See

discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/227467527

Applied Partial Differential Equations

Book · January 2003

Source: RePEc

CITATIONS READS

192 157

4 authors, including:

S. D. Howison A. A. Lacey
University of Oxford Heriot-Watt University
115 PUBLICATIONS 3,713 CITATIONS 93 PUBLICATIONS 1,971 CITATIONS

SEE PROFILE SEE PROFILE

Alexander Movchan
University of Liverpool
304 PUBLICATIONS 4,012 CITATIONS

SEE PROFILE

All content following this page was uploaded by A. A. Lacey on 14 July 2014.

The user has requested enhancement of the downloaded file. All in-text references underlined in blue
are linked to publications on ResearchGate, letting you access and read them immediately.
 An Introduction
to
Applied Partial Differential Equations

Marek Z. Elżanowski

Department of Mathematics and Statistics

Portland State University

 Contents

Preface ix

Chapter 1. Boundary Value Problems 1

1.1. Elastic Bar 1
1.2. The Green’s Function 5
1.3. Minimum Principle 13
1.4. Elastic Beam 18

Chapter 2. The Diffusion Equation 23

2.1. Heat Conduction 23
2.2. Separation of Variables 26
2.3. Uniqueness of Solutions 33
2.4. Fundamental Solutions 34
2.5. Burgers’ Equation 41

Chapter 3. Laplace’s Equation 45

3.1. Physical Interpretation 45
3.2. Separation of Variables 46
3.3. The Green’s Function 52
3.4. Minimum Principles 56

Chapter 4. Vibrations of One-Dimensional Medium 61

4.1. The Wave Equation 61
4.2. d’Alembert’s Solution 63
4.3. Nonlinear Waves 70
4.4. Classification of Linear Partial Differential Equations 75

Appendix A. Normed and Inner Product Vector Spaces 77

A.1. Inner Product Vector Spaces 77
vii
viii CONTENTS

A.2. Normed Vector Spaces 81

A.3. Complex Vector Spaces 84

Appendix B. Fourier Theory 87

B.1. Fourier Series 87
B.2. Differentiation and Integration of Fourier Series 92
B.3. Complex Fourier Series and the Delta Function 96
B.4. Convergence of the Fourier Series 99
B.5. Fourier Transform 109

Bibliography 117

Index 119
 Preface

These notes are written for a one-quarter (pilot) course in elementary partial
differential equations. It is assumed that the student has a good background in
calculus, vector calculus, and ordinary differential equations. No prior knowledge
of any partial differential equations concepts is assumed, nor any required. Some
familiarity with the elementary theory of inner vector spaces would be an asset
but is not expected. In fact, most of the needed concepts and facts are reviewed
in the Appendix.
The main objective of this presentation is to introduce basic analytic tech-
niques useful in solving most fundamental partial differential equations that arise
in the physical and engineering sciences. The emphasis are placed on the for-
mulation of a physical problem, deriving explicit analytic results, and on the
discussion of properties of solutions. Although the proofs are usually omitted,
the underlying mathematical concepts are explained and discussed at length.
The notes are divided into several short chapters and the Appendix. In Chap-
ter 1 we discuss solutions to the equilibrium equations of one-dimensional con-
tinuous systems. These are formulated as boundary-value problems for scalar
ordinary differential equations. The Green’s function technique and the mini-
mum principle are discussed. Chapter 2 deals with the diffusion equation, in
particular, the heat propagation equation. In the last section of this chapter
we briefly discuss the Burgers’ equation. Solutions to a variety of homogeneous
and inhomogeneous initial-boundary-value problems are derived using such an-
alytic techniques as the separation of variables method and the concept of the
fundamental solution. Laplace’s equation and the wave equation are dealt with
in Chapter 3 and 4, respectively. Once again, the separation of variables and
the Fourier series methods are utilized. The Green’s function technique is also
researched. d’Alembert’s solution of the wave equation is derived. An elementary
ix
x PREFACE

discussion on the propagation of one-dimensional non-linear waves is presented.

The Appendix consists of two parts. In the first part, the elements of the theory
of inner product vector spaces are reviewed. The second part contains the pre-
sentation of the theory of Fourier series, and a short section on Fourier integrals.
Although the notes are as self contained as possible, students may find use-
ful to consult some other texts like for example [Bleecker and Csordas],
[Boyce and DiPrima], [Keane], [Knobel], and [Davis], among others.

Portland. Oregon, June 2003

 CHAPTER 1

Boundary Value Problems

In this introductory chapter we discuss the equilibrium equations of one-

dimensional continuous systems. These are formulated as boundary value prob-
lems for scalar ordinary differential equations. We concentrate on deriving the
exact analytical formulae for the solutions of these equations giving us the nec-
essary inside into the physical processes we model.

1.1. Elastic Bar

By a bar we mean a finite length one-dimensional continuum that can only

be stretched or contracted (deformed, in short) in the longitudinal direction, and
is not allowed to bend in a transverse direction. Given a point x on a bar we
measure its deformation from the reference position by the displacement u(x).
That is, a material point which was originally at the position x has been moved
to the position x + u(x). We adopt the convention that u(x) > 0 means that the
material is stretched out, while u(x) < 0 describes a contraction by the amount
−u(x). We also assume that the left (top) end of the bar is fixed, i.e., u(0) = 0.
The internal forces experienced by the bar, known as stress, do not necessarily
depend on how much the bar is stretched as a whole but rather on how much
one material point is moved relative to the neighboring points. This relative
amount of elongation is measured by the strain. Consider two material points
(particles) occupying in the reference configuration positions x and x + 4x, re-
spectively, where 4x is the small section of the bar. When the bar experiences
the displacement u the section of length 4x gets stretched to the new length

[x + 4x + u(x + 4x)] − [x + u(x)] = 4x + [u(x + 4x) − u(x)] . (1.1.1)

The relative elongation of the segment 4x is

1
2 1. BOUNDARY VALUE PROBLEMS

u(x + 4x) − u(x)

. (1.1.2)
4x
Shrinking the segment of the bar to a point, we obtain the dimensionless strain
measure at the position x

u(x + 4x) − u(x) du

ε(x) ≡ lim = . (1.1.3)
4x→0 4x dx
The material the bar is made of is defined by the constitutive relation for the
stress. This constitutive law tells how the stress depends on the strain when the
bar undergoes a deformation. Here we shall only consider a linear relation which,
in fact, approximates the real case quite adequately as long as the strain is small.
If s(x) denotes the stress exerted on a material point which was at the reference
position x, we postulate

s(x) = c(x)ε(x), (1.1.4)

where c(x) measures the stiffness of the bar at material point x. If the bar is
homogeneous c(x) = c is constant.
We also postulate that the internal stresses of the deformed bar balance the
external forces imposed. That is, if f (x) denotes the external force applied at x,
we assume that

Z x+4x
s(x + 4x) − s(x) 1
+ f (s)ds = 0 (1.1.5)
4x 4x x

per unit length of a segment of the bar between positions x and x + 4x, where
f (x) > 0 if the bar gets stretched. Invoking the Mean Value Theorem and taking
the limit of the left hand side of (1.1.5) as 4x → 0 we obtain that

ds
f =− (1.1.6)
dx
everywhere along the bar. Substituting the constitutive law (1.1.4) into the
equation of balance of forces (1.1.6) and using the definition of the strain func-
tion (1.1.3) we obtain the equation of equilibrium
 1.1. ELASTIC BAR 3

µ ¶
d du
− c(x) = f (x), 0<x<l (1.1.7)
dx dx
for the linearly elastic bar of length (in the reference configuration) l. This is a
second-order ordinary differential equation for the displacement u(x). Its general
solution depends on two arbitrary constants. They can be uniquely determined
by the boundary conditions at the ends of the bar. For example, let

u(0) = 0, s(l) = c(l)ε(l) = c(l)u0 (l) = 0 (1.1.8)

as the left(top) end of the bar is fixed and the other end is assumed to be free
(stress free).

Example 1.1. Consider a homogeneous bar of unit length subjected to a

uniform force, e.g., a bar hanging from a celling and deforming under its own
weight. The equilibrium equation (1.1.7) takes the form

d2 u
−c= mg, (1.1.9)
dx2
where m denotes the mass of the bar and g is the gravitation constant. This is a
linear second order equation solution of which is

mg 2
u(x) = − x + ax + b. (1.1.10)
2c
The arbitrary integration constants a and b can be determined from the boundary
conditions (1.1.8). Namely,

mg
u(0) = b = 0, u0 (1) = −
+ a = 0. (1.1.11)
2c
The corresponding unique solution yields the parabolic displacement
µ ¶
mg x2
u(x) = x− (1.1.12)
c 2
and the linear strain

mg
ε(x) = (1 − x). (1.1.13)
c
4 1. BOUNDARY VALUE PROBLEMS

Note that the displacement is maximum at the bottom free end of the bar while
strain, and so the stress, are maximum at the fixed end. Note also that as the
boundary condition at the free end determines both the strain and the stress at
that end the equilibrium equation can be first solved uniquely for stress without
calculating the displacement. Such a mechanical configuration is known as stati-
cally determinate. This is in contrast with the problem in which the displacement
is prescribed at both ends of the beam, e.g.,

u(0) = 0, u(1) = r. (1.1.14)

The general solution to the equilibrium equation (1.1.7) takes the same parabolic
form (1.1.10). The unique solution satisfying the boundary conditions (1.1.14)
yields

mg
u(x) = (x − x2 ) + rx. (1.1.15)
2c
Once the displacement is available we can calculate the stress field

1
s(x) = mg( − x) + r. (1.1.16)
2
However, unlike as in case of the bar with the free end, the stress cannot be
determined without knowing the displacement. The equilibrium equation (1.1.7)
can be re-written in terms of stress

ds
− = mg, (1.1.17)
dx
but the integration constant in the general stress solution s(x) = −mgx + a
cannot be determined as there is no stress boundary condition available. Such a
mechanical configuration is called statically indeterminate.

Remark 1.2. Our equation (1.1.7) not only describes the mechanical equi-
librium of an elastic bar but it models also some other physical systems. For
example, this is the thermal equilibrium equation of a bar subjected to an ex-
ternal heat source. Indeed, if u(x) represents the temperature at the position x,
 1.2. THE GREEN’S FUNCTION 5

c(x) is the thermal conductivity of the material at x, and f (x) denotes the ex-
ternal heat source, then the energy conservation law yields (1.1.7)1. A boundary
condition u(l) = r corresponds to the situation when an end is kept at a fixed
temperature. u0 (l) = 0, on the other hand, describes a thermally insulated end.

1.2. The Green’s Function

The Green’s function method is one of the most important approaches to the
solution of boundary value problems. It relies on the superposition principle for
inhomogeneous linear equations. Namely, it builds the general solution out of the
solutions to a very particular set of concentrated inhomogeneities.
The superposition principle for a linear homogeneous differential equation
states that if u1 (x) and u2 (x) are solutions then every linear combination αu1 (x)+
βu2 (x) is also a solution, where α and β are arbitrary real numbers. Moreover,
if the functions f1 , . . . , fn represent the inhomogeneities (forcing terms) of the
linear differential equation

K[u] = fi , i = 1, . . . , n, (1.2.1)
where K[u] denotes the differential operator (the left hand side of an equation,
e.g., K[u] = −cu00 ), and if u1 (x), . . . , un (x) are the corresponding solutions then
the linear superposition α1 u1 (x) + α2 u2 (x) . . . + αn un (x) is a solution of

K[u] = α1 f1 + α2 f2 + . . . + αn fn (1.2.2)
for any choice of the constants α1 , . . . , αn .
Our objective here is to use this superposition principle to solve the boundary
value problem for a homogeneous elastic bar. To be able to do this we must solve
first the boundary value problem with the unit impulse as a source term. Such
a solution is called the Green’s function and it will be used later to construct a
solution to the corresponding boundary value problem with an arbitrary forcing
term. First, we shall characterize a unit impulse (a point force) concentrated at
a point of the bar by introducing the notion of the delta function.
1See the derivation of the heat conduction equation (2.1.6) in the next chapter.
6 1. BOUNDARY VALUE PROBLEMS

The Delta Function.

As the impulse is to be concentrated solely at a single point, say y, the delta
function δy (x) should be such that

δy (x) = 0, for x 6= y. (1.2.3)

Moreover, as we would like the strength of the impulse to be one, and there is no
other external force applied, we require that

Z l
δy (x)dx = 1, as long as 0 < y < l. (1.2.4)
0
Looking at both conditions δy (x) must satisfy one realizes quickly that there is
no such function.
The mathematically correct definition of such a generalized function, which
can for example be found in [Ziemer] (see also [Lang]), is well beyond the scope
of these notes. It relies on the assumption that for any bounded continuous
function u(x)

Z l
δy (x)u(x)dx = u(y), if 0 < y < l. (1.2.5)
0
Here, we will present the ”approximate” definition of the delta function which
regards δy (x), considered over the infinite domain (−∞, ∞), as the limit of a
sequence of continuous functions. To this end, let

n
fn (x; y) ≡ . (1.2.6)
π(1 + n2 (x − y)2 )
These functions are such that
Z ¯∞
∞
1 ¯
fn (x; y)dx = arctan(nx)¯¯ = 1, (1.2.7)
−∞ π −∞

and

0, x 6= 0
lim fn (x; y) = (1.2.8)
n→∞ ∞, x = 0
 1.2. THE GREEN’S FUNCTION 7

pointwise, but not uniformly2. Hence, we identify δy (x) with the limit

lim fn (x; y) = δy (x). (1.2.9)

n→∞

Note, however, that this construction of the delta function should only be viewed
as a visualization of such a generalized function, and not as its correct mathemat-
ical definition. Indeed, within the context of the Riemann’s integration theory
Z ∞ Z ∞
1 = lim fn (x; y)dx 6= lim fn (x; y)dx = 0 (1.2.10)
n→∞ −∞ −∞ n→∞

as the limit of the integral is not necessarily the integral of the limit. On the
other hand, this can be made to work if we adopt a somewhat different definition
of the limit. This can allow us to justify the formula (1.2.5) as the limit of the
approximating integrals. In these notes we will use both definitions of the delta
function interchangeably.
Let us consider now the calculus of the delta function, that is, its integration
and differentiation. Firstly, assuming that a < y and using the definition (1.2.5)
we obtain that

Z x 0, a < x < y,
δy (s)ds = σy (x) ≡ (1.2.11)
a 1, x > y > a,

is the step function. This is a function which is continuous everywhere except at

x = y, where its is not defined and experiences a jump discontinuity (1.2.15). The
value of the step function at x = y is often left undefined. Motivated by Fourier
theory3 we set σy (y) = 21 . Interestingly enough we obtain the same result using
the characterization of the delta function as the limit of the sequence fn (x; y).
Indeed, if
Z x
1 1
gn (x) ≡ fn (t, 0)dt = arctan(nx) + , (1.2.12)
−∞ π 2
then
2See Definition B.18.
3See Theorem B.4.
8 1. BOUNDARY VALUE PROBLEMS

lim gn (x) = σ(x) ≡ σ0 (x) (1.2.13)

n→∞

pointwise. In tern, the Fundamental Theorem of Calculus allows us to identify

the derivative of the step function with the delta function;

dσy (x)
= δy (x). (1.2.14)
dx
In fact, this enables us to differentiate any discontinuous function having finite
jump discontinuities at isolated points. Suppose the function f (x) is differentiable
everywhere except at a single point y at which it has a jump discontinuity

[f (y)] = f+ (y) − f− (y) ≡ lim+ f (x) − lim− f (x). (1.2.15)

x→y x→y

We can write

f (x) = g(x) + [f ]σy (x), (1.2.16)

where g(x) = f (x) − [f ]σy (x) is a continuous function, but is not necessarily
differentiable at y. Therefore,

g 0 (x), x 6= y
0
f (x) = (1.2.17)
[f ]δ (x), x = y.
y

In short, we write

f 0 (x) = g 0 (x) + [f ]δy (x). (1.2.18)

Example 1.3. Consider the function



 −x + 1, x < 0,


f (x) = 0, 0 < x < 1, (1.2.19)



 x2 , x > 1.
It has two jump discontinuities: [f ] = −1 at x = 0, and [f ] = 1 at x = 1.
Utilizing the construction (1.2.16) one gets that
 1.2. THE GREEN’S FUNCTION 9



−x + 1, x < 0,


g(x) = f (x) + σ(x) − σ1 (x) = 1, 0 < x < 1,



x2 , x>1
and



−1, x < 0,


f 0 (x) = g 0 (x) − δ(x) + δ1 (x) = −δ(x) + δ1 (x) + 0, 0 < x < 1,



2x, x > 1.

To find the derivative δy0 (x) of the delta function let us determine its effect on
a function u(x) by looking at the limiting integral

Z ∞ ¯∞ Z ∞
dfn (x; 0) ¯
lim u(x)dx = lim fn (x; 0)u(x)¯ − lim fn (x; 0)u0 (x)dx
n→∞ −∞ dx n→∞ −∞ n→∞ −∞
(1.2.20)
Z ∞
=− δ0 (x)u0 (x)dx = −u0 (0),
−∞

where the integration by parts was used and where the function u(x) is assume
continuously differentiable and bounded to guarantee that

lim fn (x; 0)u(x)|∞

−∞ = 0. (1.2.21)
n→∞

Hence, we postulate that δy0 (x) is a generalized function such that

Z l
δy0 (x)u(x)dx = −u0 (y). (1.2.22)
0
Note that this definition of the derivative of the delta function is compatible with
the formal integration by parts procedure

Z ¯l Z l
l ¯
δy0 (x)u(x)dx = δy (x)u(x)¯¯ − δy (x)u0 (x)dx = −u0 (y). (1.2.23)
0 0 0

Note also that one may view the derivative δy0 (x) as the limit of the sequence of
derivatives of the ”approximating” functions fn (x; 0). That is,
10 1. BOUNDARY VALUE PROBLEMS

dfn (x; 0) −2n3 π

δ00 (x) = lim = lim . (1.2.24)
n→∞ dx n→∞ π(1 + n2 x2 )2

These are interesting rational functions. First of all, it easy to see that the
sequence converges pointwise, but not uniformly, to 0. Also, elementary calcula-
tions revile, that the graphs of the functions consist of two increasingly concen-
1
trated symmetrically positioned at x = ± n√ 3
spikes, and that the amplitudes of
these spikes approach ∓∞, respectively, as n → ∞.

The Green’s Function.

Once we have familiarized ourselves with the delta function we may try now to
solve the boundary value problem for a homogeneous elastic bar with the delta
function (unit impulse) as its source term. As we have explained earlier, the
main idea behind this approach is to use the superposition principle to obtain
the solution for a general external force by putting together the solutions to the
impulse problems.
Consider a linearly elastic bar of the reference length l subjected to a unit
point force δy (x) applied at position 0 < y < l. The equation governing such a
system (1.1.7) takes the form
µ ¶
d du
− c(x) = δy (x), 0 < x < l. (1.2.25)
dx dx
The solution G(x; y) to the boundary value problem associated with (1.2.25) is
called the Green’s function of the problem. To illustrate how such a solution
comes about let us consider the following homogeneous boundary value problem

−u00 = δy (x), u(0) = u(1) = 0, (1.2.26)

for a bar of length l = 1 and the stiffness c = 1, where 0 < y < l. Integrating the
equation twice we obtain that

ax + b, x ≤ y,
u(x) = (1.2.27)
−(x − y) + ax + b, x ≥ y.

Taking into consideration the boundary conditions we have that

 1.2. THE GREEN’S FUNCTION 11

u(0) = b = 0, and u(1) = −(1 − y) + a + b = 0.

This implies that b = 0, a = 1 − y, and the Green’s function of this boundary
value problem is

x(1 − y), x ≤ y,
G(x; y) = (1.2.28)
y(1 − x), x ≥ y.

This is a continuous, piecewise differentiable function. Its first derivative experi-

ences a jump of magnitude −1 at x = y. In fact, this is a piecewise affine function
as its graph consists of straight line segments only. Note also that the Green’s
function, viewed as a function of two variables, is symmetric in x and y. This
symmetry has an interesting physical interpretation that the deformation of the
bar measured at position x due to the point force applied at position y is exactly
the same as the deformation of the bar at position y due to the concentrated
force being applied at position x.
Once we have the Green’s function available we can solve the general inho-
mogeneous problem

−u00 = f (x), u(0) = u(1) = 0, (1.2.29)

by the linear superposition method. To be able to do this we need first to ex-
press the forcing term f (x) as a superposition of point forces (impulses) dis-
tributed throughout the bar. The delta function comes handy again as, according
to (1.2.5), it enables us to write the external forcing term as
Z 1
f (x) = f (y)δx (y)dy. (1.2.30)
0
One may interpret the external source f as the superposition of an infinitely
many point sources f (y)δx (y) of the amplitude f (y) applied throughout the bar
at 0 < x < l. Re-writing the differential equation (1.2.29) as
Z 1
00
−u = f (y)δx (y)dy (1.2.31)
0
renders the solution u(x) as a linear combination
12 1. BOUNDARY VALUE PROBLEMS

Z 1
u(x) = f (y)G(x; y)dy (1.2.32)
0
of solutions to the unit impulse problems. We can verify by direct computation
that the formula (1.2.32) gives us the correct answer to the boundary value prob-
lem (1.2.29). Indeed, using the formula for the Green’s function (1.2.28) we may
write the solution of (1.2.29) as
Z x Z 1
u(x) = (1 − x)yf (y)dy + x(1 − y)f (y)dy. (1.2.33)
0 x
Differentiating it once gives us
Z 1 Z 1
0
u (x) = − yf (y)dy + f (y)dy.
0 x
Differentiating it once again shows that

u00 (x) = −f (x).

For a particular forcing term f (x) it may be easier to solve the problem directly
rather than by using the corresponding Green’s function. However, the advantage
of the Green’s function method is that it provides the general framework for any
and all inhomogeneous equations with the homogeneous boundary conditions.
The case of the inhomogeneous boundary value problem will be discussed in the
next chapter.

Example 1.4. Consider now a different boundary value problem for a uniform
bar of length l. Namely,

−cu00 (x) = δy (x), u(0) = 0, u0 (l) = 0, (1.2.34)

where c denotes the elastic constant. This problem models the deformation of the
bar with one end fixed and the other end free. Integrating this equation twice,
we find the general solution

1
u(x) = − ρ(x − y) + ax + b,
c
where
 1.3. MINIMUM PRINCIPLE 13


x − y, x > y,
ρ(x − y) ≡ (1.2.35)
0, x < y,
is called the (first order) ramp function. Utilizing the given boundary conditions
we find the Green’s function for this problem as

x/c, x ≤ y,
G(x; y) = (1.2.36)
y/c, x ≥ y.

Again, this function is symmetric and affine. Also, it is continuous but not
differentiable at the point of application of the external force x = y, where its
derivative experiences a −1/c magnitude jump. The formula for the solution of
the corresponding boundary value problem for the inhomogeneous equation

−cu00 (x) = f (x), u(0) = 0, u0 (l) = 0, (1.2.37)

takes the form
Z l ·Z x Z l ¸
1
u(x) = G(x; y)f (y)dy = xf (y)dy + yf (y)dy .
0 c 0 x

1.3. Minimum Principle

In this section we shall discuss how the solution to a boundary value problem
is a unique minimizer of the corresponding ”energy” functional. This minimiza-
tion property proves to be particularly significant for the design of numerical
techniques such as the finite elements method.
We start by taking a short detour to discuss the concept of an adjoint of a
linear operator on an inner product vector space 4 . Let L : U → W denote a lin-
ear operator from the inner product vector space U into another (not necessarily
different) inner product vector space W . An adjoint of the linear operator L is
the operator L? : W → U such that

hL[u]; wiW = hu; L? [u]iU for all u ∈ U, w ∈ W, (1.3.1)

4The fundamentals of Inner Product Vector Spaces are reviewed in Appendix A.
14 1. BOUNDARY VALUE PROBLEMS

where the inner products are evaluated on the respective spaces as signified by
the corresponding subscripts. Note that if U = W = Rn , with the standard dot
product, and the operator L is represented by a n × n matrix A, then the adjoint
L? can be identified with the transpose AT .
In the context of an equilibrium equation for a one-dimensional continuum the
main linear operator is the derivative D[u] = du/dx. It operates on the space of
all possible displacements U into the space of possible strains W . To evaluate its
adjoint we impose on both vector spaces the same standard L2 − inner product,
i.e.,
Z b Z b
hu; ũiU ≡ u(x)ũ(x)dx, hw; w̃iW ≡ w(x)w̃(x)dx. (1.3.2)
a a
According to (1.3.1) the adjoint D? of the operator D must satisfy

Z b Z b
du du
hD[u]; wiW = h ; wiW = w(x)dx = hu; D? [w]iU = u(x)D? [w](x)dx,
dx a dx a
(1.3.3)
for all u ∈ U and w ∈ W . Note, however, that the integration by part yields

Z b Z b
du dw
hD[u]; wiW = w(x)dx = [u(b)w(b) − u(a)w(a)] − u(x) dx (1.3.4)
a dx a dx
dw
= [u(b)w(b) − u(a)w(a)] + hu; iU .
dx
This suggests that
¶? µ
d d
=− , (1.3.5)
dx dx
provided the functions u ∈ U and w ∈ W are such that

[u(b)w(b) − u(a)w(a)] = 0. (1.3.6)

This will be possible if we impose suitable boundary conditions. In other words,
if we define the vector space

U = {u(x) ∈ C 1 [a, b] : u(a) = u(b) = 0} (1.3.7)

 1.3. MINIMUM PRINCIPLE 15

as containing all continuously differentiable functions (displacements) vanishing

at the boundary, and restrict the operator D to U , the definition of the ad-
joint (1.3.5) will hold. Obviously, these are not the only boundary conditions
guaranteeing (1.3.5). The other possible choice is the space of displacements

U = {u(x) ∈ C 1 [a, b] : u0 (a) = u0 (b) = 0}

Yet another possibility is the strain space

W = {w(x) ∈ C 1 [a, b] : w(a) = w(b) = 0}.

