# Introduction to Partial Diﬀerential Equations

John Douglas Moore
May 21, 2003
Preface
Partial diﬀerential equations are often used to construct models of the most
basic theories underlying physics and engineering. For example, the system of
partial diﬀerential equations known as Maxwell’s equations can be written on
the back of a post card, yet from these equations one can derive the entire theory
of electricity and magnetism, including light.
Our goal here is to develop the most basic ideas from the theory of partial dif-
ferential equations, and apply them to the simplest models arising from physics.
In particular, we will present some of the elegant mathematics that can be used
to describe the vibrating circular membrane. We will see that the frequencies
of a circular drum are essentially the eigenvalues from an eigenvector-eigenvalue
problem for Bessel’s equation, an ordinary diﬀerential equation which can be
solved quite nicely using the technique of power series expansions. Thus we
start our presentation with a review of power series, which the student should
have seen in a previous calculus course.
It is not easy to master the theory of partial diﬀerential equations. Unlike
the theory of ordinary diﬀerential equations, which relies on the “fundamental
existence and uniqueness theorem,” there is no single theorem which is central
to the subject. Instead, there are separate theories used for each of the major
types of partial diﬀerential equations that commonly arise.
However, there are several basic skills which are essential for studying all
types of partial diﬀerential equations. Before reading these notes, students
should understand how to solve the simplest ordinary diﬀerential equations,
such as the equation of exponential growth dy/dx = ky and the equation of
simple harmonic motion d
2
y/dx
2
+ ωy = 0, and how these equations arise in
modeling population growth and the motion of a weight attached to the ceiling
by means of a spring. It is remarkable how frequently these basic equations
arise in applications. Students should also understand how to solve ﬁrst-order
linear systems of diﬀerential equations with constant coeﬃcients in an arbitrary
number of unknowns using vectors and matrices with real or complex entries.
(This topic will be reviewed in the second chapter.) Familiarity is also needed
with the basics of vector calculus, including the gradient, divergence and curl,
and the integral theorems which relate them to each other. Finally, one needs
ability to carry out lengthy calculations with conﬁdence. Needless to say, all
of these skills are necessary for a thorough understanding of the mathematical
i
language that is an essential foundation for the sciences and engineering.
Moreover, the subject of partial diﬀerential equations should not be studied
in isolation, because much intuition comes from a thorough understanding of
applications. The individual branches of the subject are concerned with the
special types of partial diﬀerential equations which which are needed to model
diﬀusion, wave motion, equilibria of membranes and so forth. The behavior
of physical systems often suggest theorems which can be proven via rigorous
mathematics. (This last point, and the book itself, can be best appreciated by
those who have taken a course in rigorous mathematical proof, such as a course
in mathematical inquiry, whether at the high school or university level.)
Moreover, the objects modeled make it clear that there should be a constant
tension between the discrete and continuous. For example, a vibrating string
can be regarded proﬁtably as a continuous object, yet if one looks at a ﬁne
enough scale, the string is made up of molecules, suggesting a discrete model
with a large number of variables. Moreover, we will see that although a partial
diﬀerential equation provides an elegant continuous model for a vibrating mem-
brane, the numerical method used to do actual calculations may approximate
this continuous model with a discrete mechanical system with a large number of
degrees of freedom. The eigenvalue problem for a diﬀerential equation thereby
becomes approximated by an eigenvalue problem for an n n matrix where n
is large, thereby providing a link between the techniques studied in linear al-
gebra and those of partial diﬀerential equations. The reader should be aware
that there are many cases in which a discrete model may actually provide a
better description of the phenomenon under study than a continuous one. One
should also be aware that probabilistic techniques provide an additional compo-
nent to model building, alongside the partial diﬀerential equations and discrete
mechanical systems with many degrees of freedom described in these pages.
There is a major dichotomy that runs through the subject—linear versus
nonlinear. It is actually linear partial diﬀerential equations for which the tech-
nique of linear algebra prove to be so eﬀective. This book is concerned primarly
with linear partial diﬀerential equations—yet it is the nonlinear partial diﬀeren-
tial equations that provide the most intriguing questions for research. Nonlinear
partial diﬀerential equations include the Einstein ﬁeld equations from general
relativity and the Navier-Stokes equations which describe ﬂuid motion. We hope
the linear theory presented here will whet the student’s appetite for studying
the deeper waters of the nonlinear theory.
The author would appreciate comments that may help improve the next
version of this short book. He hopes to make a list of corrections available at
the web site:
http://www.math.ucsb.edu/~moore
Doug Moore
March, 2003
ii
Contents
The sections marked with asterisks are less central to the main line of discussion,
and may be treated brieﬂy or omitted if time runs short.
1 Power Series 1
1.1 What is a power series? . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Solving diﬀerential equations by means of power series . . . . . . 7
1.3 Singular points . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Bessel’s diﬀerential equation . . . . . . . . . . . . . . . . . . . . . 22
2 Symmetry and Orthogonality 29
2.1 Eigenvalues of symmetric matrices . . . . . . . . . . . . . . . . . 29
2.2 Conic sections and quadric surfaces . . . . . . . . . . . . . . . . . 36
2.3 Orthonormal bases . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.4 Mechanical systems . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.5 Mechanical systems with many degrees of freedom* . . . . . . . . 53
2.6 A linear array of weights and springs* . . . . . . . . . . . . . . . 59
3 Fourier Series 62
3.1 Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.2 Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.3 Fourier sine and cosine series . . . . . . . . . . . . . . . . . . . . 72
3.4 Complex version of Fourier series* . . . . . . . . . . . . . . . . . 77
3.5 Fourier transforms* . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 Partial Diﬀerential Equations 81
4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.2 The initial value problem for the heat equation . . . . . . . . . . 85
4.3 Numerical solutions to the heat equation . . . . . . . . . . . . . . 92
4.4 The vibrating string . . . . . . . . . . . . . . . . . . . . . . . . . 94
4.5 The initial value problem for the vibrating string . . . . . . . . . 98
4.6 Heat ﬂow in a circular wire . . . . . . . . . . . . . . . . . . . . . 103
4.7 Sturm-Liouville Theory* . . . . . . . . . . . . . . . . . . . . . . . 107
4.8 Numerical solutions to the eigenvalue problem* . . . . . . . . . . 111
iii
5 PDE’s in Higher Dimensions 115
5.1 The three most important linear partial diﬀerential equations . . 115
5.2 The Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . . 118
5.3 Initial value problems for heat equations . . . . . . . . . . . . . . 123
5.4 Two derivations of the wave equation . . . . . . . . . . . . . . . . 129
5.5 Initial value problems for wave equations . . . . . . . . . . . . . . 134
5.6 The Laplace operator in polar coordinates . . . . . . . . . . . . . 136
5.7 Eigenvalues of the Laplace operator . . . . . . . . . . . . . . . . 141
5.8 Eigenvalues of the disk . . . . . . . . . . . . . . . . . . . . . . . . 144
5.9 Fourier analysis for the circular vibrating membrane* . . . . . . 150
A Using Mathematica to solve diﬀerential equations 156
Index 163
iv
Chapter 1
Power Series
1.1 What is a power series?
Functions are often represented eﬃciently by means of inﬁnite series. Examples
we have seen in calculus include the exponential function
e
x
= 1 +x +
1
2!
x
2
+
1
3!
x
3
+ =

¸
n=0
1
n!
x
n
, (1.1)
as well as the trigonometric functions
cos x = 1 −
1
2!
x
2
+
1
4!
x
4
− =

¸
k=0
(−1)
k
1
(2k)!
x
2k
and
sin x = x −
1
3!
x
3
+
1
5!
x
5
− =

¸
k=0
(−1)
k
1
(2k + 1)!
x
2k+1
.
An inﬁnite series of this type is called a power series. To be precise, a power
series centered at x
0
is an inﬁnite sum of the form
a
0
+a
1
(x −x
0
) +a
2
(x −x
0
)
2
+ =

¸
n=0
a
n
(x −x
0
)
n
,
where the a
n
’s are constants. In advanced treatments of calculus, these power
series representations are often used to deﬁne the exponential and trigonometric
functions.
Power series can also be used to construct tables of values for these functions.
For example, using a calculator or PC with suitable software installed (such as
Mathematica), we could calculate
1 + 1 +
1
2!
1
2
=
2
¸
n=0
1
n!
1
n
= 2.5,
1
1 + 1 +
1
2!
1
2
+
1
3!
1
3
+
1
4!
1
4
=
4
¸
n=0
1
n!
1
n
= 2.70833,
8
¸
n=0
1
n!
1
n
= 2.71806,
12
¸
n=0
1
n!
1
n
= 2.71828, . . .
As the number of terms increases, the sum approaches the familiar value of the
exponential function e
x
at x = 1.
For a power series to be useful, the inﬁnite sum must actually add up to a
ﬁnite number, as in this example, for at least some values of the variable x. We
let s
N
denote the sum of the ﬁrst (N + 1) terms in the power series,
s
N
= a
0
+a
1
(x −x
0
) +a
2
(x −x
0
)
2
+ +a
N
(x −x
0
)
N
=
N
¸
n=0
a
n
(x −x
0
)
n
,
and say that the power series

¸
n=0
a
n
(x −x
0
)
n
converges if the ﬁnite sum s
N
gets closer and closer to some (ﬁnite) number as
N → ∞.
Let us consider, for example, one of the most important power series of
applied mathematics, the geometric series
1 +x +x
2
+x
3
+ =

¸
n=0
x
n
.
In this case we have
s
N
= 1 +x +x
2
+x
3
+ +x
N
, xs
N
= x +x
2
+x
3
+x
4
+x
N+1
,
s
N
−xs
N
= 1 −x
N+1
, s
N
=
1 −x
N+1
1 −x
.
If [x[ < 1, then x
N+1
gets smaller and smaller as N approaches inﬁnity, and
hence
lim
N→∞
x
N+1
= 0.
Substituting into the expression for s
N
, we ﬁnd that
lim
N→∞
s
N
=
1
1 −x
.
Thus if [x[ < 1, we say that the geometric series converges, and write

¸
n=0
x
n
=
1
1 −x
.
2
On the other hand, if [x[ > 1, then x
N+1
gets larger and larger as N ap-
proaches inﬁnity, so lim
N→∞
x
N+1
does not exist as a ﬁnite number, and neither
does lim
N→∞
s
N
. In this case, we say that the geometric series diverges. In
summary, the geometric series

¸
n=0
x
n
converges to
1
1 −x
when [x[ < 1,
and diverges when [x[ > 1.
This behaviour, convergence for [x[ < some number, and divergences for
[x[ > that number, is typical of power series:
Theorem. For any power series
a
0
+a
1
(x −x
0
) +a
2
(x −x
0
)
2
+ =

¸
n=0
a
n
(x −x
0
)
n
,
there exists R, which is a nonnegative real number or ∞, such that
1. the power series converges when [x −x
0
[ < R,
2. and the power series diverges when [x −x
0
[ > R.
We call R the radius of convergence. A proof of this theorem is given in more
advanced courses on real or complex analysis.
1
We have seen that the geometric series
1 +x +x
2
+x
3
+ =

¸
n=0
x
n
has radius of convergence R = 1. More generally, if b is a positive constant, the
power series
1 +
x
b
+

x
b

2
+

x
b

3
+ =

¸
n=0

x
b

n
(1.2)
has radius of convergence b. To see this, we make the substitution y = x/b,
and the power series becomes
¸

n=0
y
n
, which we already know converges for
[y[ < 1 and diverges for [y[ > 1. But
[y[ < 1 ⇔

x
b

< 1 ⇔ [x[ < b,
[y[ > 1 ⇔

x
b

> 1 ⇔ [x[ > b.
1
Good references for the theory behind convergence of power series are Edward D.
Gaughan, Introduction to analysis, Brooks/Cole Publishing Company, Paciﬁc Grove, 1998
and Walter Rudin, Principles of mathematical analysis, third edition, McGraw-Hill, New
York, 1976.
3
Thus for [x[ < b the power series (1.2) converges to
1
1 −y
=
1
1 −(x/b)
=
b
b −x
,
while for [x[ > b, it diverges.
There is a simple criterion that often enables one to determine the radius of
convergence of a power series.
Ratio Test. The radius of convergence of the power series
a
0
+a
1
(x −x
0
) +a
2
(x −x
0
)
2
+ =

¸
n=0
a
n
(x −x
0
)
n
is given by the formula
R = lim
n→∞
[a
n
[
[a
n+1
[
,
so long as this limit exists.
Let us check that the ratio test gives the right answer for the radius of conver-
gence of the power series (1.2). In this case, we have
a
n
=
1
b
n
, so
[a
n
[
[a
n+1
[
=
1/b
n
1/b
n+1
=
b
n+1
b
n
= b,
and the formula from the ratio test tells us that the radius of convergence is
R = b, in agreement with our earlier determination.
In the case of the power series for e
x
,

¸
n=0
1
n!
x
n
,
in which a
n
= 1/n!, we have
[a
n
[
[a
n+1
[
=
1/n!
1/(n + 1)!
=
(n + 1)!
n!
= n + 1,
and hence
R = lim
n→∞
[a
n
[
[a
n+1
[
= lim
n→∞
(n + 1) = ∞,
so the radius of convergence is inﬁnity. In this case the power series converges
for all x. In fact, we could use the power series expansion for e
x
to calculate e
x
for any choice of x.
On the other hand, in the case of the power series

¸
n=0
n!x
n
,
4
in which a
n
= n!, we have
[a
n
[
[a
n+1
[
=
n!
(n + 1)!
=
1
n + 1
, R = lim
n→∞
[a
n
[
[a
n+1
[
= lim
n→∞

1
n + 1

= 0.
In this case, the radius of convergence is zero, and the power series does not
converge for any nonzero x.
The ratio test doesn’t always work because the limit may not exist, but
sometimes one can use it in conjunction with the
Comparison Test. Suppose that the power series

¸
n=0
a
n
(x −x
0
)
n
,

¸
n=0
b
n
(x −x
0
)
n
1
and R
2
respectively. If [a
n
[ ≤ [b
n
[ for all n, then
R
1
≥ R
2
. If [a
n
[ ≥ [b
n
[ for all n, then R
1
≤ R
2
.
In short, power series with smaller coeﬃcients have larger radius of convergence.
Consider for example the power series expansion for cos x,
1 + 0x −
1
2!
x
2
+ 0x
3
+
1
4!
x
4
− =

¸
k=0
(−1)
k
1
(2k)!
x
2k
.
In this case the coeﬃcient a
n
is zero when n is odd, while a
n
= ±1/n! when
n is even. In either case, we have [a
n
[ ≤ 1/n!. Thus we can compare with the
power series
1 +x +
1
2!
x
2
+
1
3!
x
3
+
1
4!
x
4
+ =

¸
n=0
1
n!
x
n
which represents e
x
and has inﬁnite radius of convergence. It follows from the
comparison test that the radius of convergence of

¸
k=0
(−1)
k
1
(2k)!
x
2k
must be at least large as that of the power series for e
x
, and hence must also be
inﬁnite.
Power series with positive radius of convergence are so important that there
is a special term for describing functions which can be represented by such power
series. A function f(x) is said to be real analytic at x
0
if there is a power series

¸
n=0
a
n
(x −x
0
)
n
0
with positive radius of convergence R such that
f(x) =

¸
n=0
a
n
(x −x
0
)
n
, for [x −x
0
[ < R.
5
For example, the functions e
x
is real analytic at any x
0
. To see this, we
utilize the law of exponents to write e
x
= e
x0
e
x−x0
and apply (1.1) with x
replaced by x −x
0
:
e
x
= e
x0

¸
n=0
1
n!
(x −x
0
)
n
=

¸
n=0
a
n
(x −x
0
)
n
, where a
n
=
e
x0
n!
.
This is a power series expansion of e
x
0
vergence. Similarly, the monomial function f(x) = x
n
is real analytic at x
0
because
x
n
= (x −x
0
+x
0
)
n
=
n
¸
i=0
n!
i!(n −i)!
x
n−i
0
(x −x
0
)
i
by the binomial theorem, a power series about x
0
in which all but ﬁnitely many
of the coeﬃcients are zero.
In more advanced courses, one studies criteria under which functions are
real analytic. For our purposes, it is suﬃcient to be aware of the following facts:
The sum and product of real analytic functions is real analytic. It follows from
this that any polynomial
P(x) = a
0
+a
1
x +a
2
x
2
+ +a
n
x
n
is analytic at any x
0
. The quotient of two polynomials with no common factors,
P(x)/Q(x), is analytic at x
0
if and only if x
0
is not a zero of the denominator
Q(x). Thus for example, 1/(x −1) is analytic whenever x
0
= 1, but fails to be
analytic at x
0
= 1.
Exercises:
1.1.1. Use the ratio test to ﬁnd the radius of convergence of the following power
series:
a.

¸
n=0
(−1)
n
x
n
, b.

¸
n=0
1
n + 1
x
n
,
c.

¸
n=0
3
n + 1
(x −2)
n
, d.

¸
n=0
1
2
n
(x −π)
n
,
e.

¸
n=0
(7x −14)
n
, f.

¸
n=0
1
n!
(3x −6)
n
.
1.1.2. Use the comparison test to ﬁnd an estimate for the radius of convergence
of each of the following power series:
a.

¸
k=0
1
(2k)!
x
2k
, b.

¸
k=0
(−1)
k
x
2k
,
c.

¸
k=0
1
2k
(x −4)
2k
d.

¸
k=0
1
2
2k
(x −π)
2k
.
6
1.1.3. Use the comparison test and the ratio test to ﬁnd the radius of convergence
of the power series

¸
m=0
(−1)
m
1
(m!)
2

x
2

2m
.
1.1.4. Determine the values of x
0
at which the following functions fail to be real
analytic:
a. f(x) =
1
x −4
, b. g(x) =
x
x
2
−1
,
c. h(x) =
4
x
4
−3x
2
+ 2
, d. φ(x) =
1
x
3
−5x
2
+ 6x
1.2 Solving diﬀerential equations by means of
power series
Our main goal in this chapter is to study how to determine solutions to diﬀer-
ential equations by means of power series. As an example, we consider our old
friend, the equation of simple harmonic motion
d
2
y
dx
2
+y = 0, (1.3)
which we have already learned how to solve by other methods. Suppose for the
moment that we don’t know the general solution and want to ﬁnd it by means
of power series. We could start by assuming that
y = a
0
+a
1
x +a
2
x
2
+a
3
x
3
+ =

¸
n=0
a
n
x
n
. (1.4)
It can be shown that the standard technique for diﬀerentiating polynomials term
by term also works for power series, so we expect that
dy
dx
= a
1
+ 2a
2
x + 3a
3
x
2
+ =

¸
n=1
na
n
x
n−1
.
(Note that the last summation only goes from 1 to ∞, since the term with n = 0
drops out of the sum.) Diﬀerentiating again yields
d
2
y
dx
2
= 2a
2
+ 3 2a
3
x + 4 3a
4
x
2
+ =

¸
n=2
n(n −1)a
n
x
n−2
.
We can replace n by m+ 2 in the last summation so that
d
2
y
dx
2
=

¸
m+2=2
(m+ 2)(m+ 2 −1)a
m+2
x
m+2−2
=

¸
m=0
(m+ 2)(m+ 1)a
m+2
x
m
.
7
The index m is a “dummy variable” in the summation and can be replaced by
any other letter. Thus we are free to replace m by n and obtain the formula
d
2
y
dx
2
=

¸
n=0
(n + 2)(n + 1)a
n+2
x
n
.
Substitution into equation(1.3) yields

¸
n=0
(n + 2)(n + 1)a
n+2
x
n
+

¸
n=0
a
n
x
n
= 0,
or

¸
n=0
[(n + 2)(n + 1)a
n+2
+a
n
]x
n
= 0.
Recall that a polynomial is zero only if all its coeﬃcients are zero. Similarly, a
power series can be zero only if all of its coeﬃcients are zero. It follows that
(n + 2)(n + 1)a
n+2
+a
n
= 0,
or
a
n+2
= −
a
n
(n + 2)(n + 1)
. (1.5)
This is called a recursion formula for the coeﬃcients a
n
.
The ﬁrst two coeﬃcients a
0
and a
1
in the power series can be determined
from the initial conditions,
y(0) = a
0
,
dy
dx
(0) = a
1
.
Then the recursion formula can be used to determine the remaining coeﬃcients
by the process of induction. Indeed it follows from (1.5) with n = 0 that
a
2
= −
a
0
2 1
= −
1
2
a
0
.
Similarly, it follows from (1.5) with n = 1 that
a
3
= −
a
1
3 2
= −
1
3!
a
1
,
and with n = 2 that
a
4
= −
a
2
4 3
=
1
4 3
1
2
a
0
=
1
4!
a
0
.
Continuing in this manner, we ﬁnd that
a
2n
=
(−1)
n
(2n)!
a
0
, a
2n+1
=
(−1)
n
(2n + 1)!
a
1
.
8
Substitution into (1.4) yields
y = a
0
+a
1
x −
1
2!
a
0
x
2

1
3!
a
1
x
3
+
1
4!
a
0
x
4
+
= a
0

1 −
1
2!
x
2
+
1
4!
x
4

+a
1

x −
1
3!
x
3
+
1
5!
x
5

.
We recognize that the expressions within parentheses are power series expan-
sions of the functions sin x and cos x, and hence we obtain the familiar expression
for the solution to the equation of simple harmonic motion,
y = a
0
cos x +a
1
sin x.
The method we have described—assuming a solution to the diﬀerential equa-
tion of the form
y(x) =

¸
n=0
a
n
x
n
and solve for the coeﬃcients a
n
—is surprisingly eﬀective, particularly for the
class of equations called second-order linear diﬀerential equations.
It is proven in books on diﬀerential equations that if P(x) and Q(x) are well-
behaved functions, then the solutions to the “homogeneous linear diﬀerential
equation”
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0
can be organized into a two-parameter family
y = a
0
y
0
(x) +a
1
y
1
(x),
called the general solution. Here y
0
(x) and y
1
(x) are any two nonzero solutions,
neither of which is a constant multiple of the other. In the terminology used in
linear algebra, we say that they are linearly independent solutions. As a
0
and
a
1
range over all constants, y ranges throughout a “linear space” of solutions.
We say that y
0
(x) and y
1
(x) form a basis for the space of solutions.
In the special case where the functions P(x) and Q(x) are real analytic,
the solutions y
0
(x) and y
1
(x) will also be real analytic. This is the content of
the following theorem, which is proven in more advanced books on diﬀerential
equations:
Theorem. If the functions P(x) and Q(x) can be represented by power series
P(x) =

¸
n=0
p
n
(x −x
0
)
n
, Q(x) =

¸
n=0
q
n
(x −x
0
)
n
with positive radii of convergence R
1
and R
2
respectively, then any solution
y(x) to the linear diﬀerential equation
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0
9
can be represented by a power series
y(x) =

¸
n=0
a
n
(x −x
0
)
n
,
whose radius of convergence is ≥ the smallest of R
1
and R
2
.
This theorem is used to justify the solution of many well-known diﬀerential
equations by means of the power series method.
Example. Hermite’s diﬀerential equation is
d
2
y
dx
2
−2x
dy
dx
+ 2py = 0, (1.6)
where p is a parameter. It turns out that this equation is very useful for treating
the simple harmonic oscillator in quantum mechanics, but for the moment, we
can regard it as merely an example of an equation to which the previous theorem
applies. Indeed,
P(x) = −2x, Q(x) = 2p,
both functions being polynomials, hence power series about x
0
= 0 with inﬁnite
As in the case of the equation of simple harmonic motion, we write
y =

¸
n=0
a
n
x
n
.
We diﬀerentiate term by term as before, and obtain
dy
dx
=

¸
n=1
na
n
x
n−1
,
d
2
y
dx
2
=

¸
n=2
n(n −1)a
n
x
n−2
.
Once again, we can replace n by m+ 2 in the last summation so that
d
2
y
dx
2
=

¸
m+2=2
(m+ 2)(m+ 2 −1)a
m+2
x
m+2−2
=

¸
m=0
(m+ 2)(m+ 1)a
m+2
x
m
,
and then replace m by n once again, so that
d
2
y
dx
2
=

¸
n=0
(n + 2)(n + 1)a
n+2
x
n
. (1.7)
Note that
−2x
dy
dx
=

¸
n=0
−2na
n
x
n
, (1.8)
10
while
2py =

¸
n=0
2pa
n
x
n
. (1.9)
Adding together (1.7), (1.8) and (1.9), we obtain
d
2
y
dx
2
−2x
dy
dx
+ 2py =

¸
n=0
(n + 2)(n + 1)a
n+2
x
n
+

¸
n=0
(−2n + 2p)a
n
x
n
.
If y satisﬁes Hermite’s equation, we must have
0 =

¸
n=0
[(n + 2)(n + 1)a
n+2
(−2n + 2p)a
n
]x
n
.
Since the right-hand side is zero for all choices of x, each coeﬃcient must be
zero, so
(n + 2)(n + 1)a
n+2
+ (−2n + 2p)a
n
= 0,
and we obtain the recursion formula for the coeﬃcients of the power series:
a
n+2
=
2n −2p
(n + 2)(n + 1)
a
n
. (1.10)
Just as in the case of the equation of simple harmonic motion, the ﬁrst two
coeﬃcients a
0
and a
1
in the power series can be determined from the initial
conditions,
y(0) = a
0
,
dy
dx
(0) = a
1
.
The recursion formula can be used to determine the remaining coeﬃcients in
the power series. Indeed it follows from (1.10) with n = 0 that
a
2
= −
2p
2 1
a
0
.
Similarly, it follows from (1.10) with n = 1 that
a
3
=
2 −2p
3 2
a
1
= −
2(p −1)
3!
a
1
,
and with n = 2 that
a
4
= −
4 −2p
4 3
a
2
=
2(2 −p)
4 3
−2p
2
a
0
=
2
2
p(p −2)
4!
a
0
.
Continuing in this manner, we ﬁnd that
a
5
=
6 −2p
5 4
a
3
=
2(3 −p)
5 4
2(1 −p)
3!
a
1
=
2
2
(p −1)(p −3)
5!
a
1
,
11
a
6
=
8 −2p
6 5 2
a
4
=
2(3 −p)
6 5
2
2
(p −2)p
4!
a
0
= −
2
3
p(p −2)(p −4)
6!
a
0
,
and so forth. Thus we ﬁnd that
y = a
0
¸
1 −
2p
2!
x
2
+
2
2
p(p −2)
4!
x
4

2
3
p(p −2)(p −4)
6!
x
6
+

+a
1
¸
x −
2(p −1)
3!
x
3
+
2
2
(p −1)(p −3)
5!
x
5

2
3
(p −1)(p −3)(p −5)
7!
x
7
+

.
We can now write the general solution to Hermite’s equation in the form
y = a
0
y
0
(x) +a
1
y
1
(x),
where
y
0
(x) = 1 −
2p
2!
x
2
+
2
2
p(p −2)
4!
x
4

2
3
p(p −2)(p −4)
6!
x
6
+
and
y
1
(x) = x−
2(p −1)
3!
x
3
+
2
2
(p −1)(p −3)
5!
x
5

2
3
(p −1)(p −3)(p −5)
7!
x
7
+ .
For a given choice of the parameter p, we could use the power series to construct
tables of values for the functions y
0
(x) and y
1
(x). Tables of values for these
functions are found in many ”handbooks of mathematical functions.” In the
language of linear algebra, we say that y
0
(x) and y
1
(x) form a basis for the
space of solutions to Hermite’s equation.
When p is a positive integer, one of the two power series will collapse, yielding
a polynomial solution to Hermite’s equation. These polynomial solutions are
known as Hermite polynomials.
Another Example. Legendre’s diﬀerential equation is
(1 −x
2
)
d
2
y
dx
2
−2x
dy
dx
+p(p + 1)y = 0, (1.11)
where p is a parameter. This equation is very useful for treating spherically
symmetric potentials in the theories of Newtonian gravitation and in electricity
and magnetism.
To apply our theorem, we need to divide by 1 −x
2
to obtain
d
2
y
dx
2

2x
1 −x
2
dy
dx
+
p(p + 1)
1 −x
2
y = 0.
Thus we have
P(x) = −
2x
1 −x
2
, Q(x) =
p(p + 1)
1 −x
2
.
12
Now from the preceding section, we know that the power series
1 +u +u
2
+u
3
+ converges to
1
1 −u
for [u[ < 1. If we substitute u = x
2
, we can conclude that
1
1 −x
2
= 1 +x
2
+x
4
+x
6
+ ,
the power series converging when [x[ < 1. It follows quickly that
P(x) = −
2x
1 −x
2
= −2x −2x
3
−2x
5

and
Q(x) =
p(p + 1)
1 −x
2
= p(p + 1) +p(p + 1)x
2
+p(p + 1)x
4
+ .
Both of these functions have power series expansions about x
0
= 0 which con-
verge for [x[ < 1. Hence our theorem implies that any solution to Legendre’s
equation will be expressible as a power series about x
0
= 0 which converges for
[x[ < 1. However, we might suspect that the solutions to Legendre’s equation
to exhibit some unpleasant behaviour near x = ±1. Experimentation with nu-
merical solutions to Legendre’s equation would show that these suspicions are
justiﬁed—solutions to Legendre’s equation will usually blow up as x → ±1.
Indeed, it can be shown that when p is an integer, Legendre’s diﬀerential
equation has a nonzero polynomial solution which is well-behaved for all x, but
solutions which are not constant multiples of these Legendre polynomials blow
up as x → ±1.
Exercises:
1.2.1. We would like to use the power series method to ﬁnd the general solution
to the diﬀerential equation
d
2
y
dx
2
−4x
dy
dx
+ 12y = 0,
which is very similar to Hermite’s equation. So we assume the solution is of the
form
y =

¸
n=0
a
n
x
n
,
a power series centered at 0, and determine the coeﬃcients a
n
.
a. As a ﬁrst step, ﬁnd the recursion formula for a
n+2
in terms of a
n
.
b. The coeﬃcients a
0
and a
1
will be determined by the initial conditions. Use
the recursion formula to determine a
n
in terms of a
0
and a
1
, for 2 ≤ n ≤ 9.
c. Find a nonzero polynomial solution to this diﬀerential equation.
13
d. Find a basis for the space of solutions to the equation.
e. Find the solution to the initial value problem
d
2
y
dx
2
−4x
dy
dx
+ 12y = 0, y(0) = 0,
dy
dx
(0) = 1.
f. To solve the diﬀerential equation
d
2
y
dx
2
−4(x −3)
dy
dx
+ 12y = 0,
it would be most natural to assume that the solution has the form
y =

¸
n=0
a
n
(x −3)
n
.
Use this idea to ﬁnd a polynomial solution to the diﬀerential equation
d
2
y
dx
2
−4(x −3)
dy
dx
+ 12y = 0.
1.2.2. We want to use the power series method to ﬁnd the general solution to
Legendre’s diﬀerential equation
(1 −x
2
)
d
2
y
dx
2
−2x
dy
dx
+p(p + 1)y = 0.
Once again our approach is to assume our solution is a power series centered at
0 and determine the coeﬃcients in this power series.
a. As a ﬁrst step, ﬁnd the recursion formula for a
n+2
in terms of a
n
.
b. Use the recursion formula to determine a
n
in terms of a
0
and a
1
, for 2 ≤ n ≤
9.
c. Find a nonzero polynomial solution to this diﬀerential equation, in the case
where p = 3.
d. Find a basis for the space of solutions to the diﬀerential equation
(1 −x
2
)
d
2
y
dx
2
−2x
dy
dx
+ 12y = 0.
1.2.3. The diﬀerential equation
(1 −x
2
)
d
2
y
dx
2
−x
dy
dx
+p
2
y = 0,
where p is a constant, is known as Chebyshev’s equation. It can be rewritten in
the form
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0, where P(x) = −
x
1 −x
2
, Q(x) =
p
2
1 −x
2
.
14
a. If P(x) and Q(x) are represented as power series about x
0
= 0, what is the
radius of convergence of these power series?
b. Assuming a power series centered at 0, ﬁnd the recursion formula for a
n+2
in terms of a
n
.
c. Use the recursion formula to determine a
n
in terms of a
0
and a
1
, for 2 ≤ n ≤
9.
d. In the special case where p = 3, ﬁnd a nonzero polynomial solution to this
diﬀerential equation.
e. Find a basis for the space of solutions to
(1 −x
2
)
d
2
y
dx
2
−x
dy
dx
+ 9y = 0.
1.2.4. The diﬀerential equation

d
2
dx
2
+x
2

z = λz (1.12)
arises when treating the quantum mechanics of simple harmonic motion.
a. Show that making the substitution z = e
−x
2
/2
y transforms this equation into
Hermite’s diﬀerential equation
d
2
y
dx
2
−2x
dy
dx
+ (λ −1)y = 0.
b. Show that if λ = 2n+1 where n is a nonnegative integer, (1.12) has a solution
of the form z = e
−x
2
/2
P
n
(x), where P
n
(x) is a polynomial.
1.3 Singular points
Our ultimate goal is to give a mathematical description of the vibrations of a
circular drum. For this, we will need to solve Bessel’s equation, a second-order
homogeneous linear diﬀerential equation with a “singular point” at 0.
A point x
0
is called an ordinary point for the diﬀerential equation
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0 (1.13)
if the coeﬃcients P(x) or Q(x) are both real analytic at x = x
0
, or equivalently,
both P(x) or Q(x) have power series expansions about x = x
0
with positive
radius of convergence. In the opposite case, we say that x
0
is a singular point;
thus x
0
is a singular point if at least one of the coeﬃcients P(x) or Q(x) fails
to be real analytic at x = x
0
. A singular point is said to be regular if
(x −x
0
)P(x) and (x −x
0
)
2
Q(x)
15
are real analytic.
For example, x
0
= 1 is a singular point for Legendre’s equation
d
2
y
dx
2

2x
1 −x
2
dy
dx
+
p(p + 1)
1 −x
2
y = 0,
because 1 −x
2
→ 0 as x → 1 and hence the quotients
2x
1 −x
2
and
p(p + 1)
1 −x
2
blow up as x → 1, but it is a regular singular point because
(x −1)P(x) = (x −1)
−2x
1 −x
2
=
2x
x + 1
and
(x −1)
2
Q(x) = (x −1)
2
p(p + 1)
1 −x
2
=
p(p + 1)(1 −x)
1 +x
are both real analytic at x
0
= 1.
The point of these deﬁnitions is that in the case where x = x
0
is a regular
singular point, a modiﬁcation of the power series method can still be used to
ﬁnd solutions.
Theorem of Frobenius. If x
0
is a regular singular point for the diﬀerential
equation
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0,
then this diﬀerential equation has at least one nonzero solution of the form
y(x) = (x −x
0
)
r

¸
n=0
a
n
(x −x
0
)
n
, (1.14)
where r is a constant, which may be complex. If (x−x
0
)P(x) and (x−x
0
)
2
Q(x)
have power series which converge for [x −x
0
[ < R then the power series

¸
n=0
a
n
(x −x
0
)
n
will also converge for [x −x
0
[ < R.
We will call a solution of the form (1.14) a generalized power series solution.
Unfortunately, the theorem guarantees only one generalized power series solu-
tion, not a basis. In fortuitous cases, one can ﬁnd a basis of generalized power
series solutions, but not always. The method of ﬁnding generalized power series
solutions to (1.13) in the case of regular singular points is called the Frobenius
method.
2
2
For more discussion of the Frobenius method as well as many of the other techniques
touched upon in this chapter we refer the reader to George F. Simmons, Diﬀerential equations
with applications and historical notes, second edition, McGraw-Hill, New York, 1991.
16
The simplest diﬀerential equation to which the Theorem of Frobenius applies
is the Cauchy-Euler equidimensional equation. This is the special case of (1.13)
for which
P(x) =
p
x
, Q(x) =
q
x
2
,
where p and q are constants. Note that
xP(x) = p and x
2
Q(x) = q
are real analytic, so x = 0 is a regular singular point for the Cauchy-Euler
equation as long as either p or q is nonzero.
The Frobenius method is quite simple in the case of Cauchy-Euler equations.
Indeed, in this case, we can simply take y(x) = x
r
, substitute into the equation
and solve for r. Often there will be two linearly independent solutions y
1
(x) =
x
r1
and y
2
(x) = x
r2
of this special form. In this case, the general solution is
given by the superposition principle as
y = c
1
x
r1
+c
2
x
r2
.
For example, to solve the diﬀerential equation
x
2
d
2
y
dx
2
+ 4x
dy
dx
+ 2y = 0,
we set y = x
r
and diﬀerentiate to show that
dy/dx = rx
r−1
⇒ x(dy/dx) = rx
r
,
d
2
y/dx
2
= r(r −1)x
r−2
⇒ x
2
(d
2
y/dx
2
) = r(r −1)x
r
.
Substitution into the diﬀerential equation yields
r(r −1)x
r
+ 4rx
r
+ 2x
r
= 0,
and dividing by x
r
yields
r(r −1) + 4r + 2 = 0 or r
2
+ 3r + 2 = 0.
The roots to this equation are r = −1 and r = −2, so the general solution to
the diﬀerential equation is
y = c
1
x
−1
+c
2
x
−2
=
c
1
x
+
c
2
x
2
.
Note that the solutions y
1
(x) = x
−1
and y
2
(x) = x
−2
can be rewritten in the
form
y
1
(x) = x
−1

¸
n=0
a
n
x
n
, y
2
(x) = x
−2

¸
n=0
b
n
x
n
,
where a
0
= b
0
= 1 and all the other a
n
’s and b
n
’s are zero, so both of these
solutions are generalized power series solutions.
17
On the other hand, if this method is applied to the diﬀerential equation
x
2
d
2
y
dx
2
+ 3x
dy
dx
+y = 0,
we obtain
r(r −1) + 3r + 1 = r
2
+ 2r + 1,
which has a repeated root. In this case, we obtain only a one-parameter family
of solutions
y = cx
−1
.
Fortunately, there is a trick that enables us to handle this situation, the so-called
method of variation of parameters. In this context, we replace the parameter c
by a variable v(x) and write
y = v(x)x
−1
.
Then
dy
dx
= v

(x)x
−1
−v(x)x
−2
,
d
2
y
dx
2
= v

(x)x
−1
−2v

(x)x
−2
+ 2v(x)x
−3
.
Substitution into the diﬀerential equation yields
x
2
(v

(x)x
−1
−2v

(x)x
−2
+2v(x)x
−3
) +3x(v

(x)x
−1
−v(x)x
−2
) +v(x)x
−1
= 0,
which quickly simpliﬁes to yield
xv

(x) +v

(x) = 0,
v

v

= −
1
x
, log [v

[ = −log [x[ +a, v

=
c
2
x
,
where a and c
2
are constants of integration. A further integration yields
v = c
2
log [x[ +c
1
, so y = (c
2
log [x[ +c
1
)x
−1
,
and we obtain the general solution
y = c
1
1
x
+c
2
log [x[
x
.
In this case, only one of the basis elements in the general solution is a generalized
power series.
For equations which are not of Cauchy-Euler form the Frobenius method is
more involved. Let us consider the example
2x
d
2
y
dx
2
+
dy
dx
+y = 0, (1.15)
which can be rewritten as
d
2
y
dx
2
+P(x)
dy
dx
+Q(x)y = 0, where P(x) =
1
2x
, Q(x) =
1
2x
.
18
One easily checks that x = 0 is a regular singular point. We begin the Frobenius
method by assuming that the solution has the form
y = x
r

¸
n=0
a
n
x
n
=

¸
n=0
a
n
x
n+r
.
Then
dy
dx
=

¸
n=0
(n +r)a
n
x
n+r−1
,
d
2
y
dx
2
=

¸
n=0
(n +r)(n +r −1)a
n
x
n+r−2
and
2x
d
2
y
dx
2
=

¸
n=0
2(n +r)(n +r −1)a
n
x
n+r−1
.
Substitution into the diﬀerential equation yields

¸
n=0
2(n +r)(n +r −1)a
n
x
n+r−1
+

¸
n=0
(n +r)a
n
x
n+r−1
+

¸
n=0
a
n
x
n+r
= 0,
which simpliﬁes to
x
r
¸

¸
n=0
(2n + 2r −1)(n +r)a
n
x
n−1
+

¸
n=0
a
n
x
n
¸
= 0.
We can divide by x
r
, and separate out the ﬁrst term from the ﬁrst summation,
obtaining
(2r −1)ra
0
x
−1
+

¸
n=1
(2n + 2r −1)(n +r)a
n
x
n−1
+

¸
n=0
a
n
x
n
= 0.
If we let n = m+ 1 in the ﬁrst inﬁnite sum, this becomes
(2r −1)ra
0
x
−1
+

¸
m=0
(2m+ 2r + 1)(m+r + 1)a
m+1
x
m
+

¸
n=0
a
n
x
n
= 0.
Finally, we replace m by n, obtaining
(2r −1)ra
0
x
−1
+

¸
n=0
(2n + 2r + 1)(n +r + 1)a
n+1
x
n
+

¸
n=0
a
n
x
n
= 0.
The coeﬃcient of each power of x must be zero. In particular, we must have
(2r −1)ra
0
= 0, (2n + 2r + 1)(n +r + 1)a
n+1
+a
n
= 0. (1.16)
If a
0
= 0, then all the coeﬃcients must be zero from the second of these equa-
tions, and we don’t get a nonzero solution. So we must have a
0
= 0 and hence
(2r −1)r = 0.
19
This is called the indicial equation. In this case, it has two roots
r
1
= 0, r
2
=
1
2
.
The second half of (1.16) yields the recursion formula
a
n+1
= −
1
(2n + 2r + 1)(n +r + 1)
a
n
, for n ≥ 0.
We can try to ﬁnd a generalized power series solution for either root of the
indicial equation. If r = 0, the recursion formula becomes
a
n+1
= −
1
(2n + 1)(n + 1)
a
n
.
Given a
0
= 1, we ﬁnd that
a
1
= −1, a
2
= −
1
3 2
a
1
=
1
3 2
,
a
3
= −
1
5 3
a
2
= −
1
(5 3)(3 2)
, a
4
= −
1
7 4
a
3
=
1
(7 5 3)4!
,
and so forth. In general, we would have
a
n
= (−1)
n
1
(2n −1)(2n −3) 1 n!
.
One of the generalized power series solution to (1.15) is
y
1
(x) = x
0
¸
1 −x +
1
3 2
x
2

1
(5 3)(3!)
x
3
+
1
(7 5 3)4!
x
4

= 1 −x +
1
3 2
x
2

1
(5 3)(3!)
x
3
+
1
(7 5 3)4!
x
4
− .
If r = 1/2, the recursion formula becomes
a
n+1
= −
1
(2n + 2)(n + (1/2) + 1)
a
n
= −
1
(n + 1)(2n + 3)
a
n
.
Given a
0
= 1, we ﬁnd that
a
1
= −
1
3
, a
2
= −
1
2 5
a
1
=
1
2 5 3
,
a
3
= −
1
3 7
a
2
= −
1
3! (7 5 3)
,
and in general,
a
n
= (−1)
n
1
n!(2n + 1)(2n −1) 1 n!
.
20
We thus obtain a second generalized power series solution to (1.15):
y
2
(x) = x
1/2
¸
1 −
1
3
x +
1
2 5 3
x
2

1
3! (7 5 3)
x
3
+

.
The general solution to (1.15) is a superposition of y
1
(x) and y
2
(x):
y = c
1
¸
1 −x +
1
3 2
x
2

1
(5 3)(3!)
x
3
+
1
(7 5 3)4!
x
4

+c
2

x
¸
1 −
1
3
x +
1
2 5 3
x
2

1
3! (7 5 3)
x
3
+

.
We obtained two linearly independent generalized power series solutions in
this case, but this does not always happen. If the roots of the indicial equation
diﬀer by an integer, we may obtain only one generalized power series solution.
In that case, a second independent solution can then be found by variation of
parameters, just as we saw in the case of the Cauchy-Euler equidimensional
equation.
Exercises:
1.3.1. For each of the following diﬀerential equations, determine whether x = 0
is ordinary or singular. If it is singular, determine whether it is regular or not.
a. y

+xy

+ (1 −x
2
)y = 0.
b. y

+ (1/x)y

+ (1 −(1/x
2
))y = 0.
c. x
2
y

+ 2xy

+ (cos x)y = 0.
d. x
3
y

+ 2xy

+ (cos x)y = 0.
1.3.2. Find the general solution to each of the following Cauchy-Euler equations:
a. x
2
d
2
y/dx
2
−2xdy/dx + 2y = 0.
b. x
2
d
2
y/dx
2
−xdy/dx +y = 0.
c. x
2
d
2
y/dx
2
−xdy/dx + 10y = 0.
(Hint: Use the formula
x
a+bi
= x
a
x
bi
= x
a
(e
log x
)
bi
= x
a
e
ib log x
= x
a
[cos(b log x) +i sin(b log x)]
1.3.3. We want to ﬁnd generalized power series solutions to the diﬀerential
equation
3x
d
2
y
dx
2
+
dy
dx
+y = 0
21
by the method of Frobenius. Our procedure is to ﬁnd solutions of the form
y = x
r

¸
n=0
a
n
x
n
=

¸
n=0
a
n
x
n+r
,
where r and the a
n
’s are constants.
a. Determine the indicial equation and the recursion formula.
b. Find two linearly independent generalized power series solutions.
1.3.4. To ﬁnd generalized power series solutions to the diﬀerential equation
2x
d
2
y
dx
2
+
dy
dx
+xy = 0
by the method of Frobenius, we assume the solution has the form
y =

¸
n=0
a
n
x
n+r
,
where r and the a
n
’s are constants.
a. Determine the indicial equation and the recursion formula.
b. Find two linearly independent generalized power series solutions.
1.4 Bessel’s diﬀerential equation
Our next goal is to apply the Frobenius method to Bessel’s equation,
x
d
dx

x
dy
dx

+ (x
2
−p
2
)y = 0, (1.17)
an equation which is needed to analyze the vibrations of a circular drum, as we
mentioned before. Here p is a parameter, which will be a nonnegative integer
in the vibrating drum problem. Using the Leibniz rule for diﬀerentiating a
product, we can rewrite Bessel’s equation in the form
x
2
d
2
y
dx
2
+x
dy
dx
+ (x
2
−p
2
)y = 0
or equivalently as
d
2
y
dx
2
+P(x)
dy
dx
+Q(x) = 0,
where
P(x) =
1
x
and Q(x) =
x
2
−p
2
x
2
.
Since
xP(x) = 1 and x
2
Q(x) = x
2
−p
2
,
22
we see that x = 0 is a regular singular point, so the Frobenius theorem implies
that there exists a nonzero generalized power series solution to (1.17).
To ﬁnd such a solution, we start as in the previous section by assuming that
y =

¸
n=0
a
n
x
n+r
.
Then
dy
dx
=

¸
n=0
(n +r)a
n
x
n+r−1
, x
dy
dx
=

¸
n=0
(n +r)a
n
x
n+r
,
d
dx

x
dy
dx

=

¸
n=0
(n +r)
2
a
n
x
n+r−1
,
and thus
x
d
dx

x
dy
dx

=

¸
n=0
(n +r)
2
a
n
x
n+r
. (1.18)
On the other hand,
x
2
y =

¸
n=0
a
n
x
n+r+2
=

¸
m=2
a
m−2
x
m+r
,
where we have set m = n + 2. Replacing m by n then yields
x
2
y =

¸
n=2
a
n−2
x
n+r
. (1.19)
Finally, we have,
−p
2
y = −

¸
n=0
p
2
a
n
x
n+r
. (1.20)
Adding up (1.18), (1.19), and (1.20), we ﬁnd that if y is a solution to (1.17),

¸
n=0
(n +r)
2
a
n
x
n+r
+

¸
n=2
a
n−2
x
n+r

¸
n=0
p
2
a
n
x
n+r
= 0.
This simpliﬁes to yield

¸
n=0
[(n +r)
2
−p
2
]a
n
x
n+r
+

¸
n=2
a
n−2
x
n+r
= 0,
or after division by x
r
,

¸
n=0
[(n +r)
2
−p
2
]a
n
x
n
+

¸
n=2
a
n−2
x
n
= 0.
23
Thus we ﬁnd that
(r
2
−p
2
)a
0
+ [(r + 1)
2
−p
2
]a
1
x +

¸
n=2
¦[(n +r)
2
−p
2
]a
n
+a
n−2
¦x
n
= 0.
The coeﬃcient of each power of x must be zero, so
(r
2
−p
2
)a
0
= 0, [(r+1)
2
−p
2
]a
1
= 0, [(n+r)
2
−p
2
]a
n
+a
n−2
= 0 for n ≥ 2.
Since we want a
0
to be nonzero, r must satisfy the indicial equation
(r
2
−p
2
) = 0,
which implies that r = ±p. Let us assume without loss of generality that p ≥ 0
and take r = p. Then
[(p + 1)
2
−p
2
]a
1
= 0 ⇒ (2p + 1)a
1
= 0 ⇒ a
1
= 0.
Finally,
[(n +p)
2
−p
2
]a
n
+a
n−2
= 0 ⇒ [n
2
+ 2np]a
n
+a
n−2
= 0,
which yields the recursion formula
a
n
= −
1
2np +n
2
a
n−2
. (1.21)
The recursion formula implies that a
n
= 0 if n is odd.
In the special case where p is a nonnegative integer, we will get a genuine
power series solution to Bessel’s equation (1.17). Let us focus now on this
important case. If we set
a
0
=
1
2
p
p!
,
we obtain
a
2
=
−a
0
4p + 4
= −
1
4(p + 1)
1
2
p
p!
= (−1)

1
2

p+2
1
1!(p + 1)!
,
a
4
=
−a
2
8p + 16
=
1
8(p + 2)

1
2

p+2
1
1!(p + 1)!
=
=
1
2(p + 2)

1
2

p+4
1
1!(p + 1)!
= (−1)
2

1
2

p+4
1
2!(p + 2)!
,
and so forth. The general term is
a
2m
= (−1)
m

1
2

p+2m
1
m!(p +m)!
.
24
2 4 6 8 10 12 14
-0.4
-0.2
0.2
0.4
0.6
0.8
1
Figure 1.1: Graph of the Bessel function J
0
(x).
Thus we ﬁnally obtain the power series solution
y =

x
2

p

¸
m=0
(−1)
m
1
m!(p +m)!

x
2

2m
.
The function deﬁned by the power series on the right-hand side is called the
p-th Bessel function of the ﬁrst kind, and is denoted by the symbol J
p
(x). For
example,
J
0
(x) =

¸
m=0
(−1)
m
1
(m!)
2

x
2

2m
.
Using the comparison and ratio tests, we can show that the power series ex-
pansion for J
p
(x) has inﬁnite radius of convergence. Thus when p is an integer,
Bessel’s equation has a nonzero solution which is real analytic at x = 0.
Bessel functions are so important that Mathematica includes them in its
library of built-in functions.
3
Mathematica represents the Bessel functions of
the ﬁrst kind symbolically by BesselJ[n,x]. Thus to plot the Bessel function
J
n
(x) on the interval [0, 15] one simply types in
n=0; Plot[ BesselJ[n,x], ¦x,0,15¦]
and a plot similar to that of Figure 1.1 will be produced. Similarly, we can
plot J
n
(x), for n = 1, 2, 3 . . . . Note that the graph of J
0
(x) suggests that it has
inﬁnitely many positive zeros.
On the open interval 0 < x < ∞, Bessel’s equation has a two-dimensional
space of solutions. However, it turns out that when p is a nonnegative integer, a
second solution, linearly independent from the Bessel function of the ﬁrst kind,
3
For a very brief introduction to Mathematica, the reader can refer to Appendix A.
25
2 4 6 8 10 12 14
-0.2
0.2
0.4
0.6
Figure 1.2: Graph of the Bessel function J
1
(x).
cannot be obtained directly by the generalized power series method that we have
presented. To obtain a basis for the space of solutions, we can, however, apply
the method of variation of parameters just as we did in the previous section for
the Cauchy-Euler equation; namely, we can set
y = v(x)J
p
(x),
substitute into Bessel’s equation and solve for v(x). If we were to carry this out
in detail, we would obtain a second solution linearly independent from J
p
(x).
Appropriately normalized, his solution is often denoted by Y
p
(x) and called the
p-th Bessel function of the second kind. Unlike the Bessel function of the ﬁrst
kind, this solution is not well-behaved near x = 0.
To see why, suppose that y
1
(x) and y
2
(x) is a basis for the solutions on the
interval 0 < x < ∞, and let W(y
1
, y
2
) be their Wronskian, deﬁned by
W(y
1
, y
2
)(x) =

y
1
(x) y

1
(x)
y
2
(x) y

2
(x)

.
This Wronskian must satisfy the ﬁrst order equation
d
dx
(xW(y
1
, y
2
)(x)) = 0,
as one veriﬁes by a direct calculation:
x
d
dx
(xy
1
y

2
−xy
2
y

1
) = y
1
x
d
dx
(xy

2
) −y
2
x
d
dx
(xy

1
)
= −(x
2
−n
2
)(y
1
y
2
−y
2
y
1
) = 0.
26
Thus
xW(y
1
, y
2
)(x) = c, or W(y
1
, y
2
)(x) =
c
x
,
where c is a nonzero constant, an expression which is unbounded as x → 0.
It follows that two linearly independent solutions y
1
(x) and y
2
(x) to Bessel’s
equation cannot both be well-behaved at x = 0.
Let us summarize what we know about the space of solutions to Bessel’s
equation in the case where p is an integer:
• There is a one-dimensional space of real analytic solutions to (1.17), which
are well-behaved as x → 0.
• This one-dimensional space is generated by a function J
p
(x) which is given
by the explicit power series formula
J
p
(x) =

x
2

p

¸
m=0
(−1)
m
1
m!(p +m)!

x
2

2m
.
Exercises:
1.4.1. Using the explicit power series formulae for J
0
(x) and J
1
(x) show that
d
dx
J
0
(x) = −J
1
(x) and
d
dx
(xJ
1
(x)) = xJ
0
(x).
1.4.2. The diﬀerential equation
x
2
d
2
y
dx
2
+x
dy
dx
−(x
2
+p
2
)y = 0
is sometimes called a modiﬁed Bessel equation. Find a generalized power series
solution to this equation in the case where p is an integer. (Hint: The power
series you obtain should be very similar to the power series for J
p
(x).)
1.4.3. Show that the functions
y
1
(x) =
1

x
cos x and y
2
(x) =
1

x
sin x
are solutions to Bessel’s equation
x
d
dx

x
dy
dx

+ (x
2
−p
2
)y = 0,
in the case where p = 1/2. Hence the general solution to Bessel’s equation in
this case is
y = c
1
1

x
cos x +c
2
1

x
sin x.
27
1.4.4. To obtain a nice expression for the generalized power series solution to
Bessel’s equation in the case where p is not an integer, it is convenient to use
the gamma function deﬁned by
Γ(x) =

0
t
x−1
e
−t
dt.
a. Use integration by parts to show that Γ(x + 1) = xΓ(x).
b. Show that Γ(1) = 1.
c. Show that
Γ(n + 1) = n! = n(n −1) 2 1,
when n is a positive integer.
d. Set
a
0
=
1
2
p
Γ(p + 1)
,
and use the recursion formula (1.21) to obtain the following generalized power
series solution to Bessel’s equation (1.17) for general choice of p:
y = J
p
(x) =

x
2

p

¸
m=0
(−1)
m
1
m!Γ(p +m+ 1)

x
2

2m
.
28
Chapter 2
Symmetry and
Orthogonality
2.1 Eigenvalues of symmetric matrices
Before proceeding further, we need to review and extend some notions from
vectors and matrices (linear algebra), which the student should have studied
in an earlier course. In particular, we will need the amazing fact that the
eigenvalue-eigenvector problem for an n n matrix A simpliﬁes considerably
when the matrix is symmetric.
An nn matrix A is said to be symmetric if it is equal to its transpose A
T
.
Examples of symmetric matrices include

1 3
3 1

,

¸
3 −λ 6 5
6 1 −λ 0
5 0 8 −λ
¸

and

¸
a b c
b d e
c e f
¸

.
Alternatively, we could say that an n n matrix A is symmetric if and only if
x (Ay) = (Ax) y. (2.1)
for every choice of n-vectors x and y. Indeed, since x y = x
T
y, equation (2.1)
can be rewritten in the form
x
T
Ay = (Ax)
T
y = x
T
A
T
y,
which holds for all x and y if and only if A = A
T
.
On the other hand, an n n real matrix B is orthogonal if its transpose is
equal to its inverse, B
T
= B
−1
. Alternatively, an n n matrix
B = (b
1
b
2
b
n
)
29
is orthogonal if its column vectors b
1
, b
2
, . . . , b
n
satisfy the relations
b
1
b
1
= 1, b
1
b
2
= 0, , b
1
b
n
= 0,
b
2
b
2
= 1, , b
2
b
n
= 0,

b
n
b
n
= 1.
Using this latter criterion, we can easily check that, for example, the matrices

cos θ −sin θ
sin θ cos θ

, and

¸
1/3 2/3 2/3
−2/3 −1/3 2/3
2/3 −2/3 1/3
¸

are orthogonal. Note that since
B
T
B = I ⇒ (det B)
2
= (det B
T
)(det B) = det(B
T
B) = 1,
the determinant of an orthogonal matrix is always ±1.
Recall that the eigenvalues of an n n matrix A are the roots of the poly-
nomial equation
det(A−λI) = 0.
For each eigenvalue λ
i
, the corresponding eigenvectors are the nonzero solutions
x to the linear system
(A−λI)x = 0.
For a general n n matrix with real entries, the problem of ﬁnding eigenvalues
and eigenvectors can be complicated, because eigenvalues need not be real (but
can occur in complex conjugate pairs) and in the “repeated root” case, there
may not be enough eigenvectors to construct a basis for R
n
. We will see that
these complications do not occur for symmetric matrices.
Spectral Theorem.
1
Suppose that A is a symmetric n n matrix with real
entries. Then its eigenvalues are real and eigenvectors corresponding to distinct
eigenvectors are orthogonal. Moreover, there is an n n orthogonal matrix B
of determinant one such that B
−1
AB = B
T
AB is diagonal.
Sketch of proof: The reader may want to skip our sketch of the proof at ﬁrst,
returning after studying some of the examples presented later in this section.
We will assume the following two facts, which are proven rigorously in advanced
courses on mathematical analysis:
1. Any continuous function on a sphere (of arbitrary dimension) assumes its
maximum and minimum values.
2. The points at which the maximum and minimum values are assumed can
be found by the method of Lagrange multipliers (a method usually dis-
cussed in vector calculus courses).
1
This is called the “spectral theorem” because the spectrum is another name for the set of
eigenvalues of a matrix.
30
The equation of the sphere S
n−1
in R
n
is
x
2
1
+x
2
2
+ +x
2
n
= 1, or x
T
x = 1, where x =

¸
¸
¸
x
1
x
2

x
n
¸

.
We let
g(x) = g(x
1
, x
2
, . . . , x
n
) = x
T
x −1,
so that the equation of the sphere is given by the constraint equation g(x) = 0.
Our approach consists of ﬁnding of ﬁnding the point on S
n−1
at which the
function
f(x) = f(x
1
, x
2
, . . . , x
n
) = x
T
Ax
assumes its maximum values.
To ﬁnd this maximum using Lagrange multipliers, we look for “critical
points” for the function
H(x, λ) = H(x
1
, x
2
, . . . , x
n
, λ) = f(x) −λg(x).
These are points at which
∇f(x
1
, x
2
, . . . , x
n
) = λ∇g(x
1
, x
2
, . . . , x
n
), and g(x
1
, x
2
, . . . , x
n
) = 0.
In other words, these are the points on the sphere at which the gradient of f is
a multiple of the gradient of g, or the points on the sphere at which the gradient
of f is perpendicular to the sphere.
If we set
A =

¸
¸
¸
a
11
a
12
a
1n
a
21
a
22
a
2n

a
n1
a
n2
a
nn
¸

,
a short calculation shows that at the point where f assumes its maximum,
∂H
∂x
i
= 2a
i1
x
1
+ 2a
i2
x
2
+ + 2a
in
x
n
−2λx
i
= 0,
or equivalently,
Ax −λx = 0.
We also obtain the condition
∂H
∂λ
= −g(x) = 0,
which is just our constraint. Thus the point on the sphere at which f assumes
its maximum is a unit-length eigenvector b
1
, the eigenvalue being the value λ
1
of the variable λ.
Let W be the “linear subspace” of R
n
deﬁned by the homogeneous linear
equation b
1
x = 0. The intersection S
n−1
∩ W is a sphere of dimension n −2.
31
We next use the method of Lagrange multipliers to ﬁnd a point on S
n−1
∩ W
at which f assumes its maximum. To do this, we let
g
1
(x) = x
T
x −1, g
2
(x) = b
1
x.
The maximum value of f on S
n−1
∩ W will be assumed at a critical point for
the function
H(x, λ, µ) = f(x) −λg
1
(x) −µg
2
(x).
This time, diﬀerentiation shows that
∂H
∂x
i
= 2a
i1
x
1
+ 2a
i2
x
2
+ + 2a
in
x
n
−2λx
i
−µb
i
= 0,
or equivalently,
Ax −λx −µb
1
= 0. (2.2)
It follows from the constraint equation b
1
x = 0 that
b
1
(Ax) = b
T
1
(Ax) = (b
T
1
A)x = (b
T
1
A
T
)x
= (Ab
1
)
T
x = (λ
1
b
1
)
T
x = λ
1
b
1
x = 0.
Hence it follows from (2.2) that Ax−λx = 0. Thus if b
2
is a point on S
n−1
∩W
at which f assumes its maximum, b
2
must be a unit-length eigenvector for A
which is perpendicular to b
1
.
Continuing in this way we ﬁnally obtain n mutually orthogonal unit-length
eigenvectors b
1
, b
2
, . . . , b
n
. These eigenvectors satisfy the equations
Ab
1
= λ
1
b
1
, Ab
2
= λ
2
b
2
, . . . Ab
n
= λ
2
b
n
,
which can be put together in a single matrix equation

Ab
1
Ab
2
Ab
n

=

λ
1
b
1
λ
2
b
2
λ
n
b
n

,
or equivalently,
A

b
1
b
2
b
n

=

b
1
b
2
b
n

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

.
If we set
B =

b
1
b
2
b
n

,
this last equation becomes
AB = B

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

.
32
Of course, B is an orthogonal matrix, so it is invertible and we can solve for A,
obtaining
A = B

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

B
−1
, or B
−1
AB =

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

.
We can arrange that the determinant of B be one by changing the sign of one
of the columns if necessary.
A more complete proof of the theorem is presented in more advanced courses
in linear algebra.
2
In any case, the method for ﬁnding the orthogonal matrix B
such that B
T
AB is diagonal is relatively simple, at least when the eigenvalues
are distinct. Simply let B be the matrix whose columns are unit-length eigen-
vectors for A. In the case of repeated roots, we must be careful to choose a
basis of unit-length eigenvectors for each eigenspace which are perpendicular to
each other.
Example. The matrix
A =

¸
5 4 0
4 5 0
0 0 1
¸

is symmetric, so its eigenvalues must be real. Its characteristic equation is
0 =

5 −λ 4 0
4 5 −λ 0
0 0 1 −λ

= [(λ −5)
2
−16](λ −1)
= (λ
2
−10λ + 9)(λ −1) = (λ −1)
2
(λ −9),
which has the roots λ
1
= 1 with multiplicity two and λ
2
= 9 with multiplicity
one.
Thus we are in the notorious “repeated root” case, which might be expected
to cause problems if A were not symmetric. However, since A is symmetric,
the Spectral Theorem guarantees that we can ﬁnd a basis for R
3
consisting of
eigenvectors for A even when the roots are repeated.
We ﬁrst consider the eigenspace W
1
corresponding to the eigenvalue λ
1
= 1,
which consists of the solutions to the linear system

¸
5 −1 4 0
4 5 −1 0
0 0 1 −1
¸

¸
b
1
b
2
b
3
¸

= 0,
or
4b
1
+4b
2
= 0,
4b
1
+4b
2
= 0,
0 = 0.
2
There are many excellent linear algebra texts that prove this theorem in detail; one good
reference is Bill Jacob, Linear algebra, W. H. Freeman, New York, 1990; see Chapter 5.
33
The coeﬃcient matrix of this linear system is

¸
4 4 0
4 4 0
0 0 0
¸

.
Applying the elementary row operations to this matrix yields

¸
4 4 0
4 4 0
0 0 0
¸

¸
1 1 0
4 4 0
0 0 0
¸

¸
1 1 0
0 0 0
0 0 0
¸

.
Thus the linear system is equivalent to
b
1
+b
2
= 0,
0 = 0,
0 = 0.
Thus W
1
is a plane with equation b
1
+b
2
= 0.
We need to extract two unit length eigenvectors from W
1
which are perpen-
dicular to each other. Note that since the equation for W
1
is b
1
+ b
2
= 0, the
unit length vector
n =

¸
1

2
1

2
0
¸

is perpendicular to W
1
. Since
b
1
=

¸
0
0
1
¸

∈ W
1
, we ﬁnd that b
2
= n b
1
=

¸
1

2
−1

2
0
¸

∈ W
1
.
The vectors b
1
and b
2
are unit length elements of W
1
which are perpendicular
to each other.
Next, we consider the eigenspace W
9
corresponding to the eigenvalue λ
2
= 9,
which consists of the solutions to the linear system

¸
5 −9 4 0
4 5 −9 0
0 0 1 −9
¸

¸
b
1
b
2
b
3
¸

= 0,
or
−4b
1
+4b
2
= 0,
4b
1
−4b
2
= 0,
−8b
3
= 0.
The coeﬃcient matrix of this linear system is

¸
−4 4 0
4 −4 0
0 0 −8
¸

.
34
Applying the elementary row operations to this matrix yields

¸
−4 4 0
4 −4 0
0 0 −8
¸

¸
1 −1 0
4 −4 0
0 0 −8
¸

¸
1 −1 0
0 0 0
0 0 −8
¸

¸
1 −1 0
0 0 0
0 0 1
¸

¸
1 −1 0
0 0 1
0 0 0
¸

.
Thus the linear system is equivalent to
b
1
−b
2
= 0,
b
3
= 0,
0 = 0,
and we see that
W
9
= span

¸
1
1
0
¸

.
We set
b
3
=

¸
1

2
1

2
0
¸

.
Theory guarantees that the matrix
B =

¸
0
1

2
1

2
0
−1

2
1

2
1 0 0
¸

,
whose columns are the eigenvectors b
1
, b
1
, and b
3
, will satisfy
B
−1
AB =

¸
1 0 0
0 1 0
0 0 9
¸

.
Moreover, since the eigenvectors we have chosen are of unit length and perpen-
dicular to each other, the matrix B will be orthogonal.
Exercises:
2.1.1. Find a 22-matrix B such that B is orthogonal and B
−1
AB is diagonal,
where
A =

5 4
4 5

.
2.1.2. Find a 22-matrix B such that B is orthogonal and B
−1
AB is diagonal,
where
A =

3 4
4 9

.
35
2.1.3. Find a 33-matrix B such that B is orthogonal and B
−1
AB is diagonal,
where
A =

¸
5 4 0
4 5 0
0 0 1
¸

.
2.1.4. Find a 33-matrix B such that B is orthogonal and B
−1
AB is diagonal,
where
A =

¸
−1 2 0
2 0 2
0 2 1
¸

.
2.1.5. Show that the nn identity matrix I is orthogonal. Show that if B
1
and
B
2
are nn orthogonal matrices, so is the product matrix B
1
B
2
, as well as the
inverse B
−1
1
.
2.2 Conic sections and quadric surfaces
The theory presented in the previous section can be used to “rotate coordinates”
so that conic sections or quadric surfaces can be put into “canonical form.”
A conic section is a curve in R
2
which is described by a quadratic equation,
such as the equation
ax
2
1
+ 2bx
1
x
2
+cx
2
2
= 1, (2.3)
where a, b and c are constants, which can be written in matrix form as

x
1
x
2

a b
b c

x
1
x
2

= 1.
If we let
A =

a b
b c

and x =

x
1
x
2

,
we can rewrite (2.3) as
x
T
Ax = 1, (2.4)
where A is a symmetric matrix.
According to the Spectral Theorem from Section 2.1, there exists a 2 2
orthogonal matrix B of determinant one such that
B
−1
AB = B
T
AB =

λ
1
0
0 λ
2

.
If we make a linear change of variables,
x = By,
36
then since x
T
= y
T
B
T
, equation (2.3) is transformed into
y
T
(B
T
AB)y = 1,

y
1
y
2

λ
1
0
0 λ
2

y
1
y
2

= 1,
or equivalently,
λ
1
y
2
1

2
y
2
2
= 1. (2.5)
In the new coordinate system (y
1
, y
2
), it is easy to recognize the conic section:
• If λ
1
and λ
2
are both positive, the conic is an ellipse.
• If λ
1
and λ
2
have opposite signs, the conic is an hyperbola.
• If λ
1
and λ
2
are both negative, the conic degenerates to the the empty set,
because the equation has no real solutions.
In the case where λ
1
and λ
2
are both positive, we can rewrite (2.5) as
y
2
1
(

