An Introduction to

Linear Partial Differential Equations

Qing Han

To Yansu, Raymond and Tommy

Contents
Preface

vii

Chapter 1. Introduction
1.1. Notations
1.2. Well-Posed Problems
1.3. Overview

1
1
3
4

Chapter 2. First-Order Differential Equations

2.1. Non-Characteristic Hypersurfaces
2.2. The Method of Characteristics
2.3. A Priori Estimates
Exercises

7
7
11
23
34

Chapter 3. An Overview of Second-Order PDEs

3.1. Classifications of Second Order Equations
3.2. Energy Estimates
3.3. Separation of Variables
Exercises

37
37
43
49
56

Chapter 4. Laplace Equations

4.1. Dirichlet Problems
4.2. Fundamental Solutions
4.3. Mean-Value Properties
4.4. The Maximum Principle
4.5. Schauder Estimates
4.6. Weak Solutions
Exercises

59
59
65
71
76
86
91
97

Chapter 5. Heat Equations

5.1. Initial-Value Problems
5.2. Regularity of Solutions
5.3. The Maximum Principle
5.4. Harnack Inequalities
Exercises

101
101
112
117
123
129

Chapter 6. Wave Equations

6.1. One-Dimensional Wave Equations
6.2. Higher Dimensional Wave Equations
6.3. Energy Estimates
Exercises

133
133
142
153
160

vi

CONTENTS

Chapter 7. First-Order Differential Systems

7.1. Reductions to First-Order Differential Systems
7.2. Cauchy-Kowalevski Theorem
7.3. Hyperbolic Differential Systems
Exercises

165
165
171
180
188

Bibliography

191

Index

193

Preface
Is it really necessary to classify partial differential equations (PDEs) and to
employ different methods to discuss different types of equations? Why is it important to derive a priori estimates of solutions before even proving the existence
of solutions? These are only a few questions any students who just start studying
PDEs might ask. Students may find answers to these questions only at the end of a
one-semester course in basic PDEs, sometimes after they have already lost interest
in the subject. In this book, we attempt to address these issues at the beginning.
There are several notable features in this book.
First, the importance of a priori estimates is addressed at the beginning and
emphasized throughout this book. This is well illustrated by the first chapter on
first-order PDEs. Although first-order linear PDEs can be solved by the method of
characteristics, we provide a detailed analysis of a priori estimates of solutions in
sup-norms and in integral norms. To emphasize the importance of these estimates,
we demonstrate how to prove the existence of weak solutions with the help of basic
results from functional analysis. The setting here is easy, since Sobolev spaces are
not required. Meanwhile, all important ideas are in full display. In this book, we
do attempt to derive explicit expressions for solutions whenever possible. However,
these explicit expressions of solutions of special equations usually serve mostly to
suggest the correct form of estimates for solutions of general equations.
The second feature is the illustration of the necessity of classifying secondorder PDEs at the beginning. In the second chapter, immediately after classifying
second order PDEs into elliptic, parabolic and hyperbolic type, we discuss various boundary-value problems and initial/boundary-value problems for the Laplace
equation, the heat equation and the wave equation. We discuss energy methods for
proving uniqueness and find solutions in the plane by the method of separation of
variables. The explicit expressions of solutions demonstrate different properties of
solutions of different types of PDEs. Such differences clearly indicate that there is
unlikely to be a unified approach to study PDEs.
Third, we focus on simple models of PDEs and study these equations in detail.
We have chapters devoted to the Laplace equation, the heat equation and the wave
equation, and use several methods to study each equation. For example, for the
Laplace equation, we use three different methods to study harmonic functions, the
fundamental solution, the mean-value property and the maximum principle. For
each method, we indicate its advantages and its shortcomings. General equations
are not forgotten. We also discuss maximum principles for general elliptic and
parabolic equations and energy estimates for general hyperbolic equations.

vii

viii

PREFACE

The book is designed for a one-semester course at the graduate level. As a

result, many important topics have to be omitted. One topic notably missing from
this book is Sobolev spaces. It is hard, if not impossible, to introduce Sobolev
spaces and discuss applications in a one-semester course on basic PDEs. However, we seize every opportunity we can to address the importance of the missing
topic. For example, in Chapter 3 we solve initial/boundary-value problems for the
1-dimensional heat equation and for the 1-dimensional wave equation by using separation of variables. An important role is played by the eigenvalue problem for
d2
over an interval. After successfully solving 1-dimensional probthe operator
dx2
lems, we point out that, in order to solve higher dimensional versions, we need to
solve first the eigenvalue problem for the Laplace operator in a bounded smooth
domain. For that, Sobolev spaces play an essential role. Without filling in details,
we introduce Sobolev spaces and discuss how to find weak solutions of the Dirichlet
problem of the Poisson equation in Section 4.6.
The choice of topics in this book, as in any books, is influenced by personal
tastes of the author. Some topics in this book may not be viewed as basic by
others. For example, differential Harnack inequality for the heat equation usually
is not found in PDE textbooks at a comparable level. It is included here simply
because of its importance. Attempts have been made to give a balanced coverage
of different classes of partial differential equations.
Acknowledgments. This book is based on the one-semester course the author
taught at University of Notre Dame in the fall of 2007. The author would like to
thank Ms. Hairong Liu for a wonderful job of typing the manuscript.
Qing Han

CHAPTER 1

Introduction
This chapter serves as an introduction of entire book.
1.1. Notations
Most of the time, we denote by x points in Rn and write x = (x1 , , xn ) in
terms of its coordinates. For any x Rn , we denote by |x| the standard Euclidean
norm, unless otherwise stated. Namely, for any x = (x1 , , xn ), we have
n
1
X
|x| =
x2i 2 .
i=1

Sometimes, we need to distinguish one particular direction as the time direction

and write points in Rn+1 by (x, t) for x Rn and t R. In this case, we call
x = (x1 , , xn ) Rn the space variable and t R the time variable. In R2 , we
also denote points by (x, y).
Let be a domain in Rn , an open and connected subset in Rn . We denote by
C() the collection of all continuous functions in , by C m () the collection of all
functions with continuous derivatives up to order m, for any integer m 1, and by
C () the collection of all functions with continuous derivatives of arbitrary order.
For any u C m (), we denote by m u the collection of all partial derivatives of u
of order m. In particular,
u = (ux1 , , uxn ).
This is the gradient of u. For second derivatives, we usually write in the matrix
form as follows

ux1 x1 ux1 x2 ux1 xn

ux2 x1 ux2 x2 ux2 xn

2 u = .
..
.. .
..
..
.
.
.
uxn x1 uxn x2 uxn xn
This is a symmetric matrix, called the Hessian matrix of u. For derivatives of order
higher than two, we need to use multi-indices. A multi-index Zn+ is given by
= (1 , , n ) for nonnegative integers 1 , , n . We write
n
X
|| =
i .
i=1

For any vector = (1 , , n ) Rn , we denote

= 11 nn .
The partial derivative u is defined by
u = x11 xnn u,
1

1. INTRODUCTION

and its order is ||. For any positive integer m, we define

X 2 21
|m u| =
.
| u|
||=m

In particular,
|u| =

n
X

u2xi

12

i=1

and
|2 u| =

n
X

u2xi xj

21

i,j=1

A hypersurface in Rn is a surface of n1 dimension. Locally, a C m -hypersurface

can be expressed by { = 0} for a C m -function with 6= 0. Alternatively, by
a rotation, we may take (x) = xn (x1 , , xn1 ) for a C m -function of n 1
variables. A domain Rn is C m if its boundary is an C m -hypersurface.
A partial differential equation (henceforth abbreviated as PDE) in a domain
Rn is a relation of independent variables x , an unknown function u
defined in and a finite number of its partial derivatives. Solving a PDE is to find
this unknown function. The order of a PDE is the order of the highest derivative
in the relation. Hence for a positive integer m, the general form of an m-th order
PDE in a domain Rn is given by

F x, u, u(x), 2 u(x), , m u(x) = 0 for any x .

Here F is a function which is continuous in all its arguments and u is a C m function in . A C m -solution u satisfying the above equation in the pointwise
sense in is often called a classical solution. Sometimes, we need to relax regularity
requirements for solutions when classical solutions are not known to exist. Instead
of going to details, we only mention that it is an important method to establish
the existence of weak solutions first, functions with less regularity than C m and
satisfying the equation in some weak sense, and then to prove that these weak
solutions actually possess required regularity to be classical solutions.
A PDE is linear if it is linear in the unknown functions and their derivatives,
with coefficients depending on independent variables x. A general m-th order linear
PDE in is given by
X
a (x) u = f (x) for x .
||m

Here a is called the coefficient of u. A PDE of order m is quasilinear if it is linear

in derivatives of solutions of order m, with coefficients depending on independent
variables x and derivatives of solutions of order < m. In general, an m-th order
quasilinear PDE in is given by
X
a (x, u, , m1 u) u = f (x, u, , m1 u) for x .
||=m

Several PDEs involving one or more unknown functions and their derivatives
form a differential system. We define linear and quasilinear partial differential
systems accordingly.

1.2. WELL-POSED PROBLEMS

In this book, we will focus on first-order and second-order linear PDEs and
first-order linear differential systems. In a few occasions, we diverge to nonlinear
PDEs.
1.2. Well-Posed Problems
What is the meaning of solving partial differential equations? Ideally, we obtain
explicit solutions in terms of elementary functions. In practice this is only possible
for very simple PDEs or very simple solutions of more general PDEs. In general, it
is impossible to find explicit expressions of all solutions of all PDEs. In the absence
of explicit solutions, we need to seek methods to prove existence of solutions of
PDEs and discuss properties of these solutions.
A given PDE may not have solutions at all or may have many solutions. When it
has many solutions, we intend to find side conditions to pick up the most reasonable
solutions. Hadamard introduced the notion of well-posed problems. A PDE and side
conditions are called well-posed if
(i) there admits a solution;
(ii) this solution is unique;
(iii) solutions depend continuously in some suitable sense on side conditions,
i.e., solutions change a little if side conditions change a little.
Now the basic question can be formulated as follows: Given an equation, find
side conditions to have the well-posedness. We usually refer to (i), (ii) and (iii) as
the existence, uniqueness and the continuous dependence respectively.
In practice, both the uniqueness and the continuous dependence are proved by a
priori estimates. Namely, we assume solutions already exist and then derive certain
norms of solutions in terms of known functions in the equation and side conditions.
It is important to note that establishing a priori estimates is in fact the first step
in proving the existence of solutions. A closely related issue here is regularity of
solutions such as continuity and differentiability. Solutions of a particular PDE
can only be obtained if the right kind of regularity, or the right kind of norms, are
employed. Two classes of norms are used often, sup-norms and L2 -norms.
Let be a domain in Rn . For any bounded function u in , we define the
sup-norm of u in by
|u|L () = sup |u|.

Let m be a nonnegative integer. For any function u in with bounded derivatives

up to order m, we define the C m -norm of u in by
X
|u|C m () =
| u|L () .
||m

the collection of functions which

If is a bounded C -domain in R , then C m (),
m
is a Banach space equipped with the C m -norm.
are C in ,
Next, for any Lebesgue measurable function u in , we define the L2 -norm of
u in by
Z
1

u2 2 ,
kukL2 () =
m

where integration is in the Lebesgue sense. The L2 -space in is the collection

of all Lebesgue measurable functions in with finite L2 -norms and is denoted by

1. INTRODUCTION

L2 (). We learned from real analysis that L2 () is a Banach space equipped with
the L2 -norm.
Other norms will also be used. We will introduce them when needed.
In deriving a priori estimates, we follow a common practice and use the variable
constant convention. The same letter C is used to denote constants which may
change from line to line, as long as it is clear from context on what quantities the
constants depend upon. In most cases, we are not interested in the value of the
constant, but only in its existence.
1.3. Overview
There are 7 chapters in this book.
The main topic in Chapter 2 is first-order PDEs. In Section 2.1, we introduce
the basic notion of characteristic hypersurfaces. In Section 2.2, we use the method
of integral curves to study initial value problems with initial values prescribed on
non-characteristic hypersurfaces. In Section 2.3, we derive estimates of solutions in
L -norms and in L2 -norms without using explicit expressions of solutions.
Chapter 3 should be considered as an introduction to the theory of second-order
linear PDEs. In Section 3.1, we introduce the Laplace equation, the heat equation
and the wave equation. We also introduce their general forms, elliptic equations,
parabolic equations and hyperbolic equations, which will be studied in detail in
subsequent chapters. We use energy methods to discuss the uniqueness of certain
boundary value problems in Section 3.2 and use separation of variables to solve
these problems in the plane in Section 3.3.
In Chapter 4, we discuss the Laplace equation and the Poisson equation. The
Laplace equation is probably the most important PDE with the widest range of
applications. In Section 4.1, we solve Dirichlet problems for the Laplace equation in balls and derive Poisson integral formula. In the next three sections, we
study harmonic functions, (i.e., solutions of the Laplace equation), by three different methods: fundamental solutions, mean-value properties and the maximum
principle. These three sections are relatively independent of each other. In Section
4.5, we study the Poisson equation u = f and derive Schauder estimates for its
solutions. Then in Section 4.6, we briefly discuss weak solutions of the Poisson
equation.
In Chapter 5, we study the heat equation and discuss properties of its solutions.
This equation describes the temperature of a body conducting heat, when the
density is constant, In Section 5.1, we discuss the fundamental solution of the heat
equation and solve some initial-value problems. Then in Section 5.2, we use the
fundamental solution to discuss regularity of arbitrary solutions. In Section 5.3,
we discuss the maximum principle for the heat equation and its applications. In
particular, we use the maximum principle to derive interior gradient estimates. In
Section 5.4, we discuss Harnack inequalities.
In Chapter 6, we study the n-dimensional wave equation, which represents
vibrations of strings or propagation of sound waves in tubes for n = 1, waves on
the surface of shallow water for n = 2, and acoustic or light waves for n = 3. In
Section 6.1, we discuss initial-value problems and various initial/boundary-value
problems for the one-dimensional wave equation. In Section 6.2, we study initial

1.3. OVERVIEW

value problems for the wave equation in higher dimensional spaces. Then in Section
6.3, we derive energy estimates for solutions of initial-value problems. Chapter 6 is
relatively independent of Chapter 4 and Chapter 5 and can be taught after Chapter
3.
In Chapter 7, we discuss partial differential systems of first-order and focus
on non-characteristic initial-value problems. In Section 7.1, we introduce noncharacteristic hypersurfaces for partial differential equations and systems of arbitrary order. We also demonstrate that partial differential systems of arbitrary
order can always be changed to those of first-order. In Section 7.2, we discuss the
Cauchy-Kowalevski Theorem, which asserts the existence of analytic solutions of
non-characteristic initial-value problems for analytic differential systems and initial
values on analytic non-characteristic hypersurfaces. In Section 7.3, we discuss hyperbolic differential systems and derive energy estimates for symmetric hyperbolic
differential systems.
Each chapter, except this one, ends with exercises. Level of difficulty varies
considerably. Some exercises, at the most difficult level, may require efforts for
days.

CHAPTER 2

First-Order Differential Equations

The main topic in this chapter is first-order PDEs. In Section 2.1, we introduce the basic notion of non-characteristic hypersurfaces. In Section 2.2, we use
the method of characteristics to study initial value problems with initial values
prescribed on non-characteristic hypersurfaces. In Section 2.3, we derive estimates
of solutions in L -norms and in L2 -norms without using explicit expressions of
solutions.
2.1. Non-Characteristic Hypersurfaces
Let be a domain in Rn . A first-order PDE in is given by

(2.1)
F x, u, u(x) = 0 for any x ,
where F = F (x, u, p) is continuous in (x, u, p) R Rn and u is a C 1 -function
in .
We start our study with the simplest case. Given a point in , say the origin,
find a reasonable condition so that (2.1) admits a solution in a neighborhood of the
origin. Let be a hypersurface in Rn , a surface of dimension n 1, passing the
origin. We will prescribe conditions on to find a solution.
To illustrate, we consider first-order linear PDEs. Let be a domain in Rn
containing the origin and L be a first-order linear differential operator given by
Lu =

n
X

ai (x)uxi + b(x)u,

i=1

where ai are b are continuous functions in , for any i = 1, , n. Here ai and b

are called the coefficients for uxi and u respectively.
For a given function f in , we consider the equation
(2.2)

Lu = f (x) in .

The function f is called the non-homogeneous term. If f 0, (2.2) is called a

homogeneous equation.
Let be the hyperplane {xn = 0}. For x Rn , we write x = (x0 , xn ) for
0
x = (x1 , , xn1 ) Rn1 . We intend to prescribe u on to solve (2.2) in a
neighborhood of the origin. We first check whether we can find all derivatives of u
at the origin. For a given function u0 in a neighborhood of the origin in Rn1 , we
set
(2.3)

u(x0 , 0) = u0 (x0 ) for any small x0 Rn1 .

Then we can find all x0 -derivatives of u at the origin. In particular, we can

7

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

Figure 2.1. Initial values.

determine all first order derivatives of u at the origin except uxn . To find this, we
need to use the equation. If we assume
an (0) 6= 0,
then we can find uxn (0) from (2.2). In fact, we can compute all derivatives of u of
any order at the origin by using u0 and differentiating (2.2).
We usually call an initial hypersurface and u0 an initial value or a Cauchy
value. The problem of solving (2.2) together with (2.3) is called an initial-value
problem or a Cauchy problem.
More generally, consider a hypersurface given by { = 0} for a C 1 -function
in a neighborhood of the origin with 6= 0, with the origin on the hypersurface
, i.e., (0) = 0. We note that is simply a normal vector of the hypersurface .
Without loss of generality, we assume xn (0) 6= 0. Then by the implicit function
theorem, we solve = 0 around x = 0 for xn = (x1 , , xn1 ). We consider a
change of variables

x 7 y = x1 , , xn1 , (x) .
This is a well defined transform in a neighborhood of the origin with a nonsingular
Jacobian. In fact, the Jacobian J is given by

..
(y1 , , yn )

Id
.
=
J=
,
(x1 , , xn )
0
x1 xn1 xn
and hence det J(0) = xn (0) 6= 0. Now we write the operator L in new variables.
Note
n
X
uxi =
yk,xi uyk .
k=1

Therefore,
Lu =

n
n
X
X
k=1

ai (x(y))yk,xi uyk + b(x(y))u.

i=1

In new coordinates, the initial hypersurface is given by {yn = 0} and the coefficient of uyn is given by
n
X
ai (x)xi .
i=1

2.1. NON-CHARACTERISTIC HYPERSURFACES

We recall that = (x1 , , xn ) is simply a normal vector of = { = 0}.

When = {xn = 0}, or (x) = xn , then = (0, , 0, 1) and
n
X
ai (x)xi = an (x).
i=1

This reduces to the special case we just discussed.

Definition 2.1. For a linear operator L as in (2.2) defined in a neighborhood
of x0 Rn , a C 1 -hypersurface is non-characteristic at x0 if
n
X
(2.4)
ai (x0 )i 6= 0,
i=1

where = (1 , , n ) is a normal vector of at x0 . Otherwise, it is called

characteristic at x0 .

Figure 2.2. Non-characteristic hypersurfaces.

A hypersurface is non-characteristic if it is non-characteristic at every point.
Strictly speaking, a hypersurface is characteristic if it is not non-characteristic, i.e.,
if it is characteristic at some point. In this book, we will abuse this notion. When
we say a hypersurface is characteristic, we mean it is characteristic everywhere.
This should cause few confusions. When n = 2, hypersurfaces are simply curves. It
is more appropriate to call them characteristic curves and non-characteristic curves.
The non-characteristics condition has a simple geometric interpretation. If we
view a = (a1 , , an ) as a vector in Rn , then (2.4) holds if and only if a(x0 ) is
not a tangent vector to at x0 . We note that condition (2.4) is maintained under
C 1 -changes of local coordinates. This condition assures that we can compute all
derivatives of solutions at x0 .
The concept of the non-characteristics can also be generalized to linear partial
differential systems. Consider in a neighborhood of x0 Rn
n
X
Ai (x)uxi + B(x)u = f (x),
i=1

where Ai and B are N N matrices and u and f are N -column vectors. A

hypersurface is non-characteristic at x0 if
n
X

det
i Ai (x0 ) 6= 0,
i=1

10

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

where = (1 , , n ) is a normal vector of at the origin.

Next, we turn to general nonlinear partial differential equations
(2.5)

F (x, u, u) = 0.

We ask the same question as for linear equations. Given a hypersurface and an
initial value on it, can we compute all derivatives of solutions? First, we consider
the case of hyperplane {xn = 0}.
Example 2.2. Consider

n
X

u2xi = 1,

i=1

and
u(x0 , 0) = u0 (x0 ).
It is obvious that u = xi is a solution for u0 (x0 ) = xi , i = 1, , n 1. However, if
|x0 u0 (x0 )|2 > 1, there are no solutions for such an initial value.
Hence, in the case of nonlinear PDEs, the concept of non-characteristics also
depends on initial values.
Suppose an initial value is given on = {xn = 0} by
u(x0 , 0) = u0 (x0 ).
We ask whether (2.5) admits a solution in a neighborhood of the origin with the
given initial value. To answer this question, we first assume that there is a C 1 function v in a neighborhood of the origin having the given initial value and satisfying F = 0 at the origin, i.e.,

F 0, v(0), v(0) = 0.
As in the discussion of linear PDEs, we ask whether we can find uxn at the origin.
By the implicit function theorem, this is possible if

Fuxn 0, v(0), v(0) 6= 0.

Now we reconsider the equation in Example 2.2. We set
F (x, u, p) = |p|2 1

for any p Rn .

Suppose the initial value u0 satisfies

|x0 u0 (0)| < 1.
Let v = u0 +cxn for a constant to be determined. Then |v(0)|2 = |x0 u0 (0)|2 +c2 .
By choosing
p
c = 1 |x0 u0 (0)|2 6= 0,
v satisfies the equation at x = 0. For such two choices of v, we have
Fuxn (0, v(0), v(0)) = 2vxn (0) = 2c 6= 0.
Hence, the notion of non-characteristics depends on the choice of v.

2.2. THE METHOD OF CHARACTERISTICS

11

Definition 2.3. Let F = 0 be a first-order nonlinear PDE as in (2.5) in

a neighborhood of x0 Rn and be a hypersurface. With an initial value u0
prescribed on , suppose there exists a function v such that v = u0 on and
F (x0 , v(x0 ), v(x0 )) = 0. Then the hypersurface is non-characteristic at x0
with respect to the given initial value u0 and with respect to v if
n
X

Fuxi x0 , v(x0 ), v(x0 ) i 6= 0,

i=1

where = (1 , , n ) is a normal vector of at x0 .

2.2. The Method of Characteristics

In this section, we use the method of characteristics to solve first-order PDEs.
We demonstrate that solutions of any first-order PDEs with initial values prescribed
on non-characteristic hypersurfaces can be obtained by solving systems of Ordinary
Differential Equations (ODEs). This is not true for higher-order PDEs or for systems of first-order PDEs.
Let Rn be a domain. The general form of first-order PDEs in is given
by
F (x, u, u) = 0

for x ,

where F is continuous in R R . Let be a hypersurface in Rn and an initial

value be prescribed on by
u = u0

on .

In the following, we assume is a domain containing the origin and is noncharacteristic at the origin.
As shown in the previous section, for nonlinear differential equations, the noncharacteristics condition depends on initial values. For example, for a quasi-linear
PDE of the form
n
X
ai (x, u)uxi = f (x, u),
i=1

the non-characteristics condition at 0 is given by

n
X

ai 0, u0 (0) i 6= 0,

i=1

where = (1 , , n ) is a normal vector of at 0.

We first consider a simple case where the hypersurface is given by the hyperplane {xn = 0}. Obviously, a normal vector of {xn = 0} is given by (0, , 0, 1).
Our goal is to solve the following initial value problem
F (x, u, u) = 0,
u(x0 , 0) = u0 (x0 ),
for x Rn close to the origin. Here, we write x = (x0 , xn ) for x0 Rn1 . Solving
for u locally around the origin means to find values of u(x) for x close to the origin.

12

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

To motivate, we consider initial-value problems of linear homogeneous equations

n
X
ai (x)uxi = 0,
(2.6)
i=1
u(x0 , 0) = u0 (x0 ),
where ai is C 1 in a neighborhood of 0 Rn , i = 1, , n, and u0 is at least C 1 in
a neighborhood of 0 Rn1 . By introducing a = (a1 , , an ), we simply write the
equation in (2.6) as
a(x) u = 0.
Here a(x) is regarded as a vector field in Rn . Then a(x) is a directional derivative
along a(x) at x.
In the following, we assume the hyperplane {xn = 0} is non-characteristic at
the origin, i.e.,
an (0) 6= 0.
Our strategy is as follows. For any x
Rn close to the origin, we construct a
special curve along which u is constant. If such a curve starts from x
and intersects
n1
n1
R
{0} at (
y , 0) for a small y R
, then u(
x) = u0 (
y ).
To find such a curve x(t), we consider the restriction of u on it, which gives
a one-variable function u(x(t)). Now we calculate the t-derivative of this function
and have
n
X
dxi
d
u(x(t)) =
uxi
.
dt
dt
i=1
In order to have a constant value of u along this curve, we require

d
u(x(t)) = 0.
dt
A simple comparison with the equation in (2.6) yields the following sufficient condition
dxi
= ai (x) for i = 1, , n.
dt
This naturally leads to the following definition.

Figure 2.3. Integral curves.

Definition 2.4. Let x = x(t) be a C 1 -curve in Rn . It is an integral curve of
(2.6) if

dx(t)
(2.7)
= a x(t) .
dt

2.2. THE METHOD OF CHARACTERISTICS

13

The calculation preceding Definition 2.4 shows that solutions u of (2.6) are
constant along integral curves of the coefficient vector. This yields the following
method of solving (2.6). For any x
Rn near the origin, we find an integral curve
of the coefficient vector through x
by solving
dx
= a(x),
dt
(2.8)
x(0) = x
.
If it intersects the hyperplane {xn = 0} at some (
y , 0), then we let u(
x) = u0 (
y ).
It is easy to verify that, for any x Rn close to the origin, there is a unique integral curve, i.e., a solution of (2.8), starting from x and intersecting the hyperplane
{xn = 0} near the origin at a unique point.
Since (2.8) is an autonomous system (i.e., the independent variable t does not
appear explicitly), we consider the following system instead of (2.8)
(2.9)

dx
= a(x),
dt
x(0) = (y, 0).

Starting with any y Rn1 close to the origin, we expect the integral curve x(y, t)
to reach any x Rn close to the origin for small t. This is confirmed by the
following result.
Lemma 2.5. Let k 2 be an integer and a be a C k1 -vector field in a neighborhood of the origin with an (0) 6= 0. Then for any small y Rn1 and any small
t, the solution x = x(y, t) of (2.9) defines a C k -diffeomorphism in a neighborhood
of the origin in Rn .
Proof. This follows from the implicit function theorem easily. By standard
results in ordinary differential equations, (2.9) admits a C k -solution x = x(y, t) for
small (y, t) Rn1 R. We treat it as a map (y, t) 7 x and calculate its Jacobian
J at (y, t) = (0, 0). By x(y, 0) = (y, 0), we have

a1 (0)

..
x

Id
.
J(0) =
|(y,t)=(0,0) =
.
(y, t)

an1 (0)
0 0
an (0)
Hence det J(0) = an (0) 6= 0.

Hence, for any small x

, we can solve
x(y, t) = x

uniquely for small y and t. Then u(

x) = u0 (
y ) yields a solution of (2.6). Note
that t is not present in the expression of solutions. Hence the value of the solution
u(
x) depends only the initial value u0 at (
y , 0) and, meanwhile, the initial value
u0 at (
y , 0) influences the solution u along the integral curve starting from (
y , 0).
Therefore, we say the domain of dependence of the solution u(
x) on the initial value
is represented by the single point (
y , 0) and the domain of influence of the initial
value at a particular point (
y , 0) on solutions consists of the integral curve starting
from (
y , 0).

14

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

Figure 2.4. Solutions by integral curves.

For n = 2, integral curves are exactly characteristic curves. This can be seen
easily by (2.7) and Definition 2.1. Hence the ODE (2.7) is often referred to as
the characteristic ODE. This term is adopted for arbitrary dimensions. We have
demonstrated how to solve homogeneous first-order linear PDEs by using characteristics ODEs. Such a method is called the method of characteristics. Later on,
we will develop a similar method to solve general first-order PDEs.
We need to emphasize that solutions constructed by the method of characteristics are only local. In other words, they exist only in a neighborhood of the origin.
A natural question here is whether there exists a global solution for globally defined a and u0 . There are several reasons that local solutions cannot be extended
globally. First, u(
x) cannot be evaluated at x
Rn if x
is not on integral curves

Figure 2.5. Undetermined u(

x).
from the initial hypersurface, or equivalently, integral curves from x
does not intersect the initial hypersurface. Second, u(
x) cannot be evaluated at x
Rn if the
integral curve starting from x
intersects the initial hypersurface more than once. In
this case, we should have a compatibility condition for initial values and we cannot
prescribe initial values arbitrarily.
Example 2.6. Consider in R2 = {(x, y)}
uy + ux =0,
u(x, 0) =u0 (x).

2.2. THE METHOD OF CHARACTERISTICS

15

Figure 2.6. Undetermined u(

x).
Note that the vector a is given by a = (1, 1) and hence {y = 0} is non-characteristic.
Characteristic ODEs are given by
(
(
x = 1,
x|t=0 = x0 ,
and
y = 1,
y|t=0 = 0.
Hence
x =t + x0 ,
y =t.
By eliminating t, we have
x y = x0 .
Along this curve, u is constant. Hence
u(x, y) = u0 (x y).
We can interpret the fact that u is constant along the straight line x y = x0 in the
following way. If we interpret y as time, the graph of the solution represents a wave
propagating to the right with velocity 1 without changing shape. The solution u
exists globally in R2 .
Next, we discuss initial-value problems of quasi-linear PDEs. Let Rn be a
domain and consider
n
X
ai (x, u)uxi = f (x, u) in ,
(2.10)
i=1
u(x0 , 0) = u0 (x0 ) for x0 0 ,
where ai and f are C 1 -functions in R and u0 is C 1 in 0 = {x0 Rn1 ; (x0 , 0)
}. Assume the hyperplane {xn = 0} is non-characteristic at the origin, i.e.,

an 0, u0 (0) 6= 0.
Suppose (2.10) admits a solution u. We first examine integral curves

dx
=a x, u ,
dt
x|t=0 =(y, 0),

16

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

where y Rn1 . Contrary to the case of homogenous linear equations we studied

earlier, we are unable to solve this ODE since u, the unknown function we intend
to find, is present. However, viewing u as a function of t along these curves, we can
calculate how u changes. A similar calculation as before yields
du
=f (x, u),
dt
u|t=0 =u0 (y).
Hence we have an ordinary differential system for x and u. This leads to the
following method.
Consider the following ordinary differential system
dx
=a(x, u),
dt
du
=f (x, u),
dt
with initial values
x|t=0 =(y, 0),
u|t=0 =u0 (y),
where y Rn1 . In formulating this system, we treat x and u as functions of t only.
This system consists of n + 1 equations for n + 1 functions and is the characteristic
ODE of the quasilinear first-order PDE (2.10). By solving the characteristic ODE,
we have a solution given by
x = x(y, t),

u = (y, t).

As in the proof of Lemma 2.5, we can prove that the map (y, t) 7 x is a diffeomorphism. Hence, for any x
Rn close to the origin, there exists a unique y Rn1
and t R close to the origin such that
x
= x(
y , t).
Then the solution u at x
is given by
u(
x) = (
y , t).
Similarly, we discuss initial-value problems when initial hypersurfaces are not
hyperplanes. Let be a hypersurface containing the origin and u0 be a function
on . Instead of (2.10), we consider
n
X

ai (x, u)uxi = f (x, u),

i=1

u = u0

on .

We assume that is non-characteristic at 0 with respect to u0 , i.e.,

n
X
i=1

ai 0, u0 (0) i 6= 0,

2.2. THE METHOD OF CHARACTERISTICS

17

where = (1 , , n ) is a normal vector of at 0. We consider the characteristic

ODE in the following form
dx
=a(x, u),
dt
du
=f (x, u),
dt
with initial values
x|t=0 =x0 ,
u|t=0 =u0 (x0 ),
for x0 . Suppose we have a solution
x = x(x0 , t),

u = (x0 , t).

As before, the map (x0 , t) 7 x is a diffeomorphism. Hence for any given x

close

to the origin, there exists a unique (

x0 , t) R in a neighborhood of the origin
such that
x(
x0 , t) = x
.
Then define
u(
x) = (
x0 , t).
This is the desired solution.

Figure 2.7. Solutions by the method of characteristics.

Example 2.7. Consider in R2 = {(x, y)}

uy + uux =0,
u(x, 0) = x.
Note that the vector a is given by a = (u, 1) and hence {y = 0} is non-characteristic.
The characteristic ODE is given by

x = u,
x|t=0 = x0 ,
and
y = 1,
y|t=0 = 0,

u = 0,
u|t=0 = x0 .

18

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

Hence
x = tx0 + x0 ,
y =t,
u = x0 .
By eliminating t and x0 , we have
t = y,

x0 =

x
x
=
.
1t
1y

Therefore, the solution u is given by

u(x, y) =

x
.
y1

We note that this solution is not defined at y = 1. In fact, the integral curve x0
passing through the point (x0 , 0) is given by the straight line
x = x0 x0 y,
and the solution on this line is given by u = x0 . Each x0 passes the point (0, 1).

Figure 2.8. Integral curves in Example 2.7.

In general, smooth solutions of nonlinear PDEs are not expected to exist globally, as shown in Example 2.7. The equation discussed in Example 2.7 is called
Burgers equation.
We consider a more general equation
uy + g(u)ux = 0,
with an initial value is given on {y = 0} by
u(x, 0) = u0 (x).
The characteristic ODE is given by

x = g(u),
and
y = 1,

u = 0,

x|t=0 = x0 ,
y|t=0 = 0,

u|t=0 = u0 (x0 ).

First, u = u0 (x0 ), which is constant, along the integral curve starting from (x0 , 0).
Then on this integral curve,

x = g u0 (x0 ) , y = 1.

2.2. THE METHOD OF CHARACTERISTICS

19

1
. Consider x0 < x1 .
g(u0 (x0 ))
If the slope of the straight line from x1 is greater than that from x0 , these two
straight lines will intersect. For example, for g(u) = u, the slope is u01(x) . Therefore
we have a non-global solution if u0 is strictly decreasing. See Example 2.7.
So the integral curve is a straight line with the slope

Figure 2.9. Integral curves.

Example 2.8. Consider
uy + uux = 1,
u=0

on {y = x2 }.

By setting (x, y) = x2 y, we have (x, y) = (2x, 1) and hence for a = (u, 1)

a |{=0} = (u, 1) (2x, 1)|{=0} = 1 6= 0.
Therefore, {y = x2 } is non-characteristic.

x = u,
and
y = 1,

u = 1,

The characteristic ODE is given by

x|t=0 = x0 ,
y|t=0 = x20 ,

u|t=0 = 0.

Hence, we have
t2
+ x0 ,
2
y =t + x20 ,

x=

u =t.
By eliminating t and x0 , we obtain an implicit expression for u
u2 2
) .
2
So far in our discussion, initial values are prescribed on non-characteristic hypersurfaces. It becomes complicated when initial values are prescribed on characteristic hypersurfaces.
Let Rn be a domain and consider
n
X
ai (x, u)uxi = f (x, u) in ,
y = u + (x

i=1

u(x0 , 0) = u0 (x0 ) for x0 0 ,

20

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

where ai and f are C 1 in R and u0 is C 1 in 0 = {x0 Rn1 ; (x0 , 0) }. For

y0 0 , we assume the hyperplane = {xn = 0} is characteristic at y0 , i.e.,

an y0 , 0, u0 (y0 ) = 0.
Then we should have the following compatibility condition
(2.11)

n1
X

ai y0 , 0, u0 (y0 ) u0xi (y0 ) = f y0 , 0, u0 (y0 ) .

i=1

Even if y0 is the only point where {xn = 0} is characteristic, solutions may not exist
in a neighborhood of (y0 , 0) for initial values satisfying the compatibility condition
(2.11) at y0 .
Now we discuss general nonlinear first-order partial differential equations. Let
Rn be a domain containing the origin and F = F (x, u, p) be a C 1 -function in
(x, u, p) R Rn . Consider
(2.12)

F (x, u, u) = 0

for x ,

and prescribe an initial value on {xn = 0}

(2.13)

u(x0 , 0) = u0 (x0 ) for x0 with (x0 , 0) .

Here we write x = (x0 , xn ). Assume there is a scalar a0 such that

F 0, u0 (0), x0 u0 (0), a0 = 0.
We seek a solution u in a neighborhood of the origin with the given initial value
and uxn (0) = a0 . The non-characteristics condition with respect to u0 and a0 is
given by

(2.14)
Fpn 0, u0 (0), x0 u0 (0), a0 6= 0.
Note that by this condition, using the differential equation with the help of the
implicit function theorem, we can solve for a(x0 ) for small x0 Rn1 , with a(0) =
a0 , such that

F x0 , 0, u0 (x0 ), x0 u0 (x0 ), a(x0 ) = 0.

We also require
uxn (x0 , 0) = a(x0 )

for x0 small.

We start with a formal consideration. Suppose we have a solution u. Set

(2.15)

pi = uxi

for i = 1, , n.

Then
(2.16)

F (x1 , , xn , u, p1 , , pn ) = 0.

Differentiating (2.16) with respect to xi , we have

(2.17)

n
X

Fpj pi,xj = Fxi Fu uxi

for i = 1, , n,

j=1

where we used pj,xi = pi,xj . We view (2.17) as a first-order quasilinear equation

for pi , for each fixed i = 1, , n. An important feature here is that the coefficient

2.2. THE METHOD OF CHARACTERISTICS

21

for pi,xj is Fpj , independent of i. Characteristic ODEs associated with (2.17) are
given by
x j =Fpj ,

j = 1, , n,

pi = Fu pi Fxi ,
We also have
u =

n
X

uxj x j =

j=1

i = 1, , n.
n
X

pj Fpj .

j=1

Now we collect ordinary differential equations for xj , u and pi .

Consider the following ordinary differential system
x j =Fpj (x, u, p), j = 1, , n,
pi = Fu (x, u, p)pi Fxi (x, u, p),

(2.18)

u =

n
X

i = 1, , n

pj Fpj (x, u, p),

j=1

with initial values at t = 0

x|t=0 =(y, 0),
u|t=0 =u0 (y),

(2.19)

pi |t=0 =u0xi (y),

i = 1, , n 1,

pn |t=0 =a(y),
n1

where y R
. This is the characteristic ODE for general nonlinear first-order
partial differential equations. It is an ordinary differential system of 2n+1 equations
for 2n + 1 functions x, u and p. Here we view x, u and p as functions of t. Compare
this with a similar ordinary differential system of n + 1 equations for n + 1 functions
x and u for quasilinear equations. Solving (2.18) with (2.19), we have
x = x(y, t),

u = (y, t),

p = p(y, t).

We will prove that the map (y, t) 7 x is a diffeomorphism. Hence for any given x

near the origin, there exists a unique y Rn1 and t R such that
x
= x(
y , t).
Then we define u by
u(
x) = (
y , t).
Theorem 2.9. The function u defined above is a solution of (2.12)-(2.13).
We should note that this solution u depends on the choice of the scalar a0 and
the function a(x0 ).
Proof. The proof consists of several steps.
Step 1. The map (y, t) 7 x is a diffeomorphism. This is proved as in the proof
of Lemma 2.5. In fact, the Jacobian of the map (y, t) 7 x at (0, 0) is given by

..
x

Id
.
J(0) =
|y=0,t=0 =
,
(y, t)

0 0 x n (0, 0)

22

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

where
x n (0, 0) = Fpn (0, u0 (0), u0x1 (0), , u0xn1 , a0 ) 6= 0.
Hence det J(0) 6= 0 by the non-characteristics condition (2.14). By the implicit
function theorem, for any x
Rn close to the origin, we can solve x
= x(
y , t)
uniquely for y and t. Then define
u(
x) = (y, t).
We will prove that this is the desired solution and
pi (y, t) = uxi (
x(y, t)).
Step 2. We claim

F x(y, t), (y, t), p(y, t) 0.

Denote by f (t) the function at the left hand side. Then f (0) = 0 since

f (0) = F y, 0, u0 (y), x0 u0 (y), a(y)) = 0.

Next, we have by (2.18)

df (t) d
= F x(y, t), (y, t), p(y, t)
dt
dt
n
n
X
dxi
du X
dpj
=
Fxi
+ Fu
+
Fpj
dt
dt
dt
i=1
j=1
=

n
X

Fxi Fpi + Fu

i=1

n
X

pj Fpj +

j=1

n
X

Fpj (Fu pj Fxj ) = 0.

j=1

Hence f (t) 0.
Step 3. We claim
pi (y, t) = uxi (x(y, t)).
Let
wi (t) = uxi (x(y, t)) pi (y, t).
We will prove
wi (t) = 0

for any t and i = 1, , n.

First, consider t = 0. For i = 1, , n 1, wi (0) = 0 by initial values (2.19).

For i = n, we will show uxn (y, 0) = a(y). To see this, we first note by (2.18)
(2.20)

0 = u

n
X
j=1

pj Fpj =

n
X

uxj x j pj Fpj ) =

j=1

n
X

Fpj (uxj pj ).

j=1

Since wi (0) = 0 for i = 1, , n1, we have Fpn wn |t=0 = 0. This implies wn (0) = 0
since Fpn |t=0 6= 0 by the non-characteristics condition (2.14).
To prove wi (t) = 0 for any t, we show w i is a linear combination of wl , l =
1, , n, i.e.,
n
X
w i =
ail wl .
l=1

To prove this, we differentiate (2.20) with respect to xi and get

n
X
j=1

Fpj (uxi xj pj,xi ) +

n
X
(Fpj )xi wj = 0.
j=1

2.3. A PRIORI ESTIMATES

23

Then
w i =
=

n
X
j=1
n
X

uxi xj Fpj + Fu pi + Fxi

Fpj pj,xi

j=1

By Step 2,

n
X

(Fpj )xi wj + Fu pi + Fxi .

j=1

F x, u(x), p1 (x), , pn (x) = 0.

Differentiating with respect to xi , we have

n
X
Fxi + Fu uxi +
Fpj pj,xi = 0.
j=1

Hence
w i = Fxi Fu uxi

n
X

(Fpj )xi wj + Fu pi + Fxi

j=1

= Fu wi

n
X

(Fpj )xi wj ,

j=1

or
w i =

n
X

Fu ij + (Fpj )xi wj .
j=1

This ends the proof of Step 3.

Step 2 and Step 3 imply that u is the desired solution.

To end this section, we briefly compare methods we used to solve first-order

linear or quasilinear PDEs and general first-order nonlinear PDEs. In solving a
first-order quasilinear PDE, we formulate an ordinary differential system of n + 1
equations for n + 1 functions x and u. For a general first-order nonlinear PDE,
the corresponding ordinary differential system consists of 2n + 1 equations for 2n +
1 functions x, u and u. Here, we need to take into account the gradient of u
by adding n more equations for u. In other words, we may regard our firstorder nonlinear PDE as a relation for (u, p) with a constraint p = u. We should
emphasize this is a unique feature for single first-order PDEs. For PDEs of higher
order or first-order partial differential systems, nonlinear equations are dramatically
different from linear equations. In the rest of the book, we concentrate only on linear
equations.
2.3. A Priori Estimates
A priori estimates play a fundamental role in partial differential equations.
Usually, they are the starting point for the existence and regularity of solutions. To
derive a priori estimates, we first assume solutions already exist and then control
certain norms of solutions by those of known functions in equations, such as nonhomogenous terms and coefficients. Two frequently used norms are L -norms and
L2 -norms. The importance of L2 -norm estimates lies in the Hilbert space structure
of the L2 -space. Once L2 -estimates of solutions and their derivatives are derived,

24

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

standard results in Hilbert spaces such as the Riesz representation theorem can be
utilized. In many cases, this is how the existence of solutions is established.
In this section, we will use first-order linear PDEs to demonstrate how to derive
a priori estimates in L -norms and L2 -norms. We first examine briefly first-order
linear ordinary differential equations. Let be a constant and f = f (t) be a
function. Consider
du
u = f (t).
dt
A simple calculation shows
Z t
e(ts) f (s)ds.
u(t) = et u(0) +
0

For any T > 0, we have

|u(t)| et |u(0)| + T sup |f | for any t (0, T ).

[0,T ]

Here, we control the sup-norm of u in [0, T ] by the initial value u(0) and the supnorm of the nonhomogeneous term f in [0, T ].
Now we turn to PDEs. For convenience, we work in Rn R+ and denote points
by (x, t), with x Rn and t R+ . Let ai , b and f be functions in Rn R+ . We
consider
n
X
ut +
ai (x, t)uxi + b(x, t)u =f (x, t) in Rn R+ ,
(2.21)
i=1
u(x, 0) =u0 (x) in Rn ,
where a = (a1 , , an ) satisfies
1
in Rn R+ ,

for a positive constant . Obviously, the initial hypersurface {t = 0} is noncharacteristic. We may write the equation in (2.21) as

(2.22)

|a|

ut + a(x, t) x u + b(x, t)u = f (x, t).

