Cambridge PDE

Linear Partial Differential
Equations
and
Fourier Theory
Marcus Pivato
Trent University
ii DRAFT
Copyright 2009, Marcus Pivato
This is the unedited draft manuscript for a text which will be
published by Cambridge University Press later in 2009. Cambridge
UP has kindly allowed the author to make this manuscript freely
available on his website. You are free to download and/or print this
manuscript for personal use, but you are not allowed to duplicate
it for resale or other commercial use. If you have any questions or
comments, please contact the author at
marcuspivato@trentu.ca
Cambridge University Press retains the exclusive worldwide right
to produce and publish this Work, and/or to license the production
and publication of the work and any adaptation or abridgmeent of
the Work in all forms and media and in all languages throughout the
world.
Linear Partial Dierential Equations and Fourier Theory Marcus Pivato DRAFT January 31, 2009
To Joseph and Emma Pivato
for their support, encouragement,
and inspiring example.
C
h
a
p
t
e
r

D
e
p
e
n
d
e
n
c
y

L
a
t
t
i
c
e
D
e
t
a
i
l
e
d

l
i
s
t
s

o
f

p
r
e
r
e
q
u
i
s
i
t
e
s
a
p
p
e
a
r

a
t

t
h
e

s
t
a
r
t

o
f

e
a
c
h

s
e
c
t
i
o
n

o
f

t
h
e

t
e
x
t
.
R
e
q
u
i
r
e
d
R
e
c
o
m
m
e
n
d
e
d
PSfrag replacements
1
H
e
a
t
E
q
n
.
2
W
a
v
e
E
q
n
.
3
Q
u
a
n
t
u
m
4
L
i
n
e
a
r
5
A
-
5
B
I
V
P
s
5
C
-
5
D
B
V
P
s
6
F
u
n
c
.
A
n
a
l
y
s
i
s
7
F
o
u
r
i
e
r
[
0
,
8
F
o
u
r
i
e
r
[
9
M
u
l
t
i
F
o
u
r
i
e
r
1
1
R
o
d
B
V
P
1
2
B
o
x
B
V
P
1
3
C
u
b
e
B
V
P
1
4
P
o
l
a
r
B
V
P
1
6
A
-
1
6
B
S
e
p
.
V
a
r
.
1
6
C
-
1
6
D
S
e
p
.
V
a
r
.
1
7
A
-
1
7
B
A
p
p
r
.
I
d
1
7
C
-
1
7
D
I
m
p
u
l
s
e
1
7
E
P
o
i
s
s
o
n
1
7
F
P
o
i
s
s
o
n
1
9
F
o
u
r
i
e
r
T
r
a
n
s
2
0
F
o
u
r
i
e
r
S
o
l
n
1
5
E
i
g
e
n
M
e
t
h
o
d
1
0
C
o
n
v
.
P
r
o
o
f
s
1
8
C
o
m
p
l
e
x
v
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
Whats good about this book? . . . . . . . . . . . . . . . . . . . . . . xv
Suggested Twelve-Week Syllabus . . . . . . . . . . . . . . . . . . . . . xxi
I Motivating examples and major applications 1
1 Heat and diusion 3
1A Fouriers law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1A(i) ...in one dimension . . . . . . . . . . . . . . . . . . . . . 3
1A(ii) ...in many dimensions . . . . . . . . . . . . . . . . . . . . 4
1B The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1B(i) ...in one dimension . . . . . . . . . . . . . . . . . . . . . 5
1B(ii) ...in many dimensions . . . . . . . . . . . . . . . . . . . . 7
1C Laplaces equation . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1D The Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . 12
1E Properties of harmonic functions . . . . . . . . . . . . . . . . . . 16
1F

Transport and diusion . . . . . . . . . . . . . . . . . . . . . . . 18
1G

Reaction and diusion . . . . . . . . . . . . . . . . . . . . . . . 19
1H Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2 Waves and signals 23
2A The Laplacian and spherical means . . . . . . . . . . . . . . . . 23
2B The wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2B(i) ...in one dimension: the string . . . . . . . . . . . . . . . 28
2B(ii) ...in two dimensions: the drum . . . . . . . . . . . . . . . 32
2B(iii) ...in higher dimensions: . . . . . . . . . . . . . . . . . . . 34
2C The telegraph equation . . . . . . . . . . . . . . . . . . . . . . . . 34
2D Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Quantum mechanics 37
3A Basic framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3B The Schrodinger equation . . . . . . . . . . . . . . . . . . . . . . 41
3C Stationary Schrodinger equation . . . . . . . . . . . . . . . . . . 45
vi DRAFT CONTENTS
II General theory 56
4 Linear partial dierential equations 57
4A Functions and vectors . . . . . . . . . . . . . . . . . . . . . . . . 57
4B Linear operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4B(i) ...on nite dimensional vector spaces . . . . . . . . . . . . 59
4B(ii) ...on (
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4B(iii) Kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4B(iv) Eigenvalues, eigenvectors, and eigenfunctions . . . . . . . 63
4C Homogeneous vs. nonhomogeneous . . . . . . . . . . . . . . . . 64
5 Classication of PDEs and problem types 69
5A Evolution vs. nonevolution equations . . . . . . . . . . . . . . . 69
5B Initial value problems . . . . . . . . . . . . . . . . . . . . . . . . 70
5C Boundary value problems . . . . . . . . . . . . . . . . . . . . . . 71
5C(i) Dirichlet boundary conditions . . . . . . . . . . . . . . . 73
5C(ii) Neumann boundary conditions . . . . . . . . . . . . . . . 76
5C(iii) Mixed (or Robin) boundary conditions . . . . . . . . . . 81
5C(iv) Periodic boundary conditions . . . . . . . . . . . . . . . . 82
5D Uniqueness of solutions . . . . . . . . . . . . . . . . . . . . . . . 84
5D(i) Uniqueness for the Laplace and Poisson equations . . . . 86
5D(ii) Uniqueness for the heat equation . . . . . . . . . . . . . . 89
5D(iii) Uniqueness for the wave equation . . . . . . . . . . . . . . 92
5E

Classication of second order linear PDEs . . . . . . . . . . . . 95
5E(i) ...in two dimensions, with constant coecients . . . . . . 95
5E(ii) ...in general . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5F Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 98
III Fourier series on bounded domains 102
6 Some functional analysis 103
6A Inner products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6B 1
2
space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6C

More about 1
2
space . . . . . . . . . . . . . . . . . . . . . . . . 108
6C(i) Complex 1
2
space . . . . . . . . . . . . . . . . . . . . . . 109
6C(ii) Riemann vs. Lebesgue integrals . . . . . . . . . . . . . . 110
6D Orthogonality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6E Convergence concepts . . . . . . . . . . . . . . . . . . . . . . . . 116
6E(i) 1
2
convergence . . . . . . . . . . . . . . . . . . . . . . . . 117
CONTENTS vii
6E(ii) Pointwise convergence . . . . . . . . . . . . . . . . . . . . 120
6E(iii) Uniform convergence . . . . . . . . . . . . . . . . . . . . 123
6E(iv) Convergence of function series . . . . . . . . . . . . . . . 129
6F Orthogonal and orthonormal Bases . . . . . . . . . . . . . . . . 131
6G Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 133
7 Fourier sine series and cosine series 137
7A Fourier (co)sine series on [0, ] . . . . . . . . . . . . . . . . . . . 137
7A(i) Sine series on [0, ] . . . . . . . . . . . . . . . . . . . . . 137
7A(ii) Cosine series on [0, ] . . . . . . . . . . . . . . . . . . . . 141
7B Fourier (co)sine series on [0, 1] . . . . . . . . . . . . . . . . . . . 144
7B(i) Sine series on [0, 1] . . . . . . . . . . . . . . . . . . . . . 144
7B(ii) Cosine series on [0, 1] . . . . . . . . . . . . . . . . . . . . 146
7C Computing Fourier (co)sine coecients . . . . . . . . . . . . . . 147
7C(i) Integration by parts . . . . . . . . . . . . . . . . . . . . . 147
7C(ii) Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . 148
7C(iii) Step functions . . . . . . . . . . . . . . . . . . . . . . . . 151
7C(iv) Piecewise linear functions . . . . . . . . . . . . . . . . . . 155
7C(v) Dierentiating Fourier (co)sine series . . . . . . . . . . . 158
7D Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
8 Real Fourier series and complex Fourier series 161
8A Real Fourier series on [, ] . . . . . . . . . . . . . . . . . . . . 161
8B Computing real Fourier coecients . . . . . . . . . . . . . . . . 163
8B(i) Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . 163
8B(ii) Step functions . . . . . . . . . . . . . . . . . . . . . . . . 164
8B(iii) Piecewise linear functions . . . . . . . . . . . . . . . . . . 166
8B(iv) Dierentiating real Fourier series . . . . . . . . . . . . . . 168
8C Relation between (co)sine series and real series . . . . . . . . . . 168
8D Complex Fourier series . . . . . . . . . . . . . . . . . . . . . . . 172
9 Multidimensional Fourier series 179
9A ...in two dimensions . . . . . . . . . . . . . . . . . . . . . . . . . 179
9B ...in many dimensions . . . . . . . . . . . . . . . . . . . . . . . . 186
9C Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 193
10 Proofs of the Fourier convergence theorems 195
10A Bessel, Riemann and Lebesgue . . . . . . . . . . . . . . . . . . . 195
10B Pointwise convergence . . . . . . . . . . . . . . . . . . . . . . . . 197
10C Uniform convergence . . . . . . . . . . . . . . . . . . . . . . . . 204
10D 1
2
convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
10D(i) Integrable functions and step functions in L
2
[, ] . . . 208
10D(ii)Convolutions and molliers . . . . . . . . . . . . . . . . . 214
10D(iii)Proof of Theorems 8A.1(a) and 10D.1. . . . . . . . . . . . 221
viii DRAFT CONTENTS
IV BVP solutions via eigenfunction expansions 224
11 Boundary value problems on a line segment 225
11A The heat equation on a line segment . . . . . . . . . . . . . . . . 225
11B The wave equation on a line (the vibrating string) . . . . . . . . 229
11C The Poisson problem on a line segment . . . . . . . . . . . . . . 235
12 Boundary value problems on a square 239
12A The Dirichlet problem on a square . . . . . . . . . . . . . . . . . 240
12B The heat equation on a square . . . . . . . . . . . . . . . . . . . 246
12B(i) Homogeneous boundary conditions . . . . . . . . . . . . 246
12B(ii) Nonhomogeneous boundary conditions . . . . . . . . . . 251
12C The Poisson problem on a square . . . . . . . . . . . . . . . . . 255
12C(i) Homogeneous boundary conditions . . . . . . . . . . . . 255
12C(ii) Nonhomogeneous boundary conditions . . . . . . . . . . 258
12D The wave equation on a square (the square drum) . . . . . . . . 259
12E Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 262
13 Boundary value problems on a cube 265
13A The heat equation on a cube . . . . . . . . . . . . . . . . . . . . 266
13B The Dirichlet problem on a cube . . . . . . . . . . . . . . . . . . 269
13C The Poisson problem on a cube . . . . . . . . . . . . . . . . . . 272
14 Boundary value problems in polar coordinates 273
14A Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
14B The Laplace equation in polar coordinates . . . . . . . . . . . . 274
14B(i) Polar harmonic functions . . . . . . . . . . . . . . . . . . 274
14B(ii) Boundary value problems on a disk . . . . . . . . . . . . 278
14B(iii)Boundary value problems on a codisk . . . . . . . . . . . 283
14B(iv)Boundary value problems on an annulus . . . . . . . . . 286
14B(v) Poissons solution to Dirichlet problem on the disk . . . . 289
14C Bessel functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
14C(i) Bessels equation; Eigenfunctions of in Polar Coordi-
nates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
14C(ii) Boundary conditions; the roots of the Bessel function . . 296
14C(iii)Initial conditions; Fourier-Bessel expansions . . . . . . . 296
14D The Poisson equation in polar coordinates . . . . . . . . . . . . 298
14E The heat equation in polar coordinates . . . . . . . . . . . . . . 300
14F The wave equation in polar coordinates . . . . . . . . . . . . . . 302
14G The power series for a Bessel function . . . . . . . . . . . . . . . 305
14H Properties of Bessel functions . . . . . . . . . . . . . . . . . . . . 309
14I Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 315
CONTENTS ix
15 Eigenfunction methods on arbitrary domains 317
15A General solution to Poisson, heat and wave equation BVPs . . . 317
15B General solution to Laplace equation BVPs . . . . . . . . . . . . 324
15C Eigenbases on Cartesian products . . . . . . . . . . . . . . . . . 330
15D General method for solving I/BVPs . . . . . . . . . . . . . . . . 337
15E Eigenfunctions of self-adjoint operators . . . . . . . . . . . . . . 340
V Miscellaneous solution methods 351
16 Separation of variables 353
16A ...in Cartesian coordinates on 1
2
. . . . . . . . . . . . . . . . . . 353
16B ...in Cartesian coordinates on 1
D
. . . . . . . . . . . . . . . . . . 355
16C ...in polar coordinates: Bessels equation . . . . . . . . . . . . . . 357
16D ...in spherical coordinates: Legendres equation . . . . . . . . . . 359
16E Separated vs. quasiseparated . . . . . . . . . . . . . . . . . . . . 369
16F The polynomial formalism . . . . . . . . . . . . . . . . . . . . . 369
16G Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
16G(i) Boundedness . . . . . . . . . . . . . . . . . . . . . . . . . 372
16G(ii)Boundary conditions . . . . . . . . . . . . . . . . . . . . . 373
17 Impulse-response methods 375
17A Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
17B Approximations of identity . . . . . . . . . . . . . . . . . . . . . 379
17B(i) ...in one dimension . . . . . . . . . . . . . . . . . . . . . 379
17B(ii) ...in many dimensions . . . . . . . . . . . . . . . . . . . . 383
17C The Gaussian convolution solution (heat equation) . . . . . . . . 385
17C(i) ...in one dimension . . . . . . . . . . . . . . . . . . . . . 385
17C(ii) ...in many dimensions . . . . . . . . . . . . . . . . . . . . 392
17D dAlemberts solution (one-dimensional wave equation) . . . . . 393
17D(i) Unbounded domain . . . . . . . . . . . . . . . . . . . . . 393
17D(ii)Bounded domain . . . . . . . . . . . . . . . . . . . . . . . 399
17E Poissons solution (Dirichlet problem on half-plane) . . . . . . . 403
17F Poissons solution (Dirichlet problem on the disk) . . . . . . . . 406
17G
Properties of convolution . . . . . . . . . . . . . . . . . . . . . . 409

17H Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 411
18 Applications of complex analysis 415
18A Holomorphic functions . . . . . . . . . . . . . . . . . . . . . . . 415
18B Conformal maps . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
18C Contour integrals and Cauchys Theorem . . . . . . . . . . . . . 434
18D Analyticity of holomorphic maps . . . . . . . . . . . . . . . . . . 449
18E Fourier series as Laurent series . . . . . . . . . . . . . . . . . . . 454
18F
Abel means and Poisson kernels . . . . . . . . . . . . . . . . . . 461

x DRAFT CONTENTS
18G Poles and the residue theorem . . . . . . . . . . . . . . . . . . . 464
18H Improper integrals and Fourier transforms . . . . . . . . . . . . 472
18I

Homological extension of Cauchys theorem . . . . . . . . . . . . 481
VI Fourier transforms on unbounded domains 485
19 Fourier transforms 487
19A One-dimensional Fourier transforms . . . . . . . . . . . . . . . . 487
19B Properties of the (one-dimensional) Fourier transform . . . . . . 492
19C
Parseval and Plancherel . . . . . . . . . . . . . . . . . . . . . . 502

19D Two-dimensional Fourier transforms . . . . . . . . . . . . . . . . 504
19E Three-dimensional Fourier transforms . . . . . . . . . . . . . . . 506
19F Fourier (co)sine Transforms on the half-line . . . . . . . . . . . . 510
19G
Momentum representation & Heisenberg uncertainty . . . . . . 511

19H
Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . . . 515

19I Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 523
20 Fourier transform solutions to PDEs 527
20A The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . 527
20A(i) Fourier transform solution . . . . . . . . . . . . . . . . . 527
20A(ii) The Gaussian convolution formula, revisited . . . . . . . 530
20B The wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . 531
20B(i) Fourier transform solution . . . . . . . . . . . . . . . . . 531
20B(ii) Poissons spherical mean solution; Huygens principle . . 534
20C The Dirichlet problem on a half-plane . . . . . . . . . . . . . . . 537
20C(i) Fourier solution . . . . . . . . . . . . . . . . . . . . . . . 538
20C(ii) Impulse-response solution . . . . . . . . . . . . . . . . . . 539
20D PDEs on the half-line . . . . . . . . . . . . . . . . . . . . . . . . 540
20E General solution to PDEs using Fourier transforms . . . . . . . . 540
20F Practice problems . . . . . . . . . . . . . . . . . . . . . . . . . . 542
0 Appendices 545
0A Sets and functions . . . . . . . . . . . . . . . . . . . . . . . . . . 545
0B Derivatives notation . . . . . . . . . . . . . . . . . . . . . . . . 549
0C Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 551
0D Coordinate systems and domains . . . . . . . . . . . . . . . . . . 553
0D(i) Rectangular coordinates . . . . . . . . . . . . . . . . . . 554
0D(ii) Polar coordinates on 1
2
. . . . . . . . . . . . . . . . . . . 554
0D(iii) Cylindrical coordinates on 1
3
. . . . . . . . . . . . . . . 555
0D(iv) Spherical coordinates on 1
3
. . . . . . . . . . . . . . . . 556
0D(v) What is a domain ? . . . . . . . . . . . . . . . . . . . . . 556
0E Vector calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
0E(i) Gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . 557
CONTENTS xi
0E(ii) Divergence . . . . . . . . . . . . . . . . . . . . . . . . . . 558
0E(iii) The Divergence Theorem. . . . . . . . . . . . . . . . . . . 561
0F Dierentiation of function series . . . . . . . . . . . . . . . . . . 565
0G Dierentiation of integrals . . . . . . . . . . . . . . . . . . . . . 567
0H Taylor polynomials . . . . . . . . . . . . . . . . . . . . . . . . . 568
0H(i) ...in one dimension . . . . . . . . . . . . . . . . . . . . . . 568
0H(ii) ...and Taylor series . . . . . . . . . . . . . . . . . . . . . . 569
0H(iii) ...to solve ordinary dierential equations . . . . . . . . . . 571
0H(iv) ...in two dimensions . . . . . . . . . . . . . . . . . . . . . 575
0H(v) ...in many dimensions . . . . . . . . . . . . . . . . . . . . 576
Bibliography 577
Index 580
Notation 593
xii DRAFT Preface
Preface
This is a textbook for an introductory course on linear partial dierential equa-
tions (PDEs) and initial/boundary value problems (I/BVPs). It also provides a
mathematically rigorous introduction to Fourier analysis (Chapters 7, 8, 9, 10,
and 19), which is the main tool used to solve linear PDEs in Cartesian coordi-
nates. Finally, it introduces basic functional analysis (Chapter 6) and complex
analysis (Chapter 18). The rst is necessary to rigorously characterize the con-
vergence of Fourier series, and also to discuss eigenfunctions for linear dierential
operators. The second provides powerful techniques to transform domains and
compute integrals, and also oers additional insight into Fourier series.
This book is not intended to be comprehensive or encyclopaedic. It is de-
signed for a one-semester course (i.e. 36-40 hours of lectures), and it is therefore
strictly limited in scope. First, it deals mainly with linear PDEs with con-
stant coecients. Thus, there is no discussion of characteristics, conservation
laws, shocks, variational techniques, or perturbation methods, which would be
germane to other types of PDEs. Second, the book focus mainly on concrete
solution methods to specic PDEs (e.g. the Laplace, Poisson, Heat, Wave, and
Schrodinger equations) on specic domains (e.g. line segments, boxes, disks, an-
nuli, spheres), and spends rather little time on qualitative results about entire
classes of PDEs (e.g elliptic, parabolic, hyperbolic) on general domains. Only
after a thorough exposition of these special cases does the book sketch the gen-
eral theory; experience shows that this is far more pedagogically eective then
presenting the general theory rst. Finally, the book does not deal at all with
numerical solutions or Galerkin methods.
One slightly unusual feature of this book is that, from the very beginnning,
it emphasizes the central role of eigenfunctions (of the Laplacian) in the solu-
tion methods for linear PDEs. Fourier series and Fourier-Bessel expansions are
introduced as the orthogonal eigenfunction expansions which are most suitable
in certain domains or coordinate systems. Separation of variables appears rela-
tively late in the exposition (Chapter 16), as a convenient device to obtain such
eigenfunctions. The only techniques in the book which are not either implicitly
or explicitly based on eigenfunction expansions are impulse-response functions
and Greens functions (Chapter 17) and complex-analytic methods (Chapter 18).
Prerequisites and intended audience. This book is written for third-year
undergraduate students in mathematics, physics, engineering, and other math-
ematical sciences. The only prererequisites are (1) multivariate calculus (i.e.
partial derivatives, multivariate integration, changes of coordinate system) and
(2) linear algebra (i.e. linear operators and their eigenvectors).
It might also be helpful for students to be familiar with: (1) the basic the-
ory of ordinary dierential equations (specically: Laplace transforms, Frobe-
nius method); (2) some elementary vector calculus (specically: divergence and
Preface xiii
gradient operators); and (3) elementary physics (to understand the physical mo-
tivation behind many of the problems). However, none of these three things are
really required.
In addition to this background knowledge, the book requires some ability at
abstract mathematical reasoning. Unlike some applied math texts, we do not
suppress or handwave the mathematical theory behind the solution methods.
At suitable moments, the exposition introduces concepts like orthogonal basis,
uniform convergence vs. L
2
-convergence, eigenfunction expansion, and self-
adjoint operator; thus, students must be intellectually capable of understanding
abstract mathematical concepts of this nature. Likewise, the exposition is mainly
organized in a denition theorem proof example format, rather than
a problem solution format. Students must be able to understand abstract
descriptions of general solution techniques, rather than simply learn by imitating
worked solutions to special cases.
Acknowledgements. I would like to thank Xiaorang Li of Trent University,
who read through an early draft of this book and made many helpful suggestions
and corrections, and who also provided questions #6 and #7 on page 101, and
also question # 8 on page 135. I also thank Peter Nalitolela, who proofread
a penultimate draft and spotted many mistakes. I would like to thank several
anonymous reviewers who made many useful suggestions, and I would also like
to thank Peter Thompson of Cambridge University Press for recruiting these
referees. I also thank Diana Gillooly of Cambridge University Press, who was
very supportive and helpful throughout the entire publication process, especially
concerning my desire to provide a free online version of the book, and to release
the gures and problem sets under a Creative Commons license. I also thank
the many students who used the early versions of this book, especially those who
found mistakes or made good suggestions. Finally, I thank George Peschke of
the University of Alberta, for being an inspiring example of good mathematical
pedagogy.
None of these people are responsible for any remaining errors, omissions, or
other aws in the book (of which there are no doubt many). If you nd an error
or some other deciency in the book, please contact me at
marcuspivato@trentu.ca
This book would not have been possible without open source software. The
book was prepared entirely on the Linux operating system (initially RedHat
1
,
and later Ubuntu
2
). All the text is written in Leslie Lamports L
A
T
E
X2e typeset-
ting language
3
, and was authored using Richard Stallmans Emacs editor
4
. The
illustrations were hand-drawn using William Chia-Wei Chengs excellent TGIF
1
http://www.redhat.com
2
http://www.ubuntu.com
3
http://www.latex-project.org
4
http://www.gnu.org/software/emacs/emacs.html
xiv DRAFT Preface
object-oriented drawing program
5
. Additional image manipulation and post-
processing was done with GNU Image Manipulation Program (GIMP)
6
.
Many of the plots were created using GnuPlot
7
.
8
I would like to take this
opportunity to thank the many people in the open source community who have
developed this software.
Finally and most importantly, I would like to thank my beloved wife and
partner, Reem Yassawi, and our wonderful children, Leila and Aziza, for their
support and for their patience with my many long absences.
5
http://bourbon.usc.edu:8001/tgif
6
http://www.gimp.org
7
http://www.gnuplot.info
8
Many other plots were generated using Waterloo Maple (http://www.maplesoft.com),
which unfortunately is not open-source.
Whats good about this book? xv
Whats good about this book?
This text has many advantages over most other introductions to partial dier-
ential equations.
Illustrations. PDEs are physically motivated and geometrical objects; they
describe curves, surfaces and scalar elds with special geometric properties, and
the way these entities evolve over time under endogenous dynamics. To under-
stand PDEs and their solutions, it is necessary to visualize them. Algebraic
formulae are just a language used to communicate such visual ideas in lieu of
pictures, and they generally make a poor substitute. This book has over 300
high-quality illustrations, many of which are rendered in three dimensions. In
the online version of the book, most of these illustrations appear in full colour.
Also, the website contains many animations which do not appear in the printed
book.
Most importantly, on the book website, all illustrations are freely available
under a Creative Commons Attribution Noncommercial Share-Alike License
9
.
This means that you are free to download, modify, and utilize the illustrations
to prepare your own course materials (e.g. printed lecture notes or beamer
presentations), as long as you attribute the original author. Please visit
<http://xaravve.trentu.ca/pde
Physical motivation. Connecting the math to physical reality is critical: it
keeps students motivated, and helps them interpret the mathematical formalism
in terms of their physical intuitions about diusion, vibration, electrostatics,
etc. Chapter 1 of this book discusses the physics behind the Heat, Laplace,
and Poisson equations, and Chapter 2 discusses the wave equation. An unusual
addition to this text is Chapter 3, which discusses quantum mechanics and the
Schrodinger equation (one of the major applications of PDE theory in modern
physics).
Detailed syllabus. Dicult choices must be made when turning a 600+ page
textbook into a feasible twelve-week lesson plan. It is easy to run out of time
or inadvertently miss something important. To facilitate this task, this book
provides a lecture-by-lecture breakdown of how the author covers the material
(page xxi). Of course, each instructor can diverge from this syllabus to suit the
interests/background of her students, a longer/shorter teaching semester, or her
personal taste. But the prefabricated syllabus provides a base to work from, and
will save most instructors a lot of time and aggravation.
9
See http://creativecommons.org/licenses/by-nc-sa/3.0.
xvi DRAFT Whats good about this book?
Explicit prerequisites for each chapter and section. To save time, an instructor
might want to skip a certain chapter or section, but she worries that it may end
up being important later. We resolve this problem in two ways. First, page
(iv) provides a Chapter Dependency Lattice, which summarises the large-scale
structure of logical dependencies between the chapters of the book. Second,
every section of every chapter begins with an explicit list of required and rec-
ommended prerequisite sections; this provides more detailed information about
the small-scale structure of logical dependencies between sections. By tracing
backward through this lattice of dependencies, you can gure out exactly what
background material you must cover to reach a particular goal. This makes the
book especially suitable for self-study.
Flat dependency lattice. There are many paths through the twenty-chapter
Dependency Lattice on page (iv), every one of which is only seven chapters or
less in length. Thus, an instructor (or an autodidact) can design many possible
syllabi, depending on her interests, and can quickly move to advanced mate-
rial. The Recommended Syllabus on page (xxi) describes a gentle progression
through the material, covering most of the core topics in a 12 week semester,
emphasizing concrete examples and gradually escalating the abstraction level.
The Chapter Dependency Lattice suggests some other possibilities for acceler-
ated syllabi focusing on dierent themes:
Solving PDEs with impulse response functions. Chapters 1, 2, 5 and 17
only.
Solving PDEs with Fourier transforms. Chapters 1, 2, 5, 19, and 20 only.
(Pedagogically speaking, Chapters 8 and 9 will help the student understand Chap-
ter 19, and Chapters 11-13 will help the student understand Chapter 20. Also, it
is interesting to see how the impulse-response methods of Chapter 17 yield the
same solutions as the Fourier methods of Chapter 20, using a totally dierent
approach. However, strictly speaking, none of Chapters 8-13 or 17 is logically
necessary.)
Solving PDEs with separation of variables. Chapters 1, 2 and 16 only.
(However, without at least Chapters 12 and 14, the ideas of Chapter 16 will seem
somewhat articial and pointless.)
Solving I/BVPs using eigenfunction expansions. Chapters 1, 2, 4, 5, 6,
and 15.
(It would be pedagogically better to also cover Chapters 9 and 12, and probably
Chapter 14. But strictly speaking, none of these is logically necessary.)
Tools for quantum mechanics. Section 1B, then Chapters 3, 4, 6, 9, 13,
19, and 20 (skipping material on Laplace, Poisson, and wave equations in
Whats good about this book? xvii
Chapters 13 and 20, and adapting the solutions to the heat equation into
solutions to the Schrodinger equation.)
Fourier theory. Section 4A, then Chapters 6, 7, 8, 9, 10, and 19. Finally,
Sections 18A, 18C, 18E and 18F provide a complex perspective. (Section
18H also contains some useful computational tools).
Crash course in complex analysis. Chapter 18 is logically independent of
the rest of the book, and rigorously develops the main ideas in complex
analysis from rst principles. (However, the emphasis is on applications
to PDEs and Fourier theory, so some of the material may seem esoteric or
unmotivated if read in isolation from other chapters.)
Highly structured exposition, with clear motivation up front. The exposition
is broken into small, semi-independent logical units, each of which is clearly
labelled, and which has a clear purpose or meaning which is made immediately
apparent. This simplies the instructors task; she doesnt need to spend time
restructuring and summarizing the text material, because it is already structured
in a manner which self-summarizes. Instead, instructors can focus more on
explanation, motivation, and clarication.
Many practice problems (with complete solutions and source code available on-
line). Frequent evaluation is critical to reinforce material taught in class. This
book provides an extensive supply of (generally simple) computational Practice
Problems at the end of each chapter. Completely worked solutions to virtually
all of these problems are available on the book website. Also on the book web-
site, the L
A
T
E
X source code for all problems and solutions is freely available under
a Creative Commons Attribution Noncommercial Share-Alike License
10
. Thus,
an instructor can download and edit this source code, and easily create quizzes,
assignments, and matching solutions for her students.
Challenging exercises without solutions. Complex theoretical concepts can-
not really be tested in quizzes, and do not lend themselves to canned practice
problems. For a more theoretical course with more mathematically sophisti-
cated students, the instructor will want to assign some proof-related exercises
for homework. This book has more than 420 such exercises scattered through-
out the exposition; these are agged by an E _ symbol in the margin, as shown E _
here. Many of these exercises ask the student to prove a major result from the
text (or a component thereof). This is the best kind of exercise, because it rein-
forces the material taught in class, and gives students a sense of ownership of the
mathematics. Also, students nd it more fun and exciting to prove important
theorems, rather than solving esoteric make-work problems.
10
See http://creativecommons.org/licenses/by-nc-sa/3.0.
xviii DRAFT Whats good about this book?
Appropriate rigour. The solutions of PDEs unfortunately involve many tech-
nicalities (e.g. dierent forms of convergence; derivatives of innite function
series, etc.). It is tempting to handwave and gloss over these technicalities, to
avoid confusing students. But this kind of pedagogical dishonesty actually makes
students more confused; they know something is shy, but they cant tell quite
what. Smarter students know they are being misled, and may lose respect for
the book, the instructor, or even the whole subject.
In contrast, this book provides a rigorous mathematical foundation for all its
solution methods. For example, Chapter 6 contains a careful explanation of 1
2
spaces, the various forms of convergence for Fourier series, and the dierences
between them including the pathologies which can arise when one is careless
about these issues. I adopt a triage approach to proofs: The simplest proofs are
left as exercises for the motivated student (often with a step-by-step breakdown
of the best strategy). The most complex proofs I have omitted, but I provide
multiple references to other recent texts. In between are those proofs which are
challenging but still accessible; I provide detailed expositions of these proofs.
Often, when the text contains several variants of the same theorem, I prove one
variant in detail, and leave the other proofs as exercises.
Appropriate Abstraction. It is tempting to avoid abstractions (e.g. linear
dierential operators, eigenfunctions), and simply present ad hoc solutions to
special cases. This cheats the student. The right abstractions provide simple yet
powerful tools which help students understand a myriad of seemingly disparate
special cases within a single unifying framework. This book provides students
with the opportunity to learn an abstract perspective once they are ready for it.
Some abstractions are introduced in the main exposition, others are in optional
sections, or in the philosophical preambles which begin each major part of the
book.
Gradual abstraction. Learning proceeds from the concrete to the abstract.
Thus, the book begins each topic with a specic example or a low-dimensional
formulation, and only later proceed to a more general/abstract idea. This intro-
duces a lot of redundancy into the text, in the sense that later formulations
subsume the earlier ones. So the exposition is not as ecient as it could be.
This is a good thing. Eciency makes for good reference books, but lousy texts.
For example, when introducing the heat equation, Laplace equation, and
wave equation in Chapters 1 and 2, I rst derive and explain the one-dimensional
version of each equation, then the two-dimensional version, and nally, the gen-
eral, 1-dimensional version. Likewise, when developing the solution methods for
BVPs in Cartesian coordinates (Chapters 11-13), I conne the exposition to the
interval [0, ], the square [0, ]
2
and the cube [0, ]
3
, and assume all the coe-
cients in the dierential equations are unity. Then the exercises ask the student
to state and prove the appropriate generalization of each solution method for an
Whats good about this book? xix
interval/rectangle/box of arbitrary dimensions, and for equations with arbitrary
coecients. The general method for solving I/BVPs using eigenfunction expan-
sions only appears in Chapter 15, after many special cases of this method have
been thoroughly exposited in Cartesian and polar coordinates (Chapters 11-14).
Likewise, the development of Fourier theory proceeds in gradually escalating
levels of abstraction. First we encounter Fourier (co)sine series on the interval
[0, ] (7A); then on the interval [0, 1] for arbitrary 1 0 (7B). Then Chapter
8 introduces real Fourier series (i.e. with both sine and cosine terms) and
then complex Fourier series (8D). Then, in Chapter 9 introduce 2-dimensional
(co)sine series, and nally, 1-dimensional (co)sine series.
Expositional clarity. Computer scientists have long known that it is easy to
write software that works, but it is much more dicult (and important) to write
working software that other people can understand. Similarly, it is relatively easy
to write formally correct mathematics; the real challenge is to make the math-
ematics easy to read. To achieve this, I use several techniques. I divide proofs
into semi-independent modules (claims), each of which performs a simple,
clearly-dened task. I integrate these modules together in an explicit hierar-
chical structure (with subclaims inside of claims), so that their functional
interdependence is clear from visual inspection. I also explain formal steps with
parenthetical heuristic remarks. For example, in a long string of (in)equalities,
I often attach footnotes to each step, as follows:
A
()
B
()
C <
()
D. Here, () is because [...]; () follows from [...], and () is because [...].
Finally, I use letters from the same lexicographical family to denote objects
which belong together. For example: If o and T are sets, then elements of o
should be :
1
, :
2
, :
3
, . . ., while elements of T are t
1
, t
2
, t
3
, . . .. If v is a vector,
then its entries should be
1
, . . . ,
N
. If A is a matrix, then its entries should
be o
11
, . . . , o
NM
. I reserve upper-case letters (e.g. J, 1, 1, `, , . . .) for the
bounds of intervals or indexing sets, and then use the corresponding lower-case
letters (e.g. ,, /, |, :, n, . . .) as indexes. For example, n 1, 2, . . . , ,
n
:=
J
j=1
K
k=1
o
n
jk
.
Clear and explicit statements of solution techniques. Many PDEs text con-
tain very few theorems; instead they try to develop the theory through a long
sequence of worked examples, hoping that students will learn by imitation, and
somehow absorb the important ideas by osmosis. However, less gifted students
often just imitate these worked examples in a slavish and uncomprehending way.
Meanwhile, the more gifted students do not want to learn by osmosis; they
want clear and precise statements of the main ideas.
The problem is that most solution methods in PDEs, if stated as theorems
in full generality, are incomprehensible to many students (especially the non-
math majors). My solution is this: I provide explicit and precise statements of
the solution-method for almost every possible combination of (1) several major
xx DRAFT Whats good about this book?
PDEs, (2) several kinds of boundary conditions, and (3) several dierent do-
mains. I state these solutions as theorems, not as worked examples. I follow
each of these theorems with several completely worked examples. Some theo-
rems I prove, but most of the proofs are left as exercises (often with step-by-step
hints).
Of course, this approach is highly redundant, because I end up stating more
than twenty theorems which are all really special cases of three or four gen-
eral results (for example, the general method for solving the heat equation on a
compact domain with Dirichlet boundary conditions, using an eigenfunction ex-
pansion). However, this sort of redundancy is good in an elementary exposition.
Highly ecient expositions are pleasing to our aesthetic sensibilities, but they
are dreadful for pedagogical purposes.
However, one must not leave the students with the impression that the the-
ory of PDEs is a disjointed collection of special cases. To link together all the
homogeneous Dirichlet heat equation theorems, for example, I explicitly point
out that they all utilize the same underlying strategy. Also, when a proof of
one variant is left as an exercise, I encourage students to imitate the (provided)
proofs of previous variants. When the students understand the underlying simi-
larity between the various special cases, then it is appropriate to state the general
solution. The students will almost feel they have gured it out for themselves,
which is the best way to learn something.
Suggested Twelve-Week Syllabus xxi
Suggested Twelve-Week Syllabus
Week 1: Heat and Diusion-related PDEs
Lecture 1: 0A-0E Review of multivariate calculus; intro. to complex numbers
Lecture 2: 1A-1B Fouriers Law; The heat equation
Lecture 3: 1C-1D Laplace Equation; Poissons Equation
Week 2: Wave-related PDEs; Quantum Mechanics
Lecture 1: 1E; 2A Properties of harmonic functions; Spherical Means
Lecture 2: 2B-2C wave equation; telegraph equation
Lecture 3: Chap.3 The Schrodinger equation in quantum mechanics
Week 3: General Theory
Lecture 1: 4A - 4C Linear PDEs: homogeneous vs. nonhomogeneous
Lecture 2: 5A; 5B, Evolution equations & Initial Value Problems
Lecture 3: 5C Boundary conditions and boundary value problems
Week 4: Background to Fourier Theory
Lecture 1: 5D Uniqueness of solutions to BVPs; 6A Inner products
Lecture 2: 6B-6D 1
2
space; Orthogonality
Lecture 3: 6E(a,b,c) 1
2
convergence; Pointwise convergence; Uniform Con-
vergence
Week 5: One-dimensional Fourier Series
Lecture 1: 6E(d) Innite Series; 6F Orthogonal bases
7A Fourier (co/sine) Series: Denition and examples
Lecture 2: 7C(a,b,c,d,e) Computing Fourier series of polynomials, piecewise
linear and step functions
Lecture 3: 11A-11C Solution to heat equation & Poisson equation on line
segment.
Week 6: Fourier Solutions for BVPs in One and Two dimensions
Lecture 1: 11B- 12A; wave equation on line segment & Laplace equation on
a square.
Lecture 2: 9A-9B Multidimensional Fourier Series.
Lecture 3: 12B- 12C(i) Solution to heat equation & Poisson equation on a
square
Week 7: Fourier solutions for 2-dimensional BVPs in Cartesian & Polar Coordinates
Lecture 1: 12C(ii), 12D Solution to Poisson equation & wave equation on a
square
Lecture 2: 5C(iv); 8A-8B Periodic Boundary Conditions; Real Fourier Se-
ries.
xxii DRAFT Suggested Twelve-Week Syllabus
Lecture 3: 14A; 14B(a,b,c,d) Laplacian in Polar coordinates; Laplace Equa-
tion on (co)Disk.
Week 8: BVPs in Polar Coordinates; Bessel functions
Lecture 1: 14C Bessel Functions.
Lecture 2: 14D-14F Heat, Poisson, and wave equations in Polar coordinates.
Lecture 3: 14G Solving Bessels equation with the Method of Frobenius.
Week 9: Eigenbases; Separation of variables.
Lecture 1: 15A-15B Eigenfunction solutions to BVPs
Lecture 2: 15B; 16A-16B Harmonic bases. Separation of Variables in Carte-
sian coordinates.
Lecture 3: 16C-16D Separation of variables in polar and spherical coordi-
nates. Legendre Polynomials.
Week 10: Impulse Response Methods.
Lecture 1: 17A - 17C Impulse response functions; convolution. Approxima-
tions of identity. Gaussian Convolution Solution for heat equation.
Lecture 2: 17C-17F, Gaussian convolutions continued. Poissons Solutions to
Dirichlet problem on a half-plane and a disk.
Lecture 3: 14B(v); 17D Poisson solution on disk via polar coordinates; dAlembert
Solution to wave equation.
Week 11: Fourier Transforms.
Lecture 1: 19A One-dimensional Fourier Transforms.
Lecture 2: 19B Properties of one-dimensional Fourier transform.
Lecture 3: 20A ; 20C Fourier transform solution to heat equation; Dirchlet
problem on Half-plane.
Week 12: Fourier Transform Solutions to PDEs.
Lecture 1: 19D, 20B(i) Multidimensional Fourier transforms; Solution to
wave equation.
Lecture 2: 20B(ii); 20E Poissons Spherical Mean Solution; Huygens Prin-
ciple. The General Method.
Lecture 3: (Time permitting) 19G or 19H (Heisenberg Uncertainty or Laplace
transforms).
In a longer semester or a faster paced course, one could also cover parts of Chapter 10
(Proofs of Fourier Convergence) and/or Chapter 18 (Applications of Complex Analysis)
1
I Motivating examples and
major applications
A dynamical system is a mathematical model of a system evolving in time.
Most models in mathematical physics are dynamical systems. If the system has
only a nite number of state variables, then its dynamics can be encoded in
an ordinary dierential equation (ODE), which expresses the time derivative of
each state variable (i.e. its rate of change over time) as a function of the other
state variables. For example, celestial mechanics concerns the evolution of a
system of gravitationally interacting objects (e.g. stars and planets). In this
case, the state variables are vectors encoding the position and momentum of
each object, and the ODE describe how the objects move and accelerate as they
gravitationally interact.
However, if the system has a very large number of state variables, then it
is no longer feasible to represent it with an ODE. For example, consider the
ow of heat or the propagation of compression waves through a steel bar con-
taining 10
24
iron atoms. We could model this using a 10
24
-dimensional ODE,
where we explicitly track the vibrational motion of each iron atom. However,
such a microscopic model would be totally intractable. Furthermore, it isnt
necessary. The iron atoms are (mostly) immobile, and interact only with their
immediate neighbours. Furthermore, nearby atoms generally have roughly the
same temperature, and move in synchrony. Thus, it suces to consider the
macroscopic temperature distribution of the steel bar, or study the uctuation
of a macroscopic density eld. This temperature distribution or density eld
can be mathematically represented as a smooth, real-valued function over some
three-dimensional domain. The ow of heat or the propagation of sound can
then be described as the evolution of this function over time.
We now have a dynamical system where the state variable is not a nite sys-
tem of vectors (as in celestial mechanics), but is instead a multivariate function.
The evolution of this function is determined by its spatial geometry e.g. the
local steepness and variation of the temperature gradients between warmer and
cooler regions in the bar. In other words, the time derivative of the function (its
rate of change over time) is determined by its spatial derivatives (which describe
its slope and curvature at each point in space). An equation which relates the
dierent derivatives of a multivariate function in this way is a partial dieren-
tial equation (PDE). In particular, a PDE which describes a dynamical system
is called an evolution equation. For example, the evolution equation which de-
scribes the ow of heat through a solid is called the heat equation. The equation
which describes compression waves is the wave equation.
An equilibrium of a dynamical system is a state which is unchanging over time;
mathematically, this means that the time-derivative is equal to zero. An equlib-
2 DRAFT
rium of an -dimensional evolution equation satises an ( 1)-dimensional
PDE called an equilibrium equation. For example, the equilibrium equations
corresponding to the heat equation are the Laplace equation and the Poisson
equation (depending on whether or not the system is subjected to external heat
input).
PDEs are thus of central importance in the thermodynamics and acoustics
of continuous media (e.g. steel bars). The heat equation also describes chemical
diusion in uids, and also the evolving probability distribution of a particle
performing a random walk called Brownian motion. It thus nds applications
everywhere from mathematical biology to mathematical nance. When diusion
or Brownian motion is combined with deterministic drift (e.g. due to prevailing
wind or ocean currents) it becomes a PDE called the Fokker-Planck equation.
The Laplace and Poisson equations describe the equilibria of such diusion
processes. They also arise in electrostatics, where they describe the shape of
an electric eld in a vacuum. Finally, solutions of the two-dimensional Laplace
equation are good approximations of surfaces trying to minimize their elastic
potential energy that is, soap lms.
The wave equation describes the resonance of a musical instrument, the
spread of ripples on a pond, seismic waves propagating through the earths crust,
and shockwaves in solar plasma. (The motion of uids themselves is described by
yet another PDE, the Navier-Stokes equation). A version of the wave equation
arises as a special case of Maxwells equations of electrodynamics; this led to
Maxwells prediction of electromagnetic waves, which include radio, microwaves,
X-rays, and visible light. When combined with a diusion term reminiscent
of the heat equation, the wave equation becomes the telegraph equation, which
describes the propagation and degradation of electrical signals travelling through
a wire.
Finally, an odd-looking complex version of the heat equation induces wave-
like evolution in the complex-valued probability elds which describe the position
and momentum of subatomic particles. This Schrodinger equation is the starting
point of quantum mechanics, one of the two most revolutionary developments
in physics in the twentieth century. The other revolutionary development was
relativity theory. General relativity represents spacetime as a four-dimensional
manifold, whose curvature interacts with the spatiotemporal ow of mass/energy
through yet another PDE: the Einstein equation.
Except for the Einstein and Navier-Stokes equations, all the equations we
have mentioned are linear PDEs. This means that a sum of two or more solutions
to the PDE will also be a solution. This allows us to solve linear PDEs through
the method of superposition: we build complex solutions by adding together
many simple solutions. A particularly convenient class of simple solutions are
eigenfunctions. Thus, an enormously powerful and general method for linear
PDEs is to represent the solutions using eigenfunction expansions. The most
natural eigenfunction expansion (in Cartesian coordinates) is the Fourier series.
3
Chapter 1
Heat and diusion
The dierential equations of the propagation of heat express the most general conditions,
and reduce the physical questions to problems of pure analysis, and this is the proper object of
theory. Jean Joseph Fourier
1A Fouriers law
Prerequisites: 0A. Recommended: 0E.
1A(i) ...in one dimension
Figure 1A.1 depicts a material diusing through a one-dimensional domain X
(for example, X = 1 or X = [0, 1]). Let n(r, t) be the density of the material at
the point r X at time t 0. Intuitively, we expect the material to ow from
regions of greater to lesser concentration. In other words, we expect the ow of
the material at any point in space to be proportional to the slope of the curve
n(r, t) at that point. Thus, if 1(r, t) is the ow at the point r at time t, then
Greyscale
Landscape
u
(
x
)
x
Figure 1A.1: Fouriers Law of Heat Flow in one dimension
4 DRAFT Chapter 1. Heat and diusion
Grayscale Temperature
Distribution
Isothermal Contours Heat Flow Vector Field
Figure 1A.2: Fouriers Law of Heat Flow in two dimensions
we expect:
1(r, t) =
x
n(r, t)
where 0 is a constant measuring the rate of diusion. This is an example of
Fouriers Law.
1A(ii) ...in many dimensions
Prerequisites: 0E.
Figure 1A.2 depicts a material diusing through a two-dimensional domain
X 1
2
(e.g. heat spreading through a region, ink diusing in a bucket of water,
etc.). We could just as easily suppose that X 1
D
is a 1-dimensional domain.
If x X is a point in space, and t 0 is a moment in time, let n(x, t) denote
the concentration at x at time t. (This determines a function n : X1
,
1,
called a time-varying scalar eld.)
Now let

