetic energy of a
moving particle, 21 (mass)(velocity)2 , the kinetic energy of this segment is therefore
2
1 u
x (x j ) .
2 t
Summing over the segments and passing the continuum limit gives a formula for the
total kinetic energy of the string:
2
u
E K := d x.
2 0 t
The potential energy of the solution can be calculated as the energy required to
move the string from zero displacement into the configuration described by u(t, ). Let
us represent this process by scaling the displacement to su(t, ) for s [0, 1]. By (4.1)
the opposing force generated by the tension also scales proportionally to s. The work
required to shift the segment at x j from s to s +s is therefore sF(t, x j )u(t, x j )s.
The potential energy associated with this segment is
1
E P (t, x j ) := sF(t, x j )u(t, x j ) ds
0
1
= u(t, x j )F(t, x j )
2
T 2u
u(t, x j ) 2 (t, x j )x,
2 x
with a minus sign because the displacement and force are in opposing directions.
Summing over the segments and taking the continuum limit gives the total poten-
tial energy,

T 2u
E P (t) := u d x.
2 0 x 2
For comparison to the kinetic term, it is convenient to integrate by parts and rewrite
this in the form
T u 2
E P (t) = d x.
2 0 x
The total energy of the one-dimensional string at time t is given by
E = EK + E P .
For the higher-dimensional wave equation (4.30) on a domain Rn the cor-

responding definition is
2
1 u
E[u](t) := + c2 |u|2 d n x (4.46)
2 t
This is well-defined for u C 2 ([0, ) ), provided is bounded.

Theorem 4.12 Suppose Rn is a bounded domain with piecewise C 1 boundary.
If u C 2 ([0, ) ) is a solution of (4.30) with u| = 0, then the energy E[u]
defined by (4.46) is independent of t.
Proof The assumptions on u justify differentiating under the integral, so that

d u 2 u u
E[u] = + c2 u d n x.
dt t t 2 t
Under the condition u| = 0, Greens first identity (Theorem 2.10) applies to the
second term to give

u u
u d n x = u d n x. (4.47)
t t
Thus
d u 2u
E[u] = c2 u d n x,
dt t t 2
and (4.30) implies that E is constant.

Corollary 4.13 Suppose Rn is a bounded domain with piecewise C 1 boundary.
A solution u C 2 (R+ ) of the equation
4.7 Energy and Uniqueness 69
2u
c2 u = f, u| = 0,
t 2
u
u|t=0 = g, = h,
t t=0
is uniquely determined by the functions f, g, h.
Proof If u 1 and u 2 are solutions of the equation with the same initial conditions, then
w := u 1 u 2 satisfies (4.30) with the initial conditions
w
w(0, x) = 0, (0, x) = 0.
t
At time t = 0 this gives E[w] = 0, and Theorem 4.12 then implies that E[w] = 0
for all t. Since the terms in the integrand of E[w] are non-negative, they must each
vanish. This shows that w is constant, and hence w = 0 by the initial conditions.
Therefore u 1 = u 2 .
4.8 Exercises
4.1 Suppose u(t, x) is a solution of the wave equation (4.5) for x R. Let P be a
parallelogram in the (t, x) plane whose sides are characteristic lines. Show that the
value of u at each vertex of P is determined by the values at the other three vertices.
4.2 The wave equation (4.5) is an appropriate model for the longitudinal vibrations
of a spring. In this application u(t, x) represents displacement parallel to the spring.
Suppose that spring has length and is free at the ends. This corresponds to the
Neumann boundary conditions
u u
(t, 0) = (t, ) = 0, for all t 0.
x x
Assume the initial conditions are g and h as in (4.16), which also satisfy Neumann
boundary conditions on [0, ]. Determine the appropriate extensions of g and h
from [0, ] to R so that the solution u(t, x) given by (4.8) will satisfy the Neumann
boundary problem for all t.
4.3 In the derivation in Sect. 4.1, suppose we include the effect of gravity by adding
a term gx to the discrete equation of motion (4.2), where g > 0 is the constant
of gravitational acceleration. The wave equation is then modified to
2u 2 u
2
c = g. (4.48)
t 2 x 2
Assume that x [0, ], with u satisfying Dirichlet boundary conditions at the end-
points.
(a) Find an equilibrium solution u 0 (x) for (4.48), that satisfies the boundary condi-
tions but does not depend on time.
(b) Show that if u 1 is a solution of the original wave equation (4.5), also with Dirichlet
boundary conditions, then u = u 0 + u 1 solves (4.48).
(c) Given the initial conditions u(x, 0) = 0, u
t
(x, 0) = 0, find the corresponding
initial conditions for u 1 . Then apply Theorem 4.5 to find u 1 and hence solve
for u.
4.4 In Example 4.8, let the forcing term be
f (t, x) = cos(t) sin(k x),
with > 0 and

k
k := .

Find the solution u(t, x) given initial conditions g = h = 0. Include both cases
= k and = k .
4.5 The telegraph equation is a variant of the wave equation that describes the
propagation of electrical signals in a one-dimensional cable:
2u u 2 u
2
+ a + bu c = 0,
t 2 t x 2
where u(t, x) is the line voltage, c is the propagation speed, and a, b > 0 are
determined by electrical properties of the cable (resistance, inductance, etc.). Show
that the substitution
u(t, x) = eat/2 w(t, x)
reduces the telegraph equation to an ordinary wave equation for w, provided a and
b satisfy a certain condition. Find the general solution in this case. (This result has
important practical applications, in that the electrical properties of long cables can
be tuned to eliminate distortion.)
4.6 An alternative approach to the one-dimensional wave equation is to recast the

PDE as a pair of ODE. Consider the wave equation with forcing term,
2u 2 u
2
c = f.
t 2 x 2
(a) Define a vector-valued function v = (v1 , v2 ) with components

4.8 Exercises 71
u u
v1 := , v2 := .
t x
Show that v satisfies a vector equation
v v
A = b. (4.49)
t x
where b := ( f, 0) and A is the matrix

0 c2
A := .
1 0
(b) The vector equation (4.49) can be solved by diagonalizing A. Check that if we
set
1 c
T := ,
1 c
then
1 c 0
T AT = .
0 c
Then show under that the substitution
w := T v,
(4.49) reduces to a pair of linear conservation equations for the components of

w:
w1
t
c w
x
1
= f,
w2 w2 (4.50)
t
+ c x = f.
(c) Translate the initial conditions
u
u(0, x) = g(x), (0, x) = h(x),
t
into initial conditions for w1 and w2 , and then solve (4.50) using the method of
characteristics.
(d) Combine the solutions for w1 and w2 to compute v1 = u/t, and then integrate
to solve for u. Your answer should be a combination of of the dAlembert formula
(4.8) and the Duhamel formula (4.22).
4.7 The evolution of a quantum-mechanical wave function u(t, x) is governed by

the Schrdinger equation:
u
iu = 0 (4.51)
t
(ignoring the physical constants). Suppose that u(t, x) is a solution of (4.51) for
t [0, ) and x Rn , with initial condition
u(0, x) = g(x).
Assume that
|g|2 d n x < .
Rn
(a) Show that for all t 0,

|u(t, x)| d x =
2 n
|g|2 d n x.
Rn Rn
(In quantum mechanics |u|2 is interpreted as a probability density, so this identity

is conservation of total probability.)
(b) Show that a solution of Schrdingers equation is uniquely determined by the
initial condition g.
4.8 In Rn consider the wave equation
2u
c2 u = 0. (4.52)
t 2
The plane wave solutions have the form
u(t, x) = ei(kxt) , (4.53)
where R and k Rn are constants.
(a) Find the condition on = (k) for which u solves (4.52).

(b) For fixed t, R, show that x Rn ; u(t, x) = ei is a set of planes per-
pendicular to k. Show that these planes propagate, as t increases, in a direction
parallel to k with speed given by c. (Hence the term plane wave.)
4.9 The Klein-Gordon equation in Rn is a variant of the wave equation that appears
in relativistic quantum mechanics,
2u
u + m 2 u = 0, (4.54)
t 2
where m is the mass of a particle.
4.8 Exercises 73
(a) Find a formula for = (k, m) under which this equation admits plane wave
solutions of the form (4.53).
(b) Show that we can define a conserved energy E for this equation by adding a term
proportional to u 2 to the integrand in (4.46).
Chapter 5
Separation of Variables
Some PDE can be split into pieces that involve distinct variables. For example, the
equation
u
a(t)b(x)u = 0
t
could be written as
1 u
= b(x)u,
a(t) t
provided a(t) = 0. This puts all of the t derivatives and t-dependent coefficients on
the left and all of the terms involving x on the right.
Splitting an equation this way is called separation of variables. For PDE that
admit separation, it is natural to look for product solutions whose factors depend on
the separate variables, e.g., u(t, x) = v(t)(x). The full PDE then reduces to a pair
of equations for the factors. In some cases, one or both of the reduced equations is
an ODE that can be solved explicitly.
This idea is most commonly applied to evolution equations such as the heat or
wave equations. The classical versions of these PDE have constant coefficients, and
separation of variables can thus be used to split the time variable from the spatial
variables. This reduces the evolution equation to a simple temporal ODE and a spatial
PDE problem.
Separation among the spatial variables is sometimes possible as well, but this
requires symmetry in the equation that is also shared by the domain. For example,
we can separate variables for the Laplacian on rectangular or circular domains in R2 .
But if the domain is irregular or the differential operator has variable coefficients,
then separation is generally not possible.
Despite these limitations, separation of variables plays a significant role the devel-
opment of PDE theory. Explicit solutions can still yield valuable information even if
they are very special cases.

76 5 Separation of Variables
Frequency (Hz)
Intensity (dB)
Fig. 5.1 Frequency decomposition for the sound of a violin string
5.1 Model Problem: Overtones
In 1636 the mathematician Marin Mersenne published his observation that a vibrat-
ing string produces multiple pitches simultaneously. The most audible pitch corre-
sponds to the lowest frequency of vibration, called the fundamental tone of the string.
Mersenne also detected higher pitches, at integer multiples of the fundamental fre-
quency. (The relationship between frequency and pitch is logarithmic; doubling the
frequency raises the pitch by one octave.)
The higher multiples of the fundamental frequency are called overtones of the
string. Figure 5.1 shows the frequency decomposition for a sound sample of a bowed
violin string, with a fundamental frequency of 440 Hz. The overtones appear as peaks
in the intensity plot at multiples of 440.
At the time of Mersennes observations, there was no theoretical model for string
vibration that would explain the overtones. The wave equation that dAlembert sub-
sequently developed (a century later) gave the first theoretical justification. However,
this connection is not apparent in the explicit solution formula developed in Sect. 4.3.
To understand how the overtones are predicted by the wave equation, we need to
organize the solutions in terms of frequency.
5.2 Helmholtz Equation
The classical evolution equations on Rn have the form
Pt u u = 0, (5.1)
where Pt is a first- or second-order differential operator involving only the time

variable. Examples include the wave equation (Pt = 2 /t 2 ), heat equation (Pt =
/t), and Schrdinger equation (Pt = i/t).
5.2 Helmholtz Equation 77
Lemma 5.1 If u is a classical solution of (5.1) of the form
u(t, x) = v(t)(x),
for t R and x Rn , then in any region where u is nonzero there is a constant

such that the components solve the equations
Pt v = v, = . (5.2)
Proof Substituting u = v into (5.1) gives
Pt v v = 0.
Assuming that u is nonzero, we can divide by u to obtain
1 1
Pt v = .
v
The left hand-side is independent of x and the right is independent of t. We conclude

that both sides must be equal to some constant .
The two differential equations in (5.2) are analogous to eigenvalue equations from
linear algebra, with the role of the linear operator or matrix taken by the differential
operators Pt or .
Let us first focus on the spatial problem, which is usually written in the form
= . (5.3)
This is called the Helmholtz equation, after the 19th century physicist Hermann von
Helmholtz. The minus sign is included so that 0 for the most common types of
boundary conditions. Adapting the linear algebra terminology, we refer to the number
in (5.3) as an eigenvalue and the corresponding solution as an eigenfunction.
The Helmholtz equation is sometimes called the Laplacian eigenvalue equation.
We will present a general analysis of the Helmholtz problem on any bounded
domain in Rn in Chap. 11, and later in this chapter we will consider some two- or three-
dimensional cases for which further spatial separation is possible. For the remainder
of this section we restrict our attention to problems in one spatial dimensional, for
which (5.3) is an ODE.
Theorem 5.2 For C 2 [0, ] the equation
d 2
= , (0) = () = 0, (5.4)
dx2
has nonzero solutions only if
2 n2
n :=
2
for n N. Up to a constant multiple, the corresponding solutions are

n (x) := sin( n x). (5.5)
Proof Note that (5.4) implies

d 2
||2 d x = d x.
0 0 dx2
Using the Dirichlet boundary conditions we can integrate by parts on the right without
any boundary term, yielding
2
d
|| d x =
2 d x. (5.6)
dx
0 0
Assuming that is not identically zero, this shows that 0. Furthermore, = 0

implies d/d x = 0, which gives a constant solution. The only constant solution is
the trivial case = 0, because of the boundary conditions.
It therefore suffices to consider the case > 0, for which the ODE in (5.4) reduces

to the harmonic
oscillator equation, with independent solutions given by sin( x)
and cos( x). Only sine satisfies the condition (0) = 0, so the possible solutions
have the form
(x) = sin( x).
To satisfy the condition () = 0 we must have

sin( ) = 0.

For a nonzero solution this imposes the restriction that N, which gives the
claimed set of solutions.
Some of the eigenfunctions obtained in Theorem 5.2 are illustrated in Fig. 5.2. For
the sake of
application to our original string model, let us reinstate the propagation
speed c := T / and write the string equation as
2u
c2 u = 0, u(t, 0) = u(0, ) = 0. (5.7)
t 2
Fig. 5.2 The first four

eigenfunctions for a
vibrating string with fixed
ends

With the spatial solution given by the eigenfunction associated to n , the correspond-
ing temporal eigenvalue equation is also a harmonic oscillator ODE,
d 2v
= c2 n v.
dt 2
The solutions could be written in terms of sines and cosines, but for the temporal
component it is usually more convenient to use the complex exponential form. The
general complex-valued solution is
vn (t) = an ein t + bn ein t ,
with an , bn C and cn
n := c n = . (5.8)

For real-valued solutions, the coefficients are restricted by bn = an .
Combining the temporal and spatial components gives a set of solutions for the
vibrating string problem:

u n (t, x) = an ein t + bn ein t sin( n x), (5.9)
for n N.
The functions (5.9) are referred to as pure-tone solutions, because they model
oscillation at a single frequency n . In the case of visible light waves, the frequency
corresponds directly to color. For this reason the set of frequencies {n } is called
the spectrum. By association, the term spectrum is also used for sets of eigenvalues
appearing in more general problems. For example, the set {n } of eigenvalues for
which the Helmholtz problem has a nontrivial solution is called the spectrum of the
Laplacian, even though n is proportional to the square of the frequency n .
From (5.8) we can deduce the fundamental tone of the string, as predicted by
dAlemberts wave equation model. To convert frequency to the standard unit of Hz
(cycles per second), we divide 1 by 2 to obtain the formula

1 1 T
= . (5.10)
2 2
This is known as Mersennes law, published in 1637.

The wave equation model also predicts the higher frequencies n = n1 , cor-
responding to the sequence of overtones noted illustrated in Fig. 5.1. The fact that
each overtone is associated with a particular spatial eigenfunction is significant. The
waveforms for higher overtones have nodes, meaning points where the string is sta-
tionary. As we can see in Fig. 5.2, the nodes associated to the frequency n subdivide
the string into n equal segments. Touching the string lightly at one of these nodes
will knock out the lower frequencies, a practice string players refer to as playing a
harmonic.
As this discussion illustrates, the spectral analysis of the wave equation is more
directly connected to experimental observation than the explicit solution formula
(4.8). The displacement of a vibrating string is technically difficult to observe directly
because the motion is both rapid and of small amplitude. Such observations were
first achieved by Hermann von Helmholtz in the mid-19th century.
Example 5.3 The one-dimensional wave equation can be used to model for the fluc-
tuations of air pressure inside a clarinet. The interior of a clarinet is essentially a
cylindrical column, and for simplicity we can assume that the pressure is constant
on cross-sections of the cylinder, so that the variations in pressure are described by
a function u(t, x) with x [0, ], where is the length of the instrument. Pressure
fluctuations are measured relative to the fixed atmospheric background, with u = 0
for atmospheric pressure.
The maximum pressure fluctuation occurs at the mouthpiece at x = 0, where a
reed vibrates as the player blows air into the instrument. Since a local maximum
of the pressure corresponds to a critical point of u(t, ), the appropriate boundary
condition is
u
(t, 0) = 0. (5.11)
x
At the opposite end the air column is open to the atmosphere, so the pressure does
not fluctuate,
u(t, ) = 0. (5.12)
The evolution of u as a function of t is governed by the wave equation (4.5), with

c equal to the speed of sound. The corresponding Helmholtz problem is
d 2
= , (0) = 0, () = 0. (5.13)
dx2
The boundary condition at x = 0 implies that

(x) = cos( x),
and the condition at x = then requires

cos( ) = 0.
This means that the eigenvalues are given by
2
n := (n 21 )2 ,
2
for n N. Some of the resulting eigenfunctions are shown in Fig. 5.3.

Fig. 5.3 The first four eigenfunctions for pressure fluctuations in a clarinet
Hz
dB
Fig. 5.4 Clarinet frequency decomposition
The corresponding oscillation frequencies are given by

c
n = (n 21 ).

In contrast to the string, the model predicts that the clarinets spectrum will contain
only odd multiples of the fundamental frequency 1 . Figure 5.4 shows the frequency
decomposition for a clarinet sound sample. The prediction holds true for the first few
modes, but the simple model appears to break down at higher frequencies.
5.3 Circular Symmetry
In dimension greater than one, spatial separation of variables is essentially the only
way to compute explicit solutions of the Helmholtz equation (5.3), and this only
works for very special cases. The most straightforward example is a rectangular
domain in Rn , which we will discuss in the exercises.
In this section we consider the simplest non-rectangular case, based on polar
coordinates (r, ) in R2 . Separation in polar coordinates allows us to compute eigen-
functions and eigenvalues on a disk in R2 , for example.
With x = (x1 , x2 ) in R2 , polar coordinates are defined by
(x1 , x2 ) = (r cos , r sin ).

The polar form of the Laplacian is computed by writing
2 2
= +
x12 x22
and then converting the partials with respect to x1 and x2 into r and derivatives
using the chain rule. The result is

1 1 2
= r + 2 2. (5.14)
r r r r
Note that there are no mixed partials involving both r and , and that the coefficients
do not depend on . This allows separation of r and , provided the domain is defined
by specifying ranges of r and .
To solve the radial eigenvalue equation, we will use Bessel functions, named for
the astronomer Friedrich Bessel. Bessels equation is the ODE:
z 2 f (z) + z f (z) + (z 2 k 2 ) f (z) = 0, (5.15)
with k C in general. For our application k will be an integer. The standard pair of
linearly independent solutions is given by the Bessel functions Jk (z) and Yk (z).
The Bessel J-functions, a few of which are pictured in Fig. 5.5, satisfy
Jk (z) = (1)k Jk (z), (5.16)
for all k Z. Bessel represented these solutions as integrals:

1

Jk (z) := cos z sin k d.
0
One can also write Jk as a power series k N0 ,
Fig. 5.5 The first four J0

Bessel J-functions
J1
J2 J3
5.3 Circular Symmetry 83
z k
2 l
1 z
Jk (z) = . (5.17)
2 l=0
l!(k + l)! 4
Together with (5.16), this shows that Jk (z) ck z |k| as z 0 for any k Z. In
contrast, the Bessel Y-function satisfies Yk (z) ck z |k| as z 0.
A change of sign in (5.15) gives the equation
z 2 f (z) + z f (z) + (z 2 + k 2 ) f (z) = 0. (5.18)
Its standard solutions are the modified Bessel functions Ik (z) and K k (z). As z 0
these satisfy the asymptotics Ik (z) ck z |k| , as illustrated in Fig. 5.6, and K k (z)
ck z |k| .
Lemma 5.4 Suppose C 2 (R2 ) is a solution of
= ,
that factors as a product h(r )w(). Then, up to a multiplicative constant, has the
form
,k (r, ) := h k (r )eik , (5.19)
for some k Z, with

|k|
r , = 0,
h k (r ) := Jk ( r ), > 0,

Ik ( r ), < 0.
Proof Under the assumption = hw, the Helmholtz equation reduces by (5.14) to

w h h 2w
r + 2 + hw = 0.
r r r r 2
Fig. 5.6 The first four I0

modified Bessel I-functions
I1
I2
I3
With some rearranging, we can separate the r and variables,

1 2 1 2w
r h + 2 r 2 = , (5.20)
h r w 2
provided h and w are nonzero.

As in Lemma 5.1, we conclude that both sides must be equal to some constant .
The equation is
2w
= w. (5.21)
2
The function w() is assumed to be 2-periodic. By the arguments used in Theo-
rem 5.2, a nontrivial solution is possible only if = k 2 where k is an integer. A full
set of 2-periodic solutions of (5.21) is given by
wk () := eik , k Z.
Before examining the radial equation, let us note that the assumption that is
C 2 imposes
a boundary condition at r = 0. To see this, first note that the function
r = x12 + x22 is continuous at (0, 0) but not differentiable. For r > 0,
r xj
= ,
x j r
which does not have a limit as r 0. On the other hand, the functions
r ei = x1 i x2
are C . Similarly, for k Z we have

|k| ik (x1 + i x2 )k , k N0 ,
r e = (5.22)
(x1 i x2 )k , k N.
These functions are polynomial and hence C . We will see below that the solutions
of the radial equation corresponding to = k 2 satisfy h(r ) ar k as r 0, for
some constant a. The differentiability of at the origin will require the asymptotic
condition
h k (r ) ar |k| (5.23)
as r 0.
For wk () = eik , the radial component of (5.20) is

2
r h k + (r 2 k 2 )h k = 0. (5.24)
r
5.3 Circular Symmetry 85
The case = 0 is relatively straightforward to analyze. In this case (5.24) is homoge-

neous in the r variable (meaning invariant under scaling). Such equations are solved
by monomials of the form h k (r ) = r with R. If we substitute this guess into
(5.24) with = 0, the equation reduces to
2 k 2 = 0,
with solutions = k. Since a second order ODE has exactly two independent
solutions, the functions r k give a full set of solutions for k = 0. For k = 0 the two
possibilities are 1 and ln r . By the condition (5.23), the solutions ln r and r |k| must
be ruled out. The only possible solutions for = 0 are thus
h k (r ) = r |k| .
Note that the resulting solutions,

0,k (r, ) := r |k| eik ,
are precisely the polynomials (5.22).

For > 0 (5.24)
can be reduced to the Bessel form (5.15) by the change of
variables z = r . The possible solutions Yk ( r ) are ruled out because they
diverge at r = 0.On the other hand, the power series (5.17) shows that the func-
tion h k (r ) = Jk ( r ) satisfies the condition (5.23). Thus for > 0 the possible
eigenfunction with k Z is

,k (r, ) := Jk ( r )eik .
We should check that this function is at least C 2 at the origin. In fact, it follows from
the power series expansion (5.17) that ,k is C on R2 .
Similar
considerations apply for < 0, except that this time the substitution
z = r reduces (5.24) to (5.18). The condition (5.23) is satisfied only for the
solution Ik ( r ).
Example 5.5 The linear model for the vibration of a drumhead is the wave equation
(4.30). For a circular drum we can take the spatial domain to be the unit disk D :=
{r < 1} R2 . Lemma 5.1 reduces the problem of determining the frequencies of
the drum to the Helmholtz equation,
= , |D = 0. (5.25)
The possible product solutions are given by Lemma 5.4, subject to the boundary
condition h k (1) = 0. This rules out 0, because in that case h k (r ) has no zeros
for r > 0.
For > 0, we have h k (r ) = Jk ( r ), and the boundary condition takes the form

Jk ( ) = 0.
Table 5.1 Zeros of the Bessel function Jk . For each k, the spacing between zeros approaches as
m
k jk,1 jk,2 jk,3 jk,4
0 2.405 5.520 8.654 11.792
1 3.832 7.016 10.174 13.324
2 5.136 8.417 11.620 14.796
3 6.380 9.761 13.015 16.223
4 7.588 11.065 14.373 17.616

(This is analogous to the condition sin( ) = 0 from the one-dimensional string
problem.) Although Jk is not a periodic function, it does have an infinite sequence
of positive zeros with roughly evenly spacing. It is customary to write these zeros in
increasing order as
0 < jk,1 < jk,2 < . . . .
By the symmetry (5.16),

jk,m = jk,m .
Table 5.1 lists some

of these zeros.
Restricting to the set of Bessel zeros gives the set of eigenvalues
k,m = jk,m
2
,
indexed by k Z, m N. The corresponding eigenfunctions are
k,m (r, ) := Jk ( jk,m r )eik . (5.26)
The first set of these are illustrated in Fig. 5.7.

The collection of functions (5.26) yields a complete list of eigenfunctions and
eigenvalues for D, although that is not something we can prove here.
The eigenvalues calculated in Example 5.5 correspond to vibrational frequencies
k,m := cjk,m ,
for k Z and m N. The propagation speed c depends on physical properties

such as tension and density. The relative size of the frequencies helps to explain
the lack of definite pitch in the sound of a drum. The ratios of overtones above the
fundamental 0,1 are shown in Table 5.2. In contrast to the vibrating string case,
where the corresponding ratios were integers 1, 2, 3, . . ., or the clarinet model of
Example 5.3 with ratios 1, 3, 5, . . ., the frequencies of the drum are closely spaced
with no evident pattern.
5.4 Spherical Symmetry 87
m=1 m=2 m=3
k=0
k=1
k=2
Fig. 5.7 Contour plots of the spatial component of the eigenfunctions of D
Table 5.2 Frequency ratios k m k,m /0,1

for a circular drumhead
0 1 1
1 1 1.593
2 1 2.136
0 2 2.295
3 1 2.653
1 2 2.917
4 1 3.155
5.4 Spherical Symmetry
Another special case that allows separation of spatial variables is spherical symmetry
in R3 . Spherical coordinates (r, , ) are defined through the relation
(x1 , x2 , x3 ) = (r sin cos , r sin sin , r cos ).

Note that is the azimuthal angle here and the polar angle, consistent with the
notation from Sect. 5.3. (This convention is standard in mathematics; in physics the
roles are often reversed.)
As in the circular case, we can use the chain rule to translate the three-dimensional
Laplacian into spherical variables:

1 2 1 1 2
= 2 r + 2 sin + 2 2 . (5.27)
r r r r sin r sin 2
It is not immediately clear that this operator admits separation, because the coef-
ficients depend on both r and . Note, however, that we can factor r 2 out of the
angular derivative terms, to write (5.27) as

1 1
= r2 + 2 S2 , (5.28)
r 2 r r r
where
1 1 2
S2 := sin + . (5.29)
sin sin2 2
Here S2 stands for the unit sphere {r = 1} R3 , and S2 is called the spherical
Laplacian.
The expression (5.29) may look awkward at first glance, but S2 is a very natural
operator geometrically. From the fact that is invariant under rotations of R3 about
the origin, we can deduce that S2 is also invariant under rotations of the sphere. It is
possible to show that S2 is the only second-order operator with this property, up to
a multiplicative constant. The operator S2 is thus as symmetric as possible, and the
reason that (5.29) looks so complicated is that the standard coordinate system (, )
does not reflect the full symmetry of the sphere.
We will discuss the radial component of (5.28) in an example below. For now let
us focus on the Helmholtz problem on the sphere, which allows further separation
of the and variables.
The classical ODE that arises from separation of the angle variables is the asso-
ciated Legendre equation:

2
(1 z ) f (z) 2z f (z) + ( + 1)
2
f (z) = 0, (5.30)
1 z2
with parameters , C. A pair of linearly independent solutions is given by the

Legendre functions P (z) and Q (z).
In the special case where is replaced by l N0 and by a number m
{l, . . . , l}, respectively, the Legendre P-functions are given by a relatively simple
formula:
(1)m 2 m/2 d
l+m
Plm (z) = (1 z ) (z 2 1)l . (5.31)
2l l! dz l+m
Associated to this set of Legendre functions are functions of the angle variables
called spherical harmonics. These are defined by
Ylm (, ) := cl,m eim Plm (cos ), (5.32)
where cl,m is a normalization constant whose value is not important for us.
From (5.31), using z = cos and 1 z 2 = sin2 , we can see that Ylm is a
polynomial of degree l in sin and cos . This makes it relatively straightforward
to check that each Ylm (, ) is a smooth function on S2 .
Lemma 5.6 Suppose u C 2 (S2 ) is a solution of the equation
S2 u = u (5.33)
that factors as u(, ) = v()w(). Then up to a multiplicative constant, u is equal

to a spherical harmonic Ylm for l N0 and m {l, . . . , l}. The corresponding
eigenvalues depend only on l,
l = l(l + 1),
and each has multiplicity 2l + 1.

Proof By (5.29), the substitution u = vw leads to the separated equation

sin v 1 2w
sin += .
v w 2
The continuity of u requires that w be 2-periodic. Hence, for the equation
2w
= w,
2
the full set of solutions is represented by w() = eim with = m 2 for m Z.

With u(, ) = vm ()eim , the eigenvalue equation (5.33) reduces to

1 d dvm m2
sin + vm = 0.
sin d d sin2
Under the substitutions z = cos and vm () = f (cos ), this becomes

m2
(1 z 2 ) f 2z f + f = 0,
1 z2
which is recognizable as the Legendre equation (5.30) with parameters m = and

= ( + 1).
Although S2 does not have a boundary, use of the coordinate [0, ] creates
artificial boundaries at the endpoints, i.e., at the poles of the sphere. This is analogous
to the boundary at r = 0 in Lemma 5.4. We need to find solutions which will be

smooth at the poles.
(z) diverges as z 1 for any C.
It turns out that for m Z, the function Q m
Similarly, the functions Pm (z) diverge as z 1 except for the special cases
Plm (z) given by (5.31). In other words, up to a multiplicative constant vm () must
be equal to Plm (cos ) for some l N0 with l |m|. The corresponding solution u
is proportional to the spherical harmonic Ylm .
By the identification = l, the eigenvalue is given by = l(l + 1). The corre-
sponding multiplicity is the number of possible choices of m {l, . . . , l}, namely
2l + 1.
The spherical harmonics appearing in Lemma 5.6 give a complete set of eigen-
functions for S2 , in the sense that the only possible eigenvalues are l(l + 1) for
l N0 and an eigenfunction with eigenvalue l(l + 1) is a linear combination of the
Ylm for m {l, . . . , l}. To prove this requires more advanced methods than we have
available here.
Example 5.7 In 1925, Erwin Schrdinger developed a quantum model for the hydro-
gen atom in which the electron energy levels are given by the eigenvalues of the
equation
1
= (5.34)
r
on R3 . (We have omitted the physical constants.) The eigenfunctions are assumed
to be bounded near r = 0 and decaying to zero as r .
Since the term 1/r is radial, separation of the radial and angular variables is
possible in (5.34). By Lemma 5.6, the angular components are given by spherical
harmonics. A corresponding full solution has the form
(r, , ) = h(r )Ylm (, ). (5.35)
Substituting this into (5.34) and using the spherical form of the Laplacian (5.28)
gives the radial equation

1 d d l(l + 1) 1
r2 + h(r ) = h(r ). (5.36)
r 2 dr dr r2 r
One strategy used to analyze an ODE such as (5.36) is to first consider the asymptotic
behavior of solutions as r 0 or .
Suppose we assume h(r ) r as r 0. Plugging this into (5.36) and comparing
the two sides gives a leading term
( + 1)r 2 + l(l + 1)r 2
on the left side, with all other terms of order r 1 or less. This shows that h(r ) r
as r 0 is possible only if
( + 1) = l(l + 1).
The two solutions are = l or = l 1. Taking < 0 would cause h(r ) to

diverge as r 0. Therefore to obtain a solution bounded at the origin, we will
assume that
h(r ) r l
as r 0.
As r , if we consider the terms in (5.36) with coefficients of order r 0 and
drop the rest, the equation becomes
h (r ) h(r ). (5.37)
If 0 then this shows that h(r ) could not possibly decay at infinity. Hence we
assume that < 0 and set
2 := ,
with R. The asymptotic equation (5.37) implies the behavior
h(r ) cer
as r .
Determining these asymptotics allows us to make an educated guess for the form
of the solution. For an as yet undetermined function q(r ), we set
h(r ) = q(r )r l er , (5.38)
with the conditions that q(0) = 1 and q(r ) has subexponential growth as r .
The goal of setting up the solution this way is that the equation for q(r ) will simplify.
Substituting (5.38) into (5.36) leads to the equation
rq + 2(1 + l r )q + (1 2(l + 1))q = 0. (5.39)
To find solutions, we suppose q(r ) is given by a power series

q(r ) = ak r k ,
k=0
with a0 = 1. Plugging this into (5.39) gives

0= k(k 1)ak r k1 + 2(1 + l r )kak r k1 + (1 2(l + 1))ak r k .
k=0
Equating the coefficient of r k to zero then gives a recursive relation
2(k + l + 1) 1
ak+1 = ak . (5.40)
(k + 1)(k + 2l + 2)
If we assume that the numerator of (5.40) never vanishes, then the recursion
relation implies that
(2)k
ak
k!
as k . This would give q(r ) ce2r as r , making h(r ) also grow

exponentially as r .
The only way to avoid this exponential growth is for the sequence of ak to terminate
at some point, so that q is a polynomial. The numerator on the right side of (5.40)
will eventually vanish if and only if
1
= ,
2n
for some integer n l + 1. Under this assumption the sequence ak terminates at
k = n l 1. Since = 2 , this restriction on gives the set of eigenvalues
1
n := , n N.
4n 2
This is in fact the complete set of eigenvalues for this problem, given the conditions
we have imposed at r = 0 and r . With this eigenvalue calculation, Schrdinger
was able to give the first theoretical explanation of the emission spectrum of hydrogen
gas (i.e., the set of wavelengths observed when the gas is excited electrically). The
origin of these emission lines had been a mystery since their discovery by Anders
Jonas ngstrm in the mid-19th century.
Each value of n corresponds to a family of eigenfunctions given by
n,l,m (r, , ) = r l qn,l (r )e 2n Ylm (, ),

r
for l {0, . . . , n 1}, m {l, . . . , l}. Here qn,l (r ) denotes the polynomial of
degree n l 1 with coefficients specified by (5.40). To compute the multiplicity
of n , we count n 1 choices for l and then 2l + 1 choices of m for each l. The total
multiplicity is

n1
(2l + 1) = n 2 .
l=0

5.5 Exercises 93
5.5 Exercises
5.1 On the half-strip = (0, 1) (0, ) R2 , find the solutions of
u = 0
that factor as a product u(x1 , x2 ) = g(x1 )h(x2 ), under the boundary conditions
u(0, x2 ) = u(1, x2 ) = u(x1 , 0) = 0.
5.2 The linear model for vibrations of a rectangular drumhead is the wave equa-
tion (4.30) with Dirichlet boundary conditions on a rectangle R := [0, 1 ][0, 2 ]
R2 . Separation of variables leads to the corresponding Helmholtz problem
= , |R = 0.
Find the eigenfunctions of product type, (x1 , x2 ) = 1 (x1 )2 (x2 ), and the associ-
ated frequencies of vibration. For 1 = 2 , compare the ratios of these frequencies
to Table 5.2. Would a square drum do a better job of producing a definite pitch?
5.3 The one-dimensional heat equation for the temperature u(t, x) of a metal bar
of length is
u 2u
2 = 0,
t x
for t 0 and x (0, ). (We will derive this in Sect. 6.1.) If the ends of the bar are
insulated, then u should satisfy Neumann boundary conditions
u u
(t, 0) = (t, ) = 0.
x x
Find the product solutions u(t, x) = v(t)(x).
5.4 The damped wave equation on Rn is
2u u
+ u = 0, (5.41)
t 2 t
where u C 2 ([0, ) ) and 0 is a constant called the coefficient of friction.

Suppose that C 2 () satisfies the Helmholtz equation (5.3) on with eigenvalue
> 0, for some appropriate choice of boundary conditions. Show that that (5.41)
has solutions of the form
u(t, x) = (x)eit ,
and find the set of possible values of . In particular, show that Im > 0 if > 0,
which implies that the solutions decay exponentially in time. Does this decay rate
depend on the oscillation frequency?
5.5 Consider this example of a nonlinear diffusion equation:
u
(u 2 ) = 0,
t
for t 0, x Rn .
(a) Assuming a product solution of the form u(t, x) = v(t)(x), separate variables
and find the equations for v(t) and (x).
(b) Show that (x) = |x|2 solves the spatial equation, and find the corresponding
function v(t) given the initial condition v(0) = a > 0. (Observe that the solution
blows up at a finite time that depends on a.)
5.6 In polar coordinates for R2 , define the domain

= (r, ); 0 < r < 1, 0 < < /3 ,
which is a sector within the unit disk. Find the eigenvalues of on with Dirichlet
boundary conditions.
5.7 The quantum energy levels of a harmonic oscillator in Rn are the eigenvalues
of the equation

+ |x|2 = , (5.42)
under the condition that C 2 (Rn ) and 0 at infinity.