We note that a(x, t)x +t is a directional derivative along the direction (a(x, t), 1).
With (2.22), it is easy to see that the vector (a(x, t), 1) (starting from the origin)
is in fact in the cone given by
{(y, s); |y| s} Rn R.
This is a cone opening upward and with vertex at the origin.
For any point P = (X, T ) Rn R+ , consider the cone C (P ) (opening
downward) with vertex at P defined by
C (P ) = {(x, t); 0 < t < T, |x X| < T t}.
We denote by s C (P ) and C (P ) the side and bottom of the boundary respectively, i.e.,
s C (P ) = {(x, t); 0 < t T, |x X| = T t},
C (P ) = {(x, 0); |x X| T }.
We note that C (P ) is simply a closed disc centered at X with radius T / in
Rn {0}. For any (x, t) s C (P ), let a(x, t) be a vector in Rn satisfying (2.22).

2.3. A PRIORI ESTIMATES

25

Figure 2.10. The cone with the vertex at the origin.

Then the vector (a(x, t), 1), if positioned at (x, t), points outward from the inside
of the cone C (P ). Hence for a function defined only in C (P ), it makes sense to
calculate ut + a x u at (x, t). This is in particular true when (x, t) is the vertex
P.

Figure 2.11. The cone C (P ) and positions of vectors.

Now we calculate the unit outward normal vector of s C (P ) \ {P }. Set
(x, t) = |x X| (T t).
Obviously, s C (P ) \ {P } is a part of { = 0}. Then for any (x, t) s C (P ) \ {P }
xX
= (x , t ) =
,1 .
|x X|
Therefore, the unit outward normal vector of s C (P ) \ {P } at (x, t) is given by
xX
1

=
,1 .
2
+ 1 |x X|
For n = 1, the cone C (P ) is a triangle bounded by straight lines (x X) =
T t and t = 0. The side of the cone consists of two line segments
the left segment:

(x X) = T t, 0 < t T, with a normal vector (, 1),

the right segment: (x X) = T t, 0 < t T, with a normal vector (, 1).

It is easy to see that the integral curve associated with (2.21) starting from P
and going to the initial surface Rn {0} stays in C (P ). In fact, this is true for any
point (x, t) C (P ). This suggests that solutions in C (P ) should depend only

26

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

on f in C (P ) and the initial value u0 on C (P ). The following result, proved

using a maximum principle type argument, confirms this.

Figure 2.12. The domain of dependence.

Theorem 2.10. Let ai , b and f be continuous functions in Rn R+ satisfying
(2.22) and u be a C 1 -solution of (2.21) in Rn R+ . Then for any P = (X, T )
Rn R+ ,
sup |et u| sup |u0 | + T sup |et f |,
C (P )

C (P )

C (P )

where is a constant such that

b

in C (P ).

If b 0, we take = 0 and have

sup |u|
C (P )

sup

|u0 | + T sup |f |.

C (P )

C (P )

Proof. Take any positive number > and let

M=

sup

|u0 |,

C (P )

F = sup |e t f |.
C (P )

We will prove
0

|e t u(x, t)| M + tF

for any (x, t) C (P ).

For the upper bound, we consider

0

w(x, t) = e t u(x, t) M tF.

A simple calculation shows
n
X

0
wt +
ai (x, t)wxi + b(x, t) + 0 w = b(x, t) + 0 (M + tF ) + e t f F.
i=1

The expression in the right hand side is nonpositive since b + 0 > 0. Hence

wt + a x w + b + 0 )w 0 in C (P ).
Let w attain its maximum in C (P ) at (x0 , t0 ) C (P ). We prove w(x0 , t0 ) 0.
First, it is obvious if (x0 , t0 ) C (P ), since w(x0 , t0 ) = u0 (x0 ) M 0 by the
definition of M . If (x0 , t0 ) C (P ), then
(wt + a x w)|(x0 ,t0 ) = 0.

2.3. A PRIORI ESTIMATES

27

If (x0 , t0 ) s C (P ), by the position of the vector (a(x0 , t0 ), 1) relative to the cone

C (P ), we have
(wt + a x w)|(x0 ,t0 ) 0.
Hence in both cases, we obtain
(b + 0 )w|(x0 ,t0 ) 0.
Since b + 0 > 0, this implies w(x0 , t0 ) 0. (We need the positivity of b + 0 here!)
Hence w(x0 , t0 ) 0 in all three cases. Therefore, w 0 in C (P ), or,
0

u(x, t) e t (M + tF ) for any (x, t) C (P ).

We simply let 0 to get the desired upper bound. For the lower bound, we
consider
0
v(x, t) = e t u(x, t) + M + tF.
The argument is similar and hence omitted.

For n = 1, (2.21) has the form

ut + a(x, t)ux + b(x, t)u = f (x, t).
In this case, it is straightforward to see that
(wt + awx )|(x0 ,t0 ) 0,
1
if w assumes its maximum at (x0 , t0 ) s C (P ). We first note that t + x and

1
t x are directional derivatives along straight lines t t0 = (x x0 ) and

t t0 = (x x0 ) respectively. Since w assumes its maximum at (x0 , t0 ), we have

1
1
wx )|(x0 ,t0 ) 0, (wt wx )|(x0 ,t0 ) 0.

In fact, one of them is zero if (x0 , t0 ) s C (P ) \ {P }. Then we obtain

(wt +

1
|wx |(x0 , t0 ) |awx |(x0 , t0 ).
k
Theorem 2.10 implies the uniqueness of solutions of (2.21). Let u1 and u2 be
two solutions of (2.21). Then u1 u2 satisfies (2.21) with f = 0 in C (P ) and
u0 = 0 on C (P ). Hence u1 u2 = 0 in C (P ).
Theorem 2.10 also shows that the value u(P ) depends only on f in C (P ) and
u0 on C (P ). Hence C (P ) contains the domain of dependence of u(P ) on f and
C (P ) contains the domain of dependence of u(P ) on u0 . In fact, the domain of
dependence of u(P ) on f is the integral curve through P in C (P ) and the domain
of dependence of u(P ) on u0 is the intercept of this integral curve with the initial
hyperplane {t = 0}. We consider this from another point of view. For simplicity,
we assume f is identically zero and the initial value u0 at t = 0 is zero outside a
bounded domain D0 Rn . Then for any t > 0, u(, t) = 0 outside
wt (x0 , t0 )

Dt = {(x, t); |x x0 | < t for some x0 D0 }.

In other words, u0 only influences u in {t>0} Dt . This is the so-called finite-speed
propagation.
Now we use the same idea to estimate derivatives.

28

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

Figure 2.13. The domain of influence.

Theorem 2.11. Let ai , b and f be C 1 -functions in Rn R+ satisfying (2.22)
and u be a C 2 -solution of (2.21) in Rn R+ . Then for any P = (X, T ) Rn R+ ,

|u|C 1 (C (P )) C |u0 |C 1 ( C (P )) + |f |C 1 (C (P )) ,
where C is a positive constant depending only on T and the C 1 -norms of ai and b
in C (P ).
Proof. In the following, we prove

sup |x u| C
sup |x u0 | + sup |x f | + sup |u| ,
C (P )

C (P )

C (P )

C (P )

where C is a positive constant depending only on T and the C 1 -norms of ai and

b in C (P ). Then, (2.21) implies a similar estimate for ut in C (P ). Hence, we
obtain the desired estimate by Theorem 2.10. The proof for for n = 1 is easier and
we will do that first.
Proof for n = 1. In this case, the equation in (2.21) has the form
ut + a(x, t)ux + b(x, t)u = f (x, t).
By differentiating with respect to x, we have
(ux )t + a(ux )x + (b + ax )ux = fx bx u.
We view this as an equation for ux , which has the structure as the equation for u.
Then by Theorem 2.10, we obtain

sup |ux | e1 T
sup |u0x | + T sup |fx bx u| ,
C (P )

C (P )

C (P )

where 1 is a nonnegative constant satisfying

b + ax 1

in C (P ).

Proof for n > 1. Now we consider the general case. Differentiating the equation
in (2.21) with respect to xk , we obtain
(uxk )t +

n
X
i=1

ai (uxk )xi + buxk +

n
X
i=1

ai,xk uxi = fxk bxk u.

2.3. A PRIORI ESTIMATES

29

We note that this is not an equation for uxk as other derivatives of u are present.
To proceed, we set
n
1X 2
v=
uxk .
2
k=1

Then
vt =

n
X

uxk utxk ,

vx i =

k=1

n
X

uxk uxi xk .

k=1

By multiplying the equation above by uxk and summing over k = 1, , n, we

obtain
n
n
n
X
X
X
vt +
ai vxi + 2bv +
ai,xk uxi uxk =
(fxk bxk u)uxk .
i=1

i,k=1

k=1

This is not an equation for v as products of two first order derivatives of u are
present in the fourth term in the left hand side. However, any such product can be
controlled by v. In fact, the Cauchy inequality yields

1
|uxi uxk | u2xi + u2xk .
2
Hence, a summation implies
n
n
X
X
|uxi uxk | n
u2xi = 2nv.
i=1

i,k=1

Then we have
n
n
n
X
X
X
vt +
ai vxi + 2bv 2n
|x ai |L (C (P )) v
(fxk bxk u)uxk .
i=1

i=1

k=1

We apply the Cauchy inequality in a similar way to the expression in the right hand
side. Therefore, we obtain
n
n
n
X
X
1X
vt +
ai vxi + (2b 2n
|x ai |L (C (P )) 1)v
(fxk bxk u)2 .
2
i=1
i=1
k=1

This is a first order partial differential inequality rather than an equation for v. We
cannot apply Theorem 2.10 directly. However, the method in the proof still applies.
We only need an estimate on the upper bound of v since v is nonnegative. Hence,
n
X

T
sup v e1 T
sup v +
sup
(fxk bxk u)2 ,
2 C (P )
C (P )
C (P )
k=1

where 1 is a nonnegative constant such that

n
X
2b 2n
|x ai |L (C (P )) 1 1 .
i=1

With the definition of v, we get the desired estimate for x u easily.

Next, we derive an estimate of the L2 -norm of u in terms of the L2 -norms of

f and u0 . We will do this in another class of domains rather than cones. For fixed
T, t > 0, consider
D,T,t = {(x, t); |x| < t t, 0 < t < T }.

30

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

In other words,
D,T,t = C (0, t) {0 < t < T }.
If T t, then D,T,t = C (0, t).

Figure 2.14. An integral domain.

In the following, we are only interested in the case T < t. We denote by
D,T,t, s D,T,t and + D,T,t the bottom, the side and the top of the boundary,
i.e.,
D,T,t ={(x, 0); |x| < t},
s D,T,t ={(x, t); |x| = t t, 0 < t < T },
+ D,T,t ={(x, T ); |x| < t T }.
We recall the Greens formula. Let be a piecewise C 1 -domain in Rn and
= (1 , , n ) be the unit exterior normal vector to . Then for any u

C 1 () C(),
Z
Z
uxi =
ui for i = 1, , n.

Theorem 2.12. Let ai be C -functions satisfying (2.22), b and f be continuous

functions and u be a C 1 -solution of (2.21) in Rn R+ . Then for any 0 < T < t,
Z
Z
Z
2
t 2
e u
u0 +
et f 2 ,
D,T ,t

D,T ,t

D,T ,t

where is a positive constant depending only on the C 1 -norms of ai and the supnorm of b in D,T,t.
Proof. For a nonnegative , we multiply the equation in (2.21) by 2et u.
By
2et uut =(et u2 )t + et u2 ,
2ai et uuxi =(et ai u2 )xi et ai,xi u2 ,
we have
(et u2 )t +

n
n
X
X
(et ai u2 )xi + et ( + 2b
ai,xi )u2 = 2et uf.
i=1

i=1

2.3. A PRIORI ESTIMATES

31

We simply write D = D,T,t. An integration in D yields

Z
Z
Z
n
n
X
X
t 2
t
2
e u +
e (t +
ai i )u +
et ( + 2b
ai,xi )u2
+ D

s D

i=1

=
D

u20

2e

i=1

uf,

where (1 , , n , t ) is the unit exterior normal vector of s D given by

x
1
,1 .
(1 , , n , t ) =
1 + 2 |x|
First, we note
t +

n
X

ai i 0 on s D.

i=1

Next, we choose such that

+ 2b

n
X

ai,xi 2 in D.

i=1

Then

et u2

2
D

u20 +

2et uf.
D

Here we simply dropped integrals over + D and s D since they are nonnegative.
The Cauchy inequality implies
Z
Z
Z
t
t 2
2e uf
e u +
et f 2 .
D

We then have the desired result.

The domain D,T,t is important for us to discuss properties of solutions in

Rn (0, T ) for a fixed T . We note that t enters the estimate in Theorem 2.12
only through the domain D,T,t. Hence, we may let t to get the following
consequence.
Corollary 2.13. Let ai be C 1 -functions satisfying (2.22), b be continuous
functions and u be a C 1 -solution of (2.21) in Rn R+ . For any T > 0, if f
L2 (Rn (0, T )) and u0 L2 (Rn ), then
Z
Z
Z
et u2
u20 +
et f 2 ,
Rn (0,T )

Rn

Rn (0,T )

where is a positive constant depending only on the C 1 -norms of ai and the supnorm of b in Rn (0, T ).
We point out that there are no decay assumptions on u.
A natural question for anyone who just starts studying PDEs is why we need
a priori estimates. Now we use (2.21) to demonstrate that with the estimate in
Corollary 2.13 we can actually prove the existence of (weak) solutions of initial
value problems (2.21). We use the homogeneous initial value, i.e., u0 = 0, for an
illustration.

32

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

We first introduce the notion of weak solutions. In the following, we fix a T > 0
and consider functions in
DT = Rn (0, T ).
Denote by C0 (DT ) the collection of smooth functions in DT with compact supports
e (DT ) the collection of smooth functions in DT with compact
in DT and by C
0
supports in x-directions. In other words, functions in C0 (DT ) vanish for large x
e (DT ) vanish only for large x.
and for t close to 0 and T and functions in C
0
Set
n
X
(2.23)
Lu = ut +
ai (x, t)uxi + b(x, t)u in DT .
i=1

e , we integrate vLu in DT . To this end, we write

For any u, v C
0
vLu = (uv)t +

n
X

n
X

(ai uv)xi u vt +
(ai v)xi bv .

i=1

i=1

By a simple integration in DT , we have

Z
Z
Z
vLu =
uv
Rn {t=T }

DT

Z
uv

Rn {t=0}

DT

n
X

u vt +
(ai v)xi bv .
i=1

We note that there are no derivatives of u in the right-hand side.

Next, we define the adjoint differential operator L of L by
L v = vt

n
X

(ai v)xi + bv = vt

i=1

Then we have
Z
Z
vLu =
DT

n
X

ai vxi + (b

i=1

ai,xi )v.

i=1

uL v +

uv
Rn {t=T }

DT

n
X

uv
Rn {t=0}

e0 (DT ).
for any u, v C

Definition 2.14. An L2 -function u is a weak solution of Lu = f in DT if

Z
Z

uL v =
f v for any v C0 (DT ).
DT

DT

Now we are ready to prove the existence of weak solutions of (2.21) with homogeneous initial values. The proof requires the Hahn-Banach Theorem and the
Riesz Representation Theorem in functional analysis. In the following, we denote
by (, ) the L2 -inner product in DT .
We first note that, with L in (2.23), we can rewrite the estimate in Corollary
2.13 as

(2.24)
kukL2 (Rn (0,T )) C ku(, 0)kL2 (Rn ) + kLukL2 (Rn (0,T )) ,
where C is a positive constant depending only on T , the C 1 -norms of ai and the
sup-norm of b in Rn (0, T ). This holds for any u C 1 (Rn [0, T )) with
Lu L2 (Rn (0, T )) and u(, 0) L2 (Rn ).
Theorem 2.15. Let ai be C 1 -functions satisfying (2.22) and b a continuous
function in DT . Then for any f L2 (DT ), there exists a u L2 (DT ) such that
(u, L v) = (f, v)

e0 (DT ) with v = 0 on t = T.
for any v C

2.3. A PRIORI ESTIMATES

33

Moreover,
kukL2 (DT ) Ckf kL2 (DT ) ,
where C is a positive constant depending only on T , the C 1 -norms of ai and the
sup-norm of b in Rn (0, T ).
The function u in Theorem 2.15 is called a weak solution of (2.21) with u0 = 0.
Proof. We first note that L has a similar structure as L except different
signs for the principal part, the part involving derivatives. Applying (2.24) to L ,
we have
e0 (DT ) with v = 0 on t = T,
kvkL2 (D ) CkL vkL2 (D ) for any v C
T

where C is a positive constant depending only on T , the C 1 -norms of ai and the

sup-norm of b. Here we view t = T as the initial curve for the domain DT and v
b (DT ) the collection of functions
has a homogeneous initial value. We denote by C
e (DT ) with v = 0 on t = T.
vC
0
b (DT ) R given by
Consider the linear functional F : L C
b (DT ).
F (L v) = (f, v) for any v C
Then
|F (L v)| kf kL2 (DT ) kvkL2 (DT ) Ckf kL2 (DT ) kL vkL2 (DT ) .
b (DT )
Hence F is a well-defined bounded linear functional on the subspace L C
2
of L (DT ). Thus we apply the Hahn-Banach Theorem to obtain a bounded linear
extension of F defined on L2 (DT ) such that
kF k Ckf kL2 (DT ) .
Here, kF k is the norm of the linear functional F . By the Riesz Representation
Theorem, there exists a u L2 (DT ) such that
F (z) = (u, z) for any z L2 (DT ),
and
kukL2 (DT ) Ckf kL2 (DT ) .
b (DT ) to obtain the desired result.
Now we restrict z back to L C

Theorem 2.15 asserts the existence of weak solutions of (2.21) with u0 = 0. Now
we illustrate that u is indeed a classical solution if u is C 1 in DT and continuous
up to t = 0. Under these extra assumptions on u, we integrate by parts (u, L v) to
get
Z
Z
Z
e0 (DT ) with v = 0 on t = T.
vLu +
uv =
f v for any v C
DT

Rn {t=0}

DT

There are no boundary integrals on vertical sides and on the upper side since v
vanishes there. In particular,
Z
Z
vLu =
f v for any v C0 (DT ).
DT

Since

C0 (DT )

DT

is dense in L (DT ), we conclude

Lu = f

in DT .

34

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

Therefore,

Z
Rn {t=0}

or

e0 (DT ) with v = 0 on t = T,
uv = 0 for any v C
Z
u(0) = 0
Rn

for any C0 (Rn ).

Again by the density of C0 (Rn ) in L2 (Rn ), we conclude

u(, 0) = 0

on Rn .

Therefore, a crucial step in passing weak solutions to classical solutions is to improve

the regularity of weak solutions.
Now we summarize the process to establish solutions by using energy estimates.
Step 1. Prove a priori estimates;
Step 2. Prove the existence of weak solutions by methods in functional analysis;
Step 3. Improve regularities of weak solutions.
In fact, the last step is also closely related to a priori estimates of derivatives of
solutions.
Next, we make a crucial observation. The requirement that u possess continuous derivatives can be weakened. It suffices to assume that u has derivatives in
the L2 -sense so that the integration by parts can be performed. Then we conclude
Lu = f almost everywhere. This clearly suggests the need to introduce a new space
of functions, functions with only L2 -derivatives. This is the so-called Sobolev space,
which plays a fundamental role in PDEs and will be studied in details in the future.
Exercises
(1) Find solutions of the following initial-value problems in R2 :
2
(a) 2uy ux + xu = 0 with u(x, 0) = 2xex /2 .
(b) uy + (1 + x2 )ux u = 0 with u(x, 0) = arctan x.
(2) Solve the following initial-value problems:
(a) uy + ux = u2 with u(x, 0) = h(x).
(b) uz + xux + yuy = u with u(x, y, 0) = h(x, y).
1
(3) Let B1 be the unit disc in R2 and a and b be continuous functions in B
1
with a(x, y)x + b(x, y)y > 0 on B1 . Assume u is a C -solution of
1 .
a(x, y)ux + b(x, y)uy = u in B
Prove that u vanishes identically.
(4) Find an equation in R2 of the form a(x, y)ux + b(x, y)uy = 0 with smooth
coefficients such that
(a) any horizontal line {y = c} is non-characteristic;
(b) there does not exist a solution in entire R2 for any nonconstant initial
value prescribed on {y = 0}.
(5) Let be a number and h = h(x) be a continuous function in R. Consider
yux + xuy =u,
u(x, 0) =h(x).

EXERCISES

35

(a) Find all points on {y = 0} where {y = 0} is characteristic. What is

the compatibility condition of h at these points?
(b) Away from points in (a), find the solution of the initial-value problem.
What is the domain of this solution in general?
(c) For cases h(x) = x, = 1 and h(x) = x, = 3, check whether this
solution can be extended over points in (a).
(d) For each point in (a), find all characteristic curves passing it. What
is the relation of these curves and the domain in (b)?
(6) Let R be a real number and h = h(x) be continuous in R and C 1 in
R \ {0}. Consider
xux + yuy =u,
u(x, 0) =h(x).
(a) Check that the straight line {y = 0} is characteristic at each point.
(b) Find all h satisfying the compatibility condition on {y = 0}. (Consider three cases, > 0, = 0 and < 0.)
(c) For > 0, find two solutions with the given initial value on {y = 0}.
(This is easy to do simply by inspecting the equation, especially for
= 2.)
(7) In the plane, solve uy = 4u2x near the origin with u(x, 0) = x2 on {y = 0}.
(8) In the plane, find two solutions of the initial-value problem
1
xux + yuy + (u2x + u2y ) = u,
2
1
u(x, 0) = (1 x2 ).
2
(9) In the plane, find two solutions of the initial-value problem
1 2
u + uuy = u,
4 x
1
1
u(x, x2 ) = x2 .
2
2
(10) Let ai , b and f be continuous functions satisfying (2.22) and u be a C 1 solution of (2.21) in Rn R+ . Prove for any P = (X, T ) Rn R+
sup |et u|
C (P )

sup

|u0 | +

C (P )

1
sup |et f |,
+ inf C (P ) b C (P )

where is a constant such that

+ inf b > 0.
C (P )

(11) Consider the following first-order differential system for (u, v) in R R+

ut a(x, t)vx + b11 (x, t)u + b12 (x, t)v = f1 (x, t),
vt a(x, t)ux + b21 (x, t)u + b22 (x, t)v = f2 (x, t),
with
u(x, 0) = u0 (x),

v(x, 0) = v0 (x),

36

2. FIRST-ORDER DIFFERENTIAL EQUATIONS

where a satisfies

1
.

2
Derive an L -estimate of (u, v) in appropriate cones.
|a(x, t)|

CHAPTER 3

An Overview of Second-Order PDEs

This chapter should be considered as an introduction to second-order linear
PDEs. In Section 3.1, we introduce the Laplace equation, the heat equation and
the wave equation. We also introduce their general forms, elliptic equations, parabolic equations and hyperbolic equations. We use energy methods to discuss
the uniqueness of boundary-value problems and initial/boundary-value problems
in Section 3.2 and use separation of variables to solve these problems in the plane
in Section 3.3.
3.1. Classifications of Second Order Equations
The main focus in this section is second-order linear PDEs. We proceed similarly as in Section 2.1.
Let be a domain in Rn containing the origin and L be a second-order linear
differential operator given by
Lu =

n
X

aij (x)uxi xj +

i,j=1

n
X

bi (x)uxi + c(x)u in ,

i=1

where aij , bi , c are continuous in . Here aij , bi , c are called coefficients of uxi xj ,
uxi , u respectively. We usually assume that (aij ) is a symmetric matrix in . For
the operator L, we define its principal symbol by
p(x; ) =

n
X

aij (x)i j

for any x , Rn .

i,j=1

For a given function f in , we consider the equation

(3.1)

Lu = f (x) in .

The function f is called the non-homogeneous term.

Let be the hyperplane {xn = 0}. We now prescribe values of u and its
derivatives on so that we can at least find all derivatives of u at the origin.
Let u0 , u1 be functions defined in a neighborhood of the origin in Rn1 . Now we
prescribe
(3.2)

u(x0 , 0) = u0 (x0 ), uxn (x0 , 0) = u1 (x0 )

for any small x0 Rn1 .

Then we can find all x0 -derivatives of u and all x0 -derivatives of uxn at the origin.
In particular, we can determine all derivatives of u of order up to 2 at the origin
except uxn xn . To find this, we need to use the equation. If we assume
ann (0) 6= 0,
37

38

3. AN OVERVIEW OF SECOND-ORDER PDES

then we can find uxn xn (0) from (3.1). In this case, we can compute all derivatives
of u of any order at the origin by using values u0 , u1 and differentiating (3.1).
We usually call the initial hypersurface and u0 , u1 initial values or Cauchy
values. The problem of solving (3.1) together with (3.2) is called the initial-value
problem or the Cauchy problem.
More generally, consider the hypersurface given by { = 0} for a C 2 -function
in a neighborhood of the origin with 6= 0, with the origin on the hypersurface
, i.e., (0) = 0. We note that is simply a normal vector of the hypersurface .
Without loss of generality, we assume xn (0) 6= 0. Then by the implicit function
theorem, we solve = 0 around x = 0 for xn = (x1 , , xn1 ). Consider a change
of variables
x 7 y = (x1 , , xn1 , (x)).
This is a well defined transform in a neighborhood of the origin with a nonsingular
Jacobian. Now we write the operator L in new variables y. Note
n
X
uxi =
yk,xi uyk ,
k=1

and
uxi xj =

n
X

yk,xi yl,xj uyk yl +

k,l=1

n
X

yk,xi xj uyk .

k=1

Hence,
Lu =

n
n
X
X
k,l=1

aij yk,xi yl,xj )uyk yl +

i,j=1

n
X

bk +

k=1

n
X

aij yk,xi xj uyk + cu.

i,j=1

The initial hypersurface is given by {yn = 0} in new coordinates. With yn = ,

the coefficient of uyn yn is given by
n
X

aij (x)xi xj .

i,j=1

This is the principal symbol p(x, ) evaluated at = (x).

Definition 3.1. For a linear operator L as in (3.1) defined in a neighborhood
of x0 Rn , a C 1 -hypersurface passing x0 is non-characteristic at x0 if
n
X
(3.3)
aij (x0 )i j 6= 0,
i,j=1

where is a normal vector of at x0 . Otherwise, it is characteristic at x0 . A

hypersurface is non-characteristic if it is non-characteristic at every point.
When the hypersurface is given by { = 0} with 6= 0, its normal vector
is given by = (x1 , , xn ). Hence we may take = (x0 ) in (3.3). We
note that the condition (3.3) is maintained under C 2 -changes of local coordinates.
By this condition, we can find successively values of all derivatives of u at x0 , as far
as they exist. Then, we could write formal power series at x0 for solutions of initial
value problems. It would be actual representations of solutions u in a neighborhood
of x0 , if u were known to be analytic. This process can be carried out for analytic
initial values and analytic coefficients and nonhomogeneous terms, and the result

3.1. CLASSIFICATIONS OF SECOND ORDER EQUATIONS

39

is referred to as the Cauchy-Kowalevski theorem. We will discuss it later in this

book.
Now we introduce a special class of linear differential operators.
Definition 3.2. The linear differential operator L in (3.1) is elliptic at x0 if
n
X

aij (x0 )i j 6= 0

for any Rn \ {0}.

i,j=1

Hence, linear differential equations are elliptic if there are no characteristic

hypersurfaces. In other words, every hypersurface is non-characteristic. We already
assumed that (aij ) is an nn symmetric matrix. Then L is elliptic at x0 if (aij (x0 ))
is a definite matrix, positive definite or negative definite.
We now turn our attention to second-order linear differential equations in R2 ,
where complete classifications are available. Let be a domain in R2 and consider
(3.4)

Lu =

2
X
i,j=1

aij uxi xj +

2
X

bi (x)uxi + c(x)u = f (x)

in .

i=1

Here we assume (aij ) is a 2 2 symmetric matrix.

Definition 3.3. Let L be a differential operator defined in R2 as in (3.4).
(1) L is elliptic at x if det(aij (x)) > 0;
(2) L is hyperbolic at x if det(aij (x)) < 0;
(3) L is degenerate at x if det(aij (x)) = 0.
It is obvious that the ellipticity defined here coincides with that in Definition
3.2 for n = 2.
For the operator L in (3.4), the symmetric matrix (aij ) always has two eigenvalues. Then
L is elliptic if the two eigenvalues have the same sign;
L is hyperbolic if the two eigenvalues have different signs;
L is degenerate if one of the eigenvalues vanishes.
The number of characteristic curves are determined by the type of operators.
For the operator L in (3.4),
there are two characteristic curves if L is hyperbolic;
there is one characteristic curve if L is degenerate;
there are no characteristic curves if L is elliptic.
Now we study several important linear differential operators in R2 .
In R2 , the Laplace operator is give by
u = ux1 x1 + ux2 x2 .
Obviously, the Laplace operator is elliptic. In polar coordinates
x1 = r cos ,

x2 = r sin ,

the Laplace operator can be expressed by

1
1
u = urr + ur + 2 u .
r
r

40

3. AN OVERVIEW OF SECOND-ORDER PDES

The equation
u = 0
is called the Laplace equation and its solutions are called harmonic functions. By
writing x = x1 and y = x2 , we can associate with a harmonic function u(x, y) a
conjugate harmonic function v(x, y) such that u and v satisfy the first order system
of Cauchy-Riemann equations
u x = vy ,

uy = vx .

Any such a pair gives an analytic function

f (z) = u(x, y) + iv(x, y)
of the complex argument z = x + iy. Physically, (u, v) is the velocity field of an
irrotational, incompressible flow. Conversely, for any analytic function f , functions
u = Ref and v = Imf are harmonic. In this way, we can find many nontrivial
harmonic functions in the plane. For example, for any positive integer k, Re(x +
iy)k and Im(x + iy)k are homogeneous harmonic polynomials of degree k. Next,
with ez = ex+iy = ex cos y + iex sin y, we know ex cos y and ex sin y are harmonic
functions.
Although there are no characteristic curves for the Laplace operator, initial
value problems are not well-posed.
Example 3.4. Consider the Laplace equation in R2
uxx + uyy = 0.
Set
uk (x, y) =
Then u is harmonic. Moreover,

1
sin(kx)eky .
k

uk,x (x, y) = cos(kx)eky ,

uk,y (x, y) = sin(kx)eky ,

and hence
u2k,x (x, y) + u2k,y (x, y) = e2ky .
Therefore,
u2k,x (x, 0) + u2k,y (x, 0) = 1

for any x R and any k,

and
u2k,x (x, y) + u2k,y (x, y)

as k , for any x R and y > 0.

There is no continuous dependence on initial values.

In R2 , the wave operator is give by
u = ux2 x2 ux1 x1 .
Obviously, the wave operator is hyperbolic. It is actually called the one-dimensional
wave operator. This is because the wave equation u = 0 in R2 represents vibrations of strings or propagation of sound waves in tubes. Because of its physical
interpretation, we write u as a function of two independent variables x and t. The
variable x is commonly identified with position and t with time. Then the wave
equation in R2 has the form
utt uxx = 0.

3.1. CLASSIFICATIONS OF SECOND ORDER EQUATIONS

41

Two families of straight lines t = x+constant are characteristic everywhere.

In R2 , an important class of degenerate operators is given by
Lu = ux2 ux1 x1 .
This is called the heat operator. The heat equation ux2 ux1 x1 = 0 is satisfied by
the temperature distribution in a heat-conducting insulated wire. As for the wave
equation, we write u as a function of two independent variables x and t. Then the
heat equation in R2 has the form
ut uxx = 0.
Note that {t = 0} is characteristic everywhere. If we prescribe u(x, 0) = u0 (x)
in an interval of {t = 0}, then using the equation we can compute all derivatives
there. However, u0 does not determine uniquely a solution even in a neighborhood
of this interval. We will see later on that we need initial values on the entire initial
line {t = 0} to compute local solutions. Hence, the problem with Cauchy values
prescribed on characteristic hypersurfaces is not well-posed.
Definition 3.5. Let L be a differential operator defined in R2 as in (3.4).
Then L is of mixed-type if L is elliptic in a subdomain in and hyperbolic in
another subdomain.
Example 3.6. Consider the Tricomi equation
ux2 x2 + x2 ux1 x1 = f

in R2 .

It is elliptic if x2 > 0, hyperbolic if x2 < 0 and degenerate if x2 = 0.

Characteristic curves arise also naturally in connection with the propagation
of singularities. We only consider a simple case.
Theorem 3.7. Let be a domain in R2 and be a curve in . Assume
aij , bi , c, f are continuous functions in and u C 1 () C 2 ( \ ) satisfies
2
X

aij uxi xj +

i,j=1

2
X

bi (x)uxi + c(x)u = f (x)

in \ .

i=1

If 2 u has a jump across , then is a characteristic curve.

Proof. Without loss of generality, we assume divides into two parts +
and . For any function w in and any x0 , we set
w (x0 ) =

lim

xx0 , x

w(x),

w+ (x0 ) =

lim

xx0 , x+

w(x),

and
[w] = w+ w

on .

Then
[u] = [ux1 ] = [ux2 ] = 0 on .
Let (1 , 2 ) be a normal vector along . Then 2 x1 1 x2 is a directional derivative
along . Hence on
(2 x1 1 x2 )[ux1 ] = 2 [ux1 x1 ] 1 [ux1 x2 ] = 0,
(2 x1 1 x2 )[ux2 ] = 2 [ux1 x2 ] 1 [ux2 x2 ] = 0.

42

3. AN OVERVIEW OF SECOND-ORDER PDES

Rigorously, we need a limiting process to justify these two identities. By the continuity of aij , bi , c and f in , we have
a11 [ux1 x2 ] + 2a12 [ux1 x2 ] + a22 [ux2 x2 ] = 0

on .

The nontrivial vector ([ux1 x1 ], [ux1 x2 ], [ux2 x2 ]) satisfies a 3 3 homogeneous linear

system on . Hence the coefficient matrix is singular. We then have on

2 1
0
det 0
2
1 = 0,
a11 2a12 a22
or
a11 12 + 2a12 1 2 + a22 22 = 0.
This yields the desired result.

The Laplace operator, the wave operator and the heat operator can be generalized to higher dimensions.
Example 3.8. The n-dimensional Laplace operator in Rn is defined by
n
X
u =
uxi xi ,
i=1

and the Laplace equation is given by u = 0. Solutions are called harmonic functions. In this case, (aij ) is the identity matrix. Obviously, is elliptic. Note that
is invariant under rotations. If x = Ay for an orthogonal matrix A, then
n
n
X
X
uxi xi =
uyi yi .
i=1

i=1

For nonzero function f , we call equation u = f the Poisson equation.

Example 3.9. We denote points in Rn+1 by (x1 , , xn , t). The heat operator
in Rn+1 is given by
Lu = ut x u.
It is often called the n-dimensional heat operator. Its principal symbol is given by
p(x, t; , ) = ||2

for any Rn , R.

A hypersurface {(x1 , , xn , t) = 0} is non-characteristic if

|x |2 6= 0.
Likewise, a hypersurface {(x, t) = 0} is characteristic everywhere if x = 0 and
t 6= 0. For example, any horizontal hyperplane {t = t0 }, for a fixed t0 R, is
characteristic everywhere.
Example 3.10. We denote points in Rn+1 by (x1 , , xn , t). The wave operator in Rn+1 is given by
u = utt x u.
It is often called the n-dimensional wave operator. Its principal symbol is given by
p(x, t; , ) = 2 ||2

for any Rn , R.

A hypersurface {(x1 , , xn , t) = 0} is non-characteristic if

2t |x |2 6= 0.

3.2. ENERGY ESTIMATES

43

For any (x0 , t0 ) Rn R, the surface

|x x0 |2 = (t t0 )2
is characteristic everywhere except at (x0 , t0 ). We note that this surface, smooth
except at (x0 , t0 ), is the boundary of two cones.

Figure 3.1. The characteristic surface.

The heat equation and the wave equation can be generalized to parabolic equations and hyperbolic equations in arbitrary dimensions. Again, we denote points
in Rn+1 by (x1 , , xn , t). Let aij , bi , c and f be functions defined in a domain
Rn+1 , i, j = 1, , n. We assume (aij (x, t)) is an n n positive definite
matrix in . A parabolic equation in has the form
n
n
X
X
ut
aij (x, t)uxi xj +
bi (x, t)uxi + c(x, t)u = f (x, t) in ,
i=1

i=1

and a hyperbolic equation has the form

n
n
X
X
utt
aij (x, t)uxi xj +
bi (x, t)uxi + c(x, t)u = f (x, t) in .
i=1

i=1

3.2. Energy Estimates

In this section, we discuss the uniqueness of solutions of boundary-value problems for the Laplace equation and initial/boundary-value problems for the heat
equation and the wave equation. Our main tool is the energy estimates. Specifically, we derive estimates of L2 -norms of solutions in terms of those of boundary
values and/or initial values.
We first recall Greens formulas. Let be a C 1 -domain in Rn and =
(1 , , n ) be the unit exterior normal vector to . Then for any u C 1 ()

C(),
Z
Z
uxi =
ui for i = 1, , n.

and
This is referred to as the divergence theorem. Now for any w C 2 () C 1 ()
1

v C () C(), substitute u above by vwxi to get

Z
Z
(vwxi xi + vxi wxi ) =
vwxi i .

By summing up for i = 1, , n, we get the Greens formula of the following form

Z
Z

w
vw + v w =
v
,

44

3. AN OVERVIEW OF SECOND-ORDER PDES

w
where
is the normal derivative of w on . If v 1, we get the divergence
n
theorem
Z
Z
w
w =
.

n
we have another form of the Greens formula
For any v, w C 2 () C 1 (),
Z
Z

v
w
w ).
vw wv =
(v
n
n

Figure 3.2. A smooth domain in Rn .

Now we consider the Laplace equation. Assume Rn is a bounded C 1 domain. The Dirichlet problem for the Laplace equation has the following form
u =0

in ,

u = on ,
for a continuous function on . We now prove that a C 2 -solution, if exists, is
unique. In fact, the difference w of any two C 2 -solutions satisfies
w =0 in ,
w =0 on .
We multiply the homogeneous Laplace equation by w and integrate in . By the
Greens formula, we have
Z
Z
Z
w
2
0=
ww =
|w| +
w
.
n

With the homogeneous boundary value, we have

Z
|w|2 = 0,

or w = 0 in . Hence w is constant and this constant is zero since w is zero

on the boundary. Obviously, the above argument applies to Dirichlet problems for
Poisson equations u = f . In general, we have the following result.
Lemma 3.11. Assume Rn is a bounded C 1 -domain, f is a continuous
and is a continuous function on . Then
function in
u =f
u =

in ,
on ,

admits at most one solution in C 2 () C 1 ().

3.2. ENERGY ESTIMATES

45

We
By the maximum principle, the solution is in fact unique in C 2 () C().
will discuss this in Chapter 4.
For the Neumann problem, we prescribe normal derivatives on the boundary.
It has the following form
u =0

in ,

u
= on ,
n
for a continuous function on . Similarly, we can prove that solutions are unique
up to additive constants. We note that if there exists a solution of the Neumann
problem, then necessarily satisfies
Z
= 0.

This can be seen easily by the Greens formula.

Next, we derive an estimate of solutions of Dirichlet boundary value problems
of the Poisson equation. We need the following result, which is referred to as the
Poincar`e Lemma.
Lemma 3.12. Let be a bounded C 1 -domain in Rn and u be a C 1 -function in
with u = 0 on . Then
kukL2 () diam()kukL2 () .
Here diam() denotes the diameter of and is defined by
diam() = sup |x y|.
x,y
n1

x00

R
, let lx00 be the straight line through x0 and normal
Proof. For any
n1
to R
{0}. Consider those x00 such that lx00 6= . Let I be an open interval
0
on lx0 given by
I = {(x00 , xn ); a < xn < b}
such that I and (x00 , a), (x00 , b) . Since u(x00 , a) = 0, then

Figure 3.3. An integration along lx00 .

Z
u(x00 , xn )

=
a

xn

uxn (x00 , s)ds

for any x (a, b).

46

3. AN OVERVIEW OF SECOND-ORDER PDES

The Holder inequality yields

u2 (x00 , xn ) (xn a)

xn
a

u2xn (x00 , s)ds for any x (a, b).

By a simple integration along I, we have

Z b
Z
2 0
2
u (x0 , xn )dxn (b a)
a

u2xn (x00 , xn )dxn .

By adding all possible intervals like I on lx00 , we then obtain

Z
Z
2 0
2
u (x0 , xn )dxn Cx00
u2xn (x00 , xn )dxn ,
lx0

lx0

where Cx00 is the length of lx00 in . Now a simple integration over x00 yields the
desired result.

Now consider
u =f

(3.5)

in ,

u =0 on .

We note that u has a homogeneous Dirichlet boundary value on .

Theorem 3.13. Let Rn be a bounded C 1 -domain and u C 2 () C 1 ()

be a solution of (3.5). Then
kukL2 () + kukL2 () Ckf kL2 () ,
where C is a positive constant depending only on .
Proof. Multiplying the equation in (3.5) by u and integrating over , we
obtain
Z
Z
|u|2 =
uf.

The Holder inequality yields

Z
Z
Z
( |u|2 )2
u2
f 2.

By Lemma 3.12, we get

2
|u|2 diam()

Z
f 2.

With Lemma 3.12 again, we have the desired estimate.

Now we study initial/boundary-value problems for the heat equation. Suppose

is a bounded C 1 - domain in Rn , f is continuous in [0,

) and u0 is continuous
Consider
in .
ut u =f in (0, ),
(3.6)

u(, 0) =u0

in ,

u =0 on (0, ).
The geometric boundary of (0, ) consists of two parts, {0} and
(0, ). We treat {0} and (0, ) as the initial hypersurface and as
the boundary respectively. We refer to values prescribed on these two parts as
initial values and boundary values respectively. Problems of this type are usually

3.2. ENERGY ESTIMATES

47

called initial/boundary value problems or mixed problems. We note that u has a

homogeneous Dirichlet boundary value on (0, ). We derive an estimate of
the L2 -norms of solutions. In the following, we write u(t) = u(, t) as a function in
for any t 0.
[0, )) be a solution of (3.6).
Theorem 3.14. Let u C 2 ( (0, )) C(
Then
Z t
kf (s)kL2 () ds for any t > 0.
ku(t)kL2 () ku0 kL2 () +
0

Theorem 3.14 yields the uniqueness of solutions of (3.6). In fact, if f 0 and

u0 0, then u 0. We also have the continuous dependence on initial values.
Proof. Multiply the equation in (3.6) by u and integrate over for each fixed
t > 0. Note
n
X
1
uut uu = (u2 )t
(uuxi )xi + |u|2 .
2
i=1
With u(t) = 0 on , we have
Z
Z
Z
1 d
u2 (t) +
|u(t)|2 =
f (t)u(t).
2 dt

Then for any t > 0

Z
|u(t)|2 + 2

Z tZ

Z
|u|2 =

u20 + 2

Z tZ
f u.
0

Set
E(t) = ku(t)kL2 () .
Then

2
E(t) + 2

Z tZ
0

2
|u| = E(0) + 2
2

f u.
0

Differentiating with respect to t, we have

0

Z tZ

2E(t)E (t) 2E(t)E (t) + 2

|u(t)| = 2

f (t)u(t)

2kf (t)kL2 () ku(t)kL2 () .

Hence
E 0 (t) kf (t)kL2 () .
Integrating from 0 to t, we obtain the desired estimate.

Now we study initial/boundary-value problems for the wave equation. Again

suppose is a bounded C 1 -domain in Rn , f is continuous in (0, ), u0 is C 1
in and u1 is continuous in . Consider
utt u = f
(3.7)

in (0, ),

u(, 0) = u0 , ut (, ) = u1
u=0

on (0, ).

in ,

48

3. AN OVERVIEW OF SECOND-ORDER PDES

[0, )) be a solution of (3.7).

Theorem 3.15. Let u C 2 ( (0, )) C 1 (
Then
Z t
1

kut (t)k2L2 () +kx u(t)k2L2 () 2 ku1 k2L2 () +kx u0 k2L2 () 2 +

kf (s)kL2 () ds,
0

and

1
ku(t)kL2 () ku0 kL2 () + t ku1 k2L2 () + kx u0 k2L2 () 2 +

Z
0

(t s)kf (s)kL2 () ds.

Hence we also have the uniqueness and continuous dependence on initial values.
Proof. Multiply the equation by ut and integrate over for each fixed t > 0.
Note
n
X
1
ut utt ut u = (u2t + |u|2 )t
(ut uxi )xi .
2
i=1
Then

Z
Z
Z

u
1 d
ut (t) (t) =
|ut (t)|2 + |x u(t)|2
f (t)ut (t).
2 dt
n

Note ut = 0 on (0, ) since u = 0 on (0, ). Then

Z
Z

1 d
2
2
|ut (t)| + |x u(t)| =
f (t)ut (t).
2 dt

Define the energy by

Z

E(t) =
|ut (t)|2 + |x u(t)|2 .

