Sie sind auf Seite 1von 190

JUHA KINNUNEN

Partial Differential Equations

Department of Mathematics and Systems Analysis, Aalto University


2017
Contents

1 INTRODUCTION 1

2 FOURIER SERIES AND PDES 5


2.1 Periodic functions* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 The Lp space on [, ]* . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 The Fourier series* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 The best square approximation* . . . . . . . . . . . . . . . . . . . . . 19
2.5 The Fourier series on a general interval* . . . . . . . . . . . . . . . . . 24
2.6 The real form of the Fourier series* . . . . . . . . . . . . . . . . . . . . 24
2.7 The Fourier series and differentiation* . . . . . . . . . . . . . . . . . . 27
2.8 The Dirichlet kernel* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.9 Convolutions* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.10 A local result for the convergence of Fourier series* . . . . . . . . . . 32
2.11 The Laplace equation in the unit disc . . . . . . . . . . . . . . . . . . 34
2.12 The heat equation in one-dimension . . . . . . . . . . . . . . . . . . . 45
2.13 The wave equation in one-dimension . . . . . . . . . . . . . . . . . . 52
2.14 Approximations of the identity
in [, ]* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.15 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

3 FOURIER TRANSFORM AND PDES 63


3.1 The L p -space on Rn * . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.2 The Fourier transform* . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
3.3 The Fourier transform and differentiation* . . . . . . . . . . . . . . . . 67
3.4 The Fourier transform of the Gaussian* . . . . . . . . . . . . . . . . . . 71
3.5 The Fourier inversion formula* . . . . . . . . . . . . . . . . . . . . . . . 72
3.6 The Fourier transformation and convolution . . . . . . . . . . . . . . 76
3.7 Plancherels formula* . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
3.8 Approximations of the identity in Rn * . . . . . . . . . . . . . . . . . . . 79
3.9 The Laplace equation in the upper half-space . . . . . . . . . . . . 81
3.10 The heat equation in the upper half-space . . . . . . . . . . . . . . . 86
3.11 The wave equation in the upper half-space . . . . . . . . . . . . . . 91
3.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
CONTENTS ii

4 LAPLACE EQUATION 93
4.1 Gauss-Green theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
4.2 PDEs and physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.3 Boundary values and physics . . . . . . . . . . . . . . . . . . . . . . . 97
4.4 Fundamental solution of the Laplace equation . . . . . . . . . . . . 101
4.5 The Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.6 Greens function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
4.7 Greens function for the upper half-space* . . . . . . . . . . . . . . . 115
4.8 Greens function for the ball* . . . . . . . . . . . . . . . . . . . . . . . 117
4.9 Mean value formulas . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
4.10 Maximum principles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
4.11 Harnacks inequality* . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.12 Energy methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.13 Weak solutions* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.14 The Laplace equation in other coordinates* . . . . . . . . . . . . . . 133
4.15 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

5 HEAT EQUATION 140


5.1 Physical interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5.2 The fundamental solution . . . . . . . . . . . . . . . . . . . . . . . . . 142
5.3 The nonhomogeneous problem . . . . . . . . . . . . . . . . . . . . . 143
5.4 Separation of variables in Rn . . . . . . . . . . . . . . . . . . . . . . . 146
5.5 Maximum principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
5.6 Energy methods for the heat equation . . . . . . . . . . . . . . . . . 155
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

6 WAVE EQUATION 158


6.1 Physical interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
6.2 The one-dimensional wave equation . . . . . . . . . . . . . . . . . . 159
6.3 Euler-Poisson-Darboux equation . . . . . . . . . . . . . . . . . . . . . 166
6.4 The three-dimensional wave equation . . . . . . . . . . . . . . . . . 169
6.5 The two-dimensional wave equation . . . . . . . . . . . . . . . . . . 173
6.6 The nonhomogeneous problem . . . . . . . . . . . . . . . . . . . . . 177
6.7 Energy methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
6.8 Epilogue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

7 NOTATION AND TOOLS 183


Partial differential equations are not only extremely impor-
tant in applications of mathematics in physical, geometric
and probabilistic phenomena, but they also are of theoretic
interest.

Introduction
1
These notes are meant to be an elementary introduction to Partial Differential
Equations (PDEs) for undergraduate students in mathematics, the natural sci-
ences and engineering. They assume only advanced multidimensional differential
calculus including partial derivatives, integrals and the Gauss-Green formulas.
The sections denoted by * consist of additional material, which is essential in
understanding the rest of the material, but can omitted or glanced through quickly
in the first reading.
A partial differential equation is an equation involving an unknown function
of two ore more variables and its partial derivatives. Although PDE are general-
izations of ordinary differential equations (ODE), for most PDE problems it is not
possible to write down explicit formulas for solutions that are common in the ODE
theory. This means that there is greater emphasis on qualitative features. There
is no general method to solve all PDE, however, some methods have turned out
to be more useful than other. We study special cases, in which explicit solutions
and representation formulas are available and focus on features that are present
in more general situations. Qualitative aspects are also important in numerical
solutions of PDE. Without existence, uniqueness and stability results numerical
methods may give inaccurate or completely wrong solutions.
Let x , where is an open subset of Rn and t R. In these notes we study

(1) Laplaces equation


n 2 u
u = f , u =
X
u = u( x), ,
j =1 x2j

(2) the heat equation


u
u = f , u = u( x, t),
t

1
CHAPTER 1. INTRODUCTION 2

(3) and the wave equation

2 u
u = f u = u( x, t).
t2
Here we have set all physical constants equal to one. Physically, solutions of
Laplaces equation correspond to steady states for time evolutions such as heat
flow or wave motion, with f corresponding to external driving forces such as
heat sources or wave generators. A solution u = u( x) to Laplaces equation gives,
for example, the temperature at the point x and a solution u = u( x, t) to
the heat equation gives the temperature at the point x at the moment of
time t. A solution u = u( x, t) to the wave equation gives the displacement of
a body at the point x in the moment of time t. We shall later discuss the
origin and interpretation of these PDEs in more detail. If f = 0 (the function,
which is identically zero), the PDE is called homogeneous, otherwise it is said
to be inhomogeneous. All PDEs above are linear, which means that any linear
combination of solutions is a solution. More precisely, if u 1 and u 2 are solutions,
then au 1 + bu 2 , a, b R, is a solution of the corresponding equation as well.
By solving a PDE we mean that we find all functions u verifying the PDE in
a class of functions, which possibly satisfy certain auxiliary conditions. A PDE
typically has many solutions, but there may be only one solution satisfying specific
boundary or initial value conditions conditions. These conditions are motivated
by the physics and describe the physical state at a given moment or/and on the
boundary of the domain. For Laplaces equation we can describe, for example,
the temperature on the boundary . For the heat equation we can, in addition,
describe the initial temperature and for the wave equation the initial velocity at a
given moment of time. By finding a solution to a PDE we mean that we obtain
explicit representation formulas for solutions or deduce general properties that
hold true for all solutions. A PDE problem is well posed, if

(1) (existence) the problem has a solution,


(2) (uniqueness) there exist only one solution and
(3) (stability) the solution depends continuously on the data given in the
problem.

These are all desirable features when we talk about solving a PDE. The last
condition is particularly important in physical problems, since we would like that
our (unique) solution changes little when the conditions specifying the problem
change little.
There is at least one more important aspect in solving PDE. We have not yet
specified what does it mean that a function actually is a solution to a PDE. We
shall consider classical solutions, which means that all partial derivatives which
appear in the PDE exist and are continuous. In this case, we can verify by a
direct computation that a function solves the PDE. However, the PDE can be so
CHAPTER 1. INTRODUCTION 3

strong that it forces the solution to be smoother than assumed in the beginning.
A PDE may also have physically relevant weak solutions with less regularity
than classical solutions, think for example a saw tooth wave. These questions are
studied in regularity theory for PDE.
The PDE above are examples of the three most common types of linear equa-
tions: Laplaces equation is elliptic, the heat equation is parabolic and the wave
equation is hyperbolic, although general classification is somewhat useless since
it does not give any method to solve the PDE. There are many other PDE that
arise from physical problems. Let us consider, for example, Maxwells equations.
Let R3 be an open set then R is the corresponding space-time cylinder.
Maxwells equations are

div E = ,





0

div B = 0,

B
curl E = ,


t


E


curl B = 0 J + 0

,

t
where E is the electric field and B is the magnetic field (which both are maps form
R R3 ) corresponding to a charge density and a current density J (which
are functions from R R and R R3 correspondingly). Here 0 and 0 are
positive physical constants called the permittivity and permeability of free space,
respectively. Recall that the divergence of a vector field E = (E 1 , E 2 , E 3 ) is
n E
X i
div E = E =
i =1 x i

and the curl is


E 3
E 2 E 1 E 3 E 2 E 1

curl E = E = , , .
x2 x3 x3 x1 x1 x2

If there are no charges or currents in Maxwells equations, we have





div E = 0,



div B = 0,


B
curl E = ,
t



2 E 1



curl B = c , c= p .


t 0 0

Since (exercise)
curl(curl E ) = (div E ) div(E ),

where the divergence is taken componentwise, that is div(E ) = (div E 1 , div E 2 , div E 3 ),
CHAPTER 1. INTRODUCTION 4

for any C 2 -function E : R3 , we have

E = divE = (div E ) div(E )


| {z }
=0
B

= curl(curl E ) = curl
t
2 E

= (curl B) = c
t t t

and thus
2 E
c2 E = .
t2
That is, each component of E = (E 1 , E 2 .E 3 ) satisfies the wave equation with the
speed of waves c. Similarly, B satisfies the same wave equation. These are the
electromagnetic waves.
Another special case of Maxwells equations is electrostatistics. In this case
there is no current and the field is independent of the time t. Then we have
curl E = 0, which implies that E is a gradient of a function (in a simply connected
domain ). Thus E = V , where V is called the electrostatic potential. Then

div E = div (V ) = V

so that

V = .
0
That is, V is a solution to inhomogeneous Laplaces equation, called Poissons
equation. Note that V is defined only up to an additive constant, which does not
affect the negative gradient E .
Fourier series is a series representation of a function de-
fined on a bounded interval on the real axis as trigonometric
polynomials. The function does not have to be smooth, but
the convergence of a Fourier series is a delicate issue. How-

2
ever, the Fourier series gives the best square approximation
of the function and it has many other elegant and useful
properties. It also converges pointwise, if the function is
smooth enough. Solutions to several problems in partial
differential equations, including the Laplace operator, the
heat operator and the wave operator, can be obtained using
Fourier series and convolutions.

Fourier series and PDEs

Historically the study of the motion of vibrating string fixed at its end points, and
later the heat flow in a one-dimensional rod, lead to the development of Fourier
series and Fourier analysis. These physical phenomena are modeled by partial
differential equations and, as we shall see, these problems can be solved using
Fourier series. Fourier claimed that for an arbitrary function
n n
fb( j ) e i jt =
X X
S n f ( t) = fb( j ) (cos( jt) + i sin( jt)) f ( t) as n ,
j = n j = n

where
1
Z
f ( j) =
b f ( t) e i jt dt, j Z.
2
In other words, any function defined on a bounded interval on the real axis can be
represented as a Fourier series

fb( j ) e i jt .
X
f ( t) =
j =

This is somewhat analogous to Taylor series in the sense that it gives a method to
express a given function as an infinite sum of the elementary functions e j ( t) = e i jt ,
j Z,. As we shall see, the convergence of the Fourier series is a delicate issue and
it depends on in which sense the limit is taken. Fourier analytic methods play an
important role in solving linear PDEs in physics and they have many applications
in several branches of mathematics. A useful property of the functions e j ( t) from
the PDE point of view is that each basis vector is an eigenfunction of the derivative
operator in the sense that

e0j ( t) = i je j ( t), j Z.

We shall start by taking a more careful look at the Fourier series. The Fourier
series apply only for periodic functions. This is not a serious restriction, as we
shall see.

5
CHAPTER 2. FOURIER SERIES AND PDES 6

2.1 Periodic functions*


We are mainly interested in real valued functions, but complex numbers are useful
not only in Fourier analysis but also in PDEs. We say that a function f : R C is
2-periodic if for every t R we have

f ( t + 2) = f ( t). (2.1)

More generally, a function f : R C is called T -periodic , T R, T 6= 0, if

f ( t + T ) = f ( t) (2.2)

for every t R. Observe, that the period T is not unique. If f is T -periodic, then
it is also nT -periodic for every n = 1, 2, . . . . The smallest positive value of T (if
exists) for which (2.2) holds is called the fundamental period. We shall consider
functions f on [, ] with f () = f (), and assume that they are 2-periodic by
extending f periodically to the whole R. In order to study a 2-periodic function f
it is enough to do so on any interval of length 2. For this course we mainly work
with the basic interval [, ], but we could choose any other interval of length 2
as well.

T HE MORAL : Every function f : [a, b] C defined on an interval with finite


endpoints can be extended to a periodic function to the whole R. Thus it is not too
restrictive to consider periodic functions.

If f and g are T -periodic functions with a common period T , then their product
f g and linear combination a f + b g, a, b C, are also T -periodic. To prove the
latter statement, let F ( t) = a f ( t) + b g( t). Then

F ( t + T ) = a f ( t + T ) + b g( t + T ) = a f ( t) + b g( t) = F ( t).

The former statement is left as an exercise.


Examples 2.1:
(1) A constant function f : R C, f ( t) = a, a C, is T -periodic for every T R.
There is no fundamental period.
(2) The function f : R R, f ( t) = sin t is 2-periodic.
(3) Let L > 0. The functions

jt jt

f : R R, f ( t) = sin and g : R R, g( t) = cos , j = 1, 2, . . . ,
L L
2L
are j -periodic.
(4) The function f : R C, f ( t) = e i jt , j Z, is 2-periodic, since

f ( t + 2) = e i j( t+2) = e i jt |e i{z
2 j
} = f ( t)
=1
CHAPTER 2. FOURIER SERIES AND PDES 7

Figure 2.1: A graph of a periodic function.

for every t, since by Eulers formula

e i2 j = cos(2 j ) + i sin(2 j ) = 1, j Z.

However, 2 is not the fundamental period of f . In the same way as above


2
we can show that f is | j | -periodic for j 6= 0. The fundamental period of f is
2
| j| for j 6= 0.
(5) The function f : R R,
1, t Q,
f ( t) =
0, t Q,

is T -periodic for every T Q, T > 0. There is no fundamental period.

Reason. If t Q, then t + T Q and f ( t) = f ( t + T ) = 1. If t R \ Q, then


t + T R \ Q and f ( t) = f ( t + T ) = 0. Thus f ( t) = f ( t + T ) for every t R.

Lemma 2.2. Let f : R C be a T -periodic function for some T > 0. Then for every
a R we have Z T Z a+ T
f ( t) dt = f ( t) dt.
0 a

THE MORAL : The integrals of a 2-periodic function over intervals of length


2 coincide. In other words, the integral is independent of the interval.
CHAPTER 2. FOURIER SERIES AND PDES 8

Proof. If a is of the form kT for some integer k, then


Z a+ T Z ( k+1)T
f ( t) dt = f ( t) dt.
a kT

By changing variables s = t kT we have


Z a+ T Z T Z T
f ( t) dt = f ( s + kT ) ds = f ( s) ds
a 0 0

since f is T -periodic and f ( s) = f ( s + T ) = = f ( s + kT ) for every s R, k Z.


Now if a is not of the form kT there exists a unique k such that

kT a < ( k + 1)T.

This is because the intervals [ kT, ( k + 1)T ) partition the real line. Thus
Z a+ T Z ( k+1)T Z a Z a+ T
f ( t) dt = f ( t) dt f ( t) dt + f ( t) dt (2.3)
a kT kT ( k+1)T

where observe that a + T > kT + T = ( k + 1)T . By the case a = kT already considered


we have Z ( k+1)T Z T
f ( t) dt = f ( t) dt.
kT 0
For the last term in (2.3) we change variables s = t T and get
Z a+ T Z a Z a
f ( t) dt = f ( s + T ) ds = f ( s) ds
( k+1)T kT kT

by the periodicity of f . This shows that the last two terms in (2.3) cancel each
other. This proves the claim.

Remark 2.3. There is a natural connection between 2-periodic functions on R


and functions on the unit circle. A point on the unit circle is of the form e i , where
is a real number that is unique up to integer multiples of 2. If F is a function
on the unit circle, then we may define for each real number

f ( ) = F ( e i ),

and observe that with this definition, the function f is 2-periodic. Thus 2-
periodic functions on R and functions on any interval of length 2 that take on
the same value at its end points are same mathematical objects.

2.2 The L p space on [, ]*


To be able to consider functions that are not necessarily smooth, we develop the
theory of L p spaces. The most important spaces are L1 and L2 , which are needed
in the definition and properties of Fourier series.
CHAPTER 2. FOURIER SERIES AND PDES 9

Definition 2.4. Let 1 p < . A function f : [, ] C belongs to L p ([, ]), if


1
1
Z
p
k f kL p ([,]) = | f ( t)| p dt < .
2

The number k f kL p ([,]) is called the L p -norm of f .

THE MORAL : Instead of of the absolute value of the function, a power of the
absolute value of the function is required to be integrable. Geometrically this
means that the area of the graph of | f | p is finite. If p = 2, when we talk about
square integrable functions. In particular, functions belonging to L p ([, ]) do
not have to be continuous or smooth. The only requirement is that the integral
above makes sense and is finite.

Remark 2.5. Note that


Z
k f kL p ([,]) < | f ( t)| p dt < .

1 1
The factor 2 and the power p are not more than normalising parameters. For
example, if f : [, ] R, f ( t) = 1, then

k f kL p ([,]) = 1 and ka f kL p ([,]) = |a|, a R.

This shows that the definition is compatible with constant functions and scalings.
Examples 2.6:
(1) Claim: C ([, ]) L2 ([, ]).
Reason.
Z
| f ( t)|2 dt 2( max | f ( t))|)2 < .
t[,]

The converse is not true. For example, f : [, ] R,



0, t [, 0),
f ( t) =
1, t [0, ],

is not continuous, but f L2 ([, ]).


(2) Let f : [, ] R,
| t| 41 , t 6= 0,
f ( t) =
0, t = 0.
Then


1 1 p
Z Z Z
2
p
| f ( t)| dt = p dt = 2 p dt = 2 2 | t| = 4 < .
| t| 0 | t| 0

Thus f L2 ([, )) and


p 12
4 p 1
k f kL2 ([,]) = = 2 4 .
2
CHAPTER 2. FOURIER SERIES AND PDES 10

(3) Let f : [, ] R,
p1 , t 6= 0,
| t|
f ( x) =
0, t = 0.

Then 1
Z Z
| f ( t)|2 dt = dt = .
t|
|
Thus f L ([, ]). Observe, that f L1 ([, ]) so that, in general,
2

L1 ([, ]) is not contained in L2 ([, ]).


T HE MORAL : Both functions in (2) and (3) have a singularity at
t = 0. Whether the function belongs to L2 ([, ]) depends on how fast the
function blows up near t = 0. The latter function blows up too fast, since
Z
1
dt = log as a 0+ .
a t a

Next we consider vector space properties. Indeed, L2 ([, ]) is a complex


vector space with the natural addition and multiplication operations

( f + g)( t) = f ( t) + g( t) and (a f )( t) = a f ( t), a C.

Note that vectors (or elements) in L2 ([, ]) are functions. We define an inner
product in L2 ([, ]) by

1
Z
f , g = f ( t) g( t) dt
2
Z Z
1

= Re( f ( t) g( t)) dt + i Im( f ( t) g( t)) dt .
2

Here z = x i y C is the complex conjugate of z = x + i y C, where x, y R and i is


the imaginary unit.

T HE MORAL : An inner product gives a notion of an angle between vectors


and orthogonality.

This inner product is compatible with the L2 norm, since

1 1
1 1
Z Z
2 2 1
k f kL2 ([,]) = | f ( t)|2 dt = f ( t) f ( t) dt = f , f 2 .
2 2

Here we used the fact that zz = | z|2 , z C.

THE MORAL : A norm is a length of a vector.

Remark 2.7. The inner product in L2 ([, ]) satisfies the following properties:
1 1
Z Z
(1) f , f = f ( t) f ( t) dt = | f ( t)|2 dt 0.
2 2
(2) f , f = 0 if and only if f = 0 in L2 ([, ]), that is, k f kL2 ([,]) = 0.
CHAPTER 2. FOURIER SERIES AND PDES 11

1 1 1
Z Z Z
(3) f , g = f ( t) g( t) dt = f ( t) g( t) dt = f ( t) g( t) dt = g, f .
2 2 2
(4) a f , g = a f , g, a C,
(5) f + g, h = f , h + g, h.

These properties can be taken as the definition of an abstract inner product x, y,


x, y H on a complex vector space H . Let k k the norm induced by an inner
product of H , that is,
1
k xk = x, x 2 , x H.

T HE MORAL : There are many ways to define inner products depending on


the applications. We shall focus on the standard inner product in L2 ([, ]), but
several results hold true for other inner products as well.
Examples 2.8:
Pn
(1) x, y = j =1 x j y j , x = ( x1 , . . . , xn ), y = ( y1 , . . . , yn ) is an inner product in the
real vector space Rn . Moreover
v
u n
1
uX 2
k xk = x, x = t 2 xj
j =1

is the Euclidean norm in Rn . Recall that x, y = | x|| y| cos , where is the


angle between x and y.
(2) z, w = nj=1 z j w j , z = ( z1 , . . . , z n ), w = (w1 , . . . , wn ) is an inner product
P

in the complex vector space Cn . Here w is the complex conjugate of w.


Moreover v v
u n u n
1
uX uX
k zk = z, z = 2
t z j z j = t | z j |2
j =1 j =1

is a norm in Cn .

If H is a space with inner product, x, y H and x, y = 0, we say that x and y


are orthogonal. Observe that this definition is symmetric: If x, y are orthogonal
then y, x are orthogonal.

Example 2.9. Let e j : [, ] C, e j ( t) = e i jt = cos( jt) + i sin( jt), j Z (Eulers


formula). Then e j C ([, ]) and consequently e j L2 ([, ]) with
1

2 1
1 1
Z Z
2
k e j kL2 ([,]) = | e i jt |2 dt = 1 dt = 1, j = 1, 2, . . . .
2 | {z } 2
=1

The inner product of two such functions is


1 i jt ikt 1 i jt ikt
Z Z
e j , e k = e e dt = e e dt
2 2

1 i ( j k) t 1 e i ( j k) t
Z

= e dt = =0
2 2 i ( j k)
CHAPTER 2. FOURIER SERIES AND PDES 12

Figure 2.2: Polar coordinates.

provided j 6= k. On the other hand if j = k we have e i jt e i jt = | e0 |2 = 1 so that


e j , e j = 1. This shows that the set { e j } jZ is an orthonormal set in L2 ([, ])
and we summarize this as

0, j 6= k,
e j , e k =
1, j = k.

Sometimes this is denoted as e j , e k = jk , where ik is Kroneckers delta.


We shall prove the following Cauchy-Schwarz inequality with the general
properties of an inner product.

Lemma 2.10 (Cauchy-Schwartz inequality). Let H be an inner product space.


For every x, y H , we have
1 1
| x, y| x, x 2 y, y 2 .

Proof. Denote by k k the norm defined by the inner product of H , that is, k xk =
1
x, x 2 , x H . If y = 0 it is clear that the Cauchy-Schwarz holds with equality. So
let us assume that y 6= 0. We set

x, y
z = x y.
y, y

Then
x, y

z, y = x, y y, y = 0.
y, y
CHAPTER 2. FOURIER SERIES AND PDES 13

x,y
Thus vectors z and y are orthogonal. Observe, that y,y y is the projection of x to
y. Since
x, y
x= y+ z
y, y
we can use the Pythagorean theorem to write

x, y2 x, y2 x, y2
k xk2 = 2
k y k2 + k z k2 = 2
+ k z k2 .
y, y k yk k y k2

This proves the claim.

Remarks 2.11:
(1) The Cauchy-Schwarz inequality in L2 ([, ]) reads

| f , g| k f kL2 ([,]) k gkL2 ([,]) .

This implies
Z
Z 1 Z 1
2 2
| f ( t)|2 dt | g( t)|2 dt

f ( t) g( t) dt .

This is a very useful inequality for integrals, which is a special case of


Hlders inequality.
(2) By replacing f ( t) with | f ( t)| and choosing g( t) = 1, we may conclude that
L2 ([, ]) L1 ([, ]). We saw in Example 2.6 (3) that the converse
inclusion does not hold, in general.

Remark 2.12. The norm k kL2 ([,]) satisfies the following properties:

(1) k f kL2 ([,]) 0 for every f L2 ([, ]).


(2) k f kL2 ([,]) = 0 if and only if f = 0 in L2 ([, ]).
WA RNING : This does not imply that f ( t) = 0 for every t [, ]. In
fact, it implies that f ( t) = 0 for almost every t [, ] with respect to the
one-dimensional (Lebesgue) measure.
A G R E E M E N T : f = g in L2 ([, ]) if and only if

1
1
Z
2
k f gkL2 ([,]) = | f ( t) g( t)|2 dt = 0.
2

(3) ka f kL2 ([,]) = |a|k f kL2 ([,]) for every a C and f L2 ([, ]).
(4) The triangle inequality

k f + gkL2 ([,]) k f kL2 ([,]) + k gkL2 ([,])

holds for every f , g L2 ([, ]), see Remark 2.14 below.

These properties can be taken as the definition of an abstract norm k k in a vector


space H .
CHAPTER 2. FOURIER SERIES AND PDES 14

1
Lemma 2.13. If H is a space with inner product then k xk = x, x 2 , x H , is a
norm in H .

Proof. All other properties of a norm are easily verified except maybe for the
triangle inequality. To prove this, we observe that

k x + yk2 = x + y, x + y = x, x + y, x + x, y + y, y
= x, x + y, y + x, y + x, y
= k xk2 + k yk2 + 2 Re x, y.

Now the Cauchy-Schwarz inequality implies

2| Re x, y| 2| x, y| 2k xkk yk,

from which we conclude that

k x + yk2 k xk2 + k yk2 + 2k xkk yk = (k xk + k yk)2 .

Remark 2.14. The triangle inequality in L2 ([, ]) reads

k f + gkL2 ([,]) k f kL2 ([,]) + k gkL2 ([,]) .

This implies
Z 1 Z 1 Z 1
2 2 2
| f ( t) + g( t)|2 dt | f ( t)|2 dt + | g( t)|2 dt .

2.3 The Fourier series*


We begin with the definition of the Fourier series.

Definition 2.15 (Fourier series). Let f L1 ([, ]). The nth partial sum of a
Fourier series is
n
fb( j ) e i jt ,
X
S n f ( t) = n = 0, 1, 2, . . . ,
j = n

where
1
Z
fb( j ) = f , e j = f ( t) e i jt dt, j Z,
2

is the j th Fourier coefficient of f . Here e j : R C, e j ( t) = e i jt , j Z. The Fourier


series of f is the limit of the partial sums S n as n , provided the limit exists
in some reasonable sense. In this case we may write
n
fb( j ) e i jt = fb( j ) e i jt .
X X
f ( t) = lim S n f ( t) = lim
n n
j = n j =
CHAPTER 2. FOURIER SERIES AND PDES 15

T HE MORAL : This is a series approximation of a function. The point is that


this approximation also applies to functions which do not have to be smooth as
in the case of Taylor series, for example. At least in the definition, it is enough
to assume that f L1 ([, ]). As we shall see, f L2 ([, ]) space is needed to
understand the Fourier coefficients and the convergence of the Fourier series.

Remarks 2.16:
(1) In the convergence of the Fourier series we always consider symmetric
partial sums, where the indices run from n to n. Note that, in the
classical sense, convergence of the series j = z j , z j C, means that
P


X 1
X
X
zj = zj + z j,
j = j = j =0

where both series on the right-hand-side converge.


W A R N I N G : In general

X n
X
z j 6= lim z j.
n
j = j = n

Reason.
n
X
j = n + + (1) + 0 + 1 + + n = 0
j = n
P
for every n = 1, 2, . . . , but the series j = j does not converge in the
classical sense. This shows that the symmetric partial sums may converge
for a divergent series. Equality holds for absolutely convergent series.

(2) By the Cauchy-Schwarz inequality

| fb( j )| = | f , e j | k f kL2 ([,]) k e j kL2 ([,]) = k f kL2 ([,]) < , j Z.


| {z }
=1

This means that the Fourier coefficients are well defined also if f
L2 ([, ]).
(3) Since e j , j Z, is 2-periodic, the partial sum S n f ( t), n = 0, 1, 2, . . . , of
a Fourier series is 2-periodic. Consequently the limit limn S n f ( t),
whenever it exists. is 2-periodic.
T HE MORAL : We can only approximate 2-periodic functions by a
Fourier series.

Example 2.17. Let f : [, ] R, f ( t) = t. Then f L1 ([, ]). Indeed,

0
0 2 2
t t
Z Z Z
| f ( t)| dt = t dt + t dt = + = 2 < .
0 2 0 2
CHAPTER 2. FOURIER SERIES AND PDES 16

The Fourier coefficients f ( j ), j 6= 0, can be calculated by integration by parts as



1 i jt 1 te i jt 1 e i jt
Z Z
fb( j ) = te dt = dt
2 2 i j 2 i j
1 e i j e i j 1 e i jt cos(( j + 1))
Z
= dt = .
2 i j i j 2 i j ij
| {z }
=0

On the other hand,


1
Z
fb(0) = e0 dt = 0.
t |{z}
2
=1
Thus
0, j = 0,
fb( j ) =
cos(( j+1)) , j 6= 0
ij
and
n cos(( j + 1)) cos(( j + 1)) i jt
n cos(( j + 1))
i jt
X X
S n f ( t) = e e =2 sin( jt).
j =1 ij ij j =1 j

Figure 2.3: The Fourier series approximation of the saw tooth function.

Remark 2.18. We make two observations related to the previous example.

(1) The function f , extended as 2-periodic function to R, is not continuous at


the points t = k, k Z. At a point of discontinuity, for example at t = 0,
we have
1

S n f (0) = 0 = lim f ( t) + lim f ( t) .
2 t0 t0+
CHAPTER 2. FOURIER SERIES AND PDES 17

The sum of the Fourier series at a point of a jump discontinuity is the


average of the limits from the both sides. This ia a general property of the
Fourier series. Moreover, there is Gibbs phenomenon

lim ( max S n ( t)) 1, 179.


n t[,])

This means that the absolute error made in the approximation is about
18% independent of the degree of the approximation. In particular, the
error does not go to zero as n . This is an unexpected phenomenon.
2
(2) We have | fb( j )| j for j 6= 0 while fb(0) = 0. It follows that fb( j ) 0 as
| j | . Actually, this kind of decay general property of the Fourier
coefficients of any L1 ([, ]) or L2 ([, ]) function.

Example 2.19. Let f : [, ] R,



1, t [, 0),
f ( t) =
1, t [0, ].

It is clear that the function f L1 ([, ]) so that we can calculate the Fourier
coefficients fb( j ), j 6= 0, as

e i jt e i jt
Z 0 Z 0 !
1 1

i jt i jt
f ( j) =
b e dt + e dt = +
2 0 2 i j 0 i j
1 0 1
= e e i j ( e i j e0 ) = (1 cos( j ) cos( j ) + 1)
2 i j 2 i j

1 i 0, j even,
= (1 cos( j )) = (cos( j ) 1) =
i j j 2 i , j odd.
j

For j = 0 we have Z 0 Z
1

fb(0) = 1 dt + 1 dt = 0.
2 0

Note that at the points of jump discontinuity we have

1

S n f (0) = 0 = lim f ( t) + lim f ( t) .
2 t 0 t0+

We collect some easy properties of the Fourier coefficients in the following


proposition.

Lemma 2.20. Let f , g L1 ([, ]) and a, b C.

(1) (Linearity) a f + b g( j ) = a fb( j ) + b gb( j ), j Z.


1
Z
(2) (Boundedness) | fb( j )| | f ( t)| dt = k f kL1 ([,]) , j Z.
Z
2
1
(3) fb(0) = f ( t) dt.
2
CHAPTER 2. FOURIER SERIES AND PDES 18

Figure 2.4: The Fourier series approximation of the sign function.

(4) (Conjugation) f ( j ) = fb( j ), where f is the complex conjugate of f .


b

f ( t)( j ) = fb( j ), j Z.
(5) (Reflection)
( t + s)( j ) = e i js fb( j ), j Z, for a fixed s.
(6) (Shift) f
(7) (Modulation) e ikt f ( t)( j ) = fb( j k), j Z, for a fixed k Z.

Proof. (1) Property ( i ) is an immediate consequence of linearity of the integral.


1 1 1
Z Z Z
i jt
| f ( t) e i jt | dt =

(2) | f ( j )| =
b f ( t) e dt | f ( t)| dt.
2Z
2Z 2
1 1
(3) fb(0) = f ( t) |e{zi0}t dt = f ( t) dt.
2 2
=1
(6) A change of variables u = t + s gives

1
Z 1
Z + s
f
( t + s)( j ) = f ( t + s) e i jt dt = f ( u) e i j(us) du.
2 2 + s

Now using Lemma 2.2, and the fact that the function f ( u) e i j(us) is 2-periodic,
we have
1
Z + s 1
Z
i j ( u s) i js
f ( u) e du = e f ( u) e i ju du = e i js fb( j ).
2 + s 2

Other claims are left as exercises.


CHAPTER 2. FOURIER SERIES AND PDES 19

2.4 The best square approximation*


It is very instructive to consider Fourier series in terms of projections.

Lemma 2.21. The projection of the vector f L2 ([, ]) to a subspace spanned


by { e j }nj=n is

n n
fb( j ) e i jt ,
X X
S n f ( t) = f , e j e j = n = 0, 1, 2, . . . ,
j = n j = n

where
1
Z
fb( j ) = f , e j = f ( t) e i jt dt.
2

T HE MORAL : Let x = ( x1 , . . . , xn ) Rn . The projection of x to the subspace


spanned by the first k standard basis vectors e j , j = 1, . . . , k, is kj=1 x j , e j e j . The
P

previous lemma tells that the same holds true in L2 ([, ]).

Proof.
n
X
f Sn f , e k = f fb( j ) e j , e k
j = n
n
X
= f , ek fb( j ) e j , e k
j = n

= fb( k) fb( k) = 0, k = n, . . . , n,

implies
n
X n
X
f Sn f , a j e j = a j f S n f , e j = 0, (2.4)
j = n j = n

for every a j C, j = n, . . . , n. Since any vector belonging to the subspace spanned


by { e j }nj=n is a linear combination nj=n a j e j , this means that f S n f is orthog-
P

onal to the subspace spanned by { e j }nj=n .

In particular, this implies that f = S n f + ( f S n f ), where S n f and f S n f are


orthogonal. From this we have

k f k2L2 ([,]) = k f S n f + S n f k2L2 ([,])

= ( f S n f ) + S n f , ( f S n f ) + S n f
= f S n f , f S n f + f S n f , S n f + S n f , f S n f +S n f , S n f
| {z } | {z }
=0 =0

= kf S n f k2L2 ([,]) + kS n f k2L2 ([,]) .

This is the Pythagorean theorem in L2 ([, ]).


CHAPTER 2. FOURIER SERIES AND PDES 20

Figure 2.5: The least square approximation.

Since { e j } jZ is an orthonormal set in L2 ([, ]), we obtain


2
X n n
X n
X
fb( j ) e j = fb( j ) e j , fb( k) e k


j=n 2 j = n k= n
L ([,])
n n n
| fb( j )|2 .
X X X
= fb( j ) fb( k) e j , e k =
j = n k= n j = n

It follows that

k f k2L2 ([,]) = k f S n f k2L2 ([,]) + kS n f k2L2 ([,])


n (2.5)
= k f S n f k2L2 ([,]) + | fb( j )|2 .
X
j = n

THE MORAL : Note that f = S n f + ( f S n f ), where S n f is the Fourier series


approximation of f and f S n f is the error made in the approximation. The
partial sums S n f approximate f with the mean square error k f S n f kL2 ([,]) .

TERMINOLOGY:
n n n n
a j e i jt = a j ( e it ) j = a j z j,
X X X X
ajej =
j = n j = n j = n j = n

where z = e it , a j C, is called a trigonometric polynomial of degree n.


CHAPTER 2. FOURIER SERIES AND PDES 21

Example 2.22. Trigonometric polynomials are different for the standard polyno-
mials. For example,

e i jt + e i jt e i jt e i jt
cos( jt) = and sin( jt) = , j Z,
2 2i
are trigonometric polynomials.

Theorem 2.23 (Theorem of best square approximation). If f L2 ([, ]),


then
Xn
k f S n f kL2 ([,]) f a j e j

j = n

L2 ([,])

for every a j C, j = n, . . . , n.

T HE MORAL : The partial sum S n f of a Fourier series gives the best L2 -


approximation for the function f L2 ([, ]) among all trigonometric polynomials
of degree n.

Proof. Clearly
!
n
X n
X n
X
f ajej = f f ( j) e j +
b ( fb( j ) a j ) e j ,
j = n j = n j = n

where
n
X n
X
f fb( j ) e j , ( fb( j ) a j ) e j = 0,
j = n j = n
Pn
since j = n ( f ( j ) a j ) e j
b belongs to the subspace spanned by { e j }nj=n , see (2.4).
The Pythagorean theorem implies
2 2 2
Xn Xn X n
f a j e j = f fb( j ) e j + ( fb( j ) a j ) e j

j = n j = n
2 j=n 2
2 L ([,]) L ([,]) L ([,])
| {z }
0
2
Xn
f f ( j) e j .
b
j = n
2
L ([,])

Remark 2.24. Equality occurs in the previous theorem if and only if we have
equalities throughout in the proof of the theorem. This implies that the equality
occurs if and only if
X n
a j e j Sn f = 0,


j=n
L2 ([,])
Pn
that is, S n f = j = n a j e j in L2 ([, ]).