This is the case when due to the lack of support at the ends the displacement is
undetermined and the stresses vanish.
Consider now a homogeneous bar5 of length l, with material stiffness c = 1,
and suitable boundary conditions (guaranteeing (1.3.5)). Then, the equilibrium
equation (1.1.7) takes the form

K[u] = f, where K = D? ◦ D = −D2 (1.3.8)

is self-adjoint, that is, K ? = K. Indeed,

K ? = (D? ◦ D)? = D? ◦ D?? = D? ◦ D = K. (1.3.9)

Moreover, K is positive definite, where a linear operator K : U → U is said to
be positive definite if it is self-adjoint and

hK[u]; uiU > 0, for all 0 6= u ∈ U. (1.3.10)

To verify the positivity condition for K = D? ◦ D note that

hK[u]; uiU = hD? [D[u]]; uiU = hD[u]; D[u]iU = ||D[u]||2 ≥ 0, (1.3.11)

and its is positive if and only if D[u] = 0 only for u ≡ 0. This is not true in general.
However, if we impose the homogeneous boundary conditions u(0) = u(l) = 0
5To use the same framework for the analysis of a bar, or a beam, made of the inhomogeneous
material one needs to modify the inner product on the space of strains by introducing, as shown
in Section 1.4, a weighted L2 -inner product.
16 1. BOUNDARY VALUE PROBLEMS

the vanishing of the first derivative of u implies its vanishing everywhere, proving
that K is positive-definite. In fact, the same is true for the mixed boundary value
problem u(0) = u0 (l) = 0. However, this is not the case when u0 (0) = u0 (l) = 0 as
D[u] = 0 for a constant, non-vanishing displacement. The corresponding linear
operator K is self-adjoint but not positive-definite.
The minimum principle for the boundary value problem

−u00 = f, u(0) = u(l) = 0 (1.3.12)

can be formulated as the minimization of the functional

Z l· ¸
1 1 0
P[u] = ||D[u]||2 − hu; f iU = 2
(u (x)) − f (x)u(x) dx (1.3.13)
2 0 2
over the space of functions U = {u(x) ∈ C 2 [0, l] : u(0) = u(l) = 0}.

Remark 1.5. Formally, we seek a function, say u(x), from among the func-
tions belonging to the space U , such that

P[u] = min P[u].

u∈U

Suppose now that u is such a minimum and assume that

Z l
P[u] = f (x, u, u0 )dx.
0

Also, let w² (x) = u(x) + ²η(x) represent a curve of functions in the space U . Note
that this implies that the function η(x) vanishes at both ends of the interval [0, l].
Restricting the functional P to the curve w² consider the real-valued function

i(²) = P[w² ].

Since u is a minimizer of P[·] we observe that i[·] has a minimum at ² = 0.

Therefore
i0 (0) = 0.
Computing explicitly the derivative we obtain that
Z l· ¸
0 ∂f ∂f 0
i (²) = η(x) + η (x) dx.
0 ∂w ∂w0
 1.3. MINIMUM PRINCIPLE 17

Integrating the second term by parts and taking into account the boundary con-
ditions for the variation η(x), we get that
Z l · µ ¶¸
0 ∂f d ∂f
i (²) = η(x) − dx.
0 ∂w dx ∂w0
As it is valid for all variations η(x), it vanishes when
µ ¶
∂f d ∂f
− = 0. (1.3.14)
∂w dx ∂w0
This partial differential equation equation is known as the Euler-Lagrange equa-
tion. Its solution is the minimizer u. Although any minimizer of P[·] is a solution
of the corresponding Euler-Lagrange equation, the converse is not necessarily
true.

The functional P[u] represents the total potential energy of the bar due to
the deformation u(x). The first term measures the internal energy due to the
strain u0 (x) (the strain energy) while the second part is the energy due to the
external source f (x). The solution to (1.3.12) is the minimizer of P[u] over all
functions satisfying the given boundary conditions.

Example 1.6. To illustrate the importance of the positive-definiteness of

the given boundary value problem let us consider the following boundary value
problem in strains

−u00 = f, u0 (0) = u0 (l) = 0. (1.3.15)

Integrating the equation twice, we find

Z x µZ y ¶
u(x) = ax + b − f (s)ds dy. (1.3.16)
0 0
Since
Z x
0
u (x) = a − f (s)ds,
0
the boundary condition at x = 0 yields a = 0. The second boundary condition
at x = l implies that
18 1. BOUNDARY VALUE PROBLEMS

Z l
0
u (l) = f (s)ds = 0. (1.3.17)
0
This is not true in general, unless the source term has the zero mean. But even
if the distribution of external forces is such that the mean is zero, the solution
Z x µZ y ¶
u(x) = b − f (s)ds dy (1.3.18)
0 0
is not unique as the constant b remains unspecified. Physically, this corresponds
to an unstable situation. Indeed, if the ends of the bar are left free there exists
translation instability in the longitudinal direction.

1.4. Elastic Beam

In this short section we briefly discuss the use (possibly with some neces-
sary adaptation) of the methods developed in the previous sections to analyze
the deformation of an elastic (planar) beam. Here by a beam we understand
a one-dimensional continuum which in addition to being able to stretch in the
longitudinal direction is also allowed to bend in a plane, say (x, y). However, to
simplify matters, we will only consider that it can bend, neglecting its longitu-
dinal deformations. Consider therefore a beam of a reference length l, and let
y = u(x) denote the displacement in the transversal direction. As the beam bends
we postulate that its bending moment ω(x) is proportional to the curvature of
the beam

u00
κ(x) ≡ . (1.4.1)
(1 + u02 )3/2
Hence,

c(x)u00
ω(x) = c(x)κ(x) = . (1.4.2)
(1 + u02 )3/2
If we assume that u0 (x) is small, i.e., the beam does not bend too far from its
natural straight position, then the curvature is approximately equal to u00 (x), and
the linearized constitutive relation for the beam assumes the form

ω(x) = c(x)u00 (x). (1.4.3)

 1.4. ELASTIC BEAM 19

The linearized curvature κ(x) = u00 (x) plays the role of the (bending) strain
[Ogden].
Relying on the law of balance of moments of forces and using (1.4.3) we obtain
the equilibrium equation for the beam as the forth order ordinary differential
equation
µ ¶
d2 d2 u
c(x) 2 = f (x). (1.4.4)
dx2 dx
To be able to determine any particular equilibrium configuration, equation (1.4.4)
must be supplemented by a set of boundary conditions. As the equation is of
order four we need four boundary conditions; two at each end of the beam. For
example, we may assume that u(0) = ω(0) = ω(l) = ω 0 (l) = 0 which describes
the situation in which one end of the beam is simply supported while the other
is free.
Note that the balance law (1.4.4) can be viewed, with the proper choice of
the inner product and the boundary conditions, as written in the adjoint form.
To this end, let us consider the differential operator L ≡ D2 = D ◦ D, where as
d
before D ≡ dx
. The equilibrium equation (1.4.4) takes the form

L[cu00 ] = f. (1.4.5)
Let us also introduce the weighted inner product
Z l
hυ; υ̃i ≡ υ(x)υ̃(x)c(x)dx (1.4.6)
0
on the space of strains υ(x) ≡ κ(x) = u00 (x), where c(x) > 0. One can easily
check that this is indeed an inner product. To compute the adjoint operator L?
we need to evaluate
Z l Z l
d2 u
hL[u]; υi = cL[u]υdx = c υdx (1.4.7)
0 0 dx2
where differentiating by parts twice
Z · ¸¯l Z l 2
l
d2 u du d(cυ) ¯¯ d (cυ)
c 2 υdx = c υ − u ¯ + u dx. (1.4.8)
0 dx dx dx 0 0 dx2
20 1. BOUNDARY VALUE PROBLEMS

Consequently, if the functions u and υ are such that

· ¸¯l
du d(cυ) ¯¯
c υ−u = [u0 (l)ω(l) − u(l)ω 0 (l)] − [u0 (0)ω(0) − u(0)ω 0 (0)] = 0
dx dx ¯0
(1.4.9)
then

d2
L? [υ] =
(cυ). (1.4.10)
dx2
The equilibrium equation (1.4.4) can be written as

L? [υ] = f, (1.4.11)

and there is a quite a variety of possible self-adjoint boundary conditions as

determined by (1.4.9).
Although the beam operator L? ◦ L is not self-adjoint it is positive definite
for the appropriate boundary conditions, as evident from its form. As usually,
the key condition is that D2 vanishes only on the constant zero function of the
appropriate space of functions. Since D2 [u] vanishes if and only if u is affine,
the boundary conditions must be such that they force all affine functions to
vanish everywhere. For example, having one fixed end (u(0) = u0 (0) = 0) will be
sufficient, while having one simply supported end (u(0) = ω(0) = 0) and one free
end (ω(l) = ω 0 (l) = 0) will not do.
If the homogeneous boundary conditions are chosen so that the beam operator
L is positive definite, it can be shown that the solution to the boundary value
problem (1.4.11) is the unique minimizer of the corresponding energy functional

Z l· ¸
1 ¯¯ ¯¯2 1
P[u] = ¯¯D2 [u]¯¯ − hu; f i = 00
c(x)u (x) − f (x)u(x) dx. (1.4.12)
2 0 2

Example 1.7. Consider a uniform beam, with c(x) ≡ 1, of the reference

length l = 1 and such that one end is fixed and the other end is free. In case
there are no external forces the equilibrium equation (1.4.4) takes a very simple
form
 1.4. ELASTIC BEAM 21

d4 u
= 0. (1.4.13)
dx4
Its general solution

u = ax3 + bx2 + cx + d (1.4.14)

must satisfy the following boundary conditions:

u(0) = u0 (0) = ω(1) = ω 0 (1) = 0. (1.4.15)

This yields the solution

1
u = x2 (x − 3). (1.4.16)
6
To solve the forced beam problem we start by finding the appropriate Green’s
function. This means that we need to solve first the equation

d4 u
= δy (x) (1.4.17)
dx4
for the (1.4.15) boundary conditions. Integrating the equation four times and
using the fact that the integral of the delta impulse is the ramp function (1.2.35)
we obtain the general solution

 1 (x − y)3 , x > y,
u(x) = ax + bx + cx + d + 6
3 2
. (1.4.18)
0, x < y,
The boundary conditions imply that

u(0) = d = 0, u0 (0) = c = 0, ω(1) = 6a + 2b + 1 − y = 0, ω 0 (1) = 6a + 1 = 0.

(1.4.19)
Therefore, the Green’s function has the form

 1 x2 (y − x ), x < y,
G(x, y) = 2 3
(1.4.20)
 1 y 2 (x − y ), x > y.
2 3
22 1. BOUNDARY VALUE PROBLEMS

The Green’s function is again symmetric in x and y as the boundary value problem
we are dealing with is self-adjoint.
The general solution of the corresponding forced boundary value problem

d4 u
= f (x), u(0) = u0 (0) = ω(1) = ω 0 (1) = 0 (1.4.21)
dx4
is given by the superposition formula

Z 1 Z x Z 1
1 2 y 1 x
u(x) = G(x, y)f (y)dy = y (x − )f (y)dy + x2 (y − )f (y)dy.
0 2 0 3 2 x 3
(1.4.22)
 CHAPTER 2

The Diffusion Equation

In this chapter we study the one-dimensional diffusion equation

∂u ∂ 2u
= γ 2 + p(x, t),
∂t ∂x
which describes such physical situations as the heat conduction in a one-dimensional
solid body, spread of a die in a stationary fluid, population dispersion, and other
similar processes. In the last section we will also discuss the quasilinear version
of the diffusion equation, known as, the Burgers’ equation
∂u ∂u ∂ 2u
+u − γ 2 = p(x, t)
∂t ∂x ∂x
which arises in the context of modelling the motion of a viscous fluid as well as
traffic flow.
We begin with a derivation of the heat equation from the principle of the
energy conservation.

2.1. Heat Conduction

Consider a thin, rigid, heat-conducting body (we shall call it a bar) of length
l. Let θ(x, t) indicate the temperature of this bar at position x and time t, where
0 ≤ x ≤ l and t ≥ 0. In other words, we postulate that the temperature of the
bar does not vary with the thickness. We assume that at each point of the bar
the energy density per unit volume ε is proportional to the temperature, that is

ε(x, t) = c(x)θ(x, t), (2.1.1)

where c(x) is called heat capacity and where we also assumed that the mass
density is constant throughout the body and normalized to equal one. Although
the body has been assumed rigid, and with constant mass density, its material
properties, including the heat capacity, may vary from one point to another.
23
24 2. THE DIFFUSION EQUATION

To derive the ”homogeneous” heat-conduction equation we assume that there

are no internal sources of heat along the bar, and that the heat can only enter
the bar through its ends. In other words, we assume that the lateral surface
of the bar is perfectly insulated so no heat can be gained or lost through it.
The fundamental physical law which we employ here is the law of conservation
of energy . It says that the rate of change of energy in any finite part of the
bar is equal to the total amount of heat flowing into this part of the bar. Let
q(x, t) denote the heat flux that is the rate at which heat flows through the body
at position x and time t, and let us consider the portion of the bar from x to
x + 4x. The rate of change of the total energy of this part of the bar equals the
total amount of heat that flows into this part through its ends, namely
Z
∂ x+4x
c(z)θ(z, t)dz = −q(x + 4x, t) + q(x, t). (2.1.2)
∂t x
We use here commonly acceptable convention that the heat flux q(x, t) > 0 if the
flow is to the right.
In order to obtain the equation describing the heat conduction at an arbitrary
point x we shall consider the limit of (2.1.2) as 4x → 0. First, assuming that
the integrand c(z)θ(z, t) is sufficiently regular, we are able to differentiate inside
the integral. Second, dividing both sides of the equation by 4x, invoking the
Mean-Value Theorem for Integrals, and taking 4x → 0 we obtain the equation
∂θ ∂q
c(x) =− (2.1.3)
∂t ∂x
relating the rate of change of temperature with the gradient of the heat flux. We
are ready now to make yet another assumption; a constitutive assumption which
relates the heat flux to the temperature. Namely, we postulate what is known as
Fourier’s Law of Cooling, that the heat flows at the rate directly proportional to
the (spatial) rate of change of the temperature. If in addition we accept that the
heat flows, as commonly observed, from hot to cold we get that
∂θ
q(x, t) = −κ(x) . (2.1.4)
∂x
where the proportionality factor κ(x) > 0 is called the thermal conductivity.
Notice the choice of the sign in the definition of the heat flux guarantees that if
 2.1. HEAT CONDUCTION 25

the temperature is increasing with x the heat flux is negative and the heat flows
from right to left, i.e., from hot to cold.
Combining (2.1.3) and (2.1.4) produces the partial differential equation
∂θ ∂ ∂θ
c(x) = (κ(x) ), 0 < x < l, (2.1.5)
∂t ∂x ∂x
governing the heat flow in a inhomogeneous (κ is in general point dependent) one-
dimensional body. However, if the bar is made of the same material throughout,
whereby the heat capacity c(x) and the thermal conductivity κ(x) are point
independent, (2.1.5) reduces to

∂θ ∂ 2θ
= γ 2, 0 < x < l, (2.1.6)
∂t ∂x
where
κ
γ= . (2.1.7)
c

This equation is known as the heat equation, and it describes the evolution
of temperature within a finite, one-dimensional, homogeneous continuum, with
no internal sources of heat, subject to some initial and boundary conditions.
Indeed, in order to determine uniquely the temperature θ(x, t), we must specify
the temperature distribution along the bar at the initial moment, say θ(x, 0) =
g(x) for 0 ≤ x ≤ l. In addition, we must tell how the heat is to be transmitted
through the boundaries. We already know that no heat may be transmitted
through the lateral surface but we need to impose boundary conditions at the ends
of the bar. There are two particularly relevant physical types of such conditions.
We may for example assume that

θ(l, t) = α(t) (2.1.8)

which means that the right hand end of the bar is kept at a prescribed temperature
α(t). Such a condition is called the Dirichlet boundary condition. On the other
hand, the Neumann boundary condition requires specifying how the heat flows
out of the bar. This means prescribing the flux
26 2. THE DIFFUSION EQUATION

∂θ
q(l, t) = κ(l) (l, t) = β(t). (2.1.9)
∂x
at the right hand end. In particular, β(t) ≡ 0 corresponds to insulating the right
hand end of the bar. If both ends are insulated we deal with the homogeneous
Neumann boundary conditions.

Remark 2.1. Other boundary conditions like the periodic one are also pos-
sible.

2.2. Separation of Variables

The most basic solutions to the heat equation (2.1.6) are obtained by using
the separation of variables technique, that is, by seeking a solution in which the
time variable t is separated from the space variable x. In other words, assume
that
θ(x, t) = T (t)u(x), (2.2.1)

where T (t) is a x-independent function while u(x) is a time-independent function.

Substituting the separable solution into (2.1.6) and gathering the time-dependent
terms on one side and the x-dependent terms on the other side we find that the
functions T (t) and u(x) must solve an equation
T0 u00
=γ . (2.2.2)
T u
The left hand side of equation (2.2.2) is a function of time t only. The right hand
side, on the other hand, is time independent while it depends on x only. Thus,
both sides of equation (2.2.2) must be equal to the same constant. If we denote
the constant as −λ and specify the initial condition

θ(x, 0) = u(x), 0 ≤ x ≤ l, (2.2.3)

we obtain that
θ(x, t) = e−λt u(x) (2.2.4)

solves the heat equation (2.1.6) provided we are able to find u(x) and λ such that

−γu00 = λu (2.2.5)
 2.2. SEPARATION OF VARIABLES 27

along the bar. This is an eigenvalue problem for the second order differential
2
operator K ≡ −γ dtd 2 with the eigenvalue λ and the eigenfunction u(x). The
particular eigenvalues and the corresponding eigenfunctions will be determined
by the boundary conditions that u inherits from θ. Once we find all eigenval-
ues and eigenfunctions we will be able to write the general solution as a linear
combinations of basic solutions (2.2.4).

Homogeneous Boundary Conditions.

Let us consider a simple Dirichlet boundary value problem for the heat con-
duction in a (uniform) bar held at zero temperature at both ends, i.e.,

θ(0, t) = θ(l, t) = 0, t ≥ 0, (2.2.6)

where initially
θ(x, 0) = g(x), 0 < x < l. (2.2.7)
This amounts, as we have explained earlier, to finding the eigenvalues and the
eigenfunctions of (2.2.5) subject to the boundary conditions

u(0) = u(l) = 0. (2.2.8)

Notice first that as evident from the form of the equation (2.2.5) the eigenvalues
λ must be real. Also, it can be easily checked using the theory of second order
ordinary linear differential equations with constant coefficients that if λ ≤ 0,
then the boundary conditions (2.2.8) yield only the trivial solution u(x) ≡ 0.
Hence, the general solution of the differential equation (2.2.5) is a combination
of trigonometric functions

u(x) = a cos ωx + b sin ωx (2.2.9)

where we let λ = γω 2 with ω > 0. The boundary condition u(0) = 0 implies that
a = 0. Because of the second boundary condition

u(l) = b sin ωl = 0 (2.2.10)

ωl must be an integer multiple of π. Thus, the eigenvalues and the eigenfunctions

of the eigenvalue problem (2.2.5) with boundary conditions (2.2.8) are
µ ¶2
iπ iπ
λi = γ , ui (x) = sin x, i = 1, 2, 3, . . . . (2.2.11)
l l
28 2. THE DIFFUSION EQUATION

The corresponding basic solutions (2.2.4) to the heat equation are

µ ¶
γi2 π 2 iπ
θi (x, t) = exp − 2 t sin x, i = 1, 2, 3, . . . . (2.2.12)
l l
By linear superposition of these basic solutions we get a formal series
X∞ X∞ µ ¶
γi2 π 2 iπ
θ(x, t) = ai ui (x, t) = exp − 2 t sin x. (2.2.13)
i=1 i=1
l l
Assuming that the series converges we have a general series solution of the heat
equation with the initial temperature distribution
X∞
iπ
θ(x, 0) = g(x) = ai sin x. (2.2.14)
i=1
l

This is a Fourier sine series on the interval [0, l] of the initial condition g(x)1. Its
coefficients ai can be evaluated explicitly thanks to the remarkable orthogonality
property of the eigenfunctions. Indeed, it is a matter of a simple exercise on
integration by parts to show that
Z l
kπ nπ
sin x sin xdx 6= 0 (2.2.15)
0 l l
only if n = k, and that Z l
kπ l
x= .sin2 (2.2.16)
0 l 2
Multiplying the Fourier series of g(x) by the k-th eigenfunction and integrating
over the interval [0, l] one gets that
Z
2 l kπ
ak = g(x) sin xdx, k = 1, 2, 3, . . . . (2.2.17)
l 0 l
Example 2.2. Consider the initial-boundary value problem

x, 0 ≤ x ≤ 1,
θ(0, t) = θ(2, t) = 0, θ(x, 0) = g(x) = (2.2.18)
−x + 2, 1 ≤ x ≤ 2,

for the heat equation for a homogeneous bar of length 2. The Fourier coefficients
of g(x) are
8
a2k+2 ≡ 0, a2k+1 = (−1)k , k = 0, 1, 2, . . . . (2.2.19)
(2k + 1)2 π 2
1Fourier series are introduced and treated extensively in Appendix B
 2.2. SEPARATION OF VARIABLES 29

The resulting series solution is

∞
X µ ¶
(−1)i (2i + 1)2 π 2 t π
θ(x, t) = 8 2 2
exp − sin(i + )x. (2.2.20)
i=0
(2i + 1) π 4 2
Notice first that although the initial data is piecewise differentiable the solution
is smooth for any t > 0. Also, as long as the initial profile is integrable (e.g.,
piecewise continuous) on [0, 2] its Fourier coefficients are uniformly bounded,
namely:
Z 2 Z 2
|ak | ≤ |g(x) sin kπx| dx ≤ |g(x)| dx ≡ M. (2.2.21)
0 0
Consequently, the series solution (2.2.20) is bounded by an exponentially decaying
time series µ ¶
∞
X (2i + 1)2 π 2 t
|θ(x, t)| ≤ M exp − . (2.2.22)
i=0
4
This means that solution decays to the zero temperature profile, a direct conse-
quence of the fact that both ends are hold at zero temperature.

This simple example shows that in the case of homogeneous boundary con-
ditions any initial heat distributed throughout the bar will eventually dissipate
away. Moreover, as the Fourier coefficients in (2.2.20) decay exponentially as
t → ∞, the solution gets very smooth despite the fact that the initial data was
not. In fact, this is an illustration of the general smoothing property of the heat
equation.

Theorem 2.3. If u(t, x) is a solution to the heat equation with the initial
condition such that its Fourier coefficients are uniformly bounded, then for all
t > 0 the solution is an infinitely differentiable function of x. Also, u(t, x) → 0
2 t/l2
as t → ∞, in such a way that there exists K > 0 such that |u(t, x)| < Ke−γπ
for all t ≥ t0 > 0.