F(x, t) be a 1-dimensional vector describing the ow of the material
at the point x X. (This determines a time-varying vector eld

F : 1
D
1
,

1
D
.)
Again, we expect the material to ow from regions of high concentration to
low concentration. In other words, material should ow down the concentration
gradient. This is expressed by Fouriers Law of Heat Flow , which says:
F = n,
where 0 is is a constant measuring the rate of diusion.
One can imagine n as describing a distribution of highly antisocial people;
each person is always eeing everyone around them and moving in the direction
with the fewest people. The constant measures the average walking speed of
these misanthropes.
1B. The heat equation 5
t=4
t=5
t=6
t=7
u
(
x
,
0
)
u
(
x
,
1
)
u
(
x
,
2
)
u
(
x
,
3
)
t=0
t=1
t=2
t=3
x
x
x
x
x
x
x
x
u
(
x
,
4
)
u
(
x
,
5
)
u
(
x
,
6
)
u
(
x
,
7
)
Figure 1B.1: The heat equation as erosion.
1B The heat equation
Recommended: 1A.
1B(i) ...in one dimension
Prerequisites: 1A(i).
Consider a material diusing through a one-dimensional domain X (for ex-
ample, X = 1 or X = [0, 1]). Let n(r, t) be the density of the material at the
location r X at time t 1
,
, and let 1(r, t) be the ux of the material at the
location r and time t. Consider the derivative
x
1(r, t). If
x
1(r, t) 0, this
means that the ow is diverging
1
at this point in space, so the material there
is spreading farther apart. Hence, we expect the concentration at this point to
decrease. Conversely, if
x
1(r, t) < 0, then the ow is converging at this point
in space, so the material there is crowding closer together, and we expect the
concentration to increase. To be succinct: the concentration of material will
increase in regions where 1 converges, and decrease in regions where 1 diverges.
The equation describing this is:
t
n(r, t) =
x
1(r, t).
If we combine this with Fouriers Law, however, we get:
t
n(r, t) =
x
x
n(r, t),
which yields the one-dimensional heat equation:
t
n(r, t) =
2
x
n(r, t).
1
See 0E(ii) on page 558 for an explanation of why we say the ow is diverging here.
Heuristically speaking, if we imagine n(r, t) as the height of some one-dimensional
landscape, then the heat equation causes this landscape to be eroded, as if
it were subjected to thousands of years of wind and rain (see Figure 1B.1).
A) Low Frequency:
Slow decay
B) High Frequency:
Fast decay
T
i
m
e
Figure 1B.2: Under the heat equation, the exponential decay of a periodic func-
tion is proportional to the square of its frequency.
Example 1B.1. For simplicity we suppose = 1.
(a) Let n(r, t) = c
9t
sin(3r). Thus, n describes a spatially sinusoidal function
(with spatial frequency 3) whose magnitude decays exponentially over time.
(b) The dissipating wave: More generally, let n(r, t) = c
2
t
sin( r).
Then n is a solution to the one-dimensional heat equation, and looks like a
standing wave whose amplitude decays exponentially over time (see Figure
1B.2). Notice that the decay rate of the function n is proportional to the
square of its frequency.
(c) The (one-dimensional) Gauss-Weierstrass Kernel: Let
((r; t) :=
1
2
t
exp
_
r
2
4t
_
.
Then ( is a solution to the one-dimensional heat equation, and looks like
a bell curve, which starts out tall and narrow, and over time becomes
broader and atter (Figure 1B.3).
Exercise 1B.1. Verify that the functions in Examples 1B.1(a,b,c) all satisfy the E _
heat equation.
1B. The heat equation 7
T
i
m
e
Figure 1B.3: The Gauss-Weierstrass kernel under the heat equation.
All three functions in Examples 1B.1 starts out very tall, narrow, and pointy,
and gradually become shorter, broader, and atter. This is generally what the
heat equation does; it tends to atten things out. If n describes a physical
landscape, then the heat equation describes erosion.
1B(ii) ...in many dimensions
Prerequisites: 1A(ii).
More generally, if n : 1
D
1
,
1 is the time-varying density of some
material, and

F : 1
D
1
,
1 is the ux of this material, then we would
expect the material to increase in regions where

F converges, and to decrease in
regions where

F diverges.
2
In other words, we have:
t
n = div
F.
If n is the density of some diusing material (or heat), then

F is determined by
Fouriers Law, so we get the heat equation
t
n = div n = n.
Here, is the Laplacian operator
3
, dened:
n =
2
1
n +
2
2
n +. . .
2
D
n
Exercise 1B.2. (a) If 1 = 1, and n : 1 1, verify that div n(r) = n
(r) = E _
n(r), for all r 1.
2
See 0E(ii) on page 558 for a review of the divergence of a vector eld.
3
Sometimes the Laplacian is written as
2
.
(b) If 1 = 2, and n : 1
2
1, verify that div n(r, j) =
2
x
n(r, j) +
2
y
n(r, j) =
n(r, j), for all (r, j) 1
2
.
(c) For any 1 2 and n : 1
D
1, verify that div n(x) = n(x), for all x 1
D
.
By changing to the appropriate time units, we can assume = 1, so the heat

equation becomes:
t
n = n .
For example,
If X 1, and r X, then n(r; t) =
2
x
n(r; t).
If X 1
2
, and (r, j) X, then n(r, j; t) =
2
x
n(r, j; t) +
2
y
n(r, j; t).
Thus, as weve already seen, the one-dimensional heat equation is
t
n =
2
x
n
and the the two dimensional heat equation is:
t
n(r, j; t) =
2
x
n(r, j; t) +
2
y
n(r, j; t)
Example 1B.2.
(a) Let n(r, j; t) = c
25 t
sin(3r) sin(4j). Then n is a solution to the two-
dimensional heat equation, and looks like a two-dimensional grid of si-
nusoidal hills and valleys with horizontal spacing 1,3 and vertical spacing
1,4. As shown in Figure 1B.4, these hills rapidly subside into a gently un-
dulating meadow, and then gradually sink into a perfectly at landscape.
(b) The (two-dimensional) Gauss-Weierstrass Kernel: Let
((r, j; t) :=
1
4t
exp
_
r
2
j
2
4t
_
.
Then ( is a solution to the two-dimensional heat equation, and looks like
a mountain, which begins steep and pointy, and gradually erodes into a
broad, at, hill.
(c) The 1-dimensional Gauss-Weierstrass Kernel is the function ( :
1
D
1
+
1 dened
((x; t) =
1
(4t)
D/2
exp
_
|x|
2
4t
_
Technically speaking, ((x; t) is a 1-dimensional symmetric normal proba-
bility distribution with variance = 2t.
1C. Laplaces equation 9
t = 0.00 t = 0.01 t = 0.02
0
0.5
1
y
1.5
2
2.5
3
0
0.5
1
1.5
2 x
2.5
3 -1
-0.5
0
0.5
1
0
0.5
1
y
1.5
2
2.5
3
0
0.5
1
1.5
2
x
2.5
3 -0.8
-0.4
0
0.4
0.8
-0.6
-0.4
-0.2
0
0.2
0
0.4
0.6
0.5 0
0.5 1
1 1.5
1.5
2 y
2
2.5 x
2.5
3 3
-0.3
-0.2
-0.1
0
0
0.1
0.2
0.3 0.5
0
1 0.5
1 1.5
1.5
2
2 y
2.5
2.5
x
3
3
-0.1
-0.05
0
0.05 0.1
0.5
0
1
0.5
1.5 1
1.5 2
2 2.5
2.5
3
3
t = 0.04 t = 0.08
Figure 1B.4: Five snapshots of the function n(r, j; t) = c
25 t
sin(3r) sin(4j)
from Example 1B.2.
Exercise 1B.3. Verify that the functions in Examples 1B.2(a,b,c) all satisfy the E _
heat equation.
Exercise 1B.4. Prove the Leibniz rule for Laplacians: if ), p : 1
D
1 are two E _
scalar elds, and () p) : 1
D
1 is their product, then for all x 1
D
,
() p)(x) = p(x)
_
)(x)
_
+ 2
_
)(x)
_
_
p(x)
_
+ )(x)
_
p(x)
_
.
Hint: Combine the Leibniz rules for gradients and divergences (Propositions 0E.1(b)
and 0E.2(b) on pages 558 and 560).
1C Laplaces equation
Prerequisites: 1B.
If the heat equation describes the erosion/diusion of some system, then an
equilibrium or steady-state of the heat equation is a scalar eld / : 1
D
1
satisfying Laplaces Equation:
/ 0.
2
1
0
1
2
2
1
0
1
2
2
1.5
1
0.5
0
0.5
1
0.5
0
0.5
1
x
1
0.5
0
0.5
y
1
0.5
0
0.5
1
8
6
4
2
0
2
4
6
8
x
3
2
1
0
1
2
3
y
10
5
0
5
10
(A) (B) (C)
Figure 1C.1: Three harmonic functions: (A) /(r, j) = log(r
2
+ j
2
). (B)
/(r, j) = r
2
j
2
. (C) /(r, j) = sin(r) sinh(j). In all cases, note the telltale
saddle shape.
A scalar eld satisfying the Laplace equation is called a harmonic function.
Example 1C.1.
(a) If 1 = 1, then /(r) =
2
x
/(r) = /
tt
(r); thus, the one-dimensional
Laplace equation is just
/
tt
(r) = 0
Suppose /(r) = 3r + 4. Then /
t
(r) = 3, and /
tt
(r) = 0, so / is harmonic.
More generally: the one-dimensional harmonic functions are just the linear
functions of the form: /(r) = or +/ for some constants o, / 1.
(b) If 1 = 2, then /(r, j) =
2
x
/(r, j) +
2
y
/(r, j), so the two-dimensional
Laplace equation reads:
2
x
/ +
2
y
/ = 0,
or, equivalently,
2
x
/ =
2
y
/. For example:
Figure 1C.1(B) shows the harmonic function /(r, j) = r
2
j
2
.
Figure 1C.1(C) shows the harmonic function /(r, j) = sin(r) sinh(j).
Exercise 1C.1 Verify that these two functions are harmonic. E _
The surfaces in Figure 1C.1 have a saddle shape, and this is typical of
harmonic functions; in a sense, a harmonic function is one which is saddle-
shaped at every point in space. In particular, notice that /(r, j) has no maxima
or minima anywhere; this is a universal property of harmonic functions (see
Corollary 1E.2 on page 17). The next example seems to contradict this assertion,
but in fact it doesnt...
1C. Laplaces equation 11
Example 1C.2. Figure 1C.1(A) shows the harmonic function /(r, j) = log(r
2
+
j
2
) for all (r, j) ,= (0, 0). This function is well-dened everywhere except at
(0, 0); hence, contrary to appearances, (0, 0) is not an extremal point. [Verifying
that h is harmonic is problem # 3 on page 20].
When 1 3, harmonic functions no longer dene nice saddle-shaped sur-
faces, but they still have similar mathematical properties.
Example 1C.3.
(a) If 1 = 3, then /(r, j, .) =
2
x
/(r, j, .) +
2
y
/(r, j, .) +
2
z
/(r, j, .).
Thus, the three-dimensional Laplace equation reads:
2
x
/ +
2
y
/ +
2
z
/ = 0,
For example, let /(r, j, .) =
1
|(r, j, .)|
=
1
_
r
2
+j
2
+.
2
for all
(r, j, .) ,= (0, 0, 0). Then / is harmonic everywhere except at (0, 0, 0).
[Verifying that h is harmonic is problem # 4 on page 21.]
(b) For any 1 3, the 1-dimensional Laplace equation reads:
2
1
/ +. . . +
2
D
/ = 0.
For example, let /(x) =
1
|x|
D2
=
1
_
r
2
1
+ +r
2
D
_D2
2
for all x ,= 0.
Then / is harmonic everywhere everywhere in 1
D
0 (Exercise 1C.2 E _
Verify that / is harmonic on 1
D
0.)
Harmonic functions have the convenient property that we can multiply to-
gether two lower-dimensional harmonic functions to get a higher dimensional
one. For example:
/(r, j) = r j is a two-dimensional harmonic function (Exercise 1C.3 E _
Verify this).
/(r, j, .) = r(j
2
.
2
) is a three-dimensional harmonic function (Exercise 1C.4
Verify this). E _
In general, we have the following:
Proposition 1C.4. Suppose n : 1
n
1 is harmonic and : 1
m
1 is
harmonic, and dene n : 1
n+m
1 by n(x, y) = n(x) (y) for x 1
n
and
y 1
m
. Then n is also harmonic
Proof. Exercise 1C.5 Hint: First prove that n obeys a kind of Leibniz rule: E _
n(x, y) = (y) n(x) +n(x) (y). 2
The function n(x, y) = n(x) (y) is called a separated solution, and this
theorem illustrates a technique called separation of variables. The next exercise
also explores separation of variables. A full exposition of this technique appears
in Chapter 16 on page 353.
Exercise 1C.6. (a) Let j, 1 be constants, and let )(r, j) = c
x
c
y
. Suppose E _
) is harmonic; what can you conclude about the relationship between j and ? (Justify
your assertion).
(b) Suppose )(r, j) = A(r) Y (j), where A : 1 1 and Y : 1 1 are two
smooth functions. Suppose )(r, j) is harmonic
[i] Prove that
A
(r)
A(r)
=
Y
(j)
Y (j)
for all r, j 1.
[ii] Conclude that the function
A
(r)
A(r)
must equal a constant c independent of r.
Hence A(r) satises the ordinary dierential equation A
(r) = c A(r).
Likewise, the function
Y
(j)
Y (j)
must equal c, independent of j. Hence Y (j) satises
the ordinary dierential equation Y
(j) = c Y (j).
[iii] Using this information, deduce the general form for the functions A(r) and Y (j),
and use this to obtain a general form for )(r, j).
1D The Poisson equation
Prerequisites: 1C.
Imagine j(x) is the concentration of a chemical at the point x in space.
Suppose this chemical is being generated (or depleted) at dierent rates at dif-
ferent regions in space. Thus, in the absence of diusion, we would have the
generation equation
t
j(x, t) = (x),
where (x) is the rate at which the chemical is being created/destroyed at x (we
assume that is constant in time).
If we now included the eects of diusion, we get the generation-diusion
equation:
t
j = j +.
1D. The Poisson equation 13
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
0 0.5 1 1.5 2
p"(x)
p
(
x
)
p
(
x
)
-20
-15
-10
-5
0
5
10
15
20
-3 -2 -1 0 1 2 3
p(x)
p
(x
)
p
(
x
)
p(x)
j
(r) = Q(r) =
_
1 if 0 < r < 1;
0 otherwise.
j
(r) = Q(r) = 1,r

2
;
j
(r) =
_
r if 0 < r < 1;
1 otherwise.
j
(r) = 1,r + 3;
j(r) =
_
r
2
,2 if 0 < r < 1;
r
1
2
otherwise.
j(r) = log [r[ + 3r + 5.
(A) (B)
Figure 1D.1: Two one-dimensional potentials.
A steady state of this equation is a scalar eld j satisfying Poissons Equation:
j = Q.
where Q(x) =
(x)
.
Example 1D.1: One-Dimensional Poisson Equation
If 1 = 1, then j(r) =
2
x
j(r) = j
tt
(r); thus, the one-dimensional Poisson
equation is just
j
tt
(r) = Q(r).
We can solve this equation by twice-integrating the function Q(r). If j(r) =
_ _
Q(r) is some double-antiderivative of G, then j clearly satises the Poisson
equation. For example:
(a) Suppose Q(r) =
_
1 if 0 < r < 1;
0 otherwise.
. Then dene
j(r) =
_
x
0
_
y
0
(.) d. dj =
_
_
_
0 if r < 0;
r
2
,2 if 0 < r < 1;
r
1
2
if 1 < r.
(Figure 1D.1A)
(b) If Q(r) = 1,r
2
(for r ,= 0), then j(r) =
_ _
Q(r) = log [r[ + or + /
(for r ,= 0), where o, / 1 are arbitrary constants. (see Figure 1D.1B)
10
5
0
5
10
10
5
0
5
10
2
1.5
1
0.5
0
0.5
Figure 1D.2: The two-dimensional potential eld generated by a concentration
of charge at the origin.
Exercise 1D.1. Verify that the functions j(r) in Examples (a) and (b) are both E _
solutions to their respective Poisson equations.
Example 1D.2: Electrical/Gravitational Fields
Poissons equation also arises in classical eld theory
4
. Suppose, for any point
x = (r
1
, r
2
, r
3
) in three-dimensional space, that (x) is charge density at x,
and that j(x) is the the electric potential eld at x. Then we have:
j(x) = (x) ( some constant) (1D.1)
If (x) were the mass density at x, and j(x) were the gravitational potential
energy, then we would get the same equation. (See Figure 1D.2 for an example
of such a potential in two dimensions).
In particular, in a region where there is no charge/mass (i.e. where 0),
equation (1D.1) reduces to the Laplace equation j 0. Because of this,
solutions to the Poisson equation (and especially the Laplace equation) are
sometimes called potentials.
Example 1D.3: The Coulomb Potential
Let 1 = 3, and let j(r, j, .) =
1
|(r, j, .)|
=
1
_
r
2
+j
2
+.
2
. In Example
1C.3(a), we asserted that j(r, j, .) was harmonic everywhere except at (0, 0, 0),
where it is not well-dened. For physical reasons, it is reasonable to write
the equation:
j(0, 0, 0) =
0
, (1D.2)
4
For a quick yet lucid introduction to electrostatics, see [Ste95, Chap.3].
1D. The Poisson equation 15
where
0
is the Dirac delta function (representing an innite concentration
of charge at zero)
5
. Then j(r, j, .) describes the electric potential generated
by a point charge.
Exercise 1D.2. Check that j(r, j, .) =
(r, j, .)
|(r, j, .)|
3
. This is the electric eld E _
generated by a point charge, as given by Coulombs Law from classical electrostatics.
Exercise 1D.3. (a) Let : 1
3
1 be a scalar eld describing a charge density E _
distribution. If

E : 1
3
1
3
is the electric eld generated by , then Gausss law
saws div

E = , where is a constant. If j : 1
3
1 is the electric potential eld
associated with

E, then by denition,

E = j. Use these facts to derive equation (1D.1).
(b) Suppose is independent of the r
3
coordinate; that is, (r
1
, r
2
, r
3
) = Q(r
1
, r
2
)
for some function Q : 1
2
1. Show that j is also is independent of the r
3
coordinate;
that is, j(r
1
, r
2
, r
3
) = 1(r
1
, r
2
) for some function 1 : 1
2
1. Show 1 and Q satisfy
the two-dimensional version of the Poisson equation that is that 1 = Q.
(This is signicant because many physical problems have (approximate) translational
symmetry along one dimension (e.g. an electric eld generated by a long, uniformly
charged wire or plate). Thus, we can reduce the problem to two dimensions, where
powerful methods from complex analysis can be applied; see Section 18B on page 422.)
Notice that the electric/gravitational potential eld is not uniquely dened

by equation (1D.1). If j(x) solves the Poisson equation (1D.1), then so does
j(x) = j(x) + o for any constant o 1. Thus, we say that the potential eld
is well-dened up to addition of a constant; this is similar to the way in which
the antiderivative
_
Q(r) of a function is only well-dened up to some constant.
6
This is an example of a more general phenomenon:
Proposition 1D.4. Let X 1
D
be some domain, and let j : X 1 and
/ : X 1 be two functions on X. Let j(x) := j(x) + /(x) for all x X.
Suppose that / is harmonic i.e. / 0. If j satises the Poisson Equation
j , then j also satises this Poisson equation.
Proof. Exercise 1D.4 Hint: Notice that j(x) = j(x) +/(x). 2 E _
For example, if Q(r) = 1,r
2
, as in Example 1D.1(b), then j(r) = log(r) is
a solution to the Poisson equation j
tt
(r) = 1,r
2
. If /(r) is a one-dimensional
5
Equation (1D.2) seems mathematically nonsensical, but it can be made mathematically
meaningful, using distribution theory. However, this is far beyond the scope of this book, so for
our purposes, we will interpret eqn. (1D.2) as purely metaphorical.
6
For the purposes of the physical theory, this constant does not matter, because the eld
p is physically interpreted only by computing the potential dierence between two points, and
the constant a will always cancel out in this computation. Thus, the two potential elds p(x)
and p(x) = p(x) +a will generate identical physical predictions.
harmonic function, then /(r) = or +/ for some constants o and / (see Example
1C.1(a) on page 10). Thus j(r) = log(r) +or+/, and weve already seen that
these are also valid solutions to this Poisson equation.
1E Properties of harmonic functions
Prerequisites: 1C, 0H(ii). Prerequisites (for proofs): 2A, 17G, 0E(iii).
Recall that a function / : 1
D
1 is harmonic if / 0. Harmonic
functions have nice geometric properties, which can be loosely summarized as
smooth and gently curving.
Theorem 1E.1. Mean Value Theorem
Let ) : 1
D
1 be a scalar eld. Then ) is harmonic if and only if ) is
integrable, and:
For any x 1
D
, and any 1 0, )(x) =
1
(1)
_
S(x;R)
)(s) ds. (1E.1)
Here, S(x; 1) :=
_
s 1
D
; |s x| = 1
_
is the (11)-dimensional sphere of
radius 1 around x, and (1) is the (11)-dimensional surface area of S(x; 1).
Proof. Exercise 1E.1 (a) Suppose ) is integrable and statement (1E.1) is true. E _
Use the Spherical Means formula for the Laplacian (Theorem 2A.1) to show that
) is harmonic.
(b) Now, suppose ) is harmonic. Dene : 1
1 by: (1) :=
1
(1)
_
S(x;R)
)(s) ds.
Show that
(1) =
1
(1)
_
S(x;R)
)(s) ds, for some constant 1 0.