Let f L2 ([, ]). By (2.5) we have


n n
k f k2L2 ([,]) = k f S n f k2L2 ([,]) + | fb( j )|2 | fb( j )|2 ,
X X
n = 1, 2, . . . .
| {z } j=n j = n
0
CHAPTER 2. FOURIER SERIES AND PDES 22

It follows that
n
| fb( j )|2 = lim | fb( j )|2 k f k2L2 ([,]) .
X X
(2.6)
n
j = j = n

This is called Bessels inequality. Equality occurs in (2.6) if and only if

lim k f S n f k2L2 ([,]) = 0,


n

in which case we have Parsevals identity



k f k2L2 ([,]) = | fb( j )|2 .
X
(2.7)
j =

THE MORAL : Parsevals identity is the Pythagorean theorem with infinitely


many coefficients in the sense that the Fourier coefficients give the coordinates of
a function in L2 ([, ]).

Parsevals identity is equivalent with the convergence of the partial sums of


the Fourier series in the L2 -sense, which is the content of the following result.

Theorem 2.25. Let f L2 ([, ]). Then

lim k f S n f kL2 ([,]) = 0.


n

T HE MORAL : The partial sums S n f approximate f L2 ([, ]) so that the


mean square error k f S n f kL2 ([,]) goes to zero. This means that the Fourier
series always converges in L2 ([, ]). In other words, every function in L2 ([, ])
can be represented as a Fourier series.

WA RNING : The partial sums of the Fourier series of a L2 -function are only
claimed to converge in the L2 -norm. This mode of convergence is rather weak. In
particular, it does not follow in general that f ( t) = limn S n f ( t) pointwise for
every t [, ], see Example 2.17.

Proof. Let f L2 ([, ]) and > 0. By density of the trigonometric polynomials


in L2 ([, ]), there exists a trigonometric polynomial g of some degree m such
that k f gkL2 ([,]) < . We do not give the proof of this density result in this
course. Combining this with the best approximation Theorem 2.23, for n m we
have
k f S n f kL2 ([,]) k f gkL2 ([,]) < .

Here we use the fact that since g is a trigonometric polynomial of degree m and
n m, we may consider g as a trigonometric polynomial of order n with the
interpretation that some of the coefficients are zero which proves the claim.
Remarks 2.26:
(1) Theorem 2.25 implies that { e j }
j =
is an orthonormal basis for the space
L2 ([, ]) in the sense that

X n
lim fb( j ) e j f =0

n
j = n

L2 ([,])
CHAPTER 2. FOURIER SERIES AND PDES 23

for every f L2 ([, ]). This means that


n
X
X
f = lim fb( j ) e j = fb( j ) e j
n
j = n j =

in L2 ([, ]). In this sense every function in L2 ([, ]) can be represented


as a Fourier series. Since there are infinitely many vectors in the basis,
the space L2 ([, ]) is infinite dimensional.

T HE MORAL : The Fourier coefficients fb( j ), j Z, are the coordi-


nates of the function f L2 ([, ]) with respect to the orthonormal basis
{ e j } in a similar way as x = ni=1 x, e i e j = nj=1 x i e j = ( x1 , . . . , xn ) is
P P
j =
the coordinate representation of x Rn with respect to the standard basis
{ e j }nj=1 .

(2) Claim: If f L2 ([, ]), then fb( j ) 0 as | j | .

Reason. By Parsevals identity



| fb( j )|2 = k f k2L2 ([,]) < .
X
j =

This implies that the series above converges. Thus | fb( j )|2 0 and fb( j ) 0
as | j | .

1
This claim holds also for f L ([, ]), but this is out of the scope of this
course.

Parsevals identity implies a uniqueness result for the Fourier series.

Corollary 2.27 (Uniqueness). Let f , g L2 ([, ]) such that fb( j ) = gb( j ) for all
j Z. Then f = g in L2 ([, ]).

T HE MORAL : All Fourier coefficients of two functions coincide if and only if


the functions are same. Hence a function is uniquely determined by its Fourier
coefficients.

Proof. By Parsevals identity (2.7) we have



k f gk2L2 ([,]) = f g)( j )|2 = b( j )|2 = 0.
X X
|( | fb( j ) g
j = j =

This implies that f = g in L2 ([, ]).


Remarks 2.28:
(1) fb( j ) = 0 for every j Z if and only if f = 0 in L2 ([, ]).
(2) Since the definition of the Fourier series required integration, for example,
two functions which differ at finitely many points have the same Fourier
series. This shows that the equality does not hold pointwise without
additional assumptions.
(3) If f , g C ([, ]), then we can conclude that f ( x) = g( x) for every x
[, ].
CHAPTER 2. FOURIER SERIES AND PDES 24

2.5 The Fourier series on a general inter-


val*
Theory for Fourier series can be developed for functions defined on other intervals
than [, ] in the same way as above. Let f : [a, b] C be a function and assume
that f L2 ([a, b]). Then the nth partial sum of a Fourier series of f on [a, b] is
n
X n
X 2 i jt
S n f ( t) = f , e j e j = fb( j ) e ba , n = 0, 1, 2, . . . ,
j = n j = n

where the Fourier coefficients are

1
Z b 2 i jt
fb( j ) = f ( t) e ba dt, j Z.
ba a

This follows from a change of variables. For example, if f is defined on [, ],


then g( x) = f (2 x ) defined on [0, 1] and a change of variables shows that j th
Fourier coefficient of f equals j th Fourier coefficient of g.

THE MORAL : The interval [, ] does not play any special role in the Fourier
theory, but we shall mainly consider this case.

2.6 The real form of the Fourier series*


Now we describe a different way of writing the Fourier series of a function.

Theorem 2.29. Let f L1 ([, )). The nth partial sum of a Fourier series can
be written as
n
a0 X
S n f ( t) = + (a j cos( jt) + b j sin( jt)),
2 j=1

where
1
Z
aj = f ( t) cos( jt) dt, j = 0, 1, 2, . . .

and
1
Z
bj = f ( t) sin( jt) dt, j = 1, 2, . . . .

This is called the real form of the Fourier series of f . The coefficients a j are called
the Fourier cosine coefficients of f and b j are called the Fourier sine coefficients
of f . The corresponding series are called the Fourier cosine and sine series of f
correspondingly.
Conversely, any trigonometric series of the form
n
a0 X
+ (a j cos( jt) + b j sin( jt)) (2.8)
2 j=1
CHAPTER 2. FOURIER SERIES AND PDES 25

can be written in the complex form


n
c j e i jt ,
X
(2.9)
j = n

where
a ib
j 2 j , j = 1, 2, . . . , n



c j = a20 , j = 0,

a j + ib j , j = 1, 2, . . . , n.


2

T HE MORAL : If f is real valued, then the Fourier cosine and sine series
consist of real numbers. However, the use of the complex form is preferable since
it contains the same information as the cosine and sine coefficients in one complex
Fourier coefficient. Moreover, properties of the Fourier coefficients take an elegant
and easy-to-remember form in the complex notation. The real form is only for
those who are afraid of complex numbers.

Proof. Since e i jt = cos( jt) + i sin( jt) we have


X X
S n f ( t) = fb(0) + fb( j ) cos( jt) + i fb( j ) sin( jt).
j 6=0 j 6=0

For the first sum we have


X n
X n
X
fb( j ) cos( jt) = fb( j ) cos( jt) + fb( j ) cos( jt)
j 6=0 j =1 j =1
n
X n
X
= fb( j ) cos( jt) + fb( j ) cos( jt)
j =1 j =1
n
X
= ( fb( j ) + fb( j )) cos( jt)
j =1

and for the second sum


X n
X n
X
i fb( j ) sin( nt) = i fb( j ) sin( jt) + i fb( j ) sin( jt)
j 6=0 j =1 j =1
n
X n
X
=i fb( j ) sin( jt) + i fb( j ) sin( jt)
j =1 j =1
n
X
= i ( fb( j ) fb( j )) sin( jt).
j =1

By the identities

e it + e it e it e it
= cos t and = sin t, (2.10)
2 2i
CHAPTER 2. FOURIER SERIES AND PDES 26

we have
1
Z
a 0 = 2 fb(0) = f ( t) dt,

1
Z
a j = f ( j ) + f ( j ) =
b b f ( t)( e i jt + e i jt ) dt
2
1 1
Z Z
= f ( t)2 cos( jt) dt = f ( t) cos( jt) dt,
2
Z
i
b j = i ( fb( j ) fb( j )) = f ( t)( e i jt e i jt ) dt
2
Z
i 1
Z
= f ( t)(2 i ) sin( jt) dt = f ( t) sin( jt) dt.
2
Now consider the real trigonometric series as in (2.8). Using (2.10) again we
have
n
a0 X
+ (a j cos( jt) + b j sin( jt))
2 j=1
a0 Xn e i jt + e i jt X n e i jt e i jt
= + aj + bj
2 j=1 2 j =1 2i
a0 Xn 1 n 1
(a j ib j ) e i jt + (a j + ib j ) e i jt .
X
= +
2 j=1 2 j =1 2

Remarks 2.30:
(1) A function f is even, if f ( t) = f ( t) for every t and odd if f ( t) = f ( t) for
every t. For a 2-periodic odd function all Fourier cosine coefficents a j ,
j = 1, 2, . . . , in (2.8) are zero. Similarly, for a 2-periodic even function all
Fourier sine coefficents b j , j = 1, 2, . . . , in (2.8) are zero.
(2) In applications we are interested in representing by a Fourier series a
function f ( t) defined on the bounded interval 0 < t < L. There are several
ways to extend f as a periodic function to R. The even periodic extension
of f is defined by
f ( t), 0 < t < L,
f ( t) =
f ( t), L < t < 0,
and f ( t) = f ( t + 2L) otherwise. Similarly, the odd periodic extension of f is
defined by
f ( t ), 0 < t < L,
f ( t) =
f ( t), L < t < 0,
and f ( t) = f ( t + 2L) otherwise. We do not worry about the definition of ex-
tensions at the points 0, L, 2L, . . . , since that does not affect the Fourier
coefficients. The cosine coefficients of the odd periodic extension are zero
and the sine coefficients of the even periodic extension are zero. Thus we
obtain two Fourier expansions
a0 X
f ( t) = + a j cos( jt),
2 j=1
CHAPTER 2. FOURIER SERIES AND PDES 27

with
2 L jt
Z
aj = f ( t) cos dt, j = 0, 1, 2, . . .
L 0 L
and
X
f ( t) = b j sin( jt),
j =1

with
2 L jt
Z
bj = f ( t) sin dt, j = 1, 2, . . . .
L 0 L
Note carefully, that both cosine and sine series represent the same function.
Thus a function can be represented only by its sine or cosine series.

2.7 The Fourier series and differentiation*


The smoothness of the functions affects to the decay of the Fourier coefficients. In
general, the smoother the function, the faster the decay of the Fourier coefficients.

Lemma 2.31. Assume that f C 1 (R) is a 2-periodic function. Then

fb0 ( j ) = i j fb( j ), j Z.

THE MORAL : The Fourier series can be differentiated termwise and differen-
tiation becomes multiplication on the Fourier side,

fb( j ) e i jt = f 0 ( t) = i j fb( j ) e i jt .
X X
f ( t) =
j = j =

This is a very useful property in the PDE theory.

Proof. An integration by parts gives


1 0
Z
0
f ( j) =
b f ( t) e i jt dt
2
Z
1

i jt i jt
= f ( t ) e f ( t )( i je ) dt
2

Z
1 i j
= f () e f () e i j + i j f ( t) e i jt dt

2 | {z }
= f ( )
f ( ) i j ij Z
= e e i j + f ( t) e i jt dt
2 | {z } 2
=0

= i j fb( j ).

Remark 2.32. The lemma above implies

fb0 ( j )
fb( j ) = , j 6= 0,
ij
CHAPTER 2. FOURIER SERIES AND PDES 28

and consequently

| fb0 ( j )| 1 1
Z 1
| fb( j )| = | f 0 ( t)| dt max | f 0 ( t)| 0
| j| | j | 2 | j | t[,]
| {z }
<

1
as | j | . This means that fb( j ) 0 with speed | j| as | j | . We can iterate this
procedure if f has higher order derivatives and obtain a faster decay.

THE MORAL: By Remark 2.26 (2), we see that for every function f L2 ([, ])
(or f L1 ([, ])) we have fb( j ) 0 as | j | . In other words, the Fourier
coefficients of an L2 function always converge to zero, but there is no estimate for
the speed of convergence. If the function is smoother, then the Fourier coefficients
converge to zero faster with an estimate for the speed.

2.8 The Dirichlet kernel*


Next we introduce the Dirichlet kernel, which turns out to be a very important
object related to the convergence of Fourier series. Let f : R C be a continuous
2-periodic function. The partial sums of the Fourier series of f can be written as
n n 1 Z
fb( j ) e i jt = f ( s) e i js ds e i jt
X X
S n f ( t) =
j = n j = n 2
!
1 X n Z 1
Z n
i j ( t s) i j ( t s)
X
= f ( s) e ds = f ( s) e ds.
2 j=n 2 j = n

Now we define the Dirichlet kernel as


n
e i jt ,
X
D n ( t) = n = 0, 1, 2, . . . .
j = n

With this definition we have the formula


1
Z
S n f ( t) = f ( s)D n ( t s) ds, t [, ], n = 0, 1, 2, . . . . (2.11)
2
T HE MORAL : This is an integral representation of the partial sum of the
Fourier series. In fact, (2.11) is a convolution of f with the Dirichlet kernel as we
shall see soon.

Before discussing further this formula let us take a closer look at the Dirichlet
kernel.
Lemma 2.33.
sin n + 21 t

n
i jt
X
D n f ( t) = e = , t 6= 0, n = 0, 1, 2, . . . .
j = n sin 12 t

Furthermore, the Dirichlet kernel is a 2-periodic function that satisfies

D n (0) = 2 n + 1 and D n () = (1)n , n = 0, 1, 2, . . . .


CHAPTER 2. FOURIER SERIES AND PDES 29

THE MORAL : The Dirichlet kernel (2.11) can be computed explicitely.

Proof. Recall the following formula for a geometric series.


Claim: If a C \ {1} and n = 0, 1, 2, . . . , then
n an a n+1
aj =
X
S= .
j = n 1a

Reason. Clearly

S = an + an+1 + + a1 + 1 + a + + a n1 + a n
= aS = an+1 an+2 1 a a2 a n a n+1 .

By summing these up termwise we obtain


an a n+1
(1 a)S = an a n+1 S =
1a
as required.

Setting a = e it in the previous formula we get for t 6= 0 that


t
e int e i(n+1) t e i 2 e int e i(n+1) t

n
it j
X
D n ( t) = (e ) = = t
j = n 1 e it e i 2 (1 e it )
1 1 1
sin ( n + 21 ) t

e i(n+ 2 ) t e i(n+ 2 ) t 2 i sin ( n + 2 ) t
= = = .
2 i sin 2t sin 2t
t t

e i 2 e i 2
In the equality before the last we have used the latter one of the trigonometric
identities in (2.10).
For t = 0 the value D n (0) can be calculated directly since
n
X
D n (0) = 1 = 2 n + 1.
j = n

This can be recovered by the general formula as well by taking the limit as t 0.
On the other hand, for t = we have
sin n + 12


D n () =
= sin n + = cos( n) = (1)n
sin 2 2
as claimed.

The calculation of the integral of the Dirichlet kernel will turn out to be
important in some of the applications we shall discuss.
Lemma 2.34.
1
Z
D n ( t) dt = 1, n = 0, 1, 2, . . . .
2
THE MORAL : The total mass of the Dirichlet kernel is one.

Proof. A direct calculation gives


1 n 1 i jt 1 i0 t
Z X Z Z
D n ( t) dt = e dt = e dt = 1.
2 j = n 2 2
CHAPTER 2. FOURIER SERIES AND PDES 30

20

17.5

15

12.5

10

7.5

2.5

-3.2 -2.4 -1.6 -0.8 0 0.8 1.6 2.4 3.2

-2.5

-5

Figure 2.6: The graph of the Dirichlet kernel for N = 10.

2.9 Convolutions*
The notion of convolution plays a fundamental role in Fourier analysis and PDEs.
Let f , g C (R) be 2-periodic. The convolution of f and g on [, ] is the function
f g defined as
1
Z 1
Z
( f g)( t) = f ( t s) g( s) ds = f ( s) g( t s) ds, t [, ].
2 2

The second equality follows by a change of variables and the fact that f and g are
2-periodic. Indeed, setting u = t s in the defining integral we have
Z Z t Z
f ( t s) g( s) ds = f ( u) g( t u) du = f ( u) g( t u) du
t+

by Lemma 2.2. To apply this lemma we used the fact that for a fixed t the function
F ( u) = f ( u) g( t u) is 2-periodic, if f , g are 2-periodic. Indeed

F ( u + 2) = f ( u + 2) g(( t u) 2) = f ( u) g( t u) = F ( u).
CHAPTER 2. FOURIER SERIES AND PDES 31

Furthermore, the function f g is itself 2-periodic, provided that f , g are 2-


periodic, since

1
Z
( f g)( t + 2) = f (( t + 2) s) g( s) ds
2
Z
1
= f (( t s) + 2) g( s) ds = ( f g)( t).
2 | {z }
= f ( t s)

1
R
If g = 1, then f g is constant and equal to 2 f ( s) ds. Observe, that this is the
integral average of f over [, ].

T HE MORAL : Convolutions can be considered as weighted averages of the


functions.

Next we shall show that the convolution of f and g is well defined also if
f , g L1 ([, ]) or L2 ([, ]). If f , g L1 ([, ]), then

1 1
Z Z

k f gkL1 ([,]) = f ( t s) g( s) ds dt
2 2

1 1
Z Z
| f ( t s)| dt | g( s)| ds
2 2
1 1
Z Z
= | f ( t)| dt | g( s)| ds
2 2
= k f kL1 ([,]) k gkL1 ([,]) < .

For f , g L2 ([, ]), by the Cauchy-Schwarz inequality, we have


1
Z

|( f g)( t)| =
f ( t s) g( s) ds
2
1
Z
| f ( t s)|| g( s)| ds
2
1 1
1 1
Z Z
2 2
2 2
| f ( t s)| ds | g( s)| ds
2 2
= k f kL2 ([,]) k gkL2 ([,]) < , t [, ].

Thus we obtain

sup |( f g)( t)| k f kL2 ([,]) k gkL2 ([,]) .


t[,]

We gather the previous observations and other properties of the convolution


in the following lemma.

Lemma 2.35. Let f , g, h : R C be 2-periodic functions.

(1) The function f g is 2-periodic.


(2) If f , g L1 ([, ]) then f g is well defined and

k f gkL1 ([,]) k f kL1 ([,]) k gkL1 ([,]) .


CHAPTER 2. FOURIER SERIES AND PDES 32

(3) If f , g L2 ([, ]) then f g is well defined and bounded with

sup |( f g)( t)| k f kL2 ([,]) k gkL2 ([,]) .


t[,]

(4) f g = g f .
(5) f ( g h) = ( f g) h.
(6) f ( g + h) = f g + f h.
(7) a( f g) = (a f ) g = f (ag), a C.
f g)( j ) = fb( j ) gb( j ), j Z.
(8) (

Proof. (8) :

1
Z
f g)( j ) =
( ( f g)( t) e i jt dt
2
1 1
Z Z
= f ( s) g( t s) ds e i jt dt
2 2
1 1
Z Z
= f ( s) e i js g( t s) e i j( ts) dt ds
2 2
1
Z
i js 1
Z
= f ( s) e g( t) e i jt dt ds = fb( j ) gb( j ).
2 2

In the third equality we changed the order of integration and on the fourth equality
we used the fact that the integrand is 2-periodic and Lemma 2.2.
Other claims are left as exercises.

Remark 2.36. The Dirichlet formula (2.11) can be written as a convolution


1
Z
S n f ( t) = f ( s)D n ( t s) ds = (D n f )( t), t [, ], n = 0, 1, 2, . . . .
2

Thus S n f is a certain average of f with the weight D n .

T HE MORAL : Many useful quantities in Fourier analysis and PDEs can be


written as convolutions.

2.10 A local result for the convergence


of Fourier series*
We now turn to the issue of pointwise convergence of Fourier series under addi-
tional assumptions on the function.

Theorem 2.37 (Local convergence of Fourier series). Let f C ([, ]) be


a 2-periodic function which is differentiable at some point t 0 [, ]. Then

lim S n f ( t 0 ) = f ( t 0 ).
n
CHAPTER 2. FOURIER SERIES AND PDES 33

T HE MORAL : The Fourier series of a smooth function converges pointwise


everywhere.

Proof. We define the function



f ( t0 t) f ( t0 ) , if t 6= 0 and | t| < ,
t
F ( t) =
f 0 ( t ), if t = 0.
0

Since f is differentiable at t 0 and continuous everywhere we conclude that F


C ([, ]) and thus F is bounded on [, ]. We have

S n f ( t 0 ) f ( t 0 ) = (D n f )( t 0 ) f ( t 0 )
1 1
Z Z
= D n ( t) f ( t 0 t) dt D n ( t) dt f ( t 0 )
2 2
| {z }
=1
1
Z
= ( f ( t 0 t) f ( t 0 ))D n ( t) dt
2
Z
1
= tF ( t)D n ( t) dt
2
by the definition of F . Remembering the formula for the Dirichlet kernel and
using standard trigonometric identities we have

sin n + 12 t

tD n ( t) = t t
sin 2
sin( nt) cos 2t cos( nt) sin 2t

=t +t
sin 2t sin 2t

sin( nt) cos 2t



=t + t cos( nt), t 6= 0.
sin 2t

Thus
1
Z t t

1
Z
S n f ( t0 ) f ( t0 ) = t sin( nt) cos F ( t) dt + F ( t) t cos( nt) dt.
2 sin 2 2 2

Now the function


t cos( t/2)
( t ) = F ( t)
sin( t/2)
is continuous on [, ] and thus in L2 ([, ]). Thus
1
Z
|


2 ( t ) sin( nt ) dt b ( n)| 0 as n

by the Riemann-Lebesgue lemma. Similarly for the second term we see that
the function ( t) = tF ( t) is continuous and thus L2 ([, ]). Again by the
Riemann-Lebesgue lemma we conclude
1
Z
( t) cos( nt) dt |

b( n)| 0 as n .
2

The last two estimates complete the proof.


CHAPTER 2. FOURIER SERIES AND PDES 34

Remarks 2.38:
(1) It is enough to assume that f is Lipschitz continuous, that is,

| f ( t) f ( s)| L| t s| for all t, s [, ].

In this case
sup |S n f ( t) f ( t)| 0 as n .
t[,]

This means that S n f f uniformly in [, ] as n .


(2) The theorem shows that convergence of the Fourier series at a given
point depends only on the behaviour of the function in an arbitrarily
small neighbourhood of the point. This is somewhat unexpected, since the
Fourier coefficients are defined as integrals over the whole interval [, ].
(3) If f is twice continuously differentiable, denoted f C 2 (R), then the Fourier
series converges absolutely and uniformly to f . This can be seen by Remark
2.32. There are even stronger results. It can be shown, for example, that
the Fourier series of f converges uniformly, assuming only that f C 1 (R),
that is, f is continuously differentiable.

Remark 2.39. We list here some further results related to the pointwise conver-
gence of the Fourier series.

(1) Kolmogorov showed in 1926 that there exists an integrable function f


L1 ([, ]) whose Fourier series diverges at every point.
(2) There is a continuous function on [, ] such that the Fourier series
diverges on a countable, dense set of points in [, ]. A dense subset of an
interval [, ] is a set which contains sequences that approximate every
point in [, ]. For example, the rational numbers in [, ] form a dense
subset of [, ].
(3) Furthermore, Carleson in 1966 proved the (very deep) theorem that the
Fourier series of every function in L2 ([, ]) converge pointwise to the
function almost everywhere. This holds, in particular, for continuous
functions on [, ].

2.11 The Laplace equation in the unit disc


We shall consider the Laplace equation, which models heat distribution when the
system has reached thermal equilibrium. The Laplace equation occurs also in
other branches of mathematical physics. For example, in considering electrostatic
fields, the electric potential in a dielectric medium containing no electric charges
satisfies the Laplace equation. Similarly, the potential of a particle in free space
acted only by the gravitational forces satisfies the same equation. Morover, the
CHAPTER 2. FOURIER SERIES AND PDES 35

real and imaginary parts of a complex analytic function are solutions to the
Laplace equation.
We begin with considering the two-dimensional case. Let
1
= {( x, y) R2 : ( x2 + y2 ) 2 < 1}

be the unit disc in R2 and assume that g C () is a continuous function on the


boundary
1
= {( x, y) R2 : ( x2 + y2 ) 2 = 1}.

The problem is to find u C 2 () C () such that


2 2
u( x, y) = u ( x, y) + u ( x, y) = 0,

( x, y) ,

x2 y2 (2.12)
u( x, y) = g( x, y), ( x, y) .

This is called the Dirichlet problem for the Laplace equation in the unit disc.

Figure 2.7: The Dirichlet problem in the unit disc.

T HE MORAL : For a given boundary function g we want to find a solution of


the Laplace equation u = 0 inside the domain such that it assumes boundary
values g on the boundary . Physically this means that the temperature on
the boundary of the disc is given by g and that the system has reached
thermal equilibrium. In this model the temperature distribution in is given by
a solution of the Dirichlet problem for the Laplace equation u = 0. This is a time
independent problem, where all physical constants are set to be one.
CHAPTER 2. FOURIER SERIES AND PDES 36

We will solve this Dirichlet problem using Fourier series together with a
technique called separation of variables. This means that we look for solutions of
the form
u( x, y) = A ( x)B( y),

where the dependencies on x and y are separated. The problem is now reduced to
finding the functions A ( x) and B( x). The directions of the coordinate axes play a
special role in this approach and this would work, for example, for the Dirichlet
problem in a rectangular domain.
For the the unit disc we switch to polar coordinates. More precisely, any point
in the plane can be uniquely determined by its distance from the origin r and the
angle that the line segment from the origin to the point forms with the x-axis,
that is,

( x, y) = ( r cos , r sin ), ( x, y) R2 , 0 < r < , < ,


y
where r 2 = x2 + y2 and tan = x . In polar coordinates, we have

= {( r, ) : 0 < r < 1, < } and = {(1, ) : < }.

Note that the unit disc becomes a rectangular set in polar coordinates and this is
compatible with the separation of variables technique.

Figure 2.8: Transition to polar coordinates.

The next goal os to express the Laplace equation in polar coordinates as well.
This is done in the following lemma.
CHAPTER 2. FOURIER SERIES AND PDES 37

Lemma 2.40. The two-dimensional Laplace operator in polar coordinates is

2 u 1 u 1 2 u
u = + + , 0 < r < , < .
r2 r r r 2 2
Proof. Remember that x = r cos and y = r sin . We use the chain rule to express
the r and derivatives in terms of the x and y derivatives. This gives
u u x u y u u
= + = cos + sin ,
r x r y r x y
2 u u u u u

= cos cos + sin + sin cos + sin
r2 x x y y x y
2 u 2 u 2 u
= cos2 + 2 sin cos + sin2 .
x2 x y y2
Similarly
u u x u y u u
= + = r sin + r cos ,
x y x y
2 u u u u u
= r cos r sin r sin + r cos
2 x y x y
u 2 u 2 u

= r r sin r sin 2 + r cos
r x x y
2 u 2 u

+ r cos r sin + r cos 2
x y y
2
u u 2 u 2
2 u

2 2
= r + r sin 2 2 sin cos + cos 2 .
r x x y y
Thus
2 u 1 u 1 2 u 2 u 2 u
+ + = +
r2 r r r 2 2 x2 y2
as desired.

We return to the Dirichlet problem (2.12) in the unit disc. By switching to


polar coordinates we obtain
2 2
u + 1 u + 1 u = 0, 0 < r < 1,

< ,
r 2 r r r 2 2 (2.13)
u(1, ) = g( ), < ,

for u = u( r, ).

THE MORAL : This is the polar form of the Dirichlet problem (2.12) in the unit
disc. Observe that the domain becomes rectangular in polar coordinates. Solution
of (2.13) will give a solution of (2.12) in polar coordinates.

We shall derive a solution to (2.13) in four steps.

Step 1 (Separation of variables): We look for a product solution of the form

u( r, ) = A ( )B( r )
CHAPTER 2. FOURIER SERIES AND PDES 38

for this problem. Here A ( ) is a function of alone and B( r ) is a function of r


alone. By inserting this into the PDE and multiplying by r 2 , we obtain

r 2 B00 ( r ) + rB0 ( r ) A 00 ( )
r 2 A ( )B00 ( r ) + r A ( )B0 ( r ) + B( r ) A 00 ( ) = 0 = .
B( r ) A ( )
Observe that the variables have been separated in the sense that the left-hand
side depends only on r and the right-hand side only on . This can happen only if
both sides are equal to a constant, say equal to . This is called the separation
constant.

Reason. Denote
r 2 B00 ( r ) + rB0 ( r ) A 00 ( )
( r, ) = =
B( r ) A ( )
for every r and in the appropriate intervals. Then

( r, ) = 0 and ( r, ) = 0
r
and thus ( r, ) is a constant function. Another way to see this is to observe that
the term including B does not depend on and the term including A does not
depend on r . Thus we may conclude that is independent of both variables.

Thus
r 2 B00 ( r ) + rB0 ( r ) A 00 ( )
==
B( r ) A ( )
for every r and . Consequently, we may rewrite the separated equations as two
ordinary differential equations (ODEs)

A 00 ( ) + A ( ) = 0,
r 2 B00 ( r ) + rB0 ( r ) B( r ) = 0.

T HE MORAL : The PDE has been reduced to two ODEs.

As we shall see, not all values of the separation constant give nontrivial
solutions to these ODEs. However, these are simple second order ODEs with
constant coefficients.
Step 2 (Solution to the separated equations): Now we take into account
the boundary data
u(1, ) = A ( )B(1) = g( ).

Since g is defined on the circle it can be identified with a 2-periodic function and
thus A has to be 2-periodic as well.
The question is that for which values of we have nontrivial solutions to

A 00 ( ) + A ( ) = 0 A 00 ( ) = A ( ).

In other words, we are interested in eigenvalues and eigenfunctions A of the


2
second derivative operator 2 . We are mostly interested in the real valued
CHAPTER 2. FOURIER SERIES AND PDES 39

solutions, but sometimes it is useful to consider complex valued solutions as well.


In principle, then the eigenvalues can be complex numbers. However, we begin
with showing that all eigenvalues of the problem above are real.
R Let C be an eigenvalue and A the corresponding complex valued
eigenfunction. Then A 00 = A and by taking the complex conjugate of this
00
equation we obtain A = A . By the chain rule, we have
00 0
A 00 A + A A = ( A 0 A + A A )0 .

We integrate to get
Z Z

00 0 0
( A 00 A + A A ) d = ( A 0 A + A A )0 d = ( A 0 A + A A )


0 0
= A 0 () A () + A () A () ( A 0 () A () + A () A ()).

Since A is periodic, we have A () = A () and A 0 () = A 0 (), from which we


conclude that the right-hand side of the previous display is zero. Thus
Z Z Z
( ) | A |2 d = ( ) A A d = ( A A + A A ) d

Z
00
= ( A 00 A + A A ) d = 0.

R 2
Since A is not identically zero, we have | A | d > 0. Thus = 0, which
means that is a real number.
We divide the analysis into three cases depending on the value of the separa-
tion constant.
< 0 Then = 2 for some > 0 and the first equation above becomes
A 00 2 A = 0. This is a second order linear ODE with the general solution

A ( ) = c 1 e + c 2 e .

In this case the only possible 2-periodic solution is u = 0 which corresponds to


c 1 = c 2 = 0. Thus the case < 0 gives only trivial solutions.
= 0 The equation for A reduces to A 00 = 0 with the general solution A ( ) =
c 1 + c 2 . In order for this function to be 2-periodic ,we must have c 1 = 0 and
thus A ( ) = c 2 is constant. In this case the ODE for B is rB00 + B0 = 0. The general
solution of this equation is B( r ) = d 1 + d 2 ln r and thus we get

u( r, ) = c 2 ( d 1 + d 2 ln r ).

This solution becomes unbounded as r 0 which is contrary to any physical


intuition (and also not compatible with the hypothesis that our solution should be
continuous and bounded in the interior of the disc). Thus = 0 gives nonphysical
solutions, which are excluded.
CHAPTER 2. FOURIER SERIES AND PDES 40

> 0 Then = 2 , > 0, so the equation for A becomes A 00 + 2 A = 0. The


general complex solution of this equation is

A ( ) = c 1 e i + c 2 e i .

(We could consider the real solutions A ( ) = c 1 sin( ) + c 2 cos( ), but the complex
notation in more convenient for the Fourier series.) For this solution to be 2-
p
periodic we need = to be an integer. Thus we have = 2 = j 2 , j Z \ {0}
as the case j = 0 has already been considered. These are the eigenvalues and
eigenfunctions for our problem.
Now the ODE for B is
r 2 B00 + rB0 j 2 B = 0.

It can be shown that the general solution of this equation, called the Euler
equation, is
B( r ) = d 1 r j + d 2 r j , j = 1, 2, . . . .

Again, since > 0, the term with r j blows up as r 0 which contradicts the
continuity of u as well as the physical intuition. Thus we only include the solution
B( r ) = d 1 r j . Thus > 0 gives solutions of the form

u( r, ) = r | j| e i j , j Z.

Observe that these functions are special solutions of the Laplace equation in polar
coordinates with boundary values u(1, ) = e i j , j Z.
Step 3 (Fourier series solution of the entire problem): To solve the orig-
inal Dirichlet problem we try to express a general solution as a linear combination
of the special solutions in such a way that the boundary condition is satisfied.
Since the Laplace operator is a linear, any linear combination of solutions is again
a solution. We conclude that the general solution should be given in the form

a j r| j| e i j ,
X
u( r, ) = 0 r < 1, < . (2.14)
j =

This representation should be compatible with the boundary data g when r = 1.


This means that
a j e i j = g( ).
X
u(1, ) =
j =

We are already familiar with this question for the Fourier series. The question
is that can we determine coefficients a j so that the series representation above
holds? If g L2 ([, )), then by Thoerem 2.25 we see that this is possible, at least
if the equality above is interpreted in L2 -sense. If g C 1 (R), then the discussion
in Section 2.10 shows that the series above converges even uniformly on [, ]
and the coefficients a j will be the Fourier coefficients of g, that is, a j = gb( j ), j Z.

THE MORAL : The solution of the Dirichlet problem for the Laplace equation
in the unit disc is given by the Fourier series of the boundary function.
CHAPTER 2. FOURIER SERIES AND PDES 41

There are several nontrivial points related to the formula above that remain
to be discussed:

(1) Does the series in (2.14) converge?


(2) Is the function given by (2.14) really a solution to the PDE?
(3) Does the solution given by (2.14) attain the correct boundary values?
(4) Are there other solutions than given by (2.14)?

A direct computation shows that the obtained series is a solution to the Laplace
equation and by substituting r = 1 we see that the solution has the desired
boundary values. On a formal level this looks correct, but a special attention has
to be paid to switch the order of the limit and the infinite series. We return to this
question in Section 2.14. Uniqueness will be discussed in Chapter 4.
Step 4 (Explicit representation formula): By subsituting the definition of
the Fourier coefficients of g in (2.14) we obtain
1 Z
g( t) e i jt dt r | j| e i j
X
u( r, ) =
j = 2
n 1 Z
i jt
dt r | j| e i j
X
= lim g ( t) e
j = n 2
n
!
1 n
Z
| j | i j ( t )
X
= lim g ( t) r e dt.
n 2
j = n

The finite sum inside the integral is bounded by



X n
| j | i j ( t )
r | j| < ,
X
r e 0 r < 1,


j=n j=

and thus the sequence of the finite partial sums is uniformly bounded by a constant.
By switching the order of the limit and the integral, we have
!
1 n
Z
| j | i j ( t )
X
u( r, ) = g( t) lim r e dt
2 n
j = n
!
1
Z
| j | i j ( t )
X
= g ( t) r e dt.
2 j =

This can be justified by the dominated convergence theorem, which is out of the
scope of this course. We define the Poisson kernel for the disc to be the 2-periodic
function
r| j| e i j .
X
P r ( ) = P ( r, ) = (2.15)
j =

By using convolutions from Section 2.9 this means that the solution to the Dirichlet
problem can be written as

1
Z
u( r, ) = ( g P r )( ) = g( t)P r ( t) dt.
2
CHAPTER 2. FOURIER SERIES AND PDES 42

THE MORAL : This suggests that the solution of the Dirichlet problem for the
Laplace equation in the unit disc is a convolution of the boundary data with the
Poisson kernel. This is an integral representation of the solution.

From this we see that the solution is well defined whenever the convolution of
g and P r is well defined. The functions u that satisfy the Laplace equation in an
open domain are called harmonic in . Let us look at properties of the Poisson
kernel in more detail.
Lemma 2.41.

1 r2
P r ( ) = P ( r, ) = , 0 r < 1, < .
1 2 r cos + r 2
Remark 2.42. It is obvious that the Poisson kernel is a non-negative function in
the disc, since
1 r2
P ( r, ) = 0.
sin2 + (cos r )2
Furthermore
1
Z
P r ( ) d = 1
2
for every r (0, 1).

T HE MORAL : The total mass of the Poisson kernel is one implies that the
solution of the Dirichlet problem for the Laplace equation in the unit disc is an
average of the boundary function g with the weight P r .

Proof (Proof of Lemma 2.41).


n n n
r| j| e i j = 1 + r j e i j + r j e i j
X X X
j = n j =1 j =1
n n
( re i ) j + ( re i ) j
X X
= 1+
j =1 j =1

( re i )n+1 re i ( re i )n+1 re i
= 1+ + .
1 + re i re i 1
Letting n , and remembering that r < 1, the right-hand side of the display
above tends to

re i re i re i re i + 2 r 2 r 2 r cos
1+ + = 1+ = 1+2
re i 1 1 re i ( re i 1)(1 re i ) re i r 2 1 + re i
2
r cos r 1 r2
= 1+2 2 = ,
r + 1 2 r cos 1 2 r cos + r 2
which is the desired formula. The integral of P r ( ) can be calculated by integrating
(2.15) term by term (exercise).

We summarize the findings in the following theorem.