1/λ
1
)
2
+
y
2
2
(

1/λ
2
)
2
= 1,
from which we recognize that the semi-major and semi-minor axes of the ellipse
are

1/λ
1
and

1/λ
2
.
The matrix B which relates x and y represents a rotation of the plane. To
see this, note that the ﬁrst column b
1
of B is a unit-length vector, and can
therefore be written in the form
b
1
=

cos θ
sin θ

,
for some choice of θ. The second column b
2
is a unit-vector perpendicular to
b
1
and hence
b
2
= ±

−sin θ
cos θ

.
We must take the plus sign in this last expression, because det B = 1. Thus
B

1
0

=

cos θ
sin θ

, B

0
1

=

−sin θ
cos θ

,
or equivalently, B takes the standard basis vectors for R
2
to vectors which have
been rotated counterclockwise through an angle θ. By linearity,
B =

cos θ −sin θ
sin θ cos θ

must rotate every element of R
2
counterclockwise through an angle θ. Thus once
we have sketched the conic in (y
1
, y
2
)-coordinates, we can obtain the sketch in
(x
1
, x
2
)-coordinates by simply rotating counterclockwise through the angle θ.
37
Example. Let’s consider the conic
5x
2
1
−6x
1
x
2
+ 5x
2
2
= 1, (2.6)
or equivalently,

x
1
x
2

5 −3
−3 5

x
1
x
2

= 1.
The characteristic equation of the matrix
A =

5 −3
−3 5

is
(5 −λ)
2
−9 = 0, λ
2
−10λ + 16 = 0, (λ −2)(λ −8) = 0,
and the eigenvalues are λ
1
= 2 and λ
2
= 8. Unit-length eigenvectors corre-
sponding to these eigenvalues are
b
1
=

1/

2
1/

2

, b
2
=

−1/

2
1/

2

.
The proof of the theorem of Section 2.1 shows that these vectors are orthogonal
to each other, and hence the matrix
B =

1/

2 −1/

2
1/

2 1/

2

is an orthogonal matrix such that
B
T
AB =

2 0
0 8

.
Note that B represents a counterclockwise rotation through 45 degrees. If
we deﬁne new coordinates (y
1
, y
2
) by

x
1
x
2

= B

y
1
y
2

,
equation (2.6) will simplify to

y
1
y
2

2 0
0 8

y
1
y
2

= 1,
or
2y
2
1
+ 8y
2
2
=
y
2
1
(1/

2)
2
+
y
2
2
(1/

8)
2
= 1.
We recognize that this is the equation of an ellipse. The lengths of the semimajor
and semiminor axes are 1/

2 and 1/(2

2).
38
-0.4 -0.2 0.2 0.4
-0.4
-0.2
0.2
0.4
Figure 2.1: Sketch of the conic section 5x
2
1
−6x
1
x
2
+ 5x
2
2
−1 = 0.
The same techniques can be used to sketch quadric surfaces in R
3
, surfaces
deﬁned by an equation of the form

x
1
x
2
x
3

¸
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
¸

¸
x
1
x
2
x
3
¸

= 1,
where the a
ij
’s are constants. If we let
A =

¸
a
11
a
12
a
13
a
21
a
22
a
23
a
31
a
32
a
33
¸

, x =

¸
x
1
x
2
x
3
¸

,
we can write this in matrix form
x
T
Ax = 1. (2.7)
According to the Spectral Theorem, there is a 3 3 orthogonal matrix B of
determinant one such that B
T
AB is diagonal. We introduce new coordinates
y =

¸
y
1
y
2
y
3
¸

by setting x = By,
and equation (2.7) becomes
y
T
(B
T
AB)y = 1.
Thus after a suitable linear change of coordinates, the equation (2.7) can be put
in the form

y
1
y
2
y
3

¸
λ
1
0 0
0 λ
2
0
0 0 λ
3
¸

¸
y
1
y
2
y
3
¸

= 1,
39
-2
0
2
-4
-2
0
2
4
-1
-0.5
0
0.5
1
-2
0
2
Figure 2.2: An ellipsoid.
or
λ
1
y
2
1

2
y
2
2

3
y
2
3
= 1,
where λ
1
, λ
2
, and λ
3
are the eigenvalues of A. It is relatively easy to sketch the
quadric surface in the coordinates (y
1
, y
2
, y
3
).
If the eigenvalues are all nonzero, we ﬁnd that there are four cases:
• If λ
1
, λ
2
, and λ
3
are all positive, the quadric surface is an ellipsoid.
• If two of the λ’s are positive and one is negative, the quadric surface is an
hyperboloid of one sheet.
• If two of the λ’s are negative and one is positive, the quadric surface is an
hyperboloid of two sheets.
• If λ
1
, λ
2
, and λ
3
are all negative, the equation represents the empty set.
Just as in the case of conic sections, the orthogonal matrix B of determinant
one which relates x and y represents a rotation. To see this, note ﬁrst that since
B is orthogonal,
(Bx) (By) = x
T
B
T
By = x
T
Iy = x y. (2.8)
In other words, multiplication by B preserves dot products. It follows from this
that the only real eigenvalues of B can be ±1. Indeed, if x is an eigenvector for
B corresponding to the eigenvalue λ, then
λ
2
(x x) = (λx) (λx) = (Bx) (Bx) = x x,
40
so division by x x yields λ
2
= 1.
Since det B = 1 and det B is the product of the eigenvalues, if all of the
eigenvalues are real, λ = 1 must occur as an eigenvalue. On the other hand,
non-real eigenvalues must occur in complex conjugate pairs, so if there is a non-
real eigenvalue µ + iν, then there must be another non-real eigenvalue µ − iν
together with a real eigenvalue λ. In this case,
1 = det B = λ(µ +iν)(µ −iν) = λ(µ
2

2
).
Since λ = ±1, we conclude that λ = 1 must occur as an eigenvalue also in this
case.
In either case, λ = 1 is an eigenvalue and
W
1
= ¦x ∈ R
3
: Bx = x¦
is nonzero. It is easily veriﬁed that if dimW
1
is larger than one, then B must
be the identity. Thus if B = I, there is a one-dimensional subspace W
1
of R
3
which is left ﬁxed under multiplication by B. This is the axis of rotation.
Let W

1
denote the orthogonal complement to W
1
. If x is a nonzero element
of W
1
and y ∈ W

1
, it follows from (2.8) that
(By) x = (By) (Bx) = y
T
B
T
Bx = y
T
Ix = y x = 0,
so By ∈ W

1
. Let y, z be elements of W

1
such that
y y = 1, y z = 0, z z = 1;
we could say that ¦y, z¦ form an orthonormal basis for W

1
. By (2.8),
(By) (By) = 1, (By) (Bz) = 0, (Bz) (Bz) = 1.
Thus By must be a unit-length vector in W

1
and there must exist a real number
θ such that
By = cos θy + sin θz.
Moreover, Bz must be a unit-length vector in W

1
which is perpendicular to
By and hence
Bz = ±(−sin θy + cos θz).
However, we cannot have
Bz = −(−sin θy + cos θz).
Indeed, this would imply (via a short calculation) that the vector
u = cos(θ/2)y + sin(θ/2)z
is ﬁxed by B, in other words Bu = u, contradicting the fact that u ∈ W

1
.
Thus we must have
Bz = −sin θy + cos θz,
41
-1
0
1
-1
0
1
-1
-0.5
0
0.5
1
-1
0
1
-1
0
1
Figure 2.3: Hyperboloid of one sheet.
and multiplication by B must be a rotation in the plane W

1
through an angle
θ. Moreover, it is easily checked that y + iz and y − iz are eigenvectors for B
with eigenvalues
e
±iθ
= cos θ ±i sin θ.
We can therefore conclude that a 3 3 orthogonal matrix B of determinant
one represents a rotation about an axis (which is the eigenspace for eigenvalue
one) and through an angle θ (which can be determined from the eigenvalues of
B, which must be 1 and e
±iθ
).
Exercises:
2.2.1. Suppose that
A =

2 3
3 2

.
a. Find an orthogonal matrix B such that B
T
AB is diagonal.
b. Sketch the conic section 2x
2
1
+ 6x
1
x
2
+ 2x
2
2
= 1.
c. Sketch the conic section 2(x
1
−1)
2
+ 6(x
1
−1)x
2
+ 2x
2
2
= 1.
2.2.2. Suppose that
A =

2 −2
−2 5

.
42
-2
-1
0
1
2
-2
-1
0
1
2
-2
0
2
-2
-1
0
1
2
Figure 2.4: Hyperboloid of two sheets.
a. Find an orthogonal matrix B such that B
T
AB is diagonal.
b. Sketch the conic section 2x
2
1
−4x
1
x
2
+ 5x
2
2
= 1.
c. Sketch the conic section 2x
2
1
−4x
1
x
2
+ 5x
2
2
−4x
1
+ 4x
2
= −1.
2.2.3. Suppose that
A =

4 2
2 1

.
a. Find an orthogonal matrix B such that B
T
AB is diagonal.
b. Sketch the conic section 4x
2
1
+ 4x
1
x
2
+x
2
2

5x
1
+ 2

5x
2
= 0.
2.2.4. Determine which of the following conic sections are ellipses, which are
hyperbolas, etc.:
a. x
2
1
+ 4x
1
x
2
+ 3x
2
2
= 1.
b. x
2
1
+ 6x
1
x
2
+ 10x
2
2
= 1.
c. −3x
2
1
+ 6x
1
x
2
−4x
2
2
= 1.
d. −x
2
1
+ 4x
1
x
2
−3x
2
2
= 1.
2.2.5. Find the semi-major and semi-minor axes of the ellipse
5x
2
1
+ 6x
1
x
2
+ 5x
2
2
= 4.
43
2.2.6. Suppose that
A =

¸
0 2 0
2 3 0
0 0 9
¸

.
a. Find an orthogonal matrix B such that B
T
AB is diagonal.
b. Sketch the quadric surface 4x
1
x
2
+ 3x
2
2
+ 9x
2
3
= 1.
2.2.7. Determine which of the following quadric surfaces are ellipsoids, which
are hyperboloids of one sheet, which are hyperboloids of two sheets, etc.:
a. x
2
1
−x
2
2
−x
2
3
−6x
2
x
3
= 1.
b. x
2
1
+x
2
2
−x
2
3
−6x
1
x
2
= 1.
c. x
2
1
+x
2
2
+x
2
3
+ 4x
1
x
2
+ 2x
3
= 0.
2.2.8. a. Show that the matrix
B =

¸
1/3 2/3 2/3
−2/3 −1/3 2/3
2/3 −2/3 1/3
¸

is an orthogonal matrix of deteminant one. Thus multiplication by B is rotation
about some axis through some angle.
b. Find a nonzero vector which spans the axis of rotation.
c. Determine the angle of rotation.
2.3 Orthonormal bases
Recall that if v is an element of R
3
, we can express it with respect to the
standard basis as
v = ai +bj +ck, where a = v i, b = v j, c = v k.
We would like to extend this formula to the case where the standard basis
¦i, j, k¦ for R
3
is replaced by a more general “orthonormal basis.”
Deﬁnition. A collection of n vectors b
1
, b
2
, . . . , b
n
in R
n
is an orthonormal
basis for R
n
if
b
1
b
1
= 1, b
1
b
2
= 0, , b
1
b
n
= 0,
b
2
b
2
= 1, , b
2
b
n
= 0,

b
n
b
n
= 1.
(2.9)
44
From the discussion in Section 2.1, we recognize that the term orthonormal
basis is just another name for a collection of n vectors which form the columns
of an orthogonal n n matrix.
It is relatively easy to express an arbitrary vector f ∈ R
n
in terms of an
orthonormal basis b
1
, b
2
, . . . , b
n
: to ﬁnd constants c
1
, c
2
, . . . , c
n
so that
f = c
1
b
1
+c
2
b
2
+ +c
n
b
n
, (2.10)
we simply dot both sides with the vector b
i
and use (2.9) to conclude that
c
i
= b
i
f .
In other words, if b
1
, b
2
, . . . , b
n
is an orthonormal basis for R
n
, then
f ∈ R
n
⇒ f = (f b
1
)b
1
+ + (f b
n
)b
n
, (2.11)
a generalization of the formula we gave for expressing a vector in R
3
in terms
of the standard basis ¦i, j, k¦.
This formula can be helpful in solving the initial value problem
dx
dt
= Ax, x(0) = f , (2.12)
in the case where A is a symmetric n n matrix and f is a constant vector.
Since A is symmetric, the Spectral Theorem of Section 2.1 guarantees that the
eigenvalues of A are real and that we can ﬁnd an n n orthogonal matrix B
such that
B
−1
AB =

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

.
If we set x = By, then
B
dy
dt
=
dx
dt
= Ax = ABy ⇒
dy
dt
= B
−1
ABy.
Thus in terms of the new variable
y =

¸
¸
¸
y
1
y
2

y
n
¸

, we have

¸
¸
¸
dy
1
/dt
dy
2
/dt

dy
n
/dt
¸

=

¸
¸
¸
λ
1
0 0
0 λ
2
0

0 0 λ
n
¸

¸
¸
¸
y
1
y
2

y
n
¸

,
so that the matrix diﬀerential equation decouples into n noninteracting scalar
diﬀerential equations
dy
1
/dt = λ
1
y
1
,
dy
2
/dt = λ
2
y
2
,

dy
n
/dt = λ
n
y
n
.
45
We can solve these equations individually, obtaining the general solution
y =

¸
¸
¸
y
1
y
2

y
n
¸

=

¸
¸
¸
c
1
e
λ1t
c
2
e
λ2t

c
n
e
λnt
¸

,
where c
1
, c
2
, . . . , c
n
are constants of integration. Transferring back to our
original variable x yields
x = B

¸
¸
¸
c
1
e
λ1t
c
2
e
λ2t

c
n
e
λnt
¸

= c
1
b
1
e
λ1t
+c
2
b
2
e
λ2t
+ +c
n
b
n
e
λnt
,
where b
1
, b
2
, . . . , b
n
are the columns of B. Note that
x(0) = c
1
b
1
+c
2
b
2
+ +c
n
b
n
.
To ﬁnish solving the initial value problem (2.12), we need to determine the
constants c
1
, c
2
, . . . , c
n
so that
c
1
b
1
+c
2
b
2
+ +c
n
b
n
= f .
It is here that our formula (2.11) comes in handy; using it we obtain
x = (f b
1
)b
1
e
λ1t
+ + (f b
n
)b
n
e
λnt
.
Example. Suppose we want to solve the initial value problem
dx
dt
=

¸
5 4 0
4 5 0
0 0 1
¸

x, x(0) = f , where f =

¸
3
1
4
¸

.
We saw in Section 2.1 that A has the eigenvalues λ
1
= 1 with multiplicity two
and λ
2
= 9 with multiplicity one. Moreover, the orthogonal matrix
B =

¸
0
1

2
1

2
0
−1

2
1

2
1 0 0
¸

has the property that
B
−1
AB =

¸
1 0 0
0 1 0
0 0 9
¸

.
Thus the general solution to the matrix diﬀerential equation dx/dt = Ax is
x = B

¸
c
1
e
t
c
2
e
t
c
3
e
9t
¸

= c
1
b
1
e
t
+c
2
b
2
e
t
+c
3
b
3
e
9t
,
46
where
b
1
=

¸
0
0
1
¸

, b
2
=

¸
1

2
−1

2
0
¸

, b
3
=

¸
1

2
1

2
0
¸

.
Setting t = 0 in our expression for x yields
x(0) = c
1
b
1
+c
2
b
2
+c
3
b
3
.
To solve the initial value problem, we employ (2.11) to see that
c
1
=

¸
0
0
1
¸

¸
3
1
4
¸

= 4, c
2
=

¸
1

2
−1

2
0
¸

¸
3
1
4
¸

=

2,
c
3
=

¸
1

2
1

2
0
¸

¸
3
1
4
¸

= 2

2.
Hence the solution is
x = 4

¸
0
0
1
¸

e
t
+

2

¸
1

2
−1

2
0
¸

e
t
+ 2

2

¸
1

2
1

2
0
¸

e
9t
.
Exercise:
2.3.1.a. Find the eigenvalues of the symmetric matrix
A =

¸
¸
¸
5 4 0 0
4 5 0 0
0 0 4 2
0 0 2 1
¸

.
b. Find an orthogonal matrix B such that B
−1
AB is diagonal.
c. Find an orthonormal basis for R
4
consisting of eigenvectors of A.
d. Find the general solution to the matrix diﬀerential equation
dx
dt
= Ax.
e. Find the solution to the initial value problem
dx
dt
= Ax, x(0) =

¸
¸
¸
1
3
0
2
¸

.
47
2.3.2.a. Find the eigenvalues of the symmetric matrix
A =

¸
−3 2 0
2 −4 2
0 2 −5
¸

.
(Hint: To ﬁnd roots of the cubic, try λ = 1 and λ = −1.)
b. Find an orthogonal matrix B such that B
−1
AB is diagonal.
c. Find an orthonormal basis for R
3
consisting of eigenvectors of A.
d. Find the general solution to the matrix diﬀerential equation
dx
dt
= Ax.
e. Find the solution to the initial value problem
dx
dt
= Ax, x(0) =

¸
1
2
0
¸

.
2.3.3. (For students with access to Mathematica) a. Find the eigenvalues of the
matrix
A =

¸
2 1 1
1 3 2
1 2 4
¸

by running the Mathematica program
a = ¦¦2,1,1¦,¦1,3,2¦,¦1,2,4¦¦; Eigenvalues[a]
The answer is so complicated because Mathematica uses exact but complicated
formulae for solving cubics discovered by Cardan and Tartaglia, two sixteenth
century Italian mathematicians.
b. Find numerical approximations to these eigenvalues by running the program
Eigenvalues[N[a]]
The numerical values of the eigenvalues are far easier to use.
c. Use Mathematica to ﬁnd numerical values for the eigenvectors for A by
running the Mathematica program
Eigenvectors[N[a]]
and write down the general solution to the matrix diﬀerential equation
dx
dt
= Ax.
48
1
Figure 2.5: Two carts connected by springs and moving along a friction-free
track.
2.4 Mechanical systems
Mechanical systems consisting of weights and springs connected together in an
array often lead to initial value problems of the type
d
2
x
dt
2
= Ax, x(0) = f ,
dx
dt
= 0, (2.13)
where A is a symmetric matrix and f is a constant vector. These can be solved
by a technique similar to that used in the previous section.
For example, let us consider the mechanical system illustrated in Figure 2.5.
Here we have two carts moving along a friction-free track, each containing mass
m and attached together by three springs, with spring constants k
1
, k
2
and k
3
.
Let
x
1
(t) = the position of the ﬁrst cart to the right of equilibrium,
x
2
(t) = the position of the second cart to the right of equilibrium,
F
1
= force acting on the ﬁrst cart,
F
2
= force acting on the second cart,
with positive values for F
1
or F
2
indicating that the forces are pulling to the
right, negative values that the forces are pulling to the left.
Suppose that when the carts are in equilibrium, the springs are also in equi-
librium and exert no forces on the carts. In this simple case, it is possible to
reason directly from Hooke’s law that the forces F
1
and F
2
must be given by
the formulae
F
1
= −k
1
x
1
+k
2
(x
2
−x
1
), F
2
= k
2
(x
1
−x
2
) −k
3
x
2
,
but it becomes diﬃcult to determine the forces for more complicated mechanical
systems consisting of many weights and springs. Fortunately, there are some
simple principles from physics which simply the procedure for ﬁnding the forces
acting in such mechanical systems.
49
The easiest calculation of the forces is based upon the notion of work. On
the one hand, the work required to pull a weight to a new position is equal to
the increase in potential energy imparted to the weight. On the other hand, we
have the equation
Work = Force Displacement,
which implies that
Force =
Work
Displacement
= −
Change in potential energy
Displacement
.
Thus if V (x
1
, x
2
) is the potential energy of the conﬁguration when the ﬁrst cart
is located at the point x
1
and the second cart is located at the point x
2
, then
the forces are given by the formulae
F
1
= −
∂V
∂x
1
, F
2
= −
∂V
∂x
2
.
In our case, the potential energy V is the sum of the potential energies stored
in each of the three springs,
V (x
1
, x
2
) =
1
2
k
1
x
2
1
+
1
2
k
2
(x
1
−x
2
)
2
+
1
2
k
3
x
2
2
,
and hence we obtain the formulae claimed before:
F
1
= −
∂V
∂x
1
= −k
1
x
1
+k
2
(x
2
−x
1
),
F
2
= −
∂V
∂x
2
= k
2
(x
1
−x
2
) −k
3
x
2
.
It now follows from Newton’s second law of motion that
Force = Mass Acceleration,
and hence
F
1
= m
d
2
x
1
dt
2
, F
2
= m
d
2
x
2
dt
2
.
Thus we obtain a second-order system of diﬀerential equations,
m
d
2
x
1
dt
2
= −k
1
x
1
+k
2
(x
2
−x
1
) = −(k
1
+k
2
)x
1
+k
2
x
2
,
m
d
2
x
2
dt
2
= k
2
(x
1
−x
2
) −k
3
x
2
= k
2
x
1
−(k
2
+k
3
)x
2
.
We can write this system in matrix form as
m
d
2
x
dt
2
= Ax, where A =

−(k
1
+k
2
) k
2
k
2
−(k
2
+k
3
)

. (2.14)
50
Note that A is indeed a symmetric matrix. The potential energy is given by the
expression
V (x
1
, x
2
) = −
1
2

x
1
x
2

A

x
1
x
2

.
Example. Let us consider the special case of the mass-spring system in which
m = k
1
= k
2
= k
3
= 1,
so that the system (2.14) becomes
d
2
x
dt
2
=

−2 1
1 −2

x. (2.15)
To ﬁnd the eigenvalues, we must solve the characteristic equation
det

−2 −λ 1
1 −2 −λ

= (λ + 2)
2
−1 = 0,
which yields
λ = −2 ±1.
The eigenvalues in this case are
λ
1
= −1, and λ
2
= −3.
The eigenspace corresponding to the eigenvalue −1 is
W
−1
= ¦b ∈ R
2
: (A+I)b = 0¦ = . . . = span

1
1

.
It follows from the argument in Section 2.1 that the eigenspace corresponding
to the other eigenvalue is just the orthogonal complement
W
−3
= span

−1
1

.
Unit length eigenvectors lying in the two eigenspaces are
b
1
=

1/

2
1/

2

, b
2
=

−1/

2
1/

2

.
The theorem of Section 2.1 guarantees that the matrix
B =

1/

2 −1/

2
1/

2 1/

2

,
whose columns are b
1
and b
2
, will diagonalize our system of diﬀerential equa-
tions.
51
Indeed, if we deﬁne new coordinates (y
1
, y
2
) by setting

x
1
x
2

=

1/

2 −1/

2
1/

2 1/

2

y
1
y
2

,
our system of diﬀerential equations transforms to
d
2
y
1
/dt
2
= −y
1
,
d
2
y
2
/dt
2
= −3y
2
.
We set ω
1
= 1 and ω
2
=

3, so that this system assumes the familiar form
d
2
y
1
/dt
2

2
1
y
1
= 0,
d
2
y
2
/dt
2

2
2
y
2
= 0,
a system of two noninteracting harmonic oscillators.
The general solution to the transformed system is
y
1
= a
1
cos ω
1
t +b
1
sin ω
1
t, y
2
= a
2
cos ω
2
t +b
2
sin ω
2
t.
In the original coordinates, the general solution to (2.15) is
x =

x
1
x
2

=

1/

2 −1/

2
1/

2 1/

2

a
1
cos ω
1
t +b
1
sin ω
1
t
a
2
cos ω
2
t +b
2
sin ω
2
t

,
or equivalently,
x = b
1
(a
1
cos ω
1
t +b
1
sin ω
1
t) +b
2
(a
2
cos ω
2
t +b
2
sin ω
2
t).
The motion of the carts can be described as a general superposition of two
modes of oscillation, of frequencies
ω
1

and
ω
2

.
Exercises:
2.4.1.a. Consider the mass-spring system with two carts illustrated in Figure 2.5
in the case where k
1
= 4 and m = k
2
= k
3
= 1. Write down a system of second-
order diﬀerential equations which describes the motion of this system.
b. Find the general solution to this system.
c. What are the frequencies of vibration of this mass-spring system?
2.4.2.a. Consider the mass-spring system with three carts illustrated in Fig-
ure 2.5 in the case where m = k
1
= k
2
= k
3
= k
4
= 1. Show that the motion of
this system is described by the matrix diﬀerential equation
d
2
x
dt
2
= Ax, where A =

¸
−2 1 0
1 −2 1
0 1 −2
¸

.
52
Figure 2.6: Three carts connected by springs and moving along a friction-free
track.
b. Find the eigenvalues of the symmetric matrix A.
c. Find an orthonormal basis for R
3
consisting of eigenvectors of A.
d. Find an orthogonal matrix B such that B
−1
AB is diagonal.
e. Find the general solution to the matrix diﬀerential equation
d
2
x
dt
2
= Ax.
f. Find the solution to the initial value problem
d
2
x
dt
2
= Ax, x(0) =

¸
1
2
0
¸

,
dx
dt
(0) = 0.
2.4.3.a. Find the eigenvalues of the symmetric matrix
A =

¸
¸
¸
−2 1 0 0
1 −2 1 0
0 1 −2 1
0 0 1 −2
¸

.
b. What are the frequencies of oscillation of a mechanical system which is
governed by the matrix diﬀerential equation
d
2
x
dt
2
= Ax?
2.5 Mechanical systems with many degrees of
freedom*
Using an approach similar to the one used in the preceding section, we can
consider more complicated systems consisting of many masses and springs. For
53
Figure 2.7: A circular array of carts and springs.
example, we could consider the box spring underlying the mattress in a bed.
Although such a box spring contains hundreds of individual springs, and hence
the matrix A in the corresponding dynamical system constains hundreds of rows
and columns, it is still possible to use symmetries in the box spring to simplify
the calculations, and make the problem of determining the “natural frequencies
of vibration” of the mattress into a manageable problem.
To illustrate how the symmetries of a problem can make it much easier to
solve, let us consider a somewhat simpler problem, a system of n carts containing
identical weights of mass m and connected by identical springs of spring constant
k, moving along a circular friction-free track as sketched in Figure 2.6.
We choose a positive direction along the track and let x
i
denote the dis-
placement of the i-th cart out of equilibrium position in the positive direction,
for 1 ≤ i ≤ n. The potential energy stored in the springs is
V (x
1
, . . . , x
n
) =
1
2
k(x
n
−x
1
)
2
+
1
2
k(x
2
−x
1
)
2
+
1
2
k(x
3
−x
2
)
2
+. . .+
1
2
k(x
n
−x
n−1
)
2
= −
1
2
k

x
1
x
2
x
3
x
n

¸
¸
¸
¸
¸
−2 1 0 1
1 −2 1 0
0 1 −2 0

1 0 0 −2
¸

¸
¸
¸
¸
¸
x
1
x
2
x
3

x
n
¸

.
We can write this as
V (x
1
, . . . , x
n
) = −
1
2
kx
T
Ax,
where
x =

¸
¸
¸
¸
¸
x
1
x
2
x
3

x
n
¸

, A =

¸
¸
¸
¸
¸
−2 1 0 1
1 −2 1 0
0 1 −2 0

1 0 0 −2
¸

,
or equivalently as
V (x
1
, . . . , x
n
) = −
1
2
k
n
¸
i=1
n
¸
j=1
a
ij
x
i
x
j
,
54
where a
ij
denotes the (i, j)-component of the matrix A.
Just as in the preceding section, the force acting on the i-th cart can be
calculated as minus the derivative of the potential energy with respect to the
position of the i-th cart, the other carts being held ﬁxed. Thus for example,
F
1
= −
∂V
∂x
1
=
1
2
k
n
¸
j=1
a
1j
x
j
+
1
2
k
n
¸
i=1
a
i1
x
i
= k
n
¸
j=1
a
1j
x
j
,
the last step obtained by using the symmetry of A. In general, we obtain the
result:
F
i
= −
∂V
∂x
i
= k
n
¸
j=1
a
ij
x
j
,
which could be rewritten in matrix form as
F = kAx. (2.16)
On the other hand, by Newton’s second law of motion,
m
d
2
x
dt
2
= F.
Substitution into (2.16) yields
m
d
2
x
dt
2
= kAx or
d
2
x
dt
2
=
k
m
Ax, (2.17)
where A is the symmetric matrix given above. To ﬁnd the frequencies of vi-
bration of our mechanical system, we need to ﬁnd the eigenvalues of the n n
matrix A.
To simplify the calculation of the eigenvalues of this matrix, we make use of
the fact that the carts are identically situated—if we relabel the carts, shifting
them to the right by one, the dynamical system remains unchanged. Indeed,
lew us deﬁne new coordinates (y
1
, . . . , y
n
) by setting
x
1
= y
2
, x
2
= y
3
, . . . , x
n−1
= y
n
, x
n
= y
1
,
or in matrix terms,
x = Ty, where T =

¸
¸
¸
¸
¸
¸
¸
0 1 0 0
0 0 1 0
0 0 0 0

0 0 0 1
1 0 0 0
¸

.
Then y satisﬁes exactly the same system of diﬀerential equations as x:
d
2
y
dt
2
=
k
m
Ay. (2.18)
55
On the other hand,
d
2
x
dt
2
=
k
m
Ax ⇒ T
d
2
y
dt
2
=
k
m
ATy or
d
2
y
dt
2
=
k
m
T
−1
ATy.
Comparison with (2.18) yields
A = T
−1
AT, or TA = AT.
In other words the matrices A and T commute.
Now it is quite easy to solve the eigenvector-eigenvalue problem for T. In-
deed, if x is an eigenvector for T corresponding to the eigenvalue λ, the compo-
nents of x must satisfy the vector equation Tx = λx. In terms of components,
the vector equation becomes
x
2
= λx
1
, x
3
= λx
2
, . . . , x
n
= λx
n−1
, x
1
= λx
n
. (2.19)
Thus x
3
= λ
2
x
1
, x
4
= λ
3
x
1
, and so forth, yielding ﬁnally the equation
x
1
= λ
n
x
1
.
Similarly,
x
2
= λ
n
x
2
, . . . , x
n
= λ
n
x
n
.
Since at least one of the x
i
’s is nonzero, we must have
λ
n
= 1. (2.20)
This equation is easily solved via Euler’s formula:
1 = e
2πi
⇒ (e
2πi/n
)
n
= 1, and similarly [(e
2πi/n
)
j
]
n
= 1,
for 0 ≤ j ≤ n −1. Thus the solutions to (2.20) are
λ = η
j
, for 0 ≤ j ≤ n −1, where η = e
2πi/n
. (2.21)
For each choice of j, we can try to ﬁnd eigenvectors corresponding to η
j
. If
we set x
1
= 1, we can conclude from (2.19) that
x
2
= η
j
, x
3
= η
2j
, . . . , x
n
= η
(n−1)j
,
thereby obtaining a nonzero solution to the eigenvector equation. Thus for each
j, 0 ≤ j ≤ n −1, we do indeed have an eigenvector
e
j
=

¸
¸
¸
¸
¸
1
η
j
η
2j

η
(n−1)j
¸

,
for the eigenvalue η
j
. Moreover, each eigenspace is one-dimensional.
56
In the dynamical system that we are considering, of course, we need to solve
the eigenvalue-eigenvector problem for A, not for T. Fortunately, however, since
A and T commute, the eigenvectors for T are also eigenvectors for A. Indeed,
since AT = TA,
T(Ae
j
) = A(Te
j
) = A(η
j
e
j
) = η
j
(Ae
j
),
and this equation states that Ae
j
is an eigenvector for T with the same eigen-
value as e
j
. Since the eigenspaces of T are all one-dimensional, Ae
j
must be a
multiple of e
j
; in other words,
Ae
j
= λ
j
e
j
, for some number λ
j
.
To ﬁnd the eigenvalues λ
j
for A, we simply act on e
j
by A: we ﬁnd that the
ﬁrst component of the vector Ae
j
= λ
j
e
j
is
−2 +η
j

(n−1)j
= −2 + (e
2πij/n
+e
−2πij/n
) = −2 + 2 cos(2πj/n),
where we have used Euler’s formula once again. On the other hand, the ﬁrst
component of e
j
is 1, so we immediately conclude that
λ
j
= −2 + 2 cos(2πj/n).
It follows from the familiar formula
cos(2α) = cos
2
α −sin
2
α
that
λ
j
= −2 + 2[cos(πj/n)]
2
−2[sin(πj/n)]
2
= −4[sin(πj/n)]
2
.
Note that λ
j
= λ
n−j
, and hence the eigenspaces for A are two-dimensional,
except for the eigenspace corresponding to j = 0, and if n is even, to the
eigenspace corresponding to j = n/2. Thus in the special case where n is odd,
all the eigenspaces are two-dimensional except for the eigenspace with eigenvalue
λ
0
= 0, which is one-dimensional.
If j = 0 and j = n/2, e
j
and e
n−j
form a basis for the eigenspace corre-
sponding to eigenvalue λ
j
. It is not diﬃcult to verify that
1
2
(e
j
+e
n−j
) =

¸
¸
¸
¸
¸
1
cos(πj/n)
cos(2πj/n)

cos((n −1)πj/n)
¸

and
i
2
(e
j
−e
n−j
) =

¸
¸
¸
¸
¸
0
sin(πj/n)
sin(2πj/n)

sin((n −1)πj/n)
¸

57
form a real basis for this eigenspace. Let
ω
j
=

−λ
j
= 2 sin(πj/n).
In the case where n is odd, we can write the general solution to our dynamical
system (2.17) as
x = e
0
(a
0
+b
0
t) +
(n−1)/2
¸
j=1
1
2
(e
j
+e
n−j
)(a
j
cos ω
j
t +b
j
sin ω
j
t)
+
(n−1)/2
¸
j=1
i
2
(e
j
−e
n−j
)(c
j
cos ω
j
t +d
j
sin ω
j
t).
The motion of the system can be described as a superposition of several modes
of oscillation, the frequencies of oscillation being
ω
j

=

k
m
sin(πj/n)
π
.
Note that the component
e
0
(a
0
+b
0
t) =

¸
¸
¸
¸
¸
1
1
1

1
¸

(a
0
+b
0
t)
corresponds to a constant speed motion of the carts around the track. If n is
large,
sin(π/n)
π
.
=
π/n
π
=
1
n

ω
1

.
=

k
m
1
n
,
so if we were to set k/m = n
2
, the lowest nonzero frequency of oscillation would
approach one as n → ∞.
Exercise:
2.5.1. Find the eigenvalues of the matrix
T =

¸
¸
¸
¸
¸
¸
¸
0 1 0 0
0 0 1 0
0 0 0 0

0 0 0 1
1 0 0 0
¸

58
Figure 2.8: A linear array of carts and springs.
by expanding the determinant

−λ 1 0 0
0 −λ 1 0
0 0 −λ 0

0 0 0 1
1 0 0 −λ

.
2.6 A linear array of weights and springs*
Suppose more generally that a system of n−1 carts containing identical weights
of mass m, and connected by identical springs of spring constant k, are moving
along a friction-free track, as shown in Figure 2.7. Just as in the preceding
section, we can show that the carts will move in accordance with the linear
system of diﬀerential equations
d
2
x
dt
2
= Ax =
k
m
Px,
where
P =

¸
¸
¸
¸
¸
−2 1 0 0
1 −2 1 0
0 1 −2

0 0 −2
¸

.
We take k = n and m = (1/n), so that k/m = n
2
.
We can use the following Mathematica program to ﬁnd the eigenvalues of
the n n matrix A, when n = 6:
n = 6;
m = Table[Max[2-Abs[i-j],0], ¦i,n-1¦ ,¦j,n-1¦ ];
p = m - 4 IdentityMatrix[n-1];
59
a = n∧2 p
eval = Eigenvalues[N[a]]
Since
2 −[i −j[

= 2 if j = i,
= 1 if j = i ±1
≤0 otherwise,
the ﬁrst line of the program generates the (n −1) (n −1)-matrix
M =

¸
¸
¸
¸
¸
2 1 0 0
1 2 1 0
0 1 2

0 0 2
¸

.
The next two lines generate the matrices P and A = n
2
P. Finally, the last
line ﬁnds the eigenvalues of A. If you run this program, you will note that all
the eigenvalues are negative and that Mathematica provides the eigenvalues in
increasing order.
We can also modify the program to ﬁnd the eigenvalues of the matrix when
n = 12, n = 26, or n = 60 by simply replacing 6 in the top line with the new
value of n. We can further modify the program by asking for only the smallest
eigenvalue lambda[[n]] and a plot of the corresponding eigenvector:
n = 14;
m = Table[Max[2-Abs[i-j],0], ¦i,n-1¦ ,¦j,n-1¦ ];
p = m - 4 IdentityMatrix[n-1]; a = n∧2 p;
eval = Eigenvalues[N[a]]; evec = Eigenvectors[N[a]];
Print[eval[[n-1]]];
ListPlot[evec[[n-1]]];
If we experiment with this program using diﬀerent values for n, we will see
that as n gets larger, the smallest eigenvalue seems to approach −π
2
and the
plot of the smallest eigenvector looks more and more like a sine curve. Thus the
fundamental frequency of the mechanical system seems to approach π/2π = 1/2
and the oscillation in the fundamental mode appears more and more sinusoidal
in shape.
When n is large, we can consider this array of springs and weights as a model
for a string situated along the x-axis and free to stretch to the right and the left
along the x-axis. The track on which the cart runs restricts the motion of the
weights to be only in the x-direction. A more realistic model would allow the
carts to move in all three directions. This would require new variables y
i
and
z
i
such that
y
i
(t) = the y-component of displacement of the i-th weight,
z
i
(t) = the z-component of displacement of the i-th weight.
60
20 40 60 80 100
-0.14
-0.12
-0.1
-0.08
-0.06
-0.04
-0.02
Figure 2.9: Shape of the fundamental mode (n=100).
If we were to set
x =

¸
¸
¸
x
1
x
2

x
n
¸

, y =

¸
¸
¸
x
1
x
2

x
n
¸

, z =

¸
¸
¸
z
1
z
2

z
n
¸

,
then an argument similar to the one given above would show that the vectors
x, y and z would satisfy the matrix diﬀerential equations
dx
dt
=
k
m
Ax,
dy
dt
=
k
m
Ay,
dz
dt
=
k
m
Az.
and each of these could be solved just as before.
Using a similar approach, we could consider even more complicated systems
consisting of many masses and springs connected in a two- or three-dimensional
array.
61
Chapter 3
Fourier Series
3.1 Fourier series
The theory of Fourier series and the Fourier transform is concerned with dividing
a function into a superposition of sines and cosines, its components of various
frequencies. It is a crucial tool for understanding waves, including water waves,
sound waves and light waves. Suppose, for example, that the function f(t)
represents the amplitude of a light wave arriving from a distant galaxy. The
light is a superposition of many frequencies which encode information regarding
the material which makes up the stars of the galaxy, the speed with which the
galaxy is receding from the earth, its speed of rotation, and so forth. Much of
our knowledge of the universe is derived from analyzing the spectra of stars and
galaxies. Just as a prism or a spectrograph is an experimental apparatus for
dividing light into its components of various frequencies, so Fourier analysis is
a mathematical technique which enables us to decompose an arbitrary function
into a superposition of oscillations.
In the following chapter, we will describe how the theory of Fourier series can
be used to analyze the ﬂow of heat in a bar and the motion of a vibrating string.
Indeed, Joseph Fourier’s original investigations which led to the theory of Fourier
series were motivated by an attempt to understand heat ﬂow.
1
notion of dividing a function into its components with respect to an appropriate
“orthonormal basis of functions” is one of the key ideas of applied mathematics,
useful not only as a tool for solving partial diﬀerential equations, as we will see
in the next two chapters, but for many other purposes as well. For example, a
black and white photograph could be represented by a function f(x, y) of two
variables, f(x, y) representing the darkness at the point (x, y). The photograph
can be stored eﬃciently by determining the components of f(x, y) with respect
to a well-chosen “wavelet basis.” This idea is the key to image compression,
which can be used to send pictures quickly over the internet.
2
1
Fourier’s research was published in his Th´eorie analytique de la chaleur in 1822.
2
See St´ephane Mallat, A wavelet tour of signal processing, Academic Press, Boston, 1998.
62
We turn now to the basics of Fourier analysis in its simplest context. A
function f : R → R is said to be periodic of period T if it satisﬁes the relation
f(t +T) = f(t), for all t ∈ R.
Thus f(t) = sin t is periodic of period 2π.
Given an arbitrary period T, it is easy to construct examples of functions
which are periodic of period T—indeed, the function f(t) = sin(
2πt
T
) is periodic
of period T because
sin(
2π(t +T)
T
) = sin(
2πt
T
+ 2π) = sin(
2πt
T
).
More generally, if k is any positive integer, the functions
cos(
2πkt
T
) and sin(
2πkt
T
)
are also periodic functions of period T.
The main theorem from the theory of Fourier series states that any “well-
behaved” periodic function of period T can be expressed as a superposition of
sines and cosines:
f(t) =
a
0
2
+a
1
cos(
2πt
T
) +a
2
cos(
4πt
T
) +. . . +b
1
sin(
2πt
T
) +b
2
sin(
4πt
T
) +. . . .
(3.1)
In this formula, the a
k
’s and b
k
’s are called the Fourier coeﬃcients of f, and
the inﬁnite series on the right-hand side is called the Fourier series of f.
Our ﬁrst goal is to determine how to calculate these Fourier coeﬃcients. For
simplicity, we will restrict our attention to the case where the period T = 2π,
so that
f(t) =
a
0
2
+a
1
cos t +a
2
cos 2t +. . . +b
1
sin t +b
2
sin 2t +. . . . (3.2)
The formulae for a general period T are only a little more complicated, and are
based upon exactly the same ideas.
The coeﬃcient a
0
is particularly easy to evaluate. We simply integrate both
sides of (3.2) from −π to π:

π
−π
f(t)dt =

π
−π
a
0
2
dt +

π
−π
a
1
cos tdt +

π
−π
a
2
cos 2tdt +. . .
+

π
−π
b
1
sin tdt +

π
−π
b
2
sin 2tdt +. . . .
Since the integral of cos kt or sin kt over the interval from −π to π vanishes, we
conclude that

π
−π
f(t)dt = πa
0
,
63
and we can solve for a
0
, obtaining
a
0
=
1
π

π
−π
f(t)dt. (3.3)
To ﬁnd the other Fourier coeﬃcients, we will need some integral formulae.
We claim that if m and n are positive integers,

π
−π
cos nt cos mtdt =

π, for m = n,
0, for m = n,
(3.4)

π
−π
sin nt sin mtdt =

π, for m = n,
0, for m = n,
(3.5)

π
−π
sin nt cos mtdt = 0. (3.6)
Let us verify the ﬁrst of these equations. We will use the trigonometric
identities
cos((n +m)t) = cos nt cos mt −sin nt sin mt,
cos((n −m)t) = cos nt cos mt + sin nt sin mt.
cos((n +m)t) + cos((n −m)t) = 2 cos nt cos mt,
or
cos nt cos mt =
1
2
(cos((n +m)t) + cos((n −m)t)).
Hence

π
−π
cos nt cos mtdt =
1
2

π
−π
(cos((n +m)t) + cos((n −m)t))dt,
and since

π
−π
cos(kt)dt =
1
k
sin(kt)

π
−π
= 0,
if k = 0, we conclude that

π
−π
cos nt cos mtdt =

π, for m = n,
0, for m = n.
The reader is asked to verify the other two integral formulae (3.5) and (3.6)
in the exercises at the end of this section.
64
To ﬁnd the formula for the Fourier coeﬃcients a
k
for k > 0, we multiply
both sides of (3.2) by sin kt and integrate from −π to π:

π
−π
f(t) cos ktdt =

π
−π
a
0
2
cos ktdt +

π
−π
a
1
cos t cos ktdt +. . .
+. . . +

π
−π
a
k
cos kt cos ktdt +. . .
+. . . +

π
−π
b
j
sin jt cos ktdt +. . . .
According to our formulae (3.4) and (3.6), there is only one term which survives:

π
−π
f(t) cos ktdt =

π
−π
a
k
cos kt cos ktdt = πa
k
.
We can easily solve for a
k
:
a
k
=
1
π

π
−π
f(t) cos ktdt. (3.7)
A very similar argument yields the formula
b
k
=
1
π

π
−π
f(t) sin ktdt. (3.8)
Example. Let us use formulae (3.3), (3.7), and (3.8) to ﬁnd the Fourier coeﬃ-
cients of the function
f(t) =

−π, for −π < t < 0,
π, for 0 < t < π,
0, for t = 0, π,
extended to be periodic of period 2π. Then
a
0
2
= average value of f = 0,
and
a
m
=
1
π
¸
π
−π
f(t) cos mtdt

=
1
π
¸
0
−π
−π cos mtdt

+
1
π
¸
π
0
π cos mtdt

=
1
π
−π
m
[sin mt]
0
−π
+
1
π
π
m
[sin mt]
π
0
= = 0,
while
b
m
=
1
π
¸
π
−π
f(t) sin mtdt

=
1
π
¸
0
−π
−π sin mtdt

+
1
π
¸
π
0
π sin mtdt

65
-4 -2 2 4
-3
-2
-1
1
2
3
Figure 3.1: A graph of the Fourier approximation φ
5
(t) = 4 sin t +(4/3) sin 3t +
(4/5) sin 5t.
=
1
π
π
m
[cos mt]
0
−π
+
1
π
−π
m
[cos mt]
π
0
=
2
m

2
m
cos mπ
=
2
m
(1 −(−1)
m
) =

4
m
, if m is odd,
0, if m is even.
Thus we ﬁnd the Fourier series for f:
f(t) = 4 sin t +
4
3
sin 3t +
4
5
sin 5t + .
The trigonometric polynomials
φ
1
(t) = 4 sin t, φ
3
(t) = 4 sin t +
4
3
sin 3t
φ
5
(t) = 4 sin t +
4
3
sin 3t +
4
5
sin 5t
are approximations to f(t) which improve as the number of terms increases.
By the way, this Fourier series yields a curious formula for π. If we set
x = π/2, f(x) = π, and we obtain
π = 4 sin(π/2) +
4
3
sin(3π/2) +
4
5
sin(5π/2) +
from which we conclude that
π = 4(1 −
1
3
+
1
5

1
7
+
1
9
− ).
66
-4 -2 2 4
-3
-2
-1
1
2
3
Figure 3.2: A graph of the Fourier approximation φ
13
(t). The overshooting near
the points of discontinuity is known as the “Gibbs phenomenon.”
The Fourier series on the right-hand side of (3.1) is often conveniently ex-
pressed in the Σ notation,
f(t) =
a
0
2
+

¸
k=1
a
k
cos(
2πkt
T
) +

¸
k=1
b
k
sin(
2πkt
T
), (3.9)
just as we did for power series in Chapter 1.
It is an interesting and diﬃcult problem in harmonic analysis to determine
how “well-behaved” a periodic function f(t) must be in order to ensure that
it can be expressed as a superposition of sines and cosines. An easily stated
theorem, suﬃcient for many applications is:
Theorem 1. If f is a continuous periodic function of period T, with condin-
uous derivative f

, then f can be written uniquely as a superposition of sines
and cosines, in accordance with (3.9), where the a
k
’s and b
k
’s are constants.
Moreover, the inﬁnite series on the right-hand side of (3.9) converges to f(t) for
every choice of t.
However, often one wants to apply the theory of Fourier series to functions
which are not quite so well-behaved, in fact to functions that are not even be
continuous, such as the function in our previous example. A weaker suﬃcient
condition for f to possess a Fourier series is that it be piecewise smooth.
The technical deﬁnition goes like this: A function f(t) which is periodic of
period T is said to be piecewise smooth if it is continuous and has a continuous
derivative f

(t) except at ﬁnitely many points of discontinuity within the interval
[0, T], and at each point t
0
of discontinuity, the right- and left-handed limits of
67
f and f

,
lim
t→t0+
(f(t)), lim
t→t0−
(f(t)), lim
t→t0+
(f

(t)), lim
t→t0−
(f

(t)),
all exist. The following theorem, proven in more advanced books,
3
ensures
that a Fourier decomposition can be found for any function which is piecewise
smooth:
Theorem 2. If f is any piecewise smooth periodic function of period T, f
can be expressed as a Fourier series,
f(t) =
a
0
2
+

¸
k=1
a
k
cos(
2πkt
T
) +

¸
k=1
b
k
sin(
2πkt
T
)
where the a
k
’s and b
k
’s are constants. Here equality means that the inﬁnite
sum on the right converges to f(t) for each t at which f is continuous. If f is
discontinuous at t
0
, its Fourier series at t
0
will converge to the average of the
right and left hand limits of f as t → t
0
.
Exercises:
3.1.1. The function f(t) = cos
2
t can be regarded as either periodic of period
π or periodic of period 2π. Choose one of the two periods and ﬁnd the Fourier
series of f(t). (Hint: This problem is very easy if you use trigonometric identities
instead of trying to integrate directly.)
3.1.2. The function f(t) = sin
3
t is periodic of period 2π. Find its Fourier series.
3.1.3. The function
f(t) =

t, for −π < t < π,
0, for t = π,
can be extended to be periodic of period 2π. Find the Fourier series of this
extension.
3.1.4. The function
f(t) = [t[, for t ∈ [−π, π],
can be extended to be periodic of period 2π. Find the Fourier series of this
extension.
3.1.5. Find the Fourier series of the following function:
f(t) =

t
2
, for −π ≤ t < π,
f(t −2kπ), for −π + 2kπ ≤ t < π + 2kπ.
3.1.6. Establish the formulae (3.5) and (3.6), which were given in the text.
3
See, for example, Ruel Churchill and James Brown, Fourier series and boundary value
problems, 4th edition, McGraw-Hill, New York, 1987 or Robert Seeley, An introduction to
Fourier series and integrals, Benjamin, New York, 1966.
68
3.2 Inner products
There is a convenient way of remembering the formulae for the Fourier coeﬃ-
cients that we derived in the preceding section. Let V be the set of piecewise
smooth functions which are periodic of period 2π. We say that V is a vec-
tor space because elements of V can be added and multiplied by scalars, these
operations satisfying the same rules as those for addition of ordinary vectors
and multiplication of ordinary vectors by scalars. We deﬁne an “inner product”
between elements of V by means of the formula
'f, g` =
1
π

π
−π
f(t)g(t)dt.
Thus for example, if f(t) = sin t and g(t) = 2 cos t, then
'f, g` =
1
π

π
−π
2 sin t cos tdt =

π
−π
sin(2t)dt = −cos(2t)[
π
−π
= 0.
The remarkable fact is that this inner product has properties quite similar
to those of the standard dot product on R
n
:
• 'f, g` = 'g, f`, whenever f and g are elements of V .
• 'f +g, h` = 'f, h` +'g, h`.
• 'cf, g` = c'f, g`, when c is a real constant.
• 'f, f` ≥ 0, with equality holding only if f = 0 (at all points of continuity).
This suggests that we might use geometric terminology for elements of V
just as we did for vectors in R
n
. Thus, for example, we will say that an element
f of V is of unit length if 'f, f` = 1 and that two elements f and g of V are
perpendicular if 'f, g` = 0.
In this terminology, the formulae (3.4), (3.5), and (3.6) can be expressed by
stating that the functions
1/

2, cos t, cos 2t, cos 3t, . . . , sin t, sin 2t, sin 3t, . . .
are of unit length and perpendicular to each other. Moreover, by the theorem in
the preceding section, any element of of V can be written as a (possibly inﬁnite)
superposition of these functions. We will say that the functions
e
0
(t) =
1

2
, e
1
(t) = cos t, e
2
(t) = cos 2t, . . . ,
ˆ e
1
(t) = sin t, ˆ e
2
(t) = sin 2t, . . .
make up an orthonormal basis for V .
We saw in Section 2.3 that if b
1
, b
2
, . . . , b
n
is an orthonormal basis for R
n
,
then
f ∈ R
n
⇒ f = (f b
1
)b
1
+ + (f b
n
)b
n
.
69
The same formula holds when R
n
is replaced by V and the dot product is
replaced by the inner product ', `: If f is any element of V , we can write
f(t) = 'f(t), e
0
(t)`e
0
(t) +'f(t), e
1
(t)`e
1
(t) +'f(t), e
2
(t)`e
2
(t) +
+'f(t), ˆ e
1
(t)`ˆ e
1
(t) +'f(t), ˆ e
2
(t)`ˆ e
2
(t) + .
In other words,
f(t) = 'f,
1

2
`
1

2
+'f, cos t` cos t +'f, cos 2t` cos 2t +. . .
+'f, sin t` sin t +'f, sin 2t` sin 2t +. . . ,
=
a
0
2
+a
1
cos t +a
2
cos 2t +. . . +b
1
sin t +b
2
sin 2t +. . . ,
where
a
0
= 'f, 1`, a
1
= 'f, cos t`, a
2
= 'f, cos 2t`, . . . ,
b
1
= 'f, sin t`, b
2
= 'f, sin 2t`, . . . .
Use of the inner product makes the formulae for Fourier coeﬃcients almost
impossible to forget.
We say that a function f(t) is odd if f(−t) = −f(t), even if f(−t) = f(t).
Thus
sin t, sin 2t, sin 3t, . . .
are odd functions, while
1, cos t, cos 2t, cos 3t, . . .
are even functions. Let
W
odd
= ¦f ∈ V : f is odd¦, W
even
= ¦f ∈ V : f is odd¦.
Then
f, g ∈ W
odd
⇒ f +g ∈ W
odd
and cf ∈ W
odd
,
for every choice of real number c. Thus we can say that W
odd
is a linear
subspace of V . Similarly, W
even
is a linear subspace of V .
It is not diﬃcult to show that
f ∈ W
odd
, g ∈ W
even
⇒ 'f, g` = 0;
in other words, the linear subspaces W
odd
and W
even
are orthogonal to each
other. Indeed, under these conditions, fg is odd and hence
'f, g` =
1
π

π
−π
f(t)g(t)dt =
1
π

0
−π
f(t)g(t)dt +
1
π

π
0
f(t)g(t)dt
70
=
1
π

0
π
f(−t)g(−t)(−dt) +
1
π

π
0
f(t)g(t)dt
= −
1
π

π
0
−(f(t)g(t))(−dt) +
1
π

π
0
f(t)g(t)dt = 0.
The variable of integration has been changed from t to −t in the ﬁrst integral
of the second line.
It follows that if f ∈ W
odd
,
a
0
= 'f, 1` = 0, a
n
= 'f, cos nx` = 0, for n > 0.
Similarly, if f ∈ W
even
,
b
n
= 'f, sin nx` = 0.
Thus for an even or odd function, half of the Fourier coeﬃcients are auto-
matically zero. This simple fact can often simplify the calculation of Fourier
coeﬃcients.
Exercises:
3.2.1. Evaluate the inner product
'f, g` =
1
π

π
−π
f(t)g(t)dt,
in the case where f(t) = cos t and g(t) = [ sin t[.
3.2.2. Evaluate the inner product
'f, g` =

e
π
1
1
x
f(x)g(x)dx,
in the case where f(x) = cos(log(x)) and g(x) = 1, where log denotes the natural
or base e logarithm. (Hint: Use the substitution x = e
u
.)
3.2.3. Determine which of the following functions are even, which are odd, and
which are neither even nor odd:
a. f(t) = t
3
+ 3t.
b. f(t) = t
2
+[t[.
c. f(t) = e
t
.
d. f(t) =
1
2
(e
t
+e
−t
).
e. f(t) =
1
2
(e
t
−e
−t
).
f. f(t) = J
0
(t), the Bessel function of the ﬁrst kind.
71
3.3 Fourier sine and cosine series
Let f : [0, L] → R be a piecewise smooth function which vanishes at 0 and L.
We claim that we can express f(t) as the superposition of sine functions,
f(t) = b
1
sin(πt/L) +b
2
sin(2πt/L) +. . . +b
n
sin(nπt/L) +. . . . (3.10)
We could prove this using the theory of even and odd functions. Indeed, we can
extend f to an odd function
˜
f : [−L, L] → R by setting
˜
f(t) =

f(t), for t ∈ [0, L],
−f(−t), for t ∈ [−L, 0],
then to a function
ˆ
f : R → R, which is periodic of period 2L by requiring that
ˆ
f(t + 2L) =
ˆ
f(t), for all t ∈ R.
The extended function lies in the linear subspace W
odd
. It follows from the
theorem in Section 3.1 that
ˆ
f possesses a Fourier series expansion, and from the
fact that
ˆ
f is odd that all of the a
n
’s are zero. On the interval [0, L],
ˆ
f restricts
to f and the Fourier expansion of
ˆ
f restricts to an expansion of f of the form
(3.10) which involves only sine functions. We call (3.10) the Fourier sine series
of f.
A similar argument can be used to express a piecewise smooth function
f : [0, L] → R into a superposition of cosine functions,
f(t) =
a
0
2
+a
1
cos(πt/L) +a
2
cos(2πt/L) +. . . +a
n
cos(nπt/L) +. . . . (3.11)
To obtain this expression, we ﬁrst extend f to an even function
˜
f : [−L, L] → R
by setting
˜
f(t) =

f(t), for t ∈ [0, L],
f(−t), for t ∈ [−L, 0],
then to a function
ˆ
f : R → R which is periodic of period 2L, by requiring that
ˆ
f(t + 2L) =
ˆ
f(t), for all t ∈ RR.
This time the extended function lies in the linear subspace W
even
. It follows
from the theorem in Section 3.1 that
ˆ
f possesses a Fourier series expansion, and
from the fact that
ˆ
f is even that all of the b
n
’s are zero. On the interval [0, L],
ˆ
f restricts to f and the Fourier expansion of
ˆ
f restricts to an expansion of f of
the form (3.11) which involves only cosine functions. We call (3.11) the Fourier
cosine series of f.
To generate formulae for the coeﬃcients in the Fourier sine and cosine series
of f, we begin by deﬁning a slightly diﬀerent inner product space than the one
72
considered in the preceding section. This time, we let V be the set of piecewise
smooth functions f : [0, L] → R and let
V
0
= ¦f ∈ V : f(0) = 0 = f(L)¦,
a linear subspace of V . We deﬁne an inner product ', ` on V by means of the
formula
'f, g` =
2
L

L
0
f(t)g(t)dt.
This restricts to an inner product on V
0
.
Let’s consider now the Fourier sine series. We have seen that any element
of V
0
can be represented as a superpostion of the sine functions
sin(πt/L), sin(2πt/L), . . . , sin(nπt/L), . . . .
We claim that these sine functions form an orthonormal basis for V
0
with respect
to the inner product we have deﬁned. Recall the trigonometric formulae that
we used in '3.1:
cos((n +m)πt/L) = cos(nπt/L) cos(mπt/L) −sin(nπt/L) sin(mπt/L),
cos((n −m)πt/L) = cos(nπt/L) cos(mπt/L) + sin(nπt/L) sin(mπt/L).
Subtracting the ﬁrst of these from the second and dividing by two yields
sin(nπt/L) sin(mπt/L) =
1
2
(cos((n −m)πt/L) −cos((n +m)πt/L),
and hence

L
0
sin(nπt/L) sin(mπt/L)dt =
1
2

L
0
(cos((n −m)πt/L) −cos((n +m)πt/L)dt.
If n and m are positive integers, the integral on the right vanishes unless n = m,
in which case the right-hand side becomes
1
2

L
0
dt =
L
2
.
Hence
2
L

π
0
sin(nπt/L) sin(mπt/L)dt =

1, for m = n,
0, for m = n.
(3.12)
Therefore, just as in the previous section, we can evaluate the coeﬃcients of
the Fourier sine series of a function f ∈ V
0
by simply projecting f onto each
element of the orthonormal basis. When we do this, we ﬁnd that
f(t) = b
1
sin(πt/L) +b
2
sin(2πt/L) +. . . + sin(nπt/L) +. . . ,
73
where
b
n
= 'f, sin(nπt/L)` =
2
L

L
0
f(t) sin(nπt/L)dt. (3.13)
We can treat the Fourier cosine series in a similar fashion. In this case, we
can show that the functions
1

2
, cos(πt/L), cos(2πt/L), . . . , cos(nπt/L), . . .
form an orthonormal basis for V . Thus we can evaluate the coeﬃcients of the
Fourier cosine series of a function f ∈ V by projecting f onto each element of
this orthonormal basis. We will leave it to the reader to carry this out in detail,
and simply remark that when the dust settles, one obtains the following formula
for the coeﬃcient a
n
in the Fourier cosine series:
a
n
=
2
L

L
0
f(t) cos(nπt/L)dt. (3.14)
Example. First let us use (3.13) to ﬁnd the Fourier sine series of
f(t) =

t, for 0 ≤ t ≤ π/2,
π −t, for π/2 ≤ t ≤ π.
(3.15)
In this case, L = π, and according to our formula,
b
n
=
2
π
¸

π/2
0
t sin ntdt +

π
π/2
(π −t) sin ntdt
¸
.
We use integration by parts to obtain

π/2
0
t sin ntdt =
¸
−t
n
cos nt

π/2
0
+

π/2
0
1
n
cos ntdt
=
−π cos(nπ/2)
2n
+
1
n
2
[sin nt][
π/2
0
=
−π cos(nπ/2)
2n
+
sin(nπ/2)
n
2
,
while

π
π/2
(π −t) sin ntdt =
¸
−(π −t)
n
cos(nt)

π
π/2

π
π/2
1
n
cos ntdt
=
π cos(nπ/2)
2n

1
n
2
[sin nt][
π
π/2
=
π cos(nπ/2)
2n
+
sin(nπ/2)
n
2
.
Thus
b
n
=
4 sin(nπ/2)
πn
2
,
74
0.5 1 1.5 2 2.5 3
0.2
0.4
0.6
0.8
1
1.2
1.4
Figure 3.3: A graph of the Fourier sine approximations φ
1
, φ
3
, φ
5
and φ
7
.
and the Fourier sine series of f(t) is
f(t) =
4
π
sin t −
4

sin 3t +
4
25π
sin 5t −
4
49π
sin 7t +. . . . (3.16)
The trigonometric polynomials
φ
1
(t) =
4
π
sin t, φ
3
(t) =
4
π
sin t −
4

sin 3t, . . .
are better and better approximations to the function f(t).
To ﬁnd the Fourier cosine series of
f(t) =

t, for 0 ≤ t ≤ π/2,
π −t, for π/2 ≤ t ≤ π,
we ﬁrst note that
a
0
=
2
π
¸

π/2
0
tdt +

π
π/2
(π −t)dt
¸
=
2
π
¸
1
2

π
2

2
+
1
2

π
2

2

=
π
2
.
To ﬁnd the other a
n
’s, we use (3.14):
a
n
=
2
π
¸

π/2
0
t cos ntdt +

π
π/2
(π −t) cos ntdt
¸
.
This time, integration by parts yields

π/2
0
t cos ntdt =
¸
t
n
sin nt

π/2
0

π/2
0
1
n
sin ntdt
75
=
π sin(nπ/2)
2n
+
1
n
2
[cos nt][
π/2
0
=
π sin(nπ/2)
2n
+
cos(nπ/2)
n
2

1
n
2
while

π
π/2
(π −t) cos ntdt =
¸
(π −t)
n
sin(nt)

π
π/2
+

π
π/2
1
n
sin ntdt
=
−π sin(nπ/2)
2n

1
n
2
[cos nt][
π
π/2
=
−π sin(nπ/2)
2n
+
cos(nπ/2)
n
2

1
n
2
cos(nπ).
Thus when n ≥ 1,
a
n
=
2
πn
2
[2 cos(nπ/2) −1 −1(−1)
n
],
and the Fourier sine series of f(t) is
f(t) =
π
4

2
π
cos 2t −
2

cos 6t −
2
25π
cos 10t −. . . . (3.17)
Note that we have expanded exactly the same function f(t) on the interval [0, π]
as either a superposition of sines in (3.16) or as a superposition of cosines in
(3.17).
Exercises:
3.3.1.a. Find the Fourier sine series of the following function deﬁned on the
interval [0, π]:
f(t) =

4t, for 0 ≤ t < π/2,
4π −4t, for π/2 ≤ t ≤ π.
b. Find the Fourier cosine series of the same function.
3.3.2.a. Find the Fourier sine series of the following function deﬁned on the
interval [0, π]:
f(t) = t(π −t).
b. Find the Fourier cosine series of the same function.
3.3.3. Find the Fourier sine series of the following function deﬁned on the
interval [0, 10]:
f(t) =

t, for 0 ≤ t < 5,
10 −t, for 5 ≤ t ≤ 10.
76
3.3.4. Find the Fourier sine series of the following function deﬁned on the
interval [0, 1]:
f(t) = 5t(1 −t).
to the ﬁrst ten coeﬃcients of the Fourier sine series for the function
f(t) = t +t
2
−2t
3
,
deﬁned for t in the interval [0, 1], by running the following Mathematica program
f[n ] := 2 NIntegrate[(t + t∧2 - 2 t∧3) Sin[n Pi t], ¦t,0,1¦];
b = Table[f[n], ¦n,1,10¦]
3.4 Complex version of Fourier series*
We have seen that if f : R → R is a well-behaved function which is periodic of
period 2π, f can be expanded in a Fourier series
f(t) =
a
0
2
+a
1
cos t +a
2
cos(2t) +. . .
+b
1
sin t +b
2
sin(2t) +. . . .
We say that this is the real form of the Fourier series. It is often convenient to
recast this Fourier series in complex form by means of the Euler formula, which
states that
e

= cos θ +i sin θ.
It follows from this formula that
e

+e
−iθ
= 2 cos θ, e
−iθ
+e
−iθ
= 2i sin θ,
or
cos θ =
e

+e
−iθ
2
, sin θ =
e

+e
−iθ
2i
.
Hence the Fourier expansion of f can be rewritten as
f(t) =
a
0
2
+
a
1
2
(e
it
+e
−it
) +
a
2
2
(e
2it
+e
−2it
) +. . .
+
b
1
2i
(e
it
−e
−it
) +
b
2
2i
(e
2it
−e
−2it
) +. . . ,
or
f(t) = . . . +c
−2
e
−2it
+c
−1
e
−it
+c
0
+c
1
e
it
+c
2
e
2it
+. . . , (3.18)
where the c
k
’s are the complex numbers deﬁned by the formulae
c
0
=
a
0
2
, c
1
=
a
1
−ib
1
2
, c
2
=
a
2
−ib
2
2
, . . . ,
77
c
−1
=
a
1
+ib
1
2
, c
−2
=
a
2
+ib
2
2
, . . . .
If k = 0, we can solve for a
k
and b
k
:
a
k
= c
k
+c
−k
, b
k
= i(c
k
−c
−k
). (3.19)
It is not diﬃcult to check the following integral formula via direct integration:

π
−π
e
imt
e
−int
dt =

2π, for m = n,
0 for m = n.
(3.20)
If we multiply both sides of (3.18) by e
−ikt
, integrate from −π to π and apply
(3.20), we obtain

π
−π
f(t)e
−ikt
dt = 2πc
k
,
which yields the formula for coeﬃcients of the complex form of the Fourier
series:
c
k
=
1

π
−π
f(t)e
−ikt
dt. (3.21)
Example. Let us use formula (3.21) to ﬁnd the complex Fourier coeﬃcients of
the function
f(t) = t for −π < t ≤ π,
extended to be periodic of period 2π. Then
c
k
=
1

π
−π
te
−ikt
dt.
We apply integration by parts with u = t, dv = e
−ikt
dt, du = dt and v =
(i/k)e
−ikt
:
c
k
=
1

¸
(it/k)e
−ikt
[
π
−π

π
−π
(i/k)e
−ikt
dt

=
1

¸
πi
k
e
−iπk
+
πi
k
e
iπk

= i
(−1)
k
k
.
It follows from (3.19) that
a
k
= (c
k
+c
−k
) = 0, b
k
= i(c
k
−c
−k
) = −2
(−1)
k
k
.
It is often the case that the complex form of the Fourier series is far simpler to
calculate than the real form. One can then use (3.19) to ﬁnd the real form of
the Fourier series.
Exercise:
3.4.1. Prove the integral formula (3.20), presented in the text.
78
3.4.2. a. Find the complex Fourier coeﬃcients of the function
f(t) = t
2
for −π < t ≤ π,
extended to be periodic of period 2π.
b. Use (3.19) to ﬁnd the real form of the Fourier series.
3.4.3. a. Find the complex Fourier coeﬃcients of the function
f(t) = t(π −t) for −π < t ≤ π,
extended to be periodic of period 2π.
b. Use (3.19) to ﬁnd the real form of the Fourier series.
3.4.4. Show that if a function f : R → R is smooth and periodic of period 2πL,
we can write the Fourier expansion of f as
f(t) = . . . +c
−2
e
−2it/L
+c
−1
e
−it/L
+c
0
+c
1
e
it/L
+c
2
e
2it/L
+. . . ,
where
c
k
=
1
2πL