(a) First consider the case n = 1:

2
+ x 2
= , (5.43)
x 2
Substitute
(x) = q(x)ex /2
2
into (5.43) and find the corresponding ODE for q.

(b) Assume that the function q from (a) is given by a power series in x,

q(x) = ak x k ,
k=0
and find a recursive equation for ak+2 in terms of ak .

(c) Find the values of for which the power series for q from (b) truncates to a
polynomial. (The resulting functions q are called Hermite polynomials.)
5.5 Exercises 95
(d) Returning to the original problem, by reducing (5.42) to n copies of the case
(5.43), deduce the set of eigenvalues .
5.8 Let B3 R3 be the unit ball {r < 1}. Consider the Helmholtz problem
= ,
with Dirichlet boundary conditions at r = 1.

(a) Assume that
(r, , ) = h l (r )Ylm (, ),
where Ylm is the spherical harmonic introduced in Sect. 5.4. Find the radial equa-
tion for h l (r ).
(b) For l = 0 show that the radial equation is solved by

sin( r )
h 0 (r ) = .
r
What set of eigenvalues does this give?
(c) Show that the substitution,

h l (r ) = r 2 fl ( r ),
1
reduces the equation from (a) to a Bessel equation (5.15) for fl (z), with a frac-
tional value of k. Use this to write the solution h l (r ) in terms of Jk .
(d) Express the eigenvalues in terms of Bessel zeros with fractional values of k.
Chapter 6
The Heat Equation
In physics, the term heat is used to describe the transfer of internal energy within a
system of particles. When this transfer results from collective motion of particles in
a gas or fluid, the process is called convection. The continuity equation developed in
Sect. 3.1 describes convection by fluid flow, which is the special case called advection.
Another form of convection is conduction, where the heat transfer caused by random
collisions of individual particles.
The basic mathematical model for heat conduction is a PDE called the heat equa-
tion, developed by Joseph Fourier in the early 19th century. In this chapter we will
discuss the derivation and develop some basic properties of this equation, our first
example of a PDE of parabolic type.
6.1 Model Problem: Heat Flow in a Metal Rod
A metal rod that is sufficiently thin can be treated as one-dimensional system. Let
u(t, x) denote the temperature of the rod at time t and position x, with x R for
now.
There are two physical principles that govern the flow of heat in the rod. The
first is the relationship between thermal (internal) energy and temperature. Thermal
energy is proportional to a product of density and temperature, by a constant c called
the specific heat of the material. Thus, the total thermal energy in a segment [a, b]
is given by
b
U =c u d x. (6.1)
a
We will assume that the density is constant, although it could be variable in some
applications.
The second principle is Fouriers law of heat conduction, which describes how heat
flows from hotter regions to colder regions. In its one-dimensional form, Fouriers
law says that the flux of thermal energy across a given point is given by
98 6 The Heat Equation
u
q = k , (6.2)
x
where the constant k > 0 is the thermal conductivity of the material.
Assuming that the rod is thermally isolated, conservation of energy dictates the
rate of change of the thermal energy within the segment is equal to the flux across
its boundaries, i.e.,
dU
(t) = q(t, a) q(t, b). (6.3)
dt
As in our derivation of the local equation for conservation of mass, the combination
of (6.1) and (6.3) yields an integral equation

b
u q
c + d x = 0.
a t x
Since a and b were arbitrary, this implies a local conservation law,
u q
c + = 0.
t x
Using the formula for q from Fouriers law (6.2), we obtain the one-dimensional
heat equation:
u k 2u
= 0. (6.4)
t c x 2
For a rod of finite length, the solution u will satisfy boundary conditions that
depend on how the rod interacts with its environment. If the rod is parametrized by
x [0, ] and we assume that each end is held at a fixed temperature, then this fixes
the values at the endpoints,
u(t, 0) = T0 , u(t, ) = T1 (6.5)
for all t. These are inhomogeneous Dirichlet boundary conditions. In one dimension
the inhomogeneous problem can be reduced very simply to the homogeneous case
by noting that
x x
u 0 (x) := T0 + T1 1

gives an equilibrium solution to the heat equation satisfying the boundary condi-
tions (6.5). By the superposition principle, u u 0 satisfies the heat equation with
homogeneous Dirichlet conditions.
Another possible boundary assumption is that the ends are insulated, so that no
thermal energy flows in or out. This means that q vanishes at the boundary, yielding
the Neumann boundary conditions
6.1 Model Problem: Heat Flow in a Metal Rod 99
u u
(t, 0) = (t, ) = 0.
x x
Example 6.1 On the bounded interval [0, ], we can find product solutions to the heat
equation using Lemma 5.1. For the Dirichlet boundary conditions u(0) = u( ) =
0, Theorem 5.2 gives the set of sine eigenfunctions (5.5). The corresponding heat
equation solutions are
2 n2
nx
u(t, x) = e 2 t sin

for n N. Note that all of these solutions decay exponentially to 0 as t .

For insulated ends we switch to Neumann conditions and obtain the cosine modes.
The resulting set of solutions is
2 n2
nx
u(t, x) = e 2
t
cos

for n N0 . In this case the n = 0 mode yields a constant solution.

In Chap. 8 we will discuss the construction of series solutions from these trigono-
metric families.
The higher dimensional form of the heat equation can be derived by an argument
similar to that given above. In Rn , the thermal flux q is vector valued, and Fouriers
law becomes the gradient formula
q = ku.
Local conservation of energy is expressed by the continuity equation (3.20),
u
c + q = 0.
t
In combination, these yield the n-dimensional heat equation,
u k
u = 0. (6.6)
t c
The importance of the heat equation as a model extends well beyond its original
thermodynamic context. One of the most prominent examples of this is Albert Ein-
steins probabilistic derivation of the heat equation as a model for Brownian motion
in 1905, in one of the set of papers for which he was later awarded the Nobel prize.
Brownian motion is named for the botanist Robert Brown, who observed in 1827 that
minute particles ejected by pollen grains drifted erratically when suspended in water,
with a jittery motion for which no explanation was available at the time. Einstein
theorized that the motion was caused by collisions with a large number of molecules
whose velocities were distributed randomly. The existence of atoms and molecules
was still unconfirmed in 1905, and Einsteins model provided crucial supporting
evidence.
To summarize Einsteins argument, suppose that a total of n particles are distrib-
uted on the real line. In an interval of time , the position of each particle is assumed
to change by a random amount according to a distribution function . To be more
precise, the number of particles experiencing a displacement between and + d
is
dn = n( ) d.
The total number of particles is conserved, which imposes the condition

( ) d = 1. (6.7)

The distribution of displacements is assumed to be symmetric, ( ) = ( ),

meaning that particles are equally likely to move left or right.
Suppose the distribution of particles at time t is given by a density function (t, x).
Under the displacement hypothesis, the values of this density at times t and t +
are related by
(t + , x) = (t, x )( ) d. (6.8)

To find an equation for , Einstein takes the Taylor expansions of the density on both
sides of (6.8), obtaining

(t + , x) = (t, x) + (t, x) + . . . (6.9)
t
on the left, and
1 2
(t, x ) = (t, x) (t, x) + (t, x) 2 + . . .
x 2 x2
inside the integral. Integrating the latter expansion against gives, by (6.7) and the
assumption that is even (which knocks out the linear term),

1 2
(t, x )( ) d = (t, x) + (t, x) 2 ( ) d + . . . .
2 x2
Substituting this formula together with (6.9) into (6.8) and keeping the leading terms
gives
1 2
(t, x) = (t, x) 2 ( ) d.
t 2 x2
6.1 Model Problem: Heat Flow in a Metal Rod 101
Einstein then assumes that the value

1
D= 2 ( ) d (6.10)
2
is a fixed constant, so that the equation for becomes
2
D 2 = 0,
t x
i.e., the heat equation. Remarkably, the function representing the random distri-
bution of displacements plays no role in the final equation, except in the value of
the constant D. This fact is related to a fundamental result in probability called the
central limit theorem.
Diffusion models involving random motions of particles are prevalent in physics,
biology, and chemistry. The same statistical principles appear in other applications
as well, for example in models of the spread of infection in medicine, or in the study
of fluctuating financial markets. The heat equation plays a fundamental role in all of
these applications.
6.2 Scale-Invariant Solution
Let us consider the heat equation on R, with physical constants normalized to 1,
u 2u
2 = 0. (6.11)
t x
Note that the equation is invariant under the rescaling (t, x) (2 y, x), with
a nonzero
constant. This suggests a change of variables to the scale-invariant ratio
y := x/ t might simplify the equation.
Let us try to find a solution of the form u(t, x) = q(y) for t > 0. By the chain
rule,
u y 2u 1
= q , = q .
t 2t x2 t
Thus, as an equation for q, (6.11) reduces to an ODE,
y
q = q .
2
This can be solved for q by separation of variables for ODE, as described in Sect. 2.5,
q (y) = q (0)ey /4
2
.
A further integration yields

y
2
q(y) = q (0) ey /4
dy + q(0).
0
In the original coordinates this translates to

x
t
ey /4
2
u(t, x) = C1 dy + C2 .
0
It is easy to confirm that this solves (6.11) for t > 0. To see what happens as t 0,
note that

ey /4 dy = ,
2
by the computations from Exercise 2.5. Thus

C1 + C2 , x > 0,
lim u(t, x) = 0, x = 0, (6.12)
t0

C1 + C2 , x < 0.
In view of this limiting behavior, let us consider the particular solution U defined by
setting C1 = 14 , C2 = 21 ,
x
1 t 1
ey /4
2
U (t, x) := dy + . (6.13)
4 0 2
This solution is plotted for some small values of t in Fig. 6.1. By (6.12), limt0
U (t, x) = (x), the Heaviside step function defined by

1, x > 0,
(x) := 21 , x = 0,

0, x < 0.
The fact that U (t, x) has such a simple limit as t 0 can be used to derive a more
general integral formula. Suppose we want to solve (6.11) for the initial condition
Fig. 6.1 The heat solution

U (t, x) for times from t = 0
to 3
6.2 Scale-Invariant Solution 103
u(t, x) = (x)

with Ccpt (R). The key observation is that can be reproduced by integrating its
derivative against the Heaviside function,
x
(z)(x z) dz = (z) d x

= (x).
This suggests that we could solve the heat equation with initial condition by setting

u(t, x) := (z)U (t, x z) dz.

For t > 0 the function U (t, x z) is continuously differentiable, so we can integrate

by parts to rewrite this as

U
u(t, x) = (z) (t, x z) dz
z
(6.14)
1 (xz)2 /4t
= (z)e dz.
4 t
In terms of the function

1
ex /4t ,
2
Ht (x) :=
4 t
the solution is
u(t, x) = Ht (x z)(z) dz.

This integral is called the convolution of Ht with and denoted u = Ht .

In the next section, we will generalize this convolution formula to higher dimen-
sion, and check that it does yield a solution of the heat equation satisfying the desired
initial condition.
6.3 Integral Solution Formula
Consider the heat equation on Rn ,
u
u = 0 (6.15)
t
for t > 0, with initial condition
u(0, x) = g(x)
for x Rn .
Inspired by the calculations in Sect. 6.2, we define
Ht (x) := (4 t) 2 e|x| /4t

n 2
(6.16)
for t > 0. The normalizing factor (4 t) 2 is chosen so that

n

Ht (x) d n x = 1. (6.17)
Rn
(See Exercise 2.5 for the computation of integrals of this type.)

Direct differentiation shows that for t > 0,

Ht Ht = 0, (6.18)
t
so that Ht (x) is a solution of the heat equation. However, the limit of Ht (x) as t 0
is 0 for x
= 0 and for x = 0, which does not seem to make sense as a distribution
of temperatures. (We will return to discuss the interpretation of this initial condition
in Chap. 12.)
With (6.14) as motivation, our goal in this section is to show that the convolution
u = Ht g satisfies the heat equation on Rn for a continuous and bounded initial
condition g. A function that acts on other functions by convolution is an integral
kernel, and Ht is specifically called the heat kernel on Rn . It is also called the
fundamental solution of the heat equation, for reasons we will explain in Chap. 12.
Theorem 6.2 For a bounded function g C 0 (Rn ), the heat equation

u = 0, u|t=0 = g, (6.19)
t
admits a classical solution given by
u(t, x) = Ht g(x). (6.20)
Proof Explicitly, the formula (6.20) says that

u(t, x) = (4 t) 2 e|x y| /4t
n 2
g( y) d n y. (6.21)
Rn
Because the domain is infinite here, we should treat differentiation under the integral
with some care. To justify this, the key point is that the partial derivatives of Ht can be
estimated by expressions of the form c1 (t, x)ec2 (t,x)| y| , where the dependence of the
2
6.3 Integral Solution Formula 105
constants c1 and c2 on t and x is continuous. We will not go into the technical details,
but this makes it relatively straightforward to check that differentiating under the
integral works in this case. In particular, the fact that Ht (x) solves the heat equation
implies (6.19).
the initial condition, fix some x R . A change of variables to w =
n
To check
( y x)/ t in (6.21) gives

u(t, x) = (4 ) 2 e|w| g x + t 2 w d n w
n 2 1
Rn

1
= H1 (w)g x + t 2 w d n w
Rn
By (6.17) we can also write

g(x) = H1 (w)g(x) d n w.
Rn
Thus the difference we are trying to estimate is

1
u(t, x) g(x) = H1 (w) g x + t 2 w g(x) d n w. (6.22)
Rn
Given > 0, we can use the exponential decay of H1 (w) as |w| to choose
R so that
H1 (w) d n w < .
|w|R
Since g is bounded by assumption, there exists a constant M so that |g| M. The

large-w piece of (6.22) can thus be estimated by

1
H1 (w) g x + t 2 w g(x) d n w 2M. (6.23)
|w|R
On the other hand, by the continuity of g we can choose > 0 so that for y such
that for |x y| < , we have
|g (x y) g(x)| .
Thus for |w| < R and t < 2 /R 2 ,

1
g x + t 2 w g(x) .
It follows that for t < 2 /R 2 ,


1
H1 (w) g x + t 2 w g(x) d n w H1 (w) d n w
|w|<R |w|<R

(6.24)
H1 (w) d n w
Rn
= .
Combining the estimates (6.23) and (6.24) gives
|u(t, x) g(x)| (2M + 1)
for 0 < t < 2 /R 2 . Since was arbitrary, this shows that
lim u(t, x) = g(x).

t0
Without extra restrictions on u, the solution of (6.19) is not necessarily unique.

However, if we start from a bounded initial condition g, then it is physically reason-
able to assume that u is bounded over finite time intervals.
Theorem 6.3 Under the assumption that u(t, ) is bounded on [0, T ] Rn for each
T > 0, the solution of the heat equation (6.19) is unique.
We will develop tools to prove this result (maximum principles) in Chap. 9. The
statement can be improved by weakening the boundedness hypothesis to an assump-
tion of exponential growth. The counterexamples to uniqueness exhibit superexpo-
nential growth and are not considered valid as physical solutions.
In combination, Theorems 6.2 and 6.3, show that a bounded solution of the heat
equation on Rn with continuous initial data satisfies (6.21). The function Ht (x) is C
in both variables for t > 0. As we noted in the proof of Theorem 6.2, differentiation
under the integral is justified in (6.21), so this regularity can be extended to general
solutions.
Theorem 6.4 Suppose that u is a bounded solution of the heat equation (6.19) for
a bounded initial condition g C 0 (Rn ). Then u C ((0, ) Rn ).
Similar regularity results hold for the heat equation in other contexts, for example
on a bounded domain. We will discuss some of these cases later in Sect. 8.6. This
behavior, i.e., smoothness of solutions that does not depend on the regularity of the
initial data, is characteristic of parabolic equations.
Another interesting feature of the heat kernel is the fact that Ht (x) is strictly
positive for all t > 0 and x Rn . This means that if g is nonnegative and not
identically zero, then u is nonzero at all points x Rn for t > 0. Compare this to the
Huygens principle that we observed in Chap. 4, which says that for solutions of the
wave equation the range of influence of a point is limited by the (finite) propagation
speed. The heat equation exhibits infinite propagation speed.
6.3 Integral Solution Formula 107
We can see the origin of the infinite propagation speed in Einsteins diffusion
model from Sect. 6.1. In (6.10) the value D, which is assumed to be constant, is the
average squared displacement per unit time. The fact that D is fixed implies that in
the continuum limit 0, the average absolute displacement per unit time diverges.
Hence the infinite propagation speed is built in to the construction of the model. It
reflects the fact that models of diffusion are inherently statistical, and not expected
to be accurate on a microscopic scale.
6.4 Inhomogeneous Problem
Duhamels method, which was used to construct solutions of the inhomogeneous

wave equation in Sect. 4.4, was originally developed in the context of the heat equa-
tion. There are slight differences in the setup, but the basic idea of translating a
forcing term into an initial condition applies in both settings.
Consider the equation on Rn ,
u
u = f (6.25)
t
for t > 0, with initial condition u(0, x) = 0. For s 0, let s (t, x) be the solution
of the homogeneous heat equation (6.19) for t s, subject to the initial condition

s (t, x) t=s = f (s, x). (6.26)
We claim that the solution is then given by the integral

t
u(t, x) := s (t, x) ds.
0
Using the formula for s provided by Theorem 6.2, the proposed solution can be
written t
u(t, x) = Hts (x y) f (s, y) d n y ds. (6.27)
0 Rn
To justify this formula, we must investigate carefully what happens near the point
t = s.
Theorem 6.5 Assuming that f C 2 ([0, ) Rn ) and is compactly supported,
the formula (6.27) yields a classical solution to the inhomogeneous heat equation
(6.25).
Proof We can see that u is at least C 2 by changing variables in the integral formula
to obtain t
u(t, x) = Hs ( y) f (t s, x y) d n y ds.
0 Rn
Since Hs ( y) is smooth near s = t and f is compactly supported, differentiation

under the integral is justified. This gives
t
u f
(t, x) = Hs ( y) (t s, x y) d n y ds
t Rn t
0
(6.28)
f
+ Ht ( y) (0, x y) d n y,
Rn t
and t
u(t, x) = Hs ( y) x f (t s, x y) d n y ds. (6.29)
0 Rn
Our goal is to integrate by parts in these formulas, to exploit the fact that Hs solves
the heat equation. Here we must be careful, because of the singular behavior of Hs
at s = 0.
To deal with this singularity, we split the integral at s = . For the first integral in
(6.28), switching the t derivative to an s derivative and integrating by parts gives
t
f
Hs ( y) (t s, x y) d n y ds
Rn t
t
f
= Hs ( y) (t s, x y) d n y ds
Rn s
t
Hs
= ( y) f (t s, x y) d n y ds
Rn s

Ht ( y) f (0, x y) d n y + H ( y) f (t , x y) d n y
Rn Rn
The corresponding result for (6.29) has no boundary terms because of the compact
support of f , t
Hs ( y) x f (t s, x y) d n y ds
Rn
t
= y Hs ( y) f (t s, x y) d n y ds
Rn
Applying these integrations by parts to (6.28) and (6.29), and using the fact that
( s )Hs = 0, we obtain

u(t, x) = H ( y) f (t , x y) d n y
t Rn
(6.30)

+ Hs ( y) x f (t s, x y) d y ds.
n
0 Rn t
6.4 Inhomogeneous Problem 109
Since Hs > 0, the second term can be estimated by

Hs ( y) x f (t s, x y) d y ds
n
t
0 Rn

C Hs ( y) d n y ds,
0 Rn

where C is the maximum value of ( t ) f (which exists by the assumption
of compact support). By (6.17), the integral of Hs ( y) over y Rn evaluates to 1,
yielding the estimate

Hs ( y) x f (t s, x y) d n y ds C.
t
0 Rn
We can therefore take 0 in (6.30) to obtain

u(t, x) = lim H ( y) f (t , x y) d n y.
t 0 Rn
The remaining limit is very close to the limit computed in the proof of Theorem 6.2,
except that t is replaced by t . A simple modification of that argument shows that

lim H ( y) f (t , x y) d n y = f (t, x).
0 Rn
This completes the proof that u satisfies (6.25).
6.5 Exercises
6.1 Biological processes and chemical reactions are frequently described by

reaction-diffusion equations, consisting of the heat equation modified by a reaction
term. Consider the simplest such equation,
u
+ u u = 0.
t
on Rn with initial condition u(0, x) = f (x). Assuming f is continuous and bounded,

find a formula for the solution. Hint: use a substitution of the form u eat u to
reduce this to the ordinary heat equation.
6.2 For t 0, x 0, suppose that u(t, x) satisfies the one-dimensional heat
equation (6.11) with the initial condition u(0, x) = 0 for x 0 and the boundary
condition
u(t, 0) = A cos(t)
for t 0. Under the additional requirement that u(t, ) is bounded, find a solution
u(t, x). Hint: use separation of variables and assume that the temporal components
have the form eit .
6.3 Let Rn be a bounded domain with piecewise C 1 boundary. Suppose that

u(t, x) satisfies the heat equation
u
u = 0,
t
on (0, ) . Following the discussion from Sect. 6.1, we define the total thermal
energy at time t by
U[t] = u(t, x) d n x.

(a) Assume that u satisfies Neumann boundary conditions,
u
= 0,

(the insulated case). Show that U is constant.
(b) Assume that u is positive in the interior of and equals 0 on the boundary.
Show that U(t) is decreasing in this case.
6.4 Let Rn be a bounded domain with piecewise C 1 boundary. Suppose that

u(t, x) is real-valued and satisfies the heat equation
u
u = 0,
t
on (0, ) . Define
(t) := u(t, x)2 d n x. (6.31)

(a) Assume that u satisfies the Dirichlet boundary conditions:
u(t, x)| x = 0
for t 0. Show that decreases as a function of t.

(b) Use (a) to show that a solution u satisfying boundary and initial conditions
u|t=0 = g, u| x = h,
for some continuous functions g on and h on , is uniquely determined by

g and h.
Chapter 7
Function Spaces
In the preceding chapters we have seen that separation of variables can generate
families of product solutions for certain PDE. For example, we found families of
trigonometric solutions of the wave equation in Sect. 5.2 and the heat equation in
Sect. 6.1. By the superposition principle, finite linear combinations of these functions
give more general solutions.
It is natural to hope that we could push this construction farther and obtain solu-
tions by infinite series. Solutions of PDE by trigonometric series were studied exten-
sively in the 18th century by dAlembert, Euler, Bernoulli, and others. However,
notions of convergence were not well developed at that time, and many fundamental
questions were left open.
In this chapter we will introduce some basic concepts of functional analysis, which
will give us the tools to address some of these fundamental issues.
7.1 Inner Products and Norms
We assume that the reader has had a basic course in linear algebra and is familiar
with the notion of a vector space, i.e., a set equipped with the operations of addition
and scalar multiplication. The basic finite-dimensional example is the vector space
Rn . This space comes equipped with a natural inner product given by the dot product
v w for v, w Rn . The Euclidean length of a vector v Rn is v := v v.
In this section we will review the corresponding definitions for general real or
complex vector spaces, which include function spaces. One important set of exam-
ples are the spaces C m () introduced in Sect. 2.4, consisting of m-times continuously
differentiable complex-valued functions on a domain Rn . Because differentia-
bility and continuity of functions are preserved under linear combination and scalar
multiplication, C m () is naturally a complex vector space.
An inner product on a complex vector space V is a function of two variables,
u, v V u, v C,
112 7 Function Spaces
satisfying the following properties:
(I1) Positive definiteness: v, v 0 for v V , with equality only if v = 0.

(I2) Symmetry: for v, w V ,
v, w = w, v .
(I3) Linearity in the first variable: for c1 , c2 C and v1 , v2 , w V ,
c1 v1 + c2 v2 , w = c1 v1 , w + c2 v2 , w .
Together, (I2) and (I3) imply conjugate linearity in the second variable,
w, c1 v1 + c2 v2 = c1 w, v1 + c2 w, v2 .
The combination of linearity and conjugate linearity in the respective variables is

called sesquilinearity. In the real case, the complex conjugation can be omitted,
reducing sesquilinearity to bilinearity.
An inner product space is a real or complex vector space V equipped with an
inner product , . The Euclidean inner product on Cn is defined by including a
conjugation in the dot product,
v, w := v w. (7.1)
One way to define an inner product on function spaces is by integration. For example,
on C 0 [0, 1] we could take 1
f, g := f g d x.
0
Certain geometric notions are carried over from Euclidean geometry to inner product
spaces. For example, vectors u, v in an inner product space V are called orthogonal if
u, v = 0.
The analog of length for vectors in V is called a norm. A norm is a function

: V R satisfying the following properties: for all u, v V and scalar ,
(N1) Positive definiteness: u 0 with equality only if u = 0.
(N2) Homogeneity: u = || u.
(N3) Triangle inequality: u + v u + v.
For an inner product space, the definition of the Euclidean length in terms of the dot
product suggests that the function

u := u, u (7.2)
should yield a norm.

7.1 Inner Products and Norms 113
Positive definiteness of (7.2) clearly follows from positive definiteness of the

inner product, and homogeneity follows from sesquilinearity. To see that (7.2) also
satisfies the triangle inequality, we present a relation first derived in the Euclidean
case by the great 19th century analyst Augustin-Louis Cauchy; Hermann Schwarz
later generalized the result to inner product spaces.
Theorem 7.1 (Cauchy-Schwarz inequality) For an inner product space V with

defined by (7.2),
|v, w | v w
for all v, w V .
Proof For v, w V and t R, consider the function

2
q(t) := v + t v, w w ,
The claimed inequality is trivial if w = 0, so assume w = 0. By (I2), (I3), and (7.2),

q(t) = v + t v, w w, v + t v, w w
= v2 + 2t |v, w |2 + t 2 |v, w |2 w2 .
The minimum of this quadratic polynomial occurs at t0 = w2 . Since q 0,
|v, w |2
0 q(t0 ) = v2 ,
w2
which gives the claimed inequality.
The triangle inequality for (7.2) follows from the Cauchy-Schwarz inequality by
u + v2 = u + v, u + v
= u2 + 2 Re u, v + v2
u2 + 2 |u, v | + v2
u2 + 2 u v + v2
= (u + v)2 .
Thus (7.2) defines a norm associated to the inner product. This definition of the norm
is used by default on an inner product space.
It is possible to have a norm that is not associated to an inner product. For example,
this is the case for the sup norm, defined for f C 0 (), with Rn bounded, by

sup | f (x)| := sup | f (x)| ; x . (7.3)
x
We will explain how to tell that a norm does not come from an inner product in the
exercises.
7.2 Lebesgue Integration
In the early 20th century, Henri Lebesgue developed an extension of the classic
definition of the integral introduced by Bernhard Riemann in 1854. (Riemanns is
the version commonly taught in calculus courses.) Lebesgues definition agrees with
the Riemann integral when the latter exists, but extends to a broader class of integrable
functions.
A full course would be needed to develop this integration theory properly. In this
section, we present only a brief sketch of the Lebesgue theory, with the focus on the
features most relevant for applications to PDE.
The Lebesgue integral is based on a generalized notion of volume for subsets of
Rn , which can be defined in terms of approximation by rectangles. For a rectangular
subset in R Rn , let vol(R) denote the usual notion of volume, the product of
the lengths of the sides. (It is conventional to use volume as a general term when
the dimension is arbitrary.) The volume of a subset A Rn can be overestimated
by covering the set with rectangles, as illustrated in Fig. 7.1. The (n-dimensional)
measure of A is defined by taking the infimum of these overestimates,

m n (A) := inf vol(R j ); A Rj . (7.4)
j=1 j=1
For a bounded region with C 1 boundary, the definition (7.4) reproduces the notion of
volume used in multivariable calculus. Note that the concept of measure is dependent
on the dimension. The measure of a line segment in R1 is the length, but a line segment
has measure zero in Rn for n 2.
There is a major technicality in the application of (7.4). In order to make the
definition of measure consistent with respect to basic set operations, we cannot apply
it to all possible subsets of Rn . Instead, the definition is restricted to a special class
of measurable sets. Lebesgue gave a criterion for measurability that rules out certain
exotic sets for which volume is ill-defined. Fortunately, these sets are so exotic that
we are unlikely to encounter them in normal usage. All open and closed sets in Rn
are included in the measurable category, as are any sets constructed from them by
basic set operations of union and intersection.
7.2 Lebesgue Integration 115
Fig. 7.1 Covering a set with

rectangles
The characteristic function of a set A Rn is defined by

1, x A,
A (x) := (7.5)
0, otherwise.
The measure can be used to define the integral of a characteristic function,

A d n x := m n (A),
Rn
provided A is a measurable set. The integral of a general function is then built

from approximations by linear combinations of characteristic functions. In order to
construct these approximations, we need to use a restricted class of functions. A
function f : C is called measurable if the preimage f 1 (R) is a measurable
subset of for every rectangle R C. Every Riemann-integrable function is
measurable in the Lebesgue sense, so the measurable class includes all functions
encountered in a traditional calculus class. Henceforth, whenever we write f :
C or R, we will assume implicitly that f is measurable.
With this basic picture in mind, we will ask the reader to accept certain important
consequences of the Lebesgue definition without further justification. In examples
and exercises we will confine our attention to functions for which ordinary
Riemannian integrals exist.
It is standard practice when working with function spaces related to integration
to make an equivalence:
f g f = g except on a set of measure zero. (7.6)
For example, in R the characteristic functions of the intervals (a, b) and [a, b] are
equivalent. In measure theory, a property is said to hold almost everywhere if it fails
only on a set of measure zero. The equivalence (7.6) amounts to identifying functions
that agree almost everywhere.
If functions f and g satisfy

| f g| d n x = 0,
then there is no way to distinguish them in terms of integration. The definition (7.6)
is motivated by the following:
Lemma 7.2 For measurable functions f, g : C with Rn ,

| f g| d n x = 0

if and only if f g.
7.3 L p Spaces
A function f : C is defined to be integrable if its integral converges absolutely,

i.e.,
| f | d n x < .

For p 1, we define the space of p-integrable functions by

L () :=
p
f : C; |f| d x < ,
p n
(7.7)

with the understanding that functions in L p are identified according to the equivalence
(7.6). The space L p () is clearly closed under scalar multiplication. Closure under
addition is a consequence of the convexity of the function x |x| p for p 1,
which implies the inequality

f + g p | f | p + |g| p

2 2
.
Hence L p () is a complex vector space for p 1.

The L p norm is defined by
1p
f p := | f |p dn x .

To check that this is really a norm, we first note that Lemma 7.2 implies positive
definiteness (N1) because of the equivalence relation (7.6). Homogeneity (N2) is
satisfied because of cancellation between the powers p and 1/ p.
7.3 L p Spaces 117
Fig. 7.2 Step function
The triangle inequality (N3) is clear for p = 1 because | f + g| | f | + |g|. And

for p = 2 it follows from Cauchy-Schwarz inequality, because 2 is associated to
the inner product
f, g := f g d n x. (7.8)

In general the L p triangle inequality,
f + g p f p + g p ,
is called the Minkowski inequality and holds for p 1. We omit the proof because
we are mainly concerned with the cases L 1 and L 2 .
Example 7.3 To illustrate the distinction between the L p norms, consider the func-
tion
h := a[0,l] ,
for a, l > 0, as shown in Fig. 7.2.

For general p 1,
l 1/ p
h p = a pd x
0 (7.9)
= al 1/ p
.
If we think of h as a density function, then the L 1 norm gives the total mass h1 = al.
The sensitivity of p to the spread of the function decreases as p increases, as
illustrated by the fact that
lim h p = a,
p
For large p, the L p norms increasingly become measures of local concentration

rather than mass.
Example 7.3 suggests the possibility of defining a space L that is a limiting case
of the L p spaces, with a norm that generalizes the sup norm (7.3). The sup norm itself
does not respect the equivalence (7.6), so we must modify the definition to define a
norm consistent with the other L p spaces.
For a function h : R, the essential supremum is

ess-sup(h) := inf a R; {h > a} has measure zero . (7.10)
Note that {h > a} has measure zero precisely when h is equivalent to a function
bounded by a. The value ess-sup(h) is thus the least upper bound among all functions
equivalent to h. For continuous functions the essential supremum reduces to the
supremum.
For f : C, we define
f := ess-sup | f | . (7.11)
The normed vector space L () consists of functions which are essentially

bounded,
L () := f : C; f < , (7.12)
subject to the equivalence (7.6).

Collectively, the L p spaces play a vital role in the analysis of PDE. The different
norms can be thought of as a collection of measuring tools. Although the full toolkit
is needed for many applications, for this book we will rely on the cases p = 1, 2,
or .
Example 7.4 The Schrdinger equation in Rn describes the evolution of a quantum-
mechanical wave function (t, x):

i = .
t
In Exercise 4.7 we saw that solutions have constant spatial L 2 norms,
(t, )2 = (0, )2 ,
which corresponds to the conservation of total probability. On the other hand, solu-
tions also satisfy a dispersive estimate
(t, ) Ct n/2 (0, )1 ,
for all t > 0, with C a dimensional constant. The norm on the left measures the peak
amplitude of the wave. By the estimate on the right, this amplitude is bounded in
terms of the mass and decays as a function of time. In general, dispersive estimates
describe the spreading of solutions as a function of time.
It is conventional to represent elements of L p as ordinary functions, even though
each element is actually an equivalence class of functions identified under (7.6).
This usually causes no trouble because equivalent functions give the same results in
integrals.
7.3 L p Spaces 119
One point that requires clarification, however, is the issue of continuity or differ-
entiability of functions in L p . Under (7.6), a C m function is equivalent to a class of
functions which are not even continuous. To account for this technicality, we adopt
the convention that if a function in L p is equivalent to a continuous function, then the
continuous representative is used by default. This is unambiguous because the con-
tinuous representative is unique when it exists. Under this convention, the statement
that f L p is a C m function really means that f admits a continuous representative
which is C m .
7.4 Convergence and Completeness
In a normed vector space V , convergence of a sequence vn v means
lim vn v = 0. (7.13)

n
We might also write this as

v = lim vn ,
provided the choice of norm is clear.

It frequently proves useful to approximate L p functions by smooth functions. For
p 1 there is a natural inclusion

Ccpt () L p (),
because continuous functions on a compact set are bounded. The Lebesgue theory
gives the following:
Theorem 7.5 Assume 1 p < . For a function f L p () there exists an

approximating sequence {k } Ccpt (), such that
lim k f p = 0.
k
A subset W of a normed vector space V is called dense if every v V can be

obtained as a limit of a sequence in W . Theorem 7.5 thus states that Ccpt () is dense
in L () for p [1, ).
p
In PDE applications, a common method of proving the existence of a solution is to

construct a sequence of approximate solutions, and then establish convergence of this
sequence with respect to an appropriate norm. We cannot simply use the definition
(7.13) to check convergence in this situation, because the limiting function may not
exist. It is therefore crucial to be able to deduce convergence using only the sequence
itself.
Fig. 7.3 A Cauchy sequence

with respect to 1
The most useful tool for this purpose is a slightly weaker form of convergence. A
sequence {vk } V is said to be Cauchy if the difference between elements converges
to zero: given > 0 there exists an N such that k, m N implies
vk vm < .
This Cauchy condition is sometimes written as a double limit,
lim vk vm = 0.
k,m
Every convergent sequence is Cauchy. This is because the triangle inequality

implies
vk vm = vk v + v vm
vk v + v vm .
If the sequence converges then the terms on the right are arbitrarily small for k and
m sufficiently large.
In Rn , it follows from the completeness axiom for real numbers that all Cauchy
sequences are convergent. (See Theorem A.3.) This property does not necessarily
hold in a general normed vector space, as the following demonstrates.
Example 7.6 Consider the space C 0 [1, 1] equipped with the L 1 norm 1 . For
n N define the functions

1, x < n ,
1
f n (x) = nx n x n1 ,
1

1, x > n1 ,
as illustrated in Fig. 7.3.