The smoothing effect of the heat equation means that it can be effectively used
to de-noise signals by damping the high frequency modes. This, however, means
also that it is impossible to reconstruct the initial temperature by measuring the
temperature distribution at some later time. The heat equation cannot be run
backwards in time. There is no temperature distribution at t < 0 which would
30 2. THE DIFFUSION EQUATION

produce a non-smooth temperature distribution at t = 0. Had we tried to run it

backwards, we would only get noise due to the fact that the Fourier coefficients
grow exponentially as t < 0. The backwards heat equation is ill possed.

Inhomogeneous Boundary Conditions.

There is a simple homogenization transformations that converts a homoge-
neous heat equation with inhomogeneous Dirichlet boundary conditions

θ(0, t) = α(t), θ(l, t) = β(t), t ≥ 0, (2.2.23)

into an inhomogeneous heat equation with homogeneous Dirichlet boundary con-

ditions. Suppose
α(t) − β(t)
ω(x, t) = θ(x, t) − α(t) + x (2.2.24)
l
where θ(x, 0) = g(x). θ(x, t) is a solution of a homogeneous heat equation if and
only if ω(x, t) satisfies the inhomogeneous equation
∂ω ∂ 2 ω α0 − β 0
− 2 = x − α0 (2.2.25)
∂t ∂t l
subject to the initial condition
α(0) − β(0)
ω(x, 0) = g(x) − α(0) + x, (2.2.26)
l
where
ω(0, t) = ω(0, l) = 0. (2.2.27)
Note that ω(x, t) is a solution to the homogeneous heat equation if and only if
the Dirichlet boundary conditions are constant. As the homogeneous boundary
conditions are essential in being able to superpose basic solutions (eigensolutions)
the Fourier series method can be used now in conjunction with the separation of
variables to obtain solutions of (2.2.25).

Example 2.4. Consider

∂ω ∂ 2 ω
− 2 = x cos t, 0 < x < 1, t > 0, (2.2.28)
∂t ∂t
subject to the initial condition

ω(x, 0) = x, (2.2.29)
 2.2. SEPARATION OF VARIABLES 31

and the following homogeneous boundary conditions:

∂ω
ω(0, t) = (1, t) = 0, t > 0. (2.2.30)
∂x
First, let us look for a solution of the homogeneous version of (2.2.28) with the
given boundary conditions (2.2.30) using the separation of variables method. To
this end the reader can easily show that the eigenfunctions are:
(2i + 1)π
ui (x) = sin λi x, λi = , i = 0, 1, 2 . . . . (2.2.31)
2
By the analogy with the form of the solution to the homogeneous heat equation
let us suppose a solution of (2.2.28) as a series of eigenfunctions
∞
X
ω(x, t) = αi (t) sin λi x. (2.2.32)
i=0

Also, represent the right-hand side of (2.2.28) as a series of eigenfunctions. Namely,

write ̰ !
X
x cos t = bi sin λi x cos t, (2.2.33)
i=0

where
Z 1
(−1)i
bi = x sin λi xdx = . (2.2.34)
0 λ2i
Substituting the solution (2.2.32) with (2.2.33) for its right hand side we are able
to show that the unknown functions αi (t) satisfy an inhomogeneous ordinary
differential equation
dαi (−1)i
+ λ2i αi = cos t. (2.2.35)
dt λ2i
Using the method of undetermined coefficients it is easy to obtain its solution
2 cos t sin t
αi (t) = Ae−λi t + (−1)i [ 4
+ 2 ]. (2.2.36)
1 + λi λi (1 + λ4i )
From the initial condition (2.2.29) and using (2.2.32) one can calculate that
(−1)i+1 (λ4i − λ2i + 1)
A= . (2.2.37)
λ2i (λ4i + 1)
This enables us to construct the solution (2.2.32).
32 2. THE DIFFUSION EQUATION

Periodic Boundary Conditions.

Heat flow in a circular ring is governed by the same homogeneous heat equa-
tion as is heat conduction in a rod (2.1.6), however, this time subject to periodic
boundary conditions
∂θ ∂θ
θ(π, t) = θ(−π, t),
(π, t) = (−π, t), t ≥ 0, (2.2.38)
∂x ∂x
where −π < x < π is the angular variable, and where we assume that the heat
can only flow along the ring as no radiation of heat from one side of the ring to
another is permitted2.
Benefiting from the separation of variables technique we are seeking a solution
in the form θ(x, t) = e−λt u(x). Assuming for simplicity that γ = 1, we arrive, as
before, at the associated eigenvalue problem
d2 u
2
+ λu = 0, u(π) = u(−π), u0 (π) = u0 (−π). (2.2.41)
dx
Its solutions are combinations of trigonometric sine and cosine functions

ui (x) = ai cos ix + bi sin ix, i = 0, 1, 2, . . . , (2.2.42)

with the eigenvalues

λi = i2 , i = 0, 1, 2, . . . . (2.2.43)
The resulting infinite series solution is
X∞
1 2
θ(x, t) = a0 + e−i t [ai cos ix + bi sin ix] . (2.2.44)
2 i=1

If we postulate the initial condition θ(x, 0) = g(x) the coefficients ai and bi must
be such that ∞
1 X
g(x) = a0 + [ai cos ix + bi sin ix] , (2.2.45)
2 i=1

2The heat conduction equation for a heated ring can easily be derived from the two-
dimensional heat equation
∂θ ∂2θ ∂2θ
= + (2.2.39)
∂t ∂x2 ∂y 2
by rewriting its right hand side in polar coordinates (r, α)
µ ¶
∂θ 1 ∂ ∂θ 1 ∂2θ
= r + 2 , (2.2.40)
∂t r ∂r ∂r r ∂α2

and assuming that the solution is r independent.

 2.3. UNIQUENESS OF SOLUTIONS 33

which is precisely the Fourier series of the initial condition g(x) provided
Z Z
1 π 1 π
ai = g(x) cos ixdx, bi = g(x) sin ixdx, i = 0, 1, 2, . . . . (2.2.46)
π −π π −π

2.3. Uniqueness of Solutions

In this section we investigate the uniqueness of solutions to the initial-boundary

value problem for the heat equation. To this end let us consider solutions of the
homogeneous heat equation
∂θ ∂ 2θ
= , 0 < x < l, 0 < t < ∞, (2.3.1)
∂t ∂x2
with the initial condition
θ(x, 0) = g(x), (2.3.2)
and the boundary conditions

θ(0, t) = α(t), θ(l, t) = β(t). (2.3.3)

Suppose that θ1 and θ2 are two solutions of (2.3.1) both satisfying the initial
condition (2.3.2) and boundary conditions (2.3.3). As the equation (2.3.1) is
linear the function ω(x, t) ≡ θ1 − θ2 is also a solution but with the zero initial
profile and the homogeneous boundary conditions.
Let us multiply (2.3.1) by ω(x, t) and integrate the resulting equation with
respect x on the interval [0, l] to obtain
Z l Z l 2
∂ω ∂ ω
ω dx = 2
ωdx. (2.3.4)
0 ∂t 0 ∂x

Assuming that ω(x, t) is regular enough, and integrating the right-hand side by
parts we reduce the relation (2.3.4) to
Z Z l µ ¶2 Z l µ ¶2
1d l 2 ∂ω ¯¯l ∂ω ∂ω
ω dx = ω ¯ − dx = − dx ≤ 0. (2.3.5)
2 dt 0 ∂x 0 0 ∂x 0 ∂x
Let Z l
1
I(t) ≡ ω 2 dx ≥ 0. (2.3.6)
2 0
Then,
Z tZ lµ ¶2
∂ω
I(t) − I(0) = − dxdt ≤ 0. (2.3.7)
0 0 ∂x
34 2. THE DIFFUSION EQUATION

However, I(0) = 0 implying that I(t) ≤ 0. On the other hand according to its
definition I(t) ≥ 0. Hence, I(t) ≡ 0. This is possibly only if ω(x, t) ≡ 0 proving
that θ1 (x, t) = θ2 (x, t) everywhere. Note that the same technique can be used
to prove uniqueness of solutions to other boundary value problems as long as
ω ∂ω
∂x
= 0 at x = 0 and x = l.

2.4. Fundamental Solutions

The idea of the fundamental solution of a partial differential equation is an

extension of the Green’s function method for solving boundary value problems of
ordinary differential equations. To set the stage for further considerations let us
briefly review the main points of the that method3.
Consider a homogeneous boundary value problem for the linear ordinary dif-
ferential equation

L(u) = δ(x − ξ), u(0) = u(l) = 0, 0 < x < l, (2.4.1)

where L(u) denotes a linear second-order differential operator actig on the func-
tion u(x) defined on [0, l] interval, while δ(x − ξ) ≡ δξ (x) is the (Dirac) delta
function at ξ. Note that if the boundary conditions are inhomogeneous we can
use the homogenization transformation (2.2.24) to transform the problem into
one with the homogeneous boundary conditions and a different inhomogeneous
right hand side.
Let u(x, ξ) = G(x, ξ) denote the solution to (2.4.1). This is the Green’s
function of this particular boundary value problem. Once we found this solution
we can use linearity to obtain the general solution of

L(u) = f (x), u(0) = u(l) = 0, 0 < x < l, (2.4.2)

in the form of the superposition integral . Indeed, let

Z l
u(x) ≡ G(x, ξ)f (ξ)dξ. (2.4.3)
0

3Details can be found in Section 1.2

 2.4. FUNDAMENTAL SOLUTIONS 35

It is easy to see that u(x) solves the boundary value problem (2.4.2) as
Z l Z l
L(u) = Lx (G)(x, ξ)f (ξ)dξ = δ(x − ξ)f (ξ)dξ = f (x) (2.4.4)
0 0

and the boundary conditions are satisfied. Lx (u) denotes here the partial differ-
ential operator induced by L. We will try to use the same idea in the context of
the heat equation.
Consider first the initial value problem for the heat conduction in an infinite
homogeneous bar subjected initially to a concentrated unit heat source applied
at a point y. We assume for simplicity that the thermal diffusivity γ = 1. This
requires solving the heat equation with the initial condition

u(x, 0) = δ(x − y), −∞ < x < ∞. (2.4.5)

To avoid any specific boundary conditions but to guarantee the uniqueness of

solutions (see Section 2.3) we require the solution to be square integrable at all
times, that is
Z ∞
|u(x, t)|2 dx < ∞ for all t ≥ 0. (2.4.6)
−∞

This, in fact, implies that the solution vanishes at infinity.

Let us now take the complex separable solution to the heat equation
2
u(x, t) = e−k t eikx , (2.4.7)

where, as there are no boundary conditions, there are no restrictions on the choice
of frequencies k. Mimicking the Fourier series superposition solution when there
are infinitely many frequencies allowed we may combine these solutions into a
Fourier integral (see Appendix B.5)
Z ∞
1
e−k t eikx δby (k)dk
2
u(x, t) = √ (2.4.8)
2π −∞

to realize, provided we can differentiate under the integral, that it solves the heat
equation. Moreover, the initial condition is also satisfied as
Z ∞
1
u(x, 0) = √ eikx δby (k)dk = δ(x − y), (2.4.9)
2π −∞
36 2. THE DIFFUSION EQUATION

where δby (k) denotes the Fourier transform of the delta function δ(x − y), that is

1
δby (k) = √ e−iky . (2.4.10)
2π

Combining (2.4.9) with (2.4.10) we find that the fundamental solution of the heat
equation is
Z ∞
1 2 1 −(x−y)2
F (x − y, t) = e−k t eik(x−y) dk = √ e 4t . (2.4.11)
2π −∞ 2 πt

It is worth pointing out here that although the individual component of the
Fourier series (2.4.8) are not square integrable the resulting fundamental solu-
tion (2.4.11) is. Another interesting derivation of the fundamental solution based
on the concept of the similarity transformation can be found in [Kevorkian].

Remark 2.5. It is important to point out here that one of the drawbacks of
the heat equation model is - as evident from the form of the fundamental solution
- that the heat propagates at infinite speed. Indeed, a very localized heat source
at y is felt immediately at the entire infinite bar because the fundamental solution
is at all times nonzero everywhere.

With the fundamental solution F (x − y, t) at hand we can now adopt the

superposition integral formula (2.4.3) to construct the solution to the heat con-
duction problem of an infinite homogeneous bar with the an arbitrary initial
temperature distribution u(x, 0) = g(x) as
Z ∞
1 −(x−y)2
u(x, t) = √ e 4t g(y)dy. (2.4.12)
2 πt −∞

That is, the general solution is obtained by a convolution of the initial data
with the fundamental solution. In other words, the solution with the initial
temperature profile g(x) is an infinite superposition over the entire bar of the
point source solutions of the initial strength
Z ∞
g(y) = δ(x − y)g(x)dx. (2.4.13)
−∞
 2.4. FUNDAMENTAL SOLUTIONS 37

Inhomogeneous Heat Equation for the Infinite Bar .

The Green’s function method can also be used to solve the inhomogeneous
heat conduction problem
∂u ∂ 2 u
− 2 = p(x, t), −∞ < x < ∞, t > 0, (2.4.14a)
∂t ∂x
where the bar is subjected to a heat source p(x, t) which may vary in time and
along its length. We impose the zero initial condition

u(x, 0) = 0, (2.4.14b)

and some homogeneous boundary conditions. The main idea behind this method
is to solve first the heat equation with the concentrated source applied instanta-
neously at a single moment, and to use the method of superposition to obtain
the general solution with an arbitrary source term. We therefore begin by solving
the heat equation (2.4.14a) with the source term

p(x, t) = δ(x − y)δ(t − s). (2.4.15)

It represents a unit heat input applied instantaneously at time s and position y.

We postulate the same homogeneous initial and boundary conditions as in the
general case. Let
u(x, t) = G(x − y, t − s) (2.4.16)

denote the solution to this problem. We will refer to it as the general fundamental
solution or a Green’s function. Thanks to the linearity of the heat equation the
solution of the general problem is given by the superposition integral
Z tZ ∞
u(x, t) = G(x − y, t − s)p(y, s)dyds, (2.4.17)
0 −∞

where the forcing term may be also rewritten by the superposition formula as
Z ∞Z ∞
p(x, t) = p(y, s)δ(t − s)δ(x − y)dyds. (2.4.18)
0 −∞

If we replace the zero initial condition by u(x, 0) = f (x), then once again due to
the linearity of the differential equation we may write the solution as a combina-
tion of a solution to the homogeneous equation with inhomogeneous initial data
38 2. THE DIFFUSION EQUATION

and the solution with the homogeneous initial condition but a nonzero forcing
term
Z ∞ Z tZ ∞
u(x, t) = F (x − y, t)f (y)dy + G(x − y, t − s)p(y, s)dyds. (2.4.19)
−∞ 0 −∞

To find the general fundamental solution in an explicit form let us take the
Fourier transform with respect to variable x of both sides of the differential equa-
tion (2.4.14a) with the forcing term (2.4.15). Using (2.4.10) we find that
db
u 1
+ k2u
b = √ e−iky δ(t − s), (2.4.20)
dt 2π
where u
b(k, t) denotes the Fourier transform of u(x, t), and where k is viewed as
a parameter. This is an inhomogeneous first order linear ordinary differential
equation for the Fourier transform of u(x, t) with the initial condition

u
b(k, 0) = 0 for s > 0. (2.4.21)
2t
Using the integrating factor method with the integrating factor ek we obtain
that
1 2
b(k, t) = √ ek (t−s)−iky σ(t − s),
u (2.4.22)
2π
where σ(t − s) is the usual step function. The Green’s function is than obtained
by the inverse Fourier transform
Z ∞
1
G(x − y, t − s) = √ eikx u
b(k, t)dk. (2.4.23)
2π −∞

Using the formula (2.4.11) of the fundamental solution we deduce that

Z
σ(t − s) ∞ ik(x−y)+k2 (t−s)
G(x − y, t − s) = e dk (2.4.24)
2π −∞
½ ¾
σ(t − s) (x − y)2
= p exp − .
2 π(t − s) 4(t − s)
The general fundamental solution (Green’s function) is just a shift of the fun-

damental solution for the initial value problem at t = 0 to the starting time
t = s. More importantly, its form shows that the effect of a concentrated heat
source applied at the initial moment is the same as that of a concentrated initial
temperature.
 2.4. FUNDAMENTAL SOLUTIONS 39

Finally, the superposition integral (2.4.17) gives us the solution

Z tZ ∞ ½ ¾
p(y, s) (x − y)2
u(x, t) = p exp − dsdy (2.4.25)
0 −∞ 2 π(t − s) 4(t − s)
of the heat conduction problem for the infinite homogeneous bar with a heat
source.

Heat Equation for the Semi-infinite Bar .

To illustrate how the Green’s function method can be applied in the case of
the semi-infinite domain we consider the heat equation with the concentrated
forcing term
∂u ∂ 2 u
− 2 = δ(x − y)δ(t), 0 ≤ x < ∞, t>0 (2.4.26a)
∂t ∂x
and impose the zero initial condition, i.e., u(x, 0) = 0, and the homogeneous
boundary conditions
u(0, t) = 0, lim u(x, t) = 0. (2.4.26b)
x→∞

As we have remarked earlier the effect of such a concentrated instantaneous heat

source is the same as that of the concentrated initial distribution. Thus, the only
difference between this case and the case of the fundamental solution (2.4.5) is
the imposition of the boundary condition at x = 0. One possible way to tackle
this difficulty is to consider in place of this semi-infinite problem such an infinite
domain problem in which the homogeneous ”boundary” condition at x = 0 is
permanently satisfied.
Hence, consider the heat conduction problem for an infinite homogeneous bar
with δ(t)[δ(x − y) − δ(x + y)] as the forcing term, homogeneous initial condition,
and the homogeneous boundary conditions at infinities. In other words, in the
infinite domain we apply a unit strength source at x = y, and simultaneously
a negative source of unit strength at x = −y. This approach is known as the
method of images. Once again, due to the linearity of the heat equation and
that of the forcing term, the temperature profile at t > 0 will be the sum of
two fundamental solutions F (x − y, t) and F (x + y, t) each corresponding to one
of the source terms. In particular, due to the skew-symmetry of these solutions
the combined solution will always be vanishing at x = 0. Moreover since all the
40 2. THE DIFFUSION EQUATION

boundary conditions of the original problem (2.4.26a) are satisfied, and since the
second source term δ(t)δ(x + y) is outside of the original semi-infinite domain,
the Green’s function

G(x − y, t) ≡ F (x − y, t) − F (x + y, t) (2.4.27)

is the solution of (2.4.26a), where the fundamental solution F is defined by (2.4.11).

In conclusion, the solution of the inhomogeneous heat conduction problem for
a semi-infinite bar
∂u ∂ 2 u
− 2 = p(x, t), 0 ≤ x < ∞, t > 0, (2.4.28)
∂t ∂x
with the initial condition u(x, 0) = 0 and the homogeneous boundary condi-
tions (2.4.26b) has the form
Z tZ ∞ ½ · ¸ · ¸¾
p(y, s) (x − y)2 (x + y)2
u(x, t) = p exp − − exp − dyds.
0 0 2 π(t − s) 4(t − s) 4(t − s)
(2.4.29)

Example 2.6. Suppose that a semi-infinite homogeneous bar is initially heated

to a unit temperature along a finite interval [a, b], where a > 0. Assume also that
at x = 0 the temperature is held at zero (by attaching an infinite rod of this
temperature) and vanishes at infinity. This corresponds to the following initial
value problem for the heat equation:


0, if 0 < x < a,


∂u ∂ 2u
= , u(x, 0) = σ(x − a) − σ(x − b) = 1, if a < x < b, (2.4.30a)
∂t ∂x2 


0, if x > b,

with the homogeneous boundary conditions

u(0, t) = 0, lim u(x, t) = 0, t > 0.

x→∞

The method of images and the superposition formula (2.4.12) yield the solution
µZ b Z −b ¶
1 (x−y)2
− 4t
(x−y)2
− 4t
u(x, t) = √ e dy + e dy (2.4.31)
2 πt a −a
½ µ ¶ µ ¶¾ ½ µ ¶ µ ¶¾
1 x−a x+a 1 x−b x+b
= erf √ + erf √ − erf √ + erf √ ,
2 2 t 2 t 2 2 t 2 t
 2.5. BURGERS’ EQUATION 41

where the error function Z z

2 2
erfz ≡ √ e−η dη. (2.4.32)
π 0
Note that the error function is odd and that its asymptotic value at infinity is 1.

2.5. Burgers’ Equation

In this last section we will study the quasilinear version of the diffusion equa-
tion
∂u ∂u ∂2u
+u − ² 2 = 0, ² > 0, (2.5.1)
∂t ∂x ∂x
to show how the solution methods developed in previous sections for the heat
equation may be used to obtain solutions for other equations. Also, Burgers’
equation is a fundamental example of an evolution equation modelling situations
in which viscous and nonlinear effects are equally important. Moreover, it plays
somewhat important role in discussing discontinuous solutions (shocks) of the
one-dimensional conservation law
∂u ∂u
+u = 0, (2.5.2)
∂t ∂x
a topic which will not be discussed here (see for example [Knobel], [Smoller]).
We start by looking at ways at which the methods for solving the initial-
boundary value problems of heat equations can be used to solve (2.5.1).

The Cole-Hopf Transformation.

This is a change of dependent variable w = W (u) which enables us to trans-
form Burgers’ equation into the linear diffusion equation studied already in this
chapter. Let
wx
u ≡ −2² , (2.5.3)
w
where wx denotes partial differentiation. Calculating all derivatives and substi-
tuting them into (2.5.1) yields

wx (²wxx − wt ) − w(²wxx − wt )x = 0. (2.5.4)

In particular, if w(x, t) solves the diffusion equation

²wxx − wt = 0, (2.5.5)

the function u(x, t) given by (2.5.3) satisfies Burgers’ equation (2.5.1).

42 2. THE DIFFUSION EQUATION

Initial Value Problem on the Infinite Domain.

Let us consider the following initial value problem:
∂u ∂u ∂ 2u
+u − ² 2 = 0, u(x, 0) = g(x), −∞ < x < ∞, (2.5.6)
∂t ∂x ∂x
and suppose that we are looking for the solutions which satisfy the corresponding
diffusion equation (2.5.5). According to (2.5.3), the initial condition for the new
variable w(x, 0) must be such that

g(x)w(x, 0) = −2²wx (x, 0). (2.5.7)

The general solution of this linear ordinary differential equation for w(x, 0) is
· Z ¸
1 x
w(x, 0) = A exp − g(s)ds , (2.5.8)
2² 0
where A is a constant, and where we assume that the integral exists. Hence, we
essentially need to solve the following initial value problem for the homogeneous
diffusion equation with the inhomogeneous initial condition:

²wxx − wt = 0, w(x, 0) = h(x), −∞ < x < ∞. (2.5.9)

Its solution has the form of (2.4.12) :

Z ∞
1 −(x−y)2
w(x, t) = √ e 4²t h(y)dy (2.5.10)
2 π²t −∞

where we replaced t by ²t Note that the parameter ² may be eliminated from

the equation, and so from the solution, by an appropriate scaling of variables.
We retain it, however, so we one can later study the asymptotic behavior of
solutions when ² → 0. Differentiating with respect to x and using the Cole-Hopf
formula (2.5.3) we compute that
R∞ h 2
i
(x−y)
−∞ t
exp −(x−y)4²t
h(y)dy
u(x, t) = R∞ h i , (2.5.11)
−(x−y)2
−∞
exp 4²t
h(y)dy

where · Z ¸
x
1
h(y) ≡ exp − g(s)ds , (2.5.12)
2² 0
as the constant A cancels out.
 2.5. BURGERS’ EQUATION 43

Boundary Value Problem on a Finite Interval .