If f 0, then

d
E(t) = 0.
dt
Hence
Z

1
E(t) = E(0) =
|u1 |2 + |x u0 |2 .
2
This is the conservation of energy. In general,
Z tZ
E(t) = E(0) + 2
f ut .
0

To get an estimate on the energy E(t), set

1
J(t) = E(t) 2 .
Then

2
2
J(t) = J(0) + 2

Z tZ
f ut .
0

Differentiating with respect to t, we get

Z
2J(t)J 0 (t) = 2
f (t)ut (t) 2kf (t)kL2 () kut (t)kL2 () 2J(t)kf (t)kL2 () .

Hence
J 0 (t) kf (t)kL2 () .
Integrating from 0 to t, we obtain
Z t
J(t) J(0) +
kf (s)kL2 () ds.
0

3.3. SEPARATION OF VARIABLES

49

This is the desired estimate on the energy. Next, to estimate the L2 -norm of u, we
set
F (t) = ku(t)kL2 () ,
i.e.,

2
F (t) =

Z
|u(t)|2 .

A simple differentiation yields

Z
0
u(t)ut (t) 2ku(t)kL2 () kut (t)kL2 () .
2F (t)F (t) = 2

Hence

Z
F 0 (t) kut kL2 () J(0) +

kf (s)kL2 () ds.

Integrating from 0 to t, we get

Z tZ
ku(t)kL2 () ku0 kL2 () + tJ(0) +

t0

kf (s)kL2 () dsdt0 .

This implies the desired estimate on u itself easily.

There are other forms of estimates on energies. By squaring the first estimate
in Theorem 3.15, we obtain
Z
Z
Z tZ

|ut (t)|2 + |x u(t)|2 2

|u1 |2 + |x u0 |2 + 2t
|f |2 .

Integrating from 0 to t, we get

Z tZ
Z
Z tZ
2

|ut | + |x u|2 2t
|u1 |2 + |x u0 |2 + t2
|f |2 .
0

To end this section, we briefly review methods used in deriving estimates in

Theorems 3.13-3.15. In the proof of Theorems 3.13-3.14, we multiply the Laplace
equation and the heat equation by u and integrate over . While in the proof of
Theorem 3.15, we multiply the wave equation by ut and integrate over . These
strategies also work for general elliptic equations, parabolic equations and hyperbolic equations.
3.3. Separation of Variables
In this section, we use the separation of variables to solve initial/boundaryvalue problems of the heat equation and the wave equation in R R+ .
We first discuss initial/boundary-value problems for the 1-dimensional heat
equation of the following form
ut uxx = 0 for any x (0, ) and any t > 0,
(3.8)

u(x, 0) = u0 (x)

for any x (0, ),

u(0, t) = u(, t) = 0

for any t > 0.

Physically, u represents the temperature in an insulated rod with ends kept at 0

temperature.

50

3. AN OVERVIEW OF SECOND-ORDER PDES

We now use separation of variables to find solutions of

ut uxx = 0

for any x (0, ) and any t > 0,

u(0, t) = u(, t) = 0

for any t > 0.

Set
u(x, t) = a(t)w(x).
Then
a0 (t)w(x) a(t)w00 (x) = 0,
or

a0 (t) w00 (x)

= 0.
a(t)
w(x)
The first term is a function of t and the second term is a function of x. Since their
difference is zero, they should be equal to a constant. We denote this common
constant by . Then
a0 (t) + a(t) = 0

for any t > 0,

and
w00 (x) + w(x) = 0

for any x (0, ),

w(0) = w() = 0.
d2
in (0, ). It is well
dx2
2
known that k = k and wk (x) = sin kx are a sequence of solutions and {wk } forms
an orthogonal basis for Fourier series in L2 (0, ). For any v L2 (0, ), we have

The latter is the homogeneous eigenvalue problem of

v(x) =

ak sin kx,

k=1

where

2
ak =

v(x) sin kxdx.

0

The convergence for the series of v is in the L2 -sense. By the completeness of

{sin kx} in L2 (0, ), we have
kvkL2 (0,) =

X
1
a2k 2 .
2
k=1

Back to a(t), we have for each = k

2
2

a(t) = aek t ,
where a is a constant. Now we set for each integer k 1
2

uk (x, t) = ek t sin kx

for any (x, t) (0, ) (0, ).

Then uk satisfies the heat equation and the boundary value in (3.8). In order to
get a solution satisfying the equation, the boundary value and the initial value
in (3.8), we consider an infinite linear combination of uk and choose coefficients
appropriately.

3.3. SEPARATION OF VARIABLES

51

Now we are ready to solve the initial/boundary-value problem (3.8). We treat

t as a parameter and expand u(x, t) as a function of x (0, ) in its Fourier series
with respect to {sin kx}. Formally, we write
u(x, t) =

ak (t) sin kx.

k=1

Then
ut uxx =

X
0

ak (t) + k 2 ak (t) sin kx.

k=1

With
a0k (t) + k 2 ak (t) = 0

for any k = 1, 2, ,

we have
ak (t) = ak (0)ek
and hence
u(x, t) =

for any k = 1, 2, ,
2

ak (0)ek t sin kx.

k=1

Let
u0 (x) =

u0k sin kx,

k=1

where

Then ak (0) = u0k

(3.9)

Z
2
u0k =
u0 (x) sin kxdx.
0
and the solution u is formally given by
u(x, t) =

u0k ek t sin kx.

k=1

Next we prove that u in (3.9) indeed solves (3.8).

Theorem 3.16. If u0 L2 (0, ), then u given by (3.9) is smooth in [0, ]
(0, ) and satisfies
ut uxx = 0

for any x (0, ) and any t > 0,

u(0, t) = u(, t) = 0

for any t > 0,

and
lim ku(, t) u0 kL2 (0,) = 0.

t0+

Proof. By u0 L2 (0, ), we have

ku0 k2L2 (0,) =

X
|u0k |2 < .
2
k=1

For any nonnegative integers i and j, any x [0, ] and t (0, ), we have
|xi tj u(x, t)|

X
k=1

|u0k |

k i+2j
.
ek 2 t

52

3. AN OVERVIEW OF SECOND-ORDER PDES

Hence the Cauchy inequality implies

|xi tj u(x, t)|

|u0k |2

12 X
k 2i+4j 12

k=1

k=1

e2k2 t

Ci,j,t ku0 kL2 (0,) ,

where Ci,j,t is a positive constant depending only on i, j and t. This shows the
absolute convergence of series for xi tj u(x, t), for any x [0, ], t > 0 and any
nonnegative integers i and j. Hence u is smooth in [0, ] (0, ). The proof for
the convergence as t 0+ is left as an exercise.

By examining the proof, we have the following estimate. For any integer m 0
and any t0 > 0
kukC m ([0,](t0 ,)) Cku0 kL2 (0,) ,
where C is a positive constant depending only on m and t0 . This estimate controls
the C m -norm of u in (0, ) (t0 , ) for any t0 > 0 in terms of the L2 -norm of u0
on (0, ). It is referred to as an interior estimate (with respect to t). We note that
u becomes smooth immediately after t = 0 even if the initial value u0 is only L2 .
Naturally, we ask whether u in Theorem 3.16 is continuous up to t = 0, or
generally whether u is smooth up to t = 0. In order to have continuity of u up to
t = 0, first we should require u0 C[0, ]. Next, u0 should satisfy a compatibility
condition. By comparing the initial value with the homogeneous boundary value
at the corner, we have
u0 (0) = 0, u0 () = 0.
In general, in order to have smoothness of u up to t = 0, we assume u0 C [0, ].
Then from the heat equation, we have uxx (0, 0) = 0 and uxx (, 0) = 0. Hence
u000 (0) = 0,

u000 () = 0.

Continuing this process, we have a necessary condition

(3.10)

(2`)

u0

(0) = 0,

(2`)

u0

() = 0

for any ` = 0, 1, .

Now, we prove this is also a sufficient condition.

Theorem 3.17. If u0 C [0, ] and (3.10) holds, then u given by (3.9) is
smooth in [0, ] [0, ) with u(, 0) = u0 .
Proof. In order to prove the smoothness up to t = 0, we need to bound
xi tj u(x, t) independent of t for t small, for any nonnegative integers i and j. To this
end, we first improve estimates of u0k . With (3.10), we have by simple integrations
by parts
Z
Z
Z
2
2 0
cos kx
2 00
sin kx
u0k =
u0 (x) sin kxdx =
u (x)
dx =
u (x) 2 dx.
0
0 0
k
0 0
k
We note that values at end points are not present since u0 (0) = u0 () = 0 in
the first integration by parts and sin kx = 0 at x = 0 and x = in the second
integration by parts. Hence for any m 1, we continue this process with the help
of (3.10) for ` = 0, , [(m 1)/2] and obtain
Z
m1 2
cos kx
(m)
2
u0k = (1)
u (x) m dx if m is odd,
0 0
k

3.3. SEPARATION OF VARIABLES

53

and

Z
2 (m)
sin kx
u0 (x) m dx if m is even.
0
k
By the Cauchy inequality, we get for any m 0
r
2 (m)
1
|u0k |
ku0 kL2 (0,) m for any k = 1, 2, ,

k
where we used
Z
Z

sin2 kxdx =
cos2 kxdx = .
2
0
0
Now for any x [0, ] and t > 0
r

X
X
2 (m)
k i+2j
i j
i+2j
|x t (x, t)|
ku0 kL2 (0,)
< ,
|u0k |k

km
j=1
m

u0k = (1) 2

k=1

which is independent of t, if we take m = i + 2j + 2. This implies that xi tj (x, t) is

continuous in [0, ] [0, ).

If we are only interested in the continuity of u up to t = 0, we have the following

result.
Corollary 3.18. If u0 C 1 [0, ] with u000 L2 (0, ) and u0 (0) = u0 () = 0,
then u given by (3.9) is smooth in [0, ] (0, ) and continuous in [0, ] [0, )
and satisfies (3.8).
Proof. It follows from Theorem 3.16 that u is smooth in [0, ] (0, ) and
satisfies the heat equation and the homogeneous boundary condition in (3.8). The
continuity of u up to t = 0 follows from the proof of Theorem 3.17 with i = j = 0
and m = 2.

The regularity assumption on u0 in Corollary 3.18 does not seem to be optimal.

It is natural to ask whether it suffices to assume u0 C[0, ]. To answer this
question, we should analyze whether the series in (3.9) converges uniformly in t
since we need to take limit t 0. We will not pursue this issue here.
Now we provide another expression of u in (3.9). With explicit expressions of
u0k in terms of u0 , we can write
Z
(3.11)
u(x, t) =
G(x, y; t)u0 (y)dy,
0

where
G(x, y; t) =

2 X k2 t
e
sin kx sin ky

for any x, y [0, ] and t > 0.

k=1

The function G is called the Greens function associated with the initial-boundary
value problem (3.8). For each fixed t, the series for G is uniformly convergent for
x, y [0, ]. In fact, this uniform convergence justifies the exchange of summation
and integration in obtaining (3.11). The Greens function G satisfies the following
properties:
(G1) Symmetry: G(x, y; t) = G(y, x; t).
(G2) Smoothness: G(x, y; t) is smooth in x, y [0, ] and t > 0.
(G3) Solution of the heat equation: Gt Gxx = 0.

54

3. AN OVERVIEW OF SECOND-ORDER PDES

(G4) Homogeneous boundary values: G(0, y; t) = G(, y; t) = 0.

These properties follow easily from the explicit expression for G. They imply
that u in (3.11) is a smooth function for t > 0 and satisfies the heat equation
with homogeneous boundary values. We can also prove with help of the explicit
expression of G that u in (3.11) is continuous up to t = 0 and satisfies u(, 0) = u0
under appropriate assumptions on u0 . We point out that G can also be expressed
in terms of the fundamental solution of the heat equation. See Chapter 5 for
discussions of the fundamental solution.
Next, we discuss initial/boundary-value problems for the 1-dimensional wave
equation of the following form
utt uxx = 0 for any x (0, ) and any t > 0,
(3.12)

u(x, 0) = u0 (x), ut (x, 0) = u1 (x) for any x (0, ),

u(0, t) = u(, t) = 0

for any t > 0.

We proceed as for the heat equation. Set

u(x, t) =

Tk (t) sin kx.

k=1

Then
utt uxx =

Tk00 (t) + k 2 Tk (t) sin kx.

k=1

With
Tk00 (t) + k 2 Tk (t) = 0

for any k = 1, 2, ,

we have
Tk (t) = ak sin kt + bk cos kt for any k = 1, 2, .
Therefore,
(3.13)

u(x, t) =

(ak sin kt + bk cos kt) sin kx.

k=1

If we write
u0 (x) =

u0k sin kx,

u1 (x) =

k=1

u1k sin kx,

k=1

we can determine ak , bk from u0 and u1 by

Z
u1k
2
ak =
=
u1 (x) sin kxdx,
k
k 0
Z
2
u0 (x) sin kxdx.
bk = u0k =
0
Unlike the case in the heat equation, in order to get regularity now, we need to
impose similar regularity conditions on initial values. Preceding as for the heat
equation, we note that if u is a C 2 -solution, then
(3.14)

u0 (0) = 0,

u1 (0) = 0,

u000 (0) = 0,

u0 () = 0,

u1 () = 0,

u000 () = 0.

3.3. SEPARATION OF VARIABLES

55

2
1
00
2
Theorem 3.19. If u0 C 2 [0, ], u000
0 L (0, ), u1 C [0, ], u1 L (0, )
2
and u0 , u1 satisfy (3.14), then u given by (3.13) is C in [0, ] [0, ) and is a
solution of (3.12).

Proof. We first improve estimates for ak and bk . By (3.14) and integration

by parts, we have
Z
Z
2
2
ak =
u1 (x) sin kxdx = 3
u00 (x) sin kxdx,
k 0
k 0 1
Z
Z
2
2
u0 (x) sin kxdx = 3
u000 (x) cos kxdx.
bk =
0
k 0 0
In other words, {k 3 ak } is the Fourier coefficients of u001 (x) with respect to {sin kx}
and {k 3 bk } is the Fourier coefficients of u000
0 (x) with respect to {1, cos kx}. Hence

(k 6 a2k + k 6 b2k ) < .

k=1

Note for any integers i, j with 0 i + j 2 and any x [0, ] and t > 0
|xi tj u(x, t)|

k i+j (|ak | + |bk |),

k=1

and hence
|xi tj u(x, t)|

2(i+j+1)

12 X
1 12

< .
(|ak | + |bk |)
k2
2

k=1

k=1

Therefore, u is C 2 in [0, ] [0, ). Then u satisfies the wave equation and the
initial condition by a simple differentiation term by term.

By examining the proof, we have

kukC 2 ([0,](0,)) C

3
X
i=0

(i)

ku0 kL2 (0,) +

2
X

(i)
ku1 kL2 (0,) ,

i=0

where C is a positive constant independent of u.

In fact, in order to get a C 2 -solution of (3.12), it suffices to assume u0 C 2 [0, ]
and u1 C 1 [0, ]. We will prove this for a more general initial/boundary-value
problem of the wave equation in Section 6.1. See Theorem 6.3.
Now, we compare the regularity results in Theorems 3.16, 3.17 and 3.19. For
initial/boundary-value problems of the heat equation in Theorem 3.16, solutions
become smooth immediately after t = 0, even for L2 -initial values. This is the
so-called interior smoothness (with respect to time). We also proved in Theorem
3.17 that solutions are smooth up to t = 0 if initial values are also smooth with a
compatibility condition. This property is called the global smoothness. However,
solutions of initial/boundary-value problems of the wave equation exhibit a different property. We proved in Theorem 3.19 that solutions have a similar degree of
regularity as initial values. In fact, this is a general result. Solutions of the wave
equation do not have better regularity than initial values, and in some cases they
are less regular than initial values. We will see in Chapter 6 how solutions of the
wave equation depend on initial values.

56

3. AN OVERVIEW OF SECOND-ORDER PDES

To end this section, we point out that methods employed in this section to
solve initial/boundary-value problems for the one-dimensional heat equation and
wave equation can actually be generalized to higher dimensions. We use the heat
equation as an illustration. Let be a bounded smooth domain in Rn . We consider
ut u = 0
u(, 0) = u0

(3.15)

u=0

in (0, ),
in ,
on (0, ),

To solve (3.15) by separation of variables, we

for a continuous function u0 in .
need solutions of the eigenvalue problem of in with homogeneous boundary
values, i.e.,
+ = 0

(3.16)

=0

in ,
on .

Of course, this is much harder to solve than its one-dimensional counterpart. Nevertheless, a similar result still holds. In fact, solutions of (3.16) are given by a
sequence (k , k ) where k is a sequence of nondecreasing positive numbers such
which forms
that k as k and k is a sequence of smooth functions in
2
a basis in L (). Then we can use a similar method to find a solution of (3.15) of
the form

X
u(x, t) =
ak ek t k (x) for any (x, t) (0, ).
k=1

We should remark that solving (3.16) is a complicated process. We need to work in

Sobolev spaces, spaces of functions with L2 -integrable derivatives. The process is
similar as that indicated at the end of Chapter 2. First, we prove the existence of
weak solutions of (3.16) and then improve regularity of solutions. A brief discussion
can be found in Section 4.6.
Exercises
(1) Classify the following second-order PDEs:
n
X
X
(a)
uxi xi +
uxi xj = 0.
i=1

(b)

1i<jn

uxi xj = 0.

1i<jn

(2) Discuss
( the uniqueness of the following problems by energy methods:
u u3 = f in ,
(a)
u=
on .
(
R 2
u u u (y)dy = f in ,
(b)
u=
on .

EXERCISES

57

[0, T ]
(3) Let be a bounded C 1 -domain in Rn and u be a C 2 -function in
satisfying
ut u =f

in (0, ),

u(, 0) =u0

in ,

u =0 on (0, ).
Prove

0tT

|u(, t)|2 +

sup

u2t C

Z
|u0 |2 +

f2 ,

where C is a positive constant independent of u, u0 and f .

(4) Prove the convergence part in Theorem 3.16.
(5) For u0 L2 (0, ), let u be given in (3.9). Then for any nonnegative
integers i and j
sup |xi tj u(, t)| 0

as t .

[0,]

(6) Let G be defined in (3.11). Prove

1
for any x, y [0, ] and t > 0.
|G(x, y; t)|
t
(7) Solve the following problem by separation of variables
ut uxx = 0

for any 0 < x < , t > 0,

u(x, 0) = u0 (x)

for any 0 < x < ,

ux (0, t) = ux (, t) = 0

for any t > 0.

(8) For any u0 L2 (0, ) and f L2 ((0, ) (0, )), find a formal explicit
expression of a solution of the following problem
ut uxx = f

for any x (0, ) and any t > 0,

u(x, 0) = u0 (x) for any x (0, ),

u(0, t) = u(, t) = 0

for any t > 0.

(9) For any u0 , u1 L2 (0, ) and f L2 ((0, )(0, )), find a formal explicit
expression of a solution of the following problem
utt uxx = f

for any x (0, ) and any t > 0,

u(x, 0) = u0 (x), ut (x, 0) = u1 (x) for any x (0, ),

u(0, t) = u(, t) = 0

for any t > 0.

(10) Let be a bounded C 1 -domain in Rn and u be C 2 in x and C 1 in t in

[0, T ] for a constant T > 0. Suppose u satisfies

ut u = 0
u(, T ) = 0
u=0

in (0, T ),
in ,
on (0, T ).

58

3. AN OVERVIEW OF SECOND-ORDER PDES

Then u = 0 in (0, T ).
R
Hint: The function J(t) = log u2 (, t) is a decreasing convex function.

CHAPTER 4

Laplace Equations
The Laplace equation u = 0 is probably the most important PDE with the
widest range of applications. In Section 4.1, we solve Dirichlet problems for the
Laplace equation in balls and derive Poisson integral formula. In the next three
sections, we study harmonic functions, solutions of the Laplace equation, by three
different methods: fundamental solutions, mean-value properties and the maximum
principle. These three sections are relatively independent of each other. In Section
4.5, we study the Poisson equation u = f and derive Schauder estimates for its
solutions. Then in Section 4.6, we briefly discuss weak solutions of the Poisson
equation.
4.1. Dirichlet Problems
The Laplace operator is defined on a C 2 -function u in a domain in Rn by
u =

n
X

uxi xi .

i=1

The equation u = 0 is called the Laplace equation and its C 2 -solutions are called
harmonic functions.
One of important properties of the Laplace equation is its spherical symmetry.
The equation is preserved by rotations about some point a Rn . Hence, it is
plausible that there exist special solutions that are invariant under rotations about
a. We begin this section by seeking a harmonic function u in Rn which depends
only on r = |x a| for some fixed a Rn . By setting v(r) = u(x), we have
u = v 00 +
and hence

n1 0
v = 0,
r

(
c1 + c2 log r
v(r) =
c3 + c4 r2n

for n = 2,
for n 3,

where ci are constants for i = 1, 2, 3, 4. We are interested in solutions with a

singularity such that
Z
v
d = 1 for any r > 0.
Br r
Hence we set for any fixed a Rn
(
(a, x) =

1
2

log |a x|
1
2n
n (2n) |a x|
59

for n = 2,
for n 3,

60

4. LAPLACE EQUATIONS

where n is the surface area of the unit sphere in Rn . In summary, for any fixed
a Rn , (a, ) is harmonic in Rn \ {a}, i.e.,
x (a, x) = 0

for any x 6= a,

and has a singularity at x = a. Moreover,

Z

(a, x) dSx = 1
n
x
Br (a)

for any r > 0.

The function is called the fundamental solution of the Laplace operator.

Now we prove the Greens identity, which plays an important role in discussions
of harmonic functions. In the next result, we replace a, x by x, y and write (x, y)
instead of (a, x).

Theorem 4.1. Suppose is a bounded C 1 -domain in Rn and that u C 1 ()

C (). Then for any x
Z
Z

u(x) = (x, y)u(y)dy

(x, y)
(y) u(y)
(x, y) dSy .
ny
ny
2

Remark 4.2. (i) For any x , (x, ) is integrable in although it has a

singularity.
the expression in the right hand side is zero. This follows from
(ii) For x
/ ,

the Greens formula in and y (x, y) = 0 if x

/ .
(iii) By letting u = 1, we have
Z

(x, y)dSy = 1 for any x .

ny

Proof. By applying the Greens formula to u and (x, ) in \Br (x) for small
r > 0, we get
Z
Z
Z
u
u


(u u)dy =

u
dSy

u
dSy .
n
n
n
n

\Br (x)

Br (x)

Noting = 0 in \ Br (x), we have

Z
Z
Z
u

udy =

u
dSy lim
r0
n
n

u
dSy .
n
n

For n 3, we get by the definition of

Z
Z

1
u
2n
=

dS
r

y
(2 n)n

u
dSy
n

Br (x)

Br (x)

Br (x)

sup |u| 0
n 2 Br (x)
and

Z
Br (x)

1
u
dSy =
n
n rn1

as r 0,

Z
udSy u(x)

as r 0.

Br (x)

For n = 2, we get the same conclusion similarly.

4.1. DIRICHLET PROBLEMS

61

Now we use Theorem 4.1 to discuss the Dirichlet boundary-value problem

(4.1)

u =f

in ,

u = on ,

and C(). Lemma 3.11 asserts the uniqueness of a solution

for some f C()
2
1
in C ()C (). Another method to obtain this result is by the maximum principle
to be discussed later in this chapter. If u solves (4.1), then by Theorem 4.1 u can
u
be expressed in terms of f and , with one unknown term
on . We intend to
n
eliminate this term by adjusting . We need to emphasize that we cannot prescribe
u
on together with u on .
n
with
For any fixed x , we consider a function (x, ) C 2 () C 1 ()
y (x, y) = 0 in . Then by the Greens formula,
Z
Z

0=
(x, y)u(y)dy
(x, y)
(y) u(y)
(x, y) dSy .
ny
ny

By adding this to the Greens identity in Theorem 4.1, we obtain for any x
Z
Z

u(x) =
(x, y)u(y)dy
(x, y)
(y) u(y)
(x, y) dSy ,
n
n
y
y

where
(x, y) = (x, y) + (x, y).
Now we choose appropriately so that (x, ) = 0 on . This leads to the
important concept of Greens functions.
C 2 () such that
For each fixed x , we consider (x, ) C 1 ()
y (x, y) = 0 for y
(x, y) = (x, y) for y .
By Corollary 4.7, (x, ) is smooth in for each fixed x if it exists. Denote by
G(x, y) the resulting (x, y), which is called the Greens function. If such a G
exists, then the solution u of the Dirichlet problem (4.1) can be expressed by
Z
Z
G
(x, y)dSy .
(4.2)
u(x) =
G(x, y)f (y)dy +
(y)
ny

for each
Note that the Greens function G(x, y) is defined as a function of y
fixed x . Now we discuss properties of G as a function of x and y. We should
point out that the Greens function is unique. This follows from Lemma 3.11 or
Corollary 4.16 to be established, since a difference of any two Greens functions is
harmonic with a zero boundary value.
Lemma 4.3. The Greens function G(x, y) is symmetric in , i.e., G(x, y)
= G(y, x) for x 6= y .
Proof. For any x1 , x2 with x1 6= x2 , take r > 0 small such that Br (x1 )
Br (x2 ) = . Set Gi (y) = G(xi , y) and i (y) = (xi , y). We apply the Greens

62

4. LAPLACE EQUATIONS

formula in \ Br (x1 ) Br (x2 ) and get

Z
Z

G1 G2 G2 G1 =

G2
G1
G1
G2
n
n
\Br (x1 )Br (x2 )

Z
Z
G2

G1
G2
G1

G1
G2

G1
G2
.
n
n
n
n
Br (x1 )
Br (x2 )

Since Gi is harmonic for y 6= xi , i = 1, 2, and vanishes on , we have

Z
Z
G2
G2
G1
G1
G1
G1
G2
+
G2
= 0.
n
n
n
n
Br (x1 )
Br (x2 )
Now we replace G1 in the first integral by 1 and replace G2 in the second integral
by 2 . Since G1 1 is harmonic in and G2 is harmonic in \ Br (x2 ), we have
Z

G2
(G1 1 )
(G1 1 )
G2
0 as r 0.
n
n
Br (x1 )
Similarly,
Z
Br (x2 )

G1

(G2 2 )
G1
(G2 2 )
0
n
n

as r 0.

Therefore, we obtain
Z
Z
G2
2
1
G1
1
G2
+
G1
2
0,
n
n
n
n
Br (x1 )
Br (x2 )
as r 0. On the other hand, we have
Z
Z
G1
G2
0,
2
0
1
n
n
Br (x2 )
Br (x1 )
and

1
G2 (x1 ),
G2
n
Br (x1 )

Z
G1
Br (x2 )

as r 0,

2
G1 (x2 )
n

as r 0.

The two limits above can be proved similarly as the last limit in the proof of
Theorem 4.1. This implies G2 (x1 ) G1 (x2 ) = 0, or G(x2 , x1 ) = G(x1 , x2 ).

Finding Greens functions involves solving a Dirichlet problem for the Laplace
equation. Meanwhile, Greens functions are introduced to solve Dirichlet problems.
We need to break this cycle. In fact, we can construct Greens functions explicitly
for special domains. The next result yields an expression for Greens functions in
balls.
Theorem 4.4. The Greens function in the ball BR Rn is given by

R
1
|x|
G(x, y) =
log |x y| log x
y
for n = 2,
2
|x|
R
and

2n

R
1
|x|
2n

G(x, y) =
|x y|
x
y
(2 n)n
|x|
R

for n 3.

4.1. DIRICHLET PROBLEMS

63

For x = 0, we have
G(0, y) =

2n

1
|y|
R2n
(2 n)n

and
G(0, y) =

for n 3

1
log |y| log |R| for n = 2.
2

Proof. We need to adjust the fundamental solution by adding a harmonic

function so that the sum vanishes on the boundary. Fix an x 6= 0 with |x| < R, and
2
R given by X = R x. In other words, X and x are reflexive of
consider X Rn \ B
|x|2
each other with respect to the sphere BR . Note that the map x 7 X is conformal,
i.e., preserves angles. For |y| = R, we have by the similarity of triangles
|x|
R
|y x|
=
=
,
R
|X|
|y X|
which implies
(4.3)

|y x| =

|x|
|x|
R
|y X| = y
x
R
R
|x|

for any y BR .

Therefore, in order to have a zero boundary value, we take for n 3

Figure 4.1. The reflection about the sphere.

1
G(x, y) =
(2 n)n

|x y|n2

R
|x|

n2

1
|y X|n2

!
.

R . The case
We point out that the second term is smooth for y BR , since X
/B
n = 2 is similar.

Next, we calculate normal derivatives of the Greens function on spheres.

Corollary 4.5. Suppose G is the Greens function in BR as in Theorem 4.4.
Then
G
R2 |x|2
(x, y) =
for any x BR and y BR .
ny
n R|x y|n
Proof. We only consider the case n 3. With X = R2 x/|x|2 , we have

R n2
1
2n
2n
|y x|

|y X|
,
G(x, y) =
(2 n)n
|x|

64

4. LAPLACE EQUATIONS

for any x BR and y BR . Hence we get for such x and y

!
n2
1
yi xi
R
yi Xi
yi R2 |x|2
Gyi (x, y) =

=
,
n
n
n |y x|
|x|
|y X|
n R2 |x y|n
by (4.3) in the proof of Theorem 4.4. With ni =

yi
for |y| = R, we obtain
R

X
G
R2 |x|2
1

(x, y) =
ni Gyi (x, y) =
.
ny
n R |x y|n
i=1
This yields the desired result.

Denote by K(x, y) the function in Corollary 4.5, i.e.,

K(x, y) =

R2 |x|2
n R|x y|n

for any x BR , y BR .

It is called the Poisson kernel and has the following properties:

(K1) K(x, y) is smooth for any x BR and y BR .
(K2) K(x, y) > 0 for any x BR and y BR .
(K3) For any fixed x0 BR , limxx0 ,|x|<R K(x, y) = 0 uniformly in y for
|y x0 | > > 0.
(K4)
R x K(x, y) = 0 for any x BR and y BR .
(K5) BR K(x, y)dSy = 1 for any x BR .
Here (K1), (K2) and (K3) follow easily from the explicit expression for K

G(x, y) and the fact that

and (K4) follows easily from the definition K(x, y) =
ny
G(x, y) is harmonic in x. An easy derivation of (K5) is based on (4.2). By taking
R ) harmonic function u in (4.2), we conclude
a C 2 (B
Z
u(x) =
K(x, y)u(y)dSy for any |x| < R.
BR

This is called the Poisson integral formula.

u 1.

Then we have (K5) easily by taking

Now we are ready to solve the Laplace equation in balls with prescribed Dirichlet boundary values.
Theorem 4.6. For any C(BR ), the function u defined by
Z
R2 |x|2
(4.4)
u(x) =
(y)dSy for any x BR
n
BR n R|x y|
is smooth in BR and continuous up to BR and satisfies
u = 0
u=

in BR ,
on BR .

Proof. By (K1) and (K4), we conclude easily that u defined by (4.4) is smooth
and harmonic in BR . We only need to prove the continuity of u up to the boundary
BR . Let x0 BR and x BR . By (K5), we have
Z

u(x) (x0 ) =
K(x, y) (y) (x0 ) dSy = I1 + I2 ,
BR

4.2. FUNDAMENTAL SOLUTIONS

where

65

I1 =

I2 =

|yx0 |<,yBR

,
|yx0 |>,yBR

for a positive constant to be determined. For any given > 0, we choose =

() > 0 so small that
|(y) (x0 )| <

for any y BR with |y x0 | < ,

by the continuity of . Then |I1 | by (K2) and (K5). Let M = supBR ||. By
(K3), we find a 0 such that

K(x, y)
for any |x x0 | < 0 , |y x0 | > ,
2M n Rn1
where 0 depends on and = (), and hence only on . Then |I2 | < . Hence
|u(x) (x0 )| < 2 for any x BR with |x x0 | < 0 .
This shows the continuity of u at x0 BR .

For n = 2, with x = (r cos , r sin ) and y = (R cos , R sin ) in (4.4), we have

Z 2
1
R2 r 2
(R cos , R sin )d.
u(r cos , r sin ) =
2
2 0 R 2Rr cos( ) + r2
Now we discuss some properties of the solution (4.4). First, u(x) in (4.4) is
smooth for |x| < R, even if the boundary value is simply continuous. In fact,
this is a general fact. Any harmonic function is smooth. We will prove this result
in the next section.
In (4.4), by letting x = 0, we have
Z
1
u(0) =
u(y)dSy .
n Rn1 BR
This is the so-called mean-value property. The value of harmonic functions at center
of spheres is equal to their average on spheres.
Moreover, by (K2) and (K5), u in (4.4) satisfies
inf u sup

BR

BR

in BR .

In other words, harmonic functions in balls are bounded from above by their maximum on boundary and bounded from below by their minimum on boundary. Such
a result is referred to as the maximum principle. Again, this is a general fact and
will be proved for any harmonic functions in any bounded domains.
The mean-value property and the maximum principle are main topics in Section
4.3 and Section 4.4.
4.2. Fundamental Solutions
In this section, we use fundamental solutions of the Laplace equation to discuss
regularity of harmonic functions.
First, as an application of Theorem 4.1, we prove that harmonic functions are
smooth.
Lemma 4.7. Let be a domain in Rn and u C 2 () be a harmonic function
in . Then u is smooth in .

66

4. LAPLACE EQUATIONS

We note that, in its definition, a harmonic function is required only to be C 2 .

Proof. For any x0 , we take a bounded C 1 -domain 0 in such that
0 is a compact subset of . Obviously, u is C 1 in
0 and u = 0 in
x 0 and
0
. Then by Theorem 4.1, we have
Z

u(x) =
(x, y)
(y) u(y)
(x, y) dSy for any x 0 .
ny
ny
0
There is no singularity in the integrand since x 0 and y 0 . This implies
easily that u is smooth in 0 .

In the following, we will prove that harmonic functions are analytic. As the first
step, we estimate derivatives of harmonic functions. For convenience, we consider
harmonic functions in balls.
The following result is referred to as interior gradient estimates. It asserts
that derivatives of a harmonic function at any point is controlled by its maximal
value in a ball centered around the point.

R (x0 ) is harmonic in BR (x0 ) Rn . Then

Lemma 4.8. Suppose u C B
|u(x0 )|

C
max |u|,
R BR (x0 )

where C is a positive constant depending only on n.

Proof. Without loss of generality, we assume x0 = 0.
We first consider R = 1 and employ the local version of the Greens identity.
Choose a cut-off function C0 (B3/4 ) such that = 1 in B1/2 and 0 1.
For y 6= x, (x, y) is harmonic in y and hence

y (y)(x, y) = y (y)(x, y) + 2y (y) y (x, y).

This is zero for |y| < 1/2 and 3/4 < |y| < 1 since is constant there. Apply the
Greens formula to u and (x, ) in B1 \ B (x) for x B1/4 and small enough.
We proceed as in the proof of Theorem 4.1 to obtain
Z

u(x) =
u(y) y (y)(x, y) + 2y (y) y (x, y) dy,
1
3
2 <|y|< 4

for any x B1/4 . Then

Z

u(x) =
u(y) y (y)x (x, y) + 2y (y) x y (x, y) dy,
1
3
2 <|y|< 4

for any x B1/4 . There is no singularity in the integrand for |x| < 1/4 and
1/2 < |y| < 3/4. Hence, we obtain
sup |Du| C sup |u|,
B1
4

B1

where C is a constant depending only on n.

The general case follows from a simple dilation. Define
u
(x) = u(Rx) for any x B1 .

4.2. FUNDAMENTAL SOLUTIONS

67

Then u
is a harmonic function in B1 . By applying the result we just proved to u
,
we obtain
C
sup |u| sup |u|.
R BR
B1R
4

This implies the desired result.

Next, we estimate derivatives of harmonic functions of arbitrary order.

R (x0 ) is harmonic in BR (x0 ) Rn . Then for

Lemma 4.9. Suppose u C B
any multi-index with || = m
| u(x0 )|

C m em1 m!
sup |u|,
Rm
R (x0 )
B

where C is a positive constant depending only on n.

Proof. We prove by an induction on m 1. It holds for m = 1 by Lemma
4.17. We assume it holds for m and consider m+1. Let v be an arbitrary derivative
of u of order m. Obviously, it is harmonic in BR (x0 ). For 0 < < 1, we apply
Lemma 4.17 to v in B(1)R (x0 ) and get
|v(x0 )|

C
max |v|.
(1 )R B(1)R (x0 )

For any x B(1)R (x0 ), we have BR (x) BR (x0 ). By the induction assumption,
we obtain
C m em1 m!
|v(x)|
max |u| for any x B(1)R (x0 ),
(R)m BR (x)
and hence
max

(1)R (x0 )
B

|v|

C m em1 m!
max |u|.
(R)m BR (x0 )

Therefore,
|v(x0 )|
By taking =

m
, we have
m+1
1
=
(1 )m

C m+1 em1 m!
max |u|.
(1 )m Rm+1 BR (x0 )

m
1
1+
(m + 1) < e(m + 1).
m

Hence the desired result is established for any derivatives of order m + 1.

Harmonic functions are not only smooth but also analytic in their domains.
Real analytic functions will be studied in Section 7.2. Now we simply introduce the
notion. A function u : Rn R is called analytic near 0 if its Taylor series about 0
is convergent to u in a neighborhood of 0 Rn , i.e.,
X 1
u(x) =
u(0)x for any |x| < r.
!

Theorem 4.10. Harmonic functions are analytic.

68

4. LAPLACE EQUATIONS

Proof. Suppose u is a harmonic function in Rn . For any fixed x ,

we prove that u is equal to its Taylor series in a neighborhood of x. To do this, we
take B2R (x) and h Rn with |h| R. For any integer m 1, we have by the
Taylor expansion
u(x + h) = u(x) +

m1
X
i=1

i
1h
i
(h1 x1 + + hn xn ) u (x) + Rm (h),
i!

where
1
m
[(h1 x1 + + hn xn ) u] (x1 + h1 , . . . , xn + hn ),
m!
for some (0, 1). Note that x + h BR (x) for |h| < R. Hence by Lemma 4.9,
we obtain

m
m m1
1
m!
Cne|h|
m
m C e
|Rm (h)|
|h| n
max |u|
max |u|.
2R (x)
2R (x)
m!
Rm
R
B
B
Rm (h) =

Then for any h with Cne|h| < R/2, Rm (h) 0 as m .

The proof of Theorem 4.10 shows that the radius of convergence is uniform for
harmonic functions, depending only on n. In other words, any harmonic functions
defined in a fixed ball are equal to their Taylor expansions at least in balls with a
fixed radius and such a radius depends only on the dimension n and is independent of harmonic functions. It is easy to see that each homogenous polynomial of
fixed degree in Taylor expansions of harmonic functions is still harmonic, which is
called a homogenous harmonic polynomial. Homogeneous harmonic polynomials
constitute an important subject in the study of harmonic functions. Because of
the homogeneity, homogeneous harmonic polynomials in Rn can be identified with
corresponding spherical harmonics, their restrictions on the unit sphere Sn1 in Rn .
To end this section, we classify homogeneous harmonic polynomials in R2 and
3
R and study their zero sets. For any integer m 1, we denote by Hm (Rn ) the
collection of all homogeneous harmonic polynomials of degree m in Rn . Obviously,
Hm (Rn ) is a linear space.
We start with n = 2 and denote points in R2 by (x, y). A general homogeneous
polynomial Pm of degree m in R2 contains m + 1 coefficients. Then Pm is a
homogeneous polynomial of degree m2 and therefore contains m1 terms. These
terms have to vanish if Pm is a harmonic function. This gives m1 relations among
the m + 1 constants if Pm is a harmonic function; so that these constants can be
expressed linearly in terms of (m+1)(m1) or 2 of them. Therefore, Hm (R2 ) is a
linear space of dimension 2. We note that Re(x + iy)m and Im(x + iy)m are linearly
independent homogeneous harmonic polynomials and hence form a basis in Hm (R2 ).
In polar coordinates (r, ), these two homogeneous harmonic polynomials are given
by rm cos m and rm sin m. Hence, any homogeneous harmonic polynomial Pm of
degree m can be expressed by
Pm = c1 rm sin m + c2 rm cos m = arm cos(m + 0 ),
for some constants c1 , c2 , a and 0 [0, 2). We note that the nodal set, or the zero
1
set, Pm
(0) of Pm consists of m straight lines passing the origin and forming equal
angles by any two consecutive lines.

4.2. FUNDAMENTAL SOLUTIONS

69

Now we turn to n = 3. A general homogeneous polynomial Pm of degree m in

R3 contains 12 (m + 1)(m + 2) coefficients. Then Pm is a homogeneous polynomial
of degree m 2 and therefore contains 12 m(m 1) terms. These terms have to
vanish if Pm is a harmonic function. This gives 12 m(m 1) relations among the
1
2 (m + 1)(m + 2) constants if Pm is a harmonic function; so that these constants can
be expressed linearly in terms of 12 {(m + 1)(m + 2) m(m 1)} or 2m + 1 of them.
Therefore, Hm (R3 ) is a linear space of dimension 2m + 1. A basis of this linear
space is given in terms of Legendre functions. We now discuss how to construct
such a basis.
Let (x, y, z) and (r, , ) be rectangular coordinates and corresponding polar
coordinates in R3 , i.e.,
x = r sin cos ,

y = r sin sin ,

z = r cos .

Let Pm = r Qm (, ) be a homogeneous harmonic polynomial of degree m in R3 .

Then
1
Qm
1 2 Qm
sin
+
m(m + 1)Qm +
= 0.
sin

sin2 2
By writing = cos , we have

Qm
1 2 Qm
m(m + 1)Qm +
(1 2 )
+
= 0.

1 2 2
By setting Qm (, ) = f ()g(), we obtain
1
f
1 1 2g
m(m + 1) +
(1 2 )
+
= 0.
f

1 2 g 2
The first two terms in this equation are independent of g, and therefore so is the
1 2g
last. Hence, the value of
must be a constant. Since g is 2-periodic in , we
g 2
take this constant to be k 2 for an integer k. Thus
2g
= k 2 g,
2
and hence
g() = A cos k + B sin k,
where A and B are constants. Then f satisfies

f
k2
(1 2 )
+ m(m + 1)
f = 0,

1 2
which is known as Legendres associated equation. If fm,k () is a solution, then

(4.5)
rm A cos k + B sin k fm,k (cos )
is a harmonic function wherever it is defined in R3 . We are interested in only those
fm,k such that (4.5) gives a homogeneous polynomial of degree m in R3 . A lengthy
calculation then yields for k = 0, 1, , m
(4.6)

fm,k () = (1 2 ) 2

dm+k
(1 2 )m .
dm+k

We omit details here. For k = 0, we have

dm
fm,0 () =
(1 2 )m .
dm

70

4. LAPLACE EQUATIONS

This is the Legendre function. For each fixed positive integer m, the collection in
(4.5) with fm,k given by (4.6) for 0 k m consists of 2m + 1 linearly independent
homogeneous harmonic polynomials of degree m and hence forms a basis of Hm (R3 ).
(We note that there is only one function for k = 0 in (4.5). )
Now, we demonstrate that these homogenous harmonic polynomials in (4.5)
exhibit different properties for different k by plotting their nodal sets. Because of
the homogeneity, it is convenient to consider restrictions of nodal sets to the unit
sphere S2 , which is called the nodal curves.
If k = 0, the corresponding harmonic polynomial is a constant multiple of the
Legendre function fm,0 (cos ). A simple calculus argument shows that fm,0 () has
m distinct zeros between 1 and 1, arranged symmetrically about = 0. Hence
fm,0 (cos ) has m distinct zeros between 0 and , arranged symmetrically about
= /2. Accordingly on S2 , the function fm,0 (cos ) vanishes on m latitude circles.
These circles are symmetrically situated with respect to the equator = /2, and,
if m is an odd number, the equator itself is one of these circles. Similarly, level sets

Figure 4.2. Nodal curves of zonal harmonics.

of this function consist of latitude circles. Because of this division of the sphere into
zones by sets of latitude circles, the function fm,0 (cos ) and its constant multiples
are called zonal harmonics.
If 0 < k < m, the spherical harmonic is of the form

dm+k
A cos k + B sin k sink m+k (2 1)m |=cos .
d
The first factor vanishes when A cos k + B sin k = 0, i.e., when tan k = A/B,
and on S2 this corresponds to k great circles through the pole = 0, the angle
between the planes of any two consecutive great circles being /k. The second
factor vanishes at the points = 0 and = , and the third on m k latitude
circles, arranged like the corresponding circles in the case of zonal harmonics. Since
the two sets of circles intersect orthogonally, these harmonics are called tesseral
harmonics.
Finally, if k = m, the spherical harmonic is of the form

A cos m + B sin m sinm ,

which vanishes when = 0 or , or when tan m = A/B. This corresponds on
S2 to the points = 0 and = , and to m great circles through these points, the
angle between the planes of any two consecutive great circles being /m. As S2 is
thus divided up into 2m sectors, these functions called sectorial harmonics. We

4.3. MEAN-VALUE PROPERTIES

71

Figure 4.3. Nodal curves of tesseral harmonics.