We can see that the sequence { f n } is Cauchy by computing
1
f k f m 1 = | fk fm | d x
1

1 1
= .
k m
7.4 Convergence and Completeness 121
However, for f C 0 [1, 1],

0 1
lim f k f 1 = | f + 1| d x + | f 1| d x.
k 1 0
This limit equals 0 only if f (x) = 1 for x < 0 and f (x) = 1 for x > 0. That
is not possible for f continuous. Therefore the sequence { f n } does not converge in
C 0 [1, 1].
A normed vector space V is complete if all Cauchy sequences in V converge
within V . Theorem A.3 implies that Euclidean Rn is complete in this sense. For L p
spaces the Lebesgue integration theory gives the following result.
Theorem 7.7 For a domain Rn , the normed vector space L p () is complete
for each p [1, ].
In functional analysis, a complete normed vector space is called a Banach space
and a complete inner product space is called a Hilbert space. Thus Theorem 7.7 could
be paraphrased as the statement that L p () is a Banach space. The inner product
space L 2 () is a Hilbert space.
A subspace W V is closed if it contains the limit of every sequence in W that
converges in V .
Lemma 7.8 If V is a complete normed vector space and W V is a closed sub-
space, then W is complete with respect to the norm of V .
Proof Suppose {wk } W is a Cauchy sequence. The sequence is also Cauchy in
V , and so converges to some v V by the completeness of V . Since W is closed,
v W.
The L p function spaces have discrete counterparts, denoted by p , whose elements
are sequences. To a sequence (a1 , a2 , . . . ) of complex numbers we associate the
function a : N C defined by j a j . The p norm of this function is
1
p p
a p := a j ,
j=1
The corresponding vector spaces are

p (N) := a : N C; a p < , (7.14)
for p 1. It is possible to prove directly that p (N) is complete, but this can also be
deduced easily from Lemma 7.8. We interpret p (N) as a closed subspace of L p (R)
consisting of functions which are constant on each interval [ j, j + 1) for j N
and zero on (, 0). On this subspace the L p norm reduces to the p norm, so that
Lemma 7.8 implies that p (N) is complete. In particular, 2 (N) is a Hilbert space
with the inner product
a, b 2 := ajbj.
j=1
7.5 Orthonormal Bases
Let H be an infinite-dimensional complex Hilbert space. A sequence of vectors

{e1 , e2 , . . . } H is orthonormal if

1, j = k,
e j , ek = (7.15)
0, j = k,
for all j, k N. An orthonormal basis for H is an orthonormal sequence such that

each v H admits a unique representation as a convergent series,

v= cjej, (7.16)
j=1
with c j C.
As we will see in Sect. 7.6, the sets of eigenfunctions of certain differential opera-
tors naturally form orthonormal sequences with respect to the L 2 inner product. For
example the sine eigenfunctions appearing in Theorem 5.2 have this property. If a
sequence of eigenfunctions forms a basis, then we can expand general functions in
terms of eigenfunctions.
Suppose we are given an orthonormal sequence e j H , and we would like to
show that this forms a basis. To represent an element v H in the form (7.16), we
must decide how to choose the coefficients c j . This works in much the same way as
it does in finite dimensions. By the orthonormality property (7.15), we can compute
that

n
c j e j , ek = ck (7.17)
j=1

for all n k. Assuming that c j e j converges to v in H , we can take the limit
n in (7.17) to compute
v, ek = ck . (7.18)
Based on this calculation, we assign coefficients to v by setting

c j [v] := v, e j . (7.19)
7.5 Orthonormal Bases 123
The corresponding partial sums for n N are denoted by
n
Sn [v] := c j [v]e j . (7.20)
j=1
The condition that {e j } is a basis is equivalent to the convergence of Sn [v] v in

H for every v H .
Theorem 7.9 (Bessels inequality) Assume that {e j } is an orthonormal sequence in

2
an infinite-dimensional Hilbert space H . For v H , the series c j [v] converges
and the limit satisfies

c j [v]2 v2 .
j=1
Equality holds if and only if Sn [v] v in H .
Proof Using the sesquilinearity (I3) of the inner product, we can expand

v Sn [v]2 := v Sn [v], v Sn [v]
= v, v Sn [v], v v, Sn [v] + Sn [v], Sn [v] ,
for n N. By the definition (7.20) of Sn [v] and the orthonormality condition (7.15),
n

Sn [v], v = v, Sn [v] = Sn [v], Sn [v] = c j [v]2 .
j=1
We thus conclude that

n

v Sn [v]2 = v2 c j [v]2 . (7.21)
j=1
Since the left-hand side is positive, the identity (7.21) shows that
n

c j [v]2 v2 ,
j=1
2
for all n N. The partial sums of the series c j [v] are thus bounded and the
terms are all positive. Hence the series converges by the monotone sequence theorem,
to a limit satisfying the claimed bound,

c j [v]2 v2 .
j=1
To complete the proof, note that Sn [v] v in H means that the limit as n
of the left-hand side of (7.21) is zero. Hence Sn [v] v if and only if
n

lim c j [v]2 = v2 .
n
j=1
The combination of completeness and Bessels inequality leads to an alternative

characterization of a basis that is easier to apply.
Theorem 7.10 Suppose H is an infinite-dimensional Hilbert space. An orthonor-
mal sequence in H forms a basis if and only if 0 is the only element of H that is
orthogonal to all vectors in the sequence.
that {e j } forms a basis, so that every v H can be written as a

Proof Assume first
convergent sum c j [v]e j . If v is orthogonal to all of the vectors e j , then c j [v] = 0
for all j by (7.19). Hence v = 0.
To establish the converse statement, let {e j } be an orthonormal sequence. For
v H , Bessels inequality implies

c j [v]2 v2 < . (7.22)
j=1
For n m,

2
m
Sm [v] Sn [v]2 = c j [v]e j

j=n+1

m

= c j [v]2 .
j=n+1
Hence (7.22) implies that

2
lim Sm [v] Sn [v] = 0,
m,n
meaning that the sequence {Sn [v]} is Cauchy in H . By completeness of H this implies
that Sn [v] v for some v H .
7.5 Orthonormal Bases 125
Now assume that 0 is the only vector orthogonal to e j for all j. For n j we
have
v Sn [v], e j = v, e j Sn [v], e j
= c j [v] c j [v]
= 0.
Taking the limit as n with j fixed gives

e j = 0.
v v,
Thus v v is orthogonal to every e j , implying that v = v.

This proves that Sn [v] v
in H for each v H , and thus {e j } is a basis.
7.6 Self-adjointness
The process of forming a basis from eigenvectors of an operator should be familiar

from linear algebra; for a finite-dimensional matrix this is called diagonalization.
Let us briefly recall the basic facts for the finite-dimensional case. A complex n n
matrix A is self-adjoint (also called Hermitian) if the matrix is equal to its conjugate
transpose. In terms of the Euclidean inner product (7.1) this means precisely that
Au, v = u, Av (7.23)
for all u, v Cn . (In the real case self-adjoint is the same as symmetric.)
The spectral theorem in linear algebra says that for a self-adjoint matrix A there
exists an orthonormal basis for Cn consisting of eigenvectors for A, with real eigen-
values. Functional analysis allows a powerful extension of this result, that applies
in particular to certain differential operators acting on L 2 spaces. The full spectral
theorem for Hilbert spaces is too technical for us to state here, but we will prove a
version of this for the Laplacian on bounded domains later in Sect. 11.5.
Self-adjointness remains important as a hypothesis for the more general spectral
theorem, but even this condition becomes rather technical in the Hilbert space set-
ting. The issues arise from the fact that differentiable operators cannot act on the
whole space L 2 () because L 2 functions need not be differentiable. We will avoid
these complexities, by focusing on the Laplacian and restricting our attention to C 2
functions.
Lemma 7.11 Suppose that Rn is a bounded domain with C 1 boundary. If
u, v C 2 () both satisfy either Dirichlet or Neumann boundary conditions on ,
then
u, v = u, v . (7.24)
Proof By Greens first identity (Theorem 2.10),

v u
uv vu d x =
n
u v d S. (7.25)

The Dirichlet conditions require that
u| = v| = 0,
implying the vanishing of the right-hand side of (7.25). Similarly, the Neumann
conditions
u v
= =0 (7.26)

also imply that the integrand on the right vanishes.
Boundary conditions for which (7.24) holds are called self-adjoint boundary con-
ditions (for the Laplacian). Formally, (7.24) resembles the matrix condition (7.23),
but of course there is no analog of boundary conditions in the matrix case. The proper
definition of self-adjointness in functional analysis involves a more precise speci-
fication of the domain on which acts and (7.24) holds. Even without going into
these details, we can still draw some meaningful conclusions from Lemma 7.11.

Lemma 7.12 Suppose j is a sequence of eigenvalues of on a bounded
domain Rn , with eigenvectors in C 2 () subject to a self-adjoint boundary
condition. Then j R and, after possible rearrangement, the eigenvectors form an
orthonormal sequence in L 2 ().
Furthermore, j > 0 for Dirichlet conditions, and j 0 for Neumann.

Proof Suppose we have a sequence j C 2 () satisfying
j = j j .
The condition (7.24) implies that for j, k Z,

j , k = j , k .
By the eigenvalue property this reduces to

!
j k j , k = 0. (7.27)
2
For j = k the inner product equals j 2 > 0, implying that j R for all j.
can thus drop the conjugation in (7.27). If j = k , then this now implies that
We
j , k = 0.
7.6 Self-adjointness 127
If some of the j s are equal, then every linear combination of the corresponding
of eigenfunctions will still be an eigenfunction for the same value of j . Hence we
can rearrange the eigenfunctions sharing a common eigenvalue into an orthogonal
set using the Gram-Schmidt procedure from linear algebra.
By multiplying the eigenfunctions by constants we can normalize so that j 2 =
1. The divergence theorem (Theorem 2.6) then implies

j = j , j

j
= j 2 d n x j d S.

Either Dirichlet or Neumann conditions will cause the second term to vanish, imply-
ing that j 0. If j = 0 then the equation also shows that j 0, implying that
j is constant. In the Dirichlet case the only constant solution is trivial, j 0, but
for Neumann conditions a nonzero constant is possible.
Example 7.13 In Example 5.5 we found a set of eigenfunctions for a circular drum-
head modeled by the unit disk, with Dirichlet boundary conditions. The eigenfunc-
tions were given in polar coordinates by
k,m (r, ) := eik Jk ( jk,m r ), k Z, m N,
where jk,m is the mth positive zero of the Bessel function Jk . The eigenvalues of
in this case are the values jk,m . Since the only possible matches among the Bessel
zeros are jk,m = jk,m , these are the only potential non-orthogonal pairs.
Let us examine the orthogonality condition more explicitly. In polar coordinates,
the L 2 inner product of two eigenfunctions is given by

1 2
k,m , k ,m L2
= k,m (r, )k ,m (r, ) r d dr
0 0
1 2

= ei(kk ) Jk ( jk,m r )Jk ( jk ,m r ) r d dr.
0 0
Note that the eigenfunctions are clearly orthogonal when k = k , because the
integral vanishes in this case. If we set k = k , then the integral is trivial and the
inner product becomes

1
k,m , k,m L2
= 2 r Jk ( jk,m r )Jk ( jk,m r ) dr.
0
By Lemma 7.12 this integral vanishes for m = m . The cancellations occur because
of the oscillations, just as for sine functions, as Fig. 7.4 illustrates.
Fig. 7.4 Radial components =1 2 3 4

J1 ( j1,m r ) of orthogonal
eigenfunctions on the disk
7.7 Exercises
7.1 A norm on a vector space V satisfies the parallelogram law if
v + w2 + v w2 = 2 v2 + 2 w2 ,
for all v, w V .
(a) Show that a norm defined by an inner product as in (7.2) satisfies the parallelo-
gram law.
(b) In L p (R), define the functions
f (x) = [0,2] , g(x) = [0,1] [1,2] .
Use these to show that the parallelogram law fails for p if p = 2.

(c) Find an example to show that the parallelogram law fails for the sup norm (7.3).
7.2 Consider the sequence of functions on R defined by

nen x , x 0,
2
f n (x) =
0, x < 0.
Show that f n 0 in L 1 (R) but not in L 2 (R).
7.3 Consider the sequence of functions on R defined by
gn (x) = n 1 [0,n] .
Show that gn 0 in L 2 (R) but not in L 1 (R).

7.7 Exercises 129
7.4 Consider the sequence f n (x) = x n for x (0, 1). Show that f n 0 in L p (0, 1)
for each p [1, ), but not for p = .
7.5 Assume that Rn is a bounded domain. Show that there is a constant C > 0
such that for f L 2 (),
f 1 C f 2 .
This implies in particular that L 2 () L 1 (). Find an example to show that this
result does not hold for unbounded.
7.6 As an application of the Cauchy-Schwarz inequality, we can use the quantity

defined in Exercise 6.4 to show that solutions of the heat equation with fixed
boundary values are uniquely determined by the values at time t = T > 0. Under
the hypotheses from that exercise, suppose that u solves the heat equation with
u|t=T = 0, u| x = 0.
The goal is to show that these assumptions imply u = 0 for all t.
(a) Use the Cauchy-Schwarz inequality to deduce that

2
u n
(t)2 4(t) d x.

t
where is defined as in (6.31).

(b) Show that
2
u n

(t) = 4 d x,

t
so that the inequality from (a) becomes
(t)2 (t) (t). (7.28)
(c) Suppose that (0) > 0. Then by continuity log (t) is defined at least in some
neighborhood of t = 0. Using (7.28), show that
(log (t)) 0.
This implies that log (t) is bounded below by its tangent lines. In particular
(0)
log (t) log (0) + t,
(0)
which implies
(t) (0)ect ,
for c = (0)/(0) > 0. Thus if (0) > 0 then is strictly positive for all
t 0.
(d) Conclude from (c) that if (T ) = 0, then (t) = 0 for all t, and deduce that
u = 0.
7.7 Recall the radial decomposition formula (2.10). We can use this to get a basic
picture of the degree of singularity or decay at infinity that is allowed in each L p .
(a) For R consider the function

r , r 1,
g(x) :=
0, r 1.
For what values of and p [1, ] is g L p (Rn )?

(b) For R consider the function

0, r 1,
h(x) :=
r , r 1.
For what values of and p [1, ] is h L p (Rn )?
7.8 Consider the eigenfunctions given by (5.5) with = .
(a) Show that "

2
n (x) := sin(nx), n N,

defines an orthonormal sequence

# in L 2 (0, ). (Hint:$ recall the trigonometric
identity sin() sin() = 2 cos( ) cos( + ) .)
1
(b) For the function u 1, compute the corresponding expansion coefficients,
ck [1] := 1, k .
Under Theorem 7.9, what explicit summation condition corresponds to the con-
vergence Sn [1] 1 in L 2 (0, )?
Chapter 8
Fourier Series
In his study of heat flow in 1807, Fourier made the radical claim that it should be
possible to represent all solutions of the one-dimensional heat equation by trigono-
metric series. As we noted in the introduction to Chap. 7, trigonometric series had
been studied earlier by other mathematicians. Fouriers innovation was to suggest
that the general solution could be obtained this way.
This claim proved difficult to resolve, because the tools of functional analysis that
we discussed in Chap. 7 were not yet available in Fouriers time. Indeed, the difficult
problem of Fourier series convergence provided some of the strongest motivation for
the development of these tools.
In this chapter we will analyze Fourier series in more detail, and show that the
Fourier approach yields a general solution for the one-dimensional heat equation. The
primary significance of this approach to PDE is the philosophy of spectral analysis
that it inspired. The decomposition of functions with respect to the spectrum of
a differential operator is a tool with enormous applications, both theoretical and
practical.
8.1 Series Solution of the Heat Equation
Consider the heat equation

u
u = 0, (8.1)
t
on a domain Rn , with Dirichlet or Neumann boundary conditions. According

to Lemma 5.1 the product solutions of (8.1) have the form
u(t, x) = v(t)(x),

132 8 Fourier Series
where solves the Helmholtz equation
=
on . The temporal equation is a simple ODE,
v
= v,
t
with the family of solutions
v(t) = v(0)et . (8.2)
Let us assume that the equation for admits a sequence of solutions k , with
eigenvalues k . We have seen specific examples of this in Chap. 5, including the
trigonometric case in Theorem 5.2. By (8.2), the corresponding product solutions of
the heat equation are
u k (t, x) := ek t k (x).
Fouriers strategy calls for us to express the general solution as a series,

u(t, x) = an en t n (x), (8.3)
n=1
for some choice of coefficients an . To fix the coefficients an in (8.3) we assume an

initial condition u(0, x) = h(x). Setting t = 0 gives

h(x) = an n (x). (8.4)
n=1
If we can show that {n } forms an orthonormal basis of L 2 (), then this gives us
a way to assign coefficients to h such that (8.4) holds, at least in the sense of L 2
convergence.
Even if the orthonormal basis property is established, some big issues still remain.
The fact that each term u k satisfies the heat equation does not guarantee that u does,
because of the infinite series summation. Similarly, the limit of (8.3) as t 0
is not necessarily (8.4), because the limit cannot necessarily be taken inside the
summation. In this chapter we will explain how to resolve these problems in the
context of trigonometric series.
Example 8.1 Consider the case of a one-dimensional metal rod with insulated ends.
For convenience take the length to be , so that x [0, ] and the Neumann boundary
conditions are
u u
(t, 0) = (t, ) = 0.
x x
8.1 Series Solution of the Heat Equation 133
Let us consider the initial condition
h(x) = 3 x 2 2x 3 , (8.5)
as pictured on the left in Fig. 8.1.

The boundary conditions are satisfied by cosines,
n (x) = cos(nx), n N0 .
Hence the strategy outlined above calls for us to represent the initial condition as a
series

h(x) = an cos(nx). (8.6)
n=0
To choose the coefficients, we recall the discussion of basis expansion from

Sect. 7.5. The cosines satisfy an orthogonality condition with respect to the L 2 inner
product on [0, ],

0 m = n,
cos(mx) cos(nx) d x = m = n = 0, . (8.7)
0

/2 m = n 1,
This could be checked with trigonometric identities, but it is perhaps easier is to use
the complex form cos(kx) = 21 (eikx + eikx ).
Since the sequence {n } is not normalized, the coefficient formula (7.18) must be
interpreted as
1
an = h, n
.
n 2
By (8.7) the Fourier coefficients are thus given by

1
a0 = h(x) d x,
0
(8.8)
2
an = h(x) cos(nx) d x, n 1.
0
After substituting (8.5) into (8.8), integration by parts yields

3

2, n = 0,
an = n 4 , n 1, odd,
48
(8.9)

0, n 2, even.
3
h(x) = 3x2 2x3 2
48

cos x
Fig. 8.1 Comparison of the initial condition h(x) with the first two terms of its cosine series
Figure 8.1 shows a comparison between h and the partial sum S1 [h]. The close match
between these functions is clearly evident. And since the higher coefficients decay
by a factor of n 4 , convergence of this series seems quite plausible. The resulting
solution would be given by
3 48
en t cos(nx).
2
u(t, x) =
2 nN
n 4
odd
Note that the convergence rate improves dramatically as t increases.
8.2 Periodic Fourier Series
We saw examples of Fourier series based on sines in Theorem 5.2 and cosines in
Example 8.1. To account for both cases, it is convenient to consider periodic functions
on R. We define
T := R/(2 Z), (8.10)
where the quotient notation means that points separated by an integer multiple of
2 are considered equivalent. The space C m (T) consists of the functions in C m (R)
which are 2 -periodic.
Integrals of functions on T are defined by restricting the range of integration to
an arbitrary interval of length 2 in R. We will write the inner product on L 2 (T) as

f, g
= f g d x,

but the range of integration could be shifted if needed.

8.2 Periodic Fourier Series 135
The Helmholtz equation on T is
2
= ,
x2
for C 2 (T), with no need for additional boundary conditions because of the
periodicity. The eigenfunctions are the complex exponentials
k (x) := eikx , (8.11)
for k Z, with k = k 2 .
It is possible to recover cosine and sine Fourier series from the periodic case,
by restricting our attention to even or odd functions on T. We will demonstrate this
specialization in the examples and exercises.
The complex exponentials satisfy a simple orthogonality relation,

k , l
= ei(kl)x d x

2, k = l,
=
0, k = l.
We did not include a normalizing factor in (8.11), so k 2 = 2 and the Fourier

coefficients of an integrable function f L 1 (T) are defined by
1
ck [ f ] := f, k
2
(8.12)
1
= f (x)eikx d x.
2
Because the index set is Z rather than N, we define the partial sums of the periodic
Fourier series by truncating on both sides,

n
Sn [ f ](x) := ck [ f ]eikx . (8.13)
k=n
For the sequence {k }, Bessels inequality (Theorem 7.9) takes the form

ck [ f ] 2 1 f 2 , (8.14)
2
2
kZ
with equality if and only if Sn [ f ] f in L 2 (T).

In the specific example considered in Example 8.1, the Fourier series appeared
to converge very quickly. To illustrate potential complications with the convergence,
let us consider a function with a jump discontinuity.
Example 8.2 On the interval [0, ], define the function

0, x [0, 2 ],
h(x) =
1, x ( 2 , ],
as pictured on the left in Fig. 8.2. As noted above, in order to represent h as a cosine
series using the periodic eigenfunctions, we first extend h to T as an even function,
i.e.,
0, x [ 2 , 2 ] + 2 Z,
h(x) :=
1, x ( 2 , 3
2
] + 2 Z.
By (8.12), with a shift to the more convenient interval [0, 2 ], the Fourier coefficients
of h are 3
1 2
ck [h] = eikx d x
2 2

1
, k = 0,
= 2(1)k
k
k
sin 2 , k = 0.
Since ck [h] = ck [h], we can combine terms in the partial sums (8.13) to give
n
1 (1)k k
Sn [h](x) = +2 sin cos(kx).
2 k=1
k 2
Figure 8.2 shows a sample of these partial sums. In contrast to the case of Exam-
ple 8.1, where 2 terms of the Fourier series were enough to give a very convincing
approximation, we can see significant issues with convergence in the vicinity of the
jump, even with 40 terms.
The Fourier series computed in Example 8.2 makes for a good illustration of some
different notions of convergence. Consider the sequence of differences h Sn [h], as
h(x) S20 [h](x) S40 [h](x)
Fig. 8.2 Fourier series expansion for a step function

8.2 Periodic Fourier Series 137
n = 20 n = 40 n = 100
Fig. 8.3 Plots of the differences h Sn [h]
illustrated in Fig. 8.3. If {k } forms an L 2 (T) basis, then we would have

lim h Sn [h] 2 d x =
?
0.
n
It is not easy to judge such a limit visually, but this claim is true, as we will prove in
Sect. 8.6.
We could instead focus our attention the values of Sn [h](x) for some fixed x. A
sequence of functions f n is said to converge pointwise to f (assuming these functions
have a common domain) if for each fixed x in the domain,
lim f n (x) = f (x).

n
In Fig. 8.3, if we focus our attention on some point x away from the center, then the
bumps at this point do seem to be decreasing in size as n gets larger. We will verify
in Sect. 8.3 that this Fourier series converges pointwise except at x = 2 .
Another feature that is quite apparent in Fig. 8.3 is the spike near the center. It is
possible to prove that such a spike persists, with height essentially constant, for all
values of n. The historical term for this effect, which is caused by the jump discon-
tinuity, is the Gibbs phenomenon. It was actually first observed in 1848 by Henry
Wilbraham, but remained generally unknown until it was rediscovered independently
by J. Willard Gibbs in 1899.
The Gibbs phenomenon relates to yet a third definition of convergence. A sequence
of bounded functions f n is said to converge uniformly to a function f on a set W if

lim sup f n (x) f (x) = 0. (8.15)
n xW
In the cases plotted in Fig. 8.3 we can see that

1

sup h(x) Sn [h](x) .
xT 2
Since this does not decrease, uniform convergence fails for this series. However, the
sequence does converge uniformly on domains that exclude a neighborhood of the
jump point, for example on the interval [0, 2 ] for > 0.
8.3 Pointwise Convergence
The basic theory of pointwise convergence of Fourier series was worked out by
Dirichlet in the mid-19th century. In this section we will establish a criterion for
pointwise convergence of periodic Fourier series.
Theorem 8.3 Suppose f L 2 (T), and that for x T the estimate,

f (x) f (x y)
ess-sup < ,
(8.16)
y[,] y
holds for some > 0. Then
lim Sn [ f ](x) = f (x).

n
The essential supremum was defined in (7.10). The inequality (8.16) means that,
after possibly replacing f by an equivalent function in the sense of (7.6), we can
assume that
f (x) f (x y)
sup < .
(8.17)
0<|y| y
This bound holds automatically for f C 1 (T), by the estimate

f (x) f (x y) 1 x
= f (t) dt
y y xy

sup f (t) .
tT
Thus Theorem 8.3 shows that the Fourier series a C 1 function converges pointwise
on all of T. The same argument can be extended to functions on T which are merely
piecewise C 1 .
It is possible to prove pointwise convergence with a weaker hypothesis than
that of Theorem 8.3. However, there are counterexamples, discovered by Fejr and
Lebesgue, that show that pointwise convergence of the Fourier series may fail for
f C 0 (T).
Before getting into the proof of Theorem 8.3, let us consider the structure of the
partial sums in more detail. Plugging the coefficient formula (8.14) into (8.13) gives

n 2
1
Sn [ f ](x) = eikx f (y)eiky dy
k=n
2 0
(8.18)
1 ik(xy)
2 n
= f (y) e dy.
0 2 k=n
8.3 Pointwise Convergence 139
The function that appears in parentheses is called the Dirichlet kernel,
1 ikt
n
Dn (t) := e . (8.19)
2 k=n
With this definition the formula for the partial sum becomes
2
Sn [ f ](x) = f (y)Dn (x y) dy. (8.20)
0
This could be written as a convolution,
Sn [ f ] = f Dn .
Because the sum (8.19) is finite, it is clear that the Dirichlet kernel is a smooth
function on T. It is also easy to compute that
2
Dn (t) dt = 1 (8.21)
0
for n N, since only the k = 0 term in (8.19) contributes to the integral. Applying
the polynomial identity
z m+1 1
1 + z + z2 + + zm = .
z1
to (8.19) with z = eit gives the explicit formula
1 ei(n+1)t eint
Dn (t) = . (8.22)
2 eit 1
Factoring eit/2 out of the numerator and denominator reduces this to
1 sin((n + 21 )t)
Dn (t) = ,
2 sin( 21 t)
which makes it clear that Dn is real-valued.

A plot of Dn (y) for various values of n, as shown in Fig. 8.4, gives some intuition
as to why we might expect (8.23) to converge to f (x) as n . The function
Dn (y) concentrates at y = 0, and oscillates with increasing frequency away from
this point. These oscillations will cause cancellation as n , except at y = 0.
Proof of Theorem 8.3 Because both f and Dn are periodic, a change of variables
y x y allows us to rewrite the convolution in the opposite order:
D10 D20 D50
Fig. 8.4 The Dirichlet kernel for increasing values of n

Sn [ f ](x) = Dn (y) f (x y) dy. (8.23)

Thus, by (8.21) and (8.23),

f (x) Sn [ f ](x) = [ f (x) f (x y)]Dn (y) dy.

Substituting in with the explicit formula (8.22) for Dn (t) gives

1 f (x) f (x y) i(n+1)y
f (x) Sn [ f ](x) = e einy dy. (8.24)
2 e 1
i y
The crucial observation here is that if we separate the terms inside the brackets, then
this looks like a formula for Fourier coefficients.
Assuming that the hypothesis of the theorem is satisfied at x T, consider the
function
f (x) f (x y)
h(y) := , (8.25)
ei y 1
defined for y T with y = 0. We can split this into factors as
f (x) f (x y) y
h(y) = ,
y ei y 1
note that the first factor is essentially bounded near y = 0 by the assumption (8.16).
Since ei y 1 i y as y 0 by Taylors approximation, the second factor is also
bounded near y = 0. The hypothesis (8.16) thus guarantees that h(y) is equivalent to
a function that is bounded on the interval [, ]. Since f L 2 (T) and (ei y 1)1
is bounded for y [, ], we conclude from this that h L 2 (T).
We can thus interpret (8.24) in terms of Fourier coefficients,
f (x) Sn [ f ](x) = cn1 [h] cn [h]. (8.26)
Bessels inequality, which takes the form (8.14) here, implies that ck [h] 0 as
k . By (8.26) this establishes pointwise convergence at x.
8.4 Uniform Convergence 141
8.4 Uniform Convergence
According to the definition (8.15) of uniform convergence, f n f uniformly on T

if
sup f n (x) f (x) 0
xT
as n . This is closely related to the convergence with respect the L norm, as

introduced in Sect. 7.3. If a sequence converges in the L sense, then after possibly
modifying the functions on a set of measure zero we can assume that the convergence
is uniform.
Continuity is not necessarily preserved under pointwise limits. For example the
sequence enx converges pointwise on R but the limit function is discontinuous
2
at x = 0. On the other hand, uniform convergence of continuous functions does

guarantee continuity.
Lemma 8.4 Suppose { f n } C 0 () for a domain Rn . If { f n } converges
uniformly to a function f : R, then f is also continuous.
Proof The goal is to show that f ( y) can be made close to f (x) by taking y close
to x. To make use of the uniform convergence, we note that the triangle inequality
implies

| f (x) f ( y)| f (x) f n (x) + f n (x) f n ( y) + f n ( y) f ( y) . (8.27)
For n large we can control the first and third terms on the right by the assumption of
uniform convergence. To control the middle term we can use the continuity of f n .
Fix x and > 0. By uniform convergence there exists n so that

sup f n ( y) f ( y) < . (8.28)
y
The fact that f n is continuous at x means that we can find > 0 (depending on x)
such that for y satisfying |x y| < ,

f n (x) f n ( y) < . (8.29)
Combining (8.28) and (8.29) with (8.27) shows that for y satisfying
|x y| < ,
| f (x) f ( y)| < 3.
Thus f is continuous at x.
Uniform convergence is particularly easy to check for periodic Fourier series,
because the eigenfunctions k satisfy
ikx
e = 1.
Theorem 8.5 For f C 1 (T), the sequence of partial sums Sn [ f ] convergences

uniformly to f .
Proof The assumption that f C 1 (T) implies f C 0 (T). By integration by parts,

1
ck [ f ] = f (y)eiky dy
2

1 iky ik
= f (y)e + f (y)eiky dy.
2 2
The boundary term cancels by periodicity, leaving
ck [ f ] = ikck [ f ]. (8.30)
Since f L 2 (T) also, applying Bessels inequality in the form (8.14) to the
coefficients (8.30) implies that

kck [ f ] 2 < . (8.31)
kZ
Let
2 (Z\{0}) denote the discrete L 2 space on the set consisting of functions
Z\{0} C. (The
p spaces were introduced in Sect. 7.4.) The sequence
ak := |kck [ f ]|
defines an element of
2 (Z\{0}) by (8.31). If we define b
2 (Z\{0}) by bk := k 1 ,
then the sum of the coefficients ck [ f ] with k = 0 can be expressed as an
2 pairing,

|ck [ f ]| = a, b
2 .
k=0
By the Cauchy-Schwarz inequality on

2 (Z\{0}),
a, b
2 a
2 b
2 < .
Since the norms of a and b are finite, we conclude that

|ck [ f ]| < . (8.32)
kZ
Note that we already know that Sn [ f ] f pointwise by Theorem 8.10. This

implies that
Sn [ f ](x) f (x) = ck [ f ]k (x)
|k|>n
8.4 Uniform Convergence 143
for each x T. Because |k (x)| = 1,

Sn [ f ](x) f (x) ck [ f ] . (8.33)
|k|>n
The right-hand side of (8.33) is independent of x and tends to zero as n by

(8.32), proving that Sn [ f ] f uniformly.
8.5 Convergence in L 2
The uniform convergence provided by Theorem 8.5 proves to be very helpful in

resolving the L 2 basis question. This is because uniform convergence on T implies
L 2 convergence also, by the integral estimate

f n f 22 = | f n f |2 d x

2
2 sup f n (x) f (x) .
xT
Hence Theorem 8.5 gives convergence of Fourier series in L 2 for C 1 functions. In

this section we will extend this result to all of L 2 .
Theorem 8.6 The normalized periodic Fourier eigenfunctions
1
eikx , k Z,
2
form an orthonormal basis for L 2 (T).
Proof Suppose that u L 2 (T) satisfies
u, k
= 0 (8.34)
for each k Z. By Theorem 7.10 the conclusion will follow if we can deduce that
this implies u = 0.
As noted above, for C 1 (T) Theorem 8.5 implies that Sn [] in L 2 (T).
In particular, this gives
lim u, Sn []
= u,
. (8.35)
n
However, since Sn [] is a finite linear combination of the k , the assumption (8.34)

implies that
u, Sn []
= 0.
Hence from (8.35) we deduce that
u,
= 0.

Now recall Theorem 7.5, which says that Ccpt (0, 2 ) forms a dense subset of
L (0, 2 ). This implies also that C (T) is dense in L 2 (T). Therefore we can choose
2 1
a sequence {l } in C 1 (T) such that l u in L 2 (T). Thus
u22 = lim u, l
,
l
and we just showed that all terms on the right are zero under the assumption (8.34).
Therefore u = 0.
The combination of Theorems 8.6 and 7.9 immediately yields the following:
Corollary 8.7 (Parsevals identity) For f L 2 (T), the periodic Fourier coefficients
satisfy
1
|ck [ f ]|2 = f L 2 .
2
kZ
Applying Parsevals identity to f + g, where f, g L 2 (T), and separating out

the cross-term yields the corresponding result for the inner product,

f, g
= 2 ck [ f ]ck [g]. (8.36)
kZ
Example 8.8 In Example 8.2, we found for the step function h that ck [h] = k
1
for
k odd, c0 [h] = 2 , and otherwise ck [h] = 0. So for this case,
1
1 1
|ck [h]|2 = +2 .
kZ
4 kN
2k2
odd
On the other hand, h22 = , so Parsevals identity implies
1 2 1 1
+ 2 = .
4 kN k 2 2
odd
Thus we obtain the summation formula

1 2
= . (8.37)
k2 8
kNodd

8.5 Convergence in L 2 145
The space L 2 (0, 2 ) can be identified withL 2 (T) by extending functions peri-
1 ikx
odically. Hence Theorem 8.6 also implies that 2 e is an orthonormal basis for
L (0, 2 ). We can also specialize the periodic results to show that cosine or sine
2
series give orthonormal bases for L 2 (0,

) with basis functions that satisfy Dirichlet
or Neumann boundary conditions, respectively. We will discuss these cases in the
exercises.
8.6 Regularity and Fourier Coefficients
In the preceding sections we have made some progress in understanding the repre-
sentation of a function by Fourier series. However, we still have not addressed one of
the primary questions raised in Sect. 8.1: when does a Fourier series yield a classical
solution to the original PDE? In this section we will resolve this issue by studying
the relationship between the regularity of a function and the decay of its Fourier
coefficients.
The starting point for this discussion is the computation used in the proof of
Theorem 8.5,
ck [ f ] = ikck [ f ]
for f C 1 (T). Repeating this computation inductively gives the following:

Lemma 8.9 Suppose that f C m (T). Then
ck [ f (m) ] = (ik)m ck [ f ]. (8.38)
To describe the decay rates of coefficients, we introduce some convenient order

notation. For R,
|ak |
ak = o(k ) means lim = 0.
|k| |k|
This is commonly referred to as the little-o notion of order. There is a corresponding

big-O definition,
ak = O(k ) means |ak | C |k| ,
for all sufficiently large |k|, with C independent of k. Note that the little-o condition
is stronger. The content of the statement ak = o(k ) is that the ratio ak /k tends to
zero, while ak = O(k ) says only that the ratio is bounded.
Theorem 8.10 For f C m (T) with m N0 ,

2
k 2m ck [ f ] < , (8.39)
kZ
and
ck [ f ] = o(k m ).
Proof The inequality (8.39) follows immediately from a combination of Lemma 8.9
and and Bessels inequality in the form (8.14). Since the terms in a convergent series
must approach zero,
lim k m ck [ f ] = 0,
|k|
which gives the claimed decay estimate.

Example 8.11 Consider the cosine series (8.6) computed in Example 8.1 for the
function h(x) = 3 x 2 2x 3 on (0, ). Although h C (0, ), theextension of h
to T as an even function is merely C 2 . Theorem 8.10 thus implies that k 4 |ck [h]|2 <
.
The periodic Fourier coefficients corresponding to (8.9) are
3

2, k = 0,
ck [h] = k 4 , k odd,
24

0, k = 0, even.
This shows much faster decay than predicted, but not the rapid decay we would have
seen if the even periodic extension had been smooth.
Our next goal is to develop a converse to Theorem 8.10 that says that a certain
level of decay rate of Fourier coefficients guarantees a corresponding level of dif-
ferentiability for the function. In fact, the first stage of this result has already been
worked out. Suppose f L 2 (T) and its coefficients satisfy

|ck [ f ]| < . (8.40)
kZ
We know that Sn [ f ] f in the L 2 sense by Theorem 8.6. By Theorem 8.5 we also

know that {Sn [ f ]} converges uniformly, and so the limit is continuous by Lemma 8.4.
Hence we can conclude that (8.40) implies f C 0 (T). Recall from Sect. 7.3 that
when we say an L 2 function is C m we mean this only up to equivalence, i.e., the
original function might require modification on a set of measure zero to make it C m .
Theorem 8.12 Suppose f L 2 (T) has Fourier coefficients satisfying

k m ck [ f ] < , (8.41)
kZ
for m N0 . Then f C m (T).