πL
−πL
f(t)e
−ikt/L
dt.
3.5 Fourier transforms*
One of the problems with the theory of Fourier series presented in the previous
sections is that it applies only to periodic functions. There are many times when
one would like to divide a function which is not periodic into a superposition of
sines and cosines. The Fourier transform is the tool often used for this purpose.
The idea behind the Fourier transform is to think of f(t) as vanishing outside
a very long interval [−πL, πL]. The function can be extended to a periodic
function f(t) such that f(t + 2πL) = f(t). According to the theory of Fourier
series in complex form (see Exercise 3.4.4),
f(t) = . . . +c
−2
e
−2it/L
+c
−1
e
−it/L
+c
0
+c
1
e
it/L
+c
2
e
2it/L
+. . . ,
where the c
k
’s are the complex numbers.
Deﬁnition. If f : R → R is a piecewise continuous function which vanishes
outside some ﬁnite interval, its Fourier transform is
ˆ
f(ξ) =

−∞
f(t)e
−iξt
dt. (3.22)
The integral in this formula is said to be improper because it is taken from −∞
to ∞; it is best to regard it as a limit,

−∞
f(t)e
−iξt
dt = lim
L→∞

πL
−πL
f(t)e
−iξt
dt.
79
To explain (3.22), we suppose that f vanishes outside the interval [−πL, πL].
We can extend the restriction of f to this interval to a function which is periodic
of period 2πL. Then
ˆ
f(k/L) =

−∞
f(t)e
−ikt/L
dt =

πL
−πL
˜
f(t)e
−ikt/L
dt
represents 2πL times the Fourier coeﬃcient of this extension of frequency k/L;
indeed, it follows from (3.21) that we can write
f(t) = . . . +c
−2
e
−2it/L
+c
−1
e
−it/L
+c
0
+c
1
e
it/L
+c
2
e
2it/L
+. . . ,
for t ∈ [−πL, πL], where
c
k
=
1
2πL
ˆ
f(k/L),
or alternatively,
f(t) = . . . +
1
2πL
ˆ
f(−2/L)e
−2it/L
+
1
2πL
ˆ
f(−1/L)e
−it/L
+
+
1
2πL
ˆ
f(0) +
1
2πL
ˆ
f(1/T)e
it/L
+
1
2πL
ˆ
f(2/L)e
2it/L
+. . . .
In the limit as L → ∞, it can be shown that this last sum approaches an
improper integral, and our formula becomes
f(t) =
1

−∞
ˆ
f(ξ)e
iξt
dξ. (3.23)
Equation (3.23) is called the Fourier inversion formula. If we make use of
Euler’s formula, we can write the Fourier inversion formula in terms of sines
and cosines,
f(t) =
1

−∞
ˆ
f(ξ) cos ξtdξ +
i

−∞
ˆ
f(ξ) sin ξtdξ,
a superposition of sines and cosines of various frequencies.
Equations (3.22) and (3.22) allow one to pass back and forth between a given
function and its representation as a superposition of oscillations of various fre-
quencies. Like the Laplace transform, the Fourier transform is often an eﬀective
tool in ﬁnding explicit solutions to diﬀerential equations.
Exercise:
3.5.1. Find the Fourier transform of the function f(t) deﬁned by
f(t) =

1, if −1 ≤ t ≤ 1,
0, otherwise.
80
Chapter 4
Partial Diﬀerential
Equations
4.1 Overview
A partial diﬀerential equation is an equation which contains partial derivatives,
such as the equation
∂u
∂t
=

2
u
∂x
2
,
in which u is regarded as a function of x and t. Unlike the theory of ordinary
diﬀerential equations which centers upon one key theorem—the fundamental
existence and uniqueness theorem—there is no real uniﬁed theory of partial dif-
ferential equations. Instead, each type of partial diﬀerential equations exhibits
its own special features, which usually mirror the physical phenomena which
the equation was ﬁrst used to model.
Many of the foundational theories of physics and engineering are expressed
by means of systems of partial diﬀerential equations. The reader may have
heard some of these equations mentioned in previous courses in physics. Fluid
mechanics is often formulated by the Euler equations of motion or the so-called
Navier-Stokes equations, electricity and magnetism by Maxwell’s equations, gen-
eral relativity by Einstein’s ﬁeld equations. It is therefore important to develop
techniques that can be used to solve a wide variety of partial diﬀerential equa-
tions.
In this chapter, we will give two important simple examples of partial dif-
ferential equations, the heat equation and the wave equation, and we will show
how to solve them by the techniques of “separation of variables” and Fourier
analysis. Higher dimensional examples will be given in the following chapter.
We will see that just as in the case of ordinary diﬀerential equations, there is an
important dichotomy between linear and nonlinear equations. The techniques
of separation of variables and Fourier analysis are eﬀective only for linear par-
tial diﬀerential equations. Nonlinear partial diﬀerential equations are far more
81
diﬃcult to solve, and form a key topic of contemporary mathematical research.
1
Our ﬁrst example is the equation governing propagation of heat in a bar of
length L. We imagine that the bar is located along the x-axis and we let
u(x, t) = temperature of the bar at the point x at time t.
Heat in a small segment of a homogeneous bar is proportional to temper-
ature, the constant of proportionality being determined by the density and
speciﬁc heat of the material making up the bar. More generally, if σ(x) denotes
the speciﬁc heat at the point x and ρ(x) is the density of the bar at x, then the
heat within the region D
x1,x2
between x
1
and x
2
is given by the formula
Heat within D
x1,x2
=

x2
x1
ρ(x)σ(x)u(x, t)dx.
To calculate the rate of change of heat within D
x1,x2
with respect to time, we
simply diﬀerentiate under the integral sign:
d
dt
¸
x2
x1
ρ(x)σ(x)u(x, t)dx

=

x2
x1
ρ(x)σ(x)
∂u
∂t
(x, t)dx.
Now heat is a form of energy, and conservation of energy implies that if no
heat is being created or destroyed in D
x1,x2
, the rate of change of heat within
D
x1,x2
is simply the rate at which heat enters D
x1,x2
. Hence the rate at which
heat leaves D
x1,x2
is given by the exp
Rate at which heat leaves D
x1,x2
= −

x2
x1
ρ(x)σ(x)
∂u
∂t
(x, t)dx. (4.1)
(More generally, if heat is being created within D
x1,x2
, say by a chemical reac-
tion, at the rate µ(x)u(x, t) +ν(x) per unit volume, then the rate at which heat
leaves D
x1,x2
is

x2
x1
ρ(x)σ(x)
∂u
∂t
(x, t)dx +

x2
x1
(µ(x)u(x, t) +ν(x))dx.)
On the other hand, the rate of heat ﬂow F(x, t) is proportional to the partial
derivative of temperature,
F(x, t) = −κ(x)
∂u
∂x
(x, t), (4.2)
where κ(x) is the thermal conductivity of the bar at x. Thus we ﬁnd that the
rate at which heat leaves the region D
x1,x2
is also given by the formula
F(x
2
, t) −F(x
1
, t) =

x2
x1
∂F
∂x
(x, t)dx. (4.3)
1
Further reading can be found in the many excellent upper-division texts on partial diﬀer-
ential equations. We especially recommend Mark Pinsky, Partial diﬀerential equations and
boundary-value problems with applications, 2nd edition, McGraw-Hill, 1991. An excellent
but much more advanced book is Michael Taylor, Partial diﬀerential equations: basic theory,
Springer, New York, 1996.
82
Comparing the two formulae (4.1) and (4.3), we ﬁnd that

x2
x1
∂F
∂x
(x, t)dx = −

x2
x1
ρ(x)σ(x)
∂u
∂t
(x, t)dx.
This equation is true for all choices of x
1
and x
2
, so the integrands on the two
sides must be equal:
∂F
∂x
= −ρ(x)σ(x)
∂u
∂t
.
It follows from (4.2) that

∂x

−κ(x)
∂u
∂x

= −ρ(x)σ(x)
∂u
∂t
.
In this way, we obtain the heat equation
ρ(x)σ(x)
∂u
∂t
=

∂x

κ(x)
∂u
∂x

. (4.4)
In the more general case in which heat is being created at the rate µ(x)u(x, t) +
ν(x) per unit length, one could show that heat ﬂow is modeled by the equation
ρ(x)σ(x)
∂u
∂t
=

∂x

κ(x)
∂u
∂x

+µ(x)u +ν(x). (4.5)
In the special case where the bar is homogeneous, i.e. its properties are the
same at every point, ρ(x), σ(x) and κ(x) are constants, say σ and κ respectively,
and (4.4) becomes
∂u
∂t
=
κ
ρσ

2
u
∂x
2
. (4.6)
This is our simplest example of a linear partial diﬀerential equation. Although
its most basic application concerns diﬀusion of heat, it arises in many other
contexts as well. For example, a slight modiﬁcation of the heat equation was
used by Black and Scholes to price derivatives in ﬁnancial markets.
2
Exercises:
4.1.1. Study of heat ﬂow often leads to “boundary-value problems” for ordi-
nary diﬀerential equations. Indeed, in the “steady-state” case, in which u is
independent of time, equation (4.5) becomes
d
dx

κ(x)
du
dx
(x)

+µ(x)u(x) +ν(x) = 0,
2
A description of the Black-Scholes technique for pricing puts and calls is given in Paul
Wilmott, Sam Howison and Jeﬀ Dewynne, The mathematics of ﬁnancial derivatives, Cam-
bridge Univ. Press, 1995.
83
a linear ordinary diﬀerential equation with variable coeﬃcients. Suppose now
that the temperature is speciﬁed at the two endpoints of the bar, say
u(0) = α, u(L) = β.
Our physical intuition suggests that the steady-state heat equation should have
a unique solution with these boundary conditions.
a. Solve the following special case of this boundary-value problem: Find u(x),
deﬁned for 0 ≤ x ≤ 1 such that
d
2
u
dx
2
= 0, u(0) = 70, u(1) = 50.
b. Solve the following special case of this boundary-value problem: Find u(x),
deﬁned for 0 ≤ x ≤ 1 such that
d
2
u
dx
2
−u = 0, u(0) = 70, u(1) = 50.
c. Solve the following special case of this boundary-value problem: Find u(x),
deﬁned for 0 ≤ x ≤ 1 such that
d
2
u
dx
2
+x(1 −x) = 0, u(0) = 0, u(1) = 0.
d. (For students with access to Mathematica) Use Mathematica to graph the
solution to the following boundary-value problem: Find u(x), deﬁned for 0 ≤
x ≤ 1 such that
d
2
u
dx
2
+ (1 +x
2
)u = 0, u(0) = 50, u(1) = 100.
You can do this by running the Mathematica program:
a = 0; b = 1; alpha = 50; beta = 100;
sol = NDSolve[ ¦u’’[x] + (1 + x∧2) u[x] == 0,
u[a] == alpha, u[b] == beta.¦, u, ¦x,a,b¦];
Plot[ Evaluate[ u[x] /. sol], ¦x,a,b¦]
4.1.2.a. Show that the function
u
0
(x, t) =
1

4πt
e
−x
2
/4t
is a solution to the heat equation (4.6) for t > 0 in the case where κ/(ρσ) = 1.
b. Use the chain rule with intermediate variables ¯ x = x −a,
¯
t = t to show that
u
a
(x, t) =
1

4πt
e
−(x−a)
2
/4t
84
is also a solution to the heat equation.
c. Show that

−∞
u
0
(x, t)dx = 1.
Hint: Let I denote the integral on the left hand side and note that
I
2
=
1
4πt

−∞

−∞
e
−(x
2
+y
2
)/4t
dxdy.
Then transform this last integral to polar coordinates.
d. Use Mathematica to sketch u
0
(x, t) for various values of t. What can you
say about the behaviour of the function u
0
(x, t) as t → 0?
e. By diﬀerentiating under the integral sign show that if h : R → R is any
smooth function which vanishes outside a ﬁnite interval [−L, L], then
u(x, t) =

−∞
u
a
(x, t)h(a)da (4.7)
is a solution to the heat equation.
REMARK: In more advanced courses it is shown that (4.7) approaches h(x) as
t → 0. In fact, (4.7) gives a formula (for values of t which are greater than zero)
for the unique solution to the heat equation on the inﬁnite line which satisﬁes
the initial condition u(x, 0) = h(x). In the next section, we will see how to solve
the initial value problem for a rod of ﬁnite length.
4.2 The initial value problem for the heat equa-
tion
We will now describe how to use the Fourier sine series to ﬁnd the solution to
an initial value problem for the heat equation in a rod of length L which is
insulated along the sides, whose ends are kept at zero temperature. We expect
that there should exist a unique function u(x, t), deﬁned for 0 ≤ x ≤ L and
t ≥ 0 such that
1. u(x, t) satisﬁes the heat equation
∂u
∂t
= c
2

2
u
∂x
2
, (4.8)
where c is a constant.
2. u(x, t) satisﬁes the boundary condition u(0, t) = u(L, t) = 0, in other
words, the temperature is zero at the endpoints. (This is sometimes called
the Dirichlet boundary condition.)
85
3. u(x, t) satisﬁes the initial condition u(x, 0) = h(x), where h(x) is a given
function, the initial temperature of the rod.
In more advanced courses, it is proven that this initial value problem does in
fact have a unique solution. We will shortly see how to ﬁnd that solution.
Note that the heat equation itself and the boundary condition are homoge-
neous and linear—this means that if u
1
and u
2
satisfy these conditions, so does
c
1
u
1
+ c
2
u
2
, for any choice of constants c
1
and c
2
. Thus homogeneous linear
conditions satisfy the principal of superposition.
Our method makes use of the dichotomy into homogeneous and nonhomo-
geneous conditions:
Step I. We ﬁnd all of the solutions to the homogeneous linear conditions of the
special form
u(x, t) = f(x)g(t).
By the superposition principle, an arbitrary linear superposition of these solu-
tions will still be a solution.
Step II. We ﬁnd the particular solution which satisﬁes the nonhomogeneous
condition by Fourier analysis.
Let us ﬁrst carry out Step I. We substitute u(x, t) = f(x)g(t) into the heat
equation (4.8) and obtain
f(x)g

(t) = c
2
f

(x)g(t).
Now we separate variables, putting all the functions involving t on the left, all
the functions involving x on the right:
g

(t)
g(t)
= c
2
f

(x)
f(x)
.
The left-hand side of this equation does not depend on x, while the right-hand
side does not depend on t. Hence neither side can depend upon either x or t.
In other words, the two sides must equal a constant, which we denote by λ and
call the separating constant. Our equation now becomes
g

(t)
c
2
g(t)
=
f

(x)
f(x)
= λ,
which separates into two ordinary diﬀerential equations,
g

(t)
c
2
g(t)
= λ, or g

(t) = λc
2
g(t), (4.9)
and
f

(x)
f(x)
= λ, or f

(x) = λf(x). (4.10)
86
The homogeneous boundary condition u(0, t) = u(L, t) = 0 becomes
f(0)g(t) = f(L)g(t) = 0.
If g(t) is not identically zero,
f(0) = f(L) = 0.
(If g(t) is identically zero, then so is u(x, t), and we obtain only the trivial
solution u ≡ 0.)
Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the
problem requires us to ﬁnd the nontrivial solutions to a boundary value problem
for an ordinary diﬀerential equation:
f

(x) =
d
2
dx
2
(f(x)) = λf(x), f(0) = 0 = f(L). (4.11)
We will call (4.11) the eigenvalue problem for the diﬀerential operator
L =
d
2
dx
2
acting on the space V
0
of well-behaved functions f : [0, L] → R which vanish at
the endpoints 0 and L.
We need to consider three cases. (As it turns out, only one of these will
actually yield nontrivial solutions to our eigenvalue problem.)
Case 1: λ = 0. In this case, the eigenvalue problem (4.11) becomes
f

(x) = 0, f(0) = 0 = f(L).
The general solution to the diﬀerential equation is f(x) = ax +b, and the only
particular solution which satisﬁes the boundary condition is f = 0, the trivial
solution.
Case 2: λ > 0. In this case, the diﬀerential equation
f

(x) = λf(x), or f

(x) −λf(x) = 0
has the general solution
f(x) = c
1
e

λx
+c
2
e

λx
.
It is convenient for us to change basis in the linear space of solutions, using
cosh(

λx) =
1
2
(e

λx
+e

λx
), sinh(

λx) =
1
2
(e

λx
−e

λx
)
instead of the exponentials. Then we can write
f(x) = a cosh(

λx) +b sinh(

λx),
87
with new constants of integration a and b. We impose the boundary conditions:
ﬁrst
f(0) = 0 ⇒ a = 0 ⇒ f(x) = b sinh(

λx),
and then
f(L) = 0 ⇒ b sinh(

λL) = 0 ⇒ b = 0 ⇒ f(x) = 0,
so we obtain no nontrivial solutions in this case.
Case 3: λ < 0. In this case, we set ω =

−λ, and rewrite the eigenvalue
problem as
f

(x) +ω
2
f(x) = 0, f(0) = 0 = f(L).
We recognize here our old friend, the diﬀerential equation of simple harmonic
motion. We remember that the diﬀerential equation has the general solution
f(x) = a cos(ωx) +b sin(ωx).
Once again
f(0) = 0 ⇒ a = 0 ⇒ f(x) = b sin(ωx).
Now, however,
f(L) = 0 ⇒ b sin(ωL) ⇒ b = 0 or sin(ωL) = 0,
and hence either b = 0 and we obtain only the trivial solution or sin(ωL) = 0.
The latter possibility will occur if ωL = nπ, or ω = (nπ/L), where n is an
integer. In this case, we obtain
f(x) = b sin(nπx/L).
Therefore, we conclude that the only nontrivial solutions to (4.11) are constant
multiples of
f(x) = sin(nπx/L), with λ = −(nπ/L)
2
, n = 1, 2, 3, . . . .
For each of these solutions, we need to ﬁnd a corresponding g(t) solving equation
(4.9),
g

(t) = λc
2
g(t),
where λ = −(nπ/L)
2
. This is just the equation of exponential decay, and has
the general solution
g(t) = be
−(ncπ/L)
2
t
,
where a is a constant of integration. Thus we ﬁnd that the nontrivial prod-
uct solutions to the heat equation together with the homogeneous boundary
condition u(0, t) = 0 = u(L, t) are constant multiples of
u
n
(x, t) = sin(nπx/L)e
−(ncπ/L)
2
t
.
88
It follows from the principal of superposition that
u(x, t) = b
1
sin(πx/L)e
−(cπ/L)
2
t
+b
2
sin(2πx/L)e
−(2cπ/L)
2
t
+. . . (4.12)
is a solution to the heat equation together with its homogeneous boundary
conditions, for arbitrary choice of the constants b
1
, b
2
, . . . .
Step II consists of determining the constants b
n
in (4.12) so that the initial
condition u(x, 0) = h(x) is satisﬁed. Setting t = 0 in (4.12) yields
h(x) = u(x, 0) = b
1
sin(πx/L) +b
2
sin(2πx/L) +. . . .
It follows from the theory of the Fourier sine series that h can indeed be rep-
resented as a superposition of sine functions, and we can determine the b
n
’s as
the coeﬃcients in the Fourier sine series of h. Using the techniques described in
Section 3.3, we ﬁnd that
b
n
=
2
L

L
0
h(x) sin(nπx/L)dx.
Example 1. Suppose that we want to ﬁnd the function u(x, t), deﬁned for
0 ≤ x ≤ π and t ≥ 0, which satisﬁes the initial-value problem:
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = h(x) = 4 sin x+2 sin 2x+7 sin 3x.
In this case, the nonvanishing coeﬃcients for the Fourier sine series of h are
b
1
= 4, b
2
= 2, b
3
= 7,
so the solution must be
u(x, t) = 4 sin xe
−t
+ 2 sin 2xe
−4t
+ 7 sin 3xe
−9t
.
Example 2. Suppose that we want to ﬁnd the function u(x, t), deﬁned for
0 ≤ x ≤ π and t ≥ 0, which satisﬁes the initial-value problem:
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = h(x),
where
h(x) =

x, for 0 ≤ x ≤ π/2,
π −x, for π/2 ≤ x ≤ π.
We saw in Section 3.3 that the Fourier sine series of h is
h(x) =
4
π
sin x −
4

sin 3x +
4
25π
sin 5x −
4
49π
sin 7x +. . . ,
and hence
u(x, t) =
4
π
sin xe
−t

4

sin 3xe
−9t
+
4
25π
sin 5xe
−25t
−. . . .
89
Exercises:
4.2.1. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = sin 2x.
You may assume that the nontrivial solutions to the eigenvalue problem
f

(x) = λf(x), f(0) = 0 = f(π)
are
λ = −n
2
, f(x) = b
n
sin nx, for n = 1, 2, 3, . . . ,
where the b
n
’s are constants.
4.2.2. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = sin x + 3 sin 2x −5 sin 3x.
4.2.3. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = x(π −x).
(In this problem you need to ﬁnd the Fourier sine series of h(x) = x(π −x).)
4.2.4. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
1
2t + 1
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = sin x + 3 sin 2x.
4.2.5. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
(t + 1)
∂u
∂t
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0, u(x, 0) = sin x + 3 sin 2x.
4.2.6.a. Find the function w(x), deﬁned for 0 ≤ x ≤ π, such that
d
2
w
dx
2
= 0, w(0) = 10, w(π) = 50.
b. Find the general solution to the following boundary value problem for the
heat equation: Find the functions u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, such
that
∂u
∂t
=

2
u
∂x
2
, u(0, t) = 10, u(π, t) = 50. (4.13)
90
(Hint: Let v = u − w, where w is the solution to part a. Determine what
conditions v must satisfy.)
c. Find the particular solution to (4.13) which in addition satisﬁes the initial
condition
u(x, 0) = 10 +
40
π
x + 2 sin x −5 sin 2x.
4.2.7.a. Find the eigenvalues and corresponding eigenfunctions for the diﬀeren-
tial operator
L =
d
2
dt
2
+ 3,
which acts on the space V
0
of well-behaved functions f : [0, π] → R which vanish
at the endpoints 0 and π by
L(f) =
d
2
f
dt
2
+ 3f.
b. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂x
2
+ 3u, u(0, t) = u(π, t) = 0, u(x, 0) = sin x + 3 sin 2x.
4.2.8. The method described in this section can also be used to solve an initial
value problem for the heat equation in which the Dirichlet boundary condition
u(0, t) = u(L, t) = 0 is replaced by the Neumann boundary condition
∂u
∂x
(0, t) =
∂u
∂x
(L, t) = 0. (4.14)
(Physically, this corresponds to insulated endpoints, from which no heat can
enter or escape.) In this case separation of variables leads to a slightly diﬀerent
eigenvalue problem, which consists of ﬁnding the nontrivial solutions to
f

(x) =
d
2
dx
2
(f(x)) = λf(x), f

(0) = 0 = f

(L).
a. Solve this eigenvalue problem. (Hint: The solution should involve cosines
b. Find the general solution to the heat equation
∂u
∂t
=

2
u
∂x
2
subject to the Neumann boundary condition (4.14).
91
c. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, such that:
∂u
∂t
=

2
u
∂x
2
,
∂u
∂x
(0, t) =
∂u
∂x
(π, t) = 0, u(x, 0) = 3 cos x + 7 cos 2x.
4.2.9. We can also treat a mixture of Dirichlet and Neumann conditions, say
u(0, t) =
∂u
∂x
(L, t) = 0. (4.15)
In this case separation of variables leads to the eigenvalue problem which consists
of ﬁnding the nontrivial solutions to
f

(x) =
d
2
dx
2
(f(x)) = λf(x), f(0) = 0 = f

(L).
a. Solve this eigenvalue problem.
b. Find the general solution to the heat equation
∂u
∂t
=

2
u
∂x
2
subject to the mixed boundary condition (4.15).
c. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, such that:
∂u
∂t
=

2
u
∂x
2
, u(0, t) =
∂u
∂x
(π, t) = 0, u(x, 0) = 4 sin(x/2) + 12 sin(3x/2).
4.3 Numerical solutions to the heat equation
There is another method which is sometimes used to treat the initial value
problem described in the preceding section, a numerical method based upon
“ﬁnite diﬀerences.” Although it yields only approximate solutions, it can be
applied in some cases with variable coeﬃcients when it would be impossible to
apply Fourier analysis in terms of sines and cosines. However, for simplicity,
we will describe only the case where ρ and k are constant, and in fact we will
assume that c
2
= L = 1.
Thus we seek the function u(x, t), deﬁned for 0 ≤ x ≤ 1 and t ≥ 0, which
solves the heat equation
∂u
∂t
=

2
u
∂x
2
subject to the boundary conditions u(0, t) = u(1, t) = 0 and the initial con-
dition u(x, 0) = h(x), where h(x) is a given function, representing the initial
temperature.
For any ﬁxed choice of t
0
the function u(x, t
0
) is an element of V
0
, the space of
piecewise smooth functions deﬁned for 0 ≤ x ≤ 1 which vanish at the endpoints.
92
Our idea is to replace the “inﬁnite-dimensional” space V
0
by a ﬁnite-dimensional
Euclidean space R
n−1
and reduce the partial diﬀerential equation to a system
of ordinary diﬀerential equations. This corresponds to utilizing a discrete model
for heat ﬂow rather than a continuous one.
For 0 ≤ i ≤ n, let x
i
= i/n and
u
i
(t) = u(x
i
, t) = the temperature at x
i
at time t.
Since u
0
(t) = 0 = u
n
(t) by the boundary conditions, the temperature at time t
is speciﬁed by
u(t) =

¸
¸
¸
u
1
(t)
u
2
(t)

u
n−1
(t)
¸

,
a vector-valued function of one variable. The initial condition becomes
u(0) = h, where h =

¸
¸
¸
h(x
1
)
h(x
2
)

h(x
n−1
)
¸

.
We can approximate the ﬁrst-order partial derivative by a diﬀerence quo-
tient:
∂u
∂x

x
i
+x
i+1
2
, t

.
=
u
i+1
(t) −u
i
(t)
x
i+1
−x
i
=
[u
i+1
(t) −u
i
(t)]
1/n
= n[u
i+1
(t) −u
i
(t)].
Similarly, we can approximate the second-order partial derivative:

2
u
∂x
2
(x
i
, t)
.
=
∂u
∂x

xi+xi+1
2
, t

∂u
∂x

xi−1+xi
2
, t

1/n
.
= n
¸
∂u
∂x

x
i
+x
i+1
2
, t

∂u
∂x

x
i−1
+x
i
2
, t

.
= n
2
[u
i−1
(t) −2u
i
(t) +u
i+1
(t)].
Thus the partial diﬀerential equation
∂u
∂t
=

2
u
∂x
2
can be approximated by a system of ordinary diﬀerential equations
du
i
dt
= n
2
(u
i−1
−2u
i
+u
i+1
).
This is a ﬁrst order linear system which can be presented in vector form as
du
dt
= n
2
Pu, where P =

¸
¸
¸
¸
¸
−2 1 0 0
1 −2 1 0
0 1 −2

0 0 −2
¸

,
93
the last matrix having n − 1 rows and n − 1 columns. Finally, we can rewrite
this as
du
dt
= Au, where A = n
2
P, (4.16)
a system exactly like the ones studied in Section 2.5. In the limit as n → ∞
one can use the Mathematica program of '2.6 to check that the eigenvalues of A
approach the eigenvalues of d
2
/dx
2
as determined in the preceding section, and
the eigenvectors approximate more and more closely the standard orthonormal
basis of sine functions.
One could continue constructing a numerical method for solution of our ini-
tial value problem by means of another discretization, this time in the time
direction. We could do this via the familiar Cauchy-Euler method for ﬁnding
numerical solutions to the linear system (4.16). This method for ﬁnding ap-
proximate solutions to the heat equation is often called the method of ﬁnite
diﬀerences. With suﬃcient eﬀort, one could construct a computer program,
using Mathematica or some other software package, to implement it.
More advanced courses on numerical analysis often treat the ﬁnite diﬀerence
method in detail.
3
For us, however, the main point of the method of ﬁnite
diﬀerences is that it provides considerable insight into the theory behind the
heat equation. It shows that the heat equation can be thought of as arising
from a system of ordinary diﬀerential equations when the number of dependent
variables goes to inﬁnity. It is sometimes the case that either a partial diﬀeren-
tial equation or a system of ordinary diﬀerential equations with a large or even
inﬁnite number of unknowns can give an eﬀective model for the same physical
phenomenon. This partially explains, for example, why quantum mechanics
possesses two superﬁcially diﬀerent formulations, via Schr¨ odinger’s partial dif-
ferential equation or via “inﬁnite matrices” in Heisenberg’s “matrix mechanics.”
4.4 The vibrating string
Our next goal is to derive the equation which governs the motion of a vibrating
string. We consider a string of length L stretched out along the x-axis, one end
of the string being at x = 0 and the other being at x = L. We assume that the
string is free to move only in the vertical direction. Let
u(x, t) = vertical displacement of the string at the point x at time t.
We will derive a partial diﬀerential equation for u(x, t). Note that since the
ends of the string are ﬁxed, we must have u(0, t) = 0 = u(L, t) for all t.
It will be convenient to use the “conﬁguration space” V
0
described in Sec-
tion 3.3. An element u(x) ∈ V
0
represents a conﬁguration of the string at some
3
For further discussion of this method one can refer to numerical analysis books, such as
Burden and Faires, Numerical analysis, Seventh edition, Brooks Cole Publishing Company,
2000.
94
instant of time. We will assume that the potential energy in the string when it
is in the conﬁguration u(x) is
V (u(x)) =

L
0
T
2

du
dx

2
dx, (4.17)
where T is a constant, called the tension of the string.
Indeed, we could imagine that we have devised an experiment that measures
the potential energy in the string in various conﬁgurations, and has determined
that (4.17) does indeed represent the total potential energy in the string. On
the other hand, this expression for potential energy is quite plausible for the
following reason: We could imagine ﬁrst that the amount of energy in the string
should be proportional to the amount of stretching of the string, or in other
words, proportional to the length of the string. From vector calculus, we know
that the length of the curve u = u(x) is given by the formula
Length =

L
0

1 + (du/dx)
2
dx.
But when du/dx is small,
¸
1 +
1
2

du
dx

2
¸
2
= 1 +

du
dx

2
+ a small error,
and hence

1 + (du/dx)
2
is closely approximated by 1 +
1
2
(du/dx)
2
.
Thus to a ﬁrst order of approximation, the amount of energy in the string should
be proportional to

L
0
¸
1 +
1
2

du
dx

2
¸
dx =

L
0
1
2

du
dx

2
dx + constant.
Letting T denote the constant of proportionality yields
energy in string =

L
0
T
2

du
dx

2
dx + constant.
Potential energy is only deﬁned up to addition of a constant, so we can drop
the constant term to obtain (4.17).
The force acting on a portion of the string when it is in the conﬁguration
u(x) is determined by an element F(x) of V
0
. We imagine that the force acting
on the portion of the string from x to x + dx is F(x)dx. When the force
pushes the string through an inﬁnitesimal displacement ξ(x) ∈ V
0
, the total
95
work performed by F(x) is then the “sum” of the forces acting on the tiny
pieces of the string, in other words, the work is the “inner product” of F and ξ,
'F(x), ξ(x)` =

L
0
F(x)ξ(x)dx.
(Note that the inner product we use here diﬀers from the one used in Section 3.3
by a constant factor.)
On the other hand this work is the amount of potential energy lost when the
string undergoes the displacement:
'F(x), ξ(x)` =

L
0
T
2

∂u
∂x

2
dx −

L
0
T
2

∂(u +ξ)
∂x

2
dx
= −T

L
0
∂u
∂x
∂ξ
∂x
dx +

L
0
T
2

∂ξ
∂x

2
dx.
We are imagining that the displacement ξ is inﬁnitesimally small, so terms
containing the square of ξ or the square of a derivative of ξ can be ignored, and
hence
'F(x), ξ(x)` = −T

L
0
∂u
∂x
∂ξ
∂x
dx.
Integration by parts yields
'F(x), ξ(x)` = T

L
0

2
u
∂x
2
ξ(x)dx −T

∂u
∂x
ξ

(L) −T

∂u
∂x
ξ

(0).
Since ξ(0) = ξ(L) = 0, we conclude that

L
0
F(x)ξ(x)dx = 'F(x), ξ(x)` = T

L
0

2
u
∂x
2
ξ(x)dx.
Since this formula holds for all inﬁnitesimal displacements ξ(x), we must have
F(x) = T

2
u
∂x
2
,
for the force density per unit length.
Now we apply Newton’s second law, force = mass acceleration, to the
function u(x, t). The force acting on a tiny piece of the string of length dx is
F(x)dx, while the mass of this piece of string is just ρdx, where ρ is the density
of the string. Thus Newton’s law becomes
T

2
u
∂x
2
dx = ρdx

2
u
∂t
2
.
If we divide by ρdx, we obtain the wave equation,

2
u
∂t
2
=
T
ρ

2
u
∂x
2
, or

2
u
∂t
2
= c
2

2
u
∂x
2
,
96
where c
2
= T/ρ.
Just as in the preceding section, we could approximate this partial diﬀerential
equation by a system of ordinary diﬀerential equations. Assume that the string
has length L = 1 and set x
i
= i/n and
u
i
(t) = u(x
i
, t) = the displacement of the string at x
i
at time t.
Then the function u(x, t) can be approximated by the vector-valued function
u(t) =

¸
¸
¸
u
1
(t)
u
2
(t)

u
n−1
(t)
¸

of one variable, just as before. The wave equation is then approximated by the
system of ordinary diﬀerential equations
d
2
u
dt
2
= c
2
n
2
Pu,
where P is the (n−1) (n−1) matrix described in the preceding section. Thus
the diﬀerential operator
L =
d
2
dx
2
is approximated by the symmetric matrix n
2
P,
and we expect solutions to the wave equation to behave like solutions to a
mechanical system of weights and springs with a large number of degrees of
freedom.
Exercises:
4.4.1.a. Show that if f : R → R is any well-behaved function of one variable,
u(x, t) = f(x +ct)
is a solution to the partial diﬀerential equation
∂u
∂t
−c
∂u
∂x
= 0.
(Hint: Use the “chain rule.”)
b. Show that if g : R → R is any well-behaved function of one variable,
u(x, t) = g(x −ct)
is a solution to the partial diﬀerential equation
∂u
∂t
+c
∂u
∂x
= 0.
97
c. Show that for any choice of well-behaved functions f and g, the function
u(x, t) = f(x +ct) +g(x −ct)
is a solution to the diﬀerential equation

2
u
∂t
2
−c
2

2
u
∂x
2
=
¸

∂t
+c

∂x

∂u
∂t
−c
∂u
∂x

= 0.
Remark: This gives a very explicit general solution to the equation for the
vibrations of an inﬁnitely long string.
d. Show that
u(x, t) =
f(x +ct) +f(x −ct)
2
is a solution to the initial value problem

2
u
∂t
2
−c
2

2
u
∂x
2
= 0, u(x, 0) = f(x),
∂u
∂t
(x, 0) = 0.
4.4.2. Show that if the tension and density of a string are given by variable
functions T(x) and ρ(x) respectively, then the motion of the string is governed
by the equation

2
u
∂t
2
=
1
ρ(x)

∂x

T(x)
∂u
∂x

.
4.5 The initial value problem for the vibrating
string
The Fourier sine series can also be used to ﬁnd the solution to an initial value
problem for the vibrating string with ﬁxed endpoints at x = 0 and x = L.
We formulate this problem as follows: we seek a function u(x, t), deﬁned for
0 ≤ x ≤ L and t ≥ 0 such that
1. u(x, t) satisﬁes the wave equation

2
u
∂t
2
= c
2

2
u
∂x
2
, (4.18)
where c is a constant.
2. u(x, t) satisﬁes the boundary condition u(0, t) = u(L, t) = 0, i.e. the
displacement of the string is zero at the endpoints.
3. u(x, t) satisﬁes the initial conditions
u(x, 0) = h
1
(x) and
∂u
∂t
(x, 0) = h
2
(x),
where h
1
(x) and h
2
(x) are given functions, the initial position and velocity
of the string.
98
Note that the wave equation itself and the boundary condition are homogeneous
and linear, and therefore satisfy the principal of superposition.
Once again, we ﬁnd the solution to our problem in two steps:
Step I. We ﬁnd all of the solutions to the homogeneous linear conditions of the
special form
u(x, t) = f(x)g(t).
Step II. We ﬁnd the superposition of these solution which satisﬁes the nonho-
mogeneous initial conditions by means of Fourier analysis.
To carry out Step I, we substitute u(x, t) = f(x)g(t) into the wave equation
(4.18) and obtain
f(x)g

(t) = c
2
f

(x)g(t).
We separate variables, putting all the functions involving t on the left, all the
functions involving x on the right:
g

(t)
g(t)
= c
2
f

(x)
f(x)
.
Once again, the left-hand side of this equation does not depend on x, while the
right-hand side does not depend on t, so neither side can depend upon either x
or t. Therefore the two sides must equal a constant λ, and our equation becomes
g

(t)
c
2
g(t)
=
f

(x)
f(x)
= λ,
which separates into two ordinary diﬀerential equations,
g

(t)
c
2
g(t)
= λ, or g

(t) = λc
2
g(t), (4.19)
and
f

(x)
f(x)
= λ, or f

(x) = λf(x). (4.20)
Just as in the case of the heat equation, the homogeneous boundary condition
u(0, t) = u(L, t) = 0 becomes
f(0)g(t) = f(L)g(t) = 0,
and assuming that g(t) is not identically zero, we obtain
f(0) = f(L) = 0.
Thus once again we need to ﬁnd the nontrivial solutions to the boundary value
problem,
f

(x) =
d
2
dx
2
(f(x)) = λf(x), f(0) = 0 = f(L),
99
and just as before, we ﬁnd that the the only nontrivial solutions are constant
multiples of
f(x) = sin(nπx/L), with λ = −(nπ/L)
2
, n = 1, 2, 3, . . . .
For each of these solutions, we need to ﬁnd a corresponding g(t) solving
equation (4.19),
g

(t) = −(nπ/L)
2
c
2
g(t), or g

(t) + (nπ/L)
2
c
2
g(t) = 0.
This is just the equation of simple harmonic motion, and has the general solution
g(t) = a cos(ncπt/L) +b sin(ncπt/L),
where a and b are constants of integration. Thus we ﬁnd that the nontrivial
product solutions to the wave equation together with the homogeneous bound-
ary condition u(0, t) = 0 = u(L, t) are constant multiples of
u
n
(x, t) = [a
n
cos(ncπt/L) +b
n
sin(ncπt/L)] sin(nπx/L).
The general solution to the wave equation together with this boundary condition
is an arbitrary superposition of these product solutions:
u(x, t) = [a
1
cos(cπt/L) +b
1
sin(cπt/L)] sin(πx/L) (4.21)
+[a
2
cos(2cπt/L) +b
2
sin(2cπt/L)] sin(2πx/L) +. . . . (4.22)
The vibration of the string is a superposition of a fundamental mode which has
frequency

L
1

=
c
2L
=

T/ρ
2L
,
and higher modes which have frequencies which are exact integer multiples of
this frequency.
Step II consists of determining the constants a
n
and b
n
in (4.22) so that the
initial conditions
u(x, 0) = h
1
(x) and
∂u
∂t
(x, 0) = h
2
(x)
are satisﬁed. Setting t = 0 in (4.22) yields
h
1
(x) = u(x, 0) = a
1
sin(πx/L) +a
2
sin(2πx/L) +. . . ,
so we see that the a
n
’s are the coeﬃcients in the Fourier sine series of h
1
.
If we diﬀerentiate equation(4.22) with respect to t, we ﬁnd that
∂u
∂t
(x, t) = [
−cπ
L
a
1
sin(cπt/L) +

L
b
1
cos(cπt/L)] sin(πx/L)
100
+
−2cπ
L
[a
2
sin(2cπt/L) +

L
b
2
sin(2cπt/L)] cos(2πx/L) +. . . ,
and setting t = 0 yields
h
2
(x) =
∂u
∂t
(x, 0) =

L
b
1
sin(πx/L) +
2cπ
L
b
2
sin(2πx/L) +. . . .
We conclude that
ncπ
L
b
n
= the n-th coeﬃcient in the Fourier sine series of h
2
(x).
Example. Suppose that we want to ﬁnd the function u(x, t), deﬁned for 0 ≤
x ≤ π and t ≥ 0, which satisﬁes the initial-value problem:

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = 5 sin x + 12 sin 2x + 6 sin 3x,
∂u
∂t
(x, 0) = 0.
In this case, the ﬁrst three coeﬃcients for the Fourier sine series of h are
a
1
= 5, a
2
= 12, a
3
= 6,
and all the others are zero, so the solution must be
u(x, t) = 5 sin xcos t + 12 sin 2xcos 2t + 6 sin 3xcos 3t.
Exercises:
4.5.1 What happens to the frequency of the fundamental mode of oscillation of
a vibrating string when the length of the string is doubled? When the tension
on the string is doubled? When the density of the string is doubled?
4.5.2. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = sin 2x,
∂u
∂t
(x, 0) = 0.
You may assume that the nontrivial solutions to the eigenvalue problem
f

(x) = λf(x), f(0) = 0 = f(π)
are
λ = −n
2
, f(x) = b
n
sin nx, for n = 1, 2, 3, . . . ,
where the b
n
’s are constants.
101
4.5.3. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = sin x + 3 sin 2x −5 sin 3x,
∂u
∂t
(x, 0) = 0.
4.5.4. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = x(π −x),
∂u
∂t
(x, 0) = 0.
4.5.5. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes
the following conditions:

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = 0,
∂u
∂t
(x, 0) = sin x + sin 2x.
4.5.6. (For students with access to Mathematica) a. Find the ﬁrst ten coeﬃ-
cients of the Fourier sine series for
h(x) = x −x
4
by running the following Mathematica program
f[n ] := 2 NIntegrate[(x - x∧4) Sin[n Pi x], ¦x,0,1¦];
b = Table[f[n], ¦n,1,10¦]
b. Find the ﬁrst ten terms of the solution to the initial value problem for a
vibrating string,

2
u
∂t
2
=

2
u
∂x
2
, u(0, t) = u(π, t) = 0,
u(x, 0) = x −x
4
,
∂u
∂t
(x, 0) = 0.
c. Construct a sequence of sketches of the positions of the vibrating string at
the times t
i
= ih, where h = .1 by running the Mathematica program:
vibstring = Table[
102
Plot[
Sum[ b[n] Sin[n Pi x] Cos[n Pi t], ¦n,1,10¦],
¦x,0,1¦, PlotRange -> ¦-1,1¦
], ¦t,0,1.,.1¦
]
d. Select the cell containing vibstring and animate the sequence of graphics
by running “Animate selected graphics,” from the Cell menu.
4.6 Heat ﬂow in a circular wire
The theory of Fourier series can also be used to solve the initial value problem
for the heat equation in a circular wire of radius 1 which is insulated along the
sides. In this case, we seek a function u(θ, t), deﬁned for θ ∈ R and t ≥ 0 such
that
1. u(θ, t) satisﬁes the heat equation
∂u
∂t
= c
2

2
u
∂θ
2
, (4.23)
where c is a constant.
2. u(θ, t) is periodic of period 2π in the variable θ; in other words,
u(θ + 2π, t) = u(θ, t), for all θ and t.
3. u(θ, t) satisﬁes the initial condition u(θ, 0) = h(θ), where h(θ) is a given
function, periodic of period 2π, the initial temperature of the wire.
Once again the heat equation itself and the periodicity condition are homoge-
neous and linear, so they must be dealt with ﬁrst. Once we have found the
general solution to the homogeneous conditions
∂u
∂t
= c
2

2
u
∂θ
2
, u(θ + 2π, t) = u(θ, t),
we will be able to ﬁnd the particular solution which satisﬁes the initial condition
u(θ, 0) = h(θ)
by the theory of Fourier series.
Thus we substitute u(θ, t) = f(θ)g(t) into the heat equation (4.23) to obtain
f(θ)g

(t) = c
2
f

(θ)g(t)
and separate variables:
g

(t)
g(t)
= c
2
f

(θ)
f(θ)
.
103
The left-hand side of this equation does not depend on θ, while the right-hand
side does not depend on t, so neither side can depend upon either θ or t, and
we can write
g

(t)
c
2
g(t)
=
f

(θ)
f(θ)
= λ,
where λ is a constant. We thus obtain two ordinary diﬀerential equations,
g

(t)
c
2
g(t)
= λ, or g

(t) = λc
2
g(t), (4.24)
and
f

(θ)
f(θ)
= λ, or f

(θ) = λf(θ). (4.25)
The periodicity condition u(θ +π, t) = u(θ, t) becomes
f(θ + 2π)g(t) = f(θ)g(t),
and if g(t) is not identically zero, we must have
f(θ + 2π) = f(θ).
Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the
problem requires us to ﬁnd the nontrivial solutions to the problem:
f

(θ) =
d
2

2
(f(θ)) = λf(θ), f(θ + 2π) = f(θ). (4.26)
We will call (4.11) the eigenvalue problem for the diﬀerential operator
L =
d
2

2
acting on the space V of functions which are periodic of period 2π.
As before, we need to consider three cases.
Case 1: λ = 0. In this case, the eigenvalue problem (4.26) becomes
f

(θ) = 0, f(θ + 2π) = f(θ).
The general solution to the diﬀerential equation is f(θ) = a +bθ, and
a +b(θ + 2π) = a +b(θ) ⇒ b = 0.
Thus the only solution in this case is that where f is constant, and to be
consistent with our Fourier series conventions, we write f(θ) = a
0
/2, where a
0
is a constant.
Case 2: λ > 0. In this case, the diﬀerential equation
f

(θ) = λf(θ), or f

(θ) −λf(θ) = 0
104
has the general solution
f(θ) = ae
(

λθ)
+be
−(

λθ)
.
Note that
a = 0 ⇒ f(θ) → ±∞ as θ → ∞,
while
b = 0 ⇒ f(θ) → ±∞ as θ → −∞.
Neither of these is consistent with the periodicity conditions f(θ + 2π) = f(θ),
so we conclude that a = b = 0, and we obtain no nontrivial solutions in this
case.
Case 3: λ < 0. In this case, we set ω =

−λ, and rewrite the eigenvalue
problem as
f

(θ) +ω
2
f(θ) = 0, f(θ + 2π) = f(θ).
We recognize once again our old friend, the diﬀerential equation of simple har-
monic motion, which has the general solution
f(θ) = a cos(ωθ) +b sin(ωθ) = Asin(ω(θ −θ
0
)).
The periodicity condition f(θ + 2π) = f(θ) implies that ω = n, where n is an
integer, which we can assume is positive, and
f(θ) = a
n
cos(nθ) +b
n
sin(nθ).
Thus we see that the general solution to the eigenvalue problem (4.26) is
λ = 0 and f(θ) =
a
0
2
,
or
λ = −n
2
where n is a positive integer and f(θ) = a
n
cos(nθ) +b
n
sin(nθ).
Now we need to ﬁnd the corresponding solutions to (4.24)
g

(t) = λc
2
g(t),
for λ = 0, −1, −4, −9, . . . , −n
2
, . . . . As before, we ﬁnd that the solution is
g(t) = (constant)e
−n
2
c
2
t
,
where c is a constant. Thus the product solutions to the homogeneous part of
the problem are
u
0
(θ, t) =
a
0
2
, u
n
(θ, t) = [a
n
cos(nθ) +b
n
sin(nθ)]e
−n
2
c
2
t
,
where n = 1, 2, 3, . . . .
105
Now we apply the superposition principle—an arbitrary superposition of
these product solutions must again be a solution. Thus
u(θ, t) =
a
0
2
+ Σ

n=1
[a
n
cos(nθ) +b
n
sin(nθ)]e
−n
2
c
2
t
(4.27)
is a periodic solution of period 2π to the heat equation (4.23).
To ﬁnish the solution to our problem, we must impose the initial condition
u(θ, 0) = h(θ).
But setting t = 0 in (4.27) yields
a
0
2
+ Σ

n=1
[a
n
cos(nθ) +b
n
sin(nθ)] = h(θ),
so the constants a
0
, a
1
, . . . , b
1
, . . . are just the Fourier coeﬃcients of h(θ). Thus
the solution to our initial value problem is just (4.27) in which the constants a
k
and b
k
can be determined via the familiar formulae
a
k
=
1
π

π
−π
h(θ) cos kθdθ, b
k
=
1
π

π
−π
h(θ) sin kθdθ.
Note that as t → ∞the temperature in the circular wire approaches the constant
value a
0
/2.
Exercises:
4.6.1. Find the function u(θ, t), deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂θ
2
, u(θ + 2π, t) = u(θ, t), u(θ, 0) = 2 + sin θ −cos 3θ.
You may assume that the nontrivial solutions to the eigenvalue problem
f

(θ) = λf(θ), f(θ + 2π) = f(θ)
are
λ = 0 and f(θ) =
a
0
2
,
and
λ = −n
2
and f(θ) = a
n
cos nθ +b
n
sin nθ, for n = 1, 2, 3, . . . ,
where the a
n
’s and b
n
’s are constants.
4.6.2. Find the function u(θ, t), deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0, which satisﬁes
the following conditions:
∂u
∂t
=

2
u
∂θ
2
, u(θ + 2π, t) = u(θ, t),
106
u(θ, 0) = [θ[, for θ ∈ [−π, π].
4.6.3. Find the function u(θ, t), deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0, which satisﬁes
the following conditions:

2
u
∂t
2
=

2
u
∂θ
2
, u(θ + 2π, t) = u(θ, t),
u(θ, 0) = 2 + sin θ −cos 3θ,
∂u
∂t
(θ, 0) = 0.
4.7 Sturm-Liouville Theory*
We would like to be able to analyze heat ﬂow in a bar even if the speciﬁc
heat σ(x), the density ρ(x) and the thermal conductivity κ(x) vary from point
to point. As we saw in Section 4.1, this leads to consideration of the partial
diﬀerential equation
∂u
∂t
=
1
ρ(x)σ(x)

∂x

κ(x)
∂u
∂x

, (4.28)
where we make the standing assumption that ρ(x), σ(x) and κ(x) are positive.
We imagine that the bar is situated along the x-axis with its endpoints
situated at x = a and x = b. As in the constant coeﬃcient case, we expect that
there should exist a unique function u(x, t), deﬁned for a ≤ x ≤ b and t ≥ 0
such that
1. u(x, t) satisﬁes the heat equation (4.28).
2. u(x, t) satisﬁes the boundary condition u(a, t) = u(b, t) = 0.
3. u(x, t) satisﬁes the initial condition u(x, 0) = h(x), where h(x) is a given
function, deﬁned for x ∈ [a, b], the initial temperature of the bar.
Just as before, we substitute u(x, t) = f(x)g(t) into (4.28) and obtain
f(x)g

(t) =
1
ρ(x)σ(x)
d
dx

κ(x)
df
dx
(x)

g(t).
Once again, we separate variables, putting all the functions involving t on the
left, all the functions involving x on the right:
g

(t)
g(t)
=
1
ρ(x)σ(x)
d
dx

κ(x)
df
dx
(x)

1
f(x)
.
As usual, the two sides must equal a constant, which we denote by λ, and our
equation separates into two ordinary diﬀerential equations,
g

(t) = λg(t), (4.29)
107
and
1
ρ(x)σ(x)
d
dx

κ(x)
df
dx
(x)

= λf(x). (4.30)
Under the assumption that u is not identically zero, the boundary condition
u(a, t) = u(b, t) = 0 yields
f(a) = f(b) = 0.
Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the
problem, we need to ﬁnd the nontrivial solutions to the boundary value problem:
1
ρ(x)σ(x)
d
dx

κ(x)
df
dx
(x)

= λf(x), f(a) = 0 = f(b). (4.31)
We call this the eigenvalue problem or Sturm-Liouville problem for the diﬀeren-
tial operator
L =
1
ρ(x)σ(x)
d
dx

κ(x)
d
dx

,
which acts on the space V
0
of well-behaved functions f : [a, b] → R which
vanish at the endpoints a and b. The eigenvalues of L are the constants λ for
which (4.31) has nontrivial solutions. Given an eigenvalue λ, the corresponding
eigenspace is
W
λ
= ¦f ∈ V
0
: f satisﬁes (4.31)¦.
Nonzero elements of the eigenspaces are called eigenfunctions.
If the functions ρ(x), σ(x) and κ(x) are complicated, it may be impossible to
solve this eigenvalue problem explicitly, and one may need to employ numerical
methods to obtain approximate solutions. Nevertheless, it is reassuring to know
that the theory is quite parallel to the constant coeﬃcient case that we treated
in previous sections. The following theorem, due to the nineteenth century
mathematicians Sturm and Liouville, is proven in more advanced texts:
4
Theorem. Suppose that ρ(x), σ(x) and κ(x) are smooth functions which are
positive on the interval [a, b]. Then all of the eigenvalues of L are negative real
numbers, and each eigenspace is one-dimensional. Moreover, the eigenvalues
can be arranged in a sequence
0 > λ
1
> λ
2
> > λ
n
> ,
with λ
n
→ −∞. Finally, every well-behaved function can be represented on
[a, b] as a convergent sum of eigenfunctions.
Suppose that f
1
(x), f
2
(x), . . . , f
n
(x), . . . are eigenfunctions corresponding to
the eigenvalues λ
1
, λ
2
, . . . , λ
n
, . . . . Then the general solution to the heat
equation (4.28) together with the boundary conditions u(a, t) = u(b, t) = 0 is
u(x, t) =

¸
n=0
c
n
f
n
(x)e
−λnt
,
4
For further discussion, see Boyce and DiPrima, Elementary diﬀerential equations and
boundary value problems, Seventh edition, Wiley, New York, 2001.
108
where the c
n
’s are arbitrary constants.
To determine the c
n
’s in terms of the initial temperature h(x), we need a
generalization of the theory of Fourier series. The key idea here is that the
eigenspaces should be orthogonal with respect to an appropriate inner product.
The inner product should be one which makes L like a symmetric matrix. To
arrange this, the inner product that we need to use on V
0
is the one deﬁned by
the formula
'f, g` =

b
a
ρ(x)σ(x)f(x)g(x)dx.
Lemma. With respect to this inner product, eigenfunctions corresponding to
distinct eigenvalues are perpendicular.
The proof hinges on the fact that
'L(f), g` = 'f, L(g)`, for f, g ∈ V
0
,
so that if we thought of L as represented by a matrix, the matrix would be
symmetric. This identity can be veriﬁed by integration by parts; indeed,
'L(f), g` =

b
a
ρ(x)σ(x)L(f)(x)g(x)dx =

b
a
d
dx

K(x)
df
dx
(x)

g(x)
= −

b
a
K(x)
df
dx
(x)
dg
dx
(x)dx = = 'f, L(g)`,
where the steps represented by dots are just like the ﬁrst steps, but in reverse
order.
It follows that if f
i
(x) and f
j
(x) are eigenfunctions corresponding to distinct
eigenvalues λ
i
and λ
j
, then
λ
i
'f
i
, f
j
` = 'L(f), g` = 'f, L(g)` = λ
j
'f
i
, f
j
`,
and hence

i
−λ
j
)'f
i
, f
j
` = 0.
Since λ
i
−λ
j
= 0, we conclude that f
i
and f
j
are perpendicular with respect to
the inner product ', `, as claimed.
Thus to determine the c
n
’s, we can use exactly the same orthogonality tech-
niques that we have used before. Namely, if we normalize the eigenfunctions
so that they have unit length and are orthogonal to each other with respect to
', , `, then
c
n
= 'h, f
n
`,
or equivalently, c
n
is just the projection of h in the f
n
direction.
Example. We consider the operator
L = x
d
dx

x
d
dx

,
109
which acts on the space V
0
of functions f : [1, e
π
] → R which vanish at the
endpoints of the interval [1, e
π
]. To solve the eigenvalue problem, we need to
ﬁnd the nontrivial solutions to
x
d
dx

x
df
dx
(x)

= λf(x), f(1) = 0 = f(e
π
). (4.32)
We could ﬁnd these nontrivial solutions by using the techniques we have learned
for treating Cauchy-Euler equations.
However, there is a simpler approach, based upon the technique of substitu-
tion. Namely, we make a change of variables x = e
z
and note that since
dx = e
z
dz,
d
dx
=
1
e
z
d
dz
and hence x
d
dx
= e
z
1
e
z
d
dz
=
d
dz
.
Thus if we set
˜
f(z) = f(x) = f(e
z
),
the eigenvalue problem (4.32) becomes
d
2 ˜
f
dz
2
(z) = λ
˜
f(z),
˜
f(0) = 0 =
˜
f(π),
a problem which we have already solved. The nontrivial solutions are
λ
n
= −n
2
,
˜
f
n
(z) = sin nz, where n = 1, 2, . . . .
Thus the eigenvalues for our original problem are
λ
n
= −n
2
, for n = 1, 2, . . . ,
and as corresponding eigenfunctions we can take
f
n
(x) = sin(nlog x),
where log denotes the natural or base e logarithm. The lemma implies that
these eigenfunctions will be perpendicular with respect to the inner product
', `, deﬁned by
'f, g` =

e
π
1
1
x
f(x)g(x)dx. (4.33)
Exercises:
4.7.1. Show by direct integration that if m = n, the functions
f
m
(x) = sin(mlog x) and f
n
(x) = sin(nlog x)
are perpendicular with respect to the inner product deﬁned by (4.33).
110
4.7.2. Find the function u(x, t), deﬁned for 1 ≤ x ≤ e
π
and t ≥ 0, which satisﬁes
the following initial-value problem for a heat equation with variable coeﬃcients:
∂u
∂t
= x

∂x

x
∂u
∂x

, u(1, t) = u(e
π
, t) = 0,
u(x, 0) = 3 sin(log x) + 7 sin(2 log x) −2 sin(3 log x).
4.7.3.a. Find the solution to the eigenvalue problem for the operator
L = x
d
dx

x
d
dx

−3,
which acts on the space V
0
of functions f : [1, e
π
] → R which vanish at the
endpoints of the interval [1, e
π
].
b. Find the function u(x, t), deﬁned for 1 ≤ x ≤ e
π
and t ≥ 0, which satisﬁes
the following initial-value problem for a heat equation with variable coeﬃcients:
∂u
∂t
= x

∂x

x
∂u
∂x

−3u, u(1, t) = u(e
π
, t) = 0,
u(x, 0) = 3 sin(log x) + 7 sin(2 log x) −2 sin(3 log x).
4.8 Numerical solutions to the eigenvalue prob-
lem*
We can also apply Sturm-Liouville theory to study the motion of a string of
variable mass density. We can imagine a violin string stretched out along the
x-axis with endpoints at x = 0 and x = 1 covered with a layer of varnish which
causes its mass density to vary from point to point. We could let
ρ(x) = the mass density of the string at x for 0 ≤ x ≤ 1.
If the string is under constant tension T, its motion might be governed by the
partial diﬀerential equation

2
u
∂t
2
=
T
ρ(x)

2
u
∂x
2
, (4.34)
which would be subject to the Dirichlet boundary conditions
u(0, t) = 0 = u(1, t), for all t ≥ 0. (4.35)
It is natural to try to ﬁnd the general solution to (4.34) and (4.35) by sepa-
ration of variables, letting u(x, t) = f(x)g(t) as usual. Substituting into (4.34)
yields
f(x)g

(t) =
T
ρ(x)
f

(x)g(t), or
g

(t)
g(t)
=
T
ρ(x)
f

(x)
f(x)
.
111
The two sides must equal a constant, denoted by λ, and the partial diﬀerential
equation separates into two ordinary diﬀerential equations,
T
ρ(x)
f

(x) = λf(x), g

(t) = λg(t).
The Dirichlet boundary conditions (4.35) yield f(0) = 0 = f(1). Thus f must
satisfy the eigenvalue problem
L(f) = λf, f(0) = 0 = f(1), where L =
T
ρ(x)
d
2
dx
2
.
Although the names of the functions appearing in L are a little diﬀerent than
those used in the previous section, the same Theorem applies. Thus the eigen-
values of L are negative real numbers and each eigenspace is one-dimensional.
Moreover, the eigenvalues can be arranged in a sequence
0 > λ
1
> λ
2
> > λ
n
> ,
with λ
n
→ −∞. Finally, every well-behaved function can be represented on
[a, b] as a convergent sum of eigenfunctions. If f
1
(x), f
2
(x), . . . , f
n
(x), . . . are
eigenfunctions corresponding to the eigenvalues λ
1
, λ
2
, . . . , λ
n
, . . . . Then the
general solution to (4.34) and (4.35) is
u(x, t) =

¸
n=0
f
n
(x)

a
n
cos(

−λ
n
t) +b
n
sin(

−λ
n
t)

,
where the a
n
’s and b
n
’s are arbitrary constants. Each term in this sum represents
one of the modes of oscillation of the vibrating string.
In constrast to the case of constant density, it is usually not possible to ﬁnd
simple explicit eigenfunctions when the density varies. It is therefore usually
necessary to use numerical methods.
The simplest numerical method is the one outlined in '4.3. For 0 ≤ i ≤ n,
we let x
i
= i/n and
u
i
(t) = u(x
i
, t) = the displacement at x
i
at time t.
Since u
0
(t) = 0 = u
n
(t) by the boundary conditions, the displacement at time
t is approximated by
u(t) =

¸
¸
¸
u
1
(t)
u
2
(t)

u
n−1
(t)
¸

,
a vector-valued function of one variable. The partial derivative

2
u
∂t
2
is approximated by
d
2
u
dt
2
,
112
20 40 60 80 100
-0.14
-0.12
-0.1
-0.08
-0.06
-0.04
-0.02
Figure 4.1: Shape of the lowest mode when ρ = 1/(x +.1).
and as we saw in '4.3, the partial derivative

2
u
∂x
2
is approximated by n
2
Pu,
where
P =

¸
¸
¸
¸
¸
−2 1 0 0
1 −2 1 0
0 1 −2

0 0 −2
¸

.
Finally, the coeﬃcient (T/ρ(x)) can be represented by the diagonal matrix
Q =

¸
¸
¸
¸
¸
T/ρ(x
1
) 0 0 0
0 T/ρ(x
2
) 0 0
0 0 T/ρ(x
3
)

0 0 T/ρ(x
n−1
)
¸

.
Putting all this together, we ﬁnd that our wave equation with variable mass den-
sity is approximated by a second order homogeneous linear system of ordinary
diﬀerential equations
d
2
u
dt
2
= Au, where A = n
2
QP.
The eigenvalues of low absolute value are approximated by the eigenvalues of A,
while the eigenfunctions representing the lowest frequency modes of oscillation
are approximated eigenvectors corresponding to the lowest eigenvalues of A.
For example, we could ask the question: What is the shape of the lowest
mode of oscillation in the case where ρ(x) = 1/(x+.1)? To answer this question,
113
we could utilize the following Mathematica program (which is quite similar to
the one presented in Exercise 2.5.2):
n := 100; rho[x ] := 1/(x + .1);
m := Table[Max[2-Abs[i-j],0], ¦ i,n-1 ¦ ,¦ j,n-1 ¦ ];
p := m - 4 IdentityMatrix[n-1];
q := DiagonalMatrix[Table[(1/rho[i/n]), ¦ i,1,n-1 ¦ ]];
a := n∧2 q.p; eigenvec = Eigenvectors[N[a]];
ListPlot[eigenvec[[n-1]]]
If we run this program we obtain a graph of the shape of the lowest mode,
as shown in Figure 4.1. Note that instead of approximating a sine curve, our
numerical approximation to the lowest mode tilts somewhat to the left.
114
Chapter 5
PDE’s in Higher
Dimensions
5.1 The three most important linear partial dif-
ferential equations
In higher dimensions, the three most important linear partial diﬀerential equa-
tions are Laplace’s equation

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
= 0,
the heat equation
∂u
∂t
= c
2

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2

,
and the wave equation,

2
u
∂t
2
= c
2

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2

,
where c is a nonzero constant. Techniques developed for studying these equa-
tions can often be applied to closely related equations.
Each of these three equations is homogeneous linear, that is each term con-
tains u or one of its derivatives to the ﬁrst power. This ensures that the principle
of superposition will hold,
u
1
and u
2
solutions ⇒ c
1
u
1
+c
2
u
2
is a solution,
for any choice of constants c
1
and c
2
. The principle of superposition is essential
if we want to apply separation of variables and Fourier analysis techniques.
In the ﬁrst few sections of this chapter, we will derive these partial diﬀeren-
tial equations in several physical contexts. We will begin by using the divergence
115
theorem to derive the heat equation, which in turn reduces in the steady-state
case to Laplace’s equation. We then present two derivations of the wave equa-
tion, one for vibrating membranes and one for sound waves. Exactly the same
wave equation also describes electromagnetic waves, gravitational waves, or wa-
ter waves in a linear approximation. It is remarkable that the principles devel-
oped to solve the three basic linear partial diﬀerential equations can be applied
in so many contexts.
In a few cases, it is possible to ﬁnd explicit solutions to these partial dif-
ferential equations under the simplest boundary conditions. For example, the
general solution to the one-dimensional wave equation

2
u
∂t
2
= c
2

2
u
∂x
2
for the vibrations of an inﬁnitely long string, is
u(x, t) = f(x +ct) +g(x −ct),
where f and g are arbitrary well-behaved functions of a single variable.
Slightly more complicated cases require the technique of “separation of vari-
ables” together with Fourier analysis, as we studied before. Separation of vari-
ables reduces these partial diﬀerential equations to linear ordinary diﬀerential
equations, often with variable coeﬃcients. For example, to ﬁnd the explicit so-
lution to the heat equation in a circular room, we will see that it is necessary
to solve Bessel’s equation.
The most complicated cases cannot be solved by separation of variables,
and one must resort to numerical methods, together with suﬃcient theory to
understand the qualitative behaviour of the solutions.
In the rest of this section, we consider our ﬁrst example, the equation gov-
erning heat ﬂow through a region of (x, y, z)-space, under the assumption that
no heat is being created or destroyed within the region. Let
u(x, y, z, t) = temperature at (x, y, z) at time t.
If σ(x, y, z) is the speciﬁc heat at the point (x, y, z) and ρ(x, y, z) is the
density of the medium at (x, y, z), then the heat within a given region D in
(x, y, z)-space is given by the formula
Heat within D =

D
ρ(x, y, z)σ(x, y, z)u(x, y, z, t)dxdydz.
If no heat is being created or destroyed within D, then by conservation of energy,
the rate at which heat leaves D equals minus the rate of change of heat within
D, which is

d
dt
¸
D
ρσudxdydz

= −

D
ρσ
∂u
∂t
dxdydz, (5.1)
by diﬀerentiating under the integral sign.
116
On the other hand, heat ﬂow can be represented by a vector ﬁeld F(x, y, z, t)
which points in the direction of greatest decrease of temperature,
F(x, y, z, t) = −κ(x, y, z)(∇u)(x, y, z, t),
where κ(x, y, z) is the so-called thermal conductivity of the medium at (x, y, z).
Thus the rate at which heat leaves the region D through a small region in its
boundary of area dA is
−(κ∇u) NdA,
where N is the unit normal which points out of D. The total rate at which heat
leaves D is given by the ﬂux integral

∂D
(κ∇u) NdA,
where ∂D is the surface bounding D. It follows from the divergence theorem
that
Rate at which heat leaves D = −

D
∇ (κ∇u)dxdydz. (5.2)
From formulae (5.1) and (5.2), we conclude that

D
ρσ
∂u
∂t
dxdydz =

D
∇ (κ∇u)dxdydz.
This equation is true for all choices of the region D, so the integrands on the
two sides must be equal:
ρ(x, y, z)σ(x, y, z)
∂u
∂t
(x, y, z, t) = ∇ (κ∇u)(x, y, z, t).
Thus we ﬁnally obtain the heat equation
∂u
∂t
=
1
ρ(x, y, z)σ(x, y, z)
∇ (κ(x, y, z)(∇u)) .
In the special case where the region D is homogeneous, i.e. its properties are
the same at every point, ρ(x, y, z), σ(x, y, z) and κ(x, y, z) are constants, and
the heat equation becomes
∂u
∂t
=
κ
ρσ
¸

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2

.
If we wait long enough, so that the temperature is no longer changing, the
“steady-state” temperature u(x, y, z) must satisfy Laplace’s equation

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
= 0.
117
If the temperature is independent of z, the function u(x, y) = u(x, y, z) must
satisfy the two-dimensional Laplace equation

2
u
∂x
2
+

2
u
∂y
2
= 0.
Exercises:
5.1.1. For which of the following diﬀerential equations is it true that the super-
position principle holds?
a.