Using separation of variables method, we solve here the following initial-
boundary value problem:
∂u ∂u ∂ 2u
+u − ² 2 = 0, u(x, 0) = g(x), 0 < x < a, (2.5.13a)
∂t ∂x ∂x
u(0, t) = u(a, t) = 0, t > 0. (2.5.13b)
After Cole-Hope transformation we obtain the corresponding initial-boundary
value problem for the diffusion equation:

wt − ²wxx = 0, w(x, 0) = Ah(x), 0 < x < a, (2.5.14a)

wx (0, t) = wx (a, t) = 0, t > 0. (2.5.14b)

As the boundary condition are homogeneous the solution w(x, t) can easily be
derived (see page 27) as
∞
a0 X kπ 2 kπ
w(x, t) = + ak e−( a ) t cos x, (2.5.15)
2 k=1
a
where Z
2A a kπ
ak = h(x) cos xdx. (2.5.16)
aπ 0 a
From the Cole-Hope transformation formula (2.5.3) one now gets the solution to
the initial vale problem (2.5.13)
π
P∞ £ kπ 2 ¤
k=1 ka k exp −( a ) t sin kπ x
u(x, t) = 2² a0a P ∞ £ kπ 2
¤ a
kπ
(2.5.17)
2
+ k=1 ak exp −( a ) t cos a x
Example 2.7. Consider Burgers’ equation

ut + uux − ²uxx = 0, −∞ < x, ∞, (2.5.18)

with the the piecewise initial condition


1, if x < 0
u(x, 0) = 2σ(x) − 1 = (2.5.19)
−1, if x > 0.

The initial condition of the associated diffusion equation (2.5.5) may now be
obtained from (2.5.7):
1
w(x, 0) = Ae 2² |x| . (2.5.20)
44 2. THE DIFFUSION EQUATION

The solution w(x, t) takes the form of (2.4.12) with t replaced by ²t. Namely,
Z ∞
1 −(x−y)2 |y|
w(x, t) = √ e 4²t e 2² dy. (2.5.21)
2 π²t −∞
Therefore, the solution of the original initial value problem is given by (2.5.11):
R ∞ (x−y) h i |y|
−(x−y)2
−∞ t
exp 4²t
)e 2² dy
u(x, t) = R∞ h i |y| . (2.5.22)
−(x−y)2
−∞
exp 4²t
e 2² dy

Integrating independently from −∞ to 0 and from 0 to ∞, and using the substi-

tution
(x − y ± t)
η= √ ,
2 ²t
respectively, we finally obtain
³ ´ ³ ´
− x² x−t x+t
e erfc √− erfc − 2 ²t
2 ²t
√
u(x, t) = x ³ ´ ³ ´, (2.5.23)
e− ² erfc 2x−t
√
²t
+ erfc − x+t
√
2 ²t

where the complimentary error function

Z ∞
2 2
erfc(z) ≡ 1 − erf(z) = √ e−η dη. (2.5.24)
π z
 CHAPTER 3

Laplace’s Equation

In this chapter we discuss the solution techniques for the Laplace’s and Poisson
equations in two dimensions. The Laplace’s equation
∂ 2u ∂ 2u
4u ≡ + =0
∂x2 ∂y 2
and its inhomogeneous counterpart, the Poisson equation

−4u = f (x, y),

are one of the most fundamental partial differential equations. They arise in a
variety of physical and mathematical situations, ranging from heat conduction,
solid and fluid mechanics, and electromagnetism to geometry, probability, and
number theory.

3.1. Physical Interpretation

In a typical interpretation u denotes the density per unit volume of some

physical quantity in equilibrium (e.g. a thermal energy of a homogeneous mate-
rial, like in (2.1.1), or a chemical concentration). If Ω is a smooth region in R2 ,
the net flux of u through the boundary ∂Ω vanishes:
Z
q(x, y) · n dS = 0. (3.1.1)
∂Ω

q(x, y) denotes here the flux density per unit surface area and n is the unit outer
normal. Applying the Gauss-Green Theorem, we have
Z Z
q(x, y) · n dS = div q(x, y) dV = 0,
∂Ω Ω

and so
div q(x, y) = 0 in Ω, (3.1.2)
45
46 3. LAPLACE’S EQUATION

since the region Ω is arbitrary. Assuming, what is often physically reasonable,

that the flux q is proportional to the gradient ∇u, but points in the opposite
direction (the flow is from the region of higher to lower concentration)1, we obtain
Laplace’s equation
div (∇u) = 0. (3.1.3)

3.2. Separation of Variables

In order to construct an explicit solution to the Laplace’s equation, we resort

first to the method of separation of variables.
We start by writing a solution ansatz in the form

u(x, y) = w(x)v(y). (3.2.1)

Let us substitute this expression into the Laplace’s equation

4u = w00 (x)v(y) + w(x)v 00 (y) = 0

and re-write it in the form

w00 (x)v(y) = −w(x)v 00 (y).

Dividing both sides by the product w(x)v(y), while assuming that it is not iden-
tically zero, yields
w00 (x) v 00 (y)
=− (3.2.2)
w(x) v(y)
effectively separating the x and y variables on the opposite sites of the equation.
Assuming that the domain on which the equation is considered, say Ω, is con-
nected leads to the obvious conclusion that for the separation to work both sides
must be constant functions. This reduces our problem to a pair of second order
linear ordinary differential equations

w00 − λw = 0, v 00 + λv = 0, (3.2.3)

for the individual components of the separable solution u(x, y) = w(x)v(y) of the
Laplace’s equation 4u = 0. The undetermined constant λ is called the separation
constant. These are relatively simple linear ordinary differential equations. We
know how to solve them, and what form of a solution to expect depending on
1Some such examples are Fick’s Law of Diffusion and Fourier’s Law of Cooling.
 3.2. SEPARATION OF VARIABLES 47

the sign of the separation constant λ. Indeed, if λ 6= 0 the solution will be the
combinations of products of exponential functions with sine and cosine functions.
On the other hand, when λ = 0 the solutionis a second degree polynomial in x
and y.
Applying the separation of variables method to the boundary value problems
forces us to consider rather restricted geometry of the domain of solutions. Thus,
let us consider the Laplace’s equation on a rectangle, e.g.,

4u = 0 on D = {0 < x < a, 0 < y < b}, (3.2.4a)

with Dirichlet boundary conditions

u(x, 0) = f (x), u(x, b) = u(0, y) = u(a, y) = 0. (3.2.4b)

These boundary conditions are not the most general boundary conditions one
might like to consider. However, observe that an arbitrary inhomogeneous Dirich-
let boundary condition can always be decomposed into four different boundary
conditions like (3.2.4b), and use the superposition of solutions - all due to the
linearity of the Laplace’s equation. Proceeding to solve our boundary value prob-
lem (3.2.4b) let us look first at all the homogeneous conditions. To this end, we
immediately obtain that

w(0) = w(a) = 0 and v(b) = 0. (3.2.5)

Standard analysis of the associated eigenvalue problems (3.2.3) shows that the
only possible solutions are the products
nπx nπ(b − y)
un (x, y) = sin sinh , n = 1, 2, 3, . . . , (3.2.6)
a a
where the separation constants
³ nπ ´2
λn = , n = 1, 2, 3, . . . . (3.2.7)
a
It remains now to analyze the role of the inhomogeneous boundary condition
u(x, 0) = f (x). As none of the un (x, y) functions could, in general, satisfy such
an arbitrary condition we try a linear superposition of solutions in the form of
48 3. LAPLACE’S EQUATION

an infinite series
∞
X nπx nπ(b − y)
u(x, y) = cn sin sinh , (3.2.8)
n=1
a a
where the coefficients cn are to be determined from the boundary condition at
y = 0. Indeed, we have that
∞
X nπx nπb
u(x, 0) = cn sin sinh = f (x). (3.2.9)
n=1
a a

This has the form of the Fourier sine series for the function f (x). Let us therefore
take
bn
cn = (3.2.10)
sinh nπb
a
where bn are the Fourier sine coefficients on the interval (0, a) of f (x) as given
by the formula (B.1.2). This yields that the solution to the boundary value
problem (3.2.4) is the infinite series
∞
X sinh nπ(b−y)
a nπx
u(x, y) = bn nπb
sin . (3.2.11)
n=1
sinh a a

In can be shown that if the function f (x) is integrable on (0, a), i.e.,
Z a
|f (x)|dx < ∞, (3.2.12)
0

the series converges on the rectangle D.

The Unit Disk .

As the illustration of how the method of separation of variables can be used
in some other geometric situations, let us consider the boundary value problem
for the Laplace’s equation on the unit disk. Namely, let

4u = 0 on x2 + y 2 < 1, (3.2.13a)

and
u(x, y) = f (x, y) whenever x2 + y 2 = 1. (3.2.13b)
One possible physical interpretation of what the solution u(x, y) might represent
is the displacement of a circular drum of a given hight; the (circumferential) hight
of the drum cannot change but the membrane of the drum can deform.
 3.2. SEPARATION OF VARIABLES 49

When dealing with a circular geometry it seems natural to adopt polar coor-
dinates
x = r cos θ, y = r sin θ. (3.2.14)
Rewriting the boundary conditions (3.2.13b) we obtain that

u(cos θ, sin θ) = f (cos θ, sin θ) = f (θ). (3.2.15)

Elementary calculations yield the Laplacian in polar coordinates

∂ 2u ∂ 2u ∂ 2 u 1 ∂u 1 ∂ 2u
4u = + = + + . (3.2.16)
∂x2 ∂y 2 ∂r2 r ∂r r2 ∂θ2
This enables us to reformulate our boundary value problem (3.2.13) for the func-
tion u = u(r, θ) in the form
∂ 2 u 1 ∂u 1 ∂ 2u
+ + = 0, u(1, θ) = f (θ). (3.2.17)
∂r2 r ∂r r2 ∂θ2
Observe that we must also require that the solution u(r, θ), as well as the bound-
ary condition f (θ), are 2π periodic, i.e.,

u(r, θ + 2π) = u(r, θ), f (θ + 2π) = f (θ). (3.2.18)

The method of separation of variables is based on the postulate that

u(r, θ) = v(r)µ(θ). (3.2.19)

Substituting (3.2.19) into the polar version of the Laplace’s equation (3.2.17), we
obtain that
1 1
v 00 (r)µ(θ) + v 0 (r)µ(θ) + 2 v(r)µ00 (θ) = 0. (3.2.20)
r r
Multiplying both sides of the above equation by r2 and dividing by the product
vµ we are able to move all terms involving r to one side of the equation and
all the terms involving θ to the other side. Hence, the equivalent form of the
equation (3.2.20) is
r2 v 00 (r) + rv 0 (r) µ00 (θ)
=− . (3.2.21)
v(r) µ(θ)
This, as we know from our previous similar considerations, reduces to a pair of
ordinary differential equations

r2 v 00 + rv 0 − λr = 0, µ00 + λµ = 0, (3.2.22)
50 3. LAPLACE’S EQUATION

for the respective components of the separable solution vµ. The separation con-
stant λ plays here the role of the eigenvalue. The latter equation, with the added
periodicity requirement (3.2.18), was solved in Section 2.2, where we showed that
the eigenvalues are
λ = n2 , n = 0, 1, 2, . . . , (3.2.23)
with the corresponding eigenfunctions

1, sin nθ, cos nθ. (3.2.24)

The first differential equation, on the other hand, has the form of a second order
Euler equation for the function v(r), where λ = n2 . Its linearly independent
solutions are usually obtained by seeking the power solution v(r) = rk . The
resulting characteristic equation requires

k 2 − n2 = 0.

Hence,
k = ±n.
If n 6= 0 there are two linearly independent solutions,

v1 (r) = rn , v2 (r) = r−n , n = 1, 2, . . . . (3.2.25)

If, on the other hand, n = 0 we get two other solutions2

v1 (r) = 1, v2 (r) = log r. (3.2.26)

Since we require our solution u(r, θ) to be bounded and continuous - even at the
center of the disk - we are forced to discard these elementary solutions which are
singular when r → 0. Thus, the series solution takes the form
∞
a0 X n
u(r, θ) = + r [an cos nθ + bn sin nθ] . (3.2.27)
2 n=1

Taking into account the boundary condition u(1, r) = f (θ) we conclude that
Z Z
1 π 1 π
an = f (θ) cos nθdθ, bn = f (θ) sin nθdθ, (3.2.28)
π −π π −π
2These solutions can easily be obtained by the reduction of order method.
 3.2. SEPARATION OF VARIABLES 51

are precisely the Fourier coefficients (B.1.2) of the boundary value f (θ). If f (θ) is
integrable on (−π, π) then, similarly to the rectangle problem, it can be showed
that the series solution (3.2.27) converges uniformly to the solution of our bound-
ary value problem. Moreover, such a solution is analytic inside the disk.

Example 3.1. Consider the boundary value problem (3.2.17) where

u(1, θ) = θ, −π < θ < π.

Observe that the boundary condition experiences a jump at (x, y) = (−1, 0). The
Fourier series of f (θ) = θ is
∞
X sin nθ
2 (−1)n+1 .
n=1
n

Therefore, according to (3.2.27) the solution to our boundary value problem is

X∞
sin nθ
u(r, θ) = 2 (−1)n+1 rn . (3.2.29)
n=1
n

This series can be explicitly summed. Indeed, write

z n = rn einθ = rn cos nθ + irn sin nθ (3.2.30)

for the complex variable z = x + iy and represent the solution

̰ !
X zn
u(x, y) = 2Im (−1)n+1 = 2Im log(1 + z). (3.2.31)
n=1
n

If we now write 1 + z = %eiψ then the solution takes the form

y
u(x, y) = 2ψ = 2 arctan , (3.2.32)
1+x

where ψ is the angle between the line through (x, y) and (−1, 0), and the x-axis.
One can easily show that the solution given by (3.2.32) has the correct boundary
value.
52 3. LAPLACE’S EQUATION

3.3. The Green’s Function

Consider the Poisson equation

−4u = f (x, y), (3.3.1)

that is, the inhomogeneous counterpart of the Laplace’s equation. As in the

one-dimensional case, we will use the Green’s function method to solve relevant
boundary value problems. By the Green’s function we mean, the same way as
before, the solution in which the inhomogeneity f (x, y) (the forcing term) is a
concentrated unit impulse; the delta function. The solution to the equation with
the general forcing term is then obtained by using the method of superposition.
We denote the delta function concentrated at position ξ ∈ R2 by

δξ (x) ≡ δ(x − ξ). (3.3.2)

As in the one-dimensional case, we interpret it as a linear functional on the space

C(Ω) of all continuous scalar-valued function on the domain Ω ⊂ R2 , satisfying

Z Z f (ξ), ξ ∈ Ω
δξ (x)f (x)dξ = (3.3.3)
Ω 0, ξ∈/Ω
for any function f (x) and any domain Ω. If we view the delta function as the
product of two one-dimensional delta functions

δξ (x) = δξ (x)δη (y), ξ = (ξ, η), x = (x, y), (3.3.4)

and if the domain Ω is a rectangle, say Ω = {a < x < b, c < y < b}, then
Z Z Z bZ d Z b
δξ (x)f (x)dx = δξ (x)δη (y)f (x, y)dydx = δξ (x)f (x, η)dx = f (ξ, η),
Ω a c a
(3.3.5)
provided (ξ, η) ∈ Ω. If (ξ, η) ∈
/ Ω the integral is 0.
Alternatively, the delta function can be viewed as the limit of concentrated
unit densities gn (x, y) such that

lim gn (x, y) = 0, for (x, y) 6= (0, 0), (3.3.6a)

n→∞

and Z Z
gn (x, y)dxdy = 1. (3.3.6b)
Ω
 3.3. THE GREEN’S FUNCTION 53

A good example of such a sequence are the so-called radial Gauss distributions
−(x2 +y 2 )
e n
gn (x, y) = . (3.3.7)
nπ
We leave to the reader to show that
Z Z
2 2
e−(x +y )/n dxdy = nπ.
R2
Finding the Green’s function amounts to solving the equilibrium problem
subject to a concentrated forcing. For the Poisson equation, this takes the form

−4u = δξ (x)δη (y), (x, y) ∈ Ω (3.3.8)

supplemented by homogeneous boundary conditions. As we shall see later, the

only acceptable boundary conditions are pure Dirichlet or mixed boundary con-
ditions as the pure Neumann boundary value problem has a non unique solution.
The Green’s function solution to (3.3.8) - with appropriate boundary condition -
is usually denoted as
G(x, ξ) = G(x, y; ξ, η). (3.3.9)
As in the one-dimensional case, the self-adjointness of the boundary value prob-
lem implies that the Green’s function is symmetric under the change of pairs of
arguments, i.e.,
G(x, y; ξ, η) = G(ξ, η; x, y). (3.3.10)
Once the Green’s function is available, the solution to the general Poisson
boundary value problem

−4u = f (x, y), u|∂Ω = 0, (3.3.11)

is constructed by using the superposition principle. In other words, representing

Z Z
f (x, y) = δx (ξ)δy (η)f (ξ, η)dξdη (3.3.12)
Ω

and using the linearity the solution to the boundary value problem (3.3.11) is
given by the superposition. That is, regarding the forcing term as the combination
of point impulses, and the Green’s functions as the corresponding solutions, we
write the solution as
Z Z
u(x, y) = G(x, y; ξ, η)f (ξ, η)dξdη. (3.3.13)
Ω
54 3. LAPLACE’S EQUATION

It is rather unfortunate that most Green’s functions cannot be evaluated explic-

itly. However, using a bit of intuition and the fact that the general solution to
the inhomogeneous linear differential equation is a combination of a particular
solution and the general solution to its homogeneous counterpart, we shall be
able to gain some inside into the form of the Green’s function of the Poisson
equation.
Let us, once again, consider the Poisson equation with the forcing unit impulse

−4u = δ(x − ξ, y − η). (3.3.14)

The general solution to such a linear inhomogeneous equation can always be

written as
u(x, y) = up (x, y) − uh (x, y), (3.3.15)
where uh denotes an arbitrary harmonic function (the general solution to the
Laplace’s equation) while up is a particular solution to the Poisson equation,
irrespective of any boundary conditions. As usually, we shall use the solution uh
to accommodate any particular boundary conditions imposed.
We propose as a particular solution
a
up (x, y) = log[(x − ξ)2 + (y − η)2 ]. (3.3.16)
2
First of all this function is radially symmetric about the point at which the unit
external impulse is applied, a characteristic expected from a solution modelling a
homogeneous medium. Also, as we have discovered earlier in (3.2.26) it solves the
Laplace’s equation everywhere except at (ξ, η), at which it exhibits a singularity.
If this function is to work as a particular solution to our inhomogeneous problem
we must be able to determine the correct constant a. In other words, we must
find a such that
2
−4 log[(x − ξ)2 + (y − η)2 ] = δ(x − ξ, y − η). (3.3.17)
a
Interpreting the delta function as a linear functional on a space of sufficiently
regular functions g(x, y) with L2 -inner product allows us to evaluate the inner
product
2
< −4 log[(x − ξ)2 + (y − η)2 ]; g >= < δ(x − ξ, y − η); g > . (3.3.18)
a
 3.3. THE GREEN’S FUNCTION 55

By the definition of the delta function (3.3.3), the right hand side equals to
g(ξ, η), . Remembering that log[(x − ξ)2 + (y − η)2 ] satisfies Laplace’s equation
everywhere but at the center, the right hand side

< −4 log[(x − ξ)2 + (y − η)2 ]; g >

Z Z
= (−4 log[(x − ξ)2 + (y − η)2 ])g(x, y)dxdy
D²
Z Z
= g(ξ, η) (−4 log[(x − ξ)2 + (y − η)2 ])dxdy, (3.3.19)
D²

where D² is a small disk of radius ² centered at (ξ, η). Using the fact that
−4 = −∇ · ∇, and the Green’s formula (3.4.6) we get that
Z Z Z Z
2 2
4 log[(x − ξ) + (y − η) ]dxdy = (∇ · ∇ log[(x − ξ)2 + (y − η)2 ]dxdy
D² D²
I 2 2 I Z 2π
∂ log[(x − ξ) + (y − η) ] 1
= ds = ds = dθ = 2π.
∂D² ∂n ∂D² r 0
(3.3.20)
When evaluating the one before last integral we benefited from the fact that the
arc length on a circle of radius ² is ds = ²dθ. The last result shows that

< −4 log[(x − ξ)2 + (y − η)2 ]; g >= −2πg(ξ, η), (3.3.21)

proving that

−4 log[(x − ξ)2 + (y − η)2 ] = −2πδ(x − ξ, y − η). (3.3.22)

This allows us to use

1
log[(x − ξ)2 + (y − η)2 ]
up (x, y) = − (3.3.23)
4π
as a particular solution of the inhomogeneous Poisson equation (3.3.14). The
general solution to the equation (3.3.14) is
1
u(x, y) = −log[(x − ξ)2 + (y − η)2 ] + uh (x, y), (3.3.24)
4π
where uh (x, y) is an arbitrary harmonic function. Note that postulating Dirichlet
boundary conditions forces the solution uh to satisfy
1
uh (x, y) = log[(x − ξ)2 + (y − η)2 ] at all (x, y) ∈ ∂Ω. (3.3.25)
2π
56 3. LAPLACE’S EQUATION

In conclution, we have just showed that

Theorem 3.2. The Greens’s function for any positive definite boundary value
problem for the Poisson equation has the form
1
G(x, y; ξ, η) = − log[(x − ξ)2 + (y − η)2 ] + uh (x, y) (3.3.26)
4π
where uh (x, y) is the harmonic function that has the same boundary value as the
1
logarithmic potential 2π
log[(x − ξ)2 + (y − η)2 ].

3.4. Minimum Principles

The main goal of this section is to show that the solution to the Poisson
equation (3.3.1), subject to proper boundary conditions, is the unique minimizer
of the corresponding ”energy” functional. This requires presenting the Poisson
equation in the self-adjoint form, relative to the appropriate inner product and
suitable boundary conditions.
Notice first that the Laplacian
∂ 2u ∂ 2u
4u = + ,
∂x2 ∂y 2
as a linear operator on a scalar-valued function u(x, y), can be viewed as a super-
position of two operations: the gradient and the divergence. Indeed, using the
multi-variable calculus presentation of the divergence as a formal dot product of
the gradient operator with the vector field, i.e.,
∂v1 ∂v2
∇ · (v1 , v2 ) = + ,
∂x ∂y
we obtain that
∂u ∂u ∂2u ∂ 2u
∇ · ∇u = ∇ · ( , )= + .
∂x ∂y ∂x2 ∂y 2
This very fact, and the fact that in one-dimension the adjoint of the deriva-
tive is its negative (1.3.5), suggest that the adjoint of the gradient is minus the
divergence.
The gradient operator ∇ : U → V maps a suitable space of scalar-valued
functions (scalar fields) U into a space of vector-valued functions (vector fields)
 3.4. MINIMUM PRINCIPLES 57

V. According to the original definition of the adjoint (1.3.1), the operator ∇∗ :

V → U is the adjoint of the gradient if

< u; ∇∗ v >=¿ ∇u; v À, (3.4.1)

for all u ∈ U , v ∈ V, where < ·; · >, and ¿ ·; · À denote the inner products
in U and V, respectively. We assume that both inner products refer to L2 -inner
products. Namely,
Z Z Z Z
< u; w >≡ u(x, y)w(x, y)dxdy, ¿ v; w À≡ v(x, y) · w(x, y)dxy.
Ω Ω
(3.4.2)
The right hand side of the identity (3.4.1) calculates as
Z Z
¿ ∇u; v À= (∇u · v)dxdy. (3.4.3)
Ω

Recalling Green’s formula

Z Z I
∇ · (uv)dxdy = u(v · n)ds, (3.4.4)
Ω ∂Ω

where n is the outer normal to the boundary ∂Ω, and Leibniz’s rule

∇ · (uv) = u∇ · v + ∇u · v, (3.4.5)

we conclude that
Z Z I Z Z
(∇u · v)dxdy = u(v · n)ds − u(∇ · v)dxdy. (3.4.6)
Ω ∂Ω Ω

If the choice of scalar and vector fields, u and v, is such that the boundary integral
vanishes,
I
u(v · n)ds = 0, (3.4.7)
∂Ω

then according to (3.4.3) and (3.4.6)

Z Z
¿ ∇u; v À= − u(∇ · v)dxdy =< u; −∇ · v > . (3.4.8)
Ω

This shows that, subject to the boundary constraint (3.4.7), the adjoint of the
gradient is minus the divergence, ∇∗ = −∇·. What remains to be established is
what boundary conditions on u and/or v guarantee the vanishing of the boundary
58 3. LAPLACE’S EQUATION

integral (3.4.7). Clearly the boundary integral vanishes if u satisfies Dirichlet

boundary conditions, that is if

u ≡ 0 on ∂Ω. (3.4.9)

This can be interpreted, for example, as that the membrane of a drum is attached
rigidly to the edge of the drum. In heat conduction, this boundary condition
corresponds to fixing the temperature at the boundary of the region Ω. We may
also require that
v · n = 0 on ∂Ω, (3.4.10)
which simply states that v is tangent to ∂Ω at each point, and so there is no flux
through the boundary.
In the context of the Poisson equation we choose v = ∇u, and the boundary
condition (3.4.10) translates into the Neumann boundary conditions
∂u
∇u · n = = 0 on ∂Ω. (3.4.11)
∂n
When thermal equilibrium problem is considered, this means that that Ω is ther-
mally insulated. In the case of a membrane, the Neumann conditions represent
the edge of the membrane left free. Both types of boundary conditions can ob-
viously be mixed so that on one part of the boundary, say N , the Neumann
conditions are imposed, while on the other part D, disjoint from N , but such
that ∂Ω = N ∪ D, the Dirichlet conditions are assumed. Moreover, all these
boundary conditions can be generalized to the inhomogeneous form, i.e., a fixed
displacement (temperature) at the boundary and/or prescribing a flux through
the boundary.
To complete our presentation of the boundary value problem for the Pois-
son equation as a self-adjoint problem we must determine whether or not the
boundary value problem is positive definite. We already know that the operator

−4 = ∇∗ ◦ ∇ : U → U

is self-adjoint. To show that it is positive definite we must identify these boundary

conditions for which, according to (1.3.11), ∇u = 0 if and only if u ≡ 0. Relaying
on the elementary fact that the gradient ∇u of a C 1 function u(x, y) defined
on a connected domain Ω vanishes if and only if u is constant, we establish
 3.4. MINIMUM PRINCIPLES 59

that only the homogeneous Dirichlet or homogeneous mixed boundary conditions

are acceptable. Indeed, only these boundary conditions are forcing a constant
function, defined on a connected domain, to vanish identically.
In conclusion, we have properly formulated the Poisson and Laplace’s equa-
tions in positive definite self-adjoint form

−4u = ∇∗ ◦ ∇u = f,

including either homogenous Dirichlet or homogeneous mixed boundary condi-

tions. Hence, we can present now the minimization principle for the Poisson and
Laplace’s equations.