Proof As remarked above, the m = 0 case is already taken care of by Theorems 8.5
and 8.6.
8.6 Regularity and Fourier Coefficients 147
Assume that (8.41) is satisfied for m = 1. For convenience, let ck := ck [ f ]

and f n = Sn [ f ]. Since f n is a (finite) linear combination of smooth functions, the
derivatives are given by
n
f n (x) = ikck eikx .
k=n

By the m = 0 result, the sequence f n converges uniformly to some g C 0 (T).
Our goal is to show that g = f , which means
f (x + y) f (x)
g(x) = lim ,
y0 t
for every x T. To argue this we will decompose the difference quotient as

f (x + t) f (x) f n (x + y) f n (x)
g(x) = f n (x)
t y (8.42)

+ f n (x) g(x) + Rn (x, y).
The first term on the right approaches zero by the definition of f n , and the second
term approaches zero as n by the construction of g. The remainder term is
(eik(x+y) eikx )
Rn (x, y) := ck ,
|k|>n
y
which converges absolutely for each y = 0 by (8.41).

We can estimate the remainder by
iky
e 1
|Rn (x, y)|
|ck | .
|k|>n
y

By noting that eiky 1 = 2 sin(ky/2), a simple calculus estimate gives
iky
e 1
|k|
y
for all y = 0. This implies a uniform estimate on the remainder term,

|Rn (x, y)| |kck | . (8.43)
|k|>n
In particular, by the assumption (8.41) the remainder term is arbitrarily small for n
large.
Fix x T and > 0. By (8.43) and the fact that f n (x) g(x), we can pick n
so that
f (x) g(x) < and |Rn (x, y)| < .
n
for all y = 0. For this n and x, the definition of f n (x) says that we can choose such
that 0 < |y| < implies

f n (x + y) f n (x)
f n (x) < .

y
Applying these estimates to (8.42) shows that for 0 < |y| < ,

f (x + y) f (x)
g(x) 3.
y
Since was arbitrarily small, this shows that f (x) = g(x). And since g is continuous,
we conclude that f C 1 (T).
The same argument can now be repeated for higher derivatives, assuming (8.41)
holds for larger m.
The hypothesis (8.41) can be reformulated in terms of a decay condition on the

coefficients, although this gives a slightly weaker result. If ck [ f ] = O(k ), then

k m ck [ f ] C |k|+m ,
kZ kZ
and this series converges provided > m + 1. Hence Theorem 8.12 implies that
f C m (T) under the condition that
ck [ f ] = O(k m1 )
for some > 0.

Let us finally return to the one-dimensional heat equation that motivated this
discussion, first considering the periodic case.
Theorem 8.13 For h C 0 (T), the heat equation on [0, ) T,
u 2u
2 = 0,
t x
admits a solution u C ((0, ) T), defined for t > 0 by

ck [h]ek t eikx ,
2
u(t, x) := (8.44)
kZ
and satisfying
8.6 Regularity and Fourier Coefficients 149
lim u(t, x) = h(x) (8.45)

t0
for each x T.
Proof For t > 0, the Fourier coefficients of u(t, ) decay exponentially, and Theo-
rem 8.12 shows that u(t, ) C (T) for each t. The same arguments used in the
proof of that theorem apply to the t derivatives. To see this, let u n denote the partial
sum of (8.44),
n
ck ek t eikx .
2
u n (t, x) :=
k=n
where ck := ck [h]. As a finite sum, this can be differentiated directly,
u n n
(k 2 )ck ek t eikx .
2
(t, x) = (8.46)
t k=n
By Theorem 8.10 the Fourier coefficients of h satisfy ck = o(k 1 ), so that

2 k 2 t ikx
e Ckek t .
2
k ck e (8.47)
As n the series (8.46) thus converges absolutely for t > 0, allowing us to

define a function
u n
g := lim .
n t
For > 0, the estimate (8.47) shows that the convergence is uniform for t .
Lemma 8.4 shows that the limit is continuous for t . Since > 0 is arbitrary, this
implies g C 0 ((0, ) T).
We can argue that g = u/t by considering

u(t + s, x) u(t, x) u n (t + s, x) u n (t, x) u n
h(t, x) = (t, x)
s s t

u n
+ (t, x) h(t, x) + Rn (t, s, x),
t
where
ek s 1
2
k 2 t ikx
Rn (t, s, x) := ck e e .
|k|>n
s
At this point the argument becomes essentially parallel to the analysis of (8.42), so
we will omit the details. The conclusion is that u/t is continuous on (0, ) T.
The argument can be repeated for higher t derivatives, allowing us to conclude
that u C ((0, ) T). Moreover, the partial derivatives of u n converge to the
corresponding derivatives of u, pointwise on (0, )T, and uniformly if we restrict

to t for some > 0. Since u n satisfies the wave equation for each n by
construction, this shows that u satisfies the wave equation also.
At the moment we only have the tools to prove (8.45) under the stronger assump-
tion that h C 1 (T). In this case we can argue exactly in the proof of Theorem 8.5
that (8.44) converges uniformly for (t, x) [0, ) T. By Lemma 8.4 this shows
u C 0 ([0, ) T) and we can just set t = 0 to obtain (8.45).
If h is merely continuous, then this approach breaks down, because the series
(8.44) may actually diverge for t = 0. We will cover the C 0 case in Chap. 13, after
developing an alternate formula for (8.44).
The one-dimensional heat equation derived in Sect. 6.1 involved Dirichlet or Neu-
mann boundary conditions on an interval [0,
]. By rescaling the interval to [0, ] and
then extending functions to T with either even or odd symmetry, we apply the results
for periodic Fourier series results to these cases. In particular, from Theorem 8.13
we deduce the following:
Corollary 8.14 Suppose h C[0,
] and satisfies Dirichlet or Neumann boundary
conditions. The heat equation on [0, )[0,
] admits a solution u C ((0, )
[0,
]), under the same boundary condition, such that
lim u(t, x) = h(x)

t0
for each x [0,

].
The solutions obtained in Theorem 8.13 and Corollary 8.14 are uniquely deter-
mined by the initial condition h. (See Exercise 6.4.) The fact that solutions are smooth
for t > 0, even when h is merely continuous, is a characteristic property of diffu-
sion equations. In fact the smoothing phenomenon carries over to cases where h is
L 2 but not even continuous. This is illustrated in Fig. 8.5 for the case considered in
Example 8.2.
As these applications show, the Fourier series approach (and spectral analysis
in general) is well suited to analyzing the regularity of solutions. However, other
qualitative features are perhaps obscured from this viewpoint. For example, we would
expect solutions to reflect the physical principle that heat flows from hot to cold. We
can see this behavior quite clearly in the plots of Fig. 8.5, but it is not at all apparent
in the series formula (8.44).
t=0 t=.001 t=.005
Fig. 8.5 Solutions of the heat equation become smooth for t > 0
8.7 Exercises 151
8.7 Exercises
8.1 For x (0, ), let

f (x) = x.
(a) Extend f to an odd function on T and compute the periodic Fourier coeffi-
cients ck [ f ] according to (8.12). (Note that the case k = 0 needs to be treated
separately.) Show that the periodic series reduces to a sine series in this case.
(b) Show that the convergence of the Fourier series at x = 2 , which is guaranteed
by Theorem 8.3, yields the summation formula
1 1 1
= 1 + + .
4 3 5 7
(c) Show the Parseval identity (Corollary 8.7) leads to the formula

1 2
= .
k=1
k2 6
8.2 For x (0, ), let

g(x) = x.
(a) Extend g to an even function on T and compute the periodic Fourier coeffi-
cients ck [g] according to (8.12). (Note that the case k = 0 needs to be treated
separately.) Show that the periodic series reduces to a cosine series in this case.
(b) Show that the convergence of the Fourier series at x = 0, which is guaranteed
by Theorem 8.3, reproduces the formula (8.37).
(c) Show the Parseval identity (Corollary 8.7) implies the formula
1 4
= .
kN
k4 96
odd
8.3 Consider the periodic wave equation
2u 2u
=0
t 2 x2
for t R and x T. Suppose the initial conditions are
u
u(0, x) = g(x), (0, x) = h(x),
t
for g C m+1 (T) and h C m (T), for m N.

(a) Assuming that u(t, x) can be represented as a Fourier series

u(t, x) = ak (t)eikx , (8.48)
kZ
find an expression for ak (t) in terms of the Fourier coefficients of g and h.

(b) Using the assumptions on g and h, together with Theorem 8.10, show that the
coefficients ak (t) satisfy an estimate

k 2m |ak (t)|2 M < ,
kZ
uniformly for t R.
(c) By the arguments used in Theorem 8.13, (b) implies that the series (8.48) con-
verges to a solution u satisfying the initial conditions. What could you conclude
about the differentiability of u?
8.4 In L 2 (0, ) consider the sequence

2
k (x) := sin kx,

for k N.
(a) Show that {k } is an orthonormal sequence.
(b) Suppose that f L 2 (0, ) and f, k
= 0 for all k N. Show that f 0.
(Hint: extend f to an odd 2 -periodic function on R, which can be regarded as
an element of L 2 (T). Then apply Theorem 8.6.)
(c) Conclude that {k } is an orthonormal basis for L 2 (0, ).
8.5 Suppose that f L 2 (, ) satisfies

x l f (x) d x = 0

for all l N0 .

(a) Show that qm,k , f = 0 for all m N and k Z, where

m
(ikx)l
qm,k (x) := .
l=0
l!
(b) Note that

lim qm,k (x) = eikx
m
by the definition of the complex exponential. Show that this convergence is

uniform for x [, ] (with k fixed).
8.7 Exercises 153
(c) Use (a) and (b) to show that for f L 2 (, ),

eikx f (x) d x = 0

for all k Z.
(d) Conclude from Theorem 8.6 that f 0. (In other words, the monomials
1, x, x 2 , . . . form a basis for L 2 (, ), although not an orthonormal one.)
8.6 The Legendre polynomials are functions of z [1, 1] defined by
1 dk 2
Pk (z) := (z 1)k ,
2k k! dz k
for k N0 . (This corresponds to the case m = 0 in (5.31).) These are solutions of

the eigenvalue equation
L Pk = k(k + 1)Pk ,
where
d d
L := (z 2 1) .
dz dz
(a) For u, v C 2 [1, 1], check that L satisfies a formal self-adjointness condition,
u, Lv
L 2 = Lu, v
L 2 .
Conclude that the Pk s with distinct values of k are orthogonal in L 2 (1, 1).
(b) Use the result of Exercise 8.5 to show that {Pk } forms an orthogonal basis for
L 2 (1, 1). (The Pk are normalized by the condition Pk (1) = 0, rather than by
unit L 2 norm.)
Chapter 9
Maximum Principles
We saw in Sect. 4.7 that conservation of energy can be used to derive uniqueness for
solutions of the wave equation. In this chapter we will consider another approach to
issues of uniqueness and stability, based on maximum values. This method applies
generally to elliptic equations, which describe equilibrium states, and to parabolic
equations, which are generally used to model diffusion.
9.1 Model Problem: The Laplace Equation
As noted in Sect. 5.2, the classical evolution equations such as the heat or wave
equation have the form
Pt u u = 0,
where Pt denotes some combination of time derivatives. In an equilibrium state, for

which the solution is independent of time, these equations all reduce to
u = 0,
which is called the Laplace equation. A solution of the Laplace equation is also called
a harmonic function. The Laplace equation on a bounded domain is generally
formulated with an inhomogeneous Dirichlet boundary condition,
u| = f
for f : R.

156 9 Maximum Principles
The Laplace equation frequently appears in applications involving vector fields.

A conservative vector field v C 0 (; Rn ) can be represented as the gradient of a
potential function C 1 (; R),
v = ,
If the vector field v is also solenoidal ( v = 0), then the potential satisfies the
Laplace equation
= 0.
In fluid dynamics in R3 , for example, the velocity field is solenoidal for an incom-
pressible fluid, such as water, and conservative precisely when the flow is irrotational
( v = 0).
Electrostatics provides another important source of Laplace problems. In the
absence of charges, the electric field E is conservative and is commonly written
as
E =
where is the electric potential. On the other hand, Gausss law of electrostatics says
that E is proportional to the electric charge density. Hence, the electric potential
for a charge-free region satisfies the Laplace equation.
In the remainder of this section we will consider a particular classical case, the
Laplace problem on the unit disk. Circular symmetry allows us to solve the equation
explicitly using Fourier series, and the resulting formula gives some insight into the
general behavior of harmonic functions.
Let D denote the open unit disk in R2 . Given g C 0 (D), our goal is to solve
u = 0, u|D = g. (9.1)
In Sect. 5.3 we used separation of variables in polar coordinates to find the family of
harmonic functions,
k (r, ) := r |k| eik ,
for k Z. The boundary D is naturally identified with the space T := R/2Z

introduced in Sect. 8.2, parametrized by .
Consider the periodic Fourier series expansion,

g() = ck [g]eik , (9.2)
kZ
where 2
1
ck [g] := eik g() d.
2 0
9.1 Model Problem: The Laplace Equation 157
Given that
eik = k (1, ),
we might hope to construct a solution of (9.1) by setting

u(r, ) = ck [g]k (r, ). (9.3)
kZ
Theorem 8.10 shows that the sequence {ck [g]} is bounded for g continuous. Note
also that
|k (r, )| = r |k|
and
r |k| <
kZ
for r < 1 by geometric series. This implies that (9.3) converges absolutely for r < 1.
In fact the convergence is uniform on {r R} for R < 1.
We can write u(r, ) more explicitly by substituting the definition of ck [g] into
the integral,
1 2
u(r, ) = r |k| eik() g() d.
kZ
2 0
Uniform convergence in for r < 1 allows us to move the sum inside the integral,
yielding the formula
2
1
u(r, ) = Pr ( )g() d, (9.4)
2 0
where
Pr () := r |k| eik . (9.5)
kZ
This function is called the Poisson kernel. Its behavior as r 1 is illustrated in

Fig. 9.1.
Summing by geometric series gives the formula

Pr () = 1 + (r ei )k + (r ei )k
k=1 k=1
r ei r ei
=1+ + (9.6)
1 r ei 1 r ei
1r 2
= .
1 2r cos + r 2
r=.9
r=.8
r=.6
r=.4

Fig. 9.1 The Poisson kernel Pr () for a succession of radii
From the series formula (9.5) we can also deduce directly that
2
1
Pr () d = 1, (9.7)
2 0
since the only nonzero contribution comes from the term k = 0.

By periodicity, a change of variables in (9.4) gives the alternate form
2
1
u(r, ) = Pr ()g( ) d. (9.8)
2 0
In view of (9.7), this could be interpreted as a weighted average of f with a weight

function that depends on r . As r 1 this weight function becomes concentrated
at 0, as Fig. 9.1 demonstrates. This is the mechanism by which we expect to have
u(r, ) g() as r 1 .
Theorem 9.1 For f C 0 (D), the Laplace equation,
u = 0 in D, u|D = g,
admits a classical solution u C (D) C 0 (D) given by the Poisson integral (9.4).
Proof The function Pr () is smooth for r < 1, and it follows from (9.5) that
Pr () = 0,
where Pr () is interpreted as a function on D written in polar coordinates. By passing

derivatives inside the integral, we can deduce from (9.4) that u C (D) and
9.1 Model Problem: The Laplace Equation 159
u = 0.
To complete the proof we need to check that
lim u(r, ) = g() (9.9)

r 1
for every D, which will also show that u C 0 (D). Note that (9.9) is not the
same as claiming that the Fourier series for g converges, which is not necessarily
true. The difference lies in the order of the limits. In (9.9) we take the limit of the
Fourier series first for r < 1, and then the limit r 1 . This limit exists, as we will
see below, but if we first set r = 1 in (9.3) then the sum over k may diverge.
By (9.7) and (9.8) we can write

1
u(r, ) g() = Pr () g( ) g() d. (9.10)
2
The goal is to estimate the left-hand side for r close to 1. Fix D and let > 0.
Since g is continuous, there exists > 0 so that
|g( ) g()| < (9.11)
for || < . For || we can estimate
max Pr () = Pr (). (9.12)

||
Thus, splitting the integral (9.10) at || = gives

1
|u(r, ) g()| Pr ()g( ) g() d
2

1
+ Pr ()g( ) g() d
2 <||<

Pr ()
Pr ()d + g( ) g() d.
2 2 <||<
By (9.7) and the fact that Pr > 0,

1
Pr () d 1.
2
Furthermore, since g is continuous, |g( ) g()| is bounded by some constant

M for all and . This reduces the bound to
|u(r, ) g()| + M Pr ().

We can now use the fact that
lim Pr () = 0
r 1
to choose R < 1 so that

M Pr ()
for R < r < 1. We conclude that
|u(r, ) g()| 2
for R < r < 1. Since was arbitrary, this shows
lim |u(r, ) g()| = 0.

r 1

For students who know some complex analysis, we note that the formula (9.4)
could be deduced from the Cauchy integral formula, because any harmonic function
on D is the real part of a holomorphic function.
Example 9.2 For 0 < a < , suppose the boundary function is given by

||
1 , || a,
g() = a
0, a < || ,
as shown in Fig. 9.2.

This boundary condition could represent, for example, a hot spot at one point on
the edge of a metal plate. The corresponding equilibrium temperature distribution
within the plate is given by calculating the Fourier coefficients of g and substituting
into (9.3). The resulting solution,

a 2 1 cos(ka)
u(r, ) = + cos(k),
2 a k=1 k2
is illustrated in Fig. 9.3.
Fig. 9.2 Boundary function

with a triangular peak
a a
9.2 Mean Value Formula 161
Fig. 9.3 Contour plot of the

harmonic function from
Example 9.2
9.2 Mean Value Formula
Setting r = 0 in the Poisson formula (9.4) gives

2
1
u(0) = g() d, (9.13)
2 0
because P0 () = 1. In other words, the value of a harmonic function at the center of

the disk is equal to its average value on the boundary. This phenomenon is illustrated
in Fig. 9.4. In this section we will extend (9.13) to an averaging formula that works
in any dimension.
The (n 1)-dimensional volume of a sphere of radius r is
vol[ B(x 0 ; r )] = An r n1 , (9.14)
where An denotes the volume of the unit sphere in Rn , as defined in (2.13). It follows
from the radial integral formula (2.10) that
Fig. 9.4 Mean value

property of a harmonic
function
An r n
vol[B(x 0 ; r )] = . (9.15)
n
To state the mean value formula for a ball of radius R, we introduce the family of
radial functions,

1
ln( Rr ), n = 2,
G R (x) := 2 (9.16)
1
(n2)An R n2
1
r n2 , n 3.
1
The function G R is the unique solution of the equations
G R 1
= , G R
= 0. (9.17)
r An r n1 r =R
Note that G R is integrable on B(0; R), despite the singularity at the origin, because
the radial volume element is An r n1 dr by (2.10).
Theorem 9.3 (Mean value formula) Assume that u C 2 () on a domain Rn

with n 2. For R > 0 such that B(x 0 ; R) ,

1
u(x 0 ) = u(x)d S + G R (x x 0 )u(x)d n x.
An R n1 B(x 0 ;R) B(x 0 ;R)
Proof By a change of variables, it suffices to consider the case x 0 = 0. The formula

(2.15) for the radial component of the Laplacian implies that
G R (x) = 0
for x
= 0. For > 0, we can therefore apply Greens second identity (Theorem 2.11)
on the domain { < r < R} to obtain

u G R
G R u d x = n
GR u dS
{<r <R} {r =R} r r

(9.18)
u G R
GR u d S.
{r =} r r
Because G R is integrable on B(0; R), on the left-hand side of (9.18) we can take
0 to obtain

lim G R u d n x = G R u d n x. (9.19)
0 {<r <R} B(0;R)
By (9.17), the first term on the right in (9.18) reduces to

u G R 1
GR u dS = u d S. (9.20)
{r =R} r r An R n1 {r =R}
The second term on the right in (9.18) is

u G R
GR u dS
{r =} r r

u 1
= G R () dS + u d S.
{r =} r An n1 {r =}
The first of these integrals can be estimated by noting that u/r is a directional
derivative and thus bounded by the magnitude of |u|. By the assumption that u
C 2 (), |u/r | is therefore bounded by a constant C for r R, yielding the estimate
u

d S C An n1 .
{r =} r
Since the divergent term in G R () as 0 is proportional to 2n for n 3 and

log for n = 2, this implies

u
lim G R () d S = 0.
0 {r =} r
Hence

u G R 1
lim GR u d S = lim u dS .
0 {r =} r r 0 An n1 {r =}
The term in brackets is the average of u over a sphere of radius . Since u is continuous,
this average approaches u(0) as 0, so that

u G R
lim GR u d S = u(0). (9.21)
0 {r =} r r
Applying (9.19), (9.20), and (9.21) to (9.18) gives

1
G R u d n x = u(0) u d S,
B(0;R) An R n1 B(0;R)
which completes the proof.
For harmonic functions, Theorem 9.3 gives a generalization of the circle formula
(Theorem 9.3) to spherical averages in higher dimensions. As we will now show, the
mean value property can be stated in a equivalent form in terms of averages over a
ball.
Corollary 9.4 (Mean value for harmonic functions) Suppose Rn for n 2.

For u C 2 () the following properties are equivalent:
(A) The function u is harmonic on .

(B) For B(x 0 ; R) ,

1
u(x 0 ) = u d S.
An R n1 B(x 0 ;R)
(C) For B(x 0 ; R) ,

n
u(x 0 ) = u d n x.
An R n B(x 0 ;R)
Proof The fact that (A) implies (B) follows immediately by setting u = 0 in the
formula of Theorem 9.3.
To see that (B) and (C) are equivalent, fix some x 0 and define

h(r ) := u d n x,
B(x 0 ;r )
for r 0 such that B(x 0 ; r ) . As we saw in Exercise 2.4, the derivative of h(r )
is given by a surface integral

h (r ) = u d S.
B(x 0 ;r )
Hence property (B) says that
h (r ) = An r n1 u(x 0 ),
while property (C) says that
An r n
h(r ) = u(x 0 ).
n
Since h(0) = 0 by definition, these are two statements are equivalent.

Finally, we need to show that (B) implies (A). Assuming that (B) holds, Theo-
rem 9.3 gives
G R (x x 0 )u(x)d n x = 0 (9.22)
B(x 0 ;R)
provided B(x 0 ; R) . Suppose u(x 0 ) < 0 for some x 0 . Then by continuity

there exists some > 0 and > 0 such that u on B(x 0 ; ). Since G R is
strictly negative and decreasing as r 0, this implies

G R (x x 0 )u(x)d n x > G R r = > 0,
B(x 0 ;)
which contradicts (9.22). The same argument applies if u(x 0 ) > 0. We thus con-
clude that (B) implies u 0.
9.3 Strong Principle for Subharmonic Functions
A real-valued C 2 function that satisfies
u 0
is called subharmonic. The case u 0 is similarly called superharmonic. We

will focus on the subharmonic case. The results can easily be translated to the super-
harmonic case by replacing u by u.
The weak maximum principle says that for a subharmonic function the max-
imum value occurs at a boundary point. We will prove here a strong version of
this principle, which says furthermore that if that the global maximum occurs at an
interior point then the function is constant.
Theorem 9.5 (Strong maximum principle) Let Rn be a bounded domain. If

u C 2 (; R) C 0 () is subharmonic then
max u = max u.

The maximum is attained at an interior point only if u is a constant function.
Proof By the extreme value theorem (Theorem A.2), u achieves a global maximum
at some point x 0 . If x 0 then the claimed equality clearly holds. The goal
is thus to show that x 0 implies that u is constant.
Because is open, an interior point x 0 has a neighborhood contained in . We
may thus assume that B(x 0 ; R) for some R > 0. Applying Theorem 9.3 to this
ball gives

1
u(x 0 ) = u(x)d S + G R (x x 0 )u(x)d n x.
An R n1 B(x 0 ;R) B(x 0 ;R)
By the definition (9.16), G R 0 for 0 < r R. Therefore, since u 0 by

assumption,
1
u(x 0 ) u(x)d S. (9.23)
An R n1 B(x 0 ;R)
Using (9.14), we can subtract u(x 0 ) from both sides to obtain

1
[u(x) u(x 0 )] d S 0. (9.24)
An R n1 B(x 0 ;R)
By assumption u(x 0 ) is the global maximum of u, implying that the integrand of

(9.24) is nonpositive. The inequality therefore shows that the integrand vanishes, and
we conclude that u(x) = u(x 0 ) on B(x 0 ; R).
Note that the same argument works for every radius r < R, so this argument
shows that u u(x 0 ) on all of B(x 0 ; R).
To extend the conclusion to the full domain, let M denote the maximum value of
u on . We can write as a disjoint union E F, where
E := {x ; u(x) < M} ,
F := {x ; u(x) = M} .
By the argument given above, a point x F has a neighborhood B(x; R) on

which u is equal to M. Hence F is open.
On the other hand, for x E we can set = M u(x) and use the continuity of
u to find a > 0 such that
|u(x) u( y)| <
for y B(x; ). This implies in particular that u( y) < M, so that B(x; ) E.

Thus E is open also.
Recall from Sect. 2.3 that the fact that is connected means that the domain
cannot be written as a disjoint union of nonempty open sets. Since = E F with
E and F both open, one of the two sets is empty. If E is empty then u is constant on
, while if F is empty then the maximum of u is not attained in the interior.
For a superharmonic function u C 2 (; R) C 0 (), reversing the sign yields

a minimum principle,
min u = min u.

Both principles apply to a harmonic function u, which therefore satisfies
min u u(x) max u

for all x .
The maximum principle implies the following stability result for the Laplace
equation.
Corollary 9.6 Suppose that u 1 , u 2 C 2 () C 0 () are solutions of the Laplace

equation u = 0 with boundary values
u 1 | = g1 , u 2 | = g2 ,
for g1 , g2 C 0 (). Then
max |u 2 u 1 | max |g2 g1 | . (9.25)

9.3 Strong Principle for Subharmonic Functions 167
In particular, a solution the Laplace equation is uniquely determined by its boundary

data.
Proof By superposition, u 2 u 1 is a harmonic function with boundary data g2 g1 .

Theorem 9.5 applies to Re(u 2 u 1 ) as well as Im(u 2 u 1 ). Combining these
estimates yields the inequality (9.25).
Note that uniqueness of solutions of the Laplace equation also follows directly
from Greens first identity (Theorem 2.10), in the case where has piecewise C 1
boundary and u C 2 (; R). If u = 0, then setting v = u in Greens formula
gives
u
u d x =
2 n
u d S.

Thus if u = 0 on , then
u2 d n x = 0.

Since the integrand is positive, this implies u 0, so that u is constant. The

assumption u| = 0 then gives u 0 on the full domain. (This is the energy
method argument, as introduced in Sect. 4.7.)
One advantage that maximum principle has over the energy method is the explicit
stability formula (9.25). In terms of the L norm introduced in Sect. 7.3, this inequal-
ity could be written
u 2 u 1 g2 g1 .
This is an explicit formulation of the continuity requirement for well-posedness:

a small change in boundary data results in a correspondingly small change in the
solution.
9.4 Weak Principle for Elliptic Equations
Although the mean value formula gives a direct proof of the strong maximum princi-
ple, this approach applies only to the Laplacian itself. In this section we will present
an alternative approach that generalizes quite easily to operators with variable coef-
ficients.
On a domain Rn let us consider a second order elliptic operator of the form

n
2 n

L= ai j (x) + b j (x) , (9.26)
i, j=1
xi x j j=1
x j
where the coefficients ai j and b j are continuous functions on . As defined in Sect.

1.3, ellipticity means that the symmetric matrix [ai j ] is positive definite at each point.
For the maximum principle we need a stronger assumption, called uniform ellip-
ticity, that says that for some fixed constant > 0,

n
ai j (x)vi v j v2 (9.27)
i, j=1
for all x and v Rn . An equivalent way to say this is that the smallest eigenvalue
of [ai j ] is bounded below by at each point x.
Theorem 9.7 (Weak maximum principle) Suppose Rn is bounded, and L is

an operator of the form (9.26) satisfying the uniform ellipticity condition (9.27). If
u C 2 (; R) C 0 () satisfies
Lu 0
in , then
max u = max u.

Proof For the moment let u be a general function in C 2 (; R). Suppose that u has
a local maximum at x 0 . The first partial derivatives of u vanish at a local
maximum, so that
n
2u
Lu(x 0 ) = ai j (x 0 ) (x 0 ). (9.28)
i, j=1
xi x j
Furthermore, we claim that the matrix of second partials of u is negative definite at

x 0 , meaning
n
2u
(x 0 )vi v j 0
i, j=1
xi x j
for v Rn . To see this, set h(t) := u(x 0 + tv) and note that h has a local maximum
at t = 0, implying h (0) 0. Evaluating h (0) yields the inequality stated above.
The right-hand side of (9.28) could be written as tr(AB) where A and B are the
positive symmetric matrices

2u
A = ai j (x 0 ) , B= (x 0 ) .
xi x j
By switching toa basis

in which A is diagonal, tr(AB) can be written in terms of
the eigenvalues j of A as

n
tr(AB) = jbjj.
j=1
9.4 Weak Principle for Elliptic Equations 169
If the eigenvalues are ordered 1 n , then
tr(AB) 1 tr B.
The positivity of B implies tr B 0, so we conclude that
Lu(x 0 ) 0.
Thus argument shows that Lu(x 0 ) 0 for x 0 a local interior maximum. Therefore,
the strict inequality Lu < 0 implies that u cannot have a local interior maximum and
that
max u = max u. (9.29)

To complete the proof, we must relax the hypothesis to Lu 0. The strategy is

to perturb u slightly to reduce to the previous case. For M > 0, let
h(x) := e M x1 .
By the definition of L,
Lh = a11 M 2 + b1 M h.
The ellipticity condition (9.27) implies that a11 , so by choosing
1
M> max b1 ,

we can guarantee that

Lh < 0.
If we now assume now that u satisfies the hypothesis Lu 0, then
L(u + h) < 0
for > 0. Applying (9.29) to u + h gives
max(u + h) = max(u + h). (9.30)

Since h 0, clearly
max u max(u + h).

On the other hand, since is bounded, we may assume x1 < R in , for some R
sufficiently large. This implies
h eM R,
so that
max(u + h) max u + e M R .

From (9.30) we therefore conclude that
max u max u + e M R

for all > 0. Since M and R are independent of , we can take 0 to conclude
that
max u max u,

and the result follows.

The maximum principle implies in particular the only solution of Lu = 0 with
u| = 0 is u 0. Hence a solution of the equation
Lu = f, u| = g,
is uniquely determined by f and g if it exists.
9.5 Application to the Heat Equation
Fouriers law of heat conduction, as introduced in Sect. 6.1, suggests a maximum

principle for solutions of the heat equation. Because heat flows away from a spatial
maximum of the temperature, a local spatial maximum of the temperature should be
impossible at time t > 0. The global maximum of the temperature therefore must
occur either at t = 0 or on the boundary.
Although it is possible to prove the maximum principle via a mean value formula
as in the proof of Theorem 9.5, in this section we will follow the more direct approach
from Sect. 9.4, which has the advantage of generalizing to operators with variable
coefficients.
Because the heat equation is second order with respect to spatial variables and first
order in the time variable, it makes sense to define a domain for classical solutions
that takes this structure into account. For Rn , define

C heat () := u C 0 [0, ) ; R ; u(, x) C 1 (0, ), u(t, ) C 2 () .
Note that this definition includes only real-valued functions.

Theorem 9.8 Suppose Rn is a bounded domain and u C heat () satisfies
u
u 0, (9.31)
t
9.5 Application to the Heat Equation 171
on (0, T ) . Then the maximum value of u within [0, T ] occurs at a point

(t0 , x 0 ) with either t0 = 0 or x 0 .
Proof Suppose that u attains a maximum at (t0 , x 0 ) (0, T ) . By the same

calculus argument used in Sect. 9.4, this implies
u
(t0 , x 0 ) = 0, (9.32)
t
as well as
u 2u
(t0 , x 0 ) = 0, (t0 , x 0 ) 0. (9.33)
x j x 2j
In particular,
u
u (t0 , x 0 ) 0. (9.34)
t
If (9.31) were a strict inequality this would complete the proof.

To proceed we use a perturbation strategy as in the proof of Theorem 9.7. For
> 0, set
u := u + |x|2 .
Because |x|2 = 2n, the hypothesis on u gives
u u
u = u 2n < 0. (9.35)
t t
The existence of a local maximum for u within (0, T ) is ruled out by (9.34).
We conclude that u attains a global maximum at a boundary point of [0, T ] .
Let us label this point (t , x ), so that
max u = u (t , x ). (9.36)
[0,T ]
Since (t , x ) is on the boundary, either t = 0, t = T , or x .

Suppose that t = T and x . Then u (t, x ) u (T, x ) for t [0, T ],
implying that
u
(T, x ) 0.
t
By (9.35), this implies also that u (T, x ) > 0, which is ruled out by (9.33). Hence
t
= T if x .
Therefore (t , x ) lies in the set
:= ({0} ) ([0, T ] ) . (9.37)

Let R be sufficiently large so that B(0; R). This means that |x| R on , so
the inequality
u u u + R 2
holds at every point in [0, T ] . From (9.36) we can thus conclude that
max u u (t , x )
[0,T ]
(9.38)
u(t , x ) + R 2 .
This implies that

max u max u + R 2 ,
[0,T ]
because (t , x ) . Since this inequality holds for every > 0, this proves
max u max u.
[0,T ]
For a solution of the heat equation, both u satisfy the hypothesis of Theorem 9.8,
which implies that
min u u(t, x) max u,

for (t, x) (0, T ) , where is defined by (9.37). In particular this yields the
following:
Corollary 9.9 Let Rn be a bounded domain. A solution of the heat equation

u C heat () is uniquely determined by u| and u|t=0 .
The same arguments could be applied to the more general parabolic equation
u
Lu = 0,
t
where L is a uniformly elliptic operator as defined in (9.27).
In Sect. 6.3, we stated without proof a uniqueness result for solutions of the heat
equation on Rn . We now have the means to prove this, by establishing a maximum
principle for Rn as a corollary of Theorem 9.8.
Corollary 9.10 Suppose that u is a classical solution of the heat equation
u
u = 0, u|t=0 = g, (9.39)
t
on [0, ) Rn , and that u is bounded on [0, T ] Rn for T > 0. Then

9.5 Application to the Heat Equation 173
max u max g. (9.40)

[0,)Rn n R
Proof Assume that u satisfies (9.39) and also
u(t, x) M
for t [0, T ] and x Rn . For y Rn and > 0 set
|x y|2
v(t, x) := u(t, x) (T t) 2 e 4(T t) .
n
The term resembles the heat kernel defined by (6.16), except that the sign in the
exponential is reversed. Direct differentiation shows that this expression satisfies the
heat equation on (0, T ) Rn , and hence v does also.
For R > 0, let us apply the maximum principle of Theorem 9.8 to v on the domain
(0, T ) B( y; R). By construction,
v(0, x) g(x),
and for x B( y; R),
R2
v(t, x) M (T t) 2 e 4(T t)
n
M T 2 e R /4T
n 2
.
With T fixed, the right-hand side of this second inequality is arbitrarily negative for
large R. Therefore, for sufficiently large R, Theorem 9.8 implies that
max v max g.
[0,T ]B( y;R) B( y;R)
In particular, setting x = y in this inequality gives
v(t, y) max g.
nR
for t [0, T ] and y Rn . By the definition of v, this implies that
g + (T t) 2 .
n
u(t, y) max
n R
We can now take 0 and T to conclude that
u(t, y) max g
n R
for all t [0, ) and y Rn .

The argument given here can be refined to show that conclusion (9.40) holds under
the weaker growth condition
2
u(t, x) Mec|x|
for t [0, T ].
Corollary 9.10 implies Theorem 6.3 by the argument used in Corollary 9.9. That
is, if u 1 and u 2 are bounded solutions of (6.19), then (u 1 u 2 ) solves (6.19) with
g = 0. It then follows from (9.40) that u 1 = u 2 .
9.6 Exercises
9.1 Suppose that u, C 2 (; R)C 0 () on a bounded domain Rn . Assume

that u subharmonic and harmonic, with matching boundary values:
u| = | .
Show that
u
at all points of . (This is the motivation for the term subharmonic.)
9.2 Liouvilles theorem says that a bounded harmonic function on Rn is constant.

To show this, assume u C 2 (Rn ) is harmonic and satisfies
|u(x)| M
for all x Rn .
(a) For x 0 Rn , set r0 = |x 0 |. Use Corollary 9.4 at the centers 0 and x 0 to show
that
n
u(0) u(x 0 ) = ud x
n n
ud x (9.41)
An R n B(0;R) B(x 0 ;R)
for R > 0. Note that the integrals cancel on the intersection of the two balls.
(b) Show that

vol [B(0; R)\B(x 0 ; R)] vol B(0; R)\B( x20 ; R r0
2
)
An n n
= R R r20 ,
n
and the same for B(x 0 ; R)\B(0; R).

9.6 Exercises 175
(c) Apply the volume estimates and the fact that |u| M to (9.41) to estimate
r0 n
R n (R )
|u(0) u(x 0 )| 2M 2
.
Rn
Take the limit R to show that u(x 0 ) = u(0).