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
= 0.
b.

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
+u = 0.
c.

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
+u
2
= 0.
d.

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
= e
x
.
e.

2
u
∂x
2
+

2
u
∂y
2
+

2
u
∂z
2
= e
x
u.
5.1.2. Suppose that a chemical reaction creates heat at the rate
λ(x, y, z)u(x, y, z, t) +ν(x, y, z),
per unit volume. Show that in this case the equation governing heat ﬂow is
ρ(x, y, z)σ(x, y, z)
∂u
∂t
= λ(x, y, z)u(x, y, z, t) +ν(x, y, z) +∇ (κ(x, y, z)(∇u)) .
5.2 The Dirichlet problem
The reader will recall that the space of solutions to a homogeneous linear second
order ordinary diﬀerential equation, such as
d
2
u
dt
2
+p
du
dt
+qu = 0
is two-dimensional, a particular solution being determined by two constants. By
contrast, the space of solutions to Laplace’s partial diﬀerential equation

2
u
∂x
2
+

2
u
∂y
2
= 0 (5.3)
118
is inﬁnite-dimensional. For example, the function u(x, y) = x
3
− 3xy
2
is a
solution to Laplace’s equation, because
∂u
∂x
= 3x
2
−3y
2
,

2
u
∂x
2
= 6x,
∂u
∂y
= −6xy,

2
u
∂y
2
= −6x,
and hence

2
u
∂x
2
+

2
u
∂y
2
= 6x −6x = 0.
Similarly,
u(x, y) = 7, u(x, y) = x
4
−6x
2
y
2
+y
4
, and u(x, y) = e
x
sin y
are solutions to Laplace’s equation. A solution to Laplace’s equation is called
a harmonic function. It is not diﬃcult to construct inﬁnitely many linearly
independent harmonic functions of two variables.
Thus to pick out a particular solution to Laplace’s equation, we need bound-
ary conditions which will impose inﬁnitely many constraints. To see what
boundary conditions are natural to impose, we need to think of a physical
problem which leads to Laplace’s equation. Suppose that u(x, y) represents
the steady-state distribution of temperature throughout a uniform slab in the
shape of a region D in the (x, y)-plane. If we specify the temperature on the
boundary of the region, say by setting up heaters and refrigerators controlled by
thermostats along the boundary, we might expect that the temperature inside
the room would be uniquely determined. We need inﬁnitely many heaters and
refrigerators because there are inﬁnitely many points on the boundary. Spec-
ifying the temperature at each point of the boundary imposes inﬁnitely many
constraints on the harmonic function which realizes the steady-state tempera-
ture within D.
The Dirichlet Problem for Laplace’s Equation. Let D be a bounded
region in the (x, y)-plane which is bounded by a curve ∂D, and let φ : ∂D → R
be a continuous function. Find a harmonic function u : D → R such that
u(x, y) = φ(x, y), for (x, y) ∈ ∂D.
Our physical intuition suggests that the Dirichlet problem will always have a
unique solution. This is proven mathematically for many choices of boundary
in more advanced texts on complex variables and partial diﬀerential equations.
The mathematical proof that a unique solution exists provides evidence that
the mathematical model we have constructed for heat ﬂow may in fact be valid.
Our goal here is to ﬁnd the explicit solutions in the case where the region D
is suﬃciently simple. Suppose, for example, that
D = ¦(x, y) ∈ R
2
: 0 ≤ x ≤ a, 0 ≤ y ≤ b¦.
119
Suppose, moreover that the function φ : ∂D → R vanishes on three sides of ∂D,
so that
φ(0, y) = φ(a, y) = φ(x, 0) = 0,
while
φ(x, b) = f(x),
where f(x) is a given continuous function which vanishes when x = 0 and x = a.
In this case, we seek a function u(x, y), deﬁned for 0 ≤ x ≤ a and 0 ≤ y ≤ b,
such that
1. u(x, y) satisﬁes Laplace’s equation

2
u
∂x
2
+

2
u
∂y
2
= 0. (5.4)
2. u(x, y) satisﬁes the homogeneous boundary conditions u(0, y) = u(a, y) =
u(x, 0) = 0.
3. u(x, y) satisﬁes the nonhomogeneous boundary condition u(x, b) = f(x),
where f(x) is a given function.
Note that the Laplace equation itself and the homogeneous boundary con-
ditions satisfy the superposition principle—this means that if u
1
and u
2
satisfy
these conditions, so does c
1
u
1
+c
2
u
2
, for any choice of constants c
1
and c
2
.
Our method for solving the Dirichlet problem consists of two steps:
Step I. We ﬁnd all of the solutions to Laplace’s equation together with the
homogeneous boundary conditions which are of the special form
u(x, y) = X(x)Y (y).
By the superposition principle, an arbitrary linear superposition of these solu-
tions will still be a solution.
Step II. We ﬁnd the particular solution which satisﬁes the nonhomogeneous
boundary condition by Fourier analysis.
To carry out Step I, we substitute u(x, y) = X(x)Y (y) into Laplace’s equation
(5.3) and obtain
X

(x)Y (y) +X(x)Y

(y) = 0.
Next we separate variables, putting all the functions involving x on the left, all
the functions involving y on the right:
X

(x)
X(x)
= −
Y

(y)
Y (y)
.
The left-hand side of this equation does not depend on y, while the right-hand
side does not depend on x. Hence neither side can depend upon either x or y.
120
In other words, the two sides must equal a constant, which we denote by λ, and
call the separating constant, as before. Our equation now becomes
X

(x)
X(x)
= −
Y

(y)
Y (y)
= λ,
which separates into two ordinary diﬀerential equations,
X

(x) = λX(x), (5.5)
and
Y

(y) = −λY (y). (5.6)
The homogeneous boundary condition u(0, y) = u(a, y) = 0 imply that
X(0)Y (y) = X(a)Y (y) = 0.
If Y (y) is not identically zero,
X(0) = X(a) = 0.
Thus we need to ﬁnd the nontrivial solutions to a boundary value problem for
an ordinary diﬀerential equation:
X

(x) =
d
2
dx
2
(X(x)) = λX(x), X(0) = 0 = X(a),
which we recognize as the eigenvalue problem for the diﬀerential operator
L =
d
2
dx
2
acting on the space V
0
of functions which vanish at 0 and a. We have seen before
that the only nontrivial solutions to equation (5.5) are constant multiples of
X(x) = sin(nπx/a), with λ = −(nπ/a)
2
, n = 1, 2, 3, . . . .
For each of these solutions, we need to ﬁnd a corresponding Y (y) solving
equation (5.6),
Y

(y) = −λY (y),
where λ = −(nπ/a)
2
, together with the boundary condition Y (0) = 0. The
diﬀerential equation has the general solution
Y (y) = Acosh(nπy/a) +Bsinh(nπy/a),
where A and B are constants of integration, and the boundary condition Y (0) =
0 implies that A = 0. Thus we ﬁnd that the nontrivial product solutions to
Laplace’s equation together with the homogeneous boundary conditions are con-
stant multiples of
u
n
(x, y) = sin(nπx/a) sinh(nπy/a).
121
The general solution to Laplace’s equation with these boundary conditions is a
general superposition of these product solutions:
u(x, y) = B
1
sin(πx/a) sinh(πy/a) +B
2
sin(2πx/a) sinh(2πy/a) +. . . . (5.7)
To carry out Step II, we need to determine the constants B
1
, B
2
, . . . which
occur in (5.7) so that
u(x, b) = f(x).
Substitution of y = b into (5.7) yields
f(x) = B
1
sin(πx/a) sinh(πb/a) +B
2
sin(2πx/a) sinh(2πb/a)+
. . . +B
k
sin(2πk/a) sinh(kπb/a) +. . . .
We see that B
k
sinh(kπb/a) is the k-th coeﬃcient in the Fourier sine series for
f(x).
For example, if a = b = π, and f(x) = 3 sin x + 7 sin 2x, then we must have
f(x) = B
1
sin(x) sinh(π) +B
2
sin(2x) sinh(2π),
and hence
B
1
=
3
sinh(π)
, B
2
=
7
sinh(2π)
, B
k
= 0 for k = 3, . . . .
Thus the solution to Dirichlet’s problem in this case is
u(x, y) =
3
sinh(π)
sin xsinh y +
7
sinh(2π)
sin 2xsinh 2y.
Exercises:
5.2.1. Which of the following functions are harmonic?
a. f(x, y) = x
2
+y
2
.
b. f(x, y) = x
2
−y
2
.
c. f(x, y) = e
x
cos y.
d. f(x, y) = x
3
−3xy
2
.
5.2.2. a. Solve the following Dirichlet problem for Laplace’s equation in a square
region: Find u(x, y), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂x
2
+

2
u
∂y
2
= 0, u(0, y) = u(π, y) = 0,
u(x, 0) = 0, u(x, π) = sin x −2 sin 2x + 3 sin 3x.
122
0
1
2
3 0
1
2
3
-10
-5
0
5
10
0
1
2
3
Figure 5.1: Graph of u(x, y) =
3
sinh(π)
sin xsinh y +
7
sinh(2π)
sin 2xsinh 2y.
b. Solve the following Dirichlet problem for Laplace’s equation in the same
square region: Find u(x, y), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂x
2
+

2
u
∂y
2
= 0, u(0, y) = 0,
u(π, y) = sin 2y + 3 sin 4y, u(x, 0) = 0 = u(x, π).
c. By adding the solutions to parts a and c together, ﬁnd the solution to the
Dirichlet problem: Find u(x, y), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂x
2
+

2
u
∂y
2
= 0, u(0, y) = 0,
u(π, y) = sin 2y+3 sin 4y, u(x, 0) = 0, u(x, π) = sin x−2 sin 2x+3 sin 3x.
5.3 Initial value problems for heat equations
The physical interpretation behind the heat equation suggests that the following
initial value problem should have a unique solution:
Let D be a bounded region in the (x, y)-plane which is bounded by a piece-
wise smooth curve ∂D, and let h : D → R be a continuous function which
vanishes on ∂D. Find a function u(x, y, t) such that
123
1. u satisﬁes the heat equation
∂u
∂t
= c
2

2
u
∂x
2
+

2
u
∂y
2

. (5.8)
2. u satisﬁes the “Dirichlet boundary condition” u(x, y, t) = 0, for (x, y) ∈
∂D.
3. u satisﬁes the initial condition u(x, y, 0) = h(x, y).
The ﬁrst two of these conditions are homogeneous and linear, so our strategy is
to treat them ﬁrst by separation of variables, and then use Fourier analysis to
satisfy the last condition.
In this section, we consider the special case where
D = ¦(x, y) ∈ R
2
: 0 ≤ x ≤ a, 0 ≤ y ≤ b¦,
so that the Dirichlet boundary condition becomes
u(0, y, t) = u(a, y, t) = u(x, 0, t) = u(x, b, t) = 0.
In this case, separation of variables is done in two stages. First, we write
u(x, y, t) = f(x, y)g(t),
and substitute into (5.8) to obtain
f(x, y)g

(t) = c
2

2
f
∂x
2
+

2
f
∂y
2

g(t).
Then we divide by c
2
f(x, y)g(t),
1
c
2
g(t)
g

(t) =
1
f(x, y)

2
f
∂x
2
+

2
f
∂y
2

.
The left-hand side of this equation does not depend on x or y while the right
hand side does not depend on t. Hence neither side can depend on x, y, or t, so
both sides must be constant. If we let λ denote the constant, we obtain
1
c
2
g(t)
g

(t) = λ =
1
f(x, y)

2
f
∂x
2
+

2
f
∂y
2

.
This separates into an ordinary diﬀerential equation
g

(t) = c
2
λg(t), (5.9)
and a partial diﬀerential equation

2
f
∂x
2
+

2
f
∂y
2

= λf, (5.10)
124
called the Helmholtz equation. The Dirichlet boundary condition yields
f(0, y) = f(a, y) = f(x, 0) = f(x, b) = 0.
In the second stage of separation of variables, we set f(x, y) = X(x)Y (y)
and substitute into (5.10) to obtain
X

(x)Y (y) +X(x)Y

(y) = λX(x)Y (y),
which yields
X

(x)
X(x)
+
Y

(y)
Y (y)
= λ,
or
X

(x)
X(x)
= λ −
Y

(y)
Y (y)
.
The left-hand side depends only on x, while the right-hand side depends only
on y, so both sides must be constant,
X

(x)
X(x)
= µ = λ −
Y

(y)
Y (y)
.
Hence the Helmholtz equation divides into two ordinary diﬀerential equations
X

(x) = µX(x), Y

(y) = νY (y), where µ +ν = λ.
The “Dirichlet boundary conditions” now become conditions on X(x) and Y (y):
X(0) = X(a) = 0, Y (0) = Y (b) = 0.
The only nontrivial solutions are
X(x) = sin(mπx/a), with µ = −(mπ/a)
2
, m = 1, 2, 3, . . . ,
and
Y (y) = sin(nπy/b), with ν = −(nπ/b)
2
, n = 1, 2, 3, . . . .
The corresponding solutions of the Helmholtz equation are
f
mn
(x, y) = sin(mπx/a) sin(nπy/b), with λ
mn
= −(mπ/a)
2
−(nπ/b)
2
.
For any given choice of m and n, the corresponding solution to (5.9) is
g(t) = e
−c
2
λmnt
= e
−c
2
((mπ/a)
2
+(nπ/b)
2
)t
.
Hence for each choice of m and n, we obtain a product solution to the heat
equation with Dirichlet boundary condition:
u
m,n
(x, y, t) = sin(mπx/a) sin(nπy/b)e
−c
2
((mπ/a)
2
+(nπ/b)
2
)t
.
125
The general solution to the heat equation with Dirichlet boundary conditions is
an arbitrary superposition of these product solutions,
u(x, y, t) =

¸
m,n=1
b
mn
sin(mπx/a) sin(nπy/b)e
−c
2
((mπ/a)
2
+(nπ/b)
2
)t
. (5.11)
To ﬁnd the constants b
mn
appearing in (5.11), we need to apply the initial
condition u(x, y, 0) = h(x, y). The result is
h(x, y) =

¸
m,n=1
b
mn
sin(mπx/a) sin(nπy/b), (5.12)
expressing the fact that the b
mn
’s are the coeﬃcients of what is called the double
Fourier sine series of h. As in the case of ordinary Fourier sine series, the b
mn
’s
can be determined by an explicit formula. To determine it, we multiply both
sides of (5.12) by (2/a) sin(pπx/a), where p is a positive integer, and integrate
with respect to x to obtain
2
a

a
0
h(x, y) sin(pπx/a)dx
=

¸
m,n=1
b
mn
¸
2
a

a
0
sin(pπx/a) sin(mπx/a)dx

sin(nπy/b).
The expression within brackets is one if p = m and otherwise zero, so
2
a

a
0
h(x, y) sin(pπx/a)dx =

¸
n=1
b
pn
sin(nπy/b).
Next we multiply by (2/b) sin(qπy/b), where q is a positive integer, and integrate
with respect to y to obtain
2
a
2
b

a
0

b
0
h(x, y) sin(pπx/a) sin(qπy/b)dxdy
=

¸
n=1
b
pn
¸
2
b

b
0
sin(nπy/b) sin(qπy/b)dy
¸
.
The expression within brackets is one if q = n and otherwise zero, so we ﬁnally
obtain
b
pq
=
2
a
2
b

a
0

b
0
h(x, y) sin(pπx/a) sin(qπy/b)dxdy. (5.13)
Suppose, for example, that c = 1, a = b = π and
h(x, y) = sin xsin y + 3 sin 2xsin y + 7 sin 3xsin 2y.
126
In this case, we do not need to carry out the integration indicated in (5.13)
because comparison with (5.12) shows that
b
11
= 1, b
21
= 3, b
32
= 7,
and all the other b
mn
’s must be zero. Thus the solution to the initial value
problem in this case is
u(x, y, t) = sin xsin ye
−2t
+ 3 sin 2xsin ye
−5t
+ 7 sin 3xsin 2ye
−13t
.
Here is another example. Suppose that a = b = π and
h(x, y) = p(x)q(y),
where
p(x) =

x, for 0 ≤ x ≤ π/2,
π −x, for π/2 ≤ x ≤ π,
q(y) =

y, for 0 ≤ y ≤ π/2,
π −y, for π/2 ≤ y ≤ π.
In this case,
b
mn
=

2
π

2

π
0
¸
π
0
p(x)q(y) sin mxdx

sin nydy
=

2
π

2
¸
π
0
p(x) sin mxdx
¸
π
0
q(y) sin nydy

=

2
π

2
¸

π/2
0
xsin mxdx +

π
π/2
(π −x) sin mxdx
¸
¸

π/2
0
y sin nydy +

π
π/2
(π −y) sin nydy
¸
.
The integration can be carried out just like we did in Section 3.3 to yield
b
mn
=

2
π

2
¸
2
m
2
sin(mπ/2)
¸
2
n
2
sin(nπ/2)

=
16
π
2
¸
1
m
2
1
n
2
sin(mπ/2) sin(nπ/2)

.
Thus in this case, we see that
h(x, y) =
16
π
2

¸
m,n=1
¸
1
m
2
1
n
2
sin(mπ/2) sin(nπ/2)

sin mxsin my,
and hence
u(x, y, t) =
16
π
2

¸
m,n=1
¸
1
m
2
1
n
2
sin(mπ/2) sin(nπ/2)

sin mxsin mye
−(m
2
+n
2
)t
.
127
Exercises:
5.3.1. Solve the following initial value problem for the heat equation in a square
region: Find u(x, y, t), where 0 ≤ x ≤ π, 0 ≤ y ≤ π and t ≥ 0, such that
∂u
∂t
=

2
u
∂x
2
+

2
u
∂y
2
,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, y, 0) = 2 sin xsin y + 5 sin 2xsin y.
You may assume that the nontrivial solutions to the eigenvalue problem

2
f
∂x
2
(x, y) +

2
f
∂y
2
(x, y) = λf(x, y), f(x, 0) = f(x, π) = f(0, y) = f(π, y) = 0,
are of the form
λ = −m
2
−n
2
, f(x, y) = b
mn
sin mxsin ny,
for m = 1, 2, 3, . . . and n = 1, 2, 3, . . . , where b
mn
is a constant.
5.3.2. Solve the following initial value problem for the heat equation in a square
region: Find u(x, y, t), where 0 ≤ x ≤ π, 0 ≤ y ≤ π and t ≥ 0, such that
∂u
∂t
=

2
u
∂x
2
+

2
u
∂y
2
,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, y, 0) = 2(sin x)y(π −y).
5.3.3. Solve the following initial value problem in a square region: Find u(x, y, t),
where 0 ≤ x ≤ π, 0 ≤ y ≤ π and t ≥ 0, such that
1
2t + 1
∂u
∂t
=

2
u
∂x
2
+

2
u
∂y
2
,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, y, 0) = 2 sin xsin y + 3 sin 2xsin y.
5.3.4. Find the general solution to the heat equation
∂u
∂t
=

2
u
∂x
2
+

2
u
∂y
2
subject to the boundary conditions
u(x, 0, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, π, t) = sin x −2 sin 2x + 3 sin 3x.
128
5.4 Two derivations of the wave equation
We have already seen how the one-dimensional wave equation describes the
motion of a vibrating string. In this section we will show that the motion of a
vibrating membrane is described by the two-dimensional wave equation, while
sound waves are described by a three-dimensional wave equation. In fact, we
will see that sound waves arise from “linearization” of the nonlinear equations
of ﬂuid mechanics.
The vibrating membrane. Suppose that a homogeneous membrane is fas-
tened down along the boundary of a region D in the (x, y)-plane. Suppose,
moreover, that a point on the membrane can move only in the vertical direc-
tion, and let u(x, y, t) denote the height of the point with coordinates (x, y) at
time t.
If ρ denotes the density of the membrane (assumed to be constant), then
by Newton’s second law, the force F acting on a small rectangular piece of
the membrane located at (x, y) with sides of length dx and dy is given by the
expression
F = ρ

2
u
∂t
2
(x, y)dxdyk.
Suppose the force displaces the membrane from a given position u(x, y) to a
new position
u(x, y) +η(x, y),
where η(x, y) and its derivatives are very small. Then the total work performed
by the force F will be
F η(x, y)k = η(x, y)ρ

2
u
∂t
2
(x, y)dxdy.
Integrating over the membrane yields an expression for the total work performed
when the membrane moves through the displacement η:
Work =

D
η(x, y)ρ

2
u
∂t
2
(x, y)dxdy. (5.14)
On the other hand, the potential energy stored in the membrane is propor-
tional to the extent to which the membrane is stretched. Just as in the case of
the vibrating string, this stretching is approximated by the integral
Potential energy =
T
2

D
¸

∂u
∂x

2
+

∂u
∂y

2
¸
dxdy,
where T is a constant, called the tension in the membrane. Replacing u by u+η
in this integral yields
New potential energy =
T
2

D
¸

∂u
∂x

2
+

∂u
∂y

2
¸
dxdy
129
+T

D
¸
∂u
∂x

∂η
∂x

+

∂u
∂y

∂η
∂y

dxdy
+
T
2

D
¸

∂η
∂x

2
+

∂η
∂y

2
¸
dxdy.
If we neglect the last term in this expression (which is justiﬁed if η and its
derivatives are assumed to be small), we ﬁnd that
New potential energy −Old potential energy =

D
T∇u ∇ηdxdy.
It follows from the divergence theorem in the plane and the fact that η vanishes
on the boundary ∂D that

D
T∇u ∇ηdxdy +

D
Tη∇ ∇udxdy =

∂D
Tη∇u Nds = 0,
and hence
Change in potential = −

D
η(x, y)T(∇ ∇u)(x, y)dxdy. (5.15)
The work performed must be minus the change in potential energy, so it
follows from (5.14) and (5.15) that

D
η(x, y)ρ

2
u
∂t
2
(x, y)dxdy =

D
η(x, y)T(∇ ∇u)(x, y)dxdy.
Since this equation holds for all choices of η, it follows that
ρ

2
u
∂t
2
= T∇ ∇u,
which simpliﬁes to

2
u
∂t
2
= c
2
∇ ∇u, where c
2
=
T
ρ
,
or equivalently

2
u
∂t
2
= c
2

2
u
∂x
2
+

2
u
∂y
2

.
This is just the wave equation.
The equations of a perfect gas and sound waves. Next, we will describe
Euler’s equations for a perfect gas in (x, y, z)-space.
1
Euler’s equations are
expressed in terms of the quantities,
v(x, y, z, t) = (velocity of the gas at (x, y, z) at time t),
1
For a complete derivation of these equations, see Chapter 9 of Alexander Fetter and John
Walecka, Theoretical mechanics of particles and continua, McGraw-Hill, New York, 1980, or
Chapter 1 of Alexandre Chorin and Jerrold Marsden, A mathematical introduction to ﬂuid
mechanics, third edition, Springer, 1993.
130
ρ(x, y, z, t) = (density at (x, y, z) at time t),
p(x, y, z, t) = (pressure at (x, y, z) at time t).
The ﬁrst of the Euler equations is the equation of continuity,
∂ρ
∂t
+∇ (ρv) = 0. (5.16)
To derive this equation, we represent the ﬂuid ﬂow by the vector ﬁeld
F = ρv,
so that the surface integral

S
F NdA
represents the rate at which the ﬂuid is ﬂowing across S in the direction of N.
We assume that no ﬂuid is being created or destroyed. Then the rate of change
of the mass of ﬂuid within D is given by two expressions,

D
∂ρ
∂t
(x, y, z, t)dxdydz and −

S
F NdA,
which must be equal. It follows from the divergence theorem that the second of
these expressions is

D
∇ F(x, y, z, t)dxdydz,
and hence

D
∂ρ
∂t
dxdydz = −

D
∇ Fdxdydz.
Since this equation must hold for every region D in (x, y, z)-space, we conclude
that the integrands must be equal,
∂ρ
∂t
= −∇ F = −∇ (ρv),
which is just the equation of continuity.
The second of the Euler equations is simply Newton’s second law of motion,
(mass density)(acceleration) = (force density).
We make an assumption that the only force acting on a ﬂuid element is due to
the pressure, an assumption which is not unreasonable in the case of a perfect
gas. In this case, it turns out that the pressure is deﬁned in such a way that
the force acting on a ﬂuid element is minus the gradient of pressure:
Force = −∇p(x(t), y(t), z(t), t). (5.17)
131
The familiar formula Force = Mass Acceleration then yields
ρ
d
dt
(v(x(t), y(t), z(t), t)) = −∇p(x(t), y(t), z(t), t).
Using the chain rule,
dv
dt
=
∂v
∂x
dx
dt
+
∂v
∂y
dy
dt
+
∂v
∂z
dz
dt
+
∂v
∂t
dt
dt
,
we can rewrite this equation as
∂v
∂t
+ (v ∇)v = −
1
ρ
∇p. (5.18)
Note that this equation is nonlinear because of the term (v ∇)v.
To ﬁnish the Euler equations, we need an equation of state, which relates
pressure and density. The equation of state could be determined by experiment,
the simplest equation of state being
p = a
2
ρ
γ
, (5.19)
where a
2
and γ are constants. (An ideal monatomic gas has this equation of
state with γ = 5/3.)
The Euler equations (5.16), (5.18), and (5.19) are nonlinear, and hence quite
diﬃcult to solve. However, one explicit solution is the case where the ﬂuid is
motionless,
ρ = ρ
0
, p = p
0
, v = 0,
where ρ
0
and p
0
satisfy
p
0
= a
2
ρ
γ
0
.
Linearizing Euler’s equations near this explicit solution gives rise to the linear
diﬀerential equation which governs propagation of sound waves.
Let us write
ρ = ρ
0

, p = p
0
+p

, v = v

,
where ρ

, p

and v

are so small that their squares can be ignored.
Substitution into Euler’s equations yields
∂ρ

∂t

0
∇ (v

) = 0,
∂v

∂t
= −
1
ρ
0
∇p

,
and
p

= [a
2
γ(ρ
0
)
(γ−1)

= c
2
ρ

,
where c
2
is a new constant. It follows from these three equations that

2
ρ

∂t
2
= −ρ
0

∂v

∂t

= ∇ ∇p

= c
2
∇ ∇ρ

.
132
Thus ρ

must satisfy the three-dimensional wave equation

2
ρ

∂t
2
= c
2
∇ ∇ρ

= c
2

2
ρ

∂x
2
+

2
ρ

∂y
2
+

2
ρ

∂z
2

. (5.20)
If the sound wave ρ

is independent of z, (5.20) reduces to

2
ρ

∂t
2
= c
2

2
ρ

∂x
2
+

2
ρ

∂y
2

,
exactly the same equation that we obtained for the vibrating membrane.
Remark: The notion of linearization is extremely powerful because it enables
us to derive information on the behavior of solutions to the nonlinear Euler
equations, which are extremely diﬃcult to solve except for under very special
circumstances.
The Euler equations for a perfect gas and the closely related Navier-Stokes
equations for an incompressible ﬂuid such as water form basic models for ﬂuid
mechanics. In the case of incompressible ﬂuids, the density is constant, so no
equation of state is assumed. To allow for viscosity, one adds an additional term
to the expression (5.17) for the force acting on a ﬂuid element:
Force = ν(∆v)(x, y, z, t) −∇p(x(t), y(t), z(t), t).
Here the Laplace operator is applied componentwise and ν is a constant, called
the viscosity of the ﬂuid. The equations used by Navier and Stokes to model an
incompressible viscous ﬂuid (with ρ constant) are then
∇ v = 0,
∂v
∂t
+ (v ∇)v = ν∆v −
1
ρ
∇p.
It is remarkable that these equations, so easily expressed, are so diﬃcult to
solve. Indeed, the Navier-Stokes equations form the basis for one of the seven
Millenium Prize Problems, singled out by the Clay Mathematics Institute as
central problems for mathematics at the turn of the century. If you can show
that under reasonable initial conditions, the Navier-Stokes equations possess a
unique well-behaved solution, you may be able to win one million dollars. To
ﬁnd more details on the prize oﬀered for a solution, you can consult the web
Exercise:
5.4.1. Show that if the tension and density of a membrane are given by vari-
able functions T(x, y) and ρ(x, y) respectively, then the motion of the string is
governed by the equation

2
u
∂t
2
=
1
ρ(x, y)
∇ (T(x, y)∇u) .
133
5.5 Initial value problems for wave equations
The most natural initial value problem for the wave equation is the following:
Let D be a bounded region in the (x, y)-plane which is bounded by a piece-
wise smooth curve ∂D, and let h
1
, h
2
: D → R be continuous functions. Find a
function u(x, y, t) such that
1. u satisﬁes the wave equation

2
u
∂t
2
= c
2

2
u
∂x
2
+

2
u
∂y
2

. (5.21)
2. u satisﬁes the “Dirichlet boundary condition” u(x, y, t) = 0, for (x, y) ∈
∂D.
3. u satisﬁes the initial condition
u(x, y, 0) = h
1
(x, y),
∂u
∂t
(x, y, 0) = h
2
(x, y).
Solution of this initial value problem via separation of variables is very similar
to the solution of the initial value problem for the heat equation which was
presented in Section 5.3.
As before, let us suppose that D = ¦(x, y) ∈ R
2
: 0 ≤ x ≤ a, 0 ≤ y ≤ b¦, so
that the Dirichlet boundary condition becomes
u(0, y, t) = u(a, y, t) = u(x, 0, t) = u(x, b, t) = 0.
We write
u(x, y, t) = f(x, y)g(t),
and substitute into (5.21) to obtain
f(x, y)g

(t) = c
2

2
f
∂x
2
+

2
f
∂y
2

g(t).
Then we divide by c
2
f(x, y)g(t),
1
c
2
g(t)
g

(t) =
1
f(x, y)

2
f
∂x
2
+

2
f
∂y
2

.
Once again, we conclude that both sides must be constant. If we let λ denote
the constant, we obtain
1
c
2
g(t)
g

(t) = λ =
1
f(x, y)

2
f
∂x
2
+

2
f
∂y
2

.
This separates into an ordinary diﬀerential equation
g

(t) = c
2
λg(t), (5.22)
134
and the Helmholtz equation

2
f
∂x
2
+

2
f
∂y
2

= λf. (5.23)
The Dirichlet boundary condition becomes
f(0, y) = f(a, y) = f(x, 0) = f(x, b) = 0.
The Helmholtz equation is solved in exactly the same way as before, the
only nontrivial solutions being
f
mn
(x, y) = sin(mπx/a) sin(nπy/b), with λ
mn
= −(mπ/a)
2
−(nπ/b)
2
.
The corresponding solution to (5.22) is
g(t) = Acos(ω
mn
t) +Bsin(ω
mn
t), where ω
mn
= c

−λ
mn
.
Thus we obtain a product solution to the wave equation with Dirichlet boundary
conditions:
u(x, y, t) = sin(mπx/a) sin(nπy/b)[Acos(ω
mn
t) +Bsin(ω
mn
t)].
The general solution to to the wave equation with Dirichlet boundary conditions
is a superposition of these product solutions,
u(x, y, t) =

¸
m,n=1
sin(mπx/a) sin(nπy/b)[A
mn
cos(ω
mn
t) +B
mn
sin(ω
mn
t)].
The constants A
mn
and B
mn
are determined from the initial conditions.
The initial value problem considered in this section could represent the mo-
tion of a vibrating membrane. Just like in the case of the vibrating string, the
motion of the membrane is a superposition of inﬁnitely many modes, the mode
corresponding to the pair (m, n) oscillating with frequency ω
mn
/2π. The lowest
frequency of vibration or fundamental frequency is
ω
11

=
c

π
a

2
+

π
b

2
=
1
2

T
ρ
¸

1
a

2
+

1
b

2
¸
.
Exercises:
5.5.1. Solve the following initial value problem for a vibrating square membrane:
Find u(x, y, t), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂t
2
=

2
u
∂x
2
+

2
u
∂y
2
,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
135
u(x, y, 0) = 3 sin xsin y + 7 sin 2xsin y,
∂u
∂t
(x, y, 0) = 0.
5.5.2. Solve the following initial value problem for a vibrating square membrane:
Find u(x, y, t), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂t
2
= 4

2
u
∂x
2
+

2
u
∂y
2

,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, y, 0) = 0,
∂u
∂t
(x, y, 0) = 2 sin xsin y + 13 sin 2xsin y.
5.5.3. Solve the following initial value problem for a vibrating square membrane:
Find u(x, y, t), 0 ≤ x ≤ π, 0 ≤ y ≤ π, such that

2
u
∂t
2
=

2
u
∂x
2
+

2
u
∂y
2
,
u(x, 0, t) = u(x, π, t) = u(0, y, t) = u(π, y, t) = 0,
u(x, y, 0) = p(x)q(y),
∂u
∂t
(x, y, 0) = 0,
where
p(x) =

x, for 0 ≤ x ≤ π/2,
π −x, for π/2 ≤ x ≤ π,
q(y) =

y, for 0 ≤ y ≤ π/2,
π −y, for π/2 ≤ y ≤ π.
5.6 The Laplace operator in polar coordinates
In order to solve the heat equation over a circular plate, or to solve the wave
equation for a vibrating circular drum, we need to express the Laplace opera-
tor in polar coordinates (r, θ). These coordinates are related to the standard
Euclidean coordinates by the formulae
x = r cos θ, y = r sin θ.
The tool we need to express the Laplace operator
∆ =

2
∂x
2
+

2
∂y
2
in terms of polar coordinates is just the chain rule, which we studied earlier in
the course. Although straigtforward, the calculation is somewhat lengthy. Since
∂x
∂r
=

∂r
(r cos θ) = cos θ,
∂y
∂r
=

∂r
(r sin θ) = sin θ,
136
it follows immediately from the chain rule that
∂u
∂r
=
∂u
∂x
∂x
∂r
+
∂u
∂y
∂y
∂r
= (cos θ)
∂u
∂x
+ (sin θ)
∂u
∂y
.
Similarly, since
∂x
∂θ
=

∂θ
(r cos θ) = −r sin θ,
∂y
∂θ
=

∂θ
(r sin θ) = r cos θ,
it follows that
∂u
∂θ
=
∂u
∂x
∂x
∂θ
+
∂u
∂y
∂y
∂θ
= (−r sin θ)
∂u
∂x
+ (r cos θ)
∂u
∂y
.
We can write the results as operator equations,

∂r
= (cos θ)

∂x
+ (sin θ)

∂y
,

∂θ
= (−r sin θ)

∂x
+ (r cos θ)

∂y
.
For the second derivatives, we ﬁnd that

2
u
∂r
2
=

∂r
(
∂u
∂r
) =

∂r
¸
cos θ
∂u
∂x
+ sin θ
∂u
∂y

= cos θ

∂r

∂u
∂x

+ sin θ

∂r

∂u
∂y

= cos
2
θ

2
u
∂x
2
+ 2 cos θ sin θ

2
u
∂x∂y
+ sin
2
θ

2
u
∂y
2
.
Similarly,

2
u
∂θ
2
=

∂θ
(
∂u
∂θ
) =

∂θ
¸
−r sin θ
∂u
∂x
+r cos θ
∂u
∂y

= −r sin θ

∂θ

∂u
∂x

+r cos θ

∂θ

∂u
∂y

−r cos θ
∂u
∂x
−r sin θ
∂u
∂y
= r
2
sin
2
θ

2
u
∂x
2
−2r
2
cos θ sin θ

2
u
∂x∂y
+r
2
cos
2
θ

2
u
∂y
2
−r
∂u
∂r
,
which yields
1
r
2

2
u
∂θ
2
= sin
2
θ

2
u
∂x
2
−2 cos θ sin θ

2
u
∂x∂y
+ cos
2
θ

2
u
∂y
2

1
r
∂u
∂r
.
Adding these results together, we obtain

2
u
∂r
2
+
1
r
2

2
u
∂θ
2
=

2
u
∂x
2
+

2
u
∂y
2

1
r
∂u
∂r
,
or equivalently,
∆u =

2
u
∂r
2
+
1
r
2

2
u
∂θ
2
+
1
r
∂u
∂r
.
137
Finally, we can write this result in the form
∆u =
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
. (5.24)
This formula for the Laplace operator, together with the theory of Fourier
series, allows us to solve the Dirichlet problem for Laplace’s equation in a disk.
Indeed, we can now formulate the Dirichlet problem as follows: Find u(r, θ), for
0 < r ≤ 1 and θ ∈ mathbbR, such that
1. u satisﬁes Laplace’s equation,
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
= 0, (5.25)
2. u satisﬁes the periodicity condition u(r, θ + 2π) = u(r, θ),
3. u is well-behaved near r = 0,
4. u satisﬁes the boundary condition u(1, θ) = h(θ), where h(θ) is a given
well-behaved function satisfying the periodicity condition h(θ+2π) = h(θ).
The ﬁrst three of these conditions are homogeneous linear. To treat these
conditions via the method of separation of variables, we set
u(r, θ) = R(r)Θ(θ), where Θ(θ + 2π) = Θ(θ).
Substitution into (5.25) yields
1
r
d
dr

r
dR
dr

Θ +
R
r
2
d
2
Θ

2
= 0.
We multiply through by r
2
,
r
d
dr

r
dR
dr

Θ +R
d
2
Θ

2
= 0,
and divide by −RΘ to obtain

r
R
d
dr

r
dR
dr

=
1
Θ
d
2
Θ

2
.
The left-hand side of this equation does not depend on θ while the right-hand
side does not depend on r. Thus neither side can depend on either θ or r, and
hence both sides must be constant:

r
R
d
dr

r
dR
dr

=
1
Θ
d
2
Θ

2
= λ.
Thus the partial diﬀerential equation divides into two ordinary diﬀerential
equations
d
2
Θ

2
= λΘ, Θ(θ + 2π) = Θ(θ),
138
r
d
dr

r
dR
dr

= −λR.
We have seen the ﬁrst of these equations before when we studied heat ﬂow in a
circular wire, and we recognize that with the periodic boundary conditions, the
only nontrivial solutions are
λ = 0, Θ =
a
0
2
,
where a
0
is a constant, and
λ = −n
2
, Θ = a
n
cos nθ +b
n
sin nθ,
where a
n
and b
n
are constants, for n a positive integer. Substitution into the
second equation yields
r
d
dr

r
dR
dr

−n
2
R = 0.
If n = 0, the equation for R becomes
d
dr

r
dR
dr

= 0,
which is easily solved to yield
R(r) = A+Blog r,
where A and B are constants of integration. In order for this equation to be well-
behaved as r → 0 we must have B = 0, and the solution u
0
(r, θ) to Laplace’s
equation in this case is constant.
When n = 0, the equation for R is a Cauchy-Euler equidimensional equation
and we can ﬁnd a nontrivial solution by setting
R(r) = r
m
.
Then the equation becomes
r
d
dr

r
d
dr
(r
m
)

= n
2
r
m
,
and carrying out the diﬀerentiation on the left-hand side yields the characteristic
equation
m
2
−n
2
= 0,
which has the solutions m = ±n. In this case, the solution is
R(r) = Ar
n
+Br
−n
.
Once again, in order for this equation to be well-behaved as r → 0 we must
have B = 0, so R(r) is a constant multiple of r
n
, and
u
n
(r, θ) = a
n
r
n
cos nθ +b
n
r
n
sin nθ.
139
The general solution to (5.25) which is well-behaved at r = 0 and satisﬁes
the periodicity condition u(r, θ + 2π) = u(r, θ) is therefore
u(r, θ) =
a
0
2
+

¸
n=1
(a
n
r
n
cos nθ +b
n
r
n
sin nθ),
where a
0
, a
1
, . . . , b
1
, . . . are constants. To determine these constants we must
apply the boundary condition:
h(θ) = u(1, θ) =
a
0
2
+

¸
n=1
(a
n
cos nθ +b
n
sin nθ).
We conclude that the constants a
0
, a
1
, . . . , b
1
, . . . are simply the Fourier coeﬃ-
cients of h.
Exercises:
5.6.1. Solve the following boundary value problem for Laplace’s equation in a
disk: Find u(r, θ), 0 < r ≤ 1, such that
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
= 0,
u(r, θ + 2π) = u(r, θ), u well-behaved near r = 0
and
u(1, θ) = h(θ), where h(θ) = 1 + cos θ −2 sin θ + 4 cos 2θ.
5.6.2. Solve the following boundary value problem for Laplace’s equation in a
disk: Find u(r, θ), 0 < r ≤ 1, such that
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
= 0,
u(r, θ + 2π) = u(r, θ), u well-behaved near r = 0
and
u(1, θ) = h(θ),
where h(θ) is the periodic function such that
h(θ) = [θ[, for −π ≤ θ ≤ π.
5.6.3. Solve the following boundary value problem for Laplace’s equation in an
annular region: Find u(r, θ), 1 ≤ r ≤ 2, such that
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
= 0,
u(r, θ + 2π) = u(r, θ), u(1, θ) = 1 + 3 cos θ −sin θ + cos 2θ
and
u(2, θ) = 2 cos θ + 4 cos 2θ.
140
5.7 Eigenvalues of the Laplace operator
We would now like to consider the heat equation for a room whose shape is given
by a well-behaved but otherwise arbitrary bounded region D in the (x, y)-plane,
the boundary ∂D being a well-behaved curve. We would also like to consider
the wave equation for a vibrating drum in the shape of such a region D. Both
cases quickly lead to the eigenvalue problem for the Laplace operator
∆ =

2
∂x
2
+

2
∂y
2
.
∂u
∂t
= c
2
∆u, (5.26)
with the Dirichlet boundary condition,
u(x, y, t) = 0 for (x, y) ∈ ∂D.
To solve this equation, we apply separation of variables as before, setting
u(x, y, t) = f(x, y)g(t),
and substitute into (5.26) to obtain
f(x, y)g

(t) = c
2
(∆f)(x, y)g(t).
Then we divide by c
2
f(x, y)g(t),
1
c
2
g(t)
g

(t) =
1
f(x, y)
(∆f)(x, y).
The left-hand side of this equation does not depend on x or y while the right-
hand side does not depend on t. Hence both sides equal a constant λ, and we
obtain
1
c
2
g(t)
g

(t) = λ =
1
f(x, y)
(∆f)(x, y).
This separates into
g

(t) = c
2
λg(t) and (∆f)(x, y) = λf(x, y)
in which f is subject to the boundary condition,
f(x, y) = 0 for (x, y) ∈ ∂D.
The same method can be used to treat the wave equation

2
u
∂t
2
= c
2
∆u, (5.27)
141
with the Dirichlet boundary condition,
u(x, y, t) = 0 for (x, y) ∈ ∂D.
This time, substitution of u(x, y, t) = f(x, y)g(t) into (5.27) yields
f(x, y)g

(t) = c
2
(∆f)(x, y)g(t), or
1
c
2
g(t)
g

(t) =
1
f(x, y)
(∆f)(x, y).
Once again, both sides must equal a constant λ, and we obtain
1
c
2
g(t)
g

(t) = λ =
1
f(x, y)
(∆f)(x, y).
This separates into
g

(t) = λg(t) and (∆f)(x, y) = λf(x, y),
in which f is once again subject to the boundary condition,
f(x, y) = 0 for (x, y) ∈ ∂D.
In both cases, we must solve the “eigenvalue problem” for the Laplace op-
erator with Dirichlet boundary conditions. If λ is a real number, let
W
λ
= ¦smooth functions f(x, y) : ∆f = λf, f[∂D = 0¦.
We say that λ is an eigenvalue of the Laplace operator ∆ on D if W
λ
= 0.
Nonzero elements of W
λ
are called eigenfunctions and W
λ
itself is called the
eigenspace for eigenvalue λ. The dimension of W
λ
is called the multiplicity of
the eigenvalue λ. The eigenvalue problem consist of ﬁnding the eigenvalues λ,
and a basis for each nonzero eigenspace.
Once we have solved the eigenvalue problem for a given region D in the
(x, y)-plane, it is easy to solve the initial value problem for the heat equation
or the wave equation on this region. To do so requires only that we substitute
the values of λ into the equations for g. In the case of the heat equation,
g

(t) = c
2
λg(t) ⇒ g(t) = (constant)e
c
2
λt
,
while in the case of the wave equation,
g

(t) = c
2
λg(t) ⇒ g(t) = (constant) sin(c

−λ(t −t
0
)).
In the second case, the eigenvalues determine the frequencies of a vibrating
drum which has the shape of D.
Theorem. All of the eigenvalues of ∆ are negative, and each eigenvalue has
ﬁnite multiplicity. The eigenvalues can be arranged in a sequence
0 > λ
1
> λ
2
> > λ
n
> ,
142
with λ
n
→ −∞. Every well-behaved function can be represented as a convergent
sum of eigenfunctions.
2
Although the theorem is reassuring, it is usually quite diﬃcult to determine the
eigenvalues λ
1
, λ
2
, , λ
n
, explicitly for a given region in the plane.
There are two important cases in which the eigenvalues can be explicitly
determined. The ﬁrst is the case where D is the rectangle,
D = ¦(x, y) : 0 ≤ x ≤ a, 0 ≤ y ≤ b¦.
We saw how to solve the eigenvalue problem

2
f
∂x
2
+

2
f
∂y
2
= λf, f(x, y) = 0 for (x, y) ∈ ∂D,
when we discussed the heat and wave equations for a rectangular region. The
nontrivial solutions are
λ
mn
= −

πm
a

2

πn
b

2
, f
mn
(x, y) = sin

πmx
a

sin

πny
b

,
where m and n are positive integers.
A second case in which the eigenvalue problem can be solved explicitly is
that of the disk,
D = ¦(x, y) : x
2
+y
2
≤ a
2
¦,
where a is a positive number, as we will see in the next section.
3
Exercises:
5.7.1. Show that among all rectangular vibrating membranes of area one, the
square has the lowest fundamental frequency of vibration by minimizing the
function
f(a, b) =

π
a

2
+

π
b

2
subject to the constraints ab = 1, a, b > 0. Hint: One can eliminate b by setting
a = 1/b and then ﬁnd the minimum of the function
g(a) =

π
a

2
+ (πa)
2
,
when a > 0.
5.7.2. Let D be a ﬁnite region in the (x, y)-plane bounded by a smooth curve
∂D. Suppose that the eigenvalues for the Laplace operator ∆ with Dirichlet
boundary conditions on D are λ
1
, λ
2
, . . . , where
0 > λ
1
> λ
2
> ,
2
Note the similarity between the statement of this theorem and the statement of the
theorem presented in Section 4.7. In fact, the techniques used to prove the two theorems are
also quite similar.
3
A few more cases are presented in advanced texts, such as Courant and Hilbert, Methods
of mathematical physics I , New York, Interscience, 1953. See Chapter V, §16.3.
143
each eigenvalue having multiplicity one. Suppose that φ
n
(x, y) is a nonzero
eigenfunction for eigenvalue λ
n
. Show that the general solution to the heat
equation
∂u
∂t
= ∆u
with Dirichlet boundary conditions (u vanishes on ∂D) is
u(x, t) =

¸
n=1
b
n
φ
n
(x, y)e
λnt
,
where the b
n
’s are arbitrary constants.
5.7.3. Let D be a ﬁnite region in the (x, y)-plane as in the preceding problem.
Suppose that the eigenvalues for the Laplace operator ∆ on D are λ
1
, λ
2
, . . . ,
once again. Show that the general solution to the wave equation

2
u
∂t
2
= ∆u,
together with Dirichlet boundary conditions (u vanishes on ∂D) and the initial
condition
∂u
∂t
(x, y, 0) = 0,
is
u(x, t) =

¸
n=1
b
n
φ
n
(x, y) cos(

−λ
n
t),
where the b
n
’s are arbitrary constants.
5.8 Eigenvalues of the disk
To calculate the eigenvalues of the disk, it is convenient to utilize polar coordi-
nates r, θ in terms of which the Laplace operator is
1
r

∂r

r
∂f
∂r

+
1
r
2

2
f
∂θ
2
= λf, f[∂D = 0. (5.28)
Once again, we use separation of variables and look for product solutions of the
form
f(r, θ) = R(r)Θ(θ), where R(a) = 0, Θ(θ + 2π) = Θ(θ).
Substitution into (5.28) yields
1
r
d
dr

r
dR
dr

Θ +
R
r
2
d
2
Θ

2
= λRΘ.
144
We multiply through by r
2
,
r
d
dr

r
dR
dr

Θ +R
d
2
Θ

2
= λr
2
RΘ,
and divide by RΘ to obtain
r
R
d
dr

r
dR
dr

−λr
2
= −
1
Θ
d
2
Θ

2
.
Note that in this last equation, the left-hand side does not depend on θ while
the right-hand side does not depend on r. Thus neither side can depend on
either θ or r, and hence both sides must be constant:

r
R
d
dr

r
dR
dr

+λr
2
=
1
Θ
d
2
Θ

2
= µ.
Thus in the manner now familiar, the partial diﬀerential equation divides
into two ordinary diﬀerential equations
d
2
Θ

2
= µΘ, Θ(θ + 2π) = Θ(θ),
r
d
dr

r
dR
dr

−λr
2
R = −µR, R(a) = 0.
Once again, the only nontrivial solutions are
µ = 0, Θ =
a
0
2
and
µ = −n
2
, Θ = a
n
cos nθ +b
n
sin nθ,
for n a positive integer. Substitution into the second equation yields
r
d
dr

r
dR
dr

+ (−λr
2
−n
2
)R = 0.
Let x =

−λr so that x
2
= −λr
2
. Then since
r
d
dr
= x
d
dx
,
our diﬀerential equation becomes
x
d
dx

x
dR
dx

+ (x
2
−n
2
)R = 0, (5.29)
where R vanishes when x =

−λa. If we replace R by y, this becomes
x
d
dx

x
dy
dx

+ (x
2
−n
2
)y = 0, y(

−λa) = 0. (5.30)
145
The diﬀerential equation appearing in (5.30) is Bessel’s equation, which we
studied in Section 1.4.
Recall that in Section 1.4, we found that for each choice of n, Bessel’s equa-
tion has a one-dimensional space of well-behaved solutions, which are constant
multiples of the Bessel function of the ﬁrst kind J
n
(x). Here is an important
fact regarding these Bessel functions:
Theorem. For each nonnegative integer n, J
n
(x) has inﬁnitely many positive
zeros.
Graphs of the functions J
0
(x) and J
1
(x) suggest that this theorem might well be
true, but it takes some eﬀort to prove rigorously. For completeness, we sketch
the proof for the case of J
0
(x) at the end of the section.
The zeros of the Bessel functions are used to determine the eigenvalues of
the Laplace operator on the disk. To see how, note ﬁrst that the boundary
condition,
y(

−λa) = 0,
requires that

−λa be one of the zeros of J
n
(x). Let α
n,k
denote the k-th
positive root of the equation J
n
(x) = 0. Then

−λa = α
n,k
⇒ λ = −
α
2
n,k
a
2
,
and
R(r) = J
n

n,k
r/a)
will be a solution to (5.29) vanishing at r = a. Hence, in the case where n = 0,
λ
0,k
= −
α
2
0,k
a
2
,
and
f
0,k
(r, θ) = J
0

0,k
r/a)
will be a solution to the eigenvalue problem
1
r

∂r

r
∂f
∂r

+
1
r
2

2
f
∂θ
2
= λf, f[∂D = 0.
Similarly, in the case of general n,
λ
n,k
= −
α
2
n,k
a
2
,
and
f
n,k
(r, θ) = J
n

n,k
r/a) cos(nθ) or g
n,k
= J
n

n,k
r/a) sin(nθ),
for n = 1, 2, . . . , will be solutions to this eigenvalue problem.
146
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1
0
0.25
0.5
0.75
1
-1
-0.5
0
0.5
1
Figure 5.2: Graph of the function f
0,1
(r, θ).
Each of these eigenfunctions corresponds to a mode of oscillation for the
circular vibrating membrane. Suppose, for example, that a = 1. Then
α
0,1
= 2.40483 ⇒ λ
0,1
= −5.7832,
α
0,2
= 5.52008 ⇒ λ
0,2
= −30.471,
α
0,3
= 8.65373 ⇒ λ
0,3
= −74.887,
α
0,4
= 11.7195 ⇒ λ
0,3
= −139.039,
and so forth. The mode of oscillation corresponding to the function f
0,1
will
vibrate with frequency

−λ
0,1

=
α
0,1

= .38274,
while the mode of oscillation corresponding to the function f
0,2
will vibrate with
frequency

−λ
0,2

=
α
0,2

= .87855.
Similarly,
α
1,1
= 3.83171 ⇒ λ
1,1
= −14.682,
α
2,1
= 5.13562 ⇒ λ
2,1
= −26.3746,
and hence the mode of oscillation corresponding to the functions f
1,1
and g
1,1
will vibrate with frequency

−λ
1,1

=
α
1,1

= .60984,
147
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1
0
0.5
1
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1
Figure 5.3: Graph of the function f
0,2
(r, θ).
while the mode of oscillation corresponding to the functions f
2,1
and g
2,1
will
vibrate with frequency

−λ
2,1

=
α
2,1

= .81736.
A general vibration of the vibrating membrane will be a superposition of
these modes of oscillation. The fact that the frequencies of oscillation of a
circular drum are not integral multiples of a single fundamental frequency (as
in the case of the violin string) limits the extent to which a circular drum can
be tuned to a speciﬁc tone.
Proof that J
0
(x) has inﬁnitely many positive zeros: First, we make a change
of variables x = e
z
and note that as z ranges over the real numbers, the corre-
sponding variable x ranges over all the positive real numbers. Since
dx = e
z
dz,
d
dx
=
1
e
z
d
dz
and hence x
d
dx
= e
z
1
e
z
d
dz
=
d
dz
.
Thus Bessel’s equation (5.30) in the case where n = 0 becomes
d
2
y
dz
2
+e
2z
y = 0. (5.31)
Suppose that z
0
> 1 and y(z) is a solution to (5.31) with y(z
0
) = 0. We
claim that y(z) must change sign at some point between z
0
and z
0
+π.
148
-1
-0.5
0
0.5
1
-1
-0.5
0
0.5
1
-0.5
-0.25
0
0.25
0.5
-1
-0.5
0
0.5
1
Figure 5.4: Graph of the function f
1,1
(r, θ).
Assume ﬁrst that y(z
0
) > 0. Let f(z) = sin(z − z
0
). Since f

(z) = −f(z)
and y(z) satisﬁes (5.31),
d
dz
[y(z)f

(z) −y

(z)f(z)] = y(z)f

(z) −y

(z)f(z)
= −y(z)f(z) +e
2z
y(z)f(z) = (e
2z
−1)y(z)f(z).
Note that f(z) > 0 for z between z
0
and z
0
+π. If also y(z) > 0 for z between
z
0
and z
0
+π, then y(z)f(z) > 0 and
0 <

z0+π
z0
(e
2z
−1)y(z)f(z)dz = [y(z)f

(z) −y

(z)f(z)]
z0+π
z0
= y(z
0
+π)f

(z
0
+π) −y(z
0
)f(z
0
) = −y(z
0
+π) −y(z
0
) < 0,
a contradiction. Hence our assumption that y(z) be postive for z between z
0
and z
0
+π must be incorrect, y(z) must change sign at some z between z
0
and
z
0
+π, and hence y(z) must be zero at some point in this interval.
If y(z
0
) < 0, just apply the same argument to −y. In either case, we conclude
that y(z) must be zero at some point in any interval to the right of z = 1 of
length at least π. It follows that the solution to (5.31) must have inﬁnitely many
zeros in the region z > 1, and J
0
(x) must have inﬁnitely many positive zeros,
as claimed.
The fact that J
n
(x) has inﬁnitely many positive zeros could be proven in a
similar fashion.
149
Exercises:
5.8.1. Solve the following initial value problem for the heat equation in a disk:
Find u(r, θ, t), 0 < r ≤ 1, such that
∂u
∂t
=
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
,
u(r, θ + 2π, t) = u(r, θ, t), u well-behaved near r = 0,
u(1, θ, t) = 0,
u(r, θ, 0) = J
0

0,1
r) + 3J
0

0,2
r) −2J
1

1,1
r) sin θ + 4J
2

2,1
r) cos 2θ.
5.8.2. Solve the following initial value problem for the vibrating circular mem-
brane: Find u(r, θ, t), 0 < r ≤ 1, such that

2
u
∂t
2
=
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
,
u(r, θ + 2π, t) = u(r, θ, t), u well-behaved near r = 0,
u(1, θ, t) = 0,
∂u
∂t
(r, θ, 0) = 0,
u(r, θ, 0) = J
0

0,1
r) + 3J
0

0,2
r) −2J
1

1,1
r) sin θ + 4J
2

2,1
r) cos 2θ.
5.8.3. (For students with access to Mathematica) a. Run the following Mathe-
matica program to sketch the Bessel function J
0
(x):
n=0; Plot[ BesselJ[n,x], ¦x,0,15¦]
b. From the graph it is clear that the ﬁrst root of the equation J
0
(x) = 0 is
near 2. Run the following Mathematica program to ﬁnd the ﬁrst root α
0,1
of
the Bessel function J
0
(x):
n=0; FindRoot[ BesselJ[n,x] == 0 ,¦x,2¦]
Find the next two nonzero roots of the Bessel function J
0
(x).
c. Modify the programs to sketch the Bessel functions J
1
(x), . . . , J
5
(x), and
determine the ﬁrst three nonzero roots of each of these Bessel functions.
5.8.4. Which has a lower fundamental frequency of vibration, a square drum or
a circular drum of the same area?
5.9 Fourier analysis for the circular vibrating
membrane*
To ﬁnish up the solution to the initial value problem for arbitrary initial displace-
ments of the vibrating membrane, we need to develop a theory of generalized
150
Fourier series which gives a decomposition into eigenfunctions which solve the
eigenvalue problem.
Such a theory can be developed for an arbitrary bounded region D in the
(x, y)-plane bounded by a smooth closed curve ∂D. Let V denote the space of
smooth functions f : D → R whose restrictions to ∂D vanish. It is important
to observe that V is a “vector space”: the sum of two elements of V again lies
in V and the product of an element of V with a constant again lies in V .
We deﬁne an inner product ', ` on V by
'f, g` =

D
f(x, y)g(x, y)dxdy.
Lemma. With respect to this inner product, eigenfunctions corresponding to
distinct eigenvalues are perpendicular; if f and g are smooth functions vanishing
on ∂D such that
∆f = λf, ∆g = µg, (5.32)
then either λ = µ or 'f, g` = 0.
The proof of this lemma is a nice application of Green’s theorem. Indeed, it
follows from Green’s theorem that

∂D
−f
∂g
∂y
dx +f
∂g
∂x
dy =

D

∂x
[f(∂g/∂x)] −

∂y
[−f(∂g/∂y)]

dxdy
=

D
¸
∂f
∂x
∂g
∂x
+
∂f
∂y
∂g
∂y

dxdy +

D
f

2
g
∂x
2
+

2
g
∂y
2

dxdy.
Hence if the restriction of f to ∂D vanishes,

D
f∆gdxdy = −

D
∇f ∇gdxdy.
Similarly, if the restriction of g to ∂D vanishes,

D
g∆fdxdy = −

D
∇f ∇gdxdy.
Thus if f and g lie in V ,
'f, ∆g` = −

D
∇f ∇gdxdy = 'g, ∆f`.
In particular, if (5.32) holds, then
µ'f, g` = 'f, ∆g` = '∆f, g` = λ'f, g`,
and hence
(λ −µ)'f, g` = 0.
151
It follows immediately that either λ − µ = 0 or 'f, g` = 0, and the lemma is
proven.
Now let us focus on the special case in which D is a circular disk. Recall the
problem that we want to solve, in terms of polar coordinates: Find
u(r, θ, t), 0 < r ≤ 1,
so that

2
u
∂t
2
=
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
,
u(r, θ + 2π, t) = u(r, θ, t), u well-behaved near r = 0,
u(1, θ, t) = 0,
u(r, θ, 0) = h(r, θ),
∂u
∂t
(r, θ, 0) = 0.
It follows from the argument in the preceding section that the general solu-
tion to the homogeneous linear part of the problem can be expressed in terms
of the eigenfunctions
f
n,k
(r, θ) = J
n

n,k
r/a) cos(nθ), g
n,k
= J
n

n,k
r/a) sin(nθ).
Indeed, the general solution must be a superposition of the products of these
eigenfunctions with periodic functions of t of frequencies

−λ
n,k
/2π. Thus this
general solution is of the form
u(r, θ, t) =

¸
k=1
a
0,k
g
0,k
(r, θ) cos(α
0,k
t)
+

¸
n=1

¸
k=1
[a
n,k
f
n,k
(r, θ) +b
n,k
g
n,k
(r, θ)] cos(α
n,k
t). (5.33)
We need to determine the a
n,k
’s and the b
n,k
’s so that the inhomogeneous initial
condition u(r, θ, 0) = h(r, θ) is also satisﬁed. If we set t = 0 in (5.33), this
condition becomes

¸
k=1
a
0,k
f
0,k
+

¸
n=1

¸
k=1
[a
n,k
f
n,k
+b
n,k
g
n,k
] = h. (5.34)
Thus we need to express an arbitrary initial displacement h(r, θ) as a superpo-
sition of these eigenfunctions.
It is here that the inner product ', ` on V comes to the rescue. The lemma
implies that eigenfunctions corresponding to distinct eigenvalues are perpendic-
ular. This implies, for example, that
'f
0,j
, f
0,k
` = 0, unless j = k.
152
Similarly, for arbitrary n ≥ 1,
'f
n,j
, f
n,k
` = 'f
n,j
, g
n,k
` = 'g
n,j
, g
n,k
` = 0, unless j = k.
The lemma does not immediately imply that
'f
n,j
, g
n,j
` = 0.
To obtain this relation, recall that in terms of polar coordinates, the inner
product ', ` on V takes the form
'f, g` =

1
0
¸

0
f(r, θ)g(r, θ)rdθ

dr.
If we perform integration with respect to θ ﬁrst, we can conclude from the
familiar integral formula

0
sin nθ cos nθdθ = 0
that
'f
n,j
, g
n,j
` = 0,
as desired.
Moreover, using the integral formulae

0
cos nθ cos mθdθ =

π, for m = n,
0, for m = n,

0
sin nθ sin mθdθ =

π, for m = n,
0, for m = n,

0
sin nθ cos mθdθ = 0,
we can check that
'f
n,j
, f
m,k
` = 'g
n,j
, g
m,k
` = 'f
n,j
, g
m,k
` = 0,
unless m = n. It then follows from the lemma that
'f
n,j
, f
m,k
` = 'g
n,j
, g
m,k
` = 'f
n,j
, g
m,k
` = 0,
unless j = k and m = n.
From these relations, it is not diﬃcult to construct formulae for the coeﬃ-
cients of the generalized Fourier series of a given function h(r, θ). Indeed, if we
take the inner product of equation (5.34) with u
n,k
, we obtain
a
n,k
'f
n,k
, f
n,k
` = 'h, f
n,k
`,
153
or equivalently
a
n,k
=
'h, f
n,k
`
'f
n,k
, f
n,k
`
. (5.35)
Similarly,
b
n,k
=
'h, g
n,k
`
'g
n,k
, g
n,k
`
. (5.36)
Thus we ﬁnally see that the solution to our initial value problem is
u(r, θ, t) =

¸
k=1
a
0,k
f
0,k
(r, θ) cos(α
0,k
t)
+

¸
n=1

¸
k=1
[a
n,k
f
n,k
(r, θ) +b
n,k
g
n,k
(r, θ)] cos(α
n,k
t),
where the coeﬃcients are determined by the integral formulae (5.35) and (5.36).
Exercises:
5.9.1. (For students with access to Mathematica) Suppose that a circular vi-
brating drum is given the initial displacement from equilibrium described by
the function
h(r, θ) = .1(1 −r),
for 0 ≤ r ≤ 1. In order to ﬁnd the solution to the wave equation with this initial
condition, we need to expand h(r, θ) in terms of eigenfunctions of ∆. Because
of axial symmetry, we can write
h(r, θ) = a
0,1
J
0

0,1
r) +a
0,2
J
0

0,2
r) +. . . .
a. Use the following Mathematica program to determine the coeﬃcient a
0,1
in
this expansion:
d = NIntegrate[r (BesselJ[0,2.40483 r])∧2,¦r,0,1¦];
n = NIntegrate[r (BesselJ[0,2.40483 r]) .1 (1-r),¦r,0,1¦];
a01 = n/d
b. Use a modiﬁcation of the program to determine the coeﬃcients a
0,2
, a
0,3
,
and a
0,4
.
c. Determine the ﬁrst four terms in the solution to the initial value problem:
Find
u(r, θ, t), 0 < r ≤ 1,
so that

2
u
∂t
2
=
1
r

∂r

r
∂u
∂r

+
1
r
2

2
u
∂θ
2
,
154
u(r, θ + 2π, t) = u(r, θ, t), u well-behaved near r = 0,
u(1, θ, t) = 0,
u(r, θ, 0) = .1(1 −r),
∂u
∂t
(r, θ, 0) = 0.
sketch graphs of the successive positions of a membrane, vibrating with a su-
perposition of several of the lowest modes of vibration:
<< Graphics‘ParametricPlot3D‘
a01 = .2; a02 = .1; a03 = .2; a04 = .1; a11 = .3; b11 = .2;
vibmem = Table[
CylindricalPlot3D[
a01 BesselJ[0,2.4 r] Cos[2.4 t]
+ a02 BesselJ[0,5.52 r] Cos[5.52 t]
+ a03 BesselJ[0,8.65 r] Cos[8.65 t]
+ a04 BesselJ[0,11.8 r] Cos[11.8 t]
+ a11 BesselJ[1,3.83 r] Cos[u] Cos[3.83 t]
+ b11 BesselJ[1,3.83 r] Sin[u] Cos[3.83 t],
¦r,0,1¦, ¦u,0,2 Pi¦, PlotRange -> ¦-.5,.5¦
], ¦t,0,1.,.1¦
]
a. Execute this program and then animate the sequence of graphics by running
“Animate selected graphics,” from the Graphics menu. (Warning! Generating
the sequence of graphics takes lots of memory. If you have insuﬃcient memory
for the animation, try inserting PlotPoints -> 8 within the CylindricalPlot
statement.)
b. Replace the values of a01, . . . with the Fourier coeﬃcients obtained in the
preceding problem, execute the resulting program to generate a sequence of
graphics, and animate the sequence as before.
155
Appendix A
Using Mathematica to solve
diﬀerential equations
In solving diﬀerential equations, it is sometimes necessary to do calculations
which would be prohibitively diﬃcult to do by hand. Fortunately, computers
can do the calculations for us, if they are equiped with suitable software, such
as Maple, Matlab, or Mathematica. This appendix is intended to give a brief
introduction to the use of Mathematica for doing such calculations.
Most computers which are set up to use Mathematica contain an on-line
tutorial, “Tour of Mathematica,” which can be used to give a brief hands-on
introduction to Mathematica. Using this tour, it may be possible for many
students to learn Mathematica without referring to lengthy technical manuals.
However, there is a standard reference on the use of Mathematica, which can
be used to answer questions if necessary. It is The Mathematica Book, by
Stephen Wolfram, Fourth edition, Wolfram Media and Cambridge University
Press, 1999.
We give here a very brief introduction to the use of Mathematica. After
launching Mathematica, you can use it as a “more powerful than usual graphing
calculator.” For example, if you type in
(11 - 5)/3
the computer will perform the subtraction and division, and respond with
Out[1] = 2
The notation for multiplication is ∗, so if you type in
2 * (7 + 4)
the computer will respond with
Out[2] = 22
156
You can also use a space instead of ∗ to denote multiplication, so if you input
2 8
the computer will respond with
Out[3] = 16
The computer can do exact arithmetic with integers or rational numbers.
For example, since ∧ is the symbol used for raising to a power, if you type in
2∧150
the computer will calculate the 150th power of two and give the exact result:
Out[4] = 1427247692705959881058285969449495136382746624
On the other hand, the symbol N tells the computer to use an approximation
with a ﬁxed number of digits of precision, so entering
N[2∧150]
will give an approximation to the exact answer, expressed in scientiﬁc notation:
Out[5] = 1.42725 10
45
Real numbers which are not rational, such as π, have inﬁnite decimal expan-
sions which cannot be stored within a computer. However, we can approximate
a real number by a ﬁnite decimal expansion with a given number of digits of pre-
cision. For example, since Mathematica uses the name E to denote the number
e, typing in
N[E]
will tell Mathematica to give a rational approximation to e to a standard number
of digits of precision:
Out[6] = 2.71828
In principle, the computer can calculate rational approximations to an arbitrary
number of digits of precision. Since the number π is represented by Mathematica
as Pi, typing in
N[Pi,40]
will tell Mathematica to give a rational approximation to π to 40 digits of
precision:
Out[7] = 3.1415926535897932384626433832795028841972
The computer can represent not only numbers but functions. Within Math-
ematica, built-in functions are described by names which begin with capital
157
1.5 2 2.5 3
0.2
0.4
0.6
0.8
1
Figure A.1: Graph of the logarithm function.
letters. For example, Log denotes the natural base e logarithm function. Thus
entering
N[Log[2],40]
will give the logarithm of 2 with 40 digits of precision:
Out[8] = 0.693147180559945309417232121458176568076
One can also plot functions with Mathematica. For example, to plot the
logarithm function from the values 1 to 3, one simply inputs
Plot[Log[t],¦t,1,3¦]
and Mathematica will automatically produce the plot shown in Figure A.1.
We can also deﬁne functions ourselves, being careful not to capitalize them,
because capitals are reserved by Mathematica for built-in functions. Thus we
can deﬁne the function
y(t) = ce
kt
by typing
y[t ] := c * E∧(k * t);
Mathematica will remember the function we have deﬁned until we quit Mathe-
matica. We must remember the exact syntax of this example (use of the under-
line character and the colon followed by the equal sign) when deﬁning functions.
In this example c and k are parameters which can be deﬁned later. Just as in
the case of functions, we must be careful to represent parameters by lower case
letters, so they will not be confused with built-in constants. Further entry of
c = 1; k = 1; N[y[1]]
158
yields the response
Out[11] = 2.71828
while entering
Plot[y[t],¦t,0,2¦]
will give a plot of the function we have deﬁned, for t in the interval [0, 2].
We can use Mathematica to solve matrix diﬀerential equations of the form
dx
dt
= Ax, (A.1)
where A is a square matrix with constant entries.
The ﬁrst step consists of using Mathematica to ﬁnd the eigenvalues and
eigenvectors of A. To see how this works, we must ﬁrst become familiar with
the way in which Mathematica represents matrices. Since Mathematica reserves
upper case letters for descriptions of built-in functions, it is prudent to denote
the matrix A by lower case a when writing statements in Mathematica. The
matrix
A =