CHAPTER 2. FOURIER SERIES AND PDES 43

Theorem 2.43 (Solution of the Dirichlet problem on the disc). The solution
of the Dirichlet problem (2.12) in polar coordinates is

a j r| j| e i j ,
X
u( r, ) = 0 r < 1, < ,
j =

where
1
Z
a j = gb( j ) = g( ) e i j d , j Z.
2
Moreover, the solution can be written as a convolution with the Poisson kernel as
1
Z
u( r, ) = ( g P r )( ) = g( t)P r ( t) dt,
2
where
1 r2
P r ( ) = P ( r, ) = , 0 r < 1, < .
1 2 r cos + r 2
We close this section by considering two explicit examples.

Example 2.44. Consider the steady-state temperature distribution in the unit


disc, if the upper half of the circle is kept at 100 C and the lower half is kept at
0 C. Then
0, < < 0,
u(1, ) = f ( ) =
100, 0 < < .

The Fourier coefficients of f are


Z 0 Z
1

fb( j ) = 0 e i jt dt + 100 e i jt dt
2 0

100 e i jt 100 i j
= = (e e0 )
2 i j 0 2 i j

100 100 0, j even,
= (cos( j ) 1) = (1 cos( j )) =
2 i j 2 i j 100 i , j odd,
j

when j 6= 0. For j = 0 we have


Z 0 Z
1

f (0) =
b 0 dt + 100 dt = 50.
2 0

Thus after some simplifications (exercise)



fb( j ) r | j| e i j
X
u( r, ) =
j =
1
100 X
= 50 + (1 cos( j )) r j sin( j ), 0 r 1, < .
j=1 j

By setting r = 0, we see that the temperature at the center of the disc is 50 C,


which is the average temperature on the boundary of the disc. On the boundary of
the disc with r = 1 we have

200 X 1
u(1, ) = 50 + sin((2 j + 1) ), < ,
j=0 2 j + 1

which is the Fourier series of f .


CHAPTER 2. FOURIER SERIES AND PDES 44

2 u 2
Example 2.45. Consider u = x2
+ yu2 = 0 in the rectangle = [0, a] [0, b] with
the boundary conditions



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


u( x, b) = 0,

0 < x < a,



u(0, y) = 0, 0 y b,


u(a, y) = g( y), 0 y b,

where f is a continuously differentiable function on [0, b]. We look for solutions of


the form u( x, t) = A ( x)B( y). Inserting this into u = 0 gives

A 00 ( x) B00 ( y)
0 = u( x, y) = A 00 ( x)B( x) + A ( x)B00 ( x) + =0
A ( x) B( y)

for every x and y. This implies

A 00 ( x) B00 ( y)
== , R,
A ( x) B ( y)

for every x and y.


< 0 = 2 , > 0, and

B00 ( y) 2 B( y) = 0 B( y) = c 1 sinh( y) + c 2 cosh( y).

The boundary condition B(0) = B( b) = 0 implies c 1 = c 2 = 0 and thus B( y) = 0.


=0
B00 ( y) = 0 B( y) = c 1 y + c 2 .

Again B(0) = B( b) = 0 implies c 1 = c 2 = 0 and thus B( y) = 0.


> 0 = 2 , > 0, and we have

A 00 ( x) 2 A ( x) = 0 A ( x) = c 1 sinh( x) + c 2 cosh( x)

B00 ( y) + 2 B( y) = 0 B( y) = d sin( y) + d cos( y)
1 2

Now A (0) = 0 = c 2 = 0 and B(0) = 0 = d 2 = 0. Moreover, B( b) = 0 implies d 1 = 0


or sin( b) = 0. Clearly

j
sin( b) = 0 = = , j = 1, 2, . . . ,
b
and d 1 = 0 = B( y) = 0. Thus

A ( x) = c 1 sinh jx
b
B( y) = d sin j y
1 b

and
jx j y
u( x, y) = A ( x)B( y) = a j sinh sin , j = 1, 2, . . . ,
b b
CHAPTER 2. FOURIER SERIES AND PDES 45

are nontrivial special solutions. We look for the solution of the general problem in
the form
X jx j y
u( x, y) = a j sinh sin .
j =1 b b

The boundary condition



X j a j y
g( y) = u(a, y) = a j sinh sin
j =1 b b

by the Fourier sine series representation gives

j a 2 b j y
Z
a j sinh = g( y) sin d y, j = 1, 2, . . . ,
b b 0 b

and consequently

2 b j y
Z
aj = j a
g( y) sin d y, j = 1, 2, . . . .
b sinh 0 b
b

Observe, that the boundary function g( y) is extended as an odd 2 b-periodic func-


tion to R by setting g( y) = g( y), b < y < 0 and g( y) = g( y + 2 b), see Remark
2.30. Thus the Fourier cosine coefficients are zero. This gives a representation
formula for solution of the problem.

2.12 The heat equation in one-dimension


Let us now consider the time dependent heat diffusion in the one-dimensional
case. Suppose we have a ring of radius one centered at the origin. Suppose further
that the sides of the ring are perfectly insulated so that no heat passes through
them. At time t = 0 the initial temperature distribution on the ring is given by the
function g : [, ] R. We can assume the function g to be 2-periodic since it is
defined on the unit circle.
The diffusion of heat on the circle is modeled by the heat equation

u 2 u
a2 = 0,
t 2

where a2 is a positive constant known as the thermal diffusivity, which depends


only on the material from which the ring is made. Again we set a2 = 1. We are
looking for a solution u = u( , t) where describes the position on the ring and
t > 0 is time. The initial condition is u( , 0) = g( ). Thus we have the periodic
initial value problem
2
u ( , t) u ( , t) = 0, < ,

t > 0,

t 2 (2.16)
u( , 0) = g( ), < .

CHAPTER 2. FOURIER SERIES AND PDES 46

Figure 2.9: Heat distribution in a ring.

Step 1 (Separation of variables): We use the separation of variables again


and look for special solutions of the form

u( , t) = A ( )B( t).

Inserting this into the heat equation we get

B0 ( t) A 00 ( )
A ( )B0 ( t) A 00 ( )B( t) = 0 = .
B ( t) A ( )

The variables are separated in the sense that the two sides of the equation above
depend on different variables and thus both have to be constant denoted by .
Thus
A 00 ( ) = A ( ),
B 0 ( t ) = B ( t ).

Step 2 (Solution to the separated equations): The initial condition

u( , 0) = A ( )B(0) = g( )

is a 2-periodic function. We do a similar case study on the possible values of ,


as we did in the case of the Laplace equation.
> 0 Then = 2 , > 0, and thus we have A 00 ( ) 2 A ( ) = 0. This ODE has
the general solution
A ( ) = c 1 e + c 2 e ,
CHAPTER 2. FOURIER SERIES AND PDES 47

Figure 2.10: The space-time domain for the heat distribution in a ring.

which is periodic only if c 1 = c 2 = 0. This gives the trivial solution u = 0 and thus
we exclude this alternative.
= 0 Then A 00 ( ) = 0, which implies A ( ) = c 1 + c 2 . Here the only possibility
that is compatible with the periodicity of A is the solution A ( ) = c 2 is constant.
In this case we also get that B( t) = c 3 so that this gives only the constant solution.
< 0 Then = 2 , > 0 and the ODE for A ( ) is A 00 ( ) + 2 A ( ) = 0. The
general complex solution of this ODE is

A ( ) = c 1 e i + c 2 e i

for some constants c 1 and c 2 . Again for A to be periodic we need to be an integer


which gives thus, considering all cases above, that the only admissible values for
are of the form = j 2 , j Z. For these values we have B0 ( t) + j 2 B( t) = 0, which
has the general solution
2
B( t) = ce j t .

Thus we have special solutions, called the normal modes, of the form
2
u( , t) = e j t e i j , j Z.

By the construction, these functions are solutions of the one-dimensional heat


equation for every j Z.
Step 3 (Fourier series solution of the entire problem): Since the heat
equation is linear, any linear combination of the special solutions above will give
CHAPTER 2. FOURIER SERIES AND PDES 48

again a solution. Motivated by the superposition principle, we define


2
a j e j t e i j ,
X
u ( , t ) = < , t > 0.
j =

The next step is to determine coefficients a j . By the initial condition u( , 0) = g( ),


we have
a j e i j = g( )
X
j =

which identifies the coefficients a j as the Fourier coefficients gb( j ) of the initial
data g. Thus
2
gb( j ) e j t e i j .
X
u ( , t ) =
j =

This infinite series converges absolutely, since the Fourier coefficients gb( j ) are
2t
bounded if g L1 ([, ]) (see Lemma 2.20 (2)) and the term e j decays ex-
tremely fast as t > 0 and j is large. Thus
2
2
2
b( j ) e j t e i j b( j )| e j t k gkL1 ([,]) e j t .
X X X
g |g

j = j = j =
| {z }
<

The last series converges, which shows that the series defining the solution to the
heat equation converges absolutely. This allows us to differentiate the series term
by term. Thus
u 2 u 2 u

2t 2t

e j e i j e j e i j
X
( , t) ( , t) = gb( j )
t 2 j = t 2
2 2

gb( j ) ( j 2 ) e j t e i j e j t ( i j )2 e i j
X
=
j 6=0
2 2

gb( j ) ( j 2 ) e j t e i j + e j t j 2 e i j = 0.
X
=
j 6=0

This shows that u( , t) is a solution of the heat equation.


Step 4 (Explicit representation formula): Next goal is to derive an inte-
gral representation for the solution similar to the convolution with the Poisson
kernel for the Dirichlet problem for the Laplace equation in the unit disc. Using
the definition of the Fourier coefficients we obtain
2
gb( j ) e j t e i j
X
u( , t) =
j =
n
1
Z
2
i js
ds e j t e i j
X
= lim g ( s) e
n
j = n 2
!
1
Z n
j 2 t i j ( s )
X
= lim g ( s) e e ds.
n 2 j = n

We observe, as with the Poisson kernel, that we have the uniform estimate

X n
j 2 t i j ( s ) 2
e j t < .
X
e e


j=n j=
CHAPTER 2. FOURIER SERIES AND PDES 49

Here it is crucial that t > 0, so these estimates do not generalize to negative times.
By switching the order of the limit and the integral, we obtain
!
1 n
Z
j 2 t i j ( s )
X
u( , t) = g( s) lim e e ds
2 n
j = n
!
1
Z
j 2 t i j ( s )
X
= g ( s) e e ds.
2 j =

By defining the heat kernel for the circle as


2
e j t e i j ,
X
H t ( ) = < , t > 0,
j =

we see that the solution of the heat equation on the circle is given by the convolu-
tion
1
Z
u( , t) = ( g H t )( ) = g( s) H t ( s) ds, < , t > 0.
2

THE MORAL : The solution of the initial value problem for the heat equation
on the circle is a convolution of the initial temperature distribution with the heat
kernel.

The following result gives convergence of the solution to the initial data f as
time t 0+ . We shall prove this in section 2.14.

Theorem 2.46 (Solution to the heat equation on the unit circle). The solu-
tion of the (periodic) initial value problem (2.16) is
1
Z
u( , t) = ( g H t )( , t) = g( s) H t ( s) ds, < , t > 0,
2
j2 t i j
where H t ( ) =
P
j = e e .

Remark 2.47. Another way to derive the solution to a PDE problem is start with
the Fourier series expansion, insert it in the PDE and try to determine the
coefficients. This approach works for the Laplace equation, the heat equation
and the wave equation, but we shall discuss only the initial value problem (2.16)
in.detail. For every fixed t > 0, consider the Fourier series of the function 7
u( , t). This gives

c j ( t) e i j ,
X
u( , t) =
j =
where the Fourier coefficients are
1
Z
c j ( t) = u( s, t) e i js ds, j Z.
2
Observe that the Fourier coefficients depend on t. We claim that the PDE for u
implies that the Fourier coefficients satisfy an ODE as a function of t. To see this,
we switch the order of integral and derivative to have
Z
u 1 1 u
Z
i js
0
c j ( t) = u( s, t) e ds = ( s, t) e i js ds, j Z.
t 2 2 t
CHAPTER 2. FOURIER SERIES AND PDES 50

Then we use the PDE and Lemma 2.31 twice to obtain


1 u
Z
c0j ( t) = ( s, t) e i js ds
2 t
1 2 u
Z
= ( s, t) e i js ds
2 s2
1 u
Z
= ij ( s, t) e i js ds
2 s
Z
2 1
= ( i j) u( s, t) e i js ds
2
= j 2 c j ( t), j Z.

THE MORAL : The PDE becomes an ODE on the Fourier side.


2
The general solution of c0j ( t)+ j 2 c( t) = 0 is c j ( t) = a j e j t , where a j is a constant.
By the Fourier series representation
2
c j ( t) e i j = a j e j t e i j ,
X X
u( , t) = < , t > 0.
j = j =

By the initial condition u( , 0) = g( ), we have



a j e i j = g( )
X
u( , 0) =
j =

which identifies the coefficients a j as the Fourier coefficients gb( j ) of the initial
data g. Thus
2
gb( j ) e j t e i j .
X
u ( , t ) =
j =

T HE MORAL : In the this approach we inserted the Fourier series represen-


tation of the function in the PDE and determined the Fourier coefficients by the
initial condition. In the beginning of this section we derived the same formula
for the solution by separation of variables and this lead to the Fourier series.
Although the result is same, the main difference is that in the original argument
we did not have to assume in the beginning that the solution is given by the
Fourier series. However, theory of Fourier series is needed in both approaches.

Let us consider another problem for the one-dimensional heat equation. We


study the temperature distribution in a thin uniform bar of given length L > 0
with insulated lateral surface and no internal sources of heat, subject to certain
boundary and initial conditions. To describe the problem, let u( x, t), 0 < x < L, t > 0,
be the temperature at the point x at time t. Assume that the initial temperature
distribution of the bar is u( x, 0) = g( x) and assume, that the ends of the bar are
held at constant temperature 0 C, for example. This is related to the previously
considered problem of the heat distribution on a circle, since we can cut the circle
and unfold it so that it becomes a bar.
CHAPTER 2. FOURIER SERIES AND PDES 51

Figure 2.11: The heat distribution in a bar.

The following theorem gives a solution of the heat distribution in the bar of
length L > 0. We shall prove this in the exercises using the method of separation
of variables.

Theorem 2.48 (Solution of the heat equation on a bar). The solution of


2
u 2 u

a = 0, 0 < x < L, t > 0,


t x2

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



u( x, 0) = g( x), 0 x L.

is
X 2j t jx
u( x, t) = aje sin ,
j =1 L

where
2 L jx j
Z
aj = g( x) sin dx and j = a , j = 1, 2, . . . .
L 0 L L

The method of separation of variables can also be used to solve heat conduction
problems with other boundary conditions than those given above.

Example 2.49. Consider a bar of lenght is in boiling water. After reaching the
temperature 100 C throughout, the bar is taken out and immersed in a medium
with constant freezing temperature 0 C. The the bar are kept insulated and we
CHAPTER 2. FOURIER SERIES AND PDES 52

Figure 2.12: The space-time domain for the heat distribution in a bar.

assume that a2 = 1. Then the temperature distribution is


2
a j e j t sin( jx),
X
u( x, t) =
j =1

where
2
Z 200
aj = 100 sin( jx) dx = (1 cos( j )).
0 j
After some simplifications (exercise)
2
e(2 j +1) t
400 X
u( x, t) = sin((2 j + 1) x).
j=0 2 j + 1

Here we used the fact that 1 cos( j ) = 0 if j is even and 2 if j is odd.

2.13 The wave equation in one-dimension


In this section we study the motion of an elastic string of length L > 0 fixed at its
end points and allowed to vibrate freely. The elastic string may be thought of as
a violin string, for example. Suppose that the string is set in motion so that it
vibrates in a vertical plane and let u( x, t) denote the vertical displacement of the
string at the point x [0, L] at time t 0. This is modeled by the wave equation

2 u 2 u
a2 = 0.
t2 x2
CHAPTER 2. FOURIER SERIES AND PDES 53

We are looking for a solution u = u( x, t), where x [0, L] describes the position on
the string and t > 0 is time.

Figure 2.13: A vibrating string.

The initial displacement profile and the velocity of the string at time t = 0 are
given as the initial conditions

u
u( x, 0) = g( x) and ( x, 0) = h( x), 0 x L.
t
Since the end points of the string are fixed, we have the boundary conditions

u(0, t) = u(L, t) = 0 for every t 0.

The problem is to determine the solution of the wave equation that also satisfies
the the initial conditions and boundary conditions above. This can be considered
as a boundary value problem in the strip 0 < x < L, t > 0, in the xt-plane. With
this interpretation, the boundary condition is imposed on the sides of the strip
and the two initial conditions are imposed on the base of the strip. Thus we have
the problem
2 u 2 u

a2 2 = 0, 0 < x < L,


t > 0,
t x


2

u(0, t) = u(L, t) = 0, t 0,
(2.17)
u( x, 0) = g( x), 0 x L




u



( x, 0) = h( x), 0 x L.
t
CHAPTER 2. FOURIER SERIES AND PDES 54

Figure 2.14: The one-dimensional space-time domain.

Step 1 (Separation of variables): We look for special solutions of the form

u( x, t) = A ( x)B( t),

where A ( x) is a function of x alone and B( t) is a function of t alone. Plugging this


into the wave equation, we get
B00 ( t) A 00 ( x)
A ( x)B00 ( t) a2 A 00 ( x)B( t) = 0 = .
a2 B ( t) A ( x)
Again we note that the two sides of the identity above depend on different variables
and thus both have to be constant and equal to . This leads to the ODEs

A 00 ( x) = A ( x),
B00 ( t) = a2 B( t).

We could consider the complex solution of these ODE, as we did in the cases of the
Laplace and the heat equations. However, let us show here how to do the analysis
using the real Fourier series.
Step 2 (Solution to the separated equations): We do a similar case study
on the possible values of , as we did in the case of the Laplace and the heat
equations.
> 0 Then = 2 , > 0, and thus we have A 00 2 A = 0. This ODE has the
general real solution

A ( x) = c 1 cosh( x) + c 2 sinh( x).


CHAPTER 2. FOURIER SERIES AND PDES 55

We show that the only way to satisfy the boundary conditions is to take c 1 = c 2 = 0.
Indeed, A (0) = 0 implies 0 = c 1 cosh(0) + c 2 sinh(0) = c 1 so that A ( x) = c 2 sinh( x).
The condition A (L) = 0 implies c 2 sinh(L) = 0. However, L 6= 0 and so sinh(L) 6=
0. Thus c 2 = 0 and this gives the trivial solution u = 0. We exclude this alternative.
= 0 Then A 00 ( x) = 0, which implies A ( x) = c 1 x + c 2 . Here the only way to
satisfy the boundary condition is to take c 1 = c 2 = 0, which again leads to the
trivial solution u = 0.
< 0 Then = 2 for some > 0 and the ODE for A is A 00 + 2 A = 0. The
general real solution of this ODE is

A ( x) = c 1 cos( x) + c 2 sin( x)

for some constants c 1 and c 2 . Since A (0) = 0, we have c 1 = 0 and A ( x) = c 2 sin( x).
Since A (L) = 0, we have c 2 sin(L) = 0, from which we conclude that L = j ,
j = 0, 1, 2, . . . (or c 2 = 0, which again gives the trivial solution u( x, t) = 0). This
gives
j
= , j = 1, 2, . . .
L
and
jx

A ( x) = c sin , j = 1, 2, . . .
L
for any constant c that may depend on j . Note that for negative values of j we
obtain the same solutions except for a change of sign. Since the general solution
will be represented as a linear combination of the special solutions, the solutions
with negative j can be discarded without loss.
We have
j 2

2
= = , j = 1, 2, . . . .
L
With this, the ODE for B becomes
a j 2

B00 ( t) + B( t) = 0.
L
The general solution of this equation is
a jt a jt

B( t) = b 1 cos + b 2 sin , j = 1, 2, . . .
L L
for any constants b 1 and b 2 that may depend on j . Thus the product solution is
jx a jt a jt

u( x, t) = A ( x)B( t) = sin a j cos + b j sin , j = 1, 2, . . .
L L L
Observe, that we have absorbed the constant in front of sine to the constants a j
and b j .
Step 3 (Fourier series solution of the entire problem): Since the wave
equation is linear, any linear combination of these solutions will give again a
solution. We thus define
jx a jt a jt

X
u( x, t) = sin a j cos + b j sin
j =1 L L L
CHAPTER 2. FOURIER SERIES AND PDES 56

By the initial condition u( x, 0) = g( x), we have


jx

X
u( x, 0) = a j sin = g( x).
j =1 L

which identifies the coefficient a j as the Fourier sine coefficient of the initial
data f on [0, L]. Observe, that the boundary function g( x) is extended as an odd
2L-periodic function to R by setting g( x) = g( x),L < x < 0 and g( x) = g( x + 2L),
see 2.30. Thus the Fourier cosine coefficients are zero. This gives

2 L jx
Z
aj = g( x) sin dx, j = 1, 2, . . .
L 0 L

On the other hand, by differentiating the series termwise we have

u a X
jx a jt a jt

( x, t) = sin ja j sin + jb j cos ,
t L j=1 L L L

from which we obtain


u a X
jx

( x, 0) = jb j sin = h( x).
t L j =1 L

a
Consequently, L jb j must be the Fourier sine coefficient of h on [0, L]. The
boundary function h( x) is extended as an odd 2L-periodic function to R and thus
the Fourier cosine coefficients are zero. This gives

a 2 L jx
Z
jb j = h( x) sin dx, j = 1, 2, . . .
L L 0 L

or equivalently

2 L jx
Z
bj = h( x) sin dx, j = 1, 2, . . .
a j 0 L

We have now determined all the unknown coefficients in the series representation
of the solution u. We summarize our findings in here.

Theorem 2.50 (Solution of the one-dimensional wave equation). The solu-


tion of the problem (2.17) is
jx a jt a jt

X
u( x, t) = sin a j cos + b j sin ,
j =1 L L L

where
2 L jx
Z
aj = g( x) sin dx, j = 1, 2, . . . ,
L 0 L
and
2 L jx
Z
bj = h( x) sin dx, j = 1, 2, . . . .
a j 0 L
CHAPTER 2. FOURIER SERIES AND PDES 57

As in the case of the Laplace equation and the heat equation considered earlier,
this is only a formal solution of the problem. To show that the obtained formula
actually represents the solution of the problem requires further investigations
(exercise).

Remark 2.51. The solution of the vibrating string problem is an infinite sum of
the normal modes
jx a jt a jt

u j ( x, t) = sin a j cos + b j sin , j = 1, 2, . . . .
L L L

When the string vibrates according to u j , we say that it is in the j th normal mode
of vibration. The first normal mode is called the fundamental mode and the the
other modes are overtones. The quantities a j /L, j = 1, 2, . . . , are the natural
frequencies of the normal mode, which gives the number of oscillations in 2 units
of time. The factor sin( j x/L) is the displacement pattern of the string when it
vibrates at the given frequency. When the string vibrates in its normal mode,
some points of the string are fixed at all times. These are the solutions of the
equation sin( j x/L) = 0. If we do not count the ends of the string, there are n 1
equidistant points that do not vibrate in the nth normal mode.

Figure 2.15: Normal modes of a string.

Example 2.52. The ends of a string of length L = 1 are fixed at x = 0 and x = 1.


Assume that a = 1. The string is set to vibrate from rest with a initial triangular
CHAPTER 2. FOURIER SERIES AND PDES 58

profile
3x , 0 x 13 ,
10
g ( x) =
3(1 x) x, 1 x 1.
20 3

Since h( x) = 0, we have b j = 0. Using the formula for the solution and integrating
by parts, we have
Z L
aj = 2 g( x) sin( j x) dx
0
3
Z 1/3 3
Z 1
= x sin( j x) dx + (1 x) sin( j x) dx
5 0 10 1/3
j j j j j
cos 3 3 sin 3 cos 3 3 sin 3 9 sin 3
= + + + = .
5 j 5 j 2 2 5 j 10 j 2 2 102 j2

Thus
sin j

9 X 3
u( x, t) = sin( j x) cos( j t).
102 j=1 j2

2.14 Approximations of the identity


in [, ]*
There is a common theme in the expressions for the partial sums of the Fourier
series of a function, the solution to the Dirichlet problem for the Laplace equation
in the disc and the solution to the 2-periodic initial value problem for the heat
equation. Indeed, by (2.11), Theorem 2.43 and Theorem 2.46, we have
n
fb( j ) e i j , fb( j ) e i j ,
X X
S n f ( ) = ( f D n )( ) = f=
j = n j =

fb( j ) r | j| e i j ,
X
P r f ( ) = ( f P r )( ) = 0 < r < 1,
j =
2
fb( j ) e j t e i j ,
X
H t f ( ) = ( f H t )( ) = t > 0.
j =

The first formula encodes the most basic and natural question in Fourier series:
Can we recover a function f from the partial sums of its Fourier series? We have
seen that this is not always possible. The pointwise limit

lim S n f = lim (D n f )
n n

may fail to exist even if f is continuous. However, by Theorem 2.37 this limit
exists, if f is continuously differentiable. Moreover, for a continuous function f ,
we are interested in existence of the pointwise limits

lim (P r f ) = f and lim ( H t f ) = f .


r 1 t0+
CHAPTER 2. FOURIER SERIES AND PDES 59

These are related to the question in which sense the boundary or initial values
are obtained in the corresponding problems. It turns out that the Poisson kernel
and the heat kernel are just special cases of a general theory of good kernels.

Definition 2.53. A family of kernels {K }>0 which are 2-periodic functions is


said to be a family of good kernels, if it satisfies the following properties.

(1) For every > 0 we have


1
Z
K ( x) dx = 1.
2

(2) There exists a positive constant M such that


1
Z
|K ( x)| dx M
2
for every > 0.
(3) For every > 0 we have
Z
lim |K ( x)| dx = 0.
0+ <| x|

T HE MORAL : Parameter > 0 gives the scale at which we take samples of


functions. Condition (1) means that the total mass of the kernel is one at all scales
and condition (3) means that the mass concentrates to the origin at small scales.
Remarks 2.54:
(1) For nonnegative kernels K 0, property (2) is a consequence of (1) with
M = 1.
(2) Property (3) is the most crucial and describes the fact that, as 0+ , the
mass of the kernel K concentrates more and more around the origin.
(3) It can be shown that the Poisson kernel P1 and the heat kernel H t are
both good kernels (exercise). It is now natural to ask whether D n is a
good kernel, since if this were true, we would be able to conclude that the
Fourier series of a continuous function converges pointwise. Unfortunately
this is not the case. Indeed, the Dirichlet kernel is a sign changing function
and it does not satisfy condition (2) above.

The following theorem explains why good kernels are very useful. Because
of this result, the family {K }>0 is sometimes referred to as an approximation
of the identity. Note that the theorem below immediately provides the proof for
Theorem 2.43 and Theorem 2.46 for the Poisson and the heat kernel, respectively.

Theorem 2.55. Let {K }>0 be a family of good kernels and f : [, ] C be a


bounded 2-periodic function. Then

lim ( f K )( x) = f ( x)
0+

whenever f is continuous at x. If f is continuous on the whole interval [, ],


then the above limit is uniform.
CHAPTER 2. FOURIER SERIES AND PDES 60

Figure 2.16: A family of good kernels.

T HE MORAL : The fact that the boundary and initial values are obtained is
based on a general properties of the convolution with approximations of identity.
On the other hand, this can be used to approximate a given function with smoother
functions. Several fundamental solutions of PDEs give rise to an approximation
of identity.

Proof. Let > 0. Since f continuous at x there exists > 0 such that

| f ( x y) f ( x)| <

whenever | y| < . Then by property (1) of a good kernel we can write

1
Z
( f K )( x) f ( x) = ( f ( x y) f ( x))K ( y) d y.
2
Thus
1
Z
|(K f )( x) f ( x)| |K ( y)|| f ( x y) f ( x)| d y
2 | y|<
1
Z
+ |K ( y)|| f ( x y) f ( x)| d y.
2 <| y|
Z
1
Z
|K ( y)| d y + sup | f ( t)| |K ( y)| d y.
2 t[,] <| y|

Now letting 0+ the second term tends to zero by property (3) of the good
kernels. Since > 0 was arbitrary, this completes the proof of the first part of
CHAPTER 2. FOURIER SERIES AND PDES 61

the theorem. For the second part note that if f is continuous on [, ] and
2-periodic, then it is uniformly continuous. Thus the > 0 in the argument above
can be chosen independently of x, which shows that the convergence is uniform in
this case.

Example 2.56. Let


1 x
K ( x) = 1[,] , 0 < < 1,

where
1, if x [a, b],
1[a,b] ( x) =
0, otherwise,

is the indicator function (or the characteristic function) of the set [a, b].

Figure 2.17: A good kernel related to the integral average.

Observe that
1
Z 1
Z 1
K ( x) dx = dx = 1, > 0.
2 2

Thus for all > 0 the function K has the same total mass one, but as 0 this
mass is concentrated more and more around the origin. Now for any f L1 ([, ])
CHAPTER 2. FOURIER SERIES AND PDES 62

we have
1
Z
( f K )( x) = f ( x t)K ( t) dt
2
Z
1
= f ( x t) dt
2
Z x+
1
= f ( t) dt.
2 x

Thus the value ( f K )( x) is equal to the average of f on a symmetric interval of


length 2 around the point x. In this sense we can think of the convolution as an
averaging process. Observe that, if f is continuous at x, then
Z x+
1
lim( f K )( x) = lim f ( t) dt = f ( x).
0 0 2 x

This means that the pointwise value of a continuous function is the limit of the
integral averages.

2.15 Summary
The main steps in the application of the Fourier series to PDE problems are the
following.

(1) The task is to find the solution u of an initial or boundary value problem
in a rectangular domain.
(2) Separate variables and insert the product solution to the PDE.
(3) This reduces the problem to two ODE.
(4) The two ODE are solved explicitly to find nontrivial special solutions.
(5) The general solution of a problem is a linear combination of nontrivial
special solutions.
(6) The initial or boundary conditions are used to represent the coefficients
using Fourier series.
(7) It is a matter of taste whether one wants to work with the real or the
complex form of the Fourier series.
(8) The solution of the original problem is represented as a convolution of
the data with a kernel function, which is a fundamental solution of the
corresponding problem.
(9) This gives an integral representation of the solution to the original problem
and the initial or boundary values are attained by using approximations
of the unity.
Fourier transform gives a method to solve PDEs in the
higher dimensional case. It has many properties analogous
to the Fourier series, but the functions do not need to be
periodic. Convolution plays a central role in the theory.

Fourier transform and PDEs


3
The theory of the Fourier series applies to periodic functions in the one-dimensional
case. In this chapter, we develop an analogous theory for functions defined on the
whole n-dimensional Euclidean space Rn , which are not assumed to be periodic.
The Fourier series of a periodic function associates the Fourier coefficients to
that function and, under certain assumptions, a function can be represented as a
Fourier series. In this section we shall see that, under certain assumptions, a func-
tion can be represented by its Fourier transform, which is a function associated to
the original function.

3.1 The L p -space on Rn *


We consider functions defined in the n-dimensional Euclidean space Rn , n = 1, 2, . . . .
We discuss briefly some aspects of the integration of such functions. A function
f : Rn C is integrable, if Z
| f ( x)| dx < .
Rn
Here the n-dimensional integral is computed as an iterated integral
Z Z Z
f ( x) dx = f ( x1 , . . . , xn ) dx1 dxn .
Rn R R

We begin with listing some basic properties of the integral.

(1) For f : Rn C and A Rn , we have


Z Z

f ( x) dx | f ( x)| dx.

A A

(2) If A B Rn , then Z Z
| f ( x)| dx | f ( x)| dx.
A B

63
CHAPTER 3. FOURIER TRANSFORM AND PDES 64

(3) If | f | | g| and A Rn , then


Z Z
| f ( x)| dx | g( x)| dx.
A A

WA RNING : It is essential in (2) and (3) that we consider the absolute values
of the functions. The corresponding claims do not hold for real valued functions
that are allowed to change signs.

We recall the the definition of the L p (Rn ) spaces. For 1 p < , the space
L p (Rn ) consists of functions f : Rn C such that
Z 1
p
p
kf k L p (R n ) = | f ( x)| dx < .
Rn

For p = we set
k f kL (Rn ) = sup | f ( x)|.
xRn
It can be shown that L p (Rn ), 1 p , is a complete normed space (Banach
space) with the norm defined above, but this is out of the scope of this course.
Examples 3.1:
(1) Let f : Rn R, f ( x) = 1. Then k f kL (Rn ) = 1 < , so that f L (Rn ).
However, k f kL p (Rn ) = , so that f L p (Rn ) whenever 1 p < .
(2) Let f : R R,
1, | x| 1,
f ( x) =
1

| x |2
, | x | > 1.
Then

+
1 1
Z Z Z

| f ( x)| dx = 1 dx + 2 dx = 2 + 2 = 4 < .
R | x|1 1 x2 1 x
1
Thus f L (R).
1
(3) Let f : R R, f ( x) = 1+| x| .

Claim: f L (R).

Reason. For any x R we have 1 + | x| 1 and


1
| f ( x)| = 1,
1 + | x|
so that k f kL (R) = sup xR | f ( x)| 1.

Claim: f L1 (R).

Reason. For x 1 we have


1 1
.
1 + | x| 2 x
Thus
1 1
Z Z Z
| f ( x)| dx dx dx = .
R 1 1 + | x| 1 2x
CHAPTER 3. FOURIER TRANSFORM AND PDES 65

Claim: f L2 (R).

Reason.
1
Z Z
| f ( x)|2 dx = 2
dx
R R (1 + | x|)
1 1
Z Z
= 2
dx + 2
dx
| x|1 (1 + | x|) | x|>1 (1 + | x|)
1
Z Z
1 dx + 2
dx
| x|1 | x|>1 x

1
= 2 + 2 = 2 + 2 = 4 < .
1 x

THE MORAL : L2 (R) is not contained in L1 (R). Note carefully, that this
is different from the case with an interval of finite length when we have
L2 ([, ]) L1 ([, ]).
(4) Let f : R R,
p1 , if 0 < x 1,
x
f ( x) =
0, otherwise.

Then f L (R), because f is unbounded near x = 0.


Claim: f L1 (R).
Reason.
Z Z 1 1
| f ( x)| dx = p dx < .
R 0 x

Claim: f L2 (R).
Reason.
Z Z 1 1
| f ( x)|2 dx = dx =
R 0 x

T HE MORAL : L1 (R) is not contained in L2 (R).

Remark 3.2. There are essentially two reasons why a function may fail to belong
to L1 (R). The first reason is the decay of the function near infinity. For example,
the function 1/(1 + | x|)2 decays fast enough at infinity so that it belongs to L1 (R).
On the other hand, the function 1/(1 + | x|) does not decay fast enough so that it
would belong to L1 (R). A second reason is that the function may blow up at a
given point. Typical example is the function that equals 1/| x| when | x| 1 and 0
otherwise. It blows up near x = 0 too fast in order to belong to L1 (R). On the other
p
hand, the function which agrees with 1/ | x| when | x| 1 and is 0 otherwise, also
p
bows up near x = 0. However, 1/ | x| is integrable on | x| 1 and thus it belongs
to L1 (R). The borderline case for L1 (R), both at infinity and close to a point is
the function 1/| x|. This function does not decay fast enough at infinity in order to
belong to L1 (R). At the same time it blows up too fast near x = 0 in order to belong
to L1 (R).
CHAPTER 3. FOURIER TRANSFORM AND PDES 66

3.2 The Fourier transform*


Let f : Rn C be a complex valued integrable function, that is f L1 (Rn ). The
Fourier transform fb : Rn C (denoted also F ( f )) of f is defined as
Z
fb() = F ( f )() = f ( x) e ix dx,
Rn

where x = x1 1 + + xn n is the standard inner product in Rn .

T HE MORAL : The Fourier transform is a replacement of the Fourier coeffi-


cients for a function defined on the whole space.

Remark 3.3. There are several alternative definitions for the Fourier transform
in the literature, for example,
Z Z Z
n
f ( x) e ix dx, f ( x) e2 ix dx and (2) 2 f ( x) e ix dx.
Rn Rn Rn

There is an analogous theory for these definitions and, as we shall see, the factor
2 appears somewhere in each of these choices.

We gather many useful properties of the Fourier transform in the following


proposition. Note that there are many similar properties for the Fourier coeffi-
cients of a function in L1 ([, ]), compare to Lemma 2.20.

Lemma 3.4. Let f , g L1 (Rn ).

f + b g() = a fb() + b gb(), a, b C.


(1) (Linearity) a
(2) (Boundedness) For every Rn we have
Z
| fb()| | f ( x)| dx = k f kL1 (Rn ) .
Rn

This implies k fbkL (Rn ) k f kL1 (Rn ) .


Z
(3) fb(0) = f ( x) dx.
Rn
(4) (Continuity) If f L1 (Rn ), then fb is continuous.
(5) (Dilation) Let f a ( x) = a1n f ax , a > 0. Then c
f a () = fb(a).

(6) (Translation) For y Rn we have f( x + y)() = e i y fb().


(7) (Modulation) For Rn we have eix f ( x)() = fb( ).

Proof. (1) Follows from linearity of integral.


Z Z Z
ix ix
(2) | f ()| =

f ( x) e dx f ( x) e dx = | f ( x)| dx.
b
Rn Rn Rn
(3) Follows immediately by inserting = 0 in the definition of the Fourier
transform.
(4) Claim: k implies fb(k ) fb() as k .
CHAPTER 3. FOURIER TRANSFORM AND PDES 67

Reason. Since the complex exponential function is continuous, we have

lim f ( x) e ixk = f ( x) e ix .
k

This implies that


Z
lim fb(k ) = lim f ( x) e ixk dx
k k Rn
Z
= lim f ( x) e ixk dx
Rn k
Z
= f ( x) e ix dx = fb().
Rn

The fact that we can switch the order of the integral and the limit follows from
the Lebesgue dominated convergence theorem, since | f ( x) e ixk | = | f ( x)| L1 (Rn )
for every k = 1, 2, . . . .

(5)

1 x ix 1
Z Z
f a () =
c f e dx = f ( y) e ia y a n d y = fb(a).
Rn an a Rn an

Here we made a change of variables y = ax , which implies that dx = a n d y.