9.3 Suppose that Rn is bounded, with B(0; R), and assume that u
C 2 (; R) C 0 () satisfies
u = f, u| = 0,
for f C 0 ().
(a) Find a constant c > 0 (depending on f and R), such that u+c |x|2 is subharmonic
on .
(b) For this value of c, apply the maximum principle to u + c |x|2 to deduce that
max |u| C max | f | ,

where C depends only on R.

9.4 Suppose u is a harmonic function on a domain that includes B(0; 4R) for some
R > 0, and assume u 0. Show that
max u 3n min u.
B(0;R) B(0;R)
Hint: For x B(0; R), apply the maximum principle to write u(x) as an integral
over the balls B(x; R) and B(x; 3R). Then show that
B(x; R) B(0; 2R) B(x; 3R),
and use this to estimate the integrals.

9.5 Suppose u C 2 (B; R) C 0 (B) is a nonconstant subharmonic function and
assume that the maximum of u on B is attained at the point x 0 B. This automat-
ically implies that u
r
(x 0 ) 0. Hopfs lemma says that this inequality is strict,
u
(x 0 ) > 0.
r
To show this, let B := B(0; R) Rn for some R > 0, and set
A := {R/2 < |x| < R} .

(a) Consider the function

h(x) := e2n|x| /R 2
e2n .
2
Compute h and show that h is subharmonic on A.

(b) Set
m = max u, M = max u,
{r =R/2} {r =R}
and show that m < M.

(c) For > 0 set
u := u + h,
and show that by taking sufficiently small we may assume that
max u M.
A
(d) Show that u (x) M for x A, and hence that
u
(x 0 ) 0.
r
(e) By computing u /r and taking 0, conclude that
u
(x 0 ) > 0.
r
Chapter 10
Weak Solutions
In Sect. 1.2 we observed that dAlemberts formula for a solution of the wave equation
makes sense even when the initial data are not differentiable. This concept of a weak
solution that is not actually required to solve the equation literally has come up in other
contexts as well, for example in the discussion of the traffic equation in Sect. 3.4. In
this chapter we will discuss the mathematical formulation of this generalized notion
of solution.
Weak solutions first appeared in physical applications as idealized, limiting cases
of true solutions. For example, one might replace a smooth density function by a
simpler piecewise linear approximation, as illustrated in Fig. 10.1, in order to simplify
computations. (We used this idea in Example 3.9.)
Up until the late 19th century, the limiting process by which weak solutions were
obtained was understood rather loosely, and justified mainly by physical intuition.
Weak solutions proved to be extremely useful, and eventually a consistent mathe-
matical framework was developed.
10.1 Test Functions and Weak Derivatives
Consider a linear equation of the form Lu = f , where L is a differential operator

on a domain Rn . Suppose that u represents a physical quantity such as temper-
ature or density. Direct observation of such quantities at a single point is a practical
impossibility. Even the most sensitive instrument will only be able to measure the
weighted average over some small region.
To formalize this notion of a local average, we use the concept of a test function

Ccpt (). The test function defines a local measurement of a quantity u through
the integral
u d n x. (10.1)


178 10 Weak Solutions
Fig. 10.1 A smooth function and its piecewise linear approximation
The function plays the role of a experimental probe that takes a particular sample
of the values of u.
Let us consider how we would detect a derivative using test functions. Suppose
for the moment that u C 1 (R), with u = f . If we measure this derivative associated

using the test function Ccpt (R), the result is

u dx = f d x. (10.2)
R R
The fact that u = f is equivalent to the statement that (10.2) holds for all

Ccpt (R).
Note that the left-hand side could be integrated by parts, since has compact
support, yielding
u d x = f d x. (10.3)
R R
This condition now makes sense even when u fails to be differentiable. The only
requirement is that u and f be integrable on compact sets, a property we refer to as
local integrability. We can say that locally integrable functions satisfy u = f in the

weak sense provided (10.3) holds for all Ccpt (R).
To generalize this definition to a domain Rn , let us define the space of locally
integrable functions,

L 1loc () := f : C; f | K L 1 (K ) for all compact K .
The same equivalence relation (7.6) used for L p spaces applies to L 1loc , i.e., functions
that differ on a set of measure zero are considered to be the same.
Inspired by (10.3), for u and f L 1loc () we say that
u
= f
x j
as a weak derivative if

u dx = f dx (10.4)
x j

for all Ccpt (R). The condition (10.4) determines f uniquely as an element of
L loc (), by the following:
1
10.1 Test Functions and Weak Derivatives 179
Lemma 10.1 If f L 1loc () satisfies

f dn x = 0

for all Ccpt (), then f 0.
Proof It suffices to consider the case when is bounded, since a larger domain
could be subdivided into bounded pieces. For bounded the local integrability of f

implies that f L 2 (). By Theorem 7.5 we can choose a sequence k in Ccpt ()
such that k f in L (). This implies that
2
lim f, k =
f
2 .
k
The inner products f, k are zero by hypothesis, so we conclude that f 0.

Example 10.2 In R, consider the piecewise linear function

0, x < 0,
g(x) := x, 0 x 1,

1, x > 0.
If we ignore the points where g is not differentiable, then we would expect that the
derivative of g is given by

1, 0 x 1,
f (x) =
0, |x| > 1.
These functions are illustrated in Fig. 10.2.

Let us check that this works in the sense of weak derivatives. We take Ccpt (R)
and compute
1

g d x = x (x) d x + d x.
0 1
g(x) g (x)
Fig. 10.2 A piecewise linear function and its weak derivative

Using integration by parts on the first term and evaluating the second gives
1

g d x = (1) d x (1)
0

= f d x.

This verifies that g = f in the weak sense.

Example 10.3 For t R, define w L 1loc () by

w (t), t < 0
w(t) = (10.5)
w+ (t), t 0,

where w C 1 (R). For Ccpt (R),
0

w dt = w dt w+ dt
0
0

= [w+ (0) w (0)](0) + w dt + w+ dt.
0
The term proportional to (0) could not possibly come from the integral of against
a locally integrable function, because the value of the integrand at a single point does
not affect the integral. Hence w admits a weak derivative only under the matching
condition
w (0) = w+ (0).
If this is satisfied, then the derivative is

w (t), t < 0,
w (t) =
w+ (t), t > 0.

Weak derivatives of higher order are defined by an extension of (10.4). To write the
corresponding formulas, it is helpful to have a simplified notation for higher partials.
For each multi-index = (1 , . . . , n ) with j N0 , we define the differential
operator on Rn ,
1 n
D := 1 . (10.6)
x1 xnn
The order of this operator is denoted by
|| := 1 + + n .
10.1 Test Functions and Weak Derivatives 181
Repeated integration by parts introduces a minus sign for each derivative. Therefore,
a function u L 1loc () admits a weak derivative D u L 1loc () if

||
(D u) d x = (1)
n
u D d n x (10.7)

for all Ccpt ().
It might seem that we should distinguish between classical and weak derivatives
in the notation. This is made unnecessary by the following:
Theorem 10.4 (Consistency of weak derivatives) If u C m () then u is weakly

differentiable to order k and the weak derivatives equal the classical derivatives.
Conversely, if u L 1loc () admits weak derivatives D u for || m, and each

D u can be represented by a continuous function, then u is equivalent to a function
in C m () whose classical derivatives match the weak derivatives.
In one direction the argument is straightfoward. Classical derivatives satisfy the

criterion (10.7) by integration by parts, so they automatically qualify as weak deriv-
atives. Lemma 10.1 shows that weak derivatives are uniquely defined.
The argument for the converse statement, that continuity of the weak derivatives
D u implies classical differentiability, is much more technical and we will not be
able to give the details. The basic idea is to show that one can approximate u by a

sequence k Ccpt () such that D k D u uniformly on every compact subset
of for || m. The fact that the weak derivatives D u are continuous makes
this possible. Uniform convergence then allows the classical derivatives of u to be
computed as limits of the functions D k .
10.2 Weak Solutions of Continuity Equations
Consider the continuity equation on R introduced in Sect. 3.1,
u q
+ = 0, u|t=0 = g. (10.8)
t x
The flux q could depend on u as well as t and x. To allow for the nonlinear case, we
will assume that u is real-valued here.
Suppose for the moment that q is differentiable and u is a classical solution of

(10.8). Let be a test function in Ccpt ([0, ) R). Use of the closed interval [0, )
means that and its derivatives are not necessarily zero at t = 0. Pairing u t
with
and integrating by parts thus generates a boundary term,

u

dt = u
t=0 u dt.
0 t 0 t
On the other hand, the spatial integration parts has no boundary term,

q
dx = q d x.
x x
When the left-hand side of (10.8) is paired with and integrated over both t and x,
the result is thus

u q
+ d x dt = u +q d x dt
0 t x 0 t x

u|t=0 d x.

If u is a classical solution of (10.8), then

u +q d x dt + g|t=0 d x = 0 (10.9)
0 t x

for all Ccpt ([0, ) R).
The t = 0 integral in (10.9) makes sense for g L 1loc (R). Under this assumption,
we define u L 1loc ((0, ) R; R) to be a weak solution of (10.8) provided q
L 1loc ((0, ) R; R) and (10.9) holds for all test functions.
Example 10.5 Consider the linear conservation equation with constant velocity,
which means q = cu in (10.8). By the method of characteristics (Theorem 3.2),
the solution is
u(t, x) = g(x ct).
Let us check that this defines a weak solution for g L 1loc (R).

For Ccpt (R) the first term in (10.9) is

g(x ct) (t, x) c (t, x) d x dt. (10.10)
0 t x
To evaluate the integral, introduce the variables
= t, y = x ct,
and define
, y) := ( , y + c ).
(
By the chain rule,

= c . (10.11)
t x
10.2 Weak Solutions of Continuity Equations 183
The Jacobian determinant of the transformation ( , y) (t, x) is 1, so that

u c d x dt = g(y) dy d .
0 t x 0
The integration can now be done directly,

y).
( , y) d = (0,
0
This gives

y) dy
u c d x dt = g(y)(0,
0 t x

= g(y)(0, y) dy,

which verifies (10.9).
We saw in Example 10.3 that in one dimension a jump discontinuity precludes

the existence of a weak derivative. Example 10.5 shows that this is not the case in
higher dimension. For g L 1loc (R), the solution u(t, x) = g(x ct) could be highly
discontinuous. The direction of the derivative is crucial here; regularity is required
only along the characteristics.
As an application of the weak formulation (10.9), let us return to an issue that
arose in the traffic model in Sect. 3.4. For certain initial conditions the characteristic
lines crossed each other, ruling out a classical solution of the PDE. We will see that
weak solutions can still exist in this case.
Consider a one-dimensional quasilinear equation of the form
u
q(u) = 0 (10.12)
t x
with the flux q(u) a smooth function of u which is independent of t and x. As we saw
in Sect. 3.4, the characteristics are straight lines whose slope depends on the initial
conditions. Let us study the situation pictured in Fig. 3.11, where a shock forms as
characteristic lines cross at some point. For simplicity we assume that the initial
crossing occurs at the origin.
One possible way to resolve the issue of crossing characteristics is to subdivide the
(t, x) plane into two regions by drawing a shock curve C, as illustrated in Fig. 10.3.
Suppose that classical solutions u are derived by the method of characteristics above
and below this curve. We will show that this combination yields a weak solution
provided a certain jump condition is satisfied along C. The jump condition was
discovered in the 19th century by engineers William Rankine and Pierre Hugoniot,
who developed the first theories of shock waves in the context of gas dynamics.
Fig. 10.3 Shock curve with x

solutions u on either side
u+
Theorem 10.6 (Rankine-Hugoniot condition) Let C be a curve parametrized as

x = (t) with C 1 [0, ). Suppose that u is a weak solution of (10.12) given by

u (t, x), x < (t),
u(t, x) =
u + (t, x), x > (t),
where u are classical solutions. Then, at each point of C,
q(u + ) q(u ) = (u + u ) . (10.13)
Proof Since we are not concerned with the boundary conditions, we consider a test

function Ccpt ((0, ) R), for which (10.9) specializes to

u + q(u) d x dt = 0. (10.14)
0 t x
Since the solutions u are classical and is C 1 , we can separate the integral (10.14)
at the shock curve and integrate by parts on either side.
Consider first the u side. For the term involving the x derivative the integration
by parts is straightforward:
(t) (t)

q(u ) d x dt =
q(u ) d x dt
0 x 0 x

+ q(u )
x=(t) dt.
0
For the t-derivative term we start by using the fundamental theorem of calculus to
derive
10.2 Weak Solutions of Continuity Equations 185
(t)
d (t)
u
u d x = u + d x + (t)u
x=(t) .
dt t t
By the compact support of , the integral over t of the left-hand side vanishes,
yielding
(t) (t)
u
u d x dt = d x dt
0 t 0 t

(t)u
x=(t) dt.
0
Combining these integration by parts formulas gives

(t)

+ q(u )
u d x dt
0 t x
(t)
u
= q(u ) d x dt
0 t x

(t)u q(u )
dt.
0 x=(t)
The first term on the right vanishes by the assumption that u is a classical solution,
leaving
(t)
u + q(u ) d x dt = u q(u )
dt.
0 t x 0 x=(t)
The corresponding calculation on the u + side yields

u+ + q(u + ) d x dt = u + q(u + )
dt.
0 (t) t x 0 x=(t)
By (10.14) the sum of the u and u + integrals is zero, which implies

(u + u ) (q(u + ) q(u ))
dt.
0 x=(t)
Since this holds for all C ((0, ) R), we conclude that

(u + u ) (q(u + ) q(u ))
=0
x=(t)
for all t > 0.

Example 10.7 Consider the traffic equation introduced in Sect. 3.4,

u u
+ (1 2u) = 0,
t x
for which q(u) = u u 2 . For the initial condition take a step function,

a, x < 0,
g(x) := (10.15)
b, x > 0.
From 3.28 the characteristic lines are given by

x0 + (1 2a)t, x0 < 0,
x(t) =
x0 + (1 2b)t, x0 > 0.
These intersect to form a shock provided a < b.

The solutions above and below the shock line are given by constants,
u (t, x) = a, u + (t, x) = b.
The Rankine-Hugoniot condition (10.13) thus reduces to
q(b) q(a) = (b a) .
Substituting with q(u) = u u 2 reduces this condition to
= 1 b a.
Since the discontinuity starts at the origin, the shock curve is thus given by
(t) = (1 b a)t.
Hence the weak solution is

a, x < (1 b a)t,
u(t, x) =
b, x > (1 b a)t.
Some cases are illustrated in Fig. 10.4. In the plot on the left, the shock wave
propagates backwards.
For certain initial conditions, the definition (10.9) of a weak solution is not suf-
ficient to determine the solution uniquely. For example, if we had taken a > b in
(10.15), then instead of overlapping the characteristics originating from t = 0 would
separate, leaving a triangular region with no characteristic lines. An additional physi-
cal condition is required to specify the solution uniquely in this case. We will discuss
this further in the exercises.
10.3 Sobolev Spaces 187
x x
u+ = 0.5
u+ = 0.7
t t
u = 0.3
u = 0.4
Fig. 10.4 Characteristic lines meeting at the shock wave
10.3 Sobolev Spaces
Boundary values are not well defined for locally integrable functions. We were able
to avoid this issue in the discussion of the continuity equation in Sect. 10.2, because
solutions were required to be constant along characteristics. In general, the formu-
lation of boundary or initial conditions requires a class of functions with greater
regularity.
The most obvious class to consider consists of functions that admit weak higher
partial derivatives. However, it proves to be very helpful to strengthen the integrability
requirements as well. Such function spaces were introduced by Sergei Sobolev in
the mid 20th century and have since become fundamental tools of analysis.
The Sobolev spaces based on L 2 are defined by

H m () := u L 2 (); D u L 2 () for all || m ,
for m N0 , with derivatives interpreted in the weak sense. An extended family of

Sobolev spaces W m, p is given by replacing L 2 with L p in the definition. The extended
family is important in the analysis of nonlinear PDE, but our focus will be limited to
linear applications involving H m .
Sobolev spaces are useful as theoretical tools, but they also have a practical side.
For a bounded domain , the space H 1 () includes the continuous piecewise linear
functions. A function is called piecewise linear if the domain can be decomposed into
a finite number of polygonal subdomains, on which the function is linear. Figure 10.5
shows a two-dimensional example. Sobolev spaces provide a natural framework for
the approximation of solutions by computationally simple classes of functions.
The space H m () carries a natural inner product,
Fig. 10.5 Graph of a

piecewise linear H 1 function
on the unit square

u, v H m := D u, D v . (10.16)
||m
(Our convention will be that a bracket without subscript denotes the L 2 inner product.)
The corresponding norm is
21

u
H m :=
D u
22 . (10.17)
||m
Lebesgue integration theory gives us the following completeness result, analogous

to Theorem 7.7.
Theorem 10.8 For Rn and m N0 , H m () is a Hilbert space.

Recall that Theorem 7.5 says that Ccpt () is a dense subspace of L 2 (). This

means that the closure of Ccpt () with respect to the L 2 norm is L 2 (). This result
no longer holds for the Sobolev space H m () with m 1. In particular, the closure

of Ccpt () with respect to the H 1 norm defines a subspace

H01 () = u H (); lim
u k
H = 0 for k Ccpt () .
m
1 (10.18)
k
By Lemma 7.8 H01 () is also a Hilbert space with respect to the H 1 norm.
If is piecewise C 1 , then for functions in H 1 () it is possible to define bound-
ary restrictions in L 2 () that generalize the boundary restriction of a continuous
function. In this case, H01 () consists precisely of the functions whose boundary
restriction vanishes. Thus the space H01 () can be interpreted as the class of H 1
functions satisfying Dirichlet boundary conditions on .
The theory of boundary restrictions is too technical for us to cover here, but we
can at least show how this works in the one-dimensional case.
Theorem 10.9 If u H01 (a, b) then u is continuous on [a, b] and equal to zero at
the endpoints.

Proof Suppose u H01 (a, b). By definition, there exists a sequence of Ccpt (a, b)
such that
10.3 Sobolev Spaces 189
lim
k u
H 1 = 0.
k
For x [a, b], x

j (x) k (x) = [ j (t) k (t)] dt
a
The integral on the right could be expressed as an inner product on R,

x
[ j (t) k (t)] dt = j k , [a,x] ,
a
where I denotes the characteristic function of the interval I . Thus, by the Cauchy-
Schwarz inequality (Theorem 7.1),

| j (x) k (x)| x a
j k
2 .
In view of the definition of the H 1 norm, this implies the uniform bound

j k
b a
j k
H 1 (10.19)
Since {k } converges and is therefore Cauchy with respect to the H 1 norm, it

follows from (10.19) implies that the sequence {k } is also Cauchy in the uniform
sense. By the completeness of L (a, b) (Theorem 7.7) and Lemma 8.4, this implies
that k g uniformly for some g C 0 [a, b].
At this point we have k u in H 1 and k g uniformly. Uniform convergence
on a bounded interval implies convergence in L 2 , by a simple integral estimate.
Therefore k g in L 2 also, implying that u = g in L 2 . Hence u C 0 [a, b].
To show that u vanishes at the endpoints, note that
max {|u(a)|, |u(b)|} sup |k u|. (10.20)

[a,b]
because k (a) = k (b) = 0. By uniform convergence, the left-hand side of (10.20)

approaches zero as k , showing that
u(a) = u(b) = 0.
In higher dimensions, functions in H 1 are not necessarily continuous. However,

m
H does imply continuity if m is sufficiently large relative to the dimension. We will
develop this regularity theory in Sect. 10.4.
We conclude this section with an extension property that will prove useful in
Chap. 11.
Lemma 10.10 For Rn , the extension by zero of an element of H01 ()

1
gives an element of H0 ().
The weak gradient
Proof For u H01 (), let u denote the extension by zero to .
u L 2 (; Rn ) can also be extended by zero to u L 2 (; Rn ). We need to
is the weak gradient of u.
show that u This is the condition that

dn x =
u
u d n x, (10.21)

for all Ccpt ().

By the definition of H01 (), there exists a sequence of k Ccpt () such that
k u in the H norm. Since k has compact support within , integration by
1
parts gives
k d n x = k d n x. (10.22)

By the H 1 convergence k u, we can take the limit k on both sides of

(10.22) to obtain
u d x =
n
u d n x.

are equal to u and u on and vanish on

Since u and u , this is equivalent
to (10.21).

10.4 Sobolev Regularity
In this section we will consider the relationship between weak regularity, defined in
terms of Sobolev spaces, and regularity in the classical sense. This connection plays
a central role in the application of Sobolev spaces to PDE.
Theorem 10.11 (Sobolev embedding theorem) Suppose Rn is a bounded

domain. If m > k + n2 , then
H m () C k ().
This result can be sharpened and extended in various ways. One important variant
includes differentiability up to the boundary under certain conditions on . For
example, if the boundary is piecewise C 1 then it is possible to show that
H m () C k ().
These boundary results are quite important but too technically difficult for us to
include here.
10.4 Sobolev Regularity 191
The strategy we will use for Theorem 10.11 is based on the connection estab-
lished in Sect. 8.6 between regularity and the decay of Fourier coefficients. Recall
the definition T := R/2Z introduced in Sect. 8.2. To extend Fourier series to higher
dimensions we introduce the corresponding space
Tn := Rn /(2Z)n .
A function on Tn is a function on Rn which is 2-periodic in each coordinate.

The periodic Fourier series theory from can be carried over to Tn directly. For
f L 2 (Tn ) and k Zn we define the coefficients

1
ck [ f ] := ei kx f (x) d n x. (10.23)
(2)n Tn
The integral over Tn can be taken over [, ]n , or any translate of this cube. The
argument from Theorem 8.6 can be adapted, with minor notational changes, to prove
the following:
Theorem 10.12 For f L 2 (Tn ), the series

ck [ f ]ei kx
kZn
converges to f in the L 2 norm.
As a corollary, we obtain the generalization of the Parseval identity (8.36),

f, g = (2)n ck [ f ]ck [g] (10.24)
kZn
for f, g L 2 (Tn ).
Because of the periodic structure of Tn , it is not necessary to assume that test func-
tions have compact support. For f L loc
1
(T) the weak derivative D f L loc1
(T) is
defined by the condition that
2 2
||
D f d x = (1) f D d x (10.25)
0 0
for all C (T). The space H m (Tn ) consists of functions in L 2 (Tn ) which have
weak partial derivatives up to order m contained in L 2 (Tn ).
It is convenient notate powers of the components of k by analogy with D ,
k := k11 knn ,
for = (1 , . . . , n ) with j N0 . A simple computation shows that

D ei kx = (i k) ei kx .
Thus, for f H m (T), substituting ei kx into (10.25) gives
ck [D f ] = (i k) ck [ f ] (10.26)
for || m. This generalizes the integration by parts formula (8.30).
Theorem 10.13 A function f L 2 (T) lies in H m (T) for m N if and only if

|k|2m |ck [ f ]|2 < . (10.27)
kZn
Proof By (10.26) and Bessels inequality (Proposition 7.9), the condition that D f
L 2 (Tn ) implies that

k ck [ f ]
2 < . (10.28)
kZn
This holds for all || m, implying (10.27).

Conversely, if f L 2 (Tn ) satisfies (10.27), then (10.28) holds for || m. We
can therefore define functions g L 2 (Tn ) by the Fourier series

g (x) := (i k) ck [ f ]ei kx .
kZn

By Parsevals identity (10.24), the inner product of g with Ccpt (Tn ) gives

g , = (2)n (i k) ck [ f ]ck []
kZn

||
= (1) 2 ck [ f ]ck [D ]
kZ
||
= (1) f, D .
This shows that the weak derivative D f exists and is equal to g .

Theorem 10.13 makes the connection between Sobolev regularity and decay of
Fourier coefficients. Our task is now to translate this back into classical regularity.
Theorem 10.14 (Periodic Sobolev embedding) If m > q + n2 , then
H m (Tn ) C q (Tn ).
Proof Using the notation for discrete spaces introduced in Sect. 7.4, the space 2 (Zn )
is defined as the Hilbert space of functions Zn C, equipped with the inner product
10.4 Sobolev Regularity 193

, 2 := (k)(k).
kZn
Consider the function

(k) := (1 + |k|)m .
The 2 norm of can be estimated with an integral,

22 := (1 + |k|)2m
kZn

(1 + |x|)2m d n x
Rn

= An (1 + r )2m r n1 dr.
0
The integral is finite if 2m > n, implying that 2 (Zn ) for m > n2 .

By Theorem 10.13, for f H m (Tn ) we can also define an element of 2 (Zn ) by
(k) := (1 + |k|)m |ck [ f ]|,
so that
, 2 = |ck [ f ]|.
kZn
It then follows from the Cauchy-Schwarz inequality on 2 that

|ck [ f ]|
2
2 , (10.29)
kZn
which is finite for m > n2 .

Since |ei kx | = 1, the estimate (10.29) implies that the Fourier series for f con-
verges uniformly. By Lemma 8.4 the limit of this series is continuous. Thus, after
possible replacement by an equivalent function in L 2 , f is continuous.
This argument shows that
H m (Tn ) C 0 (Tn ) (10.30)
for m > n2 . To apply it to higher derivatives we note that if f H m (Tn ) for m > q + n2
then for || q the weak derivatives D f will lie in H mq (T N ). For m > q + n2
it follows from (10.30) that these derivatives are continuous. By Theorem 10.4, this
shows that u C q (Tn ).

We are now prepared to derive the Sobolev embedding result for a bounded domain
as a consequence of Theorem 10.14.
Proof of Theorem 10.11 Suppose u H m () for Rn , let x 0 . Because

is open, we can choose > 0 small enough that
B(x 0 ; ) .

Suppose that Ccpt () has support contained in B(x 0 ; ) and is equal to 1
inside B(x 0 ; /2). (Such a function can be constructed as in Example 2.2.) Since
is smooth, u H m () also. Thus, assuming < 2, we can extend u by
periodicity to a function in H m (Tn ). Theorem 10.14 then shows that u C k (Tn )
if m > k + n/2. Since u and u agree in a neighborhood of x 0 , this shows that u
is k-times continuously differentiable at x 0 . This argument applies at every interior
point of , so we conclude that u C k ().

10.5 Weak Formulation of Elliptic Equations
The Laplace equation introduced in Sect. 9.1 is the prototypical elliptic equation.
Another classic example is the Poisson equation u = f , which we will discuss
in more detail in Sect. 11.1.

If Rn is a bounded domain, then for u, Ccpt (), Greens first identity
(Theorem 2.10) gives

u d n x = u d n x. (10.31)

On the other hand, the H 1 inner product on is given by

u, H 1 := u, + u d n x. (10.32)

The right-hand side of (10.31) is thus well-defined for u H01 ().

To account for applications to the Helmholtz equation as well as the Laplace
equation, let us consider the PDE
u = u + f, u| = 0. (10.33)
For f L 2 (), we say that u H01 () constitutes a weak solution of (10.33) if

u u f d n x = 0 (10.34)

for every Ccpt (). This definition could be extended to more general elliptic
equations of the form Lu = f , but for simplicity we restrict our attention to the case
of the Laplacian. We will study the existence and regularity of solutions of (10.34)
10.5 Weak Formulation of Elliptic Equations 195
extensively in Chap. 11. For now, we consider a simple one-dimensional case to

illustrate how the definition works.
Example 10.15 On the interval [0, 2], consider the equation
u = f, u(0) = u(2) = 0,
with
x, 0 x 1,
f (x) =
1, 1 x 2.
Since f is piecewise linear, it makes sense to try using classical solutions on the two
subintervals. Imposing the boundary and continuity requirements gives a family of
possible solutions

x ax,
1 3
0 x 1,
u(x) = 6
21 x 2 + (a + 4
3
)x 2a 2
3
, 1 x 2.
To determine a we apply the weak solution condition,

2
u + f d x = 0, (10.35)
0

for Ccpt (0, 2). Using integration by parts, the first term evaluates to
2 1 2
u d x = ( 21 x 2 a) (x) d x + (x + a + 43 ) (x) d x
0 0 1
2 1 2
u d x = ( 21 a)(1) x(x) d x (a + 13 )(1) + (x) d x
0 0 1
2
= ( 16 2a)(1) f d x.
0
The weak solution condition (10.35) requires a = 1

12
. This gives

x 12
1 3 1
x, 0 x 1,
u(x) = 6
2 x + 12
1 2 17
x 56 , 1 x 2.
This result is illustrated in Fig. 10.6. Note that the condition on a corresponds to
a matching of the first derivatives at x = 1, so that u C 1 [0, 2].
We will show in Sect. 11.3 that solutions of (10.34) are unique, so the function
obtained in Example 10.15 is the only possible solution. The matching of derivatives
required for this solution is indicative of a more general regularity property for
solutions of elliptic equation, which we will discuss in detail in Sect. 11.4.
Fig. 10.6 One-dimensional

weak solution
10.6 Weak Formulation of Evolution Equations
The heat and wave equations are the primary examples of linear evolution equations.
Weak solutions for these equations can be defined by essentially the same strategy
used in Sect. 10.5. Starting from a classical solution, we pair with a test function and
use integration by parts to find the corresponding integral equation. Unfortunately,
the time dependence creates some technicalities in the definition that we are not
equipped to fully resolve here, but we can at least illustrate the basic philosophy by
working through some examples.
Consider first the wave equation on a bounded domain Rn with Dirichlet
boundary conditions,
2u
u = 0, u| x = 0, (10.36)
t 2
subject to the initial conditions
u|t=0 = g,
= h.
t t=0
Assuming u is a classical solution, pairing the wave equation for u with a test function

Ccpt ([0, ) ) gives
2
u
2 u d n x dt = 0.
0 t
Integration by parts for the u term works just as in (10.31), yielding

u d x =
n
u d n x.

In the t variable, we integrate by parts twice and pick up a boundary term each time
because the test function is not assumed to vanish at t = 0. The result is
10.6 Weak Formulation of Evolution Equations 197

2u
u
2 dt = g
dt
0 t t t=0 0 t t

2
= g
+ h|t=0 + u 2 dt.
t t=0 0 t
Combining this with the spatial integral yields

2

u 2 + u d n x dt
t
0
(10.37)

= g
dn x + h|t=0 d n x.
t t=0
As in Sect. 10.5, the Dirichlet boundary condition is imposed by assuming that

u(t, ) H01 () for all t. To make sense of the boundary and initial terms we need
to assume at least that g L 1loc () and h L 1loc [0, ). To interpret (10.37) we
also need to require that the spatial pairing of u with is integrable over t. This
condition is more technical. It turns out to be sufficient to assume that
u(t, )
H 1
is integrable as a function of t, but we will not attempt to justify this here. Instead
we will limit our discussion to examples for which the existence of the integrals in
(10.37) is clear.
Example 10.16 Consider the piecewise linear dAlembert solution for the wave
equation introduced in Sect. 1.2. On [0, 2] we take the initial conditions h = 0
and
x, 0 x 1,
g(x) :=
2 x, 1 x 2.
By Theorem 4.5, dAlemberts solution is given by extending g to an odd periodic

function on R with period 4, and then setting
1
u(t, x) = g(x + t) + g(x t) .
2
The linear components of the resulting solution are shown in Fig. 10.7. Because u is
piecewise linear and vanishes at x = 0 and 2, it is clear that u(t, ) H01 (0, 2) for
each t.
For this case the weak solution condition (10.37) specializes to
2 2 2

u
u 2 + d x dt = g
d x. (10.38)
0 0 t x x 0 t t=0
Checking this is essentially a matter of integration by parts, but the integrals must
be broken into many pieces for large t. As a sample case, let us assume that has
support in [0, 1) (0, 2).
Fig. 10.7 Piecewise linear x

wave solution u(t, x)
u=2x u=x2
u=1t
u=x u = x
The first integral in (10.38) becomes

1 1
1 2
2 1x
2 2
u d x dt = x
dt + (1 t) dt d x
0 0 t 2 0 0 t 2 1x t 2
2 1 x1
2 2
(1 t) 2 dt + (2 x) 2 dt d x
1 x1 t 0 t
1

= x (0, x) (1 x, x) d x
0 t
2

(2 x) (0, x) (x 1, x) d x
1 t
Similarly, the second term in (10.38) evaluates to

1 2
2
u
1 1t

d x dt = d x dt d x dt
0 0 x x 0 0 x 0 1+t x
1 1
= (t, 1 t) dt + (t, 1 + t) dt
0 0
1 2
= (1 x, x) d x + (x 1, x) d x.
0 1
Adding these pieces together gives

2 1 2
2 u
u + d x dt = x (0, x) d x (2 x) (0, x) d x
0 0 t 2 x x 0 t 1 t
2

= g(x) (0, x) d x,
0 t
which verifies (10.38) for this case.

Now let us consider the weak formulation of the heat equation with Dirichlet
boundary conditions,
u
u = 0, u| x = 0, u|t=0 = h. (10.39)
t
Derivation of the integral equation works just as for the wave equation, except that
there is only a single integration by parts in the time variable. Assuming that u(t, )
H01 () for each t > 0 and h L 1loc (), the weak solution condition is

u + u d x dt =
n
h|t=0 d n x (10.40)
0 t

for all Ccpt ([0, ) ).
Example 10.17 Consider the heat equation on the interval (0, ), with initial con-
dition h L 2 (0, ; R). In view of the Dirichlet boundary conditions, we use the
orthonormal basis for L 2 (0, ) developed in Exercise 8.4, given by the sine func-
tions
2
k (x) := sin(kx)

for k N. The coefficients associated to h are

ak := h(x)k (x) d x,
0

and ak k converges to h in the L 2 sense by Theorem 8.6. The corresponding heat
solution is

ak ek t k (x).
2
u(t, x) = (10.41)
k=1
For example, solution corresponding to a step function h is illustrated in Fig. 10.8.
Fig. 10.8 Heat solution with x

L 2 initial data
t
As noted in Corollary 8.14, this yields a classical solution if h is continuous. If h

is not continuous then u might not have a well-defined limit as t 0. Nevertheless,
we can check that the weak solution condition is satisfied. Given a (real-valued) test

function Ccpt ([0, )(0, ); R), define the time-dependent Fourier coefficients

bk (t) := (t, x)k (x) d x.
0
By the smoothness of , Theorem 8.10 implies that the coefficients satisfy bk (t) =
O(k ), uniformly in t, and so the series

(t, x) = bk (t)k (x)
k=1
converges uniformly as well as in L 2 . By the same principle, the series

(t, x) = bk (t)k (x)
t k=1
is also uniformly convergent. Since {k } is an orthonormal basis, we deduce from

Parsevals identity (8.36) that

ak ek t bk (t)
2
u dt = (10.42)
0 t k=1
for t 0.
Similarly, for t 0 we have L 2 convergent series

2
(t, x) = bk (t)k cos(kx),
x k=1

u 2
k 2 t
(t, x) = ak (t)e k cos(kx).
x k=1

By the Parseval identity for the cosine basis, this gives

u
k 2 ak ek t bk (t).
2
dx = (10.43)
0 x x k=1
Applying (10.42) and (10.43) to the left-hand side of (10.40) yields


u n
u + d x dt
0 0 t x x

(10.44)
ak ek t k 2 bk (t) bk (t) dt.
2
=
0 k=1
Switching the order of the summation and integration is justified if series converges
uniformly on the domain of the integral, but that is not necessarily the case here. To
check this carefully, we break the sum at some value k = N . For the finite sum there
is no convergence issue, so that
N

d k 2 t
N
ak ek k 2 bk (t) bk (t) dt =
2
t
ak e bk (t) dt
0 k=1 k=1 0 dt

N
= ak bk (0).
k=1
To estimate the tail of the sum, note that the sequence {ak } is bounded because
|ak |2 < . For bk (t), we apply repeated integration by parts to deduce
2m 2m
k (x) (t, x) d x = (t, x) k (x) d x
x x
0 0

= (1)m k 2m (t, x)k (x) d x
0
= (1)m k 2m bk (t).

for m N. Since Ccpt ([0, ) (0, )), this gives an estimate
|bk (t)| Cm k 2m ,
where Cm is independent of t. The same reasoning applies to bk (t). Combining the
m = 2 estimate for bk with the m = 1 case for bk gives

k 2 t 2

ak e k bk (t) bk (t)
Ck 2 .