Theorem 3.3. Let u(x, y) be the unique minimizer of the Dirichlet integral
Z Z
1 2 1 1
D[u] ≡ k∇uk − < u; f >= ( u2x + u2y − f u)dxdy (3.4.12)
2 Ω 2 2
among all C 1 functions that satisfy either homogeneous Dirichlet or mixed bound-
ary conditions. Then, u(x, y) is the solution to the Poisson equation (3.3.1).

Although we are not going to prove the Dirichlet minimization principle, we

will prove here, using a version of Green’s formula, that any solution to our
boundary value problem for the Poisson equation is unique. First, we show
that the only harmonic function satisfying the homogeneous Dirichlet or mixed
boundary conditions is the zero function. Starting from (3.4.6) and assuming
that v = ∇u we obtain that
Z Z I
2 ∂u
[u4u + k∇uk ]dxdy = u ds.
Ω ∂Ω ∂n
Supposing that u solves Laplace’s equation and either itself vanishes or its normal
derivative vanishes on ∂Ω, we deduce from the previous identity that
Z Z
k∇uk2 dxdy = 0.
Ω

As the integrant is nonnegative and continuous, the only way for the integral to
vanishes is if k∇uk2 ≡ 0. This implies that ∇u = 0. Therefore, u (a C 1 function)
must be constant. Hence, for the trivial boundary conditions and connected
domain Ω, the only possible choice is u ≡ 0.
60 3. LAPLACE’S EQUATION

Suppose now that u1 and u2 are two solutions to the same inhomogeneous
boundary value problem for the Poisson equation. Their difference u = u1 − u2
satisfies homogeneous boundary conditions and, by linearity, solves the Laplace’s
equation. As we have just established, u = 0, and hence the solutions to the
Poisson equation are the same, u1 = u2 .

Remark 3.4. It should be stressed here that Theorem 3.3 does not guarantee
the existence of the minimizer. It only states that if the minimizer exists then
its solves the given boundary value problem.

Remark 3.5. The Dirichlet integral (3.4.12) is minimized under the assump-
tion that the boundary conditions are homogeneous. In fact, the minimization
principle also applies to inhomogeneous boundary conditions (all due to the fact
that the kernel of the gradient operator is trivial). When the inhomogeneous
Neumann boundary conditions are considered, the minimization principle can
also be applied but with some modification. Namely, the Dirichlet integral must
be supplemented by
Z
∂u
uq ds, where =q on N ⊂ ∂Ω. (3.4.13)
N ∂n
However, the Neumann boundary value problem, in contrast with the Dirichlet
and the mixed boundary value problems, is not positive definite. As we pointed
out in the one-dimensional case - Section 1.3 - the solution is not unique, i.e., if
u(x, y) is a solution, so is u(x, y) + c, for any constant c. This contributes to the
instability of the problem. If we view u(x, y) as the vertical displacement of the
membrane, it takes no force to move the entire membrane up or down retaining
its equilibrium status. In the case of thermal equilibrium, when the Nemannn
conditions prescribe the heat flux through the boundary, the same heat flux can
be observed at different levels of temperature.
 CHAPTER 4

Vibrations of One-Dimensional Medium

4.1. The Wave Equation

We shall consider now Newton’s Law of motion for a one-dimensional medium,

say a string. This means that we will be investigating solutions to a partial
differential equation
µ ¶
∂ 2u ∂ ∂u
%(x) 2 = −K[u] = κ(x) , 0 < x < l, (4.1.1)
∂t ∂x ∂x
where κ(x) > 0 is a material characteristic such as the tension at x of the transver-
sally vibrating string or a stiffness of a longitudinally stretched bar. In fact,
equation (4.1.1) describes a number of other phenomena such as sound waves in
a column of air or water waves.
Since our equation is a partial differential equation of second order in time we
must, in order to get a unique solution, provide the initial displacement as well
as the initial velocity:
∂u
u(0, x) = u0 (x),
(0, x) = v0 (x). (4.1.2)
∂t
Moreover, we must impose some suitable boundary conditions. This means select-
ing such boundary conditions that the underlying (spatial) differential operator
µ ¶
∂ ∂u
K = −D(κ(x)D) = κ(x)
∂x ∂x
is self-adjoint. For example, homogeneous Neumann boundary conditions corre-
spond to a free end while a Dirichlet boundary conditions describe a fixed end.
Assuming that the medium is homogeneous implies that both the density and
the tension are position independent. This simplifies the equation (4.1.1) so that
we obtain the one-dimensional wave equation
∂ 2u 2
2∂ u
= c , (4.1.3)
∂t2 ∂x2
61
62 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

where the constant r

κ
c= . (4.1.4)
%
As in all the instances considered so far, let us try first a separable solution

u(t, x) = v(x) cos ωt. (4.1.5)

Substituting it into the wave equation (4.1.3) and cancelling common factors we
obtain the following ordinary differential equation
d2 v
c2 + ω 2 v = 0. (4.1.6)
dx2
If ω 6= 0, the solutions are the trigonometric functions
ω ω
cos x, sin x.
c c
The corresponding solutions to the wave equation are
ω ω
cos ωt cos x, cos ωt sin x.
c c
Two additional solutions can be obtained by replacing cosine by the sine func-
tion in the original ansatz (4.1.5). These four fundamental solutions represent
standing 2πc/ω-periodic wave solutions vibrating with the frequency ω. Due to
the linearity of the wave equation, we may use now the superposition principle
to write the general solution in the form of the trigonometric Fourier series. The
boundary conditions, which have been neglected so far, will specify particular
eigenvalues and the corresponding fundamental frequencies.
To this end, consider a string of a fixed length l with both ends fixed. In
other words, let us consider the Dirichlet boundary conditions

u(t, 0) = u(t, l) = 0.

In this case, the eigenvalues and the eigenfunctions of the eigenvalue problem (4.1.6)
are
nπc nπx
ωn = , vn (x) = sin , n = 1, 2, 3, . . .
l l
Therefore, the general solution takes the form
X∞ · ¸
nπct nπx nπct nπx
u(t, x) = an cos sin + bn sin sin . (4.1.7)
n=1
l l l l
 4.2. D’ALEMBERT’S SOLUTION 63

This can be re-written as

X∞ µ ¶
nπct nπx
dn cos + δn sin . (4.1.8)
n=1
l l

Thus, the solution is a linear combination of the various modes vibrating with
the frequencies
r
nπc nπ κ
ωn = = , n = 1, 2, 3, . . . . (4.1.9)
l l %
The Fourier coefficients an and bn can be uniquely determined by the initial
condition (4.1.2). Indeed, differentiating our series solution term by term, we
obtain that
∞
X ∞
X
nπx nπc nπx
u(0, x) = an sin = u0 (x), ut (0, x) = bn sin = v0 (x).
n=1
l n1
l l
(4.1.10)
The corresponding Fourier coefficients of the initial displacement u0 (x), are
Z
2 l nπx
an = u0 (x) sin dx, n = 1, 2, 3, . . .
l 0 l
while the scaled Fourier coefficients of the initial velocity v0 (x) take the form
Z l
2 nπx
bn = v0 (x) sin dx, n = 1, 2, 3, . . . .
nπc 0 l

4.2. d’Alembert’s Solution

Let first re-write the wave equation (4.1.3) as

∂2u 2
2∂ u
2
− c 2
= (∂t2 − c2 ∂x2 )u = 0, (4.2.1)
∂t ∂x
where a subscript indicates partial differentiation. Presenting the wave operator

¤ ≡ ∂t2 − c2 ∂x2 (4.2.2)

as a product (superposition) of two first order linear differential operator

¤u = (∂t2 − c2 ∂x2 )u = (∂t + c∂x )(∂t − c∂x )u (4.2.3)

suggests that any solution to the simpler first order equation

(∂t ± c∂x )u = ut ± cux = 0 (4.2.4)

64 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

is a solution to the wave equation (4.1.3). Equations (4.2.4) are simple enough
to be integrated explicitly. Indeed, consider the solution u(t, x) to the transport
(advection) equation 1
ut + cux = 0. (4.2.5)
We may ask ourselves if there exists a curve x(t) such that the solution, when
restricted to this curve, is constant. If such a curve exists then
d dx
u(t, x(t)) = ut + ux = 0. (4.2.6)
dt dt
Comparing (4.2.6) with the advection equation (4.2.5) we conclude easily that
the solution u(t, x) is constant on any strait line ξ = x0 − ct, where x0 denotes
the initial point at t = 0. In fact, any function

u(t, x) = f (x − ct) (4.2.7)

where f (ξ) is a completely arbitrary function, solves the advection equation (4.2.5).
Consequently, by virtue of (4.2.3), it solves also the wave equation (4.1.3). The
same can obviously be said about solutions to the sister transport equation

ut − cux = 0. (4.2.8)

In this case, however, the solution

u(t, x) = g(x + ct), (4.2.9)

where g(η), η = x + ct, is again an arbitrary function. The new variables ξ

and η are called characteristic variables and the corresponding strait lines the
characteristics of the wave equation.
Thus, we have found two different classes of solutions to the wave equation.
They are known as travelling waves moving with with the same speed c in opposite
directions.
Linearity of the wave equation implies that a combination of solutions is again
a solution. Therefore, given f (x − ct) and g(x + ct) as solutions we know that
u(t, x) = f (x − ct) + g(x + ct) is a solution. In fact, every solution to the wave
1This type of equation describes a transport process in which the changes in the value of
u(t, x) are due to movement of the medium. This is in contrast to the diffusion process.
 4.2. D’ALEMBERT’S SOLUTION 65

equation can be represented as a combination of travelling waves propagating in

opposite directions.

Theorem 4.1. The general solution to the wave equation (4.1.3) is a combi-
nation of right and left travelling wave solutions

u(t, x) = f (x − ct) + g(x + ct) (4.2.10)

for some functions f (ξ) and g(η) of the characteristic variables ξ = x − ct and
η = x + ct.

Proof. Let us first rewrite the wave equation in terms of the characteristic
variables ξ and η. Set u(x, t) = w(ξ(t, x), η(t, x)). Then

uxx = wξξ + 2wξη + wηη , utt = c2 [wξξ − 2wξη + wηη ].

This implies that

¤u = −4c2 wηξ .
Solving the wave equation reduces to solving the second order partial differential
equation
∂ 2w
= 0.
∂ξ∂η
Integrated this equation first with respect to ξ and then with respect to η, results
in
w(ξη) = f (ξ) + g(η) = f (x − ct) + g(x + ct).
Some attention need to be paid to the domain of definition of the functions
involved and the differentiability assumptions. ¤

Initial Value Problem on an Infinite Domain.

We have just shown that the general solution to the wave equation depends
on two arbitrary functions of characteristic variables. What is missing are any
initial as well as boundary conditions. Let us first consider an infinite string. In
order to solve the initial value problem

¤u = 0, u(0, x) = f (x), ut (0, x) = g(x), −∞ < x < ∞, (4.2.11)

we must find functions p(ξ) and q(η) such that

u(0, x) = p(x) + q(x) = f (x), and ut (0, x) = −cp0 (x) + cq 0 (x) = g(x).
66 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

Differentiating the first equation, multiplying it by c, and subtracting the second

equation from the first leaves us with
1
2p0 (x) = f 0 (x) − g(x).
c
Hence,
Z x
1 1
p(x) = f (x) − g(y)dy + a,
2 2c 0

where a is an integration constant. Consequently, according to the first equation,

Z
1 1 x
q(x) = f (x) − p(x) = f (x) + g(y)dy − a.
2 2c 0

Substituting both functions back into (4.2.10), we obtain that

Z
f (ξ) + f (η) 1 η
u(x, t) = p(ξ) + q(η) = + g(y)dy,
2 2c ξ

where ξ and η are the characteristic variables.

We have derived d’Alembert’s formula for the solution to the wave equation
for the infinite one-dimensional medium:
Z x+ct
f (x − ct) + f (x + ct) 1
u(t, x) = + g(y)dy. (4.2.12)
2 2c x−ct

Note that we have not yet imposed any boundary conditions. As the medium is
infinite we simply require that the solution is bounded throughout.

Example 4.2. Let us consider the initial value problem (4.2.11) where
2
u(0, x) = e−x , ut (0, x) = 0.

In other words, we consider the situation when the motion of the string is purely
the result of the initial displacement. According to (4.2.12)
1 2 1 2
u(t, x) = e−(x−ct) + e−(x+ct) .
2 2
The initial displacement gets partitioned into two waves travelling in opposite
directions with the same speed and the shape half as tall as the initial one.
 4.2. D’ALEMBERT’S SOLUTION 67

Example 4.3. Consider the wave equation with the characteristic speed c =
1, and the initial condition build out of two pulses concentrated at two different
points. For example, let
2 2
u(0, x) = e−x + e−(x+1) , ut (0, x) = 0.

The solution
1 2 1 2 1 2 1 2
u(t, x) = e−(x−t) + e−(x+1−t) + e−(x+t) + e−(x+1+t)
2 2 2 2
consists of four different pulses travelling through the medium. The first and the
second wave propagate to the left while the third and the fourth propagate to the
right. The first and the last wave run away from each other However, the second
wave and and the third wave collide2 at t = 1/2 but eventually emerge from this
collision unchanged. Indeed,
1 2 1 2 1 2 1 2
u(1, x) = e−(x−1) + e−x + e−(x+1) + e−(x+2) .
2 2 2 2
The derivation of the general solution in the d’Alemebrt’s form was possi-
ble only because we were able to show that, in general, signals propagate only
along characteristics. A nice way to see how this works is to consider the initial
condition in the form of a delta function concentrated at a point, say a,

u(0, x) = δ(x − a), ut (0, x) = 0. (4.2.13)

The corresponding solution takes the form

1 1
u(t, x) = δ(x − ct − a) + δ(x + ct − a), (4.2.14)
2 2
and it does not vanish only at x = ct + a and x = −ct + a respectively. These are
the two characteristics emerging from the initial point (0, a) on the phase plane t−
x. The displacement induced at the position x = a at time t = 0 partitions itself,
and a half of it travels along one characteristic while the other half propagates
along the other characteristic. Both signals advance at the same speed, although
2This is obviously a bit inaccurate as both pulses are not compactly supported (are nonzero
at all points of the real axis rather than only on a finite interval). They interact at all times.
On the other hand, as the exponential functions approach zero very rapidly, the bulk of the
wave is concentrated around the tallest point. Our description may therefore be considered
visually accurate.
68 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

in the opposite directions, and there are no other signals propagating anywhere
else. Observing any particular point of the string one notices that once the wave
passes through, the point returns to its original state of zero displacement. To
see that this is not always the case let us investigate the initial condition of no
displacement but nonzero velocity. In particular, let

u(0, x) = 0, ut (0, x) = δ(x − a). (4.2.15)

The d’Alembert solution (4.2.12) takes the form

Z
1 x+ct
u(t, x) = δa (y)dy = σ−ct+a (x) − σct+a (x)
2c x−ct

 1 , a − ct < x < a + ct,
= 2c (4.2.16)
0, otherwise.

The solution consists of a constant displacement of magnitude 1/2c expanding in

both directions at the constant speed c. In contrast to the concentrated initial
displacement case, the string once displaced will never return to its initial state
of no displacement but it will remain displaced by a fixed amount of 1/2c.

Boundary Value Problem on a Semi-Infinite Interval .

Suppose that a string is semi-infinite. That is, it extends over the region
0 ≤ x < ∞ with its left end at x = 0 fixed. This forms the following initial-
boundary value problem for the motion of the semi-infinite string:


 utt = c2 uxx , 0 < x < ∞, t > 0,




u(x, 0) = f (x),
(4.2.17)

 ut (x, 0) = g(x),




u(0, t) = 0.

According to Theorem 4.1 the general solution of the wave equation is a combi-
nation of travelling waves, that is

u(x, t) = F (x − ct) + G(x + ct),

where the functions F and G are constructed from the initial conditions:
Z
1 1 z
F (z) = f (z) − g(s)ds, (4.2.18a)
2 2c 0
 4.2. D’ALEMBERT’S SOLUTION 69
Z z
1 1
G(z) = f (z) + g(s)ds. (4.2.18b)
2 2c 0

Note that the initial conditions f (x) an g(x) are defined only for x ≥ 0. This
implies that that the wave equation solution (4.2.12) is valid only as long as
x − ct and x + ct remain nonnegative. The value of x + ct is always nonnegative.
However, the value of x − ct is nonnegative only if x ≥ ct. Therefore, whenever
x < ct the solution u(x, t) is not represented by the formula (4.2.12). On the other
hand, even when x < ct the solution u(x, t) must be the sum of two waves, one
travelling to the right and the other travelling to the left. The wave represented
by the function G(x + ct) is still valid, but we shall try to replace F (x − ct) by
a different travelling wave Fb(x − ct). Since the string is fixed at that end, the
boundary condition u(0, t) = 0 implies that

u(0, t) = Fb(−ct) = G(ct) = 0.

This condition must be satisfied at any time t > 0. It shows that the function Fb
is defined by the relation
Fb(z) = −G(−z)

valid for all z < 0, where z = −ct. It follows from (4.2.18b) that the value of the
solution u(x, t) when x − ct < 0 is
Z x+ct
1 1
u(x, t) = (f (x + ct) − f (ct − x)) + g(s)ds. (4.2.19)
2 2c ct−x

Putting the two cases (4.2.12) and (4.2.19) together we conclude that the solution
to the initial-boundary value problem (4.2.17) for the semi-infinite string with one
end fixed, is piecewise defined in the following form:
 R x+ct

 1 1
(f (x + ct) + f (x − ct)) + 2c g(s)ds, x ≥ ct,

 2 x−ct

u(x, t) = (4.2.20)


 1 (f (x + ct) − f (ct − x)) + 1 R x+ct g(s)ds, x < ct.

2 2c ct−x

Remark 4.4. Another way to approach the initial-boundary value problem

for the semi-infinite string with one end fixed is to replace it by the infinite
problem where the initial condition is odd, and therefore satisfies the boundary
70 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

condition u(0, t) = 0. For example, replace the initial-boundary value problem



 utt = c2 uxx , 0 < x < ∞, t > 0,




u(x, 0) = f (x),
(4.2.21)

 u (x, 0) = 0,

 t


u(0, t) = 0,

by the corresponding initial value problem for the infinite string with ut (x, 0) = 0
and u(x, 0) = fo (t), where

f (x) if x ≥ 0,
fo (t) =
−f (−x) if x < 0,

is the odd extension of f (x). It is straightforward to verify that the d’Alembert

solution
1
u(x, t) = (fo (x + ct) + fo (x − ct)) ,
2
satisfies the initial and boundary conditions of (4.2.21).

4.3. Nonlinear Waves

When investigating the wave equation we came across the transport equa-
tion (4.2.5) the solution to which proved to be the wave f (x − ct) propagating at
constant speed c along characteristics x − ct = x0 .
In this section we would like to investigate some properties of solutions to
the nonlinear version of this one-way wave equation. First, let us look at slightly
more complicated, but still linear, version of (4.2.5), that is

ut + c(x)ux = 0. (4.3.1)

The wave speed depends now on the position of the wave. It is, however, not
unreasonable to investigate how the concept of the characteristic generalizes to
this case. By analogy to the uniform case, we define the characteristic curve, as
the solution to the ordinary differential equation
dx
= c(x). (4.3.2)
dt
 4.3. NONLINEAR WAVES 71

We can show (the same way as we showed it before) that the solutions to the
wave equation equation (4.3.1) are constant along characteristic curves. Indeed,
restricting the solution to a characteristic and differentiating it with respect to
time t we obtain that
d dx
u(t, x(t)) = ut + ux = ut + c(x(t))ux = 0.
dt dt
Since u(t, x(t)) solves the equation (4.3.1), we conclude that u(t, x(t)) is constant.
Therefore, u(t, x(t)) = u(0, x(0)) along the characteristic x(t) starting at x(0).
Note that in contrast to the uniform linear case the characteristic curves are not
necessarily strait lines.

Example 4.5. Consider the equation

ut + xux = 0.

The characteristic curves are the solutions to the differential equation

dx
= x.
dt
Integrating the characteristic equation, we find that

x(t) = x0 et .

Since the solution is constant on characteristics, in order to determine the value

of the solution at an arbitrary point (t, x) we must find the characteristic passing
through this point. In particular, we must find the initial point x0 this char-
acteristic starts from. This should not be difficult. From the equation of the
characteristic we get that
x0 = xe−t .

Given the initial condition u(0, x) = u0 (x) we deduce that the solution

u(t, x) = u0 (xe−t ).

Indeed, differentiating our function with respect to t and x and substituting into
the original differential equation confirms that we have found the solution.
72 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

One of the simples, yet very important, nonlinear differential equations is the
nonlinear advection equation

ut + a(u)ux = 0. (4.3.3)

The coefficient a(u) is again interpreted as the wave speed, this time dependent
of the size of the wave.

Remark 4.6. If the coefficient a(u) in the equation (4.3.3) is such that a(u) =
f 0 (u), for some function f (u), then the transport equation takes the form of the
conservation law
ut + f (u)x = 0. (4.3.4)
A conservation law is an equation which accounts for all the ways in which a
particular quantity can change. Suppose that we have a one-dimensional medium.
Let u(t, x) measure the density at position x of some quantity at time t. The
total amount of that quantity in the finite segment [a, b] is given by the integral
Z b
u(t, x)dx.
a

The rate at which this amount changes in time is computed by the derivative
Z
d b
u(t, x)dx. (4.3.5)
dt a
Assuming that our quantity can enter the segment [a, b] only at either of its ends,
the rate at which it enters is given by the flux function. Let f (t, x) denote the
rate at which the quantity flows past position x at time t. The net rate at which
the quantity enters the segment [a, b] is measured by

f (t, a) − f (t, b). (4.3.6)

Assuming that there is no other way in which the given quantity can enter the
medium, we get an integral form of the conservation law:
Z
d b
u(t, x)dx = f (t, a) − f (t, b). (4.3.7)
dt a
This is equivalent to
Z b Z b
ut (t, x)dx = − fx (t, x)dx, (4.3.8)
a a
 4.3. NONLINEAR WAVES 73

provided we can differentiate under the integral and the flux function is regular
enough. If ut and fx are continuous, and the integral conservation law is valid
for any finite segment of the medium, then the integral law (4.3.8) is equivalent
to a conservation law in differential form (4.3.4).