(6) By a change of variable z = x + y, we have
Z
gb() = f( x + y)() = f ( x + y) e ix dx
Rn
Z Z
i ( z y)
= f ( z) e dz = f ( z) e iz+ i y dz
Rn Rn
Z
= e i y f ( z) e iz dz = e i y fb().
Rn

(7)
Z Z
ix f ( x)() =
e e ix f ( x) e ix dx = f ( x) e ix ix dx
Rn Rn
Z
= f ( x) e ix() dx = fb( ).
Rn

3.3 The Fourier transform and differentia-


tion*
The following proposition shows how the Fourier transform interacts with deriva-
tives. This turns out to be extremely useful as in many cases the Fourier transform
transforms a PDE to an ODE.
CHAPTER 3. FOURIER TRANSFORM AND PDES 68

f
Theorem 3.5. Let f C 1 (Rn ) and assume that x j
L1 (Rn ), j = 1, 2, . . . , n. We
also assume that lim| x| f ( x) = 0. Then

f
d
() = i j fb(), j = 1, 2, . . . , n.
x j

T HE MORAL : Differentiation is multiplication on the Fourier side. This is


the main reason why the Fourier transform is useful in the PDE theory. The
smoothness of f is reflected to the decay of fb() as || . Indeed, if f is very
smooth and several integrations by parts are allowed, then the theorem above
shows that powers of || multiplied by fb are bounded functions, because the
Fourier transform on the left-hand side is bounded.

Proof. Since lim| x| f ( x) = 0, an integration by parts gives

f f
Z
( x) e ix dx
d
() =
x j R n xj
f
Z Z
= ( x) e ix dx j dx1 . . . dx j1 dx j+1 . . . dxn
Rn1 R x j
Z a !
ix

= lim f ( x) e dx1 . . . dx j1 dx j+1 . . . dxn
Rn1 a x j =a


Z Z
f ( x) ( e ix ) dx j dx1 . . . dx j1 dx j+1 . . . dxn
Rn1 R x j
Z Z
= f ( x)( i j ) e ix dx = i j f ( x) e ix dx = i j fb().
Rn Rn

Remarks 3.6:
(1) In particular, the previous theorem applies to a compactly supported
smooth function f C 0 (Rn ). Recall, that a function is compactly supported,
if it is zero outside a compact (closed and bounded) set.
(2) A vector of the form = (1 , . . . , n ), where each component j is a non-
negative integer, is called a multi-index of order || = 1 + + n . For a
multi-index , we define

|| u
D u ( x) = ( x)
x1 1 . . . x
n
n

Let = (1 , . . . , n ) Cn and denote = 1 1 . . . n n . The function e : Rn
C,
e ( x) = e x = e x1 1 ++ xn n ,

belongs to C (Rn ) and


D e = e .

Consequently, every linear partial differential operator with constant


coefficients
a D : C 0 (Rn ) C 0 (Rn )
X
P = P (D ) =
|| k
CHAPTER 3. FOURIER TRANSFORM AND PDES 69

we have
a e .
X
P (D ) e =
|| k

In other words, e is an eigenvector corresponding the eigenvalue for


every operator P (D ). Note that the PDE related to the operator P is

a D u = 0.
X
P (D ) u =
|| k

The idea behind the Fourier transform is that the partial differential
operator P can be better understood if the functions on which they act
are represented as linear combinations of the eigenvectors. Observe that
the set of eigenvalues is the whole Cn and for this reason it is natural to
replace the linear combinations by integrals over . Indeed, P acts as a
scalar multiplication on each eigenvector and the scalar is the eigenvalue
corresponding the eigenvector. In practice this means that it is better to
consider the operator P on the Fourier side.

The following result deals with the case that the derivatives are on the Fourier
side.

Theorem 3.7. Suppose that f L1 (Rn ) is such that the function ix j f ( x)


L1 (Rn ). Then
fb
( ) =
ix j f ( x)(), j = 1, 2, . . . , n.
j

Proof. By the chain rule



( e ix ) = ie ix ( x )
j j

= ie ix ( x1 1 + x j j + xn n ) = ix j e ix .
j

Since ix j f ( x) L1 (Rn ) we can calculate the Fourier transform as


Z
ix j f ( x)() = ix j f ( x) e ix dx
Rn
ix
Z
= f ( x) (e ) dx
R n j

fb
Z
= f ( x) e ix dx = ().
j Rn j

The order of the limit and the integral can be changed by the Lebesgue dominated
convergence theorem.

T H E M O R A L : The decay of f ( x) as | x| is reflected to the smoothness


of fb. Indeed, if f decays rapidly as | x| , then large powers of | x| multiplied
by f are integrable and an iteration of the theorem above shows that fb can be
differentiated several times.
CHAPTER 3. FOURIER TRANSFORM AND PDES 70

Example 3.8. Let f : R R,



1, | x| < a,
f ( x) = 1(a,a) ( x) =
0, | x| a,

that is, f is the indicator function of the interval (a, a). Then the Fourier
transform of f is
a ix
a e e ia e ia 2 sin(a)
Z
fb() = f ( x) e ix dx = = = , 6= 0.
a a i i

Moreover Z
fb(0) = 1(a,a) ( x) dx = 2a.
R

Observe that lim|| fb() = 0 and fb L2 (R), but fb L1 (R). The function f has
a compact support and thus fb C (R), that is, it has derivatives of every order.
This is in accordance with the previous theorem.

1.6

1.2

0.8

0.4

-5 -4 -3 -2 -1 0 1 2 3 4 5

-0.4

Figure 3.1: The graph of the Fourier transform 1


(1,1) .

W A R N I N G : This example shows that f L1 (R) does not imply that fb L1 (R).

Example 3.9. Let f : R R,



e x , x > 0,
f ( x) =
0, x 0,

and g : R R,
xe x , x > 0,
g ( x) =
0, x 0.
CHAPTER 3. FOURIER TRANSFORM AND PDES 71

Then

a x(1+ i)
e
Z
x(1+ i )
fb() = e dx = lim
0 a
0 1 + i
ea(1+ i) 1 1
= lim = .
a 1 + i 1 + i

T HE MORAL : The Fourier transform of a real valued function may be a


complex valued function.

On the other hand,

1 1 fb
gb() =
x f ( x)() = ix f ( x)() = ()
i i
1 1 1 i 1

= = = .
i 1 + i i (1 + i ) 2 (1 + i )2

Here we used Theorem 3.7, but a direct computation works as well.

3.4 The Fourier transform of the Gaussian*


Next we calculate the Fourier transform of the Gaussian function. This will be
very useful for us later.
2
Example 3.10. Let f : R R, f ( x) = e x . Then
2
f 0 ( x) = 2 xe x = 2 x f ( x) L1 (R)

and
fc0 () = 2
x f ( x)() = 2 i
ix f ( x)().

Theorem 3.7 and Theorem 3.5 imply

fb 1 c0 1
() =
ix f ( x)() = f () = i fb() = fb().
2i 2i 2

Thus fb satisfies the ODE

fb b b 2
2
( ) + f () = 0 f () e /4 = 0 fb() = ce /4
2

for some constant c. In order to specify the constant, we observe that


Z
2
c = f (0) = e x dx.
b
R

p
Z
2
Claim: e x dx = .
R
CHAPTER 3. FOURIER TRANSFORM AND PDES 72

Reason. An integration in the polar coordinates gives


Z 2 Z Z Z
2 2 2 2 2
e x dx = e x dx e y d y = e( x + y ) dx d y
R R R R 2
Z Z Z Z
r2 2
= e dS dr = er 1 dS dr
0 B(0,r ) 0 B(0,r )
| {z }
=2 r

a
1 r2
Z
r2
= 2 re dr = 2 lim e = .
0 a 2
0
p
Thus c = fb(0) = and
p 2 /4
fb() = e .

Remark 3.11. By Lemma 3.4 (5), we have

x p p

f p () = 2 fb( 2),
2
which implies that
p 2
x2 /2 () = 2 e /2 .
e
2 /2
This shows that g( x) = e x
is an eigenfunction for the Fourier transform corre-
p
sponding to the eigenvalue 2, that is,
p
gb = 2 g.

Then we consider a higher dimensional version of the previous example.


2
Example 3.12. Let f : Rn R, f ( x) = e| x| . Then
Z Z Z Pn
2 ( x2 ix )
fb() = e| x| e ix dx = e x x ix dx = e j=1 j j j dx
R n R n R n
Z Y n Z Z Y n
x2j ix j j x2j ix j j
= e dx = e dx1 dxn
Rn j =1 R R j =1
n Z n Z
Y x2j ix j j Y x2j ix j j
= e dx j = e e dx j
j =1 R j =1 R
n n p p
x2 2 /4 n 2 /4 2
e j = ( )n e j=1 j = n/2 e|| /4
Y Y P
= e j ( j ) =
j =1 j =1

by the one-dimensional result in the previous example.

3.5 The Fourier inversion formula*


The problem we consider in this section is the following: If we are given fb can
we determine f ? The Fourier inversion theorem will state that, under certain
assumptions, we have Z
f ( x) = (2)n fb() e ix d .
Rn
CHAPTER 3. FOURIER TRANSFORM AND PDES 73

Note a beautiful analogy to the definition of the Fourier transform. This is a


deep result and it is instructive to see what happens if we try to prove directly by
substituting the formula for fb() into the integral above. If we do this, we have
Z Z Z
fb() e ix d = f ( y) e i y d y e ix d
Rn Rn Rn
Z Z
i ( x y)
= f ( y) e d d y.
Rn Rn
| {z }
=?

This does not work out, because the inner integral does not exist, that is, the
function is not integrable. Thus we have to choose another approach.

Theorem 3.13 (Fourier inversion theorem). Let f L1 (Rn ) be a bounded and


continuous function and assume that fb L1 (Rn ). Then
Z
f ( x) = (2)n
fb() e ix d
Rn
n
for all x R .

T HE MORAL : A function can be recovered from its Fourier transform. This


corresponds to the Fourier series representation of a periodic function.
2
Proof. Step 1: Let K : Rn R, K ( x) = n/2 e| x| and
1 x 2 2
K a ( x) = n K = n/2 an e| x| /a , a > 0.
a a
Then Z
b (0) = n/2 n/2 = 1,
K ( x) dx = K
Rn
b () = e||2 /4 by Example 3.12.
since K
R
Claim: Rn K a ( x) dx = 1 for all a > 0.

Reason. By the change of variables y = ax , dx = a n d y, we have


1 x 1
Z Z Z
K a ( x) dx = n
K dx = n
K ( y)a n d y = 1.
R n R n a a R n a
R
Claim: For every r > 0, we have | x|r K a ( x) dx 0 as a 0.
Reason.
1 x 1
Z Z Z
K a ( x) dx = K dx = K ( y) a n d y 0 ,
| x| r | x| r an a | y| r /a an
as a 0. Observe that
1
Z Z
n
lim K ( y ) a d y = lim K ( y)1| y|r/a ( y) d y
a0 | y| r /a a n a 0 R n
Z
= K ( y) lim 1| y|r/a ( y) d y = 0.
Rn a 0

Here the order of the limes and the integral can be switched by the Lebesgue
dominated convergence theorem.
CHAPTER 3. FOURIER TRANSFORM AND PDES 74

Figure 3.2: Kernel K a ( x) with different values of parameter a.

R
Claim: Rn f ( x)K a ( x) dx f (0) as a 0.

Reason. Since f C (Rn ), for every > 0 there exists r > 0 such that

| f ( x) f (0)| <
2
whenever | x 0| < r . This implies


Z Z Z

f ( x)K a ( x) dx f (0) = f ( x)K a ( x) dx f (0) K a ( x) dx
Rn Rn | R {z
n

}
=1
Z

= K a ( x)( f ( x) f (0)) dx
ZR
n
Z
K a ( x)| f ( x) f (0)| dx + K a ( x)| f ( x) f (0)| dx
| x|< r | x| r

Z Z
K a ( x) dx + 2 sup | f ( x)| K a ( x) dx <
2 Rn n | x| r
| {z } | xR {z }
=1 < 2

by choosing a > 0 small enough. Here we used the previous claim.


CHAPTER 3. FOURIER TRANSFORM AND PDES 75

Step 2: By changing the order of integration


Z Z Z
fb() g() d = g ( ) f ( x) e ix dx d
Rn
ZR Z Rn
n

= f ( x) g() e ix d dx
Rn Rn
Z Z Z
= f ( x) g() e i x d dx = f ( x) gb( x) dx.
Rn Rn Rn

Let g a : Rn Rn ,
2 /4
g a ( x) = (2)n e|ax| , a > 0.

Then
n
2 2 2

ca () = (2)n e
g |ax/2|2 () = (2) n e| x|
a a
n
2 2 2
= (2)n n/2 e|2/a| /4 = n/2 an e|/a| = K a ().
a
In the second equality we used Lemma 3.4 (5) and in the third equality Example
3.12.
Step 3:
Z
f (0) = lim f ( x)K a ( x) dx (Step 1)
a 0 R n
Z
= lim f ( x) gba ( x) dx (Step 2)
a 0 R n
Z
= lim fb() g a () d (Step 2)
a 0 R n
Z
= fb() lim g a () d
Rn a 0
| {z }
=(2)n
Z
= (2)n fb() d .
Rn

This proves the claim for x = 0. The general case follows by denoting F ( y) = f ( x + y).
Then
Z
f ( x) = F (0) = (2)n Fb() d
Rn
Z
= (2)n f( x + y)() d
ZR
n

= (2)n fb() e ix d .
Rn

The last equality follows from Lemma 3.4 (6).

Remark 3.14. The previous theorem holds true under the substantially weaker
hypothesis that f L1 (Rn ) and fb L1 (Rn ). This is not so surprising since the
Fourier inversion formula can be written as
Z
f ( x) = (2)n fb() e i( x) d = (2)n fb
b( x).
Rn
CHAPTER 3. FOURIER TRANSFORM AND PDES 76

Thus the function f is given by a Fourier transform of the function fb L1 (Rn ) and
it is automatically continuous and bounded.

As a corollary we get a uniqueness result for Fourier transforms. Namely, if


two functions f , g L1 (Rn ) have the same Fourier transform, then they must be
the same function.

Theorem 3.15. Let f , g L1 (Rn ) be continuous functions. If fb() = gb() for every
Rn then f ( x) = g( x) for every x Rn .

THE MORAL : A function is uniquely defined by its Fourier transform.

Proof. We note that f


g = fb g g L1 (Rn ). By the Fourier
b = 0, which implies f
inversion theorem, see Theorem 3.13, we have
Z
f ( x) g( x) = (2)n (f g)() e ix d
Rn
Z
= (2)n ( fb() gb()) e ix d = 0
Rn | {z }
=0

n
for every x R and we are done.

3.6 The Fourier transformation and con-


volution
Let f , g : Rn C. The convolution f g : Rn C is defined by
Z
( f g)( x) = f ( x y) g( y) d y,
Rn

whenever this integral exists.


T HE MORAL : The convolution on Rn has a similar role in representation
formulas for solutions of PDEs as in the one-dimensional case.
This operation is commutative and associative. To prove the commutativity,
fix x Rn and make the change of variables z = x y, which implies y = x z and
d y = dz. This implies
Z Z
( f g)( x) = f ( x y) g( y) d y = f ( z) g( x z) dz = ( g f )( x).
Rn Rn

We leave it as an exercise to show that ( f g) h = f ( g h).


Then we consider the question, that for which functions the definition makes
CHAPTER 3. FOURIER TRANSFORM AND PDES 77

sense. Assume that f , g L1 (Rn ). Then


Z Z Z

|( f g)( x)| dx = n f ( x y) g( y) d y dx

R n R n R
Z Z
| g( y)| | f ( x y)| dx d y
Rn Rn
Z Z
= | g( y)| | f ( z)| dz d y
ZR Rn
n
Z
= | f ( z)| dz | g( y)| d y,
Rn Rn

from which it follows that

k f gkL1 (Rn ) k f kL1 (Rn ) k gkL1 (Rn ) < .

In particular, this implies that f g L1 (Rn ).

WA RNING : Note that this is not obvious. In general, a product of two inte-
grable function is not necessarily integrable.

Theorem 3.16. Assume that f , g L1 (Rn ). Then

g() = fb() g
f b()

for every Rn .

THE MORAL : Convolution becomes the standard multiplication on the Fourier


side. This will be very useful when we derive representation formulas for solutions
to PDE and this is one motivation for the definition of a convolution.

Proof. Since
Z
( f g)( x) e ix = f ( y) g( x y) e ix d y
Rn
Z
= f ( y) e i y g( x y) e i( x y) d y,
Rn

we have
Z
g() =
f ( f g)( x) e ix dx
Rn
Z Z
= f ( y) e i y g( x y) e i( x y) d y dx
Rn Rn
Z Z
= f ( y) e i y g( x y) e i( x y) dx d y
Rn Rn
Z Z
i y iz
= f ( y) e g( z) e dz d y = fb() gb().
Rn Rn
| {z }
b()
=g

Next we present a corresponding result for the Fourier transform of the product.
This can be seen as a dual result of Theorem 3.16.
CHAPTER 3. FOURIER TRANSFORM AND PDES 78

Theorem 3.17. Assume that f , g, fb, gb L1 (Rn ). Then

fcg() = (2)n ( fb gb)()

for every Rn .

THE MORAL : Product becomes convolution on the Fourier side.

Proof. By the Fourier inversion theorem


Z
fcg() = f ( x) g( x) e ix dx
Rn
Z Z
= (2)n fb() e ix d g( x) e ix dx
Rn
Z Z R
n

= (2)n fb() g( x) e ix() dx d


Rn Rn
Z Z
= (2)n fb() g( x) e ix() dx d
Rn Rn
| {z }
b()
=g
n
= (2) ( fb gb)().

3.7 Plancherels formula*


Assume that f L1 (Rn ) is such a function that fb L1 (Rn ) and denote f # ( x) = f ( x).
Then
f # () = fb().
c

Define g( x) = ( f f # )( x). By Theorem 3.16 we have

gb() = f f # () = fb() fb() = | fb()|2 .


f # () = fb() c

On the other hand


Z Z Z
g(0) = f ( y) f # (0 y) d y = f ( y) f ( y) d y = | f ( y)|2 d y
Rn Rn Rn

and the Fourier inversion formula implies


Z Z Z
| f ( y)|2 d y = g(0) = (2)n gb() d = (2)n | fb()|2 d
Rn Rn Rn

T H E M O R A L : The L2 (R n ) norm of the Fourier transform is the same as the


L2 (Rn ) norm of the function up to a multiplicative constant. This can be used to
define the Fourier transform of a function an L2 (Rn ) function, but this is out of the
scope of this course. The factor (2)n appears with our definition for the Fourier
transform, but there are different scalings in the literature.
CHAPTER 3. FOURIER TRANSFORM AND PDES 79

3.8 Approximations of the identity in Rn *


In this section we study approximations of the identity in the higher dimensional
case.

Definition 3.18. Let {K }>0 be a family of functions K : Rn C, with the follow-


ing properties.

(1) For every > 0 we have


R
Rn K ( x) dx = 1.
(2) There exists some constant M > 0 such that, for every > 0 we have
Z
|K ( x)| dx M.
Rn

(3) For every > 0, we have


Z
lim |K ( x)| dx = 0.
0 | x|>

Then the family {K }>0 will be called a family of good kernels.

These families of good kernels play the same role as the good kernels in the
2-periodic case, see Section 2.14. In fact, we shall consider the analogous theory
in the higher dimensional case.
The importance of good kernels is contained in the following result.

Theorem 3.19. Let f L1 (Rn ) be bounded and continuous at x Rn and {K }>0


be a family of good kernels. Then

lim(K f )( x) = f ( x).
0

If f C 0 (Rn ) (a compactly supported continuous function) then K f f uni-


formly as 0.

THE MORAL : Approximations of the identity give an interpretation that the


convolution K f can be seen as a weighted integral average. The pointwise
value of a function is replaced with an integral average, which converges to the
value of the function as 0.

Proof. The proof is similar to the one in the 2-periodic case. Let > 0. By the
continuity of f at x there exists > 0 such that

| f ( x y) f ( x)| < when | y| < , (3.1)
2M
where M is from property (2) of a good kernel.
CHAPTER 3. FOURIER TRANSFORM AND PDES 80

For x Rn we use property (1) of the family of good kernels to have


Z Z

|(K f )( x) f ( x)| = K ( y) f ( x y) d y f ( x ) K ( y) d y
ZR R
n n

| f ( x y) f ( x)||K ( y)| d y
| y|<
Z
+ | f ( x y) f ( x)||K ( y)| d y
| y|>

= I1 + I2.

For I 1 we use (3.1) to obtain



Z
|I1| |K ( y)| d y .
2M Rn 2

For I 2 we choose small enough so that


Z
|K ( y)| d y < ,
| y|> 4k f kL (Rn )

which is possible by property (3) of the good kernel. Observe that since f is
bounded we have k f kL (Rn ) < . On the other hand, we may assume that
k f kL (Rn ) > 0, otherwise f = 0 and the claim is clear. Thus


Z
| I 2 | 2k f kL (Rn ) |K ( y)| d y < .
| y|> 2

Summing up the estimates for I 1 and I 2 we obtain

|(K f )( x) f ( x)| <

when > 0 is small enough. This proves the desired convergence at x.


If f is continuous with compact support it is bounded. Furthermore, it is
vanishing outside a compact set so it is uniformly continuous. Thus the in (3.1)
can be chosen to be the same for all x Rn which shows that the convergence is
uniform in this case.
2
Example 3.20. Let K : Rn R, K ( x) = e| x| be the Gaussian function and

1 x 2 2
K a ( x) = n
K = n/2 an e| x| /a , a > 0.
a a
Then the proof of the Fourier inversion theorem 3.13 shows that {K a }a>0 is a
family of good kernels.

Example 3.21. For t > 0, let H t : Rn R,

1 2 /(4 t)
H t ( x) = e| x| .
(4 t)n/2
CHAPTER 3. FOURIER TRANSFORM AND PDES 81

Then H t is called the heat kernel for the upper half plane and, as we shall see, it is
related to the time dependent heat equation in the upper half plane. Furthermore,
H t can be given as

1 x

H t ( x) = p4 t ( x) = p K p
( 4 t)n 4 t
1 2 1 2
= p e| x| /(4 t) = e| x| /(4 t) .
( 4 t) n (4 t ) n/2

2
where K ( x) = e| x| . Then { H t } t>0 is a family of good kernels as in Example 3.20.

Example 3.22. For r > 0, let K r : Rn R,

1B(0,r) ( x)
K r ( x) = ,
|B(0, r )|

where B( x, r ) = { y Rn : | x y| < r } is the open ball with the center x Rn and


radius r > 0. |B( x, r )| denotes the n-dimensional volume of the ball. Then {K r }r>0
is a family of good kernels and

1
Z
(K r f )( x) = f ( y)1B(0,r) ( x y) d y
|B(0, r )| Rn
1
Z
= f ( y)1B( x,r) ( y) d y
|B( x, r )| Rn
1
Z
= f ( y) d y
|B( x, r )| B( x,r)

is the integral average of f over the ball B( x, r ). Here we also used the fact that
the volume of a ball is independent of the location. The previous theorem tells
that
1
Z
lim f ( y) d y = f ( x )
r 0 |B( x, r )| B( x,r )

if f is continuous. This means that the pointwise value f ( x) is a limit of the


integral averages over balls centered at x as the radius tends to zero. It is also
possible to show this directly form the definition of continuity.

THE MORAL : Approximations of the identity can be constructed by using one


function by rescaling.

3.9 The Laplace equation in the upper


half-space
In this section we consider the Laplace equation in the upper half-space

n+1
R+ = {( x, y) : x Rn , y > 0}.
CHAPTER 3. FOURIER TRANSFORM AND PDES 82

We assume that the boundary function g is a continuous and bounded function


n+1
defined on R+ = Rn {0} and consider the following Dirichlet problem
n 2 u 2 u

u( x, y) = n+1
( x, y) + 2 ( x, y) = 0, ( x, y) R+
X
,


j =1 x 2
j
y (3.2)

u( x, 0) = g( x), x Rn .

Observe that here the Laplacian is taken with respect to all n + 1 variables.
Physically this models the case when the temperature does not change in time and
that the temperature u( x, 0) at the boundary is given by the function g( x). We look
n+1 n+1
for solutions u C 2 (R+ ) C (R+ ) of this problem. This means that u is twice
continuously differentiable (all partial derivatives of u of second order exist and
are continuous) and that u is continuous up to the boundary. The differentiability
condition guarantees that the partial derivatives in the Laplace operator make
sense and the continuity up to the boundary is needed for the boundary condition.

Figure 3.3: The Laplace equation in the upper half-space.

Step 1 (PDE on the Fourier side): To solve this problem let y > 0 be fixed
and denote Z
b(, y) =
u u( x, y) e ix dx, Rn , y > 0.
Rn
Observe that we take the Fourier transform of u in the x variable only. By Theorem
3.5, we obtain

2 u
u
d
( , y) = i j (, y) = ( i j )2 u
b(, y) = 2j u
b(, y), j = 1, . . . , n,
2
x j x j
CHAPTER 3. FOURIER TRANSFORM AND PDES 83

and, by switching the order of differentiation and integration, we have

2 u 2 u
Z
( x, y) e ix dx

( , y ) =
y2 Rn y
2

2 2 u
Z
ix
(, y).
b
= 2 u( x, y) e dx =
y R n y2

It follows that

2 u
cu(, y) = 2 u
2
1 b (, y) . . . . n u
b(, y) + ( , y)
b
y2
2 u
= ||2 u
b(, y) + (, y) = 0.
b
y2

This is the Laplace equation on the Fourier side. Note that by the Fourier inversion
theorem 3.13, we have u = 0 cu = 0. Compare this to Remark 2.47 for the
Fourier series.

THE MORAL : The Laplace equation becomes an ODE on the Fourier side.

Step 2 (Solution on the Fourier side): For a fixed the solutions of this
ODE for u
b are of the form

b(, y) = c 1 () e|| y + c 2 () e|| y .


u

Note here that, as is fixed and we solve the ODE in y, the constants c 1 and c 2
may depend on Rn . We disregard the first term on the right hand side, since
e|| y as || . This corresponds to the physically irrelevant unbounded
solution and we are left with

b(, y) = c 2 () e|| y .
u

The boundary condition on the Fourier side implies

c 2 ( ) = u
b(, 0) = g
b()

and thus
b() e|| y ,
b(, y) = g
u Rn , y > 0.

This is the candidate for a solution of the Dirichlet problem on the Fourier side.
Step 3 (Solution to the original problem): By the Fourier inversion theo-
rem 3.13
Z
u( x, y) = (2) n
b(, y) e ix d
u
Rn
Z
= (2)n e|| y gb() e ix d
Rn
Z
= (2)n b() e ix d ,
cy () g
P
Rn
CHAPTER 3. FOURIER TRANSFORM AND PDES 84

cy () = e|| y . The function P y ( x) is called the Poisson kernel for the upper
where P
half-space. Observe that at this point we only know the Fourier transform of
Poisson kernel. By Theorem 3.16, we have
Z
u( x, y) = (2)n P b() e ix d
cy () g
Rn
Z
= (2)n P y g() e ix d

Rn
Z
= (P y g)( x) = P y ( x z) g( z) dz.
Rn

Here we used the Fourier inversion formula again.

THE MORAL : The solution of the Dirichlet problem in the upper half-space is
a convolution of the boundary function with a Poisson kernel.

Step 4 (Explicit representation formula): We thus need to find, for every


cy () = e y|| . Indeed, it is possible to find
y > 0, a function P y ( x), x Rn , such that P
an explicit formula for the function P y will be the Poisson kernel for the upper
half-space. We give the following result without a proof.

Lemma 3.23. For x Rn and y > 0 we define

( n+1
2 ) y
P y ( x) = P ( x, y) = n+1 n+1
, x Rn , y > 0.
2 (| x|2 + y2 ) 2

Here ( n+1
2 ) is a dimensional constant given by the -function
Z
( s) = x s1 e x dx.
0

The Poisson kernel has the following properties.

(1) P y ( x) > 0 for every x Rn and y > 0.


(2) By the Fourier inversion theorem
Z
P y ( x) = (2)n e y|| e ix d x Rn , y > 0.
Rn

cy () = e y|| , we have
(3) Since P
Z
P y ( x) e ix dx = e y|| , Rn , y > 0.
Rn

(4) For every y > 0 we have Z


P y ( x) dx = 1.
Rn
(5) The Poisson kernel P ( x, y) = P y ( x) is a solution of the Laplace equation in
n+1
the upper half-space R+ . It is called the fundamental solution, since all
other solutions can be represented as a convolution with it.
CHAPTER 3. FOURIER TRANSFORM AND PDES 85

1.25

0.75

0.5

0.25

-5 -4 -3 -2 -1 0 1 2 3 4 5

Figure 3.4: The graph of the Poisson kernel in dimension n = 1 for y = 1 (yellow),
y = 0.5 (red) and y = 0.3 (blue).

Remark 3.24. Denote P ( x) = P1 ( x). Then

1 x

P y ( x) = n P , x Rn , y > 0.
y y

This means that the Poisson kernel P y ( x) with y > 0 can be obtained from P1 ( y) by
rescaling, compare to Example 3.20. This formula has the following consequences.
First, by the change of variables z = xy , we have

1 x
Z Z
P y ( x) dx = P dx
Rn yn Rn y
1
Z Z
= n P ( z) yn dz = P ( x) dx = 1, y > 0.
y Rn Rn

Thus one can think of the dimensional constant appearing in front of the Poisson
kernel as the constant c n such that

1
Z
cn n+1
dx = 1.
Rn (| x|2 + 1) 2

Second, from this we can see that {P y } y>0 is a family of good kernels. Indeed, all
the properties of a family of good kernels in Definition 3.18 are satisfied except
maybe the third. Thus, for every > 0 we have to show that
Z
lim |P y ( x)| dx = 0.
y0+ | x|>
CHAPTER 3. FOURIER TRANSFORM AND PDES 86

By the change of variables z = xy , we have

1 x
Z Z
|P y ( x)| dx = P dx
| x|> yn | x|> y
Z Z
= P ( z) dz = 1| x|>/ y ( x)P ( x) dx.
| z|>/ y Rn

Now observe that for every x Rn we have

lim 1| x|>/ y ( x)P ( x) = 0


y0+

because / y . Furthermore, for every y > 0 and x Rn we have

|1| x|>/ y ( x)P ( x)| |P ( x)| L1 (Rn ).

Thus the Lebesgue dominated convergence theorem gives


Z Z
lim |P y ( x)| dx = lim 1| x|>/ y ( x)P ( x) dx = 0,
y0+ | x|> Rn y0+

which shows the third property of good kernels.

By the previous remark we immediately get the following result.

Theorem 3.25. Let g C 0 (Rn ). The solution to the Dirichlet problem (3.2) is

( n+1
2 ) g( z) y
Z
u( x, y) = ( g P y )( x) = n+1 n+1
dz.
2 Rn (| x z|2 + y2 ) 2

The boundary condition is taken in the sense that

lim u( x, y) = g( x) for every x Rn .


y0+

W A R N I N G : The boundary condition cannot be verified by inserting directly


y = 0 to the formula above. Thus the limit interpretation with the approximation
of the identity is needed for the boundary values.

3.10 The heat equation in the upper half-


space
The general form of the heat equation is
u
( x, t) u( x, t) = 0,
t
where the Laplace operator is only in the x-variable
n 2 u
u( x, t) =
X
( x, t).
j =1 x2j
CHAPTER 3. FOURIER TRANSFORM AND PDES 87

We consider the initial value problem


u
( x, t) u( x, t) = 0, x Rn , t > 0,
t (3.3)
u( x, 0) = g( x), x Rn .

This is called the Cauchy problem for the heat equation. Physically this models
the case, when the initial temperature u( x, 0) at the moment t = 0 is given by the
function g( x) and we would like to know the temperature u( x, t) for t > 0. Here we
n+1 n+1
look for solutions u C 2 (R+ ) C (R+ ). This means that u has continuous second
order partial derivatives in the upper half-space Rn+1 and that u is continuous up
to the initial boundary Rn { y = 0}.

Figure 3.5: The heat equation in the upper half-space.

b(, t)
Step 1 (PDE on the Fourier side): Let t > 0 be fixed and denote by u
the Fourier transform of u( x, t) in the x-variable
Z
b(, t) =
u u( x, t) e ix dx.
Rn

On the Fourier side, the heat equation becomes

u u
d u
u(, t) = (, t)
cu(, t) = (, t) + ||2 u
b(, t) = 0.
b
t t t
This can be seen in the same way as for the Laplace equation. Compare this to
Remark 2.47 for the Fourier series.
CHAPTER 3. FOURIER TRANSFORM AND PDES 88

Step 2 (Solution on the Fourier side): The solution of this ODE for a fixed
Rn is
2
b(, t) = c() e|| t .
u

The initial condition u( x, 0) = g( x) gives

c() = u
b(, 0) = g
b()

and thus
2
b() e|| t .
b(, t) = g
u

This is the solution on the Fourier side.


Step 3 (Solution to the original problem): By the Fourier inversion theo-
rem
Z
u( x, t) = (2) n
b(, t) e ix d
u
Rn
Z
2
= (2)n e|| t gb() e ix d
Rn
Z
= (2)n b() e ix d ,
ct () g
H
Rn

ct () = e||2 t . Observe that H


where H ct is a Gaussian function. By Theorem 3.16,
we have
Z
u( x, t) = (2)n b() e ix d
ct () g
H
Rn
Z
= (2)n H t g() e ix d

Rn
Z
= ( H t g)( x, t) = H t ( x y) g( y) d y.
Rn

Here we used the Fourier inversion formula again.


Step 4 (Explicit representation formula): We claim that if

1 2 /(4 t)
H t ( x) = e| x| , x Rn , t > 0,
(4 t)n/2

ct () = e||2 t . The function H t is called the heat kernel in the upper half-
then H
space. Observe that it is a Gaussian function.

Reason.

ct () =
1 | x|2 /(4 t) () =
1 p
2
H e e(| x|/(2 t)) ()

(4 t)n/2 (4 t)n/2
1 p p
(2 t)n
2
= e| x| (2 t) (Lemma 3.4 (5))
(4 t)n/2
p
n/2 n/2 |2 t|2 /4 2
= e = e|| t . (Example 3.12)
CHAPTER 3. FOURIER TRANSFORM AND PDES 89

THE MORAL : The solution of the initial value problem for the heat equation
in the upper half-space is a convolution of the initial value function with the heat
kernel.

The family { H t } t>0 is a family of good kernels related to approximations of the


identity, see Example 3.21. Moreover, the heat kernel H ( x, t) = H t ( x) is a solution
n+1
of the heat equation in the upper half-space R+ . It is called the fundamental
solution, since all other solutions can be represented as a convolution with it.

Theorem 3.26. Assume that g C 0 (Rn ). The solution to the Cauchy problem
(3.3) is
1
Z
2 /(4 t)
u( x, t) = ( H t g)( x) = e| x y| g( y) d y, x Rn , t > 0.
(4 t)n/2 Rn

The initial values are attained in the sense

lim u( x, t) = g( x) for every x Rn .


t0+

W A R N I N G : The initial condition cannot be verified by inserting directly t = 0


to the formula above. Thus the limit interpretation with the approximation of the
identity is needed for the initial values.

Remark 3.27. We consider the nonhomogeneous initial value problem


u
( x, t) u( x, t) = f ( x), x Rn , t > 0,
t (3.4)
u( x, 0) = g( x), x Rn .

On the Fourier side, the equation becomes

u
(, t) + ||2 u
b(, t) = fb()
b
t
2t
and after a multiplication through by e|| we have

2t u 2 2
e || (, t) + e|| t ||2 u
b(, t) = e|| t fb().
b
t
For the left-hand side, we observe, that

2t u 2 2

e | | (, t) + e|| t ||2 u
b(, t) = e | | t u
b(, t) .
b
t t
Thus by the fundamental theorem of calculus, we have

2
Z t 2
e || t u
b(, t) + u
b(, 0) = e|| s fb(, s) ds
0

and so Z t
2 2 ( t s)
b(, t) = e|| t g
u b() + e|| fb(, s) ds.
0
CHAPTER 3. FOURIER TRANSFORM AND PDES 90

By the Fourier inversion theorem


Z
u( x, t) = (2)n fb(, t) e ix d
Rn
Z Z Z t
2 2
= (2)n e|| t gb() e ix d + (2)n e|| ( ts) fb(, s) ds e ix d
ZR Rn
n 0
n ix
= (2) ct () g
H b() e d ,
Rn
1
Z
2
= e| x y| /(4 t) g( y) d y
(4 t)n/2 Rn
Z t
1
Z
2
+ e| x y| /(4( ts)) f ( y, s) d y ds
0 (4( t s))
n /2
R n
Z t
= ( H t g)( x) + ( H ts f (, s))( x) ds.
0

Thus the solution of the inhomogeneous problem can be represented using the
solutions of the homogeneous problem given by Theorem 3.26. This is so-called
Duhamels principle. We shall return to this later in Section 5.3.

Example 3.28. Let us consider the initial value problem for the Schrdinger equa-
tion u
i ( x, t) + u( x, t) = 0, x Rn , t > 0,
t
u( x, 0) = g( x), x Rn .

Here u and f are complex valued functions. If v( x, t) = u( x, it), where u is as in


the representation formula above, we have
1
Z
2
v( x, t) = e i| x y| /(4 t) g( y) d y, x Rn , t > 0.
(4 it) n /2
R n

Here
1 1 1 i i
i 2 = e 2 log i = e 2 (log | i|+ i arg i) = e 2 arg i = e 4 .