This shows that

k 2 t

ak e k bk (t) bk (t)
C N 1 ,
2
(10.45)

k=N +1
independently of t.
Now fix M > 0 so the support of is contained in [0, M]. Applying (10.45) to
the integral gives

M

k 2 t

ak e k bk (t) bk (t) dt
C M N 1 .
2

0
k=N +1
Returning to (10.44), our analysis of the sum over k now gives

N
u
u + d n x dt = ak bk (0) + O(N 1 ).
0 0 t x x k=1
By taking N , we deduce

u
u + d n x dt = ak bk (0).
0 0 t x x k=1
On the other hand Parsevals identity gives

h(x)(0, x) d x = ak bk (0),
0 k=1
so the weak solution condition (10.40) is satisfied.
10.7 Exercises
10.1 On R consider the ordinary differential equation
du
x = 1.
dx
(a) Develop a weak formulation of this ODE in terms of pairing with a test function

Ccpt (R)
(b) Show that u(x) = log |x| is locally integrable and solves the equation in the
weak sense.
10.2 In Exercise 3.6 we studied Burgers equation,
u u
+u = 0,
t x
with the initial condition
10.7 Exercises 203

a, x 0,
u(0, x) = a(1 x) + bx, 0 < x < 1,

b, x 1.
For a > b a shock forms at some positive time. Assuming u is a weak solution, find
the equation of the shock curve starting from this point.
10.3 Consider the traffic equation
u u
+ (1 2u) =0
t x
with initial data
0, x > 0,
g(x) =
1, x < 0.
(a) Sketch the characteristic lines for this initial condition, and show that they leave
a triangular region uncovered.
(b) Show that the constant solution u(t, x) = g(x) satisfies the Rankine-Hugoniot
condition for the shock curve (t) = 0 and thus gives a weak-solution of the
traffic equation. (This solution is considered unphysical because characteristic
lines emerge from the shock line to fill the triangular region.)
(c) The physical solution is specified by an entropy condition that says that charac-
teristic lines may only intersect when followed forwards in time. Show that the
continuous function

0, x > t,
u(t, x) = 2 2t , t < x < t,
1 x

1, x < t,
satisfies the weak solution condition (10.9) (with q = u u 2 ), as well as this

entropy condition. (This type of solution is called a rarefaction wave.)
10.4 Define f L 1loc (Rn ) by

f + (x), xn > 0,
f (x) =
f (x), xn < 0.
with f C 1 (Rn ).
f
(a) For j = 1, . . . , n 1, show that f has weak partial derivatives given by x j
for
xn > 0.
f f
(b) Show that the weak partial x n
exists and is given by xn
for xn > 0 only if f
extends to a continuous function at xn = 0.
10.5 Let D R2 be the unit disk {r < 1} with r := |x|. Consider the function
u(x) = r with R constant.
u
(a) Compute the ordinary partial derivatives x j
, j = 1, 2, for r = 0.
(b) Show that for > 1 these partials lie in L 1 (D) and define weak derivatives.
(c) For what values of is u H 1 (D)?
10.6 In R3 consider the equation

1, r a,
u =
0, r > a,
with r := |x| and a > 0. (With appropriate physical constants this is the equation
for the gravitational potential of a spherical planet of radius a.)
(a) Assuming that u depends only on r , formulate a weak solution condition in terms

of pairing with a test function (r ) with Ccpt [0, ).
(b) Find the unique solution which is smooth at r = 0 and for which u(r ) 0 as
r .
10.7 Let Rn be bounded with piecewise C 1 . For u C 2 () and f

C 0 (), suppose that
u f d n x = 0. (10.46)

for all C (). Show that u satisfies the Poisson equation with Neumann bound-
ary condition,
u
u = f,
= 0. (10.47)

(Thus (10.46) allows a weak formulation of (10.47) for u H 1 ().)

Chapter 11
Variational Methods
Recall the formulas for the kinetic and potential energy of a solution of the wave
equation derived in Sect. 4.7. At equilibrium the kinetic energy is zero, and by phys-
ical reasoning the system should occupy a state of minimum potential energy. This
suggests a strategy of reformulating the Laplace equation, which models the equi-
librium state, as a minimization problem for the energy.
In this application, the potential energy term from the wave equation is called the
Dirichlet energy. For a bounded domain Rn and w C 2 (), let

1
E[w] := |w|2 d n x. (11.1)
2
The term functional is used to describe functions such as E[], to indicate that the
domain is a function space.
To see how minimization of energy is related to the Laplace equation, let us
suppose that u C 2 (; R) satisfies
E[u] E[u + ]

for all Ccpt (; R). This implies that for t R the function t E[u + t]
achieves a global minimum at t = 0. Hence

d
E[u + t] = 0. (11.2)
dt t=0
Differentiation under the integral in the definition of E[u + t] gives

206 11 Variational Methods

d 1 d
E[u + t] = |u|2 + 2tu + t 2 ||2 d n x
dt t=0 2 dt
t=0
= u d x.
n

By Greens first identity (Theorem 2.10) and the fact that vanishes on ,

u d x = n
u d n x.

Thus (11.2) is equivalent to

u d n x = 0.

This holds for all Ccpt (; R) if and only if u = 0 on .
11.1 Model Problem: The Poisson Equation
The empirical law describing the electric field in the presence of a charge distribution
was formulated by Gauss in the mid-19th century. Gausss law states that the outward
flux of the electric field through a closed surface is proportional to the total electric
charge contained within the region bounded by the surface. More specifically, if
R3 is a bounded domain with piecewise C 1 boundary and is the charge
density within , then

E d S = 4k d 3 x,

where k is called Coulombs constant.

Using the divergence theorem (Theorem 2.6), we can rewrite the flux integral as

E dS = E d 3 x,

so that Gausss law becomes

E d 3 x = 4k d 3 x.

This holds for an arbitrary region if and only if
E = 4k (11.3)
(the differential form of Gausss law).

11.1 Model Problem: The Poisson Equation 207
In Sect. 9.1 we noted that the electric potential and electric field E are related
by
E = .
Substituting this into (11.3) gives the Poisson equation
= 4k.
A common electrostatics problem is to find the electric potential caused by a distribu-

tion of charges within a region bounded by a conducting material. The electric field
must be perpendicular to a conducting surface, implying that the boundary restriction
| is constant. If the boundary is connected then we can set this constant to 0, so
the potential satisfies the Poisson equation with Dirichlet boundary conditions.
Other forms of the Poisson equation appear in various contexts. For example in
Newtonian gravity the relationship between gravitational potential and the mass
density is
= 4G,
where G is Newtons gravitational constant. In this application the domain is R3 ,

and the physical assumption that 0 at infinity plays the role of a boundary
condition.
11.2 Dirichlets Principle
To solve Poissons equation using a minimization argument, it proves to be very

helpful to use the weak formulation of the equation. This is because H01 () is a
Hilbert space. The completeness of H01 () with respect to the H 1 norm will play an
essential role in establishing the existence of a minimizing function.
For Rn bounded, and f C 0 (), the classical Poisson problem is to find
a function u C 2 () C 0 () so that
u = f, u| = 0. (11.4)
The weak formulation of (11.4) is a special case of (10.34), We take f L 2 () and

the goal is to find u H01 () such that

u f d n x = 0, (11.5)

for every Ccpt ().
For convenience let us consider real-valued functions. (In the complex case we
could split the Poisson problem into real and imaginary parts.) In view of (11.5), we
define the functional
D f [w] := E[w] f, w
, (11.6)
for f L 2 (; R) and w H01 (; R), where E[] is the Dirichlet energy (11.1).
Theorem 11.1 (Dirichlets principle) Suppose Rn is a bounded domain and

f L 2 (; R). If u H01 (; R) satisfies
D f [u] D f [w]
for all w H01 (; R), then u is a weak solution of the Poisson equation, in the sense
of (11.5).

Proof Since Ccpt (; R) H01 (), the assumption on u implies
D f [u] D f [u + t]

for Ccpt (; R) and t R. Therefore
d
0= D f [u + t]t=0
dt
d
= E[u + t] f, u + t

dt t=0
n
= u f d x.

Dirichlets principle is a classic example of a variational method. This terminol-

ogy refers to the family of variations u + t used to derive the PDE from the min-
imization problem. In Sect. 11.3 we will show that D f [] attains a minimum within
H01 (), guaranteeing the existence of a weak solution. Furthermore, in Sect. 11.4 we
will see that the weak solution is actually a classical solution under certain conditions.
11.3 Coercivity and Existence of a Minimum
The functional D f defined in (11.6) consists of a quadratic term plus a linear term.
The Dirichlet minimization problem is thus analogous to minimizing the polynomial
ax 2 + bx for x R. This polynomial obviously has a minimum if and only if a > 0.
For the Dirichlet case the analogous condition is a lower bound on the quadratic term
E[]. The original form of this result was proven by Henri Poincar.
Theorem 11.2 (Poincar Inequality) For a bounded domain Rn , there is a

constant > 0, depending only on , such that
11.3 Coercivity and Existence of a Minimum 209
u22 2 E[u]
for all u H01 ().

Proof Because Ccpt () is dense in H01 (), we can restrict our attention to smooth
compactly supported functions at first and extend to the general case later. To illustrate
the core argument we start with the n = 1 case.

Consider a bounded interval (a, b) R. For Ccpt (a, b), our goal is to compare

the size of (x) to values of its derivative . The obvious connection between them
comes from the fundamental theorem of calculus:
x
(x) = (t) dt.
a
The right-hand side can be written as an L 2 pairing with the characteristic function
of [a, x],
(x) = , [a,x] .
The Cauchy-Schwarz inequality (Theorem 7.1) then gives
|(x)|2 [a,x] 22 22
(b a)E[],
for all x [a, b]. We can integrate this estimate over x to obtain
22 (b a)2 E[]

for Ccpt (a, b).
Now let us consider the higher dimensional case Rn . The domain is assumed
to be bounded, so
R := [M, M]n

for some large M. Functions in Ccpt () can be extended by zero to smooth functions

on R, so it suffices to derive the Poincar inequality for Ccpt (R). Following
the one-dimensional case, we apply the fundamental theorem of calculus in the x1
variable to write x1

(x) = (y, x2 , . . . , xn ) dy.
M x 1
By the Cauchy-Schwarz inequality on L 2 (M, M),

2
M
|(x)|2 2M (y, x , . . . , x )
n dy,
x 2
M 1
for all x R. Integrating this estimate over x R yields

2

22 4M 2

. (11.7)
x1 2
By the definition (11.1), the energy is given by

1
E[] = d n x
2 R

1 2 2
= + + dn x
2 R x1 xn
1

n

2
=

.
2 j=1 x j 2
Thus the bound (11.7) implies the estimate,
22 8M 2 E[], (11.8)

for Ccpt ().
To complete the argument, suppose that u H01 (). By the definition of H01

there exists an approximating sequence {k } Ccpt () such that k u in the H 1
norm. By (11.8) the inequality
k 22 8M 2 E[k ] (11.9)
holds for each k N. Our goal is thus to take the limit k on both sides.
For the energy side note that
n
u u k k
E[u] E[k ] = , ,
j=1
x j x j x j x j
n
(u k ) u k (u k )
= , + ,
j=1
x j x j x j x j
Hence by the Cauchy-Schwarz inequality
n

(u k )

u

k
|E[u] E[k ]|

+

.
j=1
x j 2 x j 2 x j 2
By the definition of the H 1 norm this yields
|E[u] E[k ]| u k H 1 (u H 1 + k H 1 ) .
In particular, k u in H 1 implies that
lim E[k ] = E[u].

k
Thus taking the limit k in (11.9) gives
u22 8M 2 E[u].
The Poincar constant () for a bounded domain is defined to be the optimal

choice of in Theorem 11.2. In other words,
u2
() := sup 1 .
uH01 ()\{0} E[u] 2
Our proof gives the rough estimate

() 8 diam(),
which is rather poor compared to the best known bounds. In Sect. 11.5 we will estab-
lish a direct relationship between the Poincar constant and the lowest eigenvalue of
on .
Since our goal is to minimize D f [] over H01 , it is useful to express the conclusion
of Theorem 11.2 in terms of the H 1 norm. Because
u2H 1 = u22 + E[u], (11.10)
(11.8) is equivalent to
u2H 1 (2 + 1)E[u] (11.11)
for all u H01 (). A quadratic functional on a Hilbert space is called coercive if its
ratio to the norm squared is bounded below. The Poincar inequality thus states that
E[] is coercive on H01 ().
The identity (11.10) also gives an upper bound,
E[u] u2H 1 . (11.12)
A quadratic functional is called bounded if its ratio to the norm squared is bounded
above. For the energy this condition is automatic.
We are now prepared to tackle the minimization problem for D f [], by exploiting
the fact that E[] is both coercive and bounded.
Theorem 11.3 For a bounded domain Rn and f L 2 () there is a unique

function u H01 () such that
D f [u] D f [w]
for all w H01 (), where D f [] is defined by (11.6).
Proof By the triangle inequality,
D f [w] E[w] | f, w
| .
Applying (11.11) to the energy and the Cauchy-Schwarz inequality to the inner
product gives
1
D f [w] 2 w2H 1 f 2 w2
+1
(11.13)
1
2 w2H 1 f 2 w H 1 .
+1
The right-hand side has the form cx 2 bx where c = 1/(2 + 1), b = f 2 , and
x = w H 1 . According to the minimization formula for a quadratic polynomial,
b
min(cx 2 bx) =
xR 4c
for c > 0. Applying this to (11.13) gives
2 + 1
D f [w] f 22 , (11.14)
4
for w H01 ().

If we set
d0 := inf D f [w],
wH01 ()
then (11.14) shows that d0 > . By Lemma 2.1, there exists a sequence of
wk H01 () so that
lim D f [wk ] = d0 . (11.15)
k
Our strategy is to argue that the sequence {wk } is Cauchy in H01 (), and therefore
converges by completeness.
The quadratic structure of E[] implies that

u+v 1 1 1
E = E[u] + E[v] E[u v] (11.16)
2 2 2 4
for all u, v H01 (). This allows us to compute

wk + wm 1 1 1 1
Df = E[wk ] + E[wm ] E[wk wm ] f, wk + wm
2 2 2 4 2
1 1 1
= D f [wk ] + D f [wm ] E[wk wm ].
2 2 4
Because d0 is the infimum of D f [], this implies
1 1 1
d0 D f [wk ] + D f [wm ] E[wk wm ]. (11.17)
2 2 4
Turning this inequality around gives
E[wk wm ] 2D f [wk ] + 2D f [wm ] 4d0 .
By (11.15),
lim 2D f [wk ] + 2D f [wm ] 4d0 = 0,
k,m
and since E[] 0 this yields
lim E[wk wm ] = 0. (11.18)

k,m0
Using the coercivity estimate (11.11), it follows from (11.18) that
lim wk wm H 1 = 0,
k,m0
i.e., the sequence {wk } is Cauchy in H01 (). Therefore, by completeness, there exists
a function u H01 () such that
u := lim wk .
k
Convergence in H 1 implies that
D f [u] = lim D f [wk ] = d0 .

k
Therefore u minimizes D f [].

To see that the minimizing function is unique, suppose that both u 1 and u 2 satisfy
D f [u j ] = d0 . By the same reasoning used to derive (11.17) we have
1 1 1
d0 = D f [u 1 ] + D f [u 2 ] E[u 1 u 2 ].
2 2 4
By assumption both D f [u 1 ] and D f [u 2 ] are equal to d0 , so this implies
E[u 1 u 2 ] = 0.
Theorem 11.2 then implies u 1 u 2 2 = 0, so that u 1 u 2 .
Note the crucial role that completeness plays in the proof of Theorem 11.3. If we
had taken the domain of the Dirichlet energy to C 2 (), then there would be no way
to deduce convergence of the sequence {wk } from the energy limit (11.18).
Corollary 11.4 For f L 2 () the weak formulation (11.5) of the Poisson equation
admits a unique solution u H01 ().
Proof Existence of the solution follows from Theorems 11.1 and 11.3. To prove
uniqueness, suppose that u 1 and u 2 both satisfy (11.5). Subtracting the equations
gives
(u 1 u 2 ) d n x = 0

for all Ccpt (). By the definition (10.18) of H01 () we can take a sequence of
k H0 () such that k u 1 u 2 in the H 1 norm. This implies in particular that
1
k (u 1 u 2 ) in the L 2 sense, so that

E[u 1 u 2 ] = lim (u 1 u 2 ) k d n x
k
= 0.
It follows from the Poincar inequality (Theorem 11.2) that u 1 u 2 = 0, hence

u1 u2.
11.4 Elliptic Regularity
If u = f in the classical sense, then f has a level of differentiability 2 orders

below that of u. Our goal in this section is to develop a converse to this statement
that allows us to deduce the Sobolev regularity of a weak solution u from that of f .
This type of regularity result holds for elliptic equations in general, but we will focus
on the Laplacian for simplicity.
To avoid complications near the boundary of the domain, we introduce local
versions of the Sobolev spaces. For Rn let

m
Hloc () := u L 1loc (); u H m () for all Ccpt () .
11.4 Elliptic Regularity 215
Theorem 11.5 (Interior regularity) Suppose that u Hloc

1
() is a weak solution of
u = f.
If f Hloc
m
() for m 0 then u Hloc
m+2
().
The Sobolev embedding theorem (Theorem 10.11) gives

m
Hloc () C k ()
for k < m n2 . Thus Theorem 11.5 shows that the weak solution of the Poisson
equation obtained in Corollary 11.4 is a classical solution provided f Hloc m
()

with m > 2 . Furthermore, if f C () then u C () also.
n
It is possible to include regularity up to the boundary, although this is technically

much more difficult. For example, if is C and u H01 () is a weak solution
of u = f for f C (), then u C ().
Our proof of Theorem 11.5 makes use of the Fourier analysis on Tn that we
introduced in Sect. 10.4.
Lemma 11.6 Suppose that u H 1 (Tn ) solves u = f in the weak sense. If
f H m (Tn ) for m N0 , then u H m+2 (Tn ).
Proof For u H 1 (Tn ) and f L 2 (Tn ), we assume that

[u f ] d n x = 0, (11.19)
Tn
for all C (T). Setting (x) = ei kx in this equation gives

[i k u(x) f (x)] ei kx d n x = 0.
Tn
Using 10.26, we can translate this into a relation between the Fourier coefficients,

n
u
ck [ f ] = ik j ck
j=1
x j (11.20)
= |k| ck [u].
2
According to Theorem 10.13, if f H m (Tn ) then

|k|2m |ck [ f ]|2 < .
kZn
By (11.20) this implies that


|k|2m+4 |ck [u]|2 < ,
kZn
which gives u H m+2 (Tn ) by Theorem 10.13.
We can now deduce the interior elliptic regularity result from Lemma 11.6 by
localizing.
Proof of Theorem 11.5. Suppose that u = f in the weak sense of (11.5), with
u H01 () and f L 2loc (). By rescaling and translating, if necessary, we can
assume that [0, 2]n , which allows us to identify with a subset of Tn . For

Ccpt () we can extend u by zero to a function on [0, 2]n . This allows us
to consider u as a function in H 1 (Tn ). Our goal is to apply Lemma 11.6 to the
localized function u.
We must first show that u satisfies a weak Poisson equation on Tn . Given

C (Tn ) we can use the test function Ccpt () in the weak solution condition
(11.5) to obtain
[u () f ] d n x = 0.
Tn
Using the product rule for the gradient, we can rewrite this as

0= [u + u f ] d n x
Tn
(11.21)
= [(u) u + u f ] d x. n
Tn
In order to interpret this as a weak equation for u, we need to rewrite the second
term as an integral involving rather than .

Since the components of u are in Ccpt (), the definition of the weak derivative
u gives

u () d n x = u () d n x

= [u + u] d n x.

Using this formula on the term u in (11.21) gives

0= (u) + 2u + u f d n x. (11.22)
Tn
This holds for all C (Tn ), implying that u is a weak solution, in the sense of
(11.19), of the equation
(u) = 2u u + f . (11.23)
11.4 Elliptic Regularity 217

Because Ccpt (), the right-hand side of (11.23) lies in L 2 (Tn ) by the
assumptions on u and f . Hence u H 2 (Tn ) by Lemma 11.6. This holds for

all Ccpt (), implying that u Hloc
2
(Tn ).
q
We can now apply the same argument inductively. If u Hloc (), and f
q1
Hloc (), then the right-hand side of (11.23) lies in H (T ) for Ccpt
q1 n
().
q+1
Lemma 11.6 then gives u Hloc ().
11.5 Eigenvalues by Minimization
In Sect. 7.6, we mentioned that the spectral theorem for finite-dimensional matrices
has an extension to certain differential operators. In this section, we will prove this
result in the classical setting of the Laplacian on a bounded domain with Dirichlet
boundary conditions.
Theorem 11.7 (Spectral theorem for the Dirichlet Laplacian) Let Rn be a

bounded domain. There exists an orthonormal basis {k }kN for L 2 () such that
k = k k
with k H01 (; R) C (). Furthermore, k > 0 for all k and
lim k = .
k
In the case of L 2 [0, ], the sequence of Dirichlet eigenfunctions is given by

k (x) = sin(kx) for k N, with eigenvalues k = k 2 . If follows from Theorem 8.6
that this sequence yields a basis, as shown in Exercise 8.4.
We can solve the eigenvalue problem in the general case by adapting the vari-
ational method used for the Poisson equation earlier in this chapter. This method
gives more than just existence of the basis; it also suggests a natural strategy for
numerical approximation of eigenvalues and eigenfunctions, which we will explore
in Sect. 11.7.
The weak formulation of the eigenvalue equation = with Dirichlet
boundary conditions is a special case of (10.34). For H01 (), the condition is
that
[ ] d n x = 0 (11.24)

for all Ccpt ().
If we substitute in place of then the second term in (11.24) is times the L 2
inner product ,
. By the same token we could interpret the first term as an inner
product form of the Dirichlet energy (11.1). We will denote this by

E[u, v] := u v d n x,

so that E[u] = E[u, u]. The L 2 and H 1 inner products are related by
u, v
H 1 = u, v
+ E[u, v].
With this convention we can write the weak eigenvalue equation (11.24) in an equiv-
alent form as
E[, ] = ,
(11.25)

for all Ccpt ().
For the minimization argument, it will prove convenient to enlarge the space of test

functions from Ccpt () to H01 (). To justify this, note that by definition a function

v H0 () can be approximated by a sequence of k Ccpt
1
() with respect to the
1
H norm. This implies in particular that
lim E[, k ] = E[, v], lim , k

= , v
.
k k
We can thus conclude that solves (11.24) if and only if
E[, v] = , v
(11.26)
for all v H01 ().

The formulation of the eigenvalue equation as a minimization problem is known
as Rayleighs principle, after the physicist Lord Rayleigh. For v H01 (), v = 0,
the ratio
E[v]
R[v] :=
v22
is called the Rayleigh quotient for v. Note also that if satisfies (11.26) then
R[] = . (11.27)
Furthermore, the Poincar inequality (Theorem 11.2) shows that
1
R[v] >0 (11.28)
2
for v H01 (), v = 0. This suggests that the smallest eigenvalue is related to the
Poincar constant, and that we can locate it by minimizing R[].
The argument for existence of a minimum is a little trickier than the analysis of
the Dirichlet principle in Sect. 11.3. To understand why, note that
R[cv] = R[v]
11.5 Eigenvalues by Minimization 219
for c C\{0}, so the minimizing function is not unique. Therefore it is quite possible
to have a sequence that minimizes the Rayleigh quotient but does not converge in H 1 .
The tool that allows us to resolve this issue was developed by Franz Rellich in the
early 20th century.
Theorem 11.8 (Rellichs theorem) Suppose Rn is a bounded domain and {vk }
is a sequence in H01 () that satisfies a uniform bound
vk H 1 M
for all k N. Then {vk } has a subsequence that converges in L 2 ().

It is important to note that Rellichs theorem refers to two different norms. The
sequence is assumed bounded in the H 1 norm, and the subsequence is guaranteed
to converge with respect to the L 2 norm. This is crucial to the result and not a mere
technicality. We will defer the proof of Rellichs theorem to Sect. 11.6. The remainder
of this section is devoted to its application to the Rayleigh minimization scheme.
Theorem 11.9 (First eigenvalue) There exists 1 H01 (; R) C () satisfying
1 = 1 1
for 1 > 0, such that

1 R[v] (11.29)
for all v H01 (), v = 0.

Proof By (11.28) the values of R[] are bounded below by the Poincar constant.
Therefore, the infimum
1 := inf R[v]
vH01 ()\{0}
exists and is strictly positive.

By Lemma 2.1 there exists a sequence {vk } H01 ()\ {0} such that
lim R[vk ] = 1 . (11.30)

k
After rescaling each vk by a constant we can assume that vk 2 = 1, so that
R[vk ] = E[vk ].
The sequence of energies E[vk ] is bounded by (11.30). This also implies that the
sequence {vk } is bounded with respect to the H 1 norm, because
vk 2H 1 = 1 + E[vk ]
by the relation (11.10).

According to Theorem 11.8 there exists a subsequence of {vk } that converges in

the L 2 sense to some function 1 L 2 . By restrict our attention to this subsequence
and relabeling if needed, we can assume that
lim vk 1 2 = 0. (11.31)
k
The next goal is to improve this to a statement of convergence in H 1 . We use the

same strategy as in Sect. 11.3, starting from

vk + vm 1 1 1
E = E[vk ] + E[vm ] E[vk vm ]. (11.32)
2 2 2 4
By the definition of 1 ,
vk + vm
R 1 .
2
Therefore
vk + vm 1
E vk + vm 22 .
2 4
Using this in (11.32) gives
E[vk vm ] 2E[vk ] + 2E[vm ] 1 vk + vm 22 . (11.33)
By (11.31), we have
lim vk + vm 22 = 21 22 = 4,
k,m
and by construction E[vk ] 1 . Hence (11.33) implies that
lim E[vk vm ] 0.
k,m
Since we already know that {vk } converges in the L 2 norm, we conclude that
lim vk vm H 1 0.
k,m
That is, the sequence {vk } is Cauchy with respect to the H 1 norm.
By completeness {vk } converges with respect to the H 1 norm to some u H01 ().
Since the L 2 norm is bounded above by the H 1 norm, this means vk u in the
L 2 sense also. Hence u 1 , which proves that 1 H01 (). It then follows from
(11.30) that
E[1 ] = R[1 ] = 1 .
To see that this implies the weak solution condition, suppose w H01 (). Using
the inner product form of E[], we can expand
E[1 + tw] = 1 + 2t Re E[1 , w] + t 2 E[w], (11.34)
for t R. Similarly,
1 + tw22 = 1 + 2t Re 1 , w
+ t 2 w22 . (11.35)
By the definition of 1 , the function t R[1 + tw] has a minimum at t = 0, so

that
d
R[1 + tw] = 0.
dt t=0
Computing this derivative using (11.34) and (11.35) gives

d
R[1 + tw] = 2 Re E[1 , w] 21 Re 1 , w
.
dt t=0
We thus conclude that

Re E[1 , w] = 1 Re 1 , w
for all w H01 (). By replacing w by iw, we can deduce also that
Im E[1 , w] = 1 Im 1 , w
for all w H01 (). In combination these give (11.26), so 1 is a weak solution of
1 = 1 1 .
In principle, 1 could be complex-valued at this point, but since its real and imag-
inary parts each satisfy (11.26) separately, we can select one of these to specialize
to the real-valued case.
To deduce the regularity of 1 , we apply Theorem 11.5 with f = 1 . The fact
that 1 H01 () then implies that 1 Hloc
3
(). Starting from Hloc 3
() then gives
q
1 Hloc (), and so on. This inductive argument shows that 1 Hloc () for each
5

q N. We conclude that 1 C () by Theorem 10.11.
It is clear that the 1 produced in Theorem 11.9 is the smallest eigenvalue, since
all eigenvalues occur as values of the Rayleigh quotient by (11.27). An example of 1
is shown in Fig. 11.1. We will see in the exercises that the first eigenfunction cannot
have zeros in . This can be used to show that the first eigenfunction is unique up to
a multiplicative constant, i.e., 1 has multiplicity one.
To find other eigenvalues, the strategy is to restrict to subspaces and then apply the
same construction used for 1 . For a subset A L 2 () the orthogonal complement is
Fig. 11.1 The first

eigenfunction of a equilateral
triangle domain

A := w L 2 (); w, v
= 0 for all v A .
The orthogonal complement of a set is a subspace, by the linearity of the inner

product.
In the argument below we will consider subspaces of H01 () of the form
W = A H01 (),
where A is a finite list of eigenfunctions. We claim that W is closed as a subspace

of H01 (). To see this, suppose that wk w in the H 1 norm, with wk W . Since
wk w2 wk w H 1 by (11.10), this implies that wk w with respect to the
L 2 norm also. Thus for v A,
w, v
= lim wk , v
= 0.
k
This shows that w W . Therefore W is closed. By Lemma 7.8, this implies that W
is a Hilbert space with respect to the H 1 inner product
Proof of Theorem 11.7 Let 1 H01 (; R) be the eigenvector obtained in
Theorem 11.9, normalized so that 1 2 = 1. The subspace
W1 := {1 } H01 ()
is a Hilbert space with respect to the H 1 norm, by the remarks above.

Applying the minimization procedure used in Theorem 11.9 to the restriction of
the Rayleigh quotient to W1 gives 2 W1 such that 2 2 = 1 and
2 := R[2 ] R[w] (11.36)
for all w W1 \ {0}. By the same variational argument used for 1 , this implies that
E[2 , w] = 2 2 , w
(11.37)
for all w W1 .
To extend this formula to the full weak solution condition, note that 2 , 1
= 0
because 2 W1 . The fact that 1 satisfies (11.26) also gives
E[1 , 2 ] = 1 1 , 2
= 0. (11.38)
Now consider a general v H01 (). For c := v, 1

we have
v c1 , 1
= 0,
so that v c1 W1 . After setting w := v c1 , we can expand

E[2 , v] 2 2 , v
L 2 = c E[2 , 1 ] 2 2 , 1
+ E[2 , w] 2 2 , w
The first line on the right is zero by (11.38) and the second is zero by (11.37). Thus
(11.26) is satisfied for all v H01 (), showing that 2 = 2 2 in the weak
sense. By taking the real or imaginary part we can assume that 2 is real-valued.
Subsequent eigenvalues are obtained by repeating this process inductively. After
k eigenfunctions have been found, we set
Wk := {1 , . . . , k } H01 (), (11.39)
and minimize the Rayleigh quotient over Wk to find k+1 and k+1 . Note that Wk
always contains nonzero vectors, because H01 () is infinite-dimensional. The reg-
ularity argument from the end of the proof of Theorem 11.9 applies to any solution
of (11.26), so that k C () for each k.
This process produces an orthonormal sequence of eigenfunctions {k } with eigen-
values satisfying
1 2 3 . . . .
To see that k , note that k 2 = 1 and R[k ] = k by construction, so that
k 2H 1 = 1 + k . (11.40)
Suppose the sequence {k } is bounded. Then (11.40) shows that the sequence {k } is
bounded with respect to the H 1 norm. Theorem 11.8 then implies that a subsequence
of {k } converges in L 2 (). But the k are orthonormal with respect to the L 2 norm,
so that
k m 2 = 2
for all k = m. Convergence of a subsequence of {k } is therefore impossible in

L 2 (). This contradiction shows that {k } cannot be bounded. Since the sequence
is increasing, this implies
lim k = .
k
The final claim is that {k } forms an orthonormal basis of L 2 (). After obtaining
the full sequence from the inductive procedure, let us set
W := {1 , 2 , . . . } H01 ().
Suppose that W contains a nonzero vector. Applying the Rayleigh quotient mini-
mization as above produces yet another eigenvalue . Since the k s were constructed
by minimizing the Rayleigh quotient on subspaces Wk W , this new eigenvalue
satisfies k for all k N. This is impossible because k . Hence W = {0}.
In other words, the only vector in H01 () that is orthogonal to all of the k is

0. Since Ccpt () H01 () and Ccpt () is dense in L 2 () by Theorem 7.5, this
implies that the only vector in L 2 () that is orthogonal to all of the k is 0. Hence
{k } is a basis by Theorem 7.10.
11.6 Sequential Compactness
In this section we take up the proof of Rellichs theorem (Theorem 11.8). Results
of this type, that force convergence of an approximating sequence, are a crucial
component of variational strategies for PDE.
In a normed vector space, a subset A is said to be sequentially compact if every
sequence within A contains a subsequence converging to a limit in A. A fundamental
result in analysis called the Bolzano-Weierstrass theorem (Theorem A.1) says that in
Rn this is equivalent to the definition of compact given in Sect. 2.3. That is, a subset
of Rn is sequentially compact if and only if it is closed and bounded.
Rellichs theorem could be paraphrased as the statement that a closed and bounded
subset of H01 () is sequentially compact as a subset of L 2 (), provided we are
careful about the two different norms referenced in this statement. Our strategy will
be to reduce Rellichs theorem to an application of Bolzano-Weierstrass using Fourier
series. We start with the periodic case.

Theorem 11.10 Suppose that v j is a sequence in H 1 (Tn ) that satisfies a uniform
bound
v j H 1 M (11.41)
for all j N. Then {v j } has a subsequence that converges in L 2 (Tn ).
Proof The argument is essentially the in any dimension, so let us take n = 1 to

same
simplify the notation. Suppose that v j is a sequence in H 1 (T) satisfying (11.41).
The periodic Fourier coefficients are defined by

1
ck [v j ] := v j (x)eikx d x (11.42)
2
11.6 Sequential Compactness 225
for k Z, with corresponding partial sums

m
Sm [v j ] = ck [v j ]eikx .
k=m
Applying the Cauchy-Schwarz inequality (Theorem 7.1) to (11.42) gives

ck [v j ] 1 v j 2 .
2
The assumption (11.41) implies also that the L 2 norms v j 2 are bounded by M, so
that

ck [ f j ] M
2
for all j N and k Z.

We start the process of finding a convergent subsequence by making the first
few coefficients converge. The collection of points (c1 [v j ], c0 [v j ], c1 [v j ]) forms
a bounded sequence in C3 . Applying Bolzano-Weierstrass (Theorem A.1) to this
sequence gives a point (a1 , a0 , a1 ) C3 and a subsequence

v (1)
j vj ,
such that
lim ck v (1)
j = ak
j
for k = 1, 0, 1. Since the partial sum S1 involves only these three coefficients, this
implies the uniform convergence

1
lim S1 v (1)
j = ak eikx ,
j
k=1
which also gives convergence in the L 2 sense.

The same reasoning can be applied to the Fourier coefficients with k = 2, . . . , 2
to obtain a subsequence
v (2)
j v (1)
j ,
such that
lim ck v (2)
j = ak
j
for k = 2, . . . ,2. This

process can be continued inductively to produce a family
(l)
of subsequences v j such that
l
lim Sl v (l)
j = ak eikx (11.43)
j
k=l
in L 2 (T).
To complete the proof, set
( j)
w j := v j .
Because w j is an element of the lth subsequence for l j, we deduce from (11.43)

that
m
lim Sm w j = ak eikx (11.44)
j
k=m
in L 2 (T) for all m N.