Here,
)(s) is the outward normal derivative of ) at the point s on the sphere (see
page 564 for an abstract denition; see 5C(ii) on page 76 for more information).
(c) Let B (x; 1) :=
_
b 1
D
; |b x| 1
_
be the ball of radius 1 around x. Apply
Greens Formula (Theorem 0E.5(a) on page 564) to show that
(1) =
1
(1)
_
B(x;R)
)(b) db.
(d) Deduce that, if ) is harmonic, then must be constant.
(e) Use the fact that ) is continuous to show that lim
r0
(:) = )(x). Deduce that
(:) = )(x) for all : 0. Conclude that, if ) is harmonic, then statement (1E.1)
must be true. 2
1E. Properties of harmonic functions 17
Corollary 1E.2. Maximum Principle for harmonic functions
Let X 1
D
be a domain, and let n : X 1 be a nonconstant harmonic
function. Then n has no local maximal or minimal points anywhere in the interior
of X.
If X is bounded (hence compact), then n does obtain a maximum and mini-
mum, but only on the boundary of X.
Proof. (by contradiction). Suppose x was a local maximum of n somewhere in
the interior of X. Let 1 0 be small enough that S(x; 1) X, and such that
n(x) n(s) for all s S(x; 1), (1E.2)
where this inequality is strict for at least one s
0
S(x; 1).
Claim 1: There is a nonempty open subset Y S(x; 1) such that n(x)
n(y) for all y in Y.
Proof. We know that n(x) n(s
0
). But n is continuous, so there must be
some open neighbourhood Y around s
0
such that n(x) n(y) for all y in
Y.
Claim 1
Equation (1E.2) and Claim 1 imply that
)(x)
1
(1)
_
S(x;R)
)(s) ds.
But this contradicts the Mean Value Theorem. By contradiction, x cannot be
a local maximum. (The proof for local minima is analogous). 2
A function 1 : 1
D
1 is spherically symmetric if 1(x) depends only
on the norm |x| (i.e. 1(x) = ) (|x|) for some function ) : 1
,
1). For
example, the function 1(x) := exp(|x|
2
) is spherically symmetric.
If /, 1 : 1
D
1 are two integrable functions, then their convolution is
the function / 1 : 1
D
1 dened by
/ 1(x) :=
_
1
D
/(y) 1(x y) dy, for all x 1
D
(if this integral converges). We will encounter convolutions in 10D(ii) on
page 214 (where they will be used to prove the 1
2
convergence of a Fourier series)
and again in Chapter 17 (where they will be used to construct impulse-response
solutions for PDEs). For now, we state the following simple consequence of the
Mean Value Theorem:
Lemma 1E.3. If / : 1
D
1 is harmonic and 1 : 1
D
1 is integrable
and spherically symmetric, then / 1 = 1 /, where 1 1 is some constant.
Proof. Exercise 1E.2 2 E _
Proposition 1E.4. Smoothness of harmonic functions
If / : 1
D
1 is a harmonic function, then / is innitely dierentiable.
Proof. Let 1 : 1
D
1 be some innitely dierentiable, spherically sym-
metric, integrable function. For example, we could take 1(x) := exp(|x|
2
).
Then Proposition 17G.2(f) on page 410 says that / 1 is innitely dieren-
tiable. But Lemma 1E.3 implies that / 1 = 1/ for some constant 1 1;
thus, / is also innitely dierentiable.
(For another proof, see Theorem 6 on p. 28 of [Eva91, 2.2].) 2
Actually, we can go even further than this. A function / : X 1 is
analytic if, for every x X, there is a multivariate Taylor series expansion for
/ around x with a nonzero radius of convergence.
7
Proposition 1E.5. Harmonic functions are analytic
Let X 1
D
be an open set. If / : X 1 is a harmonic function, then / is
analytic on X.
Proof. For the case 1 = 2, see Corollary 18D.2 on page 451. For the general
case 1 2, see Theorem 10 on p. 31 of [Eva91, 2.2]. 2
1F
Transport and diusion

Prerequisites: 1B, 6A.
If n : 1
D
1 is a mountain, then recall that n(x) points in the
direction of most rapid ascent at x. If v 1
D
is a vector, then the dot product
vn(x) measures how rapidly you would be ascending if you walked in direction
v.
Suppose n : 1
D
1 describes a pile of leafs, and there is a steady wind
blowing in the direction v 1
D
. We would expect the pile to slowly move in
the direction v. Suppose you were an observer xed at location x. The pile is
moving past you in direction v, which is the same as you walking along the pile
in direction v; thus, you would expect the height of the pile at your location
to increase/decrease at rate v n(x). The pile thus satises the Transport
Equation:
t
n = v n.
Now, suppose that the wind does not blow in a constant direction, but instead
has some complex spatial pattern. The wind velocity is therefore determined
by a vector eld

V : 1
D
1
D
. As the wind picks up leafs and carries them
around, the ux of leafs at a point x X is then the vector

F(x) = n(x)

V(x).
7
See Appendices 0H(ii) and 0H(v) on pages 569 and 576.
1G.
Reaction and diusion 19

But the rate at which leafs are piling up at each location is the divergence of the
ux. This results in Liouvilles Equation:
t
n = div
F = div (n

V)
()

V n n div

V.
Here, () is by the Leibniz rule for divergence (Proposition 0E.2(b) on page 560).
Liouvilles equation describes the rate at which n-material accumulates when
it is being pushed around by the

V-vector eld. Another example:

V(x) de-
scribes the ow of water at x, and n(x) is the buildup of some sediment at
x.
Now suppose that, in addition to the deterministic force

V acting on the
leafs, there is also a random component. In other words, while being blown
around by the wind, the leafs are also subject to some random diusion. To
describe this, we combine Liouvilles Equation with the heat equation, to obtain
the Fokker-Plank equation:
t
n = n

V n n div

V.
1G
Reaction and diusion

Prerequisites: 1B.
Suppose ,1 and C are three chemicals, satisfying the chemical reaction:
2+1 = C
As this reaction proceeds, the and 1 species are consumed, and C is produced.
Thus, if o, /, c are the concentrations of the three chemicals, we have:
t
c = 1(t) =
t
/ =
1
2
t
o,
where 1(t) is the rate of the reaction at time t. The rate 1(t) is determined by
the concentrations of and 1, and by a rate constant . Each chemical reaction
requires 2 molecules of and one of 1; thus, the reaction rate is given by
1(t) = o(t)
2
/(t)
Hence, we get three ordinary dierential equations, called the reaction kinetic
equations of the system:
t
o(t) = 2 o(t)
2
/(t)
t
/(t) = o(t)
2
/(t)
t
c(t) = o(t)
2
/(t)
_
_
_
(1G.1)
Now, suppose that the chemicals , 1 and C are in solution, but are not uni-
formly mixed. At any location x X and time t 0, let o(x, t) be the con-
centration of chemical at location x at time t; likewise, let /(x, t) be the
concentration of 1 and c(x, t) be the concentration of C. (This determines three
time-varying scalar elds, o, /, c : 1
3
1 1.) As the chemicals react, their
concentrations at each point in space may change. Thus, the functions o, /, c will
obey the equations (1G.1) at each point in space. That is, for every x 1
3
and
t 1, we have
t
o(x; t) 2 o(x; t)
2
/(x; t)
etc. However, the dissolved chemicals are also subject to diusion forces. In other
words, each of the functions o, / and c will also be obeying the heat equation.
Thus, we get the system:
t
o =
a
o(x; t) 2 o(x; t)
2
/(x; t)
t
/ =
b
/(x; t) o(x; t)
2
/(x; t)
t
c =
c
c(x; t) + o(x; t)
2
/(x; t)
where
a
,
b
,
c
0 are three dierent diusivity constants.
This is an example of a reaction-diusion system. In general, in a
reaction-diusion system involving distinct chemicals, the concentrations of
the dierent species is described by a concentration vector eld u : 1
3
1
1
N
, and the chemical reaction is described by a rate function 1 : 1
N
1
N
.
For example, in the previous example, u(x, t) =
_
o(x, t), /(x, t), c(x, t)
_
, and
1(o, /, c) =
_
2o
2
/, o
2
/, o
2
/
.
The reaction-diusion equations for the system then take the form
t
n
n
=
n
n
n
+ 1
n
(u),
for n = 1, ...,
1H Practice problems
1. Let ) : 1
4
1 be a dierentiable scalar eld. Show that div )(r
1
, r
2
, r
3
, r
4
) =
)(r
1
, r
2
, r
3
, r
4
).
2. Let )(r, j; t) = exp(34t) sin(3r +5j). Show that )(r, j; t) satises the
two-dimensional heat equation:
t
)(r, j; t) = )(r, j; t).
3. Let n(r, j) = log(r
2
+j
2
). Show that n(r, j) satises the (two-dimensional)
Laplace Equation, everywhere except at (r, j) = (0, 0).
Remark: If (r, j) 1
2
, recall that |(r, j)| :=
_
r
2
+j
2
. Thus, log(r
2
+
j
2
) = 2 log |(r, j)|. This function is sometimes called the logarithmic
potential.
1H. Practice problems 21
4. If (r, j, .) 1
3
, recall that |(r, j, .)| :=
_
r
2
+j
2
+.
2
. Dene
n(r, j, .) =
1
|(r, j, .)|
=
1
_
r
2
+j
2
+.
2
Show that n satises the (three-dimensional) Laplace equation, everywhere
except at (r, j, .) = (0, 0, 0).
Remark: Observe that n(r, j, .) =
(r, j, .)
|(r, j, .)|
3
. What force eld does
this remind you of? Hint: n(r, j, .) is sometimes called the Coulomb
potential.
5. Let n(r, j; t) =
1
4t
exp
_
|(r, j)|
2
4t
_
=
1
4t
exp
_
r
2
j
2
4t
_
be the
(two-dimensional) Gauss-Weierstrass Kernel. Show that n satises the
(two-dimensional) heat equation,
t
n = n.
6. Let and be real numbers, and let /(r, j) = sinh(r) sin(j).
(a) Compute /(r, j).
(b) Suppose / is harmonic. Write an equation describing the relation-
ship between and .
Further reading
An analogy of the Laplacian can be dened on any Riemannian manifold, where
it is sometimes called the Laplace-Beltrami operator. The study of harmonic
functions on manifolds yields important geometric insights [War83, Cha93].
The reaction diusion systems from 1G play an important role in modern
mathematical biology [Mur93].
The heat equation also arises frequently in the theory of Brownian motion
and other Gaussian stochastic processes on 1
D
[Str93].
23
Chapter 2
Waves and signals
There is geometry in the humming of the strings. Pythagoras
2A The Laplacian and spherical means
Prerequisites: 0A, 0B, 0H(v). Recommended: 1B.
Let n : 1
D
1 be a function of 1 variables. Recall that the Laplacian
of n is dened:
n =
2
1
n +
2
2
n +. . .
2
D
n.
In this section, we will show that n(x) measures the discrepancy between n(x)
and the average of n in a small neighbourhood around x.
Let S(c) be the 1-dimensional sphere of radius c around 0. For example:
If 1 = 1, then S(c) is just a set with two points: S(c) = c, +c.
If 1 = 2, then S(c) is the circle of radius c: S(c) =
_
(r, j) 1
2
; r
2
+j
2
= c
2
_
If 1 = 3, then S(c) is the 3-dimensional spherical shell of radius c:
S(c) =
_
(r, j, .) 1
3
; r
2
+j
2
+.
2
= c
2
_
.
More generally, for any dimension 1,
S(c) =
_
(r
1
, r
2
, . . . , r
D
) 1
D
; r
2
1
+r
2
2
+. . . +r
2
D
= c
2
_
.
Let
be the surface area of the sphere. For example:

If 1 = 1, then S(c) = c, +c is a nite set, with two points, so we say
= 2.
If 1 = 2, then S(c) is the circle of radius c; the perimeter of this circle is
2c, so we say
= 2c.
24 DRAFT Chapter 2. Waves and signals
x x+ x-
f(x)
f(x+ )
f(x- )

f(x+) f(x-) +
2
M
f(x) =
M
f(x)
f(x)
M
f(x) - f(x)
Figure 2A.1: Local averages: )(r) vs. M
)(r) :=
f(x)+f(x+)
2
.
If 1 = 3, then S(c) is the sphere of radius c, which has surface area 4c
2
.
Let M
)(0) :=
1
_
S()
)(s) ds; then M
)(0) is the average value of )(s) over all

s on the surface of the c-radius sphere around 0, which is called the spherical
mean of ) at 0. The interpretation of the integral sign
_
depends on the
dimension 1 of the space. For example,
_
represents a surface integral if 1 = 3,
a line integral if 1 = 2, and simple two-point sum if 1 = 1. Thus:
If 1 = 1, then S(c) = c, +c, so that
_
S()
)(s) ds = )(c) + )(c);
thus,
M
) =
)(c) +)(c)
2
.
If 1 = 2, then any point on the circle has the form
_
c cos(), c sin()
_
for
some angle [0, 2). Thus,
_
S()
)(s) ds =
_
2
0
)
_
c cos(), c sin()
_
c d,
so that
M
) =
1
2c
_
2
0
)
_
c cos(), c sin()
_
c d =
1
2
_
2
0
)
_
c cos(), c sin()
_
d,
Likewise, for any x 1
D
, we dene M
)(x) :=
1
_
S()
)(x + s) ds to be
the average value of ) over an c-radius sphere around x. Suppose ) : 1
D
1
2A. The Laplacian and spherical means 25
is a smooth scalar eld, and x 1
D
. One interpretation of the Laplacian is this:
)(x) measures the disparity between )(x) and the average value of ) in the
immediate vicinity of x. This is the meaning of the next theorem:
Theorem 2A.1.
(a) If ) : 1 1 is a smooth scalar eld, then (as shown in Figure 2A.1), for
any r 1,
)(r) = lim
0
2
c
2
_
M
)(x) )(x)
_
= lim
0
2
c
2
_
)(r c) +)(r +c)
2
)(r)
_
.
(b)
1
If ) : 1
D
1 is a smooth scalar eld, then for any x 1
D
,
)(x) = lim
0
C
c
2
_
M
)(x) )(x)
_
= lim
0
C
c
2
_
1
_
S()
)(x +s) ds )(x)
_
(Here C is a constant determined by the dimension 1).
Proof. (a) Using Taylors theorem (see 0H(i) on page 568), we have:
)(r +c) = )(r) + c)
t
(r) +
c
2
2
)
tt
(r) + O(c
3
)
and )(r c) = )(r) c)
t
(r) +
c
2
2
)
tt
(r) + O(c
3
).
Here, )
t
(r) =
x
)(r) and )
tt
(r) =
2
x
)(r). The expression O(c) means
some function (we dont care which one) such that lim
0
O(c) = 0.
2
Like-
wise, O(c
3
) means some function (we dont care which one) such that
lim
0
O(c
3
)
c
2
= 0. Summing these two equations, we get
)(r +c) +)(r c) = 2)(r) + c
2
)
tt
(r) + O(c
3
).
Thus,
)(r c) 2)(r) +)(r +c)
c
2
= )
tt
(r) +O(c).
[because O(c
3
),c
2
= O(c).] Now take the limit as c 0, to get
lim
0
)(r c) 2)(r) +)(r +c)
c
2
= lim
0
)
tt
(r)+O(c) = )
tt
(r) = )(r),
1
Part (b) of Theorem 2A.1 is not necessary for the physical derivation of the wave equation
which appears later in this chapter. However, part (b) is required for to prove the Mean Value
Theorem for harmonic functions (Theorem 1E.1 on page 16).
2
Actually, O() means slightly more than this see 0H(i). However, for our purposes,
this will be sucient.
as desired.
(b) Dene the Hessian derivative matrix of ) at x:
D
2
)(x) =
_
2
1
)
1
2
) . . .
1
D
)
1
)
2
2
) . . .
2
D
)
.
.
.
.
.
.
.
.
.
.
.
.
1
)
D
2
) . . .
2
D
)
_
_
Then, for any s S(c), the Multivariate Taylors theorem (see 0H(v) on
page 576) says:
)(x +s) = )(x) + s )(x) +
1
2
s D
2
)(x) s + O(c
3
).
Now, if s = (:
1
, :
2
, . . . , :
D
), then sD
2
)(x) s =
D
c,d=1
:
c
:
d

c
d
)(x). Thus,
for any c 0, we have
)(x) =
_
S()
)(x +s) ds
=
_
S()
)(x) ds +
_
S()
s )(x) ds
+
1
2
_
S()
s D
2
)(x) s +
_
S()
O(c
3
) ds
=
)(x) + )(x)
_
S()
s ds
+
1
2
_
S()
_
_
D
c,d=1
:
c
:
d

c
d
)(x)
_
_
ds + O(c
D+2
)
=
)(x) + )(x) 0
. .
()
+
1
2
D
c,d=1
_
d
)(x)
_
_
S()
:
c
:
d
ds
__
+ O(c
D+2
)
=
)(x) +
1
2
D
d=1
_
2
d
)(x)
_
_
S()
:
2
d
ds
__
. .
()
+ O(c
D+2
)
=
)(x) +
1
2
)(x) c
D+1
1 + O(c
D+2
),
where 1 :=
_
S(1)
:
2
1
ds. Here, () is because
_
S()
s ds = 0, because the centre-
of-mass of a sphere is at its centre, namely 0. () is because, if c, d [1...1],
2B. The wave equation 27
Fixed
endpoint
Fixed
endpoint
mass m
-T cos()
T

s
i
n
(
)
T
T cos()
T

s
i
n
(
)
T
2

T

s
i
n
(
y
Figure 2B.1: A bead on a string
and c ,= d, then
_
S()
:
c
:
d
ds = 0 (Exercise 2A.1 Hint: Use symmetry). Thus, E _
)(x)
)(x) =
c
D+1
1
2
)(x) + O(c
D+2
),
so M
)(x) )(x) =
c
D+1
1
2
)(x) +
1
O(c
D+2
)
()
c
D+1
1
2
1
c
D1
)(x) + O
_
c
D+2
c
D1
_
=
c
2
1
2
1
)(x) +O(c
3
),
where () is because
=
1
c
D1
. Thus,
2
1
1 c
2
_
M
)(x) )(x)
_
= )(x) +O(c).
Now take the limit as c 0, and set C :=
2
1
1
, to prove part (b). 2
Exercise 2A.2. Let ) : 1
D
1 be a smooth scalar eld, such that M
)(x) = E _
)(x) for all x 1
D
. Show that ) is harmonic.
2B The wave equation
Prerequisites: 2A.
2B(i) ...in one dimension: the string
We want to mathematically describe vibrations propagating through a stretched
elastic cord. We will represent the cord with a one-dimensional domain X; either
X = [0, 1] or X = 1. We will make several simplifying assumptions:
(W1) The cord has uniform thickness and density. Thus, there is a constant
linear density 0, so that a cord-segment of length / has mass /.
(W2) The cord is perfectly elastic; meaning that it is innitely exible and does
not resist bending in any way. Likewise, there is no air friction to resist
the motion of the cord.
(W3) The only force acting on the cord is tension, which is force of magnitude
T pulling the cord to the right, balanced by an equal but opposite force
of magnitude T pulling the cord to the left. These two forces are in
balance, so the cord exhibits no horizontal motion. The tension T must be
constant along the whole length of the cord. Thus, the equilibrium state
for the cord is to be perfectly straight. Vibrations are deviations from this
straight position.
3
(W4) The vibrational motion of the cord is entirely vertical; there is no horizon-
tal component to the vibration. Thus, we can describe the motion using a
scalar-valued function n(r, t), where n(r, t) is the vertical displacement of
the cord (from its at equilibrium) at point r at time t. We assume that
n(r, t) is relatively small relative to the length of the cord, so that the cord
is not signicantly stretched by the vibrations
4
.
For simplicity, lets rst imagine a single bead of mass : suspended at the mid-
point of a (massless) elastic cord of length 2c, stretched between two endpoints.
Suppose we displace the bead by a distance j from its equilibrium, as shown in
Figure 2B.1. The tension force T now pulls the bead diagonally towards each
endpoint with force T. The horizontal components of the two tension forces are
equal and opposite, so they cancel, so the bead experiences no net horizontal
force. Suppose the cord makes an angle with the horizontal; then the vertical
component of each tension force is T sin(), so the total vertical force acting on
the bead is 2T sin(). But = arctan(c,j) by the geometry of the triangles in
3
We could also incorporate the force of gravity as a constant downward force. In this case,
the equilibrium position for the cord is to sag downwards in a catenary curve. Vibrations are
then deviations from this curve. This doesnt change the mathematics of this derivation, so we
will assume for simplicity that gravity is absent and the cord is straight.
4
If u(x, t) was large, then the vibrations stretch the cord, and a restoring force acts against
this stretching, as described by Hookes Law. By assuming that the vibrations are small, we
are assuming we can neglect Hookes Law.
Figure 2B.2: Each bead feels a negative force proportional to its deviation from the local
average.
Figure 2B.1, so sin() =
j
_
j
2
+c
2
. Thus, the vertical force acting on the bead
is
1 = 2T sin() =
2Tj
_
j
2
+c
2
(2B.1)
Now we return to our original problem of the vibrating string. Imagine
that we replace the string with a necklace made up of small beads of mass
: separated by massless elastic strings of length c. Each of these beads, in
isolation, behaves like the bead on a string in Figure 2B.1. However, now, the
vertical displacement of each bead is not computed relative to the horizontal,
but instead relative to the average height of the two neighbouring beads. Thus,
in eqn.(2B.1), we set j := n(r) M
n(r), where n(r) is the height of bead

r, and where M
n :=
1
2
[n(r c) + n(r + c)] is the average of its neighbours.
Substituting this into eqn.(2B.1), we get
1
(r) =
2T[n(r) M
n(r)]
_
[n(r) M
n(r)]
2
+c
2
(2B.2)
(Here, the c subscript in 1
is to remind us that this is just an c-bead

approximation of the real string). Each bead represents a length-c segment of
the original string, so if the string has linear density , then each bead must have
mass :
:= c. Thus, by Newtons law, the vertical acceleration of bead r must

be
o
(r) =
1
(r)
:
=
2T[n(r) M
n(r)]
c
_
[n(r) M
n(r)]
2
+c
2
=
2T[n(r) M
n(r)]
c
2
_
[n(r) M
n(r)]
2
,c
2
+ 1
(2B.3)
Now, we take the limit as c 0, to get the vertical acceleration of the string at
r:
o(r) = lim
0
o
(r) =
T
lim
0
2
c
2
_
n(r) M
n(r)
_
lim
0
1
_
[n(r) M
n(r)]
2
,c
2
+ 1
()
T

2
x
n(r)
1
lim
0
_
c
2

2
x
n(r)
2
+ 1
()
T

2
x
n(r). (2B.4)
Here, () is because Theorem 2A.1(a) on page 25 says that lim
0
2
c
2
[n(r) M
n(r)] =
2
x
n(r). Finally, () is because, for any value of n
tt
1, we have lim
0
_
c
2
n
tt
+ 1 =
1. We conclude that
o(r) =
T

2
x
n(r) =
2
2
x
n(r),
where :=
_
T,. Now, the position (and hence, velocity and acceleration) of
the cord is changing in time. Thus, o and n are functions of t as well as r. And
of course, the acceleration o(r, t) is nothing more than the second derivative of
n with respect to t. Hence we have the one-dimensional Wave Equation:
2
t
n(r, t) =
2

2
x
n(r, t).
This equation describes the propagation of a transverse wave along an idealized
string, or electrical pulses propagating in a wire.
T=0
T=1
T=2
T=3 T=7
T=6
T=5
T=4
Figure 2B.3: A one-dimensional standing wave.
Example 2B.1. Standing Waves
(a) Suppose
2
= 4, and let n(r; t) = sin(3r) cos(6t). Then n satises the
Wave Equation and describes a standing wave with a temporal frequency
of 6 and a wave number (or spatial frequency) of 3. (See Figure 2B.3)
(b) More generally, x 0; if n(r; t) = sin(r)cos(t), Then n satises
the wave equation and describes a standing wave of temporal frequency
and wave number .
Exercise 2B.1. Verify examples (a) and (b) above. E _
T=0
T=1
T=2
(A)
T=0
T=1
T=2
(B)
T=3
Figure 2B.4: (A) A one-dimensional sinusoidal travelling wave. (B) A general
one-dimensional travelling wave.
Example 2B.2. Travelling Waves
(a) Suppose
2
= 4, and let n(r; t) = sin(3r 6t). Then n satises the
Wave Equation and describes a sinusoidal travelling wave with temporal
frequency 6 and wave number 3. The wave crests move rightwards along
the cord with velocity 2. (Figure 2B.4A).
(b) More generally, x 1 and let n(r; t) = sin( r t). Then
n satises the wave equation and describes a sinusoidal travelling wave of
wave number . The wave crests move rightwards along the cord with
velocity .
(c) More generally, suppose that ) is any function of one variable, and dene
n(r; t) = ) (r t). Then n satises the wave equation and describes a
travelling wave, whose shape is given by ), and which moves rightwards
along the cord with velocity (see Figure 2B.4B).
Exercise 2B.2. Verify examples 2B.2(a,b,c) above. E _
Exercise 2B.3. According to Example 2B.2(c), you can turn any function into a E _
travelling wave. Can you turn any function into a standing wave? Why or why not?
2B(ii) ...in two dimensions: the drum
Now, suppose 1 = 2, and imagine a two-dimensional rubber sheet. Suppose
n(r, j; t) is the the vertical displacement of the rubber sheet at the point (r, j)
1
2
at time t. To derive the two-dimensional wave equation, we approximate this
rubber sheet as a two-dimensional mesh of tiny beads connected by massless,
tense elastic strings of length c. Each bead (r, j) feels a net vertical force 1 =
1
x
+1
y
, where 1
x
is the vertical force arising from the tension in the r direction,
and 1
y
is the vertical force from the tension in the j direction. Both of these
are expressed by a formula similar to eqn.(2B.2). Thus, if bead (r, j) has mass
:
, then it experiences acceleration o = 1,:
= 1
x
,:
+ 1
y
,:
= o
x
+ o
y
,
where o
x
:= 1
x
,:
and o
y
:= 1
y
,:
, and each of these is expressed by a formula

similar to eqn.(2B.3). Taking the limit as c 0 as in eqn.(2B.4), we deduce that
o(r, j) = lim
0
o
x,
(r, j) + lim
0
o
y,
(r, j) =
2
2
x
n(r, j) +
2
2
y
n(r, j),
where is a constant determined by the density and tension of the rubber
membrane. Again, we recall that n and o are also functions of time, and that
o(r, j; t) =
2
t
n(r, j; t). Thus, we have the two-dimensional Wave Equation:
2
t
n(r, j; t) =
2