We can check by a direct calculation, that u is a solution of the Schrdinger


equation above. Indeed, if v( x, t) = u( x, it), then
v u
( x, t) = ( u( x, it)) = i ( x, it) = i u( x, it) = i v( x, t).
t t t
The function
1 2 /(4 t)
( x, t) = e i | x| , x Rn , t 6= 0,
(4 it)n/2
is called the fundamental solution of the Schrdinger equation. Note that the
formula u = f makes sense for all t 6= 0. Thus this gives a solution to the
problem
i u ( x, t) + u( x, t) = 0, x Rn , t R,
t
u( x, 0) = g( x), x Rn .
In particular, the Schrdinger equation is reversible in time, whereas the heat
equation is not. We shall return to this point later.
CHAPTER 3. FOURIER TRANSFORM AND PDES 91

3.11 The wave equation in the upper half-


space
The general form of the wave equation is

2 u
( x, t) u( x, t) = 0,
t2
where the Laplace operator is only in the x-variable
n 2 u
u( x, t) =
X
( x, t).
j =1 x2j

One can think of the wave equation as describing the displacement of a vibrat-
ing string (in dimension n = 1), a vibrating membrane (in dimension n = 2) or an
elastic solid (dimension n = 3). In dimension n = 3 this equation also determines
the behavior of electromagnetic waves in vacuum and the propagation of sound
waves.
The general goal is to find a solution of the Cauchy problem for the wave
equation 2
u u = 0, x Rn , t > 0,

t2

u( x, 0) = g( x), x Rn ,

u ( x, 0) = h( x), x Rn .


t

b(, t)
Step 1 (PDE on the Fourier side): Let t > 0 be fixed and denote by u
the Fourier transform of u( x, t) in the x-variable
Z
b(, t) =
u u( x, t) e ix dx.
Rn

Step 2 (Solution on the Fourier side): On the Fourier side, the wave
equation becomes


2u 2 u
2 u
u(, t) = 2 (, t)
cu(, t) = (, t) + ||2 u
b(, t) = 0.
b
t 2 t t2
The solution of this ODE for a fixed Rn is

b(, t) = c 1 () cos(|| t) + c 2 () sin(|| t)


u

for some functions c 1 () and c 2 (). Taking the Fourier transforms of the initial
conditions we obtain

b() = u (, 0) = || c 2 ().
d
gb() = u
b(, 0) = c 1 () and h
t
Thus the solution on the Fourier side can be written in the form

b(, t) = g
u b() sin(|| t) .
b() cos(|| t) + h
| |
CHAPTER 3. FOURIER TRANSFORM AND PDES 92

Step 3 (Solution to the original problem): By the Fourier inversion for-


mula
sin(|| t) ix
Z
u( x, t) = (2)n gb() cos(|| t) + h
b() e d . (3.5)
Rn | |
Denote
sin(|| t) ct () = cos(|| t) = t ().
c

ct () = and
| | t
Then
Z
u( x, t) = (2)n gb() ct () e ix d
b ( )
ct () + h
ZR
n

= (2)n t () e ix d
g t () + h
(Theorem 3.16)
Rn

= ( g t )( x) + ( h t )( x). (Fourier inversion formula)

Now the problem is how to determine functions t and t and what is the
interpretation of the representation formula above. This is a hard problem and
we shall return to this later. A direct calculation shows that the u given by
the formula above does indeed solve the Cauchy problem for the wave equation.
Furthermore, the solution to the Cauchy problem is unique, but we will not prove
this here.

3.12 Summary
The main steps in the application of the Fourier transform to PDE problems are
the following.

(1) The task is to find a solution of an initial or boundary value problem in


the half-space.
(2) Take the Fourier transform of the PDE and of the initial conditions, with
respect to the space variables.
(3) This reduces the problem to an ODE.
(4) The ordinary ODE is solved on the Fourier side.
(5) The initial or boundary conditions are used to determine the free parame-
ters.
(6) The Fourier inversion formula gives the solution of the original problem.
(7) The solution of the original problem is represented as a convolution of the
data with the fundamental solution.
(8) This gives a solution to the original problem and the initial or boundary
values are attained by using approximations of the unity.
Boundary value problems for the Laplace equation in sub-
domains of the higher dimensional Euclidean space appear
frequently in natural sciences and engineering. We de-
rive representation formulas and study general properties

4
of solutions to the Laplace equation. In this process we
shall encounter fundamental solutions, Greens functions,
mean value property, Harnacks inequality and maximum
principles.

Laplace equation

The n-dimensional Laplace equation

u = 0

and the Poisson equation


u = f

appear frequently in natural sciences and engineering. Recall that


n 2 u
u =
X
.
j =1 x2j

Let be an open subset of Rn . The problem is to find a function u C 2 () such


that it is a solution to the Laplace or Poisson equation in . Physically, solutions
of the Poisson equation correspond to steady states for time evolutions such as
heat flow or wave motion, with f corresponding to external driving forces such as
heat sources or wave generators.

Definition 4.1. A function u C 2 (), which satisfies u = 0 in , is called a


harmonic function in .

We have already seen examples of harmonic functions, see Theorem 3.25. In


this section we take a more systematic approach to the Laplacian.

4.1 Gauss-Green theorem


We shall need certain integral formulas to be able to study the Laplacian. Here
we assume that is a bounded and open subset of Rn and that the boundary
is smooth. This means that the boundary can be locally represented as a graph
of a smooth function. The closure of the domain is a union of the domain and its

93
CHAPTER 4. LAPLACE EQUATION 94

boundary, that is, = . We say that u C 1 (), if u C 1 () is such that u


u
and all partial derivatives x j
, j = 1, . . . , n, can be extended continuously up to the
boundary .

Theorem 4.2 (Gauss-Green theorem). Assume that u C 1 (). Then


u
Z Z
( x) dx = u( x) j ( x) dS ( x), j = 1, . . . , n.
x j

Here ( x) = (1 ( x), . . . , n ( x)) is the outward pointing unit normal vector on ,


u
( x) = u( x) ( x), x ,

is the outward normal derivative of u and
u u

u ( x) = ( x ), . . . , ( x) , x ,
x1 xn
is the gradient of u.

Remark 4.3. Another way to write the Gauss-Green theorem is


Z Z
div F ( x) dx = F ( x) ( x) dS ( x),

where F = (F1 , . . . , F n ) is a vector field, whose component functions satisfy the


assumption in the Gauss-Green theorem. Recall, that
n F
X j
div F ( x) = ( x)
j =1 x j

is the divergence of F .

Reason.
Z n F
Z X n Z F
j X j
div F ( x) dx = ( x) dx = ( x) dx
j =1 x j j =1 x j
n
XZ Z X n
= F j ( x) j ( x) dS ( x) = F j ( x) j ( x) dS ( x)
j =1 j =1
Z
= F ( x) ( x) dS ( x).

T HE MORAL : Gauss-Green theorem gives information about the divergence


of a vector field inside the domain by its values on the boundary of the domain.
More precisely, the integral of the divergence of a vector field over a domain is
equal to the total flow through the boundary. This is useful in boundary value
problems for PDEs.

Theorem 4.4 (Integration by parts). Assume that u, v C 1 (). Then


u v
Z Z Z
( x)v( x) dx = ( x) u( x) dx + u( x)v( x) j ( x) dS ( x), j = 1, . . . , n.
x j x j
CHAPTER 4. LAPLACE EQUATION 95

Proof. Apply the Gauss-Green theorem for uv (exercise).

T HE MORAL : The Gauss-Green theorem can be seen as an integration by


parts formula in Rn .

Theorem 4.5 (Greens identities). Assume that u, v C 1 (). Then


v
Z Z Z
(1) u( x) v( x) dx = u( x)v( x) dx + ( x) u( x) dS ( x),

v u
Z Z
(2) ( u( x)v( x) v( x) u( x)) dx = u( x) ( x) v( x) ( x) dS ( x),

Z
u
Z
(3) u( x) dx = ( x) dS ( x).

These are called Greens first, second and third identities, respectively.

v
Proof. (1) By replacing v with x j
in Theorem 4.4, we have

v u 2 v v
Z Z Z
( x) ( x) dx = u ( x) ( x) dx + u ( x) ( x) j ( x) dS ( x),
x j x j
2
x j x j

j = 1, . . . , n. The claim follows by summing over j = 1, . . . , n.


(2) Switch u and v in (1) and subtract (exercise).
u
(3) By replacing u with x j
and v = 1 in Theorem 4.4, we have

2 u u
Z Z
( x) dx = ( x) j ( x) dS ( x), j = 1, . . . , n.

2
x j x j

The claim follows by summing over j = 1, . . . , n, since


Z n Z 2 u n Z
u
u( x) dx =
X X
( x) dx = ( x) j ( x) dS ( x)

2
j =1 x j
j =1 x j
u
Z Z
= u( x) ( x) dS ( x) = ( x) dS ( x).

Remark 4.6. Greens third identity tells that the integral of the Laplacian of
a function over a domain is equal to the total flow through the boundary. In
particular, if u is harmonic in , then

u
Z
( x) dS ( x) = 0
V

for every subdomain V for which V . This means that the total flow is zero
through the boundary of any subdomain V . Physically this means that there are
no heat sources or electric charges in the domain.
CHAPTER 4. LAPLACE EQUATION 96

4.2 PDEs and physics


In a typical case u is a function that denotes the density of some quantity in
steady state. For example u may denote temperature, chemical concentration or
electrostatic potential. If V is any smooth subdomain of , the total flux through
the boundary V is zero Z
F ( x) ( x) dS ( x) = 0,
V
where F = (F1 , . . . , F ) is the flux density and is the unit outer normal of V .

By the Gauss-Green theorem we have


Z Z
div F ( x) dx = F ( x) ( x) dS ( x) = 0.
V V

Since this holds for every subdomain V of , we have

div F ( x) = 0 for every x .

It is physically reasonable to assume that the flux F is proportional to the gradient


u but in the opposite direction, since the flow is from regions of high temperature
to regions of low temperature or high concentration to low concentration. Thus

F ( x ) = a u ( x ), a > 0.

This gives

div F ( x) = adiv u( x) = a u( x) = 0 for every x ,


CHAPTER 4. LAPLACE EQUATION 97

which implies u = 0 in . A similar argument can be done for the Poisson


equation. In this case, the function f describes the heat source or an electric
charge distribution.
Examples 4.7:
(1) If u is chemical concentration, then u = 0 is Ficks law of diffusion.
(2) If u is temperature, then u = 0 is Fouriers law of heat conduction (steady
state).
(3) If u is electrostatic potential, then u = 0 is Ohms law of electrical con-
duction.
(4) Maxwells equations for the electric field E with a source f are

div E = f ,
curl E = 0.

Now curl E = 0 implies that E = u + c. Thus div E = f implies that

div u = u = f ,

which is the Poisson equation.

4.3 Boundary values and physics


The Dirichlet problem
Assume that is a bounded open set in Rn with smooth boundary . We shall
consider two different boundary value problems. The Dirichlet problem for the
Laplace equation is
u = 0 in ,
u = g on .

Here u C 2 () C () and g is referred to as boundary data. It is useful to


keep in mind the physical significance of the Dirichlet problem. Let the function
u describe the steady state temperature distribution in a homogeneous isotropic
body in the interior, which is the domain , and suppose that the temperature
distribution g is given on the boundary . In electrostatics the Dirichlet bound-
ary condition specifies the values of the potential u on the boundary , which
induces the electric field E = u in .
Remarks 4.8:
(1) If u is harmonic in and u = 0 on , then u = 0 in .
CHAPTER 4. LAPLACE EQUATION 98

Figure 4.1: The Dirichlet problem.

Reason. By Greens first identity


Z Z
2
| u( x)| dx = u( x) u( x) dx

u
Z Z
= u( x) u( x) dx + ( x) u( x) dS ( x) = 0.
| {z } |{z}
=0 =0

This implies that | u( x)| = 0 and thus u( x) = c for every x . Since


u( x) = 0 for x , the constant c has to be zero and we have u( x) = 0 for
every x .

(2) If u and v are harmonic functions in and u( x) = v( x) for every x ,


then u( x) = v( x) for every x .

Reason. Let w = u v. Then

w = ( u v) = u v = 0 in

and w = u v = 0 on . (1) implies w = 0 and thus u = v in .

T HE MORAL : Greens formulas imply that the solution of the Dirichlet


problem for the Laplace equation in a bounded domain is unique. Observe that
this result is not based on representation formulas for solutions and holds in all
domains and boundary values.
CHAPTER 4. LAPLACE EQUATION 99

Example 4.9. Consider the Dirichlet problem



u = 0 in ,
u = c on ,

where c R is a given constant. It is obvious that u( x) = c is a solution to this


problem. Since the solution is unique, it is the only solution to this problem.

Example 4.10. Let be the unit disc in R2 and consider the Dirichlet problem in
polar coordinates
u = 0 in ,
u(1, ) = cos , < .

It is easy to check that u( x, y) = x, or u( r, ) = r cos in polar coordinates, is the


only solution to this problem (exercise).

The Neumann problem


The Neumann problem for the Laplace equation is

u = 0 in ,
u = h on .

Physically the Neumann problem describes the steady state temperature distribu-
tion in when the heat flow through is given by h.

Example 4.11. It is obvious that all constant functions f ( x) = c, where c R, are


solutions to the Neumann problem above. Thus the problem has infinitely many
solutions.

THE MORAL : The solution of the Neumann problem for the Laplace equation
is not unique.

We need one more definition before the next claim. An open set is connected,
if every pair of points in can be connected by a piecewise linear path in .
Remarks 4.12:
u
(1) If u is harmonic in a connected domain and
= 0 on , then u = c in
.

Reason. By Greens first identity


Z Z
| u( x)|2 dx = u( x) u( x) dx

u
Z Z
= u( x) u( x) dx + ( x) u( x) dS ( x) = 0.
| {z } |
=0 {z }
=0

This implies that | u( x)| = 0 and thus u( x) = c for every x .


CHAPTER 4. LAPLACE EQUATION 100

Figure 4.2: The Neumann problem.

u v
(2) If u and v are harmonic in a connected domain and
=
on , then
u = v + c in .

Reason. Let w = u v. Then

w = ( u v) = u v = 0 in

and
w
= ( u v) = 0 on .

(1) implies w = c and thus u = v + c in .

THE MORAL : The Greens formulas imply that the solution of the Neumann
problem is unique up to an additive constant in a connected domain. If the
domain has several components, the solution may be different constant in every
component of the domain. Observe that this result is not based on representation
formulas for solutions and holds in all domains and boundary values.

Remarks 4.13:
(1) By Greens third identity, we obtain the following compatibility condition
of the Neumann problem
u
Z Z Z
u dx =
0 = |{z} dS = h dS.

=0

Note that the solution does not exist without the condition that the total
heat flow through the boundary is zero.
CHAPTER 4. LAPLACE EQUATION 101

(2) It is also possible to consider a combination of the Dirichlet and Neumann


boundary conditions

u
a( x) ( x) + b( x) u( x) = g( x), x ,

but we shall not do this here.
(3) We may define the Dirichlet and Neumann problems for the Poisson equa-
tion u = f . In this case, the function f describes the heat source.
However, it is enough to consider boundary value problems in which ei-
ther the equation is homogeneous ( u = 0) or the boundary condition is
homogeneous ( g = 0 or h = 0). For example, to solve

u = f in ,
u = g on ,

we may write u = u 1 + u 2 with



u 1 = f in ,
u = 0 on ,
1

and
u 2 = 0 in ,
u = g
2 on .

The solution of the original problem is a sum of solution these problems.

THE MORAL : It is enough to consider the nonhomogeneous PDE with


zero boundary values and the homogeneous PDE with nonzero boundary
values.

4.4 Fundamental solution of the Laplace


equation
In Section 3.9 we derived the fundamental solution of the Laplace equation in the
upper half-space using the Fourier transform. In this section, we derive a formula
for the fundamental solution of the Laplace equation in Rn . Since the equation
is linear, any linear combination, or integral, of fundamental solutions will be a
solution to the Laplace equation as well. This will give us a method to represent
all other solutions as integrals, or convolutions, with the fundamental solution.
We assume that = Rn and we are looking for special solutions of the form

u( x) = v(| x|) = v( r ( x)),


CHAPTER 4. LAPLACE EQUATION 102

where r ( x) = | x| = ( x12 + + x2n )1/2 . This means that we are looking for radial
solutions, that is, solutions that only depend on the distance from the origin.
Let us see what the Laplace equation is for the radial solutions. By the chain
rule
r 1 1 xj xj
( x) = (| x|) = ( x12 + + x2n ) 2 2 x j = = , j = 1, . . . , n, x 6= 0.
x j x j 2 | x| r ( x)
Again, by the chain rule
u r xj
( x) = (v( r ( x))) = v0 ( r ( x)) ( x) = v0 ( r ( x))
x j x j x j r ( x)
and

2 u x2j

xj

( x) = v00 ( r ( x)) + v0 ( r ( x))
x j2 r ( x )2 x j r ( x)

x2j x2j
!
00 01
= v ( r ( x)) + v ( r ( x)) , j = 1, . . . , n, x 6= 0.
r ( x )2 r ( x ) r ( x )3
Thus
n 2 u
u ( x) =
X
( x)
j =1 x2j
n x2j v0 ( r ( x)) n x2j
= v00 ( r ( x)) 0
X X
2
+ n v ( r ( x )) 3
j =1 r ( x) r ( x) j =1 r ( x)
| {z } | {z }
=1 =1/ r ( x)

00 n1 0
= v ( r ( x)) + v ( r ( x)) = 0, x 6= 0.
r ( x)
Hence for a radial function
n1 0
u( x) = 0, x 6= 0 v00 ( r ) + v ( r ) = 0, r > 0.
r
T HE MORAL : This is the radial version of the Laplace equation. Note that
the Laplace equation for a radial function becomes an ODE.

We solve the ODE and obtain


v00 n1 1 0

(ln v0 )0 = = = ( n 1)(ln r )0
= ln
v0 r r n1
which gives
c
v0 ( r ) =
r n1
for some constant c. Note also that we can assume that n 2 otherwise the ODE
is completely trivial v00 = 0. An integration gives

a ln r + b, n = 2,
v( r ) =
c + d, n 3,
r n2

where a, b, c, d are constants. These functions give radial solutions of the Laplace
equation in Rn \ {0}.
CHAPTER 4. LAPLACE EQUATION 103

Definition 4.14. The function : Rn \ {0} R,

1


ln | x|, n = 2,
( x ) = 2
1 1

, n 3,
n( n 2)( n) | x|n2

is called the fundamental solution of the Laplace equation. Here we denote the
volume of the unit ball in Rn by ( n) = |B(0, 1)|.

T HE MORAL : The goal is to represent all solutions to boundary value prob-


lems for the Laplace and Poisson equations using the fundamental solution.
Physically the fundamental solution is the potential induced by a unit point mass
at the origin.

Remarks 4.15:
(1) Note that is harmonic in Rn \ {0}, that is, ( x) = 0 for every x Rn \ {0}
(exercise). Moreover, has a singularity at the origin in the sense that
is unbounded in every neighbourhood of the origin.
(2) The choice of the scaling constants for the fundamental solution is a
normalization that gives


Z
dS = 1 for every r > 0.
B(0,r )

Essentially this normalization has the same role as the normalization


R
K ( x) dx = 1 for a family of good kernels so that f K will converge to f
instead of c f for some constant c. We shall return to this soon.

4
Example 4.16. Let n = 3. Then the volume of the unit ball in R3 is (3) = 3 and

1 1
( x ) = , x R3 , x 6= 0,
4 | x |

is the Coulomb (or Newton) potential. Thus the electric field is

1 1 x 1 x
E ( x) = ( x) = (| x|1 ) = | x|2 = , x R3 , x 6= 0.
4 4 | x | 4 | x | 3

4.5 The Poisson equation


We derive a representation formula for a solution to the nonhomogeneous Laplace
equation, called the Poisson equation,

u = f in Rn ,

where f C 0 (Rn ), that is, f is a compactly supported smooth function. Recall,


that the support of a function is compact, if the function is identically zero outside
CHAPTER 4. LAPLACE EQUATION 104

a closed and a bounded set. In practice, this set can be chosen to be a ball. For
a fixed y Rn , the function x 7 ( x y) f ( y) is harmonic in Rn \ { y}. Here is
the fundamental solution to the Laplace equation. Since the Laplace equation is
linear, we could think that the convolution
Z
u ( x) = ( x y) f ( y) d y
Rn

is also a solution of the Laplace equation.

W A R N I N G : This is wrong. We are not allowed switch the order of differentia-


tion and integration and conclude that
Z
u ( x) = x ( x y) f ( y) d y = 0.
Rn

The problem is that the second order partial derivatives of behave as c/| x|n ,
which is not an integrable function.

However, we have the following result, which shows that the function above is
a solution to the Poisson equation.

Theorem 4.17 (Solution to the Poisson equation in the whole space). Let
f C 0 (Rn ) and define
Z
u( x) = ( f )( x) = f ( y)( x y) d y,
Rn

where is the fundamental solution of the Laplace equation. Then u C 2 (Rn )


and u = f in Rn .

W A R N I N G : The problem above does not have a unique solution, since we can
always add a function v with v = 0 to the solution. This is not a serious problem,
since the solution above will be used as a tool in representation formulas.

THE MORAL : A convolution of the source term with the fundamental solution
is a solution to the Poisson equation in the whole space. Physically f describes a
charge density, that is, a distribution of electric charges and u is the potential of
the electric field induced by f . Observe that the potential is harmonic outside the
support of f .

Examples 4.18:
(1) In the plane R2 we have the logarithmic potential

1
Z
u ( x) = f ( y) ln | x y| d y.
2 R2

(2) In the space R3 we have the Newton (or Coulomb) potential

1 f ( y)
Z
u ( x) = d y.
4 R3 | x y|
CHAPTER 4. LAPLACE EQUATION 105

Remark 4.19. The theorem gives a solution u in the whole space without a speci-
fication of the boundary values. However, if is an open and bounded subset of
Rn and v is a solution of the Dirichlet problem

v = 0 in ,
v = u on .

Then w = u + v is a solution to the problem



w = f in ,
w = 0 on .

Thus we have reduced the Dirichlet problem for the Poisson equation to the
Dirichlet problem for the Laplace equation. This will be useful later in the
construction of Greens function.

Proof. Step 1: Let us first calculate the partial derivatives of u. The j th difference
quotient is

u( x + he j ) u( x) f ( x y + he j ) f ( x y)
Z
= ( y) dy
h Rn h

where h 6= 0 and e j = (0, . . . , 0, 1, 0, . . . , 0) with 1 in the j th slot. Since f is twice


continuously differentiable this difference quotient converges and

f ( x y + he j ) f ( x y) f
lim = ( x y).
h 0 h x j

This partial derivative is bounded and has compact support since f C 0 (Rn ). The
Lebesgue dominated convergence theorem allows us to interchange the limit and
the integral to get

u u( x + he j ) u( x)
( x) = lim
x j h 0 h
f ( x y + he j ) f ( x y)
Z
= lim ( y) dy
h 0 R n h
f ( x y + he j ) f ( x y)
Z
= ( y) lim dy
R n h 0 h
f f
Z
= ( y) ( x y) d y = ( x), j = 1, . . . , n.
Rn x j x j

Arguing in exactly the same way, which is possible because f has second order
continuous derivatives with compact support, we get

2 u 2 f 2 f
Z
( x) = ( y) ( x y) d y = ( x), j, k = 1, . . . , n.
x j xk Rn x j xk x j xk

Thus u C 2 (Rn ) and we have a formula for calculating the derivatives of u.


CHAPTER 4. LAPLACE EQUATION 106

T HE MORAL : This argument shows that the differentiation can be taken


inside the convolution. This is why the convolution inherits smoothness properties
of the functions.

Step 2: To show that u is a solution to the Poisson equation, we would like


to argue in the same way, pass the derivatives inside the integral, apply them
to instead of f and use the fact that is harmonic. The warning before the
statement of the theorem shows that we have to be careful here. We thus split the
convolution integral into two parts, one close to the origin and the other far away
from the origin.

In order to carry out this plan, let 0 < < 12 and write
Z Z
u ( x) = x ( y) f ( x y) d y = ( y) x f ( x y) d y
Z R Rn
n
Z
= ( y) x f ( x y) d y + ( y) x f ( x y) d y
B(0,) Rn \B(0,)

= I + J .

Step 3: First we estimate I .


Z
|I| |( y)|| x f ( x y)| d y
B(0,)
Z Z
sup | x f ( x y)| |( y)| d y sup | f ( x)| |( y)| d y,
yB(0,) B(0,) xRn B(0,)

where sup xRn | f ( x)| < since f C 0 (Rn ). Next we shall compute the remaining
integral.
CHAPTER 4. LAPLACE EQUATION 107

n=2

1 1
Z Z Z Z
|( y)| d y = ln | y| d y = ln | y| dS ( y) dr
B(0,) 2 B(0,) 2 0 B(0,r)
1 1
Z Z Z
= ln r 1 dS ( y) dr = 2 r ln r dr
2 0 B(0,r ) 2 0

1 2 1 1 1

= r ln r r 2 = 2 ln 2
2 0 4 2 4
1
= 2 | ln | + 2 .
4
| ln |
Since < 1/2 | ln | > | ln 1/2| 1 < | ln 1/2| the previous estimate implies that
Z
|( y)| d y c2 | ln |
B(0,)

for some constant c.


n3
Z Z Z Z
|( y)| d y = c | y|2n d y = c | y|2n dS ( y) dr
B(0,) B(0,) 0 B(0,r )
Z Z Z
=c r 2n 1 dS ( y) dr = c r 2n |B(0, r )| dr,
0 B(0,r ) 0

where c is a constant that depends only on the dimension. Now the ( n 1)-
dimensional volume of the sphere B(0, r ) is |B(0, r )| = cr n1 for some constant c
that depends only on dimension. Thus
Z Z
|( y)| d y = c r dr = c2
B(0,) 0

for some constant c that depends only on the dimension. In both cases we see that

lim I = 0.
0

Step 4: Then we compute J .


Z Z
J ( x) = ( y) x ( f ( x y)) d y = ( y) y ( f ( x y)) d y.
Rn \B(0,) Rn \B(0,)

This equality follows from

2 2
( f ( x y)) = ( f ( x y)), ( f ( x y)) = ( f ( x y)), j = 1, . . . , n.
x j yj 2
x j y2j

Greens first identity with = Rn \ B(0, ) is applied so that the exterior unit vector
points inwards to the ball
y y
( y) = = , y B(0, ).
| y|
CHAPTER 4. LAPLACE EQUATION 108

We apply Greens first identity to the functions ( y) and y 7 f ( x y) with x fixed.


This gives


Z Z
J ( x) = ( y) ( f ( x y)) dS ( y) y ( y) y ( f ( x y)) d y = K + L .
B(0,) Rn \B(0,)

Step 5: We estimate K .

Z
|( y)| ( f ( x y)) dS ( y)

|K |
B(0,)
Z
= |( y)|| f ( x y) ( x y)| dS ( y)
B(0,)
Z
|( y)|| f ( x y)| |( x y)| dS ( y) (Cauchy-Schwarz inequality)
B(0,) | {z }
=1
Z
sup | f ( y)| |( y)| dS ( y),
yRn B(0,)

where sup yRn | f ( y)| < because f C 0 (Rn ). Now we estimate the remaining
integral. Since is a radial function and | x| = on B(0, ) we get

Z 2| ln |, n = 2,
|( y)| dS ( y) = |()||B(0, )|
B(0,) c2n n1 = c, n 3.

From this we conclude that


lim K = 0.
0

Step 6: We compute L . Greens first identity gives


Z
L = y ( y) y ( f ( x y)) d y
Rn \B(0,)

Z Z
= y ( y) f ( x y) d y f ( x y) ( y) dS ( y)
Rn \B(0,) | {z } B(0,)
=0, y6=0

Z
= f ( x y) ( y) dS ( y)
B(0,)

Let us calculate the normal derivative



( y) = ( y) ( y).

n3

1
( y) = (2 n)| y|1n (| y|) ( y)
n( n 2)( n)
1 y y 1 | y| 2

= =
n( n) | y|n | y| n( n) | y|n+1
1 1
= | y|1n = 1n , y B(0, ).
n( n) n( n)
CHAPTER 4. LAPLACE EQUATION 109

Figure 4.3: The outward normal of Rn \ B(0, ).

This implies
1
Z
L = f ( x y) dS ( y)
n( n)n1 B(0,)
1
Z
= f ( x y) dS ( y)
|B(0, )| B(0,)
1
Z
= f ( z) dS ( z)
|B( x, )| B( x,)
by the change of variables z = x y. Recall that |B( x, )| denotes the ( n 1)-
dimensional volume of the sphere B( x, r ). Thus the last quantity is an integral
average of f over a sphere of radius centered at x. Let us take a closer look at
this average.
1
Z
Claim: lim f ( z) dS ( z) = f ( x).
0 |B( x, )| B( x,)

Reason. We have
1 1
Z Z
f ( z) dS ( z) f ( x) = ( f ( z) f ( x)) dS ( z).
|B( x, )| B( x,) |B( x, )| B( x,)
Let > 0. By continuity there exists > 0 such that

| z x| | f ( x) f ( z)| .

This implies

1 1
Z Z
| f ( z) f ( x)| dS ( z) ,


|B( x, )| f ( z) dS ( z) f ( x)

B( x,) |B( x, )| B( x,)
CHAPTER 4. LAPLACE EQUATION 110

which proves the claim. Observe, that this is the same argument as in approxima-
tion of the identity.

This implies that


lim L = f ( x).
0

n = 2 Observe that
1
( y) = ( y) ( y) = (ln | y|) ( y)
2
1 1 y y 1 1

= = = , y B(0, ), j = 1, . . . , n.
2 | y| | y| | y| 2| y| 2
This is precisely the same formula as in the case n 3 for n = 2. Thus for
y B(0, ) we have
1 1
( y) =
=
2 |B(0, )|
and the rest is exactly the same as in the higher dimensional case. Thus

lim L = f ( x)
0

also when n = 2.
Step 7: Gathering all estimates together we see that

u( x) = I + J = I + K + L f ( x) as 0
|{z} |{z} |{z}
0 0 f ( x)

and we find that u solves the Poisson equation u = f .

Remark 4.20. We collect here basic results on volumes related to balls and spheres.
Since the n-dimensional volume with n 2 of a ball B( x, r ) is translation invariant
and scales to the power n, we have

|B( x, r )| = |B(0, r )| = r n |B(0, 1)| = ( n) r n ,

where ( n) is the n-dimensional volume of the unit ball B(0, 1). On the other hand,
since the ( n 1)-dimensional volume of the sphere B( x, r ) is translation invariant
and scales to the power n 1, we have

|B( x, r )| = |B(0, r )| = r n1 |B(0, 1)| = ( n) r n1 ,

where ( n) is the ( n 1)-dimensional volume of the unit sphere B(0, 1). Moreover,
( n)
Z Z 1Z Z 1
( n) = 1 dx = 1 dS dr = ( n) r n1 dr =
B(0,1) 0 B(0,r ) 0 n
Thus
|B(0, 1)| = ( n) = n( n) = n|B(0, 1)|.
It can be shown that n
2
|B(0, 1)| = n ,
2 +1
so that the volume of a ball can be represented in terms of the gamma function.
CHAPTER 4. LAPLACE EQUATION 111

4.6 Greens function


In this section we show how we can use the fundamental solution of the Laplace
equation in the whole space to solve a Dirichlet problem in a subdomain. Let
Rn be an open and bounded set with a smooth boundary and assume that
f C 0 (). We consider the Dirichlet problem for the Poisson equation

u = f in ,
u = g on .

Our goal is to derive a general representation formula for the solution of this
problem using potential functions and so called Greens function.
Assuming that u C 2 ()). Let > 0 be small enough so that B( x, ) . Recall
that is the fundamental solution of the Laplace equation in Defintion 4.14. By
Greens second identity, we have
Z
u( y) y (( y x)) ( y x) y u( y) d y

\B( x,) | {z }
=0
u
Z Z
= u( y) ( y x) dS ( y) ( y x) ( y) dS ( y) (4.1)
B( x,) B( x,)
Z Z Z Z
= dS ( y) + dS ( y) dS ( y) dS ( y)
B( x,) B( x,)

= I 1 () + I 2 I 3 () I 4 .

Here denotes, as usual, the outer unit normal vector on B( x, ). Let us


consider the terms above separately.
I 3 ()

u
Z
| I 3 ()| = ( y x) ( y) dS ( y)

B( x,)
Z
|( y x)|| u( y)||( y)| dS ( y)
B( x,)
Z
sup | u( y)| |()| d y
y B( x,)

sup
y
| u( y)|| ln |, n = 2,

c sup | u( y)|n1 2n , n 3.
y

Thus
lim I 3 () = 0.
0

I 1 () By the calculation in the proof of Theorem 4.17, we obtain

1 1
( y x) = = , y B( x, ).
n( n)| y x|n1 n( n)n1
CHAPTER 4. LAPLACE EQUATION 112

A substitution of this in I 1 () gives



Z
I 1 () = u ( y) ( y x) dS ( y)
B( x,)
1
Z
= u( y) dS ( y)
n( n)n1 B( x,)
1
Z
= u( y) dS ( y) u( x) as 0.
|B( x, )| B( x,)

Thus
lim I 1 () = u( x).
0
Finally we recall that is harmonic away from zero so that y ( y x) = 0 for
y 6= x. Thus by rearranging terms in (4.1) and letting 0 we conclude that

u( x) = lim I 1 ()
0
Z
= lim I 3 () + I 4 I 2 lim ( y x) y u( y) d y
0 0 \B( x,)
| {z }
=0
u
Z Z
= ( y x ) ( y) dS ( y) u( y) ( y x) dS ( y) (4.2)

Z
( y x) y u( y) d y, x .

T HE MORAL : This representation formula holds for any function u C 2 ().


In particular, the function u does not need to be a solution to a PDE. This gives
a representation formula for a function inside the domain by its values on the
boundary of the domain. More precisely, this allows us to determine u if we know
the value of u in as well as the value of u and the normal derivative on the
boundary . This is useful in boundary value problems for PDEs.

Remarks 4.21:
(1) If u = 0 in , then by (4.2), we have

u
Z Z
u ( x) = ( y x ) ( y) dS ( y) u( y) ( y x) dS ( y), x . (4.3)

The first integral on the right-hand side is called the single layer potential
u
with charge density
( y) and the second integral is called the double
layer potential with dipole moment density u( y). The latter represents the
potential induced by a double layer of charges of opposite sign on and
the former represents the potential induced by a single layer of charges on
. Note that these potentials are harmonic in and in Rn \ .
(2) Representation formula (4.3) implies that if u C 2 () is a harmonic func-
tion in , then u C (). This means that every harmonic function
is smooth. Indeed, since there is no singularity in the integrand, the
derivatives can be taken inside the integral.
CHAPTER 4. LAPLACE EQUATION 113

(3) By choosing u = 1 in (4.3), we have


Z
( y x) dS ( y) = 1

for every x . This is related to the normalization of the fundamental


solution in Defintion 4.14.

Let us look at the representation formula above in connection with the Dirich-
let problem. We require that u = f in . The boundary condition specifies the
values of u = g on the boundary , but the normal derivative of u is unknown.
We solve this problem as follows. For a fixed x , let x = x ( y) be a corrector
function, which is a solition the Dirichlet problem

y x ( y) = 0 , y ,
x ( y) = ( y x), y .

T HE MORAL : We solve a Dirichlet problem with boundary values given by


the fundamental solution of the Laplace equation in Rn .

Applying Greens second identity and using the fact that x = 0 in we have

Z Z
x ( y) u( y) d y = u( y) x ( y) x ( y) u( y) d y
| {z }
=0
x u
Z
= u ( y) ( y) x ( y)
( y) dS ( y) (4.4)

x u
Z
= u ( y) ( y) ( y x) ( y) dS ( y).

In the last equality we used the fact that x ( y) = ( y x) for y .

Definition 4.22. Greens function for is

G ( x, y) = ( y x) x ( y), x, y , x 6= y.

With the definition of Greens function and adding (4.2) to (4.4), we have

u
Z
u ( x) = ( y) (( y x) ( y x)) dS ( y)
| {z }
=0
x
Z
u ( y) ( y x) ( y) dS ( y)

Z
x
+ u( y)( ( y) ( y x)) d y

G
Z Z
= u( y) ( x, y) dS ( y) u( y)G ( x, y) d y

CHAPTER 4. LAPLACE EQUATION 114

where
G
( x, y) = y G ( x, y) ( y)

and is the exterior unit normal on .

T HE MORAL : This representation formula holds for any function u C 2 ().


This allows us to determine u if we know the value of u in , the value of u on
the boundary and Greens function for . In contrast with (4.2), there is no
normal derivative on the boundary.

Theorem 4.23 (Solution of the Poisson equation in a subdomain). Suppose


that u solves the boundary value problem

u = f in ,
u = g on ,

then
G
Z Z
u ( x) = g( y) ( x, y) dS ( y) + f ( y)G ( x, y) d y,

where G is Greens function for .

T HE MORAL : This is an explicit representation formula solution for the


Poisson equation in a bounded subdomain. Note that u = u 1 + u 2 , where u 1 is a
solution to the corresponding homogeneous Dirichlet problem and u 2 is a solution
to a nonhomogeneous problem with zero boundary values as in Remark 4.13 (3).
The representation formula depends on the construction of Greens function for
, which is a difficult task and depends heavily on the geometry of . We shall
derive Greens functions of some relatively simple domains soon.

Remarks 4.24:
(1) Let us symbolically write x for the generalized function (distribution) that
has the property
Z
( y) x ( y) d y = ( x) for all C 0 (Rn ).
Rn

This generalized function is called Diracs delta mass at x. We have showed


that f solves the Poisson equation u = f on Rn . Formally we have
= 0 and
Z Z
u( x) = ( y) f ( x y) d y = 0 ( y) f ( x y) d y = f ( x).
Rn Rn

The generalized function 0 , roughly speaking, can be interpreted as


an object that has mass one at the point 0 and is zero everywhere else.
Physically this means that we set a unit charge at the origin and the
fundamental solution is the induced potential.
CHAPTER 4. LAPLACE EQUATION 115

(2) For a fixed x , consider Greens function G ( x, y) as a function of y. Then



y G ( x, y) = x , y ,
G ( x, y) = 0, y ,

where x a Dirac mass at x . In particular, Greens function G ( x, y) is


a harmonic function of y in \ { x} and G ( x, y) has zero boundary values
on . The boundary condition follows from the fact that x ( y) = ( y x)
for y . Physically this means that we set a unit charge at the point
x and require that the induced potential is zero on the boundary, that is,
the boundary is grounded. Observe, that Greens function depends only on
the domain. It can be also shown mathematically, that Greens function
exists for any smooth bounded domain , but this is out of the scope of
this course.