We define the characteristic curves x(t) as solutions to the ordinary differen-

tial equation
dx
= a(u(t, x(t)). (4.3.9)
dt
As the right hand side is solution dependent we cannot determine the charac-
teristic curve without knowing the solution. We can, however, show that the
solution is still constant on characteristics. Assume that x(t) is the character-
istic. Restrict the solution u(t, x) to x(t), and differentiate u(t, x(t)) along the
characteristic. Thus,

d dx
u(t, x(t)) = ut (t, x(t))+ ux (t, x(t)) = ut (t, x(t))+a(u(t, x(t)))ux (t, x(t)) = 0,
dt dt
since u(t, x) is the solution to equation (4.3.3). This implies, in turn, that the
characteristic curves are strait lines

x = a(u)t + x0 , (4.3.10)

where u is the value of the solution on the characteristic as determined by the

value of the initial condition at x0 . Indeed, given the initial condition u0 (x), we
draw through each point (0, x) the characteristic line ξ = a(u0 (x))t + x. The
solution u(t, x(t)) = u0 (x) on the entire line.
Although all characteristic curves are strait lines they slopes a(u) are depen-
dent on the size of the wave. Thus the characteristic slopes depend on the form
of the initial condition and the properties of the nonlinearity a(u). In fact, they
are parallel only if the initial condition is constant. Otherwise they may intersect
at some future time tc . When characteristics cross the solution is no longer well
defined, not at least in the classical sense. Indeed, at the same point in space
and the same time the solution is supposed to have two different values, one
corresponding to each characteristic.
74 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

Example 4.7. Consider the initial value problem for the inviscid Burgers’
equation
ut + uux = 0, u(0, x) = u0 (x). (4.3.11)

The equation (4.3.11) can also be written as the conservation law

µ ¶
1 2
ut + u = 0. (4.3.12)
2 x

The characteristic curves take the form

x = ut + x0 = u0 (x0 )t + x0 . (4.3.13)

Consider now two different characteristic curves emerging at t = 0 from x0 and

x0 + 4x, respectively. These two strait lines will cross if at some future time
t = tc
u0 (x0 )tc + x0 = u0 (x0 + 4x)tc + x0 + 4x.

This yields
4x
tc = − .
u0 (x0 + 4x) − u0 (x0 )
Taking the limit of the right had side when 4x → 0 we obtain that the character-
istic line emanating from (0, x) intersects with another characteristic at critical
time
1
tc = − . (4.3.14)
u00 (x)
Moreover, since tc > 0, the characteristics intersect only if the initial condition
u0 (x) is a decreasing function. If, on the other hand, the initial condition is
nondecreasing, the characteristics never cross and the solution is well defined for
all times. The characteristics emanating from the x-axis create a fan of strait
lines. This solution represents a rarefaction wave which gradually spreads out as
time progresses.

For the general nonlinear wave equation (4.3.3) the critical time at which a
characteristic intersect with another characteristic is given by
1
tc = − 0
. (4.3.15)
a0 (u 0 )u0 (x)
 4.4. CLASSIFICATION OF LINEAR PARTIAL DIFFERENTIAL EQUATIONS 75

Thus, given the initial condition u0 (x) the classical solution exits only for times

0 < t < min tc . (4.3.16)

x

The critical time tc depends on the monotonicity of the initial condition as well as
on the monotonicity of the wave speed a(u0 ). This, in the case of the conservation
law, translates into convexity of the flux f (x).

4.4. Classification of Linear Partial Differential Equations

We have already come across three different linear, second order, partial dif-
ferential equations for functions in two variables. These are:

(1) The heat equation: ut − γuxx = 0,

(2) Laplace equation: uxx + uyy = 0,
(3) The wave equation: utt − c2 uxx = 0.

They represent three fundamentally different, as we explain shortly, classes

of partial differential equations; parabolic, elliptic, and hyperbolic. As we saw
it already, they have distinct types of solutions, possessing completely different
properties, and different numerical and analytical solution techniques. Hyper-
bolic and parabolic equations describe dynamical processes while equations of
equilibrium such as Laplace equation involve only spatial variables and are ellip-
tic. Also, elliptic partial differential equations lead to boundary value problems
whereas hyperbolic and parabolic involve solving initial-boundary value problems.
Although the classification we discuss here concerns only planar differential equa-
tions, the terminology and, if fact, the fundamental properties of solutions carries
over to higher dimensions.
Consider the most general linear, second order partial differential equation in
two variables. Namely,

Auxx + Buxy + Cuyy + Dux + Euy + F u = f (x, y), (4.4.1)

where all coefficients may, in general, be functions of both variables, x and y.

This equation is said to be homogeneous if and only if f ≡ 0. Also, in order to
consider only genuine second order equations, we assume that at least one of the
coefficients A, B, C is nonzero.
76 4. VIBRATIONS OF ONE-DIMENSIONAL MEDIUM

In order to determine the type of the differential equation we calculate its

discriminant
∆(x, y) ≡ B 2 − 4AC. (4.4.2)
The terminology which we use to classify partial differential equations is moti-
vated by the underlining classification of planar conic sections. That is, at a point
(x,y), a linear, second order equation is called:
• elliptic if ∆(x, y) < 0,
• parabolic if ∆(x, y) = 0,
• hyperbolic if ∆(x, y) > 0.

Remark 4.8. Note that as the coefficients of the partial differential equation
are allowed to vary from one point to another, the type of the equation can change
respectively. One such example is the Tricomi equation

yuxx = uyy . (4.4.3)

Its discriminant is ∆ = 4y. Therefore, its type changes with y.

 APPENDIX A

Normed and Inner Product Vector Spaces

A.1. Inner Product Vector Spaces

The concepts of a norm and an inner product on a vector space formalize,

and the same time generalize, the Euclidean notions of length and angle. In
this chapter we make a brief and necessarily rather sketchy presentation of these
concepts, and show how they appear in the context of vector spaces, both finite
and infinite-dimensional. We start our presentation by introducing the notion of
an inner product on a vector space.
Let V denote a real vector space. An inner product on V is a real-valued
function
h·; ·i : V × V → R (A.1.1)
which is bilinear (linear in each argument), symmetric and positive definite. Being
positive definite means that

hv; vi ≥ 0, for any vector v ∈ V, (A.1.2)

and
hv; vi = 0 if and only if v = 0. (A.1.3)
A vector space with a given inner product is called an inner product vector space.
Be aware that any particular vector space may admit several different inner prod-
ucts.

Example A.1. Consider V = Rn . The standard Euclidean inner product is

defined as
n
X
hv; wi = vi wi , (A.1.4)
i=1
where v = (v1 , . . . , vn ) and w = (w1 , . . . , wn ). This is not the only possible inner
product on Rn , although most important. For example, let c1 , . . . , cn be a set of
77
78 A. NORMED AND INNER PRODUCT VECTOR SPACES

positive numbers. Define the weighted inner product

n
X
hv; wi = ci vi wi . (A.1.5)
i=1

Example A.2. Let [a, b] be a bounded, close interval on R. Consider the

vector space C[a, b] of all continuous functions f : [a, b] → R, and define an inner
product as
Z b
hf ; gi = f (x)g(x)dx. (A.1.6)
a
Notice that the assumption that all our functions are continuous on [a, b] is essen-
tial for the given product to be definite. Relaxing the continuity assumption and
extending the selection of functions to include functions which are, for example,
piecewise continuous creates problems. Indeed, consider a function h which is
zero everywhere on [a, b] except at a single point, at which it is 1. It it obvious
that hh; hi = 0 despite the fact that h 6= 0.
The inner product function spaces play essential role in the development of
the theory of Fourier series and the solution to the boundary value problems.

Given the inner product vector space V there exists a norm associated with
the given inner product. Such a norm is defined as the square root of an inner
product of a vector with itself:
p
k v k= hv; vi. (A.1.7)

The positive definiteness of the inner product implies that the norm (A.1.7) is
nonnegative, and that it vanishes if and only if v ∈ V is the zero vector.
In general, a norm on the real vector space V is a non-negative , real-valued
function
k · k : V → R+ (A.1.8)
which is homogeneous:

k cv k= |c| k v k, for any c ∈ R, and v ∈ V, (A.1.9)

satisfies triangle inequality:

k v + w k≤k v k + k w k for any v, w ∈ V, (A.1.10)

 A.1. INNER PRODUCT VECTOR SPACES 79

and is positive definite:

k v k= 0 if and only if v = 0. (A.1.11)

Example A.3. The Euclidean inner product (A.1.4) induces on Rn the norm
v
u n
uX
k v k= hv; vi = t vi2 . (A.1.12)
i=1

The corresponding weighted norm takes the form

v
u n
uX
k v k= t c2i vi2 . (A.1.13)
i=1

The L2 -norm on C[a, b] is:

s
Z b
k f k= f (x)2 dx. (A.1.14)
a

On of the most important relations between the inner product of two vector
and the associated norms of these vectors is the Cauchy-Schwarz inequality.

Theorem A.4. Every inner product satisfies the Cauchy-Schwarz inequality

|hv; wi ≤k v kk w k, v, w ∈ V. (A.1.15)

Equality holds if and only if the vectors v and w are parallel1.

Proof. First, observe that if any of the vectors v, w is zero the inequality is
trivially satisfied as both sides vanish. We, therefore, assume that both vectors
are different from the zero vector. Second, consider the function

f (t) =k v + tw k2

where t ∈ R is an arbitrary scalar. Using the bilinearity of the inner product and
the definition of the norm (A.1.7) we obtain that

f (t) =k w k2 t2 + 2hv; wit+ k v k .

1Two vectors are considered parallel if there exists a scalar such that one vector is the
multiple of the other. According to this convention the zero vector is parallel to every other
vector.
80 A. NORMED AND INNER PRODUCT VECTOR SPACES

The function f (t) in nonnegative and it attains a minimum as its leading coeffi-
cient is positive. The minimum is taken at a point at which the derivative f 0 (t)
vanishes. Namely,
f 0 (t) = 2 k w k2 t + 2hv; wi2 = 0,
when
hv; wi
t=− .
k w k2
Substituting this into the definition of the function we obtain that
hv; wi2
k v k2 ≥ , or equivalently k v k2 k w k2 ≥ hv; wi2 .
k w k2
Taking the square root of both sides completes the proof. ¤

Two vectors v, w ∈ V , of an inner product vector space V , are called orthogo-

nal if hv; wi = 0. In Rn (with the Euclidean inner product) orthogonality means
geometric perpendicularity. In spaces of functions such geometric analogies are
not available. Also, being orthogonal with respect to one inner product does not
imply being orthogonal with respect to another inner product.

Example A.5. Consider the space C[0, π] with the L2 -inner product. There,
the function sin x is orthogonal to cos x. Indeed,
Z π ¯π
1 2 ¯¯
hsin x; cos xi = sin x cos xdx = sin x¯ = 0.
0 2 0

Example A.6. Let P 2 [0, 1] be the space of all polynomials of degree not
bigger than 2 defined on the interval [0, 1]. It is elementary to show that the
polynomials 1 and x are orthogonal with respect to the standard inner product,

ha0 + a1 x + a2 x2 ; b0 + b1 x + b2 x2 i = a0 b0 + a1 b1 + a2 b2 .

However, the same two polynomials are not orthogonal in the L2 -inner product
as ¯1
Z 1
1 2 ¯¯ 1
h1; xi = xdx = x ¯ = .
0 2 0 2
One of the consequences of the Cauchy-Schwarz inequality is the triangle
inequality (A.1.10).
 A.2. NORMED VECTOR SPACES 81

Theorem A.7. The norm associated with an inner product satisfies the tri-
angle inequality
k v + w k≤k v k + k w k (A.1.16)
for every pair of vectors v, w ∈ V . The equality holds if and only if v and w are
parallel.

Proof. Consider k v + w k2 and use the Cauchy-Schwarz inequality (A.1.15)

to show that

k v + w k2 =k v k2 +2hv; wi+ k w k2
≤k v k2 +2 k v kk w k + k w k2 = (k v k + k w k)2 .

Taking the square root of both sides of the inequality completes the proof. ¤

A.2. Normed Vector Spaces

An inner product vector space has the associated inner product norm, as
we showed in the previous section. On the other hand, a vector space may be
equipped with a norm which does not come from any inner product.

Remark A.8. Given a norm in a vector space we may define the notion of a
distance between vectors:

d(v, w) =k v − w k . (A.2.1)

Realize that this function possess all the properties we expect from a function
measuring distance. It is symmetric, vanishes if and only if v = w, and satisfies
the triangle inequality

d(v, w) ≤ d(v, u) + d(u, w) (A.2.2)

for any choice of vectors u, v, w ∈ V .

Let us look now at some norms which are not associated with any inner
product.

Example A.9. Consider again Rn . The 1-norm of a vector is defined as:

k v k1 = |v1 | + · · · + |vn |. (A.2.3)

82 A. NORMED AND INNER PRODUCT VECTOR SPACES

The p-norm, for any integer p ≥ 1, is

v
u n
uX
k v kp = t
p
|vi |p . (A.2.4)
i=1

Taking p to ∞ yields ∞-norm

k v k∞ = sup {|vi |}. (A.2.5)

1≤i≤n

We leave as an exercise to verify that these are indeed well defined norms2.

The norms in Rn have they counterparts in the space of functions.

Example A.10. Consider once again C[a, b], the space of all continuous func-
tions on a closed, bounded interval [a, b]. The Lp -norm is defined as
µZ b ¶ p1
p
k f kp = |f (x)| x . (A.2.6)
a

Respectively, the L∞ -norm is defined as

k f k∞ = sup {|f (x)|}. (A.2.7)

a≤x≤b

The L2 -norm is the only norm associated with an inner product.

Having all these different norms available one may wounder if they are equiv-
alent in some sense. The answer to this question depends on whether the vector
space is finite-dimensional like Rn or infinite-dimensional like the space C[a, b].
That is, any two norms in a finite-dimensional space are equivalent in the follow-
ing sense:

Theorem A.11. Let k · k1 and k · k2 be any two norms in Rn . Then, there

exist positive constants c and C such that

c k v k1 ≤k v k2 ≤ C k v k1 (A.2.8)

for every v ∈ Rn .

The proof of this fact can be found in [Lang].

2The triangle inequality for the p-norm, known as the Hölder inequality, is non-trivial,
[Taylor].
 A.2. NORMED VECTOR SPACES 83

Example A.12. Let us compare the 1-norm and the ∞-norm on Rn . To this
end, note first that k v k∞ = sup{|vi |} ≤ |v1 | + · · · + |vn | =k v k1 . Thus, we may
1 1
choose in (A.2.8) C = 1. Also, k v k∞ = sup{|vi |} ≥ n
(|v1 | + · · · + |vn |) = n
k
1
v k1 . Therefore, c = n
and we have that
1
k v k1 ≤k v k∞ ≤k v k1 . (A.2.9)
n
The important consequence of the equivalence of norms in any finite-dimensional
vector space, for example Rn , is that the convergence of sequences is norm inde-
pendent.
We say that the sequence {v(k) } ⊂ V converges in the norm to v̂ ∈ V if the se-
quence of scalars k v(k) − v̂ k converges to 0. It is evident from Theorem A.11 that
the convergence with respect to one norm implies the convergence with respect
to any other norm. Also, the convergence in the norm implies the convergence
(k)
of the individual components, i.e., vi → v̂i . The converse is obviously true as
well. This is, in general, not true in infinite-dimensional spaces.

Example A.13. Consider the space C[0, 1]. The sequence of continuous func-
tions 
−nx + 1, 0 ≤ x ≤ 1
n
fn (x) =
0, 1
≤x≤1 n
∞
is such that its L -norm

k fn k∞ = sup {|fn (x)|} = 1.

0≤x≤1

On the other hand, the L2 -norm is

ÃZ 1 ! 21
n 1
k fn k 2 = (1 − nx)2 dx =√ .
0 3n
It approaches zero when n → ∞. Therefore, there is no constant C such that

k fn k∞ ≤ C k fn k2

for this choice of functions, and so for all functions in C[0, 1]. This proves that
the norms L∞ and L2 are not equivalent in C[0, 1]. Notice, however, that

k f k2 ≤k f k∞ (A.2.10)
84 A. NORMED AND INNER PRODUCT VECTOR SPACES

for any f ∈ C[0, 1]. The convergence in the L∞ -norm implies the convergence in
the L2 -norm but not vice versa.

A.3. Complex Vector Spaces

In this section we present a few basic facts about complex vector spaces that is
the vector spaces in the definition of which the set of real scalars is replaced by the
set of complex numbers C. In this set of notes, where we deal with real quantities
like the measure of deformation or temperature, we use complex numbers and
complex vector spaces primarily to simplify presentation of periodic phenomena.
There are, however, physical theories, e.g., quantum mechanics, where complex
valued functions are intrinsic as they describe basic physical quantities.
The fundamental example of the complex vector space is the space Cn consist-
ing of n-tuples of complex numbers (u1 , u2 , · · · , un ), where u1 , . . . , un ∈ C. We
can write any vector u ∈ Cn as a linear combination of two real vectors x, y ∈ Rn ,
namely u = x + iy. Its complex conjugate u is obtained by taking the complex
conjugate of its coordinates, that is u = x − iy.
Most vector space concepts curry over from the real case to the complex realm.
The only notable exception is the concept of the inner product. Motivated by
the desire to have the real associated norm3 on a complex vector we define the
inner product on Cn as

v · w = v1 w 1 + v2 w 2 + · · · + vn w n , (A.3.1)

where v, w ∈ Cn . This construction is known as the Hermitian inner product on

Cn . For example, if

v = (i, −1), and w = (1 + i, i),

then
v · w = i(1 − i) + (−1)(−i) = 1 + 2i.
However,
w · v = (1 + i)(−i) + i(−1) = 1 − 2i,
3Complex numbers cannot be ordered. Therefore, it does not make any sense to have a
complex number z > 0.
 A.3. COMPLEX VECTOR SPACES 85

which shows that the Hermitian dot product is not symmetric. It conjugates,
however, under reversal of arguments. Indeed,

w · v = v · w, (A.3.2)

for any v, w ∈ Cn . Also, the dot product is ”sesquilinear ”, rather than bilinear,
as

(cv) · w = c(v · w), while v · (cw) = c(v · w). (A.3.3)

On the other hand, the associated norm looks the same as in the real case:
√ p
k v k= v·v = |v1 |2 + · · · + |vn |2 , (A.3.4)

and it is positive for any v 6= 0. | · | denotes here the complex modulus.

The general definition of an inner product on a complex vector space is based
on the Hermitian dot product on Cn and it states that an inner product on a
complex vector space V is a complex-valued function h·; ·i : V × V → C which is
sesquilinear, positive definite and such that

hv; wi = hw; vi

for any pair of vectors v, w ∈ V .

One can show that the Cauchy-Schwarz inequality (A.1.15), in which the
absolute value is replaced by the complex modulus, holds for any inner product
complex vector space.

Example A.14. Let C[−π, π] denote the space of all complex valued contin-
uous function on the interval [−π, π] ⊂ R. The Hermitian L2 -inner product is
defined as
Z π
1
hf ; gi = f (x)g(x)dx. (A.3.5)
2π −π

The associated with this inner product norm is

µZ π ¶ 12
2
k f k= |f (x)| dx . (A.3.6)
−π
86 A. NORMED AND INNER PRODUCT VECTOR SPACES

For example, the set of complex exponential eikx , where k is any integer, is an
orthonormal system of functions. Indeed,

Z π Z π 1, k = m,
1 1
heikx ; eimx i = eikx e−imx dx = ei(k−m)x dx = (A.3.7)
2π −π 2π −π 0, k =
6 m,
k
where we utilized the fact that eikπ = (eiπ ) = (cos π + i sin π)k = (−1)k . The
orthogonality property (A.3.7) of complex exponentials is of significants in the
Fourier theory, as we will see in the next chapter.
 APPENDIX B

Fourier Theory

This chapter serves as a brief introduction to the Fourier theory and related
topics.

B.1. Fourier Series

The basic object of study of the Fourier theory is an infinite trigonometric

series
∞
a0 X
+ [ak cos kx + bk sin kx] , (B.1.1)
2 k=1
called a Fourier series. In general, such a series may not converge without ad-
ditional assumptions about its coefficients. We would like to know if a Fourier
series can converge to a function, and whether a function can be represented by
a Fourier series.
A Fourier series of the function f (x) is a Fourier series (B.1.1) where
Z
1 π
ak = f (x) cos kxdx, k = 0, 1, 2, . . . , (B.1.2a)
π −π
Z
1 π
bk = f (x) sin kxdx, k = 1, 2, 3, . . . , (B.1.2b)
π −π
and where we assume that the integrals are well defined. In fact, as we will see
later, these integrals are well defined for a quite a broad class of functions f (x).
This obviously does not guarantee the convergence of the series, and certainly
not to the function f (x). The choice of the coefficients is, however, dictated by
our desire to be able to represent a function by a Fourier series. Indeed, suppose
that f (x) is well approximated over the interval [−π, π] by a Fourier series, that
is
∞
a0 X
f (x) ∼
= + [ak cos kx + bk sin kx] (B.1.3)
2 k=1
87
88 B. FOURIER THEORY

for some choice of the coefficients ak and bk . Multiply both sides of (B.1.3) by
cos lx and use integration by parts to show that
Z Z
1 π 1 π
cos kx cos lxdx = sin kx cos lxdx = 0 if k 6= l, , (B.1.4a)
π −π π −π
and Z Z
1 π 2 1 π
cos lxdx = sin2 lxdx = 1. (B.1.4b)
π −π π −π
This implies immediately the choice of the Fourier coefficients (B.1.2).

Example B.1. Consider the function f (x) = x. Computing its Fourier coef-
ficients directly we obtain that:
Z Z
1 π 1 π
a0 = xdx = 0, ak = x cos kxdx = 0, (B.1.5)
π −π π −π

and
Z π · ¸
1 1 x cos kx sin kx π 2
bk = x sin kxdx = − + 2
|−π = (−1)k+1 . (B.1.6)
π −π π k k k
The vanishing of the coefficients ak is a consequence of the fact that the function
x is odd while cos kx is an even function. Therefore, the Fourier series of f (x) = x
is
X∞
sin kx
2 (−1)k+1 . (B.1.7)
k=1
k
It is not an elementary exercise to determine the convergence of this series. But,
even if we overcome this obstacle and determine that it converges, one does
not know what it converges to. For example, it certainly does not converge to
f (x) = x. Indeed, at x = π the series (B.1.7) converges to 0, as every term in
the series vanishes at π, while f (π) = π 6= 0.

As our example shows the convergence of Fourier series is not a simple matter.
The standard tests used to analyze the convergence of power series fail. Also,
power series always converge at least at 0 or on an interval (possibly infinite)
centered at 0. Fourier series, on the other hand, may converge on rather un-
usual sets. Moreover, as all components of a Fourier series are 2π periodic the
series will, in general, converge to 2π periodic functions. The power series and
the Fourier series differ also by what they converge to. Indeed, if a power series
 B.1. FOURIER SERIES 89

converges on an interval it converges to an infinitely differentiable function as

all its derivative series converge to the corresponding derivatives. Such functions
are called analytic. A Fourier series, on the other hand, may converge to dis-
continuous functions and even to such a ”bizarre” function as the Dirac delta
function.
We will look first at the issue of periodicity of the limit of a Fourier series.
Given a function f (x) for −π < x ≤ π let us define its periodic extension as a
function f˜(x) defined everywhere and such that f˜(x + 2π) = f˜(x), and f˜(x) =
f (x) for −π < x ≤ π. To this end, let x ∈ R. Thus, there exists a unique integer
m such that (2m − 1)π < x ≤ (2m + 1)π. We therefore postulate that

f˜(x) = f (x − 2mπ) for (2m − 1)π < x ≤ (2m + 1)π.

Such a function is obviously unique, 2π periodic, and coincides with f (x) on

−π < x ≤ π. In most cases such a periodic extension is not a continuous
function. It is, however, ”piecewise continuous”.

Definition B.2. A function f (x) is called piecewise continuous on the inter-

val [a, b] if it is defined and continuous except possibly a finite number of point.
Moreover, at each ”point of discontinuity”, say xi , the left and the right hand
limits
f+ (xi ) ≡ lim+ f (x), f− (xi ) ≡ lim− f (x),
x→xi x→xi

exist. A function defined on all of R is piecewise continuous if it is piecewise

continuous on every bounded interval.