Figure 4.4. Nodal curves of sectorial harmonics.

note that sectorial harmonics are simply linear combinations of Re(x + iy)m and
Im(x + iy)m .
In summary, zonal harmonics, tesseral harmonics and sectorial harmonics exhibit different nodal curves. Generally, we do not expect a simple pattern for nodal
sets of arbitrary spherical harmonics.
Finally, we point out that it is beyond the reach of this book to classify homogeneous harmonic polynomials in arbitrary dimensions.

4.3. Mean-Value Properties

As a simple consequence of the Poisson integral formula, harmonic functions are
equal to their mean values over arbitrary spheres. This is the so-called mean-value
property. In this section, we will use the mean-value property to discuss harmonic
functions. The fundamental solution is not used throughout this section.
We first prove by a simple calculation that harmonic functions satisfy the meanvalue property.
Theorem 4.11. Let be a domain in Rn and u C 2 () be harmonic in .
Then
Z
1
u(x) =
u(y)dSy for any Br (x) ,
n rn1
Br (x)

where n is the area of the unit sphere B1 in Rn .

72

4. LAPLACE EQUATIONS

Proof. Take any ball Br (x) . For any (0, r), we apply the Greens
formula in B (x) and get
Z
Z
Z
u
u
n1
4u =
dS =
(x + w)dSw
B (x)
B (x) n
B1
(4.7)
Z

= n1
u(x + w)dSw .
B1
Hence for the harmonic function u, we have for any (0, r)
Z

u(x + w) dSw = 0.
B1
Integrating from 0 to r, we obtain
Z
Z
u(x + rw) dSw =
B1

u(x) dSw = u(x)n ,

B1

or

Z
1
u(x) =
u(x + rw) dSw .
n B1
A simple change of variables yields the desired result.

Theorem 4.11 asserts that harmonic functions equal to their mean values on
spheres. There are two versions of mean-value properties, mean values on spheres
and mean values on balls.
Definition 4.12. We assume that is a domain in Rn . For u C(), (i) u
satisfies the first mean-value property if
Z
1
u(y)dSy for any Br (x) ;
u(x) =
n rn1
Br (x)

(ii) u satisfies the second mean-value property if

Z
n
u(x) =
u(y)dy for any Br (x) ,
n rn
Br (x)

where n denotes the surface area of the unit sphere in Rn .

We note that n rr1 is the surface area of the sphere Br (x) and n rn /n is
the volume of the ball Br (x).
These two definitions are equivalent. In fact, if we write (i) as
Z
1
u(x)rn1 =
u(y)dSy ,
n
Br (x)

we integrate with respect to r to get (ii). If we write (ii) as

Z
n
n
u(y)dy,
u(x)r =
n
Br (x)

we differentiate to get (i).

By a change of variables, we also write mean-value properties in the following
equivalent forms:

4.3. MEAN-VALUE PROPERTIES

73

(i) u satisfies the first mean-value property if

Z
1
u(x) =
u(x + ry)dSy for any Br (x) ;
n
B1

(ii) u satisfies the second-mean value property if

Z
n
u(x) =
u(x + ry) dy for any Br (x) .
n B1
For a function u satisfying mean-value properties, u is required only to be
continuous. However, a harmonic function is required to be C 2 . We now prove
these two are equivalent. We already proved that harmonic functions satisfy the
mean-value property in Theorem 4.11. We now prove that any function with the
mean-value property is harmonic.
Theorem 4.13. If u C() has the mean-value property in , then u is
smooth and harmonic in .
Proof. For the smoothness, we prove that u is equal to the convolution of
itself Rwith some smooth function. To this end, we choose a function C0 (B1 )
with B1 = 1 and (x) = (|x|), i.e.,
Z 1
n
rn1 (r) dr = 1.
0

1 x
The existence of such a function can be verified easily. We define (x) = n

for any > 0. Obviously, supp B . We claim

Z
u(x) =
u(y) (y x)dy for any x with d(x, ) > .

Then it follows that u is smooth. Moreover, by (4.7) in the proof of Theorem 4.11
and the mean-value property, we have for any Br (x)
Z
Z

u = rn1
u(x + rw)dSw = rn1
n u(x) = 0.
r B1
r
Br (x)
This implies u = 0 in .
Now we prove the claim. For any x and < dist (x, ), we have
Z
Z
Z
y
1
u(x + y)
dy
u(y) (y x)dy =
u(x + y) (y)dy = n
B

B
Z
=
u(x + y)(y)dy
B1
1

Z
=

Z
r

n1

dr

Z
=

u(x + rw)(rw)dSw
B1

(r)rn1 dr

= u(x)n

u(x + rw)dSw
B1

(r)rn1 dr = u(x),

where in the last equality we used the mean-value property.

74

4. LAPLACE EQUATIONS

By combining Theorems 4.11 and 4.13, we conclude the following result.

Corollary 4.14. Harmonic functions are smooth and satisfy the mean-value
property.
Now we prove the maximum principle for harmonic functions.
be harmonic in . Then u assumes its
Theorem 4.15. Let u C 2 () C()
maximum and minimum only on unless u is constant.
Proof. We prove only for the maximum. Set

D = x ; u(x) = M max u .

It is obvious that D is relatively closed; namely, for any sequence {xm } D, if

xm x , then x D. This follows easily from the continuity of u. Next we
r (x0 ) for some r > 0. By the
show that D is open. For any x0 D, take B
mean-value property, we have
Z
Z
n
n
M = u(x0 ) =
u(y)dy M
dy = M.
n rn
n rn
Br (x0 )

Br (x0 )

This implies u = M in Br (x0 ). Hence D is both relatively closed and open in .

Therefore either D = or D = . A similar argument can be used to prove for the
minimum.

A consequence is the uniqueness of solutions of the Dirichlet problem in a

bounded domain.
Corollary 4.16. For any f C() and C(), there exists at most one
of the following problem
solution u C 2 () C()
u = f in ,
u = on .
Proof. Let w be the difference of any two solutions. Then w = 0 in and
w = 0 on . Theorem 4.15 implies w = 0 in .

Compare Corollary 4.16 with Lemma 3.11, where the uniqueness was proved

by energy estimates for solutions u C 2 () C 1 ().

In general, the uniqueness does not hold for unbounded domains. Consider the
Dirichlet problem in an unbounded domain
u = 0 in
u = 0 on .
First, we consider the case = {x Rn ; |x| > 1}. We have a nontrivial solution u
given by
(
log |x|
for n = 2;
u(x) =
2n
|x|
1 for n 3.
Note that u as |x| for n = 2 and u is bounded for n 3. Next,
we consider the upper half space = {x Rn ; xn > 0}. Then u(x) = xn is a
nontrivial solution, which is unbounded.

4.3. MEAN-VALUE PROPERTIES

75

Next, we employ the mean-value property to prove interior gradient estimates.

R (x0 ) is harmonic in BR (x0 ). Then

Lemma 4.17. Suppose u C B
n
|u(x0 )|
max |u|.
R BR (x0 )
We note that Lemma 4.17 improves Lemma 4.8.
R (x0 )). Since u is smooth, then
Proof. For simplicity, we assume u C 1 (B
(uxi ) = 0, i.e., uxi is also harmonic in BR (x0 ). Hence uxi satisfies the mean-value
property. By the divergence theorem, we have
Z
Z
n
n
uxi (x0 ) =
u
(y)dy
=
u(y) i dSy ,
x
n Rn BR (x0 ) i
n Rn BR (x0 )
and hence
|uxi (x0 )|

n
n
max |u| n Rn1
max |u|.
n Rn BR (x0 )
R BR (x0 )

This yields the desired result.

When u is nonnegative, we can improve Lemma 4.17.

R (x0 ) is a nonnegative harmonic function in

Lemma 4.18. Suppose u C B
BR (x0 ). Then
n
|u(x0 )| u(x0 ).
R
This result is often referred to as the differential Harnack inequality. It has
many important consequences.
Proof. As in the proof of Lemma 4.17, by the divergence theorem and the
nonnegativeness of u, we have
Z
n
n
|uxi (x0 )|
u(y) dSy = u(x0 ),
n
n R BR (x0 )
R
where in the last equality we used the mean-value property.

As an application, we prove the following result which is referred to as the

Liouville Theorem.
Corollary 4.19. A harmonic function in Rn bounded from above or below is
constant.
Proof. Suppose u is a harmonic function in Rn with u 0. Then for any
x Rn , we apply Lemma 4.18 to u in BR (x) and then let R . We conclude
u(x) = 0 for any x Rn and hence u is constant.

As another application, we prove the Harnack inequality, which asserts that

nonnegative harmonic functions have comparable values in compact subsets.
Lemma 4.20. Suppose u is a nonnegative harmonic function in BR (x0 ). Then
u(x) Cu(y)

for any x, y B R (x0 ),

where C is a positive constant depending only on n.

76

4. LAPLACE EQUATIONS

Proof. By considering u + for any > 0, we assume u > 0 in BR (x0 ). For

any x BR/2 (x0 ), we apply Lemma 4.18 to u in BR/2 (x) and get
|u(x)|

2n
u(x),
R

or
| log u(x)|
For any x, y BR/2 (x0 ),
u(x)
=
log
u(y)

1
0

2n
.
R

d
log u(ty + (1 t)x)dt.
dt

With ty +(1 t)x BR/2 (x0 ) for any t [0, 1] and |x y| R, a simple integration
yields
Z 1
u(x)
2n
log
|x y|
| log u(ty + (1 t)x)|dt
|x y| 2n.
u(y)
R
0
Therefore
u(x) e2n u(y).
This yields the desired result.

In fact, Lemma 4.20 can be proved directly by the mean-value property.

Another proof of Lemma 4.20. Consider any B4r (
x) BR (x0 ). We first
claim
u(x) 2n u(y) for any x, y Br (
x).
To see this, we note that Br (x) B2r (y) B4r (
x) for any x, y Br (
x). Then the
mean-value properties imply
Z
Z
n
n
u(x) =
u
u 2n u(y).
n rn Br (x)
n rn B2r (y)
With r = R/8, we choose finitely many x
1 , , x
N BR/2 (x0 ) such that {Br (
xi )}
covers BR/2 (x0 ). Obviously, B4r (
xi ) BR (x0 ) and N is a constant depending only
on n. Then we have the desired result by choosing C = 2nN .

4.4. The Maximum Principle

One of the important tools in studying harmonic functions is the maximum
principle. In this section, we discuss the maximum principle for a class of elliptic
differential equations slightly more general than the Laplace equation and derive a
priori estimates for boundary-value problems, interior gradient estimates and the
Harnack inequality.
Throughout this section, we assume is a bounded domain in Rn and c is a
Sometimes, we assume c 0 in . For u C 2 () C()

continuous function in .
and f C(), we consider
(4.8)

u + c(x)u = f (x) in .

Obviously, u is harmonic if c = f = 0. A C 2 -function u is called a subsolution

(or supersolution) of (4.8) if u + cu f (or u + cu f ). If c = 0 and
f = 0, subsolutions (or supersolutions) are called subharmonic (or superharmonic)

4.4. THE MAXIMUM PRINCIPLE

77

functions. In other words, a C 2 -function u is subharmonic (or superharmonic) if

u 0 (or u 0).
Now we start to prove the maximum principle without using mean-value properties. We prove a simple result first.
satisfies u + cu > 0 in with
Lemma 4.21. Suppose u C 2 () C()
then u cannot attain this
c(x) 0 in . If u has a nonnegative maximum in ,
maximum in .

Proof. Suppose u attains its nonnegative

2
maximum of at x0 . Then
u(x0 ) = 0 and the Hessian matrix u(x0 ) is nonpositive definite. By taking a
trace, we have
u(x0 ) = tr(2 u(x0 )) 0,
and hence
(u + cu)(x0 ) 0,
which is a contradiction.

Remark 4.22. If c(x) 0, the requirement for the nonnegativeness on u can

be removed. This remark holds for many results in the rest of this section.
Remark 4.23. Lemma 4.21 in fact holds for more general elliptic differential
and f C(), consider
equations. For u C 2 () C()
n
X

Lu

aij (x)uxi xj +

i,j=1

n
X

bi (x)uxi + c(x)u = f (x)

in .

i=1

and that L
We assume that aij , bi and c are continuous and hence bounded in
is uniformly elliptic in in the following sense
n
X
aij (x)i j ||2 for any x and any Rn ,
i,j=1

for some positive constant . The uniform ellipticity means a uniform positive lower
bound of all eigenvalues of (aij ) in . Many results in this section hold for Lu = f .
We now prove the maximum principle for subsolutions.
satisfies u + cu 0 in with
Theorem 4.24. Suppose u C 2 () C()
i.e.,
c(x) 0 in . Then u attains on its nonnegative maximum in ,
sup u sup u+ .

always attains its maximum in .

Theorem 4.24
A continuous function in
asserts that any subsolution continuous up to the boundary attains its nonnegative
maximum on the boundary , possibly also in . Theorem 4.24 is often called
the weak maximum principle. A stronger version asserts that subsolutions attain
their maximum only on the boundary. We will prove the strong maximum principle
later.
Proof. Without loss of generality, we assume is contained in a ball of radius
R and centered at the origin, i.e., BR . For any > 0, consider
u (x) = u(x) (R2 |x|2 ).

78

4. LAPLACE EQUATIONS

By c 0 and |x| < R in , we have

u + cu = u + cu + 2n c(R2 |x|2 ) 2n > 0 in .
By Lemma 4.21, u attains its nonnegative maximum only on , i.e.,
sup u sup u+
.

Then we obtain
2
+
2
sup u sup u + R2 sup u+
+ R sup u + R .

We finish the proof by letting 0.

The following result is often referred to as the comparison principle.

satisfy
Corollary 4.25. Suppose c 0 in and u, v C 2 () C()
u + cu v + cv
u v

in

on .

Then u v in .
Proof. The difference w = u v satisfies w + cw 0 in and w 0 on
. Theorem 4.24 implies w 0 in .

The comparison principle provides a reason that functions u satisfying u +

cu f are called subsolutions. They are less than solutions v of v + cv = f with
the same boundary values.
A consequence of the comparison principle is the uniqueness of solutions of
Dirichlet problems.
be
Corollary 4.26. Let be a bounded domain in Rn and u C 2 () C()
a solution of
u + cu = f
u=

in
on ,

and C(). If c(x) 0 in , then u is the unique solution.

for some f C()
The boundedness of domains is essential, since it guarantees the existence
The uniqueness does not hold if domains
of maximum and minimum of u in .
are unbounded. Examples are given in Section 4.3. Equally important is the
nonpositiveness of the coefficient c. For example, u = sin x is a nontrivial solution
of the problem
u00 + u = 0

in (0, )

u(0) = u() = 0.
The weak maximum principle asserts subsolutions of elliptic differential equations with nonpositive zero order coefficients attain on boundary their nonnegative
maximum. In fact, they can attain their nonnegative maximum only on boundary,
unless they are constant. This is the strong maximum principle. To prove this,
we need the following Hopf lemma. Let u be a subsolution in which attains its
u
maximum on , say at x0 . Then
(x0 ) 0. The Hopf lemma asserts that
n
it is in fact positive.

4.4. THE MAXIMUM PRINCIPLE

79

Theorem 4.27. Let B be an open ball in Rn with x0 B. Suppose u

C (B) C 1 (B {x0 }) satisfies u + cu 0 in B with c(x) 0 in B. Assume in
addition that
2

u(x) < u(x0 )

for any x B and u(x0 ) 0.

Then
u
(x0 ) > 0.
n
Proof. We assume that B is centered at the origin with radius r. We assume
and u(x) < u(x0 ) for any x B\{x

further that u C(B)

0 }, since we can construct
a ball B B tangent to B at x0 .
Set D = B Br/2 (x0 ) and for some to be determined
2

h(x) = e|x| er .
A direct calculation yields

2
2
h + ch =e|x| 42 |x|2 2n + c cer

2
e|x| 42 |x|2 2n + c ,
where we used c 0 in B. By |x| r/2 in D and choosing sufficiently large, we
conclude
h + ch > 0

in D.

Consider v(x) = u(x) + h(x). Then,

v + cv = u + cu + (h + ch) > 0

in D,

for any > 0. By Lemma 4.21, v cannot attain its nonnegative maximum in D.
Next, we prove, for some small > 0, v attains at x0 its nonnegative maximum.
Consider v on the boundary D.
(i) For x DB, since u(x) < u(x0 ), so u(x) < u(x0 ) for some > 0. Take
small such that h < on D B. Hence, for such an , we have v(x) < u(x0 )
for any x D B.
(ii) On D B, h(x) = 0 and u(x) < u(x0 ) for x 6= x0 . Hence v(x) < u(x0 )
on D B \ {x0 } and v(x0 ) = u(x0 ).
Therefore, we conclude
v
(x0 ) 0,
n
or
2
h
u
(x0 ) (x0 ) = 2rer > 0.
n
n
This yields the desired result.

Now, we are ready to prove the strong maximum principle.

satisfy u + cu 0 with c(x) 0 in
Theorem 4.28. Let u C 2 () C()
can be attained only on unless u
. Then the nonnegative maximum of u in
is a constant.

80

4. LAPLACE EQUATIONS

Set D = {x
Proof. Let M be the nonnegative maximum of u in .
; u(x) = M }. It is relatively closed in . We prove D = by contradiction.
If D is a proper subset of , we can find an open ball B \ D with a point on
its boundary belonging to D. In fact, we may choose a point p \ D such that
d(p, D) < d(p, ) first, construct a small ball in \ D of center p and then extend
this ball. It hits D before hitting . Suppose x0 B D. Obviously, we have
u + cu 0 in B and
u(x) < u(x0 )

for any x B and u(x0 ) = M 0.

u
(x0 ) > 0, where n is the outward normal vector at x0 to
n
the ball B. While x0 is an interior maximal point of , we have u(x0 ) = 0. This
leads to a contradiction.

Theorem 4.27 implies

The following result improves Corollary 4.25.

satisfies u + cu 0 in with
Corollary 4.29. Suppose u C 2 () C()
c(x) 0 in . If u 0 on , then either u < 0 in or u 0 in .
In order to discuss boundary-value problems with general boundary conditions,
we need the following result, which is a corollary of Theorem 4.27 and Theorem
4.28.
Corollary 4.30. Suppose is a bounded C 1 -domain in Rn and that u
satisfies u + cu 0 in with c(x) 0. Assume u attains its
C () C 1 ()
Then x0 and
nonnegative maximum at x0 .
u
(x0 ) > 0,
n

unless u is a constant in .
2

Now we discuss the uniqueness of solutions of a class of boundary-value problems with general boundary conditions.
Corollary 4.31. Suppose is a bounded C 1 -domain in Rn , c is a continuous
be a
function in and is a continuous function on . Let u C 2 () C 1 ()
solution of the boundary-value problem
u + cu =f

in

u
+ u = on ,
n
and C(). Assume in addition that c 0 in and 0
for some f C()
on . Then u is the unique solution if c 6 0 or 6 0. If c 0 and 0, u is
unique up to additive constants.
Proof. Suppose u is a solution of the homogeneous problem
u + cu = 0

in

u
+ u = 0 on .
n

Case 1. c
6 0 or 6 0. Suppose that u has a positive maximum at x0 .
If u is a positive constant, there leads to a contradiction to the condition c 6 0 in

4.4. THE MAXIMUM PRINCIPLE

81

u
or 6 0 on . Otherwise x0 and n
(x0 ) > 0 by Corollary 4.30, which
contradicts the boundary value. Therefore u 0.
Case 2. c 0 and 0. Suppose u is a nonconstant solution. Then its
is assumed only on by Theorem 4.28, say at x0 . Again
maximum in
u
(x0 ) > 0. This contradiction shows that u is a constant.
Corollary 4.30 implies n

Proving the uniqueness of solutions of boundary-value problems is an important application of maximum principles. Equally important or more important is
to derive a priori estimates. In derivations of a priori estimates, essential steps
consist of constructions of auxiliary functions. Next, we derive a priori estimates
for solutions of Dirichlet problems.
satisfies
Theorem 4.32. Suppose u C 2 () C()
u + cu = f
u=

in ,
on ,

and C(). If c(x) 0, then

for some f C()
max |u| max || + C max |f |,

where C is a positive constant depending only on n and diam().

If = BR (x0 ), then
max |u| max || +

BR (x0 )

BR (x0 )

R2
max |f |.
2n BR (x0 )

This follows from the proof below easily.

Proof. We construct an auxiliary function w in such that
(i) ( + c)(w u) = ( + c)w f 0, or ( + c)w f in ;
(ii) w u = w 0, or w on .
Set
F = max |f |,

= max ||.

We need
w + cw F

in ,

w on .
Without loss of generality, we assume is contained in a ball of radius R and
centered at the origin, i.e., BR . Set
F
(R2 |x|2 ).
w =+
2n
Then by c 0 and |x| R in , we have
cF 2
(R |x|2 ) F.
2n
We also have w on . Hence w satisfies (i) and (ii). By Corollary 4.25, the
comparison principle, we conclude w u w in , and in particular,
1
|u(x)| +
(R2 |x|2 )F for any x .
2n
w + cw = F + c +

82

4. LAPLACE EQUATIONS

This yields the desired result.

Now we consider a class of general boundary-value problems.

satisfies
Theorem 4.33. Suppose u C 2 () C 1 ()
u + cu = f

in ,

u
+ u =
n

on ,

and C(). If c(x) 0 in and (x) 0 on for a

for some f C()
positive constant 0 , then

max |u| C max || + max |f | ,

where C is a positive constant depending only on 0 and diam().

If c 0 in and 0 on , the homogeneous problem in Theorem 4.33
(with f 0 in and 0 on ) admits a nontrivial solution u 1 in . Hence
there does not hold a sup-norm estimate in this case.
Proof. By assuming BR for some R > 0, we prove
sup |u|

1
1 1 + R2
max || +
+ R2 max |f |.

0
2n
0

Set
= max ||,

and
v(x) =

F = max |f |,

1
1 1 + R2
+
+ R2 |x|2 F.
0
2n
0

Then
v + cv = F + cv F

in ,

and

F
1 + R2
+ v = +
2x n + (
+ R2 |x|2 )
n
0
2n
0
F
(2x n + 1 + R2 ) on ,
+
2n
where n is the unit exterior normal vector of . With
w = v u,
we have
w + cw 0
w
+ w 0
n
It is easy to prove w 0 in , or
|u| v
This yields the desired result.

in ,
on .

in .

4.4. THE MAXIMUM PRINCIPLE

83

In the following, we derive gradient estimates, estimates of derivatives. The

basic method to derive gradient estimates, the so-called the Bernstein method,
involves a differentiation of the equation with respect to xk , k = 1, . . . , n, followed
by a multiplication by uxk and summation over k. The maximum principle is then
applied to the resulting equation in the function v = |u|2 , possibly with some
modifications. There are two kinds of gradient estimates, global gradient estimates
and interior gradient estimates. We use the Laplace equation to demonstrate ideas
of proving interior gradient estimates. Compare with Lemma 4.17.
1 ) is harmonic in B1 . Then
Theorem 4.34. Suppose u C(B
sup |u| C sup |u|,
B1

B1

where C is a positive constant depending only on n.

Proof. A direct calculation shows
n
n
n
X
X
X
4(|u|2 ) = 2
u2xi xj + 2
uxi (4u)xi = 2
u2xi xj ,
i,j=1

i=1

i,j=1

where we used 4u = 0 in B1 . Hence |u|2 is a subharmonic function. To get

interior estimates, we need to introduce a cut-off function. For any nonnegative
function C01 (B1 ), we have
4(|u|2 ) = (4)|u|2 + 4

n
X

xi uxj uxi xj + 2

i,j=1

n
X

u2xi xj .

i,j=1

By the Cauchy inequality, we have

4|xi uxj uxi xj | 2u2xi xj +

2 2 2
u .
xi xj

Then

2||2
4(|u|2 )
|u|2 .

To get a bounded ||2 / in B1 , we take = 2 for some C01 (B1 ). Moreover,

we require 1 in B1/2 . Then

4( 2 |u|2 ) 24 6||2 |u|2 C|u|2 ,

where C is a positive constant depending only on and n. Note

4(u2 ) = 2|u|2 + 2u4u = 2|u|2 ,
since u is harmonic. By taking a constant large enough, we get
4( 2 |u|2 + u2 ) (2 C)|u|2 0.
We apply Theorem 4.24, the maximum principle, to get
sup( 2 |u|2 + u2 ) sup( 2 |u|2 + u2 ).
B1

B1

This implies the desired result easily since = 0 on B1 and = 1 in B1/2 .

Next we derive the differential Harnack inequality for harmonic functions.
Compare this with Lemma 4.18.

84

4. LAPLACE EQUATIONS

Theorem 4.35. Suppose u is a positive harmonic function in B1 . Then

sup | log u| C,
B1
2

where C is a positive constant depending only on n.

Proof. Set v = log u. A direct calculation shows
4v = |v|2 .
Next, we prove an interior gradient estimate for v. By setting w = |v|2 , we get
4w + 2

n
X

vxi wxi = 2

i=1

n
X

vx2i xj .

i,j=1

As before, we need to introduce a cut-off function. First note

n
X

(4.9)

vx2i xj

i,j=1

n
X

vx2i xi

i=1

w2
|v|4
1
(4v)2 =
=
.
n
n
n

Take a nonnegative function C01 (B1 ). As in the proof of Theorem 4.34, we have
4(w) + 2

n
X

vxi (w)xi

i=1

=2

n
X
i,j=1

n
X
i,j=1

n
X

vx2i xj + 4

vx2i xj

xi vxj vxi xj + 2w

i,j=1

n
X

xi vxi + (4)w

i=1

4||2
2|||v| + 4
|v|2 .

Here we keep one vx2i xj instead of dropping it entirely in the proof of Theorem 4.34.
To get a bounded ||2 / in B1 , we take = 4 for some C01 (B1 ). We obtain
by (4.9)
4( 4 w) + 2

n
X

vxi ( 4 w)xi

i=1

1
4 |v|4 8 3 |||v|3 + 4 2 (4 13||2 )|v|2 .
n
Note

1 4
t 8||t3 + 4(4 13||2 )t2 C for any t R,
2n
where C is a positive constant depending only on n and . Hence with t = |v|,
we get
n
X
1 4 2
w C.
4( 4 w) + 2
vxi ( 4 w)xi
2n
i=1
where C is a positive constant depending only on n and .
Suppose 4 w attains its maximum at x0 B1 . Then ( 4 w) = 0 and
4( 4 w) 0 at x0 . Hence
4 w2 (x0 ) C.

4.4. THE MAXIMUM PRINCIPLE

85

If w(x0 ) 1, then 4 w(x0 ) C. Otherwise 4 w(x0 ) 4 (x0 ). In both cases we

conclude
4 w C in B1 ,
where C is a positive constant depending only on n and . With the help of the
definition of w and = 1 in B1/2 , this implies the desired result.

The following result is called the Harnack inequality. Compare it with Lemma
4.20.
Corollary 4.36. Suppose u is a nonnegative harmonic function in B1 . Then
u(x1 ) Cu(x2 )

for any x1 , x2 B 21 ,

where C is a positive constant depending only on n.

The proof is identical to the first proof of Lemma 4.20 and is omitted.
To end this section, we discuss isolated singularities of harmonic functions.
We note that the fundamental solution of the Laplace operator has an isolated
singularity and is harmonic elsewhere. The next result asserts that an isolated
singularity of harmonic functions, if better than that of the fundamental solution,
can be removed.
Theorem 4.37. Suppose u is harmonic in BR \ {0} Rn and satisfies
(
o(log |x|), n = 2
u(x) =
as |x| 0.
o(|x|2n ), n 3
Then u can be defined at 0 so that it is C 2 and harmonic in BR .
Proof. Assume u is continuous in 0 < |x| R. Let v solve
v = 0

in BR ,

v = u on BR .
The existence of v is guaranteed by Theorem 4.6. By the maximum principle,
|v| M

in BR ,

where M = maxBR |u|. We prove u = v in BR \ {0}. Set w = v u in BR \ {0}

and Mr = maxBr |w| for any r < R. We only consider the case n 3. First, we
have
rn2
|w(x)| Mr n2 for any x Br .
|x|
1
Note that w and
are harmonic in BR \ Br with w = 0 on BR . Hence the
|x|n2
maximum principle implies
|w(x)| Mr

rn2
for any x BR \ Br ,
|x|n2

where
Mr = max |v u| max |v| + max |u| M + max |u|.
Br

Br

Br

Br

Then for each fixed x 6= 0,

|w(x)|

rn2
1
M + n2 rn2 max |u| 0 as r 0.
Br
|x|n2
|x|

86

4. LAPLACE EQUATIONS

This implies w = 0 in BR \ {0}.

4.5. Schauder Estimates

In this section, we discuss regularity of solutions of the Poisson equation. Let
be a domain in Rn and f be a continuous function in . Then the Poisson equation
has the form
(4.10)

u = f.
2

The Laplace operator acts on C -functions and u is continuous for any C 2 function u. Conversely, we ask whether u is C 2 if f is continuous. The following
example provides a negative answer.
Example 4.38. For any R < 1, consider in BR R2

x2 x2
1
4
(4.11)
u = 2 2 1
+
.
2|x|
( log |x|)1/2
2( log |x|)3/2
By letting f (x) be the function in the right hand side of (4.11), we note that f is
R if we set it equal to zero at the origin. The function
continuous in B
R ) C (B
R \ {0})
u(x) = (x21 x22 )( log |x|)1/2 C(B
satisfies (4.11) in BR \ {0}. Note that u is not in C 2 (BR ) since limx0 ux1 x1 = .
In fact, (4.11) has no C 2 -solutions. Assume on the contrary that there exists a
C 2 -solution v in BR . Then the function w = u v is harmonic and bounded in
BR \ {0}. By Theorem 4.37, w may be redefined at the origin so that 4w = 0 in
BR and therefore belongs to C 2 (BR ). In particular, the (finite) limit limx0 ux1 x1
exists, which is a contradiction.
It turns out that u fails to be C 2 because there is no control on the module of
continuity of f . If there is a better assumption than the continuity of f , we can
control the module of continuity of 2 u. Here, controlling the module of continuity
means imposing conditions on how fast functions approach constants. The simplest
way to do this is known by Holder norms.
In the following, we discuss the Schauder estimates for Poisson equations.
Schauder estimates are among the most important results in the theory of elliptic
partial differential equations. They give Holder regularity estimates for solutions
with Holder continuous data and form the basis of general existence theorems.
There are two classes of Schauder estimates, interior estimates and boundary estimates. In this section, we only discuss interior Schauder estimates for Poisson
equations. In its simplest form, the interior Schauder estimate asserts that the
C -norm of any second-order derivatives of u in a ball is estimated by the sum of
the C -norm of u and the sup-norm of u in a larger ball. It is an optimal result
in the sense that u is as regular as the data allow. Even though u is just one
particular combination of second-order derivatives of u, the Holder continuity of
u implies the same regularity for all second-order derivatives.
We first define pointwise Holder continuity.
Definition 4.39. Let k be a positive integer and (0, 1) be a constant.
Suppose u is a function defined in B1 .

4.5. SCHAUDER ESTIMATES

87

(1) A function u is C at 0 if
|u(x) u(0)| c|x|

for any x B1 ,

for a positive constant c. The Holder semi-norm of u at 0 is defined by

|u(x) u(0)|
.
|x|
|x|1

[u]C (0) sup

(2) A function u is C k, at 0 if there exists a polynomial p(x) of degree k such

that
|u(x) p(x)| c|x|k+ for any x B1 ,
for a positive constant c.
We note that the definition of Holder continuity here gives a quantitative speed
of how fast functions approach constant or a polynomial. We are only interested
in asymptotic behavior of u as x 0. The ball B1 can be replaced by any Br
for small r. This affects only the definition of Holder semi-norms. The asymptotic
behavior of u as x 0 remains the same.
We also note that the polynomial p in (2) is unique if it exists. We point out
that u in Definition 4.39(2) is not necessarily C k in a neighborhood of 0. If u is,
then p is evidently the Taylor expansion of order k of u at 0. In particular, for
k = 2,
1
p(x) =u(0) + u(0) x + xT 2 u(0)x
2
n
n
X
1 X
=u(0) +
uxi (0)xi +
ux x (0)xi xj .
2 i,j=1 i j
i=1
We first state a global estimate on solutions in terms of nonhomogeneous terms
and boundary values.
1 ) satisfies
Lemma 4.40. Suppose u C 2 (B1 ) C(B
u = f in B1 ,
for some f C(B1 ). Then
sup |u| sup |u| +
B1

B1

1
sup |f |.
2n B1

This is a special case of Theorem 4.32.

Now we begin to estimate second derivatives of solutions.
Lemma 4.41. Suppose u C 2 (B1 ) satisfies
u = f in B1 ,
for some f C(B1 ). Then for any (0, 1), there exist constants c0 > 0, (0, 1)
and 0 > 0, depending only on n and , such that, if |u| 1 and |f | 0 in B1 ,
then there exists a second order harmonic polynomial
1
p(x) = xT Ax + B x + C,
2
satisfying
|u(x) p(x)| 2+ for any |x| ,

88

4. LAPLACE EQUATIONS

and
|A| + |B| + |C| c0 .
Here, A is an n n matrix, B is a vector and C is a constant.
1 ). Suppose v is the
Proof. Without loss of generality, we assume u C(B
harmonic function satisfying
v = 0 in B1 ,
v = u on B1 .
The existence of v is guaranteed by Theorem 4.6. Clearly |v| 1 in B1 , by the
maximum principle since |u| 1 on B1 . Then Lemma 4.9 implies
3
X

|k v|L (B 1 ) c(n),
2

k=0

where c(n) is a positive constant depending only on n, which is often called a

universal constant. Hence the second order Taylor polynomial of v at 0
1 T
x Ax + B x + C
2
has universal bounded coefficients and is harmonic. Now the function u v satisfies
p(x) =

(u v) = f in B1 ,
u v = 0 on B1 .
By Lemma 4.40, we have
|u(x) v(x)|

1
sup |f |
2n B1

for any x B1 .

By the mean value theorem, we have for any x B1/2

|u(x) p(x)| |v(x) p(x)| +

1
1
sup |f | c|x|3 sup |3 v| +
sup |f |
2n B1
2n
B1
B1
2

1
c(n)|x|3 +
sup |f |.
2n B1
Now take small enough such that
c(n)1

1
,
2

and then take 0 such that

0 2+ .
We then have
|u(x) p(x)|

1 2+ 1

+ 0 2+
2
2

This is the desired estimate.

Now we prove the pointwise Schauder estimate.

for any x B .

4.5. SCHAUDER ESTIMATES

89

Theorem 4.42. Suppose u C 2 (B1 ) satisfies

u = f in B1 ,
where f is H
older continuous at 0. Then, u is C 2, at 0. Moreover, there exists a
second order polynomial
1
p(x) = xT Ax + Bx + C
2
satisfying

|u(x) p(x)| c0 |x|2+ |u|L + |f (0)| + [f ]C (0) for any x B1 ,

and

|A| + |B| + |C| c0 |u|L + |f (0)| + [f ]C (0) ,

where c0 is a positive constant depending only on n and .
Obviously, for u C 2 (B1 ), we have

1
p(x) = u(0) + u(0) x + xT 2 u(0)x.
2
Proof. Without loss of generality, we assume f (0) = 0. In general, we set
1
v = u 2n
f (0)|x|2 . Then v = f f (0) in B1 . Furthermore, we assume |u| 1 and
[f ]C (0) 0 for small 0 > 0. The general case can be recovered by considering
u
.
|u|L + 10 [f ]C (0)
First we claim that there exist harmonic polynomials for any k = 1, 2, ,
1
Pk (x) = xT Ak x + Bk x + Ck ,
2
satisfying
|u(x) Pk (x)| (2+)k for any |x| k ,
and
|Ak Ak+1 | ck ,
|Bk Bk+1 | c(+1)k ,
|Ck Ck+1 | c(+2)k ,
where (0, 1) and c are positive constants depending only on n and .
Note that the case k = 1 corresponds to Lemma 4.41. Assume it holds for some
k 1. Set
(u Pk )(k y)
w(y) =
for any y B1 .
(2+)k
Then
f (k y)
w(y) =
for any y B1 ,
k
with
|f (k y)|
sup
[f ]C 0 (0) 0 .
k
yB1
By Lemma 4.41, there is a harmonic polynomial p0 with bounded coefficients such
that
|w(y) p0 (y)| 2+ for |y| .

90

4. LAPLACE EQUATIONS

Now we scale back to get

(2+)k

|u(x) Pk (x)

p0

x
k

| (k+1)(2+)

Clearly we proved the (k + 1)-th step by letting

Pk+1 (x) = Pk (x) + (2+)k p0

x
k

for |x| k+1 .

It is easy to see that Ak , Bk and Ck converge and the limiting polynomial

1
p(x) = xT A x + B x + C
2
satisfies

|Pk (x) p(x)| c |x|2 k + |x|(+1)k + (+2)k c(2+)k ,

for any |x| k . Hence
|u(x) p(x)| |u(x) Pk (x)| + |Pk (x) p(x)| c(+2)k

for any |x| k .

For any x B1 , there exists a nonnegative integer k such that k+1 |x| k .
This implies
|u(x) p(x)| c|x|2+ for any x B1 ,
which yields the desired result.

Next, we put Theorem 4.42 in the classical form of Schauder estimates. We

first introduce Hoder continuous functions in domains.
Definition 4.43. Let k be a nonnegative integer, (0, 1) be a constant and
be a domain in Rn .
(1) A function u is C in , denoted by u C (), if
|u(x) u(y)| c|x y|

for any x, y ,

for some constant c. The Holder semi-norm of u in is defined by

|u(x) u(y)|
.
|x y|
x,y

[u]C () sup

(2) A function u is C k, in , denoted by u C k, (), if u C k () and

u C ().
(3) The C k, -norm in is defined by
k

|u|C k, () = |u|C k () + [k u]C () .

and define C k, ()
similarly. It is
In Definition 4.43, we may replace by
k,
easy to check that C () is a Banach space equipped with the C k, -norm.
For u C (), it is easy to see that u is C at x for any x as in Definition
4.39. In fact, the converse is also true. If u is C at x for any x as in Definition
4.39, then u C (). In other words, the (global) Holder continuity is equivalent
to the everywhere pointwise Holder continuity. We will leave verification as an
exercise.
The following result is often referred to as the interior Schauder regularity or
the interior Schauder estimate.

4.6. WEAK SOLUTIONS

91

Theorem 4.44. Let (0, 1) be a constant and u C 2 (B1 ) satisfy

u = f in B1 .
If f C (B1 ), then u C 2, (B1 ). Moreover,

|u|C 2, (B 1 ) c |u|L (B1 ) + |f |C (B1 ) ,

2

where c is a positive constant depending only on n and .

Proof. We only sketch the proof. For any x B1/2 , we apply Theorem 4.42
in B(1|x|) (x) to obtain a quadratic polynomial px such that
the absolute values of coefficients of px are bounded by c(|u|L (B1 ) + |f |C (B1 ) )
and
|u(y) px (y x)| c(|u|L (B1 ) + |f |C (B1 ) )|y x|2+

for any y B 1|x| (x),

2

where c is a positive constant depending only on n and . Then we have the desired
result easily by the equivalence of the (global) Holder continuity and the everywhere
pointwise Holder continuity.

To end this section, we mention briefly another approach for Schauder estimates. Let be a domain in Rn and f be a continuous function in . We consider
the Poisson equation (4.10) in .
If u C0 () is a solution of (4.10), then f evidently has a compact support
in and hence by Theorem 4.1
Z
u(x) =
(x, y)f (y)dy,

where (x, y) is the fundamental solution of in Rn .

Now let be a bounded domain in Rn and f be a bounded function in . We
define
Z
(4.12)
uf (x) =
(x, y)f (y)dy.

This is called the potential of f in . If uf is C 2 and satisfies (4.10) in , i.e.,

uf = f in , then any solution of (4.10) differs from uf by an addition of a
harmonic function. Since harmonic functions are analytic, regularity of arbitrary
solutions of (4.10) is determined by those of uf defined in (4.12). The classical
approach for the Schauder theory is to prove uf C 2, () and uf = f in if
f C () for some (0, 1). We will not provide proofs here. Example 4.38
shows that uf is not necessarily C 2 if f is only continuous.
4.6. Weak Solutions
In this section, we discuss briefly how to extend the notion of classical solutions
of the Poisson equation to less regularized solutions, solutions with derivatives only
in integral sense, the so-called weak solutions. The same process can be applied to
general linear elliptic equations of second order, even nonlinear elliptic equations, of
divergence form. This section serves only as an introduction to this important advanced topic in the theory of elliptic differential equations. A complete presentation
will constitute a book much thicker than this one.

92

4. LAPLACE EQUATIONS

To introduce weak solutions, we make use of divergence structure or variation

structure of the Laplace operator. Namely, we write the Laplace operator as
u = div(u).
Then the Laplace equation can be considered as an Euler equation of the Dirichlet
energy.
For any domain Rn , we define the Dirichlet energy of u in by
Z
1
|u|2 .
E(u) =
2
Suppose u is a C 1 -minimizer of E(u) among all C 1 -functions in with the same
boundary value. In particular, a small perturbation of u in any subregion of
increases the energy. Hence, for any C 1 -function with a compact support in ,
Z
Z
1
1
|(u + )|2
|u|2 for any .
2
2
In other words,
F ()

1
2

Z
|(u + )|2

has a minimum at = 0. This implies F (0) = 0, or

Z
u = 0 for any C01 ().

In general, for any f L2 (), we consider

Z
Z
1
(4.13)
J(u) =
|u|2 +
uf.
2

If u is a C 1 -minimizer, we obtain similarly

Z
Z

u =
f for any C01 ().

This suggests the following definition.

Definition 4.45. Let f L2 (). Then u C 1 () is a weak solution of
u = f in if
Z
Z

u =
f for any C01 ().

In Definition 4.45, is called a test function. If u is a weak C 2 -solution, we

obtain by a simple integration by parts
Z
Z
u =
f for any C01 ().

This implies easily that

u = f

a.e. in .

If f is continuous in , then u = f in . In other words, a weak C 2 -solution of

the Poisson equation is a classical solution.

4.6. WEAK SOLUTIONS

93

Next, we illustrate how to find a weak solution of the Poisson equation with a
homogeneous Dirichlet boundary value, i.e.,
u =f

(4.14)

in ,

u =0 on .

We intend to minimize J(u) given in (4.13) in the space

u = 0 on }.
C = {u C 1 () C();
We first demonstrate that J has a lower bound in C. First, by the Cauchy inequality,
we have
Z
Z
Z
1

1
f2 2 .
u2 2
|uf |

By Lemma 3.12, the Poincar`e inequality, we have for any u C

Z
Z
2
u C
|u|2 ,

where C is a positive constant depending only on . Then

Z
Z
Z
Z
Z
1
1
1
|uf | C
|u|2 2
|u|2 + C
f 2.
f2 2
4

Hence for any u C

(4.15)

J(u)

1
4

Z
|u|2 C

f 2,

and in particular

Z
f 2.

J(u) C

Therefore, J has a lower bound in C. We set

J0 = inf{J(u); u C}.
If J0 is realized by some u C, then u is a minimizer of J and hence a weak solution
of u = f .
To discuss whether J realizes J0 in C, we consider a minimizing sequence {uk }
C with J(uk ) J0 as k and ask whether there exists a function u C such
that J(uk ) J(u). If such a u exists, then J0 is realized by u by the convergence of
J(uk ). Here we do not require that {uk } converge to u in the C 1 -norm. However, in
order to construct u, {uk } has to converge in some sense. So what is a reasonable
norm to provide such a convergence? To answer this question, we examine the
expression of J(u) in (4.13). First, we have by (4.15)
Z
Z
2
|uk | 4J(uk ) + 4C
f 2.

Since J(uk ) is bounded, there is a uniform bound on the L2 -norms of gradients of

uk . Hence a convergence of uk , if exists, should be related to the L2 -norm. For
this, we need to introduce Sobolev norms.
For any u C01 (), we define the H01 -norm of u by
Z
1

kukH01 =
|u|2 2 .

94

4. LAPLACE EQUATIONS

Here in H01 , the super-index 1 indicates the order of differentiation and the sub-index
0 refers to functions with compact support in . We then complete the space C01 ()
with the H01 -norm. The resulting space is called the Sobolev space H01 () and its
elements are called H01 -functions in . By the Poincar`e inequality, we can prove
H01 () L2 (). In general, functions in H01 () may not have classical derivatives.
However, it has weak derivatives in an integral sense. For any u H01 () and any
i = 1, , n, there exists a function vi L2 () such that
Z
Z
vi for any C01 ().
uxi =

Here vi is called a weak xi -derivative of u and is often denoted by uxi , the same
notion of classical derivatives. Weak derivatives are unique by the density of continuous functions in the L2 -space. Hence, kukH01 can be computed for any u H01 ()
by using weak derivatives of u. By a simple integration by parts, classical derivatives of C 1 -functions are weak derivatives. We also point out that H01 () is in fact
a Hilbert space with the inner product
Z
(u, v) =
u v for any u, v H01 (),

where u and v are weak gradients of u and v.