−2 5
1 −3

can be entered into Mathematica as a collection of row vectors,
a = ¦¦-2,5¦,¦1,-3¦¦
with the computer responding by
Out[1] = ¦¦-2,5¦,¦1,-3¦¦
Thus a matrix is thought of as a “vector of vectors.” Entering
MatrixForm[a]
will cause the computer to give the matrix in the familiar form
Out[2] = −2 5
1 −3
To ﬁnd the eigenvalues of the matrix A, we simply type
Eigenvalues[a]
and the computer will give us the exact eigenvalues
−5 −

21
2
,
−5 +

21
2
,
which have been obtained by using the quadratic formula. Quite often numerical
approximations are suﬃcient, and these can be obtained by typing
eval = Eigenvalues[N[a]]
159
the response this time being
Out[4] = ¦-4.79129, -0.208712¦
Deﬁning eval to be the eigenvalues of A in this fashion, allows us to refer to
the eigenvalues of A later by means of the expression eval.
We can also ﬁnd the corresponding eigenvectors for the matrix A by typing
evec = Eigenvectors[N[a]]
and the computer responds with
Out[5] = ¦¦-0.873154, 0.487445¦, ¦0.941409, 0.337267¦¦
Putting this together with the eigenvalues gives the general solution to the
original linear system (A.1) for our choice of the matrix A:
x(t) = c
1

−0.873154
0.487445

e
−4.79129t
+c
2

0.941409
0.337267

e
−0.208712t
.
Mathematica can also be used to ﬁnd numerical solutions to nonlinear diﬀer-
ential equations. The following Mathematica programs will use Mathematica’s
diﬀerential equation solver (which is called up by the command NDSolve), to
ﬁnd a numerical solution to the initial value problem
dy/dx = y(1 −y), y(0) = .1,
give a table of values for the solution, and graph the solution curve on the
interval 0 ≤ x ≤ 6. The ﬁrst step
sol := NDSolve[¦ y’[x] == y[x] (1 - y[x]), y[0] == .1 ¦,
y, ¦x,0,6¦]
generates an “interpolation function” which approximates the solution and calls
it sol, an abbreviation for solution. We can construct a table of values for the
interpolation function by typing
Table[Evaluate[y[x] /. sol], ¦x,0,6,.1¦];
or graph the interpolation function by typing
Plot[Evaluate[y[x] /. sol], ¦x,0,6¦]
This leads to a plot like that shown in Figure A.2.
Readers can modify these simple programs to graph solutions to initial value
problems for quite general diﬀerential equations of the canonical form
dy
dx
= f(x, y).
All that is needed is to replace the ﬁrst argument of NDSolve with the diﬀerential
equation one wants to solve, remembering to replace the equal signs with double
equal signs, as in the example.
160
1 2 3 4 5 6
0.2
0.4
0.6
0.8
1
Figure A.2: Solution to y

= y(1 −y), y(0) = .1.
0.5 0.75 1.25 1.5 1.75 2
0.1
0.2
0.3
0.4
Figure A.3: A parametric plot of a solution to dx/dt = −xy, dy/dt = xy −y.
161
In fact, it is no more diﬃcult to treat initial value problems for higher order
equations or systems of diﬀerential equations. For example, to solve the initial
value problem
dx
dt
= −xy,
dy
dt
= xy −y, x(0) = 2, y(0) = .1. (A.2)
one simply types in
sol := NDSolve[¦ x’[t] == - x[t] y[t], y’[t] == x[t] y[t] - y[t],
x[0] == 2, y[0] == .1 ¦,
¦x,y¦, ¦t,0,10¦]
Once we have this solution, we can print out a table of values for y by entering
Table[Evaluate[y[t] /. sol], ¦t,0,2,.1¦]
We can also plot y(t) as a function of t by entering
Plot[Evaluate[y[t] /. sol], ¦t,0,10¦]
Figure A.3 shows a parametric plot of (x(t), y(t)) which is generated by entering
ParametricPlot[Evaluate[¦ x[t], y[t]¦ /. sol], ¦t,0,10¦]
162
Index
Bessel’s equation, 22, 146
Bessel function, 25
boundary value problem, 83, 119
Cauchy-Euler equation, 17, 110, 139
Chebyshev’s equation, 14
comparison test, 5
conic section, 36
convergence, 2
Dirichlet boundary conditions, 85
Dirichlet problem, 119
double Fourier sine series, 126
eigenfunctions, 108, 142
eigenspace, 33, 108, 142
eigenvalue problem, 87, 108, 142
eigenvalues, 30, 108, 142
eigenvectors, 30
ellipse, 37
ellipsoid, 40
Euler equations for a gas, 132
even function, 70
ﬁnite diﬀerences, 94
Fourier coeﬃcients, 63
Fourier cosine series, 72, 91
Fourier series, 63
Fourier sine series, 72, 89
Fourier transform, 79
frequencies of oscillation, 52, 58, 135
Frobenius method, 16
fundamental mode, 100
fundamental frequency, 135
gamma function, 28
geometric series, 2
general solution, 9
generalized Fourier series, 151
generalized power series, 16
Gibbs phenomenon, 67
heat equation, 83, 115, 117
heat ﬂow in a bar, 82
heat ﬂow in space, 116
Hermite’s equation, 10
Hooke’s law, 49
hyperbola, 37
hyperboloid, 40
initial value problem, 85, 98, 103,
124, 134
inner product, 68, 151
Laplace’s equation, 115
Legendre’s equation, 12
linearization of PDE, 132
mechanical systems, 49
modes of oscillation, 52, 58, 135
modiﬁed Bessel equation, 27
Navier-Stokes equations, 133
Neumann boundary conditions, 91
Newton’s second law, 50, 96
odd function, 70
ordinary point, 13
orthogonal matrix, 29
orthonormal basis, 44, 69
partial diﬀerential equation, 81
periodic function, 63
piecewise smooth, 67
potential energy, 50, 95, 129
power series, 1
ratio test, 4
real analytic, 5
recursion formula, 8
regular singular point, 13
separating constant, 86
separation of variables, 86
singular point, 13
sound waves, 130
163
spectral theorem, 30
Sturm-Liouville problem, 108
superposition principle, 86, 115
symmetric matrix, 29
Theorem of Frobenius, 16
variation of parameters, 18
vibrating membrane, 129
vibrating string, 94
wave equation, 96, 115
164

Preface
Partial diﬀerential equations are often used to construct models of the most basic theories underlying physics and engineering. For example, the system of partial diﬀerential equations known as Maxwell’s equations can be written on the back of a post card, yet from these equations one can derive the entire theory of electricity and magnetism, including light. Our goal here is to develop the most basic ideas from the theory of partial differential equations, and apply them to the simplest models arising from physics. In particular, we will present some of the elegant mathematics that can be used to describe the vibrating circular membrane. We will see that the frequencies of a circular drum are essentially the eigenvalues from an eigenvector-eigenvalue problem for Bessel’s equation, an ordinary diﬀerential equation which can be solved quite nicely using the technique of power series expansions. Thus we start our presentation with a review of power series, which the student should have seen in a previous calculus course. It is not easy to master the theory of partial diﬀerential equations. Unlike the theory of ordinary diﬀerential equations, which relies on the “fundamental existence and uniqueness theorem,” there is no single theorem which is central to the subject. Instead, there are separate theories used for each of the major types of partial diﬀerential equations that commonly arise. However, there are several basic skills which are essential for studying all types of partial diﬀerential equations. Before reading these notes, students should understand how to solve the simplest ordinary diﬀerential equations, such as the equation of exponential growth dy/dx = ky and the equation of simple harmonic motion d2 y/dx2 + ωy = 0, and how these equations arise in modeling population growth and the motion of a weight attached to the ceiling by means of a spring. It is remarkable how frequently these basic equations arise in applications. Students should also understand how to solve ﬁrst-order linear systems of diﬀerential equations with constant coeﬃcients in an arbitrary number of unknowns using vectors and matrices with real or complex entries. (This topic will be reviewed in the second chapter.) Familiarity is also needed with the basics of vector calculus, including the gradient, divergence and curl, and the integral theorems which relate them to each other. Finally, one needs ability to carry out lengthy calculations with conﬁdence. Needless to say, all of these skills are necessary for a thorough understanding of the mathematical

i

language that is an essential foundation for the sciences and engineering. Moreover, the subject of partial diﬀerential equations should not be studied in isolation, because much intuition comes from a thorough understanding of applications. The individual branches of the subject are concerned with the special types of partial diﬀerential equations which which are needed to model diﬀusion, wave motion, equilibria of membranes and so forth. The behavior of physical systems often suggest theorems which can be proven via rigorous mathematics. (This last point, and the book itself, can be best appreciated by those who have taken a course in rigorous mathematical proof, such as a course in mathematical inquiry, whether at the high school or university level.) Moreover, the objects modeled make it clear that there should be a constant tension between the discrete and continuous. For example, a vibrating string can be regarded proﬁtably as a continuous object, yet if one looks at a ﬁne enough scale, the string is made up of molecules, suggesting a discrete model with a large number of variables. Moreover, we will see that although a partial diﬀerential equation provides an elegant continuous model for a vibrating membrane, the numerical method used to do actual calculations may approximate this continuous model with a discrete mechanical system with a large number of degrees of freedom. The eigenvalue problem for a diﬀerential equation thereby becomes approximated by an eigenvalue problem for an n × n matrix where n is large, thereby providing a link between the techniques studied in linear algebra and those of partial diﬀerential equations. The reader should be aware that there are many cases in which a discrete model may actually provide a better description of the phenomenon under study than a continuous one. One should also be aware that probabilistic techniques provide an additional component to model building, alongside the partial diﬀerential equations and discrete mechanical systems with many degrees of freedom described in these pages. There is a major dichotomy that runs through the subject—linear versus nonlinear. It is actually linear partial diﬀerential equations for which the technique of linear algebra prove to be so eﬀective. This book is concerned primarly with linear partial diﬀerential equations—yet it is the nonlinear partial diﬀerential equations that provide the most intriguing questions for research. Nonlinear partial diﬀerential equations include the Einstein ﬁeld equations from general relativity and the Navier-Stokes equations which describe ﬂuid motion. We hope the linear theory presented here will whet the student’s appetite for studying the deeper waters of the nonlinear theory. The author would appreciate comments that may help improve the next version of this short book. He hopes to make a list of corrections available at the web site: http://www.math.ucsb.edu/~moore Doug Moore March, 2003

ii

. . . .2 Inner products . . . . . . . 3. . 2. . . . of freedom* . . . 2. . . . . . . . . . . . . . . . . .Contents The sections marked with asterisks are less central to the main line of discussion. . . . . . . . . . . 4. . . . . . 4. . . . . . . . . . . .3 Numerical solutions to the heat equation . . 107 . .5 Fourier transforms* . . . . . . . . . . . . . . 4. 111 iii . . . . . . . . . . . . . 1. . . . . . . . . . . . .2 Solving diﬀerential equations by means 1. .6 Heat ﬂow in a circular wire . . . . . . . . . . . . . . . . . . . . . . . . .1 What is a power series? . . . . . . . . . . . . . . . . . . 81 . . . . . . . . . . . . . . . . . . . . . . 98 . . . . . . . . . . . . 2 Symmetry and Orthogonality 2. . of power series . 3. . 92 .3 Singular points . . . . . 2. . . . . . . . . . . . . . . . . .4 Mechanical systems . . . .1 Eigenvalues of symmetric matrices . 1 1 7 15 22 29 29 36 44 49 53 59 62 62 69 72 77 79 . . . . . . . . . . . . . . 4. . . . . . . . . 85 . . . . . . . . . . . . . . . . . . .7 Sturm-Liouville Theory* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Orthonormal bases . . . . . . . . . 4. . 4 Partial Diﬀerential Equations 4. . . . . .6 A linear array of weights and springs* 3 Fourier Series 3. . . . . . . . . . .3 Fourier sine and cosine series . . . . . . 1 Power Series 1. . . . .5 The initial value problem for the vibrating string 4. . . . . . .2 The initial value problem for the heat equation . . .4 Bessel’s diﬀerential equation . . . . . . 4. . . . 3. . . . . . .5 Mechanical systems with many degrees 2. . . . . . . . . . . .4 Complex version of Fourier series* 3. . . . . . 1. . 2. . . . . . . . . . 103 . . . . . .4 The vibrating string . . . . . . . .8 Numerical solutions to the eigenvalue problem* . . . 81 . . . 94 . . . . . . . . . . . .1 Overview . . . . and may be treated brieﬂy or omitted if time runs short. . .1 Fourier series . . . . . . . . . . . . . . . . . . .2 Conic sections and quadric surfaces . . . . . . . . . . . . . . . . .

. . .5 PDE’s in Higher Dimensions 5. . . . . . . . . . . . . . 5. . . . . . . . . .8 Eigenvalues of the disk . . . . . . . . 115 115 118 123 129 134 136 141 144 150 156 163 iv . . . . . . . . 5. . . . . . . . . .3 Initial value problems for heat equations . 5. . . . . . . . . 5. 5. . . . . . . . . . . . . . . . . A Using Mathematica to solve diﬀerential equations Index . . . . . .7 Eigenvalues of the Laplace operator . . . . . . . . . . . . . . . . . . .4 Two derivations of the wave equation . . . . . . 5. . . .5 Initial value problems for wave equations . . .2 The Dirichlet problem .9 Fourier analysis for the circular vibrating membrane* . . 5. . . . . . . . . . . . .6 The Laplace operator in polar coordinates . .1 The three most important linear partial diﬀerential equations 5. .

1) as well as the trigonometric functions cos x = 1 − and sin x = x − 1 3 1 x + x5 − · · · = 3! 5! 1 2 1 x + x4 − · · · = 2! 4! ∞ (−1)k k=0 1 2k x (2k)! (−1)k k=0 1 x2k+1 . a power series centered at x0 is an inﬁnite sum of the form ∞ a0 + a1 (x − x0 ) + a2 (x − x0 )2 + · · · = n=0 an (x − x0 )n . (2k + 1)! An inﬁnite series of this type is called a power series. these power series representations are often used to deﬁne the exponential and trigonometric functions. In advanced treatments of calculus. where the an ’s are constants. 2! 3! n! n=0 ∞ ∞ (1.1 What is a power series? Functions are often represented eﬃciently by means of inﬁnite series. Power series can also be used to construct tables of values for these functions. For example. we could calculate 1+1+ 1 2 1 n 1 = 1 = 2. To be precise.Chapter 1 Power Series 1.5. Examples we have seen in calculus include the exponential function ex = 1 + x + 1 n 1 2 1 x + x3 + · · · = x . 2! n! n=0 1 2 . using a calculator or PC with suitable software installed (such as Mathematica).

1+1+
8

1 n 1 2 1 1 1 + 13 + 14 = 1 = 2.70833, 2! 3! 4! n! n=0 1 n 1 = 2.71828, n! n=0
12

4

1 n 1 = 2.71806, n! n=0

...

As the number of terms increases, the sum approaches the familiar value of the exponential function ex at x = 1. For a power series to be useful, the inﬁnite sum must actually add up to a ﬁnite number, as in this example, for at least some values of the variable x. We let sN denote the sum of the ﬁrst (N + 1) terms in the power series,
N

sN = a0 + a1 (x − x0 ) + a2 (x − x0 )2 + · · · + aN (x − x0 )N =
n=0

an (x − x0 )n ,

and say that the power series

an (x − x0 )n
n=0

converges if the ﬁnite sum sN gets closer and closer to some (ﬁnite) number as N → ∞. Let us consider, for example, one of the most important power series of applied mathematics, the geometric series

1 + x + x2 + x3 + · · · =
n=0

xn .

In this case we have sN = 1 + x + x2 + x3 + · · · + xN , sN − xsN = 1 − xN +1 , xsN = x + x2 + x3 + x4 · · · + xN +1 , sN = 1 − xN +1 . 1−x

If |x| < 1, then xN +1 gets smaller and smaller as N approaches inﬁnity, and hence lim xN +1 = 0.
N →∞

Substituting into the expression for sN , we ﬁnd that
N →∞

lim sN =

1 . 1−x

Thus if |x| < 1, we say that the geometric series converges, and write

xn =
n=0

1 . 1−x

2

On the other hand, if |x| > 1, then xN +1 gets larger and larger as N approaches inﬁnity, so limN →∞ xN +1 does not exist as a ﬁnite number, and neither does limN →∞ sN . In this case, we say that the geometric series diverges. In summary, the geometric series

xn
n=0

converges to

1 1−x

when |x| < 1,

and diverges when |x| > 1. This behaviour, convergence for |x| < some number, and divergences for |x| > that number, is typical of power series: Theorem. For any power series

a0 + a1 (x − x0 ) + a2 (x − x0 ) + · · · =
2 n=0

an (x − x0 )n ,

there exists R, which is a nonnegative real number or ∞, such that 1. the power series converges when |x − x0 | < R, 2. and the power series diverges when |x − x0 | > R. We call R the radius of convergence. A proof of this theorem is given in more advanced courses on real or complex analysis.1 We have seen that the geometric series

1 + x + x2 + x3 + · · · =
n=0

xn

has radius of convergence R = 1. More generally, if b is a positive constant, the power series x x 1+ + b b
2

x + b

3

+ ··· =
n=0

x b

n

(1.2)

has radius of convergence b. To see this, we make the substitution y = x/b, ∞ and the power series becomes n=0 y n , which we already know converges for |y| < 1 and diverges for |y| > 1. But |y| < 1 |y| > 1 ⇔ ⇔ x <1 b x >1 b ⇔ ⇔ |x| < b, |x| > b.

1 Good references for the theory behind convergence of power series are Edward D. Gaughan, Introduction to analysis, Brooks/Cole Publishing Company, Paciﬁc Grove, 1998 and Walter Rudin, Principles of mathematical analysis, third edition, McGraw-Hill, New York, 1976.

3

Thus for |x| < b the power series (1.2) converges to 1 1 b = = , 1−y 1 − (x/b) b−x while for |x| > b, it diverges. There is a simple criterion that often enables one to determine the radius of convergence of a power series. Ratio Test. The radius of convergence of the power series

a0 + a1 (x − x0 ) + a2 (x − x0 )2 + · · · =
n=0

an (x − x0 )n

is given by the formula R = lim so long as this limit exists.
n→∞

|an | , |an+1 |

Let us check that the ratio test gives the right answer for the radius of convergence of the power series (1.2). In this case, we have an = 1 , bn so |an | bn+1 1/bn = n = b, = |an+1 | 1/bn+1 b

and the formula from the ratio test tells us that the radius of convergence is R = b, in agreement with our earlier determination. In the case of the power series for ex , 1 n x , n! n=0 in which an = 1/n!, we have |an | 1/n! (n + 1)! = = = n + 1, |an+1 | 1/(n + 1)! n! and hence R = lim
n→∞ ∞

|an | = lim (n + 1) = ∞, |an+1 | n→∞

so the radius of convergence is inﬁnity. In this case the power series converges for all x. In fact, we could use the power series expansion for ex to calculate ex for any choice of x. On the other hand, in the case of the power series

n!xn ,
n=0

4

If |an | ≥ |bn | for all n. we have |an | n! 1 = = . 1 + 0x − 1 2 1 x + 0x3 + x4 − · · · = 2! 4! ∞ (−1)k k=0 1 2k x . the radius of convergence is zero. (2k)! In this case the coeﬃcient an is zero when n is odd. and the power series does not converge for any nonzero x. then R1 ≥ R2 . Suppose that the power series ∞ ∞ an (x − x0 )n . and hence must also be inﬁnite.in which an = n!. A function f (x) is said to be real analytic at x0 if there is a power series ∞ an (x − x0 )n n=0 about x0 with positive radius of convergence R such that ∞ f (x) = n=0 an (x − x0 )n . 5 . for |x − x0 | < R. then R1 ≤ R2 . If |an | ≤ |bn | for all n. Power series with positive radius of convergence are so important that there is a special term for describing functions which can be represented by such power series. but sometimes one can use it in conjunction with the Comparison Test. we have |an | ≤ 1/n!. Consider for example the power series expansion for cos x. while an = ±1/n! when n is even. power series with smaller coeﬃcients have larger radius of convergence. In short. |an+1 | (n + 1)! n+1 R = lim n→∞ |an | = lim |an+1 | n→∞ 1 n+1 = 0. The ratio test doesn’t always work because the limit may not exist. In either case. n=0 n=0 bn (x − x0 )n have radius of convergence R1 and R2 respectively. It follows from the comparison test that the radius of convergence of ∞ (−1)k k=0 1 2k x (2k)! must be at least large as that of the power series for ex . In this case. Thus we can compare with the power series ∞ 1 2 1 n 1 3 1 4 1 + x + x + x + x + ··· = x 2! 3! 4! n! n=0 which represents ex and has inﬁnite radius of convergence.

is analytic at x0 if and only if x0 is not a zero of the denominator Q(x). (2k)! ∞ b.1) with x replaced by x − x0 : ex = ex0 1 an (x − x0 )n . but fails to be analytic at x0 = 1. 2n n=0 1 (3x − 6)n .1. n=0 (7x − 14)n . 1 (x − π)n . (−1)n xn .For example. the functions ex is real analytic at any x0 . In more advanced courses. Thus for example. n! n=0 ∞ ∞ e. a power series about x0 in which all but ﬁnitely many of the coeﬃcients are zero. k=0 ∞ 1 2k x . Use the ratio test to ﬁnd the radius of convergence of the following power series: ∞ ∞ 1 a. b. k=0 . Exercises: 1. n! This is a power series expansion of ex about x0 with inﬁnite radius of convergence. 1 (x − π)2k . For our purposes. it is suﬃcient to be aware of the following facts: The sum and product of real analytic functions is real analytic. 22k c. k=0 1 (x − 4)2k 2k 6 d. Use the comparison test to ﬁnd an estimate for the radius of convergence of each of the following power series: ∞ a. To see this. f. (x − x0 )n = n! n=0 n=0 ∞ ∞ where an = ex0 . one studies criteria under which functions are real analytic. 1. xn .1. The quotient of two polynomials with no common factors. k=0 ∞ (−1)k x2k .2. we utilize the law of exponents to write ex = ex0 ex−x0 and apply (1. It follows from this that any polynomial P (x) = a0 + a1 x + a2 x2 + · · · + an xn is analytic at any x0 . 3 (x − 2)n . P (x)/Q(x). 1/(x − 1) is analytic whenever x0 = 1. n+1 n=0 n=0 c. Similarly. n+1 n=0 ∞ ∞ d.1. the monomial function f (x) = xn is real analytic at x0 because n n! xn = (x − x0 + x0 )n = xn−i (x − x0 )i i!(n − i)! 0 i=0 by the binomial theorem.

d. x−4 x −1 4 1 c. g(x) = 2 . the equation of simple harmonic motion d2 y + y = 0. Suppose for the moment that we don’t know the general solution and want to ﬁnd it by means of power series. 2 (m!) 2 m=0 1.1.4. Determine the values of x0 at which the following functions fail to be real analytic: 1 x a. = a1 + 2a2 x + 3a3 x2 + · · · = dx n=1 (Note that the last summation only goes from 1 to ∞. dx2 m+2=2 m=0 7 ∞ ∞ ∞ ∞ . Use the comparison test and the ratio test to ﬁnd the radius of convergence of the power series ∞ x 2m 1 (−1)m . f (x) = . since the term with n = 0 drops out of the sum. (1.3.3) which we have already learned how to solve by other methods.1.2 Solving diﬀerential equations by means of power series Our main goal in this chapter is to study how to determine solutions to diﬀerential equations by means of power series.4) It can be shown that the standard technique for diﬀerentiating polynomials term by term also works for power series. h(x) = 4 . φ(x) = 3 x − 3x2 + 2 x − 5x2 + 6x 1.1. We could start by assuming that ∞ y = a0 + a1 x + a2 x2 + a3 x3 + · · · = n=0 an xn .) Diﬀerentiating again yields d2 y = 2a2 + 3 · 2a3 x + 4 · 3a4 x2 + · · · = n(n − 1)an xn−2 . b. we consider our old friend. As an example. so we expect that dy nan xn−1 . dx2 n=2 We can replace n by m + 2 in the last summation so that d2 y = (m + 2)(m + 2 − 1)am+2 xm+2−2 = (m + 2)(m + 1)am+2 xm . dx2 (1.

3) yields ∞ ∞ ∞ (n + 2)(n + 1)an+2 xn + n=0 n=0 an xn = 0.5) with n = 0 that a2 = − a0 1 = − a0 . n=0 Recall that a polynomial is zero only if all its coeﬃcients are zero. it follows from (1.5) with n = 1 that a3 = − and with n = 2 that a4 = − a2 1 1 1 = a0 = a0 . Thus we are free to replace m by n and obtain the formula d2 y = (n + 2)(n + 1)an+2 xn . dy (0) = a1 . a power series can be zero only if all of its coeﬃcients are zero. 2·1 2 Similarly. Similarly.5) This is called a recursion formula for the coeﬃcients an . (2n)! a2n+1 = 8 . dx2 n=0 Substitution into equation(1. (2n + 1)! a1 1 = − a1 . y(0) = a0 . Indeed it follows from (1. or ∞ [(n + 2)(n + 1)an+2 + an ]xn = 0. dx Then the recursion formula can be used to determine the remaining coeﬃcients by the process of induction. (n + 2)(n + 1) (1. The ﬁrst two coeﬃcients a0 and a1 in the power series can be determined from the initial conditions. or an+2 = − an . It follows that (n + 2)(n + 1)an+2 + an = 0. 4·3 4·32 4! (−1)n a1 . 3·2 3! Continuing in this manner. we ﬁnd that a2n = (−1)n a0 .The index m is a “dummy variable” in the summation and can be replaced by any other letter.

neither of which is a constant multiple of the other. It is proven in books on diﬀerential equations that if P (x) and Q(x) are wellbehaved functions. In the special case where the functions P (x) and Q(x) are real analytic. which is proven in more advanced books on diﬀerential equations: Theorem. As a0 and a1 range over all constants. We say that y0 (x) and y1 (x) form a basis for the space of solutions. The method we have described—assuming a solution to the diﬀerential equation of the form ∞ y(x) = n=0 an xn and solve for the coeﬃcients an —is surprisingly eﬀective. particularly for the class of equations called second-order linear diﬀerential equations. Here y0 (x) and y1 (x) are any two nonzero solutions.Substitution into (1. This is the content of the following theorem. y = a0 cos x + a1 sin x.4) yields y = a0 + a1 x − = a0 1 − 1 1 1 a0 x2 − a1 x3 + a0 x4 + · · · 2! 3! 4! + a1 x − 1 3 1 x + x5 − · · · 3! 5! . called the general solution. the solutions y0 (x) and y1 (x) will also be real analytic. then the solutions to the “homogeneous linear diﬀerential equation” d2 y dy + Q(x)y = 0 + P (x) dx2 dx can be organized into a two-parameter family y = a0 y0 (x) + a1 y1 (x). In the terminology used in linear algebra. 1 2 1 x + x4 − · · · 2! 4! We recognize that the expressions within parentheses are power series expansions of the functions sin x and cos x. Q(x) = n=0 qn (x − x0 )n with positive radii of convergence R1 and R2 respectively. we say that they are linearly independent solutions. If the functions P (x) and Q(x) can be represented by power series ∞ ∞ P (x) = n=0 pn (x − x0 )n . then any solution y(x) to the linear diﬀerential equation d2 y dy + Q(x)y = 0 + P (x) dx2 dx 9 . y ranges throughout a “linear space” of solutions. and hence we obtain the familiar expression for the solution to the equation of simple harmonic motion.

6) where p is a parameter.7) (1. both functions being polynomials. but for the moment. P (x) = −2x. It turns out that this equation is very useful for treating the simple harmonic oscillator in quantum mechanics. = dx n=1 ∞ d2 y = n(n − 1)an xn−2 . we write ∞ y= n=0 an xn . We diﬀerentiate term by term as before. Indeed. = dx n=0 ∞ ∞ ∞ ∞ (1.8) 10 . dx2 n=2 ∞ Once again. so that d2 y = (n + 2)(n + 1)an+2 xn . As in the case of the equation of simple harmonic motion. hence power series about x0 = 0 with inﬁnite radius of convergence.can be represented by a power series ∞ y(x) = n=0 an (x − x0 )n . Example. dx2 m+2=2 m=0 and then replace m by n once again. Q(x) = 2p. we can regard it as merely an example of an equation to which the previous theorem applies. dx2 dx (1. dx2 n=0 Note that −2x dy −2nan xn . This theorem is used to justify the solution of many well-known diﬀerential equations by means of the power series method. whose radius of convergence is ≥ the smallest of R1 and R2 . Hermite’s diﬀerential equation is d2 y dy − 2x + 2py = 0. we can replace n by m + 2 in the last summation so that d2 y = (m + 2)(m + 2 − 1)am+2 xm+2−2 = (m + 2)(m + 1)am+2 xm . and obtain dy nan xn−1 .

dy y(0) = a0 . each coeﬃcient must be zero.8) and (1. + 2py = dx2 dx n=0 n=0 If y satisﬁes Hermite’s equation. 2·1 Similarly.while ∞ 2py = n=0 2pan xn . so (n + 2)(n + 1)an+2 + (−2n + 2p)an = 0.9) Adding together (1. we ﬁnd that a5 = 6 − 2p 2(3 − p) 2(1 − p) 22 (p − 1)(p − 3) a3 = a1 = a1 .7). Since the right-hand side is zero for all choices of x.10) with n = 1 that a3 = and with n = 2 that a4 = − 4 − 2p 2(2 − p) −2p 22 p(p − 2) a2 = a0 = a0 . it follows from (1. dx The recursion formula can be used to determine the remaining coeﬃcients in the power series. and we obtain the recursion formula for the coeﬃcients of the power series: an+2 = 2n − 2p an . (n + 2)(n + 1) (1. the ﬁrst two coeﬃcients a0 and a1 in the power series can be determined from the initial conditions. we obtain d2 y dy − 2x (n + 2)(n + 1)an+2 xn + (−2n + 2p)an xn . Indeed it follows from (1. (0) = a1 .10) with n = 0 that a2 = − 2p a0 . 4·3 4·3 2 4! 2 − 2p 2(p − 1) a1 = − a1 . (1. we must have ∞ ∞ ∞ 0= n=0 [(n + 2)(n + 1)an+2 (−2n + 2p)an ]xn .10) Just as in the case of the equation of simple harmonic motion. (1.9). 5·4 5·4 3! 5! 11 . 3·2 3! Continuing in this manner.

This equation is very useful for treating spherically symmetric potentials in the theories of Newtonian gravitation and in electricity and magnetism. we need to divide by 1 − x2 to obtain d2 y p(p + 1) 2x dy + − y = 0. Thus we ﬁnd that a6 = y = a0 1 − 2p 2 22 p(p − 2) 4 23 p(p − 2)(p − 4) 6 x + x − x + ··· 2! 4! 6! +a1 x − 2(p − 1) 3 22 (p − 1)(p − 3) 5 x + x 3! 5! − 23 (p − 1)(p − 3)(p − 5) 7 x + ··· . These polynomial solutions are known as Hermite polynomials. we say that y0 (x) and y1 (x) form a basis for the space of solutions to Hermite’s equation. Another Example. To apply our theorem.8 − 2p 2(3 − p) 22 (p − 2)p 23 p(p − 2)(p − 4) a4 = a0 = − a0 . one of the two power series will collapse. − 2x dx2 dx (1.” In the language of linear algebra. 6·5·2 6·5 4! 6! and so forth. yielding a polynomial solution to Hermite’s equation. 7! We can now write the general solution to Hermite’s equation in the form y = a0 y0 (x) + a1 y1 (x). 1 − x2 . Legendre’s diﬀerential equation is (1 − x2 ) d2 y dy + p(p + 1)y = 0. Tables of values for these functions are found in many ”handbooks of mathematical functions. where y0 (x) = 1 − and y1 (x) = x − 2(p − 1) 3 22 (p − 1)(p − 3) 5 23 (p − 1)(p − 3)(p − 5) 7 x + x − x +··· . 3! 5! 7! 2p 2 22 p(p − 2) 4 23 p(p − 2)(p − 4) 6 x + x − x + ··· 2! 4! 6! For a given choice of the parameter p. When p is a positive integer. 2 2 dx dx 1−x 1 − x2 Thus we have P (x) = − 2x . we could use the power series to construct tables of values for the functions y0 (x) and y1 (x). 1 − x2 12 Q(x) = p(p + 1) .11) where p is a parameter.

b. 1 − x2 Both of these functions have power series expansions about x0 = 0 which converge for |x| < 1. Indeed. for 2 ≤ n ≤ 9. ﬁnd the recursion formula for an+2 in terms of an . However. we can conclude that 1 = 1 + x2 + x4 + x6 + · · · . dx2 dx which is very similar to Hermite’s equation. a.2. Legendre’s diﬀerential equation has a nonzero polynomial solution which is well-behaved for all x. 1 − x2 the power series converging when |x| < 1. and determine the coeﬃcients an . Experimentation with numerical solutions to Legendre’s equation would show that these suspicions are justiﬁed—solutions to Legendre’s equation will usually blow up as x → ±1.1. It follows quickly that P (x) = − and 2x = −2x − 2x3 − 2x5 − · · · 1 − x2 p(p + 1) = p(p + 1) + p(p + 1)x2 + p(p + 1)x4 + · · · . Use the recursion formula to determine an in terms of a0 and a1 . it can be shown that when p is an integer. We would like to use the power series method to ﬁnd the general solution to the diﬀerential equation d2 y dy − 4x + 12y = 0. Q(x) = Exercises: 1. If we substitute u = x2 . c. As a ﬁrst step. 13 .Now from the preceding section. we know that the power series 1 + u + u2 + u3 + · · · converges to 1 1−u for |u| < 1. we might suspect that the solutions to Legendre’s equation to exhibit some unpleasant behaviour near x = ±1. Hence our theorem implies that any solution to Legendre’s equation will be expressible as a power series about x0 = 0 which converges for |x| < 1. a power series centered at 0. So we assume the solution is of the form ∞ y= n=0 an xn . The coeﬃcients a0 and a1 will be determined by the initial conditions. but solutions which are not constant multiples of these Legendre polynomials blow up as x → ±1. Find a nonzero polynomial solution to this diﬀerential equation.

e. The diﬀerential equation (1 − x2 ) d2 y dy −x + p2 y = 0. dy (0) = 1. d. a. dx y= n=0 an (x − 3)n . in the case where p = 3. 1 − x2 . ﬁnd the recursion formula for an+2 in terms of an . 1 − x2 Q(x) = p2 . Find a basis for the space of solutions to the equation. + P (x) 2 dx dx where 14 P (x) = − x . dx2 dx 1. − 2x 2 dx dx where p is a constant. b. dx2 dx f. It can be rewritten in the form d2 y dy + Q(x)y = 0. is known as Chebyshev’s equation. for 2 ≤ n ≤ 9.2. 2 dx dx it would be most natural to assume that the solution has the form ∞ y(0) = 0. We want to use the power series method to ﬁnd the general solution to Legendre’s diﬀerential equation (1 − x2 ) d2 y dy − 2x + p(p + 1)y = 0. As a ﬁrst step. Find the solution to the initial value problem d2 y dy − 4x + 12y = 0. Use the recursion formula to determine an in terms of a0 and a1 . c. dx2 dx d2 y dy + 12y = 0. Use this idea to ﬁnd a polynomial solution to the diﬀerential equation d2 y dy − 4(x − 3) + 12y = 0. To solve the diﬀerential equation d2 y dy − 4(x − 3) + 12y = 0.2.3.d. Find a nonzero polynomial solution to this diﬀerential equation.2. Find a basis for the space of solutions to the diﬀerential equation (1 − x2 ) 1. dx2 dx Once again our approach is to assume our solution is a power series centered at 0 and determine the coeﬃcients in this power series.

The diﬀerential equation − d2 + x2 z = λz dx2 (1. Show that making the substitution z = e−x Hermite’s diﬀerential equation 2 /2 y transforms this equation into d2 y dy − 2x + (λ − 1)y = 0. e. A singular point is said to be regular if (x − x0 )P (x) and 15 (x − x0 )2 Q(x) . A point x0 is called an ordinary point for the diﬀerential equation d2 y dy + P (x) + Q(x)y = 0 2 dx dx (1. Use the recursion formula to determine an in terms of a0 and a1 . a.12) has a solution 2 of the form z = e−x /2 Pn (x). In the opposite case. Find a basis for the space of solutions to (1 − x2 ) 1. what is the radius of convergence of these power series? b.13) if the coeﬃcients P (x) or Q(x) are both real analytic at x = x0 . where Pn (x) is a polynomial. In the special case where p = 3. 2 dx dx b. Show that if λ = 2n+1 where n is a nonnegative integer. 1.a. ﬁnd a nonzero polynomial solution to this diﬀerential equation.3 Singular points Our ultimate goal is to give a mathematical description of the vibrations of a circular drum. we say that x0 is a singular point. ﬁnd the recursion formula for an+2 in terms of an . we will need to solve Bessel’s equation. −x dx2 dx arises when treating the quantum mechanics of simple harmonic motion.2. both P (x) or Q(x) have power series expansions about x = x0 with positive radius of convergence.4. thus x0 is a singular point if at least one of the coeﬃcients P (x) or Q(x) fails to be real analytic at x = x0 . For this. (1. c. or equivalently. a second-order homogeneous linear diﬀerential equation with a “singular point” at 0. for 2 ≤ n ≤ 9. Assuming a power series centered at 0. d. If P (x) and Q(x) are represented as power series about x0 = 0.12) d2 y dy + 9y = 0.

New York. not a basis.13) in the case of regular singular points is called the Frobenius method . Simmons.14) a generalized power series solution. Diﬀerential equations with applications and historical notes. 1991.2 2 For more discussion of the Frobenius method as well as many of the other techniques touched upon in this chapter we refer the reader to George F. (1. but not always. one can ﬁnd a basis of generalized power series solutions. We will call a solution of the form (1. but it is a regular singular point because (x − 1)P (x) = (x − 1) and (x − 1)2 Q(x) = (x − 1)2 p(p + 1) p(p + 1)(1 − x) = 1 − x2 1+x are both real analytic at x0 = 1.are real analytic. The point of these deﬁnitions is that in the case where x = x0 is a regular singular point. If x0 is a regular singular point for the diﬀerential equation d2 y dy + P (x) + Q(x)y = 0. dx2 1 − x2 dx 1 − x2 because 1 − x2 → 0 as x → 1 and hence the quotients 2x 1 − x2 and p(p + 1) 1 − x2 −2x 2x = 2 1−x x+1 blow up as x → 1. which may be complex. 16 . Theorem of Frobenius. 2 dx dx then this diﬀerential equation has at least one nonzero solution of the form ∞ y(x) = (x − x0 )r n=0 an (x − x0 )n .14) where r is a constant. If (x−x0 )P (x) and (x−x0 )2 Q(x) have power series which converge for |x − x0 | < R then the power series ∞ an (x − x0 )n n=0 will also converge for |x − x0 | < R. The method of ﬁnding generalized power series solutions to (1. a modiﬁcation of the power series method can still be used to ﬁnd solutions. In fortuitous cases. For example. the theorem guarantees only one generalized power series solution. Unfortunately. x0 = 1 is a singular point for Legendre’s equation d2 y p(p + 1) 2x dy + − y = 0. second edition. McGraw-Hill.

the general solution is given by the superposition principle as y = c1 xr1 + c2 xr2 .13) for which p q P (x) = . Often there will be two linearly independent solutions y1 (x) = xr1 and y2 (x) = xr2 of this special form. Indeed. + 4x 2 dx dx we set y = xr and diﬀerentiate to show that dy/dx = rxr−1 d2 y/dx2 = r(r − 1)xr−2 ⇒ x(dy/dx) = rxr . 17 . so both of these solutions are generalized power series solutions. substitute into the equation and solve for r. n=0 y2 (x) = x−2 ∞ bn xn . For example. Substitution into the diﬀerential equation yields r(r − 1)xr + 4rxr + 2xr = 0. and dividing by xr yields r(r − 1) + 4r + 2 = 0 or r2 + 3r + 2 = 0. so the general solution to the diﬀerential equation is y = c1 x−1 + c2 x−2 = c1 c2 + 2. The Frobenius method is quite simple in the case of Cauchy-Euler equations. in this case. x x Note that the solutions y1 (x) = x−1 and y2 (x) = x−2 can be rewritten in the form y1 (x) = x−1 ∞ an xn . so x = 0 is a regular singular point for the Cauchy-Euler equation as long as either p or q is nonzero. The roots to this equation are r = −1 and r = −2. x x where p and q are constants. ⇒ x2 (d2 y/dx2 ) = r(r − 1)xr . we can simply take y(x) = xr . Note that xP (x) = p and x2 Q(x) = q are real analytic. Q(x) = 2 . to solve the diﬀerential equation x2 d2 y dy + 2y = 0. This is the special case of (1. In this case.The simplest diﬀerential equation to which the Theorem of Frobenius applies is the Cauchy-Euler equidimensional equation. n=0 where a0 = b0 = 1 and all the other an ’s and bn ’s are zero.

only one of the basis elements in the general solution is a generalized power series.On the other hand. there is a trick that enables us to handle this situation. v x log |v | = − log |x| + a. dx d2 y = v (x)x−1 − 2v (x)x−2 + 2v(x)x−3 . x where a and c2 are constants of integration. which has a repeated root. we replace the parameter c by a variable v(x) and write y = v(x)x−1 . and we obtain the general solution y = c1 1 log |x| + c2 . Then dy = v (x)x−1 − v(x)x−2 . if this method is applied to the diﬀerential equation x2 we obtain r(r − 1) + 3r + 1 = r2 + 2r + 1. 2x Q(x) = 1 . In this context. which quickly simpliﬁes to yield xv (x) + v (x) = 0.15) . In this case. In this case. x x so y = (c2 log |x| + c1 )x−1 . the so-called method of variation of parameters. v 1 =− . For equations which are not of Cauchy-Euler form the Frobenius method is more involved. v = c2 . 2x d2 y dy + + y = 0. + 3x 2 dx dx Substitution into the diﬀerential equation yields x2 (v (x)x−1 − 2v (x)x−2 + 2v(x)x−3 ) + 3x(v (x)x−1 − v(x)x−2 ) + v(x)x−1 = 0. A further integration yields v = c2 log |x| + c1 . Let us consider the example 2x which can be rewritten as d2 y dy + P (x) + Q(x)y = 0. 2 dx dx (1. we obtain only a one-parameter family of solutions y = cx−1 . dx2 d2 y dy + y = 0. Fortunately. dx2 dx where 18 P (x) = 1 .

then all the coeﬃcients must be zero from the second of these equations. obtaining (2r − 1)ra0 x−1 + ∞ ∞ (2n + 2r + 1)(n + r + 1)an+1 xn + n=0 n=0 an xn = 0. We begin the Frobenius method by assuming that the solution has the form ∞ ∞ y = xr n=0 an xn = n=0 an xn+r . this becomes (2r − 1)ra0 x−1 + ∞ ∞ (2m + 2r + 1)(m + r + 1)am+1 xm + m=0 n=0 an xn = 0. Finally. If we let n = m + 1 in the ﬁrst inﬁnite sum. So we must have a0 = 0 and hence (2r − 1)r = 0. Then dy (n + r)an xn+r−1 . We can divide by xr . The coeﬃcient of each power of x must be zero. = dx n=0 and ∞ ∞ d2 y = (n + r)(n + r − 1)an xn+r−2 dx2 n=0 ∞ d2 y 2x 2 = 2(n + r)(n + r − 1)an xn+r−1 . we replace m by n. and we don’t get a nonzero solution.16) If a0 = 0. In particular. which simpliﬁes to ∞ ∞ xr n=0 (2n + 2r − 1)(n + r)an xn−1 + n=0 an xn = 0. we must have (2r − 1)ra0 = 0. 19 . and separate out the ﬁrst term from the ﬁrst summation. dx n=0 ∞ ∞ Substitution into the diﬀerential equation yields ∞ 2(n + r)(n + r − 1)an xn+r−1 + n=0 n=0 (n + r)an xn+r−1 + n=0 an xn+r = 0. obtaining (2r − 1)ra0 x−1 + ∞ ∞ (2n + 2r − 1)(n + r)an xn−1 + n=1 n=0 an xn = 0. (1.One easily checks that x = 0 is a regular singular point. (2n + 2r + 1)(n + r + 1)an+1 + an = 0.

(2n + 1)(n + 1) 1 1 a2 = − . the recursion formula becomes an+1 = − 1 1 an = − an . (2n + 2)(n + (1/2) + 1) (n + 1)(2n + 3) Given a0 = 1. we would have an = (−1)n 1 . (2n + 2r + 1)(n + r + 1) for n ≥ 0.16) yields the recursion formula an+1 = − 1 an . n!(2n + 1)(2n − 1) · · · 1 · n! 20 . We can try to ﬁnd a generalized power series solution for either root of the indicial equation. r2 = 1 . If r = 0. 5·3 (5 · 3)(3 · 2) and so forth. a3 = − a2 = − 1 1 a1 = . it has two roots r1 = 0. 3·2 3·2 a4 = − 1 1 a3 = . 2 The second half of (1.15) is y1 (x) = x0 1 − x + =1−x+ 1 2 1 1 x − x3 + x4 − · · · 3·2 (5 · 3)(3!) (7 · 5 · 3)4! 1 2 1 1 x − x3 + x4 − · · · . In this case. the recursion formula becomes an+1 = − Given a0 = 1. an = (−1)n a2 = − 1 1 a1 = . 3·7 3! · (7 · 5 · 3) 1 . we ﬁnd that 1 a1 = − . 3 a3 = − and in general. 2·5 2·5·3 1 1 a2 = − . we ﬁnd that a1 = −1.This is called the indicial equation. In general. 3·2 (5 · 3)(3!) (7 · 5 · 3)4! If r = 1/2. (2n − 1)(2n − 3) · · · 1 · n! One of the generalized power series solution to (1. 7·4 (7 · 5 · 3)4! 1 an .

2. x2 d2 y/dx2 − xdy/dx + y = 0. We want to ﬁnd generalized power series solutions to the diﬀerential equation d2 y dy 3x 2 + +y =0 dx dx 21 . In that case.3. but this does not always happen.3. y + xy + (1 − x2 )y = 0.) 1. Find the general solution to each of the following Cauchy-Euler equations: a. x3 y + 2xy + (cos x)y = 0. y + (1/x)y + (1 − (1/x2 ))y = 0. d. c. determine whether x = 0 is ordinary or singular. If it is singular. just as we saw in the case of the Cauchy-Euler equidimensional equation. b. Exercises: 1. b. determine whether it is regular or not. 1. For each of the following diﬀerential equations. c.15) is a superposition of y1 (x) and y2 (x): y = c1 1 − x + 1 2 1 1 x − x3 + x4 − · · · 3·2 (5 · 3)(3!) (7 · 5 · 3)4! √ 1 1 1 +c2 x 1 − x + x2 − x3 + · · · .15): 1 1 1 y2 (x) = x1/2 1 − x + x2 − x3 + · · · . x2 d2 y/dx2 − 2xdy/dx + 2y = 0. x2 d2 y/dx2 − xdy/dx + 10y = 0. we may obtain only one generalized power series solution. 3 2·5·3 3! · (7 · 5 · 3) We obtained two linearly independent generalized power series solutions in this case. a.1.3. 3 2·5·3 3! · (7 · 5 · 3) The general solution to (1. If the roots of the indicial equation diﬀer by an integer.3. (Hint: Use the formula xa+bi = xa xbi = xa (elog x )bi = xa eib log x = xa [cos(b log x) + i sin(b log x)] to simplify the answer.We thus obtain a second generalized power series solution to (1. a second independent solution can then be found by variation of parameters. x2 y + 2xy + (cos x)y = 0.

4 Bessel’s diﬀerential equation d dx dy dx + (x2 − p2 )y = 0. a. Here p is a parameter. 1 x and Q(x) = x2 − p2 . Our procedure is to ﬁnd solutions of the form ∞ ∞ y = xr n=0 an xn = n=0 an xn+r . b. which will be a nonnegative integer in the vibrating drum problem. x x (1. 1. as we mentioned before.by the method of Frobenius. dx2 dx where P (x) = Since xP (x) = 1 and 22 x2 Q(x) = x2 − p2 .17) an equation which is needed to analyze the vibrations of a circular drum. where r and the an ’s are constants. Find two linearly independent generalized power series solutions.3. Our next goal is to apply the Frobenius method to Bessel’s equation. we can rewrite Bessel’s equation in the form x2 or equivalently as d2 y dy + P (x) + Q(x) = 0. where r and the an ’s are constants.4. To ﬁnd generalized power series solutions to the diﬀerential equation 2x d2 y dy + + xy = 0 dx2 dx ∞ by the method of Frobenius. Find two linearly independent generalized power series solutions. 1. Determine the indicial equation and the recursion formula. we assume the solution has the form y= n=0 an xn+r . Using the Leibniz rule for diﬀerentiating a product. Determine the indicial equation and the recursion formula. a. b. x2 d2 y dy +x + (x2 − p2 )y = 0 dx2 dx .

∞ −p y = − 2 n=0 p2 an xn+r . ∞ ∞ [(n + r)2 − p2 ]an xn + n=0 n=2 an−2 xn = 0. or after division by xr .20) Adding up (1. ∞ ∞ ∞ (n + r)2 an xn+r + n=0 n=2 an−2 xn+r − n=0 p2 an xn+r = 0.19).19) Finally.18). This simpliﬁes to yield ∞ ∞ [(n + r)2 − p2 ]an xn+r + n=0 n=2 an−2 xn+r = 0.18) ∞ x2 y = n=0 an xn+r+2 = m=2 am−2 xm+r . (1. ∞ Then dy (n + r)an xn+r−1 . so the Frobenius theorem implies that there exists a nonzero generalized power series solution to (1. d dx x dy dx ∞ = n=0 (n + r)2 an xn+r . and (1. (1. Replacing m by n then yields ∞ x2 y = n=2 an−2 xn+r .17). ∞ ∞ x dy (n + r)an xn+r . = dx n=0 x dy dx ∞ = n=0 (n + r)2 an xn+r−1 . 23 . To ﬁnd such a solution. = dx n=0 d dx and thus x On the other hand.we see that x = 0 is a regular singular point.17). (1. we start as in the previous section by assuming that ∞ y= n=0 an xn+r . we have. where we have set m = n + 2.20). (1. we ﬁnd that if y is a solution to (1.

1!(p + 1)! a4 = = 1 2 p+2 1 = 1!(p + 1)! 1 2 p+4 1 = (−1)2 1!(p + 1)! 1 . 2 p! we obtain a2 = −a0 1 1 =− = (−1) 4p + 4 4(p + 1) 2p p! 1 −a2 = 8p + 16 8(p + 2) 1 2(p + 2) 1 2 p+4 p+2 1 2 1 . The general term is a2m = (−1)m 1 2 p+2m 1 . [(r +1)2 −p2 ]a1 = 0.17). m!(p + m)! 24 . In the special case where p is a nonnegative integer. Let us focus now on this important case. [(n+r)2 −p2 ]an +an−2 = 0 for n ≥ 2. 2!(p + 2)! and so forth. which implies that r = ±p. Since we want a0 to be nonzero.21) ⇒ [n2 + 2np]an + an−2 = 0. 2np + n2 (1. so (r2 −p2 )a0 = 0. [(n + p)2 − p2 ]an + an−2 = 0 which yields the recursion formula an = − 1 an−2 . r must satisfy the indicial equation (r2 − p2 ) = 0. we will get a genuine power series solution to Bessel’s equation (1.Thus we ﬁnd that ∞ (r2 − p2 )a0 + [(r + 1)2 − p2 ]a1 x + n=2 {[(n + r)2 − p2 ]an + an−2 }xn = 0. The coeﬃcient of each power of x must be zero. If we set 1 a0 = p . Then [(p + 1)2 − p2 ]a1 = 0 Finally. Let us assume without loss of generality that p ≥ 0 and take r = p. The recursion formula implies that an = 0 if n is odd. ⇒ (2p + 1)a1 = 0 ⇒ a1 = 0.

we can plot Jn (x). However. ∞ x 2m 1 J0 (x) = (−1)m .2 -0. .x]. .0. it turns out that when p is a nonnegative integer. Plot[ BesselJ[n.1 0.6 0. Thus when p is an integer. Bessel functions are so important that Mathematica includes them in its library of built-in functions. Bessel’s equation has a nonzero solution which is real analytic at x = 0.4 0.x].3 Mathematica represents the Bessel functions of the ﬁrst kind symbolically by BesselJ[n. For example. the reader can refer to Appendix A. Thus to plot the Bessel function Jn (x) on the interval [0.4 Figure 1. {x. for n = 1. .1 will be produced. Note that the graph of J0 (x) suggests that it has inﬁnitely many positive zeros. On the open interval 0 < x < ∞. Bessel’s equation has a two-dimensional space of solutions. we can show that the power series expansion for Jp (x) has inﬁnite radius of convergence. 3 For a very brief introduction to Mathematica.1: Graph of the Bessel function J0 (x). Thus we ﬁnally obtain the power series solution y= x 2 p ∞ m=0 4 6 8 10 12 14 (−1)m 1 m!(p + m)! x 2 2m . Similarly. 2. a second solution. 25 . 3 . (m!)2 2 m=0 Using the comparison and ratio tests.15}] and a plot similar to that of Figure 1. 15] one simply types in n=0. linearly independent from the Bessel function of the ﬁrst kind.8 0. The function deﬁned by the power series on the right-hand side is called the p-th Bessel function of the ﬁrst kind .2 2 -0. and is denoted by the symbol Jp (x).

and let W (y1 .4 0.2 2 -0. To see why. this solution is not well-behaved near x = 0. y2 (x) y2 (x) This Wronskian must satisfy the ﬁrst order equation d (xW (y1 . we can set y = v(x)Jp (x). substitute into Bessel’s equation and solve for v(x). apply the method of variation of parameters just as we did in the previous section for the Cauchy-Euler equation.0.2: Graph of the Bessel function J1 (x). y2 ) be their Wronskian. dx as one veriﬁes by a direct calculation: x d d d (xy1 y2 − xy2 y1 ) = y1 x (xy2 ) − y2 x (xy1 ) dx dx dx = −(x2 − n2 )(y1 y2 − y2 y1 ) = 0.6 0. If we were to carry this out in detail.2 4 6 8 10 12 14 Figure 1. Appropriately normalized. cannot be obtained directly by the generalized power series method that we have presented. Unlike the Bessel function of the ﬁrst kind. namely. suppose that y1 (x) and y2 (x) is a basis for the solutions on the interval 0 < x < ∞. we would obtain a second solution linearly independent from Jp (x). y2 )(x)) = 0. y2 )(x) = y1 (x) y1 (x) . deﬁned by W (y1 . 26 . however. we can. his solution is often denoted by Yp (x) and called the p-th Bessel function of the second kind . To obtain a basis for the space of solutions.

(Hint: The power series you obtain should be very similar to the power series for Jp (x). y2 )(x) = c.4. which are well-behaved as x → 0. and 1 y2 (x) = √ sin x x in the case where p = 1/2.17). The diﬀerential equation x2 d2 y dy − (x2 + p2 )y = 0 +x dx2 dx and d (xJ1 (x)) = xJ0 (x).4. Using the explicit power series formulae for J0 (x) and J1 (x) show that d J0 (x) = −J1 (x) dx 1. • This one-dimensional space is generated by a function Jp (x) which is given by the explicit power series formula x Jp (x) = 2 p ∞ m=0 (−1)m 1 m!(p + m)! x 2 2m .Thus c . Find a generalized power series solution to this equation in the case where p is an integer. It follows that two linearly independent solutions y1 (x) and y2 (x) to Bessel’s equation cannot both be well-behaved at x = 0.) 1. or W (y1 . dx is sometimes called a modiﬁed Bessel equation.2. x x 27 . an expression which is unbounded as x → 0.1. Let us summarize what we know about the space of solutions to Bessel’s equation in the case where p is an integer: xW (y1 . Hence the general solution to Bessel’s equation in this case is 1 1 y = c1 √ cos x + c2 √ sin x.4. x where c is a nonzero constant. Show that the functions 1 y1 (x) = √ cos x x are solutions to Bessel’s equation x d dx x dy dx + (x2 − p2 )y = 0. y2 )(x) = • There is a one-dimensional space of real analytic solutions to (1. Exercises: 1.3.

b. Use integration by parts to show that Γ(x + 1) = xΓ(x). Show that Γ(1) = 1.4.1. and use the recursion formula (1. it is convenient to use the gamma function deﬁned by ∞ Γ(x) = 0 tx−1 e−t dt.4. Set a0 = 1 2p Γ(p + 1) .21) to obtain the following generalized power series solution to Bessel’s equation (1. c. d. To obtain a nice expression for the generalized power series solution to Bessel’s equation in the case where p is not an integer. when n is a positive integer. a. Show that Γ(n + 1) = n! = n(n − 1) · · · 2 · 1.17) for general choice of p: y = Jp (x) = x 2 p ∞ m=0 (−1)m 1 m!Γ(p + m + 1) x 2 2m . 28 .

1−λ 0  . B T = B −1 .1 Eigenvalues of symmetric matrices Before proceeding further.1) for every choice of n-vectors x and y. (2. In particular. an n × n real matrix B is orthogonal if its transpose is equal to its inverse. equation (2. an n × n matrix B = (b1 b2 · · · bn ) 29 . An n × n matrix A is said to be symmetric if it is equal to its transpose AT . since x · y = xT y. we need to review and extend some notions from vectors and matrices (linear algebra). Alternatively. Examples of symmetric matrices include     3−λ 6 5 a b c 1 3  6 b d e  . which the student should have studied in an earlier course. and 3 1 5 0 8−λ c e f Alternatively.1) can be rewritten in the form xT Ay = (Ax)T y = xT AT y. we will need the amazing fact that the eigenvalue-eigenvector problem for an n × n matrix A simpliﬁes considerably when the matrix is symmetric. we could say that an n × n matrix A is symmetric if and only if x · (Ay) = (Ax) · y.Chapter 2 Symmetry and Orthogonality 2. which holds for all x and y if and only if A = AT . Indeed. On the other hand.

· · · . · · · . Moreover. b2 · bn = 0. 2. because eigenvalues need not be real (but can occur in complex conjugate pairs) and in the “repeated root” case. Spectral Theorem. for example. · bn · bn = 1. and sin θ cos θ 2/3 −2/3 1/3 are orthogonal. there may not be enough eigenvectors to construct a basis for R n . we can easily check that. . We will assume the following two facts.is orthogonal if its column vectors b1 . 1 This is called the “spectral theorem” because the spectrum is another name for the set of eigenvalues of a matrix. bn satisfy the relations b1 · b1 = 1. Sketch of proof: The reader may want to skip our sketch of the proof at ﬁrst. 30 . b2 · b2 = 1. the matrices   1/3 2/3 2/3 cos θ − sin θ −2/3 −1/3 2/3 . . b2 . Any continuous function on a sphere (of arbitrary dimension) assumes its maximum and minimum values. We will see that these complications do not occur for symmetric matrices. Note that since BT B = I ⇒ (det B)2 = (det B T )(det B) = det(B T B) = 1. . returning after studying some of the examples presented later in this section. Using this latter criterion. there is an n × n orthogonal matrix B of determinant one such that B −1 AB = B T AB is diagonal. b1 · b2 = 0. the determinant of an orthogonal matrix is always ±1. .1 Suppose that A is a symmetric n × n matrix with real entries. the problem of ﬁnding eigenvalues and eigenvectors can be complicated. which are proven rigorously in advanced courses on mathematical analysis: 1. For a general n × n matrix with real entries. b1 · bn = 0. For each eigenvalue λi . Recall that the eigenvalues of an n × n matrix A are the roots of the polynomial equation det(A − λI) = 0. The points at which the maximum and minimum values are assumed can be found by the method of Lagrange multipliers (a method usually discussed in vector calculus courses). Then its eigenvalues are real and eigenvectors corresponding to distinct eigenvectors are orthogonal. the corresponding eigenvectors are the nonzero solutions x to the linear system (A − λI)x = 0.

λ) = f (x) − λg(x). . . In other words. xn . x2 . . . . Let W be the “linear subspace” of R n deﬁned by the homogeneous linear equation b1 · x = 0. λ) = H(x1 . . and g(x1 . . x2 . 1 2 n or xT x = 1. so that the equation of the sphere is given by the constraint equation g(x) = 0. . we look for “critical points” for the function H(x. . . ∂λ which is just our constraint. . x2 . The intersection S n−1 ∩ W is a sphere of dimension n − 2. We also obtain the condition ∂H = −g(x) = 0. A =  21  · · · ·  an1 an2 · ann a short calculation shows that at the point where f assumes its maximum. 31 . the eigenvalue being the value λ1 of the variable λ. these are the points on the sphere at which the gradient of f is a multiple of the gradient of g.The equation of the sphere S n−1 in R n is  x1 x  x =  2 . xn ) = xT x − 1. . . . xn ) = xT Ax assumes its maximum values. Our approach consists of ﬁnding of ﬁnding the point on S n−1 at which the function f (x) = f (x1 . If we set   a11 a12 · a1n a a22 · a2n  . xn ) = 0. Ax − λx = 0. Thus the point on the sphere at which f assumes its maximum is a unit-length eigenvector b1 . . . ∂xi or equivalently. where We let g(x) = g(x1 . x2 . ∂H = 2ai1 x1 + 2ai2 x2 + · · · + 2ain xn − 2λxi = 0. or the points on the sphere at which the gradient of f is perpendicular to the sphere. . These are points at which ∇f (x1 . . To ﬁnd this maximum using Lagrange multipliers. .  ·  xn  x2 + x2 + · · · + x2 = 1. x2 . . . . . xn ) = λ∇g(x1 . xn ). x2 . .

b2 .2) that Ax − λx = 0.We next use the method of Lagrange multipliers to ﬁnd a point on S n−1 ∩ W at which f assumes its maximum. Thus if b2 is a point on S n−1 ∩ W at which f assumes its maximum.2) which can be put together in a single matrix equation Ab1 or equivalently. . A b1 b2 · bn = b1 b2 · bn If we set B = b1 this last equation becomes λ1 0 AB = B  · 0 32  0 λ2 · 0  · 0 · 0 . λ. These eigenvectors satisfy the equations Ab1 = λ1 b1 . Ab2 = λ2 b2 . . b2 must be a unit-length eigenvector for A which is perpendicular to b1 .. . . The maximum value of f on S n−1 ∩ W will be assumed at a critical point for the function H(x. . g2 (x) = b1 · x. It follows from the constraint equation b1 · x = 0 that b1 · (Ax) = bT (Ax) = (bT A)x = (bT AT )x 1 1 1 = (Ab1 )T x = (λ1 b1 )T x = λ1 b1 · x = 0. Ax − λx − µb1 = 0. we let g1 (x) = xT x − 1. Abn = λ2 bn .  λ1 0  · 0 0 λ2 · 0  · 0 · 0 . diﬀerentiation shows that ∂H = 2ai1 x1 + 2ai2 x2 + · · · + 2ain xn − 2λxi − µbi = 0. bn . (2. . Continuing in this way we ﬁnally obtain n mutually orthogonal unit-length eigenvectors b1 . ∂xi or equivalently. · ·  · λn b2 · bn . · ·  · λn Ab2 · Abn = λ1 b1 λ2 b2 · λn bn . Hence it follows from (2.. To do this. This time. µ) = f (x) − λg1 (x) − µg2 (x).

However. which might be expected to cause problems if A were not symmetric. The matrix 5 A = 4 0 5−λ 4 0 4 5−λ 0 0 0 1−λ  4 5 0  0 0 1 is symmetric.Of course. Example. Thus we are in the notorious “repeated root” case. Linear algebra. Simply let B be the matrix whose columns are unit-length eigenvectors for A. 1990. 2 There are many excellent linear algebra texts that prove this theorem in detail. the Spectral Theorem guarantees that we can ﬁnd a basis for R 3 consisting of eigenvectors for A even when the roots are repeated. or B AB =  A=B · · · · ·  0 0 · λn 0 we can solve for A. = 0.2 In any case. which has the roots λ1 = 1 with multiplicity two and λ2 = 9 with multiplicity one. so its eigenvalues must be real. 0 0 1−1 b3 or 4b1 4b1 +4b2 +4b2 0 = 0. one good reference is Bill Jacob. = 0. 0 λ2 · 0  · 0 · 0 . In the case of repeated roots. we must be careful to choose a basis of unit-length eigenvectors for each eigenspace which are perpendicular to each other. B is an orthogonal matrix. see Chapter 5. the method for ﬁnding the orthogonal matrix B such that B T AB is diagonal is relatively simple. · ·  · λn We can arrange that the determinant of B be one by changing the sign of one of the columns if necessary. 33 . so it is invertible and obtaining    λ1 0 · 0 λ1  0 λ2 · 0  −1 0 −1  B . Freeman. We ﬁrst consider the eigenspace W1 corresponding to the eigenvalue λ1 = 1. New York. which consists of the solutions to the linear system    5−1 4 0 b1  4 5−1 0   b2  = 0. W. H. Its characteristic equation is 0= = [(λ − 5)2 − 16](λ − 1) = (λ2 − 10λ + 9)(λ − 1) = (λ − 1)2 (λ − 9). since A is symmetric. at least when the eigenvalues are distinct. A more complete proof of the theorem is presented in more advanced courses in linear algebra.

The coeﬃcient matrix of this linear system is   4 4 0  4 4 0 . 0 0 0 Applying the elementary row operations    4 4 0 1 1  4 4 0 → 4 4 0 0 0 0 0 Thus the linear system is equivalent to b1 + b2 0 0 = 0, = 0, = 0. to this matrix   0 1 0 → 0 0 0 yields 1 0 0  0 0 . 0

Thus W1 is a plane with equation b1 + b2 = 0. We need to extract two unit length eigenvectors from W1 which are perpendicular to each other. Note that since the equation for W1 is b1 + b2 = 0, the unit length vector  √  1 n=
2 1 √ 2

0 is perpendicular to W1 . Since    √  1 0 2 −1  b1 =  0  ∈ W1 , we ﬁnd that b2 = n × b1 =  √2 ∈ W1 . 1 0 The vectors b1 and b2 are unit length elements of W1 which are perpendicular to each other. Next, we consider the eigenspace W9 corresponding to the eigenvalue λ2 = 9, which consists of the solutions to the linear system    5−9 4 0 b1  4 5−9 0   b2  = 0, 0 0 1−9 b3 or −4b1 4b1 +4b2 −4b2 = 0, = 0, = 0.

−8b3

The coeﬃcient matrix of this linear system is   −4 4 0  4 −4 0  . 0 0 −8 34

Applying the elementary row operations to this matrix yields      −4 4 0 1 −1 0 1 −1  4 −4 0  →  4 −4 0  →  0 0 0 0 −8 0 0 −8 0 0     1 −1 0 1 −1 0 →  0 0 0  →  0 0 1 . 0 0 1 0 0 0 Thus the linear system is equivalent to b1 − b2 b3 0 and we see that = 0, = 0, = 0, 

 0 0  −8

 1 W9 = span  1  . 0  b3 = 
1 √ 2 1 √ 2

We set

 .