Note that according to this definition a piecewise discontinuous function may

not be defined at a ”point of discontinuity”. Even if it is defined its value f (xi ) is
not necessarily equal to either right or left hand limit. Such a point xi is called a
jump discontinuity and we say that the function f (x) experiences a jump of the
magnitude
[f (xi )] ≡ f+ (xi ) − f− (xi ). (B.1.8)
If the function f (x) is piecewise continuous its (formal) Fourier series is well
defined as the function is integrable on any closed interval. The convergence of
90 B. FOURIER THEORY

such a series, even if the function is continuous on [−π, π] interval, is still not
guaranteed.

Definition B.3. A function f (x) is called piecewise C1 on the interval [a, b] if

it is piecewise continuous on [a, b], and continuously differentiable on [a, b] except
possibly a finite number of points. If xi is such a point then the left and right
hand limits
f+0 (xi ) ≡ lim+ f 0 (x), f−0 (xi ) ≡ lim− f 0 (x), (B.1.9)
x→xi x→xi

exist.

Note that a piecewise C1 function may have points at which either the function
experiences a jump but the left and right hand derivatives exist or it is continuous
but its derivative experiences a jump. The periodic extension of the function
f (x) = x has jump discontinuities of the first type at π + 2mπ while the absolute
value |x| is continuous everywhere but its derivative experiences a jump at 0. We
are now in the position to state the convergence theorem for Fourier series. The
proof will be presented in Section B.4.

Theorem B.4. Let f˜(x) be a periodic extension of a piecewise C1 function

on [−π, π], then its Fourier series converges at all x to:

f˜(x) whenever f˜ is continuous,

1h˜ i
f+ (x) + f˜− (x) if x is a point of discontinuity.
2

In other words, if in the definition of the periodic extension we replace the

values at any discontinuity point x by
1h˜ i
f˜(x) = f+ (x) + f˜− (x) ,
2
the Fourier series of a piecewise C1 periodic extension f˜(x) will converge to f˜(x)
everywhere.
 B.1. FOURIER SERIES 91

Example B.5. Let us revisit example B.1, but let the periodic extension of
f (x) = x be such that

f˜((2m + 1)π) = 0, for any integer m.

Then the Fourier series ∞

X sin kx
2 (−1)k+1
k=1
k
converges everywhere to the ”new” periodic extension, and the convergence prob-
lem we previously encountered at π is solved. Indeed, for any x = (2m + 1)π all
terms in our Fourier series vanish.

Example B.6. Let f (x) = |x|. Observe that as |x| is an even function its
periodic extension is continuous. Its Fourier coefficients can easily be computed:
Z Z
1 π 2 π
a0 = |x|dx = xdx = π,
π −π π 0
Z
2 π 4
ak = x cos xdx = − 2 , if k is odd, and otherwise 0.
π 0 k π
Z
1 π
bk = |x| sin xdx = 0,
π −π
as the function |x| sin x is odd. Therefore, the Fourier series of f (x) = |x| is
∞
π 4 X cos(2k + 1)
− . (B.1.10)
2 π k=0 (2k + 1)2
According to Theorem B.4, it converges to the periodic extension of f (x) = |x|.
In particular, if x = 0 we obtain that
∞
π 4X 1
= . (B.1.11)
2 π k=0 (2k + 1)2
This series can be used to obtain an approximation to the number π.

As we have noted in Examples B.5 and B.6 the coefficients of the Fourier
cosine series of the function f (x) = x, and the sine coefficients of the function
f (x) = |x|, are 0. This is not a coincidence, but rather a consequence of the fact
that x is an odd function while |x| is an even function. Indeed, the following
proposition, whose proof is elementary and will be left to the reader, generalizes
our observations.
92 B. FOURIER THEORY

Proposition B.7. If f (x) is an even function then its Fourier series sine
coefficients bk all vanish, and f can be represented by a Fourier cosine series
∞
a0 X
f (x) ∼
= + ak cos kx. (B.1.12)
2 k=1

If f (x) is odd, then its Fourier series cosine coefficients ak all vanish. Thus, f
can be represented by a Fourier sine series
∞
X
f (x) ∼
= ak sin kx. (B.1.13)
k=1

Conversely, a convergent Fourier cosine (sine) series always represents an even

(odd) function.

B.2. Differentiation and Integration of Fourier Series

Knowing that power series can be differentiated and integrated term by term,
and that these two operations do not change (except for the end points of the
interval of convergence) the convergence of these series, it make sense to inves-
tigate whether a similar property holds for Fourier series. The main difference
between these two cases is that the power series converge to analytic functions,
and hence can be freely differentiated and integrated, while the Fourier series may
converge to functions of very different degrees of smoothness. Thus, investigating
differentiation and integration of Fourier series we must pay careful attention to
the regularity of its limits.

Integration. When attempting to integrate a Fourier series we are faced with

the fundamental problem that, in general, the integral of the periodic function is
not periodic. However, a Fourier series consists mostly of sine and cosine terms
which when integrated are also period functions. The only term we are evidently
going to have a problem with is the free term a0 /2. Hence, we will first try to
integrate Fourier series with the zero constant term:
Z
1 π
a0 = f (x)dx = 0.
π −π
This shows that a function has no constant term in its Fourier series if its average
on the interval [−π, π] is zero. The periodic zero average functions are the once
 B.2. DIFFERENTIATION AND INTEGRATION OF FOURIER SERIES 93

that remain periodic upon integration. Indeed, using the definition of a periodic
function one can easily confirm the following property:

Lemma B.8. Let f (x) be 2π periodic. Then, the integral

Z x
g(x) ≡ f (x)dx
0

is a 2π periodic function if and only if f has zero average on the interval [−π, π].

Furthermore, simple integration by parts shows that:

Theorem B.9. If f (x) is a piecewise continuous, 2π periodic function, and

has zero average, then its Fourier series can be integrated term by term to produce
the Fourier series of
Z x X∞ · ¸
∼ ak bk
g(x) = f (y)dy = m + sin kx − cos kx , (B.2.1)
0 k=1
k k
where Z π
1
m≡ g(x)dx
2π −π
is the average value of g(x).

Note that integrating the Fourier series (B.2.1) on the interval [−π, π], and
observing that the average of any odd function is zero we obtain that
X∞
bk
m= . (B.2.2)
k=1
k
This provides a convenient alternative derivation of the sine coefficients of a
Fourier series.

Example B.10. Let us consider again the function f (x) = x. This function
is odd hence, has zero average. Integrating its Fourier series (B.1.7) from 0 to x
we obtain the series ∞
X (−1)k−1
2 (1 − cos kx) .
k=1
k2
The constant term of this series is the average of x2 /2. As in (B.2.2):
X∞ X∞ Z π 2
bk (−1)k−1 1 x π2
m= =2 2
= dx = .
k=1
k k=1
k 2π −π 2 6
94 B. FOURIER THEORY

This immediately yields the Fourier series of the function x2 :

X∞
2∼ π2 cos kx
x = −4 (−1)k+1 2
.
3 k=1
k

As x2 is an even function its Fourier series converges everywhere to its periodic

extension.

Differentiation. As much as integration is making functions nicer, differen-

tiation does the opposite. Therefore, in order to secure the convergence of the
derivative of a Fourier series we must start with the sufficiently nice function;
but how nice. If the Theorem B.4 is to be applicable the derivative f 0 (x) must
be piecewise C1 . This means that the function f (x) must be at least continuous
and piecewise C2 .

Theorem B.11. If f (x) is continuous, piecewise C2 and 2π periodic, then

the term by term differentiation of its Fourier series produces the Fourier series
for the derivative
∞
X
g(x) = f (x) ∼
0
= k [bk cos kx − ak sin kx] . (B.2.3)
k=1

Example B.12. Consider again f (x) = |x|. Differentiating its Fourier se-
ries (B.1.10) we obtain
∞
4 X sin kx
. (B.2.4)
π k=1 k
On the other hand, the derivative of the absolute value |x| can be represented as
a difference of two step functions

 1, x > 0,
σ(x) − σ(−x) =
−1, x < 0.

A simple calculation shows that the Fourier series of the step function σ(x) is
∞
1 2 X sin kx
+ . (B.2.5)
2 π k=1 k
 B.2. DIFFERENTIATION AND INTEGRATION OF FOURIER SERIES 95

It is now easy to see that the derivative of the Fourier series of |x| is indeed the
Fourier series of the difference σ(x)−σ(−x). Moreover, according to Theorem B.4
it converges to it.

Remark B.13. So far, we have defined a Fourier series on the interval [−π, π]
only. This procedure can obviously be easily adopted to any other 2π length
interval. In many applications the functions we deal with are defined on intervals
of other lengths. Therefore, it would help to show how the formulas we developed
so far change if we change the length of the interval. First, let us note that any
symmetric interval [−l, l] can be rescaled to the interval [−π, π] by the simple
change of variables
π
y= x.
l
The rescaled function µ ¶
b l
f (y) ≡ f y ,
π
which lives on [−π, π] interval, has the standard Fourier series
∞
a0 X
fb(y) ∼
= + [ak cos ky + bk sin ky] .
2 k=1

Going back to the original variable x we deduce that

∞ · ¸
∼ a0 X kπx kπx
f (x) = + ak cos + bk sin , (B.2.6)
2 k=1
l l
where, thanks to the our change of variables, the Fourier coefficients ak and bk
have the modified formulas
Z Z
1 l kπx 1 l
kπx
ak = f (x) cos dx, bk = f (x) sin dx. (B.2.7)
l −l l l −l l
According to our convergence Theorem B.4 the given Fourier series converges to
the 2l periodic extension of f (x) with the midpoint values at jump discontinuities,
provided f (x) is piecewise C1 on [−l, l].
If the function f (x) is defined on an arbitrary interval [a, b] then the first step
in the process of developing its Fourier series is to rescale it into a symmetric
interval. This can easily be done by the translation
1
y = x − (a + b).
2
96 B. FOURIER THEORY
£1 ¤
The new interval 2
(a − b), 12 (b − a) is symmetric of length b − a, and the latter
procedure can be utilized.

B.3. Complex Fourier Series and the Delta Function

A natural, and often very convenient, generalization of Fourier series is its

complex counterpart, where exponential functions with complex exponents re-
place sine and cosine functions. Indeed, using Euler’s formula

eikx = cos kx + i sin kx, (B.3.1)

we can represent trigonometric functions sine and cosine as

eikx + e−ikx eikx − e−ikx
cos kx = , and sin kx = . (B.3.2)
2 2i
Substituting these relations into a Fourier series of a piecewise continuous (real
or complex valued) function we obtain a new Fourier series representation
∞
X ∞
X k=∞
X
−ikx ikx
c−k e + ck e = ck eikx , (B.3.3)
k=1 k=0 k=−∞

where
1
ck = (ak + ibk ), if k ≤ 0, (B.3.4a)
2
and
1
ck = (ak − ibk ), if k ≥ 0. (B.3.4b)
2
This and Euler’s formula (B.3.1) show, in fact, that the complex Fourier coeffi-
cients can be evaluated directly as
Z π
1
ck = f (x)e−ikx dx. (B.3.5)
2π −π

Remark B.14. This result is deeper than it seems. It is a consequence (as

in the real case) of the orthonormality of the complex exponential functions with
respect to the Hermitian inner product
Z π
1
hf ; gi ≡ f (x)g(x)dx, (B.3.6)
2π −π
 B.3. COMPLEX FOURIER SERIES AND THE DELTA FUNCTION 97

where g(x) denotes the conjugate. Indeed, evoking Euler’s formula (B.3.1) one
can easily show that

Z π 1,
­ ikx ilx ® 1 k = l,
e ;e = ei(k−l)x dx = (B.3.7)
2π −π 0, k 6= l.

Therefore, multiplying the complex Fourier series (B.3.3) by eilx and integrating
term by term we obtain the formula for the complex Fourier coefficients (B.3.5).

Example B.15. Let us develop the complex Fourier series for the function
f (x) = x. Its complex Fourier coefficients are
Z π
1 (−1)k+1 (−1)k i
ck = xe−ikx dx = = .
2π −π ik k
Consequently, the complex Fourier series of f (x) = x is
X∞
(−1)k i ikx
e .
k=−∞
k

The reader is asked to show, using once again Euler’s formula (B.3.1), that this
is exactly the same series as the sine Fourier series (B.1.7).

Example B.16 (Dirac delta function). In this example we shall investigate

a Fourier series representation of the delta function to prove that a Fourier series
can converge to a generalized function, i.e., a function which although is a limit
of a sequence of piecewise continuous functions is not a standard function itself.
This example shows also the benefits of using complex Fourier series rather than
its real counterpart. First, let us compute complex Fourier coefficients:
Z π Z ∞
1 −ikx 1 1 −ik0 1
ck = δ(x)e dx = δ(x)e−ikx dx = e = . (B.3.8)
2π −π 2π −∞ 2π 2π
Therefore, the complex Fourier series of the delta function has the form
k=∞
1 X ikx
e . (B.3.9)
2π k=−∞

Observe that this series has the form of an infinite (in both directions) geometric
series with the ratio r = eix . This is in contrast to the real Fourier series of the
98 B. FOURIER THEORY

delta function which, as easily deduced directly from its definition, is

∞
1 1X
+ cos kx. (B.3.10)
2π π k=1
These are obviously two different representations of the same series. However, in
order to show that such a series converges to the delta function we will use the
complex version.
The complex Fourier series of the delta function (B.3.9) has a form of the geo-
metric series, as we have indicated earlier. In order to determine its convergence
let us consider its nth partial sum
k=n
1 X ikx
sn (x) = e .
2π k=−n

This is the 2n+1 partial sum of the geometric series with the initial term a = e−inx
and the ratio r = eix . It can be therefore computed exactly that
k=n · ¸ · ¸
1 X ikx 1 −inx ei(2n+1)x − 1 1 ei(n+1)x − e−inx
sn (x) = e = e =
2π k=−n 2π eix − 1 2π eix − 1
· x ¸ " 1 1
#
1 e−i 2 (ei(n+1)x − e−inx ) 1 ei(n+ 2 )x − e−i(n+ 2 )x
= x = x x
2π e−i 2 (eix − 1) 2π ei 2 − e−i 2
1 sin(n + 12 )x
= ,
2π sin x2
(B.3.11)
where the representation (B.3.2) of the sine function in terms of complex expo-
nentials was utilized. The sequence of partial sums sn (x) converges at x = 0 to
infinity, as easily attested from the (B.3.9). Moreover,
Z π Z π Xk=n Z π X
k=n
1 ikx 1
sn (x)dx = e dx = [cos kx + i sin kx] dx = 1
−π 2π −π k=−n 2π −π k=−n
(B.3.12)
as required for the convergence to the delta function. At any other point x the
sequence sn (x) does not converge to zero. In fact, one can see from (B.3.11) that
it oscillates faster and faster. It appears, however, when added up over a large
domain, that the oscillations cancel out. The Fourier series of the delta function
 B.4. CONVERGENCE OF THE FOURIER SERIES 99

e point-wise. It may be shown

δ(x) does not converge to its periodic extension δ(x)
that it converges in a weak sense, i.e., in the sense of some integral average.

B.4. Convergence of the Fourier Series

In this section we present basic convergence results for Fourier series. A

proper, although inevitably limited, presentation of these fundamental results
requires some familiarity with the concepts and methods of advanced mathemat-
ical analysis. In our effort to make this presentation as self contained as possible
we introduce the necessary tools as they become necessary for the analysis of
the convergence of Fourier series. We start our presentation by analyzing the
pointwise convergence of a Fourier series.

Pointwise Convergence.

Proof of Theorem B.4. Our objective is to show that, given a 2π periodic

piecewise C1 function f (x), the limit of the sequence of partial Fourier sums sn (x)
is the arithmetic average of the left hand and the right hand limits of this function,
i.e.,
1
lim sn (x) = [f+ (x) + f− (x)] .
n→∞ 2
Consider
n
X Xn µ Z π ¶
ikx 1 −iky
sn (x) = ck e = f (y)e dy eikx
k=−n k=−n
2π −π
Z π n
X
1
= f (y) eik(x−y) dy,
2π −π k=−n

where the formula (B.3.5) for the complex Fourier coefficients was used. Using
the summation formula (B.3.11) and the periodicity of the functions involved we
calculate that
Z Z
1 π
sin(n + 12 )(x − y) 1 π
sin(n + 12 )y
sn (x) = f (y) 1 dy = f (x + y) dy,
2π −π sin 2 (x − y) 2π −π sin 12 y
where to obtain the last integral we changed the variables from y to x + y. If we
could now show that
Z
1 π
sin(n + 12 )y
lim f (x + y) dy = f+ (x), (B.4.1a)
n→∞ π 0 sin 12 y
100 B. FOURIER THEORY

and Z
1 0
sin(n + 12 )y
lim f (x + y) dy = f− (x) (B.4.1b)
n→∞ π −π sin 21 y
the proof would be complete. To this end, recalling (B.3.12) allows us to claim
that Z
1 π
sin(n + 12 )y
dy = 1.
π 0 sin 12 y
Consequently, the statement (B.4.1a) can be replaced by the equivalent statement
Z π
f (x + y) − f+ (x) 1
lim y sin(n + )ydy = 0. (B.4.2)
n→∞ 0 sin 2 2
Let
f (x + y) − f+ (x)
g(y) ≡ . (B.4.3)
sin y2
Then, using the trigonometric identity for sin(n + 12 ), and Riemann’s Lemma 1 we
notice that
Z π
f (x + y) − f+ (x) 1
lim y sin(n + )ydy
n→∞ 0 sin 2 2
Z π³ Z
1 y´ 1 π³ y´
= lim g(y) sin cos nydy + lim g(y) cos sin nydy = 0
n→∞ π 0 2 n→∞ π 0 2

as long as the function g(y) sin y2 is piecewise continuous. This completes our
proof provided we can show that g(y) is piecewise continuous on [0, π]. Looking
at the definition of the function g(y) we can easily see that this is true except
possibly at y = 0. However,
y
f (x + y) − f+ (x)
lim+ g(y) = lim+ y
2
= 2f+0 (x),
y→0 y→0
2
sin y2
as the function f (x) is piecewise C1 . This fact confirms that g(y) is piecewise
continuous. Identical arguments prove (B.4.1b). ¤
1Riemann’s Lemma states that if a function f (x) is piecewise continuous on [a, b] then
Z b
lim f (x) cos αxdx = 0,
α→∞ a

and similarly if cosine is replaced by sine. For the proof of this fact see for example [Lang].
 B.4. CONVERGENCE OF THE FOURIER SERIES 101

Uniform Convergence. Let us consider a finite-dimensional vector space

V , e.g., Rn . In fact, all finite-dimensional spaces of the same dimension are
essentially identical (isomorphic), as we stated earlier. In order to be able to
talk about convergence of sequences in vector spaces we need a measure of a
distance between vectors. Hence, let V be a normed vector space with the norm
k · k : V → R (A.1.8).

Definition B.17. Let W be an normed vector space. We say that a sequence

vn ∈ W converges in the norm to w ∈ W if ||vn − w|| → 0 as n → ∞.

In a finite-dimensional vector space this is equivalent to a convergence of

individual components of uk = (uk1 , . . . , ukn ) to the corresponding components of
u = (u1 , . . . , un ). This is one way of saying that in a finite-dimensional space
all norms are equivalent (see Theorem A.11), i.e., a convergence in one norm
quarantines the convergence in any other norm. This, however, is not true in
general in the infinite-dimensional vector spaces. All because of the fact that
there are many, not necessarily equivalent, norms in such spaces - Example A.13.
Also, in infinite dimensional vector spaces the convergence in the norm does not
imply - as we will see later - pointwise convergence

lim uki (x) = ui (x), i = 1, . . . , ∞ for all x, (B.4.4)

k→∞

of the sequence of values of the functions uki (x).

In addition to the convergence in the norm and the pointwise convergence
mechanisms there exists yet another form of convergence; the uniform conver-
gence.

Definition B.18. A sequence of real valued function fk (x) converges uni-

formly to a function f (x) on an interval I ⊂ R if, given ² > 0 there exists an
integer n such that
|fk (x) − f (x)| < ²
for all x ∈ I and all k ≥ n.

The name uniform converges is a reflection of the fact that the sequence of
functions converges ”the same way” at all points of the interval I. The choice
102 B. FOURIER THEORY

of the integer n depends only on ² and not on a point x ∈ I. Note also, that
although uniform convergence implies pointwise convergence the converse may
1
not be true. For example, consider the sequence of functions fk (x) = kx
on
the interval (0, 1). The sequence converges pointwise to f (x) ≡ 0 at all points
x ∈ (0, 1) but it does not converge uniformly on (0, 1) as given ² there is no
possible choice of the integer n. One can always select a point x so close to 0
1
that nx
is larger than ².
The greatest advantage of uniform convergence is that it preserves continuity.

Theorem B.19. If a sequence of continuous functions fk (x) converges uni-

formly on I then the limit f (x) is a continuous function on I.

As we are interested here primarily in the convergence of series the following

test proves to be particularly useful:

Theorem B.20 (Weierstrass test). Suppose that the function fk (x) are bounded,
i.e.,
|fk (x)| ≤ mk for all x ∈ I,
where mk are positive constants. If, in addition, the series
∞
X
mk
k=1

converges, then the series

∞
X
fk (x)
k=1
converges uniformly on the interval I to a function f (x). In particular, if the
partial sums sn (x) are continuous functions so is the limit f (x).

We advise the reader to prove that one can integrate and differentiate a uni-
formly convergent series term by term to obtain a uniformly convergent series,
provided the series of derivatives is also uniformly convergent.

Proposition B.21. If the series

∞
X
fk (x) = f (x)
k=1
 B.4. CONVERGENCE OF THE FOURIER SERIES 103

is uniformly convergent then

Z x ÃX∞
! ∞ Z
X x Z x
fk (y) dy = fk (y)dy = f (y)dy
a k=1 k=1 a a

is uniformly convergent. Also, if the series of derivatives

∞
X
fk0 (x) = g(x)
k=1

is uniformly convergent then the original series converges uniformly and g(x) =
f 0 (x).

Let us now examine a uniform convergence of a complex Fourier series

∞
X
ck e−ikx . (B.4.5)
k=−∞

Since x is real the magnitude

p
|e−ikx | = | cos kx + i sin kx| = cos2 kx + sin2 kx = 1,

and
|ck eikx | ≤ |ck | for all x.
Applying the Weierstrass test (Theorem B.20) we immediately deduce that
1
Theorem B.22. If the coefficients ck = (a
2 k
± ibk ) of a complex Fourier
series (B.4.5) are such that
∞
X X∞ q
1
|ck | = a0 + a2k + b2k < ∞ (B.4.6)
k=−∞
2 k=1

then the Fourier series converges uniformly to a continuous function f (x). More-
over, the coefficients ck are equal to the Fourier coefficients of the function f (x).

Although Theorem B.22 gives conditions guaranteeing the convergence of a

Fourier series, it does not tell if the limit f (x) is the original function f˜(x) for
which the Fourier series was derived. Indeed, knowing that the coefficients are
derived by integration one may suspect that the functions f (x) and f˜(x) differ
at a finite, or even a countable, number of point. In fact, f˜(x) may not be
continuous. However, it can be shown that if a function is periodic and regular
104 B. FOURIER THEORY

enough its Fourier series converges uniformly to the very function it got derived
from.

Theorem B.23. Let f (x) be 2π periodic and piecewise C1 . If f (x) is contin-

uous on an interval (a, b) then its Fourier series converges uniformly to f (x) on
any subinterval [a + δ, b − δ] where δ > 0.

In other words, as long as we stay away from the discontinuities of f˜(x) its
Fourier series converges uniformly.
The uniform convergence of a Fourier series, as sanctioned by Theorem B.22,
requires that the Fourier coefficients ck approach 0 when k → ∞. However, the
convergence of the Fourier coefficients to zero is only a necessary condition. The
coefficients need to converge fast enough to guarantee that the sum is finite. As
we know from elementary calculus these coefficients must converge to zero faster
then k1 . If, for example,
M
|ck | ≤ for all k sufficiently large, where β > 1, (B.4.7)
|k|β
then using the standard ratio test we can conclude that the series of coeffi-
cients ck converges absolutely. This yields, according to the uniform convergence
test (B.4.6), the uniform convergence of the Fourier series.
Note also that the faster the Fourier coefficients converge to zero the smoothes
the limit function become. Indeed, suppose that for some positive integer n
M
|ck | ≤ β+n for all k sufficiently large, where β > 1. (B.4.8)
|k|
Then, we can differentiate the Fourier series up to n-times obtaining always,
according to Proposition B.21, a uniformly convergent series.