The following compactness result, called Rellichs Theorem, plays an important
role in the theory of Sobolev spaces. Any sequence in H01 () with bounded H01 ()norms has a subsequence convergent in the L2 ()-norm to some function in H01 ().
We emphasize that the limit function is in H01 () although the convergence is only
in L2 . In general, we do not have a convergence in the H01 -norm.
It is easy to see that the functional J(u) in (4.13) can be extended to functions
u H01 (), with classical gradients replaced by weak gradients. Instead of minimizing J in C01 (), we minimize J in H01 (). One advantage of H01 () over C01 () is
that H01 () is complete under the H01 ()-norm, which is more naturally associated
with the functional J(u) in (4.13). Another important aspect of the Sobolev space
H01 () in studies of elliptic differential equations lies in the following simple fact.
The concept of weak solutions in Definition 4.45 can be extended to H01 -functions.
We also point out that the Poincar`e inequaltiy holds for H01 ()-functions by a simple approximation of H01 ()-functions by C01 ()-functions. Hence, (4.15) also holds
for H01 ()-functions.
We return to our minimizing problem. First, it is easy to see
J0 =

inf

uH01 ()

J(u).

Let {uk } be a minimizing sequence of J(u) in H01 (). Then (4.15) shows that
kuk kH01 is uniformly bounded. By Rellichs theorem, there exists a subsequence
{uk0 } and a u H01 () such that
uk0 u in the L2 -norm as k 0 .
Next, with the help of the Hilbert space structure of H01 () it is not difficult to
prove
J(u) lim
inf J(u0k ).
0
k

4.6. WEAK SOLUTIONS

95

This implies J(u) = J0 . In fact, we can also prove uk0 u in the H01 ()-norm
as k 0 . Now we try to interpret in what sense u solves (4.14). First, u is a
minimizer of J in H01 (), i.e.,
J(u + ) J(u) for any C01 ().
Then by the same analysis preceding Definition 4.45, we have
Z
Z

u =
f for any C01 ().

Here u is the weak gradient of u. Hence, u is a weak solution of u = f .

Concerning the boundary value, we point out that u is not defined in the pointwise
sense. We cannot conclude u = 0 at each point on . Here the boundary condition
u = 0 on is interpreted precisely by the fact u H01 (), i.e., u is the limit of
a sequence of C01 ()-functions in the H01 ()-norm. One consequence is that u|
is a well-defined zero function in L2 (). This finishes the existence part of the
variational approach. We have found a weak solution of (4.14), the Poisson equation
with a homogeneous Dirichlet boundary value.
Now we ask whether u possesses a better regularity. The answer is yes. With
f L2 (), the solution u has weak derivatives of second order in the following
sense. For any 1 i, j n, there exists a function vij L2 () such that
Z
Z
uxi xj =
vij for any C02 ().

We also denote vij by uxi xj . The solution u is a so-called H 2 -function in . With

weak derivatives uxi xj defined as L2 -functions, we also conclude
u = f

a.e. in .

In fact, if f is an H k -function for any k 1, a function with weak derivatives of

order up to k as L2 -functions, then u is an H k+2 -function. In particular, if f is
smooth, then u is smooth. We omit details of discussions.
With similar methods, we can solve the eigenvalue problem of the Laplace
operator in a smooth bounded domain in Rn . Consider
u =u in ,
u =0
First, set
1 = inf

on .

|u|2
; u H01 () \ {0} .
2
u

With a minimizing sequence, weRcan prove that 1 is achieved in H01 (), say by
u1 H01 (). We further assume u21 = 1. Then for any C01 (),
R
|u1 + |2
F () = R
(u + )2
1
has a minimum at = 0. Hence F 0 (0) = 0, or
Z
Z
u1 = 1
u1 for any C01 ().

96

4. LAPLACE EQUATIONS

It is easy to check that 1 > 0. Here 1 is the first eigenvalue of the Laplace
operator and u1 is a corresponding eigenfunction in the weak sense. In fact, we can
and satisfies
prove u1 is a smooth function in
u1 =1 u1

in ,

u1 =0 on
in the classical sense. The definition of 1 yields
R
|u|2
R
1 for any u H01 (),
2
u

or
Z
Z
1
u2
|u|2 for any u H01 ().

Therefore, 1/1 is the best constant in the Poincar`e inequality. See Lemma 3.12.
To find the next eigenvalue, we set
R
Z
|u|2

1
R
2 = inf
; u H0 () \ {0},
uu1 = 0 .
2
u

Obviously, 2 1 . With a similar minimizing

R process, we can prove that 2
1
1
()
with
u u = 0. We further assume
(),
say
by
u

H
is
achieved
in
H
2
0
0
1 2
R 2
=
1.
We
can
check
similarly
u
2
Z
Z
u2 = 2
u2 for any C01 ().

Next, we set
3 = inf

|u|2
; u H01 () \ {0},
2
u

uui = 0, i = 1, 2 .

By continuing this process,

we obtain an increasing sequence {i } and a sequence
R
{ui } H01 () with ui uj = ij such that
Z
Z
ui = i
ui for any C01 ().

for any i = 1, 2, , and

We can prove that ui in fact is a smooth function in
satisfies
ui =i ui

in ,

ui =0 on
in the classical sense. Moreover, {ui } forms a complete sequence in L2 (). In other
words, for any v L2 ()

X
v=
(v, ui )L2 () ui ,
i=1
2

where (, )
is the L -inner product in and the convergence of this series is
in the L2 -norm in .
With such a sequence of eigenvalues and eigenfunctions in a bounded smooth
domain, we can solve initial/boundary-value problems of the heat equation and the
wave equation in arbitrary dimension, as described at the end of Section 3.3.
L2 ()

EXERCISES

97

Exercises
(1) Suppose u(x) is harmonic in some domain in Rn . Then
x
v(x) = |x|2n u
|x|2
is also harmonic in a suitable domain.
(2) For n = 2, find the Greens function for the Laplace operator on the first
quadrant.
(3) Find the Greens function for the Laplace operator in the upper half space
{xn > 0} and then derive a formal integral representation for a solution
of the Dirichlet problem
u = 0
u=

in {xn > 0},

on {xn = 0}.

(4) Let be a positive constant. Find the fundamental solution for 3 u+u =
0 in R3 .
(5) (a) Prove by the Poisson formula the following Harnack inequality: Suppose u is harmonic in BR (x0 ) and u 0. Then
n2

n2

Rr
R
R+r
R
u(x0 ) u(x)
u(x0 ),
R+r
R+r
Rr
Rr
where r = |x x0 | < R.
(b) Prove by (a) the Liouville Theorem: If u is a harmonic function in
Rn and bounded above or below, then u const.
R
(6) Let u be a harmonic function in Rn with Rn |u|p < for some p (1, ).
Then u 0.
(7) Let u be a harmonic function in Rn and u(x) = O(|x|m ) as |x| for
a positive integer m. Then u is a polynomial of degree m.
+ ) is harmonic in B + = {x B1 ; xn > 0} with u = 0
(8) Suppose u C(B
1
1
on {xn = 0} B1 . Then the odd extension of u in B1 is harmonic in B1 .
(9) Suppose u is harmonic in the open upper half space {xn > 0} and continuous in its closure. If u = 0 on {xn = 0} and u(x) = O(|x|) as |x| ,
then u(x) = cxn for some constant c.
satisfies u + u = 0 in
(10) Let be a domain in Rn and u C 2 () C()
for a constant > 0. Then u is analytic in .
(11) Suppose is a domain in Rn and {um } is a sequence of uniformly bounded
harmonic functions in . Prove that {um } has a subsequence convergent
to a function u uniformly on any compact subsets of and that u is
harmonic in .
(12) Prove that the square of a nonnegative subharmonic function is subharmonic. Give an example to show that the condition nonnegative cannot
be omitted.

98

4. LAPLACE EQUATIONS

(13) Suppose u is harmonic in the open upper half space {xn > 0} and continuous and bounded in its closure. Then
sup u = sup u.
{xn 0}

{xn =0}

(14) Let u be a C 2 -solution of

u = 0
u=0

in Rn \ BR ,
on BR .

Prove that u 0 if
u(x)
= 0 for n = 2,
|x| ln |x|
lim u(x) = 0 for n 3.
lim

|x|

satisfies
(15) Suppose u C 2 () C 1 ()
u + u3 = 0

in ,

u
+ u = on ,
n
for continuous functions and on with (x) 0 > 0. Prove
1
max ||.
max |u|

0
(16) Let be a smooth bounded domain in Rn with = 1 2 , c be a
continuous function in with c 0 and be a continuous function on
with 0. Discuss the uniqueness of the following problem
u + cu = f
u
+ u =
n
u=

in ,
on 1 ,
on 2 .

(17) Let k and > 0 be constants and B1+ be the upper half disc in R2 .
+ ) is a solution of the following problem
Suppose u C 2 (B1+ ) C(B
1
k
u + uy u = f in B1 ,
y
u = 0 on B1+ .
Then
max
|u|
+

B
1

1
max |f |.
B1+

(18) Let u be a nonzero harmonic function in B1 Rn and set

R
r Br |u|2
N (r) = R
for any r (0, 1).
u2
Br
(a) Prove N (r) is a nondecreasing function in r (0, 1) and identify
lim N (r).

r0+

EXERCISES

99

(b) Prove for any 0 < r < R < 1

2N (R)
Z
Z
R
1
1
2
u2 .
u

Rn1 BR
r
rn1 Br
Remark: The quantity N (r) is called the frequency. The estimate in (b)
is referred to as the doubling condition for R = 2r.
(19) Let k be a nonnegative integer, (0, 1) be a constant and u be a C k function in [0, 1]. Suppose for any x [0, 1]
|u(y) Tk (y; x)| C|y x|k+

for any y [0, 1],

where Tk (; x) is the k-th Taylor expansion of u at x. Then u C k, [0, 1]

and [k u]C [0,1] C.

CHAPTER 5

Heat Equations
The n-dimensional heat equation is given by ut u = 0 for functions u =
u(x, t), with x Rn and t R. Here, x is called the space variable and t the time
variable. The heat equation models the temperature of a body conducting heat,
when the density is constant. In this chapter, we discuss properties of solutions
of the heat equation. In Section 5.1, we discuss the fundamental solution of the
heat equation and solve initial-value problems. Then in Section 5.2, we use the
fundamental solution to discuss regularity of arbitrary solutions. In Section 5.3,
we discuss the maximum principle for the heat equation and its applications. In
particular, we use the maximum principle to derive interior gradient estimates. In
Section 5.4, we discuss Harnack inequalities.
5.1. Initial-Value Problems
In this section, we discuss initial-value problems of the heat equation and derive
an explicit expression for their solutions.
We first examine the equation itself. The n-dimensional heat equation is given
by
(5.1)

ut u = 0,
n

for u = u(x, t) with x R and t R. We note that (5.1) is not preserved by

the change t 7 t. This indicates that the heat equation describes an irreversible
process and distinguishes between past and future. This fact will be well illustrated
by Harnack inequalities which we will derive in Section 5.4. Next, (5.1) is preserved
under linear transforms x0 = x and t0 = 2 t for any nonzero constant , which
leave the quotient |x|2 /t invariant. Because of this, the expression |x|2 /t appears
frequently in connection with the heat equation (5.1). In fact, the fundamental
solution we will derive has such an expression.
If u is a solution of (5.1) in a domain in Rn R, then for any (x0 , t0 ) in this
domain and any r > 0
u
(x, t) = u(x0 + rx, t0 + r2 t)
is a solution of (5.1) in an appropriate domain in Rn R.
Let u be a polynomial in Rn R. It is of p-homogeneous degree d if
u(x, 2 t) = d u(x, t) for any (x, t) Rn R and > 0.
Now we let P be a homogeneous polynomial of degree d in Rn and proceed to find
a p-homogeneous polynomial of degree d such that
ut u =0
u(, 0) =P
101

in Rn R,
in Rn .

102

5. HEAT EQUATIONS

To do this, we simply expand u as a power function of t with coefficients as functions

of x, i.e.,

X
u(x, t) =
ak (x)tk .
k=0

Then a simple calculation shows that

a0 = P,

ak =

1
ak1
k

for any k 1.

Since P is a polynomial of degree d, then [d/2]+1 P = 0, where [d/2] is the integral

part of d/2, i.e., [d/2] = d/2 if d is an even integer and [d/2] = (d 1)/2 if d is an
odd integer. Hence
d

u(x, t) =

[2]
X

k P (x)tk .

k=0

For n = 1, let ud be a p-homogeneous polynomial of degree d in R R satisfying

the heat equation and ud (x, 0) = xd . The first five such polynomials are given by
u1 (x, t) =x,
u2 (x, t) =x2 + 2t,
u3 (x, t) =x3 + 6xt,
u4 (x, t) =x4 + 12x2 t + 12t2 ,
u5 (x, t) =x5 + 20x3 t + 60xt2 .
In the following, we discuss initial-value problem of the heat equation and derive
an explicit expression for its solution.
We first introduce briefly Fourier transforms. Fourier transforms are an important subject and has a close connection with many fields in mathematics, especially
with partial differential equations.
By allowing functions to be complex valued, we define the Schwartz class S as
the collection of all functions u C (Rn ; C) such that
m

(1 + |x|2 ) 2 x u(x) is bounded in Rn for any Zn+ and m Z+ .

In other words, the Schwartz class consists of smooth functions whose arbitrary
2
derivatives decay faster than any polynomials. For example, u(x) = e|x| is in the
Schwartz class.
For any u S, define a new function u
in Rn by
Z
1
u
() =
eix u(x)dx for any Rn .
n
(2) 2 Rn
The operator u 7 u
is called the Fourier transform operator and u
is called the
Fourier transform of u.
We now state some properties of Fourier transforms.
(F1) Linearity: The Fourier transform operator is linear, i.e., for any u1 , u2 S
and c1 , c2 C
(cu1 + c2 u2 ) = c1 u
1 + c2 u
2 .

5.1. INITIAL-VALUE PROBLEMS

103

(F2) Differentiation: For any u S and any j = 1, , n, we have by integration by parts

Z
1
d
u()
=
eix xj u(x)dx = ij u
(),
n
j
(2) 2 Rn
and hence for any multi-index Zn+

d
().
x u() = (i) u

For any polynomial P () on Rn given by

X
a ,
P () =
||m

we define a differential operator P () by

X
P ()u =
a u.
||

Then

P\
()u() = P (i)
u().

(F3) Multiplication by polynomials: It is easy to see that u

C (Rn ) for any
u S. Then for any j = 1, , n
Z
i
eix xj u(x)dx = id
xj u(),
j u
() =
n
(2) 2 Rn
and hence for any multi-index Zn+
u().
u
() = (i)|| xd

As a consequence, we have u
S for u S, i.e., | u
()| is bounded in Rn
n
for any , Z+ . To see this, we note by (F2) and (F3)
d
\
u() = (i)||+|| (i) x
u() = (i)||+||

u
() =(i)|| xd
x (x u)()
Z

1
||+||
=
eix x x u(x) dx.
n (i)
(2) 2
Rn
Hence

1
|x x u(x) |dx < ,
n
(2) 2 Rn
Rn
since each term in the integrand decays faster than any polynomials.
(F4) Translation and dilation: For any u S, a Rn and k R,
sup | u
()|

\
u(
a)() = eia u
(),
and

[
u(k)()
=
u
( ).
|k|n k
(F5) Convolution: For any u, v S, it is easy to check u v S, where u v
is the convolution of u and v defined by
Z
(u v)(x) =
u(x y)v(y)dy.
Rn

We then have

u[
v() = (2) 2 u
()
v ().

104

5. HEAT EQUATIONS

We leave verification of Properties (F1)-(F5) as an exercise.

The following example will be useful in discussions of the heat equation. We
first note
Z

2
ex dx = .

Example 5.1. Consider u(x) = e|x| in Rn . Then u S and a direct calculation yields
2
1
1
u
() = n e 4 || .
22
2

Therefore by (F4), for f (x) = eA|x| with A > 0,

f() =

||2
1
4A
.
n e
(2A) 2

The Fourier transform is an important subject in mathematics. Now we attempt to use it to solve certain classes of partial differential equations. Let P be
a polynomial in Rn of degree m. Consider the following linear partial differential
equation of degree m
P ()u = f in Rn .
By applying Fourier transforms and (F2), we have
P (i)
u() = f(),
and then
u
() =

f()
.
P (i)

Hence a solution u is a function whose Fourier transform is f()/P (i). This

naturally leads to inverse Fourier transforms.
Theorem 5.2. Suppose u S. Then
Z
1
eix u
()d.
u(x) =
n
(2) 2 Rn
The right hand side is simply the Fourier transform of u
evaluated at x.
(x).
Hence, u(x) = u
Proof. First, by the definition of the Fourier transform, we have

Z Z
1
ix iy
R.H.S =
e
e
u(y)dy
d.
(2)n Rn
Rn
We cannot simply change the order of integrations since there is no absolute convergence. Instead, we introduce a limiting process to have the absolute convergence.
Then
Z Z
||2
1
R.H.S. =
lim
ei(xy) e u(y)dyd
n
(2) Rn Rn
Z Z
||2
1
i(xy)
lim
e
u(y)ddy,
=
(2)n Rn Rn

5.1. INITIAL-VALUE PROBLEMS

105

where we changed the order of integrations in the last step. A simple calculation
of the exponential factor yields
Z
Z
(xj yj ) j 2
P
1
n
+

|xy|2
j=1 i
2
4

ddy.
lim
u(y)e
e
R.H.S. =
(2)n Rn
Rn
For any a R, we first note
Z
Z
2

2
Z 2

i a 2 +

d =
e
e
d = e
d = .
R

This integral is independent of a. Hence

Z
(xj yj ) j 2
(x y ) j 2
n Z
P
Y
n
n
i j 2j
+
+
j=1 i
2

e
d =
e
dj = () 2 .
Rn

j=1

Then

1
R.H.S. =
lim
n
(4) 2

Z
n

2 e

|xy|2
4

u(y)dy.

Rn

Let

|z|2
n
1
2
4
.
n e
(4) 2
R
Then K > 0 in Rn and Rn K = 1. It is straightforward to prove
Z
R.H.S. = lim
K (x y)u(y)dy = u(x).

K (z) =

Rn

We will provide detail in the proof of Theorem 5.5.

Next, we set

Z
(u, v)L2 (Rn ) =

u
v for any u, v L2 (Rn ).
Rn

Here u and v may be complex-valued. The next result is referred to as the Parsevals
identity.
Theorem 5.3. For any u, v S,
(u, v)L2 (Rn ) = (
u, v)L2 (Rn ) .
In particular, for any u S
kukL2 (Rn ) = k
ukL2 (Rn ) .
Proof. We note

(
u, v)L2 (Rn ) =
v ()d =
v()eix u(x)dxd
u
()
Rn Rn
Rn
Z Z
Z
=
v()eix u(x)ddx =
u(x)
v (x)dx = (u, v)L2 (Rn ) ,
Rn

Rn

Rn

where we applied Theorem 5.2 to v. This yields the desired result.

We now extend Fourier transforms to a larger class of functions than S. Note

that the Fourier transform is a linear operator from S to S. Since S is dense in
L2 (Rn ), we can complete the space S under k kL2 (Rn ) to get L2 (Rn ). Hence, the
Fourier transform can be extended to a bounded linear operator in L2 (Rn ). Since
it is an isometry in S with respect to the L2 -norm, it is also an isometry in L2 (Rn ).

106

5. HEAT EQUATIONS

We note that the Fourier transform also extends to functions in L1 (Rn ). For
any u S,
Z
1
u
() =
eix u(x)dx.
n
(2) 2 Rn
Hence
Z
1
|
u()|
|u(x)|dx,
n
(2) 2 Rn
or
1
sup |
u|
n kukL1 (Rn ) .
2
n
(2)
R
1
n
Note that S is dense in L (R ). By completing S under the L1 (Rn ), we conclude
that the Fourier transform operator is a bounded linear operator from L1 (Rn ) to
the set of bounded continuous functions in Rn .
Now we are ready to derive an expression for solutions of general initial-value
problems of the heat equation. We consider
(5.2)

ut u = 0
u(, 0) = u0

in Rn (0, ),
in Rn .

Here we require u C 2 (Rn (0, )) C(Rn [0, )). Although called an initialvalue problem, (5.2) is not the type of initial-value problems we discussed in Section
3.1. The heat equation is of the second order, while only one condition is prescribed
on the initial hypersurface {t = 0}, which is characteristic.
We first use Fourier transforms to derive a formal solution. In the following,
we employ the Fourier transform of u with respect to x Rn and write
Z
1
eix u(x, t)dx.
u
(, t) =
n
(2) 2 Rn
Then by (F2)
u
t + ||2 u
=0
u
(, 0) = u
0

in Rn (0, ),
in Rn .

This is an initial-value problem for an ODE with as a parameter and its solution
is given by
2
u
(, t) = u
0 ()e|| t .
Now we treat t as a parameter. Let K(x, t) satisfy
t) =
K(,

1
||2 t
n e
2
(2)

for t > 0.

Then
n
t)
u
(, t) = (2) 2 K(,
u0 ().

Therefore Theorem 5.2 and (F5) imply

Z
u(x, t) =
K(x y, t)u0 (y)dy.
Rn

This is called the Poisson integral formula. By Theorem 5.2 and Example 5.1, we
have
Z
2
1
K(x, t) =
eix e|| t d,
(2)n Rn

5.1. INITIAL-VALUE PROBLEMS

107

or
(5.3)

K(x, t) =

|x|2
1
4t
n e
(4t) 2

for any x Rn , t > 0,

which is called the fundamental solution of the heat equation.

Now we verify directly that the Poisson integral formula defines a solution under
suitable conditions on u0 . The proof follows the same line as that of the Poisson
integral formula for the Laplace equation in Theorem 4.6.
We first note that the fundamental solution K satisfies the following properties:
(K1) K(x, t) is smooth for any x Rn and t > 0.
(K2) K(x, t) > 0 for any x Rn and t > 0.
n
(K3) (
R t x )K(x, t) = 0 for any x R and t > 0.
(K4) Rn K(, t) = 1 for any t > 0.
(K5) For any > 0,
Z
K(x, t)dx = 0.
lim
t0+

|x|>

Figure 5.1. Graphs of fundamental solutions for t2 > t1 > 0.

Here (K1) and (K2) follow from the explicit expression for K in (5.3). We may
also get (K3) from (5.3) by a straightforward calculation. However, an easy way to
get (K3) is to use the integral expression of K preceding (5.3). For (K4) and (K5),
we simply note
Z
Z
2
1
K(x, t)dx = n
e|| d.
2

|x|>
||>
2

This implies (K4) for = 0 and (K5) for > 0.

Now we are ready to solve initial-value problems of the heat equation.
Theorem 5.4. For any bounded u0 C(Rn ), the function u defined by
Z
(5.4)
u(x, t) =
K(x y, t)u0 (y)dy
Rn

is smooth in R (0, ) and continuous up to t = 0 and satisfies

ut u = 0
u(, 0) = u0

in Rn (0, ),
in Rn .

108

5. HEAT EQUATIONS

Proof. Step 1. We first prove that u(x, t) is smooth for any x Rn and t > 0.
It suffices to check
Z
|xy|2
|x y|m e 4t |u0 (y)|dy < ,
Rn

for any m 0. This follows from the exponential decay of the integrand. Then u is
a solution of the heat equation by a simple differentiation. We point out for future
references that we only use the boundedness of u0 .
Step 2. For any fixed x0 Rn , we claim
lim

(x,t)(x0 ,0)

u(x, t) = u0 (x0 ).

For simplicity, we assume x0 = 0. By (K4), we have

Z

K(x y, t) u0 (y) u0 (0) dy.

u(x, t) u0 (0) =
Rn

Then

Z
|xy|2
1
e 4t |u0 (y) u0 (0)|dy.
n
(4t) 2 Rn
For any > 0, by the continuity of u0 , there exists a positive constant = ()
such that
|u0 (y) u0 (0)| < for any |y| .
|u(x, t) u0 (0)|

Next, we assume |u0 | M by the boundedness of u0 . Then by (K2) and (K4)

|u(x, t) u0 (0)|
Z
Z

K(x y, t)|u0 (y) u0 (0)|dy +

K(x y, t)|u0 (y) u0 (0)|dy
B
Rn \B
Z
+ 2M
K(x y, t)dy.
Rn \B

We note |x y| /2 for any y Rn \ B and x B/2 . Hence

Z
|u(x, t) u0 (0)| + 2M
K(z, t)dz for any x B .
2

Rn \B

By (K5), there exists a t0 > 0 depending only on , M and such that

|u(x, t) u0 (0)| 2

for any x B , t (0, t0 ).

2

We then have the desired result.

Now we discuss a result more general than Theorem 5.4. The boundedness
assumption on u0 in Theorem 5.4 can be relaxed. To seek a reasonable assumption
on initial values, we examine the expression of the fundamental solution K. We note
that K in (5.3) has an exponential decay in space variables with a large decay rate
for small time. This suggests that we can allow an exponential growth for initial
values. In the convolution (5.4), a fixed exponential growth from initial values can
be offset by the fast exponential decay in the fundamental solution at least for a
short time period. To see this clearly, we consider an example. For any > 0, set
G(x, t) =

1
14t |x|
n e
(1 4t) 2

for any x Rn , t <

1
.
4

5.1. INITIAL-VALUE PROBLEMS

109

It is straightforward to check
for any x Rn , t <

Gt G = 0
Note

1
.
4

G(x, 0) = e|x| for any x Rn .

Hence, G has an exponential growth initially for t = 0, and in fact for any t < 1/4.
The growth rate becomes arbitrarily large as t approaches 1/4 and G does not
exist beyond t = 1/4.

Figure 5.2. Graphs of G for different time.

Now we formulate a general result. If u0 is continuous and has an exponential
growth, then (5.4) still defines a solution of the initial-value problem in a short
period of time.
Theorem 5.5. Suppose u0 C(Rn ) satisfy
2

|u0 (x)| M eA|x|

for any x Rn ,

for some constants M, A 0. Then u defined in (5.4) is a C (Rn (0, T ))

C(Rn [0, T ))-solution of
in Rn (0, T ],

ut u = 0
u(, 0) = u0
for any T <

1
4A .

in Rn ,

Moreover,
|u(x, t)| M1 eA1 |x|

for any x Rn , 0 < t T,

where A1 0 and M1 > 0 are constants depending only on A, M and T .

We first prove the following lemma.
Lemma 5.6. Suppose f, g C(Rn ) satisfy

eA|x| ,
|f (x)| M eA|x| , |g(x)| M
> 0 and some A, A.
If A + A < 0, then
for some constants M, M
Z
h(x) =
g(x y)f (y)dy
Rn

is continuous in Rn and

eA|x| ,
|h(x)| M
and A are constants depending only on M, M
, A and A.

where M

110

5. HEAT EQUATIONS

Proof. With A + A < 0, we have

Z
2
2

eA|y| +A|xy| dy
|h(x)| M M
n
R
Z

2
2
AA
A

|x|
e A+
A
= MM
e(A+A)|y A+A x| dy
Rn

n AA |x|2
2 e A+
A
= MM
.
A + A
This yields the desired estimate.

Now we prove Theorem 5.5. The proof follows the same line as that of Theorem
5.4.
Proof of Theorem 5.5. Step 1. The case A = 0 is covered by Theorem 5.4.
1
1
We only consider A > 0 and take T such that 4T
+ A < 0 or T < 4A
. Then
n
Lemma 5.6 implies u is continuous in R (0, T ]. To show that u has continuous
derivatives of arbitrary order in Rn (0, T ], we only need to verify
Z
|xy|2
|x y|m e 4t |u0 (y)|dy < ,
Rn

for any m 0. This can be proved similarly as the proof of Lemma 5.6.
Step 2. For any fixed x0 Rn , we claim
lim

(x,t)(x0 ,0)

u(x, t) = u0 (x0 ).

For simplicity, we assume x0 = 0. By (K4), we have

Z

|xy|2
1
u(x, t) u0 (0) =
e 4t u0 (y) u0 (0) dy.
n
(4t) 2 Rn
Set v(y) = u0 (y) u0 (0). Then
1
|u(x, t) u0 (0)|
n
(4t) 2

Z
e

|xy|2
4t

|v(y)|dy.

Rn

For any > 0, by the continuity of v, there exists a = () such that

|v(y)| < for any |y| .
e
Note |v(y)| M
B > 0 such that

2
A|y|

and A.
Then there exists a constant
for some constants M
2

|v(y)| eB|y|
In fact, we can choose B such that e

for any |y| .

. Then by eB|y|2 1, we have
M

2
(BA)

|v(y)| eB|y|

for any y Rn .

Therefore,

Z
|xy|2
2

|u(x, t) u0 (0)|
e 4t eB|y| dy.
n
(4t) 2 Rn
A direct calculation then yields
2
B
1

14Bt |x|
for any t <
.
|u(x, t) u0 (0)|
n e
4B
(1 4Bt) 2
We have the desired result by letting (x, t) (0, 0).

5.1. INITIAL-VALUE PROBLEMS

111

Now we discuss some properties of the solution (5.4) of the initial-value problem
(5.2). First from (5.4), u(x, t) for t > 0 depends on values of u0 at all points.
Equivalently, values of u0 near a point x0 Rn affect the value of u(x, t) at all x
as long as t > 0. Thus values here travel at infinite speed.
Next, the function u(x, t) in (5.4) becomes smooth for t > 0, even if the initial
value u0 is simply bounded. This is well illustrated in Step 1 in the proof of Theorem
5.4. We did not use any regularity assumption on u0 there. Compare with Theorem
3.16. Later on, we will prove a general result that any solutions of the heat equation
in a domain in Rn (0, ) are smooth away from the boundary. Moreover, for
each fixed time, any solutions of the heat equation are analytic in space variables
away from the boundary.
We need to point out that (5.4) represents only one of infinitely many solutions.
Solutions are not unique without further conditions on u, like the boundedness
assumption or the exponential growth assumption. In fact, there exists a nontrivial
solution u C (Rn R) of ut u = 0, with u 0 for t 0. In the following,
we construct such a solution of the 1-dimensional heat equation due to Tychonov.
Proposition 5.7. There exists a nonzero smooth function u C (R [0, ))
satisfying
ut uxx =0
u(, 0) =0

in R [0, ),
in R.

Proof. We construct a smooth function in R R such that ut uxx = 0 in

R R and u 0 for t < 0. We treat x = 0 as the initial curve and attempt to find
a smooth solution of the following initial-value problem
ut uxx = 0

in R R,

u(0, t) = a(t), ux (0, t) = 0

for any t R,

for an appropriate function a in R. We write u as a power series in x as

X
u(x, t) =
ak (t)xk .
k=1

By a simple substitution in ut = uxx and a comparison of coefficients of power of

x, we have
a0 = a, a1 = 0, a0k = (k + 1)(k + 2)ak+2 for any k 0.
Therefore, we have a formal solution

X
u(x, t) =
k=0

1 (k)
a (t)x2k .
(2k)!

We choose a(t) appropriately so that u(x, t) defined above is a smooth function and
is identically zero for t < 0. To this end, we define
( 1
e t2 for t > 0,
a(t) =
0
for t 0.
Then it is straightforward to verify that the series defining u is absolutely convergent
in RR. This implies that u is continuous. Similarly, we can prove that u is smooth.
We skip details and leave the rest of the proof as an exercise.

112

5. HEAT EQUATIONS

To end this section, we discuss briefly terminal-value problems. For a fixed

constant T > 0, we consider
ut uxx = 0

in R (0, T ),

u(, T ) = in R.
Here the initial value is prescribed at the terminal time T . We are interested in
the well-posedness of this problem. First, we ask whether there exists a function a
in R such that the solution u of the initial-value problem
ut uxx = 0

in R (0, T ),

u(, 0) = a in R
satisfies
u(, T ) =

in R.

Obviously, a necessary condition is given by C (R). In fact, has to be

analytic, as will be shown by Corollary 5.11.
Now we consider an example. For any positive integer m, we set
m (x) = um (x, T ) = sin(mx).
Then, a solution is given by
um (x, t) = em
with

(T t)

sin(mx),
2

am (x) = vm (x, 0) = em

sin(mx).

We note
max |m | = ,
R

and
max |am | = em
R

as m .

There is no continuous dependence on initial values (prescribed at the terminal

time T ).
5.2. Regularity of Solutions
In this section, we discuss regularity of solutions of the heat equation with the
help of the fundamental solution.
We often discuss the heat equation ut u = 0 in cylinders of the form
(t0 , t1 ], where is a domain in Rn and t0 < t1 are two scalars. The parabolic
boundary p ( (t0 , t1 ]) is defined by

p (t0 , t1 ] = {t0 } [t0 , t1 ) .

In other words, parabolic boundary consists of the bottom and the side of the
boundary. For any (x0 , t0 ) Rn R and any r > 0, we define
Qr (x0 , t0 ) = Br (x0 ) (t0 r2 , t0 ].
We note that Qr (x0 , t0 ) for the heat equation play the same role as balls for the
Laplace equation. If u is a solution of the heat equation ut u = 0 in Qr (0), then
ur (x, t) = u(rx, r2 t)

5.2. REGULARITY OF SOLUTIONS

113

Figure 5.3. The region QR (x0 , t0 ).

is a solution of the heat equation in Q1 (0). For any (x, t) and (y, s), the parabolic
distance is defined by

1
dP (x, t), (y, s) = |x y|2 + |t s| 2 .
In the following, we denote by C 2,1 the collection of functions which are C 2 in
x and C 1 in t. We first have the following regularity result for solutions of the heat
equation.
Theorem 5.8. Let u be a C 2,1 -solution of ut u = 0 in QR (x0 , t0 ) for some
R > 0. Then u is smooth in QR (x0 , t0 ).
Proof. We claim for any (x, t) QR (x0 , t0 )
Z

u(x, t) =
K x y, t (t0 R2 ) u(y, t0 R2 )dy
Z

BR (x0 )

+
t0 R2

u
K
K(x y, t s)
(y, s) u(y, s)
(x y, t s) dSy ds.
n
n
y
y
BR (x0 )

This implies the smoothness of u in QR (x0 , t0 ) easily since the boundary integrals
in the right-hand side are only on the parabolic boundary of BR (x0 ) (t0 R2 , t]
and there is no singularity for integrands for (x, t) QR (x0 , t0 ).
We denote by (y, s) points in QR (x0 , t0 ). Set
|xy|2
1

K = K(x y, t s) =
n2 e 4(ts)
4(t s)

for s < t.

Then
Ks + y K = 0,
and hence
0 =K(us y u) = (uK)s +

n
X

(uKyi Kuyi )yi u(Ks + y K)

i=1

=(uK)s +

n
X
i=1

(uKyi Kuyi )yi .

114

5. HEAT EQUATIONS

For any fixed (x, t) QR (x0 , t0 ) and any > 0, we integrate with respect to (y, s)
in BR (x0 ) (t0 R2 , t ). Then
Z
Z

K(x y, )u(y, t )dy =

K x y, t (t0 R2 ) u(y, t0 R2 )dy
BR (x0 )
Z t

+
t0 R2

BR (x0 )

K
u
(y, s) u(y, s)
(x y, t s) dSy ds.
K(x y, t s)
ny
ny
BR (x0 )

Now it suffices to prove

lim

0+

K(x y, )u(y, t ) = u(x, t).

BR (x0 )

The proof is almost identical to Step 2 in the proof of Theorem 5.4. The integral
over a finite domain here introduces few changes. We omit details.

Now we prove the interior gradient estimates by a similar method.

Theorem 5.9. Let u be a C 2,1 -solution of ut u = 0 in QR (x0 , t0 ) for some
R > 0. Then
c
|x u(x0 , t0 )|
sup |u|,
R QR (x0 ,t0 )
where c is a positive constant depending only on n.

Figure 5.4. Domains for interior gradient estimates.

Proof. We first modify the proof of Theorem 5.8 to express u using the fundamental solution and cutoff functions. We denote by (y, s) points in QR (x0 , t0 ).
For any smooth function v is BR (x0 ) (t0 R2 , t0 ], we have
n
X
0 = v(us y u) = (uv)s +
(uvyi vuyi )yi u(vs + y v).
i=1

We first choose a cutoff function C (QR (x0 , t0 )) such that

supp Q 43 R (x0 , t0 ),

1 in Q 21 R (x0 , t0 ).

Let K be the fundamental solution of the heat equation as in (5.3) and set for any
fixed (x, t) QR/4 (x0 , t0 )
v = K = (y, s)K(x y, t s)

for s < t.

For any > 0, we integrate with respect to (y, s) in BR (x0 ) (t0 R2 , t ).

We first note that there is no boundary integral over the parabolic boundary of

5.2. REGULARITY OF SOLUTIONS

115

BR (x0 ) (t0 R2 , t ) since the support of has an empty intersection with this
part of the boundary. Hence
Z
Z
(u)(y, t )K(x y, )dy =
u(s + y )(K)dyds.
BR (x0 )(t0 R2 ,t)

BR (x0 )

Then similarly as in the proof of Theorem 5.4, with 0 we have

Z
(x, t)u(x, t) =
u(s + y )(K)dyds.
BR (x0 )(t0 R2 ,t)

With
K=

1
4(t s)

n2 e

|xy|2

4(ts)

for s < t,

we note
Ks + y K = 0.
Hence for any (x, t) QR/4 (x0 , t0 )
Z
u(x, t) =
u(s K + y K + 2y y K)dyds.
BR (x0 )(t0 R2 ,t)

We note that each term in the integrand involves a derivative of , which is zero in
QR/2 (x0 , t0 ) since 1 there. Then the integral domain D is given by
3
1
D = B 34 R (x0 ) (t0 ( R)2 , t] \ B 12 R (x0 ) (t0 ( R)2 , t].
4
2
Hence the distance between any (y, s) D and any (x, t) QR/4 (x0 , t0 ) has a
positive lower bound. Therefore, the integrand has no singularity in the integral
domain. (This gives an alternate proof of the smoothness of u in QR/4 (x0 , t0 ).)
Next, we have for any (x, t) QR/4 (x0 , t0 )
Z

x u(x, t) =
u (s + y )x K + 2y x y K dyds.
D

For the cutoff function in QR (x0 , t0 ), we require further

c
c
|y | , |s | + |2y | 2 ,
R
R
where c is a positive constant depending only on n. With the explicit expression of
K, we have
2
|x y| |xy|
4(ts) ,
|x K| c
e
n
(t s) 2 +1
and
2
|x y|2 + (t s) |xy|
4(ts) .
|2x K| c
e
n
(t s) 2 +2
Obviously, for any (x, t) QR/4 (x0 , t0 ) and any (y, s) D,
|x y| R,

0 < t s R2 .

Therefore, for any (x, t) QR/4 (x0 , t0 ),

Z
|xy|2
|xy|2

R
1
4(ts)
4(ts)
e
e
|x u(x, t)| c
+
|u(y, s)|dyds.
n
n
+1
(t s) 2 +2
D R(t s) 2

116

5. HEAT EQUATIONS

Now we claim

Z
D

|xy|2
c
1
e 4(ts) dyds 2i2
n
+i
R
(t s) 2

for i = 1, 2.

Then we obtain easily

|x u(x, t)|

c
sup |u| for any (x, t) Q R (x0 , t0 ).
4
R QR (x0 ,t0 )

Next, we prove the claim. We decompose D into two parts

3
1
D1 =B R (x0 ) t0 ( R)2 , t0 ( R)2 ,
2
4
2

3
D2 = B 34 R (x0 ) \ B 12 R (x0 ) t0 ( R)2 , t .
4
For any (x, t) QR/4 (x0 , t0 ) and (y, s) D1 , we have
ts

1 2
R .
4

Hence for any (x, t) QR/4 (x0 , t0 )

Z
Z
|xy|2
1
2n+2i
c
4(ts)
dyds

e
dyds 2i2 ,
n
+i
n+2i
2
R
D1 (t s)
D1 R
where we used the fact that the volume of D1 has an order of Rn+2 . Now we
consider D2 . For any (x, t) QR/4 (x0 , t0 ) and (y, s) D2 , we have
|y x|
and hence

1
R,
4

2
|xy|2
1
1
R
e 4(ts)
e 43 (ts) .
n
n
+i
+i
(t s) 2
(t s) 2

Then for any (x, t) QR/4 (x0 , t0 ) and (y, s) D2 ,

Z
Z t
R2
|xy|2
1
1
43 (ts)
4(ts)
n
I
e
dyds

cR
e
ds.
n
n
+i
+i
D2 (t s) 2
t0 ( 43 R)2 (t s) 2
With t s = R2 , we obtain
I

c
R2i2

( 34 )2

1 31
c
e 4 d 2i2 .
n
R
2 +i

This finishes the proof of the claim.

Next, we estimate derivatives of arbitrary order.

Corollary 5.10. Let u be a C 2,1 -solution of ut u = 0 in QR (x0 , t0 ) for
some R > 0. Then
|tk m
x u(x0 , t0 )|

cm+2k k m+2k1
n e
(m + 2k)! sup |u|,
Rm+2k
QR (x0 ,t0 )

where c is a positive constant depending only on n.

5.3. THE MAXIMUM PRINCIPLE

117

Proof. For x-derivatives, we proceed as in the proof of Theorem 4.9 and

obtain for any multi-index with || = m
|x u(x0 , t0 )|

cm em1 m!
sup |u|.
Rm
QR (x0 ,t0 )

For t-derivatives, we have ut = u and hence tk u = k u for any nonnegative

integer k. We note that there are nk terms of x-derivatives of u of order 2k in k u.
Hence
k
|tk m
max |x u(x0 , t0 )|.
x u(x0 , t0 )| n
||=m+2k

This implies the desired result easily.

Now we prove the analyticity of solutions of the heat equation on any time
slice.
Corollary 5.11. Let u be a C 2,1 -solution of ut u = 0 in QR (x0 , t0 ) for
some R > 0. Then u(, t) is analytic in BR (x0 ) for any t (t0 R2 , t0 ]. Moreover,
for each nonnegative integer k, tk u(, t) is analytic in BR (x0 ) for any t (t0
R2 , t0 ].
The proof is identical to that of Theorem 4.10 and is omitted. As shown in
Proposition 5.7, solutions of ut u = 0 are not analytic in t in general.
5.3. The Maximum Principle
In this section, we discuss the maximum principle for the heat equation and its
applications. Suppose Rn is a bounded smooth domain. For any T > 0, set
QT = (0, T ] = {(x, t); x , 0 < t T },
and
p QT = ( {t = 0}) ( [0, T ]).
As in the previous section, p QT is called the parabolic boundary of QT . We denote
by C 2,1 (QT ) the collection of functions which are C 2 in x and C 1 in t in QT .
Without confusion, we simply write Q instead QT .
We first prove the maximum principle, which asserts that any subsolutions of
the heat equation attains their maximum on the parabolic boundary.
if ut u 0, then the maximum
Theorem 5.12. For u C 2,1 (Q) C(Q),

of u in Q is attained on p Q, i.e.,
max u = max u.

p Q

Proof. 1) We first consider a special case. If ut u < 0, we prove by

cannot be attained in Q. Suppose there
contradiction that the maximum of u in Q
exists a point P0 Q such that
u(P0 ) = max u.

Then uxi (P0 ) = 0 and {uxi xj (P0 )} is a nonpositive definite matrix. Moreover,
ut (P0 ) = 0 for t (0, T ) and ut (P0 ) 0 for t = T . Hence ut u 0 at P0 , which
leads to a contradiction.

118

5. HEAT EQUATIONS

2) For any > 0, we consider an auxiliary function

v(x, t) = u(x, t) t.
Then
vt v = ut u < 0.
By 1), v cannot attain its maximum inside Q. Hence
max v = max v.

p Q

Then
max u(x, t) = max(v(x, t) + t)) max v(x, t) + T

= max v(x, t) + T max u(x, t) + T.

p Q

p Q

Letting 0, we get the desired result.

The following result is referred to as the comparison principle.

satisfy ut u vt v
Corollary 5.13. Suppose u, v C 2,1 (Q) C(Q)

in Q and u v on p Q. Then u v in Q.
Next, we consider a slightly more general operator given by
Lu ut u + cu = f

in Q.

We prove the following weak maximum principle for L.

satisfies Lu 0 with c(x, t) 0.
Theorem 5.14. Suppose u C 2,1 (Q) C(Q)
Then the nonnegative maximum of u must be attained on p Q, i.e.,
max u max{0, u}.

p Q

The proof is a simple modification of that of Theorem 5.12 and is omitted.

Now, we consider a more general case.
Theorem 5.15. Suppose c(x, t) c0 in Q for a nonnegative constant c0 and
satisfies Lu 0. If u 0 on p Q, then u 0 in Q.
u C 2,1 (Q) C(Q)
always have global minimum. Therefore, c c0
Continuous functions in Q
Such a
always holds in Q for some nonnegative constant c0 if c is continuous in Q.
condition is introduced to emphasize the role of the minimum of c.
Proof. Let v(x, t) = ec0 t u(x, t). Then
vt v + (c(x, t) + c0 )v = ec0 t Lu 0.
Since c(x, t) + c0 0, by Theorem 5.14, we get
max v(x, t) max{0, v(x, t)} max{0, ec0 t u(x, t)} 0.
Q

Hence u(x, t) 0 in Q.

p Q

p Q

The following result is referred to as the comparison principle, which generalizes

Corollary 5.13.