0 Theory guarantees that the matrix  1 0 √2 −1 B =  0 √2 1 0
1 √ 2 1 √ 2

 ,

0

whose columns are the eigenvectors b1 , b1 , and b3 , will satisfy   1 0 0 B −1 AB =  0 1 0  . 0 0 9 Moreover, since the eigenvectors we have chosen are of unit length and perpendicular to each other, the matrix B will be orthogonal. Exercises: 2.1.1. Find a 2 × 2-matrix B such that B is orthogonal and B −1 AB is diagonal, where 5 4 A= . 4 5 2.1.2. Find a 2 × 2-matrix B such that B is orthogonal and B −1 AB is diagonal, where 3 4 A= . 4 9 35

2.1.3. Find a 3 × 3-matrix B such that B is orthogonal and B −1 AB is diagonal, where   5 4 0 A =  4 5 0 . 0 0 1 2.1.4. Find a 3 × 3-matrix B such that B where  −1 A= 2 0 is orthogonal and B −1 AB is diagonal,  2 0 0 2 . 2 1

2.1.5. Show that the n × n identity matrix I is orthogonal. Show that if B1 and B2 are n × n orthogonal matrices, so is the product matrix B1 B2 , as well as the −1 inverse B1 .

2.2

The theory presented in the previous section can be used to “rotate coordinates” so that conic sections or quadric surfaces can be put into “canonical form.” A conic section is a curve in R 2 which is described by a quadratic equation, such as the equation ax2 + 2bx1 x2 + cx2 = 1, 1 2 where a, b and c are constants, which can be written in matrix form as x1 If we let A= we can rewrite (2.3) as xT Ax = 1, (2.4) x2 a b b c x1 x2 = 1. (2.3)

a b b c

and

x=

x1 x2

,

where A is a symmetric matrix. According to the Spectral Theorem from Section 2.1, there exists a 2 × 2 orthogonal matrix B of determinant one such that B −1 AB = B T AB = If we make a linear change of variables, x = By, 36 λ1 0 0 λ2 .

then since xT = yT B T . we can rewrite (2. The second column b2 is a unit-vector perpendicular to b1 and hence − sin θ b2 = ± . the conic is an ellipse. In the case where λ1 and λ2 are both positive. cos θ or equivalently. • If λ1 and λ2 have opposite signs. equation (2. because det B = 1. sin θ B 0 1 = − sin θ . cos θ We must take the plus sign in this last expression. 2 2 λ1 y1 + λ2 y2 = 1. y2 )-coordinates. it is easy to recognize the conic section: • If λ1 and λ2 are both positive. To see this. the conic degenerates to the the empty set. x2 )-coordinates by simply rotating counterclockwise through the angle θ. The matrix B which relates x and y represents a rotation of the plane. sin θ for some choice of θ. B takes the standard basis vectors for R2 to vectors which have been rotated counterclockwise through an angle θ. • If λ1 and λ2 are both negative. By linearity. Thus B 1 0 = cos θ . Thus once we have sketched the conic in (y1 . note that the ﬁrst column b1 of B is a unit-length vector. 37 . and can therefore be written in the form b1 = cos θ . the conic is an hyperbola. y2 ). from which we recognize that the semi-major and semi-minor axes of the ellipse are 1/λ1 and 1/λ2 . y1 y2 λ1 0 0 λ2 y1 y2 = 1. we can obtain the sketch in (x1 .3) is transformed into yT (B T AB)y = 1.5) In the new coordinate system (y1 . or equivalently. because the equation has no real solutions. (2.5) as 2 y1 ( 1/λ1 )2 + 2 y2 ( 1/λ2 )2 = 1. B= cos θ sin θ − sin θ cos θ must rotate every element of R 2 counterclockwise through an angle θ.

Unit-length eigenvectors corresponding to these eigenvalues are √ √ 1/√2 −1/ 2 √ b1 = . (λ − 2)(λ − 8) = 0. and the eigenvalues are λ1 = 2 and λ2 = 8. The characteristic equation of the matrix A= is (5 − λ)2 − 9 = 0. 1 2 or equivalently.6) = 1.Example. 2 (1/ 2) (1/ 8)2 We recognize that this is the equation of √ ellipse. λ2 − 10λ + 16 = 0. If we deﬁne new coordinates (y1 .6) will simplify to y1 or 2 2 2y1 + 8y2 = =B y1 y2 . x1 x2 5 −3 −3 5 x1 x2 (2. 38 . 1/ 2 1/ 2 The proof of the theorem of Section 2. The lengths of the semimajor an √ and semiminor axes are 1/ 2 and 1/(2 2). b2 = . y2 ) by x1 x2 equation (2.1 shows that these vectors are orthogonal to each other. and hence the matrix √ √ 1/√2 −1/ 2 √ B= 1/ 2 1/ 2 is an orthogonal matrix such that B T AB = 2 0 0 8 . y2 2 0 0 8 y1 y2 = 1. Let’s consider the conic 5x2 − 6x1 x2 + 5x2 = 1. 5 −3 −3 5 Note that B represents a counterclockwise rotation through 45 degrees. 2 2 y1 y2 √ √ + = 1.

surfaces deﬁned by an equation of the form    a11 a12 a13 x1 x1 x2 x3  a21 a22 a23   x2  = 1.2 -0. x3  we can write this in matrix form xT Ax = 1.2 0. (2. y3 0 0 λ3 39 . We introduce new coordinates   y1 by setting x = By.4 Figure 2.2 -0.4 0. there is a 3 × 3 orthogonal matrix B of determinant one such that B T AB is diagonal.7) According to the Spectral Theorem. a31 a32 a33 x3 where the aij ’s are constants. Thus after a suitable linear change of coordinates.  a11 A =  a21 a31 If we let a12 a22 a32  a13 a23  .2 -0.1: Sketch of the conic section 5x2 − 6x1 x2 + 5x2 − 1 = 0. 1 2 The same techniques can be used to sketch quadric surfaces in R 3 .0.7) becomes yT (B T AB)y = 1. the equation (2. y = y2  y3 and equation (2.7) can be put in the form    λ1 0 y1 0 y1 y2 y3  0 λ2 0   y2  = 1.4 -0.4 0. a33  x1 x =  x2  .

y3 ). or 2 2 2 λ1 y1 + λ2 y2 + λ3 y3 = 1. we ﬁnd that there are four cases: • If λ1 . It follows from this that the only real eigenvalues of B can be ±1.8) In other words. To see this. if x is an eigenvector for B corresponding to the eigenvalue λ. the quadric surface is an ellipsoid . and λ3 are the eigenvalues of A. the quadric surface is an hyperboloid of two sheets. and λ3 are all positive.5 -1 -2 0 2 -4 -2 2 0 Figure 2.2: An ellipsoid. • If λ1 . and λ3 are all negative. y2 . note ﬁrst that since B is orthogonal.5 0 -0. λ2 . • If two of the λ’s are negative and one is positive. λ2 . (2. multiplication by B preserves dot products. the orthogonal matrix B of determinant one which relates x and y represents a rotation. Indeed. λ2 . 40 . (Bx) · (By) = xT B T By = xT Iy = x · y. It is relatively easy to sketch the quadric surface in the coordinates (y1 . the quadric surface is an hyperboloid of one sheet. Just as in the case of conic sections. the equation represents the empty set. If the eigenvalues are all nonzero. where λ1 . • If two of the λ’s are positive and one is negative.4 1 0. then λ2 (x · x) = (λx) · (λx) = (Bx) · (Bx) = x · x.

(By) · (By) = 1. we conclude that λ = 1 must occur as an eigenvalue also in this case. In this case. non-real eigenvalues must occur in complex conjugate pairs. Indeed. so if there is a nonreal eigenvalue µ + iν. 1 = det B = λ(µ + iν)(µ − iν) = λ(µ2 + ν 2 ). (Bz) · (Bz) = 1.so division by x · x yields λ2 = 1. 41 . λ = 1 is an eigenvalue and W1 = {x ∈ R 3 : Bx = x} is nonzero. Since det B = 1 and det B is the product of the eigenvalues. Bz must be a unit-length vector in W1 which is perpendicular to By and hence Bz = ±(− sin θy + cos θz). On the other hand. z} form an orthonormal basis for W1 . λ = 1 must occur as an eigenvalue. Thus if B = I. (By) · (Bz) = 0. then B must be the identity. ⊥ ⊥ so By ∈ W1 . It is easily veriﬁed that if dim W1 is larger than one. In either case. we cannot have Bz = −(− sin θy + cos θz). z be elements of W1 such that y · y = 1. this would imply (via a short calculation) that the vector u = cos(θ/2)y + sin(θ/2)z ⊥ is ﬁxed by B. it follows from (2. ⊥ Let W1 denote the orthogonal complement to W1 . If x is a nonzero element ⊥ of W1 and y ∈ W1 . Since λ = ±1. By (2. This is the axis of rotation. in other words Bu = u. if all of the eigenvalues are real. ⊥ Moreover. then there must be another non-real eigenvalue µ − iν together with a real eigenvalue λ.8) that (By) · x = (By) · (Bx) = yT B T Bx = yT Ix = y · x = 0. ⊥ we could say that {y. ⊥ Thus By must be a unit-length vector in W1 and there must exist a real number θ such that By = cos θy + sin θz. z · z = 1. However. Let y. there is a one-dimensional subspace W1 of R 3 which is left ﬁxed under multiplication by B. Thus we must have Bz = − sin θy + cos θz.8). y · z = 0. contradicting the fact that u ∈ W1 .

Exercises: 2.5 0 -0. Suppose that A= 2 −2 −2 5 . 42 .1 0 -1 1 0. We can therefore conclude that a 3 × 3 orthogonal matrix B of determinant one represents a rotation about an axis (which is the eigenspace for eigenvalue one) and through an angle θ (which can be determined from the eigenvalues of B. it is easily checked that y + iz and y − iz are eigenvectors for B with eigenvalues e±iθ = cos θ ± i sin θ. Sketch the conic section 2x2 + 6x1 x2 + 2x2 = 1.2. which must be 1 and e±iθ ). a. ⊥ and multiplication by B must be a rotation in the plane W1 through an angle θ.3: Hyperboloid of one sheet. Find an orthogonal matrix B such that B T AB is diagonal. Moreover. Sketch the conic section 2(x1 − 1)2 + 6(x1 − 1)x2 + 2x2 = 1. b.2.5 -1 -1 0 1 Figure 2. 2 2. Suppose that A= 2 3 3 2 .1.2. 1 2 c.

1 2 c. a. Determine which of the following conic sections are ellipses. −3x2 + 6x1 x2 − 4x2 = 1. Find the semi-major and semi-minor axes of the ellipse 5x2 + 6x1 x2 + 5x2 = 4.3. 1 2 43 . x2 + 6x1 x2 + 10x2 = 1. etc. 1 2 2.4: Hyperboloid of two sheets. Suppose that A= 4 2 2 1 . Find an orthogonal matrix B such that B T AB is diagonal. 1 2 d. Sketch the conic section 4x2 + 4x1 x2 + x2 − 5x1 + 2 5x2 = 0. −x2 + 4x1 x2 − 3x2 = 1. Sketch the conic section 2x2 − 4x1 x2 + 5x2 − 4x1 + 4x2 = −1. Sketch the conic section 2x2 − 4x1 x2 + 5x2 = 1. which are hyperbolas. 1 2 c.2.2. √ √ b. a. 1 2 b.4.: a. x2 + 4x1 x2 + 3x2 = 1. Find an orthogonal matrix B such that B T AB is diagonal. 1 2 2. b.5.2. 1 2 2.2 2 0 -2 -2 -1 0 1 2 -2 -1 0 1 Figure 2.

2. Sketch the quadric surface 4x1 x2 + 3x2 + 9x2 = 1. c. a. which are hyperboloids of two sheets. (2. k} for R 3 is replaced by a more general “orthonormal basis.8. b. Show that the matrix   2/3 2/3 1/3 1/3 B = −2/3 2/3 2/3 −1/3 −2/3 is an orthogonal matrix of deteminant one. b = v · j. etc. . Find an orthogonal matrix B such that B T AB is diagonal.2. b1 · b2 = 0.: a.9) 44 . where a = v · i. x2 + x2 − x2 − 6x1 x2 = 1.7. which are hyperboloids of one sheet. b1 · bn = 0. b. A collection of n vectors b1 .2. We would like to extend this formula to the case where the standard basis {i. Determine which of the following quadric surfaces are ellipsoids. 2 3 2. 1 2 3 2. Find a nonzero vector which spans the axis of rotation. 1 2 3 c.3 Orthonormal bases Recall that if v is an element of R 3 . we can express it with respect to the standard basis as v = ai + bj + ck. Thus multiplication by B is rotation about some axis through some angle. b2 · b2 = 1. . Determine the angle of rotation. · · · . bn in R n is an orthonormal basis for R n if b1 · b1 = 1.6.” Deﬁnition. 1 2 3 b. j. . · bn · bn = 1. x2 − x2 − x2 − 6x2 x3 = 1. · · · . b2 · bn = 0.2. Suppose that  0 2 0 A =  2 3 0 . 2. x2 + x2 + x2 + 4x1 x2 + 2x3 = 0. b2 . . c = v · k. 0 0 9  a.

b2 .From the discussion in Section 2. . 45 . dt   y1 0 0   y2   . This formula can be helpful in solving the initial value problem dx = Ax. . then f ∈ Rn ⇒ f = (f · b1 )b1 + · · · + (f · bn )bn . . It is relatively easy to express an arbitrary vector f ∈ R n in terms of an orthonormal basis b1 . · dyn /dt = λn yn . . ·  ·  λn yn Thus in terms of the new variable      y1 dy1 /dt λ1  y2   dy2 /dt   0    y =  .1 guarantees that the eigenvalues of A are real and that we can ﬁnd an n × n orthogonal matrix B such that   λ1 0 · · · 0  0 λ2 · · · 0  . if b1 . k}. bn : to ﬁnd constants c1 .12) in the case where A is a symmetric n × n matrix and f is a constant vector. (2. the Spectral Theorem of Section 2. we recognize that the term orthonormal basis is just another name for a collection of n vectors which form the columns of an orthogonal n × n matrix. . bn is an orthonormal basis for R n . dy2 /dt = λ2 y2 . dt x(0) = f . . j. cn so that f = c1 b1 + c2 b2 + · · · + cn bn . then B dy dx = = Ax = ABy dt dt ⇒ dy = B −1 ABy. we have ·  · = · dyn /dt 0 yn 0 λ2 · 0 ··· ··· ··· so that the matrix diﬀerential equation decouples into n noninteracting scalar diﬀerential equations dy1 /dt = λ1 y1 .10) we simply dot both sides with the vector bi and use (2. (2. In other words. c2 . . b2 . Since A is symmetric. .1. .9) to conclude that ci = bi · f . B −1 AB =  · · ·  0 0 · · · λn If we set x = By. . . . (2.11) a generalization of the formula we gave for expressing a vector in R 3 in terms of the standard basis {i.

9 Thus the general solution to the matrix diﬀerential equation dx/dt = Ax is   c1 et x = B  c2 et  = c1 b1 et + c2 b2 et + c3 b3 e9t . c2 .12). where f= 0 1 4 We saw in Section 2. Example. obtaining the general solution    λ t c1 e 1 y1  y2   c2 eλ2 t   y= =  ·   · . c2 . . . . . 5 0 x. Note that x(0) = c1 b1 + c2 b2 + · · · + cn bn . . cn so that c1 b1 + c2 b2 + · · · + cn bn = f . . x(0) = f . To ﬁnish solving the initial value problem (2. Suppose  5 dx  = 4 dt 0 we want to solve the initial value problem    4 0 3 1 . we need to determine the constants c1 . b2 . cn eλn t where b1 . Transferring back to our original variable x yields  λ t c1 e 1  c2 eλ2 t  λ1 t λ2 t λn t  x=B  ·  = c1 b1 e + c2 b2 e + · · · + cn bn e . . . bn are the columns of B. . cn are constants of integration. .We can solve these equations individually. . . the orthogonal matrix  1 1  0 √2 √2 −1 1 B =  0 √2 √2  1 0 0 has the property that 1 B −1 AB =  0 0  0 1 0  0 0 .11) comes in handy. It is here that our formula (2. c3 e9t 46 . using it we obtain x = (f · b1 )b1 eλ1 t + · · · + (f · bn )bn eλn t .1 that A has the eigenvalues λ1 = 1 with multiplicity two and λ2 = 9 with multiplicity one. yn cn eλn t where c1 . Moreover.

To solve the initial value problem.1.3. c2 =  √2  · 1 = 2. c.where   0 b1 = 0 . 1 4 4 0 √    1 3 2 √ 1    c3 =  √2 · 1 = 2 2. Find the solution to the initial value problem   1 3 dx = Ax. Find the eigenvalues of the symmetric matrix   5 4 0 0 4 5 0 0  A= 0 0 4 2 . we employ (2. 0 0 2 1 b. 1 0 0 Exercise: 2. 0 dt 2 47 . 4 0 Hence the solution is   √  √  1 1 0 2 2 √ √ −1 1 x = 4 0 et + 2  √2  et + 2 2  √2  e9t .a. d. 2 √  1 1 b3 =  √2  . 1 √  1 b2 = 0 2 −1 √  . Find an orthonormal basis for R 4 consisting of eigenvectors of A. 0 2 Setting t = 0 in our expression for x yields x(0) = c1 b1 + c2 b2 + c3 b3 . x(0) =   . Find the general solution to the matrix diﬀerential equation dx = Ax. Find an orthogonal matrix B such that B −1 AB is diagonal.11) to see that     √    1 0 3 3 2 √ −1 c1 = 0 · 1 = 4. dt e.

2.3.2.a. Find the eigenvalues of the symmetric matrix   −3 2 0 A =  2 −4 2  . 0 2 −5 (Hint: To ﬁnd roots of the cubic, try λ = 1 and λ = −1.) b. Find an orthogonal matrix B such that B −1 AB is diagonal. c. Find an orthonormal basis for R 3 consisting of eigenvectors of A. d. Find the general solution to the matrix diﬀerential equation dx = Ax. dt e. Find the solution to the initial value problem   1 dx = Ax, x(0) = 2 . dt 0 2.3.3. (For students with access to Mathematica) a. Find the eigenvalues of the matrix   2 1 1 A = 1 3 2 1 2 4 by running the Mathematica program a = {{2,1,1},{1,3,2},{1,2,4}}; Eigenvalues[a] The answer is so complicated because Mathematica uses exact but complicated formulae for solving cubics discovered by Cardan and Tartaglia, two sixteenth century Italian mathematicians. b. Find numerical approximations to these eigenvalues by running the program Eigenvalues[N[a]] The numerical values of the eigenvalues are far easier to use. c. Use Mathematica to ﬁnd numerical values for the eigenvectors for A by running the Mathematica program Eigenvectors[N[a]] and write down the general solution to the matrix diﬀerential equation dx = Ax. dt 48

1

Figure 2.5: Two carts connected by springs and moving along a friction-free track.

2.4

Mechanical systems

Mechanical systems consisting of weights and springs connected together in an array often lead to initial value problems of the type d2 x = Ax, dt2 x(0) = f , dx = 0, dt (2.13)

where A is a symmetric matrix and f is a constant vector. These can be solved by a technique similar to that used in the previous section. For example, let us consider the mechanical system illustrated in Figure 2.5. Here we have two carts moving along a friction-free track, each containing mass m and attached together by three springs, with spring constants k1 , k2 and k3 . Let x1 (t) x2 (t) F1 F2 = the position of the ﬁrst cart to the right of equilibrium, = the position of the second cart to the right of equilibrium, = force acting on the ﬁrst cart, = force acting on the second cart,

with positive values for F1 or F2 indicating that the forces are pulling to the right, negative values that the forces are pulling to the left. Suppose that when the carts are in equilibrium, the springs are also in equilibrium and exert no forces on the carts. In this simple case, it is possible to reason directly from Hooke’s law that the forces F1 and F2 must be given by the formulae F1 = −k1 x1 + k2 (x2 − x1 ), F2 = k2 (x1 − x2 ) − k3 x2 ,

but it becomes diﬃcult to determine the forces for more complicated mechanical systems consisting of many weights and springs. Fortunately, there are some simple principles from physics which simply the procedure for ﬁnding the forces acting in such mechanical systems.

49

The easiest calculation of the forces is based upon the notion of work . On the one hand, the work required to pull a weight to a new position is equal to the increase in potential energy imparted to the weight. On the other hand, we have the equation Work = Force × Displacement, which implies that Force = Work Change in potential energy =− . Displacement Displacement

Thus if V (x1 , x2 ) is the potential energy of the conﬁguration when the ﬁrst cart is located at the point x1 and the second cart is located at the point x2 , then the forces are given by the formulae F1 = − ∂V , ∂x1 F2 = − ∂V . ∂x2

In our case, the potential energy V is the sum of the potential energies stored in each of the three springs, V (x1 , x2 ) = 1 1 1 k1 x2 + k2 (x1 − x2 )2 + k3 x2 , 1 2 2 2 2

and hence we obtain the formulae claimed before: F1 = − ∂V = −k1 x1 + k2 (x2 − x1 ), ∂x1 ∂V = k2 (x1 − x2 ) − k3 x2 . ∂x2

F2 = −

It now follows from Newton’s second law of motion that Force = Mass × Acceleration, and hence d2 x1 d2 x2 , F2 = m 2 . dt2 dt Thus we obtain a second-order system of diﬀerential equations, F1 = m m d2 x1 = −k1 x1 + k2 (x2 − x1 ) = −(k1 + k2 )x1 + k2 x2 , dt2

d2 x2 = k2 (x1 − x2 ) − k3 x2 = k2 x1 − (k2 + k3 )x2 . dt2 We can write this system in matrix form as m m d2 x = Ax, dt2 where A= −(k1 + k2 ) k2 k2 −(k2 + k3 ) 50 . (2.14)

The eigenvalues in this case are λ1 = −1. 1/ 2 1/ 2 . and λ2 = −3.1 guarantees that the matrix √ √ 1/√2 −1/ 2 √ B= .14) becomes d2 x = dt2 −2 1 1 x. −2 (2. Let us consider the special case of the mass-spring system in which m = k1 = k2 = k3 = 1. It follows from the argument in Section 2. . will diagonalize our system of diﬀerential equations. 51 .1 that the eigenspace corresponding to the other eigenvalue is just the orthogonal complement W−3 = span −1 1 . . The potential energy is given by the expression 1 x V (x1 . The eigenspace corresponding to the eigenvalue −1 is W−1 = {b ∈ R2 : (A + I)b = 0} = . = span 1 1 . −2 − λ 1 1 −2 − λ = (λ + 2)2 − 1 = 0. x2 ) = − x1 x2 A 1 . Unit length eigenvectors lying in the two eigenspaces are √ √ 1/√2 −1/ 2 √ b1 = . b2 = 1/ 2 1/ 2 The theorem of Section 2. so that the system (2. x2 2 Example. we must solve the characteristic equation det which yields λ = −2 ± 1.Note that A is indeed a symmetric matrix.15) To ﬁnd the eigenvalues. whose columns are b1 and b2 .

b. = −3y2 . where A =  1 −2 1  .a.4. The general solution to the transformed system is y1 = a1 cos ω1 t + b1 sin ω1 t. the general solution to (2. 3. The motion of the carts can be described as a general superposition of two modes of oscillation. Show that the motion of this system is described by the matrix diﬀerential equation   −2 1 0 d2 x = Ax. x = b1 (a1 cos ω1 t + b1 sin ω1 t) + b2 (a2 cos ω2 t + b2 sin ω2 t). y2 = a2 cos ω2 t + b2 sin ω2 t.5 in the case where m = k1 = k2 = k3 = k4 = 1.Indeed.1. x2 y2 1/ 2 1/ 2 our system of diﬀerential equations transforms to d2 y1 /dt2 d2 y2 /dt2 We set ω1 = 1 and ω2 = √ = −y1 . a system of two noninteracting harmonic oscillators. 2π . Find the general solution to this system.2. Consider the mass-spring system with two carts illustrated in Figure 2. What are the frequencies of vibration of this mass-spring system? 2. 0. dt2 0 1 −2 52 and ω2 . . c.4.15) is √ √ x1 1/√2 −1/ 2 a1 cos ω1 t + b1 sin ω1 t √ x= = a2 cos ω2 t + b2 sin ω2 t x2 1/ 2 1/ 2 or equivalently. In the original coordinates. so that this system assumes the familiar form 2 d2 y1 /dt2 + ω1 y1 2 2 2 d y2 /dt + ω2 y2 = = 0. of frequencies ω1 2π Exercises: 2.a.5 in the case where k1 = 4 and m = k2 = k3 = 1. Write down a system of secondorder diﬀerential equations which describes the motion of this system. Consider the mass-spring system with three carts illustrated in Figure 2. y2 ) by setting √ √ x1 1/√2 −1/ 2 y1 √ = . if we deﬁne new coordinates (y1 .

dt2 0 dx (0) = 0. b. dt2 f.5 Mechanical systems with many degrees of freedom* Using an approach similar to the one used in the preceding section. For 53 . Find the solution to the initial value problem   1 2 d x = Ax. Find an orthonormal basis for R 3 consisting of eigenvectors of A.3. What are the frequencies of oscillation of a mechanical system which is governed by the matrix diﬀerential equation d2 x = Ax? dt2 2. A= 0 1 −2 1  0 0 1 −2 b. Find the general solution to the matrix diﬀerential equation d2 x = Ax. Find the eigenvalues of the symmetric matrix   −2 1 0 0  1 −2 1 0 . c. d. e.6: Three carts connected by springs and moving along a friction-free track. dt 2. Find an orthogonal matrix B such that B −1 AB is diagonal. x(0) = 2 . we can consider more complicated systems consisting of many masses and springs.4.Figure 2.a. Find the eigenvalues of the symmetric matrix A.

. . . a system of n carts containing identical weights of mass m and connected by identical springs of spring constant k. 2 i=1 j=1 54 n n .   1 =− k 2 x1 We can write this as    x=   or equivalently as where 1 V (x1 . .Figure 2. and make the problem of determining the “natural frequencies of vibration” of the mattress into a manageable problem. . . . we could consider the box spring underlying the mattress in a bed. example. xn ) = 1 1 1 1 k(xn −x1 )2 + k(x2 −x1 )2 + k(x3 −x2 )2 +. . .6.7: A circular array of carts and springs. x2 x3 · xn     · · · ··· ·  ·  1 0 0 · · · −2 xn 1 V (x1 . . moving along a circular friction-free track as sketched in Figure 2. 2 x1 x2 x3 · xn    . . Although such a box spring contains hundreds of individual springs. xn ) = − kxT Ax. let us consider a somewhat simpler problem. and hence the matrix A in the corresponding dynamical system constains hundreds of rows and columns. We choose a positive direction along the track and let xi denote the displacement of the i-th cart out of equilibrium position in the positive direction. .      A=   −2 1 0 · 1 1 −2 1 · 0 0 1 −2 · 0 ··· ··· ··· ··· ··· 1 0 0 · −2    .+ k(xn −xn−1 )2 2 2 2 2    −2 1 0 ··· 1 x1  1 −2 1 · · · 0   x2      0 1 −2 · · · 0   x3  . To illustrate how the symmetries of a problem can make it much easier to solve. xn ) = − k aij xi xj . . . it is still possible to use symmetries in the box spring to simplify the calculations. The potential energy stored in the springs is V (x1 . for 1 ≤ i ≤ n.

xn−1 = yn . To simplify the calculation of the eigenvalues of this matrix.    x2 = y3 . dt2 (2. . On the other hand. 2 dt m 55 (2. .16) where A is the symmetric matrix given above. . we need to ﬁnd the eigenvalues of the n × n matrix A. . . ∂xi j=1 which could be rewritten in matrix form as F = kAx. x = T y. Just as in the preceding section.     T =    0 0 0 · 0 1 1 0 0 · 0 0 0 1 0 · 0 0 ··· ··· ··· ··· ··· ··· 0 0 0 · 1 0     . . In general. ∂x1 2 j=1 2 i=1 j=1 n n n the last step obtained by using the symmetry of A. Indeed. xn = y1 .17) d2 x = F.16) yields m d2 x = kAx dt2 or d2 x k = Ax. the force acting on the i-th cart can be calculated as minus the derivative of the potential energy with respect to the position of the i-th cart. by Newton’s second law of motion. F1 = − ∂V 1 1 = k a1j xj + k ai1 xi = k a1j xj . the dynamical system remains unchanged. j)-component of the matrix A. the other carts being held ﬁxed. we make use of the fact that the carts are identically situated—if we relabel the carts.18) . where Then y satisﬁes exactly the same system of diﬀerential equations as x: d2 y k = Ay. or in matrix terms. we obtain the result: n ∂V Fi = − =k aij xj . Thus for example. dt2 m (2.. To ﬁnd the frequencies of vibration of our mechanical system. shifting them to the right by one.where aij denotes the (i. m Substitution into (2. lew us deﬁne new coordinates (y1 . yn ) by setting x1 = y2 ..

the components of x must satisfy the vector equation T x = λx. where η = e2πi/n . the vector equation becomes x2 = λx1 . xn = η (n−1)j . x1 = λxn .. x3 = η 2j . This equation is easily solved via Euler’s formula: 1 = e2πi ⇒ (e2πi/n )n = 1. If we set x1 = 1. . or T A = AT. we must have λn = 1. if x is an eigenvector for T corresponding to the eigenvalue λ. we can conclude from (2. . we can try to ﬁnd eigenvectors corresponding to η j .19) that x2 = η j . xn = λxn−1 . (2. x3 = λx2 . d2 x k = Ax 2 dt m ⇒ T d2 y k = AT y 2 dt m or d2 y k = T −1 AT y. In terms of components. Since at least one of the xi ’s is nonzero.     · η (n−1)j for the eigenvalue η j . x2 = λn x2 .. 0 ≤ j ≤ n − 1. . yielding ﬁnally the equation x1 = λn x1 . we do indeed have an eigenvector   1   ηj   ej =  η 2j  . . Similarly. (2.. each eigenspace is one-dimensional. and similarly [(e2πi/n )j ]n = 1. (2. Thus the solutions to (2.19) Thus x3 = λ2 x1 . .. .20) are λ = ηj . Moreover. . xn = λn xn . Thus for each j.On the other hand. for 0 ≤ j ≤ n − 1.21) For each choice of j. x4 = λ3 x1 . 56 . . thereby obtaining a nonzero solution to the eigenvector equation. Indeed. and so forth. 2 dt m Comparison with (2.20) for 0 ≤ j ≤ n − 1. In other words the matrices A and T commute.18) yields A = T −1 AT. Now it is quite easy to solve the eigenvector-eigenvalue problem for T .

Aej = λj ej . Since the eigenspaces of T are all one-dimensional. however. and this equation states that Aej is an eigenvector for T with the same eigenvalue as ej . which is one-dimensional. Aej must be a multiple of ej . On the other hand. If j = 0 and j = n/2. where we have used Euler’s formula once again. for some number λj . since A and T commute. the eigenvectors for T are also eigenvectors for A. not for T . and if n is even. and hence the eigenspaces for A are two-dimensional. in other words. we need to solve the eigenvalue-eigenvector problem for A. It follows from the familiar formula cos(2α) = cos2 α − sin2 α that λj = −2 + 2[cos(πj/n)]2 − 2[sin(πj/n)]2 = −4[sin(πj/n)]2 . Indeed. the ﬁrst component of ej is 1. To ﬁnd the eigenvalues λj for A. Thus in the special case where n is odd. ej and en−j form a basis for the eigenspace corresponding to eigenvalue λj . since AT = T A. Note that λj = λn−j . of course. to the eigenspace corresponding to j = n/2. Fortunately.In the dynamical system that we are considering. except for the eigenspace corresponding to j = 0. all the eigenspaces are two-dimensional except for the eigenspace with eigenvalue λ0 = 0. It is not diﬃcult to verify that   1   cos(πj/n)   1  cos(2πj/n)  (ej + en−j ) =   2   · cos((n − 1)πj/n) and  0   sin(πj/n)   i  sin(2πj/n)  (ej − en−j ) =   2   · sin((n − 1)πj/n) 57  . so we immediately conclude that λj = −2 + 2 cos(2πj/n). we simply act on ej by A: we ﬁnd that the ﬁrst component of the vector Aej = λj ej is −2 + η j + η (n−1)j = −2 + (e2πij/n + e−2πij/n ) = −2 + 2 cos(2πj/n). T (Aej ) = A(T ej ) = A(η j ej ) = η j (Aej ).

the frequencies of oscillation being ωj = 2π Note that the component   1 1   e0 (a0 + b0 t) = 1 (a0 + b0 t)   · 1 corresponds to a constant speed motion of the carts around the track. = = ⇒ = .17) as (n−1)/2 x = e0 (a0 + b0 t) + j=1 1 (ej + en−j )(aj cos ωj t + bj sin ωj t) 2 i (ej − en−j )(cj cos ωj t + dj sin ωj t).form a real basis for this eigenspace. sin(π/n) . π π n 2π mn so if we were to set k/m = n2 .5.1. m π ··· ··· ··· ··· ··· ··· 58 . In the case where n is odd. Let ωj = −λj = 2 sin(πj/n). If n is large. Find the eigenvalues of the matrix  0 1 0 0 0 1  0 0 0 T = · · ·  0 0 0 1 0 0  0 0  0  ·  1 0 k sin(πj/n) . the lowest nonzero frequency of oscillation would approach one as n → ∞. Exercise: 2. we can write the general solution to our dynamical system (2. 2 (n−1)/2 + j=1 The motion of the system can be described as a superposition of several modes of oscillation. π/n k 1 1 ω1 .

Figure 2. · 1 −λ 2.7. m = Table[Max[2-Abs[i-j]. by expanding the determinant −λ 0 0 · 0 1 1 −λ 0 · 0 0 0 1 −λ · 0 0 ··· ··· ··· ··· ··· ··· 0 0 0 . {i. so that k/m = n2 . 2 dt m where    P =   −2 1 0 · 0 1 −2 1 · 0 0 1 −2 · · ··· ··· ··· ··· ··· 0 0 · · −2    . we can show that the carts will move in accordance with the linear system of diﬀerential equations d2 x k = Ax = P x. We can use the following Mathematica program to ﬁnd the eigenvalues of the n × n matrix A. and connected by identical springs of spring constant k. when n = 6: n = 6.4 IdentityMatrix[n-1].n-1} ]. are moving along a friction-free track. Just as in the preceding section. p = m .{j.0].   We take k = n and m = (1/n).6 A linear array of weights and springs* Suppose more generally that a system of n − 1 carts containing identical weights of mass m.8: A linear array of carts and springs. as shown in Figure 2.n-1} . 59 .

m = Table[Max[2-Abs[i-j]. n = 26. This would require new variables yi and zi such that yi (t) = the y-component of displacement of the i-th weight. Print[eval[[n-1]]]. p = m . If we experiment with this program using diﬀerent values for n. ListPlot[evec[[n-1]]]. We can also modify the program to ﬁnd the eigenvalues of the matrix when n = 12. or n = 60 by simply replacing 6 in the top line with the new value of n. the (n − 1) × (n − 1)-matrix  0 ··· 0 1 ··· 0   2 ··· · .  2 − |i − j| = 1 if j = i ± 1   ≤0 otherwise. {i. A more realistic model would allow the carts to move in all three directions.{j.0]. we will see that as n gets larger.  · ··· ·  · ··· 2 the ﬁrst line of the program generates  2 1  1 2  M = 0 1   · · 0 0 The next two lines generate the matrices P and A = n2 P . 60 .a = n∧2 p eval = Eigenvalues[N[a]] Since  = 2 if j = i. you will note that all the eigenvalues are negative and that Mathematica provides the eigenvalues in increasing order. We can further modify the program by asking for only the smallest eigenvalue lambda[[n]] and a plot of the corresponding eigenvector: n = 14. The track on which the cart runs restricts the motion of the weights to be only in the x-direction.n-1} . zi (t) = the z-component of displacement of the i-th weight. evec = Eigenvectors[N[a]]. a = n∧2 p.n-1} ]. the smallest eigenvalue seems to approach −π 2 and the plot of the smallest eigenvector looks more and more like a sine curve. Thus the fundamental frequency of the mechanical system seems to approach π/2π = 1/2 and the oscillation in the fundamental mode appears more and more sinusoidal in shape. we can consider this array of springs and weights as a model for a string situated along the x-axis and free to stretch to the right and the left along the x-axis. the last line ﬁnds the eigenvalues of A.4 IdentityMatrix[n-1]. eval = Eigenvalues[N[a]]. If you run this program. When n is large. Finally.

04 -0.12 -0.02 -0.14 40 60 80 100 Figure 2. dt m and each of these could be solved just as before. we could consider even more complicated systems consisting of many masses and springs connected in a two. · zn  then an argument similar to the one given above would show that the vectors x.08 -0.06 -0. y and z would satisfy the matrix diﬀerential equations dx k = Ax. dt m dy k = Ay. 61 .  ·  xn   z1 z  z =  2 .20 -0.  ·  xn   x1 x  y =  2 .9: Shape of the fundamental mode (n=100).or three-dimensional array.1 -0. If we were to set  x1 x  x =  2 . dt m dz k = Az. Using a similar approach.

2 1 Fourier’s 2 See research was published in his Th´orie analytique de la chaleur in 1822. useful not only as a tool for solving partial diﬀerential equations. which can be used to send pictures quickly over the internet. A wavelet tour of signal processing. The light is a superposition of many frequencies which encode information regarding the material which makes up the stars of the galaxy. Just as a prism or a spectrograph is an experimental apparatus for dividing light into its components of various frequencies. as we will see in the next two chapters. that the function f (t) represents the amplitude of a light wave arriving from a distant galaxy. its components of various frequencies. Boston.” This idea is the key to image compression. Indeed. a black and white photograph could be represented by a function f (x. we will describe how the theory of Fourier series can be used to analyze the ﬂow of heat in a bar and the motion of a vibrating string. y) of two variables. e 62 . Joseph Fourier’s original investigations which led to the theory of Fourier series were motivated by an attempt to understand heat ﬂow.1 Fourier series The theory of Fourier series and the Fourier transform is concerned with dividing a function into a superposition of sines and cosines. and so forth. so Fourier analysis is a mathematical technique which enables us to decompose an arbitrary function into a superposition of oscillations.Chapter 3 Fourier Series 3. Suppose. The photograph can be stored eﬃciently by determining the components of f (x. e St´phane Mallat. sound waves and light waves.1 Nowadays. f (x. It is a crucial tool for understanding waves. y). y) representing the darkness at the point (x. including water waves. For example. Much of our knowledge of the universe is derived from analyzing the spectra of stars and galaxies. 1998. the notion of dividing a function into its components with respect to an appropriate “orthonormal basis of functions” is one of the key ideas of applied mathematics. y) with respect to a well-chosen “wavelet basis. the speed with which the galaxy is receding from the earth. but for many other purposes as well. In the following chapter. its speed of rotation. Academic Press. for example.

+ b1 sin t + b2 sin 2t + . The main theorem from the theory of Fourier series states that any “wellbehaved” periodic function of period T can be expressed as a superposition of sines and cosines: f (t) = a0 2πt 4πt 2πt 4πt + a1 cos( ) + a2 cos( ) + . 2 T T T T (3. for all t ∈ R . the ak ’s and bk ’s are called the Fourier coeﬃcients of f . + −π b1 sin tdt + −π b2 sin 2tdt + . Given an arbitrary period T . The coeﬃcient a0 is particularly easy to evaluate. Our ﬁrst goal is to determine how to calculate these Fourier coeﬃcients. . . . . For simplicity. . the function f (t) = sin( 2πt ) is periodic T of period T because sin( 2π(t + T ) 2πt 2πt ) = sin( + 2π) = sin( ). and are based upon exactly the same ideas.. (3. . . . the functions cos( 2πkt ) T and sin( 2πkt ) T are also periodic functions of period T .2) from −π to π: π π f (t)dt = −π −π a0 dt + 2 π π −π π a1 cos tdt + π −π a2 cos 2tdt + . Since the integral of cos kt or sin kt over the interval from −π to π vanishes. + b1 sin( ) + b2 sin( ) + . it is easy to construct examples of functions which are periodic of period T —indeed.2) 2 The formulae for a general period T are only a little more complicated. . . Thus f (t) = sin t is periodic of period 2π. A function f : R → R is said to be periodic of period T if it satisﬁes the relation f (t + T ) = f (t)..We turn now to the basics of Fourier analysis in its simplest context. we will restrict our attention to the case where the period T = 2π. T T T More generally. so that a0 f (t) = + a1 cos t + a2 cos 2t + . if k is any positive integer.1) In this formula. . We simply integrate both sides of (3. . 63 . . we conclude that π −π f (t)dt = πa0 . and the inﬁnite series on the right-hand side is called the Fourier series of f .

−π (3. cos((n − m)t) = cos nt cos mt + sin nt sin mt. we obtain cos((n + m)t) + cos((n − m)t) = 2 cos nt cos mt. 64 . for m = n. for m = n. We will use the trigonometric identities cos((n + m)t) = cos nt cos mt − sin nt sin mt. for m = n.6) in the exercises at the end of this section. (3. for m = n. for m = n. we conclude that π cos nt cos mtdt = −π π. (3.and we can solve for a0 .5) π sin nt cos mtdt = 0. We claim that if m and n are positive integers.3) To ﬁnd the other Fourier coeﬃcients. π and since π cos(kt)dt = −π 1 sin(kt) k = 0. −π if k = 0.5) and (3. 0. obtaining a0 = 1 π π f (t)dt.6) Let us verify the ﬁrst of these equations. 2 1 2 π −π cos nt cos mtdt = −π (cos((n + m)t) + cos((n − m)t))dt. 0. or cos nt cos mt = Hence π 1 (cos((n + m)t) + cos((n − m)t)). 0.4) π sin nt sin mtdt = −π π. π cos nt cos mtdt = −π π. we will need some integral formulae. The reader is asked to verify the other two integral formulae (3. for m = n. Adding these together. −π (3.

8) Example. .. .To ﬁnd the formula for the Fourier coeﬃcients ak for k > 0. .2) by sin kt and integrate from −π to π: π π f (t) cos ktdt = −π −π a0 cos ktdt + 2 π π −π a1 cos t cos ktdt + . extended to be periodic of period 2π. Then a0 = average value of f = 0. we multiply both sides of (3.8) to ﬁnd the Fourier coeﬃfor −π < t < 0. . for 0 < t < π..   0. there is only one term which survives: π π f (t) cos ktdt = −π −π ak cos kt cos ktdt = πak .  f (t) = π. + −π According to our formulae (3. (3.4) and (3. −π 0 π m πm 1 π 0 −π f (t) sin mtdt = −π −π sin mtdt + 1 π π π sin mtdt 0 65 . bj sin jt cos ktdt + . . +. . +. Let us use formulae (3..7). and (3.. −π (3. for t = 0.6).3). We can easily solve for ak : ak = 1 π π f (t) cos ktdt.7) A very similar argument yields the formula bk = 1 π π f (t) sin ktdt. + −π π ak cos kt cos ktdt + . . 2 and am = 1 π π 0 −π π f (t) cos mtdt = −π 1 π −π cos mtdt + 1 π π cos mtdt 0 = while bm = 1 π π 1 −π 1 π [sin mt]0 + [sin mt]π = · · · = 0. −π (3. π. cients of the function  −π.

1 π 1 −π 2 2 [cos mt]0 + [cos mt]π = − cos mπ −π 0 πm π m m m = 2 (1 − (−1)m ) = m 4 m. By the way.3 2 1 -4 -2 -1 -2 -3 2 4 Figure 3. f (x) = π.1: A graph of the Fourier approximation φ5 (t) = 4 sin t + (4/3) sin 3t + (4/5) sin 5t. if m is even. If we set x = π/2. φ3 (t) = 4 sin t + φ5 (t) = 4 sin t + 4 sin 3t 3 4 4 sin 3t + sin 5t + · · · . if m is odd. this Fourier series yields a curious formula for π. Thus we ﬁnd the Fourier series for f : f (t) = 4 sin t + The trigonometric polynomials φ1 (t) = 4 sin t. = 0. 3 5 7 9 4 4 sin(3π/2) + sin(5π/2) + · · · 3 5 66 . 3 5 4 4 sin 3t + sin 5t 3 5 are approximations to f (t) which improve as the number of terms increases. and we obtain π = 4 sin(π/2) + from which we conclude that π = 4(1 − 1 1 1 1 + − + − · · · ).

Moreover. If f is a continuous periodic function of period T .9). The technical deﬁnition goes like this: A function f (t) which is periodic of period T is said to be piecewise smooth if it is continuous and has a continuous derivative f (t) except at ﬁnitely many points of discontinuity within the interval [0.3 2 1 -4 -2 -1 -2 -3 2 4 Figure 3. T ]. suﬃcient for many applications is: Theorem 1.1) is often conveniently expressed in the Σ notation.” The Fourier series on the right-hand side of (3.9) converges to f (t) for every choice of t. in fact to functions that are not even be continuous. the right. where the ak ’s and bk ’s are constants. in accordance with (3. then f can be written uniquely as a superposition of sines and cosines.and left-handed limits of 67 . f (t) = a0 + 2 ∞ ak cos( k=1 2πkt )+ T ∞ bk sin( k=1 2πkt ). A weaker suﬃcient condition for f to possess a Fourier series is that it be piecewise smooth. often one wants to apply the theory of Fourier series to functions which are not quite so well-behaved. However. T (3. The overshooting near the points of discontinuity is known as the “Gibbs phenomenon. It is an interesting and diﬃcult problem in harmonic analysis to determine how “well-behaved” a periodic function f (t) must be in order to ensure that it can be expressed as a superposition of sines and cosines.9) just as we did for power series in Chapter 1. with condinuous derivative f . An easily stated theorem. such as the function in our previous example. and at each point t0 of discontinuity.2: A graph of the Fourier approximation φ13 (t). the inﬁnite series on the right-hand side of (3.

Here equality means that the inﬁnite sum on the right converges to f (t) for each t at which f is continuous.1. Find its Fourier series. t→t0 + lim (f (t)).1. The function f (t) = t. 3. McGraw-Hill. which were given in the text. If f is discontinuous at t0 . Benjamin. for −π < t < π.3 ensures that a Fourier decomposition can be found for any function which is piecewise smooth: Theorem 2. for t ∈ [−π. 3 See.1. 0.f and f . can be extended to be periodic of period 2π. The function f (t) = cos2 t can be regarded as either periodic of period π or periodic of period 2π. Find the Fourier series of this extension.5.) 3. t→t0 − lim (f (t)). New York.1. The following theorem. f (t) = a0 + 2 ∞ ak cos( k=1 2πkt )+ T ∞ bk sin( k=1 2πkt ) T where the ak ’s and bk ’s are constants. can be extended to be periodic of period 2π. t→t0 + lim (f (t)). An introduction to Fourier series and integrals.3.4. 1987 or Robert Seeley. (Hint: This problem is very easy if you use trigonometric identities instead of trying to integrate directly. New York. The function f (t) = sin3 t is periodic of period 2π. 68 . Establish the formulae (3. Find the Fourier series of the following function: f (t) = t2 .5) and (3.1. f (t − 2kπ). for −π + 2kπ ≤ t < π + 2kπ. for t = π. all exist. 4th edition.6). If f is any piecewise smooth periodic function of period T . Exercises: 3.1.1. Ruel Churchill and James Brown. f can be expressed as a Fourier series.6.2. 3. π]. 3. Find the Fourier series of this extension. Fourier series and boundary value problems. its Fourier series at t0 will converge to the average of the right and left hand limits of f as t → t0 . 1966. for −π ≤ t < π. Choose one of the two periods and ﬁnd the Fourier series of f (t). The function f (t) = |t|. proven in more advanced books. t→t0 − lim (f (t)). 3. for example.

. ˆ e2 (t) = sin 2t. • f + g. . these operations satisfying the same rules as those for addition of ordinary vectors and multiplication of ordinary vectors by scalars.5). f = 1 and that two elements f and g of V are perpendicular if f. e1 (t) = sin t. when c is a real constant.. .. We say that V is a vector space because elements of V can be added and multiplied by scalars.2 Inner products There is a convenient way of remembering the formulae for the Fourier coeﬃcients that we derived in the preceding section. h = f. g . . . sin 3t. . . the formulae (3. . h . (3. This suggests that we might use geometric terminology for elements of V just as we did for vectors in R n . any element of of V can be written as a (possibly inﬁnite) superposition of these functions.4). In this terminology. g = g. 2 e1 (t) = cos t. . if f (t) = sin t and g(t) = 2 cos t. . sin t. 69 . g = 1 π π π 2 sin t cos tdt = −π −π sin(2t)dt = − cos(2t)|π = 0. Thus. by the theorem in the preceding section. cos 2t. −π The remarkable fact is that this inner product has properties quite similar to those of the standard dot product on R n : • f.3. b2 . We saw in Section 2. we will say that an element f of V is of unit length if f. • cf. • f. g = 1 π π f (t)g(t)dt. then f ∈ R n ⇒ f = (f · b1 )b1 + · · · + (f · bn )bn . . then f. . f . e2 (t) = cos 2t. . g = 0. h + g. cos t. whenever f and g are elements of V .3 that if b1 .6) can be expressed by stating that the functions √ 1/ 2. We deﬁne an “inner product” between elements of V by means of the formula f. Moreover. sin 2t. . with equality holding only if f = 0 (at all points of continuity). bn is an orthonormal basis for R n . and (3. Let V be the set of piecewise smooth functions which are periodic of period 2π. ˆ make up an orthonormal basis for V . g = c f.. for example. cos 3t. −π Thus for example. We will say that the functions 1 e0 (t) = √ . . f ≥ 0. are of unit length and perpendicular to each other.

f g is odd and hence f.. a2 = f. cos 2t . in other words. Weven is a linear subspace of V . a1 = f. sin 2t . cos 2t cos 2t + . sin 2t sin 2t + . are odd functions. . Similarly. Thus we can say that Wodd is a linear subspace of V . . g ∈ Wodd ⇒ f + g ∈ Wodd and cf ∈ Wodd . are even functions.The same formula holds when R n is replaced by V and the dot product is replaced by the inner product ·. It is not diﬃcult to show that f ∈ Wodd . . e2 (t) e2 (t) + · · · . 1 1 f (t) = f. under these conditions.. e1 (t) e1 (t) + f (t). a0 = + a1 cos t + a2 cos 2t + . . sin 3t. cos t. Indeed. e0 (t) e0 (t) + f (t). sin t sin t + f. Thus sin t. . We say that a function f (t) is odd if f (−t) = −f (t). e2 (t) e2 (t) + · · · ˆ ˆ ˆ + f (t). sin 2t. Let Wodd = {f ∈ V : f is odd}. g ∈ Weven ⇒ f.. cos t .. . . b2 = f. + b1 sin t + b2 sin 2t + . . 2 2 + f. .. Then f. . · : If f is any element of V . cos t cos t + f. g = 1 π π f (t)g(t)dt = −π 1 π 0 f (t)g(t)dt + −π 1 π π f (t)g(t)dt 0 70 . while 1. Use of the inner product makes the formulae for Fourier coeﬃcients almost impossible to forget. b1 = f. we can write f (t) = f (t). . e1 (t) e1 (t) + f (t). 1 . √ √ + f. cos 2t. Weven = {f ∈ V : f is odd}. even if f (−t) = f (t).. 2 where a0 = f.. . ˆ In other words. . g = 0. sin t .. . for every choice of real number c. cos 3t. the linear subspaces Wodd and Weven are orthogonal to each other. . .

= 1 π 0 f (−t)g(−t)(−dt) + π 1 π π f (t)g(t)dt 0 =− 1 π π −(f (t)g(t))(−dt) + 0 1 π π f (t)g(t)dt = 0. 3. the Bessel function of the ﬁrst kind.2. b. for n > 0.2. if f ∈ Weven . d. −π in the case where f (t) = cos t and g(t) = | sin t|. x in the case where f (x) = cos(log(x)) and g(x) = 1. bn = f. and which are neither even nor odd: a. Exercises: 3.2. 0 The variable of integration has been changed from t to −t in the ﬁrst integral of the second line. f (t) = 1 (et + e−t ).3. sin nx = 0. a0 = f. f (t) = t2 + |t|. This simple fact can often simplify the calculation of Fourier coeﬃcients. Thus for an even or odd function. f (t)g(t)dt.) 3. which are odd. 71 .1. c. 2 f. 2 e. g = 1 1 f (x)g(x)dx. half of the Fourier coeﬃcients are automatically zero. f (t) = t3 + 3t. cos nx = 0. f (t) = 1 (et − e−t ). f (t) = et . (Hint: Use the substitution x = eu . Determine which of the following functions are even. Evaluate the inner product eπ f. Similarly. 1 = 0. Evaluate the inner product f. It follows that if f ∈ Wodd . where log denotes the natural or base e logarithm.2. f (t) = J0 (t). g = 1 π π an = f.

0]. L]. for t ∈ [−L. L] → R into a superposition of cosine functions. (3. This time the extended function lies in the linear subspace Weven . L]. ˆ then to a function f : R → R.3.10) the Fourier sine series of f . To generate formulae for the coeﬃcients in the Fourier sine and cosine series of f . .11) the Fourier cosine series of f .10) We could prove this using the theory of even and odd functions. . . which is periodic of period 2L by requiring that ˆ ˆ f (t + 2L) = f (t). . The extended function lies in the linear subspace Wodd . for all t ∈ R . (3. f restricts ˆ fact that f ˆ to f and the Fourier expansion of f restricts to an expansion of f of the form (3. On the interval [0. . −f (−t). We claim that we can express f (t) as the superposition of sine functions. we ﬁrst extend f to an even function f : [−L. ˜ f (t) = f (−t). for t ∈ [0. and from the ˆ is odd that all of the an ’s are zero. It follows from the ˆ theorem in Section 3. Indeed. ˆ then to a function f : R → R which is periodic of period 2L. ˆ ˆ f restricts to f and the Fourier expansion of f restricts to an expansion of f of the form (3. It follows ˆ from the theorem in Section 3.1 that f possesses a Fourier series expansion. . L]. we can ˜ extend f to an odd function f : [−L. f (t) = a0 + a1 cos(πt/L) + a2 cos(2πt/L) + . .3 Fourier sine and cosine series Let f : [0. 0]. We call (3. + an cos(nπt/L) + . we begin by deﬁning a slightly diﬀerent inner product space than the one 72 . .11) 2 ˜ To obtain this expression. On the interval [0.10) which involves only sine functions. f (t) = b1 sin(πt/L) + b2 sin(2πt/L) + . . L] → R by setting f (t). and ˆ from the fact that f is even that all of the bn ’s are zero. for t ∈ [−L. . L] → R be a piecewise smooth function which vanishes at 0 and L. A similar argument can be used to express a piecewise smooth function f : [0. + bn sin(nπt/L) + . We call (3. for all t ∈ R R.1 that f possesses a Fourier series expansion. by requiring that ˆ ˆ f (t + 2L) = f (t). L] → R by setting ˜ f (t) = f (t).11) which involves only cosine functions. L]. for t ∈ [0.

for m = n. L 0 This restricts to an inner product on V0 . . we ﬁnd that f (t) = b1 sin(πt/L) + b2 sin(2πt/L) + . sin(nπt/L). + sin(nπt/L) + . Recall the trigonometric formulae that we used in §3.. g = f (t)g(t)dt. We have seen that any element of V0 can be represented as a superpostion of the sine functions sin(πt/L). . cos((n − m)πt/L) = cos(nπt/L) cos(mπt/L) + sin(nπt/L) sin(mπt/L). · on V by means of the formula 2 L f.. This time.. sin(2πt/L). just as in the previous section.1: cos((n + m)πt/L) = cos(nπt/L) cos(mπt/L) − sin(nπt/L) sin(mπt/L). . We claim that these sine functions form an orthonormal basis for V0 with respect to the inner product we have deﬁned. 2 sin(nπt/L) sin(mπt/L)dt = 0 1. . .considered in the preceding section. . When we do this. 73 . We deﬁne an inner product ·. 2 sin(nπt/L) sin(mπt/L)dt = 1 2 L (cos((n − m)πt/L) − cos((n + m)πt/L)dt. 0. Subtracting the ﬁrst of these from the second and dividing by two yields sin(nπt/L) sin(mπt/L) = and hence 0 L 1 (cos((n − m)πt/L) − cos((n + m)πt/L). L] → R and let V0 = {f ∈ V : f (0) = 0 = f (L)}. for m = n.12) Therefore. . 0 If n and m are positive integers. Let’s consider now the Fourier sine series. in which case the right-hand side becomes 1 2 Hence 2 L π L dt = 0 L . (3.. a linear subspace of V . we can evaluate the coeﬃcients of the Fourier sine series of a function f ∈ V0 by simply projecting f onto each element of the orthonormal basis. . . the integral on the right vanishes unless n = m. we let V be the set of piecewise smooth functions f : [0.

Thus we can evaluate the coeﬃcients of the Fourier cosine series of a function f ∈ V by projecting f onto each element of this orthonormal basis. . πn2 74 .. + 2 [sin nt]|0 = + 2n n 2n n2 −(π − t) cos(nt) n π π (π − t) sin ntdt = π/2 − π/2 π/2 1 cos ntdt n = Thus π cos(nπ/2) π cos(nπ/2) sin(nπ/2) 1 . We use integration by parts to obtain π/2 t sin ntdt = 0 −t cos nt n π/2 π/2 + 0 0 1 cos ntdt n = while π −π cos(nπ/2) −π cos(nπ/2) sin(nπ/2) 1 π/2 . π − t.15) In this case.. We will leave it to the reader to carry this out in detail. L = π. − 2 [sin nt]|π = + π/2 2n n 2n n2 bn = 4 sin(nπ/2) . and simply remark that when the dust settles. . 0 (3. 0 (3.14) Example. form an orthonormal basis for V .where bn = f. First let us use (3. . cos(nπt/L)..13) to ﬁnd the Fourier sine series of f (t) = t. 2 cos(πt/L). one obtains the following formula for the coeﬃcient an in the Fourier cosine series: an = 2 L L f (t) cos(nπt/L)dt. bn = 2 π π/2 π t sin ntdt + 0 π/2 (π − t) sin ntdt . sin(nπt/L) = 2 L L f (t) sin(nπt/L)dt. In this case. (3. we can show that the functions 1 √ . for 0 ≤ t ≤ π/2.. and according to our formula.13) We can treat the Fourier cosine series in a similar fashion. cos(2πt/L). for π/2 ≤ t ≤ π.

for π/2 ≤ t ≤ π. π 9π are better and better approximations to the function f (t).5 2 2.2 1 0. .4 0.3: A graph of the Fourier sine approximations φ1 .8 0.1. t. To ﬁnd the Fourier cosine series of f (t) = we ﬁrst note that a0 = 2 π π/2 π for 0 ≤ t ≤ π/2.16) The trigonometric polynomials φ1 (t) = 4 sin t. π φ3 (t) = 4 4 sin t − sin 3t.14): an = 2 π π/2 π t cos ntdt + 0 π/2 (π − t) cos ntdt .5 1 1.5 3 Figure 3. integration by parts yields π/2 t cos ntdt = 0 t sin nt n 75 π/2 π/2 − 0 0 1 sin ntdt n . 2 To ﬁnd the other an ’s. .2 0. tdt + 0 π/2 (π − t)dt = 2 1 π 2 π 2 2 + 1 2 π 2 2 = π . .4 1. . we use (3. and the Fourier sine series of f (t) is f (t) = 4 4 4 4 sin t − sin 3t + sin 5t − sin 7t + . . π 9π 25π 49π (3. . φ3 . π − t. φ5 and φ7 .6 0. This time.

for π/2 ≤ t ≤ π. . for 0 ≤ t < π/2.a. π]: 4t. b.= 1 π sin(nπ/2) π/2 + 2 [cos nt]|0 2n n π sin(nπ/2) cos(nπ/2) 1 + = − 2 2n n2 n (π − t) sin(nt) n π π while π (π − t) cos ntdt = π/2 + π/2 π/2 1 sin ntdt n −π sin(nπ/2) 1 − 2 [cos nt]|π π/2 2n n −π sin(nπ/2) cos(nπ/2) 1 + = − 2 cos(nπ). 76 .3. Find the Fourier sine series of the following function deﬁned on the interval [0. 10]: t. π] as either a superposition of sines in (3.3. Find the Fourier cosine series of the same function.a.1. Find the Fourier sine series of the following function deﬁned on the interval [0.2. for 0 ≤ t < 5. 3.17). Exercises: 3. b. . f (t) = 4π − 4t. 2 2n n n = Thus when n ≥ 1.17) Note that we have expanded exactly the same function f (t) on the interval [0.16) or as a superposition of cosines in (3. Find the Fourier sine series of the following function deﬁned on the interval [0. an = 2 [2 cos(nπ/2) − 1 − 1(−1)n ]. 4 π 9π 25π (3.3. . f (t) = 10 − t. π]: f (t) = t(π − t). πn2 and the Fourier sine series of f (t) is f (t) = π 2 2 2 − cos 2t − cos 6t − cos 10t − . 3. for 5 ≤ t ≤ 10.3. Find the Fourier cosine series of the same function.

2 t∧3) Sin[n Pi t]. by running the following Mathematica program f[n ] := 2 NIntegrate[(t + t∧2 . (3.18) b1 it b2 (e − e−it ) + (e2it − e−2it ) + . 2i 2i . 1]: f (t) = 5t(1 − t). which states that eiθ = cos θ + i sin θ. . {t. . . . We say that this is the real form of the Fourier series.1. or e−iθ + e−iθ = 2i sin θ. It follows from this formula that eiθ + e−iθ = 2 cos θ.. . . .0. f can be expanded in a Fourier series f (t) = a0 + a1 cos t + a2 cos(2t) + . It is often convenient to recast this Fourier series in complex form by means of the Euler formula. . deﬁned for t in the interval [0. . 2 +b1 sin t + b2 sin(2t) + .3.. sin θ = .10}] 3. b = Table[f[n]. . 2 77 c2 = a2 − ib2 . 2 2i Hence the Fourier expansion of f can be rewritten as cos θ = f (t) = a0 a1 a2 + (eit + e−it ) + (e2it + e−2it ) + . 3.(For students with access to Mathematica) Find numerical approximations to the ﬁrst ten coeﬃcients of the Fourier sine series for the function f (t) = t + t2 − 2t3 .5.3. . . . . eiθ + e−iθ eiθ + e−iθ . 1]. . 2 2 2 + or f (t) = . .3.4. 2 . Find the Fourier sine series of the following function deﬁned on the interval [0. 2 c1 = a1 − ib1 . where the ck ’s are the complex numbers deﬁned by the formulae c0 = a0 . {n. + c−2 e−2it + c−1 e−it + c0 + c1 eit + c2 e2it + .4 Complex version of Fourier series* We have seen that if f : R → R is a well-behaved function which is periodic of period 2π.1}].

du = dt and v = (i/k)e−ikt : π 1 ck = (i/k)e−ikt dt (it/k)e−ikt |π − −π 2π −π = It follows from (3. Then ck = 1 2π π −π te−ikt dt.20) If we multiply both sides of (3. c−2 = .19) It is not diﬃcult to check the following integral formula via direct integration: π −π eimt e−int dt = 2π. 2 2 If k = 0.19) to ﬁnd the real form of the Fourier series. integrate from −π to π and apply (3.1. Let us use formula (3. We apply integration by parts with u = t. . 0 for m = n. . (3. which yields the formula for coeﬃcients of the complex form of the Fourier series: ck = 1 2π π −π f (t)e−ikt dt. for m = n.21) Example.4. bk = i(ck − c−k ) = −2 (−1)k . bk = i(ck − c−k ).. 2π k k k It is often the case that the complex form of the Fourier series is far simpler to calculate than the real form. presented in the text.a1 + ib1 a2 + ib2 .18) by e−ikt . extended to be periodic of period 2π. we can solve for ak and bk : c−1 = ak = ck + c−k . dv = e−ikt dt.21) to ﬁnd the complex Fourier coeﬃcients of the function f (t) = t for −π < t ≤ π. we obtain π −π f (t)e−ikt dt = 2πck . Prove the integral formula (3.20). (3. 78 .19) that ak = (ck + c−k ) = 0. k 1 πi −iπk πi iπk (−1)k =i + e e . (3.. Exercise: 3. One can then use (3.20).

Find the complex Fourier coeﬃcients of the function f (t) = t(π − t) extended to be periodic of period 2π. There are many times when one would like to divide a function which is not periodic into a superposition of sines and cosines. f (t) = . 3. Deﬁnition.19) to ﬁnd the real form of the Fourier series. it is best to regard it as a limit.2. If f : R → R is a piecewise continuous function which vanishes outside some ﬁnite interval. . b. . (3. for −π < t ≤ π. Use (3. its Fourier transform is ˆ f (ξ) = ∞ −∞ f (t)e−iξt dt. .4.5 Fourier transforms* One of the problems with the theory of Fourier series presented in the previous sections is that it applies only to periodic functions. a.3. b. . 3. Show that if a function f : R → R is smooth and periodic of period 2πL.4. we can write the Fourier expansion of f as f (t) = . Use (3. The Fourier transform is the tool often used for this purpose. . ∞ −∞ f (t)e−iξt dt = lim πL −πL L→∞ f (t)e−iξt dt. where the ck ’s are the complex numbers.4. .4). . .3. a. f (t)e−ikt/L dt. 3. πL]. 79 . According to the theory of Fourier series in complex form (see Exercise 3. + c−2 e−2it/L + c−1 e−it/L + c0 + c1 eit/L + c2 e2it/L + .4.4. where ck = 1 2πL πL −πL for −π < t ≤ π.22) The integral in this formula is said to be improper because it is taken from −∞ to ∞.19) to ﬁnd the real form of the Fourier series. . . The function can be extended to a periodic function f (t) such that f (t + 2πL) = f (t). Find the complex Fourier coeﬃcients of the function f (t) = t2 extended to be periodic of period 2π. + c−2 e−2it/L + c−1 e−it/L + c0 + c1 eit/L + c2 e2it/L + . The idea behind the Fourier transform is to think of f (t) as vanishing outside a very long interval [−πL.

where ck = or alternatively. for t ∈ [−πL.22). 2πL 2πL 2πL In the limit as L → ∞. we can write the Fourier inversion formula in terms of sines and cosines.To explain (3. . . Like the Laplace transform. the Fourier transform is often an eﬀective tool in ﬁnding explicit solutions to diﬀerential equations. and our formula becomes f (t) = 1 2π ∞ −∞ ˆ f (ξ)eiξt dξ. .21) that we can write f (t) = . Exercise: 3. . (3. Find the Fourier transform of the function f (t) deﬁned by f (t) = 1. . it follows from (3. otherwise.1. . We can extend the restriction of f to this interval to a function which is periodic of period 2πL. Then ˆ f (k/L) = ∞ −∞ f (t)e−ikt/L dt = πL −πL ˜ f (t)e−ikt/L dt represents 2πL times the Fourier coeﬃcient of this extension of frequency k/L. . πL]. 80 . + + 1 ˆ 1 ˆ f (−2/L)e−2it/L + f (−1/L)e−it/L + 2πL 2πL 1 ˆ f (k/L).22) and (3.23) Equation (3. + c−2 e−2it/L + c−1 e−it/L + c0 + c1 eit/L + c2 e2it/L + . .22) allow one to pass back and forth between a given function and its representation as a superposition of oscillations of various frequencies.5. we suppose that f vanishes outside the interval [−πL. 0. πL]. If we make use of Euler’s formula.23) is called the Fourier inversion formula. f (t) = . indeed. . 2πL 1 ˆ 1 ˆ 1 ˆ f (0) + f (1/T )eit/L + f (2/L)e2it/L + . Equations (3. if −1 ≤ t ≤ 1. it can be shown that this last sum approaches an improper integral. f (t) = 1 2π i ˆ f (ξ) cos ξtdξ + 2π −∞ ∞ ∞ −∞ ˆ f (ξ) sin ξtdξ. . a superposition of sines and cosines of various frequencies.