2
x
n(r, j; t) +
2

2
y
n(r, j; t) (2B.5)
or, more abstractly:
2
t
n =
2
n.
This equation describes the propagation of wave energy through any medium
with a linear restoring force. For example:
Transverse waves on an idealized rubber sheet.
Ripples on the surface of a pool of water.
Acoustic vibrations on a drumskin.
Example 2B.3. Two-dimensional Standing Waves
(a) Suppose
2
= 9, and let n(r, j; t) = sin(3r)sin(4j)cos(15t). This describes
a two-dimensional standing wave with temporal frequency 15.
(b) More generally, x = (
1
,
2
) 1
2
and let = ||
2
=
_
2
1
+
2
2
.
Then the function
n(x; t) := sin (
1
r) sin (
2
j) cos ( t)
satises the 2-dimensional wave equation and describes a standing wave
with temporal frequency .
Figure 2B.5: A two-dimensional travelling wave.
(c) Even more generally, x = (
1
,
2
) 1
2
and let = ||
2
=
_
2
1
+
2
2
,
as before.
Let oC
1
(r) = either sin(r) or cos(r);
let oC
2
(j) = either sin(j) or cos(j);
and let oC
t
(t) = either sin(t) or cos(t).
Then
n(x; t) = oC
1
(
1
r) oC
2
(
2
j) oC
t
( t)
satises the 2-dimensional wave equation and describes a standing wave
with temporal frequency .
Exercise 2B.4. Check examples (a), (b) and (c) above. E _
Example 2B.4. Two-dimensional Travelling Waves
(a) Suppose
2
= 9, and let n(r, j; t) = sin(3r + 4j + 15t). Then n satises
the two-dimensional wave equation, and describes a sinusoidal travelling
wave with wave vector = (3, 4) and temporal frequency 15. (see Figure
2B.5).
(b) More generally, x = (
1
,
2
) 1
2
and let = ||
2
=
_
2
1
+
2
2
.
Then
n(x; t) = sin
_
1
r +
2
j + t
_
and (x; t) = cos
_
1
r +
2
j + t
_
both satisfy the two-dimensional wave equation, and describe sinusoidal
travelling waves with wave vector and temporal frequency .
Exercise 2B.5. Check examples (a) and (b) above. E _
2B(iii) ...in higher dimensions:
The same reasoning applies for 1 3. For example, the 3-dimensional wave
equation describes the propagation of (small amplitude
5
) sound-waves in air or
water. In general, the wave equation takes the form
2
t
n =
2
n,
where is some constant (determined by the density, elasticity, pressure, etc. of
the medium) which describes the speed-of-propagation of the waves.
By a suitable choice of time units, we can always assume that = 1. Hence,
from now on, we will consider the simplest form of the wave equation:
2
t
n = n.
For example, x = (
1
, . . . ,
D
) 1
D
and let = ||
2
=
_
2
1
+. . . +
2
D
.
Then
n(x; t) = sin
_
1
r
1
+
2
r
2
+. . . +
D
r
D
+ t
_
= sin
_
x + t
_
satises the 1-dimensional wave equation and describes a transverse wave of
with wave vector propagating across 1-dimensional space. (Exercise 2B.6
Check this.) E _
2C The telegraph equation
Recommended: 2B(i), 1B(i).
Imagine a signal propagating through a medium with a linear restoring force
(e.g. an electrical pulse in a wire, a vibration on a string). In an ideal universe,
the signal obeys the Wave Equation. However, in the real universe, damping
eects interfere. First, energy might leak out of the system. For example, if a
wire is imperfectly insulated, then current can leak out into surrounding space.
Also, the signal may get blurred by noise or frictional eects. For example, an
electric wire will pick up radio waves (crosstalk) from other nearby wires, while
losing energy to electrical resistance. A guitar string will pick up vibrations from
the air, while losing energy to friction.
Thus, intuitively, we expect the signal to propagate like a wave, but to be
gradually smeared out and attenuated by noise and leakage (Figure 2C.6). The
model for such a system is the telegraph equation:
2
t
n +
1
t
n +
0
n = n
5
At large amplitudes, nonlinear eects become important and invalidate the physical argu-
ment used here.
2D. Practice problems 35
T=1
T=2
T=3
T=4
Figure 2C.6: A solution to the telegraph equation propagates like a wave, but it also diuses
over time due to noise, and decays exponentially in magnitude due to leakage.
(where
2
,
1
,
0
, 0 are constants).
Heuristically speaking, this equation is a sum of two equations. The rst,
2
t
n =
1
n
is a version of the wave equation, and describes the ideal signal, while the
second,
t
n =
0
n +
2
n
describes energy lost due to leakage and frictional forces.
2D Practice problems
1. By explicitly computing derivatives, show that the following functions sat-
isfy the (one-dimensional) wave equation
2
t
n =
2
x
n.
(a) n(r, t) = sin(7r) cos(7t).
(b) n(r, t) = sin(3r) cos(3t).
(c) n(r, t) =
1
(xt)
2
(for r ,= t).
(d) n(r, t) = (r t)
2
3(r t) + 2.
(e) (r, t) = (r t)
2
.
2. Let ) : 1 1 be any twice-dierentiable function. Dene n : 11 1
by n(r, t) := )(r t), for all (r, t) 1 1.
Does n satises the (one-dimensional) wave equation
2
t
n = n? Justify
your answer.
3. Let n(r, t) be as in 1(a) and let (r, t) be as in 1(e), and suppose n(r, t) =
3n(r, t)2(r, t). Conclude that n also satises the wave equation, without
explicitly computing any derivatives of n.
4. Suppose n(r, t) and (r, t) are both solutions to the wave equation, and
n(r, t) = 5n(r, t) + 2(r, t). Conclude that n also satises the wave
equation.
5. Let n(r, t) =
_
x+t
xt
cos(j) dj = sin(r + t) sin(r t). Show that n
satises the (one-dimensional) wave equation
2
t
n = n.
6. By explicitly computing derivatives, show that the following functions sat-
isfy the (two-dimensional) wave equation
2
t
n = n.
(a) n(r, j; t) = sinh(3r) cos(5j) cos(4t).
(b) n(r, j; t) = sin(r) cos(2j) sin(
5t).
(c) n(r, j; t) = sin(3r 4j) cos(5t).
37
Chapter 3
Quantum mechanics
[M]odern physics has denitely decided in favor of Plato. In fact the smallest units
of matter are not physical objects in the ordinary sense; they are forms, ideas which can be
expressed unambiguously only in mathematical language. Werner Heisenberg
3A Basic framework
Prerequisites: 0C, 1B(ii).
Near the beginning of the twentieth century, physicists realized that elec-
tromagnetic waves sometimes exhibited particle-like properties, as if light was
composed of discrete photons. In 1923, Louis de Broglie proposed that, con-
versely, particles of matter might have wave-like properties. This was conrmed
in 1927 by C.J. Davisson and L.H. Germer, and independently, by G.P. Thomp-
son, who showed that an electron beam exhibited an unmistakable diraction
pattern when scattered o a metal plate, as if the beam was composed of elec-
tron waves. Systems with many interacting particles exhibit even more curious
phenomena. Quantum mechanics is a theory which explains these phenomena.
We will not attempt here to provide a physical justication for quantum
mechanics. Historically, quantum theory developed through a combination of
vaguely implausible physical analogies and wild guesses motivated by inexpli-
cable empirical phenomena. By now, these analogies and guesses have been
overwhelmingly vindicated by experimental evidence. The best justication for
quantum mechanics is that it works, by which we mean that its theoretical
predictions match all available empirical data with astonishing accuracy.
Unlike the heat equation in 1B and the Wave Equation in 2B, we cannot
derive quantum theory from rst principles, because the postulates of quantum
mechanics are the rst principles. Instead, we will simply state the main as-
sumptions of the theory, which are far from self-evident, but which we hope you
will accept because of the weight of empirical evidence in their favour. Quan-
tum theory describes any physical system via a probability distribution on a
certain statespace. This probability distribution evolves over time; the evolution
38 DRAFT Chapter 3. Quantum mechanics
is driven by a potential energy function, as described by a partial dierential
equation called the Schrodinger equation. We will now examine each of these
concepts in turn.
Statespace: A system of interacting particles moving in 3 dimensional space
can be completely described using the 3-dimensional state space X := 1
3N
.
An element of X consists of list of ordered triples:
x = (r
11
, r
12
, r
13
; r
21
, r
22
, r
23
; . . . r
N1
, r
N2
, r
N3
) 1
3N
,
where (r
11
, r
12
, r
13
) is the spatial position of particle #1, (r
21
, r
22
, r
23
) is the
spatial position of particle #2, and so on.
Example 3A.1. (a) Single electron A single electron is a one-particle system,
so it would be represented using a 3-dimensional statespace X = 1
3
. If the
electron was conned to a two-dimensional space (e.g. a conducting plate),
we would use X = 1
2
. If the electron was conned to a one-dimensional space
(e.g. a conducting wire), we would use X = 1.
(b) Hydrogen Atom: The common isotope of hydrogen contains a single proton
and a single electron, so it is a two-particle system, and would be represented
using a 6-dimensional state space X = 1
6
. An element of X has the form
x = (r
p
1
, r
p
2
, r
p
3
; r
e
1
, r
e
2
, r
e
3
), where (r
p
1
, r
p
2
, r
p
3
) are the coordinates of the
proton, and (r
e
1
, r
e
2
, r
e
3
) are those of the electron.
Readers familiar with classical mechanics may be wondering how momentum is
represented in this statespace. Why isnt the statespace 6-dimensional, with
3 position and 3 momentum coordinates for each particle? The answer, as we
will see later, is that the momentum of a quantum system is implicitly encoded
in the wavefunction which describes its position (see 19G on page 511).
Potential Energy: We dene a potential energy (or voltage) function \ :
X 1, which describes which states are prefered by the quantum system.
Loosely speaking, the system will avoid states of high potential energy, and
seek states of low energy. The voltage function is usually dened using reasoning
familiar from classical physics.
Example 3A.2: Electron in ambient eld
Imagine a single electron moving through an ambient electric eld

E. The
statespace for this system is X = 1
3
, as in Example 3A.1(a). The potential
function \ is just the voltage of the electric eld; in other words, \ is any
scalar function such that
e
E = \ , where
e
is the charge of the electron.
For example:
3A. Basic framework 39
(a) Null eld: If

E 0, then \ will be a constant, which we can assume is
zero: \ 0.
(b) Constant eld: If

E (1, 0, 0), for some constant 1 1, then \ (r, j, .) =
e
1r +c, where c is an arbitrary constant, which we normally set to zero.
(c) Coulomb eld: Suppose the electric eld

E is generated by a (stationary)
point charge Q at the origin. Let c
0
be the permittivity of free space. Then
Coulombs law says that the electric voltage is given by
\ (x) :=

e
Q
4c
0
[x[
, for all x 1
3
.
In SI units,
e
1.60 10
19
C, and c
0
8.85 10
12
C,N m
2
. However, for
simplicity, we will normally adopt atomic units of charge and eld strength,
where
e
= 1 and 4c
0
= 1. Then the above expression becomes \ (x) =
Q,[x[.
(d) Potential well: Sometimes we conne the electron to some bounded region
B 1
3
, by setting the voltage equal to positive innity outside B. For
example, a low-energy electron in a cube made of conducting metal can move
freely about the cube, but cannot leave
1
the cube. If the subset B represents
the cube, then we dene \ : X [0, ] by
\ (x) =
_
0 if x B;
+ if x , B.
(if + makes you uncomfortable, then replace it with some really big num-
ber).
Example 3A.3: Hydrogen atom:
The system is an electron and a proton; the statespace of this system is X = 1
6
as in Example 3A.1(b). Assuming there is no external electric eld, the voltage
function is dened
\ (x
p
, x
e
) :=

2
e
4c
0
[x
p
x
e
[
, for all (x
p
, x
e
) 1
6
.
where x
p
is the position of the proton, x
e
is the position of the electron, and
e
is the charge of the electron (which is also the charge of the proton, with
reversed sign). If we adopt atomic units where
e
:= 1 and 4c
0
= 1, then
this expression simplies to
\ (x
p
, x
e
) :=
1
[x
p
x
e
[
, for all (x
p
, x
e
) 1
6
,
1
Cannot leave of course really means is very highly unlikely to leave.
Probability and Wavefunctions. Our knowledge of the classical properties
of a quantum system is inherently incomplete. All we have is a time-varying
probability distribution : X 1 1
,
which describes where the particles
are likely or unlikely to be at a given moment in time.
As time passes, the probability distribution evolves. However, itself
cannot exhibit the wavelike properties of a quantum system (e.g. destructive
interference), because is a nonnegative function (and we need to add negative
to positive values to get destructive interference). So, we introduce a complex-
valued wavefunction : X 1 C. The wavefunction determines via
the equation:
t
(x) := [
t
(x)[
2
, for all x X and t 1.
(Here, as always in this book, we dene
t
(x) := (x; t) and
t
(x) := (x; t);
subscripts do not indicate derivatives). Now,
t
is supposed to be a probability
density function, so
t
must satisfy the condition
_
X
[
t
(x)[
2
dx = 1, for all t 1. (3A.1)
It is acceptable (and convenient) to relax condition (3A.1), and instead simply
require
_
X
[
t
(x)[
2
dx = \ < , for all t 1. (3A.2)
where \ is some nite constant, independent of t. In this case, we dene
t
(x) :=
1
W
[
t
(x)[
2
for all x X. It follows that any physically meaningful solution to the
Schrodinger equation must satisfy condition (3A.2). This excludes, for example,
solutions where the magnitude of the wavefunction grows exponentially in the x
or t variables.
For any xed t 1, condition (3A.2) is usually expressed by saying that
t
is square-integrable. Let L
2
(X) denote the set of all square-integrable functions
on X. If
t
L
2
(X), then the 1
2
-norm of is dened
|
t
|
2
:=
_
X
[
t
(x)[
2
dx.
Thus, a fundamental postulate of quantum theory is:
Let : X 1 C be a wavefunction. To be physically mean-
ingful, we must have
t
L
2
(X) for all t 1. Furthermore, |
t
|
2
must be constant in time.
We refer the reader to 6B on page 105 for more information on 1
2
-norms and
1
2
-spaces.
3B. The Schrodinger equation 41
3B The Schr odinger equation
Prerequisites: 3A. Recommended: 4B.
The wavefunction evolves over time in response to the potential eld \ .
Let / be the rationalized Planck constant
/ :=
/
2

1
2
6.6256 10
34
J s 1.0545 10
34
J s.
Then the wavefunctions evolution is described by the Schrodinger Equation:
i/
t
= H, (3B.1)
where H is a linear dierential operator called the Hamiltonian operator,
dened by:
H
t
(x) :=
/
2
2

t
(x) + \ (x)
t
(x), for all x X and t 1. (3B.2)
Here,
t
is like the Laplacian of
t
, except that the components for each
particle are divided by the rest mass of that particle. The potential function \ :
X 1 encodes all the exogenous aspects of the system we are modelling (e.g.
the presence of ambient electric elds). Substituting eqn.(3B.2) into eqn.(3B.1),
we get
i/
t
=
/
2
2
+ \ , (3B.3)
In atomic units, / = 1, so the Schrodinger equation (3B.3) becomes
i
t
t
(x) =
1
2

t
(x) + \ (x)
t
(x), for all x X and t 1.
Example 3B.1. (a) Free Electron: Let :
e
9.1110
31
kg be the rest mass of
an electron. A solitary electron in a null electric eld (as in Example 3A.2(a))
satises the free Schrodinger equation:
i/
t
t
(x) =
/
2
2 :
e
t
(x). (3B.4)
(In this case =
1
me
, and \ 0 because the ambient eld is null). In
atomic units, we set :
e
:= 1 and / := 1, so eqn.(3B.4) becomes
i
t
=
1
2
=
1
2
_
2
1
+
2
2
+
2
3
_
. (3B.5)
(b) Electron vs. point charge: Consider the Coulomb electric eld, generated
by a (stationary) point charge Q at the origin, as in Example 3A.2(c). A
solitary electron in this electric eld satises the Schrodinger equation
i/
t
t
(x) =
/
2
2 :
e
t
(x) +

e
Q
4c
0
[x[

t
(x).
In atomic units, we have :
e
:= 1,
e
:= 1, etc. Let

Q = Q,
e
be the charge Q
converted in units of electron charge. Then the previous expression simplies
to
i
t
t
(x) =
1
2

t
(x) +

Q
[x[

t
(x).
(c) Hydrogen atom: (see Example 3A.3) An interacting proton-electron pair
(in the absence of an ambient eld) satises the two-particle Schrodinger equa-
tion
i/
t
t
(x
p
, x
e
) =
/
2
2 :
p
t
(x
p
, x
e
) +
/
2
2 :
e
t
(x
p
, x
e
) +

2
e

t
(x
p
, x
e
)
4c
0
[x
p
x
e
[
,
(3B.6)
where
p
:=
2
x
p
1
+
2
x
p
2
+
2
x
p
3
is the Laplacian in the proton position
coordinates, and :
p
1.672710
27
kg is the rest mass of a proton. Likewise,
e
:=
2
x
e
1
+
2
x
e
2
+
2
x
e
3
is the Laplacian in the electron position
coordinates, and :
e
is the rest mass of the electron. In atomic units, we have
4c
0
= 1,
e
= 1, and :
e
= 1. If :
p
1864 is the ratio of proton mass to
electron mass, then 2 :
p
3728, and eqn.(3B.6) becomes
i
t
t
(x
p
, x
e
) =
1
3728

p

t
(x
p
, x
e
) +
1
2

e

t
(x
p
, x
e
) +

t
(x
p
, x
e
)
[x
p
x
e
[
.
The major mathematical problems of quantum mechanics come down to nd-

ing solutions to the Schrodinger equations for various physical systems. In gen-
eral it is very dicult to solve the Schrodinger equation for most realistic po-
tential functions. We will conne ourselves to a few toy models to illustrate the
essential ideas.
Example 3B.2: Free Electron with Known Velocity (Null Field)
Consider a single electron in a null electromagnetic eld. Suppose an ex-
periment has precisely measured the classical velocity of the electron, and
determined it to be v = (
1
, 0, 0). Then the wavefunction of the electron is
given
2
t
(x) = exp
_
i
/
:
e
2
1
2
t
_
exp
_
i
/
:
e
1
r
1
_
. (see Figure 3B.1)
(3B.7)
This satises the free Schrodinger equation (3B.4). [See practice problem
# 1 on page 54 of 3D.]
3B. The Schrodinger equation 43
Time: t=0
Time: t=T/4
Time: t=T/2
Time: t=3T/4
Figure 3B.1: Four successive snapshots of the wavefunction of a single electron in a zero
potential, with a precisely known velocity. Only one spatial dimension is shown. The angle of
the spiral indicates complex phase.
Exercise 3B.1. (a) Check that the spatial wavelength of the function is E _
given =
2/
j
1
=
/
:
e
. This is the so-called de Broglie wavelength of an electron

with velocity .
(b) Check that the temporal period of is T :=
2/
:
e
2
.
(c) Conclude the phase velocity of (i.e. the speed at which the wavefronts propagate
through space) is equal to .
More generally, suppose the electron has a precisely known velocity v =
(
1
,
2
,
3
), with corresponding momentum vector p := :
e
v. Then the wave-
function of the electron is given
t
(x) = exp
_
i
/
1
k
t
_
exp
_
i
/
p x
_
, (3B.8)
where 1
k
:=
1
2
:
e
[v[
2
is kinetic energy, and px := j
1
r
1
+j
2
r
2
+j
3
r
3
. If we
convert to atomic units, then 1
k
=
1
2
[v[
2
and p = v, and this function takes
the simpler form
t
(x) = exp
_
i [v[
2
t
2
_
exp (i v x) .
2
We will not attempt here to justify why this is the correct wavefunction for a particle with
this velocity. It is not obvious.
This satises the free Schrodinger equation (3B.5). [See practice problem
# 2 on page 54 of 3D.]
The wavefunction (3B.8) represent a state of maximal uncertainty about the
position of the electron. This is an extreme manifestation of the infamous
Heisenberg Uncertainty Principle; by assuming that the electrons velocity was
precisely determined, we have forced its position to be entirely undetermined
(see 19G for more information).
Indeed, the wavefunction (3B.8) violates our fundamental postulate the
function
t
is not square-integrable, because [
t
(x)[ = 1 for all x 1, so
_
1
3
[
t
(x)[
2
dx = . Thus, wavefunction (3B.8) cannot be translated into
a probability distribution, so it is not physically meaningful. This isnt too
surprising, because wavefunction (3B.8) seems to suggest that the electron is
equally likely to be anywhere in the (innite) universe 1
3
! It is well known
that the location of a quantum particle can be dispersed over some region of
space, but this seems a bit extreme. There are two solutions to this problem.
Let B(1) 1
3
be a ball of radius 1, where 1 is much larger than the
physical system or laboratory apparatus we are modelling (e.g. 1 = 1
lightyear). Dene the wavefunction
(R)
t
(x) by (3B.8) for all x B(1),
and set
(R)
t
(x) = 0 for all x , B(1). This means that the position of
the electron is still extremely dispersed (indeed, innitely dispersed for
the purposes of any laboratory experiment), but the function
(R)
t
is still
square-integrable. Note that the function
(R)
t
violates the Schrodinger
equation at the boundary of B(1), but this boundary occurs very far from
the physical system we are studying, so it doesnt matter. In a sense, the
solution (3B.8) can be seen as the limit of
(R)
as 1 .
Reject the wavefunction (3B.8) as physically meaningless. Our starting
assumption an electron with a precisely known velocity has led to a
contradiction. Our conclusion: a free quantum particle can never have a
precisely known classical velocity. Any physically meaningful wavefunc-
tion in a vacuum must contain a mixture of several velocities.
Remark. (The meaning of phase) At any point x in space and moment t in

time, the wavefunction
t
(x) can be described by its amplitude
t
(x) := [
t
(x)[
and its phase
t
(x) :=
t
(x),
t
(x). We have already discussed the physical
meaning of the amplitude: [
t
(x)[
2
is the probability that a classical measurement
will produce the outcome x at time t. What is the meaning of phase?
The phase
t
(x) is a complex number of modulus one an element of the unit
circle in the complex plane (hence
t
(x) is sometimes called the phase angle).
The oscillation of the wavefunction over time can be imagined in terms of
3C. Stationary Schrodinger equation 45
the rotation of
t
(x) around the circle. The wavelike properties of quantum
systems (e.g. interference patterns) occur because wavefunctions with dierent
phases will partially cancel one another when they are superposed. In other
words, it is because of phase that the Schrodinger Equation yields wave-like
phenomena, instead of yielding diusive phenomena like the heat equation.
However, like potential energy, phase is not directly physically observable.
We can observe the phase dierence between wavefunction and wavefunction
(by observing cancelation between and ), just as we can observe the potential
energy dierence between point and point 1 (by measuring the energy released
by a particle moving from point to point 1). However, it is not physically
meaningful to speak of the absolute phase of wavefunction , just as it is not
physically meaningful to speak of the absolute potential energy of point .
Indeed, inspection of the Schrodinger equation (3B.3) on page 41 will reveal
that the speed of phase rotation of a wavefunction at point x is determined by
the magnitude of the potential function \ at x. But we can arbitrarily increase
\ by a constant, without changing its physical meaning. Thus, we can arbitrarily
accelerate the phase rotation of the wavefunction without changing the physical
meaning of the solution.
3C Stationary Schrodinger equation
Prerequisites: 3B. Recommended: 4B(iv).
A stationary state of a quantum system is one where the probability density
does not change with time. This represents a physical system which is in some
kind of long-term equilibrium. Note that a stationary quantum state does not
mean that the particles are not moving (whatever moving means for quanta).
It instead means that they are moving in some kind of regular, conned pattern
(i.e. an orbit) which remains qualitatively the same over time. For example, the
orbital of an electron in a hydrogen atom should be a stationary state, because
(unless the electron absorbs or emits energy) the orbital should stay the same
over time.
Mathematically speaking, a stationary wavefunction yields a time-invariant
probability density function : X 1 such that, for any t 1,
[
t
(x)[
2
= (x), for all x X.
The simplest way to achieve this is to assume that has the separated form
t
(x) = (t)
0
(x), (3C.1)
where
0
: X C and : 1 C satisfy the conditions
[(t)[ = 1, for all t 1, and [
0
(x)[ =
_
(x), for all r X. (3C.2)
Lemma 3C.1. Suppose
t
(x) = (t)
0
(x) is a separated solution to
the Schrodinger equation, as in eqn.(3C.1) and eqn.(3C.2). Then there is some
constant 1 1 so that
(t) = exp(i1t,/), for all t 1.
H
0
= 1
0
; in other words
0
is an eigenfunction
3
of the Hamiltonian
operator H, with eigenvalue 1.
Thus,
t
(x) = c
iEt//

0
(x), for all x X and t 1.
Proof. Exercise 3C.1 Hint: use separation of variables.
4
2 E _
Physically speaking, 1 corresponds to the total energy (potential + kinetic)
of the quantum system
5
. Thus, Lemma 3C.1 yields one of the key concepts of
quantum theory:
Eigenfunctions of the Hamiltonian correspond to stationary quantum
states. The eigenvalues of these eigenfunctions correspond to the
energy level of these states.
Thus, to get stationary states, we must solve the stationary Schrodinger
equation:
H
0
= 1
0
,
where 1 1 is an unknown constant (the energy eigenvalue), and
0
: X C
is an unknown wavefunction.
Example 3C.2: The Free Electron
Recall free electron of Example 3B.2. If the electron has velocity , then
the function in eqn.(3B.7) yields a solution to the stationary Schrodinger
equation, with eigenvalue 1 =
1
2
:
e
2
. [See practice problem # 3 on page 54 of
3D]. Observe that 1 corresponds to the classical kinetic energy of an electron
with velocity .
3
See 4B(iv) on page 63.
4
See Chapter 16 on page 353.
5
This is not obvious, but its a consequence of the fact that the Hamiltonian H measures
the total energy of the wavefunction . Loosely speaking, the term

2
2
represents the kinetic
energy of , while the term V represents the potential energy.
L
0
V
V
0
0
Figure 3C.1: The (stationary) wavefunction of an electron in a one-dimensional square
potential well, with nite voltage gaps.
Example 3C.3: One-dimensional square potential well; nite voltage
Consider an electron conned to a one-dimensional environment (e.g. a long
conducting wire). Thus, X := 1, and the wavefunction
0
: 11 C obeys
the one-dimensional Schrodinger equation
i
t
0
=
1
2

2
x
0
+\
0
,
where \ : 1 1 is the potential energy function, and we have adopted
atomic units. Let \
0
0 be some constant, and suppose that
\ (r) =
_
0 if 0 r 1;
\
0
if r < 0 or 1 < r.
Physically, this means that \ denes a potential energy well, which tries
to conne the electron in the interval [0, 1], between two walls, which are
voltage gaps of height \
0
(see Figure 3C.1). The corresponding stationary
Schrodinger equation is:
1
2

2
x
0
+\
0
= 1
0
, (3C.3)
where 1 0 is an (unknown) eigenvalue which corresponds to the energy of
the electron. The function \ only takes two values, so we can split eqn.(3C.3)
into two equations, one inside the interval [0, 1], and one outside it:
1
2

2
x
0
(r) = 1
0
(r), for r [0, 1];
1
2

2
x
0
(r) = (1 \
0
)
0
(r), for r , [0, 1].
(3C.4)
Assume that 1 < \
0
. This means that the electrons energy is less than the
voltage gap, so the electron has insucient energy to escape the interval (at
least in classical theory). The (physically meaningful) solutions to eqn.(3C.4)
have the form
0
(r) =
_
_
_
C exp(c
t
r), if r (, 0];
sin(cr) +1cos(cr), if r [0, 1];
1exp(c
t
r), if 1 [1, ).
(3C.5)
(See Figure 3C.1.) Here, c :=

21 and c
t
:=

21 2\
0
, and , 1, C, 1 C
are constants. The corresponding solution to the full Schrodinger equation is:
t
(r) =
_
_
_
Cc
i(EV
0
)t
exp(c
t
r), if r (, 0];
c
iEt
(sin(cr) +1cos(cr)) , if r [0, 1] ;
1c
i(EV
0
)t
exp(c
t
r), if 1 [1, ).
for all t 1.
This has two consequences:
(a) With nonzero probability, the electron might be found outside the interval
[0, 1]. In other words, it is quantumly possible for the electron to escape
from the potential well, something which is classically impossible
6
. This
phenomenon called quantum tunnelling (because the electron can tunnel
through the wall of the well).
(b) The system has a physically meaningful solution only for certain values
of 1. In other words, the electron is only allowed to reside at certain
discrete energy levels; this phenomenon is called quantization of energy.
To see (a), recall that the electron has probability distribution
(r) :=
1
\
[
0
(r)[
2
, where \ :=
_

[
0
(r)[
2
dr.
Thus, if C ,= 0, then (r) ,= 0 for r < 0, while if 1 ,= 0, then (r) ,= 0 for
r 1. Either way, the electron has nonzero probability of tunnelling out of
the well.
To see (b), note that we must choose , 1, C, 1 so that
0
is continuously
dierentiable at the boundary points r = 0 and r = 1. This means we must
have
1 = sin(0) +1cos(0) =
0
(0) = C exp(0) = C
c = c cos(0) 1c sin(0) =
t
0
(0) = c
t
C exp(0) = c
t
C
sin(c1) +1cos(c1) =
0
(1) = 1exp(c
t
1)
c cos(c1) 1c sin(c1) =
t
0
(1) = c
t
1exp(c
t
1)
(3C.6)
6
Many older texts observe that the electron can penetrate the classically forbidden region,
which has caused mirth to generations of physics students.
Clearly, we can satisfy the rst two equations in (3C.6) by setting 1 := C :=
/
. The third and fourth equations in (3C.6) then become
c
/
L
_
sin(c1) +
c
c
t
cos(c1)
_
= 1 =
c
c
t
c
/
L
_
cos(c1)
c
c
t
sin(c1)
_
,
(3C.7)
Cancelling the factors c
/
L
and from both sides and substituting c :=

21
and c
t
:=

21 2\
0
, we see that eqn.(3C.7) is satisable if and only if
sin
_
21 1
_
+
1 cos(
21 1)
1 \
0
=
1 cos
_
21 1
_
1 \
0
+
1 sin(
21 1)
1 \
0
.
(3C.8)
Hence, eqn.(3C.4) has a physically meaningful solution only for those values
of 1 which satisfy the transcendental equation (3C.8). The set of solutions
to eqn.(3C.8) is an innite discrete subset of 1; each solution for eqn.(3C.8)
corresponds to an allowed energy level for the physical system.
L
0
V
0
Figure 3C.2: The (stationary) wavefunction of an electron in an innite potential well.
Example 3C.4: One-dimensional square potential well; innite voltage
We can further simplify the model of Example 3C.3 by setting \
0
:= +,
which physically represents a huge voltage gap that totally connes the elec-
tron within the interval [0, 1] (see Figure 3C.2). In this case, c
t
= , so
exp(c
t
r) = 0 for all r < 0 and exp(c
t
r) = 0 for all r 1. Hence, if
0
is as
in eqn.(3C.5), then
0
(r) 0 for all r , [0, 1], and the constants C and 1
are no longer physically meaningful; we set C = 0 = 1 for simplicity. Also,
we must have
0
(0) = 0 =
0
(1) to get a continuous solution; thus, we must
set 1 := 0 in eqn.(3C.5). Thus, the stationary solution in eqn.(3C.5) becomes
0
(r) =
_
0 if r , [0, 1];
sin(
21 r) if r [0, 1],
where is a constant, and 1 satises the equation
sin(
21 1) = 0. (Figure 3C.2) (3C.9)

Assume for simplicity that 1 := . Then eqn.(3C.9) is true if and only if
21 is an integer, which means 21 0, 1, 4, 9, 16, 25, . . ., which means

1 0,
1
2
, 2,
9
2
, 8,
25
2
, . . .. Here we see the phenomenon of quantization of
energy in its simplest form.
The set of eigenvalues of a linear operator is called the spectrum of that
operator. For example, in Example 3C.4, the spectrum of the Hamiltonian op-
erator H is the set 0,
1
2
, 2,
9
2
, 8,
25
2
, . . .. In quantum theory, the spectrum of
the Hamiltonian is the set of allowed energy levels of the system.
Example 3C.5: Three-dimensional square potential well; innite voltage
We can easily generalize Example 3C.4 to three dimensions. Let X := 1
3
,
and let B := [0, ]
3
be a cube with one corner at the origin, having sidelength
1 = . We use the potential function \ : X 1 dened
\ (x) =
_
0 if x B;
+ if x , B.
Physically, this represents an electron conned within a cube of perfectly con-
ducting material with perfectly insulating boundaries
7
. Suppose the electron
has energy 1. The corresponding stationary Schrodinger equation is
1
2

0
(x) = 1
0
(x) for x B;
1
2

0
(x) =
0
(x) for x , B;
(3C.10)
(in atomic units). By reasoning similar Example 3C.4, we nd that the phys-
ically meaningul solutions to eqn.(3C.10) have the form
0
(x) =
_

2
3/2
sin(n
1
r
1
) sin(n
2
r
2
) sin(n
3
r
3
) if x = (r
1
, r
2
, r
3
) B;
0 if x , B.
(3C.11)
where n
1
, n
2
, and n
3
are arbitrary integers (called the quantum numbers of
the solution), and 1 =
1
2
(n
2
1
+n
2
2
+n
2
3
) is the associated energy eigenvalue.
7
Alternately, it could be any kind of particle, conned in a cubical cavity with impenetrable
boundaries.
The corresponding solution to the full Schrodinger equation for all t 1 is
t
(x) =
_

2
3/2
c
i(n
2
1
+n
2
2
+n
2
3
)t/2
sin(n
1
r
1
) sin(n
2
r
2
) sin(n
3
r
3
) if x B;
0 if x , B.
Exercise 3C.2. (a) Check that eqn.(3C.11) is a solution for eqn.(3C.10). E _

(b) Check that := [[
2
is a probability density, by conrming that
_
X
[
0
(x)[
2
dx =
2
3
_

0
_

0
_

0
sin(n
1
r
1
)
2
sin(n
2
r
2
)
2
sin(n
3
r
3
)
2
dr
1
dr
2
dr
3
= 1,
(this is the reason for using the constant
3/2
).
1
0.8
0.6
0.4
0.2
r
3 2.5 2 1 0.5 0 1.5
y
2
-2
3
1
3
x
0
-1
-1 -2 1 2
-3
0 -3
(A) (B)
Figure 3C.3: The groundstate wavefunction for a hydrogen atom. (A) Probability density as
a function of distance from the nucleus. (B) Probability density visualized in three dimensions.
Example 3C.6: Hydrogen Atom
In Example 3A.3 on page 39, we described the hydrogen atom as a two-
particle system, with a six-dimensional state space. However, the correspond-
ing Schrodinger equation (Example 3B.1(c)) is already too complicated for us
to solve it here, so we will work with a simplied model.
Because the proton is 1864 times as massive as the electron, we can treat
the proton as remaining eectively immobile while the electron moves around
it. Thus, we can model the hydrogen atom as a one-particle system: a single
electron moving in a Coulomb potential well, as described in Example 3B.1(b).
The electron then satises the Schrodinger equation
i/
t
t
(x) =
/
2
2 :
e
t
(x) +