5.3. THE MAXIMUM PRINCIPLE

119

Corollary 5.16. Suppose c(x, t) c0 in Q for a nonnegative constant c0

satisfy Lu Lv in Q and u v on p Q. Then u v
and u, v C 2,1 (Q) C(Q)

in Q.
There is also a strong maximum principle for parabolic differential equations.
We will not discuss it here. In the following, we discuss applications of maximum
principles.
We first derive an estimate in sup-norms of solutions of initial/boundary-value
problems with Dirichlet boundary values. Compare this with the estimates in
integral norms in Theorem 3.14.
is a solution of
Theorem 5.17. Suppose u C 2,1 (Q) C(Q)
Lu ut u + cu =f

in QT ,

u(, 0) =u0
u =

in ,
on (0, T ),

u0 C()
and C( [0, T ]). If c(x, t) c0 in Q for a
for some f C(Q),
nonnegative constant c0 , then

max |u| ec0 T max{sup |u0 |, sup ||} + T sup |f | .

(0,T )

Proof. Set
B = max{sup |u0 |,

sup

||},

F = sup |f |,
Q

(0,T )

and
v(x, t) = ec0 t (F t + B).
Then
Lv = (c0 + c)ec0 t (F t + B) + ec0 t F F Lu in Q,
and
v B u
Here we used c + c0 0 and e

c0 t

on p Q.

1 in Q. By Corollary 5.16, we obtain

v u in Q,

or
|u(x, t)| ec0 t (F t + B)

for any (x, t) QT .

This implies the desired estimate.

In the following, we study a priori estimates for initial-value problems.

Theorem 5.18. Suppose u is a bounded C 2 -solution of
in Rn (0, T ],

Lu ut u + cu =f
u(, 0) =u0

in Rn ,

for some bounded f C(Rn (0, T ]) and u0 C(Rn ). If c c0 for a nonnegative

constant c0 , then
sup
Rn (0,T )

|u| ec0 T ( sup

Rn (0,T )

|u0 | + T

sup
Rn (0,T )

|f |).

120

5. HEAT EQUATIONS

Proof. Set
H = Rn (0, T ] = {(x, t); x Rn , t (0, T ]},
and
F = sup |f |,
H

B = sup |u0 |.
Rn

We assume supH |u| M for a positive constant M since u is bounded. For any
R > 0, set
QR = BR (0, T ].
Consider an auxiliary function
w(x, t) = ec0 t (F t + B) + vR (x, t) u(x, t) in QR ,
where vR is a function to be chosen. Then
Lw = (c + c0 )ec0 t (F t + B) + ec0 t F Lu + LvR LvR

in QR .

Here we used c + c0 0 and ec0 t 1. Moreover,

w(, 0) = B + vR (, 0) u0

in BR ,

and
w vR u

on BR (0, T ].

We require
LvR 0 in QR ,
vR (, 0) 0 in BR ,
vR u 0 on BR (0, T ].
To construct such a vR , we consider
M c0 t
e (2nt + |x|2 ).
R2
Obviously, vR 0 for t = 0 and vR M on |x| = R. Next,
vR (x, t) =

LvR =

M c0 t
e (c + c0 )(2nt + |x|2 ) 0
R2

in QR .

With such a vR , we have

Lw 0
w0

in QR ,
on p QR .

Hence Corollary 5.16 yields w 0 in QR . Therefore for any fixed (x, t), by taking
R large such that (x, t) QR , we have
|u(x, t)| ec0 t (F t + B) +

M c0 t
e (2nt + |x|2 ).
R2

By letting R +, we have
|u(x, t)| ec0 t (F t + B) for any (x, t) H.
This yields the desired estimate.

Next, we prove the uniqueness of solutions of the heat equation under an assumption of exponential growth.

5.3. THE MAXIMUM PRINCIPLE

121

Theorem 5.19. Let u C 2,1 (Rn (0, T )) C(Rn [0, T ]) satisfy

ut u =0
u(, 0) =0

in Rn (0, T ),
in Rn .

If
|u(x, t)| M eA|x|

for any (x, t) Rn (0, T ),

for some positive constants M and A, then u 0 in Rn (0, T ).

Proof. For any constant > A, we prove
1
).
4
Our strategy is as follows. We construct an appropriate auxiliary function vR in
BR (0, 1/4) so that
1
|u(x, t)| vR (x, t) for any (x, t) BR 0,
,
4
and for any fixed (x, t) BR (0, 1/4)
u(x, t) = 0

for any (x, t) Rn (0,

vR (x, t) 0 as R .
The function vR is constructed with the help of the fundamental solution. Now we
start our proof.
For any constant R > 0, consider
2

vR (x, t) =

1
for (x, t) BR 0,
.
4

|x|2
M e(A)R 14t
n e
2
(1 4t)

We note that vR is simply the example we discussed preceding Theorem 5.5. Then
1
t vR vR = 0 in BR 0,
.
4
Obviously,
vR (, 0) 0 = u(, 0) in BR .
Next,
2

vR (x, t) M e(A)R eR = M eAR u(x, t)

1
for any (x, t) BR 0,
.
4
By Corollary 5.13, the comparison principle, we have
1
u(x, t) vR (x, t) for any (x, t) BR 0,
,
4
or
1
|u(x, t)| vR (x, t) for any (x, t) BR 0,
.
4
n
For any fixed (x, t) R (0, 1/4), we choose R > |x|. We note vR (x, t) 0 as
R , by > A. This implies easily u(x, t) = 0.

As the final application of the maximum principle, we provide another proof of

interior gradient estimates by the maximum principle. We do this only for solutions
of the heat equation. Recall for any r
Qr = Br (r2 , 0].

122

5. HEAT EQUATIONS

1 ) satisfies ut u = 0 in Q1 .
Theorem 5.20. Suppose u C 2,1 (Q1 ) C(Q
Then
sup |x u| C sup |u|,
Q1

p Q1

where C is a positive constant depending only on n.

Proof. Set v = |x u|2 . Then
vt v = 2

n
X

u2xi xj + 2

i,j=1

n
X

uxi (ut 4u)xi = 2

i=1

n
X

u2xi xj .

i,j=1

To get interior estimates, we intorduce a cut-off function. For any smooth function
in C (Q1 ) with supp Q3/4 , we have
n
X

(v)t (v) = (t 4)|x u|2 4

xi uxj uxi xj 2

i,j=1

n
X

u2xi xj .

i,j=1

By taking = 2 for some C (Q1 ) with 1 in Q1/2 and supp Q3/4 , we

have
( 2 |x u|2 )t 4( 2 |x u|2 )
2

n
X

=(2t 24 2|x | )|x u| 8

xi uxj uxi xj 2

i,j=1

n
X

u2xi xj .

i,j=1

By the Cauchy inequality, we obtain

8|xi uxj uxi xj | 4x2i u2xj + 2 2 u2xi xj ,
and hence

2
( |x u| )t 4( |x u| ) 2t 24 + 6|x | |x u|2 C|x u|2 ,
2

where C is a positive constant depending only on and n. Note

(u2 )t 4(u2 ) = 2|x u|2 + 2u(ut 4u) = 2|x u|2 .
By taking a constant large enough, we get
(t )( 2 |x u|2 + u2 ) (C 2)|x u|2 0.
We apply Corollary 5.13 to get
sup( 2 |x u|2 + u2 ) sup ( 2 |x u|2 + u2 ).
Q1

p Q1

This implies the desired result since = 0 on p Q1 and = 1 in Q1/2 .

To end our discussions, we compare maximum principles for elliptic and parabolic equations of the following forms
Le u = u + c(x)u in ,
and
Lp u = ut u + c(x, t)u in QT (0, T ).

5.4. HARNACK INEQUALITIES

123

If c 0, then
Le u 0 implies u attains its nonnegative maximum on ,
Lp u 0 implies u attains its nonnegative maximum on p Q.
If c 0, the nonnegativity condition can be removed. For c 0, comparison
principles can be stated as follows:
Le u Le w in , u v on u v in ,
Lp u Lp w in QT , u v on p QT u v in QT .
We should note that comparison principles for parabolic equations hold for c c0 ,
for a nonnegative constant c0 .
In practice, we need to construct auxiliary functions for comparisons. Usually,
2
we take |x|2 or e|x| for elliptic equations and Kt + |x|2 for parabolic equations.
Sometimes, auxiliary functions are constructed with the help of the fundamental
solutions for the Laplace equation and the heat equation.
5.4. Harnack Inequalities
For positive harmonic functions, the Harnack inequality asserts that values in
compact subsets are comparable. In this section, we study the Harnack inequality
for positive solutions of the heat equation. In order to investigate a proper form of
the Harnack inequality for solutions of the heat equation, we begin our discussion
with the fundamental solution of the heat equation.
Consider for any Rn
|x|2
1
4t
u(x, t) =
for any x Rn , t > 0.
n e
(4t) 2
Then u satisfies the heat equation ut u = 0 in Rn (0, ). Hence, for any
(x1 , t1 ) and (x2 , t2 ) Rn (0, )
n2
|x2 |2
|x1 |2
u(x1 , t1 )
t2
e 4t2 4t1 .
=
u(x2 , t2 )
t1
Note

(p + q)2
p2
q2

+
for any a, b > 0,
a+b
a
b
where equality holds if and only if bp = aq. This implies for any t2 > t1 > 0

|x2 |2
|x2 x1 |2
|x1 |2

+
,
t2
t2 t1
t1
where equality holds if and only if
t2 x1 t1 x2
.
=
t2 t1
Therefore,
n2
|x2 x1 |2
t2
u(x1 , t1 )
e 4(t2 t1 ) u(x2 , t2 ) for any t2 > t1 > 0,
t1
where equality holds if is chosen as above. This simple calculation suggests that
the Harnack inequality for the heat equation has an evolution feature: the value
of a positive solution at a certain time is controlled from above by the value at

124

5. HEAT EQUATIONS

Figure 5.5. Values of the fundamental solution at different time.

a later time. In general, we cannot estimate the value of a positive solution at
a certain time from above by values at a previous time. Hence if we attempt to
establish the following estimate
u(x1 , t1 ) Cu(x2 , t2 ),
the constant C should depend on t2 /t1 , |x2 x1 | and most importantly (t2 t1 )1 (>
0). Usually, we choose following regions for points (x1 , t1 ) D1 and (x2 , t2 ) D2 .

Figure 5.6. Domains for Harnack inequalities.

Suppose u is a positive function and set v = log u. In order to derive an estimate
for the quotient
u(x1 , t1 )
,
u(x2 , t2 )
it suffices to get an estimate for the difference
v(x1 , t1 ) v(x2 , t2 ).
To this end, we need an estimate of vt and |v|. For a hint of proper forms, we
again turn our attention to the fundamental solution of the heat equation.
Consider
|x|2
1
4t
u(x, t) =
for any x Rn , t > 0.
n e
(4t) 2
Then
|x|2
n
,
v(x, t) = log u(x, t) = log(4t)
2
4t
and hence
n
vt = + |v|2 .
2t
We have the following result for arbitrary positive solutions of the heat equation.

5.4. HARNACK INEQUALITIES

125

Theorem 5.21. Suppose u C 2,1 (Rn (0, T ]) satisfies

ut = u,

u>0

in Rn (0, T ].

Then v = log u satisfies

(5.5)

vt +

n
|v|2
2t

in Rn (0, T ].

The inequality (5.5) is referred to as the differential Harnack inequality.

first prove a corollary.

We

Corollary 5.22. Suppose u C 2,1 (Rn (0, T ]) satisfies

ut = u,

u>0

in Rn (0, T ].

Then for any (x1 , t1 ), (x2 , t2 ) Rn (0, T ] with t2 > t1 > 0

n2
u(x1 , t1 )
t2
|x2 x1 |2

}.
(5.6)
exp{
u(x2 , t2 )
t1
4(t2 t1 )

Proof. Along any path t, x(t) with x(ti ) = xi , i = 1, 2, we have by Theorem

5.21
dx
dx
n
d
v(x(t), t) = vt + v
|v|2 + v
.
dt
dt
dt
2t
By completing square, we obtain
2
d
1 dx
n
v(x(t), t) ,
dt
4 dt
2t
and hence

Z 2
n
t2
1 t2 dx
v(x2 , t2 ) v(x1 , t1 ) log
dt.
2
t1
4 t1 dt
To seek an optimal path which makes the last integral minimal, we require
d2 x
=0
dt2
along the path. Hence we set
for some a, b Rn .

x = at + b

By xi = ati + b, i = 1, 2, we take
t2 x1 t1 x2
x2 x1
and b =
.
a=
t2 t1
t2 t1
Then,
Z t 2 2
2
dx
dt = |x2 x1 | .
dt
t2 t1
t1
Therefore, we obtain
v(x2 , t2 ) v(x1 , t1 )
or
u(x2 , t2 )

u(x1 , t1 )
This is the desired estimate.

t1
t2

t2
1 |x2 x1 |2
n
log
,
2
t1
4 t2 t1

n2
exp{

|x2 x1 |2
}.
4(t2 t1 )

126

5. HEAT EQUATIONS

Now we start to prove the differential Harnack inequality. The basic idea is to
apply the maximum principle to an appropriate combination of derivatives of v. In
our case, we consider |v|2 vt and intend to derive an upper bound. First, we
need to derive a parabolic equation satisfied by |v|2 vt . A careful analysis shows
there are uncontrollable terms in this equation. Therefore, we should introduce a
parameter (0, 1) and consider |v|2 vt instead. After we apply the maximum
principle, we let 1. Since the proof below is quite involved, we divide it into
several steps.
Proof of Theorem 5.21. Step 1. We first derive some equations involving
derivatives of v = log u. A simple calculation shows
vt = v + |v|2 .
Note that the differential Harnack inequality asserts
n
0,
v +
2t
which implies v is almost concave.
Consider w = v. Then we have
wt = vt = (v + |v|2 ) = w + |v|2 .
Since
|v|2 = 2|2 v|2 + 2v (v) = 2|2 v|2 + 2v w,
then
wt w 2v w = 2|2 v|2 .

(5.7)

Note that the uncontrolled expression v appears as a coefficient of the equation

in (5.7). It is convenient to derive an equation for it. Set w
= |v|2 . Then,
w
t = 2v vt = 2v (v + |v|2 )
= 2v (v) + 2v w

= |v|2 2|2 v|2 + 2v w

= w
+ 2v w
2|2 v|2 .
Therefore, w
satisfies
w
t w
2v w
= 2|2 v|2 .

(5.8)

Note by the Cauchy inequality

n
n
n
X
X
1
1 X
|2 v|2 =
vx2i xj
vxi xi )2 = (v)2 .
vx2i xi (
n
n
i,j=1
i=1
i=1
So (5.7) implies
wt w 2v w

2 2
w .
n

Step 2. Set for a constant (0, 1]

F = |v|2 vt .
Then
F = |v|2 v |v|2 = v (1 )|v|2 = w (1 )w,

5.4. HARNACK INEQUALITIES

127

and hence by (5.7) and (5.8)

Ft F 2v F = 2|2 v|2 .
Next, we estimate |2 v|2 by F . Note
2
1
1
1
|v|2 vt = ((1 )|v|2 + F )2
|2 v|2 (v)2 =
n
n
n

1 2
2
4
=
F + (1 ) |v| + 2(1 )|v|2 F
n

1 2
F + 2(1 )|v|2 F .

n
We obtain

2 2
(5.9)
Ft F 2v F
F + 2(1 )|v|2 F .
n
We should point out that the term |v|2 at the right-hand side plays an important
role later on.
Step 3. Now we introduce a cut-off function. Set
G = tF ,
where 0 and C0 (Rn ). We derive a differential inequality for G. Note
Gt = F + tFt ,
G = tF + tF ,
G = tF + 2tF + tF .
Then,
G
,
t

G,
tF = G

tF = G G
G
G

||2

= G 2
G + 2 2 G
G.

tFt = Gt

Multiplying (5.9) by t2 2 and substituting above inequalities, we obtain

t(Gt G) + 2t( v) G

2
||2
4(1 )
2
G
G+t 2

|v| 2v
.
n

n
To eliminate |v| from the right-hand side, we complete square for the last two
terms. Here we need < 1! Otherwise, we cannot control the expression 2v
in the right-hand side. Then we have
t(Gt G) + 2t( v) G

2
||2
n
||2
G
G+t 2

.
n

4(1 )

128

5. HEAT EQUATIONS

We point out that there are no unknown expressions in the right-hand side except
G. By choosing = 2 for some 0 and C0 (Rn ), we get
t 2 (Gt G) + 2t(2 2 v) G

2
n
G 2
G + t 6||2 2
||2
.
n
(1 )
Now we fix a cut-off function 0 C0 (B1 ), with 0 < 0 1 in B1 and 0 = 1
in B1/2 . For an arbitrary R 1, we consider (x) = 0 (x/R). Then we obtain in
BR (0, T )

C t
2
2
2
G+ 2 ,
t (Gt G) + 2t(2 v) G G 1
n
R
where C is a positive constant depending only on and 0 .
Step 4. We claim
2
C t
(5.10)
1
G + 2 0 in BR (0, T ).
n
R
Note that G vanishes on the parabolic boundary of BR (0, T ) since G = tF .
2
C t
Suppose (5.10) does not hold. Then 1
G + 2 has a negative minimum at
n
R
(x0 , t0 ) BR (0, T ]. Moreover, we have
G(x0 , t0 ) > 0,
and
Gt 0, G = 0, G 0

at (x0 , t0 ).

Then at (x0 , t0 ), we get

0 t 2 (Gt G) + 2t(2 2 v) G

C t
2
G + 2 < 0.
G 1
n
R
This is a contradiction. Hence (5.10) holds in BR (0, T ).
Therefore, we obtain
2 2
C t
(5.11)
1
t (|v|2 vt ) + 2 0 in BR (0, T ).
n
R
n
For any fixed (x, t) R (0, T ), choose R > |x|. Recall (x) = 0 (x/R) and
0 = 1 in B1/2 . Letting R +, we obtain
2
t(|v|2 vt ) 0.
n
We then let 1 and get the desired estimate.
1

By a simple modification of the proof above, we obtain the following differential

Harnack inequality for positive solutions in a finite region.
Theorem 5.23. Suppose u satisfies
ut u = 0, u > 0

in B1 (0, 1).

Then for any (0, 1), v = log u satisfies

n
vt |v|2 +
+C 0
2t

in B 12 (0, 1),

EXERCISES

129

where C is a positive constant depending only on n and .

Proof. We simply take R = 1 in (5.11).

Now we state the Harnack inequality in a finite region.

Corollary 5.24. Suppose u satisfies
ut u = 0, u > 0

in B1 (0, 1).

Then for any (x1 , t1 ), (x2 , t2 ) B1/2 (0, 1) with t2 > t1 ,

u(x1 , t1 ) Cu(x2 , t2 ),
where C is a positive constant depending only on t2 /t1 and (t2 t1 )1 .
The proof is left as an exercise.
Exercises
(1) Prove the following statements by straightforward calculations.
n

(a) K(x, t) = t 2 e

|x|2
4t

satisfies the heat equation for t > 0.

n

|x|2

(b) For any > 0, G(x, t) = (14t) 2 e 14t satisfies the heat equation
for t < 1/4.
(2) For u0 C0 (Rn ) and u defined in (5.4), prove limt u(x, t) = 0 uniformly in x.
(3) For u0 C0 (Rn ) and u defined in (5.4), find an appropriate condition on
u0 so that u is C k up to t = 0 and give a proof.
(4) Let u0 be a bounded and continuous function in [0, ) with u0 (0) = 0.
Find an integral representation for the solution of the following problem
ut uxx =0 for x > 0, t > 0,
u(x, 0) =u0 (x)
u(0, t) =0

for x > 0,

for t > 0.

(5) Prove that u constructed in the proof of Proposition 5.7 is smooth in

R R.
(6) Suppose the initial temperature u0 of an infinite rod is given by
(
1 x [0, 1],
u0 =
0 otherwise.
At what time does a point with coordinate x attain the highest temperature? In particular, consider points x = 1/2 and x = 3/2.
T ) satisfies
(7) Suppose u C 2 (QT ) C(Q
ut u + c(x, t)u = u2
u|t=0 =u0 ,
u|(0,T ) =0,

in QT = (0, T ),

130

5. HEAT EQUATIONS

T ) with c c0 and u0 C()

with u0 0. Prove
where c(x, t) C(Q
0 u ec0 t sup |u0 |

in .

T ) satisfies
(8) Suppose u C 2 (QT ) C(Q
ut u =eu f (x)
u|t=0 =u0

in QT ,

in ,

u|(0,T ) =.
Prove
M u eM T + M

in ,

where
M = T sup |f | + sup{sup |u0 |,

sup

||}.

(0,T )

satisfy
(9) Let Q = (0, l) (0, ) and u C 3 (Q) C 1 (Q)
ut uxx =0 in Q,
u|t=0 = in (0, l)
u|x=0 = 0, u|x=l =0 for any t > 0,
for a function C 1 [0, l] with (0) = (l). Prove
sup |ux | sup |0 |.
Q

[0,l]

(10) Let h(t) be a continuous increasing function with h(0) 0 and

QT = {(x, t); 0 x h(t), 0 t T }.
Let g be an increasing nonnegative function in [0, T ] and be a decreasing
C 1 -function in [0, h(0)] with where (0) = g(0) and (h(0)) = 0. Suppose
u C 2,1 (Q) is a solution of
ut uxx =0

in Q,

u(x, 0) =(x)

for x [0, h(0)],

u(0, t) =g(t),
u(h(t), t) =0,
Prove x u 0 in QT .

Figure 5.7. A domain.

EXERCISES

131

(11) Let Q = (0, l) (0, ) and uh = uh (x, t) be a solution of

ut uxx =0
u(x, 0) =0

in Q,
in (0, l),

u|x=l =0,

x u + h(u0 u) |x=0 =0,

for positive constants u0 and h. Prove

(a) 0 u u0 in (0, l) (0, );
(b) uh is monotone increasing for h.
(12) Let u1 , , um be C 2,1 -functions in Q = (0, T ] Rn R+ satisfying
t ui = ui

in Q, for i = 1, , m.

Suppose f is a convex function in Rm . Then

sup f (u1 , , um ) sup f (u1 , , um ).
Q

p Q

(13) Let u C 2,1 (Rn (0, T ]) satisfy

ut u = 0

in Rn (0, T ].

Suppose u and u are bounded in Rn (0, T ]. Prove

1
sup |u| sup |u(, 0)|.
n
2t Rn
R (0,T ]
Hint: With |u| M at t = 0, consider
w = u2 + 2t|u|2 M 2 .
(14) Prove Corollary 5.24.

CHAPTER 6

Wave Equations
In this chapter, we study the wave equation in Rn R+ . The wave equation
represents vibrations of strings or propagation of sound waves in tubes for n = 1,
waves on the surface of shallow water for n = 2, and acoustic or light waves for
n = 3. In Section 6.1, we discuss initial-value problems and initial/boundary-value
problems for the one-dimensional wave equation. In Section 6.2, we study initialvalue problems for the wave equation in higher dimensional spaces. Then in Section
6.3, we derive energy estimates for solutions of initial-value problems.
6.1. One-Dimensional Wave Equations
In this section, we discuss initial-value problems and various initial/boundaryvalue problems for the one-dimensional wave equation. We first study initial-value
problems.
For f C(R R+ ), C 2 (R) and C 1 (R), we seek a solution u
C 2 (R R+ ) of the following problem
(6.1)

utt uxx = f

in R R+ ,

u(, 0) = , ut (, 0) =

on R.

We will use different methods to derive expressions of solutions in various special

cases.
As discussed in Section 3.1, characteristic curves for the one-dimensional wave
equation are given by straight lines t = x+c. In particular, for any (
x, t) RR+ ,

there are two characteristic curves through (

x, t) given by
t x = t x
and

t + x = t + x
.

We first consider the case f 0. By setting

= x t, = x + t,
we have
u = 0,
and hence
u(, ) = g() + h(),
for some functions g and h in R. Therefore,
(6.2)

u(x, t) = g(x t) + h(x + t).

With initial values, we have

(x) = g(x) + h(x),

(x) = g 0 (x) + h0 (x).

133

134

6. WAVE EQUATIONS

Then

Z
1
1 x
(s)ds + c
g(x) = (x)
2
2 0
Z x
1
1
h(x) = (x) +
(s)ds c,
2
2 0

for a constant c. Hence,

(6.3)

1
1
u(x, t) = (x t) + (x + t) +
2
2

x+t

(s)ds.
xt

This is called the dAlemberts formula. Such a formula clearly shows that regularity
of u(, t) for any t > 0 is the same as that of the initial value u(, 0). There is no
improvement of regularity.
We see from (6.3) that u(x, t) is determined uniquely by initial values in the
interval [x t, x + t] of the x-axis whose end points are cut out by the characteristic
curves through the point (x, t). This interval represents the domain of dependence
for the solution at the point (x, t). Conversely, the initial values at a point (x0 , 0)
of the x-axis influence u(x, t) at points (x, t) in the wedge-shaped region bounded
by characteristic curves through (x0 , 0), i.e., for x0 t < x < x0 + t.

Figure 6.1. The domain of influence and the domain of dependence.

Now we derive an important formula for the solution of the wave equation. Let
u be a C 2 -solution of
utt uxx = 0

in R R+ ,

and A, B, C, D be four vertices of a parallelogram bounded by four characteristic

curves in R R+ as follows. (This parallelogram is in fact a rectangle.) Then
(6.4)

u(A) + u(D) = u(B) + u(C).

In other words, the sums of the values of u in opposite vertices are equal. This
follows easily from (6.2).
Next, we use the method of characteristics to solve (6.1) if f 0 and 0.
By writing
utt uxx = (t + x )(t x )u,
we decompose (6.1) to two initial-value problems for first order PDEs as follows
(6.5)

ut ux = v
u(, 0) = 0

in R R+ ,
on R,

6.1. ONE-DIMENSIONAL WAVE EQUATIONS

135

Figure 6.2. A rectangle.

and
vt + vx = 0

(6.6)

in R R+ ,

v(, 0) =

on R.

We note that integral curves are exactly characteristic curves for first order linear
PDEs in the plane. First, we discuss (6.6). Its associated integral curves are given
by x = t + c. Along these curves, the equation in (6.6) becomes
dv(x(t), t)
= 0.
dt
Hence
v(x(t), t) = v(x(0), 0) = (x(0)).
For any (
x, t), we have x(t) = t + (
x t) and hence
v(
x, t) = v(x(t), t) = (x(0)) = (
x t).
Therefore, (6.6) has a solution
v(x, t) = (x t).
For (6.5), its integral curves are given by x = t + c and hence x(t) = t + x
+ t

for any (
x, t). Along these curves, the equation in (6.5) becomes
du(x(t), t)
= v(x(t), t) = (x(t) t).
dt
Then

u x(t), t =

x( ) d,

and hence

u(
x, t) =

x( ) d =
0

1
(
x + t 2 )d =
2

Therefore,

x
+t
x
t

(s)ds.

Z
1 x+t
(s)ds.
2 xt
This is simply a special case of the DAlemberts formula.
Now we derive an expression of solutions in the general case. For any (x, t)
R R+ , consider the triangle
u(x, t) =

C1 (x, t) = {(y, ); |y x| < t , > 0}.

136

6. WAVE EQUATIONS

This is exactly the cone we introduced in Section 2.3 for n = 1. The boundary of
C1 (x, t) consists of three parts
L+ = {(y, ); = y + x + t, 0 < < t},
L = {(y, ); = y x + t, 0 < < t},
and
L0 = {(y, 0); x t < y < x + t}.
We note that L+ and L are parts of the characteristic curves through (x, t). Let
= (1 , 2 ) be the unit exterior normal vector of C1 (x, t). Then

2
on L+ ,
(1, 1)/
= (1, 1)/ 2 on L ,

(0, 1)
on L0 .
Hence by the Greens formula, we have

Figure 6.3. Integral curves.

Z

Z
f=

C1 (x,t)
Z x+t

Z
(utt uxx ) =

C1 (x,t)

ut (s, 0)ds +
xt

L+

(ut 2 ux 1 )
C1 (x,t)

1
(ut ux ) +
2

Z
L

1
(ut + ux ),
2

where the
orientation of integrals on L+ and L is counterclockwise. Note that
(t x )/ 2 is a directional derivative along L+ with unit length and with the
direction matching the orientation of the integral on L+ . Hence
Z
1
(ut ux ) = u(x, t) u(x + t, 0).
2
L+

On the other hand, (t + x )/ 2 is a directional derivative along L with unit

length and with the opposite direction matching the orientation of the integral on
L . Hence
Z

1
(ut + ux ) = u(x t, 0) u(x, t) .
2
L
Therefore, a simple substitution yields
Z
Z
Z x+(t )
1 x+t
1
1 t
(6.7) u(x, t) = (x+t)+(xt) +
(s)ds+
d
f (y, )dy.
2
2 xt
2 0
x(t )

6.1. ONE-DIMENSIONAL WAVE EQUATIONS

137

Theorem 6.1. Suppose C 2 (R), C 1 (R) and f C 1 (R [0, )). Then

u in (6.7) is a C 2 -solution of (6.1).
The proof is a straightforward calculation and is omitted. Obviously, C 2 solutions of (6.1) are unique.
The formula (6.7) illustrates that the value u(x, t) is determined by f in the
triangle C1 (x, t) and by and on the interval [x t, x + t] {0}. This has the
same feature as solutions of initial value problems of first-order linear differential
equations discussed in Section 2.3.
Without using the explicit expression of solutions in (6.7), we can also derive
energy estimates, the estimates of the L2 -norms of solutions of (6.1) and their
derivatives. For any constants 0 < T < t, we use the following domain for energy
estimates
{(x, t); |x| < t t, 0 < t < T }.
We postpone derivation until the final section of this chapter.

Figure 6.4. A domain of integration.

In the rest of this section, we study initial/boundary-value problems. For simplicity, we only discuss homogeneous wave equations.
First, we study half-space problems. Assume C 2 [0, ), C 1 [0, ) and
C 2 [0, ), and consider
utt uxx = 0
(6.8)

in [0, ) (0, ),

u(, 0) = , ut (, 0) =
u(0, t) = (t)

on [0, ),

for t > 0.

We construct a C 2 -solution under appropriate compatibility conditions. If (6.8)

admits a solution which is C 2 in [0, ) [0, ), a simple calculation shows
(6.9)

(0) = (0), (0) = 0 (0), 00 (0) = 00 (0).

We first consider the case 0 and use the method of extension to solve (6.8).
The compatibility condition (6.9) has the following form
(0) = 0,

(0) = 0,

00 (0) = 0.

Now we assume this holds and proceed to construct a C 2 -solution of (6.8). We

extend and as odd functions in R. The extended and are C 2 and C 1 in R

138

6. WAVE EQUATIONS

respectively. Let u be the unique C 2 -solution of

utt uxx = 0

in R R+ ,

u(, 0) = , ut (, 0) =

in R.

In fact, u is given by the DAlemberts formula (6.3). We now prove that u(x, t) is
a solution of (6.8) when we restrict x [0, ). We only need to prove
u(0, t) = 0

for any t > 0.

In fact, for v(x, t) = u(x, t), a simple calculation yields

vtt vxx = 0

in R R+ ,

v(, 0) = , vt (, 0) =

in R.

By the uniqueness, u(x, t) = v(x, t) = u(x, t) and hence u(0, t) = 0.

Now we consider the general case of (6.8) and use the method of reflection to
construct a solution in [0, )[0, ). To do this, we divide [0, )[0, ) into two
parts by the straight line t = x and construct u in each region. First, we consider
1 = {(x, t); x > t > 0}.
By (6.3), the solution u1 in 1 is given by
u1 (x, t) =

1
1
(x + t) + (x t) +
2
2

Set
(x) = u1 (x, x) =

x+t

(s)ds for any (x, t) 1 .

xt

1
1
(2x) + (0) +
2
2

2x

(s)ds.
0

Next, we set
2 = {(x, t); t > x > 0},
and consider
utt uxx = 0

in 2

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

We denote its solution by u2 . For any (x, t) 2 , consider a rectangle formed by
tx
t+x t+x
(x, t), (0, t x), ( tx
2 , 2 ) and ( 2 , 2 ). By (6.4), we have
x + t
x+t
) + (
)
2
2
Z
1 x+t
1
(s)ds.
= (x + t) + (x + t) (x + t) +
2
2 x+t

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

Set u = u1 in 1 and u = u2 in 2 . Now we check u, ut , ux , utt , uxx , utx are

continuous along t = x. By a direct calculation, we have
u1 (x, t)|t=x u2 (x, t)|t=x = (0) (0) = (0) (0),
x u1 (x, t)|t=x x u2 (x, t)|t=x = (0) + 0 (0),
x2 u1 (x, t)|t=x x2 u2 (x, t)|t=x = 00 (0) + 00 (0).
Then (6.9) implies
u1 = u2 ,

x u1 = x u2 ,

x2 u1 = x2 u2

on {t = x}.

6.1. ONE-DIMENSIONAL WAVE EQUATIONS

139

Figure 6.5. The method of reflection.

It is easy to get t u1 = t u2 on {t = x} by u1 = u2 and x u1 = x u2 on {t = x}.
Similarly, we get xt u1 = xt u2 and tt u1 = tt u2 on {t = x}. Therefore, u is C 2
across t = x. Hence, we obtain the following result.
Theorem 6.2. Suppose C 2 [0, ), C 1 [0, ), C 2 [0, ) and satisfy the compatibility condition (6.9). Then there exists a C 2 -solution of (6.8) in
[0, ) [0, ).
As for solutions of initial-value problems, we can also derive a priori energy
estimates for solutions of (6.8). For any constants 0 < T < t, we use the following
domain for energy estimates
{(x, t); 0 < x < t t, 0 < t < T }.

Figure 6.6. A domain of integration.

Now we consider the initial/boundary-value problem discussed in Section 3.3.
For a positive l > 0, assume C 2 [0, l], C 1 [0, l] and , C 2 [0, ). Consider
utt uxx = 0
(6.10)

in [0, l] (0, ),

u(, 0) = , ut (, 0) =

on [0, l],

u(0, t) = (t), u(l, t) = (t)

for t > 0.

The compatibility condition is given by

(6.11)

(0) = (0), (0) = 0 (0), 00 (0) = 00 (0),

(l) = (0), (l) = 0 (0), 00 (l) = 00 (0).

140

6. WAVE EQUATIONS

For = 0, we can use the method of extension to construct solutions. We

first extend (x) = (x) for any x [0, l] and then extend (x + 2l) = (x)
for any x. So is odd and 2l-periodic in R. We extend similarly. The extended
functions and are C 2 and C 1 on R respectively. Let u be the unique solution
of
utt uxx = 0

in R R+ ,

u(, 0) = , ut (, 0) =

in R.

We now prove that u(x, t) is a solution of (6.10) when we restrict x [0, l], i.e.,
u(0, t) = 0, u(l, t) = 0

for any t > 0.

The proof is similar to that of Theorem 6.2. We use v(x, t) = u(x, t) to prove
u(0, t) = 0 and w(x, t) = u(2l x, t) to prove u(l, t) = 0.

Figure 6.7. Extensions of initial values.

Next we use the method of reflection to construct a solution of (6.10) in the
general case. We divide [0, l] [0, l] into four regions by t = x and t = x + l and
construct u in each region. The process is similar and hence is omitted. Therefore,

Figure 6.8. A method of reflection.

we obtain the following result.
Theorem 6.3. Suppose C 2 [0, ), C 1 [0, ), , C 2 [0, ) and satisfy the compatibility condition (6.11). Then there exists a C 2 -solution of (6.10) in
[0, ) [0, ).

6.1. ONE-DIMENSIONAL WAVE EQUATIONS

141

Theorem 6.3 includes Theorem 3.19 in Chapter 2 as a special case.

Now we summarize various problems discussed in this section. We emphasize
that characteristic curves play an important role in studies of the 1-dimensional
wave equation.
First, presentation of problems depends on characteristic curves. Let be a
piecewise smooth domain in R2 whose boundary is not characteristic. We intend
to prescribe appropriate boundary values to ensure the well-posedness for the wave
equation. To do this, we take an arbitrary point on the boundary and examine
characteristic curves through this point. We then count how many characteristic
curves enter the domain in the positive t-direction. In this section, we discussed
cases where is given by the upper half space R R+ , the first quadrant R+ R+
and I R+ for a finite integral I. We note that the number of boundary values is
the same as the number of characteristic curves entering the domain in the positive
t-direction. In summary, we have
u|t=0 = , ut |t=0 = for initial-value problems;
u|t=0 = , ut |t=0 = , u|x=0 = for half-space problems;
u|t=0 = , ut |t=0 = , u|x=0 = , u|x=l = for initial/boundary-value problems.

Figure 6.9. Characteristic directions.

Second, characteristic curves determine the domain of influence and the domain of dependence. In fact, as illustrated by (6.7), initial values propagate along
characteristic curves.

Figure 6.10. The domain of influence and the domain of dependence.

Last, characteristic curves also determine domains for energy estimates. We
will explore this in detail in the final section of this chapter.

142

6. WAVE EQUATIONS

6.2. Higher Dimensional Wave Equations

In this section, we discuss initial-value problems for the wave equation in higher
dimensions. Let and be C 2 and C 1 functions in Rn respectively and f be a
continuous function in Rn R+ . Consider
(6.12)

utt u = f

in Rn R+ ,
in Rn .

u(, 0) = , ut (, 0) =

We intend to derive an expression for C 2 -solution u in Rn [0, ). To do this, we

decompose (6.12) to three problems
(6.13)

(6.14)

utt u = 0

in Rn R+ ,
in Rn ,

u(, 0) = , ut (, 0) = 0
utt u = 0

in Rn R+ ,
in Rn ,

u(, 0) = 0, ut (, 0) =

and
(6.15)

utt u = f

in Rn R+ ,
in Rn .

u(, 0) = 0, ut (, 0) = 0

Obviously, a sum of solutions of (6.13)-(6.15) yields a solution of (6.12).

The next result is referred to as the Duhamels Principle.
Theorem 6.4. Suppose u2 = M (x, t) is a solution of (6.14). Then
u1 (x, t) = t M (x, t),
Z t
u3 (x, t) =
Mf (x, t )d

where f = f (, )

are solutions of (6.13) and (6.15) respectively.

The proof is based on straightforward calculations.
Proof. First, M (x, t) satisfies
tt M M = 0

in Rn R+ ,

M |t=0 = 0, t M |t=0 =

in Rn .

Then
tt u1 u1 = (tt )t M = t (tt M M ) = 0
u1 |t=0 = t M (x, t)|t=0 =

in Rn R+ ,

in Rn ,

t u1 |t=0 = (tt M )|t=0 = (M )|t=0 = 0

in Rn .

Next, w(x, t) = Mf (x, t ) satisfies

wtt w = 0

in Rn R+ ,

w|t= = 0, t w|t= = f (, )
Then
t u3 = Mf (x, t )| =t +

Z
0

in Rn .
Z

t Mf (x, t )d =

t Mf (x, t )d,

6.2. HIGHER DIMENSIONAL WAVE EQUATIONS

and

Z
tt u3 = t Mf (x, t )| =t +
Z
= f (x, t) +
0

143

tt Mf (x, t )d
Z

Mf (x, t )d = f (x, t) +

t
0

Mf (x, t )d

= f (x, t) + u3 .
Hence tt u3 u3 = f in Rn R+ and u3 |t=0 = 0, t u3 |t=0 = 0 in Rn .

In the following, we concentrate on (6.14). Suppose u is a C 2 -solution of (6.14)

in Rn [0, ), i.e.,
utt u = 0

in Rn R+ ,

u(, 0) = 0, ut (, 0) =

in Rn .

We solve this initial-value problem by the method of spherical average due to Poisson. Set for any t > 0 and r > 0
Z
1
I(r, t) =
u(y, t)dSy ,
n rn1 |yx|=r
where n is the surface area of the unit sphere in Rn . Obviously, I(r, t) is the
average of u(, t) over the sphere Br (x) and also depends on x. First, I(r, t) is
defined for r > 0. By changing it to the form
Z
1
I(r, t) =
u(x + r, t)dS ,
n ||=1
we note that I(r, t) is defined for all r (, ) and is even for r, since replacing
r by r does not change the value of the integral. It is clear that u can be recovered
from I, since
Z
1
I(0, r) = lim I(r, t) =
u(x, t)dS = u(x, t).
r0
n ||=1
Next, we change the equation of u into an equation of I(r, t). We claim that
I(r, t) satisfies the following Euler-Poisson-Darboux equation
Itt = Irr +

n1
Ir
r

for r > 0, t > 0,

with initial values

I(r, 0) = 0, It (r, 0) = (r) for r > 0,
where

Z
1
(x + r)dS for r > 0.
n ||=1
To verify the equation satisfied by I, we first have by differentiating under the
integral sign
Z
Z
1
1
u
u
Ir (r, t) =
(x + r, t)dS =
(y, t)dSy
n ||=1 n
n rn1 |yx|=r n
Z
1
=
u(y, t)dy.
n rn1 |yx|r
(r) =

144

6. WAVE EQUATIONS

Then

1
n

rn1 Ir =
Hence
(r

n1

1
Ir ) r =
n

Z
u(y, t)dy =
|yx|r

1
n

1
tt
utt (y, t)dSy =

n
|yx|=r

utt (y, t)dy.

|yx|r

Z
u(y, t)dSy = rn1 Itt .
|yx|=r

For initial values, we simply note for any r > 0

Z
1
I(r, 0) =
u(x + r, 0)dS = 0,
n ||=1
Z
1
It (r, 0) =
(x + r)dS .
n ||=1
In general, it is a tedious process to solve initial-value problems for the EulerPoisson-Darboux equation for general dimension. However, this process is relatively
easy for n = 3. If n = 3, we have
2
Itt = Irr + Ir ,
r
or
(rI)tt = (rI)rr

for r > 0, t > 0,

with
(rI)(r, 0) = 0, (rI)t (r, 0) = r(r)

for r > 0.

This is a half-space problem for rI(r, t) studied in the previous section for the onedimensional wave equation. We note (r)|r=0 = 0 and (rI)|r=0 = 0. Hence the
compatibility condition in Theorem 6.2 is satisfied. In order to find a solution rI,
we extend r(r) to r (, ) as an odd function. (In fact, (r) as the average
of over Br (x) can be extended as an even function for r (, ) as before.)
Then
Z
Z
1 r+t
1 r+t
s(s)ds =
s(s)ds,
(rI)(r, t) =
2 rt
2 tr
where we changed r t to t r for the lower integral limit by the oddness of the
integrand. Hence,
Z t+r
1
I(r, t) =
s(s)ds.
2r tr
Note that the area of the unit sphere in R3 is 4. Then we have
Z
Z
1
t
(x + t)dS =
(y)dSy .
lim I(r, t) = t(t) =
r0
4 ||=1
4t |yx|=t
Therefore, we obtain formally an expression of a solution of (6.14) in R3 R+ given
by
Z
1
(6.16)
u(x, t) =
(y)dSy .
4t |yx|=t
We note that u(x, t) depends on the initial value in the sphere Bt (x).
Any C 2 -solution u of the initial-value problem (6.14) in R3 R+ is given by
the formula (6.16), hence is unique. Now we prove it yields the classical solution.

6.2. HIGHER DIMENSIONAL WAVE EQUATIONS

145

Theorem 6.5. For any C 2 (R3 ), u in (6.16) is a C 2 -solution of (6.14) in

R R+ .
3

Proof. By setting
1
(x, t) =
4

Z
(x + t)dS ,
||=1

we have by (6.16)
u(x, t) = t(x, t).
Then
ut = + tt ,

utt = 2t + ttt .

We calculate derivatives of as follows

Z
Z
1

t =
(x + t)dS =
(y)dSy
2
4 ||=1 n
4t |yx|=t n
Z
Z t Z
1
1
=
(y)dy =
d
(y)dSy ,
4t2 |yx|t
4t2 0
|yx|=
Z
Z
1
1
tt (t) =
(y)dy +
(y)dSy
2t3 |yx|t
4t2 |yx|=t
Z
1
2
y (y)dSy .
= t (t) +
t
4t2 |yx|=t
By y = x + t, we have
Z
Z
t
t
utt =

(y)dS
=
x (x + t)dS = u,
y
y
4t2 |yx|=t
4 ||=1
where we used x (x + t) = y (y). It is easy to see u(, 0) = 0 and ut (, 0) =
in R3 .

Now we consider the initial-value problem (6.12) of the wave equation in R3 as

follows
utt u = f

in R3 R+ ,

u(, 0) = , ut (, 0) =

in R3 .