Theorem B.24. If the Fourier coefficients are such that

X∞
k n |ck | < ∞, (B.4.9)
k=−∞

for some nonnegative integer n, then the Fourier series converges to a 2π periodic
function f˜(x) which is n times continuously differentiable. Moreover, for any
m ≤ n, the m times differentiated Fourier series converges uniformly to the
corresponding derivative f˜(m) (x).
 B.4. CONVERGENCE OF THE FOURIER SERIES 105

Hence, we may analyze the smoothness of a function by looking at how fast

its Fourier coefficients approach zero. For example, if they converge to zero
exponentially the function is infinitely differentiable but not necessarily analytic.

Convergence in the Norm. In order to be able to discuss this important

aspect of convergence of Fourier series we need to take a short detour introducing
the concept of a Hilbert space. The precise definition of a Hilbert space, which is
rather technical, is beyond the scope of these notes. We will therefore present here
a ”working” version which will enable us to deal with Fourier series of functions
such as, for example, the delta function.

Definition B.25. A complex-valued function f (x) is said to be square-

integrable on the [−π, π] if
µ Z π ¶ 12
1 2
||f || ≡ |f (x)| dx < ∞. (B.4.10)
2π −π

In particular, but not only, every piecewise continuous function is square-

integrable. Other, more singular, functions may also be square-integrable, e.g.,
x−r is as long as r < 12 . Note also, that a function may not be square-integrable
on some intervals while it is on some others.
The set of all square-integrable functions on [−π, π], which is usually denoted
as L2 [−π, π], is a complex vector space with the pointwise operations of addition
and multiplication by a scalar. This is a strait forward consequence of the triangle
inequality
|af (x) + bg(x)| ≤ |a||f (x)| + |b||g(x)|.
Moreover, it is a normed vectors space with the L2 -norm given by (B.4.10) pro-
vided, we identify as one any two functions which differ at most on a measure
zero set (for example, at a finite number of points of [−π, π])2. L2 space is a very
special normed vector space as the L2 -norm is based on the (Hermitian) inner
product Z π
1
hf ; gi ≡ f (x)g(x)dx. (B.4.11)
2π −π

2Such an identification is necessary as otherwise the L2 -norm would be zero for other than
the constant zero function. In fact, it would not be a norm according to its definition A.1.8.
106 B. FOURIER THEORY

Indeed, hf ; f i = ||f ||2 . The inner product is well defined and finite for any two
square-integrable functions thanks to the Cauchy-Schwarz inequality.

Example B.26. Consider a square-integrable function f (x) on [−π, π]. Its

Fourier coefficients are inner products of f with the exponential functions:
Z π
1
ck = f (x)e−ikx dx = hf ; eikx i.
2π −π
Note also that as shown in (B.3.7)

||eikx ||2 = heikx ; eikx i = 1,

while
heikx ; eilx i = 0 if k 6= l.
We say that the infinite system of complex exponentials is an orthonormal system
of functions (vectors) in L2 [−π, π], i.e., all elements have unit length and are
mutually orthogonal as the scaler product of any two different elements vanish.

The sequence of square-integrable functions fn converges in the L2 -norm to

g(x) ∈ L2 [−π, π] if
Z π
2 1
||fn (x) − g(x)|| = |fn (x) − g(x)|2 dx → 0 as n → ∞. (B.4.12)
2π −π

Hence, let us consider the convergence of the sequence of partial sums sn (x) of
the Fourier series of a square-integrable function f (x). As each element of the
sequence sn (x) is obviously square-integrable we can evaluate ||f − sn ||2 . Using
the definition of the sequence sn (x), and the linearity and symmetry of the inner
product we obtain that
n
X
2 2 2 2
||f − sn || = ||f || − 2hf ; sn i + ||sn || = ||f || − 2 ck hf ; eikx i + ||sn ||2
k=−n
n
X n
X n
X
2 2 2
= ||f || − 2 ck ck + ||sn || = ||f || − 2 ck ck + |ck |2
k=−n k=−n k=−n
2 2
= ||f || − ||sn || .

(B.4.13)
 B.4. CONVERGENCE OF THE FOURIER SERIES 107

As the left hand side is always nonnegative

n
X
|ck |2 ≤ ||f ||2
k=−n

for any integer n. We have therefore proved the following:

Proposition B.27 (Bessel’s inequality). If the function f is square-integrable

on [−π, π] then its Fourier coefficients are such that
∞ ∞
1 2 1 X¡ 2 ¢ X
a0 + ak + b2k = |ck |2 ≤ ||f ||2 . (B.4.14)
4 2 k=1 −∞

This immediately implies that the individual Fourier coefficients (both com-
plex and real) of a square-integrable function approach zero as |k| approaches
infinity. The convergence of the series (B.4.14) requires, in fact, more. The
Fourier coefficients must tend to zero fast enough. If we postulate, as we did
before in (B.4.7), that
M
|ck | ≤ for sufficiently large k, (B.4.15)
|k|β
1
then selecting β > 2
guarantees the convergence already. Note this is a slower
rate of decay than the one needed for the uniform convergence. One may have a
Fourier series which converges in the L2 -norm but not uniformly.
If the Fourier series of a square-integrable function f (x) converges in the L2 -
norm to the very function it got derived from the Bessel’s inequality becomes
equality. Indeed, as evident from (B.4.13)

||f − sn ||2 = ||f ||2 − ||sn ||2 (B.4.16)

for any Fourier partial sum sn (x). Therefore,

lim ||f − sn || = 0,
n→∞

if and only if
∞
X
2 2
||f || = lim ||sn || = |ck |2 . (B.4.17)
n→∞
k=−∞

This formula becomes a convenient criterion for the L2 -convergence of square-

integrable functions.
108 B. FOURIER THEORY

Proposition B.28 (Plancherel formula). A Fourier series of a square in-

tegrable function f ∈ L2 [−π, π] converges in the L2 -norm to f (x) if and only
if
∞
X
2
||f || = |ck |2 . (B.4.18)
k=−∞

Before we finally show that the Fourier series of every L2 -function f (x) con-
verges in the L2 -norm to f (x) let us note that one of the immediate consequences
of the Plancheler formula is that a function is uniquely determined by its Fourier
coefficients.

Corollary B.29. Two functions f, g ∈ L2 [−π, π] have the same Fourier

coefficients if and only if f = g.

The main result of the theory of Fourier series is the following convergence
theorem.

Theorem B.30. Let sn (x) denote the nth partial sum of the Fourier series of
the square-integrable function f (x) ∈ L2 [−π, π], then

lim ||f − sn || = 0. (B.4.19)

n→∞

In other words, any square-integrable function on [−π, π] can be uniquely rep-

resented by the infinite system3 of complex exponentials eikx , k = 0, ±1, ±2, . . . .

Proof of Theorem B.30. We shall provide here the prove for continuous
functions only. Proving the validity of this statement for all square-integrable
functions requires some extra work, see [Carrier, Krook and Pearson], and
also [Brown and Churchill], and [Kammler]. According to Theorem B.23, if
a function f (x) is piecewise C1 and continuous, its Fourier series converges to
f (x) everywhere, i.e.,
∞
X
f (x) = ck eikx .
k=−∞

3This property of an infinite system of elements of a Hilbert space is known as the com-
pleteness of such a system. As the exponentials functions are orthonormal in L2 [−π, π], this is
the completeness of an orthonormal system of vectors.
 B.5. FOURIER TRANSFORM 109

Therefore, utilizing the fact that we are allowed to multiply and integrate uni-
formly convergent series term by term, we have
Z π Z π X∞ Z π
2 1 2
||f || = |f (x)| dx = f (x)f (x) = f (x)ck e−ikx dx
2π −π −π k=−∞ −π
∞
X ∞
X
= ck ck = |ck |2 ,
k=−∞ k=−∞

and Plancherel’s identity (B.4.18) holds proving the convergence of the Fourier
series. ¤

B.5. Fourier Transform

The Fourier transform may be viewed as a generalization of the Fourier series

to an infinite interval. That is, take a Fourier series on a symmetric interval [−l, l],
and consider taking the limit as the length l → ∞. The result of such a process
is a Fourier series on an infinite interval. The corresponding representation of
a function is given the name of the Fourier transform. Indeed, let us look for
the representation of a f (x) by a rescaled complex Fourier series (B.3.3) on an
interval [−l, l] in the following form
∞ r
X π fˆ(kr ) ikr x
e (B.5.1)
r=−∞
2 l

where the sum is over the discrete set of frequencies

rπ
kr = , r = 0, ±1, ±2, . . . , (B.5.2)
l
to incorporate all trigonometric functions of period 2l. Therefore, based on the
form of Fourier coefficients (B.3.5) we get that
Z l
1
fˆ(kr ) = √ f (x)e−ikr x dx. (B.5.3)
2π −l
This allows us to pass to the limit l → ∞, and to get the infinite integral
Z ∞
ˆ 1
f (k) = √ f (x)e−ikx dx (B.5.4)
2π −∞
110 B. FOURIER THEORY

known as the Fourier transform of the function f (x). If the function f (x) is
piecewise continuous and decays to 0 reasonably fast as x → ±∞, the Fourier
transform fˆ(k) is defined for all frequencies −∞ < k < ∞.

Remark B.31. Note that the discrete frequencies (B.5.2) used to represent
the function f (x) on the interval [−l, l] are equally spaced as
π
4k = kr+1 − kr = .
l
As l → ∞, the spacing between frequencies 4k → 0, and the frequencies become
more and more densely distributed. That is why in the Fourier transform limit
we anticipate that all frequencies participate in representing a function.

Taking into account the fact that the discrete frequencies kr are equally spaced
allows us to re-write the Fourier series (B.5.1) as
∞
1 X ˆ
√ f (kr )eikr x 4 k. (B.5.5)
2π r=−∞

This has the form of a Riemann sum for the function fˆ(k)eikx . Assuming that it
converges as 4k → 0 we obtain
Z ∞
1
√ fˆ(k)eikx dk. (B.5.6)
2π −∞

The series (B.5.5) becomes a Fourier integral which reconstructs the function
f (x) as a continuous superposition of complex exponential functions eikx . For
example, if the function f (x) is piecewise continuously differentiable everywhere
and it decays reasonably fast as |x| → ∞, then the inverse Fourier integral (B.5.6)
converges to f (x) at all points of continuity. At the jump discontinuities it
converges to the midpoint 12 [f+ (x) − f− (x)]. Indeed, it can be shown4 that

Theorem B.32. If the function f (x) is piecewise continuous and square-

integrable (for −∞ < x < ∞), then its Fourier transform fb(k) is well defined
and square-integrable. Moreover, if the right and left hand limits f− (x), f+ (x)
and f−0 (x), f+0 (x) exist, then the Fourier integral (B.5.6) converges to the average
4A rigorous proof of a more general statement can be found
in [Carrier, Krook and Pearson].
 B.5. FOURIER TRANSFORM 111

value 12 [f− (x) + f+ (x)]. In particular, if f (x) is continuously differentiable at a

point x, then the Fourier integral equals the value of f (x).

Example B.33. Let us consider an exponentially decaying pulse

f (x) = e−a|x| . (B.5.7)

Its Fourier transform

Z ∞ Z 0 Z ∞
ˆ 1 −a|x| −ikx 1 (a−ik)x 1
f (k) = √ e e dx = √ e dx + √ e−(a+ik)x dx
2π −∞ 2π −∞ 2π 0
¯0 ¯∞ r
1 e(a−ik)x ¯¯ 1 e−(a+ik)x ¯¯ 2 a
=√ − √ = .
2π a − ik ¯ −∞ 2π a + ik ¯ 0 π k 2 + a2
(B.5.8)

Notice that the Fourier transform of our pulse f (x) which is real and even, is
itself real and even. The inverse Fourier transform gives
Z Z
−a|x| 1 ∞ aeikx 1 ∞ a cos kx
e = dk = dk. (B.5.9)
π −∞ k 2 + a2 π −∞ k 2 + a2
Here the imaginary part of the integral vanishes as the integrant is odd.

We have been discussing the Fourier transform of the pulse function not with-
out reason. First, it is interesting to notice that as a → 0 the pulse approaches the
constant function g(x) ≡ 1. Moreover, the limit of its Fourier transform (B.5.8)
is 
r 0,
2 2a k 6= 0
lim = . (B.5.10)
a→0 π k 2 + a2 ∞, k = 0

Comparing this with the original construction (1.2.8) of the delta function as the
limit of approximating functions we notice that
n a
δ(x) = lim 2 2
= lim , (B.5.11)
n→∞ π(1 + n x ) a→0 π(a + x2 )
2

provided n = 1/a. This, in turn, allows us to write the Fourier transform of the
constant function f (x) ≡ 1 as
√
fˆ(k) = 2πδ(k). (B.5.12)
112 B. FOURIER THEORY

Equivalently, this should mean that

Z ∞
1
δ(k) = e−ikx dx. (B.5.13)
2π −∞

The infinite integral on the right hand side does not converge5. However, from
the definition of the delta function we have
Z ∞
δ(k)e−ikx dk = eik0 = 1, (B.5.14)
−∞

which implies that the Fourier transform of the delta function is a constant func-
tion
Z ∞
1 e−ik0 1
δ̂(k) = √ δ(x)e−ikx dx = √ =√ . (B.5.15)
2π −∞ 2π 2π
To determine the Fourier transform of the delta function concentrated at an
arbitrary position y, namely δy (x) we cite the following theorem:

Theorem B.34.
• If the Fourier transform of the function f (x) is fˆ(k), then the transform
of fˆ(x) is f (−k).
• If the function f (x) has the Fourier transform fˆ(k), then the Fourier
transform of the shifted function f (x − y) is e−iky fˆ(k). By analogy,
the transform of the product einx f (x) is the shifted Fourier transform
fˆ(k − n).

Therefore, according to (B.5.15) the Fourier transform of the delta function

at y 6= 0 is
e−iky
δ̂y (k) = √ . (B.5.16)
2π
Thus, combining the definitions of the Fourier transform and its inverse, as well
as using the basic properties of the delta function leads to the identity
Z ∞Z ∞
1
f (y) = f (x)e−ik(x−y) dxdk (B.5.17)
2π −∞ −∞
valid for suitable functions.
5It may, however, be interpreted in the context of generalized functions.
 B.5. FOURIER TRANSFORM 113

Differentiation and Integration.

As our main objective in dealing with Fourier transforms is to show how they
can be used to solve differential equations we embark now on the analysis of the
process of differentiation and integration of these functions.
Let us consider the Fourier transform of the derivative of f (x)
Z ∞
b 1
f (k) = √
0 f 0 (x)e−ikx dx, (B.5.18)
2π −∞
assuming that f 0 (x) is behaving well enough so the integral exists. Integrating
by parts, and taking into account that f (x) approaches 0 at infinities rapidly
enough, we obtain that
Z ∞
ik
fb0 (k) = √ f (x)e−ikx dx. (B.5.19)
2π −∞

Thus, we have that

Proposition B.35. The Fourier transform of the derivative f 0 (x) of a func-

tion is obtained by multiplying its Fourier transform by ik:

fb0 (k) = ik fˆ(k). (B.5.20)

Iterating the formula (B.5.20) yields that

Corollary B.36. The Fourier transform of f (n) (x) is (ik)n fˆ(k).

Consider now Z x
g(x) = f (y)dy.
−∞
We are interested in finding the Fourier transform ĝ(k). To this end notice first
that Z ∞
lim g(x) = 0, while lim = f (y)dy = c.
x→−∞ x→+∞ −∞
Therefore, consider the function h(x) = g(x) − cσ(x) which decays to 0 at both
±∞. As the Fourier transform of the step function σ(x) is
r
2 i
σ̂(k) = δ(k) − √ (B.5.21)
π k 2π
we obtain that r
2 ic
ĥ(k) = ĝ(k) − c δ(k) + √ . (B.5.22)
π k 2π
114 B. FOURIER THEORY

On the other hand,

h0 (x) = f (x) − cδ(x).

Applying Proposition B.35 we conclude after some manipulations that

r
fˆ(k) π
ĝ(k) = +c δ(k). (B.5.23)
ik 2
Example B.37. Consider the boundary value problem

−u00 + ω 2 u = h(x), −∞ < x < ∞, (B.5.24)

where ω > 0, and the solution is assumed square-integrable implying that u

approaches 0 at ±∞. Taking the Fourier transform of both sides of the differential
equation renders an algebraic equation relating the Fourier transforms of u and
h. Indeed,
k2u b=b
b(k) + ω 2 u h(k). (B.5.25)

The transformed equation can now be solved for the Fourier transform of the
solution
b
h(k)
u
b(k) = . (B.5.26)
k + ω2
2

Using the Fourier formula (B.5.6) we reconstruct the solution as

Z ∞ b
1 h(k)eikx
u(x) = √ 2 2
dk. (B.5.27)
2π −∞ k + ω

One of the most important and interesting cases is the one in which h(x) = δy (x),
that is when the forcing term is in the form of the impulse concentrated at y.
The corresponding solution is the Green’s function u(x) = G(x, y). According
to (B.5.26) and Theorem B.34, its Fourier transform with respect to x is
−iky
b y) = e
G(k, . (B.5.28)
k2 + ω2
Notice first that according to Example B.33 the inverse Fourier transform of the
reciprocal of k 2 + ω 2 is
e−ωx
.
2ω
 B.5. FOURIER TRANSFORM 115

Secondly, the exponential term in the numerator implies a shift. Therefore, the
Green’s function for our boundary-value problem is
e−ω|x−y|
G(x, y) = .
2ω
The Green’s function satisfies the differential equation everywhere except at x = y
where it has a jump discontinuity of unit magnitude. It also satisfies the boundary
conditions as it decays to 0 as |x| → ∞. Invoking the general superposition prin-
ciple for Green’s function we obtain the solution to our boundary-value problem
with an arbitrary forcing term as
Z ∞ Z ∞
1
u(x) = G(x, y)h(y)dy = e−ω|x−y| h(y)dy. (B.5.29)
−∞ 2ω −∞

Note that the Green’s function G(x, y) depends only on the difference x − y
and that the solution u(x) takes the form of the convolution. Namely,
Z ∞
u(x) = g(x − y)h(y)dy = g(x) ∗ h(x). (B.5.30)
−∞

On the other hand, as we saw earlier in (B.5.26), the Fourier transform of the
solution to our boundary value problem is a product of the Fourier transforms of
the Green’s function and the forcing term. Indeed, we have that:

Theorem B.38. The Fourier transform of the convolution u(x) = g(x) ∗ h(x)
of two functions is, up to multiple, the product of their Fourier transforms
√
b(k) = 2πb
u g (k)b
h(k). (B.5.31)

By symmetry, the Fourier transform of the product h(x) = f (x)g(x) is a multiple

of the convolution of their Fourier transforms
b 1
h(k) = √ fb(k) ∗ gb(k). (B.5.32)
2π
 Bibliography

[Bleecker and Csordas] , D. Bleeker and G. Csordas, Basic Partial Differential Equations, In-
ternational Press, Cambridge, Massachusetts, 1996.
[Boyce and DiPrima] , W. E. Boyce and R. C. DiPrima, Elementary Differential Equations and
Boundary Value Problems, 5th edition, John Wiley & Sons Inc., New York, 1992.
[Brown and Churchill] , J. W. Brown, and R. V. Churchill, Fourier Series and Boundary Value
Problems, McGraw-Hill, New York, 1993.
[Carrier, Krook and Pearson] , G. F. Carrier, M. Krook, and C. E. Pearson, Functions of a
Complex Variable, Theory and Technique, McGraw-Hill, New York, 1966.
[Davis] , Mathematics of Wave Propagation, Princeton University Press, Princeton, NJ, 2000.
[Kammler] D. W. Kammler, A First Course in Fourier Analysis, Prentice Hall, Upper Saddler
River, NJ, 2000.
[Keane] , M. K. Keane, A Very Applied Course in Partial Differential Equations, Prentice Hall,
Upper Saddler River, NJ, 2002.
[Kevorkian] J. Kevorkian, Partial Differential Equations, Second Edition, Springer-Verlag, New
York, 2000.
[Knobel] R. Knobel, An Introduction to the Mathematical Theory of Waves, American Math-
ematical Society, Providance, 2000.
[Lang] S. Lang, Undergraduate Analysis, Springer-Verlag, New York, 1983.
[Ogden] R. W. Ogden, Non-Linear Elastic Deformations, Dover Publications, Mineola, N.Y.,
1997.
[Smoller] J. Smoller, Shock Waves and Reaction-Diffusion Equations, Springer-Verlag, New
York, 1983.
[Taylor] M. E. Taylor, Partial Differential Equations (Basic Theory), Springer-Verlag, New
York, 1996.
[Ziemer] W. P. Ziemer, Weakly Differentiable Functions, Springer-Verlag, New York, 1989.

117
 Index

Adjoint Constitutive relation, 2

of the linear operator, 13 Convergence
of the gradient, 57 in the norm, 83
Advection equation, 64 pointwise of a Fourier series, 99
uniform, 101
Bar, 1 uniform of a Fourier series, 101
equation of equilibrium of, 3 weak, 99
Green’s function of, 11, 13 Convolution, 115
Beam, 18
Green’s function of, 21
Delta function, 5
Bending moment, 18
derivative of, 9
Bessel’s inequality, 107
Fourier series of, 97
Boundary conditions,
Fourier transform of, 112
Dirichlet, 25
Diffusion,
Neumann, 25
equation, 23
self-adjoint, 20
d’Alembert’s formula, 66
Burgers’ equation, 23
Dirichlet
inviscid, 74
boundary condition, 25
quasilinear, 41
integral, 59
Characteristics, 64 Displacement, 1
Characteristic curve, 70, 73
Characteristic variables, 64 Eigenfunction, 27
Cauchy-Schwarz inequality, 79 Eigenvalue, 27
Cole-Hopf transformation, 41 Error function, 41
Complex Fourier series, 96 complimentary, 44
Complex vector space, 84 Euler equation, 50
Conservation law, 72 characteristic equation of, 50
Conservation of energy law, 24 Euler-Lagrange equation, 17
119
120 INDEX

Flux function, 72 Inner product, 77

Fourier coefficients, 88 Euclidean, 77
Complex Fourier coefficients, 96 Hermitian L2 , 85
Fourier series, 87 sesquilinear, 85
differentiation of, 94 weighted, 78
integration of, 92 Inner product vector space, 77
pointwise convergence of, 99 complex, 84
Fourier’s Law of Cooling, 24
Fourier integral, 110 Jump discontinuity, 8, 89
Fourier transform, 110
derivative of, 113 Laplace equation, 45
Function in polar coordinates, 49
2
analytic, 89 L −inner product, 14
average of, 92,
generalized, 6, 97, Method of images, 39

periodic extension of, 89 Minimum Principle, 13

piecewise continuous, 89
Neumann
square-integrable, 105
boundary condition, 25
homogeneous boundary conditions, 26
Generalized function, 6
Norm, 78, 78
Green’s formula, 57
1-norm, 81
Green’s function, 5, 10
p-norm, 82
∞-norm, 82
Harmonic function, 54 Lp -norm, 82
Heat capacity, 23 L∞ -norm, 82
Heat equation, 25
smoothing property of, 29 Operator
uniqueness of solutions, 33 self-adjoint, 15
fundamental solution, 36, 37 positive definite, 15
inhomogeneous, 37 eigenvalue problem of, 27
periodic boundary conditions of, 32
in polar coordinates, 49 Phase plane, 67
Heat flux, 24 Plancherel formula, 108
Hermitian inner product, 84, 96 Pointwise convergence, 101
Hölder inequality, 82 Poisson equation, 45, 52
Hook’s Law, 2 Greens’s function of, 56
Homogenization transformation, 30 the minimization principle for, 59
 INDEX 121

Radial Gauss distributions, 53

Ramp function, 13
Rarefaction wave, 74
Riemann’s Lemma, 100

Separation of variables, 26
Statically determinate configuration, 4
Statically indeterminate configuration, 4
Step function, 7
derivative of, 8
Stiffness, 2
Strain, 1
bending, 19
Strain energy, 17
Stress, 1
Superposition integral, 34
Superposition principle, 5

Thermal conductivity, 5, 24
Transport equation, 64
Travelling waves, 64
Triangle inequality, 78

Uniform convergence, 101

Wave equation, 61
Weierstrass test, 102