Remark 4.25. The solution of the Dirichlet problem



u = 0 in ,
u = g on ,

is Z
u ( x) = H ( x, y) g( y) dS ( y),

where
G
H ( x, y) = ( x, y)

is called the Poisson kernel in .

4.7 Greens function for the upper half-


space*
n+1
Let R+ be the upper half-space

n+1
R+ = {( x0 , xn+1 ) : x0 Rn , xn+1 > 0}.

Although this domain is unbounded, and the arguments in the previous section
do not directly apply, we shall determine its Green function by the method of
reflection. After we have done that we have to check whether representation
formula really gives a solution to the problem. Remember that in order to construct
n+1
Greens function, for every x R+ , we need to construct the corrector function
x such that
x ( y) = 0, y Rn+1 ,
+
x ( y) = ( y x), y Rn+1 = Rn .
+
CHAPTER 4. LAPLACE EQUATION 116

n+1
Greens function for R+ will then be G ( x, y) = ( y x) x ( y).
n+1
Let x = ( x1 , . . . , xn , xn+1 ) R+ . Reflecting the vector x across the boundary
n+1
R+ gives the point x = ( x1 , . . . , xn , xn+1 ). Observe that x belongs to the lower
n+1
half-space of Rn+1 . For x, y R+ we set

x ( y) = ( y x )

where is the fundamental solution of the Laplace equation. The singularity at


n+1 n+1
x R+ is reflected to x . Observe that, if y R+ , then

x ( y) = ( y x ) = ( y x )

since is a radial function. Furthermore, since x R+


n+1
, we have y x 6= 0 and
thus is harmonic in a neighborhood of y x . This shows that x is a solution of
the desired boundary value problem. This is called the method of reflections.
n+1
Theorem 4.26. Greens function for the upper half-space R+ is

n+1
G ( x, y) = ( y x) ( y x ), x, y R+ , x 6= y.

We consider the boundary value problem



u = 0 in Rn+1 ,
+
(4.5)
u = g on Rn+1 .
+

The representation formula in Remark 4.25 gives

G
Z
u ( x) = ( x, y) g( y) dS ( y).
n +1
R+

G
We derive an explicit expression for
( x, y) in the upper half-space. Recalling the
definition of the fundamental solution for n 3, we have

1 1 1

G ( x, y) = .
n( n 2)( n) | x y|n2 | x y|n2

Thus
G G
( x, y) = G ( x, y) ( y) = ( x, y)
yn+1

= ( y x) + ( y x )
yn+1 yn+1
1 yn+1 xn+1 yn+1 + xn+1 2 xn+1

= = .
n( n) | x y| n +1 | x y| n +1 n( n) | x y|n+1

This holds also when n = 2 (exercise). By insreting this into the representation
formula above, we have

2 xn+1
Z
n+1
u ( x) = g( y) d y, x R+ , y Rn = R+
n+1
.
n( n) Rn | x y|n+1
CHAPTER 4. LAPLACE EQUATION 117

The function
2 xn+1 n+1
K ( x, y) = , x R+ , y Rn = R+
n+1
.
n( n) | x y|n+1
n+1
is the Poisson kernel of R+ and

2 xn+1 g ( y)
Z Z
n+1
u ( x) = dy = K ( x, y) g( y) d y, x R+ , y Rn .
n( n) Rn | x y|n+1 Rn

Observe that this is a convolution of g with the Poisson kernel.

Theorem 4.27 (Poisson formula in the upper half-space). The solution of the
Dirichlet problem (4.5) is given by

2 xn+1 g( y)
Z
u ( x) = d y. (4.6)
n( n) Rn | x y|n+1

T HE MORAL : This is an explicit representation formula for the solution of


the Dirichlet problem in the upper half-space. The Poisson kernel for the upper
half-space can be computed using Greens function.

Example 4.28. For a harmonic function in the upper half plane R2+ , formula (4.6)
gives
y g( z)
Z
u( x, y) = dz,
R ( x z)2 + y2
where < x < and y > 0.

Remark 4.29. Let us write ( x, y) instead of ( x0 , xn+1 ). Then (4.6) becomes

2y g( z)
Z
n+1
u( x, y) = dz, ( x, y) R+ .
n( n) Rn (| x z|2 + y2 ) n+2 1

We just rediscovered our familiar Poisson kernel for the upper half plane by means
of Greens function for the upper half plane. Previously, we derived the same
formula using the Fourier transform, see Theorem 3.25. We already know that
the u given above solves the Dirichlet problem for the Laplace equation in the
upper half-space with boundary data g by the approximation of the identity.

4.8 Greens function for the ball*


Let to be the unit ball B(0, 1) Rn . We shall again use the method of reflection
to construct Greens function. We need to find the corrector function x , for every
x B(0, 1), such that

x ( y) = 0, y B(0, 1),
x ( y) = ( y x), y B(0, 1).
CHAPTER 4. LAPLACE EQUATION 118

Greens function for B(0, 1) will then be G ( x, y) = ( y x) x ( y). For any point
x B(0, 1) we will again reflect the point with across to the boundary B(0, 1) as
follows. For x B(0, 1), we define
x
x = .
| x |2
1
Observe that | x| < 1 implies that | x | = | x| > 1 and thus x B(0, 1). The following
calculations are essentially the same when n = 2 so we give the details only for
the case n 3. Set
x ( y) = (| x|( y x )).

Then
1 1 1
x ( y) = = n2 ( y x ).
n( n 2)( n) | x| n 2
|y x | n 2 | x|
Since y 6= x the function x ( y) is harmonic in B(0, 1) (as a function of y) as long
as x 6= 0. Let us check what happens on the boundary. For y B(0, 1) and x 6= 0
we have
x x x x 2x y

| x |2 | y x | 2 = | x |2 y 2 y 2 = | x |2 y y + 2 2
| x| | x| | x| | x| | x |2
1 2x y

()
= | x |2 1 + 2 = | x |2 + 1 2 x y
| x| | x |2
= ( x y) ( x y) = | x y|2 ,

where in (*) we used that y y = | y|2 = 1 since y B(0, 1). The previous calculation
implies | x|| y x | = | y x| when y B(0, 1) so that

1 1
x ( y) = (| x|( y x )) = = ( x y)
n( n 2)( n) (| x|| y x |)n2

when y B(0, 1). Thus x is the corrector function for B(0, 1) and Greens function
for B(0, 1) becomes

G ( x, y) = ( y x) x ( y) = ( y x) (| x|( y x )), x, y B(0, 1), x 6= y, x 6= 0.

Theorem 4.30. Greens function for the ball B(0, 1) is

G ( x, y) = ( y x) (| x|( y x )), x, y B(0, 1), x 6= y, x 6= 0.

Let us now consider the Dirichlet problem



u = 0 in B(0, 1),
u = g on B(0, 1).

The general formula for the solution (see Remark 4.25) is

G
Z
u ( x) = ( x, y) g( y) dS ( y)
B(0,1)
CHAPTER 4. LAPLACE EQUATION 119

where ( y) = y/| y| is the exterior unit normal. Clearly

G y
( x, y) = y G ( x, y) ( y) = y G ( x, y) .
| y|

We calculate the partial derivatives of G (again for n 3), when y B(0, 1), and
have
G
(| x|( y x )

( x, y) = ( y x)
yj yj yj
| x|( y j xj )
!
1 1 yj x j 1
= | x|
n( n) | y x|n1 | y x| || x|( y x )|n1 | x|| y x |
!
1 y j x j | x |2 y j x j 1 1 | x |2
= = y j , j = 1, . . . , n.
n( n) | y x| n | y x| n n( n) | x y|n

Thus
1 1 | x |2
y G ( x, y) = y
n( n) | x y|n
and

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

since | y|2 = 1 when y B(0, 1). We conclude that

1 1 | x |2
Z
u ( x) = g( y) dS ( y)
n( n) B(0,1) | x y|
n

is the solution for the Dirichlet problem for the Laplace equation in the unit ball
B(0, 1).
The Dirichlet problem for B(0, r ), r > 0, is

u = 0 in B(0, r ),
u = g on B(0, r ).

Then by a change of variables, we have

r 2 | x |2 g ( y)
Z
u ( x) = dS ( y), x B(0, r ).
n( n) r B(0,r ) | x y|
n

The function

r 2 | x |2 1
K ( x, y) = , x B(0, r ), y B(0, r ),
n( n) r | x y|n

is called the Poisson kernel for the ball B(0, r ) and

r 2 | x |2 g ( y)
Z Z
u ( x) = dS ( y) = K ( x, y) g( y) dS ( y), x B(0, r ),
n( n) r B(0,r ) | x y|
n
B(0,r )

is the solution of the Dirichlet problem above.


CHAPTER 4. LAPLACE EQUATION 120

Remark 4.31. The Poisson kernel has the following properties (exercise).

(1) K ( x, y) is a smooth function of x B(0, r ) for any fixed y B(0, r ).


(2) K ( x, y) > 0 for every x B(0, r ) and y B(0, r ).
(3) For any fixed x0 B(0, r ) and > 0,

lim K ( x, y) = 0
x x0 ,xB(0,r )

for every y B(0, r ) \ B( x0 , ).


(4) x K ( x, y) = 0 for every x B(0, r ) and y B(0, r ).
Z
(5) K ( x, y) dS ( y) = 1 for every x B(0, r ).
B(0,r )

Note that these properties are analogous to the properties of an approximation of


the identity, but now the kernel function is defined on the sphere.

Theorem 4.32 (Poisson formula in the ball). The solution of the problem

u = 0 in B(0, r ),
u = g on B(0, r ).

is given by
r 2 | x |2 g ( y)
Z
u ( x) = dS ( y).
n( n) r B(0,r ) | x y|
n

T HE MORAL : This is an explicit representation formula for the solution of


the Dirichlet problem in a ball. The Poisson kernel for the ball can be computed
using Greens function for the corresponding ball.

Remark 4.33. In dimension n = 2 the unit ball B(0, 1) is just the disc. Let us
write z = ( r cos , r sin ) in polar coordinates. Since y B(0, 1) we can write
y = (cos , sin ) and integrate in instead. We can calculate

| z y|2 = ( r cos cos )2 + ( r sin sin )2


= r 2 + 1 2 r (cos cos + sin sin )
= 1 2 r cos( ) + r 2

Thus the previous formula becomes

1 r2
Z g(cos , sin )
u( r, ) = d ,
2 1 2 r cos( ) + r 2

where, in the last equality, we abuse notation and write g() for g(cos , sin ).
Again, we recover our familiar Poisson kernel for the disc in two dimensions. This
discussion motivates the definition of a Poisson kernel for the ball B(0, 1).
CHAPTER 4. LAPLACE EQUATION 121

4.9 Mean value formulas


Let Rn be an open set and assume that u is a harmonic function in . An
important property of harmonic functions is that their value at every point x
equals the average of the function over balls B( x, r ) or spheres B( x, r ) whenever
B( x, r ) . In one dimension, the Laplace equation is u00 = 0 and the harmonic
functions are of the form u( x) = ax + b, where a, b R. The value of this function
at the midpoint of a finite interval is the arithmetic average of the values at the
endpoints and the integral average over the interval. This is the one-dimensional
version of the mean value formula.

Theorem 4.34 (Mean value formulas for harmonic functions). Let u C 2 ()


be harmonic in . Then for every ball B( x, r ) such that B( x, r ) we have

1 1
Z Z
u ( x) = u( y) d y = u( y) dS ( y).
|B( x, r )| B( x,r) |B( x, r )| B( x,r)

THE MORAL : The value of a harmonic function at a point is the equal to the
average of its values over any ball or sphere centered at that point.

Proof. Set

1
Z
( r ) = u( y) dS ( y)
|B( x, r )| B( x,r)
1
Z
= u( x + rz) r n1 dS ( z)
n( n) r n1 B(0,1)
1
Z
= u( x + rz) dS ( z),
n( n) B(0,1)

where we used the change of variables y = x + rz, so that z = 1r ( y x) and dS ( y) =


r n1 dS ( z). Recall the formulas for the measure of the ball and the sphere from
Remark 4.20. This change of variables takes the sphere B( x, r ) to the sphere
B(0, 1). By differentiating with respect to r under the integral sign and using the
chain rule, we get

1
Z
0 ( r ) = u( x + rz) z dS ( z)
n( n) B(0,1)
1 y x
Z
= u( y) dS ( y)
n( n) r n1 B( x,r) r
1
Z
= u( y) ( y) dS ( y),
n( n) r n1 B( x,r)

where we used the change of variables y = x + rz, so that so that z = 1r ( y x) and


dS ( z) = r 1n dS ( y). Observe that this change of variables takes B(0, 1) back to
B( x, r ). Note that ( y) = ( y x)/ r is the outer normal unit vector on the sphere
CHAPTER 4. LAPLACE EQUATION 122

B( x, r ). By the Gauss-Green formula

1 1 u
Z Z
u( y) ( y) dS ( y) = ( y) dS ( y)
n( n) r n1 B( x,r) n( n) r n1 B( x,r)
1
Z
= div u( y) d y
n( n) r n1 B( x,r)
1
Z
= u ( y) d y = 0 ,
n( n) r n1 B( x,r)

since u is harmonic in B( x, r ) . Thus 0 ( r ) = 0 which implies that is a constant


function. Recall that
1
Z
lim u( y) dS ( y) = u( x)
t0 |B( x, t)| B( x,t)

whenever u is continuous, as in the proof of Theorem 4.17. Since is constant, we


have
1
Z
( r ) = lim ( t) = lim u( y) dS ( y) = u( x).
t 0 t0 |B( x, t)| B( x,t)

This shows that


1
Z
u ( x) = u( y) dS ( y).
|B( x, r )| B( x,r )

On the other hand,


Z Z rZ Z r
u ( y) d y = u( y) dS ( y) ds = n( n) s n1 u( x) ds
B( x,r ) 0 B( x,s) 0
n
r
= n( n) u( x) = ( n) r n u( x) = |B( x, r )| u( x),
n
which shows the mean value formula for balls as well.

Remark 4.35. The mean value property of harmonic functions follows also from
the Poisson formula for the ball in Theorem 4.32 (exercise).

We now state a converse to the mean value property. It states that the mean
value property characterizes harmonic functions.

Theorem 4.36. If u C () satisfies


1
Z
u ( x) = u( y) dS ( y)
|B( x, r )| B( x,r )

for all balls B( x, r ) , then u is harmonic in .

T HE MORAL : The mean value property is a special property of harmonic


functions and it does not hold as such for other PDE.

Proof. We shall prove the claim under additional assumption that u C 2 (). A
convolution approximation will give the general result, but this will be omitted
here (exercise). If u( x) 6= 0 for some x then there is a ball B( x, r ) such
that u( y) > 0 for every y B( x, r ). As in the previous proof we have 0 ( r ) = 0,
CHAPTER 4. LAPLACE EQUATION 123

because of the hypothesis that the averages with respect to balls centered at x are
constant and equal to u( x). Thus we have

1
Z
0 = 0 ( r ) = u( y) d y > 0,
n( n) r n1 B( x,r )

which is a contradiction.

Remark 4.37. It follows from the proof above that a continuous function u is
harmonic if and only if for every point in the domain of definition the mean value
property holds true for small enough balls centered at the point.

4.10 Maximum principles


Recall, that a continuous function attains its maximum and minimum values on
a closed and bounded set. The maximum principle asserts that if, in addition,
the function is harmonic, then it must attain its maximum and minimum at the
boundary of the set. Physically, the Laplace equation governs the steady state
temperature distribution of a body. If the body is in thermal equilibrium, there
cannot be no internal hot or cold spots, since otherwise the heat energy would
flow from hot to cold. In other words, the temperature cannot have maximum or
minimum inside the body unless the temperature is constant throughout the entire
body. The maximum and minimum principles have several useful applications in
PDE. Recall, that an open set is connected, if every pair of points in can be
connected by a piecewise linear path in .

Theorem 4.38. Let Rn be an open and bounded set and assume that u
C 2 () C () is a harmonic function in .

(1) (Weak maximum principle) Then

max u( x) = max u( x).


x x

(2) (Strong maximum principle) If is a connected set and there exists x0


such that
u( x0 ) = max u( x),
x

then u is constant in .

Remark 4.39. By replacing u by u we get the minimum principles with min


replacing max in the maximum principle (exercise).

THE MORAL : According to the weak maximum principle, a harmonic function


cannot have a strict interior maximum or minimum point. It attains the minimum
and maximum on the boundary. The strong maximum principle asserts that a
CHAPTER 4. LAPLACE EQUATION 124

harmonic function can attain an interior minimum or maximum only if it is a


constant function. Observe that this result holds for all bounded domains without
any regularity assumption on the boundary.

Proof. (2) Suppose that there exists x0 such that

u( x0 ) = max u( x) = M.
x

Let 0 < r < dist( x0 , ). Then B( x0 , r ) and the mean value property implies
that
1 1
Z Z
M = u ( x0 ) = u ( y) d y M d y = M.
|B( x0 , r )| B( x0 ,r ) | B ( x0 , r )| B( x0 ,r )

It follows that an equality holds throughout and thus


Z
( M u( y)) d y = 0.
B( x0 ,r )

Since M u( y) 0, we conclude that u( y) = M for every y B( x0 , r ).


Now consider any point x . Since is connected, we can find a sequence of
balls B( x0 , r ), B( x1 , r ), . . . , B( x N , r ) such that

x j+1 B( x j , r ) for every j = 0, 1, . . . , N 1

and x = x N .

Figure 4.4: The chaining argument.


CHAPTER 4. LAPLACE EQUATION 125

Since u( x) = M for every x B( x0 , r ) and B( x0 , r ) contains the center of B( x1 , r )


we conclude that u( x1 ) = M . Repeating the same argument as above we have
u( x) = M for every x B( x1 , r ) as well. Continuing the same way we see that u = M
on all balls B( x0 , r ), B( x1 , r ), . . . , B( x N , r ). Since x B( x N , r ) we have u( x) = M .
Since x was an arbitrary point in , we have showed that u( x) = M for every x .
(1) We may assume that is connected by considering every component
separately. Observe that we always have

max u( x) max u( x).


x x

Let us assume (for contradiction) that this inequality is strict, that is,

max u( x) > max u( x).


x x

Then there is x0 such that

u( x0 ) = max u( y).
y

The claim (2) implies that u is constant in and since u C () we conclude that
u is constant in as well. Thus

max u( x) = max u( x),


x x

which is a contradiction with our assumption.

Remark 4.40. Let be a bounded, open and connected set and u C 2 () C ().
Suppose that u is a solution to the Dirichlet problem

u = 0 in ,
u = g on .

Assume that g 0. If g( x) > 0 for some x , then the strong minimum principle
implies that u > 0 everywhere in .

Reason. By the weak minimum principle

u( x0 ) min u( x) = min u( x) = min g( x) 0


x x x

for every x0 . If u( x0 ) = 0 for some x0 , then

u( x0 ) = min u( x) = 0
x

and by the strong minimum principle u = 0 in . Since u C (), we have u = 0 in


. This implies that u = g = 0 on , which is a contradiction.

Remark 4.41. The weak maximum principle can be also proved without the mean
value property.
CHAPTER 4. LAPLACE EQUATION 126

Reason. Assume first that u( x) > 0 for every x . If u has a maximum in at a


point x0 , then u( x0 ) = 0 and

2 u
( x0 ) 0 , j = 1, . . . , n,
x2j

(think of a function of one variable: if f has a local maximum at x0 and f is


twice differentiable, then f 00 ( x0 ) 0). By summing up, we have u( x0 ) 0. This
is impossible, thus u cannot have a maximum point in . Since u C (), we
conclude
max u( x) = max u( x).
x x

Then we consider the general case u( x) 0 for every x . Consider the auxiliary
function v ( x) = u( x) + | x|2 with > 0. A direct calculation shows that

v ( x) = ( u( x) + | x|2 ) = u( x) + (| x|2 ) > 0 for every x .

Note that the only property of the function | x|2 that we use here is that its Laplace
operator is strictly positive. Any other function with this property would do as
well. Since v C 2 () C (), by the beginning of the proof we have

max v ( x) = max v ( x).


x x

This implies

max u( x) + min | x|2 max( u( x) + | x|2 )


x x x

= max( u( x) + | x|2 )
x

max u( x) + max | x|2 .


x x

By letting 0, we have
max u( x) max u( x).
x x

On the other hand, since , we have

max u( x) max u( x)
x x

Thus
max u( x) = max u( x).
x x

As immediate consequences we obtain a comparison principle, a stability


and uniqueness results for the Dirichlet problem for the Poisson equation. The
following results hold, in particular, for harmonic functions.

Theorem 4.42 (Comparison principle). Let Rn be open and bounded and


assume that u, v C 2 () C () are solutions to the Poisson equation u = f in
. If u v on , then u v in .
CHAPTER 4. LAPLACE EQUATION 127

Proof. Since u v on we have u v 0 on . On the other hand,

( u v) = u v = f + f = 0 in

The maximum principle implies that u v 0 in and the claim follows.

Theorem 4.43 (Stability). Let Rn be open and bounded and assume that
u, v C 2 () C () are solutions to the Poisson equation u = f in . If | u v|
on , then | u v| in .

THE MORAL : This shows that the solution of the Dirichlet problem depends
continuously on the boundary data on a bounded set.

Proof. Since | u v| on we have u v on . Since u v is harmonic


in , the minimum and maximum principles imply that u v in and
the claim follows.

Example 4.44. The following Hadamards example shows that stability fails in
unbounded domains. Consider the Cauchy problem
2 u 2 u

u = 2 + 2 = 0 in R2+ ,


x y




u( x, 0) = 0, x R,
u

1


( x, 0) = sin( jx), x R, j = 1, 2, . . . .
y

j
Then
1
u( x, y) = sinh( j y) sin( jx)
j2
is a solution of the problem. Recall that sinh( x) = 12 ( e x e x ). Observe that this
function is a solution both to the Dirichlet and Neumann problems. By taking j
sufficiently large, the absolute value of the boundary data can be be everywhere
arbitrarily small, while the solution takes arbitrarily large values even at points
( x, y) with | y| as small as we wish.

An important application of the maximum principle is uniqueness of the


Dirichlet boundary value problem for the Poisson equation.

Theorem 4.45 (Uniqueness). Let Rn be a bounded open set. Then there


exists at most one solution u C 2 () C () to the boundary value problem

u = f in ,
u = g on .

Proof. Let u and v be two solutions for the Dirichlet problem. Then

( u v) = u v = f + f = 0 in

and uv = g g = 0 on . By the maximum principle we have uv 0 everywhere


in and by the minimum principle we have u v 0 everywhere in . Thus
u v = 0 everywhere in
CHAPTER 4. LAPLACE EQUATION 128

Remarks 4.46:
(1) Another way to show uniqueness in a smooth domain is to apply Gauss-
Green formula as in Remark 4.8.
(2) The assumption that is bounded is essential. Consider, for example, the
Dirichlet problem
u = 0 in Rn+1 ,
+
u = 0 on Rn+1 = Rn ,
+

n+1
in the unbounded open and connected unbounded set R+ . Every function
u( x0 , xn+1 ) = axn+1 , where x0 Rn , xn+1 R+ and a R, is a solution to the
problem. Thus the uniqueness of the boundary valued problem fails in this
case. Indeed, there are infinitely many solutions to the Dirichlet problem
with zero boundary value.
(3) Consider the Dirichlet problem

u = 0 in ,
u = 0 on .

If is a bounded set, then u = 0 is the unique solution of the problem.


However, we may have nontrivial solutions for an undounded . Let
= { x Rn : | x| > 1}. Then

| x|2n 1, n 3,
u ( x) =
log | x|, n = 2,

is a nontrivial solution to the problem.

4.11 Harnacks inequality*


Harnacks inequality can be seen as a quantitative version of the maximum
principle.

Theorem 4.47 (Harnacks inequality). Let Rn be an open set and V


be a connected bounded open set such that V . Then there exists a constant
c > 0, depending only on V , such that

sup u( x) c inf u( x)
xV xV

for all nonnegative harmonic functions u in .

Remark 4.48. Harnacks inequality gives the pointwise estimate

1
u( y) u( x) cu( y)
c
CHAPTER 4. LAPLACE EQUATION 129

for all points x, y V . This means that the values of nonnegative harmonic
functions in V are comparable. Thus if u is small (or large) somewhere in V it is
small (or large) everywhere in V . In particular, if u( y) = 0 for some y , then
u( x) = 0 for every x . The assumption that u 0 is essential in the result.

Proof. Since V is a compact set inside , we have dist(V, ) > 0. Let


1
r= dist(V, ) .
4
Let x, y V such that | x y| < r . Then B( y, r ) B( x, 2 r ). By the mean value
property
1 1
Z Z
u ( x) = u( z) dz = n u( z) dz
|B( x, 2 r )| B( x,2r) 2 ( n) r n B( x,2r)
1 1 1 1
Z Z
n u( z) dz = n u( z) dz = n u( y).
2 ( n) r n B( y,r) 2 |B( y, r )| B( y,r) 2

Now V is connected and V is compact so we can cover V by finitely many balls


{B( x j , r /2)} N
j =1
such that
r r
B x j , B x j+1 , 6= ; for j = 1, . . . , N 1.
2 2
Then
1
u ( x) u ( y)
(2n ) N
for any x, y V . By switching the roles of x and y, we have
1
u ( y) u( x).
(2n ) N
This implies that
1
u( y) u( x) (2n ) N u( y).
(2n ) N

4.12 Energy methods


We now characterize the solution of the Dirichlet problem for the Poisson equation
as a minimizer of an appropriate energy functional.

Definition 4.49. Let w be a function in the class

A = {w C 2 () C () : w = g on }.

This is the class of admissible functions for the Dirichlet problem with the given
boundary values. The energy functional for the Poisson equation u = f is

1
Z
2
I ( w) = |w| w f dx.
2
CHAPTER 4. LAPLACE EQUATION 130

Theorem 4.50 (Dirichlets principle). Suppose that u C 2 () C () solves


the Dirichlet problem for the Poisson equation

u = f in ,
u = g on .

Then
I ( u) = min I (w).
wA
Conversely, if u A satisfies I ( u) = minwA I (w), then u is a solution of the
Dirichlet problem above for the Poisson equation.

T ERMINOLOGY: The Poisson equation is said to be the Euler-Lagrange


equation for the energy (or variational) integral above.

THE MORAL : A solution of the Poisson (and Laplace) equation is a minimizer


of an energy integral. Conversely, every minimizer of the energy integral above
is a solution of the Poisson equation. Thus a function is a solution to the Poisson
equation if and only if it is a minimizer of the energy integral. Observe that
there are only first order partial derivatives in the energy functional, but the
corresponding Poisson equation involves second order partial derivatives. For
example, the finite element method to compute solutions numerically is based on
this approach.

Proof. Let w A . By Greens first identitty


Z
0 = ( u f )( u w) dx
| {z }
=0
u
Z Z
= ( u w) dx + ( u ( u w) f ( u w)) dx
| {z }
=0
Z Z Z
= | u|2 dx u w dx f ( u w) dx

where we used the fact that u w = g g = 0 on . Thus


Z Z
(| u|2 u f ) dx = ( u w f w) dx

1 1
Z
|w|2 + | u|2 f w dx,
2 2
where the last inequality follows from the Cauchy-Schwarz inequality since
1 1
| u w| | u||w| | u|2 + |w|2 .
2 2
In the last step we used the fact that ab 12 a2 + 12 b2 . Thus we have showed that
I ( u) I (w) for every w A . Since u A , we have

I ( u) = min I (w).
wA
CHAPTER 4. LAPLACE EQUATION 131

Assume that I ( u) = minwA I (w). Let v C 0 () and define

() = I ( u + v), R.

Observe that for every R we have u + v A , because v vanishes on . Since


u minimizes the energy functional I we see that has a minimum at = 0. Thus
we must have 0 (0) = 0 provided the derivative exists. However,

1
Z
() = | u + v|2 ( u + v) f dx
2
1
Z
= ( u + v) ( u + v) ( u + v) f dx
2
1 2
Z
= | u|2 + |v|2 + u v u f v f dx.
2 2
By differentiating under the integeral, we obtain
Z
0 ( ) = |v|2 + u v v f dx.

By Greens first identity


Z
0
0 = (0) = ( u v v f ) dx

u
Z Z
= v dS v( u + f ) dx

Z
= v( u + f ) dx,

since v = 0 on . We have thus proved that


Z
( u + f )v dx = 0 for every v C 0 ().

Claim: This implies that u + f = 0 everywhere in .

Reason. Let x and C 0 (B(0, 1)), which satisfies Rn ( y) d y = 1. Define


R

1 y
( y) = , > 0.
n
If > 0 is small enough, then ( x y) is supported in a small neighborhood of
x and thus ( x ) C 0 (). Thus for all > 0 small enough
Z
0 = ( u( y) + f ( y)) ( x y) d y = (( u + f ) )( x)

But { }>0 is a family of good kernels so letting 0 we have

u( x) f ( x) = lim(( u + f ) )( x) = 0
0

for all x as long as f satisfies some mild assumptions, for example, f integrable,
bounded, and continuous in .
CHAPTER 4. LAPLACE EQUATION 132

4.13 Weak solutions*


In this section we consider another point of view to the energy methods. If
u C 2 () is a solution to the Laplace equation u = 0 and C 0 (), that is, is
a compactly supported smooth function on . Then by Greens first identity

u
Z Z Z
0 = u dx = u dx + dS,
|{z}
=0

so that Z
u dx = 0 (4.7)

for every C 0 (). On the other hand, if (4.7) holds, then the computation above
shows that Z
u dx = 0

for every C 0 (). This implies that

u = 0 in

if and only if (4.7) holds for every C 0 ().


On the other hand, by Greens first identity


Z Z Z
0 = u dx = u dx + u dS,

|{z}
=0

so that Z
u dx = 0 (4.8)

for every C 0 (). As above,
Z Z
u dx = u dx = 0

for every C 0 (). This shows that (4.8) holds for for every C 0 () if and
only if
u = 0 in .

Formulas (4.7) and (4.8) give two possible ways to define a generalized solution
to the Laplace equation. Observe, that in (4.7) we only need to assume that u
has the fist order derivatives. In (4.7) it is enough to assume that u is integrable.
With this interpretation, the function u does not have to have any derivatives at
all. This is in contrast with the standard definition of the Laplacian, where u has
to be twice differentiable.
CHAPTER 4. LAPLACE EQUATION 133

4.14 The Laplace equation in other coor-


dinates*
In this section we consider the three-dimensional space R3 . We have solved the
two-dimensional Laplace equation over rectangular domains and on the disc by
switching to polar coordinates. Among the most useful alternatives to Carte-
sian coordinates in R3 there are two coordinate systems that generalize polar
coordinates in R2 .

Cylindrical coordinates
Cylindrical coordinates ( r, , z) in R3 are defined by

x = r cos , y = r sin , z = z.

Figure 4.5: Cylindrical coordinates.

The 3-dimensional Laplace equation takes the form

2 u 1 u 1 2 u 2 u
+ + + =0
r2 r r r 2 2 z2
in cylindrical coordinates.
CHAPTER 4. LAPLACE EQUATION 134

Reason. By a direct computation of the derivatives, we have


x y x y
= cos , = sin , = r sin , = r cos .
r r
By the chain rule
u u x u y u u
= + = cos + sin ,
r x r y r x y
and again by the chain rule

2 u u u

= cos + sin
r2 r x r y
u x u y u x u y

= cos + + sin +
x x r y x r x y r y y r
2 u 2 u 2 u 2 u
= cos2 + sin cos + sin cos + sin2 .
x2 x y x y y2
With respect to , we have
u u x u y u u
= + = r sin + r cos .
x y x y
By the product and the chain rules for derivatives

2 u u u

= r sin + r cos
2 x y
u u x u y

= r cos r sin +
x x x y x
u u x u y

r sin + r cos +
y x y y y
u 2 u 2 u
= r cos + r 2 sin2 r 2 sin cos
x x2 x y
2
u u 2 u
r sin r 2 sin cos + r 2 cos2 .
y x y y2
Thus
2 u 1 u 1 2 u
+ +
r 2 r r r 2 2
2
2 u 2 u 2
2 u cos u sin u

= cos 2 + 2 sin cos + sin 2 + +
x x y y r x r y
2 2 2
cos u 2 u u sin u 2 u

+ + sin 2 2 sin cos + cos 2
r x x x y r y y
2
u 2 2 2
u u u 2
= cos2 + sin2 + sin + cos2

= + ,
x2 y2 x 2 y2
which gives

2 u 1 u 1 2 u 2 u 2 u 2 u 2 u
+ + + = + + = u = 0.
r2 r r r 2 2 z2 x2 y2 z2
CHAPTER 4. LAPLACE EQUATION 135

Spherical coordinates
Spherical coordinates ( r, , ) in R3 are defined by

x = r cos sin , y = r sin sin , z = r cos .

Figure 4.6: Spherical coordinates.

The 3-dimensional Laplace operator takes the form

2 u 2 u 1 2 u 1 2 u cot u
+ + 2 + + 2 =0
r2 r r r (sin )2 2 r 2 2 r
in spherical coordinates.

Reason. A direct differentiation gives


x x x
= cos sin , = r sin sin , = r cos cos ,
r
y y
= sin sin , = r cos sin , = r sin cos ,
r y
z z z
= cos , = 0, = r sin .
r

By the chain rule


u u x u u u z
= + +
u x r y r z r
u u u
= cos sin + sin sin + cos ,
x y z
CHAPTER 4. LAPLACE EQUATION 136

from which, again by the chain rule, we obtain

2 u u u u

= cos sin + sin sin + cos
r2 r x r y r z
u x u y u z

= cos sin + +
x x r y x r z x r
u x u y u z

+ sin sin + +
x y r y y r z y r
u x u y u z

+ cos + +
x z r y z r z z r
2 u 2 u 2 u
= cos2 sin2 + sin cos sin2 + cos sin cos
x2 x y x z
2 u 2 u 2 u
+ sin cos sin2 + sin2 sin2 + sin sin cos
x y y2 y z
2
u 2 u
2 u
+ cos sin cos + sin sin cos + cos2 2 .
x z y z z

With respect to , we have

u u x u y u z u u
= + + = r sin sin + r cos sin u y + 0,
x y z x y

and thus

2 u u u

= r sin sin + r cos sin u y
2 x y
u u x u y u z

= r cos sin r sin sin + +
x x x y x z x
u u x u y u z

r sin sin + r cos sin + +
y x y y y z y
u 2 u 2 u
= r cos sin + r 2 sin2 sin2 r 2 sin cos sin2 +0
x x2 x y
u 2 u 2 u
r sin sin r 2 sin cos sin2 + r 2 cos2 sin2 + 0.
y x y y2

With respect to , we have

u u x u y
u z
= + +
x y z
u u u
= r cos cos + r sin cos r sin ,
x y z
CHAPTER 4. LAPLACE EQUATION 137

from which we have


2 u u u u

= r cos cos + r sin cos r sin
2 x y z
u u x u y u z

= r cos sin + r cos cos + +
x x x y x z x
u x u y u z

r sin sin + r sin cos + +
x y y y z y
u u x u y u z

r cos r sin + +
z x z y z z z
u 2 u 2 u
= r cos sin + r 2 cos2 cos2 + r 2 sin cos cos2
x x2 x y
2
u
r 2 cos sin cos
x z
u 2 u 2 u
r sin sin + r 2 sin cos cos2 + r 2 sin2 cos2
y x y y2
2
u
r 2 sin sin cos
y z
u 2 u 2 u 2 u
r cos r 2 cos sin cos r 2 sin sin cos + r 2 sin2 .
z x z y z z2
Using the obtained expressions for the partial derivatives, we collect form the
formula
2 u 2 u 1 2 u 1 2 u cot u
+ + 2 + 2 + 2
r2 r r r (sin )
2 2 r 2 r
u
partial derivatives with respect to x, y and z. For the x
-terms, we have

2 1 1
0 + (cos sin ) + 2 ( r cos sin ) + 2 ( r cos sin )
r r (sin )2 r
cot
+ 2 ( r cos cos )
r
cos
= (2 sin2 1 sin2 + cos2 ) = 0.
r sin
u
Similarly, for the y
-terms, we have

2 1 1
0 + (sin sin ) + 2 ( r sin sin ) + 2 ( r sin sin )
r r (sin )2 r
cot
+ 2 ( r sin cos )
r
sin
= (2 sin2 1 sin2 + cos2 ) = 0
r sin
u
and for the z
-terms we have

2 1 1 cot
0+ cos + 2 0 + 2 ( r cos ) + 2 ( r sin )
r r (sin )2 r r
1
= (2 cos cos cos ) = 0.
r
CHAPTER 4. LAPLACE EQUATION 138

2 u
For the x2
-terms, we have

2 1 1 cot
cos2 sin2 + 0+ 2 ( r 2 sin2 sin2 ) + 2 ( r 2 cos2 cos2 ) + 2 0
r r (sin )2 r r
= cos2 (sin2 + cos2 ) + sin2 = 1.

2 u
Similarly, for the y2
-terms, we have

2 1 1 cot
sin2 sin2 + 0+ 2 ( r 2 cos2 sin2 ) + 2 ( r 2 sin2 cos2 ) + 2 0
r r (sin )2 r r
= sin2 (sin2 + cos2 ) + cos2 = 1

2 u
and for the z2
-terms, we have

2 1 1 cot
cos2 + 0+ 2 0 + 2 ( r 2 sin2 ) + 2 0 = 1.
r r (sin )2 r r

2 u
Moreover, for the x y
-terms, we have

2 1
2 sin cos sin2 + 0+ 2 (2 r 2 sin cos sin2 )
r r (sin )2
1 cot
+ 2 (2 r 2 sin cos cos2 ) + 2 0
r r
= 2 sin cos (sin2 1 + cos2 ) = 0.