By Theorem 6.4, we have

u(x, t) =t
(6.17)

1
4t

1
4

Z
(y)dSy
|yx|=t

|yx|t

1
+
4t

Z
(y)dSy
|yx|=t

f (y, t |y x|)
dy.
|y x|

146

6. WAVE EQUATIONS

In fact, we only need to verify the final term in (6.17). We note

Z t
Z
1
(t )
f (x + (t ), )dS d (with t = 0 )
4 0
||=1
Z t Z
1
=

f (x + , t )dS d
4 0
||=1
Z tZ
1
f (y, t )
=
dSy d
4 0 |yx|=

Z
1
f (y, t |y x|)
=
dy.
4 |yx|t
|y x|
Hence we obtain the following result.
Theorem 6.6. For any C 3 (R3 ), C 2 (R3 ) and f C 2 (R3 R+ ), the
function u given by (6.17) is a C 2 -solution of (6.12) in R3 R+ .
We can also express the final term in (6.17) as
Z t
Z
1
1
f (y, )dSy d.
4 0 t |yx|=t
Hence the value of the solution u at (x, t) depends only on the values of f in points
(y, ) with
|y x| = t , 0 < < t.
This is exactly the backward characteristic cone with vertex (x, t).
We now use the method of descent to solve initial-value problems of the wave
equation in R2 R+ . As for three dimensional case, we first consider (6.14). Let
be C 2 in R2 and u(x, t) be C 2 in R2 [0, ) and satisfy (6.14), i.e.,
utt u = 0

in R2 R+ ,

u(, 0) = 0, ut (, 0) =

in R2 .

Any solutions in R2 can be viewed as solutions of the same problem in R3 which

are independent of the third space variable. Then for any (x, t) R2 R+ (with
x3 = 0), we have by (6.16)
Z
1
(y)dSy,
u(x, t) =
4t |yx|=t
where y = (y1 , y2 , y3 ) = (y, y3 ). The integral here is a surface integral in R3 . Now
we evaluate it as a 2-dimensional integral by eliminating y3 . The sphere |
y x| = t
in R3 has two pieces given by
p
y3 = t2 |y x|2 ,
and its surface area element is

1
t
dSy = 1 + (y1 y3 )2 + (y2 y3 )2 2 dy1 dy2 = p
dy1 dy2 .
2
t |y x|2
Therefore, we obtain
(6.18)

u(x, t) =

1
2

(y)

p
|yx|t

t2 |y x|2

dy.

6.2. HIGHER DIMENSIONAL WAVE EQUATIONS

147

We note that u(x, t) depends on initial values in the solid disc Bt (x).
Theorem 6.7. For any C 2 (R2 ), u in (6.18) is a C 2 -solution of (6.14) in
R R+ .
2

Next, we discuss briefly how to obtain explicit expressions for solutions of the
wave equation in arbitrary dimension. For odd dimension, we seek an appropriate
combination of I(r, t) and its derivatives to satisfy the 1-dimensional wave equation
and then proceed as for n = 3. For even dimension, we again use the method of
descent.
Let n 3 be an odd integer. Then the spherical average I(r, t) satisfies
n1
(6.19)
Itt = Irr +
Ir for r > 0, t > 0.
r
First, we write (6.19) as

1
Itt = rIrr + (n 1)Ir .
r
By
(rI)rr = rIrr + 2Ir ,
we obtain

1
Itt = (rI)rr + (n 3)Ir ,
r
or
(6.20)

(rI)tt = (rI)rr + (n 3)Ir .

If n = 3, then rI satisfies the 1-dimensional wave equation. This is how we solved

the wave equation in 3-dimension. By differentiating (6.19) with respect to r, we
have

n1
1
n1
Irr
Ir = 2 r2 Irrr + (n 1)rIrr (n 1)Ir .
Irtt = Irrr +
2
r
r
r
By
(r2 Ir )rr = r2 Irrr + 4rIrr + 2Ir ,
we obtain

1
Irtt = 2 (r2 I)rr + (n 5)rIrr (n + 1)Ir ,
r
or
(6.21)

(r2 Ir )tt = (r2 Ir )rr + (n 5)rIrr (n + 1)Ir .

The second term in the right-hand side of (6.21) has a coefficient n 5, which is
2 less than the coefficient of the second term in the right-hand side of (6.20). Also
the third term involving Ir in the right-hand side of (6.21) has a similar expression
as the second term in the right-hand side of (6.20). Therefore an appropriate
combination of (6.20) and (6.21) eliminates terms involving Ir . In particular, for
n = 5, we have
(r2 Ir + 3rI)tt = (r2 Ir + 3rI)rr .
In other words, r2 Ir + 3rI satisfies the 1-dimensional wave equation. We can
continue this process to obtain appropriate forms for all odd dimensions. Next, we
note
1
r2 Ir + 3rI = (r3 I)r .
r

148

6. WAVE EQUATIONS

It turns out that the correct combination of I and its derivatives for arbitrary odd
dimension n is given by

n3
2
n2

1
J(r, t) =
r
I(r, t) .
r r
We leave derivation as an exercise.
We need to point out that the method of spherical average is by no means the
only way to derive explicit expressions of solutions of the wave equation. Other
methods are available. Refer to an exercise for an alternative approach to solving
three-dimensional wave equation.
Now we compare several formulas we obtained in the previous section and in
this section. Let u be a C 2 -solution of the initial-value problem
(6.22)

utt u = 0

in Rn R+ ,

u(, 0) = , ut (, 0) =

in Rn .

We write un for dimension n. Then

Z
1 x+t
1
(x + t) + (x t) +
(y)dy,
2
2 xt

!
Z
Z
1
(y)
1
(y)
p
p
u2 (x, t) = t
dy +
dy,
2 |yx|t t2 |y x|2
2 |yx|t t2 |y x|2

!
Z
Z
1
1
u3 (x, t) = t
(y)dSy +
(y)dSy .
4t |yx|=t
4t |yx|=t
u1 (x, t) =

These formulas display many important properties of solutions u.

According to these expressions, the value u at (x, t) depends on the values of
and on the interval [x t, x + t] for n = 1, on the solid disc Bt (x) of center
x and radius t for n = 2, and on the sphere Bt (x) of center x and radius t for
n = 3. These regions are the domain of dependence of solutions on initial values.
Conversely, initial values and at a point x0 on the initial hypersurface t = 0

Figure 6.11. The domain of dependence.

influence u in points (x, t) in the solid cone |x x0 | t for n = 2 and only the cone
|x x0 | = t for n = 3 at a later time t.

6.2. HIGHER DIMENSIONAL WAVE EQUATIONS

149

Figure 6.12. The domain of influence.

The central issue here is that the solution at a given point is determined by
initial values in a proper subset of the initial hypersurface. An important consequence is that the process of solving initial-value problems for the wave equation
can be localized in space. Specifically, changing initial values outside domain of
dependence of a point does not change values of solutions there. This is a unique
property of the wave equation which distinguishes it from the heat equation.
Before exploring the difference between n = 2 and n = 3, we first note that
it takes time (literally) for initial values to make influences. Suppose the initial
values , have their support contained in a ball Br (x0 ). Then at a later time t,
the support of u(, t) is contained in the union of all balls Bt (
x) for x
Br (x0 ). It
is easy to see that such a union is in fact the ball with center x0 and radius r + t.
The support of u spreads at a finite speed. To put it in another perspective, we fix
an x
/ Br (x0 ). Then u(x, t) = 0 for t < |x x0 | r. This is the so called finite
speed propagation.
For n = 2, if the support of and takes the entire disc Br (x0 ), then the
support of u(, t) will take the entire disc Br+t (x0 ) in general. The influence from
initial values never disappears in a finite time at any particular point, like the
surface waves arising from a stone dropped into water.
For n = 3, the behavior of solutions is different. Again, we assume the support
of and are contained in a ball Br (x0 ). Then at a later time t, the support of
u(, t) is in fact contained in the union of all spheres Bt (
x) for x
Br (x0 ). Such a
union is a ball Bt+r (x0 ) for t r, as in the two-dimensional case, and an annular
region with center x0 and outer and inner radius t + r and t r respectively for
t > r. Such an annular region has a thickness 2r and spreads at a finite speed. In
other words, u(x, t) is not zero only if
t r < |x x0 | < t + r,
or
|x x0 | r < t < |x x0 | + r.
Therefore, for a fixed x R3 , u(x, t) = 0 for t < |x x0 | r (corresponding to the
finite speed propagation) and for t > |x x0 | + r. So, influence from initial values
lasts only for an interval of length 2r in time. This phenomenon is called Huygens
principle for the wave equation.
In fact, Huygens principle holds for the wave equation in every odd space
dimension n except n = 1 and does not hold in even space dimension.

150

6. WAVE EQUATIONS

Figure 6.13. A comparison between n = 2 and n = 3.

Next, we discuss whether solutions decay as t . In this aspect, there is a
sharp difference between 1-dimension and higher dimensions. By the dAlemberts
formula, it is obvious that solutions to the 1-dimensional homogeneous wave equation do not decay as t . However, solutions in higher dimensions have a
different behavior.
Theorem 6.8. Let be a smooth function with compact support in Rn and u
be a solution of
in Rn R+ ,

utt u = 0

u(, 0) = 0, ut (, 0) =

on Rn .

Then for any t > 1

C
|u(, t)|L (Rn ) kkL1 (Rn ) + kkL1 (Rn )
t

for n = 2,

and

C
kkL1 (Rn ) for n = 3,
t
where C is a positive constant independent of u and .
|u(, t)|L (Rn )

We note that decay rates vary according to dimensions. In fact, solutions to the
n1
n-dimensional wave equation have a decay rate of t 2 . These decay estimates
play an important role in study of global solutions of nonlinear wave equations.
Proof. We first consider n = 3. By (6.16), the solution u is given by
Z
t
u(x, t) =
(x + t)dS .
4 ||=1
Since has compact support, we have
Z
(x + t) =

Then
u(x, t) =

t
4

Z
t

(x + s)ds.
s

Z
||=1

(x + s)dS ds.
s

For s t, we write t s2 /t and hence

Z Z
1
1
|u(x, t)|
|(x + s)|dS ds
s2
kkL1 (R3 ) .
4t t
4t
||=1

6.2. HIGHER DIMENSIONAL WAVE EQUATIONS

151

This implies the desired result for n = 3. In fact, it holds for any t > 0.
Now we consider n = 2. By (6.18) and a change of variables, we have
Z t
Z
r
1

u(x, t) =
(x + r)dS dr.
2 0
t2 r2 ||=1
For some > 0 to be determined, we write u as
Z
Z t

1
1 t
+
=
I1 + I2 .
u(x, t) =
2 0
2
t
For I1 , we have
Z

r
(x + r)dS dr
2
2
t r ||=1
0
Z t Z
1
p
r
|(x + r)|dS dr
t2 (t )2 0
||=1
1

kkL1 (R2 ) .
2t 2

|I1 | =

Next, for r (t , t), as in the proof for n = 3, we have

Z
Z Z

(x + r)dS =
(x + s)dS ds,
s
||=1
r
||=1
and hence
Z

(x + r)dS

||=1

s
r

Then

|I2 | =

t
t
t

2
t r2

|(x + s)|dS ds kkL1 (R2 ) .

(x + r)dS dr

||=1

Z
!
1

dr sup
r
(x + r)dS
t2 r2
r[t,t]
||=1

||=1

t2

1
dr kkL1 (R2 ) .
r2

A simple calculation yields

Z t
Z t
Z t
1
1
1
1
2

dr =
dr
dr = .
t t t r
t
t2 r 2
(t + r)(t r)
t
t
Therefore, we obtain
|u(x, t)|

1
2

2
1
kkL1 (R2 ) + kkL1 (R2 ) .
t
2t 2

For any t > 1, we take = 1/2 and obtain the desired result.

Decay estimates in Theorem 6.8 are optimal for large t. We have the following
result for arbitrary t.

152

6. WAVE EQUATIONS

Theorem 6.9. Let be a smooth function with compact support in Rn and u

be a solution of
utt u = 0

in Rn R+ ,

u(, 0) = 0, ut (, 0) =

on Rn .

Then for any t > 0

|u(, t)|L (Rn ) CkkL1 (Rn )

for n = 2,

|u(, t)|L (Rn ) Ck2 kL1 (Rn )

for n = 3,

and
where C is a positive constant independent of u and .
The proof is left as an exercise.
Now we compare regularity of solutions for n = 1 and n = 3. For n = 1, the
regularity of u is clearly the same as u(, 0) and one order better than ut (, 0). In
other words, u C m and ut C m1 initially at t = 0 guarantee u C m at a later
time. However, such a result does not hold for n = 3. By writing spherical averages
as integrals over the unit sphere || = 1 and differentiating under integral sign, we
have
Z
3
X

1
(y)
+
yi (y)(yi xi ) + t(y) dSy .
u3 (x, t) =
2
4t |yx|=t
i=1
This formula indicates that u can be less regular than initial values. There is a
possible loss of one order of differentiability. Namely, u C m and ut C m1
initially at t = 0 only guarantee u C m1 at a later time.
Next, we construct a solution of the 3-dimensional homogeneous wave equation
with a loss of differentiation described above. Let u = u(r, t) be a spherically
symmetric solution of the wave equation in R3 R+ . Then
2
utt urr ur = 0,
r
or
(ru)tt (ru)rr = 0.
In other words, ru(r, t) is a solution of the 1-dimensional wave equation. Hence
1
u(r, t) = (f (t + r) + g(t r)) for any r 6= 0,
r
for some functions f and g on R. In the following, we consider f = g. For any
positive integer m, we define
(
0
for r 1,
f (r) =
2m+1
(r 1)
for 1 < r < 2.
We extend f smoothly for r > 1. We note that f is C 2m in R and not C 2m+1 at
r = 1. Then for t = 0,
1
1
u(r, 0) = (f (r) + f (r)) , ut (r, 0) = (f 0 (r) + f 0 (r)) .
r
r
2m
2m1
Hence u(, 0) C (R), ut (, 0) C
(R), and u(r, 0) is not C 2m+1 and ut (r, 0)
2m
is not C
at r = 1. We now claim that u(r, t) is not C 2m at (r, t) = (0, 1). The

6.3. ENERGY ESTIMATES

153

intuitive idea is that the singularity of initial values at r = 1 propagates along a

characteristic hypersurface and focuses at t = 1. We note that (r, t) = (0, 1) is the
vertex of the characteristic hypersurface {(x, t); t < |x|} which intersects t = 0 on
r = 1. First, u(r, 0) = ut (r, 0) = 0 for r 1. Then u(x, t) = 0 for 0 < t |x|. In
particular,
lim t2m u(0, t) = 0.
t1

2m

If u were C
at (r, t) = (0, 1), then m u = t2m u = 0 at (0, 1). Next for r > 0
sufficiently small, we have 1 + r > 1 and 1 r < 1, and hence
1
f (1 + r) = r2m = |x|2m .
r
Therefore for r > 0 sufficiently small,

m
2
m
x u(r, 1) = rr + r
r2m = (2m + 1)!.
r
u(r, 1) =

We conclude that u is not C 2m at (r, t) = (0, 1).

This example demonstrates that solutions of the higher dimensional wave equation do not have good point-wise behavior. However, the differentiability is preserved in the L2 -sense. Before we move on to energy estimates in the next section,
we show by a simple case what is involved.
Suppose u is a C 2 -solution of (6.22) for general n. We assume that and have
compact support. By finite speed propagation, u(, t) also has compact support for
any t > 0. We multiply the equation in (6.22) by ut and integrate in BR (0, t).
Here we choose R sufficiently large such that the support of u(, s) is contained in
BR for any s (0, t). Note
ut utt ut u =

n
X
1 2
(ut + |x u|2 )t
(ut uxi )xi .
2
i=1

Then a simple integration in BR (0, t) yields

Z
Z
1
1
(u2t + |x u|2 ) =
(u2 + |x u|2 ).
2 Rn {t}
2 Rn {0} t
This is the conservation of energy: the L2 -norm of derivatives at each time slice is
constant, independent of time.
6.3. Energy Estimates
In this section, we derive energy estimates of solutions of initial-value problems
for a slightly more general hyperbolic equations instead of just wave equations. Fix
a positive number T . Let a, c and f be continuous functions in Rn [0, T ] and
and be continuous functions in Rn . We consider the following initial-value
problem
(6.23)

utt au + cu = f

in Rn [0, T ],

u(, 0) = , ut (, 0) =

in Rn .

We assume a is a positive function satisfying

(6.24)

a(x, t)

for any (x, t) Rn [0, T ] and Rn ,

154

6. WAVE EQUATIONS

for some positive constants and . If the principal part of the equation is given
by the wave operator
utt u, then a = 1 and we can choose = = 1 in (6.24).
With = 1/ , we introduce for fixed t > T
D,T,t = {(x, t); |x| < t t, 0 < t < T }.
We denote by D,T,t, s D,T,t and + D,T,t the bottom, the side and the top of
the boundary. See discussions in Section 2.3 for details.

Figure 6.14. An integral domain.

Theorem 6.10. Assume a is C 1 , c and f are continuous in Rn [0, T ], is

C in Rn and is continuous in Rn . Let (6.24) be assumed and u be a C 2 -solution
of (6.23). Then for any 0
Z
Z
2
t 2
2
et (u2 + u2t + a|u|2 )
e (u + ut + a|u| ) + ( 0 )
1

+ D,T ,t

D,T ,t

Z
(2 + 2 + a||2 ) +

D,T ,t

et f 2 ,
D,T ,t

where 0 is a positive constant depending only on , the C 1 -norm of a and the

L -norm of c in D,T,t.
Proof. In the following, we simply set D = D,T,t. We multiply the equation
in (6.23) by 2et ut and integrate in D, for a scalar to be determined. First, we
note
2et ut utt = et (u2t )t + et u2t ,
and
2et aut u =

n
X

2(et aut uxi )xi + 2et auxi utxi + 2et axi ut uxi

i=1

n
X

i=1

2(et aut uxi )xi + (et au2xi )t + 2et axi ut uxi

+ et au2xi et at u2xi ,

6.3. ENERGY ESTIMATES

155

where we used 2uxi utxi = (u2xi )t . Therefore, we obtain

n
X
(et ut uxi )xi + et (u2t + a|u|2 )

(et u2t + et a|u|2 )t 2

i=1
n
X

2et axi ut uxi et at |u|2 + 2et cuut = 2et ut f.

i=1

To control the final term in the left hand side involving uut , we note
(et u2 )t + et u2 2et uut = 0.
Then a simple addition yields

u2t

(u +

n
X

+ a|u| ) t
2(et aut uxi )xi + et (u2 + u2t + a|u|2 )
2

i=1

= 2et

n
X

axi ut uxi + et at |u|2 2et (c 1)uut + 2et ut f.

i=1

The first three terms in the right hand side is quadratic in ut , uxi and u. Then by
(6.24) and the Cauchy inequality applied to each term in the right hand side, we
have
n
X
t 2

e (u + u2t + a|u|2 ) t 2
(et aut uxi )xi
i=1

+ ( 0 )et (u2 + u2t + a|u|2 ) et f 2 ,

where 0 is a positive constant depending only on , the C 1 -norm of a and the
L -norm of c. By integrating over D, we obtain
Z
Z
t 2
2
2
e (u + ut + a|u| ) + ( 0 )
et (u2 + u2t + a|u|2 )
+ D

s D

(u +
D

i=1

Z
2

n
X
2

2
2
(u + ut + a|u| )t 2
aut uxi i

u2t

+ a|u| ) +

t 2

f ,

where = (1 , , n , t ) is the unit exterior normal vector along s D. We only

need to prove that the integrand for s D is nonnegative. We claim
(u2t + a|u|2 )t 2

n
X

aut uxi i 0 on s D.

i=1

To prove this, we first note by the Cauchy inequality

|

n
X
i=1

uxi i |

n
X

u2xi ) 2

n
X

i=1

i2

12

q
= |u| 1 t2 .

i=1

By the explicit expression of (see Section 2.3), we have

t =

1
,
1 + 2

156

6. WAVE EQUATIONS

and hence with (6.24) and = 1/

n
X
2

1
(u2t + a|u|2 )t 2
aut uxi i
ut + a|u|2 2a|ut | |u|
2
1+
i=1

ut + a|u|2 2 a|ut | |u| ,

2
1+
which is nonnegative. Therefore, the boundary integral on s D is nonnegative.
A consequence of Theorem 6.10 is the uniqueness of solutions of (6.23). We
can also discuss domain of dependence and domain of influence as in the previous
section.
Next, we consider (6.23) in a general domain
D = {(x, t); h (x) < t < h+ (x), x },
where is a bounded domain in Rn and h and h+ are two piecewise C 1 -functions
in with h = h+ on . We now perform a similar integration in D as in the

Figure 6.15. A general domain.

above proof to get
Z
n
X

et (u2 + u2t + a|u|2 )+t 2

aut uxi +i
+ D

+ ( 0 )
D

i=1

et (u2 + u2t + a|u|2 )

n
X

et (u2 + u2t + a|u|2 )t 2

aut uxi i +
i=1

Z
et f 2 ,
D

where = (1 , , n , t ) are unit normal vector pointing to positive tdirection along D. We are interested in whether the integrand for + D is nonnegative. As in the proof above, we have by the Cauchy inequality
n
n
n
q
X
X
X
12
1
2
2 .
|
uxi +i |
u2xi ) 2
+i
= |u| 1 +t
i=1

i=1

i=1

Then it is easy to see

n
X

(u2t + a|u|2 )+t 2

aut uxi +i

i=1

(u2t + a|u|2 )+t 2

2 )
a(1 +t

a|u| 0

on + D

6.3. ENERGY ESTIMATES

if

157

q
2 ),
a(1 +t

+t
or

(6.25)

+t

a
1+a

on + D.

With (6.25), we obtain

Z
Z
et (u2 + u2t + a|u|2 )
et u2 +t + ( 0 )
D

+ D

n
X

et (u2 + u2t + a|u|2 )t 2

aut uxi i +
i=1

Z
et f 2 .
D

In particular, if u = ut = 0 on D and f = 0 in D, then u = 0 in D.

Now we introduce the notion of space-like and time-like surfaces.
Definition 6.11. Let be a C 1 -hypersurface in Rn R+ and = (x , t ) be
a unit normal vector with t 0. Then is space-like at (x, t) if
s
a(x, t)
t (x, t) >
;
1 + a(x, t)
is time-like at (x, t) if

s
t (x, t) <

a(x, t)
.
1 + a(x, t)

If the hypersurface is given by t = t(x), it is easy to check that is space-like

at (x, t(x)) if
1
|t(x)| < p
.
a(x, t(x))
For the wave equation
(6.26)

utt u = f,

we have a = 1. Then the hypersurface is space-like at (x, t) if t (x, t) > 1/ 2. If

is given by t = t(x), then is space-like at (x, t(x)) if
|t(x)| < 1.
In the following, we use the wave equation to discuss the importance of space-like
surfaces.
Let be a space-like surface. Then for any (x0 , t0 ) , the domain of influence
of (x0 , t0 ) is given by the cone {(x, t); t t0 > |x x0 |} and hence is always above
. It follows that prescribing initial values on space-like surfaces yields a wellposed problem. In fact, integral domains for energy estimates can be constructed
accordingly. However, we cannot prescribe initial values on time-like hypersurfaces
in general. This can be explained as follows. Let be a time-like hypersurface.
For some point p on , the domain of influence of p includes other points on .
Values of solutions at those points, say q, depend on those at p. Hence, initial
values cannot be prescribed arbitrarily at points like q.
Of course, the above illustration is valid only for n > 1. The one-dimensional
wave equation utt uxx = 0 has a symmetric form with respect to x and t, and

158

6. WAVE EQUATIONS

Figure 6.16. Space-like and time-like surfaces.

Figure 6.17. An integral domain for space-like initial hypersurfaces.

we can always turn around the influence domain if we are allowed to exchange the
role of space and time. We emphasize that this works only for initial-value prob-

Figure 6.18. An illustration for n = 1.

lems. For initial/boundary-value problems for the one-dimensional wave equation
we discussed in Section 6.1, we do not have freedom to exchange the role of space
and time. The space and time variables are already fixed. The initial curve is not
necessarily given by t = 0. It can be any space-like curve, i.e., any curve t = t(x)
with |t0 (x)| < 1. The two vertical lines are time-like. We point out that we did not
prescribe initial values there. (Recall that initial values consist of two conditions
for a second-order PDE.) Instead, we prescribed Dirichlet boundary values u = 0
there.

6.3. ENERGY ESTIMATES

159

Figure 6.19. Initial/boundary-value problems.

To end our discussion of space-like and time-like surfaces, we consider an example of initial-value problems with initial values prescribed on a time-like hypersurface. Consider
utt = uxx + uyy

for x > 0, y, t R,

u
1
1
=
sin my on x = 0.
sin my,
2
m
n
m
Here we treat {x = 0} as the initial hypersurface. We note that {x = 0} is time-like.
A solution is given by
1
um (x, y) = 2 emx sin my.
m
Note
um
um 0 on x = 0,
0 on x = 0 as m .
n
Meanwhile,
u=

sup |um (x, )| =

R2

1 mx
e as m for any x > 0.
m2

Therefore, there is no continuous dependence on initial values.

Now we return to Theorem 6.10. As in Theorem 2.12 and Corollary 2.13, we
note that t enters the estimate only through the domain D,T,t. Hence, we let
t and obtain the following result.
Theorem 6.12. Assume a is C 1 , c and f are continuous in Rn [0, T ], is
C in Rn and is continuous in Rn . Let (6.24) be assumed and u be a C 2 -solution
of (6.23). If f L2 (Rn (0, T )) and , x , L2 (Rn ), then for any 0
Z
Z
t 2
2
2
e (u + ut + a|u| ) + ( 0 )
et (u2 + u2t + a|u|2 )
Rn {T }
Rn (0,T )
Z
Z

(2 + 2 + a||2 ) +
et f 2 ,
1

Rn {0}

Rn (0,T )

where 0 is a positive constant depending only on , the C 1 -norm of a and the

L -norm of c in Rn [0, T ].
With Theorem 6.12, we can prove the existence of weak solutions of (6.23) by
a similar process used in the proof of Theorem 2.15. However, there is a significant
difference. The weak solutions in Theorem 2.15 are in L2 because estimates of
the L2 -norms of solutions are established in Corollary 2.13. In the present situation, Theorem 6.12 establishes an estimate of the L2 -norms of solutions and their

160

6. WAVE EQUATIONS

derivatives. This naturally leads to a new norm defined by

Z
kukH 1 (Rn (0,T )) =

(u +
Rn (0,T )

u2t

! 12
2

+ |x u| )

The superscript 1 in H 1 indicates the order of derivatives. With such a norm, we

can define the Sobolev space H 1 (Rn (0, T )) as the completion of smooth functions
of finite H 1 -norms with respect to the H 1 -norm. Obviously, H 1 (Rn (0, T )) defined
in this way is complete. In fact, it is a Hilbert space, since the H 1 -norm is naturally
induced by an H 1 -inner product given by
Z
(uv + ut vt + x u x v).
(u, v)H 1 (Rn (0,T )) =
Rn (0,T )

Then proceeding as in the proof of Theorem 2.15, we can prove (6.23) admits a
weak H 1 -solution in Rn (0, T ) if = = 0. We will not provide details here.
In fact, we did not define the notion of weak H 1 -solutions. The purpose of the
short discussion here, similar as that at the end of Section 3.3, is to demonstrate
the importance of Sobolev spaces in PDEs. We will study Sobolev spaces in detail
in the future.
Exercises
(1) Let l be a positive number and consider
utt uxx = 0

in (0, l) (0, ),

u(, 0) = , ut (, 0) =

in [0, l],

u(0, t) = 0, ux (l, t) = 0 for t > 0.

Find a compatibility condition and prove the existence of a C 2 -solution
under such a condition.
(2) Let 1 and 2 be functions defined in {x < 0} and {x > 0} respectively.
Consider the characteristic initial-value problem
utt uxx = 0

for t > |x|,

u(x, x) = 1 (x)

for x < 0,

u(x, x) = 2 (x) for x > 0.

Solve and find the domain of dependence for any point (x, t) with t > |x|.
(3) Let 1 and 2 be functions defined in {x > 0}. Consider the Goursat
problem
utt uxx = 0

for 0 < t < x,

u(x, 0) = 1 (x), u(x, x) = 2 (x) for x > 0.

Solve and find the domain of dependence for any point (x, t) with 0 < t <
x.

EXERCISES

161

(4) Let be a constant and and be C 2 -functions on R+ which vanish

near x = 0. Consider
utt uxx = 0

for x > 0, t > 0,

u(x, 0) = (x), ut (x, 0) = (x)

for x > 0,

ut (0, t) = ux (0, t) for t > 0.

Find a solution for 6= 1 and prove that there exist no solutions in
general for = 1.
(5) Let be a positive constant. Use the method of descent to solve the
following initial-value problems for n = 2
utt = u + 2 u in R2 R+ ,
u(, 0) = 0, ut (, 0) =

in R2 ,

and
utt = u 2 u in R2 R+ ,
u(, 0) = 0, ut (, 0) =

in R2 .

Hint: You need to use complex functions temporarily to solve the second
problem.
(6) Let u be a solution of the following initial-value problem
utt u = 0

in Rn R+ ,

u(, 0) = 0, ut (, 0) =

on Rn .

(a) Let n 3 be odd and I(r, t) be the spherical average of u at (x, t).
Set
n3

2
n2

1
r
I(r, t) .
J(r, t) =
r r
Prove J(r, t) satisfies the 1-dimensional wave equation, i.e., Jtt
Jrr = 0.
(b) Let n 3 be odd. Derive a formula for u(x, t).
(c) Let n 2 be even. Derive a formula for u(x, t) by the method of
descent.
(7) Let be a smooth function with compact support in Rn and u be a
solution of
utt u = 0

in Rn R+ ,

u(, 0) = 0, ut (, 0) =

on Rn .

Prove for any t > 1

n2

|u(, t)|L (Rn ) Ct

n1
2

2
X

ki kL1 (Rn )

for even n 2,

i=0

and
|u(, t)|L (Rn ) Ct

n1
2

n1
2

kL1 (Rn )

for odd n 3,

where C is a positive constant depending only on n.

162

6. WAVE EQUATIONS

(8) Let be a smooth function with compact support in Rn and u be a

solution of
utt u = 0

in Rn R+ ,

u(, 0) = 0, ut (, 0) =

on Rn .

Prove for any n 2 and for any t > 0

|u(, t)|L (Rn ) Ckn1 kL1 (Rn ) ,
where C is a positive constant depending only on n.
(9) Let u be a solution of the following initial-value problem
utt u = 0

in R3 R+ ,

u(, 0) = , ut (, 0) =

on R3 .

(a) For any fixed (x0 , t0 ) R3 R+ , set for any x Bt0 (x0 ) \ {x0 }

x u(x, t)
x x0
x x0

v(x) =
+
u(x,
t)
+
u
(x,
t)
.
t

3
2
|x x0 |
|x x0 |
|x x0 |
t=t0 |xx0 |
Prove divv = 0.
(b) Derive an expression of u(x0 , t0 ) in terms of and by integrating
divv in Bt0 (x0 ) \ B (x0 ) and then letting 0.
Remark: This exercise gives an alternative approach to solving the 3dimensional wave equation.
(10) Let a be a positive constant and u be a C 2 -solution of the following characteristic initial-value problem
utt u = 0 in {(x, t) R3 R+ ; t > |x| > a},
u(x, |x|) = 0

for any |x| > a.

(a) For any fixed (x0 , t0 ) R3 R+ with t0 > |x0 | > a, integrate divv
(introduced in the previous problem) in a region bounded by |x
x0 | + |x| = t0 , |x| = a and |x x0 | = . By letting 0, express
u(x0 , t0 ) by an integral over Ba .
(b) For any S2 and > 0, prove the limit

lim ru(r, r + )
r

exists and the convergence is uniform for S2 and (0, 0 ] for

any fixed 0 > 0.
Remark: The limit in (b) is called the radiation field.
(11) Set QT = {(x, t); 0 < x < l, 0 < t < T }. Consider the equation
Lu 2utt + 3utx + uxx = 0.
(a) Give a correct presentation of the boundary-value problem in QT ;
(b) Find an explicit expression of a solution with prescribed boundary
values;
(c) Derive an estimate of the integral of u2x + u2t in QT .

EXERCISES

163

Hint: For (b), divide QT into three regions separated by characteristic

curves from (0, 0). For (c), integrate an appropriate linear combination of
ut Lu and ux Lu to make integrands on [0, l] {t} and {l} [0, t] positive
definite.
(12) For some constant a > 0, consider the following characteristic initial-value
problem for the wave equation
utt u =f (x, t) in a < |x| < t + a,
u =(x, t) on |x| > a, t = |x| a,
u =(x, t)
1

on |x| = a, t > 0,

where f is a C -function in a < |x| < t + a, is a C 1 -function on

r0 < |x| = t a and is a C 1 -function on |x| = a and t > 0. Derive an
energy estimate in an appropriate domain in a < |x| < t + a.

CHAPTER 7

First-Order Differential Systems

In this chapter, we discuss partial differential systems of first-order and focus
on non-characteristic initial-value problems. In Section 7.1, we introduce noncharacteristic hypersurfaces for partial differential equations and systems of arbitrary order. We also demonstrate that partial differential systems of arbitrary
order can always be changed to those of first-order. In Section 7.2, we discuss the
Cauchy-Kowalevski Theorem, which asserts the existence of analytic solutions of
non-characteristic initial-value problems for analytic differential systems and initial
values on analytic non-characteristic hypersurfaces. In Section 7.3, we discuss hyperbolic differential systems and derive energy estimates for symmetric hyperbolic
differential systems.
7.1. Reductions to First-Order Differential Systems
The main focus in this section is linear partial differential systems. We start
with linear PDEs of arbitrary order and proceed similarly as in Section 2.1 and
Section 3.1.
Let be a domain in Rn containing the origin and L be an m-th order linear
differential operator given by
X
Lu =
a (x) u in ,
||m

where a is continuous in , for any Zn+ . Here a is called the coefficient for
u.
Definition 7.1. The principal part L0 and the principal symbol p of L are
defined by
X
a (x) u in ,
L0 u =
||=m

and
p(x; ) =

a (x)

for any x and Rn .

||=m

The principal symbol p plays an important role in prescribing appropriate conditions associated with L. We usually write first-order and second-order linear
differential operators in following forms respectively
Lu =

n
X

ai (x)uxi + b(x)u,

i=1
165

166

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

and
Lu =

n
X

aij (x)uxi xj +

i,j=1

n
X

bi (x)uxi + c(x)u.

i=1

Their principal symbols are given by

p(x; ) =

n
X

ai (x)i ,

i=1

and
p(x; ) =

n
X

aij (x)i j .

i,j=1

For second-order differential operators, we usually assume that (aij ) is a symmetric

matrix in .
For a given function f in , we consider the equation
(7.1)

Lu = f (x) in .

The function f is called the non-homogeneous term.

Let be the hyperplane {xn = 0}. We now prescribe values of u and its
derivatives on so that we can at least find all derivatives of u at the origin.
Let u0 , u1 , , um1 be functions defined in a neighborhood of the origin in Rn1 .
First, we prescribe
u(x0 , 0) = u0 (x0 ) for any small x0 Rn1 .
Then we can find all x0 -derivatives of u at the origin. Similarly, for any j =
1, , m 1, we prescribe
xj n u(x0 , 0) = uj (x0 )

for any small x0 Rn1 .

Then we can find all x0 -derivatives of xj n u at the origin for j = 1, , m1. Hence
with the help of u0 (x0 ), u1 (x0 ), , um1 (x0 ) on , we can determine all derivatives
of u of order up to m at the origin except xmn u. To find this, we need to use the
equation. If we assume
a(0, ,0,m) (0) 6= 0,
m
then we can find xn u(0) from (7.1). In this case, we can compute all derivatives
of u of any order at the origin by using values u0 , u1 , , um1 and differentiating
(7.1).
Now we summarize these conditions on = {xn = 0} by writing
(7.2)

xj n u(x0 , 0) = uj (x0 )

for any small x0 Rn1 and j = 0, 1, , m 1.

We usually call the initial hypersurface and u0 , , um1 initial values or Cauchy
values. The problem of solving (7.1) together with (7.2) is called the initial value
problem or the Cauchy problem.
More generally, consider the hypersurface given by { = 0} for a C m -function
in a neighborhood of the origin with 6= 0, with the origin on the hypersurface
, i.e., (0) = 0. We note that is simply a normal vector of the hypersurface .
Without loss of generality, we assume xn (0) 6= 0. Then by the implicit function
theorem, we solve = 0 around x = 0 for xn = (x1 , , xn1 ). Consider a change
of variables
x 7 y = (x1 , , xn1 , (x)).

7.1. REDUCTIONS TO FIRST-ORDER DIFFERENTIAL SYSTEMS

167

This is a well defined transform in a neighborhood of the origin with a nonsingular

Jacobian. Now we write the operator L in new variables y as
X
Lu =
a
(y)y u.
||m

The initial hypersurface is given by {yn = 0} in new coordinates. We are interested in the coefficient a
(0, ,0,m) of ymn u. Note
uxi =

n
X

yk,xi uyk = xi uyn + terms not involving uyn ,

k=1

and hence
n m
m
1
x u = x11 xnn u =
x1 xn yn u + terms not involving yn u.

Therefore,
Lu =

a (x)x u =

||m

1
n m
a x(y)
x1 xn yn u

||=m

+ terms not involving ymn u.

The coefficient of ymn u is given by
X

1
n
a x(y)
x1 xn .
||=m

Back to x-variables, we note that this is simply

p x; (x) ,
where p is the principal symbol of the operator L.
Definition 7.2. For a linear operator L as in (7.1) defined in a neighborhood
of x0 Rn , a C 1 -hypersurface passing x0 is non-characteristic at x0 if
X
(7.3)
p(x0 ; ) =
a (x0 ) 6= 0,
||=m

where is a normal vector of at x0 . Otherwise, it is characteristic at x0 . A

hypersurface is non-characteristic if it is non-characteristic at every point.
When the hypersurface is given by { = 0} with 6= 0, its normal vector
is given by = (x1 , , xn ). Hence we may take = (x0 ) in (7.3). We
note that the condition (7.3) is maintained under C m -changes of local coordinates.
By this condition, we can find successively values of all derivatives of u at x0 , as
far as they exist. Then, we could write formal power series at x0 for solutions
of initial-value problems. It would be actual representations of solutions u in a
neighborhood of x0 , if u were known to be analytic. This process can be carried
out for analytic initial values and analytic coefficients and nonhomogeneous terms
and the result is referred to as the Cauchy-Kowalevski theorem. We will discuss it
in Section 7.2.
Now we introduce a special class of linear differential operators.
Definition 7.3. The linear differential operator L in (7.1) is elliptic at x0 if
p(x0 ; ) 6= 0 for any Rn \ {0}, where p is the principal symbol of L.

168

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Hence, linear differential equations are elliptic if there are no characteristic

hypersurfaces. In other words, every hypersurface is non-characteristic.
For example, consider a first-order linear differential operator of the form
Lu =

n
X

ai (x)uxi + b(x)u in Rn .

i=1

Its principal symbol is given by

p(x; ) =

n
X

for any x and any Rn \ {0}.

ai (x)i

i=1

Hence first-order linear differential equations with real coefficients are never elliptic.
Complex coefficients may yield elliptic equations. For example, we take a1 = 1/2
and a2 = 1/2i in R2 . Then z = (x + iy )/2 is elliptic.
For second-order linear differential operators, we write
n
X

Lu =

aij (x)uxi xj +

i,j=1

n
X

bi (x)uxi + c(x)u

in Rn .

i=1

Its principal symbol is given by

p(x; ) =

n
X

aij (x)i j

for any x and any Rn \ {0}.

i,j=1

Then L is elliptic at x if
n
X

aij (x)i j 6= 0

for any Rn \ {0}.

i,j=1

If (aij (x)) is a real-valued n n symmetric matrix, L is elliptic at x if (aij (x)) is a

definite matrix at x, positive definite or negative definite.
The concept of non-characteristics can also be generalized to linear partial
differential systems. Let Rn be a domain. Consider a linear partial differential
systems of N equations for N unknowns as follows
X
(7.4)
Lu
A (x) u = f in ,
||m

where A are N N matrices and u and f are N -column vectors.

We define principle parts, principle symbols and non-characteristic hypersurfaces similarly as for single differential equations.
Definition 7.4. The principal part L0 and the principal symbol p of L are
defined by
X
L0 u =
A (x) u in ,
||=m

and
p(x; ) = det

X
||=m

A (x)

for any x and Rn .

7.1. REDUCTIONS TO FIRST-ORDER DIFFERENTIAL SYSTEMS

169

Definition 7.5. For a linear operator L as in (7.4) defined in a neighborhood

of x0 Rn , a C 1 -hypersurface passing x0 is non-characteristic at x0 if
X

p(x0 ; ) = det
A (x0 ) 6= 0,
||=m

where is a normal vector of at x0 . Otherwise, it is characteristic at x0 . A

hypersurface is non-characteristic if it is non-characteristic at every point.
Let be a non-characteristic hypersurface. We prescribe initial values on as
follows. Let be a normal vector on . We prescribe
j u
= uj on for j = 0, 1, , m 1,
j
where u0 , u1 , , um1 are N -column vectors on .
We now demonstrate that we can always reduce the order of differential systems
to 1 by increasing the number of equations and the number of components of
solution vectors.
(7.5)

Proposition 7.6. Let be a non-characteristic hypersurface at x0 . Then

the initial-value problem (7.4)-(7.5) in a neighborhood of x0 is equivalent to an
initial-value problem of a first-order differential system with initial values prescribed
on .
Proof. We assume x0 is the origin. In the following, we write x = (x0 , xn )
and = (0 , n ).
Step 1. Straightening initial hypersurfaces. Assume is given by { = 0} for
a C m -function in a neighborhood of the origin with xn 6= 0. Then we introduce
a change of coordinates x = (x0 , xn ) 7 (x0 , (x)). In the new coordinates, still
denoted by x, the hypersurface is given by {xn = 0} and the initial condition
(7.5) is given by
xj n u(x0 , 0) = uj (x0 )

for j = 0, 1, , m 1.

Step 2. Reductions to canonical forms and zero initial values. In the new
coordinates, {xn = 0} is non-characteristic at 0. Then, the coefficient matrix
A(0, ,0,m) (x) is a nonsingular matrix at x = 0 and hence also in a neighborhood of
the origin. Multiplying the partial differential system by the inverse of this matrix,
we assume that A(0, ,0,m) is the identity matrix in a neighborhood of the origin.
Next, we may assume
uj (x0 ) = 0

for j = 0, 1, , m 1.

To this end, we introduce a function v such that

u=v+

m1
X
j=0

1
uj (x0 )xjn .
j!

Then the differential system for v is the same as that for u with f replaced by
f (x)

m1
X

j=0 ||m

A (x)

uj (x0 )xjn .
j!

170

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

With Step 1 and Step 2, we assume (7.4)-(7.5) have the following form
xmn u +

m1
X

n =0

|0 |mn

A u = f,

with
xj n u(x0 , 0) = 0

for j = 0, 1, , m 1.

Step 3. Lowering the order. We now replace the above system by an equivalent
one of order m 1. Introduce new functions
U0 = u,

Ui = uxi , i = 1, , n,

and
U = (U0 , U1 , , Un )T .
Then
U0,xn = Un ,

Ui,xn = Un,xi , i = 1, , n 1.

Hence
xm1
U0 xm2
Un = 0,
n
n
xm1
Ui xi xm2
Un = 0,
n
n

i = 1, , n 1.

To get an (m 1)-th order differential equation for Un , we write the equation for
u as
m1
X
X
X
0
xmn u +
A u +
A(0 ,0) ( ,0) u = f.
n =1 |0 |mn

|0 |m

We substitute Un = uxn in the first two terms in the right-hand side to get
xm1
Un
n

m2
X

A Un +

n =0 |0 |mn 1

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

|0 |m

In the last summation in the left hand side, any m-th order term of u can be
changed to an (m 1)-th order term for Ui for some i = 1, , n 1, since no
derivatives with respect to xn are involved. Now we can write a differential system
for U in the following form
xm1
U+
n

m2
X
n =0

X
|0 |m

(1)
.
A(1)
U =F
n 1

The initial value for U is given by

xj n U (x0 , 0) = 0

for j = 0, 1, , m 2.

Now we prove that the initial-value problem for the new (m1)-th order differential
system is equivalent to that for the original m-th order differential system. In
particular, if U is a solution of the new (m 1)-th order differential system, then
U0 is a solution of the original m-th order differential system. To this end, we prove
Ui = U0,xi , for i = 1, , n. First, by the first equation in the system for U and
initial conditions for U , we have
xm2
(Un U0,xn ) = 0,
n

7.2. CAUCHY-KOWALEVSKI THEOREM

171

and on t = 0
xj n (Un U0,xn ) = 0

for j = 0, , m 3.

This implies easily Un = U0,xn . Next, for i = 1, , n 1,

xm1
Ui xi xm2
Un = xm1
Ui xi xm1
U0 = xm1
(Ui xi U0 ).
n
n
n
n
n
By the second equation in the system for U and initial conditions, we have
(Ui xi U0 ) = 0,
xm1
n
and on t = 0
xj n (Ui xi U0 ) = 0

for j = 0, , m 2.