We now claim that the sequence w j converges in L 2 (T). In order to deduce this
from (11.44) we need to control the rate at which Sm [w j ] converges to w j as m .
This is where the H 1 bound (11.41) becomes crucial. In terms of Fourier coefficients,
the H 1 norm could be written

f 2H 1 = (1 + k 2 ) |ck [ f ]|2 .
k=
Hence (11.41) implies that

2
(1 + k 2 ) ck [w j ] M 2
k=
for all j. This leads to an estimate:

w j Sm [w j ]22 = ck [w j ]2
|k|>m
1 + k2
ck [w j ]2 (11.45)
|k|>m
1+m 2
M2
,
1 + m2
independent of j.
By the triangle inequality,
wi w j 2 wi Sm [wi ]2 + Sm [wi ] Sm [w j ]2 + w j Sm [w j ]2 .
Given > 0, fix m large enough that

11.6 Sequential Compactness 227
M
< .
1 + m2
By (11.45) this reduces our triangle estimate to
wi w j 2 < Sm [wi ] Sm [w j ]2 + 2. (11.46)
Since Sm [w j ] converges in L 2 (T) by (11.44), we can choose N sufficiently large so

that
Sm [wi ] Sm [w j ] L 2 <
for all i, j N . By (11.46) this implies that
wi w j 2 < 3

for i. j N . This shows that the subsequence w j is Cauchy with respect to the L 2
norm. By completeness the subsequence converges in L 2 (T).
Proof of Theorem 11.8 We can assume (, )n , after rescaling if needed. By

Lemma 10.10, elements of H01 () can be extended by zero to H 1 ([, ]n ). We
can then make these functions periodic to give the inclusion
H01 () H 1 (Tn ). (11.47)

Given a sequence v j H01 () that is uniformly bounded in the H 1 norm,applying

Theorem 11.10 to the extension given by (11.47) gives a subsequence w j that
converges in L 2 (Tn ). By construction the restriction of w j to [, ]n vanishes
outside , so this also gives a convergent subsequence in L 2 ().
11.7 Estimation of Eigenvalues
As we saw in Sect. 11.5, the Rayleigh principle for the eigenvalue problem is very
useful as a theoretical tool. It also leads to some very practical applications in terms
of estimating eigenvalues or calculating them numerically.
The basic strategy is to exploit the formula for eigenvalues that appeared in the
proof of Theorem 11.7. Assuming the Dirichlet eigenvalues {k } of are written in
increasing order and repeated according to multiplicity,
k = min R[w], (11.48)

wWk1 \{0}
where
Wk1 := {1 , . . . , k1 } H01 (),
with k the eigenfunction corresponding to k The only problem with this formula
is that determining the kth eigenfunction requires knowledge of the first k 1 eigen-
functions. This issue is resolved by the following:
Theorem 11.11 (Minimax principle) For a bounded domain Rn , let k denote
the set of all k-dimensional subspaces of H01 (). Let {k } denote the sequence of
Dirichlet eigenvalues of in increasing order. Then

k = min max R[u] (11.49)
V k uV \{0}
for each k N.
Proof Let {k } H01 () denote the eigenfunction basis. By the weak solution
condition (11.25), orthonomality in L 2 () implies also that

i , i = j,
E[i , j ] = (11.50)
0, i = j.
Let us set
Veig = [1 , . . . , k ] H01 (),
where [. . . ] denotes the linear span of a collection of vectors. For u Veig , expanded
as
k
u= c j k ,
j=1
we see from (11.50) that

k
2
E[u] = j c j .
j=1
The Rayleigh quotient is thus

k 2

j=1 j c j
R[u] = k 2 .
j=1 c j
Since 1 k by assumption, it is clear that
R[u] k
for u Veig \ {0}. Moreover, R[k ] = k , so that
max R[u] = k . (11.51)

uVeig \{0}
11.7 Estimation of Eigenvalues 229
Now consider a general subspace V k . Since dim[1 , . . . , k1 ] = k 1,

there exists a nonzero vector
w V [1 , . . . , k1 ] .
Since w [1 , . . . , k1 ] , it follows from (11.48) that
k R[w].
Hence,
k max R[u] (11.52)
uV \{0}
for V k .
In combination, (11.51) and (11.52) show that the minimum on the right-hand
side of (11.49) exists and is equal to k .
As a sample application, we can use the minimax principle to compare eigenvalues

of nested domains. This is possible because of the inclusion

H01 () H01 (), (11.53)
provided by Lemma 10.10.
Corollary 11.12 Consider two bounded domains in Rn satisfying . Assum-

ing the Dirichlet eigenvalue sequences are arranged in increasing order,
k ()
k ()
for all k N.
Proof Let {k } H01 () be the sequence of eigenfunctions of . By (11.53),

If
we can consider [1 , . . . , k ] to be a subspace of H01 (). k denotes the set of
k-dimensional subspaces of H01 (), then this implies

min max R[u] max R[u]
k
W uW \{0} u[1 ,...,k ]\{0}
while the right-hand side equals

By Theorem 11.11, the left-hand side is k (),
k () by (11.51).
Another way to make use the Rayleigh and minimax principles is to approxi-
mate eigenvalues by restricting subspaces of computationally simple functions within
H01 (). This approach can give surprisingly good estimates even when the subspaces
are small.
Example 11.13 On the unit disk D = {r < 1} R2 , consider the family of functions
w (r ) := 1 r
for > 0. Because w is radial, the energy can be computed by

w 21
E[w ] = 2r dr
0 r
1
= 22 r 21 dr
0
= .
The L 2 norm is 1
w 22 = (1 r )2 2r dr
0
2
= .
2 + 3 + 2
Hence the Rayleigh quotient gives the bound
2
1 R[w ] = +3+

for > 0. The optimal choice is = 2, which gives
.
1 3 + 2 2 = 5.828. (11.54)
Compare this to the exact value computed in Example 5.5 in terms of zeros of the
Bessel J -function,
.
1 = j0,1
2
= 5.783. (11.55)
The optimal choice of w gives a reasonable approximation to the true eigenfunction,

as Fig. 11.2 demonstrates.
Fig. 11.2 The Bessel

function for the first
J0 (j0,1 r)
eigenfunction and w2 2
1r
We can estimate higher eigenvalues and improve the accuracy by using a larger
subspace. This computational strategy was introduced by Walter Ritz in 1909, and
is referred to as the Rayleigh-Ritz method. Given a finite dimensional subspace A
H01 (), we let k (A) denote the k-dimensional subspaces of A. The approximate
eigenvalues associated to A are then given by

k := min max R[u] , (11.56)
W k (A) uW \{0}
for k = 1, . . . , dim A.
Since A is finite dimensional, the calculation of (11.56) can be recast as a matrix
eigenvalue problem. By the same arguments used in the proof of Theorem 11.7,
the values k are associated to vectors vk A satisfying the approximate weak
eigenvalue equation,
E[vk , w] = k vk , w
(11.57)
for all w A.
To interpret (11.57) as a matrix eigenvalue equation we fix a basis {w j }mj=1 for A.
In terms of this basis, the energy functional and L 2 inner product define matrices

E i j := E[wi , w j ], Fi j := wi , w j . (11.58)
If vk is expanded as

m
vk = cjwj,
j=1
then (11.57) is equivalent to

m
ci (E i j k Fi j ) = 0 (11.59)
i=1
for j = 1, . . . , m. This equation has a nontrivial solution only if the rows of the
matrix E k F are linearly dependent, which is equivalent to the vanishing of the
determinant. The values k can thus be calculated as the roots of a polynomial
{1 , . . . , k } = { : det(E F) = 0} . (11.60)
In other words, the j are the eigenvalues of E F 1 .
Example 11.14 Consider D as in Example 11.13, but now take the subspace A =
[w1 , w2 , w3 ], where
w j (r ) := 1 r 2 j .
Straightforward computations give the matrices (11.58) as

8
5 9

2 3
3 3 12 20
5 7
E = 8
3
4 24
5
, F= 12 8
15 12
.
24 9 7 9
3 5 6 20 12 4
The roots of det(E F) are

. . .
1 = 5.783, 2 = 30.712, 3 = 113.505.
The estimate 1 matches the exact value (11.55) very closely; in fact,
|1 1 | 106 .
The second value 2 is a reasonable approximation to

.
3 = j0,2
2
= 30.471.
However, we missed the second eigenvalue

.
2 = j1,1
2
= 14.682.
The problem is that the space A consists entirely of radial functions, so that the
second eigenfunction, &
2 (r, ) = J1 ( 2 r )ei ,
is orthogonal to A.
The missing eigenvalue in Example 11.14 illustrates a potential flaw in the

Rayleigh-Ritz scheme. We need to make sure the subspace A covers H01 () suf-
ficiently well in order to catch all low-lying eigenvalues. At the same time, we also
need a means of producing this subspace efficiently. The finite element method is an
approach that addresses both of these concerns.
To set up the finite element method we subdivide into small polygonal domains,
producing a mesh (or triangulation). Figure 11.3 illustrates a mesh for a two-
dimensional domain. To each interior vertex is associated a piecewise linear function
called an element that is positive at the vertex and decays linearly to zero on the
neighboring faces. The span of these elements defines a subspace A with dimension
equal to the number of interior vertices.
Example 11.15 Let us consider the domain (0, ), for which the exact Dirichlet
eigenfunctions are k (x) = sin kx for k N, with k = k 2 . Define a mesh by
subdividing (0, ) into m + 1 intervals of length /(m + 1). To the jth vertex we
associate the element
Fig. 11.3 Discrete mesh for

an oval domain in R2
Fig. 11.4 Piecewise linear

elements for (0, ) with w1 w2 w3
m=3

x m+1
j1
, j1
x j
,
m+1 m+1
w j (x) = x + m+1 , m+1 x
j+1 j j+1
,

m+1
0, otherwise,
for j = 1, . . . m. These elements are illustrated in Fig. 11.4.

The energy and inner product matrices can be computed by straightforward inte-
grals,

2 3
= 3(m+1)3 i = j,

2
m+1
i j,
3
Ei j = m+1 |i j| = 1, Fi j = |i j| = 1,

6(m+1)3
0 otherwise, 0 otherwise.
Table 11.1 shows the resulting approximate eigenvalues as a function of the number
of elements m.
The finite element method produces approximate eigenfunctions as well as eigen-

values. Once the eigenvalue k is determined, the coefficients c j can be computed
1 T
from m represent the eigenvectors of (E F ) in the basis
(11.59). These coefficients
w j . The function vk = j=1 c j w j is best approximation within the subspace A
to the true eigenfunction k . Figure 11.5 shows some approximate eigenfunctions
determined by the calculations in Example 11.15.
The Rayleigh-Ritz strategy proves to be quite adaptable to more complicated
problems. The procedure consists of: (1) generating a mesh for a given domain,
(2) computing the matrix entries E i j and Fi j corresponding to the elements of the
Table 11.1 Approximate Dirichlet eigenvalues for (0, ) computed using m elements
m 3 10 25 50 exact
1 1.05 1.01 1.00 1.00 1
2 4.86 4.11 4.02 4.01 4
3 12. 84 9.56 9.10 9.03 9
4 17.80 16.31 16.08 16
5 29.45 25.77 25.20 25
m=10 m=20 m=30
Fig. 11.5 Piecewise linear approximations to the eigenfunction 3 (x) = cos 3x
Fig. 11.6 A approximation

of the eigenfunction 6 for a
star-shaped domain
mesh, and (3) solving a finite-dimensional eigenvalue problem. A two-dimensional

example is shown in Fig. 11.6.
11.8 Euler-Lagrange Equations
In the mid-18th century, Lagrange and Euler jointly developed a framework for
expressing problems in classical mechanics in terms of the minimization of an action
functional. Euler coined the term calculus of variations to describe this approach,
which proved adaptable to a great variety of problems.
In the original classical mechanics setting, the action functional was the integral
of a Lagrangian function, defined as kinetic energy minus potential energy. These
might be energies of a single particle or a system.
11.8 Euler-Lagrange Equations 235
In a typical PDE application on a bounded domain Rn , we take the

Lagrangian L( p, w, x) to be a smooth function
L : Rn R R.
The action functional is defined by

S[w] := L(w(x), w(x), x) d n x

for w C 1 (; R). Suppose that u is a critical point of S, in the sense that

d
S[u + t] = 0
dt t=0

for every Ccpt (). This implies a (possibly nonlinear) PDE for u, called the
Euler-Lagrange equation of L.
The Dirichlet principle gives the most basic example of this setup. For
| p|2
L( p, w, x) := (11.61)
2
the action functional is the Dirichlet energy E[w], and the Euler-Lagrange equation is
the Laplace equation. To formulate the Poisson equation, we modify the Lagrangian
to include the forcing term f L 2 (; R),
| p|2
L( p, w, x) := f w. (11.62)
2
In this case the action is the functional D f [w] defined in (11.6).
A classic nonlinear example is the surface area minimization problem. For
R2 , the graph of a function w : R defines a surface in R3 . According to (2.8),
the surface area of this patch of surface is given by
+
A[w] := 1 + |w|2 d 2 x.

We can interpret this as an action functional corresponding to the Lagrangian

+
L( p, w, x) := 1 + | p|2 .
For f : R, the problem of minimizing A[w] under the constraint w| = f

was first studied by Lagrange. But historically this is called the Plateau problem
after the 19th century physicist Joseph Plateau who conduct experiments on minimal
surfaces using soap films.
Let us work out the Euler-Lagrange equation for the surface area functional. For

Ccpt (), we have
+
d d
A[u + t] = 1 + |u + t|2 d 2 x
dt t=0 dt t=0

u
= & d 2 x.
1 + |u|2
By Greens first identity (Theorem 2.10), and the fact that vanishes near the bound-
ary, , -

d u
A[u + t] = & d 2 x.
dt t=0 1 + |u|2
Setting this equal to zero for all gives the Euler-Lagrange equation
, -
u
& = 0. (11.63)
1 + |u|2
This nonlinear PDE is called the minimal surface equation.

There is a well-developed existence and regularity theory for general Euler-
Lagrange equations, but this is too technical for us to go into here. In many cases, the
finite element method can be used to effectively reduce the numerical approximation
of solutions to linear algebra. Figure 11.7 shows a solution of the minimal surface
equation calculated using finite elements.
Fig. 11.7 The minimal

surface over the unit disk
associated to the boundary
function f () = cos 6
11.9 Exercises 237
11.9 Exercises
11.1 For a finite-dimensional subspace A H01 (), suppose we approximate the

solution of the Poisson equation (11.4) for f L 2 () by setting
u = min D f [w].
wA
m
Given a basis w1 , . . . wm for A (not necessarily orthonormal), set u = i=1 ci wi .
Find equation an equation for (c1 , . . . , cm ) in terms of f and the matrices E and F
defined in (11.58).
11.2 To demonstrate the role that ellipticity plays in Theorem 11.5, consider the
operator L = r 2 on the unit disk D = {r < 1} R2 , where r := |x|2 . For
f L 2 (D), a weak solution of the equation Lu = f with Dirichlet boundary
conditions is defined as a function u H01 (D) satisfying

u (r 2 ) + f d 2 x = 0 (11.64)
D

for all Ccpt (D).
(a) Compute the weak partial derivatives of the function log r and show that log r
H01 (D).
(b) Show that u(x) = log r satisfies (11.64) with f = 0. Note that even though D
and f are C , the solution u is not even locally H 2 . (It is not a coincidence that
the singularity of u occurs at the point where ellipticity of the operator fails.)

11.3 Let R2 be the equilateral triangle with vertices (0, 0), (2, 0), and (1, 3).
Define w H01 () to be the piecewise linear function
whose graph forms a tetrahe-
dron over , with the top vertex at (1, 1/ 3, 1/ 3). Approximate the first eigen-
value by computing the Rayleigh quotient R[w]. (For comparison, the exact value
2
is 1 = 43 ; this corresponds to the eigenfunction shown in Fig. 11.1.)
11.4 Let B3 denote the unit ball {r < 1} R3 . Find an upper bound on the first
eigenvalue by computing the Rayleigh quotient of the radial function w(r ) = 1 r .
(Compare your answer to the exact value 1 = 2 found in Exercise 5.8.)
11.5 For a bounded domain Rn , let 1 H01 (; R) C () be the first

eigenfunction as obtained in Theorem 11.9, normalized so that 1 2 = 1. In this
problem we will show that 1 has no zeros in and that this eigenfunction is unique
up to a multiplicative constant.
(a) Set
(x) := max {1 (x), 0}
for x , and show that

1 = +
with 0 and + 0.
(b) Note Lemma 10.10 implies that H01 (; R). Assuming that 1 is normal-
ized 1 2 = 1, show that
+ 2 + 2 = 1,
E[+ ] + E[ ] = 1 .
(c) Since 1 minimizes the Rayleigh quotient,
E[ ] 1 2 .
Use this, together with the fact that E[1 ] = 1 , to deduce that
E[ ] = 1 2 .
Hence, by the proof of Theorem 11.9, C () and
= 1 . (11.65)
(d) Use the strong maximum principle of Theorem 9.5 to deduce from (11.65) that
if has a zero within then 0. Conclude that 1 has no zeros in .
(e) If u H01 (; R) C () is some other eigenfunction with eigenvalue 1 ,
then u c1 is also an eigenfunction for each c R. Show that c can be chosen
so that u c1 has a zero in . Conclude that u c1 .
11.6 Determine the Euler-Lagrange equations on Rn corresponding to the

following Lagrangians:
1 1
(a) L( p, w, x) = | p|2 + a(x)w 2 for a C 0 ().
2 2
1
n
(b) L( p, w, x) = ai j (x) pi p j with ai j C 1 ().
2 i, j=1
(c) L( p, w, x) = | p|.
1
(d) L( p, w, x) = | p|2 + F(w) for F C 1 (R).
2
Chapter 12
Distributions
To define weak derivatives in Chap. 10, we measured the values of a function f

L 1loc () by integrating against test functions. One way to interpret this process is

that f defines a functional Ccpt () C given by

f d n x.

A distribution on Rn is a more general functional Ccpt () C, not necessarily
expressible as an integral. To qualify as a distribution, a functional is required to
satisfy conditions that insure that weak derivatives and other basic operations are
well defined.
As with weak derivatives, the concept of a distribution was inspired by idealized
situations in physics. Indeed, the term distribution was inspired by charge distrib-
utions in electrostatics, an example that we will discuss in Sect. 12.1. Distributions
generalize the notion of weak solutions, in the sense that every function in L 1loc ()
also defines a distribution. The trade-off for the increased generality is that some
basic operations for functions cannot be applied to distributions. The product of two
distributions is not generally well defined, for example.
There are some technicalities in the mathematical theory of distributions that
require more background on the topology of function spaces than we assume for this
text. We will treat these technicalities rather lightly; our focus will be on exploring
the PDE applications.
12.1 Model Problem: Coulombs Law
Coulombs law of electrostatics is an empirical observation developed by 18th century

physicist Charles-Augustin de Coulomb. It says that a particle with electric charge
q0 , located at the origin, generates an electric field given by

240 12 Distributions
kq0 x
E(x) = , (12.1)
|x|3
where k (Coulombs constant) depends on the properties of the medium surrounding

the charges.
In Sect. 11.1 we discussed another important empirical law of electrostatics,
Gausss law. With the same convention for physical constants as in (12.1), the dif-
ferential form of the law says that
E = 4k, (12.2)
where is the charge per unit volume as a function of position.

These two empirical laws present something of a mathematical conundrum, in
that the field specified by Coulomb is not differentiable at x = 0, not even weakly.
On the other hand, for x = 0,
x x 3x
= 3 4 r
r3 r r
3 3x x (12.3)
= 3 4
r r r
=0
This is consistent with (12.2), in that Coulomb assumes the charge density is zero
for x = 0. However, if a function in L 1loc vanishes except at a single point, then
that function is zero by the equivalence (7.6). Thus a point charge density has no
meaningful interpretation as a locally integrable function.
To reconcile (12.1) with Gausss law, let us consider the weak form of (12.2),

E d 3 x = 4k d 3 x (12.4)
R3 R3

for all Ccpt (R3 ). The left side of (12.4) is well defined because the components
of E are locally integrable.
Since the Coulomb field is smooth away from the origin, we can integrate by parts
as long as we exclude the origin from the region of integration by writing the integral
as a limit,
E d x = lim
3
E d 3 x. (12.5)
R3 0 {r }
The region {r } has boundary given by the sphere {r = }. In this case the
outward unit normal is a radial unit vector pointing towards the origin,
x
= . (12.6)
r r =
By the divergence theorem (Theorem 2.6), and the fact that E = 0 for r > 0 by
(12.3),
12.1 Model Problem: Coulombs Law 241

E d 3 x = E d S.
{r } {r =}
Hence by (12.1) and (12.6),

kq0
E d x = 3
dS
{r } {r =} 2
Taking 0 now gives

kq0
E d 3 x = lim d S. (12.7)
R3 0 {r =} 2
Because is continuous, the average of over the sphere {r = } approaches (0)

as 0, i.e.,
1
lim d S = (0).
0 42 {r =}
Applying this to (12.7) gives

kq x
d 3 x = 4kq0 (0). (12.8)
R3 r3
The weak condition (12.4) thus requires that

d 3 x = q0 (0),
R3

for every Ccpt (R3 ). This is consistent with the physical interpretation of as a
charge located exactly at the origin.
The concept of a point density was widely used in physics applications in the
18th and 19th centuries. In a 1930 book on quantum mechanics, the physicist Paul
Dirac described such densities in terms of a delta function (x), whose defining
property is that
f (x)(x) d n x := f (0), (12.9)
Rn
for a continuous function f . This terminology and notation are potentially mislead-
ing, because is not a function and (12.9) is not actually an integral. However, Diracs
formulation hints at the proper mathematical interpretation, which is that should
be understood as a functional f f (0).
If we accept the intuitive definition of the delta function for the moment, then we
can interpret the calculation (12.8) as showing that
x
= 4. (12.10)
r3
12.2 The Space of Distributions

A distribution on a domain Rn is a continuous linear functional Ccpt () C.
The map defined by a distribution u is usually written as a pairing of u with a test
function, i.e.,
(u, ) C (12.11)

for Ccpt (). Linearity means that
(u, c1 1 + c2 2 ) = c1 ( f, 1 ) + c2 ( f, 2 ),

for all c1 , c2 C and 1 , 2 Ccpt ().
The definition of distribution also includes the word continuous. To define con-

tinuity for functionals we must first specify what convergence means in Ccpt (). The

standard definition is that for a sequence {k } to converge to in Ccpt () means that
all k have support in some fixed compact set K , and the sequence of functions
and all sequences of partial derivatives converge uniformly on K . Continuity of the
functional (12.11) is then defined by the condition that convergence of a sequence

k in Ccpt () implies that
lim (u, k ) = (u, ). (12.12)

k
In finite dimensions continuity is implied by linearity. That is not the case here, but
in practice it is quite difficult to come up with a functional that is linear but not
continuous.
The set of distributions on forms a vector space denoted by D
(). Linear
combinations of distributions are defined in the obvious way by
(c1 u 1 + c2 u 2 , ) := c1 (u 1 , ) + c2 (u 2 , ),
for u 1 , u 2 D
() and c1 , c2 C. The mathematical theory of distributions was
developed independently in the mid-20th century by Sergei Sobolev and Laurent

Schwartz. Schwartz used D as a notation for Ccpt , and the prime accent on D
comes
from the notation for the dual of a vector space in linear algebra.
A locally integrable function f L 1loc () defines a distribution through the
integral pairing
( f, ) := f d n x. (12.13)

Under this convention there is an inclusion
L 1loc () D
().
In particular, all L p functions can be interpreted as distributions.

12.2 The Space of Distributions 243
As we saw with the point charge density in Sect. 12.1, not all distributions are
given by functions. We use the notation x for the delta function centered at x ,
defined by
( x , ) := (x). (12.14)
By convention the subscript is dropped for x = 0, i.e., := 0 .

Multiplication by smooth functions preserves the space Ccpt (). Therefore it
makes sense to multiply a distribution u D () by a function f C (). The

product distribution is defined by
( f u, ) := (u, f ).
It does not make sense, however, to multiply two distributions together. This fact was
intuitively clear in early applications: the product of two charge densities makes no
physical sense.
Convergence of a sequence of distributions is defined in a very straightforward
way. We say that u k u in D
() if
lim (u k , ) = (u, )
k

for all Ccpt (). All distributions can in fact be approximated by smooth functions
by such a limit, although we are not equipped to prove that here. We will present
one useful special case, a construction of the delta function as a limit of integrable
functions.
Lemma 12.1 Given f L 1 (Rn ) satisfying

f d n x = 1, (12.15)
Rn
define the rescaled function,

f a (x) := a n f (ax)
for a > 0. Then

lim f a = , (12.16)
a
as a distributional limit.

Proof For Ccpt (Rn ) we can evaluate the pairing with fa using a change variables,

( f a , ) = a n f (ax)(x) d n x
Rn

= f (x)(x/a) d n x.
Rn
By the assumption (12.15), we can also write

(0) = f (x)(0) d n x,
Rn
which gives the estimate

( f a , ) (0) | f (x)|(x/a) (0) d n x. (12.17)
Rn
Given > 0, the fact that f is integrable implies that exists R sufficiently large
so that
| f | d n x < . (12.18)
|x|R
By the continuity of we can also choose > 0 so that
|(x) (0)| <
for |x| < . For a R/ this implies that

(x/a) (0) < (12.19)
for all |x| R. Using (12.18) and (12.19) to estimate the difference (12.17) gives

( f a , ) (0) 2 | f (x)| d n x + | f (x)| d n x
|x|R |x|R

2 + f 1 ,
for a R/. Since was arbitrary, this shows that
lim ( f a , ) = (0).
a
The rescaling used in Lemma 12.1 is illustrated in Fig. 12.1. Note that this looks
very similar to Fig. 9.1, and in fact the proof of Lemma 12.1 uses essentially the
same argument as that of Theorem 9.1. We saw another case of this construction in
the proof of Theorem 6.2. Indeed, we can now interpret the result of Theorem 6.2 as
a distributional limit of the heat kernel,
lim Ht = ,
t0
where Ht was defined in (6.16).

12.3 Distributional Derivatives 245
Fig. 12.1 Rescaled

f16
functions f a for
f (x) = 21 e|x|
f8
f4
f2
f1
12.3 Distributional Derivatives
The distributional derivative extends the concept of the weak derivative introduced
in Sect. 10.1. By analogy with (10.7), for u D
() and we define the distribution
D u by
(D u, ) := (1)|| (u, D ), (12.20)
with
1 n
D := 1
x1 xnn
and || := 1 + +n , as before. The pairing (12.20) is well defined as a distribution

because D is both linear and continuous as a map Ccpt
() Ccpt
().
The terms distributional and weak are frequently used interchangeably to
describe derivatives, since the definitions overlap to a considerable extent. The only
difference is that a weak derivative is representable as a locally integrable function.
Weak derivatives may not exist, whereas all distributions are infinitely differentiable.
Example 12.2 Let us reconsider Example 10.3, where we considered the derivative
of w L 1loc () defined by

w (t), t < 0,
w(t) =
w+ (t), t 0,
where w C 1 (R). As part of that calculation we showed that

w
dt = w+ (0) w (0) (0) + h dt, (12.21)

where h is the piecewise derivative


w (t), t < 0,
h(t) :=
w+ (t), t > 0.
The left-hand side of (12.21) is the pairing (w

, ) by the definition (12.20). From
the right-hand side we can thus see that the distributional derivative is

w
= h + w+ (0) w (0) .
Example 12.3 For x D

(Rn ), the derivatives D x are easily computed from the

definition (12.20). For Ccpt (Rn ),
(D x , ) = (1)|| ( x , D )
= (1)|| D (x).
In other words, the distribution D x evaluates the derivative of the test function at
the point x, up to a sign.
Example 12.4 The function ln |x| is locally integrable on R and so defines a distri-
bution in D
(R). Therefore (ln |x|)
exists in the distribution sense. This is puzzling
because
d 1
ln |x| =
dx x
for x = 0, and x 1 is not locally integrable.

To understand what is happening here, we must return to the distributional defin-
ition,
(ln |x|)
, := (ln |x|,
)

=
(x) ln |x| d x

for Ccpt (R). To compute this we avoid the singularity at 0 by writing

(ln |x|)
, = lim
(x) ln |x| d x. (12.22)
0 |x|
Integration by parts gives

(x)

(x) ln |x| d x = (x) ln |x| + dx
x

(x)
= () ln + d x,
x
12.3 Distributional Derivatives 247
and similarly

(x)
(x) ln |x| d x = () ln + d x.
x
After combining these two halves, we obtain

(x)
(x) ln |x| d x = () () ln + d x.
|x| |x| x
By the definition of the derivative,
() ()
lim =
(0).
0 2
Therefore
lim () () ln = 2
(0) lim ln
0 0
= 0.
Hence (12.22) reduces to

(x)
(ln |x|)
, = lim d x. (12.23)
0 |x| x

The limit on the right exists for Ccpt (R), even though x 1 is not integrable,
because the limit is taken symmetrically. This limiting procedure defines a distribu-
tion called the principal value of x 1 , written as PV[x 1 ]. We could rephrase (12.23)
as
d
ln |x| = PV x 1 .
dx
Example 12.5 Let us reinterpret the discussion from Sect. 12.1 in terms of distrib-
utional derivatives. We already noted that the components of x/r 3 are locally inte-
grable, so we can consider the Coulomb formula (12.1) for E as the definition of a
vector-valued distribution. The distributional divergence of x/r 3 is defined by the
condition that
x
x
3 , := 3
d 3 x,
r R 3 r

for Ccpt (R3 ). The derivation of (12.8) thus shows that
x
= 4. (12.24)
r3
We can also consider the corresponding result for the Coulomb electric potential
1
(x) = .
r
(ignoring the physical constants). The gradient of exists in the weak sense and is
given by
1 x
= 3.
r r
Since = , we deduce from (12.24) that

1
= . (12.25)
4r
12.4 Fundamental Solutions
Because the Poisson equation is linear, it makes sense to construct a solution with
a continuous density by superimposing a field of point sources. With a change of
variables, we can see from (12.25) that the potential function corresponding to a
point source at y R3 is
1
y (x) := .
4|x y|
Weighting the point sources by the density and summing them with an integral
gives
1 (y) 3
u(x) = d y. (12.26)
4 R3 |x y|
This formula, which is often stated as the integral form of Coulombs law, does
indeed yield a solution of the Poisson equation on R3 under certain conditions. For
example if Ccpt
1
(R3 ) then one can confirm that u = by direct computation.
1
The C condition is stronger than necessary here, but continuity alone would not
be sufficient. (The precise notion of regularity needed for this problem is something
called Hlder continuity.)
This idea of constructing of general solutions by superposition of point sources is
the inspiration for the concept of a fundamental solution. For a constant-coefficient
differential operator L acting on Rn , of the form

L= a D ,
||m
12.4 Fundamental Solutions 249
with a C, a fundamental solution is a distribution D

(Rn ) such that
L = . (12.27)
For example, in the Coulomb case the calculation (12.25) gives the fundamental
solution of on R3 . Fundamental solutions are especially important for classical
problems involving the Laplacian.
The solution formula (12.26) resembles the convolution used to solve the heat
equation in Sect. 6.3. For f, g L 1 (Rn ) the convolution is defined as

f g(x) := f (y)g(x y) d n y.
Rn
A simple change of variables shows that this product is symmetric,
f g = g f.
In order to produce solution formulas from fundamental solutions, we need to under-

stand how to take convolutions with distributions.

For f, g L 1 (Rn ), the distributional pairing of f g with Ccpt () gives

( f g, ) = f (y)g(x y)(x) d n y d n x. (12.28)
Rn Rn
The x integration looks almost like the convolution of with g, except with the
argument switched from y x to x y. With the reflection defined by
g (x) := g(x),
we have

g (y) = g(x y)(x) d n x.
Rn
Thus (12.28) reduces to

( f g, ) := ( f, g ).

If , Ccpt (Rn ), then it is easy to check that Ccpt

(Rn ) also. Moreover,
the map is linear and continuous. We can thus define u for u D
(Rn )

and Ccpt (Rn ) by
(u , ) := (u, ) (12.29)

for Ccpt (Rn ).
The distribution plays a special role with regard to convolutions. By the definition
(12.29),
( , ) := (, )
= (0)

= (x)(x) d n x

= (, ).
This shows that

= . (12.30)
In other words, convolution by is the identity map.

Let be the fundamental solution for the constant coefficient operator L. Our
goal is to show that the equation Lu = f is solved by the convolution u = f , at

least for f Ccpt (Rn ). To check this, we need to know how to evaluate derivatives
of the convolution.
Lemma 12.6 For w D
(Rn ) and Ccpt

(Rn ),
D (w f ) = (D w) = w (D f ).

Proof For , Ccpt (), we compute directly that

D ( )(x) = (x y)(y) d n y

= D (x y)(y) d n y

= (D ) (x).
Since the convolution is symmetric, the same formula holds with and switched.
Thus the formula
D ( ) = (D ) = (D ) (12.31)
holds for test functions.

For w D
(R) and Ccpt

(Rn ), it follows from the definitions that
(D ( f ), ) = (1)|| ( f, (D )).
By (12.31) this gives
(D ( f ), ) = (1)|| ( f, D ( ))
= (D f, )
= ((D f ) , ),
12.4 Fundamental Solutions 251
and also
(D ( f ), ) = (1)|| ( f, (D ) ))
= ( f (D ), ).
Theorem 12.7 If L is a constant coefficient operator on Rn with fundamental solu-

tion , then for f Ccpt (Rn ) the equation
Lu = f
is solved by
u = f.
Proof By Lemma 12.6,

L( f ) = a D ( f )
||m

= a (D ) f
||m
= (L) f.
Note that the second step only works because the coefficients a are assumed to be
constant. Since L = , we see from (12.30) that
L( f ) = f.
A result called the Malgrange-Ehrenpreis theorem, proven in the 1950s, says that
every constant coefficient differential operator on Rn admits a fundamental solution.
The fundamental solution of the Laplacian, which we will now work out for any
dimension, is the most important case.
Theorem 12.8 On Rn the operator has the fundamental solution

2
1
ln r, n = 2,
(x) = (12.32)
1
(n2)An r n2
, n 3,
where An denotes the volume of the unit sphere in dimension n.
Proof We start from the distributional derivative,
(, ) = (, )

for Ccpt (Rn ). To evaluate this, it is useful to first compute the gradient,
x
(x) = .
An r n
The function x/r n is locally integrable in Rn and has compactly support. Therefore
we can deduce from Greens first identity (Theorem 2.10) that

d x =
n
d n x
Rn Rn

1 x
= d n x
A n Rn r n

1 1 n
= d x.
A n Rn r n1 r
The integral can be evaluated using radial coordinates as in (2.10):

1
d n x = dr d S
Rn An Sn1 r
0
1
= (0) d S
An Sn1
= (0).
This shows that

(, ) = (0),
hence = .
12.5 Greens Functions
Although fundamental solutions are defined only for the domain Rn , one of their
principle applications is to boundary value problems on a bounded domain Rn .
The connection comes from a integral formula introduced in 1828 by George Green.
For this section, let denote the fundamental solution of the Laplacian on Rn ,
as given by (12.32). For y Rn we set
y (x) := (x y). (12.33)
Theorem 12.9 (Greens representation formula) Suppose that Rn is a bounded

domain with piecewise C 1 boundary. For u C 2 (),
12.5 Greens Functions 253

u y
u(y) = y u d n x + y u dS

for y .
Proof Because the point y is fixed, for notational convenience we can change
variables to assume y = 0. For > 0 set
B := B(0; ),
and assume that is small enough that B .

On B , is smooth and satisfies = 0. Therefore, applying Greens
second identity (Theorem 2.11) on this domain with v = gives

u u
u d n x = u dS + u d S.
B B
(12.34)
Because u is continuous and is locally integrable,

lim u d x =
n
u d n x,
0 B

To prove the representation formula we must therefore show that

u
lim u d S = u(0). (12.35)
0 B

r r
To handle the first term in (12.35), note that

u
|u|,
r
for r > 0. Therefore, since u is continuous by assumption, we have a bound

u
max M
B r
for > 0, with M independent of . Using the fact that vol( B ) = An n1 and the
formula (12.32) for , we can estimate

u ln , n = 2,
d S M
r
, n 3.
B n2
This shows that

u
lim d S = 0.
0 B

r
For the second term in (12.35), we use the fact that
1
= ,
r An r n1
for r > 0, to compute

1
u dS = u d S.
B r A n n1
B
The right-hand side is the average value of u over the sphere B . By continuity,

1
lim u d S = u(0).
0 vol( B ) B
This proves (12.35), and thus establishes the representation formula.
The representation formula of Theorem 12.9 has many applications. The original
goal that Green had in mind was a solution formula for the Poisson problem with
inhomogeneous Dirichlet boundary conditions, which we will now describe.
Suppose there exists a family of functions Hy C 2 (), for y , satisfying

Hy = 0, Hy = y . (12.36)
Then the Greens function of is
G y := y Hy . (12.37)
It is possible to show that Hy exists under general regularity conditions on , but

this is too technical for us to get into here. We will focus on cases where Hy can be
computed explicitly, which requires the geometry of to be very simple.
Theorem 12.10 Suppose Rn is a bounded domain with piecewise C 1 boundary

that admits a Greens function G y , Then the Poisson problem on ,

u = f, u = g,
for f C 0 (), g C 0 (), is solved by the function

G y
u(y) = f Gy dn x g d S.

Proof Setting v = Hy in Greens second identity (Theorem 2.11) gives

u
Hy u uHy d n y = Hy d S. (12.38)

The result then follows by subtracting (12.38) from the representation formula of
Theorem 12.9.
Example 12.11 The Greens function for the unit disk D R2 can be derived using
a trick from electrostatics called the method of images. This involves placing charges
outside the domain in order to solve the boundary value problem. For the unit disk,
in order to find Hy we consider a charge placed at the point y given by reflecting
y C\{0} across the unit circle, i.e.,
y
y := . (12.39)
|y|2
Note that y is harmonic on D because y

/ D.
For x D, let denote the angle from y to x, as shown in Fig. 12.2. By the law
of cosines on the triangle made by 0, y and x,
|x y|2 = 1 + |y|2 2|y| cos .
If y is replaced by y,
the corresponding formula is
|x y|
2 = 1 + |y|
2 2|y|
cos
= 1 + |y|2 2|y|1 cos .
Solving for cos in these expressions gives the relation
|x y|2 1 |y|2 2 1 |y|2

|x y|
= ,
2|y| 2|y|1
which simplifies to
|x y|
|x y|
= (12.40)
|y|
for x D.
Fig. 12.2 Geometry for the x

method of images on the unit
disk
y
y
0
Since (x) = 2
1
ln |x| in R2 , taking the logarithm of (12.40) gives
1
y (x) = y (x) + ln |y|
2
for y = 0 or x. Thus we can solve (12.36) for y = 0 by setting
1
Hy := y ln |y|.
2
For y = 0 the obvious solution is H0 := 0 because |D = 0.
The Greens function is thus

|xy|
2
1
ln |x y|
|y|
, y = 0
G y (x) =
2 ln |x|,
1
y = 0.
To apply this in the solution formula, we need the radial derivative of G. For y fixed
and r := |x| we compute

ln |x y| = x ln |x y|
r
(x y)
=x
|x y|2
1xy
= .
|x y|2
Applying the corresponding result for |x y|

and subtracting gives

G y 1 1xy 1 x y
(x) = ,
r 2 |x y|2 |x y|
2
which by (12.40) simplifies to
G y 1 |y|2
(x) = . (12.41)
r 2|x y|2
Conveniently, the calculation for y = 0 leads to the same expression.