2
e
4c
0
[x[

t
(x), x 1
3
. (3C.12)
(Recall that :
e
is the mass of the electron,
e
is the charge of both electron
and proton, c
0
is the permittivity of free space, and / is the rationalized
Plank constant.) Assuming the electron is in a stable orbital, we can replace
eqn.(3C.12) with the stationary Schrodinger equation
/
2
2 :
e
0
(x) +

2
e
4c
0
[x[

0
(x) = 1
0
(x), x 1
3
, (3C.13)
where 1 is the energy level of the electron. One solution to this equation is
(x) =
/
3/2
exp(/[x[), where / :=
:
2
e
4c
0
/
2
, (3C.14)
with corresponding energy eigenvalue
1 =
/
2
2:
/
2
=
:
4
e
32
2
c
2
0
/
2
(3C.15)
Exercise 3C.3. (a) Verify that the function
0
in eqn.(3C.14) is a solution to E _
eqn.(3C.13), with 1 given by eqn.(3C.15).
(b) Verify that the function
0
denes a probability density, by checking that
_
X
[[
2
=
1.
There are many other, more complicated solutions to eqn.(3C.13). However,
eqn.(3C.14) is the simplest solution, and has the lowest energy eigenvalue 1
of any solution. In other words, the solution (3C.13) describes an electron in
the ground state: the orbital of lowest potential energy, where the electron is
closest to the nucleus.
This solution immediately yields two experimentally testable predictions:
(a) The ionization potential for the hydrogen atom, which is the energy re-
quired to ionize the atom, by stripping o the electron and removing it
to an innite distance from the nucleus.
(b) The Bohr radius of the hydrogen atom that is, the most probable
distance of the electron from the nucleus.
To see (a), recall that 1 is the sum of potential and kinetic energy for the
electron. We assert (without proof) that there exist solutions to the stationary
Schrodinger equation (3C.13) with energy eigenvalues arbitrarily close to zero
(note that 1 is negative). These zero-energy solutions represent orbitals where
the electron has been removed to some very large distance from the nucleus,
and the atom is essentially ionized. Thus, the energy dierence between these
ionized states and
0
is 10 = 1, and this is the energy necessary to ionize
the atom when the electron is in the orbital described by
0
.
By substituting in numerical values
e
1.6010
19
C, c
0
8.8510
12
C,N m
2
,
:
e
9.1110
31
kg, and / 1.054510
34
J s, the reader can verify that, in
fact, 1 2.1796 10
18
J 13.605 eV, which is very close to 13.595 eV,
the experimentally determined ionization potential for a hydrogen atom.
8
To see (b), observe that the probability density function for the distance : of
the electron from the nucleus is given by
1(:) = 4:
2
[(:)[
2
= 4/
3
:
2
exp(2/[x[).
(Exercise 3C.4). The mode of the radial probability distribution is the max- E _
imal point of 1(:); if we solve the equation 1
t
(:) = 0, we nd that the mode
occurs at
: :=
1
/
=
4c
0
/
2
:
e
2
e
5.29172 10
11
m.
The Balmer Lines. Recall that the spectrum of the Hamiltonian operator H
is the set of all eigenvalues of H. Let c = 1
0
< 1
1
< 1
2
< . . . be the spectrum
of the Hamiltonian of the hydrogen atom from Example 3C.6, with the elements
listed in increasing order. Thus, the smallest eigenvalue is 1
0
13.605, the
energy eigenvalue of the aforementioned ground state
0
. The other, larger
eigenvalues correspond to electron orbitals with higher potential energy.
When the electron falls from a high energy orbital (with eigenvalue 1
n
,
for some n N) to a low energy orbital (with eigenvalue 1
m
, where : < n),
it releases the energy dierence, and emits a photon with energy (1
n
1
m
).
Conversely, to jump from a low 1
m
-energy orbital to a higher 1
n
-energy orbital,
the electron must absorb a photon, and this photon must have exactly energy
(1
n
1
m
).
Thus, the hydrogen atom can only emit or absorb photons of energy [1
n

1
m
[, for some n, : N. Let c
t
:= [1
n
1
m
[ ; n, : N. We call c
t
the
energy spectrum of the hydrogen atom.
Plancks law says that a photon with energy 1 has frequency ) = 1,/, where
/ 6.626 10
34
J s is Plancks constant. Thus, if T = 1,/ ; 1 c
t
, then a
hydrogen atom can only emit/absorb a photon whose frequency is in T; we say
T is the frequency spectrum of the hydrogen atom.
8
The error of 0.01 eV is mainly due to our simplifying assumption of an immobile proton.
Here lies the explanation for the empirical observations of 19th century physi-
cists such as Balmer, Lyman, Rydberg, and Paschen, who found that an ener-
gized hydrogen gas has a distinct emission spectrum of frequencies at which it
emits light, and an identical absorption spectrum of frequencies which the gas
can absorb. Indeed, every chemical element has its own distinct spectrum; as-
tronomers use these spectral signatures to measure the concentrations of chem-
ical elements in the stars of distant galaxies. Now we see that
The (frequency) spectrum of an atom is determined by the (eigen-
value) spectrum of the corresponding Hamiltonian.
Further reading
Unfortunately, most other texts on partial dierential equations do not discuss
the Schrodinger equation; one of the few exceptions is the excellent text [Asm05].
For an lucid, fast, yet precise introduction to quantum mechanics in general, see
[McW72]. For a more comprehensive textbook on quantum theory, see [Boh79].
A completely dierent approach to quantum theory uses Feynmans path in-
tegrals; for a good introduction to this approach, see [Ste95], which also con-
tains excellent introductions to classical mechanics, electromagnetism, statistical
physics, and special relativity. For a rigorous mathematical approach to quantum
theory, an excellent introduction is [Pru81]; another source is [BEH94].
1. Let
1
1 be a constant. Consider the function : 1
3
1 C dened:
t
(r
1
, r
2
, r
3
) = exp
_
i
/
:
e
2
1
2
t
_
exp
_
i
/
:
e
1
r
1
_
.
Show that satises the (free) Schrodinger equation: i/
t
t
(x) =
/
2
2 :
e
t
(x).
2. Let v := (
1
,
2
,
3
) be a three-dimensional velocity vector, and let [v[
2
=
2
1
+
2
2
+
2
3
. Consider the function : 1
3
1 C dened:
t
(r
1
, r
2
, r
3
) = exp
_
i [v[
2
t,2
_
exp (i v x) .
Show that satises the (free) Schrodinger equation: i
t
=
1
2
.
3. Consider the stationary Schrodinger equation for a null potential:
H
0
= 1
0
, where H =
/
2
2:
e
.
Let 1 be a constant. Consider the function
0
: 1
3
C dened:
0
(r
1
, r
2
, r
3
) = exp
_
i
/
:
e
1
r
1
_
.
Show that
0
is a solution to the above stationary Schrodinger equation,
with eigenvalue 1 =
1
2
:
e
2
.
4. Exercise 3C.2(a) (page 51).
5. Exercise 3C.3(a) (page 52).
56
II General theory
57
Chapter 4
Linear partial dierential
equations
The Universe is a grand book which cannot be read until one rst learns the language in
which it is composed. It is written in the language of mathematics. Galileo Galilei
4A Functions and vectors
Prerequisites: 0A.
Vectors: If v =
_
_
2
7
3
_
_
and w =
_
_
1.5
3
1
_
_
, then we can add these two vectors
componentwise:
v +w =
_
_
2 1.5
7 + 3
3 + 1
_
_
=
_
_
0.5
10
2
_
_
.
4
2
1
-1
2
-3
0
4
2
2
.
9
1
.
2
3
.
8
1
3
.
2
5
1
.
1
0
.
7
0
-
1
.
8
-
3
.
1
-
3
.
8
-
0
.
6
8
1
.
1
3
.
0
(4, 2, 2.9, 1.2, 3.81, 3.25, 1.1, 0.7,
0, -1.8, -3.1, -3.8, -0.68, 1.1, 3.0)
(4, 2, 1, 0, -1,-3, 2)
N=7
N=15
N=oo
Figure 4A.1: We can think of a function as an innite-dimensional vector
58 DRAFT Chapter 4. Linear partial dierential equations
u = (4, 2, 1, 0, 1, 3, 2)
4
2
1
2
0
1
3
v = (1, 4, 3, 1, 2, 3, 1)
4
1
2
1
3
3
1
w = u + v = (5, 6, 4, 1, 3, 6, 3)
6
4
3
1
3
6
5
f(x) = x
g(x) = x - 3x + 2
h(x) = f(x) + g(x) = x - 2x +2
2
2
(A)
(B)
Figure 4A.2: (A) We add vectors componentwise: If u = (4, 2, 1, 0, 1, 3, 2)
and v = (1, 4, 3, 1, 2, 3, 1), then the equation w = v + w means that
w = (5, 6, 4, 1, 3, 6, 3). (B) We add two functions pointwise: If )(r) = r, and
p(r) = r
2
3r+2, then the equation / = )+p means that /(r) = )(r)+p(r) =
r
2
2r + 2 for every r.
In general, if v, w 1
3
, then u = v +w is dened by:
n
n
=
n
+n
n
, for n = 1, 2, 3 (4A.1)
(see Figure 4A.2A) Think of v as a function : 1, 2, 3 1, where (1) = 2,
(2) = 7, and (3) = 3. If we likewise represent w with n : 1, 2, 3 1
and u with n : 1, 2, 3 1, then we can rewrite eqn.(4A.1) as n(n) =
(n) + n(n) for n = 1, 2, 3. In a similar fashion, any -dimensional vector
u = (n
1
, n
2
, . . . , n
N
) can be thought of as a function n : [1...] 1.
Functions as Vectors: Letting go to innity, we can imagine any function
) : 1 1 as a sort of innite-dimensional vector (see Figure 4A.1). Indeed,
if ) and p are two functions, we can add them pointwise, to get a new function
/ = ) +p, where
/(r) = )(r) +p(r), for all r 1 (4A.2)
(see Figure 4A.2B) Notice the similarity between formulae (4A.2) and (4A.1),
and the similarity between Figures 4A.2A and 4A.2B.
One of the most important ideas in the theory of PDEs is that functions
are innite-dimensional vectors. Just as with nite vectors, we can add them
together, act on them with linear operators, or represent them in dierent co-
ordinate systems on innite-dimensional space. Also, the vector space 1
D
has
4B. Linear operators 59
a natural geometric structure; we can identify a similar geometry in innite
dimensions.
Let X 1
D
be some domain. The vector space of all continuous functions
from X into 1
m
is denoted ((X; 1
m
). That is:
((X; 1
m
) := ) : X 1
m
; ) is continuous .
When X and 1
m
are obvious from context, we may just write (.
Exercise 4A.1. Show that ((X; 1
m
) is a vector space. E _
A scalar eld ) : X 1 is innitely dierentiable (or smooth) if, for ev-
ery 0 and every i
1
, i
2
, . . . , i
N
[1...1], the th derivative
i
1
i
2

i
N
)(x)
exists at each x X. A vector eld ) : X 1
m
is innitely dieren-
tiable (or smooth) if )(x) := ()
1
(x), . . . , )
m
(x)), where each of the scalar elds
)
1
, . . . , )
m
: X 1 is innitely dierentiable. The vector space of all smooth
functions from X into 1
m
is denoted (
(X; 1
m
). That is:
(
(X; 1
m
) := ) : X 1
m
; ) is innitely dierentiable .
When X and 1
m
are obvious from context, we may just write (
.
Example 4A.1.
(a) (
(1
2
; 1) is the space of all smooth scalar elds on the plane (i.e. all
functions n : 1
2
1).
(b) (
(1; 1
3
) is the space of all smooth curves in three-dimensional space.
Exercise 4A.2. Show that (
(X; 1
m
) is a vector space, and thus, a linear subspace E _
of ((X; 1
m
).
4B Linear operators
Prerequisites: 4A.
4B(i) ...on nite dimensional vector spaces
Let v :=
_
2
7
and w :=
_
1.5
3
, and let u := v+w =

_
0.5
10
. If A :=
_
1 1
4 0
, then
A u = A v +A w. That is:
_
1 1
4 0
_
_
0.5
10
_
=
_
9.5
2
_
=
_
5
8
_
+
_
4.5
6
_
=
_
1 1
4 0
_
_
2
7
_
+
_
1 1
4 0
_
_
1.5
3
_
;
Also, if x = 3v =
_
6
21
, then Ax = 3Av. That is:

_
1 1
4 0
_
_
6
21
_
=
_
15
24
_
= 3
_
5
8
_
= 3
_
1 1
4 0
_
_
2
7
_
.
In other words, multiplication by the matrix A is a linear operator on the
vector space 1
2
. In general, a function 1 : 1
N
1
M
is linear if:
For all v, w 1
N
, we have 1(v +w) = 1(v) +1(w)
For all v 1
N
and : 1, we have 1(: v) = : 1(v).
Every linear function from 1
N
to 1
M
corresponds to multiplication by some
` matrix.
Example 4B.1.
(a) Dierence Operator: Suppose 1 : 1
5
1
4
is the function:
1
_
_
r
0
r
1
r
2
r
3
r
4
_
_
=
_
_
r
1
r
0
r
2
r
1
r
3
r
2
r
4
r
3
_
_
.
Then 1 corresponds to multiplication by the matrix
_
_
1 1
1 1
1 1
1 1
_
_
.
(b) Summation operator: Suppose o : 1
4
1
5
is the function:
o
_
_
r
1
r
2
r
3
r
4
_
_
=
_
_
0
r
1
r
1
+r
2
r
1
+r
2
+r
3
r
1
+r
2
+r
3
+r
4
_
_
Then o corresponds to multiplication by the matrix
_
_
0 0 0 0
1 0 0 0
1 1 0 0
1 1 1 0
1 1 1 1
_
_
.
(c) Multiplication operator: Suppose ` : 1
5
1
5
is the function
`
_
_
r
1
r
2
r
3
r
4
r
5
_
_
=
_
_
3 r
1
2 r
2
5 r
3
3
4
r
4
2 r
5
_
_
Then ` corresponds to multiplication by the matrix
_
_
3
2
5
3
4

2
_
_
.
Remark Notice that the transformation 1 is a left-inverse to the transforma-

tion o. That is, 1 o = Id. (However, 1 is not a right-inverse to o, because if
x = (r
0
, r
1
, . . . , r
4
), then o 1(x) = x (r
0
, r
0
, . . . , r
0
).
4B(ii) ...on (
Recommended: 1B, 1C, 2B.

A transformation L : (
is called a linear operator if, for any two

dierentiable functions ), p (
, we have L() +p) = L()) + L(p), and, for any

real number : 1, we have L(: )) = : L()).
Example 4B.2.
(a) Dierentiation: If ), p : 1 1 are dierentiable functions, and / =
) +p, then we know that, for any r 1,
/
t
(r) = )
t
(r) +p
t
(r).
Also, if / = : ), then /
t
(r) = : )
t
(r). Thus, if we dene the operation
D : (
(1; 1) (
(1; 1) by D[)] = )
t
, then D is a linear transformation
of (
(1; 1). For example, sin and cos are elements of (
(1; 1), and we

have
D[sin] = cos, and D[cos] = sin .
More generally, if ), p : 1
D
1 and / = ) +p, then for any i [1..1],
j
/ =
j
) +
j
p.
Also, if / = : ), then
j
/ = :
j
). In other words, the transformation
j
: (
(1
D
; 1) (
(1
D
; 1) is a linear operator.
(b) Integration: If ), p : 1 1 are integrable functions, and / = ) + p,
then we know that, for any r 1,
_
x
0
/(j) dj =
_
x
0
)(j) dj +
_
x
0
p(j) dj.
Also, if / = : ), then
_
x
0
/(j) dj = :
_
x
0
)(j) dj.
Thus, if we dene the operation S : (
(1; 1) (
(1; 1) by
S[)](r) =
_
x
0
)(j) dj, for all r 1.
then S is a linear transformation. For example, sin and cos are elements of
(
(1; 1), and we have

S[sin] = 1 cos, and S[cos] = sin .
(c) Multiplication: If : 1
D
1 is a scalar eld, then dene the operator
: (
by: [)] = ). In other words, for all x 1

D
, [)](x) =
(x) )(x). Then is a linear function, because, for any ), p (
,
[) +p] = [) +p] = ) + p = [)] + [p].
Remark. Notice that the transformation D is a left-inverse for the transfor-
mation S, because the Fundamental Theorem of Calculus says that D S()) = )
for any ) (
(1). However, D is not a right-inverse for S, because in general

S D()) = ) c, where c = )(0) is a constant.
Exercise 4B.1. Compare the three linear transformations in Example 4B.2 with E _
those from Example 4B.1. Do you notice any similarities?
Remark. Unlike linear transformations on 1
N
, there is in general no way
to express a linear transformation on (
in terms of multiplication by some

matrix. To convince yourself of this, try to express the three transformations
from example 4B.2 in terms of matrix multiplication.
Any combination of linear operations is also a linear operation. In particular,
any combination of dierentiation and multiplication operations is linear. Thus,
for example, the second-derivative operator D
2
[)] =
2
x
) is linear, and the
Laplacian operator
) =
2
1
) +. . . +
2
D
)
is also linear; in other words, [) +p] = ) +p.
A linear transformation that is formed by adding and/or composing multi-
plications and dierentiations is called a linear dierential operator . For
example, the Laplacian is a linear dierential operator.
4B(iii) Kernels
If 1 is a linear function, then the kernel of 1 is the set of all vectors v such that
1(v) = 0.
Example 4B.3.
(a) Consider the dierentiation operator
x
on the space (
(1; 1). The kernel

of
x
is the set of all functions n : 1 1 such that
x
n 0 in other
words, the set of all constant functions.
(b) The kernel of
2
x
is the set of all functions n : 1 1 such that
2
x
n 0
in other words the set of all at functions of the form n(r) = or+/.
Many partial dierential equations are really equations for the kernel of some
dierential operator.
Example 4B.4.
(a) Laplaces equation n 0 really just says: n is in the kernel of
.
(b) The heat equation
t
n = n really just says: n is in the kernel of
the operator L =
t
.
4B(iv) Eigenvalues, eigenvectors, and eigenfunctions
If L is a linear operator on some vector space, then an eigenvector of L is a
vector v such that
L(v) = v,
for some constant C, called the associated eigenvalue.
Example 4B.5. If 1 : 1
2
1
2
is dened by the matrix
_
0 1
1 0
and v =
_
1
1
,
then 1(v) =
_
1
1
_
= v, so v is an eigenvector for 1, with eigenvalue
= 1.
If L is a linear operator on (
, then an eigenvector of L is sometimes called an

eigenfunction.
Example 4B.6. Let n, : N. Dene n(r, j) = sin(n r) sin(: j). Then
n(r, j) = (n
2
+:
2
) sin(n r) sin(: j) = n(r, j),
where = (n
2
+:
2
). Thus, n is an eigenfunction of the linear operator ,
with eigenvalue . (Exercise 4B.2 Verify the these claims.) E _
Eigenfunctions of linear dierential operators (particularly, eigenfunctions of )
play a central role in the solution of linear PDEs. This is implicit in Chapters
11-14 and 20, and is made explicit in Chapter 15.
4C Homogeneous vs. nonhomogeneous
Prerequisites: 4B.
If L : (
is a linear dierential operator, then the equation Ln 0

is called a homogeneous linear partial dierential equation.
Example 4C.1. The following are linear homogeneous PDEs. Here X 1
D
is some domain.
(a) Laplaces Equation
1
: Here, (
= (
(X; 1), and L = .

(b) heat equation
2
: (
= (
(X 1; 1), and L =
t
.
(c) wave equation
3
: (
= (
(X 1; 1), and L =
2
t
.
(d) Schrodinger Equation
4
: (
= (
(1
3N
1; C), and, for any (
and (x; t) 1
3N
1, L (x; t) :=
/
2
2
(x; t) + \ (x) (x; t)
i/
t
(x; t). (Here, \ : 1
3N
1 is some potential function, and is
like a Laplacian operator, except that the components for each particle are
divided by the rest mass of that particle.)
(e) Fokker-Plank
5
: (
= (
(X 1; 1), and, for any n (
,
L(n) =
t
n n +

V n + n div

V.
Linear homogeneous PDEs are nice because we can combine two solutions
together to obtain a third solution.
Example 4C.2.
(a) Let n(r; t) = 7 sin[2t + 2r] and (r; t) = 3 sin[17t + 17r] be two travelling
wave solutions to the wave equation. Then n(r; t) = n(r; t) + (r; t) =
7 sin(2t +2r) +3 sin(17t +17r) is also a solution (see Figure 4C.1). To use
a musical analogy: if we think of n and as two pure tones, then we can
think of n as a chord.
1
See 1C on page 9.
2
See 1B on page 5.
3
See 2B on page 27.
4
See 3B on page 41.
5
See 1F on page 18.
4C. Homogeneous vs. nonhomogeneous 65
1
0.5
0
-0.5
-1
x
3 2 1 0 -1 -2 -3 +
1
0.5
0
-0.5
-1
x
3 2 1 0 -1 -2 -3 =
1
0.5
0
-0.5
-1
x
3 2 1 0 -1 -2 -3
n(r, t) = 7 sin(2t + 2r) (r; t) = 3 sin(17t + 17r) n(r, t) = n(r; t) +(r; t)
Figure 4C.1: Example 4C.2(a).
(b) Let )(r; t) =
1
2
t
exp
_
r
2
4t
_
, p(r; t) =
1
2
t
exp
_
(r 3)
2
4t
_
,
and /(r; t) =
1
2
t
exp
_
(r 5)
2
4t
_
be one-dimensional Gauss-Weierstrass
kernels, centered at 0, 3, and 5, respectively. Thus, ), p, and / are all so-
lutions to the heat equation. Then, 1(r) = )(r) + 7 p(r) + /(r) is also
a solution to the heat equation. If a Gauss-Weierstrass kernel models the
erosion of a single mountain, then the function 1 models the erosion of
a little mountain range, with peaks at 0, 3, and 5, and where the middle
peak is seven times higher than the other two.
These examples illustrate a general principle:
Theorem 4C.3. Superposition Principle for homogeneous Linear PDEs
Suppose L is a linear dierential operator, and n
1
, n
2
(
are solutions to the

homogeneous linear PDE Ln = 0. Then, for any c
1
, c
2
1, n = c
1
n
1
+c
2
n
2
is also a solution.
Proof. Exercise 4C.1 2 E _
If (
is some xed nonzero function, then the equation Lj is

called a nonhomogeneous linear partial dierential equation.
Example 4C.4. The following are linear nonhomogeneous PDEs
(a) The antidierentiation equation j
t
= is familiar from rst year
calculus. The Fundamental Theorem of Calculus says that one solution to
this equation is the integral function j(r) =
_
x
0
(j) dj.
(b) The Poisson Equation
6
, j = , is a nonhomogeneous linear PDE.
Recall Examples 1D.1 and 1D.2 on page 14, where we obtained new solutions
to a nonhomogeneous equation by taking a single solution, and adding solutions
of the homogeneous equation to this solution. These examples illustrates a gen-
eral principle:
Theorem 4C.5. Subtraction Principle for nonhomogeneous linear PDEs
Suppose L is a linear dierential operator, and (
. Let j
1
(
be a
solution to the nonhomogeneous linear PDE Lj
1
= . If / (
is any solution
to the homogeneous equation (i.e. L/ = 0), then j
2
= j
1
+/ is another solution
to the nonhomogeneous equation. In summary:
_
Lj
1
= ; L/ = 0; and j
2
= j
1
+/.
_
=
_
Lj
2
=
_
.
If P : (
is not a linear operator, then a PDE of the form Pn 0

or Pn p is called a nonlinear PDE. For example, if 1 : 1
D
1
D
is
some nonlinear rate function describing chemical reactions, then the reaction-
diusion equation
7
t
u = u + 1(u),
is a nonlinear PDE, corresponding to the nonlinear dierential operator P(u) :=
t
u u 1(u).
The theory of linear partial dierential equations is relatively simple, because
solutions to linear PDEs interact in very nice ways, as shown by Theorems 4C.3
and 4C.5. The theory of nonlinear PDEs is much more complicated; furthermore,
many of the methods which do exist for solving nonlinear PDEs involve somehow
approximating them with linear ones. In this book we shall concern ourselves
only with linear PDEs.
1. For each of the following equations: n is an unknown function; is always
some xed, predetermined function; and is always a constant.
In each case, is the equation linear? If it is linear, is it homogeneous?
Justify your answers.
6
See 1D on page 12
7
See 1G on page 19
(a) heat equation:
t
n(x) = n(x).
(b) Poisson Equation: n(x) = (x).
(c) Laplace Equation: n(x) = 0.
(d) Monge-Ampère Equation: (r, j) = det
_

2
x
n(r, j)
x
y
n(r, j)
y
n(r, j)
2
y
n(r, j)
_
.
(e) Reaction-Diusion
t
n(x; t) = n(x; t) +
_
n(x; t)
_
.
(f) Scalar conservation Law
t
n(r; t) =
x
( n)(r; t).
(g) Helmholtz Equation: n(x) = n(x).
(h) Airys Equation:
t
n(r; t) =
3
x
n(r; t).
(i) Beam Equation:
t
n(r; t) =
4
x
n(r; t).
(j) Schrodinger Equation:
t
n(x; t) = i n(x; t) + (x; t) n(x; t).
(k) Burgers Equation:
t
n(r; t) = n(r; t)
x
n(r; t).
(l) Eikonal Equation: [
x
n(r)[ = 1.
2. Which of the following are eigenfunctions for the 2-dimensional Laplacian
=
2
x
+
2
y
? In each case, if n is an eigenfunction, what is the eigenvalue?
(a) n(r, j) = sin(r) sin(j) (Figure 5F.1(A) on page 100)
(b) n(r, j) = sin(r) + sin(j) (Figure 5F.1(B) on page 100)
(c) n(r, j) = cos(2r) + cos(j) (Figure 5F.1(C) on page 100)
(d) n(r, j) = sin(3r) cos(4j).
(e) n(r, j) = sin(3r) + cos(4j).
(f) n(r, j) = sin(3r) + cos(3j).
(g) n(r, j) = sin(3r) cosh(4j).
(h) n(r, j) = sinh(3r) cosh(4j).
(i) n(r, j) = sinh(3r) + cosh(4j).
(j) n(r, j) = sinh(3r) + cosh(3j).
(k) n(r, j) = sin(3r + 4j).
(l) n(r, j) = sinh(3r + 4j).
(m) n(r, j) = sin
3
(r) cos
4
(j).
(n) n(r, j) = c
3x
c
4y
.
(o) n(r, j) = c
3x
+c
4y
.
(p) n(r, j) = c
3x
+c
3y
.
69
Chapter 5
Classication of PDEs and
problem types
If one looks at the dierent problems of the integral calculus which arise naturally when
one wishes to go deep into the dierent parts of physics, it is impossible not to be struck by
the analogies existing. Whether it be electrostatics or electrodynamics, the propogation of heat,
optics, elasticity, or hydrodynamics, we are led always to dierential equations of the same
family. Henri Poincare
5A Evolution vs. nonevolution equations
Recommended: 1B, 1C, 2B, 4B.
An evolution equation is a PDE with a distinguished time coordinate,
t. In other words, it describes functions of the form n(x; t), and the equation
has the form:
D
t
n = D
x
n
where D
t
is some dierential operator involving only derivatives in the t variable
(e.g.
t
,
2
t
, etc.), while D
x
is some dierential operator involving only derivatives
in the x variables (e.g.
x
,
2
y
, , etc.)
Example 5A.1. The following are evolution equations:
(a) The heat equation
t
n = n of 1B.
(b) The wave equation
2
t
n = n of 2B.
(c) The telegraph equation
2
2
t
n +
1
t
n =
0
n + n of 2C.
(d) The Schrodinger equation
t
=
1
i/
H of 3B (here H is a Hamiltonian
operator).
70 DRAFT Chapter 5. Classication of PDEs and problem types
(e) Liouvilles Equation, the Fokker-Plank equation, and Reaction-Diusion
Equations.
Nonexample 5A.2. The following are not evolution equations:
(a) The Laplace Equation n = 0 of 1C.
(b) The Poisson Equation n = of 1D.
(c) The Helmholtz Equation n = n (where C is a constant i.e. an
eigenvalue of ).
(d) The Stationary Schrodinger equation H
0
= 1
0
(where 1 C is a
constant eigenvalue).
In mathematical models of physical phenomena, most PDEs are evolution equa-
tions. Nonevolutionary PDEs generally arise as stationary state equations for
evolution PDEs (e.g. Laplaces equation) or as resonance states (e.g. Sturm-
Liouville, Helmholtz).
Order: The order of the dierential operator
2
x
3
y
is 2+3 = 5. More generally,
the order of the dierential operator
k
1
1

k
2
2
. . .
k
D
D
is the sum /
1
+ . . . + /
D
.
The order of a general dierential operator is the highest order of any of its
terms. For example, the Laplacian is second order. The order of a PDE is the
highest order of the dierential operator that appears in it. Thus, the Transport
Equation, Liouvilles Equation, and the (nondiusive) Reaction Equation is rst
order, but all the other equations we have looked at (the heat equation, the wave
equation, etc.) are of second order.
5B Initial value problems
Prerequisites: 5A.
Let X 1
D
be some domain, and let L be a dierential operator on (
(X; 1).
Consider evolution equation
t
n = L n, (5B.1)
for an unknown function n : X 1
,
1. An initial value problem (IVP)
for equation (5B.1) is the following problem:
Given some function )
0
: X 1 (the initial conditions), nd
a continuous function n : X 1
,
1 which satises (5B.1) and
also satises n(x, 0) = )
0
(x), for all x X.
5C. Boundary value problems 71
For example, suppose the domain X is an iron pan being heated on a gas
ame stove. You turn o the ame (so there is no further heat entering the
system) and then throw some vegetables into the pan. Thus, (5B.1) is the Heat
Equation, and )
0
describes the initial distribution of heat: cold vegetables in a
hot pan. The initial value problem asks: How fast do the vegetables cook? How
fast does the pan cool?
Next, consider the second order-evolution equation
2
t
n = L n, (5B.2)
for a unknown function n : X1
,
1. An initial value problem (or IVP,
or Cauchy problem) for (5B.2) is as follows:
Given a function )
0
: X 1 (the initial position), and/or
another function )
1
: X 1 (the initial velocity), nd a con-
tinuously dierentiable function n : X 1
,
1 which satises
(5B.2) and also satises n(x, 0) = )
0
(x) and
t
n(x, 0) = )
1
(x), for
all x X.
For example, suppose (5B.1) is the wave equation on X = [0, 1]. Imagine
[0, 1] as a vibrating string. Thus, )
0
describes the initial displacement of the
string, and )
1
its initial momentum.
If )
0
, 0, and )
1
0, then the string is initially at rest, but is released from
a displaced state in other words, it is plucked (e.g. in a guitar or a harp).
Hence, the initial value problem asks: How does a guitar string sound when it
is plucked?
On the other hand, if )
0
0, and )
1
, 0, then the string is initially at, but
is imparted with nonzero momentum in other words, it is struck (e.g. by the
hammer in the piano). Hence, the initial value problem asks: How does a piano
string sound when it is struck?
5C Boundary value problems
Prerequisites: 0D, 1C. Recommended: 5B.
If X 1
D
is a nite domain, then X denotes its boundary. The interior
of X is the set int (X) of all points in X not on the boundary.
Example 5C.1.
(a) If " = [0, 1] 1 is the unit interval, then " = 0, 1 is a two-point set,
and int (") = (0, 1).
(b) If X = [0, 1]
2
1
2
is the unit square, then int (X) = (0, 1)
2
. and
X = (r, j) X ; r = 0 or r = 1 or j = 0 or j = 1 .
(c) In polar coordinates on 1
2
, let | = (:, ) ; : 1, [, ) be the
unit disk. Then | = (1, ) ; [, ) is the unit circle, and
int (|) = (:, ) ; : < 1, [, ).
(d) In spherical coordinates on 1
3
, let B =
_
x 1
3
; |x| 1
_
be the 3-
dimensional unit ball in 1
3
. Then B = S :=
_
x 1
D
; |x| = 1
_
is
the unit sphere, and int (B) =
_
x 1
D
; |x| < 1
_
.
(e) In cylindrical coordinates on 1
3
, let X = (:, , .) ; : 1, , 0 . 1
be the nite cylinder in 1
3
. Then X = (:, , .) ; : = 1 or . = 0 or . = 1.
A boundary value problem (BVP) is a problem of the following kind:

Find a continuous function n : X 1 such that
1. n satises some PDE at all x in the interior of X.
2. n also satises some other equation (maybe a dierential equa-
tion) for all s on the boundary of X.
The condition n must satisfy on the boundary of X is called a boundary
condition. Note that there is no time variable in our formulation of a BVP;
thus, typically the PDE in question is an equilibrium equation, like the Laplace
equation or the Poisson equation.
If we try to solve an evolution equation with specied initial conditions and
specied boundary conditions, then we are confronted with an initial/boundary
value problem. Formally, an initial/boundary value problem (I/BVP) is a
problem of the following kind:
Find a continuous function n : X 1
,
1 such that
1. n satises some (evolution) PDE at all x in the interior of
X 1
,
.
2. n satises some boundary condition for all (s; t) in (X) 1
,
.
3. n(x; 0) also satises some initial condition (as described in 5B)
for all x X.
We will consider four kinds of boundary conditions: Dirichlet, Neumann,
Mixed, and Periodic. Each of these boundary conditions has a particular physical
interpretation, and yields particular kinds of solutions for a partial dierential
equation.
0
0.05
0.1
0.15
0.2
0.25
0.2 0.4 0.6 0.8 1
x
Figure 5C.1: )(r) = r(1 r) satises homogeneous Dirichlet boundary condi-
tions on the interval [0, 1].
5C(i) Dirichlet boundary conditions
Let X be a domain, and let n : X 1 be a function. We say that n satises
homogeneous Dirichlet boundary conditions (HDBC) on X if:
For all s X, n(s) 0.
Physical interpretation.
Thermodynamic. (Heat equation, Laplace Equation, or Poisson Equation) In
this case, n represents a temperature distribution. We imagine that the
domain X represents some physical object, whose boundary X is made
out of metal or some other material which conducts heat almost perfectly.
Hence, we can assume that the temperature on the boundary is always equal
to the temperature of the surrounding environment.
We further assume that this environment has a constant temperature T
E
(for example, X is immersed in a bath of some uniformly mixed uid),
which remains constant during the experiment (for example, the uid is
present in large enough quantities that the heat owing into/out of X
does not measurably change it). We can then assume that the ambient
temperature is T
E
0, by simply subtracting a constant temperature of
T
E
o the inside and the outside. (This is like changing from measuring
temperature in degrees Kelvin to measuring in degrees Celsius; youre just
adding 273
o
to both sides, which makes no mathematical dierence.)
Electrostatic. (Laplace equation or Poisson Equation) In this case, n repre-
sents an electrostatic potential. The domain X represents some compart-
ment or region in space, whose boundary X is made out of metal or some
other perfect electrical conductor. Thus, the electrostatic potential within
the metal boundary is a constant, which we can normalize to be zero.
Acoustic. (Wave equation) In this case, n represents the vibrations of some vi-
brating medium (e.g. a violin string or a drum skin). Homogeneous Dirich-
1
0
1
x
0.5
0
0.5
1
y
0
0.2
0.4
0.6
0.8
1
z
1
0.5
0
0.5
1
x
1
0.5
0
0.5
y
0
0.2
0.4
0.6
0.8
1
z
(A) (B)
Figure 5C.2: (A) )(:, ) = 1 : satises homogeneous Dirichlet boundary con-
ditions on the disk | = (:, ) ; : 1, but is not smooth at zero. (B)
)(:, ) = 1 :
2
satises homogeneous Dirichlet boundary conditions on the disk
| = (:, ) ; : 1, and is smooth everywhere.
let boundary conditions mean that the medium is xed on the boundary
X (e.g. a violin string is clamped at its endpoints; a drumskin is pulled
down tightly around the rim of the drum).
The set of innitely dierentiable functions from X to 1 which satisfy homoge-
neous Dirichlet Boundary Conditions will be denoted (
0
(X; 1) or (
0
(X). Thus,
for example
(
0
[0, 1] =
_
) : [0, 1] 1; ) is smooth, and )(0) = 0 = )(1)
_
The set of continuous functions from X to 1 which satisfy homogeneous Dirichlet
Boundary Conditions will be denoted (
0
(X; 1) or (
0
(X).
Example 5C.2.
(a) Suppose X = [0, 1], and ) : X 1 is dened by )(r) = r(1 r). Then
)(0) = 0 = )(1), and ) is smooth, so ) (
0
[0, 1]. (See Figure 5C.1).
(b) Let X = [0, ].
1. For any n N, let S
n
(r) = sin(n r) (see Figure 6D.1 on page 113).
Then S
n
(
0
[0, ].
2. If )(r) = 5 sin(r) 3 sin(2r) + 7 sin(3r), then ) (
0
[0, ]. More
generally, any nite sum
N
n=1
1
n
S
n
(r) (for some constants 1
n
) is in
(
0
[0, ].
3. If )(r) =
n=1
1
n
S
n
(r) is a uniformly convergent Fourier sine series
1
,
then ) (
0
[0, ].
(c) Let | = (:, ) ; : 1 be the unit disk. Let ) : | 1 be the cone in
Figure 5C.2(A), dened: )(:, ) = (1:). Then ) is continuous, and ) 0
on the boundary of the disk, so ) satises Dirichlet boundary conditions.
Thus, ) (
0
(|). However, ) is not smooth (it is nondierentiable at zero),
so ) , (
0
(|).
(d) Let ) : | 1 be the dome in Figure 5C.2(B), dened )(:, ) = 1 :
2
.
Then ) (
0
(|).
(e) Let X = [0, ] [0, ] be the square of sidelength .
1. For any (n, :) N
2
, let S
n,m
(r, j) = sin (n r) sin (: j). Then
S
n,m
(
0
(X). (see Figure 9A.2 on page 181).
2. If )(r) = 5 sin(r) sin(2j) 3 sin(2r) sin(7j) + 7 sin(3r) sin(j), then
) (
0
(X). More generally, any nite sum
N
n=1
M
m=1
1
n,m
S
n,m
(r) is in
(
0
(X).
3. If ) =
n,m=1
1
n,m
S
n,m
is a uniformly convergent two dimensional
Fourier sine series
2
, then ) (
0
(X).
Exercise 5C.1. (i) Verify examples (b) to (e) above E _

(ii) Show that (
0
(X) is a vector space.
(iii) Show that (
0
(X) is a vector space.
Arbitrary nonhomogeneous Dirichlet boundary conditions are im-
posed by xing some function / : X 1, and then requiring:
n(s) = /(s), for all s X. (5C.3)
For example, the classical Dirichlet Problem is to nd a continuous function
n : X 1 satisfying the Dirichlet condition (5C.3), such that n also satises
Laplaces Equation: n(x) = 0 for all x int (X).
1
See 7B on page 144.
2
See 9A on page 179.
Physical interpretations.
Thermodynamic. n describes a stationary temperature distribution on X,
where the temperature is xed on the boundary. Dierent parts of the
boundary may have dierent temperatures, so heat may be owing through
the region X from warmer boundary regions to cooler boundary regions.
But the actual temperature distribution within X is in equilibrium.
Electrostatic. n describes an electrostatic potential eld within the region X.
The voltage level on the boundaries is xed (e.g. boundaries of X are wired
up to batteries which maintain a constant voltage). However dierent parts
of the boundary may have dierent voltages (the boundary is not a perfect
conductor).
Minimal surface. n describes a minimal-energy surface (e.g. a soap lm).
The boundary of the surface is clamped in some position (e.g. the wire
frame around the soap lm); the interior of the surface must adapt to
nd the minimal energy conguration compatible with these boundary
conditions. Minimal surfaces of low curvature are well-approximated by
harmonic functions.
For example, if X = [0, 1], and /(0) and /(1) are two constants, then the
Dirichlet Problem is to nd n : [0, 1] 1 such that
n(0) = /(0), n(1) = /(1), and
2
x
n(r) = 0, for 0 < r < 1. (5C.4)
That is, the temperature at the left-hand endpoint is xed at /(0), and at the
right-hand endpoint is xed at /(1). The unique solution to this problem is the
function n(r) =
_
/(1) /(0)
_
r,1 +/(0). (Exercise 5C.2). E _
5C(ii) Neumann boundary conditions
Suppose X is a domain with boundary X, and n : X 1 is some function.
Then for any boundary point s X, we use
n(s) to denote the outward

normal derivative
3
of n on the boundary. Physically,
n(s) is the rate of change

in n as you leave X by passing through X in a perpendicular direction.
Example 5C.3.
(a) If X = [0, 1], then
n(0) =
x
n(0) and
n(1) =
x
n(1).
(b) Suppose X = [0, 1]
2
1
2
is the unit square, and (r, j) X. There are
four cases:
3
This is sometimes indicated as
u
n
or
u
, or as u

N, or as u n.
If r = 0 (left edge), then
n(0, j) =
x
n(0, j).
If r = 1 (right edge), then
n(1, j) =
x
n(1, j).
If j = 0 (top edge), then
n(r, 0) =
y
n(r, 0).
If j = 1 (bottom edge), then
n(r, 1) =
y
n(r, 1).
(If more than one of these conditions is true for example, at (0, 0) then
(r, j) is a corner, and
n(r, j) is not well-dened).

(c) Let | = (:, ) ; : < 1 be the unit disk in the plane. Then | is the set
(1, ) ; [, ), and for any (1, ) |,
n(1, ) =
r
n(1, ).
(d) Let | = (:, ) ; : < 1 be the disk of radius 1. Then | = (1, ) ; [, ),
and for any (1, ) |,
n(1, ) =
r
n(1, ).
(e) Let B = (:, , ) ; : < 1 be the unit ball in 1
3
. Then B = (:, , ) ; : = 1
is the unit sphere. If n(:, , ) is a function in polar coordinates, then for
any boundary point s = (1, , ),
n(s) =
r
n(s).
(f) Suppose X = (:, , .) ; : 1, 0 . 1, < , is the nite cylin-
der, and (:, , .) X. There are three cases:
If : = 1 (sides), then
n(1, , .) =
r
n(1, , .).
If . = 0 (bottom disk), then
n(:, , 0) =
z
n(:, , 0).
If . = 1 (top disk), then
n(:, , 1) =
z
n(:, , 1).
We say that n satises homogeneous Neumann boundary conditions

if
n(s) = 0 for all s X. (5C.5)

Physical Interpretations.
Thermodynamic. (Heat, Laplace, or Poisson equation) Suppose n represents
a temperature distribution. Recall that Fouriers Law of Heat Flow ( 1A
on page 3) says that n(s) is the speed and direction in which heat is
owing at s. Recall that
n(s) is the component of n(s) which is per-

pendicular to X. Thus, homogeneous Neumann BC means that n(s) is
parallel to the boundary for all s X. In other words no heat is crossing
the boundary. This means that the boundary is a perfect insulator.
If n represents the concentration of a diusing substance, then n(s) is the
ux of this substance at s. Homogeneous Neumann Boundary conditions
mean that the boundary is an impermeable barrier to this substance.
Electrostatic. (Laplace or Poisson equation) Suppose n represents an electric
potential. Thus n(s) is the electric eld at s. Homogeneous Neumann
BC means that n(s) is parallel to the boundary for all s X; i.e. no
eld lines penetrate the boundary.
The set of continuous functions fromX to 1 which satisfy homogeneous Neumann
boundary conditions will be denoted (
(X). The set of innitely dierentiable

functions from X to 1 which satisfy homogeneous Neumann boundary conditions
will be denoted (
(X). Thus, for example

(
[0, 1] =
_
) : [0, 1] 1; ) is smooth, and )
t
(0) = 0 = )
t
(1)
_
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.2 0.4 0.6 0.8 1
x
1
0
1
x
0.5
0
0.5
1
y
0
0.2
0.4
0.6
0.8
1
z
(A) (B)
Figure 5C.3: (A) f(x) =
1
2
x
2
1
3
x
3
satses homogeneous Neumann boundary conditions
on the interval [0, 1]. (B) f(r, ) = (1 r)
2
satises homogeneous Neumann boundary
conditions on the disk | = |(r, ) ; r 1, but is not dierentiable at zero.
Example 5C.4.
(a) Let X = [0, 1], and let ) : [0, 1] 1 be dened by )(r) =
1
2
r
2
1
3
r
3
(See
Figure 5C.3(A)). Then )
t
(0) = 0 = )
t
(1), and ) is smooth, so ) (
[0, 1].
(b) Let X = [0, ].
1. For any n N, let C
n
(r) = cos (n r) (see Figure 6D.1 on page 113).
Then C
n
(
[0, ].
2. If )(r) = 5 cos(r) 3 cos(2r) + 7 cos(3r), then ) (
[0, ]. More
generally, any nite sum
N
n=1
n
C
n
(r) (for some constants
n
) is in
(
[0, ].
3. If )(r) =
n=1
n
C
n
(r) is a uniformly convergent Fourier cosine se-
ries
4
, and the derivative series )
t
(r) =
n=1
n
n
S
n
(r) is also uni-
formly convergent, then ) (
[0, ].
(c) Let | = (:, ) ; : 1 be the unit disk.
1. Let ) : | 1 be the witchs hat of Figure 5C.3(B), dened:
)(:, ) := (1 :)
2
. Then
) 0 on the boundary of the disk, so )

satises Neumann boundary conditions. Also, ) is continuous on |;
hence ) (
(|). However, ) is not smooth (it is nondierentiable at

zero), so ) , (
(|).
1
0.5
0
0.5
1
x
1
0
y
0
0.2
0.4
0.6
0.8
1
z
1 0.5 0 0.5 1
1
0
0
0.5
1
1.5
2
(A) (B)
Figure 5C.4: (A) f(r, ) = (1 r
2
)
2
satises homogeneous Neumann boundary conditions
on the disk, and is smooth everywhere. (B) f(r, ) = (1 + cos()
2
) (1 (1 r
2
)
4
) does
not satisfy homogeneous Neumann boundary conditions on the disk, and is not constant on the
boundary.
2. Let ) : | 1 be the bell of Figure 5C.4(A), dened: )(:, ) :=
(1:
2
)
2
. Then
) 0 on the boundary of the disk, and ) is smooth

everywhere on |, so ) (
(|).
3. Let ) : | 1 be the ower vase of Figure 5C.4(B), dened
)(:, ) := (1 +cos()
2
) (1 (1 :
2
)
4
). Then
) 0 on the bound-
ary of the disk, and ) is smooth everywhere on |, so ) (
(|).
Note that, in this case, the angular derivative is nonzero, so ) is not
constant on the boundary of the disk.
(d) Let X = [0, ] [0, ] be the square of sidelength .
1. For any (n, :) N
2
, let C
n,m
(r, j) = cos(nr) cos(:j) (see Fig-
ure 9A.2 on page 181). Then C
n,m
(
(X).
4
See 7B on page 144.
2. If )(r) = 5 cos(r) cos(2j) 3 cos(2r) cos(7j) + 7 cos(3r) cos(j), then
) (
(X). More generally, any nite sum

N
n=1
M
m=1
n,m
C
n,m
(r) (for
some constants
n,m
) is in (
(X).
3. More generally, if ) =
n,m=0
n,m
C
n,m
is a uniformly convergent two
dimensional Fourier cosine series
5
, and the derivative series
x
)(r, j) =
n,m=0
n
n,m
sin(nr) cos(:j)
y
)(r, j) =
n,m=0
:
n,m
cos(nr) sin(:j)
are also uniformly convergent, then ) (
(X).
Exercise 5C.3 Verify examples (b) to (d) E _
Arbitrary nonhomogeneous Neumann Boundary conditions are imposed
by xing a function / : X 1, and then requiring
n(s) = /(s) for all s X. (5C.6)

For example, the classical Neumann Problem is to nd a continuously dif-
ferentiable function n : X 1 satisfying the Neumann condition (5C.6), such
that n also satises Laplaces Equation: n(x) = 0 for all x int (X).
Physical Interpretations.
Thermodynamic. Here n represents a temperature distribution, or the con-
centration of some diusing material. Recall that Fouriers Law ( 1A
on page 3) says that n(s) is the ux of heat (or material) at s. Thus,
for any s X, the derivative
n(s) is the ux of heat/material across

the boundary at s. The nonhomogeneous Neumann Boundary condition
n(s) = /(s) means that heat (or material) is being pumped across the
boundary at a constant rate described by the function /(s).
Electrostatic. Here, n represents an electric potential. Thus n(s) is the elec-
tric eld at s. Nonhomogeneous Neumann boundary conditions mean that
the eld vector perpendicular to the boundary is determined by the func-
tion /(s).
5
See 9A on page 179.
5C(iii) Mixed (or Robin) boundary conditions
These are a combination of Dirichlet and Neumann-type conditions obtained as
follows: Fix functions / : X 1, and /, /
: X 1. Then (/, /
, /)-
mixed boundary conditions are given:
/(s) n(s) + /
(s)
n(s) = /(r) for all s X. (5C.7)

For example:
Dirichlet Conditions corresponds to / 1 and /
0.
Neumann Conditions corresponds to / 0 and /
1.
No boundary conditions corresponds to / /
0.
Newtons Law of Cooling reads:
n = c (n T
E
) (5C.8)
This describes a situation where the boundary is an imperfect conductor
(with conductivity constant c), and is immersed in a bath with ambient
temperature T
E
. Thus, heat leaks in or out of the boundary at a rate
proportional to c times the dierence between the internal temperature n
and the external temperature T
E
. Equation (5C.8) can be rewritten:
c n
n = /,
where / = c T
E
. This is the mixed boundary equation (5C.7), with / c
and /
1.
Homogeneous mixed boundary conditions take the form:
/ n + /
n 0.
The set of functions in (
(X) satisfying this property will be denoted

(
h,h
(X). Thus, for example, if X = [0, 1], and /(0), /
(0), /(1) and

/
(1) are four constants, then

(
h,h
[0, 1] =
_
) : [0, 1] 1; ) is dierentiable, /(0))(0) /
(0))
t
(0) = 0
and /(1))(1) + /
(1))
t
(1) = 0.
_
Remarks. (a) Note that there is some redundancy in this formulation. Equa-
tion (5C.7) is equivalent to
/ /(s) n(s) + / /
(s)
n(s) = / /(s),
for any constant / ,= 0. Normally we chose / so that at least one of the coecients
/ or /
is equal to 1.
(b) Some authors (e.g. [Pin98]) call this general boundary conditions,
and, for mathematical convenience, write this as
cos()n + 1 sin()
n = T. (5C.9)
where and and T are parameters. Here, the cos(), sin() coecients
of (5C.9) are just a mathematical gadget to concisely express any weighted
combination of Dirichlet and Neumann conditions. An expression of type (5C.7)
can be transformed into one of type (5C.9) as follows: Let := arctan
_
/
1 /
_
(if / = 0, then set =

2
) and let T := /
cos() +1sin()
/ +/
. Going the other way

is easier; simply dene / := cos(), /
:= 1 sin(), and T := /.
5C(iv) Periodic boundary conditions
Periodic boundary conditions means that function n looks the same on oppo-
site edges of the domain. For example, if we are solving a PDE on the interval
[, ], then periodic boundary conditions are imposed by requiring
n() = n() and n
t
() = n
t
().
Interpretation #1: Pretend that n is actually a small piece of an innitely
extended, periodic function n : 1 1, where, for any r 1 and n Z, we
have:
n(r + 2n) = n(r).
Thus n must have the same value and the same derivative at r and r+2n,
for any r 1. In particular, n must have the same value and derivative at
and . This explains the name periodic boundary conditions.
Interpretation #2: Suppose you glue together the left and right ends of the
interval [, ] (i.e. glue to ). Then the interval looks like a a circle (where
and actually become the same point). Thus n must have the same value
and the same derivative at and .
Example 5C.5.
(a) n(r) = sin(r) and (r) = cos(r) have periodic boundary conditions.
(b) For any n N, the functions S
n
(r) = sin(nr) and C
n
(r) = cos(nr) have
periodic boundary conditions. (See Figure 6D.1 on page 113.)
(c) sin(3r) + 2 cos(4r) has periodic boundary conditions.
Figure 5C.5: If we glue the opposite edges of a square together, we get a torus.
(d) If n
1
(r) and n
2
(r) have periodic boundary conditions, and c
1
, c
2
are any
constants, then n(r) = c
1
n
1
(r) + c
2
n
2
(r) also has periodic boundary con-
ditions.
Exercise 5C.4 Verify these examples. E _
On the square [, ] [, ], periodic boundary conditions are imposed
by requiring:
(P1) n(r, ) = n(r, ) and
y
n(r, ) =
y
n(r, ), for all r [, ].
(P2) n(, j) = n(, j) and
x
n(, j) =
x
n(, j) for all j [, ].
Interpretation #1: Pretend that n is actually a small piece of an innitely
extended, doubly periodic function n : 1
2
1, where, for every (r, j) 1
2
,
and every n, : Z, we have:
n(r + 2n, j + 2:) = n(r, j).
Exercise 5C.5. Explain how conditions (P1) and (P1) arise naturally from this E _
interpretation.
Interpretation #2: Glue the top edge of the square to the bottom edge, and
the right edge to the left edge. In other words, pretend that the square is really
a torus (Figure 5C.5).
Example 5C.6.
(a) The functions n(r, j) = sin(r) sin(j) and (r, j) = cos(r) cos(j) have peri-
odic boundary conditions. So do the functions n(r, j) = sin(r) cos(j) and
n(r, j) = cos(r) sin(j)
(b) For any (n, :) N
2
, the functions S
n,m
(r) = sin(nr) sin(:j) and C
n,m
(r) =
cos(nr) cos(:r) have periodic boundary conditions. (See Figure 9A.2 on
page 181.)
(c) sin(3r) sin(2j) + 2 cos(4r) cos(7j) has periodic boundary conditions.
(d) If n
1
(r, j) and n
2
(r, j) have periodic boundary conditions, and c
1
, c
2
are
any constants, then n(r, j) = c
1
n
1
(r, j) + c
2
n
2
(r, j) also has periodic
boundary conditions.
Exercise 5C.6 Verify these examples. E _
On the 1-dimensional cube [, ]
D
, we require, for d = 1, 2, . . . , 1 and all
r
1
, . . . , r
D
[, ], that
n(r
1
, . . . , r
d1
, , r
d+1
, . . . , r
D
) = n(r
1
, . . . , r
d1
, , r
d+1
, . . . , r
D
)
and
d
n(r
1
, . . . , r
d1
, , r
d+1
, . . . , r
D
) =
d
n(r
1
, . . . , r
d1
, , r
d+1
, . . . , r
D
).
Again, the idea is that we are identifying [, ]
D
with the 1-dimensional torus.
The space of all functions satisfying these conditions will be denoted (
per
[, ]
D
.
Thus, for example,
(
per
[, ] =
_
) : [, ] 1; ) is dierentiable,
)() = )() and )
t
() = )
t
()
_
(
per
[, ]
2
=
_
) : [, ][, ] 1; ) is dierentiable,
and satises (P1) and (P2) above
_
5D Uniqueness of solutions
Prerequisites: 1B, 2B, 1C, 5B, 5C.
Prerequisites (for proofs): 1E, 0E(iii), 0G.
Dierential equations are interesting primarily because they can be used to
express the laws governing physical phenomena (e.g. heat ow, wave motion,
electrostatics, etc.). By specifying particular initial conditions and boundary
conditions, we try to mathematically encode the physical conditions, constraints
and external inuences which are present in a particular situation. A solution
to the dierential equation which satises these initial/boundary conditions thus
constitutes a prediction about what will occur under these physical conditions.
However, this strategy can only succeed if there is a unique solution to the
dierential equation with particular initial/boundary conditions. If there are
many mathematically correct solutions, then we cannot make a clear prediction
about which of them will really occur. Sometimes we can reject some solutions as
being unphysical (e.g. they are nondierentiable, or discontinuous, or contain
unacceptable innities, or predict negative values for a necessarily positive quan-
tity like density). However, these notions of unphysicality really just represent
further mathematical constraints which we are implicitly imposing on the solu-
tion. If multiple solutions still exist, we should try to impose further constraints
5D. Uniqueness of solutions 85
(i.e. construct a more detailed or well-specied model) until we get a unique
solution. Thus, the question of uniqueness of solutions is extremely important
in the general theory of dierential equations (both ordinary and partial). In
this section, we will establish sucient conditions for the uniqueness of solutions
to I/BVPs for the Laplace, Poisson, Heat, and wave equations.
Let o 1
D
. We say that o is a smooth graph if there is an open subset
U 1
D1
, a function ) : U 1, and some d [1...1], such that o looks
like the graph of the function ), plotted over the domain U, with the value of )
plotted in the dth coordinate. In other words:
o = (n
1
, . . . , n
d1
, j, n
d
, . . . , n
D1
) ; (n
1
, . . . , n
D1
) U, j = )(n
1
, . . . , n
D1
) .
Intuitively, this means that o looks like a smooth surface (oriented roughly
perpendicular to the dth dimension). More generally, if o 1
D
, we say that o
is a smooth hypersurface if, for each s o, there exists some c 0 such that
B(s, c) o is a smooth graph.
Example 5D.1.
(a) Let P 1
D
be any (1 1)-dimensional hyperplane; then P is a smooth
hypersurface.
(b) Let S
1
:=
_
s 1
2
; [s[ = 1
_
be the unit circle in 1
2
. Then S
1
is a smooth
hypersurface in 1
2
.
(c) Let S
2
:=
_
s 1
3
; [s[ = 1
_
be the unit sphere in 1
3
. Then S
2
is a smooth
hypersurface in 1
3
.
(d) Let S
D1
:=
_
s 1
D
; [s[ = 1
_
be the unit hypersphere in 1
D
. Then S
D1
is a smooth hypersurface in 1
D
.
(e) Let o 1
D
be any smooth hypersurface, and let U 1
D
be an open set.
Then o U is also a smooth hypersurface (if it is nonempty).
Exercise 5D.1 Verify these examples. E _
A domain X 1
D
has piecewise smooth boundary if X is a nite
union of smooth hypersurfaces. If n : X 1 is some dierentiable function,
then this implies that the normal derivative
n(s) is well-dened for s X,

except for those s on the (negligible) regions where two or more of these smooth
hypersurfaces intersect. This means that it is meaningful to impose Neumann
boundary conditions on n. It also means that certain methods from vector
calculus can be applied to n (see 0E(iii) on page 561).
Example 5D.2. Every domain in Example 5C.1 on page 71 has a piecewise
smooth boundary. (Exercise 5D.2 Verify this.) E _
Indeed, every domain we will consider in this book will have a piecewise
smooth boundary, as does any domain which is likely to arise in any physically
realistic model. Hence, it suces to obtain uniqueness results for such domains.
5D(i) Uniqueness for the Laplace and Poisson equations
Let X 1
D
be a domain and let n : X 1. We say that n is continuous
and harmonic on X if n is continuous on X and n(x) = 0 for all x int (X).
Lemma 5D.3. (Solution uniqueness for Laplace equation; homogeneous BC)
Let X 1
D
be a bounded domain, and suppose n : X 1 is continuous and
harmonic on X. Then various homogeneous boundary conditions constrain the
solution as follows:
(a) (Homogeneous Dirichlet BC) If n(s) = 0 for all s X, then n must be
the constant 0 function: i.e. n(x) = 0, for all x X.
(b) (Homogeneous Neumann BC) Suppose X has a piecewise smooth bound-
ary. If
n(s) = 0 for all s X, then n must be a constant: i.e.

n(x) = C, for all x X.
(c) (Homogeneous Robin BC) Suppose X has a piecewise smooth boundary,
and let /, /
: X 1
,
be two other continuous nonnegative functions
such that /(s) +/
(s) 0 for all s X. If /(s)n(s) +/
(s)
n(s) = 0
for all s X, then n must be a constant function.
Furthermore, if / is nonzero somewhere on X, then n(x) = 0, for all
x X.
Proof. (a) If n : X 1 is harmonic, then the Maximum Principle (Corollary
1E.2 on page 17) says that any maximum/minimum of n occurs somewhere on
X. But n(s) = 0 for all s X; thus, max
X
(n) = 0 = min
X
(n); thus, n 0.
(If X has a piecewise smooth boundary, then another proof of (a) arises by
setting / 1 and /
0 in part (c).)
To prove (b), set / 0 and /
1 in part (c).
To prove (c), we will use Greens Formula. We begin with the following claim.
Claim 1: For all s X, we have n(s)
n(s) 0.
Proof. The homogeneous Robin boundary conditions say /(s)n(s)+/
(s)
n(s) =
0. Multiplying by n(s), we get
/(s)n
2
(s) +n(s)/
(s)
n(s) = 0. (5D.1)
If /
(s) = 0, then /(s) must be nonzero, and equation (5D.1) reduces to

/(s)n
2
(s) = 0, which means n(s) = 0, which means n(s)
n(s) 0, as
desired.
If /
(s) ,= 0, then we can rearrange equation (5D.1) to get

n(s)
n(s) =
/(s)n
2
(s)
/
(s)

()
0,
where () is because because /(s), /
(s) 0 by hypothesis, and of course

n
2
(s) 0. The claim follows.
Claim 1
Now, if n is harmonic, then n is innitely dierentiable, by Proposition 1E.4
on page 18. Thus, we can apply vector calculus techniques from Appendix
0E(iii). We have
0
()
_
X
n(s)
n(s) ds
()
_
X
n(x) n(x) + [n(x)[
2
dx
()
_
X
[n(x)[
2
dx
()
0. (5D.2)
Here, () is by Claim 1, () is by Greens Formula (Theorem 0E.5(b) on page
564), () is because n 0, and () is because [n(x)[
2
0 for all x X.
The inequalities (5D.2) imply that
_
X
[n(x)[
2
dx = 0.
But this implies that [n(x)[ = 0 for all x X, which means n 0, which
means n is a constant on X, as desired.
Now, if n 0, then clearly
n(s) = 0 for all s X. Thus, the Robin

boundary conditions reduce to /(s)n(s) = 0. If /(s) ,= 0 for some s X,
then we get n(s) = 0. But since n is a constant, this means that n 0. 2
One of the nice things about linear dierential equations is that linearity
enormously simplies the problem of solution uniqueness. First we show that
the only solution satisfying homogeneous boundary conditions (and, if applicable,
zero initial conditions) is the constant zero function (as in Lemma 5D.3 above).
Then it is easy to deduce uniqueness for arbitrary initial/boundary conditions.
Corollary 5D.4. (Solution uniqueness: Laplace equation, nonhomogeneous BC)
Let X 1
D
be a bounded domain, and let / : X 1 be continuous.
(a) There exists at most one continuous, harmonic function n : X 1 which
satises the nonhomogeneous Dirichlet BC n(s) = /(s) for all s X.
(b) Suppose X has a piecewise smooth boundary.
[i] If
_
X
/(s) ds ,= 0, then there is no continuous harmonic function n :
X 1 which satises the nonhomogeneous Neumann BC
n(s) = /(s)
for all s X.
[ii] Suppose
_
X
/(s) ds = 0. If n
1
, n
2
: X 1 are two continuous
harmonic functions which both satisfy the nonhomogeneous Neumann BC
n(s) = /(s) for all s X, then n

1
= n
2
+C for some constant C.
(c) Suppose X has a piecewise smooth boundary, and let /, /
: X 1
,
be
two other continuous nonnegative functions such that /(s) +/
(s) 0 for
all s X. If n
1
, n
2
: X 1 are two continuous harmonic functions which
both satisfy the nonhomogeneous Robin BC /(s)n(s) + /
(s)
n(s) =
/(s) for all s X, then n
1
= n
2
+ C for some constant C. Furthermore,
if / is nonzero somewhere on X, then n
1
= n
2
.
Proof. Exercise 5D.3 Hint: for (a), (c), and (b)[ii], suppose that n
1
, n
2
: X 1 E _
are two continuous harmonic functions with the desired nonhomogeneous boundary
conditions. Then (n
1
n
2
) is a continuous harmonic function satisfying homogeneous
boundary conditions of the same kind; now apply the appropriate part of Lemma
5D.3 to conclude that (n
1
n
2
) is zero or a constant.
For (b)[i], use Greens Formula (Theorem 0E.5(a) on page 564). 2
Exercise 5D.4. Let X = | = (:, ) ; [, ), : 1 be the closed unit disk E _
(in polar coordinates). Consider the function / : | 1 dened by /(:, ) = log(:).
In Cartesian coordinates, / has the form /(r, j) = log(r
2
+j
2
) (see Figure 1C.1(A) on
page 10). In Example 1C.2 we observed that / is harmonic. But / satises homogeneous
Dirichlet BC on |, so it seems to be a counterexample to Lemma 5D.3(a). Also,
/(r) = 1 for all r |, so / seems to be a counterexample to Corollary 5D.4(b)[i].

Why is this function not a counterexample to Lemma 5D.3 or Corollary 5D.4(b)[i].?
Theorem 5D.5. (Solution uniqueness: Poisson equation, Nonhomogeneous BC)

Let X 1
D
be a bounded domain with a piecewise smooth boundary. Let
: X 1 be a continuous function (e.g. describing an electric charge or
heat source), and let / : X 1 be another continuous function (a bound-
ary condition). Then there is at most one continuous function n : X 1
satisfying the Poisson Equation n = , and satisfying either of the following
nonhomogeneous boundary conditions:
(a) (Nonhomogeneous Dirichlet BC) n(s) = /(s) for all s X.
(b) (Nonhomogeneous Robin BC) /(s))n(s) + /
(s)
n(s) = /(s) for all

s X, where /, /
: X 1
,
are two other nonnegative functions, and
/ is nontrivial.
Furthermore, if n
1
and n
2
are two functions satisfying n = , and also satisfy-
ing:
(c) (Nonhomogeneous Neumann BC)
n(s) = /(s) for all s X.