2 u
Similarly, for the x z
-terms

2 1 1 cot
2 cos sin cos + 0+ 2 0 + 2 (2 r 2 cos sin cos ) + 2 0 = 0
r r (sin ) 2 r r

2 u
and for the y z
-terms

2 1 1 cot
2 sin sin cos + 0+ 2 0 + 2 (2 r 2 sin sin cos ) + 2 0 = 0.
r r (sin )2 r r

Collecting everything together gives

2 u 2 u 1 2 u 1 2 u cot u 2 u 2 u 2 u
+ + 2 + 2 + 2 = + + = 0.
r2 r r r (sin )
2 2 r 2 r x2 y2 z2

4.15 Summary
We have considered four aspects of the PDE theory for the Laplace operator.

(1) Existence. We have derived several representation formulas for the


solutions. The Fourier series and the Fourier transform can be used to
solve boundary value problems in special cases. The solution to the Poisson
CHAPTER 4. LAPLACE EQUATION 139

equation in the whole space can be represented as a convolution of the


source term with the fundamental solution by Theorem 4.17. Remark 4.19
shows how this can be used to solve a Dirichlet problem in a subdomain.
Theorem 4.23 gives a solution of a boundary value problems for the Poisson
equation in a subdomain in terms of Greens function for .
(2) Uniqueness. Remark 4.8 shows how we can conclude uniqueness of a
solution to the Dirichlet problem from Greens formulas. On the other hand,
uniqueness can be shown using the maximum principle as in Theorem
4.45.
(3) Stability. Theorem 4.43 shows that the solution of the Dirichlet problem
depends continuously on the boundary data. This can be concluded also
from the representation formulas whenever they are available.
(4) Regularity. Representation formulas for solutions are integrals which can
be used to show that the solutions are smooth. This can be also concluded
from the mean value property. It is possible to consider generalized or
energy solutions for the Laplace equations to obtain existence under very
general conditions. It can be shown that the generalized solutions are
smooth as well.
The heat equation governs the diffusion of heat in a body.
The heat kernel will be the fundamental solution of the
heat equation. The fundamental solution can be used to
derive representation formulas for solutions of nonhomo-

5
geneous problems. For subdomains the solutions can be
constructed through separation of variables, which leads
to an eigenvalue problem for the Laplace operator. We also
discuss the maximum principle to study uniqueness of the
solution.

Heat equation

In this chapter we study the heat equation


u
u = 0
t
and the nonhomogeneous heat equation
u
u = f
t
with appropriate initial and boundary conditions. Here
n 2 u
u =
X
,
j =1 x2j

that is, the Laplace is taken with respect to the spatial variable x. Let be an
open subset of Rn and T > 0. The problem is to find a function u = u( x, t) such that
it is a solution to the heat equation in (0, T ).
Physically, a solution u = u( x, t) of the heat equation represents the tempera-
ture of the body at the point x and time t. Observe that any solution v = v( x) of
the Laplace equation induces a time independent solution u = u( x, t) = v( x) of the
heat equation. In practice this suggests that for every claim about solutions to
the Laplace equation there should be a corresponding claim for the solutions of
the heat equation. However, the dependence in time leads to new phenomena and
challenges which are not visible in the stationary case.
The appropriate initial condition is

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

This describes the initial temperature distribution at the time t = 0. In addition,


we may have a Dirichlet type boundary condition

u( x, t) = g( x, t) x , t > 0,

140
CHAPTER 5. HEAT EQUATION 141

which describes the temperature on the boundary, or a Neumann type boundary


condition
u
( x, t) = x u( x, t) ( x) = h( x, t), x , t > 0,

which describes the heat flow through the boundary. Here, as usual, ( x) =
(1 ( x), . . . , n ( x)) is the unit outer normal on .

Remark 5.1. The more general equation

u
a u = 0,
t
where a R is the heat conductivity, can be reduced to the standard heat equation
by the change of variables t 7 at. This means that it is enough to consider the
case a = 1.

5.1 Physical interpretation


The heat equation, also known as the diffusion equation, governs the propagation
or diffusion in time of the density of some quantity such as heat or chemical
concentration. If V is any smooth subdomain of , the rate of change of the total
quantity of heat in V equals the negative of the net flux through the boundary V


Z Z
u( x, t) dx = F ( x, t) ( x) dS ( x) = 0,
t V V

where F = (F1 , . . . , F n ) is the flux density and is the unit outer normal of V . By
the Gauss-Green theorem we have
Z Z
div x F ( x, t) dx = F ( x, t) ( x) dS ( x)
V V
u
Z Z
= u( x, t) dx = ( x, t) dx.
t V V t

The last equality follows by switching the order of the derivative and the integral.
Since this holds for every subdomain V of , we have

u
div x F ( x, t) = ( x, t).
t
It is physically reasonable to assume that the flux F is proportional to the gradient
u but in the opposite direction, since the flow is usually from regions of high tem-
perature to regions of low temperature or high concentration to low concentration.
Thus
F ( x, t) = a u( x, t), a > 0.

This gives
u
( x, t) = div x F ( x, t) = adiv x u( x, t) = a u( x, t),
t
which implies that u is a solution of the general heat equation.
CHAPTER 5. HEAT EQUATION 142

5.2 The fundamental solution


In section 3.10 we have already derived a formula for the solution of the Cauchy
problem
u
u = 0 in Rn (0, ),
t
u = g on Rn { t = 0},

where g C 0 (Rn ). Indeed, by Theorem 3.26

1
Z Z
2
u( x, t) = ( x y, t) g( y) d y = e| x y| /(4 t) g( y) d y (5.1)
Rn (4 t)n/2 Rn
is a solution to the heat equation in Rn (0, ). Here
1

2
e| x| /(4 t) , x Rn , t > 0,


( x, t) = (4 t ) n/2
0, x Rn , t 0,

is the fundamental solution of the heat equation (or the heat kernel). Note that
is a solution to the heat equation in the upper half-space and that unbounded in
any neighbourhood of (0, 0). The fact that the solution assumes the initial values
g is a consequence of Theorem 3.19, which is the general result about families of
good kernels and the fact that the fundamental solution induces a family of good
kernels, see Example 3.20. This gives the existence of a solution to the Cauchy
problem, but the uniqueness may fail in unbounded domains, as we shall see later.
Note also that the solution given by (5.1) is smooth, since the fundamental solution
is smooth in the upper half-space and the convolution inherits smoothness.
Remarks 5.2:
(1) We can also write

= 0 in Rn (0, ),
t
= 0 in Rn { t = 0},

where 0 is the Dirac measure in Rn giving unit mass to the point 0. In


particular, is a solution to the heat equation in Rn (0, ).
(2) The heat equation has a natural symmetry, which may be used to derive
the formula for the fundamental solution. We may use the similarity
method to find , > 0 and v : Rn R so that the function

1 x

u( x, t) = v , x Rn , t > 0,
t t
is a solution to the heat equation. We can insert this in the heat equation
and assume that v is a radial function. This gives an ODE for v which
can be solved. This is an alternative way to derive the formula for the
fundamental solution, which applies to other PDEs as well.
CHAPTER 5. HEAT EQUATION 143

(3) If g C (Rn ) is bounded, g 0 and there is a point y Rn such that g( y) > 0,


then u > 0 everywhere in the upper half-space. This means that the heat
equation has infinite speed of propagation of disturbances. If the initial
temperature is nonnegative and positive somewhere, the temperature is
positive at any moment of time (no matter how small). Note that u( x, t) for
any t > 0 depends on the initial values of g( y) for all y Rn .
(4) By (3.26), we have

1
Z
| x y|2 /(4 t)

| u( x, t)| =
e g( y) d y
(4 t) n /2
R n

1
Z
| x y|2 /(4 t)
e g( y) d y

(4 t) n /2
Rn
1
Z
| g( y)| d y
(4 t)n/2 Rn

for every x Rn and t > 0. This gives the decay estimate

1
Z
sup | u( x, t)| | g( y)| d y, t > 0.
xRn (4 t)n/2 Rn

In particular, this implies that the solution tends to zero as time goes to
infinity. The same argument can also be used to study stability of solutions
in the upper half-space.

5.3 The nonhomogeneous problem


Let us consider the nonhomogeneous Cauchy problem
u
u = f in Rn (0, ),
t (5.2)
u = 0 in Rn { t = 0}.

By (5.1), for every fixed s, with 0 < s < t, the function


Z
u( x, t; s) = ( x y, t s) f ( y, s) d y
Rn

solves the translated initial value problem


u
( x, t; s) u( x, t; s) = 0, x Rn , t > s,
t
u( x, s; s) = f ( x, s), x Rn .

Observe that this is a initial value problem for the homogeneous heat equation
with the time t = 0 replaced by the time t = s. This is called the translation
principle.
CHAPTER 5. HEAT EQUATION 144

Figure 5.1: Nonhomogeneous problem for the heat equation.

So called Duhamels principle, see Remark 3.27, suggests that we can build a
solution to the nonhomogeneous problem by integrating solutions u( x, t; s) over
s (0, t) and have
Z t
u( x, t) = u( x, t; s) ds, x Rn , t > 0.
0

T HE MORAL : Duhamels principle is a process of expressing the solution of


a nonhomogeneous problem as an integral of the solutions to the homogeneous
problem in the way that the source term is interpreted as the initial condition.

To verify that this function is a solution to the heat equation, we have to


differentiate a function defined by an integral, as illustrated by the following
lemma.
Lemma 5.3.
u t u
Z
( x, t) = u( x, t; t) + ( x, t; s) ds.
t 0 t
CHAPTER 5. HEAT EQUATION 145

Figure 5.2: Duhamels principle.

Proof. A direct formal calculation shows that

u
Z t+h Z t
1

( x, t) = lim u( x, t + h; s) ds u( x, t; s) ds
t h 0 h 0 0
Z t+h Z t+ h
1

= lim u( x, t + h; s) ds u( x, t; s) ds
h 0 h 0 0
Z t+h Z t
1

+ lim u( x, t; s) ds u( x, t; s) ds
h 0 h 0 0
Z t+ h
1 1 t+ h
Z
= lim ( u( x, t + h; s) u( x, t; s)) ds + lim u( x, t; s) ds
h 0 0 h h0 h t
u
Z t
= u( x, t; t) + ( x, t; s) ds.
0 t

The following theorem claims that this gives a solution to the initial value
problem for the nonhomogeneous heat equation.

Theorem 5.4. Let f C 0 (Rn (0, )). The solution of the nonhomogeneous
Cauchy problem (5.2) is
Z tZ
u( x, t) = ( x y, t s) f ( y, s) d y ds
0 Rn
Z t | x y|2
1
Z
4( ts)
= e f ( y, s ) d y ds. (5.3)
0 (4( t s))
n/2
Rn
CHAPTER 5. HEAT EQUATION 146

Proof. We give a formal argument here. By the previous lemma, we have


u u
Z t
( x, t) = u( x, t; t) + ( x, t; s) ds
t 0 t
Z t
= f ( x, t) + u( x, t; s) ds
0
Z t
= f ( x, t) + u( x, t; s) ds
0

= f ( x, t) + u( x, t).

Moreover, by letting t 0 in (5.3), we see that the boundary condition is satisfied


using the approximation of the identity. The detailed proof is rather similar to the
proof of Theorem 4.17 and it will be omitted here.
Remarks 5.5:
(1) To solve the general problem
u
u = f in Rn (0, ),
t
u = g in Rn { t = 0},

we can use the same approach is in Remark 4.13 (3) for the Laplace
equation and write u = u 1 + u 2 with
u
1 u 1 = f in Rn (0, ),
t
u 1 = 0 in Rn { t = 0},

and u
2 u 2 = 0 in Rn (0, ),
t
u 2 = g in Rn { t = 0}.

We can combine (5.1) and (5.3) to find that


Z Z tZ
u( x, t) = ( x y, t) g( y) d y + ( x y, t s) f ( y, s) d y ds.
Rn 0 Rn
(2) Duhamels principle does not depend on the specific structure of the equa-
tion and it applies to other linear ODEs and PDE as well.

5.4 Separation of variables in Rn


In the previous section we derived a solution to an initial value problem for
the heat equation in the whole space Rn . The goal of this section is to derive
corresponding solutions in a subdomain. Let Rn be a bounded open set with a
smooth boundary. Consider the initial and boundary value problem
u

u = 0 in (0, ),


t

u = 0 on (0, ),

u = g on { t = 0}.


CHAPTER 5. HEAT EQUATION 147

Step 1 (Separation of variables): We separate variables and look for a


solution in the form
u( x, t) = v( t)w( x), x , t > 0. (5.4)

Then
u
( x, t) = v0 ( t)w( x) and u( x, t) = v( t)w( x)
t
and
u
0= ( x, t) u( x, t) = v0 ( t)w( x) v( t)w( x).
t
It follows that
v0 ( t) w( x)
=
v( t) w( x)
for every x and t > 0 such that w( x) 6= 0 and v( t) 6= 0. Since the left hand side
depends only on t and the right hand side depends only on x, both sides have to
be the same constant. Thus we arrive at
0
v ( t)
= , t > 0,



v( t)
w ( x )
= , x .



w( x)

Here the negative sign is a convention that will be apparent soon.

THE MORAL : The problem for the heat equation has been transformed to an
ODE and an eigenvalue problem for the Laplace equation.

Step 2 (Solution of the separated equations): The general solution of


0
v = v is
v( t) = ce t ,

where c is a constant.
Consider then the other equation w = w. We say that is an eigenvalue
of the (negative) Laplacian in , if there exists a solution w of the problem

w = w in ,
w = 0 on ,

so that w is not identically zero in . The function w is the eigenfunction corre-


sponding to the eigenvalue . Observe that Greens first identity gives
Z Z
|w|2 dx = w w dx

w
Z Z
= ww dx + w dS

|{z}
=0
Z
= w2 dx.

CHAPTER 5. HEAT EQUATION 148

This implies that 0 and this is explains the choice of the negative sign above.
In fact, we have > 0, since if = 0, then the Dirichlet problem above only has
the trivial solution w = 0. From (5.4) we conclude that

u( x, t) = ce t w( x)

is a solution to u
u = 0 in (0, ),
t (5.5)
u = 0 on (0, ),

with the initial condition u( x, 0) = cw( x) for x .


Step 3 (Solution of the entire problem): If j , j = 1, 2, . . . , is an eigenvalue
and w j , j = 1, 2, . . . , is the corresponding eigenfunction, then the linear combination

c j e j t w j ( x)
X
u( x, t) =
j =1

should be a solution to (5.5) with the initial condition



X
u( x, 0) = c j w j ( x).
j =1

If we can determine the coefficients c j , j = 1, 2, . . . , so that



X
c j w j ( x ) = g ( x ), (5.6)
j =1

then u is a solution of the original problem.


It is known (but out of the scope of this couse) that there are infinite number of
eigenvalues j > 0, j = 1, 2, . . . , with j as j . We arrange the eigenvalues
in the increasing order so that 0 < 1 < 2 < . . . . Moreover, the corresponding
eigenfunctions {w j }
j =1
can be chosen to be an orthonormal basis in L2 (). By
results in Chapter 1, this means that if g L2 (), then
Z
c j = g, w j = g( y)w j ( y) d y, j = 1, 2, . . . .

Here , denotes the standard inner product in L2 (). The coefficients c j , j =


1, 2, . . . , can be seen as the Fourier coefficients of g L2 () and the series

X
X
g= c jwj = g, w j w j
j =1 j =1

converges in L2 (), as in Chapter 1. In addition, the series



g, w j e j t w j ( x)
X
u( x, t) = (5.7)
j =1
CHAPTER 5. HEAT EQUATION 149

gives the solution to the original problem in L2 (). Thus we have the representa-
tion formula

g, w j e j t w j ( x)
X
u( x, t) =
j =1
Z
g( y)w j ( y) d y e j t w j ( x)
X
=
j =1
Z X

!
j t
= e w j ( x)w j ( y) g( y) d y
j =1
Z
= K ( x, y, t) g( y) d y,

where
e j t w j ( x)w j ( y)
X
K ( x, y, t) =
j =1

is the heat kernel in .

T HE MORAL : Existence of a solution can be proved using separation of


variables and eigenfunction expansions in bounded domains.

This is precisely the same strategy as in the separation of variables in Chapter


1. Note that this approach depends on whether we are able to find many enough
eigenvalues and eigenfunctions so that we can represent the initial value as an
infinite linear combination of the eigenfunctions. Moreover, we have to verify that
the series above converges in an appropriate sense and that the representation
formula (5.7) above really is a solution to the original problem.

Remark 5.6. For time dependent problems it is also relevant to study the be-
haviour of solutions as t .
Claim:
k u(, t)kL2 () e1 t k gkL2 () , t > 0,

where 1 > 0 is the first (and the smallest) eigenvalue of Laplacian.

Reason. Since

X
j t
| u( x, t)| = g, w j e w j ( x )

j=1

| g, w j | e j t |w j ( x)|
X

j =1

e1 t
X
| g, w j ||w j ( x)|, (0 < 1 < 2 < . . . )
j =1
CHAPTER 5. HEAT EQUATION 150

by Parsevals equality we have



X
1 t
k u(, t)kL2 () e g, w j w j ()

j=1
L 2 ( )
!1/2

= e1 t | g, w j |2
X
j =1

= e1 t k gkL2 () .

In particular, the solution tends to zero as time goes to infinity. This method can
be used to study stability as well.

Remark 5.7. Similar argument also applies, for example, to the Neumann problem
u

u = 0 in (0, ),
t



u
= 0 on (0, ),



u = g on { t = 0}.

Physically this models an insulated boundary problem. This leads to the con-
struction of orthonormal eigenfunctions for the Laplacian with the zero Neumann
boundary conditions.

5.5 Maximum principle


The heat equation resembles the Laplace equation in the sense that its solutions
satisfy a maximum principle. To formulate this principle, let us introduce the
space-time cylinder
T = (0, T ),

where Rn is open and bounded and 0 < T < . By the maximum principle for
the solutions of the Laplace equation u = 0 we know that harmonic functions
achieve their maximum on the boundary . For the heat equation, the result
states that the maximum is achieved on certain part of the boundary, which is
called the parabolic boundary

T = ( [0, T ]) ( { t = 0}).

Observe that T consists of the base and the lateral sides of the cylinder T .
In this section we consider functions u C (T ), such that the second order
partial derivatives in space and the time derivative are continuous in T .

Theorem 5.8 (Weak maximum principle). Let Rn be open and bounded


and suppose that u is a solution to the heat equation in T . Then

max u( x, t) = max u( x, t).


( x,t)T ( x,t)T
CHAPTER 5. HEAT EQUATION 151

Figure 5.3: The maximum principle for the heat operator.

Remark 5.9. If we replace u by u we get the corresponding statement with min


replacing max. Observe that u solves the heat equation whenever u solves the
heat equation.

THE MORAL : A solution to the heat equation attains maximum and minimum
in T on the parabolic boundary T . This result holds for any bounded domain
without regularity assumptions on the boundary. Observe carefully that at a fixed
time t > 0 the function x 7 u( x, t) does not have to have maximum or minimum at
.

Proof. We shall prove the claim in two steps.


Step 1: Fix > 0 and set v( x, t) = u( x, t) t. We will first show that

max v( x, t) = max v( x, t)
( x,t)T ( x,t)T

and then conclude the claim by letting 0.


Now let us assume, to the contrary, that there is a maximum point in (0, T ],
that is,
max v( x, t) > max v( x, t).
( x,t)T ( x,t)T

Then there exists ( x0 , t 0 ) T \ T = (0, T ] such that

v( x0 , t 0 ) = max v( x, t).
( x,t)T
CHAPTER 5. HEAT EQUATION 152

In particular, we have v( x0 , t) v( x0 , t 0 ) for all t < t 0 . This implies that

v
( x0 , t 0 ) 0,
t
since otherwise the maximum would not be achieved at t = t 0 as a function of t.
On the other hand, v( x, t 0 ) v( x0 , t 0 ) for all x , since ( x0 , t 0 ) is a maximum of v
as a function of x. This implies that v( x0 , t 0 ) = 0 and

2 v
( x0 , t 0 ) 0 , j = 1, . . . , n,
x2j

(think of a function of one variable: if f has a local maximum at x0 and f is twice


differentiable, then f 00 ( x0 ) 0). By summing up, we have v( x0 , t 0 ) 0. Thus

v
( x0 , t 0 ) v( x0 , t 0 ) 0.
t
This is a contradiction with
v u
v = u = < 0.
t t
Note that the auxiliary function v was introduced only to get a strict inequality
here.
Step 2:

max u( x, t) = max (v( x, t) + t) ( v = u t)


( x,t)T ( x,t)T

max v( x, t) + T
( x,t)T

= max v( x, t) + T (Step 1)
( x,t)T

max u( x, t) + T. (v u)
( x,t)T

Letting 0 we have
max u( x, t) max u( x, t).
( x,t)T ( x,t)T

On the other hand, we always have

max u( x, t) max u( x, t)
( x,t)T ( x,t)T

and thus
max u( x, t) = max u( x, t).
( x,t)T ( x,t)T

Remark 5.10. The maximum principle gives a comparison principle and a stability
result as in the case of the Laplace equation (exercises).
CHAPTER 5. HEAT EQUATION 153

Theorem 5.11 (Uniqueness for bounded domains). Assume that T is bounded,


g C (T ) and f C (T ). Then there exists at most one solution u of the initial
and boundary value problem
u
u = f in T ,
t
u = g on T .

T HE MORAL : The initial and boundary values can be given only on the
parabolic boundary T . The equation determines the values uniquely inside T
and on the top { t = T }, as in the representation formula (5.1).

Proof. Let u, v be two solutions to the problem. Then w = u v satisfies


w
w = 0 in T ,
t
w = 0 on T .

By the weak maximum principle applied to w we see that w 0 in T . Now


applying the weak maximum principle to w we get the opposite inequality w 0
in T . Thus w = 0 on T , which implies u = v in T .

Example 5.12. Let u : R [0, ) R,


x2 j j

2
( e1/ t ), x R,
X

t > 0,
u( x, t) = j=0 (2 j )! t
j

0, t = 0.

Then
x2 j j 2
( e1/ t )
X
lim u( x, t) = lim
t 0 t0 j =0 (2 j )! t j

x2 j j 2
lim j ( e1/ t ) = 0.
X
=
j =0 (2 j )! t
|
0 t {z }
=0

On the other hand,

2 u
X x2 j2 j 1/ t2
( x, t) = 2 j (2 j 1) (e )
x 2
j =1 (2 j )! t j
X x2 j2 j 1/ t2
= (e )
j =1 (2 j 2)! t
j

and
u x2 j j +1 2
( e1/ t )
X
( x, t) =
t j =0 (2 j )! t j +1


X x2( j1) j 1/ t2
= (e ).
j =1 (2( j 1))! t
j
CHAPTER 5. HEAT EQUATION 154

Thus u is a solution to

u t u xx = 0 in R (0, ),
u = 0 on R { t = 0}.

Since au is a solution to the same problem for every a R, we see that the
uniqueness fails for unbounded domains without extra assumptions. In this case,
the problem has infinitely many solutions. Observe that u = 0 is the physically
reasonable solution, but there are many nonphysical solutions as well. The
growth condition in the previous theorem excludes these nonphysical solutions.
The nonphysical solutions grow extremely fast as | x| .
Note that the definition of u( x, t) also works for t < 0, so that the backward
Cauchy problem
u t u xx = 0 in R (, 0),
u = 0 on R { t = 0}.
does not have a unique solution in general.
Remarks 5.13:
(1) There is also a version of the strong maximum principle for solutions of
the heat equation. Let be an open, bounded and connected set in Rn . If
there exists a point ( x0 , t 0 ) (0, T ] such that

u( x0 , t 0 ) = max u( x, t)
( x,t)T

then u is constant in t0 . The proof of the strong maximum principle is


beyond the scope of this course.
(2) By applying the strong maximum principle to the function u we get
a similar statement with min replacing max. The strong maximum (or
minimum) principle asserts that if u attains its maximum or minimum at
an interior point, then u is constant at all earlier times. This is physically
intuitive, since the solution will be constant on the time interval [0, t 0 ] if
the initial and boundary conditions are constant. However, the solution
may change at later times t > t 0 if the boundary conditions change. This
happens, for example, if we start heating the boundary.
(3) Changing t to t does not preserve the heat equation and the PDE distin-
guishes between solutions forward and backward in time. This corresponds
to the physical fact that heat conduction is, in general, irreversible, given
an initial temperature we may predict future temperatures, but we cannot
in general determine the thermal status that generated that particular
temperature distribution. The backward in time problem
u
u = 0 in Rn (, 0),
t
u = g on Rn { t = 0}.

CHAPTER 5. HEAT EQUATION 155

is illposed in the sense that, unlike the forward in time problem, it is not
solvable, in general, within the class of bounded solutions. Indeed, if a
bounded, continuous solution did exist for every choice g C (Rn ), then the
representation formula (5.1) gives
Z
g( x) = u( x, 0) = ( x y, T ) u( y, T ) d y
ZR
n

1 2
= e| x y| /(4T ) u( y, T ) d y.
(4T )n/2 Rn

This implies that g is smooth, which cannot be the case if g is only assumed
to be continuous.
Moreover, the backward in time problem is not stable. To see this, let > 0
and define x
2
u( x, t) = e t/ sin , x R, t < 0.

Then u is a solution to the problem
2
u u = 0 in R (, 0),


t x2
u = g on R { t = 0},

with x
g( x) = sin .

Then
max | g( x)| 0 as 0,
xR
but
max | u( x, t)| as 0.
xR
This means that the solution does not depend continuously on the boundary
values.

5.6 Energy methods for the heat equa-


tion
In this section we discuss how one can use energy methods to show, for example,
uniqueness of solutions to the heat equation. This discussion is analogous to the
energy methods applications for harmonic functions.
Let Rn be open and bounded. We consider the initial and boundary value
problem for the heat equation
u
u = 0 in T ,
t (5.8)
u = g on T .

CHAPTER 5. HEAT EQUATION 156

Theorem 5.14 (Uniqueness by energy methods). There exists at most one


solution for (5.8).

Proof. Let u, v be two solutions to (5.8). Then w = u v satisfies



w w = 0 in ,
t T
w = 0 on .
T

For 0 t T we define the energy


Z
e ( t) = w( x, t)2 dx.

Then

Z Z
e 0 ( t) = w( x, t)2 dx = (w( x, t)2 ) dx
t t
w
Z Z
= 2w dx = 2ww dx (w satisfies the heat equation)
t
w
Z Z
=2 w dS 2 |w|2 dx (Greens first identity)

Z
2
= 2 |w| dx 0. (w = 0 on )

We then conclude that e0 ( t) 0, which implies that e( t) is a decreasing function, so


that 0 e( t) e(0) for all 0 t T . However e(0) = 0 since w = 0 on T . Thus
Z
e( t) = w( x, t)2 dx = 0 for all 0 t T

and thus w = 0 on T .

5.7 Summary
(1) We have derived several representation formulas for the solutions. The
Fourier series and the Fourier transform can be used to solve boundary
value problems in special cases. The solution to the Cauchy problem for
heat equation in the whole space can be represented as a convolution of
the source term with the fundamental solution.
(2) The nonhomogeneous Cauchy problem can be solved using Duhamels
principle by Theorem 5.4.
(3) Separation of variables and eigenvalue problems can be used to derive a
representation formula for the solutions in a subdomain, see Section 5.4.
(4) The heat equation has infinite propagation speed for disturbances.
(5) The heat equation smoothens the boundary data.
(6) The boundary and initial values can be given only on the parabolic bound-
ary of a bounded space time cylinder. The same applies to the maximum
principle, uniqueness and stability results.
CHAPTER 5. HEAT EQUATION 157

(7) The uniqueness fails for the Cauchy problem without extra growth as-
sumptions.
(8) Energy decreases and solutions decay to zero as t .
(9) Energy methods can be used to give a short proof of the uniqueness in a
bounded space time cylinder.
(10) If the direction of time changed, then the obtained backward in time heat
equation does not have a unique solution and the solution does not depend
continuously on the boundary data.
We study the wave equation in all dimensions, but with par-
ticular focus on the physically relevant cases of dimensions
one, two and three. dAlemberts formula gives a solution
of the Cauchy problem in the one-dimensional case. The

6
three-dimensional Cauchy problem is solved by a spherical
mean reduction to the one-dimensional case. Finally, the
two-dimensional problem is solved by Hadamards decent
method from three dimensions. Duhamels principle ap-
plies to the nonhomogeneous problem. Energy methods can
be used to prove the existence.

Wave equation

In this chapter we study the n-dimensional wave equation with particular atten-
tion to the physically relevant cases of dimensions one, two and three. It turns
out that the properties of the solutions depend on the dimension. We can think
of the wave equation as describing the displacement of a vibrating string (in
dimension n = 1), a vibrating membrane (in dimension n = 2) or an elastic solid
(dimension n = 3). In dimension n = 3 this equation also determines the behaviour
of electromagnetic waves in vacuum and the propagation of sound waves.
The ndimensional wave equation is

2 u
u = 0
t2
and the nonhomogeneous wave equation is

2 u
u = f .
t2
Here
n 2 u
u =
X
,
j =1 x2j

that is, the Laplace is taken with respect to the spatial variable x.

Remark 6.1. The general equation

2 u
c 2 u = 0,
t2
where c R, can be reduced to the standard wave equation by the change of
variables (exercise).

158
CHAPTER 6. WAVE EQUATION 159

6.1 Physical interpretation


If V is any smooth subdomain of , the acceleration in V is

2
Z Z
u( x, t) dx = F ( x, t) ( x) dS ( x) = 0,
t2 V V

where F = (F1 , . . . , F n ) is the net contact force acting on V through V and is the
unit outer normal of V . This follows from Newtons law, which asserts that the
mass times the acceleration is the force. The mass density is assumed to be one.
By the Gauss-Green theorem we have
Z Z
div x F ( x, t) dx = F ( x, t) ( x) dS ( x)
V V
2 2 u
Z Z
= u( x, t) dx = ( x, t) dx.
t2 V V t2

The last equality follows by switching the order of the derivative and the integral.
Since this holds for every subdomain V of , we have

2 u
div x F ( x, t) = ( x, t).
t2
It is physically reasonable to assume that the force F is proportional to the
gradient u but in the opposite direction. Thus

F ( x, t) = c2 u( x, t), c R.

This gives
2 u
( x, t) = div x F ( x, t) = c2 div x u( x, t) = c2 u( x, t),
t2
which implies that u is a solution of the general wave equation.

6.2 The one-dimensional wave equation


We have showed using the Fourier transform that the solution of the Cauchy
problem

2 u
u = 0 in Rn (0, ),



t2


u = g in Rn { t = 0}, (6.1)
u



= h in Rn { t = 0}.


t

for the n-dimensional wave equation is given by formula (3.5), which states that

b() sin(|| t) e ix d .
Z
u( x, t) = (2)n gb() cos(|| t) + h
Rn | |
CHAPTER 6. WAVE EQUATION 160

Remark 6.2. To consider initial value problems for the wave equation in a bounded
domain T = (0, ), we can use eigenfunction expansions as in section 5.4.

In the one-dimensional case with n = 1, we have

1 b() sin(|| t) e ix d
Z
u( x, t) = gb() cos(|| t) + h
2 R | |
1 sin( t) ix
Z
= gb() cos( t) + hb() e d
2 R
1 1
Z
= gb() ( e i t + e i t ) e ix d
2 R 2
1 b() 1 ( e i t e i t ) e ix d .
Z
+ h
2 R 2 i

Here we used the symmetry of the trigonometric functions and trigonometric


identities (2.10). Theorem 3.4 (6) and the Fourier inversion theorem 3.13 imply

1 1
Z
gb() ( e i t + e i t ) e ix d
2 R 2
1 1 1 1
Z Z
= e i t gb() e ix d + e i t gb() e ix d
2 2 R 2 2 R
1 1 1 1
Z Z
= g( x + t)() e ix d +
g( x t)() e ix d
2 2 R 2 2 R
1
= ( g( x + t) + g( x t)).
2
On the other hand,

1 b() 1 ( e i t e i t ) e ix d
Z
h
2 R 2 i
1 1 1 1
Z Z
= e i t H
b () e ix d e i t Hb () e ix d
2 2 R 2 2 R
1 1 1 1
Z Z
= H
( x + t)() e ix d H ( x t)() e ix d
2 2 R 2 2 R
1
= ( H ( x + t) H ( x t)),
2
where
b ( )
h
b () =
H .
i
Claim: H 0 ( x) = h( x) and
Z x Z x
0
H ( x) = H ( y) d y = h( y) d y

Reason. Theorem 3.5 implies

b() = i H
h b () = H
c0 ()

and thus by the Fourier inversion theorem 3.13, we obtain H 0 ( x) = h( x).


CHAPTER 6. WAVE EQUATION 161

We conclude that
1 1
u( x, t) = ( g( x + t) + g( x t)) + ( H ( x + t) H ( x t))
2 2
1 1 x+ t
Z
= ( g( x + t) + g( x t)) + h( y) d y.
2 2 x t
This is dAlemberts formula for the one-dimensional wave equation.

Theorem 6.3 (dAlemberts formula). Assume that g C 2 (R), h C 1 (R), and


define u by
1 1
Z x+ t
u( x, t) = ( g( x + t) + g( x t)) + h( y) d y.
2 2 x t
Then u is a solution to the Cauchy problem (6.1) in the one-dimensional case.

Proof. The proof is a straightforward calculation (exercise).

Example 6.4. If

jx

g( x) = sin , j = 1, 2, . . . , and h( x) = 0,
L
then Theorem 2.50 gives

jx jx

u( x, t) = sin cos .
L L
On the other hand, dAlemberts formula above gives

1 j ( x t ) j ( x t )

u( x, t) = sin + sin .
2 L L

Since sin a cos b = 21 (sin(a + b) + sin(a b)) the two solutions are the same.

Example 6.5. Assume that L = 1, g( x) = 0 and h( x) = x, 0 < x < 1. We take the odd
2 periodic extension of g( x) to the whole real line. By dAlemberts formula

1 x+ t 1
Z
u( x, t) = h( y) d y = ( H ( x + t) H ( x t)),
2 x t 2
where H is an antiderivative of h. For example, we can take
Z x
H ( x) = h( y) d y, 1 < x < 1.
1

Since h( y) = y, 1 < y < 1, we have


Z x
1 1
H ( x) = y d y = x2 , 1 < x < 1.
1 2 2
The solution of the problem is
1
u( x, t) = (( x + t)2 ( x t)2 ) = xt, 1 < x < 1.
4
We can take the 2-periodic extension of this function to the whole R.
CHAPTER 6. WAVE EQUATION 162

Remarks 6.6:
(1) dAlemberts formula gives existence. The solution u is of the form

u( x, t) = F ( x + t) + G ( x t)

for appropriate functions F and G . Conversely, every function of this


form is a solution to u tt u xx = 0 in R (0, ) (exercise). The solution is a
superposition of traveling waves. The function F ( x + t) can be thought of
as a wave traveling in time with the speed one. In other words, think of
the graph of the function F ( x) as a wave at time t = 0. Then at time t the
wave has moved to left with speed one and this is the graph of F ( x + t). On
the other hand, G ( x t) is a wave traveling to right with the speed one.

Figure 6.1: dAlemberts solution with h = 0.

(2) dAlemberts formula gives uniqueness.

Reason. Assume that u and v are solutions with the same initial values.
Then u v is a solution with the zero initial values and dAlemberts
formula implies u v = 0.

(3) dAlemberts formula gives stability.

Reason. Assume that u is a solution with the initial values g 1 and h 1 and
that v is a solution with the initial values g 2 and h 2 . dAlemberts formula
CHAPTER 6. WAVE EQUATION 163

gives

1 1
| u( x, t) v( x, t)| | g 1 ( x + t) g 2 ( x + t)| + | g 1 ( x t) g 2 ( x t)|
2 2
1 x+ t
Z
+ | h 1 ( y) h 2 ( y)| d y
2 x t
sup | g 1 ( y) g 2 ( y)| + t sup | h 1 ( y) h 2 ( y)|
yR yR

+ t = (1 + t),

if
sup | g 1 ( y) g 2 ( y)| and sup | h 1 ( y) h 2 ( y)| .
yR yR

This means that a small change in the initial data g and h will affect the
solution u only a small amount. Thus the solution depends continously on
the boundary data.
(4) If g C k (R) and h C k1 (R), then u C k (R (0, )), but is not in general
smoother. Thus the wave equation does not smoothen the solution as in
the case of the Laplace or the heat equation.
(5) The solution at the point ( x, t) depends only on the values of g and h on
the interval [ x t, x + t]. This is called the domain of dependence of ( x, t).
Thus the value u( x, t) is not affected by the choice of g and h outside that
interval. Conversely, for every x0 R, there is a conical region called the
range of influence of x0 . Physically this means that the disturbances or
signals propagate with a finite speed. Initial disturbance at x0 will not be
felt at a point x until time t = | x x0 |.
(6) dAlemberts formula makes sense also with initial values that are not
necessary even continuous. This is analogous to the mean value principle
for the harmonic functions, see section 4.9. Then the function u( x, t) is no
longer differentiable and it may not be a solution of the wave equation
in the classical sense. However, it is possible to consider so-called weak
solutions, but this is out of the scope of this course.

T HE MORAL : dAlemberts formula gives a complete answer in the one-


dimesional case. However, at this point it is not clear what happens in higher
dimensions.

Remark 6.7. Consider the initial and boundary value problem on the first quad-
rant
u u xx = 0 in R+ (0, ),
tt



u = g, u t = h on R+ { t = 0},



u = 0 on { x = 0} (0, ),
CHAPTER 6. WAVE EQUATION 164

Figure 6.2: The domain of dependence and the range of influence.

where g(0) = h(0) = 0. We solve this problem by the method of reflection and
extend u, g and h to R by an odd reflection

u( x, t), x 0, t 0,
u
e( x, t) =
u( x, t), x 0, t 0,

g( x), x 0,
ge( x) =
g( x), x 0,
and
h( x), x 0,
h
e ( x) =
h( x), x 0.

Then the problem becomes



u e xx = 0 in R (0, ),
e tt u
ue=g e, u e on R { t = 0}.
et = h

dAlemberts formula gives


1 1
Z x+ t
e( x, t) = ( g
u e( x + t) + g
e( x t)) + h
e( y) d y.
2 2 x t
By the definitions of the odd reflections, we have
1 x+ t

1
Z
( g( x + t) + g( x t)) + h( y) d y, x t 0,


u( x, t) = 2 2 Z x t
x + t (6.2)
1 1
( g( x + t) g( t x)) + h( y) d y, 0 x t.


2 2 x+ t
CHAPTER 6. WAVE EQUATION 165

This formula means that the initial displacement g splits into two parts, one
moving to right with speed one and one moving left with speed one. The latter
reflects off the point x = 0, where the string is held fixed.