Higher dimensional examples will be given in the following chapter. Nonlinear partial diﬀerential equations are far more 81 . and we will show how to solve them by the techniques of “separation of variables” and Fourier analysis. such as the equation ∂u ∂2u = . Fluid mechanics is often formulated by the Euler equations of motion or the so-called Navier-Stokes equations. electricity and magnetism by Maxwell’s equations. Instead.Chapter 4 Partial Diﬀerential Equations 4. In this chapter.1 Overview A partial diﬀerential equation is an equation which contains partial derivatives. each type of partial diﬀerential equations exhibits its own special features. Many of the foundational theories of physics and engineering are expressed by means of systems of partial diﬀerential equations. It is therefore important to develop techniques that can be used to solve a wide variety of partial diﬀerential equations. general relativity by Einstein’s ﬁeld equations. which usually mirror the physical phenomena which the equation was ﬁrst used to model. The reader may have heard some of these equations mentioned in previous courses in physics. there is an important dichotomy between linear and nonlinear equations. we will give two important simple examples of partial differential equations. The techniques of separation of variables and Fourier analysis are eﬀective only for linear partial diﬀerential equations. Unlike the theory of ordinary diﬀerential equations which centers upon one key theorem—the fundamental existence and uniqueness theorem—there is no real uniﬁed theory of partial differential equations. the heat equation and the wave equation. We will see that just as in the case of ordinary diﬀerential equations. ∂t ∂x2 in which u is regarded as a function of x and t.

t)dx. To calculate the rate of change of heat within Dx1 . t) − F (x1 . t)dx.x2 is given by the exp Rate at which heat leaves Dx1 . ∂t (4. F (x. then the rate at which heat leaves Dx1 . t)dx = x1 x1 ρ(x)σ(x) ∂u (x. 82 . Partial diﬀerential equations: basic theory. Heat in a small segment of a homogeneous bar is proportional to temperature.x2 = x1 ρ(x)σ(x)u(x. then the heat within the region Dx1 . t) = x2 x1 ∂F (x. Thus we ﬁnd that the rate at which heat leaves the region Dx1 . t)dx + ∂t x2 (µ(x)u(x. Partial diﬀerential equations and boundary-value problems with applications. and conservation of energy implies that if no heat is being created or destroyed in Dx1 .x2 . We especially recommend Mark Pinsky.x2 is also given by the formula F (x2 . and form a key topic of contemporary mathematical research. More generally. the rate of change of heat within Dx1 .1) (More generally. 2nd edition. t)dx. t) = −κ(x) ∂u (x.x2 with respect to time. ∂x (4. if σ(x) denotes the speciﬁc heat at the point x and ρ(x) is the density of the bar at x.x2 = − x2 ρ(x)σ(x) x1 ∂u (x. Hence the rate at which heat leaves Dx1 . 1991. the constant of proportionality being determined by the density and speciﬁc heat of the material making up the bar.x2 . t) + ν(x))dx. Springer.3) 1 Further reading can be found in the many excellent upper-division texts on partial diﬀerential equations. at the rate µ(x)u(x. if heat is being created within Dx1 .diﬃcult to solve. t)dx. t). ∂x (4. McGraw-Hill. the rate of heat ﬂow F (x. we simply diﬀerentiate under the integral sign: d dt x2 x2 ρ(x)σ(x)u(x. ∂t Now heat is a form of energy.x2 . t) = temperature of the bar at the point x at time t.x2 between x1 and x2 is given by the formula x2 Heat within Dx1 .x2 is simply the rate at which heat enters Dx1 . say by a chemical reaction.x2 is − x2 ρ(x)σ(x) x1 ∂u (x. t) is proportional to the partial derivative of temperature.) x1 On the other hand. t) + ν(x) per unit volume. We imagine that the bar is located along the x-axis and we let u(x. An excellent but much more advanced book is Michael Taylor. New York.1 Our ﬁrst example is the equation governing propagation of heat in a bar of length L. 1996.2) where κ(x) is the thermal conductivity of the bar at x.

Press. (4.1. ρ(x). one could show that heat ﬂow is modeled by the equation ρ(x)σ(x) ∂u ∂ = ∂t ∂x κ(x) ∂u ∂x + µ(x)u + ν(x).5) In the special case where the bar is homogeneous. Indeed. 1995. in which u is independent of time. i. Study of heat ﬂow often leads to “boundary-value problems” for ordinary diﬀerential equations. t) + ν(x) per unit length. ∂t In this way. = ∂t ρσ ∂x2 (4. Although its most basic application concerns diﬀusion of heat. t)dx. The mathematics of ﬁnancial derivatives. ∂x ∂t It follows from (4. so the integrands on the two sides must be equal: ∂F ∂u = −ρ(x)σ(x) . and (4. ∂t This equation is true for all choices of x1 and x2 .1. its properties are the same at every point.1) and (4. equation (4. in the “steady-state” case.2 Exercises: 4. t)dx = − ∂x x2 ρ(x)σ(x) x1 ∂u (x.6) This is our simplest example of a linear partial diﬀerential equation.4) In the more general case in which heat is being created at the rate µ(x)u(x. (4. Sam Howison and Jeﬀ Dewynne.3). σ(x) and κ(x) are constants. it arises in many other contexts as well. For example. Cambridge Univ. dx 2 A description of the Black-Scholes technique for pricing puts and calls is given in Paul Wilmott. 83 . we obtain the heat equation ρ(x)σ(x) ∂u ∂ = ∂t ∂x κ(x) ∂u ∂x .5) becomes d dx κ(x) du (x) + µ(x)u(x) + ν(x) = 0.e. a slight modiﬁcation of the heat equation was used by Black and Scholes to price derivatives in ﬁnancial markets.2) that ∂ ∂x −κ(x) ∂u ∂x = −ρ(x)σ(x) ∂u .4) becomes ∂u κ ∂2u . say σ and κ respectively. we ﬁnd that x2 x1 ∂F (x.Comparing the two formulae (4.

{x.}. deﬁned for 0 ≤ x ≤ 1 such that d2 u = 0. u[b] == beta. Our physical intuition suggests that the steady-state heat equation should have a unique solution with these boundary conditions. Solve the following special case of this boundary-value problem: Find u(x). u(1) = 50. alpha = 50.a.b}] 4. u(1) = 0. Show that the function u0 (x. b = 1. b. dx2 u(0) = 70. deﬁned for 0 ≤ x ≤ 1 such that d2 u + x(1 − x) = 0.1. u.a. t = t to show that ¯ ua (x. sol = NDSolve[ {u’’[x] + (1 + x∧2) u[x] == 0.b}]. say u(0) = α. Use the chain rule with intermediate variables x = x − a. Solve the following special case of this boundary-value problem: Find u(x). deﬁned for 0 ≤ x ≤ 1 such that d2 u − u = 0. u(L) = β. Suppose now that the temperature is speciﬁed at the two endpoints of the bar.6) for t > 0 in the case where κ/(ρσ) = 1. Plot[ Evaluate[ u[x] /.a. Solve the following special case of this boundary-value problem: Find u(x). beta = 100. u[a] == alpha. dx2 u(0) = 70. {x. c. ¯ b. u(1) = 50.a linear ordinary diﬀerential equation with variable coeﬃcients. You can do this by running the Mathematica program: a = 0. dx2 u(0) = 50.2. t) = √ 1 −(x−a)2 /4t e 4πt 84 . (For students with access to Mathematica) Use Mathematica to graph the solution to the following boundary-value problem: Find u(x). u(1) = 100. d. dx2 u(0) = 0. deﬁned for 0 ≤ x ≤ 1 such that d2 u + (1 + x2 )u = 0. a. sol]. t) = √ 1 −x2 /4t e 4πt is a solution to the heat equation (4.

7) gives a formula (for values of t which are greater than zero) for the unique solution to the heat equation on the inﬁnite line which satisﬁes the initial condition u(x. t) satisﬁes the heat equation ∂u ∂2u = c2 2 . u(x. t) = −∞ ua (x. deﬁned for 0 ≤ x ≤ L and t ≥ 0 such that 1.2 The initial value problem for the heat equation We will now describe how to use the Fourier sine series to ﬁnd the solution to an initial value problem for the heat equation in a rod of length L which is insulated along the sides. then ∞ u(x. 4. Show that ∞ −∞ u0 (x. In fact. By diﬀerentiating under the integral sign show that if h : R → R is any smooth function which vanishes outside a ﬁnite interval [−L. u(x. Hint: Let I denote the integral on the left hand side and note that I2 = 1 4πt ∞ −∞ ∞ −∞ e−(x 2 +y 2 )/4t dxdy. t). t)dx = 1. 2. In the next section. the temperature is zero at the endpoints. t) satisﬁes the boundary condition u(0. t) as t → 0? e. t) = 0.7) approaches h(x) as t → 0. t) for various values of t. d. (4.) 85 (4. c. 0) = h(x). What can you say about the behaviour of the function u0 (x. We expect that there should exist a unique function u(x. Use Mathematica to sketch u0 (x. t)h(a)da (4. in other words. Then transform this last integral to polar coordinates.is also a solution to the heat equation. (This is sometimes called the Dirichlet boundary condition. ∂t ∂x where c is a constant.8) .7) is a solution to the heat equation. L]. REMARK: In more advanced courses it is shown that (4. we will see how to solve the initial value problem for a rod of ﬁnite length. t) = u(L. whose ends are kept at zero temperature.

Thus homogeneous linear conditions satisfy the principal of superposition. We will shortly see how to ﬁnd that solution.9) 86 . Our method makes use of the dichotomy into homogeneous and nonhomogeneous conditions: Step I.10) or g (t) = λc2 g(t). (4. In other words. (4. Now we separate variables. t) satisﬁes the initial condition u(x. Let us ﬁrst carry out Step I.3. We substitute u(x. u(x. for any choice of constants c1 and c2 . putting all the functions involving t on the left. g (t) = λ. We ﬁnd all of the solutions to the homogeneous linear conditions of the special form u(x. 2 g(t) c f (x) which separates into two ordinary diﬀerential equations. the two sides must equal a constant. the initial temperature of the rod. Hence neither side can depend upon either x or t. where h(x) is a given function. c2 g(t) and f (x) = λ. Note that the heat equation itself and the boundary condition are homogeneous and linear —this means that if u1 and u2 satisfy these conditions. an arbitrary linear superposition of these solutions will still be a solution.8) and obtain f (x)g (t) = c2 f (x)g(t). By the superposition principle. t) = f (x)g(t) into the heat equation (4. it is proven that this initial value problem does in fact have a unique solution. so does c1 u1 + c2 u2 . Step II. while the right-hand side does not depend on t. all the functions involving x on the right: g (t) f (x) = c2 . t) = f (x)g(t). Our equation now becomes g (t) f (x) = = λ. which we denote by λ and call the separating constant. 0) = h(x). f (x) or f (x) = λf (x). g(t) f (x) The left-hand side of this equation does not depend on x. In more advanced courses. We ﬁnd the particular solution which satisﬁes the nonhomogeneous condition by Fourier analysis.

only one of these will actually yield nontrivial solutions to our eigenvalue problem. the diﬀerential equation f (x) = λf (x). L] → R which vanish at the endpoints 0 and L.11) We will call (4. 2 √ √ 1 √ sinh( λx) = (e λx − e− λx ) 2 instead of the exponentials. t). the trivial solution. Then we can write √ √ f (x) = a cosh( λx) + b sinh( λx). dx2 f (0) = 0 = f (L). using √ √ 1 √ cosh( λx) = (e λx + e− λx ). In this case. In this case. f (0) = 0 = f (L). the eigenvalue problem (4. We need to consider three cases. Case 2: λ > 0. If g(t) is not identically zero. has the general solution √ or f (x) − λf (x) = 0 √ f (x) = c1 e λx + c2 e− λx . t) = u(L.The homogeneous boundary condition u(0. and the only particular solution which satisﬁes the boundary condition is f = 0.11) the eigenvalue problem for the diﬀerential operator L= d2 dx2 acting on the space V0 of well-behaved functions f : [0. f (0) = f (L) = 0. then so is u(x. and we obtain only the trivial solution u ≡ 0. The general solution to the diﬀerential equation is f (x) = ax + b.) Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the problem requires us to ﬁnd the nontrivial solutions to a boundary value problem for an ordinary diﬀerential equation: f (x) = d2 (f (x)) = λf (x). 87 . (If g(t) is identically zero.) Case 1: λ = 0. (As it turns out. (4. It is convenient for us to change basis in the linear space of solutions. t) = 0 becomes f (0)g(t) = f (L)g(t) = 0.11) becomes f (x) = 0.

t) = 0 = u(L. √ Case 3: λ < 0. where a is a constant of integration. where n is an integer. n = 1. f (L) = 0 ⇒ b sin(ωL) ⇒ b = 0 or sin(ωL) = 0. 2. This is just the equation of exponential decay.11) are constant multiples of f (x) = sin(nπx/L). ⇒ a=0 ⇒ f (x) = b sin(ωx). . g (t) = λc2 g(t). and hence either b = 0 and we obtain only the trivial solution or sin(ωL) = 0. or ω = (nπ/L). and then f (L) = 0 ⇒ √ b sinh( λL) = 0 ⇒ b=0 ⇒ f (x) = 0. In this case. We remember that the diﬀerential equation has the general solution f (x) = a cos(ωx) + b sin(ωx). For each of these solutions. 2 88 . and has the general solution 2 g(t) = be−(ncπ/L) t . We impose the boundary conditions: ﬁrst √ f (0) = 0 ⇒ a = 0 ⇒ f (x) = b sinh( λx). we need to ﬁnd a corresponding g(t) solving equation (4. . We recognize here our old friend. however. f (0) = 0 = f (L). the diﬀerential equation of simple harmonic motion. and rewrite the eigenvalue problem as f (x) + ω 2 f (x) = 0. t) = sin(nπx/L)e−(ncπ/L) t . we set ω = −λ. we obtain f (x) = b sin(nπx/L). In this case.9). so we obtain no nontrivial solutions in this case. . Once again f (0) = 0 Now. Therefore. we conclude that the only nontrivial solutions to (4. t) are constant multiples of un (x. 3. The latter possibility will occur if ωL = nπ.with new constants of integration a and b. where λ = −(nπ/L)2 . Thus we ﬁnd that the nontrivial product solutions to the heat equation together with the homogeneous boundary condition u(0. with λ = −(nπ/L)2 . .

the nonvanishing coeﬃcients for the Fourier sine series of h are b1 = 4. .3 that the Fourier sine series of h is h(x) = and hence u(x. t) = 0. t) = b1 sin(πx/L)e−(cπ/L) t + b2 sin(2πx/L)e−(2cπ/L) t + . deﬁned for 0 ≤ x ≤ π and t ≥ 0. u(x. π − x. . t) = u(π. . . = ∂t ∂x2 where h(x) = u(0. t) = 0. . t). In this case. b2 . deﬁned for 0 ≤ x ≤ π and t ≥ 0. It follows from the theory of the Fourier sine series that h can indeed be represented as a superposition of sine functions. which satisﬁes the initial-value problem: ∂u ∂2u . x.3. . = ∂t ∂x2 u(0. π 9π 25π 49π . . which satisﬁes the initial-value problem: ∂u ∂2u . Example 2.12) yields h(x) = u(x. and we can determine the bn ’s as the coeﬃcients in the Fourier sine series of h. Suppose that we want to ﬁnd the function u(x.12) so that the initial condition u(x. π 9π 25π 89 4 4 4 4 sin x − sin 3x + sin 5x − sin 7x + . for π/2 ≤ x ≤ π. Suppose that we want to ﬁnd the function u(x. . We saw in Section 3. b2 = 2. 0) = h(x) = 4 sin x+2 sin 2x+7 sin 3x. . t) = 4 sin xe−t + 2 sin 2xe−4t + 7 sin 3xe−9t . . for arbitrary choice of the constants b1 . . t) = 4 4 4 sin xe−t − sin 3xe−9t + sin 5xe−25t − . Step II consists of determining the constants bn in (4. Setting t = 0 in (4.12) is a solution to the heat equation together with its homogeneous boundary conditions. for 0 ≤ x ≤ π/2. .It follows from the principal of superposition that u(x. 0) = h(x) is satisﬁed. 0) = h(x). . t) = u(π. so the solution must be u(x. 2 2 (4. . u(x. b3 = 7. 0 Example 1. we ﬁnd that bn = 2 L L h(x) sin(nπx/L)dx. Using the techniques described in Section 3. 0) = b1 sin(πx/L) + b2 sin(2πx/L) + . . t).

b. u(0. = 2t + 1 ∂t ∂x2 u(0. t) = u(π. Find the function u(x. for n = 1. t).3. t). deﬁned for 0 ≤ x ≤ π and t ≥ 0. t).2. t). Find the function u(x.2. .4. u(π. (4. 4.2. which satisﬁes the following conditions: ∂u ∂2u = .2. deﬁned for 0 ≤ x ≤ π. t) = u(π. 3. 0) = x(π − x). w(π) = 50. . Find the function w(x). 0) = sin x + 3 sin 2x.a.Exercises: 4. Find the general solution to the following boundary value problem for the heat equation: Find the functions u(x. 0) = sin 2x. u(x. u(x. t) = 0. t). Find the function u(x. 4. t) = 50. Find the function u(x.2. 0) = sin x + 3 sin 2x. u(x. t) = u(π. u(0. where the bn ’s are constants. t) = 0. . f (0) = 0 = f (π) 4. f (x) = bn sin nx. 2. t) = u(π. = ∂t ∂x2 u(0. are λ = −n2 . 4. = ∂t ∂x2 You may assume that the nontrivial solutions to the eigenvalue problem f (x) = λf (x). u(x. deﬁned for 0 ≤ x ≤ π and t ≥ 0. Find the function u(x. t) = 0. which satisﬁes the following conditions: ∂u ∂2u .6. which satisﬁes the following conditions: (t + 1) ∂u ∂2u . deﬁned for 0 ≤ x ≤ π and t ≥ 0. such that ∂u ∂2u . = ∂t ∂x2 (In this problem you need to ﬁnd the Fourier sine series of h(x) = x(π − x). deﬁned for 0 ≤ x ≤ π and t ≥ 0.2. t) = 0. ∂t ∂x2 u(0. which satisﬁes the following conditions: ∂u ∂2u . such that d2 w = 0. .1.) 4.5. t) = 0. which satisﬁes the following conditions: 1 ∂u ∂2u . 0) = sin x + 3 sin 2x − 5 sin 3x. t). dx2 w(0) = 10. = ∂t ∂x2 u(0. t) = u(π.13) 90 . deﬁned for 0 ≤ x ≤ π and t ≥ 0.2. deﬁned for 0 ≤ x ≤ π and t ≥ 0. u(x. t) = 10.

91 . 0) = 10 + x + 2 sin x − 5 sin 2x. Solve this eigenvalue problem. t) = u(L. 4. (Hint: The solution should involve cosines instead of sines. π 4. Find the function u(x.) b. which consists of ﬁnding the nontrivial solutions to f (x) = d2 (f (x)) = λf (x).13) which in addition satisﬁes the initial condition 40 u(x.) c. dt2 b. t) = 0. Determine what conditions v must satisfy. t) = u(π. deﬁned for 0 ≤ x ≤ π and t ≥ 0. t).7. ∂x ∂x (4. t) = (L. t) = 0 is replaced by the Neumann boundary condition ∂u ∂u (0.) In this case separation of variables leads to a slightly diﬀerent eigenvalue problem. dx2 f (0) = 0 = f (L). where w is the solution to part a. u(x. t) = 0. The method described in this section can also be used to solve an initial value problem for the heat equation in which the Dirichlet boundary condition u(0. Find the eigenvalues and corresponding eigenfunctions for the diﬀerential operator d2 L = 2 + 3.(Hint: Let v = u − w. dt which acts on the space V0 of well-behaved functions f : [0. = ∂t ∂x2 u(0.14).14) (Physically.8. from which no heat can enter or escape. Find the particular solution to (4. this corresponds to insulated endpoints.2.a. π] → R which vanish at the endpoints 0 and π by L(f ) = d2 f + 3f. Find the general solution to the heat equation ∂u ∂2u = ∂t ∂x2 subject to the Neumann boundary condition (4.2. which satisﬁes the following conditions: ∂u ∂2u + 3u. a. 0) = sin x + 3 sin 2x.

t) = 0. Find the function u(x. say u(0. such that: ∂u ∂2u . t) = 0 and the initial condition u(x. deﬁned for 0 ≤ x ≤ π and t ≥ 0. 0) = 4 sin(x/2) + 12 sin(3x/2). t) = (π. a. t) = u(1.3 Numerical solutions to the heat equation There is another method which is sometimes used to treat the initial value problem described in the preceding section. ∂x (4. the space of piecewise smooth functions deﬁned for 0 ≤ x ≤ 1 which vanish at the endpoints. However. t) = ∂u (π. = ∂t ∂x2 u(0. t).2.15). t0 ) is an element of V0 .9. We can also treat a mixture of Dirichlet and Neumann conditions. representing the initial temperature. Find the function u(x. ∂x ∂x u(x. ∂x u(x. dx2 f (0) = 0 = f (L). which solves the heat equation ∂u ∂2u = ∂t ∂x2 subject to the boundary conditions u(0. for simplicity. and in fact we will assume that c2 = L = 1. 0) = 3 cos x + 7 cos 2x. Find the general solution to the heat equation ∂u ∂2u = ∂t ∂x2 subject to the mixed boundary condition (4. t) = 0. a numerical method based upon “ﬁnite diﬀerences. t) = ∂u (L.15) In this case separation of variables leads to the eigenvalue problem which consists of ﬁnding the nontrivial solutions to f (x) = d2 (f (x)) = λf (x). b. such that: ∂u ∂2u . c. where h(x) is a given function. t) = 0. 4. it can be applied in some cases with variable coeﬃcients when it would be impossible to apply Fourier analysis in terms of sines and cosines. t). For any ﬁxed choice of t0 the function u(x.c. 92 . t). = ∂t ∂x2 ∂u ∂u (0. deﬁned for 0 ≤ x ≤ π and t ≥ 0. 4. 0) = h(x). deﬁned for 0 ≤ x ≤ 1 and t ≥ 0.” Although it yields only approximate solutions. we will describe only the case where ρ and k are constant. Thus we seek the function u(x. Solve this eigenvalue problem.

t) = the temperature at xi at time t. ui+1 (t) − ui (t) = . let xi = i/n and ui (t) = u(xi .t 2 . the temperature at time t is speciﬁed by   u1 (t)  u (t)  .t − 2 ∂x xi−1 + xi . we can approximate the second-order partial derivative: ∂2u . =n ∂x ∂u ∂x xi +xi+1 . The initial condition becomes   h(x1 )  h(x2 )  . u(0) = h. = n2 [ui−1 (t) − 2ui (t) + ui+1 (t)]. Thus the partial diﬀerential equation ∂u ∂2u = ∂t ∂x2 can be approximated by a system of ordinary diﬀerential equations dui = n2 (ui−1 − 2ui + ui+1 ).t 2 − ∂u ∂x xi−1 +xi . This corresponds to utilizing a discrete model for heat ﬂow rather than a continuous one. u(t) =  2   · un−1 (t) a vector-valued function of one variable. Since u0 (t) = 0 = un (t) by the boundary conditions. For 0 ≤ i ≤ n. dt This is a ﬁrst order linear system which can be presented in vector form as   −2 1 0 ··· 0  1 −2 1 · · · 0    du 2  0 1 −2 · · · · . t) = ∂x2 ∂u .Our idea is to replace the “inﬁnite-dimensional” space V0 by a ﬁnite-dimensional Euclidean space R n−1 and reduce the partial diﬀerential equation to a system of ordinary diﬀerential equations. (xi . where h=   · h(xn−1 ) We can approximate the ﬁrst-order partial derivative by a diﬀerence quotient: ∂u xi + xi+1 [ui+1 (t) − ui (t)] .t 2 1/n xi + xi+1 ∂u . ∂x 2 xi+1 − xi 1/n Similarly.t = = n[ui+1 (t) − ui (t)]. = n P u. where P =   dt  · · · ··· ·  0 0 · · · · −2 93 .

An element u(x) ∈ V0 represents a conﬁguration of the string at some 3 For further discussion of this method one can refer to numerical analysis books. With suﬃcient eﬀort. This method for ﬁnding approximate solutions to the heat equation is often called the method of ﬁnite diﬀerences.3 For us.16). We will derive a partial diﬀerential equation for u(x. t). Finally.3. for example.16) a system exactly like the ones studied in Section 2. t) = vertical displacement of the string at the point x at time t.4 The vibrating string Our next goal is to derive the equation which governs the motion of a vibrating string. Seventh edition. to implement it. and the eigenvectors approximate more and more closely the standard orthonormal basis of sine functions. It is sometimes the case that either a partial diﬀerential equation or a system of ordinary diﬀerential equations with a large or even inﬁnite number of unknowns can give an eﬀective model for the same physical phenomenon. 2000. however. One could continue constructing a numerical method for solution of our initial value problem by means of another discretization. We could do this via the familiar Cauchy-Euler method for ﬁnding numerical solutions to the linear system (4. why quantum mechanics possesses two superﬁcially diﬀerent formulations. via Schr¨dinger’s partial difo ferential equation or via “inﬁnite matrices” in Heisenberg’s “matrix mechanics. Numerical analysis. we must have u(0. Brooks Cole Publishing Company. this time in the time direction.5. More advanced courses on numerical analysis often treat the ﬁnite diﬀerence method in detail. the main point of the method of ﬁnite diﬀerences is that it provides considerable insight into the theory behind the heat equation. such as Burden and Faires. Note that since the ends of the string are ﬁxed. It shows that the heat equation can be thought of as arising from a system of ordinary diﬀerential equations when the number of dependent variables goes to inﬁnity.6 to check that the eigenvalues of A approach the eigenvalues of d2 /dx2 as determined in the preceding section. In the limit as n → ∞ one can use the Mathematica program of §2. This partially explains.the last matrix having n − 1 rows and n − 1 columns. 94 . t) for all t. one end of the string being at x = 0 and the other being at x = L. We assume that the string is free to move only in the vertical direction. we can rewrite this as du = Au. dt where A = n2 P. using Mathematica or some other software package.” 4. (4. It will be convenient to use the “conﬁguration space” V0 described in Section 3. t) = 0 = u(L. We consider a string of length L stretched out along the x-axis. one could construct a computer program. Let u(x.

we could imagine that we have devised an experiment that measures the potential energy in the string in various conﬁgurations. We will assume that the potential energy in the string when it is in the conﬁguration u(x) is L V (u(x)) = 0 T 2 du dx 2 dx. so we can drop the constant term to obtain (4. proportional to the length of the string. 2 du dx 2 2 =1+ du dx 2 + a small error. When the force pushes the string through an inﬁnitesimal displacement ξ(x) ∈ V0 .17) where T is a constant. We imagine that the force acting on the portion of the string from x to x + dx is F (x)dx. Potential energy is only deﬁned up to addition of a constant. From vector calculus. But when du/dx is small.17) does indeed represent the total potential energy in the string. or in other words. Indeed. 1 1+ 2 and hence 1 + (du/dx)2 1 is closely approximated by 1 + (du/dx)2 . The force acting on a portion of the string when it is in the conﬁguration u(x) is determined by an element F (x) of V0 . and has determined that (4. called the tension of the string.17). Thus to a ﬁrst order of approximation. we know that the length of the curve u = u(x) is given by the formula L Length = 0 1 + (du/dx)2 dx. the total 95 . Letting T denote the constant of proportionality yields L energy in string = 0 T 2 du dx 2 dx + constant. the amount of energy in the string should be proportional to L 1+ 0 1 2 du dx 2 L dx = 0 1 2 du dx 2 dx + constant. (4. this expression for potential energy is quite plausible for the following reason: We could imagine ﬁrst that the amount of energy in the string should be proportional to the amount of stretching of the string. On the other hand.instant of time.

∂x2 Since this formula holds for all inﬁnitesimal displacements ξ(x). The force acting on a tiny piece of the string of length dx is F (x)dx.work performed by F (x) is then the “sum” of the forces acting on the tiny pieces of the string. L F (x). We are imagining that the displacement ξ is inﬁnitesimally small. force = mass × acceleration. we must have F (x) = T ∂2u . and hence L ∂u ∂ξ F (x). t). ∂2u T ∂2u = . 2 ∂x ∂t If we divide by ρdx. we obtain the wave equation. ∂t2 ρ ∂x2 or 96 ∂2u ∂2u = c2 2 . Thus Newton’s law becomes T ∂2u ∂2u dx = ρdx 2 . to the function u(x. the work is the “inner product” of F and ξ. while the mass of this piece of string is just ρdx. so terms containing the square of ξ or the square of a derivative of ξ can be ignored.3 by a constant factor. ξ(x) = 0 F (x)ξ(x)dx. (Note that the inner product we use here diﬀers from the one used in Section 3. ∂x Since ξ(0) = ξ(L) = 0. we conclude that L L F (x)ξ(x)dx = F (x). in other words. Now we apply Newton’s second law. ξ(x) = T 0 0 ∂2u ξ(x)dx. ξ(x) = T 0 ∂2u ξ(x)dx − T ∂x2 ∂u ξ (L) − T ∂x ∂u ξ (0). ∂x2 for the force density per unit length.) On the other hand this work is the amount of potential energy lost when the string undergoes the displacement: L F (x). where ρ is the density of the string. ∂t2 ∂x . ξ(x) = 0 L T 2 ∂u ∂x 2 L dx − 0 L 0 T 2 ∂ξ ∂x ∂(u + ξ) ∂x 2 2 dx = −T 0 ∂u ∂ξ dx + ∂x ∂x T 2 dx. ξ(x) = −T dx. 0 ∂x ∂x Integration by parts yields L F (x).

Show that if f : R → R is any well-behaved function of one variable. just as before.4. t) can be approximated by the vector-valued function   u1 (t)  u (t)   u(t) =  2   · un−1 (t) of one variable. t) = the displacement of the string at xi at time t. Then the function u(x. ∂t ∂x (Hint: Use the “chain rule. and we expect solutions to the wave equation to behave like solutions to a mechanical system of weights and springs with a large number of degrees of freedom. u(x.1. Assume that the string has length L = 1 and set xi = i/n and ui (t) = u(xi .”) b. we could approximate this partial diﬀerential equation by a system of ordinary diﬀerential equations. Exercises: 4. Show that if g : R → R is any well-behaved function of one variable. Thus the diﬀerential operator L= d2 dx2 is approximated by the symmetric matrix n2 P. t) = f (x + ct) is a solution to the partial diﬀerential equation ∂u ∂u −c = 0. t) = g(x − ct) is a solution to the partial diﬀerential equation ∂u ∂u +c = 0. u(x. The wave equation is then approximated by the system of ordinary diﬀerential equations d2 u = c2 n2 P u. ∂t ∂x 97 .a.where c2 = T /ρ. dt2 where P is the (n − 1) × (n − 1) matrix described in the preceding section. Just as in the preceding section.

Remark: This gives a very explicit general solution to the equation for the vibrations of an inﬁnitely long string. the displacement of the string is zero at the endpoints.18) where h1 (x) and h2 (x) are given functions. 0) = f (x). u(x. t) = f (x + ct) + g(x − ct) is a solution to the diﬀerential equation ∂2u ∂2u ∂ ∂ − c2 2 = +c 2 ∂t ∂x ∂t ∂x ∂u ∂u −c ∂t ∂x = 0. Show that 4.5 The initial value problem for the vibrating string The Fourier sine series can also be used to ﬁnd the solution to an initial value problem for the vibrating string with ﬁxed endpoints at x = 0 and x = L. t) = ∂2u ∂2u − c2 2 = 0. u(x. Show that for any choice of well-behaved functions f and g. t). then the motion of the string is governed by the equation ∂2u 1 ∂ ∂u = T (x) . 2. t) = u(L. i.c. t) = 0. We formulate this problem as follows: we seek a function u(x. 2 ∂t ∂x u(x.4. ∂t2 ∂x where c is a constant. 3. 98 . the initial position and velocity of the string. 0) = h1 (x) and ∂u (x. 0) = h2 (x). the function u(x. 0) = 0. t) satisﬁes the initial conditions u(x. ∂t d. Show that if the tension and density of a string are given by variable functions T (x) and ρ(x) respectively.2.e. u(x. t) satisﬁes the boundary condition u(0. 2 ∂t ρ(x) ∂x ∂x 4. f (x + ct) + f (x − ct) 2 is a solution to the initial value problem u(x. ∂u (x. ∂t (4. deﬁned for 0 ≤ x ≤ L and t ≥ 0 such that 1. t) satisﬁes the wave equation ∂2u ∂2u = c2 2 .

Note that the wave equation itself and the boundary condition are homogeneous and linear , and therefore satisfy the principal of superposition. Once again, we ﬁnd the solution to our problem in two steps: Step I. We ﬁnd all of the solutions to the homogeneous linear conditions of the special form u(x, t) = f (x)g(t). Step II. We ﬁnd the superposition of these solution which satisﬁes the nonhomogeneous initial conditions by means of Fourier analysis. To carry out Step I, we substitute u(x, t) = f (x)g(t) into the wave equation (4.18) and obtain f (x)g (t) = c2 f (x)g(t). We separate variables, putting all the functions involving t on the left, all the functions involving x on the right: g (t) f (x) = c2 . g(t) f (x) Once again, the left-hand side of this equation does not depend on x, while the right-hand side does not depend on t, so neither side can depend upon either x or t. Therefore the two sides must equal a constant λ, and our equation becomes g (t) f (x) = = λ, 2 g(t) c f (x) which separates into two ordinary diﬀerential equations, g (t) = λ, c2 g(t) and f (x) = λ, f (x) or f (x) = λf (x). (4.20) or g (t) = λc2 g(t), (4.19)

Just as in the case of the heat equation, the homogeneous boundary condition u(0, t) = u(L, t) = 0 becomes f (0)g(t) = f (L)g(t) = 0, and assuming that g(t) is not identically zero, we obtain f (0) = f (L) = 0. Thus once again we need to ﬁnd the nontrivial solutions to the boundary value problem, d2 f (x) = 2 (f (x)) = λf (x), f (0) = 0 = f (L), dx 99

and just as before, we ﬁnd that the the only nontrivial solutions are constant multiples of f (x) = sin(nπx/L), with λ = −(nπ/L)2 , n = 1, 2, 3, . . . .

For each of these solutions, we need to ﬁnd a corresponding g(t) solving equation (4.19), g (t) = −(nπ/L)2 c2 g(t), or g (t) + (nπ/L)2 c2 g(t) = 0.

This is just the equation of simple harmonic motion, and has the general solution g(t) = a cos(ncπt/L) + b sin(ncπt/L), where a and b are constants of integration. Thus we ﬁnd that the nontrivial product solutions to the wave equation together with the homogeneous boundary condition u(0, t) = 0 = u(L, t) are constant multiples of un (x, t) = [an cos(ncπt/L) + bn sin(ncπt/L)] sin(nπx/L). The general solution to the wave equation together with this boundary condition is an arbitrary superposition of these product solutions: u(x, t) = [a1 cos(cπt/L) + b1 sin(cπt/L)] sin(πx/L) +[a2 cos(2cπt/L) + b2 sin(2cπt/L)] sin(2πx/L) + . . . . (4.21) (4.22)

The vibration of the string is a superposition of a fundamental mode which has frequency T /ρ cπ 1 c = = , L 2π 2L 2L and higher modes which have frequencies which are exact integer multiples of this frequency. Step II consists of determining the constants an and bn in (4.22) so that the initial conditions u(x, 0) = h1 (x) and ∂u (x, 0) = h2 (x) ∂t

are satisﬁed. Setting t = 0 in (4.22) yields h1 (x) = u(x, 0) = a1 sin(πx/L) + a2 sin(2πx/L) + . . . , so we see that the an ’s are the coeﬃcients in the Fourier sine series of h1 . If we diﬀerentiate equation(4.22) with respect to t, we ﬁnd that ∂u −cπ cπ (x, t) = [ a1 sin(cπt/L) + b1 cos(cπt/L)] sin(πx/L) ∂t L L 100

−2cπ cπ [a2 sin(2cπt/L) + b2 sin(2cπt/L)] cos(2πx/L) + . . . , L L and setting t = 0 yields + h2 (x) = ∂u 2cπ cπ (x, 0) = b1 sin(πx/L) + b2 sin(2πx/L) + . . . . ∂t L L

We conclude that ncπ bn = the n-th coeﬃcient in the Fourier sine series of h2 (x). L Example. Suppose that we want to ﬁnd the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes the initial-value problem: ∂2u ∂2u = , ∂t2 ∂x2 u(0, t) = u(π, t) = 0,

∂u (x, 0) = 0. ∂t In this case, the ﬁrst three coeﬃcients for the Fourier sine series of h are u(x, 0) = 5 sin x + 12 sin 2x + 6 sin 3x, a1 = 5, a2 = 12, a3 = 6,

and all the others are zero, so the solution must be u(x, t) = 5 sin x cos t + 12 sin 2x cos 2t + 6 sin 3x cos 3t. Exercises: 4.5.1 What happens to the frequency of the fundamental mode of oscillation of a vibrating string when the length of the string is doubled? When the tension on the string is doubled? When the density of the string is doubled? 4.5.2. Find the function u(x, t), deﬁned for 0 ≤ x ≤ π and t ≥ 0, which satisﬁes the following conditions: ∂2u ∂2u = , ∂t2 ∂x2 u(0, t) = u(π, t) = 0,

∂u (x, 0) = 0. ∂t You may assume that the nontrivial solutions to the eigenvalue problem u(x, 0) = sin 2x, f (x) = λf (x), are λ = −n2 , f (x) = bn sin nx, for n = 1, 2, 3, . . . , where the bn ’s are constants. 101 f (0) = 0 = f (π)

5.1}]. ∂t u(x. which satisﬁes the following conditions: ∂2u ∂2u = . 0) = 0. Find the function u(x. ∂t2 ∂x2 u(0. t) = u(π.3. ∂t2 ∂x2 u(0. t). 0) = 0. c. deﬁned for 0 ≤ x ≤ π and t ≥ 0. 0) = sin x + sin 2x. Find the function u(x.0. 2 ∂t ∂x2 u(0. ∂t u(x. ∂t u(x. ∂u (x.1 by running the Mathematica program: vibstring = Table[ 102 .6. t) = 0. 0) = sin x + 3 sin 2x − 5 sin 3x. Construct a sequence of sketches of the positions of the vibrating string at the times ti = ih. Find the function u(x. {x. t). deﬁned for 0 ≤ x ≤ π and t ≥ 0. which satisﬁes the following conditions: ∂2u ∂2u = .5. t) = 0. which satisﬁes the following conditions: ∂2u ∂2u = . t) = 0. ∂t u(x. 0) = x(π − x).5. ∂u (x. ∂2u ∂2u = . Find the ﬁrst ten terms of the solution to the initial value problem for a vibrating string. Find the ﬁrst ten coeﬃcients of the Fourier sine series for h(x) = x − x4 by running the following Mathematica program f[n ] := 2 NIntegrate[(x . 0) = 0. (For students with access to Mathematica) a. t) = u(π. b = Table[f[n].4.4. 4. where h = . 2 ∂t ∂x2 u(0. t) = u(π. t) = u(π. {n.x∧4) Sin[n Pi x].10}] b. 0) = x − x4 . 4.5. ∂u (x. ∂u (x.5.1. 0) = 0. t). t) = 0. 4. deﬁned for 0 ≤ x ≤ π and t ≥ 0.

0. the initial temperature of the wire. {t.Plot[ Sum[ b[n] Sin[n Pi x] Cos[n Pi t]. t) = f (θ)g(t) into the heat equation (4. t) satisﬁes the initial condition u(θ. 4. (4. t).0. ∂t ∂θ where c is a constant. for all θ and t. {x. Once we have found the general solution to the homogeneous conditions ∂u ∂2u = c2 2 .. 0) = h(θ).1.” from the Cell menu. so they must be dealt with ﬁrst. we will be able to ﬁnd the particular solution which satisﬁes the initial condition u(θ. In this case. 2.23) 3. t) satisﬁes the heat equation ∂u ∂2u = c2 2 . in other words. we seek a function u(θ. ∂t ∂θ u(θ + 2π. where h(θ) is a given function. t) = u(θ. periodic of period 2π. t).1}.. PlotRange -> {-1. Select the cell containing vibstring and animate the sequence of graphics by running “Animate selected graphics. Once again the heat equation itself and the periodicity condition are homogeneous and linear.1} ]. {n. g(t) f (θ) 103 . t) = u(θ. t). u(θ.1} ] d.10}].1.23) to obtain f (θ)g (t) = c2 f (θ)g(t) and separate variables: g (t) f (θ) = c2 . u(θ + 2π.6 Heat ﬂow in a circular wire The theory of Fourier series can also be used to solve the initial value problem for the heat equation in a circular wire of radius 1 which is insulated along the sides. u(θ. Thus we substitute u(θ. t) is periodic of period 2π in the variable θ. 0) = h(θ) by the theory of Fourier series. deﬁned for θ ∈ R and t ≥ 0 such that 1. u(θ.

(4. f (θ + 2π) = f (θ).11) the eigenvalue problem for the diﬀerential operator L= d2 dθ2 acting on the space V of functions which are periodic of period 2π. where a0 is a constant. As before. and a + b(θ + 2π) = a + b(θ) ⇒ b = 0. dθ2 f (θ + 2π) = f (θ). and to be consistent with our Fourier series conventions.25) or g (t) = λc2 g(t). The general solution to the diﬀerential equation is f (θ) = a + bθ.26) becomes f (θ) = 0. In this case. 2 g(t) c f (θ) where λ is a constant.26) We will call (4. t) becomes f (θ + 2π)g(t) = f (θ)g(t). or f (θ) − λf (θ) = 0 104 . In this case. (4. while the right-hand side does not depend on t. the eigenvalue problem (4. (4. c2 g(t) and f (θ) = λ. and if g(t) is not identically zero. Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the problem requires us to ﬁnd the nontrivial solutions to the problem: f (θ) = d2 (f (θ)) = λf (θ).The left-hand side of this equation does not depend on θ.24) The periodicity condition u(θ + π. so neither side can depend upon either θ or t. Thus the only solution in this case is that where f is constant. we need to consider three cases. Case 1: λ = 0. t) = u(θ. the diﬀerential equation f (θ) = λf (θ). We thus obtain two ordinary diﬀerential equations. we must have f (θ + 2π) = f (θ). we write f (θ) = a0 /2. g (t) = λ. and we can write g (t) f (θ) = = λ. f (θ) or f (θ) = λf (θ). Case 2: λ > 0.

We recognize once again our old friend. . Neither of these is consistent with the periodicity conditions f (θ + 2π) = f (θ). Thus the product solutions to the homogeneous part of the problem are u0 (θ. 2 un (θ. 105 . . . The periodicity condition f (θ + 2π) = f (θ) implies that ω = n. where n is an integer. −n2 . and f (θ) = an cos(nθ) + bn sin(nθ). . −4. . and we obtain no nontrivial solutions in this case. −9. Thus we see that the general solution to the eigenvalue problem (4. t) = a0 . . . In this case. .has the general solution f (θ) = ae( Note that a = 0 ⇒ f (θ) → ±∞ as while b = 0 ⇒ f (θ) → ±∞ as θ → −∞. . .24) g (t) = λc2 g(t). and f (θ) = a0 . t) = [an cos(nθ) + bn sin(nθ)]e−n 2 2 c t . . so we conclude that a = b = 0.26) is λ=0 or λ = −n2 where n is a positive integer and f (θ) = an cos(nθ) + bn sin(nθ). 2. √ λθ) + be−( √ λθ) . and rewrite the eigenvalue problem as f (θ) + ω 2 f (θ) = 0. which has the general solution f (θ) = a cos(ωθ) + b sin(ωθ) = A sin(ω(θ − θ0 )). 3. where c is a constant. −1. Now we need to ﬁnd the corresponding solutions to (4. where n = 1. for λ = 0. √ Case 3: λ < 0. . 2 θ → ∞. we set ω = −λ. f (θ + 2π) = f (θ). the diﬀerential equation of simple harmonic motion. which we can assume is positive. we ﬁnd that the solution is g(t) = (constant)e−n 2 2 c t . As before.

. deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0. t) = 2 2 a0 + Σ∞ [an cos(nθ) + bn sin(nθ)]e−n c t n=1 2 (4.6. . . ∂t ∂θ2 u(θ + 2π. −π bk = 1 π π h(θ) sin kθdθ. t).1. . But setting t = 0 in (4. n=1 2 so the constants a0 . To ﬁnish the solution to our problem.27) is a periodic solution of period 2π to the heat equation (4. .2. are λ=0 and λ = −n2 and f (θ) = an cos nθ + bn sin nθ. Thus the solution to our initial value problem is just (4. and f (θ) = f (θ + 2π) = f (θ) a0 . Thus u(θ. . .Now we apply the superposition principle—an arbitrary superposition of these product solutions must again be a solution. 4. t). 3. a1 . . You may assume that the nontrivial solutions to the eigenvalue problem f (θ) = λf (θ). Exercises: 4.6. t). for n = 1. are just the Fourier coeﬃcients of h(θ). ∂t ∂θ2 u(θ + 2π. we must impose the initial condition u(θ. Find the function u(θ.27) in which the constants ak and bk can be determined via the familiar formulae ak = 1 π π h(θ) cos kθdθ. t) = u(θ. Find the function u(θ. which satisﬁes the following conditions: ∂2u ∂u = . deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0. 2.27) yields a0 + Σ∞ [an cos(nθ) + bn sin(nθ)] = h(θ). . 0) = 2 + sin θ − cos 3θ. . 2 where the an ’s and bn ’s are constants. u(θ. −π Note that as t → ∞ the temperature in the circular wire approaches the constant value a0 /2. 0) = h(θ). .23). 106 . b1 . t). t) = u(θ. which satisﬁes the following conditions: ∂u ∂2u = .

putting all the functions involving t on the left. ∂t u(θ.u(θ. we separate variables. and our equation separates into two ordinary diﬀerential equations. b]. 2. As in the constant coeﬃcient case. 0) = 2 + sin θ − cos 3θ. 107 (4. which we denote by λ. we expect that there should exist a unique function u(x. ∂u (θ.3. 2 ∂t ∂θ2 u(θ + 2π. π]. deﬁned for a ≤ x ≤ b and t ≥ 0 such that 1. we substitute u(x. As we saw in Section 4. deﬁned for 0 ≤ θ ≤ 2π and t ≥ 0. t) satisﬁes the boundary condition u(a. the two sides must equal a constant. t). g (t) = λg(t).28). Just as before. which satisﬁes the following conditions: ∂2u ∂2u = . 3.28) where we make the standing assumption that ρ(x). t) satisﬁes the heat equation (4. We imagine that the bar is situated along the x-axis with its endpoints situated at x = a and x = b. 0) = h(x). all the functions involving x on the right: g (t) 1 d = g(t) ρ(x)σ(x) dx κ(x) df (x) dx 1 . 4. t) = u(θ. t) = 0. u(x. u(x. t). the density ρ(x) and the thermal conductivity κ(x) vary from point to point. t) = f (x)g(t) into (4. u(x.29) .7 Sturm-Liouville Theory* We would like to be able to analyze heat ﬂow in a bar even if the speciﬁc heat σ(x). t) satisﬁes the initial condition u(x.28) and obtain f (x)g (t) = 1 d ρ(x)σ(x) dx κ(x) df (x) g(t). Find the function u(θ. t). (4.1. 0) = 0. 4. σ(x) and κ(x) are positive. dx Once again. f (x) As usual. the initial temperature of the bar.6. t) = u(b. this leads to consideration of the partial diﬀerential equation ∂u 1 ∂ = ∂t ρ(x)σ(x) ∂x κ(x) ∂u ∂x . for θ ∈ [−π. 0) = |θ|. deﬁned for x ∈ [a. where h(x) is a given function.

. 108 . The following theorem. t) = u(b. b].30) Under the assumption that u is not identically zero. Suppose that f1 (x). Then all of the eigenvalues of L are negative real numbers. f2 (x). . Nonzero elements of the eigenspaces are called eigenfunctions. σ(x) and κ(x) are smooth functions which are positive on the interval [a. t) = n=0 4 For cn fn (x)e−λn t . Wiley. . Nevertheless. .31) has nontrivial solutions.28) together with the boundary conditions u(a. New York. ρ(x)σ(x) dx dx which acts on the space V0 of well-behaved functions f : [a. . Seventh edition. b] as a convergent sum of eigenfunctions. we need to ﬁnd the nontrivial solutions to the boundary value problem: 1 d ρ(x)σ(x) dx κ(x) df (x) dx = λf (x). . 2001. Thus to ﬁnd the nontrivial solutions to the homogeneous linear part of the problem. λ2 . . Elementary diﬀerential equations and boundary value problems. is proven in more advanced texts:4 Theorem. due to the nineteenth century mathematicians Sturm and Liouville. b] → R which vanish at the endpoints a and b. λn . see Boyce and DiPrima. σ(x) and κ(x) are complicated. t) = u(b. (4. Given an eigenvalue λ. the corresponding eigenspace is Wλ = {f ∈ V0 : f satisﬁes (4. .31) We call this the eigenvalue problem or Sturm-Liouville problem for the diﬀerential operator 1 d d L= κ(x) . . it is reassuring to know that the theory is quite parallel to the constant coeﬃcient case that we treated in previous sections. . Moreover. . and one may need to employ numerical methods to obtain approximate solutions. f (a) = 0 = f (b). (4. . every well-behaved function can be represented on [a. the eigenvalues can be arranged in a sequence 0 > λ 1 > λ2 > · · · > λ n > · · · . fn (x). t) = 0 yields f (a) = f (b) = 0. Suppose that ρ(x). . . with λn → −∞. . Finally. further discussion. it may be impossible to solve this eigenvalue problem explicitly. t) = 0 is ∞ u(x. and each eigenspace is one-dimensional.31)}.and 1 d ρ(x)σ(x) dx κ(x) df (x) dx = λf (x). If the functions ρ(x). Then the general solution to the heat equation (4. The eigenvalues of L are the constants λ for which (4. the boundary condition u(a. are eigenfunctions corresponding to the eigenvalues λ1 .

b b L(f ). dx dx where the steps represented by dots are just like the ﬁrst steps. the matrix would be symmetric.where the cn ’s are arbitrary constants. we need a generalization of the theory of Fourier series. L(g) = λj fi . Example. L(g) . The inner product should be one which makes L like a symmetric matrix. or equivalently. g ∈ V0 . ·. indeed. With respect to this inner product. g = f. for f. then cn = h. if we normalize the eigenfunctions so that they have unit length and are orthogonal to each other with respect to . It follows that if fi (x) and fj (x) are eigenfunctions corresponding to distinct eigenvalues λi and λj . fj . This identity can be veriﬁed by integration by parts. as claimed. fj = 0. fn . L(g) . · . we conclude that fi and fj are perpendicular with respect to the inner product ·. eigenfunctions corresponding to distinct eigenvalues are perpendicular. Lemma. Namely. but in reverse order. the inner product that we need to use on V0 is the one deﬁned by the formula b f. fj = L(f ). cn is just the projection of h in the fn direction. g = a ρ(x)σ(x)f (x)g(x)dx. To determine the cn ’s in terms of the initial temperature h(x). To arrange this. Since λi − λj = 0. · . then λi fi . and hence (λi − λj ) fi . so that if we thought of L as represented by a matrix. . The proof hinges on the fact that L(f ). g = a ρ(x)σ(x)L(f )(x)g(x)dx = a b d dx K(x) df (x) g(x) dx =− a K(x) df dg (x) (x)dx = · · · = f. Thus to determine the cn ’s. The key idea here is that the eigenspaces should be orthogonal with respect to an appropriate inner product. We consider the operator L=x d dx 109 x d dx . g = f. we can use exactly the same orthogonality techniques that we have used before.

x (4. 110 . the functions fm (x) = sin(m log x) and fn (x) = sin(n log x) are perpendicular with respect to the inner product deﬁned by (4. . where log denotes the natural or base e logarithm. g = 1 1 f (x)g(x)dx.33).which acts on the space V0 of functions f : [1.33) Exercises: 4. Thus if we set 1 d d = z dx e dz and hence x d d 1 d = ez z = . 2. 2 dz ˜ ˜ f (0) = 0 = f (π). . 2. dx e dz dz ˜ f (z) = f (x) = f (ez ). To solve the eigenvalue problem.7. Show by direct integration that if m = n. . . a problem which we have already solved. deﬁned by eπ f. . eπ ] → R which vanish at the endpoints of the interval [1. we make a change of variables x = ez and note that since dx = ez dz. .1. f (1) = 0 = f (eπ ).32) becomes ˜ d2 f ˜ (z) = λf (z). based upon the technique of substitution. there is a simpler approach. The nontrivial solutions are λn = −n2 . and as corresponding eigenfunctions we can take fn (x) = sin(n log x). · . The lemma implies that these eigenfunctions will be perpendicular with respect to the inner product ·. . ˜ fn (z) = sin nz. where n = 1. eπ ]. for n = 1. we need to ﬁnd the nontrivial solutions to x d dx x df (x) dx = λf (x).32) We could ﬁnd these nontrivial solutions by using the techniques we have learned for treating Cauchy-Euler equations. (4. the eigenvalue problem (4. Thus the eigenvalues for our original problem are λn = −n2 . Namely. . However.

u(x. We can imagine a violin string stretched out along the x-axis with endpoints at x = 0 and x = 1 covered with a layer of varnish which causes its mass density to vary from point to point. 0) = 3 sin(log x) + 7 sin(2 log x) − 2 sin(3 log x). t) = 0.34) It is natural to try to ﬁnd the general solution to (4.35) (4.7.2. which acts on the space V0 of functions f : [1. Substituting into (4. If the string is under constant tension T . its motion might be governed by the partial diﬀerential equation ∂2u T ∂2u = . u(1.4.7.35) by separation of variables. u(1. or = .34) and (4. t) = u(eπ . for all t ≥ 0. ∂t2 ρ(x) ∂x2 which would be subject to the Dirichlet boundary conditions u(0. which satisﬁes the following initial-value problem for a heat equation with variable coeﬃcients: ∂u ∂ =x ∂t ∂x x ∂u ∂x − 3u. t) = f (x)g(t) as usual. We could let ρ(x) = the mass density of the string at x for 0 ≤ x ≤ 1. (4. t) = 0 = u(1. t). Find the function u(x. t) = 0. letting u(x. Find the solution to the eigenvalue problem for the operator L=x d dx x d dx − 3.34) yields T g (t) T f (x) f (x)g (t) = f (x)g(t). deﬁned for 1 ≤ x ≤ eπ and t ≥ 0. deﬁned for 1 ≤ x ≤ eπ and t ≥ 0. u(x. which satisﬁes the following initial-value problem for a heat equation with variable coeﬃcients: ∂u ∂ =x ∂t ∂x x ∂u ∂x .3. 0) = 3 sin(log x) + 7 sin(2 log x) − 2 sin(3 log x). Find the function u(x.8 Numerical solutions to the eigenvalue problem* We can also apply Sturm-Liouville theory to study the motion of a string of variable mass density. eπ ]. t). eπ ] → R which vanish at the endpoints of the interval [1. t) = u(eπ . 4. 4. t).a. ρ(x) g(t) ρ(x) f (x) 111 . b.

. T f (x) = λf (x). denoted by λ. λ2 . are eigenfunctions corresponding to the eigenvalues λ1 . we let xi = i/n and ui (t) = u(xi . fn (x). the same Theorem applies. Each term in this sum represents one of the modes of oscillation of the vibrating string. ρ(x) g (t) = λg(t). where L= T d2 . every well-behaved function can be represented on [a.35) yield f (0) = 0 = f (1).35) is ∞ u(x. . f2 (x). the displacement at time t is approximated by   u1 (t)  u (t)  . the eigenvalues can be arranged in a sequence 0 > λ1 > λ2 > · · · > λ n > · · · . Then the general solution to (4. In constrast to the case of constant density. it is usually not possible to ﬁnd simple explicit eigenfunctions when the density varies. The partial derivative ∂2u ∂t2 is approximated by d2 u . . f (0) = 0 = f (1). . If f1 (x). b] as a convergent sum of eigenfunctions. where the an ’s and bn ’s are arbitrary constants. The simplest numerical method is the one outlined in §4. . The Dirichlet boundary conditions (4. Thus the eigenvalues of L are negative real numbers and each eigenspace is one-dimensional. . . λn .3. u(t) =  2   · un−1 (t) a vector-valued function of one variable. Finally.34) and (4. t) = the displacement at xi at time t. . . dt2 112 . For 0 ≤ i ≤ n. with λn → −∞. . . It is therefore usually necessary to use numerical methods. Since u0 (t) = 0 = un (t) by the boundary conditions. . . .The two sides must equal a constant. . Thus f must satisfy the eigenvalue problem L(f ) = λf. t) = n=0 fn (x) an cos( −λn t) + bn sin( −λn t) . ρ(x) dx2 Although the names of the functions appearing in L are a little diﬀerent than those used in the previous section. and the partial diﬀerential equation separates into two ordinary diﬀerential equations. Moreover.

04 -0. 0 0 T /ρ(x3 ) · · · · Q=     · · · ··· · 0 0 · · · · T /ρ(xn−1 ) Putting all this together. while the eigenfunctions representing the lowest frequency modes of oscillation are approximated eigenvectors corresponding to the lowest eigenvalues of A. The eigenvalues of low absolute value are approximated by the eigenvalues of A. and as we saw in §4.1: Shape of the lowest mode when ρ = 1/(x + .   Finally.1).02 -0.1 -0.08 -0. dt2 where A = n2 QP.1)? To answer this question. 0 0 · · −2    .3. 113 . For example. the coeﬃcient (T /ρ(x)) can be represented by the diagonal matrix   0 0 ··· 0 T /ρ(x1 )   0 T /ρ(x2 ) 0 ··· 0   .06 -0.20 -0. we could ask the question: What is the shape of the lowest mode of oscillation in the case where ρ(x) = 1/(x+. the partial derivative ∂2u ∂x2 where    P =   is approximated by −2 1 0 · 0 1 −2 1 · 0 0 1 −2 · · ··· ··· ··· ··· ··· n2 P u.14 40 60 80 100 Figure 4. we ﬁnd that our wave equation with variable mass density is approximated by a second order homogeneous linear system of ordinary diﬀerential equations d2 u = Au.12 -0.

p := m .we could utilize the following Mathematica program (which is quite similar to the one presented in Exercise 2.1).2): n := 100. { i.1.n-1 } .{ j. rho[x ] := 1/(x + . a := n∧2 q. { i.p. as shown in Figure 4. our numerical approximation to the lowest mode tilts somewhat to the left. m := Table[Max[2-Abs[i-j]. Note that instead of approximating a sine curve.n-1 } ]].n-1 } ].0].4 IdentityMatrix[n-1]. q := DiagonalMatrix[Table[(1/rho[i/n]). eigenvec = Eigenvectors[N[a]].1. 114 . ListPlot[eigenvec[[n-1]]] If we run this program we obtain a graph of the shape of the lowest mode.5.

This ensures that the principle of superposition will hold. Techniques developed for studying these equations can often be applied to closely related equations. ∂2u = c2 ∂t2 ∂2u ∂2u ∂2u + 2 + 2 ∂x2 ∂y ∂z . In the ﬁrst few sections of this chapter. we will derive these partial diﬀerential equations in several physical contexts. We will begin by using the divergence 115 . ∂2u ∂2u ∂2u + 2 + 2 ∂x2 ∂y ∂z .Chapter 5 PDE’s in Higher Dimensions 5. that is each term contains u or one of its derivatives to the ﬁrst power. The principle of superposition is essential if we want to apply separation of variables and Fourier analysis techniques. u1 and u2 solutions ⇒ c1 u1 + c2 u2 is a solution. for any choice of constants c1 and c2 . ∂x2 ∂y ∂z the heat equation ∂u = c2 ∂t and the wave equation. where c is a nonzero constant. the three most important linear partial diﬀerential equations are Laplace’s equation ∂2u ∂2u ∂2u + 2 + 2 = 0. Each of these three equations is homogeneous linear .1 The three most important linear partial differential equations In higher dimensions.

z. together with suﬃcient theory to understand the qualitative behaviour of the solutions. or water waves in a linear approximation. z)u(x. t) = temperature at (x. the rate at which heat leaves D equals minus the rate of change of heat within D. z) at time t. We then present two derivations of the wave equation. then by conservation of energy. y. Slightly more complicated cases require the technique of “separation of variables” together with Fourier analysis. where f and g are arbitrary well-behaved functions of a single variable. y. t)dxdydz. y. it is possible to ﬁnd explicit solutions to these partial differential equations under the simplest boundary conditions. is u(x. as we studied before. It is remarkable that the principles developed to solve the three basic linear partial diﬀerential equations can be applied in so many contexts. y. often with variable coeﬃcients. y.theorem to derive the heat equation. Separation of variables reduces these partial diﬀerential equations to linear ordinary diﬀerential equations. z) is the density of the medium at (x. we will see that it is necessary to solve Bessel’s equation. ∂t (5. the general solution to the one-dimensional wave equation ∂2u ∂2u = c2 2 ∂t2 ∂x for the vibrations of an inﬁnitely long string. which is − d dt ρσudxdydz = − D D ρσ ∂u dxdydz. t) = f (x + ct) + g(x − ct). z)-space is given by the formula Heat within D = D ρ(x. Exactly the same wave equation also describes electromagnetic waves. which in turn reduces in the steady-state case to Laplace’s equation. For example. z). z)σ(x. 116 . In a few cases. to ﬁnd the explicit solution to the heat equation in a circular room. z) is the speciﬁc heat at the point (x. For example. y. y. gravitational waves. If no heat is being created or destroyed within D. we consider our ﬁrst example. y. In the rest of this section. the equation governing heat ﬂow through a region of (x. z. y. The most complicated cases cannot be solved by separation of variables. then the heat within a given region D in (x. z) and ρ(x. Let u(x. and one must resort to numerical methods. one for vibrating membranes and one for sound waves. y. under the assumption that no heat is being created or destroyed within the region. y. z)-space. If σ(x.1) by diﬀerentiating under the integral sign.

σ(x. F(x. z) and κ(x. z) ∂u (x. ∂t ρσ ∂x2 ∂y ∂z If we wait long enough.2) From formulae (5. y. It follows from the divergence theorem that Rate at which heat leaves D = − D ∇ · (κ∇u)dxdydz. t).On the other hand. y. t) which points in the direction of greatest decrease of temperature. y. z. y. heat ﬂow can be represented by a vector ﬁeld F(x. where ∂D is the surface bounding D. y. y.1) and (5. (5. z). The total rate at which heat leaves D is given by the ﬂux integral − ∂D (κ∇u) · NdA. z)σ(x. z) must satisfy Laplace’s equation ∂2u ∂2u ∂2u + 2 + 2 = 0. y. z) is the so-called thermal conductivity of the medium at (x. z. t) = −κ(x. so the integrands on the two sides must be equal: ρ(x. y. Thus the rate at which heat leaves the region D through a small region in its boundary of area dA is −(κ∇u) · NdA. z) are constants. we conclude that ρσ D ∂u dxdydz = ∂t ∇ · (κ∇u)dxdydz. y. z. ∂t ρ(x. i. where N is the unit normal which points out of D. z)σ(x. y. y. t). ∂x2 ∂y ∂z 117 . D This equation is true for all choices of the region D. y. y. t) = ∇ · (κ∇u)(x. ∂t Thus we ﬁnally obtain the heat equation ∂u 1 = ∇ · (κ(x. y. y. its properties are the same at every point.e. z.2). y. z) In the special case where the region D is homogeneous. z). y. z)(∇u)) . and the heat equation becomes ∂u κ ∂2u ∂2u ∂2u = + 2 + 2 . the “steady-state” temperature u(x. where κ(x. so that the temperature is no longer changing. z)(∇u)(x. ρ(x. z.

2 The Dirichlet problem The reader will recall that the space of solutions to a homogeneous linear second order ordinary diﬀerential equation.If the temperature is independent of z. Show that in this case the equation governing heat ﬂow is ρ(x. y) = u(x. z)u(x. y. ∂x2 ∂y ∂z ∂2u ∂2u ∂2u + 2 + 2 + u = 0. y. ∂x2 ∂y Exercises: 5. z) + ∇ · (κ(x.1. ∂x2 ∂y ∂z ∂2u ∂2u ∂2u + 2 + 2 = ex u. ∂t ∂2u ∂2u ∂2u + 2 + 2 = ex . z) must satisfy the two-dimensional Laplace equation ∂2u ∂2u + 2 = 0. c. 5. z) ∂u = λ(x. y. y. 5.1. ∂x2 ∂y ∂z ∂2u ∂2u ∂2u + 2 + 2 + u2 = 0. y. y. such as d2 u du + qu = 0 +p dt2 dt is two-dimensional. y. z)u(x. t) + ν(x. ∂x2 ∂y ∂z d. the space of solutions to Laplace’s partial diﬀerential equation ∂2u ∂2u + 2 =0 ∂x2 ∂y 118 (5. Suppose that a chemical reaction creates heat at the rate λ(x.1. For which of the following diﬀerential equations is it true that the superposition principle holds? a. By contrast. z)(∇u)) . z). the function u(x. per unit volume.2. z. z)σ(x. t) + ν(x. ∂2u ∂2u ∂2u + 2 + 2 = 0. Explain your answers. y. y. y.3) . e. a particular solution being determined by two constants. ∂x2 ∂y ∂z b. z.

the function u(x. For example. for example. Our goal here is to ﬁnd the explicit solutions in the case where the region D is suﬃciently simple. ∂x ∂u = −6xy. say by setting up heaters and refrigerators controlled by thermostats along the boundary. This is proven mathematically for many choices of boundary in more advanced texts on complex variables and partial diﬀerential equations. we might expect that the temperature inside the room would be uniquely determined. y)-plane. Thus to pick out a particular solution to Laplace’s equation. A solution to Laplace’s equation is called a harmonic function. ∂y 2 ∂2u ∂2u + 2 = 6x − 6x = 0.is inﬁnite-dimensional. that D = {(x. are solutions to Laplace’s equation. and let φ : ∂D → R be a continuous function. 119 . y) = ex sin y Similarly. y) = 7. If we specify the temperature on the boundary of the region. y) = φ(x. The mathematical proof that a unique solution exists provides evidence that the mathematical model we have constructed for heat ﬂow may in fact be valid. y)-plane which is bounded by a curve ∂D. The Dirichlet Problem for Laplace’s Equation. To see what boundary conditions are natural to impose. Let D be a bounded region in the (x. for (x. It is not diﬃcult to construct inﬁnitely many linearly independent harmonic functions of two variables. Our physical intuition suggests that the Dirichlet problem will always have a unique solution. y) ∈ ∂D. u(x. because ∂u = 3x2 − 3y 2 . we need to think of a physical problem which leads to Laplace’s equation. Suppose that u(x. ∂y and hence ∂2u = 6x. Suppose. and u(x. y) represents the steady-state distribution of temperature throughout a uniform slab in the shape of a region D in the (x. y) = x3 − 3xy 2 is a solution to Laplace’s equation. ∂x2 ∂y u(x. y) ∈ R 2 : 0 ≤ x ≤ a. Specifying the temperature at each point of the boundary imposes inﬁnitely many constraints on the harmonic function which realizes the steady-state temperature within D. 0 ≤ y ≤ b}. y) = x4 − 6x2 y 2 + y 4 . ∂x2 ∂2u = −6x. Find a harmonic function u : D → R such that u(x. we need boundary conditions which will impose inﬁnitely many constraints. y). We need inﬁnitely many heaters and refrigerators because there are inﬁnitely many points on the boundary.

we seek a function u(x. 120 . y) = φ(x. where f (x) is a given continuous function which vanishes when x = 0 and x = a. b) = f (x). Our method for solving the Dirichlet problem consists of two steps: Step I. We ﬁnd all of the solutions to Laplace’s equation together with the homogeneous boundary conditions which are of the special form u(x. Hence neither side can depend upon either x or y. 3. Step II. all the functions involving y on the right: X (x) Y (y) =− . y) satisﬁes the nonhomogeneous boundary condition u(x. for any choice of constants c1 and c2 . y) = u(x. Note that the Laplace equation itself and the homogeneous boundary conditions satisfy the superposition principle—this means that if u1 and u2 satisfy these conditions. while φ(x. deﬁned for 0 ≤ x ≤ a and 0 ≤ y ≤ b. In this case. y) = X(x)Y (y). so that φ(0. while the right-hand side does not depend on x. Next we separate variables.Suppose.3) and obtain X (x)Y (y) + X(x)Y (y) = 0. moreover that the function φ : ∂D → R vanishes on three sides of ∂D. y) satisﬁes Laplace’s equation ∂2u ∂2u + 2 = 0. y) = X(x)Y (y) into Laplace’s equation (5. we substitute u(x.4) 2. ∂x2 ∂y (5. By the superposition principle. u(x. To carry out Step I. y) = u(a. putting all the functions involving x on the left. u(x. X(x) Y (y) The left-hand side of this equation does not depend on y. We ﬁnd the particular solution which satisﬁes the nonhomogeneous boundary condition by Fourier analysis. y). such that 1. where f (x) is a given function. y) = φ(a. so does c1 u1 + c2 u2 . y) satisﬁes the homogeneous boundary conditions u(0. b) = f (x). 0) = 0. u(x. 0) = 0. an arbitrary linear superposition of these solutions will still be a solution.

and Y (y) = −λY (y). dx2 d2 dx2 X(0) = 0 = X(a).6). . as before. which we denote by λ. The homogeneous boundary condition u(0. .5) which we recognize as the eigenvalue problem for the diﬀerential operator L= acting on the space V0 of functions which vanish at 0 and a. and the boundary condition Y (0) = 0 implies that A = 0. y) = sin(nπx/a) sinh(nπy/a). where A and B are constants of integration.In other words. We have seen before that the only nontrivial solutions to equation (5. For each of these solutions. Thus we need to ﬁnd the nontrivial solutions to a boundary value problem for an ordinary diﬀerential equation: X (x) = d2 (X(x)) = λX(x). Our equation now becomes X (x) Y (y) =− = λ. 121 . X (x) = λX(x). X(0) = X(a) = 0. The diﬀerential equation has the general solution Y (y) = A cosh(nπy/a) + B sinh(nπy/a). and call the separating constant. . where λ = −(nπ/a)2 . Thus we ﬁnd that the nontrivial product solutions to Laplace’s equation together with the homogeneous boundary conditions are constant multiples of un (x. together with the boundary condition Y (0) = 0. the two sides must equal a constant. 2. Y (y) = −λY (y). . If Y (y) is not identically zero.6) (5. n = 1. with λ = −(nπ/a)2 . we need to ﬁnd a corresponding Y (y) solving equation (5.5) are constant multiples of X(x) = sin(nπx/a). y) = 0 imply that X(0)Y (y) = X(a)Y (y) = 0. X(x) Y (y) which separates into two ordinary diﬀerential equations. y) = u(a. (5. 3.