....then n
1
= n
2
+C, where C is a constant.
Proof. Suppose n
1
and n
2
were two continuous functions satisfying one of (a)
or (b), and such that n
1
= = n
2
. Let n = n
1
n
2
. Then n is continuous,
harmonic, and satises one of (a) or (c) in Lemma 5D.3. Thus, n 0. But
this means that n
1
n
2
. Hence, there can be at most one solution. The proof
for (c) is Exercise 5D.5 . 2 E _
5D(ii) Uniqueness for the heat equation
Throughout this section, if n : X 1
,
1 is a time-varying scalar eld, and
t 1
,
, then dene the function n
t
: X 1 by n
t
(x) := n(x; t), for all x X.
(Note: n
t
does not denote the time-derivative).
If ) : X 1 is any integrable function, then the 1
2
-norm of ) is dened
|)|
2
:=
__
X
[)(x)[
2
dx
_
1/2
.
(See 6B for more information). We begin with a result which reinforces our
intuition that the heat equation resembles melting or erosion.
Lemma 5D.6. (1
2
-norm decay for heat equation)
Let X 1
D
be a bounded domain with a piecewise smooth boundary. Suppose
that n : X 1
,
1 satises the following three conditions:
(a) (Regularity) n is continuous on X 1
,
, and
t
n and
2
1
n, . . . ,
2
D
n are
continuous on int (X) 1
+
;
(b) (Heat equation)
t
n = n;
(c) (Homogeneous Dirichlet/Neumann BC) For all s X and t 1
,
, either
n
t
(s) = 0 or
n
t
(s) = 0.
6
Dene the function 1 : 1
,
1
,
by
1(t) := |n
t
|
2
2
=
_
X
[n
t
(x)[
2
dx, for all t 1
,
. (5D.3)
Then 1 is dierentiable and nonincreasing that is, 1
t
(t) 0 for all t 1
,
.
6
Note that this allows dierent boundary points to satisfy dierent homogeneous boundary
conditions at dierent times.
Proof. For any x X and t 1
,
, we have
t
[n
t
(x)[
2
()
2n
t
(x)
t
n
t
(x)
()
2n
t
(x) n
t
(x), (5D.4)
where () is the Leibniz rule, and () is because n satises the heat equation
by hypothesis (b). Thus,
1
t
(t)
()
_
X
t
[n
t
(x)[
2
dx
()
2
_
X
n
t
(x) n
t
(x) dx, (5D.5)
Here () comes from dierentiating the integral (5D.3) using Proposition 0G.1
on page 567. Meanwhile, () is by eqn.(5D.4).
Claim 1: For all t 1
,
,
_
X
n
t
(x) n
t
(x) dx =
_
X
|n
t
(x)|
2
dx.
Proof. For all s X, either n
t
(s) = 0 or
n
t
(s) = 0 by hypothesis (c).
But
n
t
(s) = n
t
(s)

N(s) (where

N(s) is the unit normal vector at s),
so this implies that n
t
(s) n
t
(s)

N(s) = 0 for all s X. Thus,
0 =
_
X
n
t
(s) n
t
(s)

N(s) ds
()
_
X
div (n
t
n
t
)(x) dx
()
_
X
_
n
t
div n
t
+ n
t
n
t
_
(x) dx
()
_
X
n
t
(x) n
t
(x) dx +
_
X
|n
t
(x)|
2
dx.
Here, () is the Divergence Theorem 0E.4 on page 563, () is by the Leibniz
rule for divergences (Proposition 0E.2(b) on page 560) and () is because
div n = n, while n
t
n
t
= |n
t
(x)|
2
. We thus have
_
X
n
t
n
t
+
_
X
|n
t
|
2
= 0.
Rearranging this equation yields the claim.
Claim 1
Applying Claim 1 to equation (5D.5), we get
1
t
(t) = 2
_
X
|n
t
(x)|
2
dx 0.
because |n
t
(x)|
2
0 for all x X. 2
Lemma 5D.7. (Solution uniqueness for heat equation; homogeneous I/BC)
Let X 1
D
that n : X 1
,
1 satises the following four conditions:
,
, and
t
n and
2
1
n, . . . ,
2
D
n are
+
;
(b) (Heat equation)
t
n = n;
(c) (Zero initial condition) n
0
(x) = 0 for all x X;
(d) (Homogeneous Dirichlet/Neumann BC) For all s X and t 1
,
, either
n
t
(s) = 0 or
n
t
(s) = 0.
7
Then n must be the constant 0 function: n 0.
Proof. Dene 1 : 1
,
1
,
as in Lemma 5D.6. Then 1 is a nonincreasing
function. But 1(0) = 0, because n
0
0 by hypothesis (c). Thus, 1(t) = 0 for
all t 1
,
. Thus, we must have n
t
0 for all t 1
,
. 2
Theorem 5D.8. (Uniqueness: forced heat equation, nonhomogeneous I/BC)
Let X 1
D
1 : X 1 be a continuous function (describing an initial condition), and let
/ : X1
,
1, and /, /
: X1
,
1 be three other continuous functions
(describing time-varying boundary conditions). Let ) : int (X) 1
,
1 be
another continuous function (describing exogenous heat being forced into or out
of the system). Then there is at most one solution function n : X 1
,
1
satisfying the following four conditions:
,
, and
t
n and
2
1
n, . . . ,
2
D
n are
+
;
(b) (Heat equation with forcing)
t
n = n +);
(c) (Initial condition) n(x, 0) = 1(x) for all x X;
(d) (Nonhomogeneous Mixed BC) /(s, t) n
t
(s) + /
(s, t)
n
t
(s) = /(r, t),
for all s X and t 1
,
.
8
7
Note that this allows dierent boundary points to satisfy dierent homogeneous boundary
conditions at dierent times.
8
Note that this includes nonhomogeneous Dirichlet BC (set h
0) and nonhomogeneous
Neumann BC (set h 0) as special cases. Also note that by varying h and h
, we can allow
dierent boundary points to satisfy dierent nonhomogeneous boundary conditions at dierent
times.
Proof. Suppose n
1
and n
2
were two functions satisfying all of (a)-(d). Let
n = n
1
n
2
. Then n satises all of (a)-(d) in Lemma 5D.7. Thus, n 0. But
this means that n
1
n
2
. Hence, there can be at most one solution. 2
5D(iii) Uniqueness for the wave equation
Throughout this section, if n : X 1
,
1 is a time-varying scalar eld, and
t 1
,
, then dene the function n
t
: X 1 by n
t
(x) := n(x; t), for all x X.
(Note: n
t
does not denote the time-derivative). For all t 0, the energy of n is
dened:
1(t) :=
1
2
_
X
[
t
n
t
(x)[
2
+|n
t
(x)|
2
dx. (5D.6)
We begin with a result which has an appealing physical interpretation.
Lemma 5D.9. (Conservation of Energy for wave equation)
Let X 1
D
n : X 1
,
1 satises the following three conditions:
,
, and n (
2
(int (X) 1
+
);
(b) (Wave equation)
2
t
n = n;
(c) (Homogeneous Dirichlet/Neumann BC) For all s X, either n
t
(s) = 0
for all t 0, or
n
t
(s) = 0 for all t 0.
9
Then 1 is constant in time that is,
t
1(t) = 0 for all t 0.
Proof. The Leibniz rule says that
t
[
t
n[
2
= (
2
t
n) (
t
n) + (
t
n) (
2
t
n)
= 2 (
t
n) (
2
t
n), (5D.7)
and
t
|n|
2
= (
t
n) (n) + (n) (
t
n)
= 2 (n) (
t
n)
= 2 (n) (
t
n). (5D.8)
Thus,
t
1
()
1
2
_
X
_
t
[
t
n[
2
+
t
|n|
2
_
()
_
X
_
t
n
2
t
n + (n) (
t
n)
_
. (5D.9)
9
This allows dierent boundary points to satisfy dierent homogeneous boundary conditions;
but each particular boundary point must satisfy the same homogeneous boundary condition at
all times.
Here () comes from dierentiating the integral (5D.6) using Proposition 0G.1
on page 567). Meanwhile, () comes from substituting (5D.7) and (5D.8).
Claim 1: Fix s X and let

N(s) be the outward unit normal vector to
X at s. Then
t
n
t
(s) n
t
(s)

N(s) = 0, for all t 0.
Proof. By hypothesis (c), either
n
t
(s) = 0 for all t 0, or n
t
(s) = 0 for
all t 0. Thus, either n
t
(s)

N(s) = 0 for all t 0, or
t
n
t
(s) = 0 for all
t 0. In either case,
t
n
t
(s) n
t
(s)

N(s) = 0 for all t 0.
Claim 1
Claim 2: For any t 1
,
,
_
X
n
t

t
n
t
=
_
X
t
n
t
n
t
.
Proof. Integrating Claim 1 over X, we get
0 =
_
X
t
n
t
(s) n
t
(s)

N(s) ds
()
_
X
div (
t
n
t
n
t
) (x) dx
()
_
X
_
t
n
t
div n
t
+
t
n
t
n
t
_
(x) dx
()
_
X
(
t
n
t
n
t
) (x) dx +
_
X
(
t
n
t
n
t
) (x) dx.
Here, () is the Divergence Theorem 0E.4 on page 563, () is by the Leibniz
rule for divergences (Proposition 0E.2(b) on page 560) and () is because
div n
t
= n
t
. We thus have
_
X
n
t

t
n
t
+
_
X
t
n
t
n
t
= 0.
Rearranging this equation yields the claim.
Claim 2
Putting it all together, we get:
t
1
()
_
X
t
n
2
t
n +
_
X
(n) (
t
n)
()
_
X
t
n
2
t
n
_
X
t
n n =
_
X
t
n
_
2
t
n n
_
()
_
X
t
n 0 = 0,
as desired. Here, () is by equation (5D.9), () is by Claim 2, and () is because
2
t
n n 0 because n satises the wave equation by hypothesis (b). 2
Physical interpretation. 1(t) can be interpreted as the total energy in the
system at time t. The rst term in the integrand of (5D.6) measures the kinetic
energy of the wave motion, while the second term measures the potential energy
stored in the deformation of the medium. With this physical interpretation,
Lemma 5D.9 simply asserts the principle of Conservation of Energy: 1 must be
constant in time, because no energy enters or leaves the system, by hypotheses
(b) and (c).
Lemma 5D.10. (Solution uniqueness for wave equation; homogeneous I/BC)
Let X 1
D
n : X 1
,
1 satises all ve of the following conditions:
,
, and n (
2
(int (X) 1
+
);
(b) (Wave equation)
2
t
n = n;
(c) (Zero initial position) n
0
(x) = 0, for all x X;
(d) (Zero initial velocity)
t
n
0
(x) = 0 for all x X;
(e) (Homogeneous Dirichlet/Neumann BC) For all s X, either n
t
(s) = 0
for all t 0, or
n
t
(s) = 0 for all t 0.
10
Then n must be the constant 0 function: n 0.
Proof. Let 1 : 1
,
1
,
be the energy function from Lemma 5D.9. Then 1
is a constant. But 1(0) = 0 because n
0
0 and
t
n
0
0, by hypotheses (c)
and (d). Thus, 1(t) = 0 for all t 0. But this implies that [
t
n
t
(x)[
2
= 0,
and hence
t
n
t
(x) = 0, for all x X and t 0. Thus, n is constant in time.
Since n
0
0, we conclude that n
t
0 for all t 0, as desired. 2
Theorem 5D.11. (Uniqueness: forced wave equation, nonhomogeneous I/BC)
Let X 1
D
1
0
, 1
1
: X 1 be continuous functions (describing initial position and velocity).
Let / : X1
,
1 be another continuous function (describing a time-varying
boundary condition). Let ) : int (X) 1
,
1 be another continuous function
(describing exogenous vibrations being forced into the system). Then there is
at most one solution function n : X1
,
1 satisfying all ve of the following
conditions:
,
, and n (
2
(int (X) 1
+
);
10
This allows dierent boundary points to satisfy dierent homogeneous boundary conditions;
but each particular boundary point must satisfy the same homogeneous boundary condition at
all times.
5E.
Classication of second order linear PDEs 95

(b) (Wave equation with forcing)
2
t
n = n +);
(c) (Initial position) n(x, 0) = 1
0
(x) for all x X.
(d) (Initial velocity)
t
n(x, 0) = 1
1
(x) for all x X.
(e) (Nonhomogeneous Dirichlet/Neumann BC) For all s X, either n(s, t) =
/(s, t) for all t 0, or
n(s, t) = /(s, t) for all t 0.

Proof. Suppose n
1
and n
2
were two functions satisfying all of (a)-(e). Let
n = n
1
n
2
. Then n satises all of (a)-(e), in Lemma 5D.10. Thus, n 0.
But this means that n
1
n
2
. Hence, there can be at most one solution. 2
Remark. (a) Earlier, we observed that the initial position problem for the
(unforced) wave equation represents a plucked string (e.g. in a guitar), while the
initial velocity problem represents a struck string (e.g. in a piano). Continuing
the musical analogy, the forced wave equation represents a rubbed string (e.g. in
a violin or cello), as well as any other musical instrument driven by an exogenous
vibration (e.g. any wind instrument).
(b) Notice Theorems 5D.5, 5D.8, and 5D.11 apply under much more general
conditions than any of the solution methods we will actually develop in this book
(i.e. they work for almost any reasonable domain, we allow for possible forcing,
and we even allow the boundary conditions to vary in time). This is a recurring
theme in dierential equation theory; it is generally possible to prove qualitative
results (e.g. about existence, uniqueness, or general properties of solutions) in
much more general settings than it is possible to get quantitative results (i.e.
explicit formulae for solutions). Indeed, for most nonlinear dierential equations,
qualitative results are pretty much all you can ever get.
5E
Classication of second order linear PDEs

Prerequisites: 5A. Recommended: 1B, 1C, 1F, 2B.
5E(i) ...in two dimensions, with constant coecients
Recall that (
(1
2
; 1) is the space of all dierentiable scalar elds on the plane
1
2
. In general, a second-order linear dierential operator L on (
(1
2
; 1) with
constant coecients looks like:
Ln = o
2
x
n + /
x
y
n + c
2
y
n + d
x
n + c
y
n + ) n (5E.1)
where o, /, c, d, c, ) are constants. Dene:
= ), =
_
d
c
_
and =
_
o
1
2
/
1
2
/ c
_
=
_

11

12
21

22
_
.
Then we can rewrite (5E.1) as:
Ln =
2
c,d=1
c,d

c
d
n +
2
d=1
d

d
n + n,
Any 2 2 symmetric matrix denes a quadratic form G : 1
2
1 by
G(r, j) = [r j]
_

11

12
21

22
_
_
r
j
_
=
11
r
2
+
_
12
+
21
_
rj +
22
j
2
.
We say is positive denite if, for all r, j 1, we have:
G(r, j) 0;
G(r, j) = 0 if and only if r = 0 = j.
Geometrically, this means that the graph of G denes an elliptic paraboloid in
1
2
1, which curves upwards in every direction. Equivalently, is positive
denite if there is a constant 1 0 such that
G(r, j) 1 (r
2
+j
2
)
for every (r, j) 1
2
. We say is negative denite if is positive denite.
The dierential operator L from equation (5E.1) is called elliptic if the ma-
trix is either positive denite or negative denite.
Example 5E.1. If L = , then =
_
1
1
_
is just the identity matrix. while
= 0 and = 0. The identity matrix is clearly positive denite; thus, is
an elliptic dierential operator.
Suppose that L is an elliptic dierential operator. Then:
An elliptic PDE is one of the form: Ln = 0 (or Ln = p). For example, the
Laplace equation is elliptic.
A parabolic PDE is one of the form:
t
= Ln. For example, the two-
dimensional heat equation is parabolic.
A hyperbolic PDE is one of the form:
2
t
= Ln. For example, the two-
dimensional wave equation is hyperbolic.
(See Remark 16F.4 on page 371 for a partial justication of this terminology).
5E.
Classication of second order linear PDEs 97

Exercise 5E.1. Show that is positive denite if and only if 0 < det() = oc
1
4
/
2
. E _
In other words, L is elliptic if and only if 4oc /
2
0.
5E(ii) ...in general
Recall that (
(1
D
; 1) is the space of all dierentiable scalar elds on 1-dimensional
space. The general second-order linear dierential operator on (
(1
D
; 1) has
the form
Ln =
D
c,d=1
c,d

c
d
n +
D
d=1
d

d
n + n, (5E.2)
where : 1
D
1 1 is some time-varying scalar eld, (
1
, . . . ,
D
) = :
1
D
1 1
D
is a time-varying vector eld, and
c,d
: 1
D
1 1 are
functions such that, for any x 1
D
and t 1, the matrix
(x; t) =
_
11
(x; t) . . .
1D
(x; t)
.
.
.
.
.
.
.
.
.
D1
(x; t) . . .
DD
(x; t)
_
_
is symmetric (i.e.
cd
=
dc
).
Example 5E.2.
(a) If L = , then 0, = 0, and Id =
_
_
_
_
_
1 0 . . . 0
0 1 . . . 0
.
.
.
.
.
.
.
.
.
.
.
.
0 0 . . . 1
_
_
.
(b) The Fokker-Plank Equation (see 1F on page 18) has the form
t
n = Ln,
where = div

V(x), (x) =
V(x), and Id. (Exercise 5E.2) E _
If the functions
c,d
,
d
and are independent of x, then we say L is spatially
homogeneous. If they are also independent of t, we say that L has constant
coecients.
Any symmetric matrix denes a quadratic form G : 1
D
1 by
G(x) = [r
1
...r
D
]
_
11
. . .
1D
.
.
.
.
.
.
.
.
.
D1
. . .
DD
_
_
_
_
r
1
.
.
.
r
D
_
_ =
D
c,d=1
c,d
r
c
r
d
is called positive denite if, for all x 1
D
, we have:
G(x) 0;
G(x) = 0 if and only if x = 0.
Equivalently, is positive denite if there is a constant 1
0 such that G(x)

1
|x|
2
for every x 1
D
. On the other hand, is negative denite if
is positive denite.
The dierential operator L from equation (5E.2) is elliptic if the matrix
(x; t) is either positive denite or negative denite for every (x; t) 1
D
1
,
,
and furthermore, there is some 1 0 such that 1
(x;t)
1 for all (x; t)
1
D
1
,
. For example, the Laplacian and the Fokker-Plank operator are both
elliptic. (Exercise 5E.3) E _
Suppose that L is an elliptic dierential operator. Then:
An elliptic PDE is one of the form: Ln = 0 (or Ln = p).
A parabolic PDE is one of the form:
t
= Ln.
A hyperbolic PDE is one of the form:
2
t
= Ln.
Example 5E.3.
(a) Laplaces Equation and Poissons Equation are elliptic PDEs.
(b) The heat equation and the Fokker-Plank Equation are parabolic.
(c) The wave equation is hyperbolic.
Parabolic equations are generalized heat equations, describing diusion
through an inhomogeneous
11
, anisotropic
12
medium with drift. The terms in
(x; t) describe the inhomogeneity and anisotropy of the diusion
13
, while the
vector eld describes the drift.
Hyperbolic equations are generalized wave equations, describing wave prop-
agation through an inhomogeneous, anisotropic medium with drift for exam-
ple, sound waves propagating through an air mass with variable temperature
and pressure and wind blowing.
5F Practice problems
Evolution equations and initial value problems. For each of the following
equations: n is an unknown function; is always some xed, predetermined
function; and is always a constant. In each case, determine the order of the
equation, and decide: is this an evolution equation? Why or why not?
11
Homogeneous means, Looks the same everywhere in space, whereas inhomogeneous
is the opposite.
12
Isotropic means looks the same in every direction; anisotropic means the opposite.
13
If the medium is homogeneous, then is constant. If the medium is isotropic, then = Id.
5F. Practice problems 99
1. heat equation:
t
n(x) = n(x).
2. Poisson Equation: n(x) = (x).
3. Laplace Equation: n(x) = 0.
4. Monge-Ampère Equation: (r, j) = det
_

2
x
n(r, j)
x
y
n(r, j)
y
n(r, j)
2
y
n(r, j)
_
.
5. Reaction-Diusion
t
n(x; t) = n(x; t) +
_
n(x; t)
_
.
6. Scalar conservation Law
t
n(r; t) =
x
( n)(r; t).
7. Helmholtz Equation: n(x) = n(x).
8. Airys Equation:
t
n(r; t) =
3
x
n(r; t).
9. Beam Equation:
t
n(r; t) =
4
x
n(r; t).
10. Schrodinger Equation:
t
n(x; t) = i n(x; t) + (x; t) n(x; t).
11. Burgers Equation:
t
n(r; t) = n(r; t)
x
n(r; t).
12. Eikonal Equation: [
x
n(r)[ = 1.
Boundary value problems.
1. Each of the following functions is dened on the interval [0, ], in Carte-
sian coordinates. For each function, decide: Does it satisfy homogeneous
Dirichlet BC? Homogeneous Neumann BC? Homogeneous Robin
14
BC?
Periodic BC? Justify your answers.
(a) n(r) = sin(3r).
(b) n(r) = sin(r) + 3 sin(2r) 4 sin(7r).
(c) n(r) = cos(r) + 3 sin(3r) 2 cos(6r).
(d) n(r) = 3 + cos(2r) 4 cos(6r).
(e) n(r) = 5 + cos(2r) 4 cos(6r).
2. Each of the following functions is dened on the interval [, ], in Carte-
14
BC?
(a) n(r) = sin(r) + 5 sin(2r) 2 sin(3r).
(b) n(r) = 3 cos(r) 3 sin(2r) 4 cos(2r).
0
0
0
0.2
0.5
0.4
0.5
0.6
1
0.8
1
1
1.5 1.5 x 2 y 2
2.5 2.5
3 3
0
0.5
0
0
1
0.5
1
0.5
1.5
2
1.5
1
x
1.5
y
2
2
2.5
2.5 3
3
0
0.5
1
y
1.5
2
2.5
30
0.5
1
x
1.5
2
2.5
3
-2
-1
0
1
2
(A) f(x, y) = sin(x) sin(y) (B) g(x, y) = sin(x) + sin(y) (C) h(x, y) = cos(2x) + cos(y).
Figure 5F.1: Problems #3a, #3b and #3c
(c) n(r) = 6 + cos(r) 3 cos(2r).
3. Each of the following functions is dened on the box [0, ]
2
. in Carte-
14
BC?
(a) )(r, j) = sin(r) sin(j) (Figure 5F.1(A))
(b) p(r, j) = sin(r) + sin(j) (Figure 5F.1(B))
(c) /(r, j) = cos(2r) + cos(j) (Figure 5F.1(C))
(d) n(r, j) = sin(5r) sin(3j).
(e) n(r, j) = cos(2r) cos(7j).
4. Each of the following functions is dened on the unit disk
| = (:, ) ; 0 : 1, and [0, 2)
in polar coordinates. For each function, decide: Does it satisfy homoge-
neous Dirichlet BC? Homogeneous Neumann BC? Homogeneous Robin
14
BC? Justify your answers.
(a) n(:, ) = (1 :
2
).
(b) n(:, ) = 1 :
3
.
(c) n(:, ) = 3 + (1 :
2
)
2
.
14
Here, Robin B.C. means nontrivial Robin B.C. i.e. not just homogenous Dirichlet or
Neumann.
(d) n(:, ) = sin()(1 :
2
)
2
.
(e) n(:, ) = cos(2)(c c
r
).
5. Each of the following functions is dened on the 3-dimensional unit ball
B =
_
(:, , ) ; 0 : 1, [0, 2), and
_
2
,

2
__
in spherical coordinates. For each function, decide: Does it satisfy homo-
geneous Dirichlet BC? Homogeneous Neumann BC? Homogeneous Robin
14
BC? Justify your answers.
(a) n(:, , ) = (1 :)
2
.
(b) n(:, , ) = (1 :)
3
+ 5.
6. Which Neumann BVP has solution(s) on the domain X = [0, 1]?
(a) n
tt
(r) = 0, n
t
(0) = 1, n
t
(1) = 1.
(b) n
tt
(r) = 0, n
t
(0) = 1, n
t
(1) = 2.
(c) n
tt
(r) = 0, n
t
(0) = 1, n
t
(1) = 1.
(d) n
tt
(r) = 0, n
t
(0) = 1, n
t
(1) = 2.
7. Which BVP of Laplace equation on the unit disk | has a solution? Which
BVP has more than one solution?
(a) n = 0, n(1, ) = 0, for all [, ).
(b) n = 0, n(1, ) = sin, for all [, ).
(c) n = 0,
n(1, ) = sin(), for all [, ).

(d) n = 0,
n(1, ) = 1 + cos(), for all [, ).

102
III Fourier series on bounded
domains
Any complex sound is a combination of simple pure tones of dierent fre-
quencies. For example, a musical chord is a superposition of three (or more)
musical notes, each with a dierent frequency. In fact, a musical note itself is
not really a single frequency at all; a note consists of a fundamental frequency,
plus a cascade of higher frequency harmonics. The energy distribution of these
harmonics is part of what gives each musical instrument its distinctive sound.
The decomposition of a sound into separate frequencies is sometimes called its
power spectrum. A crude graphical representation of this power spectrum is
visible on most modern stereo systems (the little jiggling red bars).
Fourier theory is based on the idea that a real-valued function is like a sound,
which can be represented as a superposition of pure tones (i.e. sine waves
and/or cosine waves) of distinct frequencies. This provides a coordinate system
for expressing functions, and within this coordinate system, we can express the
solutions for many partial dierential equations in a simple and elegant way.
Fourier theory is also an essential tool in probability theory and signal analysis
(although we will not discuss these applications in this book).
The idea of Fourier theory is simple, but to make this idea rigorous enough
to be useful, we must deploy some formidable mathematical machinery. So we
will begin by developing the necessary background concerning inner products,
orthogonality, and the convergence of functions.
103
Chapter 6
Some functional analysis
Mathematical science is in my opinion an indivisible whole, an organism whose vitality
is conditioned upon the connection of its parts. David Hilbert
6A Inner products
Prerequisites: 4A.
Let x, y 1
D
, with x = (r
1
, . . . , r
D
) and y = (j
1
, . . . , j
D
). The inner
product
1
of x, y is dened:
x, y) := r
1
j
1
+r
2
j
2
+. . . +r
D
j
D
.
The inner product describes the geometric relationship between x and y, via the
formula:
x, y) := |x| |y| cos()
where |x| and |y| are the lengths of vectors x and y, and is the angle between
them. (Exercise 6A.1 Verify this). In particular, if x and y are perpendicular, E _
then =
2
, and then x, y) = 0; we then say that x and y are orthogonal.
For example, x =
_
1
1
and y =
_
1
1
_
are orthogonal in 1
2
, while
u =
_
_
1
0
0
0
_
_
, v =
_
_
0
0
1
2
1
2
_
_
, and w =
_
_
0
1
0
0
_
_
are all orthogonal to one another in 1
4
. Indeed, u, v, and w also have unit norm;
we call any such collection an orthonormal set of vectors. Thus, u, v, w is
an orthonormal set. However, x, y is orthogonal but not orthonormal (because
|x| = |y| =

2 ,= 1).
1
This is sometimes this is called the dot product, and denoted x y.
104 DRAFT Chapter 6. Some functional analysis
The norm of a vector satises the equation:
|x| =
_
r
2
1
+r
2
2
+. . . +r
2
D
_
1/2
= x, x)
1/2
.
If x
1
, . . . , x
N
are a collection of mutually orthogonal vectors, and x = x
1
+. . . +
x
N
, then we have the generalized Pythagorean formula:
|x|
2
= |x
1
|
2
+|x
2
|
2
+. . . +|x
N
|
2
.
(Exercise 6A.2 Verify the Pythagorean formula.) E _
An orthonormal basis of 1
D
is any collection of mutually orthogonal vec-
tors v
1
, v
2
, . . ., v
D
, all of norm 1, such that, for any w 1
D
, if we dene
d
= w, v
d
) for all d [1..1], then:
w =
1
v
1
+
2
v
2
+. . . +
D
v
D
.
In other words, the set v
1
, v
2
, . . . , v
D
denes a coordinate system for 1
D
,
and in this coordinate system, the vector w has coordinates (
1
,
2
, . . . ,
D
). If
x 1
D
is another vector, and
d
= x, v
d
) all d [1..1], then we also have
x =
1
v
1
+
2
v
2
+. . . +
D
v
D
.
We can then compute w, x) using Parsevals Equality:
w, x) =
1
1
+
2
2
+
D
D
.
(Exercise 6A.3 Prove Parsevals equality.) In particular, if x = w, we get the E _
the following version of the generalized Pythagorean formula:
|w|
2
=
2
1
+
2
2
+. . . +
2
D
.
Example 6A.1.
(a)
_
_
_
_
_
_
_
_
_
_
_
_
_
_
1
0
.
.
.
0
_
_
,
_
_
_
_
_
0
1
.
.
.
0
_
_
, . . .
_
_
_
_
_
0
0
.
.
.
1
_
_
_
_
_
_
_
_
_
_
_
is an orthonormal basis for 1
D
.
(b) If v
1
=
_
3,2
1,2
_
and v
2
=
_
1,2
3,2
_
, then v
1
, v
2
is an orthonormal
basis of 1
2
.
If w =
_
2
4
_
, then
1
=

3 + 2 and
2
= 2
3 1, so that
_
2
4
_
=
1
v
1
+
2
v
2
=
_
3 + 2
_
_
3,2
1,2
_
+
_
2
3 1
_
_
1,2
3,2
_
.
Thus, |w|
2
2
= 2
2
+ 4
2
= 20, and also, by Parsevals equality, 20 =
2
1
+
2
2
=
_
3 + 2
_
2
+
_
1 2
3
_
2
. (Exercise 6A.4 Verify these claims.) E _
6B. 1
2
space 105
f (x) dx
2
f =
2
2
f (x)
f (x)
2
f (x) dx
2
f =
2
Figure 6A.1: The 1
2
norm of ): |)|
2
=
_
_
X
[)(r)[
2
dr
6B L
2
space
The ideas of section 6A generalize to spaces of functions. Suppose X 1
D
is some bounded domain, and let ` :=
_
X
1 dx be the volume
2
of the domain X.
(The second column of Table 6.1 provides examples of ` for various domains.)
Domain M Inner Product
Unit interval X = [0, 1] 1 length M = 1 f, g) =
_
1
0
f(x) g(x) dx
interval X = [0, ] 1 length M = f, g) =
1
_

0
f(x) g(x) dx
Unit square X = [0, 1] [0, 1] 1
2
area M = 1 f, g) =
_
1
0
_
1
0
f(x, y) g(x, y) dx dy
square X = [0, ] [0, ] 1
2
area M =
2
f, g) =
1
2
_

0
_

0
f(x, y) g(x, y) dx dy
Unit Disk
(polar coords)
X = |(r, ) ; r 1 1
2
area M = f, g) =
1
_
1
0
_

f(r, ) g(r, ) r d dr
Unit cube X = [0, 1] [0, 1] [0, 1] 1
3
volume M = 1 f, g) =
_
1
0
_
1
0
_
1
0
f(x, y, z) g(x, y, z) dx dy dz
Table 6.1: Inner products on various domains.
If ), p : X 1 are integrable functions, then the inner product of ) and
p is dened:
), p) :=
1
`
_
X
)(x) p(x) dx. (6B.1)
2
Or length, if D = 1, or area if D = 2....
Example 6B.1.
(a) Suppose X = [0, 3] = r 1 ; 0 r 3. Then ` = 3. If )(r) = r
2
+ 1
and p(r) = r for all r [0, 3], then
), p) =
1
3
_
3
0
)(r)p(r) dr =
1
3
_
3
0
(r
3
+r) dr =
27
4
+
3
2
.
(b) The third column of Table 6.1 provides examples of ), p) for various other
domains.
The 1
2
-norm of an integrable function ) : X 1 is dened
|)|
2
:= ), ))
1/2
=
_
1
`
_
X
)
2
(x) dx
_
1/2
. (6B.2)
(See Figure 6A.1. Of course, this integral may not converge.) The set of all
integrable functions on X with nite 1
2
-norm is denoted L
2
(X), and is called
1
2
-space. For example, any bounded, continuous function ) : X 1 is in
L
2
(X).
Example 6B.2. (a) Suppose X = [0, 3], as in Example 6B.1, and let )(r) = r+1.
Then ) L
2
[0, 3], because
|)|
2
2
= ), )) =
1
3
_
3
0
(r + 1)
2
dr
=
1
3
_
3
0
r
2
+ 2r + 1 dr =
1
3
_
r
3
3
+r
2
+r
_
x=3
x=0
= 7,
hence |)|
2
=