Example 6.8. By Hadamards Example 4.44, the function

1
u ( x1 , x2 ) = sin( jx1 ) sinh( jx2 ), j = 1, 2, . . . ,
j2

is a solution to the Dirichlet problem





u( x1 , x2 ) = 0 in ( x1 , x2 ) R2 ,


u( x1 , 0) = 0, x1 R,
u 1


( x1 , 0) = sin( jx1 ), x1 R.


x2

j

Observe that for x2 > 0, we have

1
lim sinh( jx2 ) = ,
j j2

which shows that the solutions blows up even if the boundary values tend to zero
as j . Recall that sinh( x) = 12 ( e x e x ). This shows that stability fails for this
problem.
Next we modify Hadamards example so that it applies to the wave equation.
Consider u as a function on Rn (0, ) that is independent of variables x3 , . . . , xn
and t. This is the trivial extension of u to the upper half-space. Then

u tt u = 0 in Rn (0, )

with the initial conditions

u( x1 , 0, x3 , . . . , xn , t) = 0

and
u 1
( x1 , 0, x3 , . . . , xn , t) = sin( jx1 ).
x2 j
By taking j sufficiently large, the absolute value of the boundary data can be be
everywhere arbitrarily small, while the solution takes arbitrarily large values
even at points with x2 6= 0 as small as we wish.

T HE MORAL : This shows that stability fails for the Cauchy problem for to
the wave operator when n 2.

Remark 6.9. It is enough to solve the Cauchy problem only in the upper half-space
t 0. The reason is that the problem for the lower half-space t 0 can be reduced
to the problem in the upper half-space by switching t to t. Note that the wave
equation remains unchanged in this change of variables (exercise).
CHAPTER 6. WAVE EQUATION 166

6.3 Euler-Poisson-Darboux equation


In the higher dimensional case, the solutions of the wave equation do not have as
simple expression as dAlemberts formula above. In this section we consider the
method of spherical averages and use it to solve the Cauchy problem (6.1).

Figure 6.3: The Cauchy problem for the wave equation.

We shall use the method of spherical means and define the appropriate aver-
ages over spheres
1
Z
U ( x; r, t) = u( y, t) dS ( y),
|B( x, r )| B( x,r)
1
Z
G ( x; r ) = g( y) dS ( y),
|B( x, r )| B( x,r)
1
Z
H ( x; r ) = h( y) dS ( y).
|B( x, r )| B( x,r)

THE MORAL : The pointwise values of the functions are replaced by integral
averages over spheres.

Remember that these averages converge to the corresponding function as


r 0, when the function is continuous, that is,
1
Z
lim U ( x; r, t) = lim u( y, t) dS ( y) = u( x, t), (6.3)
r 0 r 0 |B( x, r )| B( x,r )

if u is continuous, see the proof of Theorem 4.17. Here |B( x, r )| = n( n) r n1 is


the ( n 1)-dimensional volume of the sphere B( x, r ), see Remark 4.20. Note that
CHAPTER 6. WAVE EQUATION 167

these integral averages are continuous functions of x and r 0. In this sense


the integral averages above can be seen as approximations of the functions u, g
and h. It can be shown by differentiating under the integral, that if u is a C k
as a function of x, then U has the same property. In other words, smoothness is
preserved in the integral averages.
It turns out that U ( x; r, t) satisfies a one-dimensional equation which can be
further transformed to the one-dimensional wave equation.

Lemma 6.10 (Euler-Poisson-Darboux equation). Let u be a solution of the


Cauchy problem (6.1). Then for every fixed x Rn , we have

U tt Urr n 1 Ur = 0 in R+ (0, ),

r (6.4)
U = G, U t = H in R+ { t = 0}.

Figure 6.4: The Euler-Poisson-Darboux equation.

T HE MORAL : The original Cauchy problem for the n-dimensional wave


equation is transformed to a PDE in two variables r and t. If we can solve (6.4) at
every point x, then we can obtain the solution of (6.1) by letting r 0 as in (6.3).

Remark 6.11. Recall from Section 4.4 that the Laplace operator applied to a radial
function u( x) = v(| x|) = v( r ) is
n1 0
u( x) = v00 ( r ) + v ( r ).
r
Observe that this term also appears in the Euler-Poisson-Darboux equation above.
For a radial function u( x, t) = v(| x|, t) = v( r, t), we have U ( x; r, t) = v( r, t) and we
CHAPTER 6. WAVE EQUATION 168

have the Laplace equation in the spherical coordinates in the Euler-Poisson-


Darboux equation.

Proof. As in the proof of the mean value property for harmonic functions, see
Theorem 4.34, we have
r 1
Z
Ur ( x; r, t) = u( y, t) d y
n |B( x, r )| B( x,r)
1
Z
= u( y, t) d y
n( n) r n1 B( x,r)
1
Z
= u tt ( y, t) d y. ( u = u tt )
n( n) r n1 B( x,r)
This implies
rn
Z
r n1Ur ( x; r, t) = u tt ( y, t) d y
n|B( x, r )| B( x,r)
1
Z
= u tt ( y, t) d y
n( n) B( x,r)
Z rZ
1
= u tt ( y, t) dS ( y) d .
n( n) 0 B( x, )
Thus
1
Z
n1
(r Ur ( x; r, t)) = u tt ( y, t) dS ( y)
r n( n) B( x,r )

r n1
Z
= u tt ( y, t) dS ( y)
|B( x, r )| B( x,r )

= r n1U tt ( x; r, t),

where in the last equality we switched the order of the integral and the derivative.
On the other hand, by the product rule

( r n1Ur ( x; r, t)) = ( n 1) r n2Ur ( x; r, t) + r n1Urr ( x; r, t).
r
This gives

r n1U tt ( x; r, t) = ( n 1) r n2Ur ( x; r, t) + r n1Urr ( x; r, t),

which implies the Euler-Poisson-Darboux equation.


The claims for the initial values are relatively clear. Indeed,
1
Z
U ( x; r, 0) = u( y, 0) dS ( y)
|B( x, r )| B( x,r)
1
Z
= g( y) dS ( y) = G ( x; r )
|B( x, r )| B( x,r)
and
1
Z
U t ( x; r, t) = u( y, t) dS ( y)
t |B( x, r )| B( x,r)
1 u
Z
= ( y, t) dS ( y),
|B( x, r )| B( x,r) t
CHAPTER 6. WAVE EQUATION 169

from which it follows that


1 u
Z
U t ( x; r, 0) = ( y, 0) dS ( y)
|B( x, r )| B( x,r) t
1
Z
= h( y) dS ( y) = H ( x; r ).
|B( x, r )| B( x,r)

6.4 The three-dimensional wave equation


In this section we assume that n = 3 and that u is a solution to the Cauchy
problem (6.1). We recall the definitions of the integral averages U , G and H from
the previous section and denote

U
e = rU, G
e = rG and H
e = rH.

T HE MORAL : These are modified integral averages of the functions over


spheres.

Then

U
ett U
err = rU tt (U + rUr )r

= rU tt (Ur + Ur + rUrr )
= rU tt 2Ur rUrr
2
= r (U tt Urr Ur ) = 0
r
in R+ (0, ) by the Euler-Darboux-Poisson equation (6.4) with n = 3. Furthermore
the initial conditions become
r
Z
U
e ( x; r, 0) = rU ( x; r, 0) = u( y, 0) dS ( y)
|B( x, r )| B( x,r)
r
Z
= g( y) dS ( y) = rG ( x; r ) = G
e ( x; r )
|B( x, r )| B( x,r)

and, in the same way,

r
Z
U
et ( x; r, t) = rU t ( x; r, t) = u( y, t) dS ( y)
t |B( x, r )| B( x,r)
r u
Z
= ( y, t) dS ( y),
|B( x, r )| B( x,r) t

from which it follows that


r u
Z
U
et ( x; r, 0) = ( y, 0) dS ( y)
|B( x, r )| B( x,r) t
r
Z
= h( y) dS ( y) = rH ( x; r ) = H
e ( x ; r ).
|B( x, r )| B( x,r)
CHAPTER 6. WAVE EQUATION 170

Moreover,

r
Z
U
e ( x; 0, t) = lim u( y, t) dS ( y) = u( x, t) lim r = 0.
r 0 |B( x, r )| B( x,r ) r 0
| {z }
=0

Thus we have reduced the problem to the wave equation in the first quadrant

Ue U err = 0 in R+ (0, ),
tt


Ue = G,
e U e on R+ { t = 0},
et = H



Ue = 0 on { r = 0} (0, ).

T HE MORAL : The original Cauchy problem for the three-dimensional wave


equation is transformed to a problem for the one-dimensional wave equation. We
can solve this problem by dAlemberts formula at every point x and we can obtain
the solution to the original problem by considering

rU ( x; r, t) U
e ( x; r, t)
u( x, t) = lim U ( x; r, t) = lim = lim .
r 0 r 0 r r 0 r

Figure 6.5: The one-dimensional problem for U


e.

Applying the formula (6.2), we obtain

1 r+ t e

1 e
Z
(G ( x; r + t) + G ( x; r t)) + H ( x; y) d y, r t 0,

e
Ue ( x; r, t) = 2 2 Zr t
r+ t
1 1
(Ge ( x; r + t) G
e( x; t r )) + H
e ( x; y) d y, 0 r t.


2 2 r+ t
CHAPTER 6. WAVE EQUATION 171

Here we are interested in the solution for 0 r t since we want to eventually


pass r 0. Thus we have

Ue ( x; r, t)
u( x, t) = lim
r 0 r
!
G
e ( x; r + t) G
e ( x; t r ) 1 t+ r e
Z
= lim + H ( x; y) d y
r 0 2r 2 r t r
=Ge t ( x; t) + He ( x; t)
t t
Z Z
= g( y) dS ( y) + h( y) dS ( y).
t |B( x, t)| B( x,t) |B( x, t)| B( x,t)
On the other hand,
1 1
Z Z
g( y) dS ( y) = g( y) dS ( y) (|B( x, t)| = 4 t2 )
|B( x, t)| B( x,t) 4 t2 B( x,t)
1
Z
= g( x + tz) t2 dS ( z)
4 t2 B(0,1)
( y = x + tz, dS ( y) = t2 dS ( z))
1
Z
= g( x + tz) dS ( z) (|B(0, 1)| = 4)
|B(0, 1)| B(0,1)
and thus
1 1
Z Z
g( y) dS ( y) = ( g( x + tz)) dS ( z)
t |B( x, t)| B( x,t) |B(0, 1)| B(0,1) t
1
Z
= g( x + tz) z dS ( z)
|B(0, 1)| B(0,1)
1 y x
Z
= g ( y) dS ( y).
|B( x, t)| B( x,t) t
( y = x + tz, dS ( z) = t2 dS ( y))

This gives
1 1
Z Z
u( x, t) = g( y) dS ( y) + t g( y) dS ( y)
|B( x, t)| B( x,t) t |B( x, t)| B( x,t)
t
Z
+ h( y) dS ( y) (6.5)
|B( x, t)| B( x,t)
1
Z
= ( th( y) + g( y) + g( y) ( y x)) dS ( y), x R3 , t > 0.
|B( x, t)| B( x,t)
This is Kirchhoffs formula for the solution of the Cauchy problem (6.1) for the
three-dimensional wave equation.
Remarks 6.12:
(1) To compute u( x, t) we only need information on the data on the sphere
B( x, r ), not the entire ball B( x, r ). In other words, the domain of depen-
dence of a point ( x, t) is the surface of the sphere B( x, t) in R3 . Similarly,
the range of influence of a point x0 R3 is the surface of the (light) cone

{( x, t) R3 (0, ) : | x x0 | = t}.
CHAPTER 6. WAVE EQUATION 172

Figure 6.6: The domain of dependence in the three-dimensional case.

Physically this is a finite propagation speed of the disturbances, and more


specifically, the existence of sharp signals for three-dimensional waves
such as light or sound. For example, an initial disturbance near x = 0 will
be observed at x0 R3 only at the time t = | x0 | and not after that. This is
called Huygens principle. This is on contrast with the one-dimensional
case and also with the two-dimensional case, as we shall see in the next
section.
(2) One way to express the finite speed of propagation us that the solution at
a given point is determined by the initial values in a bounded subset. An
important consequence is that the process of solving initial value problems
for the wave equation can be localized in space. This is in a strict contrast
with the heat equation, since it has infinite speed of propagation.
(3) Another difference to the one-dimensional case is the loss of regularity.
Indeed, if g C 2 (R3 ) and h C 1 (R3 ), then Kirchhoffs formula gives only
that u is C 1 -function with respect to the space variable. This is due to the
focusing of the initial irregularities to a smaller set.
CHAPTER 6. WAVE EQUATION 173

Figure 6.7: The range of influence in the three-dimensional case.

6.5 The two-dimensional wave equation


In this section we consider the Cauchy problem (6.1) for the wave equation when
n = 2. We shall use Hadamards method of descent and view the two-dimensional
wave equation as a special case of the three-dimensional problem in which the
third spatial variable x3 does not appear.
To set this up, let us assume that u is solution of the initial value problem
(6.1) when n = 2. Let us write

u
e( x1 , x2 , x3 , t) = u( x1 , x2 , t), ( x1 , x2 , x3 , t) R3 (0, ).

We also define ge( x1 , x2 , x3 ) = g( x1 , x2 ) and h


e( x1 , x2 , x3 ) = h( x1 , x2 ) in the same fash-
ion. Since u satisfies (6.1), we see that u
e satisfies

ue tt u
e = 0 in R3 (0, ),
(6.6)
ue=g e, u e =he on R3 { t = 0}.
t

T HE MORAL : We add a dummy variable to the two-dimensional problem to


obtain a three-dimensional problem so that we may use Kirchhoffs formula.

For x = ( x1 , x2 ) R2 let us write e


x = ( x1 , x2 , 0) R3 . Then, Kirchhoffs formula
(6.5) for the three-dimensional problem gives
1 t
Z Z
u( x, t) = u x, t) =
e( e t ge( e
y) dS ( e
y) + h y) dS ( e
e( e y).
t | B x, t)| Be(ex,t)
e( e | B x, t)| Be(ex,t)
e( e
CHAPTER 6. WAVE EQUATION 174

Figure 6.8: Propagation of disturbances near the origin in the three-dimensional


case.

Here B x, t) denotes the ball in R3 , centered at e


e( e x R3 and having radius t. We can
simplify the previous formula by observing that
1 1
Z Z
ge( e
y) dS ( e
y) = ge( e
y) dS ( e
y)
| B x, t)| Be(ex,t)
e( e 4 t2 Be(ex,t)

yB
Now if e y = ( y, y3 ) with y R2 , y3 R, this means that
x, t) and e
e( e

x| = t | y x|2 + y32 = t2 y32 = t2 | x y|2 .


ye
|e

Consider the upper hemisphere of B x, t) in R3 and let us denote it by B


e( e x, t)+ .
e( e
y in the hemisphere can be given in the form
Then every point e

y = ( y1 , y2 , ( y1 , y2 )),
e where ( y1 , y2 ) = ( t2 | x y|2 )1/2 .

Setting y = ( y1 , y2 ) R2 as before the previous display can be written as e


y = ( y, ( y))
when ey B x, t)+ . Then we have the following (general) formula for change of
e( e
variables Z Z
F(e
y) dS ( e
y) = F ( y, ( y))(1 + |( y)|2 )1/2 d y
B x,t)+
e( e B( x,t)
Applying this to our function ge we get
1 1
Z Z
g ( y) dS ( y ) = ge( y, ( y))(1 + |( y)|2 )1/2 d y.
4 t2 Be(ex,t)+ 4 t2 B( x,t)
e e e

1
Z
= g( y)(1 + |( y)|2 )1/2 d y,
4 t2 B( x,t)
CHAPTER 6. WAVE EQUATION 175

since ge is independent of the third variable. For the lower hemisphere we have
that y3 = ( t2 | x y|2 )1/2 = ( y). Since |( y)| = |(( y))|, we obtain

1 2
Z Z
g ( y ) dS ( y) = g( y)(1 + |( y)|2 )1/2 d y.
4 t2 Be(ex,t) 4 t2 B( x,t)
e e e

The factor two enters since B x, t) consists of two hemispheres.


e( e
We have
x j yj
( y) = , j = 1, 2,
yj ( t2 | x y|2 )1/2
so that
| x y|
|( y)| =
( t2 | x y|2 )1/2
and
1/2
| x y| 2 t

(1 + |( y)|2 )1/2 = 1 + = .
t | x y| 2
2 ( t2 | x y|2 )1/2
From this we conclude that
1 1 g ( y)
Z Z
ge( e
y) dS ( e
y) = dy
| B ( e
e x, t)| B(ex,t)
e 2 t B( x,t) ( t | x y|2 )1/2
2

t 1 g ( y)
Z
= d y.
2 |B( x, t)| B( x,t) ( t2 | x y|2 )1/2

In the same way, we have

1 t 1 h ( y)
Z Z
h y) dS ( e
e( e y) = 2 2 1/2
d y.
| B ( e
e x, t)| B(ex,t)
e 2 | B ( x, t )| B( x,t) ( t | x y| )

Thus the formula for the solution becomes


1 2 1 g ( y)
Z
u( x, t) = t d y
2 t |B( x, t)| B( x,t) ( t2 | x y|2 )1/2
t2 1 h( y)
Z
+ dy
2 |B( x, t)| B( x,t) ( t2 | x y|2 )1/2
= I1 + I2.

A final calculation will allow us to further simplify the formula given above. For
I 1 we change variables y = x + tz, d y = t2 dz, to get

1 g( y) t2 g( x + tz) 2
Z Z
t2 dy = t dz
|B( x, t)| B( x,t) ( t2 | x y|2 )1/2 t B(0,1) t(1 | z|2 )1/2
2

t g( x + tz)
Z
= dz.
B(0,1) (1 | z|2 )1/2

The product and chain rules together with a change of variables give

1 g( x + tz) t g( x + tz) z
Z Z
I1 = dz + dz
2 B(0,1) (1 | z|2 )1/2 2 B(0,1) (1 | z|2 )1/2
t g( y) t g ( y) ( y x )
Z Z
= dy+ d y.
2|B( x, t)| B( x,t) ( t2 | x y|2 )1/2 2|B( x, t)| B( x,t) ( t2 | x y|2 )1/2
CHAPTER 6. WAVE EQUATION 176

We thus arrive at the formula


1 1 tg( y) + t2 h( y) + t g( y) ( y x)
Z
u( x, t) = d y, x R2 , t > 0.
2 |B( x, t)| B( x,t) ( t2 | x y|2 )1/2
This is Poissons formula for the solution of the two-dimensional wave equation.

Remark 6.13. To compute u( x, t) we only need information on the data on the


entire ball B( x, r ), not just on the sphere B( x, r ) as in the two-dimensional case.
The domain of dependence of a point ( x, t) is the disk B( x, t) in R2 . Similarly, the
range of influence of a point x0 R2 is the interior of the cone

{( x, t) R2 (0, ) : | x x0 | t}.

Physically this is a finite propagation speed, and more specifically, the absence of
sharp signals for two-dimensional waves such as water waves. For example, an
initial disturbance near x = 0 will be felt at x0 R2 after the time t = | x0 | for ever.

Example 6.14. Bessel functions are special functions that arise, for example, in
the study of vibrating circular membranes. As we have seen Section 5.4 for the
heat equation, the separation variables leads to the eigenvalue problem

w = w

with an appropriate boundary condition. If = { x R2 : | x| < 1} is the unit disc, by


switching to polar coordinates (see Section 2.11), we obtain

2 u 1 u 1 u
+ + = u, 0 < r < 1, < ,
r2 r r r 2 2
for u = u( r, ). We separate variables as in Section 2.11 and look for a product
solution of the form
u( r, ) = A ( )B( r ).

Inserting this into the PDE and multiplying by r 2 , we obtain

r 2 A ( )B00 ( r ) + r A ( )B0 ( r ) + B( r ) A 00 ( ) = A ( )B( r ),

and so we have
r 2 B00 ( r ) + rB0 ( r ) + r 2 B( r ) A 00 ( )
= .
B( r ) A ( )
Consequently, we may rewrite the separated equations as two ODE

A 00 ( ) + A ( ) = 0,
r 2 B00 ( r ) + rB0 ( r ) + r 2 B( r ) B( r ) = 0,

where is the separation constant. The latter equation is the parametric version
of the modified Bessel equation of order zero and its solutions are called Bessel
functions. We do not develop this issue futher here.
CHAPTER 6. WAVE EQUATION 177

Figure 6.9: The domain of dependence and the range of influence in the two-
dimensional case.

6.6 The nonhomogeneous problem


In this section we study the nonhomogeneous Cauchy problem

u tt u = f in Rn (0, ),
(6.7)
u = 0, u = 0 in Rn { t = 0}.
t

The goal is to apply Duhamels principle as in Section 5.3 for the heat equation.
Then, for every fixed s, let the function u( x, t; s) be the solution of the Cauchy
problem
u tt ( x, t; s) u( x, t; s) = 0, x Rn , t > s,
u( x, s; s) = 0, u t ( x, s; s) = f ( x, s), x Rn .
Observe that this is just the homogeneous initial value problem for the wave
equation with the time t = 0 replaced by the time t = s. Duhamels principle sug-
gests that we can build a solution to the nonhomogeneous problem by integrating
solutions u( x, t; s) over s (0, t)
Z t
u( x, t) = u( x, t; s) ds, x Rn , t > 0.
0

A formal calculation shows that


Z t Z t
u t ( x, t) = u( x, t; t) + u t ( x, t; s) ds = u t ( x, t; s) ds
| {z } 0 0
=0
CHAPTER 6. WAVE EQUATION 178

and
Z t
u tt ( x, t) = u t ( x, t; t) + u tt ( x, t; s) ds
0
Z t
= f ( x, t) + u tt ( x, t; s) ds
0
Z t
= f ( x, t) + u( x, t; s) ds
0
Z t
= f ( x, t) + u( x, t; s) ds
0

= f ( x, t) + u( x, t).

Moreover, inserting t = 0 in the formula for u, we obtain u( x, 0) = u t ( x, 0) = 0. Thus


u defined as above is a solution to (6.7).

Remark 6.15. The nonhomogeneous Cauchy problem with more general boundary
values can be solved by summing up the solutions to (6.1) and (6.7) as in Remark
5.5 (1) for the heat equation and Remark 4.13 (3) for the Laplace equation.

T HE MORAL : Duhamels principle is a process of expressing the solution of


a nonhomogeneous problem as an integral of the solutions to the homogeneous
problem in the way that the source term is interpreted as the initial condition.
This gives a method to solve a nonhomogeneous PDE with nonzero boundary
values as a sum of the nonhomogeneous PDE with zero boundary values and the
homogeneous PDE with nonzero boundary values.

Examples 6.16:
(1) For n = 1 dAlemberts formula in Theorem 6.3 gives

1
Z x+( t s)
u( x, t; s) = f ( y, s) d y.
2 x( t s)

Thus
Z t
u( x, t) = u( x, t; s) ds
0
1 t x+ t s
Z Z
= f ( y, s) d y ds
2 0 x t+ s
Z t Z x+ r
1
= f ( y, t r ) d y dr (s = t r)
2 0 x r
1 t x+ s
Z Z
= f ( y, t s) d y ds.
2 0 x s

(2) For n = 3 Kirchhoff s formula (6.5) gives

ts
Z
u( x, t; s) = f ( y, s) dS ( y).
|B( x, t s)| B( x,ts)
CHAPTER 6. WAVE EQUATION 179

Thus
Z t
u( x, t) = u( x, t; s) ds
0
Z t ts
Z
= f ( y, s) dS ( y) ds
0 |B( x, t s)| B( x,t s)
Z tZ
1 f ( y, s)
= dS ( y) ds (|B( x, t s)| = 4( t s)2 )
4 0 B( x,ts) t s
1 t f ( y, t r )
Z Z
= dS ( y) dr ( r = t s)
4 0 B( x,r) r
1 f ( y, t | y x|)
Z
= d y.
4 B( x,t) | y x|

Example 6.17. We solve



u t u xx = e t cos x, x R, t > 0,
u( x, 0) = 0, u ( x, 0) = 0, x R.
t

By Duhamels principle, we consider solution u = u( x, t; s) to the auxiliary problem



u u xx = 0, x R, t > 0,
tt



u( x, s; s) = 0, x R,


u ( x, s; s) = es cos x, x R.

t

where s > 0 is fixed and t > s. By dAlemberts formula

es x+( ts)
Z
u( x, t; s) = cos y d y
2 x( ts)
es
= (sin( x + ( t s)) sin( x ( t s))
2
= es sin( t s) cos x.

The solution to the original problem is


Z t
u( x, t) = es sin( t s) cos x ds
0
Z t
= cos x es sin( t s) ds
0
1
= cos x( e t cos t + sin t).
2

6.7 Energy methods


Let Rn be open and bounded with a smooth boundary and 0 < T < . Recall
that we denote T = (0, T ) and T = ( [0, T ]) ( { t = 0}). Define the
energy
1
Z
( u t )2 + | u|2 dx,

e ( t) = 0 t T.
2
CHAPTER 6. WAVE EQUATION 180

This energy measures the first order regularity of a function. If the solution
develops a singularity so that the first order derivatives become unbounded, we
might expect the energy to become unbounded. On the other hand, if the energy
is constant, then such singularities become concentrated to a smaller and smaller
sets. We shall show that the wave equation conserves energy.
By differenting the energy, we have
2
1 u
Z
e 0 ( t) = + | u|2 dx
t
2 t
1 u 2 u
Z
= 2 + 2 u u dx
2 t t2 t
u 2 u u
Z
= u dx (Greens first identity)
t t2 t
Z
= u t ( u tt u) dx = 0,

provided
u tt u = 0 in T ,
u = 0 on T .

Thus e0 ( t) = 0, 0 t T , which implies that e( t) = e(0), 0 t T .

THE MORAL : The energy is preserved in the wave equation.

Theorem 6.18. There exists at most one solution to the problem



u u = f in T ,
tt



u = g on T ,


u = h in { t = 0}.

t

Proof. If u and v are solutions, then w = u v is a solution to



w w = 0 in T ,
tt



w = 0 on T ,


w = 0 in { t = 0}.

t

Then
1
Z
(w t )2 + |w|2 dx = e(0) = 0,

e ( t) =
2
from which it follows that
w
=0 and w = 0 in T .
t
This shows that w is constant in every component of T . Since the w = 0 on
{ t = 0} we conclude w = u v = 0 in T . This proves the claim.
CHAPTER 6. WAVE EQUATION 181

6.8 Epilogue
We have seen that nonhomogeneous linear PDEs can be solved by Duhamels
principle if we can solve the corresponding homogeneous PDEs. Thus we may first
focus on homogeneous problems only. Let us return to homogeneous Maxwells
equations


div E = 0,



div B = 0,


B
curl E = ,
t



2 E 1



curl B = c , c= p ,


t 0 0
discussed in Introduction. We have seen that each component of electric field
E = (E 1 , E 2 .E 3 ) satisfies the wave equation

2 E
c2 E = 0.
t2
Similarly, magnetic field B satisfies the same wave equation

2 B
c2 B = 0.
t2

Consider initial conditions E ( x, 0) = E 0 ( x) and B( x, 0) = B0 ( x), where the vector


fields E 0 ( x) and B0 ( x) are otherwise arbitrary except for the conditions

div E 0 = 0 and div B0 = 0,

which are the first two equations in Maxwells equations.


We shall derive a solution to homogeneous Maxwells equations in the three-
dimensional space by using Kirchhoffs formula for the solution of the wave
equation. Indeed, in the first set of problems we have seen that Maxwells equa-
tions can be solved by the wave equation. By Maxwells equations, electric field E
satisfies the initial conditions
E
E ( x, 0) = E 0 ( x) and ( x, 0) = c2 curl B0 ( x)
t
and magnetic field B satisfies the initial conditions

B
B( x, 0) = B0 ( x) and ( x, 0) = curl E 0 ( x).
t
Thus we have the Cauchy problems

2 E
c2 E = 0 in R3 (0, ),



t2


E = E 0 in R3 { t = 0},
E



= c2 curl B0 in R3 { t = 0}.


t

CHAPTER 6. WAVE EQUATION 182

and
2 B
c2 B = 0 in R3 (0, ),



t2


B = B0 in R3 { t = 0},
B



= curl E 0 in R3 { t = 0}.


t

By a change of variables and Kirchhoffs formula (6.5) we see that

1
Z
u( x, t) = ( ct h( y) + g( y) + g( y) ( y x)) dS ( y),
|B( x, ct)| B( x,ct)

x R3 , t > 0 and c > 0, is a solution to



u tt c2 u = 0 in R3 (0, ),
u = g, u = h in R3 { t = 0}.
t

Thus we have
1
Z
tc3 curl B0 ( y) + E 0 ( y) + E 0 ( y) ( y x) dS ( y)

E ( x, t) =
|B( x, ct)| B( x,ct)

for x R3 and t > 0. Similarly

1
Z
tc curl E 0 ( y) + B0 ( y) + B0 ( y) ( y x) dS ( y)

B( x, t) =
|B( x, ct)| B( x,ct)

for x R3 and t > 0. These formulas give a solution to Maxwells equations in R3 .

6.9 Summary
(1) We have derived several representation formulas for the solutions. The
Fourier series and the Fourier transform can be used to solve boundary
value problems in special cases.
(2) The behaviour of the solutions to the wave equation is different in dimen-
sions one, two and three.
(3) The nonhomogenerous Cauchy problem can be solved using Duhamels
principle.
(4) The wave equation has finite propagation speed for disturbances.
(5) There is no maximum principle for the wave equation.
(6) The wave equation does not smoothen the boundary data.
(7) The direction of time can be changed in the wave equation.
(8) The energy is preserved and thus there is no decay as t .
(9) The energy is preserved in the wave equation and this can be used to give
a short proof of the uniqueness in a bounded space time cylinder.
Notation and tools
7
(1) Rn = { x = ( x1 , . . . , xn ) : x j R, j = 1, 2, . . . , n} is the n-dimensional Euclidean
space and R = R1 is the real line.
(2) C = { x + i y : x, y R} is the complex plane. Here i is the imaginary unit for
which i 2 = 1. The real part of z = x + i y C is Re( z) = y and the imaginary
part is Im( z) = y. The complex conjugate of z = x + i y is z = x i y. Observe
that zz = | z|2 .
(3) x y = nj=1 x j y j , x = ( x1 , . . . , xn ), y = ( y1 , . . . , yn ) is the standard inner prod-
P

uct in Rn .
1
qP
n
(4) k xk = ( x x) 2 = 2
j =1 x j is the norm of x Rn .
(5) e j = (0, . . . , 0, 1, 0, . . . , 0) is the j th standard basis vector of Rn .
n+1
(6) R+ = {( x, t) : x Rn , t > 0} is the upper half space in Rn+1 .
(7) B( x, r ) = { y Rn : | x y| < r } is the open ball with the center x Rn and
radius r > 0.
(8) is an open subset of Rn , if for every x there is r > 0 such that
B( x, r ) , that is, every point x in is an interior point of .
(9) A point x belongs to the boundary of , denoted , if B( x, r ) 6= ; and
B( x, r ) (Rn \ ) 6= ; for every r > 0, that is, every ball with a positive radius
centered at the boundary intersects both the set and its complement.
(10) is smooth, if the boundary can be locally represented as a graph of a
smooth function.
(11) = is the closure of , that is, the closure of an open set is the
union of the set and its boundary.
(12) An open set is connected, if every pair of points in can be connected
by a piecewise linear path in .
(13) Rn is bounded if there is M < such that k xk M for every x .

183
CHAPTER 7. NOTATION AND TOOLS 184

Equivalently, there is r > 0 such that B(0, r ), that is, is bounded if it


is contained in a ball with a finite radius centered at the origin.
(14) The j th partial derivative of u is
u u( x + he j ) u( x)
( x) = lim , j = 1, 2, . . . , n,
x j h 0 h
provided the limit exists.

(15) u( x) = xu1 ( x), . . . , xun ( x) is the gradient of u at point x.
(16) Higher order derivatives are denoted by
|| u
( x), || = 1 + + n ,
x1 1 . . . x
n
n

where each component j is a nonnegative integer.


(17) If ( x) = (1 ( x), . . . , n ( x)) is a unit vector, then
u
( x ) = u ( x ) ( x )

is the derivative of u in the direction at point x. Note that
u
( x) = u ( x) e j , j = 1, 2, . . . , n.
x j
(18) C () = { u : R : u is continuous}.
(19) C () is the class of continuous functions in , which can be extended
continuously to .
(20) C k () = { u : R : u is k times continuously differentiable}.
(21) C k () the class of is k times continuously differentiable functions in ,
whose all partial derivatives up to order k can be extended continuously to
.
(22) C () = C k () is the class of smooth functions.
T
k=0
(23) The support of a function u : R is the set { x : u( x) 6= 0}.
(24) A function u : R is compactly supported in , if the support of u is
a bounded (and thus compact) subset of . Equivalently, u is compactly
supported in , if there exists an open an bounded open set 0 such that
0 and u( x) = 0 for every x \ 0 .
(25) C 0 (), C 0k () and C 0 () denote functions in the corresponding classes
with compact support.
(26) u C 0 (Rn ) is called a compactly supported smooth function in Rn . u
C 0 (Rn ), if u C (Rn ) and there is r > 0 such that u( x) = 0 for every
x Rn \ B(0, r ), that is, the function is are zero outside a ball with a finite
radius centered at the origin.
(27) For 1 p < , the space L p () consists of functions f : C such that
Z 1
p
p
k f k L p ( ) = | f ( x)| dx < .

CHAPTER 7. NOTATION AND TOOLS 185

For a continuous function f we set k f kL () = sup x | f ( x)|.


(28) A sequence of functions f j : C, j = 1, 2, . . . , converges to f in L p ()
1 p < , if k f i f kL p () 0 as j , that is,
Z
lim | f j ( x) f ( x)| p dx = 0.
j

(29) A sequence of continuous functions f j : C, j = 1, 2, . . . , converges uni-


formly in to f , if k f j f kL () 0 as j , that is,

lim sup | f i ( x) f ( x)| = 0.


j x

(30) Let f j : C, j = 1, 2, . . . , be a sequence of integrable functions such that


f i ( x) f ( x) for every x as j . If there exists an integrable function
g such that | f j ( x)| g( x) for every x and for every j = 1, 2, . . . , then the
Lebesgue dominated convergence theorem asserts that f is integrable and
Z Z Z
f ( x) dx = lim f j ( x) dx = lim f j ( x) dx.
j j

Theis means that the order of taking limits and integral can be switched if
there is an integrable majorant function g. Observe that the same g has
to do for all functions f j .
(31) The Lebesgue dominated convergence theorem can be used to show that
in certain ceases we can switch the order of integrals and limits. Assume
thar I R is an interval. Suppose that for every fixed t I there exists an
integrable function on . Thus we have f : I R, f = f ( x, t). For each
t I we denote Z
F ( t) = f ( x, t) dx.

Assume that or every y I , the function x 7 f ( x, t) is integrable in ,
the function t 7 f ( x, t) is continuous for every x at t 0 I and there
exists g L1 () such that | f ( x, t)| g( x) for every ( x, t) I . Then F is
continuous at t 0 , that is,
Z
lim F ( t) = lim f ( x, t) dx
t t 0 t t 0
Z Z
= lim f ( x, t) dx = f ( x, t) dx.
t t 0

(32) The Lebesgue dominated convergence theorem can be used to show that
in certain ceases we can switch the order of integrals and derivatives. We
use the same notation as in the previous item. Assume that for every
t I , the function x 7 f ( x, t) is integrable in , the function t 7 f ( x, t) is
differentiable for every x at every point t I and there exists h L1 ()
f

such that t ( x, t) h( x) for every ( x, t) I . Then F is differentiable at

every point t I and



Z Z
0
F ( t) = f ( x, t) dx = f ( x, t) dx.
t t
CHAPTER 7. NOTATION AND TOOLS 186

Z
(33) f , g = f ( x) g( x) dx is the standard inner product in L2 ().

Z 1 Z 1
1 2 2
(34) k f kL2 () = f , f 2 = f ( x) f ( x) dx = | f ( x)|2 dx is the L2 norm in

.
n
2
(35) |B( x, r )| = |B(0, r )| = r n |B(0, 1)| = ( n) r n , where ( n) = is the n-
( n2 +1)
R
dimensional volume of the unit ball B(0, 1). Here ( s) = 0 x s1 e x dx
is the gamma function
(36) The boundary of the ball B( x, r ) is the sphere B( x, r ) = { y Rn : | x y| = r }.
(37) |B( x, r )| = |B(0, r )| = r n1 |B(0, 1)| = ( n) r n1 , where ( n) = n( n) is the
( n 1)-dimensional volume of the unit sphere B(0, 1).
(38) The convolution f g : Rn C is
Z
( f g)( x) = f ( x y) g( y) d y,
Rn

whenever this integral exists.


(39) The space-time cylinder is T = (0, T ), where Rn is open and
bounded and 0 < T < .
(40) The parabolic boundary of T is T = ( [0, T ]) ( { t = 0}). Observe
that the parabolic boundary consists of the base and the lateral sides of
the cylinder T .
Bibliography

[1] N.H. Asmar, Partial Differential Equations with Fourier Series and
Boundary Value Problems, Second Edition, Pearson Prentice Hall, 2005.

[2] W.E. Boyce and R.C. DiPrima, Elementary Differential Equations and
Boundary Value Problems, 9th Edition, John Wiley & Sons, 2010.

[3] L.C. Evans, Partial Differential Equations, American Mathematical


Society, Second Edition, 2010.

[4] E. DiBenedetto, Partial Differential Equations, Birkhuser 1995.

[5] Q. Han, A Basic Course in Partial Differential Equations, American


Mathematical Society, 2011.

[6] R. McOwen, Partial Differential Equations. Methods and Applications.


Prentice-Hall, 1996.

[7] E.M. Stein and R. Sakarchi, Fourier Analysis. An Introduction. Prince-


ton University Press, 2003.

[8] A. Vasy, Partial Differential Equations. An Accessible Route through


Theory and Applications. American Mathematical Society, 2015.

187