Hence, Ui = U0,xi , for i = 1, , n1. Substituting Ui = U0,xi in the final equation

in the system for U , we conclude that U0 is a solution for the original m-th order
differential system.
Now, we can repeat the procedure to reduce m to 1.

We point that straightening initial hypersurfaces and reducing initial values to

zero are frequently used techniques in discussions of initial-value problems.
7.2. Cauchy-Kowalevski Theorem
For any given first-order linear partial differential system in a neighborhood
of x0 Rn and an initial value u0 prescribed on a hypersurface passing x0 , we
first intend to find a solution u formally. To this end, we need to determine all
derivatives of u at x0 , in terms of the initial value u0 and all known functions in the
equation. Obviously, all tangential derivatives (with respect to ) of u are given by
derivatives of u0 . In order to find derivatives of u involving the normal direction, we
need help of the equation. It has been established that, if is non-characteristic at
x0 , the initial-value problem leads to evaluations of all derivatives of u at x0 . This
is clearly a necessary first step to the determination of a solution of the initial-value
problem. If the coefficient matrices and initial values are analytic, a Taylor series
solution could be developed for u. The Cauchy-Kowalevski Theorem asserts the
convergence of this Taylor series in a neighborhood of x0 .
To motivate our discussion, we study an example of first-order partial differential systems which may not have solutions in any neighborhood of the origin, unless
initial values prescribed on analytic non-characteristic hypersurfaces are analytic.
Example 7.7. Let g = g(x) be a real valued function in R. Consider the
following system in R2+ = {(x, y); y > 0}
(7.6)

ux vy =0,
uy + vx =0,

with prescribed initial values given by

u = g(x), v = 0

on y = 0.

Note that (7.6) is simply the Cauchy-Riemann equation in C = R2 . It can be

written in the matrix form as follows

1 0
u
0 1
u
0
+
=
.
0 1
v y
1 0
v x
0

172

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Note that {y = 0} is non-characteristic. In fact, there are no characteristic curves.

To see this, we take any = (1 , 2 ) R2 and then have

1 0
0 1
1 2
+ 2
=
.
1
0 1
1 0
2 1
Its determinant is 12 + 22 , which is not zero for any 6= 0. Therefore there are
no characteristic curves. We now write (7.6) in complex form. Suppose we have a
solution (u, v) for (7.6) with the given initial values and let w = u + iv. Then
wx + iwy = 0 for y > 0 with w|y=0 = g(x).
Therefore, w is (complex) analytic in the upper half plane and its imaginary part is
zero on the x-axis. By the Schwartz reflection principle, w can be extended across
y = 0 to an analytic function in C = R2 . This implies in particular that g is (real)
analytic since w|y=0 = g. We conclude that (7.6) does not admit any solutions with
the given initial value on {y = 0} unless g is real analytic.
Example 7.7 naturally leads to discussions of analytic solutions. We introduced
real analyticity in Section 4.3. We now discuss this subject in detail.
To introduce (real) analytic functions, we need to study convergence of multiple
infinite series of the form
X
c ,

where c are real numbers defined for all multi-indices Zn+ . Throughout this
section,
Hence, a
P to absolute convergence.
P the term convergence always refers
series c is convergent if and only if |c | < . Here, the summation is over
all multi-indices Zn+ .
Definition 7.8. A function u : Rn R is called analytic near 0 if there exists
an r > 0 and constants {u } such that
X
u(x) =
u x for any |x| < r.

If u is analytic near 0, then u is smooth near 0. Moreover, the constants u

are given by
1
u = u(0) for any Zn+ .
!
Thus u equals its Taylor series about 0, i.e.,
X 1
u(x) =
u(0)x for any |x| < r.
!

Now we give an important analytic function.

Example 7.9. For r > 0, set
u(x) =

r
r (x1 + + xn )

r
for any |x| < .
n

7.2. CAUCHY-KOWALEVSKI THEOREM

173

Then

1 X

x1 + + xn k
x1 + + xn
u(x) = 1
=
r
r
k=0

X ||!
X
1 X k
=
x =
x .
k
|| !

r
r

k=0
||=k

This power series is absolutely convergent for |x| < r/ n since

X ||!
X
|x1 | + + |xn | k

|x
|
=
< ,
||
r
r !

k=0

for |x1 | + + |xn | |x| n < r. We also note

||!
for any Zn+ .
r||
We point out that all derivatives of u at 0 are positive.
u(0) =

An effective method to prove analyticity of functions is to control their derivatives by derivatives of functions known to be analytic. For this, we first introduce
the following terminology.
Definition 7.10. Let u and v be two C -functions defined in a neighborhood
of the origin in Rn . Then v majorizes u, denoted by v u or u v, if
v(0) | u(0)|

for any Zn+ .

We also call v a majorant of u.

We have the following simple method to verify analyticity.
Lemma 7.11. If v u and v is analytic near 0, then u is analytic near 0.
Proof. We prove that the Taylor series of u about 0 converges for |x| < r if
the Taylor series of v about 0 converges for |x| < r. We simply note
X 1
X 1
| u(0)x |
v(0)|x | < for any |x| < r.
!
!

We hence have the desired convergence for u.

Next, we prove that every analytic function has a majorant.

Lemma 7.12. If the Taylor series of u is convergent for |x| < r and 0 < s n <

r, then u has an analytic majorant for |x| s/ n.

Proof. By setting y = s(1, , 1), we have |y| = s n < r and hence

X 1
u(0)y
!

is a convergent series. Then there exists a constant C such that

1
| u(0)y | C for any Zn+ ,
!
and in particular,
C
1
||!
| u(0)| 1
C || .
!
y1 ynn
s !

174

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Now set

X ||!
Cs
.
=C
s (x1 + + xn )
s|| !

Then v majorizes u for |x| s/ n.

v(x)

So far, our discussions are limited to scalar-valued functions. All results can be
generalized to vector-valued functions easily. For example, a vector-valued function
u = (u1 , , uN ) is analytic if each of its component is analytic.
We have following results for compositions of functions.
Lemma 7.13. Let u, v be C -functions of a neighborhood of 0 Rn into Rm
and f, g be C -functions of a neighborhood of 0 Rm into RN , with u(0) = 0,
f (0) = 0, u v and f g. Then f u g v.
Lemma 7.14. Let u be an analytic function of a neighborhood of 0 Rn into
R and f be an analytic function of a neighborhood of u(0) Rm into RN . Then
f u is an analytic function in a neighborhood of 0 Rn .
m

We leave proof as an exercise.

Now we are ready to discuss real analytic solutions of non-characteristic initialvalue problems for real analytic equations and initial values. We study first-order
quasilinear partial differential systems of N equations for N unknowns with initial
values prescribed on non-characteristic hyperplanes.
For convenience, we study initial-value problems in Rn+1 = {(x, t)} with initial
values prescribed on {t = 0}. Consider
(7.7)

ut =

n
X

Aj (x, t, u)uxj + F (x, t, u),

j=1

with
(7.8)

u = u0

on {t = 0},

where A1 , , An are N N matrices in (x, t) and F are N -column vectors. We

assume all functions are analytic in their arguments.
The next result is referred to as the Cauchy-Kowalevski theorem.
Theorem 7.15. Let A1 , , An be analytic N N matrices and F be an analytic N -column vector near (0, 0, u0 (0)) Rn+1+N and u0 be an analytic function
near 0 Rn . Then (7.7)-(7.8) admits an analytic solution u near 0 Rn+1 .
Proof. Without loss of generality, we assume u0 (x) = 0. To this end, we
introduce an analytic function v by v(x, t) = u(x, t) u0 (x). Then the differential
system for v is similar as that for u. Next, we add t as an additional component
of u by introducing uN +1 such that uN +1,t = 1 and uN +1 |t=0 = 0. This increases
the number of equations and the number of components of solution vectors in (7.7)
by 1 and meanwhile deletes t from A1 , , An and F . We still denote by N the
number of equations and the number of components of solution vectors.
In the following, we study
n
X
(7.9)
ut =
Aj (x, u)uxj + F (x, u),
j=1

7.2. CAUCHY-KOWALEVSKI THEOREM

175

with
(7.10)

u=0

on {t = 0},

where A1 , , An are analytic N N matrices and F is an analytic N -column

vector in a neighborhood of the origin in Rn+N . We seek an analytic solution u
in a neighborhood of the origin in Rn+1 . To this end, we will compute u(0) in
terms of derivatives of A1 , , An and F at (0, 0) Rn+N and then prove that the
Taylor series of u at 0 converges in a neighborhood of 0 Rn+1 . We note that t
does not appear explicitly in the right hand side of (7.9).
Since u = 0 on {t = 0}, we have
x u(0) = 0

for any Zn+ .

Fix i = 1, , n and differentiate (7.9) with respect to xi to get

uxi t =

n
X
(Aj uxi xj + Aj,xi uxj + Aj,u uxi uxj ) + Fu uxi + Fxi .
j=1

In view of (7.10), we have

uxi t (0) = Fxi (0, 0).
More generally, we obtain by induction
x t u(0) = x F (0, 0)

for any Zn+ .

Next, for any Zn+ , we have

x t2 u =x (ut )t = x

n
X

Aj uxj + F

j=1

=x

n
X

(Aj uxj t + Aj,u ut uxj ) + Fu ut .

j=1

Here we used the fact that Aj and F are independent of t. Thus,

x t2 u(0) = x

n
X

(Aj uxj t + Aj,u ut uxj ) + Fu ut x=0,t=0,u=0 .

j=1

The expression on the right hand side can be worked out to be a polynomial
with nonnegative coefficients involving various derivatives of A1 , , An and F
and derivatives x tl u with || + l || + 2 and l 1.
More generally, for any Zn+ and k 0, we have

(7.11)
x tk u(0) = p,k (x u A1 , , x u An , x u F, x tl u)x=0,t=0,u=0 ,
where p,k is a polynomial with nonnegative coefficients and indices , , range
over , Zn+ and ZN
+ with ||+|| ||+k 1, ||+l ||+k and l k 1.
We need to point out that p,k (x u A1 , ) is considered as a polynomial of
components of x u A1 , . We denote by p,k (|x u A1 |, ) the value of p,k
when all components of x u A1 , are replaced by their absolute values. Since
p,k has nonnegative coefficients, we conclude

p,k (x u A1 , , x u An , x u F, x tl u)

x=0,t=0,u=0
(7.12)

p,k (|x u A1 |, , |x u An |, |x u F |, |x tl u|)x=0,t=0,u=0 .

176

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

We now consider a new differential system

n
X
vt =
Bj (x, v)vxi + G(x, v)
(7.13)
j=1
v =0

on {t = 0},

where B1 , , BN are analytic N N matrices and G is an analytic N -column

vector in a neighborhood of the origin in Rn+N . We will choose B1 , , Bn and G
such that
(7.14)

Bj Aj for j = 1, , n

and

G F.

and

x u G(0) |x u F (0)|.

Hence, for any (, ) Zn+N

,
+
x u Bj (0) |x u Aj (0)| for j = 1, , n

The above inequalities should be understood as for each components.

Let v be a solution of (7.13). We now claim
|x tk u(0)| x tk v(0)

for any (, k) Zn+1

+ .

The proof is by induction on the order of t-derivatives. The general step follows
since

|x tk u(0)| =|p,k (x u A1 , , x u An , x u F, x tl u)x=0,t=0,u=0 |

p,k (|x u A1 |, , |x u An |, |x u F |, |x tl u|)x=0,t=0,u=0

p,k (x u B1 , , x u Bn , x u G, x tl v)x=0,t=0,v=0
=x tk v(0),
where we used (7.11), (7.12) and the fact that p,k has positive coefficients. Thus
(7.15)

v u.

It remains to prove the Taylor series of v at 0 converges in a

0 Rn+1 .
To this end, we consider

1
Cr
..
B1 = = Bn =

r (x1 + + xn + u1 + + uN ) .
1
and
G=

Cr
r (x1 + + xn + u1 + + uN )

neighborhood of

..
.

1
.. ,
.
1

1
..
. .
1

for positive constants C and r with |x| + |u| < r/ n + N . As demonstrated in the
proof of Lemma 7.12, we may choose C sufficiently large and r sufficiently small
such that (7.14) holds.
Set

1
..
v = w . ,
1

7.2. CAUCHY-KOWALEVSKI THEOREM

177

for some scalar-valued function w in a neighborhood of 0 Rn+1 . Then (7.13) is

reduced to
n
X
Cr
wt =
(N
wxi + 1),
r (x1 + + xn + N w)
i=1
w =0

on {t = 0}.

This is a (single) first-order quasilinear partial differential equation and has a solution of the form
w(x1 , , xn , t) = w(x
1 + + xn , t).
Then w
= w(z,
t) satisfies
Cr
(nN w
z + 1),
r z Nw
w
=0 on {t = 0}.

w
t =

By using the method of characteristics as in Section 2.2, we have an explicit solution

1
1
w(z,
t) =
r z [(r z)2 2Cr(n + 1)N t] 2 .
(n + 1)N
and hence
w(x, t) =

n
n
X
X

1
1
r
xi [(r
xi )2 2Cr(n + 1)N t] 2 .
(n + 1)N
i=1
i=1

This expression is analytic for |(x, t)| < s, for sufficiently small s > 0. Hence, the
corresponding solution v of (7.13) is analytic for |(x, t)| < s. By Lemma 7.11 and
(7.15), uPis analytic for |(x, t)| < s. Since the Taylor series of the analytic functions
n
ut and j=1 Aj (x, u)uxj + F (x, u) have the same coefficients at the origin, they
agree throughout the region |(x, t)| < s.

At the beginning of the proof, we introduced an extra component for the solution vector to get rid of t in coefficient matrices of the differential system. Had we
chosen to preserve t, we would have to solve the initial-value problem
Cr
(nN w
z + 1),
r z t Nw
w
=0 on {t = 0}.

w
t =

It is difficult, if not impossible, to find an explicit expression of the solution w.

The solution given in Theorem 7.15 is the only analytic solution since all derivatives of the solution are computed at the origin and they uniquely determine the
analytic solution. A natural question is whether non-analytic solution is also unique.
For this, we discuss initial value problems of linear differential systems as follows
(7.16)

A0 (x, t)ut +

n
X

Aj (x, t)uxj + B(x, t)u =F (x, t),

j=1

u(x, 0) = 0,
where A0 , A1 , , An , B are analytic N N matrices and F is an analytic N -column
vector in a neighborhood of the origin in Rn+1 .
The next result is referred to as the local Holmgren uniqueness theorem.

178

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Theorem 7.16. Let A0 , A1 , , An , B be analytic N N matrices and F be an

analytic N -column vector near the origin in Rn+1 and u0 be an analytic function
near 0 Rn . If {t = 0} is non-characteristic at the origin, then any C 1 -solution of
(7.16) is analytic in a sufficiently small neighborhood of the origin in Rn+1 .
In the proof, we need a notion of adjoint operators. Let L be a differential
operator defined by
Lu = A0 (x, t)ut +

n
X

Ai (x, t)uxi + B(x, t)u.

i=1

For any vector-valued functions u and v, we write

v T Lu = (v T A0 u)t +

n
X

n
X

T
(v T Ai u)xi AT0 vt +
(ATi v)xi B T v u.

i=1

i=1

We define the adjoint operator L of L by

L v = (AT0 v)t

n
X

(ATi v)xi + B T v

i=1

= AT0 vt

n
X

ATi vxi + (B T AT0,t

i=1

n
X

ATi,xi )v.

i=1

Then
v T Lu = (v T A0 u)t +

n
X

T
(v T Ai u)xi + L v u.

i=1

Proof. We introduce an analytic change of coordinates so that the initial

hypersurface = {t = 0} becomes a paraboloid
t=

n
X

x2i .

i=1
1

We will prove that any C -solution u of Lu = 0 with a zero initial value on is in

fact zero. For any > 0, we set
= {(x, t);

n
X

x2i < t < }.

i=1

We will prove u = 0 in for a sufficiently small . In the following, we denote by

+ and the upper and lower boundary of respectively, i.e.,
+ ={(x, t);

n
X

x2i < t = },

i=1

={(x, t);

n
X

x2i = t < }.

i=1

We note that det(A0 (0)) 6= 0 since is non-characteristic at the origin. Hence

A0 is nonsingular in a neighborhood of the origin. By multiplying the equation in
(7.16) by A1
0 , we assume A0 = I.

7.2. CAUCHY-KOWALEVSKI THEOREM

179

For any vector-valued function v in a neighborhood of the origin containing ,

we have
Z
Z
Z
0=
v T Lu =
uT L v +
uv T .

There is no boundary integral on since u = 0 there. Let Pk = Pk (x) be

an arbitrary polynomial in Rn , k = 1, , n, and form P = (P1 , , PN ). We
consider the initial value problem
L v =0

in Br ,

v =P

on Br {t = },

where Br is the ball with center at the origin and radius r in Rn+1 . The principal
part of L is the same as that of L, except a different sign and a transpose. We fix
r so that {t = } Br is non-characteristic for L , for each small . By Theorem
7.15, an analytic solution v exists in Br for any small. We need to point out
that the domain of convergence of v is independent of P , whose components are
polynomials. We choose small such that Br . Then we have
Z
uP T = 0.
+

By Weierstrass approximation theorem, any continuous function in a compact domain can be approximated in the L -norm by a sequence of polynomials. Therefore, u = 0 on + for any small and hence in .

Theorem 7.15 guarantees the existence of solutions of initial-value problems in

the analytic setting. As the next example shows, we do not expect any estimates
of solutions in terms of initial values.
Example 7.17. In R2 , consider the first-order homogeneous linear differential
system (7.6)
ux vy =0,
uy + vx =0.
Note that all coefficients are constant. As shown in Example 7.7, there are no
characteristic curves. Consider for any integer k 1
uk (x, y) = sin(kx)eky ,

vk (x, y) = cos(kx)eky .

Then uk and vk satisfy (7.6) and on {y = 0}

uk (x, 0) = sin(kx),

vk (x, 0) = cos(kx).

Obviously,
u2 (x, 0) + v 2 (x, 0) = 1,
and
u2k (x, y) + vk2 (x, y) = e2ky as k for any y > 0.
Therefore, there is no continuous dependence on initial values. This illustrates that
initial-value problems are not well posed for (7.6).

180

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

7.3. Hyperbolic Differential Systems

In this section, we study the non-characteristic initial-value problem for partial
differential systems without analyticity and introduce an important class of firstorder partial differential systems hyperbolic differential systems. The basic idea
underlying the hyperbolicity of a differential system is that the initial-value problem
should be well-posed. Specifically, initial values on non-characteristic hypersurface
are sufficient to determine unique solutions that depend continuously on values
specified at points on .
We consider a first-order linear differential system in Rn R = {(x, t)} in the
following form
(7.17)

A0 (x, t)ut +

n
X

Ak (x, t)uxk + B(x, t)u = f (x, t),

k=1

where u and f are column vectors with N elements and A0 , A1 , , An and B are
N N matrices. We prescribe an initial value on the hyperplane {t = 0} by
(7.18)

u(x, 0) = u0 (x).

In the following, we always assume that A0 (x, t) is nonsigular for any (x, t), i.e.,
det(A0 (x, t)) 6= 0.
Hence, the hypersurface {t = 0} is non-characteristic.
For (7.17), principal symbols defined in Definition 7.4 have the following form
(7.19)

p(x, t; , ) = det

n
X

k Ak (x, t) + A0 (x, t) ,

k=1
n+1

n+1

for any (x, t) R

and (, ) R
. For any Rn , p(x, t; , ) is a polynomial
of of degree N . If A0 = I, p(x, t; , ) is simply the characteristic polynomial of
the matrix
n
X

k Ak (x, t).
k=1

Now, we start to investigate solvability of initial value problems associated with

(7.17). As discussed in Section 7.2, we can find all derivatives of u at t = 0, in terms
of A0 , A1 , , An , B, f and u0 . This determines a formal solution for small t. If the
coefficient matrices and initial values are analytic, a Taylor series solution could be
developed for u. The Cauchy-Kowaleski Theorem asserts the convergence of this
Taylor series at least in a neighborhood of any point (x, 0).
In this section, we discuss the non-analytic case. Extra assumptions are needed
for the well-posedness of initial-value problems.
As a motivation, we study a constant coefficient system
(7.20)

A0 ut +

n
X

Ak uxk = 0.

k=1

For some constant C and constant vectors Rn and v CN , we seek a

solution u of the form
u(x, t) = ei(x+t) v.

7.3. HYPERBOLIC DIFFERENTIAL SYSTEMS

181

An easy calculation shows that u is a solution of (7.20) if and only if

n
X

(7.21)

k Ak + A0 v = 0.

k=1

Suppose (7.21) holds for some C \ R, Rn \ {0} and v CN , with Im < 0.

We consider a family of solutions
1
1
ur (x, t) = Re{eir(x+t) v} = etIm Re{eir(x+tRe) v},
r
r
for r > 0. Then,
ur (x, 0) 0

as r ,

and
ur (x, t)

as r for any t > 0.

This simple example illustrates that the initial-value problem for (7.20) is not wellposed if Im 6= 0.
In summary, a necessary condition for initial-value problems (7.17) to be wellposed is that the principal symbol p(x, t; , ) admits only real roots for any
Rn \ {0}.
We are led naturally to the following definition of hyperbolicity.
Definition 7.18. The differential system (7.17) is hyperbolic at (x, t) if for any
unit vector Rn the principal symbol p(x, t; , ) has N real zeros 1 , , N and
there exist N linearly independent v1 , , vN in RN satisfying
n
X

k Ak (x, t) + i A0 (x, t) vi = 0, i = 1, , N.

k=1

The differential system (7.17) is strictly hyperbolic at (x, t) if p(x, t; , ) has N

distinct real zeros 1 , , N .
The notion of hyperbolicity is invariant under the change of coordinates of the
form
x = x(y, s), t = s.
An important class of hyperbolic differential system is symmetric hyperbolic
differential systems.
Definition 7.19. The differential system (7.17) is symmetric hyperbolic at
(x, t) if A0 (x, t), A1 (x, t), , An (x, t) are symmetric and A0 (x, t) is positive definite.
Obviously, symmetric hyperbolic differential systems are indeed hyperbolic.
The canonical form of (7.17) is given when A0 = I. In general, we can obtain
canonical forms easily. Since A0 is assumed to be nonsingular, we may multiply
(7.17) by A1
0 to get an equivalent system
ut +

n
X
k=1

t)u = f(x, t).

Ak (x, t)uxk + B(x,

182

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Its principal symbol is simply given by

p(x, t; , ) = det

n
X

k Ak (x, t) + I .

k=1

If the differential system (7.17) is symmetric hyperbolic, we wish to preserve the

symmetry. Since A0 is positive definite, we may write
A0 = M t M,
where M (x, t) is an N N nonsingular matrix. Introduce a new dependent variable
vector v by
v = M u.
Then v satisfies
vt +

n
X

t)v = f(x, t),

Ak (x, t)vxk + B(x,

k=1

where Ak = M t Ak M 1 is still symmetric.

Now we study solvability of the initial-value problem (7.17)-(7.18). If coefficient matrices and nonhomogeneous terms in (7.17) and initial values in (7.18) are
analytic, then analytic solutions exist locally in any neighborhood of (x, 0). This is
the Cauchy-Kawaleski Theorem. No hyperbolicity condition is required. However,
we do not have any estimates for solutions in general, as Example 7.17 suggests.
If N = 1, the system (7.17) is reduced to a single equation for a scalar-valued
function u. The hyperbolicity condition is redundant. The initial-value problem
(7.17)-(7.18) can be solved by integrating along integral curves. An energy estimate
was derived in Section 2.3.
If N > 1, the symmetry plays an essential role in solving the initial-value problem (7.17)-(7.18). Symmetric hyperbolic differential systems in general dimensions
behave like single differential equations of a similar form. An energy estimate can
be derived similarly and hence the problem (7.17)-(7.18) can be solved in weak
sense. Now, we derive such an estimate.
We consider the initial-value problem for a first-order linear differential system
in Rn R = {(x, t)} in the following form

(7.22)

A0 (x, t)ut +

n
X

Ak (x, t)uxk + B(x, t)u =f (x, t),

k=1

u(x, 0) =u0 (x),

where u and f are column vectors with N elements and A1 , , An and B are
N N matrices. In the following, we always assume
A0 , A1 , , An are symmetric in Rn (0, T ),
and
A0 is positive definite.
We take positive constants a and such that
(7.23)

A0 aI

in Rn R+ ,

7.3. HYPERBOLIC DIFFERENTIAL SYSTEMS

183

and
(7.24)

A0 +

n
X

i Ai 0

in Rn R+ ,

i=1

for any Rn with || 1.

We recall the domain D,T,t defined in Section 2.3
D,T,t = {(x, t); |x| < t t, 0 < t < T },
for fixed T, t > 0. Its boundary consists of several pieces given by
D,T,t ={(x, 0); |x| < t},
s D,T,t ={(x, t); |x| = t t, 0 < t < T },
+ D,T,t ={(x, T ); |x| < t T }.
Theorem 7.20. Let A0 , A1 , , An be C 1 symmetric N N matrices satisfying
(7.23)-(7.24), B a continuous N N matrix, f a continuous function and u be a
C 1 -solution of (7.22) in Rn R+ . Then for any 0 < T < t,
(Z
)
Z
Z
t
2
2
t
2
e |u| C
|u0 | +
e |f | ,
D,T ,t

D,T ,t

D,T ,t

where and C are positive constants depending only on a in (7.23), the C 1 -norms
of Ai and the sup-norm of B in D,T,t.
The estimate in Theorem 7.20 implies the uniqueness of solutions and the
property of finite speed propagation.
Before proving Theorem 7.20, we point out the role of symmetry. The proof
proceeds similarly as that for scalar equations Section 2.3. We take an inner product
of the equation (7.22) with 2u and hence need to analyze terms such as 2uT A0 ut
and similar terms for x-derivatives. If A0 is symmetric, then
2uT A0 ut = uTt A0 u + uT A0 ut = (uT A0 u)t uT A0,t u.
It is not necessary to assume symmetry for B since no derivatives of u are involved
with B. We simply note
2uT Bu = uT (B + B T )u,
and B T + B is always symmetric.
Proof. For a nonnegative , we multiply the equation in (7.22) by 2et uT .
By
2et uT A0 ut =(et uT A0 u)t + et |u|2 et uT A0,t u,
2et uT Ai uxi =(et uT Ai u)xi et uT Ai,xi u,
we have
(et uT A0 u)t +

n
X
(et uT Ai u)xi
i=1

+et uT (A0

n
X
i=1

Ai,xi + B + B T )u = 2et uT f.

184

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

We write D = D,T,t. An integration in D yields

Z
Z
n
X
et uT A0 u +
et uT (t +
Ai i )u
+ D

s D

i=1

Z
et uT (A0

+
D

Ai,xi + B + B T )u

i=1

2et uT f,

u A0 u +

n
X

where (1 , , n , t ) is the unit exterior normal vector of s D given by

x
1
(1 , , n , t ) =
,1 .
1 + 2 |x|
First, we note by (7.24)
n
n
X
X
1
xi
t +
Ai ) 0 on s D.
ai i =
(A0 +
2
|x|
1+
i=1
i=1
Next, we choose such that
n
X
A0
Ai,xi + B + B T 2aI

in D.

i=1

Then

Z
et |u|2 sup |A0 |

2a
D

Z
|u0 |2 +

2et uT f.
D

Here we simply dropped integrals over + D and s D since they are nonnegative.
The Cauchy inequality implies
Z
Z
Z
1
t T
t
2
2e u f a
e |u| +
et |f |2 .
a D
D
D
We then have the desired result.

By letting t , we have the following result.

Theorem 7.21. Let A0 , A1 , , An be C 1 symmetric N N matrices satisfying
(7.23)-(7.24), B a continuous N N matrix, f a continuous functions and u be
a C 1 -solution of (7.22) in Rn R+ . For any T > 0, if f L2 (Rn (0, T )) and
u0 L2 (Rn ), then
(Z
)
Z
Z
t
2
2
t
2
e |u| C
|u0 | +
e |f | ,
Rn (0,T )

Rn

Rn (0,T )

where and C are positive constants depending only on a, the C 1 -norms of Ai and
the sup-norm of B in Rn (0, T ).
Proceeding as in Section 2.3, we can discuss weak solutions of (7.22). For a
fixed T > 0, we consider functions in
DT = Rn (0, T ).
Denote by C0 (DT ) the collection of smooth functions in DT with compact supports
e (DT ) the collection of smooth functions in DT with compact
in DT and by C
0
supports in x-directions.

7.3. HYPERBOLIC DIFFERENTIAL SYSTEMS

185

Set
(7.25)

Lu = A0 (x, t)ut +

n
X

Ai (x, t)uxi + B(x, t)u in DT .

i=1

We introduce the adjoint operator L of L by

n
X
L v = (A0 v)t
(Ai v)xi + B T v
i=1

= A0 vt

n
X

Ai vxi + (B T A0,t

i=1

n
X

Ai,xi )v.

i=1

e , we have
For any u, v C
0
v T Lu = (v T A0 u)t +

n
X

T
(v T Ai u)xi L v u.

i=1

By a simple integration in DT , we have

Z
Z
Z
T
T
v Lu =
u L v+
DT

Rn {t=T }

DT

Z
T

v A0 u

Rn {t=0}

v T A0 u

e0 (DT ).
for any u, v C
We note that there are no derivatives of u in the right-hand side.
Definition 7.22. An L2 -function u is a weak solution of Lu = f in DT if
Z
Z
uT L v =
f T v for any v C0 (DT ).
DT

DT

Now we can prove the existence of weak solutions of (7.22) with homogeneous
initial values by the Hahn-Banach Theorem and the Riesz Representation Theorem
as in Section 2.3.
Theorem 7.23. Let A0 , A1 , , An be C 1 symmetric N N matrices satisfying
(7.23)-(7.24) and B a continuous N N matrix. Then for any f L2 (DT ), there
exists a u L2 (DT ) such that
Z
Z
T
e0 (DT ) with v = 0 on t = T.
u L v=
f T v for any v C
DT

DT

Moreover,
kukL2 (DT ) Ckf kL2 (DT ) ,
where C is a positive constant depending only on a, the C 1 -norms of Ai and the
sup-norm of B in Rn (0, T ).
The symmetry plays an essential role in Theorem 7.20, Theorem 7.21 and
Theorem 7.23. The remaining question is whether we can change an arbitrary
hyperbolic differential system into a symmetric hyperbolic differential system. The
discussion of this question in general case is beyond the scope of this book. Now
we study a special case.
When the space dimension is one, we can change strictly hyperbolic differential
systems to symmetric hyperbolic differential systems easily. For n = 1, we consider
(7.26)

ut + A(x, t)ux + B(x, t)u = f (x, t).

186

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

The principal symbol is given by

p(x, t; , ) = det A(x, t) + I .

The hyperbolicity condition means that all eigenvalues of A are real and eigenvectors span RN . If the system (7.26) is a constant coefficient homogeneous system of
the form
ut + Aux = 0,
then solution can be written explicitly. Suppose A has N real eigenvalues 1 , , N
and corresponding eigenvectors v1 , , vN , which span RN . Then solution u is
given by
N
X
u(x, t) =
k (x k t)vk ,
k=1

where functions 1 , , N are determined by the initial condition

u(x, 0) =

N
X

k (x)vk .

k=1

Back to the general system (7.26), we introduce a new solution vector v by

u = M v,
where M is an N N nonsingular matrix. Then a straightforward calculation yields
vt + M 1 AM vx + (M 1 BM + M 1 AMx + M 1 Mt )v = M 1 f.
This is a symmetric hyperbolic differential system if M 1 AM is symmetric. Now
we take M whose columns consist of eigenvectors of A. Then M 1 AM is a diagonal matrix with eigenvalues of A as diagonal entries, which is obviously symmetric.
Therefore, v satisfies a symmetric hyperbolic differential system. If M is C 1 , an
energy estimate can be derived. In fact, we can obtain a solution v by integrating
along characteristic curves. However, the matrix M consisting of eigenvectors of
A may not be C 1 when eigenvalues of high multiplicity are present. One condition to ensure regularity of eigenvectors is that all eigenvalues are simple, i.e., the
differential system (7.26) is strictly hyperbolic.
To end this section, we demonstrate that second-order hyperbolic differential
equations can always be changed to symmetric hyperbolic differential systems. We
consider
(7.27)

wtt

n
X
i,j=1

aij (x, t)wxi xj

n
X

bi (x, t)wxi c(x, t)w = f,

i=1

where (aij ) is positive definite in Rn [0, T ]. Here we assume aij , bi and c are
smooth, with
(7.28)

I (aij ) I

in Rn [0, T ],

for some positive constants . For convenience, we write

A = (aij ),

7.3. HYPERBOLIC DIFFERENTIAL SYSTEMS

187

and
ai =(ai1 , , ain )

for any i = 1, , n,

b =(b1 , , bn ).
Set

w
u1
u0 wt

U u1 = wx1 .
. .
.. ..
wxn
un

(7.29)

This is an (n + 2)-vector. We write the equation (7.27) as

u1,t = u0 ,
u0,t =

n
X

aij uj,xi +

i,j=1
n
X

aij uj,t =

j=1

n
X

n
X

bi ui + cu1 + f,

i=1

aij u0,xj ,

i = 1, , n.

j=1

Then we put it in a matrix form

(7.30)

A0 Ut +

n
X

Ai Uxi + BU = F,

i=1

where
A0 =

I22

0 0
0
ai ,
Ai = 0 0
0 aTi 0nn

0 1
0
B = c 0
b ,
0 0 0nn
and

FT = 0

i = 1, , n,

0 .

Obviously, (7.30) is symmetric hyperbolic since A0 , A1 , , An are symmetric and

A0 is positive definite.
We now consider the initial-value problem
(7.31)

wtt

n
X
i,j=1

aij wxi xj

n
X

bi wxi cw =f

in Rn (0, T ),

i=1

w(, 0) = w0 , wt (, 0) =w1

on Rn .

Then U defined in (7.29) satisfies (7.30) and initial conditions

(7.32)

U = (w0 , w1 , w0,x1 , , w0,xn )T

on Rn .

188

7. FIRST-ORDER DIFFERENTIAL SYSTEMS

Now we verify that the initial-value problem (7.31) is equivalent to the initialvalue problem (7.30) and (7.32). We only need to prove that a solution U =
(u1 , u0 , u1 , , un )T of (7.30) and (7.32) yields a solution w = u1 of (7.31). First,
by the first equation in the system (7.30) and the corresponding initial condition,
we have u0 = u1,t . Hence we have u1,t = w1 on t = 0. Next, by the last n
equations in (7.30), we have uj,t = u0,xj = u1,xj t , or (uj u1,xj )t = 0. Then we
get uj = u1,xj since they match on t = 0. Hence w = u1 is a solution of (7.31).
We then have the following result as a consequence of Theorem 7.21.
Theorem 7.24. Let aij be C 1 and bi , c be continuous in Rn (0, T ), (aij ) satisfy
(7.28) and w be a C 2 -solution of (7.31) in Rn (0, T ). If f L2 (Rn (0, T )) and
w0 , w0 , w1 L2 (Rn ), then
Z
Z

t
2
2
2
e (w + |w| + wt ) C
(|w0 |2 + |w0 |2 + w12 )
Rn (0,T )
Rn
Z

+
et |f |2 ,
Rn (0,T )

where and C are positive constants depending only on , , the C 1 -norms of aij
and the sup-norm of bi , c in Rn (0, T ).
Theorem 6.12 in Section 6.3 is a special case of Theorem 7.24.
Exercises
(1) Classify the following 4-th order equation in R3
2x4 u + 2x2 y2 u + y4 u 2x2 z2 u + z4 u = f.
(2) Prove Lemma 7.13 and Lemma 7.14.
(3) Consider the initial-value problem
utt uxx u = 0

in R R+ ,

u(x, 0) = x, ut (x, 0) = x.
Find a solution in a power series expansion about the origin and identify
this solution.
(4) Consider the initial-value problem for u : R (0, T ) RN as follows
ut + A(x, t)ux = f (x, t, u)

in R (0, T ),

with
u(, 0) = 0
1

on R,

where A is an N N diagonal C -matrix on R(0, T ) and f : R(0, T )

RN RN is a C 2 -map. Under appropriate conditions on f , prove the
above initial-value problem admits a C 1 -solution by using the contraction
mapping principle.
Hint: It may be helpful to write in a system of equations instead of in a
matrix form.

EXERCISES

189

(5) Consider in D = {(x, t); x > 0, t > 0} R2

ut + aux + b11 u + b12 v =f
vx + b12 u + b22 v =g,
with the condition
u(x, 0) = (x) for x > 0 and v(0, t) = (t) for t > 0.
Assume a is C 1 and bij are continuous in D.
(a) Assume a 0 in D. Derive an energy estimate for (u, v) in an
appropriate domain in D.
(b) Assume a 0 in D. For sufficiently small T , derive an estimate for
sup[0,T ] |u(0, )| in terms of sup-norms of f, g, and .
(c) Discuss whether similar estimates can be derived if a is positive somewhere along {(0, t); t > 0}.
(6) Consider in a neighborhood of the origin in R2 = {(x, t)}
ut + aux + b11 u + b12 v =f
vx + b12 u + b22 v =g,
with the condition
u(x, 0) = (x) and v(0, t) = (t).
Assume a, bij are analytic in a neighborhood of 0 R2 and , are
analytic in a neighborhood of 0 R.
(a) Prove all derivatives of u and v at 0 can be expressed in terms of
derivatives of a, bij , and at 0.
(b) Prove there exists an analytic solution (u, v) in a neighborhood of
0 R2 .

Bibliography
[1] Arnolds, V. I., Lectures on Partial Differential Equations, Universitext, Springer, 2004.
[2] Caffarelli, L., Cabr
e, X., Fully Nonlinear Elliptic Equations, AMS Colloquium Publications,
Vol. 43, American Math. Society, Providence, RI, 1993.
[3] Evans, L., Partial Differential Equations, Graduate Studies in Mathematics, Vol. 19, American Math. Society, Providence, RI, 1998.
[4] Friedman, A., Partial Differential Equations of Parabolic Type, Prentice-Hall, Englewood
Cliffs, 1964.
[5] Gilbarg, D., Trudinger, N., Elliptic Partial Differential Equations of Second Order, (2nd ed.),
Grundlehren der Mathematischen Wisenschaften, Vol. 224, Spinger-Verlag, Berlin-New York,
1983.
[6] Han, Q., Lin, F.-H., Elliptic Partial Differential Equations, Courant Institute Lecture Notes,
Volume 1, American Math. Society, Providence, RI, 2000.
[7] John, F., Partial Differential Equations, (4th ed.), Applied Math. Sciences, Vol. 1, SpringerVerlag, New York, 1991.
[8] MacRobert, T. M., Spherical Harmonics, An Elementary Treatise on Harmonic Functions
with Applications, Pergamon Press, Oxford-New York-Toronto, 1967.
[9] Taylor, M., Partial Differential Equations I: Basic Theory, Applied Math. Sciences, Vol. 115,
Springer-Verlag, New York, 1996.

191

Index

elliptic differential equations, 39, 168

energy estimates
first-order PDEs, 31
second-order hyperbolic equations, 159
symmetric hyperbolic differential
systems, 184
wave equations, 47
Euclidean norms, 1
Euler-Poisson-Darboux equation, 143

a priori estimates, 3
adjoint differential operators, 32, 178
analytic functions, 67, 172
auxiliary functions, 81, 120
Bernstein method, 83
Burgers equation, 18
canonical forms
first-order systems, 181
Cauchy problems, 8, 38, 166
Cauchy values, 8, 38, 166, 169
Cauchy-Kowalevski theorem, 174
characteristic curves, 9
characteristic hypersurfaces, 9, 38
non-characteristic hypersurfaces, 167, 169
characteristic ODEs, 14, 16, 21
compact supports, 32
comparison principles, 78, 80, 118
compatibility conditions, 20, 52, 54, 137,
139
conservation of energies, 48, 153
convergence of series, 67, 172
absolute convergence, 172
convolutions, 103

finite speed propagation, 149

finite-speed propagation, 27
first-order linear differential systems, 180
canonical forms, 181
initial-value problems, 180
first-order linear PDEs, 7
initial-value problems, 24
Fourier series, 50
Fourier transforms, 102
inverse Fourier transforms, 104
frequency, 99
fundamental solutions, 107
Laplace equations, 60
gradient estimates
interior gradient estimates, 66, 75, 83,
114, 121
gradients, 1
Greens formula, 30, 43
Greens function, 53, 61
Greens function in balls, 62
Greens identity, 60

dAlemberts formula, 134

decay estimates, 150
degenerate differential equations, 39
diameters, 45
differential Harnack inequalities, 75
Dirichlet energy, 92
Dirichlet problems, 44, 61, 74
Greens function, 61
divergence theorem, 43
domain of influence, 13
domains, 1
domains of dependence, 13, 27, 134, 148
domains of influence, 27, 134, 148
doubling condition, 99
Duhamels principle, 142

H
older continuity, 90
pointwise H
older continuity, 86
half-space problems, 137
harmonic functions, 40, 59
conjugate harmonic functions, 40
converegence of Taylor series, 67
differential Harnack inequalities, 75, 83
doubling condition, 99
frequency, 99
Harnack inequalities, 75, 85

eigenvalue problems, 50, 56

193

194

interior gradient estimates, 66, 75, 83

Liouville Theorem, 75
mean-value properties, 72
nodal sets, 68
removable singularity, 85
superharmonic functions, 77
Harnack inequalities, 75, 85, 129
differential Harnack inequalities, 75, 83,
125, 128
heat equation
fundamental solutions, 107
heat equations
1 dimension, 41
analyticity of solutions, 117
differential Harnack inequalities, 125, 128
Harnack inequalities, 129
initial/boundary-value problems, 46, 49
interior gradient estimates, 114, 121
n dimensions, 42
Poisson integral formula, 106
Hessian matrices, 1
Holmgren uniqueness theorem, 177
Hopf lemma, 78
Huygens principle, 149
hyperbolic differential equations, 39, 43
hyperbolic differential systems, 181
strictly hyperbolic systems, 181
symmetric hyperbolic systems, 181
hypersurfaces, 2
initial hypersurfaces, 8, 38, 166
initial values, 8, 38, 166, 169
initial-value problems, 166
first-order PDEs, 8, 11
first-order systems, 180
second-order PDEs, 38
wave equations, 133, 142
initial/boundary-value problems
heat equations, 46, 49
wave equations, 47, 54, 139
integral curves, 13
Laplace equations, 39, 42
fundamental solutions, 60
Greens indentity, 60
Poisson integral formula, 64
Poisson kernel, 64
Legendre equations, 69
Legendre functions, 70
linear differential systems
m-th order, 168
first-order, 180
linear PDEs, 2
first-order, 7
m-th order, 165
second-order, 37

INDEX

Liouville Theorem, 75
loss of differentiations, 152
majorants, 173
maximum principles, 74
strong maximum principles, 79
weak maximum principles, 77, 117, 118
mean-value properties, 72
method of characteristics, 14
method of descent, 146
method of extensions, 137, 140
method of reflections, 138, 140
method of spherical averages, 143
mixed problems, 47
multi-indices, 1
Neumann problems, 45
nodal curves, 70
nodal sets, 68
non-characteristic curves, 9
non-characteristic hypersurfaces, 9, 11, 38
non-homogeneous terms, 7, 37, 166
parabolic boundaries, 112, 117
parabolic differential equations, 43
parabolic distance, 113
Parsevals identity, 105
partial differential equations (PDEs), 2
linear PDEs, 2
quasilinear PDEs, 2
PDEs of mixed type, 41
Poincar`
e Lemma, 45
Poisson equations, 42, 86
weak solutions, 92
Poisson integral formula, 64, 106
Poisson kernel, 64
principal parts, 165, 168
principal symbols, 37, 165, 168, 180
propagation of singularities, 41
quasilinear PDEs, 2
radiation field, 162
Rellichs theorem, 94
removable singularity, 85
Schauder estimates, 90
pointwise Schauder estimates, 88
Schwartz class, 102
second-order linear PDEs, 37
in the plane, 39
separation of variables, 49
space variables, 1
space-like surfaces, 157
spherical harmonics, 68
sectorial harmonics, 70
tesseral harmonics, 70

INDEX

zonal harmonics, 70
strictly hyperbolic systems, 181
subsolutions, 77
subharmonic functions, 77
supersolutions, 77
superharmonic functions, 77
symmetric hyperbolic systems, 181
Taylor series, 67, 172
terminal value problems, 112
time variables, 1
time-time surfaces, 157
Tricomi equation, 41
uniform ellipticity, 77
wave equations
1 dimension, 40, 133
2 dimensions, 146
3 dimensions, 144
Huygens principle, 149
decay estimates, 150
energy estimates, 153
half-space problems, 137
initial-value problems, 133, 142
initial/boundary-value problems, 47, 54,
139
n dimensions, 42, 142
radiation field, 162
weak solutions, 32, 92, 159, 185
Weierstrass approximation theorem, 179
well-posed problems, 3

195