. sinh(2π) Bk = 0 for k = 3. y). Which of the following functions are harmonic? a. f (x.2. . then we must have f (x) = B1 sin(x) sinh(π) + B2 sin(2x) sinh(2π). We see that Bk sinh(kπb/a) is the k-th coeﬃcient in the Fourier sine series for f (x). u(0. y) = x2 + y 2 .2.7) yields f (x) = B1 sin(πx/a) sinh(πb/a) + B2 sin(2πx/a) sinh(2πb/a)+ . π) = sin x − 2 sin 2x + 3 sin 3x. . 122 . 0 ≤ x ≤ π.7) To carry out Step II. c. . 5.1. y) = B1 sin(πx/a) sinh(πy/a) + B2 sin(2πx/a) sinh(2πy/a) + . . d. . ∂x2 ∂y u(x. . y) = x2 − y 2 . sinh(π) B2 = 7 . such that ∂2u ∂2u + 2 = 0. . and hence B1 = 3 . . (5. sinh(π) sinh(2π) Exercises: 5. + Bk sin(2πk/a) sinh(kπb/a) + . Solve the following Dirichlet problem for Laplace’s equation in a square region: Find u(x. b. . if a = b = π. u(x. B2 . f (x.7) so that u(x.The general solution to Laplace’s equation with these boundary conditions is a general superposition of these product solutions: u(x. . y) = u(π. . For example. .2. y) = 3 7 sin x sinh y + sin 2x sinh 2y. y) = x3 − 3xy 2 . 0) = 0. f (x. which occur in (5. and f (x) = 3 sin x + 7 sin 2x. a. Thus the solution to Dirichlet’s problem in this case is u(x. y) = ex cos y. y) = 0. f (x. we need to determine the constants B1 . b) = f (x). 0 ≤ y ≤ π. . . Substitution of y = b into (5.

π). such that ∂2u ∂2u + 2 = 0. y) = sin 2y+3 sin 4y. y) = 0. y)-plane which is bounded by a piecewise smooth curve ∂D. ∂x2 ∂y u(π. Find a function u(x. ﬁnd the solution to the Dirichlet problem: Find u(x. By adding the solutions to parts a and c together. π) = sin x−2 sin 2x+3 sin 3x. 0 ≤ x ≤ π. u(0. y) = sin 2y + 3 sin 4y. u(x. Solve the following Dirichlet problem for Laplace’s equation in the same square region: Find u(x. 0) = 0. 0 ≤ x ≤ π. such that ∂2u ∂2u + 2 = 0. y) = 0. u(x. y).3 Initial value problems for heat equations The physical interpretation behind the heat equation suggests that the following initial value problem should have a unique solution: Let D be a bounded region in the (x. 0 ≤ y ≤ π.1: Graph of u(x. c. 5. b. u(x. 0 ≤ y ≤ π. ∂x2 ∂y u(π. and let h : D → R be a continuous function which vanishes on ∂D. y).10 5 0 -5 -10 0 1 2 3 Figure 5. y. u(0. t) such that 123 . 0) = 0 = u(x. y) = 3 sinh(π) 3 2 1 0 7 sinh(2π) sin x sinh y + sin 2x sinh 2y.

we write u(x. or t. y)g(t). Hence neither side can depend on x. for (x. In this section. 0 ≤ y ≤ b}. y. t) = u(x. t) = u(x. 0) = h(x. 1 1 g (t) = 2 g(t) c f (x. separation of variables is done in two stages. y. u satisﬁes the initial condition u(x. and substitute into (5. y). and then use Fourier analysis to satisfy the last condition. The left-hand side of this equation does not depend on x or y while the right hand side does not depend on t. so our strategy is to treat them ﬁrst by separation of variables. y)g(t). y.8) 2. we obtain 1 c2 g(t) g (t) = λ = 1 f (x. y. (5. y)g (t) = c2 Then we divide by c2 f (x. so both sides must be constant. y. we consider the special case where D = {(x. The ﬁrst two of these conditions are homogeneous and linear. u satisﬁes the heat equation ∂u = c2 ∂t ∂2u ∂2u + 2 ∂x2 ∂y . (5. 3. 0. t) = u(a. If we let λ denote the constant. This separates into an ordinary diﬀerential equation g (t) = c2 λg(t). y) ∂2f ∂2f + 2 ∂x2 ∂y . y) ∈ R 2 : 0 ≤ x ≤ a. t) = 0. u satisﬁes the “Dirichlet boundary condition” u(x. ∂2f ∂2f + 2 ∂x2 ∂y g(t). First. and a partial diﬀerential equation ∂2f ∂2f + 2 2 ∂x ∂y 124 = λf.8) to obtain f (x. y) ∈ ∂D. so that the Dirichlet boundary condition becomes u(0. y.10) (5. b.1. y) ∂2f ∂2f + 2 2 ∂x ∂y .9) . t) = f (x. t) = 0. In this case.

n = 1. For any given choice of m and n. the corresponding solution to (5. . while the right-hand side depends only on y. y. 0) = f (x. 125 . . The corresponding solutions of the Helmholtz equation are fmn (x. which yields X (x) Y (y) + = λ.10) to obtain X (x)Y (y) + X(x)Y (y) = λX(x)Y (y). and Y (y) = sin(nπy/b). X(x) Y (y) Hence the Helmholtz equation divides into two ordinary diﬀerential equations X (x) = µX(x).n (x. X(x) Y (y) The left-hand side depends only on x. 3. 2. . The “Dirichlet boundary conditions” now become conditions on X(x) and Y (y): X(0) = X(a) = 0. X (x) Y (y) =µ=λ− . m = 1. . b) = 0. . Hence for each choice of m and n. . 2. In the second stage of separation of variables. y) = f (a. with λmn = −(mπ/a)2 − (nπ/b)2 . so both sides must be constant. with µ = −(mπ/a)2 . . 3. y) = X(x)Y (y) and substitute into (5. The only nontrivial solutions are X(x) = sin(mπx/a).9) is g(t) = e−c 2 λmn t = e−c 2 ((mπ/a)2 +(nπ/b)2 )t . X(x) Y (y) or X (x) Y (y) =λ− . t) = sin(mπx/a) sin(nπy/b)e−c 2 ((mπ/a)2 +(nπ/b)2 )t . Y (y) = νY (y). we obtain a product solution to the heat equation with Dirichlet boundary condition: um. where µ + ν = λ. we set f (x. y) = f (x. Y (0) = Y (b) = 0. y) = sin(mπx/a) sin(nπy/b). with ν = −(nπ/b)2 .called the Helmholtz equation. . The Dirichlet boundary condition yields f (0.

that c = 1.n=1 bmn sin(mπx/a) sin(nπy/b)e−c 2 ((mπ/a)2 +(nπ/b)2 )t .12) expressing the fact that the bmn ’s are the coeﬃcients of what is called the double Fourier sine series of h. y) sin(pπx/a) sin(qπy/b)dxdy. for example. 126 .11).13) Suppose. y. 0 The expression within brackets is one if q = n and otherwise zero. y) sin(pπx/a)dx = 0 n=1 bpn sin(nπy/b). we multiply both sides of (5. t) = m. To determine it. and integrate with respect to x to obtain 2 a ∞ a h(x.The general solution to the heat equation with Dirichlet boundary conditions is an arbitrary superposition of these product solutions. y) = m. (5. y) sin(pπx/a)dx 0 = m.n=1 bmn sin(mπx/a) sin(nπy/b). the bmn ’s can be determined by an explicit formula. and integrate with respect to y to obtain 22 ab a 0 0 ∞ b h(x. y). (5. ∞ u(x. so 2 a a ∞ h(x. The result is ∞ h(x. As in the case of ordinary Fourier sine series. where q is a positive integer. (5. y) sin(pπx/a) sin(qπy/b)dxdy 2 b b = n=1 bpn sin(nπy/b) sin(qπy/b)dy .11) To ﬁnd the constants bmn appearing in (5. a = b = π and h(x. we need to apply the initial condition u(x. Next we multiply by (2/b) sin(qπy/b).12) by (2/a) sin(pπx/a). 0) = h(x.n=1 bmn 2 a a sin(pπx/a) sin(mπx/a)dx sin(nπy/b). y. 0 The expression within brackets is one if p = m and otherwise zero. so we ﬁnally obtain bpq = 22 ab a 0 0 b h(x. y) = sin x sin y + 3 sin 2x sin y + 7 sin 3x sin 2y. where p is a positive integer.

y) = p(x)q(y). π 2 m2 n 2 Thus in this case. y) = and hence u(x. we do not need to carry out the integration indicated in (5. q(y) = y.13) because comparison with (5. y.In this case. t) = sin x sin ye−2t + 3 sin 2x sin ye−5t + 7 sin 3x sin 2ye−13t . p(x)q(y) sin mxdx sin nydy = = p(x) sin mxdx 0 q(y) sin nydy (π − x) sin mxdx π/2 π x sin mxdx + y sin nydy + 0 π/2 (π − y) sin nydy . for π/2 ≤ x ≤ π. b21 = 3. we see that h(x.3 to yield bmn = 2 π 2 2 sin(mπ/2) m2 = 2 sin(nπ/2) n2 16 1 1 sin(mπ/2) sin(nπ/2) . π − y. b32 = 7. for 0 ≤ x ≤ π/2.n=1 1 1 sin(mπ/2) sin(nπ/2) sin mx sin my. The integration can be carried out just like we did in Section 3. t) = 16 π2 ∞ m. for π/2 ≤ y ≤ π. where p(x) = In this case. for 0 ≤ y ≤ π/2. Thus the solution to the initial value problem in this case is u(x. Here is another example.12) shows that b11 = 1. bmn = 2 π 2 π 2 π 2 0 π/2 2 0 2 0 π/2 π π π 0 π π x. m2 n 2 16 π2 ∞ m. and all the other bmn ’s must be zero.n=1 2 2 1 1 sin(mπ/2) sin(nπ/2) sin mx sin mye−(m +n )t . π − x. m2 n 2 127 . y. Suppose that a = b = π and h(x.

2t + 1 ∂t ∂x2 ∂y u(x. such that ∂u ∂2u ∂2u + 2. y. 0. t) = u(0. where 0 ≤ x ≤ π.3. . 0) = 2 sin x sin y + 5 sin 2x sin y. Solve the following initial value problem in a square region: Find u(x. 0. where 0 ≤ x ≤ π.3. y) = f (π. t). 0) = 2(sin x)y(π − y). y. t) = u(π. u(x. y. π. 0) = 2 sin x sin y + 3 sin 2x sin y. y. t) = 0. π. 3. = ∂t ∂x2 ∂y u(x. 0 ≤ y ≤ π and t ≥ 0. . = ∂t ∂x2 ∂y u(x. such that ∂u ∂2u ∂2u + 2. t). . . y. t) = u(x. y. t) = 0. t) = u(π. t) = u(π. y. . π. t) = u(x. t) = u(0. t) = u(x. t) = u(0. t) = 0. and n = 1. 0) = f (x. u(x. y. y) + 2 (x. π. . y) = bmn sin mx sin ny. u(x. 128 . . 5. f (x. 2 ∂x ∂y are of the form λ = −m2 − n2 .4. y) = λf (x. t) = 0. 5.3. 2. f (x. y. y). t) = sin x − 2 sin 2x + 3 sin 3x. 3. y) = 0. y. where bmn is a constant.Exercises: 5.2. t) = u(0.1. 0 ≤ y ≤ π and t ≥ 0. where 0 ≤ x ≤ π. You may assume that the nontrivial solutions to the eigenvalue problem ∂2f ∂2f (x. u(x. 0 ≤ y ≤ π and t ≥ 0. 5. t) = u(π.3. y. such that 1 ∂u ∂2u ∂2u = + 2. y. 0. y. t). for m = 1. Solve the following initial value problem for the heat equation in a square region: Find u(x. y. 0. π) = f (0. Find the general solution to the heat equation ∂u ∂2u ∂2u = + 2 ∂t ∂x2 ∂y subject to the boundary conditions u(x.3. 2. Solve the following initial value problem for the heat equation in a square region: Find u(x.

y)ρ ∂2u (x. y) and its derivatives are very small.5. y) + η(x. ∂t Suppose the force displaces the membrane from a given position u(x. where η(x. while sound waves are described by a three-dimensional wave equation. y)ρ ∂2u (x. ∂t2 (5. y)dxdy. y)dxdy. y)-plane. In fact. y) at time t. then by Newton’s second law. y)k = η(x. the force F acting on a small rectangular piece of the membrane located at (x. this stretching is approximated by the integral Potential energy = T 2 ∂u ∂x 2 + D ∂u ∂y 2 dxdy. ∂t2 Integrating over the membrane yields an expression for the total work performed when the membrane moves through the displacement η: Work = D η(x. The vibrating membrane. y)dxdyk. y). called the tension in the membrane. Just as in the case of the vibrating string. Then the total work performed by the force F will be F · η(x. the potential energy stored in the membrane is proportional to the extent to which the membrane is stretched. y) with sides of length dx and dy is given by the expression ∂2u F = ρ 2 (x. If ρ denotes the density of the membrane (assumed to be constant). y) to a new position u(x. Suppose. In this section we will show that the motion of a vibrating membrane is described by the two-dimensional wave equation. where T is a constant. t) denote the height of the point with coordinates (x. and let u(x.4 Two derivations of the wave equation We have already seen how the one-dimensional wave equation describes the motion of a vibrating string. Replacing u by u+η in this integral yields New potential energy = T 2 ∂u ∂x 2 + D ∂u ∂y 2 dxdy 129 . y. moreover. that a point on the membrane can move only in the vertical direction.14) On the other hand. Suppose that a homogeneous membrane is fastened down along the boundary of a region D in the (x. we will see that sound waves arise from “linearization” of the nonlinear equations of ﬂuid mechanics.

z. y)dxdy.+T D ∂u ∂x ∂η ∂x ∂η ∂x 2 + + ∂u ∂y ∂η ∂y 2 ∂η ∂y dxdy. 1980. dxdy + T 2 D If we neglect the last term in this expression (which is justiﬁed if η and its derivatives are assumed to be small). Next. we ﬁnd that New potential energy − Old potential energy = D T ∇u · ∇ηdxdy. we will describe Euler’s equations for a perfect gas in (x. ∂t2 or equivalently ∂2u = c2 ∂t2 This is just the wave equation. 1 For a complete derivation of these equations. and hence Change in potential = − D η(x. v(x. It follows from the divergence theorem in the plane and the fact that η vanishes on the boundary ∂D that T ∇u · ∇ηdxdy + D D T η∇ · ∇udxdy = ∂D T η∇u · Nds = 0. y)dxdy. third edition. y)dxdy = ∂t2 η(x. 130 . z)-space.15) that η(x. 1993. ρ where c2 = ∂2u ∂2u + 2 ∂x2 ∂y .14) and (5. it follows that ρ which simpliﬁes to ∂2u = c2 ∇ · ∇u. see Chapter 9 of Alexander Fetter and John Walecka. y. Springer. y. A mathematical introduction to ﬂuid mechanics. New York. so it follows from (5. McGraw-Hill. y)T (∇ · ∇u)(x.1 Euler’s equations are expressed in terms of the quantities. y. ∂2u = T ∇ · ∇u.15) The work performed must be minus the change in potential energy. The equations of a perfect gas and sound waves. y)ρ D ∂2u (x. ∂t2 T . y)T (∇ · ∇u)(x. or Chapter 1 of Alexandre Chorin and Jerrold Marsden. (5. z) at time t). t) = (velocity of the gas at (x. D Since this equation holds for all choices of η. Theoretical mechanics of particles and continua.

Then the rate of change of the mass of ﬂuid within D is given by two expressions. z(t). z. z) at time t). z. z) at time t). y. z. It follows from the divergence theorem that the second of these expressions is − D ∇ · F(x.16) represents the rate at which the ﬂuid is ﬂowing across S in the direction of N. an assumption which is not unreasonable in the case of a perfect gas. The ﬁrst of the Euler equations is the equation of continuity. y.17) 131 . ∂ρ dxdydz = − ∂t and hence D ∇ · Fdxdydz. t)dxdydz ∂t and − S F · NdA. ∂t which is just the equation of continuity. y. ∂ρ (x. we conclude that the integrands must be equal. ∂ρ + ∇ · (ρv) = 0.ρ(x. We make an assumption that the only force acting on a ﬂuid element is due to the pressure. y. (mass density)(acceleration) = (force density). y(t). ∂t To derive this equation. y. z)-space. ∂ρ = −∇ · F = −∇ · (ρv). t). so that the surface integral F · NdA S (5. t)dxdydz. D Since this equation must hold for every region D in (x. z. D which must be equal. The second of the Euler equations is simply Newton’s second law of motion. In this case. it turns out that the pressure is deﬁned in such a way that the force acting on a ﬂuid element is minus the gradient of pressure: Force = −∇p(x(t). We assume that no ﬂuid is being created or destroyed. y. (5. we represent the ﬂuid ﬂow by the vector ﬁeld F = ρv. t) = (density at (x. p(x. t) = (pressure at (x. y.

which relates pressure and density.16). z(t).) The Euler equations (5. where c2 is a new constant. we need an equation of state. v = v . (5. To ﬁnish the Euler equations. 132 . Substitution into Euler’s equations yields ∂ρ + ρ0 ∇ · (v ) = 0. 0 Linearizing Euler’s equations near this explicit solution gives rise to the linear diﬀerential equation which governs propagation of sound waves. (5.19) are nonlinear. dv ∂v dx ∂v dy ∂v dz ∂v dt = + + + . ∂t ρ (5. ∂t ∂v 1 = − ∇p .The familiar formula Force = Mass × Acceleration then yields ρ d (v(x(t). Let us write ρ = ρ0 + ρ . ρ = ρ0 . (An ideal monatomic gas has this equation of state with γ = 5/3. p and v are so small that their squares can be ignored. However. v = 0. t)) = −∇p(x(t). The equation of state could be determined by experiment. dt Using the chain rule. one explicit solution is the case where the ﬂuid is motionless. and (5. the simplest equation of state being p = a2 ργ . p = p0 . It follows from these three equations that ∂2ρ ∂v = −ρ0 ∇ · ∂t2 ∂t = ∇ · ∇p = c2 ∇ · ∇ρ . t). y(t).18). where ρ . z(t). dt ∂x dt ∂y dt ∂z dt ∂t dt we can rewrite this equation as ∂v 1 + (v · ∇)v = − ∇p. where ρ0 and p0 satisfy p0 = a2 ργ . ∂t ρ0 and p = [a2 γ(ρ0 )(γ−1) ]ρ = c2 ρ . y(t).18) Note that this equation is nonlinear because of the term (v · ∇)v. and hence quite diﬃcult to solve.19) where a2 and γ are constants. p = p 0 + p .

the density is constant. then the motion of the string is governed by the equation ∂2u 1 = ∇ · (T (x. singled out by the Clay Mathematics Institute as central problems for mathematics at the turn of the century. so easily expressed. Show that if the tension and density of a membrane are given by variable functions T (x. ∂t ρ It is remarkable that these equations. y)∇u) . In the case of incompressible ﬂuids. t) − ∇p(x(t). the Navier-Stokes equations form the basis for one of the seven Millenium Prize Problems.org/millennium/ Exercise: 5. ∂v 1 + (v · ∇)v = ν∆v − ∇p. y) respectively. t). one adds an additional term to the expression (5. (5. you may be able to win one million dollars. z. Indeed. y. (5.20) If the sound wave ρ is independent of z. you can consult the web address: http://www. so no equation of state is assumed. y(t). z(t).4. y) 133 . called the viscosity of the ﬂuid. To allow for viscosity. exactly the same equation that we obtained for the vibrating membrane. To ﬁnd more details on the prize oﬀered for a solution. The Euler equations for a perfect gas and the closely related Navier-Stokes equations for an incompressible ﬂuid such as water form basic models for ﬂuid mechanics. If you can show that under reasonable initial conditions. are so diﬃcult to solve.claymath.1.Thus ρ must satisfy the three-dimensional wave equation ∂2ρ = c2 ∇ · ∇ρ = c2 ∂t2 ∂2ρ ∂2ρ ∂2ρ + + 2 2 ∂x ∂y ∂z 2 . Here the Laplace operator is applied componentwise and ν is a constant.17) for the force acting on a ﬂuid element: Force = ν(∆v)(x.20) reduces to ∂2ρ = c2 ∂t2 ∂2ρ ∂2ρ + 2 ∂x ∂y 2 . ∂t2 ρ(x. Remark: The notion of linearization is extremely powerful because it enables us to derive information on the behavior of solutions to the nonlinear Euler equations. y) and ρ(x. the Navier-Stokes equations possess a unique well-behaved solution. The equations used by Navier and Stokes to model an incompressible viscous ﬂuid (with ρ constant) are then ∇ · v = 0. which are extremely diﬃcult to solve except for under very special circumstances.

t) = u(x. As before. h2 : D → R be continuous functions. y. y. we obtain 1 1 g (t) = λ = 2 g(t) c f (x. we conclude that both sides must be constant.21) 2. t) = 0. ∂2f ∂2f + 2 ∂x2 ∂y g(t). t) = u(a. y. and substitute into (5. y) ∈ R 2 : 0 ≤ x ≤ a. u satisﬁes the “Dirichlet boundary condition” u(x. y. y) ∂2f ∂2f + 2 2 ∂x ∂y . y. u satisﬁes the initial condition u(x. Find a function u(x. let us suppose that D = {(x. for (x. y)-plane which is bounded by a piecewise smooth curve ∂D. y) ∂2f ∂2f + 2 ∂x2 ∂y . t) = u(x. so that the Dirichlet boundary condition becomes u(0. ∂u (x. 0) = h1 (x. t) = f (x. y)g(t).22) . u satisﬁes the wave equation ∂2u = c2 ∂t2 ∂2u ∂2u + 2 ∂x2 ∂y . y.21) to obtain f (x. 1 c2 g(t) g (t) = 1 f (x.3. We write u(x. y). and let h1 . 0 ≤ y ≤ b}. y) ∈ ∂D. 3. y)g (t) = c2 Then we divide by c2 f (x. y). 134 (5. (5. 0.5. t) such that 1.5 Initial value problems for wave equations The most natural initial value problem for the wave equation is the following: Let D be a bounded region in the (x. Once again. 0) = h2 (x. b. If we let λ denote the constant. This separates into an ordinary diﬀerential equation g (t) = c2 λg(t). y. t) = 0. y)g(t). ∂t Solution of this initial value problem via separation of variables is very similar to the solution of the initial value problem for the heat equation which was presented in Section 5.

the mode corresponding to the pair (m. √ where ωmn = c −λmn . . 0. such that ∂2u ∂2u ∂2u = + 2. The initial value problem considered in this section could represent the motion of a vibrating membrane. y) = f (a. ∞ u(x. Just like in the case of the vibrating string.1. The Helmholtz equation is solved in exactly the same way as before. b) = 0. the only nontrivial solutions being fmn (x. 135 π a 2 + π b 2 = 1 2 T ρ 1 a 2 + 1 b 2 . y) = sin(mπx/a) sin(nπy/b). y) = f (x. t). y.and the Helmholtz equation ∂2f ∂2f + 2 ∂x2 ∂y The Dirichlet boundary condition becomes f (0. (5. t) = u(π.5. = λf. y. π. The lowest frequency of vibration or fundamental frequency is ω11 c = 2π 2π Exercises: 5. The corresponding solution to (5. 0 ≤ y ≤ π. t) = m. t) = 0. t) = u(x. t) = sin(mπx/a) sin(nπy/b)[A cos(ωmn t) + B sin(ωmn t)]. 0 ≤ x ≤ π. with λmn = −(mπ/a)2 − (nπ/b)2 . 0) = f (x.22) is g(t) = A cos(ωmn t) + B sin(ωmn t). The constants Amn and Bmn are determined from the initial conditions.n=1 sin(mπx/a) sin(nπy/b)[Amn cos(ωmn t) + Bmn sin(ωmn t)]. y. ∂t2 ∂x2 ∂y u(x. the motion of the membrane is a superposition of inﬁnitely many modes.23) Thus we obtain a product solution to the wave equation with Dirichlet boundary conditions: u(x. t) = u(0. Solve the following initial value problem for a vibrating square membrane: Find u(x. n) oscillating with frequency ωmn /2π. y. y. The general solution to to the wave equation with Dirichlet boundary conditions is a superposition of these product solutions.

t) = u(x. t) = u(π. The tool we need to express the Laplace operator ∆= ∂2 ∂2 + 2 ∂x2 ∂y in terms of polar coordinates is just the chain rule. π − x. t) = u(π.2. u(x.5. or to solve the wave equation for a vibrating circular drum. y. t) = 0. such that ∂2u =4 ∂t2 ∂2u ∂2u + 2 ∂x2 ∂y . y. the calculation is somewhat lengthy. These coordinates are related to the standard Euclidean coordinates by the formulae x = r cos θ. such that ∂2u ∂2u ∂2u = + 2. θ). y. u(x. we need to express the Laplace operator in polar coordinates (r. π. t). for 0 ≤ x ≤ π/2. y. Solve the following initial value problem for a vibrating square membrane: Find u(x. t) = 0. where p(x) = x. ∂t 5. y. ∂u (x. ∂t2 ∂x2 ∂y u(x. y. ∂u (x. y. y. ∂r ∂r ∂y ∂ = (r sin θ) = sin θ. y. t) = u(0. π − y. q(y) = y. for π/2 ≤ y ≤ π. 0 ≤ y ≤ π. 0. y. 0 ≤ y ≤ π. y. t). u(x. t) = u(0. Although straigtforward. Solve the following initial value problem for a vibrating square membrane: Find u(x. y.5.3. Since ∂x ∂ = (r cos θ) = cos θ. ∂r ∂r 136 . ∂t 5. which we studied earlier in the course. 0) = 3 sin x sin y + 7 sin 2x sin y. t) = u(x. y = r sin θ. 0) = 2 sin x sin y + 13 sin 2x sin y.u(x. 0 ≤ x ≤ π. ∂t 5. 0 ≤ x ≤ π.6 The Laplace operator in polar coordinates In order to solve the heat equation over a circular plate. π. 0) = 0. 0) = 0. for 0 ≤ y ≤ π/2. 0. ∂u (x. 0) = p(x)q(y). for π/2 ≤ x ≤ π. 0) = 0.

∂θ ∂θ it follows that ∂u ∂u ∂x ∂u ∂y ∂u ∂u = + = (−r sin θ) + (r cos θ) . we obtain ∂2u 1 ∂2u ∂ 2 u ∂ 2 u 1 ∂u + 2 2 = + 2 − . + 2 2 + 2 ∂r r ∂θ r ∂r 137 . ∂r ∂x ∂r ∂y ∂r ∂x ∂y Similarly. ∂x2 ∂x∂y ∂y = r2 sin2 θ which yields ∂2u ∂2u ∂2u ∂u − 2r2 cos θ sin θ + r2 cos2 θ 2 − r . ∆u = ∂2u 1 ∂ 2 u 1 ∂u . 2 ∂r r ∂θ ∂x2 ∂y r ∂r or equivalently. = sin2 θ 2 − 2 cos θ sin θ r2 ∂θ2 ∂x ∂x∂y ∂y r ∂r Adding these results together. 2 ∂x ∂x∂y ∂y ∂r 1 ∂2u ∂2u ∂2u ∂ 2 u 1 ∂u + cos2 θ 2 − . ∂ ∂ ∂ = (cos θ) + (sin θ) . ∂θ ∂θ For the second derivatives. ∂2u ∂ ∂u ∂u ∂ ∂u ( )= −r sin θ + r cos θ = ∂θ2 ∂θ ∂θ ∂θ ∂x ∂y = −r sin θ ∂ ∂θ ∂u ∂x + r cos θ ∂ ∂θ ∂u ∂y − r cos θ ∂u ∂u − r sin θ ∂x ∂y ∂u ∂x + sin θ ∂ ∂r ∂u ∂y ∂2u ∂2u ∂2u + 2 cos θ sin θ + sin2 θ 2 . since ∂x ∂ = (r cos θ) = −r sin θ. ∂r ∂x ∂y ∂ ∂ ∂ = (−r sin θ) + (r cos θ) . we ﬁnd that ∂2u ∂ ∂u ∂u ∂ ∂ ∂u ( )= cos θ + sin θ = cos θ = 2 ∂r ∂r ∂r ∂r ∂x ∂y ∂r = cos2 θ Similarly. ∂θ ∂x ∂θ ∂y ∂θ ∂x ∂y We can write the results as operator equations.it follows immediately from the chain rule that ∂u ∂u ∂x ∂u ∂y ∂u ∂u = + = (cos θ) + (sin θ) . ∂θ ∂x ∂y ∂y ∂ = (r sin θ) = r cos θ.

1 ∂ r ∂r r ∂u ∂r + 1 ∂2u = 0. θ). u satisﬁes the boundary condition u(1.25) 2. u is well-behaved near r = 0. we can now formulate the Dirichlet problem as follows: Find u(r. θ + 2π) = u(r. 3.Finally.25) yields 1 d r dr We multiply through by r2 . dθ2 r dR dr Θ+ R d2 Θ = 0. Indeed. where h(θ) is a given well-behaved function satisfying the periodicity condition h(θ+2π) = h(θ). Θ(θ + 2π) = Θ(θ). r2 ∂θ2 (5. r2 ∂θ2 (5. dθ2 138 .24) This formula for the Laplace operator. The ﬁrst three of these conditions are homogeneous linear. together with the theory of Fourier series. 4. Θ dθ2 Thus the partial diﬀerential equation divides into two ordinary diﬀerential equations d2 Θ = λΘ. θ). we can write this result in the form ∆u = 1 ∂ r ∂r r ∂u ∂r + 1 ∂2u . θ) = R(r)Θ(θ). θ) = h(θ). u satisﬁes Laplace’s equation. such that 1. r d dr r dR dr Θ+R d2 Θ = 0. u satisﬁes the periodicity condition u(r. To treat these conditions via the method of separation of variables. Substitution into (5. allows us to solve the Dirichlet problem for Laplace’s equation in a disk. for 0 < r ≤ 1 and θ ∈ mathbbR. Θ dθ2 The left-hand side of this equation does not depend on θ while the right-hand side does not depend on r. we set u(r. r2 dθ2 where Θ(θ + 2π) = Θ(θ). Thus neither side can depend on either θ or r. and divide by −RΘ to obtain − r d R dr r dR dr = 1 d2 Θ . and hence both sides must be constant: − r d R dr r dR dr = 1 d2 Θ = λ.

and carrying out the diﬀerentiation on the left-hand side yields the characteristic equation m2 − n2 = 0. the equation for R becomes d dr which is easily solved to yield R(r) = A + B log r. 2 where an and bn are constants. Θ= a0 . where A and B are constants of integration. dr dr If n = 0. in order for this equation to be well-behaved as r → 0 we must have B = 0. r dR dr = 0. In order for this equation to be wellbehaved as r → 0 we must have B = 0. Once again. We have seen the ﬁrst of these equations before when we studied heat ﬂow in a circular wire. and λ = −n2 . and the solution u0 (r. Θ = an cos nθ + bn sin nθ. θ) = an rn cos nθ + bn rn sin nθ. In this case.r d dr r dR dr = −λR. the solution is R(r) = Arn + Br−n . for n a positive integer. the equation for R is a Cauchy-Euler equidimensional equation and we can ﬁnd a nontrivial solution by setting R(r) = rm . When n = 0. Then the equation becomes r d dr r d m (r ) dr = n2 r m . so R(r) is a constant multiple of rn . where a0 is a constant. θ) to Laplace’s equation in this case is constant. which has the solutions m = ±n. 139 . and un (r. the only nontrivial solutions are λ = 0. and we recognize that with the periodic boundary conditions. Substitution into the second equation yields d dR r r − n2 R = 0.

1 ≤ r ≤ 2. where h(θ) = 1 + cos θ − 2 sin θ + 4 cos 2θ. such that 1 ∂ r ∂r and u(1. θ) = a0 + (an cos nθ + bn sin nθ). . r2 ∂θ2 u(r.The general solution to (5. . Solve the following boundary value problem for Laplace’s equation in a disk: Find u(r. u well-behaved near r = 0 5.1. θ) = 1 + 3 cos θ − sin θ + cos 2θ . . such that 1 ∂ r ∂r u(r. . . θ) is therefore u(r. . θ + 2π) = u(r. . . r2 ∂θ2 u(1. r ∂u ∂r + 1 ∂2u = 0. . 2 n=1 ∞ We conclude that the constants a0 . . . θ). r ∂u ∂r + 1 ∂2u = 0. . θ). + 2 n=1 ∞ where a0 .2. To determine these constants we must apply the boundary condition: h(θ) = u(1. θ). 140 r ∂u ∂r + 1 ∂2u = 0. θ + 2π) = u(r. a1 . Solve the following boundary value problem for Laplace’s equation in an annular region: Find u(r. θ). θ) = a0 (an rn cos nθ + bn rn sin nθ). and u(2. r2 ∂θ2 u(r.3. .6. θ) = 2 cos θ + 4 cos 2θ. θ + 2π) = u(r.6. θ + 2π) = u(r. θ). such that 1 ∂ r ∂r and u(1. b1 . θ) = h(θ). Exercises: 5. 0 < r ≤ 1. u well-behaved near r = 0 5. for −π ≤ θ ≤ π. are simply the Fourier coeﬃcients of h. 0 < r ≤ 1. θ) = h(θ).25) which is well-behaved at r = 0 and satisﬁes the periodicity condition u(r. a1 . b1 . Solve the following boundary value problem for Laplace’s equation in a disk: Find u(r. . where h(θ) is the periodic function such that h(θ) = |θ|.6. are constants. θ).

y)g(t). y.5.7 Eigenvalues of the Laplace operator We would now like to consider the heat equation for a room whose shape is given by a well-behaved but otherwise arbitrary bounded region D in the (x. y) This separates into g (t) = c2 λg(t) and (∆f )(x. Then we divide by c2 f (x. y. The same method can be used to treat the wave equation ∂2u = c2 ∆u. u(x. y) ∈ ∂D. and we obtain 1 1 g (t) = λ = (∆f )(x. y)g (t) = c2 (∆f )(x. and substitute into (5. 1 1 g (t) = (∆f )(x. y)g(t). y) ∈ ∂D. y). c2 g(t) f (x.27) . Both cases quickly lead to the eigenvalue problem for the Laplace operator ∆= ∂2 ∂2 + 2. f (x. 2 g(t) c f (x. ∂t with the Dirichlet boundary condition. We would also like to consider the wave equation for a vibrating drum in the shape of such a region D. y). y) = 0 for (x.26) to obtain f (x. ∂t2 141 (5. the boundary ∂D being a well-behaved curve. Hence both sides equal a constant λ. 2 ∂x ∂y Let’s start with the heat equation ∂u = c2 ∆u. t) = 0 for (x. we apply separation of variables as before. y) The left-hand side of this equation does not depend on x or y while the righthand side does not depend on t. y)g(t). (5. y) = λf (x. y) in which f is subject to the boundary condition. setting u(x. t) = f (x. y)-plane.26) To solve this equation.

it is easy to solve the initial value problem for the heat equation or the wave equation on this region. y) = λf (x. y)g(t). substitution of u(x. To do so requires only that we substitute the values of λ into the equations for g. y) = 0 for (x. The dimension of Wλ is called the multiplicity of the eigenvalue λ. In both cases. the eigenvalues determine the frequencies of a vibrating drum which has the shape of D. 142 2 λt . y) ∈ ∂D. y). y) This separates into g (t) = λg(t) and (∆f )(x. The eigenvalue problem consist of ﬁnding the eigenvalues λ. √ g (t) = c2 λg(t) ⇒ g(t) = (constant) sin(c −λ(t − t0 )). let Wλ = {smooth functions f (x. f |∂D = 0}. or 1 1 g (t) = (∆f )(x. c2 g(t) f (x. . y)g(t) into (5. Once we have solved the eigenvalue problem for a given region D in the (x. in which f is once again subject to the boundary condition. y)-plane. Nonzero elements of Wλ are called eigenfunctions and Wλ itself is called the eigenspace for eigenvalue λ.with the Dirichlet boundary condition. g (t) = c2 λg(t) ⇒ g(t) = (constant)ec while in the case of the wave equation. y) : ∆f = λf. c2 g(t) f (x. t) = 0 for (x. f (x. t) = f (x. and each eigenvalue has ﬁnite multiplicity. we must solve the “eigenvalue problem” for the Laplace operator with Dirichlet boundary conditions. y) ∈ ∂D.27) yields f (x. and we obtain 1 1 g (t) = λ = (∆f )(x. y)g (t) = c2 (∆f )(x. We say that λ is an eigenvalue of the Laplace operator ∆ on D if Wλ = 0. u(x. and a basis for each nonzero eigenspace. All of the eigenvalues of ∆ are negative. y). In the case of the heat equation. y. This time. y. Theorem. The eigenvalues can be arranged in a sequence 0 > λ1 > λ2 > · · · > λ n > · · · . y) Once again. In the second case. If λ is a real number. both sides must equal a constant λ. y).

. D = {(x. We saw how to solve the eigenvalue problem ∂2f ∂2f + 2 = λf. There are two important cases in which the eigenvalues can be explicitly determined. λ2 . D = {(x. . Interscience.7.2.with λn → −∞. In fact. where a is a positive number. y) : 0 ≤ x ≤ a.7. Every well-behaved function can be represented as a convergent sum of eigenfunctions. Hint: One can eliminate b by setting a = 1/b and then ﬁnd the minimum of the function g(a) = when a > 0. λ2 . · · · explicitly for a given region in the plane.1. b) = + a b subject to the constraints ab = 1. Methods of mathematical physics I . 2 Note the similarity between the statement of this theorem and the statement of the theorem presented in Section 4. y) = sin πny πmx sin . y) : x2 + y 2 ≤ a2 }. fmn (x. a. Let D be a ﬁnite region in the (x. §16. y) ∈ ∂D. 3 A few more cases are presented in advanced texts. as we will see in the next section. · · · . . Suppose that the eigenvalues for the Laplace operator ∆ with Dirichlet boundary conditions on D are λ1 . The ﬁrst is the case where D is the rectangle.3. See Chapter V. such as Courant and Hilbert. New York. when we discussed the heat and wave equations for a rectangular region. The nontrivial solutions are λmn = − πm a 2 − πn b 2 . A second case in which the eigenvalue problem can be solved explicitly is that of the disk. where 0 > λ1 > λ2 > · · · . Show that among all rectangular vibrating membranes of area one. it is usually quite diﬃcult to determine the eigenvalues λ1 . π a 2 + (πa) .2 Although the theorem is reassuring. 2 143 . . a b where m and n are positive integers. y)-plane bounded by a smooth curve ∂D. 1953. the square has the lowest fundamental frequency of vibration by minimizing the function π 2 π 2 f (a. y) = 0 for (x.3 Exercises: 5. 0 ≤ y ≤ b}. the techniques used to prove the two theorems are also quite similar. b > 0. 5. 2 ∂x ∂y f (x. λn .7.

Suppose that the eigenvalues for the Laplace operator ∆ on D are λ1 . 144 . . Θ(θ + 2π) = Θ(θ). . λ2 . y) is a nonzero eigenfunction for eigenvalue λn . y.7. where the bn ’s are arbitrary constants. r2 ∂θ2 f |∂D = 0. y)eλn t . t) = n=1 bn φn (x. Suppose that φn (x. t) = n=1 bn φn (x. Substitution into (5. 5. ∂t2 together with Dirichlet boundary conditions (u vanishes on ∂D) and the initial condition ∂u (x.28) yields 1 d r dr r dR dr Θ+ R d2 Θ = λRΘ. θ in terms of which the Laplace operator is 1 ∂ r ∂r r ∂f ∂r + 1 ∂2f = λf. ∂t is ∞ u(x. 0) = 0. y) cos( −λn t). (5.each eigenvalue having multiplicity one. 5. where the bn ’s are arbitrary constants.8 Eigenvalues of the disk To calculate the eigenvalues of the disk. Show that the general solution to the heat equation ∂u = ∆u ∂t with Dirichlet boundary conditions (u vanishes on ∂D) is ∞ u(x.3. we use separation of variables and look for product solutions of the form f (r.28) Once again. r2 dθ2 where R(a) = 0. y)-plane as in the preceding problem. it is convenient to utilize polar coordinates r. once again. θ) = R(r)Θ(θ). . Show that the general solution to the wave equation ∂2u = ∆u. . Let D be a ﬁnite region in the (x.

Then since r d d =x . R(a) = 0. Θ dθ2 Thus in the manner now familiar.30) + (x2 − n2 )y = 0. Substitution into the second equation yields r Let x = √ d dr r dR dr + (−λr2 − n2 )R = 0. for n a positive integer. r d dr r dR dr Θ+R d2 Θ = λr2 RΘ. and hence both sides must be constant: − r d R dr r dR dr + λr2 = 1 d2 Θ = µ. the only nontrivial solutions are µ = 0. and µ = −n2 . Thus neither side can depend on either θ or r. Θ= a0 2 −λr so that x2 = −λr2 . (5. If we replace R by y. − λr2 R = −µR. (5. Θ = an cos nθ + bn sin nθ. the partial diﬀerential equation divides into two ordinary diﬀerential equations d2 Θ = µΘ. dθ2 and divide by RΘ to obtain r d R dr r dR dr − λr2 = − 1 d2 Θ . this becomes √ y( −λa) = 0. 145 . dθ2 r d dr r dR dr Θ(θ + 2π) = Θ(θ).We multiply through by r2 . the left-hand side does not depend on θ while the right-hand side does not depend on r. Once again. dr dx our diﬀerential equation becomes x d dx x √ dR dx + (x2 − n2 )R = 0.29) where R vanishes when x = x d dx x dy dx −λa. Θ dθ2 Note that in this last equation.

Bessel’s equation has a one-dimensional space of well-behaved solutions. θ) = J0 (α0. we sketch the proof for the case of J0 (x) at the end of the section. which are constant multiples of the Bessel function of the ﬁrst kind Jn (x). . Jn (x) has inﬁnitely many positive zeros. √ y( −λa) = 0.k r/a) sin(nθ). 2 α0. which we studied in Section 1. λn. . 146 .k .The diﬀerential equation appearing in (5.k r/a) cos(nθ) or gn. Hence. Here is an important fact regarding these Bessel functions: Theorem.k . Then √ and R(r) = Jn (αn.k denote the k-th positive root of the equation Jn (x) = 0. in the case of general n. Graphs of the functions J0 (x) and J1 (x) suggest that this theorem might well be true. r2 ∂θ2 f |∂D = 0.29) vanishing at r = a.k = − and f0. √ requires that −λa be one of the zeros of Jn (x). note ﬁrst that the boundary condition. .4. a2 −λa = αn. a2 for n = 1.k r/a) will be a solution to (5.k (r.k (r.k ⇒ λ=− 2 αn.k r/a) will be a solution to the eigenvalue problem 1 ∂ r ∂r r ∂f ∂r + 1 ∂2f = λf.k = Jn (αn. but it takes some eﬀort to prove rigorously. a2 Similarly.4. λ0. .k .k = − and fn. in the case where n = 0. 2.30) is Bessel’s equation. will be solutions to this eigenvalue problem. The zeros of the Bessel functions are used to determine the eigenvalues of the Laplace operator on the disk. Recall that in Section 1. θ) = Jn (αn. we found that for each choice of n. For completeness. 2 αn. Let αn. To see how. For each nonnegative integer n.

1 = −26.1 = 5.75 0.87855. α0.1 (r.38274.2 = −30.5 0.7832.1 will vibrate with frequency −λ1.2 α0. θ).25 0 -1 -0. and so forth.1 will vibrate with frequency −λ0.5 0 1 0.5 Figure 5. The mode of oscillation corresponding to the function f0. that a = 1. Then α0.4 = 11. for example. α1.60984.3746.83171 ⇒ λ1.5 0 0.1 0.5 1 -1 -0.471.1 = 2.3 = −139. 2π 2π while the mode of oscillation corresponding to the function f0. 2π 2π Similarly.1 α1. Suppose.40483 ⇒ λ0.2 will vibrate with frequency −λ0.7195 ⇒ λ0.039.682.52008 ⇒ λ0.2 = = .1 = −14. α0.65373 ⇒ λ0.1 and g1.1 α0.3 = −74.1 = 3.1 = −5. 2π 2π 147 .2: Graph of the function f0.3 = 8.887. α2. Each of these eigenfunctions corresponds to a mode of oscillation for the circular vibrating membrane.2 = 5. and hence the mode of oscillation corresponding to the functions f1.1 = = .1 = = .13562 ⇒ λ2. α0.

1 and g2. Proof that J0 (x) has inﬁnitely many positive zeros: First. dx e dz dz Thus Bessel’s equation (5. we make a change of variables x = ez and note that as z ranges over the real numbers.1 = = .1 will vibrate with frequency −λ2. We claim that y(z) must change sign at some point between z0 and z0 + π.5 0 -0.31) with y(z0 ) = 0.5 -1 1 0.31) Suppose that z0 > 1 and y(z) is a solution to (5.5 0 -1 -0.30) in the case where n = 0 becomes d2 y + e2z y = 0.5 0 0. the corresponding variable x ranges over all the positive real numbers. 2π 2π A general vibration of the vibrating membrane will be a superposition of these modes of oscillation. dz 2 (5. while the mode of oscillation corresponding to the functions f2.81736. d 1 d = z dx e dz and hence x 1 d d d = ez z = .2 (r.5 1 Figure 5. Since dx = ez dz.3: Graph of the function f0. 148 . θ).1 α2.1 0. The fact that the frequencies of oscillation of a circular drum are not integral multiples of a single fundamental frequency (as in the case of the violin string) limits the extent to which a circular drum can be tuned to a speciﬁc tone.

31).1 (r. 149 . and hence y(z) must be zero at some point in this interval. d [y(z)f (z) − y (z)f (z)] = y(z)f (z) − y (z)f (z) dz = −y(z)f (z) + e2z y(z)f (z) = (e2z − 1)y(z)f (z). In either case.31) must have inﬁnitely many zeros in the region z > 1.5 0 Figure 5. It follows that the solution to (5.5 1 0. then y(z)f (z) > 0 and z0 +π 0< z0 (e2z − 1)y(z)f (z)dz = [y(z)f (z) − y (z)f (z)]z0 0 z +π = y(z0 + π)f (z0 + π) − y(z0 )f (z0 ) = −y(z0 + π) − y(z0 ) < 0. θ). Let f (z) = sin(z − z0 ). If also y(z) > 0 for z between z0 and z0 + π.5 0. a contradiction. The fact that Jn (x) has inﬁnitely many positive zeros could be proven in a similar fashion. and J0 (x) must have inﬁnitely many positive zeros.5 0 0. Since f (z) = −f (z) and y(z) satisﬁes (5. just apply the same argument to −y.0.25 0 -0. Assume ﬁrst that y(z0 ) > 0. y(z) must change sign at some z between z0 and z0 + π. we conclude that y(z) must be zero at some point in any interval to the right of z = 1 of length at least π.5 1 -1 -0.25 -0.5 -1 -0. as claimed.4: Graph of the function f1. Note that f (z) > 0 for z between z0 and z0 + π. If y(z0 ) < 0. Hence our assumption that y(z) be postive for z between z0 and z0 + π must be incorrect.

2}] Find the next two nonzero roots of the Bessel function J0 (x). θ. . we need to develop a theory of generalized 150 .1 r) sin θ + 4J2 (α2. Plot[ BesselJ[n. θ. 5. θ. 0) = J0 (α0. r ∂u ∂r + 1 ∂2u .2 r) − 2J1 (α1. u(1.8. θ + 2π. FindRoot[ BesselJ[n. t) = u(r. 5.1 r) + 3J0 (α0.x]. ∂t u(r. 0 < r ≤ 1.1 r) cos 2θ. . (For students with access to Mathematica) a.3. 0 < r ≤ 1. 0) = 0. .15}] b. t) = 0.2 r) − 2J1 (α1.{x. such that ∂2u 1 ∂ = ∂t2 r ∂r u(r. θ.9 Fourier analysis for the circular vibrating membrane* To ﬁnish up the solution to the initial value problem for arbitrary initial displacements of the vibrating membrane. θ.8. θ + 2π. Solve the following initial value problem for the vibrating circular membrane: Find u(r.1 r) + 3J0 (α0. u(1. t) = 0. Which has a lower fundamental frequency of vibration. θ.2. and determine the ﬁrst three nonzero roots of each of these Bessel functions. 5.Exercises: 5. r ∂u ∂r + 1 ∂2u . t).4. t). u(r.x] == 0 . ∂u (r. such that ∂u 1 ∂ = ∂t r ∂r u(r. r2 ∂θ2 u well-behaved near r = 0. t). Modify the programs to sketch the Bessel functions J1 (x). .1 of the Bessel function J0 (x): n=0.1 r) sin θ + 4J2 (α2. Run the following Mathematica program to sketch the Bessel function J0 (x): n=0. θ. {x. t). θ. t) = u(r. r2 ∂θ2 u well-behaved near r = 0. Solve the following initial value problem for the heat equation in a disk: Find u(r. c. From the graph it is clear that the ﬁrst root of the equation J0 (x) = 0 is near 2.0.1.8.8. J5 (x). Run the following Mathematica program to ﬁnd the ﬁrst root α0. 0) = J0 (α0.1 r) cos 2θ. θ. a square drum or a circular drum of the same area? 5.

Fourier series which gives a decomposition into eigenfunctions which solve the eigenvalue problem. Indeed. Such a theory can be developed for an arbitrary bounded region D in the (x. g . ∆f . and hence (λ − µ) f. Thus if f and g lie in V . if (5. We deﬁne an inner product . g = 0. It is important to observe that V is a “vector space”: the sum of two elements of V again lies in V and the product of an element of V with a constant again lies in V . g = 0. if the restriction of g to ∂D vanishes. With respect to this inner product. Let V denote the space of smooth functions f : D → R whose restrictions to ∂D vanish. eigenfunctions corresponding to distinct eigenvalues are perpendicular. In particular. g = λ f. Lemma. then µ f. ∆g = ∆f. (5. it follows from Green’s theorem that −f ∂D ∆g = µg. y)-plane bounded by a smooth closed curve ∂D. ∆g = − D ∇f · ∇gdxdy = g. then either λ = µ or f. g = f. Similarly. if f and g are smooth functions vanishing on ∂D such that ∆f = λf. The proof of this lemma is a nice application of Green’s theorem.32) holds. f. g∆f dxdy = − D D ∇f · ∇gdxdy. f ∆gdxdy = − D D ∇f · ∇gdxdy. Hence if the restriction of f to ∂D vanishes. on V by f. g = D f (x. y)dxdy.32) ∂g ∂g dx + f dy = ∂y ∂x = D D ∂ ∂ [f (∂g/∂x)] − [−f (∂g/∂y)] dxdy ∂x ∂y f D ∂f ∂g ∂f ∂g + dxdy + ∂x ∂x ∂y ∂y ∂2g ∂2g + 2 ∂x2 ∂y dxdy. 151 . y)g(x.

t). the general solution must be a superposition of the products of these eigenfunctions with periodic functions of t of frequencies −λn. θ) is also satisﬁed. ∂t It follows from the argument in the preceding section that the general solution to the homogeneous linear part of the problem can be expressed in terms of the eigenfunctions u(r. so that ∂2u 1 ∂ = ∂t2 r ∂r u(r. 0) = h(r. r2 ∂θ2 r + u well-behaved near r = 0. θ) as a superposition of these eigenfunctions.k = Jn (αn. If we set t = 0 in (5. This implies. Recall the problem that we want to solve. Indeed.k gn.k ’s so that the inhomogeneous initial condition u(r. g = 0. 0 < r ≤ 1.k f0. t) = u(r. and the lemma is proven. θ. θ) cos(α0. θ) + bn.k = 0. The lemma implies that eigenfunctions corresponding to distinct eigenvalues are perpendicular.k g0.k (r. in terms of polar coordinates: Find u(r. θ. It is here that the inner product . (5. θ) = Jn (αn.k gn. θ)] cos(αn. 0) = 0.k r/a) cos(nθ).34) Thus we need to express an arbitrary initial displacement h(r. θ. 0) = h(r. t).k (r. fn.k (r.k ’s and the bn. this condition becomes ∞ ∞ ∞ a0.k ] = h. θ. ∂u ∂r 1 ∂2u . on V comes to the rescue. . θ.It follows immediately that either λ − µ = 0 or f.k + bn. u(1.33) We need to determine the an.k t) + n=1 k=1 [an. t) = k=1 ∞ ∞ a0.j . for example. θ. t) = 0. ∂u (r. θ).k /2π.k fn. Thus this general solution is of the form ∞ u(r. (5.k (r. θ.33). that f0. 152 unless j = k. θ + 2π. Now let us focus on the special case in which D is a circular disk.k r/a) sin(nθ).k fn. gn. f0.k t).k + k=1 n=1 k=1 [an.

k = 0.k = gn.j . 0 we can check that fn. If we perform integration with respect to θ ﬁrst. for arbitrary n ≥ 1. fn. gn. g = 0 0 f (r. gn.j .j . The lemma does not immediately imply that fn.j . gm. unless m = n.j .k = 0. gn.j . using the integral formulae 2π cos nθ cos mθdθ = 0 2π π.Similarly. f. Moreover. θ)g(r. the inner product . gm. for m = n.j = 0.k = h. we can conclude from the familiar integral formula 2π sin nθ cos nθdθ = 0 0 that fn.j . sin nθ sin mθdθ = 0 2π sin nθ cos mθdθ = 0.j .j = 0. fm. It then follows from the lemma that fn. for m = n.k . θ)rdθ dr. 0.k fn. π. for m = n. To obtain this relation. it is not diﬃcult to construct formulae for the coeﬃcients of the generalized Fourier series of a given function h(r.k = fn. gm. From these relations. for m = n. Indeed. recall that in terms of polar coordinates. fn.j . 0.j . fm. we obtain an.k . fn. if we take the inner product of equation (5.j . unless j = k and m = n. 153 . gm. fn.k = 0.34) with un.k = fn.k = gn. θ). on V takes the form 1 2π unless j = k.k = gn. as desired.k . gn.k = fn.

1 (1-r).40483 r]) .k (r. we can write h(r.k (r. . for 0 ≤ r ≤ 1. . 0 < r ≤ 1. a0.1}]. Determine the ﬁrst four terms in the solution to the initial value problem: Find u(r.40483 r])∧2. fn. gn.2. gn. and a0. where the coeﬃcients are determined by the integral formulae (5.k . a01 = n/d b.k . Because of axial symmetry.1(1 − r).1 in this expansion: d = NIntegrate[r (BesselJ[0. so that ∂2u 1 ∂ = ∂t2 r ∂r r ∂u ∂r + 1 ∂2u .k gn. In order to ﬁnd the solution to the wave equation with this initial condition.k t) + n=1 k=1 [an.k t).2 r) + . bn.36). θ) + bn. Exercises: 5.or equivalently an. θ) in terms of eigenfunctions of ∆. gn. fn.3 .9.k fn. .k .{r. n = NIntegrate[r (BesselJ[0.2 .k = h.2 J0 (α0.{r.k = Similarly. Use a modiﬁcation of the program to determine the coeﬃcients a0. θ)] cos(αn. fn. t) = k=1 ∞ ∞ a0.k (5.k f0.36) h.1}]. θ.35) Thus we ﬁnally see that the solution to our initial value problem is ∞ u(r.1.2. t).0.1 r) + a0. θ) = a0. θ) cos(α0. a. θ.k (5. θ) = .4 . we need to expand h(r. c.0.35) and (5.k . Use the following Mathematica program to determine the coeﬃcient a0. r2 ∂θ2 154 . (For students with access to Mathematica) Suppose that a circular vibrating drum is given the initial displacement from equilibrium described by the function h(r.1 J0 (α0.k (r.

.8.65 t] + a04 BesselJ[0. 0) = .2 Pi}. vibmem = Table[ CylindricalPlot3D[ a01 BesselJ[0. 5.1} ] a.2.u(r. θ. u(r.) b. PlotRange -> {-.83 t].2. {r.83 t] + b11 BesselJ[1.1. {u. (Warning! Generating the sequence of graphics takes lots of memory. ∂t u(1. b11 = .52 t] + a03 BesselJ[0.3.. If you have insuﬃcient memory for the animation. θ + 2π. ..4 t] + a02 BesselJ[0. with the Fourier coeﬃcients obtained in the preceding problem. t) = 0.5.3. a02 = .5.83 r] Cos[u] Cos[3. (For students with access to Mathematica) The following program will sketch graphs of the successive positions of a membrane. Replace the values of a01.” from the Graphics menu.4 r] Cos[2.2. t) = u(r.0. vibrating with a superposition of several of the lowest modes of vibration: << Graphics‘ParametricPlot3D‘ a01 = . {t. 155 .. θ.0. ∂u (r.3. 0) = 0. θ.1. a03 = . a04 = .8 t] + a11 BesselJ[1.1}.52 r] Cos[5. and animate the sequence as before. a11 = . t). θ.1. Execute this program and then animate the sequence of graphics by running “Animate selected graphics.9.1(1 − r). execute the resulting program to generate a sequence of graphics. u well-behaved near r = 0.5} ]. .2.65 r] Cos[8.11.83 r] Sin[u] Cos[3.0.2.8 r] Cos[11. try inserting PlotPoints -> 8 within the CylindricalPlot statement.

and respond with Out[1] = 2 The notation for multiplication is ∗.Appendix A Using Mathematica to solve diﬀerential equations In solving diﬀerential equations. you can use it as a “more powerful than usual graphing calculator. 1999. It is The Mathematica Book . Matlab.5)/3 the computer will perform the subtraction and division. computers can do the calculations for us. such as Maple. or Mathematica. Fortunately. there is a standard reference on the use of Mathematica.” For example. Fourth edition.” which can be used to give a brief hands-on introduction to Mathematica. After launching Mathematica. which can be used to answer questions if necessary. We give here a very brief introduction to the use of Mathematica. it may be possible for many students to learn Mathematica without referring to lengthy technical manuals. “Tour of Mathematica. However. This appendix is intended to give a brief introduction to the use of Mathematica for doing such calculations. if they are equiped with suitable software. by Stephen Wolfram. if you type in (11 . so if you type in 2 * (7 + 4) the computer will respond with Out[2] = 22 156 . Wolfram Media and Cambridge University Press. Using this tour. Most computers which are set up to use Mathematica contain an on-line tutorial. it is sometimes necessary to do calculations which would be prohibitively diﬃcult to do by hand.

typing in N[Pi. so if you input 2 8 the computer will respond with Out[3] = 16 The computer can do exact arithmetic with integers or rational numbers. so entering N[2∧150] will give an approximation to the exact answer. For example.42725 1045 Real numbers which are not rational. Since the number π is represented by Mathematica as Pi. we can approximate a real number by a ﬁnite decimal expansion with a given number of digits of precision. For example. However.71828 In principle.You can also use a space instead of ∗ to denote multiplication. since Mathematica uses the name E to denote the number e.1415926535897932384626433832795028841972 The computer can represent not only numbers but functions. the computer can calculate rational approximations to an arbitrary number of digits of precision. built-in functions are described by names which begin with capital 157 .40] will tell Mathematica to give a rational approximation to π to 40 digits of precision: Out[7] = 3. such as π. typing in N[E] will tell Mathematica to give a rational approximation to e to a standard number of digits of precision: Out[6] = 2. since ∧ is the symbol used for raising to a power. have inﬁnite decimal expansions which cannot be stored within a computer. expressed in scientiﬁc notation: Out[5] = 1. the symbol N tells the computer to use an approximation with a ﬁxed number of digits of precision. Within Mathematica. if you type in 2∧150 the computer will calculate the 150th power of two and give the exact result: Out[4] = 1427247692705959881058285969449495136382746624 On the other hand.

8 0.6 0. being careful not to capitalize them. Thus entering N[Log[2]. Thus we can deﬁne the function y(t) = cekt by typing y[t ] := c * E∧(k * t).693147180559945309417232121458176568076 One can also plot functions with Mathematica. N[y[1]] 158 .1. We must remember the exact syntax of this example (use of the underline character and the colon followed by the equal sign) when deﬁning functions. In this example c and k are parameters which can be deﬁned later. Mathematica will remember the function we have deﬁned until we quit Mathematica.5 2 2. For example. letters. so they will not be confused with built-in constants. k = 1.4 0. because capitals are reserved by Mathematica for built-in functions.5 3 Figure A.1: Graph of the logarithm function.40] will give the logarithm of 2 with 40 digits of precision: Out[8] = 0.{t. Just as in the case of functions. Further entry of c = 1.2 1. For example.1. one simply inputs Plot[Log[t].1 0. We can also deﬁne functions ourselves. to plot the logarithm function from the values 1 to 3. Log denotes the natural base e logarithm function. we must be careful to represent parameters by lower case letters.3}] and Mathematica will automatically produce the plot shown in Figure A.

2].2}] will give a plot of the function we have deﬁned.1) where A is a square matrix with constant entries. a = {{-2.{t. it is prudent to denote the matrix A by lower case a when writing statements in Mathematica. dt (A.” Entering MatrixForm[a] will cause the computer to give the matrix in the familiar form Out[2] = −2 1 5 −3 To ﬁnd the eigenvalues of the matrix A. The matrix −2 5 A= 1 −3 can be entered into Mathematica as a collection of row vectors.yields the response Out[11] = 2. The ﬁrst step consists of using Mathematica to ﬁnd the eigenvalues and eigenvectors of A. To see how this works.0.5}.{1.{1. . Quite often numerical approximations are suﬃcient.-3}} Thus a matrix is thought of as a “vector of vectors.-3}} with the computer responding by Out[1] = {{-2. we simply type Eigenvalues[a] and the computer will give us the exact eigenvalues √ √ −5 − 21 −5 + 21 . 2 2 which have been obtained by using the quadratic formula. We can use Mathematica to solve matrix diﬀerential equations of the form dx = Ax.5}.71828 while entering Plot[y[t]. Since Mathematica reserves upper case letters for descriptions of built-in functions. we must ﬁrst become familiar with the way in which Mathematica represents matrices. and these can be obtained by typing eval = Eigenvalues[N[a]] 159 . for t in the interval [0.

873154.941409.337267}} Putting this together with the eigenvalues gives the general solution to the original linear system (A.208712} Deﬁning eval to be the eigenvalues of A in this fashion.6}] generates an “interpolation function” which approximates the solution and calls it sol. remembering to replace the equal signs with double equal signs.6.2. y(0) = .1 }. -0. Readers can modify these simple programs to graph solutions to initial value problems for quite general diﬀerential equations of the canonical form dy = f (x. y). The following Mathematica programs will use Mathematica’s diﬀerential equation solver (which is called up by the command NDSolve). {x. to ﬁnd a numerical solution to the initial value problem dy/dx = y(1 − y).6}] This leads to a plot like that shown in Figure A. We can construct a table of values for the interpolation function by typing Table[Evaluate[y[x] /. as in the example.0.0.1) for our choice of the matrix A: x(t) = c1 −0. y.79129. 0..1}].the response this time being Out[4] = {-4. {x. allows us to refer to the eigenvalues of A later by means of the expression eval. 0.941409 −0. 0. y[0] == . or graph the interpolation function by typing Plot[Evaluate[y[x] /. an abbreviation for solution.487445}. 160 . We can also ﬁnd the corresponding eigenvectors for the matrix A by typing evec = Eigenvectors[N[a]] and the computer responds with Out[5] = {{-0.208712t e e + c2 .337267 Mathematica can also be used to ﬁnd numerical solutions to nonlinear diﬀerential equations.0. give a table of values for the solution. and graph the solution curve on the interval 0 ≤ x ≤ 6. {0.1.873154 −4. The ﬁrst step sol := NDSolve[{ y’[x] == y[x] (1 . sol]. sol]. {x.487445 0.y[x]).79129t 0. dx All that is needed is to replace the ﬁrst argument of NDSolve with the diﬀerential equation one wants to solve.

1 0.1.75 1.1 0.5 0.4 0. 0. dy/dt = xy − y. 161 . y(0) = .2: Solution to y = y(1 − y).25 1.3 0.3: A parametric plot of a solution to dx/dt = −xy.2 0.5 1.75 2 Figure A.6 0.4 0.2 1 2 3 4 5 6 Figure A.8 0.

x[t] y[t].1.10}] Once we have this solution. {t.0. y’[t] == x[t] y[t] .2) We can also plot y(t) as a function of t by entering Plot[Evaluate[y[t] /.0. {t. y(0) = . sol]. y[t]} /. to solve the initial value problem dx = −xy.3 shows a parametric plot of (x(t).0. sol]. x[0] == 2.10}] Figure A.0. {x.10}] 162 . dt x(0) = 2. {t. sol]. For example. we can print out a table of values for y by entering Table[Evaluate[y[t] /. y[0] == .y[t]. it is no more diﬃcult to treat initial value problems for higher order equations or systems of diﬀerential equations.. (A.2.In fact. {t. y(t)) which is generated by entering ParametricPlot[Evaluate[{ x[t].1 }. dt one simply types in sol := NDSolve[{ x’[t] == .1}] dy = xy − y.y}.

70 ﬁnite diﬀerences. 13 orthogonal matrix. 91 Newton’s second law. 135 Frobenius method. 129 power series. 117 163 heat ﬂow in a bar. 133 Neumann boundary conditions. 5 recursion formula. 89 Fourier transform. 82 heat ﬂow in space. 130 . 16 fundamental mode. 29 orthonormal basis. 63 Fourier sine series. 25 boundary value problem. 135 modiﬁed Bessel equation. 119 double Fourier sine series. 142 eigenvalues. 119 Cauchy-Euler equation. 146 Bessel function. 72. 115 Legendre’s equation. 27 Navier-Stokes equations. 110. 50. 40 initial value problem. 94 Fourier coeﬃcients. 40 Euler equations for a gas. 83. 132 mechanical systems. 33. 58. 68. 142 eigenvalue problem. 13 separating constant. 85 Dirichlet problem. 70 ordinary point. 22. 52. 63 piecewise smooth. 72. 139 Chebyshev’s equation. 8 regular singular point. 108. 134 inner product. 108. 9 generalized Fourier series. 44. 95. 2 Dirichlet boundary conditions. 50. 49 modes of oscillation. 87. 108. 98. 142 eigenspace. 86 singular point. 79 frequencies of oscillation. 69 partial diﬀerential equation. 86 separation of variables. 30. 132 even function. 96 odd function. 91 Fourier series. 124. 85. 81 periodic function. 13 sound waves. 10 Hooke’s law. 1 quadric surfaces. 14 comparison test. 151 Laplace’s equation. 83. 17. 37 hyperboloid. 67 potential energy. 49 hyperbola. 116 Hermite’s equation. 39 radius of convergence. 100 fundamental frequency. 4 real analytic. 67 heat equation. 151 generalized power series. 58. 115. 28 geometric series. 135 gamma function. 63 Fourier cosine series. 12 linearization of PDE. 37 ellipsoid.Index Bessel’s equation. 142 eigenvectors. 36 convergence. 2 general solution. 16 Gibbs phenomenon. 103. 126 eigenfunctions. 3 ratio test. 52. 5 conic section. 30 ellipse. 108.

129 vibrating string. 94 wave equation. 18 vibrating membrane. 96. 115 symmetric matrix.spectral theorem. 16 variation of parameters. 86. 30 Sturm-Liouville problem. 29 Theorem of Frobenius. 108 superposition principle. 115 164 .