In the case of the Laplace equation on D, Theorem 12.10 gives the formula for a
harmonic function u with boundary value g as

1 1 |y|2
u(y) = g(x) d S(x).
2 D |x y|2
This will look more familiar in polar coordinates. With y = (r cos , r sin ) and
x = (cos , sin ), the formula becomes

1 2
1 r2
u(r, ) = g(cos , sin ) d.
2 0 1 + r 2 2r cos( )
This is the classical Poisson formula (9.4) that we derived from Fourier series.
12.6 Time-Dependent Fundamental Solutions
To adapt the concept of a fundamental solution to evolution equations, we need

to consider time-dependent distributions on Rn . We will use a subscript to denote
the time dependence, to avoid confusion with the spatial variables. Thus a map
R D
(Rn ) will be written
t wt .

For Ccpt (Rn ) the pairing (wt , ) is a complex-valued function of t.
The function t wt is differentiable with respect to time if there exists a family
of distributions w
t
t
D
(Rn ) such that

d wt
(wt , ) = , , (12.42)
dt t

for all Ccpt (). Higher derivatives are defined in the same way.
Example 12.12 In R, consider the derivatives of t , the delta function supported at

t. By definition,
(t , ) := (t),
so that
n
t , = (n) (t).
t n
Compare this to the spatial derivatives, defined according to (12.20),

n
t , := (1)n t , (n)
x n
= (1)n (n) (t).
We conclude that
n n
n
t = (1) t . (12.43)
t n x n
Let us try to deduce the fundamental solution for the one-dimensional wave equa-
tion from dAlemberts formula (4.8),

1 1 x+t
u(t, x) = g(x + t) + g(x t) + h( ) d , (12.44)
2 2 xt
The second term in (12.44) could be interpreted as a convolution

x+t
h( ) d = [t,t] (x )h( ) d
xt
= [t,t] h(x),
where I denotes a characteristic function as in (7.5). Therefore it makes sense to

define this component of the fundamental solution as
1
Wt := (t,t) . (12.45)
2
The time derivatives of Wt are computed from the pairing
t
1
(Wt , ) = dx
2 t

for Ccpt (R). By the fundamental theorem of calculus,
d 1
(Wt , ) = (t) + (t) , (12.46)
dt 2
which shows that
Wt 1
= t + t . (12.47)
t 2
Differentiating again using (12.43) gives
2 Wt 1

= t + t . (12.48)
t 2 2
On the other hand, x-derivatives of Wt are defined by (12.20). In particular,

2 Wt
, := Wt ,
x 2
This can be evaluated by direct integration,

2 Wt 1 t
, := dx
x 2 2 t
1
=
(t)
(t) .
2
12.6 Time-Dependent Fundamental Solutions 259
This shows that

2 Wt 1

= t + t . (12.49)
x 2 2
By (12.48) and (12.49), Wt is a distributional solution of the wave equation,

2 2
Wt = 0.
t 2 x 2
In contrast to the definition (12.27) of a fundamental solution in the spatial case, Wt

satisfies a homogeneous equation. The delta function appears only in the boundary
conditions,
Wt
W0 = 0, = .
t t=0
The distribution Wt , which is analogous to a fundamental solution, is called the
wave kernel. By (12.47), the g component of the dAlembert solution formula (12.44)
could be written in terms of Wt as
1 Wt
g(x + t) + g(x t) = g(x).
2 t
Thus the full convolution formula for the solution reads
Wt
u(t, ) = g + Wt h.
t
12.7 Exercises
12.1 Define the distribution u D

(R) by

1
(x) (0) (x)
(u, ) := dx + d x.
1 x |x|1 x
Show that u = PV[x 1 ].
12.2 Let f L 1loc (R) be the function

log x, x > 0,
f (x) =
log(x), x < 0.
For x = 0, f
(x) = |x|1 , but this is not locally integrable. Show that the distribu-
tional derivative is

1
(x) (0) (x)
( f , ) = + d x.
1 |x| |x|1 |x|
12.3 Let H denote the upper half-plane {x2 > 0} R2 . The goal of this problem is
to show that the Laplace equation on H,
u = 0, u(, 0) = g,
has the solution

1 y2
u(y) = g(x) d x
(x y1 )2 + y22

for g Ccpt (R).
(a) Derive this formula from Theorem 12.10 using the method of images as in
Example 12.11. In this case the reflection of y H is given by (y1 , y2 ) =
(y1 , y2 ) (the complex conjugate).
(b) Show that the fact that u(, 0) = g could also be derived by using Lemma 12.1
to deduce that y
lim = (y).
x0 (x + y 2 )
2
12.4 In R3 show that

eikr
( k 2 ) =
4r
for all k R.
12.5 For n 3 let Bn denote the unit ball {r < 1} Rn .

(a) Apply the method of images as in Example 12.11 to derive the solution Hy of
(12.36) for y Bn . and compute the Greens function. (Note that the formulas
(12.39) and (12.40) remain valid in any dimension.)
(b) Show that the radial derivative of the Greens function satisfies
G y 1 |y 2 |
(x) = .
r An |x y|n1
(c) Find the resulting solution formula from Theorem 12.10, and show that this gen-
eralizes the mean value formula for harmonic functions obtained in Theorem 9.3.
Chapter 13
The Fourier Transform
For a bounded domain Rn , Theorem 11.7 shows that we can effectively

diagonalize the Laplacian by choosing an orthonormal basis for L 2 () consisting
of eigenfunctions. Such a result is not possible on Rn itself; the Laplacian has no
eigenfunctions in L 2 (Rn ).
The closest analog to eigenfunctions on Rn are the spatial components of the plane
wave solutions introduced in Exercise 4.8,
(x) := eix ,
associated to a frequency vector Rn . These functions satisfy a convenient dif-

ferentiation formula,
D = (i) ,
and in particular
eix = ||2 eix .
The appropriate generalization of the Fourier series to L 2 (Rn ) is an integral trans-

form based on these plane waves. Although the technical details are quite different
from Fourier series, the transform serves a similar purpose in that it exchanges the
roles of differentiation and multiplication.
13.1 Fourier Transform
The Fourier transform of a function in f L 1 (Rn ) is a function of the frequency

Rn defined by

f () := eix f (x) d n x. (13.1)
Rn

262 13 The Fourier Transform
Note that the integral is well defined by the integrability of f , and in fact

f () f L 1 (13.2)
for all Rn . As a map the transform is denoted by
F : f f.
To develop the properties of the Fourier transform, it proves convenient to intro-

duce a particular class of test functions, called Schwartz functions. The space S con-
sists of smooth functions which, along with all derivatives, decay rapidly at infinity.
The precise meaning of rapid is faster than any power of r . An alternate form of
this definition is

S(Rn ) := f C (Rn ); x D f < for all , , (13.3)
where is the sup norm introduced in Sect. 7.3.

A basic example of a Schwartz function is a Gaussian function of the form
f (x) = ea|x| ,
2
with a > 0. We also have

Ccpt (Rn ) S(Rn ),
because compactly supported functions obviously satisfy the decay requirement.

As in the discrete case, the Fourier transform interchanges the operations of dif-
ferentiation and multiplication in a convenient way. For this statement, we let D x
and D denote partial derivatives with respect to x or , respectively.
Lemma 13.1 For S(Rn ),

F[D x ]() = (i) (), (13.4)
and

F[x ]() = (i D ) (). (13.5)
Proof The first identity follows from integration by parts,

F[D ]() = eix D x (x) d n x
Rn

= (x)(i D ) (eix ) d n x
Rn

= (x)(i) eix d n x
Rn

= (i) ().
13.1 Fourier Transform 263
The second is also a direct computation,

F[x f ]() = eix x f (x) d n x
Rn

= (i D ) (eix ) f (x) d n x
Rn
= (i D ) f().
Pulling the differentiation outside the integral in the final step is justified by the
smoothness and decay assumptions on .
In Lemma 13.1 we can see Schwartzs motivation for the definition of S. Under
the Fourier transform, smoothness translates to rapid decay, and vice versa. These
properties are balanced in the definition of S, which leads to the following result.
Lemma 13.2 The Fourier transform F maps S(Rn ) S(Rn ).
Proof Suppose that f S. In order to show that f is Schwartz, we need to produce
a bound on the function D f for each , . By (13.4) and (13.5),

D f() = i ||+|| eix x D x f (x) d n x. (13.6)
Rn
To estimate, we set

M N ,, := (1 + |x|2 ) N x D x f ,

which is finite by the definition (13.3). Because (1 + |x|2 )N is integrable for N

sufficiently large, we can estimate (13.6) by

1
D f () M N ,, d n x.
Rn (1 + |x|2 ) N
The right-hand side is independent of , so this yields the required estimate.

Example 13.3 Consider the one-dimensional Gaussian function
(x) := eax
2
for a > 0. Note that satisfies the ODE
d
= 2ax.
dx
Taking the Fourier transform of both sides and applying Lemma 13.1 gives
d
i = 2ai ,
d
which reduces to
d
= .
d 2a
Separating variables and integrating yields the solution
2
/4a
()
= (0)e
.
To fix the constant, we can use (2.19) with n = 1 to compute

eax d x
2
(0)
=

1
ex d x
2
= (13.7)
a

= .
a
Thus,
2 /4a
()
= e .
a
The computation from Example 13.3 can be generalized to Rn by factoring the

integrals,

F ea|x| () = ei xa|x| d n x
2 2
Rn
n

ei x j j ax j d x j
2
=
j=1
n

2j /4a
= e .
j=1
a
Thus n2
F ea|x| () = e|| /4a
2 2
(13.8)
a
for a > 0.
For f, g S(Rn ), consider the integral

f (x)ei xy g(y) d n x d n y. (13.9)
Rn Rn
The integrals over x and y can be taken in either order, yielding the useful identity:

f g d n x = fg d n y (13.10)
Rn Rn
for f, g S(Rn ),
Theorem 13.4 The Fourier transform on S(Rn ) has an inverse F 1 given by

f (x) = (2) n
eix f() d n . (13.11)
Rn
Proof In (13.10) let us set g = eax for a > 0, By (13.8), this implies
2
n2
x 2 /4a
f(y)eay d n y.
2
f (x)e d x=
n
(13.12)
a Rn Rn
On the left-hand side we can use the same argument as in the proof of Lemma 12.1
to show that
n
2 x 2 /4a n

f ( ax)ex /4 d n x
n 2
lim f (x)e d x = lim
2
a0 a Rn a0 Rn
= (2)n f (0).
We claim that the corresponding limit on the right-hand side of (13.12) is

f(y)eay d n y = f(y) d n y.
2
lim (13.13)
a0 Rn Rn
Because the convergence is not uniform, we will check this carefully. The difference
of the two sides can be estimated by

f(y)eay d n y
2
n
f (y) d y f (y) 1 eay d n y.
2

Rn Rn Rn
Given > 0 we can choose R large enough that

n
f (y) d y < ,
|x|>R
since f is integrable. Splitting the integral at |x| = R gives the estimate

ay 2
f (y) 1 eay d n y
2
f (y) 1 e d y +
n
Rn B(0;R)

+ 1 ea R f1 .
2
The second term approaches zero as a 0, so that

f (y) 1 eay d n y < 2,
2
Rn
for a sufficiently small. This establishes (13.13).

By these calculations, the limit of (13.12) as a 0 yields

(2) f (0) =
n
f(x) d n x. (13.14)
Rn
This is a special case of the desired formula.

The general inverse formula can be deduced from (13.14) by a simple translation
argument. For w Rn , define the translation operator Tw on S(Rn ) by
Tw f (y) := f (y + w).
A change of variables shows that

Tw f (x) = ei xy f (y + w) d n y
Rn

= ei x(yw) f (y) d n y
Rn
= ei xw f(x).
Since Tw f (0) = f (w), plugging Tw f into (13.14) gives

(2)n f (w) = ei xw f(x) d n x.
Rn

The pairing formula (13.10) suggests that the L 2 inner product will behave natu-
rally under the Fourier transform. Indeed, by Theorem 13.4 we can compute

f()g()
dn = f (x)ei x d n x g()
dn
Rn Rn Rn

= ei x g()
dn f (x) d n x
Rn Rn

= (2) n
f (x)g(x) d n x,
Rn
for f, g S(Rn ). In other words,

f, g
= (2)n
f, g. (13.15)
The integral (13.1) defining the Fourier transform does not necessarily converge
for f L 2 (Rn ), but the identity (13.15) makes it possible to define transforms on
L 2 by taking limits.
Theorem 13.5 (Plancherels theorem) The Fourier transform extends from S(Rn )
to an invertible map on L 2 (Rn ), such that (13.15) holds for all f, g L 2 (Rn ).
Proof First note that Theorem 7.5 implies that S(Rn ) is dense in L 2 (Rn ) because
it includes the compactly supported smooth functions. Hence for f L 2 (Rn ) there
exists a sequence of Schwartz functions k f in L 2 . As a convergent sequence,
{k } is automatically Cauchy, i.e.,
lim k m 2 = 0.
k,m
By (13.15),

k m = (2)n/2 k m 2 ,
2
implying that { k } is also Cauchy in L 2 (Rn ). Since L 2 (Rn ) is complete by Theorem

7.7, this implies convergence, and we can then define
f := lim k ,
k
with the limit taken in the L 2 sense.

To show that (13.15) extends to L 2 , suppose for f, g L 2 that k f and m
g are approximating sequences of Schwartz functions. By the property (13.15),

k , m = (2)n
k , m .
Taking the limit k, m then gives

f, g
= (2)n
f, g.
The same argument can be used to show that f is independent of the choice of
approximating sequence.
13.2 Tempered Distributions
Since F maps S(Rn ) to itself, to extend the Fourier transform to distributions it is

natural replace Ccpt (Rn ) by S(Rn ) as the space of test functions. The result is the
space of tempered distributions

S (Rn ) := continuous linear functionals S(Rn ) C .
Here the word tempered refers to a restriction on the growth at infinity. Because the
Schwartz functions decay rapidly, a locally integrable function is essentially required
to have a polynomial growth rate at infinity in order to define an element of S (Rn ).
The definition of continuity of a functional on S(Rn ) depends on a notion of
convergence for Schwartz
functions. A sequence {k } S(Rn ) converges if the
sequences x D k converge uniformly for each , . To say that u S (Rn ) is

continuous means that (u, k ) (u, ) whenever k in S(Rn ).

The delta function x and its derivatives are clearly tempered distributions. We
claim also that
L p (Rn ) S (Rn ),
for p [1, ]. This follows fairly directly from the fact that S(Rn ) L p (Rn ) for
p [1, ].
The pairing formula (13.10) gives the prescription for extending F to the tempered
distributions. For u S (Rn ), we define u by
(u,
) := (u, ) (13.16)
for S(Rn ). To justify this definition one needs to check that the Fourier trans-
form is continuous as a map S(Rn ) S(Rn ). This essentially follows from the
calculations in the proof of Lemma 13.2.
As an example, consider the function u = 1 as an element of S (Rn ). For
S(Rn ),

) := (1, )
(1,

=
(x) d n x.
Rn
According to the inverse Fourier transform formula (13.11),

(x) d n x = (2)n (0).
Rn
Therefore
1 = (2)n . (13.17)
Physicists often express this fact by writing

(x) = (2)n ei x d n ,
Rn
with the understanding that the integral on the right is not to be taken literally.
The Fourier transform of is a similar calculation. For S,
13.2 Tempered Distributions 269
) := (, )
(,

= (0)
(13.18)
= (x) d n x.
Rn
Therefore
= 1. (13.19)
Because differentiation and multiplication by polynomials are continuous opera-

tions on S(Rn ), they extend to tempered distributions. From Lemma 13.1 we imme-
diately derive the following:
Lemma 13.6 For u S (Rn ),
F[D x u] = (i) u,

and
F[x u] = (i D) u.

The Fourier transform on S (Rn ) is particularly useful in the construction of

fundamental solutions. Consider the constant coefficient operator

L= a D
||m
with a C. According to Lemma 13.6 and (13.19), the Fourier transform of the
equation
L =
is
P() = 1,
where
P() := a (i) .
||m
If the reciprocal of P() makes sense as a tempered distribution then we can set

() = 1/P() and take the inverse Fourier transform to construct a fundamental
solution as an element of S (Rn ).
Example 13.7 The polynomial corresponding to on Rn is P() = ||2 . For

n 3 the function ||2 is locally integrable and decays at infinity, so this defines a
tempered distribution. Hence we should be able to compute as the inverse Fourier
transform of ||2 .
Because ||2 is not globally integrable, we cannot apply the formula (13.11)
directly. A trick to get around this is based on the fact that

et dt = 1
0
for > 0. Setting = ||2 gives

||2 = et|| dt
2
for = 0. We can pair both sides with a Schwartz function () and integrate to
show that

F 1 ||2 = F 1 et|| dt.
2
(13.20)
0
Setting a = 1/(4t) in (13.8) gives

F 1 et|| (x) = (4t) 2 e|x| /4t ,
2 n 2
so that (13.20) reduces to

F 1 ||2 = (4t) 2 e|x| /4t
n 2
dt.
0
To evaluate the integral we substitute s = |x|2 /4t to obtain

n2
1
2 |x|2 |x|2
F || = es ds
0 s 4s 2

1 n
= 2 |x|2n s 2 2 es ds.
n
4 0
In terms of the gamma function (2.17) this calculation gives the fundamental solution
1 n n
(x) = 2 ( 2 1) |x|2n . (13.21)
4
This agrees with the formula for from Theorem 12.8, because
n
2 2
An =
( n2 )
and
( n2 )
( n2 1) = .
n
2
1
13.3 The Wave Kernel 271
13.3 The Wave Kernel
Following the discussion in Sect. 12.6, we can try to define the wave kernel Wt on
Rn by solving the distributional equations

2 Wt
Wt = 0, W0 = 0, = . (13.22)
t 2 t t=0
If we assume that Wt S (Rn ), then the spatial Fourier transform allows us to

analyze this equation by turning it into a simple ODE.
For each t define W t S (Rn ) by the (spatial) distributional transform (13.16).
By Lemma 13.6 and (13.19), (13.22) transforms to

2 W t
+ ||2
W t = 0, W 0 = 0, = 1.
t 2 t t=0
The unique solution to this ODE is

|| , = 0,
sin(t||)
W t () = (13.23)
t, = 0.
The function W t is smooth and bounded, and therefore defines a tempered distribution
on Rn . The inverse Fourier transform Wt S (Rn ) thus yields a general solution
formula for the wave equation on Rn . For initial conditions g, h S(Rn ),
Wt
u(t, ) = g + Wt h. (13.24)
t
The direct computation of the inverse Fourier transform of (13.23) is rather tricky,
but we can check this formula against the results we already know. For n = 1 we
have Wt = 21 [t,t] from the dAlembert formula. Since this is integrable the Fourier
transform can be computed directly:

1
W t () = [t,t] (x) ei x d x
2
t
1
= ei x d x
2 t

sin(t)

, = 0,
=
t, = 0.
For n = 3, the Kirchhoff formula from Theorem 4.10 shows that the wave kernel
is the distribution defined by

1
(Wt , ) := d S,
4t B(0;t)
for S(R3 ). By definition, the Fourier transform is given by

1
W t , := (x) d S(x)
4t B(0;t)

1
= ei x () d 3 d S(x).
4t B(0;t) R3
Since (y) has rapid decay as y and the x integral is restricted to a sphere,
we can switch the order of integration and conclude that

1
W t () = ei x d S(x).
4t B(0;t)
To compute this surface integral, note that we could rotate the x coordinate without
changing the result of the integration. It therefore suffices to consider the case where
is parallel to the x3 axis. If we then use the spherical coordinates (r, , ) for the
x variables, this gives
x = || r cos .
For the surface integral at radius r = t,
d S(x) = t 2 sin d d.
The Fourier transform is thus

2
1
W t () = eit|| cos t 2 sin d d
4t 0
0
t it|| cos
= e sin d.
2 0
With the substitution u = cos this becomes

1
t
W t () := eit||u du
2 1

|| , = 0,
sin(t||)
=
t, = 0.
Hence the Kirchhoff formula agrees with the transform solution (13.23).
13.4 The Heat Kernel 273
13.4 The Heat Kernel
By analogy with (13.22), the heat kernel Ht is defined as the solution of the distrib-
utional equation

Ht = 0, H0 = . (13.25)
t
Assuming Ht S (Rn ), let H t denote the spatial Fourier transform of Ht . By Lemma

13.6 and (13.19), (13.25) transforms to

+ || H t = 0,
2
H 0 = 1.
t
This simple ODE has the unique solution
H t () = et|| .
2
(13.26)
Because H t is a Schwartz function for t > 0, we can compute the inverse Fourier
transform by the direct integral formula (13.11), which gives

n
eix et|| d n .
2
Ht (x) = (2)
Rn
According to (13.8), this inverse transform is
Ht (x) = (4t) 2 e|x| /4t

n 2
. (13.27)
In Sect. 6.3 we guessed this formula from a calculation in the one-dimensional case.
The Fourier transform allows for a systematic derivation.
13.5 Exercises
13.1 Let H R2 denote the upper half space {x2 > 0}. The Poisson kernel on H is
the distributional solution of the equation
P = 0, P|x2 =0 = .
x2 ) denote the distributional Fourier transform of P with respect to the

(a) Let P(,

x1 variable. Find the corresponding equation for P.
x2 ) S (R) is the function
(b) Show that the unique solution of the ODE with P(,
x2 ) = ex2 || .
P(,
(c) Compute the inverse transform to show that

x2
P(x) = .
(x12 + x22 )
(d) For f S(R), use P to write an integral formula for the solution of the Laplace
problem on H:
u = 0, u|x2 =0 = f.
13.2 For S(R), the Poisson summation formula says that

(k) =
(2m).
k= m=
Derive this formula using the steps below.

(a) Define a periodic function f C (T) (where T := R/2Z as in Sect. 8.2) by

averaging ,

f (x) := + 2m).
(x
m=
Show that
ck [ f ] = (k).
(b) Obtain the summation formula by comparing f to its Fourier series expansion
at x = 0.
13.3 Recall that the heat equation on T was solved by Fourier series in Theorem
8.13.
(a) Use the solution formula (8.44) to show that the heat kernel on T is given by the
series

1 k 2 t+ikx
h t (x) := e .
2 k=
(b) Use the Poisson summation formula from Exercise 13.2 to show that the periodic
heat kernel h t and the heat kernel Ht on R are related by averaging

h t (x) = Ht (x + 2m)
m=
for t > 0. (Note that this shows h t (x) > 0 for all x T, t > 0, which is not
clear in the formula from (a).)
13.5 Exercises 275
13.4 The Schrdinger equation on Rn ,
u
i u = 0, u|t=0 = g,
t
was introduced in Exercise 4.7.
(a) Assuming that g S(Rn ), find a formula for the spatial Fourier transform u(t,
).
(b) Show that the result from Exercise 4.7,

|u(t, x)| d x =
2 n
|g|2 d n x
Rn Rn
for all t 0, follows from the Plancherel theorem (Theorem 13.5).

Appendix A
Analysis Foundations
In this section we will develop some implications of the completeness axiom for R
which are referenced in the text.
The fundamental result from which the others follow is the equivalence of com-
pactness and sequential compactness for subsets of Rn . Recall from Sect. 11.6 that
a set A is sequentially compact if every sequence within A contains a subsequence
converging to a limit in A. The equivalence was first proven by Bernard Bolzano in
the early 19th century, and later rediscovered by Karl Weierstrass.
Theorem A.1 (Bolzano-Weierstrass) In Rn a subset is sequentially compact if and

only if it is closed and bounded.

Proof If A Rn is unbounded, then there exists a sequence of points x j A
with x j . Any subsequence has the same property, so x j has no convergent
subsequence. If A is not closed, then there is some w / A which is a boundary
point of A. Every
neighborhood of w thus includes points of
A, so there exists a
sequence x j A converging to w. All subsequences of x j also converge to
w, and therefore no subsequence converges in A. We conclude that a sequentially
compact subset of Rn is closed and bounded.
For
the converse argument, let us first consider the one-dimensional case. Let
x j be a sequence in a bounded set A R. For each n the real number
bn := sup {xk ; k n} (A.1)
exists by the completeness axiom. The sequence {bn } is decreasing, because the
supremum is taken over successively smaller sets, and also bounded by the hypothesis
on A. Therefore the number
:= inf bn
nN
is well-defined in R. The fact that {bn } is decreasing implies bn for all n.

We claim that a subsequence of xk converges to . For this purpose it suffices to
show that the interval ( , + ) contains infinitely many xk for each > 0.
Universitext, DOI 10.1007/978-3-319-48936-0
278 Appendix A: Analysis Foundations
If this were not the case, then for some n we would have xk / ( , + ) for
all k n. This would imply either bn or bn + , both of which are
impossible by the definition of .
This proves the existence of a subsequence converging to . The fact that A is
closed implies A, so this completes the argument that a closed bounded subset
of R is sequentially compact.
To extend this argument to higher dimensions, consider a sequence {x k } in a
compact subset A Rn . The sequence of first coordinates of the x k is a bounded
sequence in R, so the above argument yields a subsequence such that the first coor-
dinates converge. We can then restrict our attention to this subsequence and apply
the same reasoning to the second coordinate, and so on. After n steps this procedure
produces a subsequence which converges to an element of A.
Bolzano used sequential compactness to prove the following result, which serves
as the foundation for applications of calculus to optimization problems.
Theorem A.2 (Extreme value theorem) For a compact set K Rn , a continuous

function K R achieves a maximum and minimum value on K .
Proof Assume that f : K R is continuous. We will show first that f is bounded.
Suppose there is a sequence x j K such that f (x j ) . By Theorem A.1,
after restricting to a subsequence if necessary, we can assume that x j w
K . Continuity implies f (x j ) f (w), but this is impossible if f (x j ) .
Therefore a continuous function on K is bounded.
Since f (K ) is a bounded subset of R, b := sup f (K ) exists in R by the com-
pleteness axiom. To prove that f achieves a maximum, we need to show b f (K ).
If b
/ f (K ) then the function
1
h(x) :=
b f (x)
is continuous on K , and therefore bounded by the above argument. However, h(x)

M for x K would imply that sup f (K ) b 1/M, contradicting the definition
of b. Therefore b f (K ), so f achieves a maximum. A similar argument applies to
the minimum.
The final result is the completeness of Rn as a normed vector space, as noted in

Sect. 7.4.
Theorem A.3 In Rn a sequence converges if and only if it is Cauchy.
Proof We have already noted that a convergent sequence is Cauchy in a normed

vector space. Suppose that {x k } is a Cauchy sequence in Rn . This implies in particular
that the sequence is bounded. Therefore, by Theorem A.1, there exists a subsequence
converging to some w Rn .
Appendix A: Analysis Foundations 279
By the definition of Cauchy, for > 0 there exists N sufficiently large such that

x j xk <
for all j, k N . We can also choose an element x l in the subsequence such that
l N and
|x l w| < .
The triangle inequality then gives

x j w x j x l + |x l w|
< 2,
for all j N . Since the choice of was arbitrary, this shows that the full sequence
converges to w.
References
For additional analysis background
Apostol, T. M. (1974). Mathematical analysis (2nd ed.). Addison-Wesley.

Edwards, R. E. (1979). Fourier series. A modern introduction (2nd ed.). Graduate Texts in Mathe-
matics 64. Springer.
Folland, G. B. (1999). Real analysis (2nd ed.). Pure and Applied Mathematics. Wiley.
Rudin, W. (1976). Principles of mathematical analysis (3rd ed.). McGraw-Hill.
Stein, E. M., & Shakarchi, R. (2003). Fourier analysis. Princeton Lectures in Analysis I: Princeton.
Stein, E. M., & Shakarchi, R. (2005). Real analysis. Princeton Lectures in Analysis III: Princeton.
Stein, E. M., & Shakarchi, R. (2011). Functional analysis. Princeton Lectures in Analysis IV:
Princeton.
Strichartz, R.S. (2003). A Guide to Distribution Theory and Fourier Transforms. World Scientific.
For further study of PDE
DiBenedetto, E. (2010). Partial differential equations (3nd ed.). Birkhuser.

Evans, L. C. (2010). Partial differential equations (2nd ed.). Graduate Studies in Mathematics 19.
American Mathematical Society.
Gilbarg, D., & Trudinger, N. S. (2001). Elliptic partial differential equations of second order.
Classics in Mathematics: Springer.
Jost, J. (2013). Partial differential equations (3rd ed.). Graduate Texts in Mathematics 214. Springer.
Strauss, W. A. (2008). Partial differential equations. An introduction (2nd ed.). Wiley.
Taylor, M. E. (2011). Partial differential equations I. Basic theory, Applied Mathematical Sciences
115. Springer.
Vasy, A. (2015). Partial differential equations, Graduate Studies in Mathematics 169. American
Mathematical Society.

Index
A support, 14
acoustic waves, 59 completeness, 9, 119, 276
action functional, 232 conduction, 95
advection, 27, 95 conjugate, 10
almost everywhere, 113 connected, 13
continuity equation, 27, 33
convection, 95
B convergence
ball, 12 pointwise, 135
Banach space, 119 uniform, 135, 139
Bessel convolution, 101, 137, 247
functions, 80 Coulomb, 237
inequality, 121, 133
zeros, 84
Bolzano-Weierstrass theorem, 275 D
boundary conditions Darboux, 61
Dirichlet, 5 delta function, 239, 241
Neumann, 5 dense, 117
self-adjoint, 124 Dirichlet, 136
boundary point, 12 eigenvalues, 215, 226
Burgers equation, 43 energy, 203
kernel, 137
Dirichlets principle, 205
C dispersive estimate, 116
calculus of variations, 232 distribution, 237, 240
Cauchy-Schwarz inequality, 111 tempered, 265
Cauchy sequence, 118 distributional derivative, 243
characteristic, 27, 33 divergence-free, 33
characteristic function, 113 divergence theorem, 20
classical solution, 1 domain, 13
closed, 13, 119 domain of dependence, 54
closure, 13, 186 Duhamels method, 54
coercive, 209
compact, 14 E
sequential, 222 eigenfunction, 75
sequentially, 222, 275 eigenvalue, 75

284 Index
elliptic, 4, 153, 165, 192, 235 initial conditions, 5

regularity, 212 inner product, 109
uniformly, 165, 170 integrable, 114
energy, 66, 72, 95, 165 integral kernel, 102
equilibrium, 153 interior point, 12
essential supremum, 116 irrotational, 154
Euler, 60, 232
formula, 11
Euler-Lagrange equation, 233 K
evolution equations, 4 Kirchhoff, 62
Klein-Gordon equation, 72
F
finite element method, 7, 231
flux, 25, 32 L
forcing term, 42, 54 Lagrange, 28, 232
Fourier, 95 Lagrangian derivative, 28, 34, 60
series, 132, 189, 213 Lagrangian function, 233
transform, 259 Laplace equation, 153
functional, 203 Laplacian, 4, 18, 59
fundamental solution, 246 eigenvalue equation, 75
polar coordinate, 80
spherical, 86
G Legendre functions, 86
gamma function, 24, 268 linear conservation equation, 27, 33
Gauss, 204 linear equation, 3
Gaussian function, 260 Liouvilles theorem, 172
gradient, 18 local integrability, 176
Greens function, 252
Greens identities, 23
M
material derivative, 28
H maximum, 10
Hamilton-Jacobi equation, 43 measurable function, 113
harmonic function, 153 measure, 112
harmonic ODE, 17
Mersenne, 74, 77
harmonic oscillator, 76
method of characteristics, 29, 47
heat equation, 91, 95, 102, 129, 146, 148,
method of descent, 64
168
method of images, 253
heat kernel, 102, 271
minimal surface equation, 234
Heaviside step function, 100
minimax principle, 226
Helmholtz equation, 75, 130, 192
minimum, 10
Hermite polynomials, 92
Minkowski inequality, 115
Hilbert space, 119
multi-index, 178
Hlder continuity, 246
Hopfs lemma, 173
Huygens principle, 50, 64
hyperbolic, 4 N
Navier-Stokes, 7
neighborhood, 12
I Newton, 205
infimum, 9 nodes, 77
infinite propagation speed, 104 norm, 110
inflow boundary, 42 normal vector, 19
Index 285
O Schwartz functions, 260

open set, 12 separation of variables, 73
order, 1 sequential compactness, 222
order notation, 143 sequentially compact, 222, 275
orthogonal, 110 sesquilinear, 110
complement, 219 sesquilinearity, 110
orthonormal, 120 shock, 41, 43
basis, 120 curve, 181
overtones, 74 smooth, 13
Sobolev
embedding theorem, 188
P
regularity, 188
parabolic, 4, 153, 170
spaces, 185
parallelogram law, 126
solenoidal, 33, 154
piecewise C 1 boundary, 15
spectral
piecewise linear, 185
plane wave, 72, 259 analysis, 129
Plateau problem, 234 theorem, 123, 215
Poincar constant, 209 spectrum, 77
Poisson spherical harmonics, 87
equation, 204, 212 stability, 5
integral formula, 65 subharmonic, 163
kernel, 155, 271 sup norm, 112
summation formula, 272 superharmonic, 163
polar coordinates, 23, 79 superposition principle, 4
potential function, 154 support, 14
principal value, 245 supremum, 9
propagation speed, 47, 49 surface integral, 19
finite, 66
T
Q
telegraph equation, 70
quasilinear equation, 36
tempered distribution, 265
test function, 175
R traffic equation, 37
range of influence, 56, 64 transport equation, 27
Rankine-Hugoniot condition, 182
rapidly, 260
rarefaction wave, 201 V
Rayleigh variation of parameters, 54
principle, 216 vector space, 109
quotient, 216
Rayleigh-Ritz method, 229
Reaction-diffusion equations, 107
W
reaction term, 42
regularity wave equation, 2, 47, 83, 269
elliptic, 212 acoustic, 60
Rellichs theorem, 217, 222 damped, 91
resonance, 58 wave kernel, 257, 269
Weak derivative, 176
weak solution, 1, 51, 175
S Weierstrass, 275
Schrdinger equation, 71 well posed, 5, 165

PDF

Uploaded by

Copyright:

Available Formats

PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

PDF

Uploaded by

Copyright:

Available Formats

Universitext

Universitext is a series of textbooks that presents material from a wide variety of

More information about this series at http://www.springer.com/series/223

ISSN 0172-5939 ISSN 2191-6675 (electronic)

Springer International Publishing AG 2016

Printed on acid-free paper

This Springer imprint is published by Springer Nature

I owe a great debt of gratitude to my colleague Alessandro Veneziani, with

Atlanta, USA David Borthwick

10.5 Weak Formulation of Elliptic Equations . . . . . . . . . . . . . . . . . . . 194

1.1 Partial Differential Equations

Springer International Publishing AG 2016 1

1.2 Example: dAlemberts Wave Equation

u(t, x) := f 1 (x + t) + f 2 (x t). (1.3)

Fig. 1.1 Initial state of a

Fig. 1.2 Evolution of the

1.3 Types of Equations

where f is some function independent of u, and L is a differential operator. Many of

where the coefficients ai j , b j , and c are functions of x. The second-order coefficients

named after the mathematician and physicist Pierre-Simon Laplace.

1.4 Well Posed Problems

example in fluid mechanics, well-posedness remains a difficult unsolved conjecture.

2.1 Real Numbers

and similarly for inf A.

Proof If A is not bounded above then there exists a sequence of xk A with xk .

2.2 Complex Numbers

The absolute value on C is the vector absolute value from Euclidean R2 ,

This can be written in terms of conjugation,

which shows in particular that the absolute value is multiplicative,

It follows from the completeness axiom of R that absolute convergence of a series

cos( + ) = cos cos sin sin ,

The Euclidean distance is used to define limits: limk x k = w means that

The ball of radius R > 0 centered at a point x 0 Rn is

A small ball centered at x 0 is called a neighborhood of x 0 . If x A has a neighbor-

For example, the boundary of the ball B(x 0 ; R) is the sphere

A set is open if and only if it contains no boundary points.

Fig. 2.1 Open rectangle

A subset of Rn is connected if any two points in the set can be joined by a

in the set or a boundary point.

The space of continuous, complex-valued functions on a domain Rn which

an adjective, short for m-times continuously differentiable.

The support of f C 0 () is defined as

Example 2.2 To demonstrate the existence of compactly supported smooth func-

where qm denotes a polynomial of degree m. As x 1, the term (1 x 2 )m blows

Fig. 2.2 Compactly

Certain problems require regularity of the boundary of . For example, many

j = 1, . . . , n 1, are linearly independent at each point of .

2.5 Ordinary Differential Equations

Example 2.3 Consider the equation

Higher-order ODE are generally analyzed by reducing to a system of first-order

First-order systems can be solved generally by the strategy of Picard iteration,

For the construction, we set u0 (t) := w 0 and define a sequence of functions by

Example 2.5 The harmonic ODE is the equation

y(t) = c1 eit + c2 eit .

for k N. In the limit k this gives

y(t) = a cos(t) + b sin(t).

2.6 Vector Calculus

For a vector-valued function v C 1 (; Rn ) with components (v1 , . . . , vn ), the

The Laplacian operator introduced in (1.7) is the divergence of the gradient

Substitute t = x 2 to evaluate the one-dimensional integral in terms of ( 21 ).