7 < .
(b) Let X = (0, 1], and suppose ) (
(0, 1] is dened )(r) := 1,r. Then

|)|
2
= , so) , L
2
(0, 1].
Remark. Some authors dene the inner product as ), p) :=
_
X
)(x)p(x) dx,
and dene the 1
2
-norm as |)|
2
:=
__
X
)
2
(x) dx
_
1/2
. In other words, these
authors do not divide by the volume ` of the domain. This yields a mathemati-
cally equivalent theory. The advantage of our denition is greater computational
convenience in some situations. (For example, if 11
X
is the constant 1-valued
function, then in our denition, |11
X
|
2
= 1.) When comparing formulae from
dierent books, you should always check their respective denitions of 1
2
norm.
6B. 1
2
space 107
1
2
space on an innite domain. Suppose X 1
D
is a region of innite
volume (or length, area, etc.). For example, maybe X = 1
,
is the positive half-
line, or perhaps X = 1
D
. In this case, ` = , so it doesnt make any sense to
divide by `. If ), p : X 1 are integrable functions, then the inner product
of ) and p is dened:
), p) :=
_
X
)(x) p(x) dx (6B.3)
Example 6B.3. Suppose X = 1. If )(r) = c
[x[
and p(r) =
_
1 if 0 < r < 7
0 otherwise
,
then
), p) =
_

)(r)p(r) dr =
_
7
0
c
x
dr = (c
7
c
0
) =
1
1
c
7
.
The 1
2
-norm of an integrable function ) : X 1 is dened
|)|
2
= ), ))
1/2
=
__
X
)
2
(x) dx
_
1/2
. (6B.4)
Again, this integral may not converge. Indeed, even if ) is bounded and contin-
uous everywhere, this integral may still equal innity. The set of all integrable
functions on X with nite 1
2
-norm is denoted L
2
(X), and called 1
2
-space. (You
may recall that on page 40 of 3A, we discussed how 1
2
-space arises naturally
in quantum mechanics as the space of physically meaningful wavefunctions.)
Proposition 6B.4. Properties of the inner product
Whether it is dened using equation (6B.1) or (6B.3), the inner product has the
following properties.
Bilinearity. For any )
1
, )
2
, p
1
, p
2
L
2
(X), and any constants :
1
, :
2
, :
1
, :
2
1,
:
1
)
1
+:
2
)
2
, :
1
p
1
+:
2
p
2
) = :
1
:
1
)
1
, p
1
) +:
1
:
2
)
1
, p
2
) +:
2
:
1
)
2
, p
1
) +:
2
:
2
)
2
, p
2
) .
Symmetry. For any ), p L
2
(X), ), p) = p, )).
Positive-denite. For any ) L
2
(X), ), )) 0. Also, ), )) = 0 if and only
if ) = 0.
Proof. Exercise 6B.1 2 E _
If v, w 1
D
, recall that v, w) = |v| |w| cos(), where is the angle
between v to w. In particular, this implies that
v, w)
|v| |w| . (6B.5)

If ), p L
2
(X) are two functions, then it doesnt make sense to talk about the
angle between ) and p [as vectors in L
2
(X)]. But an inequality analogous to
(6B.5) is still true.
Theorem 6B.5. (Cauchy-Bunyakowski-Schwarz Inequality)
Let ), p L
2
(X). Then
), p)
|)|
2
|p|
2
.
Proof. Let = |p|
2
2
, 1 := ), p), and C := |)|
2
2
; thus, we are trying to show
that 1
C. Dene : 1 1 by (t) := |) t p|
2
2
. Then
(t) = ) t p, ) t p)
()
), )) t ), p) t p, )) +t
2
p, p)
= |)|
2
2
+ 2 ), p) t + |p|
2
2
t
2
= C + 21t +t
2
, (6B.6)
a quadratic polynomial in t. (Here, step (.) is by Proposition 6B.4(a)).
Now, (t) = |) t p|
2
2
0 for all t 1; thus, (t) has at most one root, so
the discriminant of the quadratic polynomial (6B.6) is not positive. That is
41
2
4C 0. Thus, 1
2
C, and thus, 1
C, as desired. 2
Note. The CBS inequality involves three integrals: ), p), |)|
2
, and |p|
2
.
But the proof of Theorem 6B.5 does not involve any integrals at all. Instead, it
just uses simple algebraic manipulations of the inner product operator. In par-
ticular, this means the same proof works whether we dene the inner product
using (6B.1) or using (6B.3). Indeed, the CBS inequality is not really about
1
2
spaces, per se it is actually a theorem about a much broader class of ab-
stract geometric structures, called inner product spaces. An enormous amount of
knowledge about L
2
(X) can be obtained from this abstract geometric approach,
usually through simple algebraic arguments like the proof of Theorem 6B.5 (i.e.
without lots of messy integration technicalities). This is the beginning of a beau-
tiful area of mathematics called Hilbert space theory (see [Con90] for an excellent
introduction).
6C
More about L
2
space
Prerequisites: 6B, 0C.
This section contains some material which is not directly germane to the
solution methods we present later in the book, but may be interesting to some
students who want a broader perspective.
6C.
More about 1
2
space 109
6C(i) Complex L
2
space
6B introduced the inner product for real-valued functions. The inner product
for complex-valued functions is slightly dierent. For any . = r + ji C, let
. := r ji denote the complex conjugate of .. Let X 1
D
be some domain,
and let ), p : X C be complex-valued functions. We dene
), p) :=
_
X
)(x) p(x) dx. (6C.1)
If p is real-valued, then p = p, and then eqn.(6C.1) is equivalent to eqn.(6B.4).
For any . C, recall that . . = [.[
2
. Thus, if ) is a complex-valued
function, then )(r))(r) = [)(r)[
2
. It follows that we can dene the 1
2
-norm of
an integrable function ) : X C just as before:
|)|
2
= ), ))
1/2
=
__
X
[)[
2
(x) dx
_
1/2
,
and this quantity will always be a real number (when the integral converges).
We dene L
2
(X; C) to be the set of all integrable functions ) : X C such
that |)|
2
< .
3
Proposition 6C.1. Properties of the complex inner product
The inner product on L
2
(X; C) has the following properties.
Sesquilinearity. For any )
1
, )
2
, p
1
, p
2
L
2
(X; C), and any constants /
1
, /
2
, c
1
, c
2

C,
/
1
)
1
+/
2
)
2
, c
1
p
1
+c
2
p
2
) = /
1
c
1
)
1
, p
1
) +/
1
c
2
)
1
, p
2
) +/
2
c
1
)
2
, p
1
) +/
2
c
2
)
2
, p
2
) .
Hermitian. For any ), p L
2
(X; C), ), p) = p, )).
Positive-denite. For any ) L
2
(X; C), ), )) is a real number and ), ))
0. Also, ), )) = 0 if and only if ) = 0.
CBS Inequality. For any ), p L
2
(X; C),
), p)
|)|
2
|p|
2
.
Proof. Exercise 6C.1 Hint: Imitate the proofs of Proposition 6B.4 and Theorem E _
6B.5. In your proof of the CBS inequality, dont forget that ), p)+), p) = 2Re [), p)].
2
3
We are using L
2
(X) to refer to only real-valued functions. In more advanced books, the
notation L
2
(X) denotes the set of complex-valued L
2
functions; if one wants to refer only to
real-valued L
2
functions, one must use the notation L
2
(X; 1).
6C(ii) Riemann vs. Lebesgue integrals
We have dened L
2
(X) to be the set of all integrable functions on X with nite
1
2
-norm, but we have been somewhat vague about what we mean by integrable.
The most familiar and elementary integral is the Riemann integral. For example,
if X = [o, /], and ) : X 1, then the Riemann integral of ) is dened
_
b
a
)(r) dr := lim
N
/ o
n=1
)
_
o +
n(/ o)
_
(6C.2)
A similar (but more complicated) denition can be given if X is an arbitrary
domain in 1
D
. We say ) is Riemann integrable if the limit (6C.2) exists and is
nite.
However, this is not what we mean here by integrable. The problem is that
the limit (6C.2) only exists if the function ) is reasonably nice (e.g. piecewise
continuous). We need an integral which works even for extremely nasty func-
tions (e.g. functions which are discontinuous everywhere; functions which have a
fractal structure, etc.). This object is called the Lebesgue integral; its denition
is similar to (6C.2) but much more complicated.
Loosely speaking, the Riemann sum in (6C.2) chops the interval [o, /] up into
equal subintervals. The corresponding sum in the Lebesgue integral allows us
to chop [o, /] into any number of Borel-measurable subsets. A Borel-measurable
subset is any open set, any closed set, any (countably innite) union or intersec-
tion of open or closed sets, any (countably innite) union or intersection of these
sets, etc. Clearly measurable subsets can become very complex. The Lebesgue
integral is obtained by taking a limit over all possible Riemann sums obtained
using such measurable partitions of [o, /]. This is a very versatile and powerful
construction, which can integrate incredibly bizarre and pathological functions.
(See Remark 10D.3 on page 211 for further discussion of Riemann vs. Lebesgue
integration).
You might ask, Why would I want to integrate bizarre and pathological
functions? Indeed, the sorts of functions which arise in applied mathematics
are almost always piecewise continuous, and for them, the Riemann integral
works just ne. To answer this, consider the dierence between the following
two equations:
(a) r
2
=
16
9
; (b) r
2
= 2.
Both equations have solutions, but they are dierent. The solutions to (a) are
rational numbers, for which we have an exact expression r = 4,3. The solutions
to (b) are irrational numbers, for which we have only approximate expressions:
r =
2 1.414213562....
Irrational numbers are pathological: they do not admit nice, simple, exact
expressions like 4,3. We might be inclined to ignore such pathological objects
in our mathematics to pretend they dont exist. Indeed, this was precisely
6C.
More about 1
2
space 111
the attitude of the ancient Greeks, whose mathematics was based entirely on
rational numbers. The problem is: in this ancient Greek mathematical universe,
equation (b) has no solution. This is not only inconvenient, it is profoundly
counterintuitive; after all,

2 is simply the length of the hypotenuse of a right
angle triangle whose other sides both have length 1. And surely the sidelength
of a triangle should be a number.
Furthermore we can nd rational numbers which seem to be arbitrarily good
approximations to a solution of equation (b). For example,
_
1, 414
100
_
2
= 1.999396;
_
1, 414, 213
100, 000
_
2
= 1.999998409;
_
141, 421, 356
10, 000, 000
_
2
= 1.999999993;
.
.
.
.
.
.
.
.
.
It certainly seems like this sequence of rational numbers is converging to some-
thing. Our name for that something is

2. In fact, this is the only way we can
ever specify

2. Since we cannot express

2 as a fraction or some simple deci-
mal expansion, we can only say,
2 is the number to which the above sequence

of rational numbers seems to be converging.
But how do we know that any such number exists? Couldnt there just be a
hole in the real number line where we think

2 is supposed to be? The answer
is that the set 1 is complete that is, any sequence in 1 which looks like it
is converging
4
does, in fact, converge to some limit point in 1. Because 1 is
complete, we are condent that

2 exists, even though we can never precisely
specify its value.
Now lets return to L
2
(X). Like the real line 1, the space L
2
(X) has a
geometry: a notion of distance dened by the 1
2
-norm ||
2
. This geometry
provides us with a notion of convergence in L
2
(X) (see 6E(i) on page 117). Like
1, we would like L
2
(X) to be complete, so that any sequence of functions which
looks like it is converging does, in fact, converge to some limit point in L
2
(X).
Unfortunately, a sequence of perfectly nice functions in L
2
(X) can converge
to a totally pathological limit function, the same way that a sequence of nice
rational numbers can converge to an irrational number. If we exclude the patho-
logical functions from L
2
(X), we will be like the ancient Greeks, who excluded
irrational numbers from their mathematics. We will encounter situations where
a certain equation should have a solution, but doesnt, just as the Greeks dis-
covered that the equation r
2
= 2 had no solution in their mathematics.
4
Technically, any Cauchy sequence.
Thus our denition of L
2
(X) must include some pathological functions. But
if these pathological functions are in L
2
(X), and L
2
(X) is dened as the set of
elements with nite norm, and the norm |)|
2
is dened using an integral like
(6B.2), then we must have a way of integrating these pathological functions.
Hence the necessity of the Lebesgue integral.
Fortunately, all the functions we will encounter in this book are Riemann
integrable. For the purposes of solving PDEs, you do not need to know how to
compute the Lebesgue integral. But it is important to know that it exists, and
that somewhere in the background, its presence is making all the mathematics
work properly.
6D Orthogonality
Prerequisites: 6A.
Two functions ), p L
2
(X) are orthogonal if ), p) = 0. Intuitively, this
means that ) and p are perpendicular vectors in the innite-dimensional vector
space L
2
(X).
Example 6D.1. Treat sin and cos as elements of L
2
[, ]. Then they are
orthogonal:
sin, cos) =
1
2
_

sin(r) cos(r) dr = 0. (Exercise 6D.1). E _

An orthogonal set of functions is a set )
1
, )
2
, )
3
, . . . of elements in L
2
(X)
such that )
j
, )
k
) = 0 whenever , ,= /. If, in addition, |)
j
|
2
= 1 for all ,, then
we say this is an orthonormal set of functions. Fourier analysis is based on
the orthogonality of certain families of trigonometric functions. Example 6D.1
was an example of this, which generalizes as follows....
Proposition 6D.2. Trigonometric Orthogonality on [, ]
For every n N, dene the functions S
n
, C
n
: [, ] 1 by S
n
(r) :=
sin (nr) and C
n
(r) := cos (nr), for all r [, ]. (See Figure 6D.1). Then the
set C
0
, C
1
, C
2
, . . . ; S
1
, S
2
, S
3
, . . . is an orthogonal set of functions for L
2
[, ].
In other words:
(a) S
n
, S
m
) =
1
2
_

sin(nr) sin(:r) dr = 0 , whenever n ,= :.

(b) C
n
, C
m
) =
1
2
_

cos(nr) cos(:r) dr = 0, whenever n ,= :.

(c) S
n
, C
m
) =
1
2
_

sin(nr) cos(:r) dr = 0, for any n and :.

6D. Orthogonality 113
1
0.5
0.5
1
3 2 1 1 2 3
x
C
1
(r) = cos (r)
1
0.5
0.5
1
3 2 1 1 2 3
x
C
2
(r) = cos (2r)
1
0.5
0.5
1
3 2 1 1 2 3
x
C
3
(r) = cos (3r)
1
0.5
0.5
1
3 2 1 1 2 3
x
C
4
(r) = cos (4r)
1
0.5
0.5
1
3 2 1 1 2 3
x
S
1
(r) = sin(r)
1
0.5
0.5
1
3 2 1 1 2 3
x
S
2
(r) = sin(2r)
1
0.5
0.5
1
3 2 1 1 2 3
x
S
3
(r) = sin(3r)
1
0.5
0.5
1
3 2 1 1 2 3
x
S
4
(r) = sin(4r)
Figure 6D.1: C
1
, C
2
, C
3
, and C
4
; S
1
, S
2
, S
3
, and S
4
(d) However, these functions are not orthonormal, because they do not have
unit norm. Instead, for any n ,= 0,
|C
n
|
2
=
1
2
_

cos(nr)
2
dr =
1
2
, and |S
n
|
2
=
1
2
_

sin(nr)
2
dr =
1
2
.
Proof. Exercise 6D.2 Hint: Use the trigonometric identities: 2 sin() cos() = E _
sin(+)+sin(), 2 sin() sin() = cos()cos(+), and 2 cos() cos() =
cos( +) + cos( ). 2
Remark. Notice that C
0
(r) = 1 is just the constant function.
It is important to remember that the statement, ) and p are orthogonal
depends upon the domain X which we are considering. For example, compare
the following theorem to the preceeding one...
Proposition 6D.3. Trigonometric Orthogonality on [0, 1]
Let 1 0, and, for every n N, dene the functions S
n
, C
n
: [0, 1] 1 by
S
n
(r) := sin
_
nr
1
_
and C
n
(r) := cos
_
nr
1
_
, for all r [0, 1].
(a) The set C
0
, C
1
, C
2
2
[0, 1]. In
other words: C
n
, C
m
) =
1
1
_
L
0
cos
_
n
1
r
_
cos
_
:
1
r
_
dr = 0,
whenever n ,= :.
However, these functions are not orthonormal, because they do not have
unit norm. Instead, for any n ,= 0, |C
n
|
2
=
1
1
_
L
0
cos
_
n
1
r
_
2
dr =
1
2
.
(b) The set S
1
, S
2
, S
3
2
[0, 1]. In
other words: S
n
, S
m
) =
1
1
_
L
0
sin
_
n
1
r
_
sin
_
:
1
r
_
dr = 0,
whenever n ,= :.
However, these functions are not orthonormal, because they do not have
unit norm. Instead, for any n ,= 0, |S
n
|
2
=
1
1
_
L
0
sin
_
n
1
r
_
2
dr =
1
2
.
(c) The functions C
n
and S
m
are not orthogonal to one another on [0, 1].
Instead:
S
n
, C
m
) =
1
1
_
L
0
sin
_
n
1
r
_
cos
_
:
1
r
_
dr =
_
_
0 if n +: is even
2n
(n
2
:
2
)
if n +: is odd.
Proof. Exercise 6D.3 . 2 E _
Remark. The trigonometric functions are just one of several important orthog-
onal sets of functions. Dierent orthogonal sets are useful for dierent domains
or dierent applications. For example, in some cases, it is convenient to use
a collection of orthogonal polynomial functions. Several orthogonal polynomial
families exist, including the Legendre Polynomials (see 16D on page 359), the
Chebyshev polynomials (see Exercise 14B.1(e) on page 278 of 14B(i)), the Her-
mite polynomials and the Laguerre polynomials. See [Bro89, Chap.3] for a good
introduction.
In the study of partial dierential equations, the following fact is particularly
important:
6D. Orthogonality 115
H
1
H
2
H
4
H
3
1
/
21
1
/
4
1
/
2
3
/
41
1
/
8
1
/
4
3
/
8
1
/
2
5
/
8
3
/
4
7
/
81
1
/
1
6
1
/
8
3
/
1
6
1
/
4
5
/
1
6
3
/
8
7
/
1
6
1
/
2
9
/
1
6
5
/
8
1
1
/
1
6
3
/
4
1
3
/
1
6
7
/
8
1
5
/
1
6
1
Figure 6D.2: Four Haar basis elements: H
1
, H
2
, H
3
, H
4
Let X 1
D
be any domain. If ), p : X C are two eigenfunc-
tions of the Laplacian with dierent eigenvalues, then ) and p are
orthogonal in L
2
(X).
(See Proposition 15E.9 on page 345 for a precise statement of this.) Because
of this, we can get orthogonal sets whose members are eigenfunctions of the
Laplacian (see Theorem 15E.12 on page 347). These orthogonal sets are the
building blocks with which we can construct solutions to a PDE satisfying
prescribed initial conditions or boundary conditions. This is the basic strategy
behind the solution methods of Chapters 11-14.
Exercise 6D.4. Figure 6D.2 portrays the The Haar Basis. We dene H
0
1, E _
and for any natural number N, we dene the th Haar function H
N
: [0, 1] 1
by:
H
N
(r) =
_
_
1 if
2n
2
N
r <
2n + 1
2
N
, for some n
_
0...2
N1
_
;
1 if
2n + 1
2
N
r <
2n + 2
2
N
, for some n
_
0...2
N1
_
.
(a) Show that the set H
0
, H
1
, H
2
, H
3
, . . . is an orthonormal set in L
2
[0, 1].
(b) There is another way to dene the Haar Basis. First recall that any number r [0, 1]
has a unique binary expansion of the form
r =
r
1
2
+
r
2
4
+
r
3
8
+
r
4
16
+ +
r
n
2
n
+
where r
1
, r
2
, r
3
, r
4
, . . . are all either 0 or 1. Show that, for any n 1,
H
n
(r) = (1)
xn
=
_
1 if r
n
= 0;
1 if r
n
= 1.
Exercise 6D.5 Figure 6D.3 portrays a Wavelet Basis. We dene W

0
1, E _
1/2 1 3/4 1/4
1/2 1 3/4 1/4 1/2 1 3/4 1/4
1/2 1 3/4 1/4 1/2 1 3/4 1/4
1/2 1
3/4
1/4 1/2 1 3/4 1/4
1;0
W
2;0
W
2;1
W
3;0
W
3;1
W
3;2
W
3;3
W
Figure 6D.3: Seven Wavelet basis elements: W
1,0
; W
2,0
, W
2,1
;
W
3,0
, W
3,1
, W
3,2
, W
3,3
and for any N and n
_
0...2
N1
_
, we dene
W
n;N
(r) =
_
_
1 if
2n
2
N
r <
2n + 1
2
N
;
1 if
2n + 1
2
N
r <
2n + 2
2
N
;
0 otherwise.
Show that the the set
W
0
; W
1,0
; W
2,0
, W
2,1
; W
3,0
, W
3,1
, W
3,2
, W
3,3
; W
4,0
, . . . , W
4,7
; W
5,0
, . . . , W
5,15
; . . .
is an orthogonal set in L
2
[0, 1], but is not orthonormal: for any and n, we have
|W
n;N
|
2
=
1
2
(N1)/2
.
6E Convergence concepts
Prerequisites: 4A.
If r
1
, r
2
, r
3
, . . . is a sequence of numbers, we know what it means to say
lim
n
r
n
= r. We can think of convergence as a kind of approximation.
6E. Convergence concepts 117
f (x)
f
2
1
(x)
f
2
2
1
1
f (x)
f
2
2
(x)
f
2
2
2
2
f (x)
f
3
2
(x)
f
2
2
3
3
f (x)
f
4
2
(x)
f
2
2
4
4
0 =
2
2 0 =0
2
0
0
Figure 6E.1: The sequence )
1
, )
2
, )
3
, . . . converges to the constant 0 function
in L
2
(X).
Heuristically speaking, if the sequence r
n
n=1
converges to r, then, for very
large n, the number r
n
is a good approximation of r.
If )
1
, )
2
, )
3
, . . . was a sequence of functions, and ) was some other function,
then we might want to say that lim
n
)
n
= ). We again imagine convergence
as a kind of approximation. Heuristically speaking, if the sequence )
n
n=1
converges to ), then, for very large n, the function )
n
is a good approximation of
).
However, there are several ways we can interpret good approximation, and
these in turn lead to several dierent notions of convergence. Thus, conver-
gence of functions is a much more subtle concept that convergence of numbers.
We will deal with three kinds of convergence here: 1
2
-convergence, pointwise
convergence, and uniform convergence.
6E(i) L
2
convergence
Let X 1
D
be some domain, and dene
` :=
_
_
_
_
X
1 dx if X is a nite domain;
1 if X is an innite domain.
If ), p L
2
(X), then the 1
2
-distance between ) and p is just
|) p|
2
:=
_
1
`
_
X
[)(x) p(x)[
2
dx
_
1/2
,
If we think of ) as an approximation of p, then |) p|
2
measures the root-
mean-squared error of this approximation.
Lemma 6E.1. ||
2
is a norm. That is:
(a) For any ) : X 1 and : 1, |: )|
2
= [:[ |)|
2
.
(b) (Triangle Inequality) For any ), p : X 1, |) +p|
2
|)|
2
+|p|
2
.
(c) For any ) : X 1, |)|
2
= 0 if and only if ) 0.
If )
1
, )
2
, )
3
, . . . is a sequence of successive approximations of ), then we
say the sequence converges to ) in 1
2
if lim
n
|)
n
)|
2
= 0 (sometimes
this is called convergence in mean square). See Figure 6E.1. We then write
) = L
2
lim
n
)
n
.
Example 6E.2. In each of the following examples, let X = [0, 1].
(a) Suppose )
n
(r) =
_
1 if 1,n < r < 2,n
0 otherwise
(Figure 6E.2A). Then |)
n
|
2
=
1
n
(Exercise 6E.2). Hence, lim
n
|)
n
|
2
= lim
n
1
n
= 0, so the se- E _
quence )
1
, )
2
, )
3
, . . . converges to the constant 0 function in L
2
[0, 1].
(b) For all n N, let )
n
(r) =
_
n if 1,n < r < 2,n;
0 otherwise
(Figure 6E.2B).
Then |)
n
|
2
=

n (Exercise 6E.3). Hence, lim
n
|)
n
|
2
= lim
n
n = E _
, so the sequence )
1
, )
2
, )
3
, . . . does not converge to zero in L
2
[0, 1].
(c) For each n N, let )
n
(r) =
_
1 if

1
2
r
<
1
n
;
0 otherwise
. Then the sequence
)
n
n=1
converges to 0 in 1
2
. (Exercise 6E.4 ) E _
(d) For all n N, let )
n
(r) =
1
1 +n
r
1
2
. Figure 6E.3 portrays elements

)
1
, )
10
, )
100
, and )
1000
; these picture strongly suggest that the sequence
is converging to the constant 0 function in L
2
[0, 1]. The proof of this is
Exercise 6E.5 . E _
(e) Recall the Wavelet functions from Example 6D.4(b). For any N
and n
_
0..2
N1
_
, we had |W
N,n
|
2
=
1
2
(N1)/2
. Thus, the sequence of
wavelet basis elements converges to the constant 0 function in L
2
[0, 1].
119
1
1
1
1
1
1
1
1
1
1
1
1
f
2
f
3
f
4
f
5
f
6
f
7
2
1
3
1
4
1
5
1
7
1
6
1
f
2
f
3
f
4
f
5
f
6
f
7
LabelA LabelB
PSfrag replacements
1
2
1
2
1
2
1
2
1
3
1
3
1
3
1
3
1
4
1
4
1
5
1
5
1
6
1
6
1
7
1
7
2
3
2
3
2
5
2
5
2
7
2
7
Figure 6E.2: (A) Examples 6E.2(a), 6E.5(a), and 6E.9(a); (B) Examples
6E.2(b) and 6E.5(b).
0.7
0.75
0.8
0.85
0.9
0.95
1
0 0.2 0.4 0.6 0.8 1
x
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
x
)
1
(r) =
1
1+1[x
1
2
[
)
10
(r) =
1
1+10[x
1
2
[
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
x
0
0.2
0.4
0.6
0.8
0.2 0.4 0.6 0.8 1
x
)
100
(r) =
1
1+100[x
1
2
[
)
1000
(r) =
1
1+1000[x
1
2
[
Figure 6E.3: Examples 6E.2(c) and 6E.5(c): If )
n
(r) =
1
1+n[x
1
2
[
, then the
sequence )
1
, )
2
, )
3
, . . . converges to the constant 0 function in L
2
[0, 1].
Note that, if we dene p
n
= ) )
n
for all n N, then
_
)
n

n
) in 1
2
_

_
p
n

n
0 in 1
2
_
Hence, to understand 1
2
-convergence in general, it is sucient to understand
1
2
-convergence to the constant 0 function.
Lemma 6E.3. The inner product function , ) is continuous with respect
to 1
2
convergence. That is: if )
1
, )
2
, )
3
, . . . and p
1
, p
2
, p
3
, . . . are two se-
quences of functions in L
2
(X), and L
2
lim
n
)
n
= ) and L
2
lim
n
p
n
= p, then
lim
n
)
n
, p
n
) = ), p).
6E(ii) Pointwise convergence
Convergence in 1
2
only means that the average approximation error gets small.
It does not mean that lim
n
)
n
(x) = )(x) for every x X. If this equation is
true, then we say that the sequence )
1
, )
2
, . . . converges pointwise to ) (see
Figure 6E.4). We then write ) lim
n
)
n
. Pointwise convergence is generally
considered stronger than 1
2
convergence because of the following result:
Theorem 6E.4. Let X 1
D
be a bounded domain, and let )
1
, )
2
, . . . be a
sequence of functions in L
2
(X). Suppose:
(a) All the functions are uniformly bounded that is, there is some ` 0
such that [)
n
(r)[ < ` for all n N and all r X.
w x
y
z
f
1
(
w
)
f
4
(
w
)
f
1
(
x
)
f
2
(
x
)
f
4
(
x
)
f
1
(
x
)
f
4
(
x
)
f
3
(
x
)
f
1
(
z
)
f
2
(
z
)
f
3
(
x
)
f
2
(
w
)
f
2
(
x
)
f
3
(
w
)
f
4
(
z
)
f
3
(
z
)
Figure 6E.4: The sequence )
1
, )
2
, )
3
, . . . converges pointwise to the constant
0 function. Thus, if we pick some random points n, r, j, . X, then we see that
lim
n
)
n
(n) = 0, lim
n
)
n
(r) = 0, lim
n
)
n
(j) = 0, and lim
n
)
n
(.) = 0.
(b) The sequence )
1
, )
2
, . . . converges pointwise to some function ) L
2
(X).
Then the sequence )
1
, )
2
, . . . also converges to ) in 1
2
. 2
Proof. Exercise 6E.7 Hint: You may use the following special case of Lebesgues E _
Dominated Convergence Theorem:
5
Let p
1
, p
2
, . . . be a sequence of integrable functions on the domain X.
Let p : X 1 be another such function. Suppose that
(a) There is some some 1 0 such that [p
n
(r)[ < 1 for all n N and
all r X.
(b) For all r X, lim
n
p
n
(r) = p(r).
Then lim
n
_
X
p
n
(r) dr =
_
X
p(r) dr.
5
See [Fol84, Thm.2.24, p.53] or [KF75, 30.1, p.303].
0.7
0.75
0.8
0.85
0.9
0.95
1
0 0.2 0.4 0.6 0.8 1
x
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1 x
p
1
(r) =
1
1+1[x
1
2
[
p
5
(r) =
1
1+5[x
1
10
[
0
0.2
0.4
0.6
0.8
1
0.2 0.4 0.6 0.8 1
x
0
0.2
0.4
0.6
0.8
0.2 0.4 0.6 0.8 1
x
p
10
(r) =
1
1+10[x
1
20
[
p
15
(r) =
1
1+15[x
1
30
[
0
0.2
0.4
0.6
0.2 0.4 0.6 0.8 1
x
0
0.1
0.2
0.3
0.4
0.5
0.6
0.2 0.4 0.6 0.8 1
x
p
30
(r) =
1
1+30[x
1
60
[
p
50
(r) =
1
1+50[x
1
100
[
Figure 6E.5: Examples 6E.5(d) and 6E.9(d): If p
n
(r) =
1
1+n[x
1
2n
[
, then the
sequence p
1
, p
2
, p
3
, . . . converges pointwise to the constant 0 function on [0, 1].
Let p
n
:= [) )
n
[
2
for all n N, and let p = 0. Apply the Dominated Convergence
Theorem. 2
(a) As in Example 6E.2(a), for each n N, let )
n
(r) =
_
1 if 1,n < r < 2,n;
0 otherwise
.
(Fig.6E.2A). The sequence )
n
n=1
converges pointwise to the constant
0 function on [0, 1]. Also, as predicted by Theorem 6E.4, the sequence
)
n
n=1
converges to the constant 0 function in 1
2
(see Example 6E.2(a)).
(b) As in Example 6E.2(b), for each n N, let )
n
(r) =
_
n if 1,n < r < 2,n;
0 otherwise
(Fig.6E.2B). Then this sequence converges pointwise to the constant 0
function, but does not converge to zero in L
2
[0, 1]. This illustrates the
importance of the boundedness hypothesis in Theorem 6E.4.
(c) As in Example 6E.2(c), for each n N, let )
n
(r) =
_
1 if

1
2
r
<
1
n
;
0 otherwise
.
Then the sequence )
n
n=1
does not converges to 0 in pointwise, although
it does converge in 1
2
.
f (x)
f (x)
f
oo
Figure 6E.6: The uniform norm of ) is dened: |)|
:= sup
xX
[)(r)[.
(d) Recall the functions )
n
(r) =
1
1 +n
r
1
2
from Example 6E.2(d). This

sequence of functions converges to zero in L
2
[0, 1], however, it does not
converge to zero pointwise (Exercise 6E.8 ). E _
(e) For all n N, let p
n
(r) =
1
1 +n
r
1
2n
. Figure 6E.5 on the facing

page portrays elements p
1
, p
5
, p
10
, p
15
, p
30
, and p
50
; These picture strongly
suggest that the sequence is converging pointwise to the constant 0 function
on [0, 1]. The proof of this is Exercise 6E.9 . E _
(f) Recall from Example 6E.2(e) that the sequence of Wavelet basis elements
W
N;n
converges to zero in L
2
[0, 1]. Note, however, that it does not
converge to zero pointwise (Exercise 6E.10 ). E _
n
= ) )
n
for all n N, then
_
)
n

n
) pointwise
_

_
p
n

n
0 pointwise
_
Hence, to understand pointwise convergence in general, it is sucient to under-
stand pointwise convergence to the constant 0 function.
6E(iii) Uniform convergence
There is an even stronger form of convergence. If ) : X 1 is a function, then
the uniform norm of ) is dened:
|)|
:= sup
xX
)(x)
.
This measures the farthest deviation of the function ) from zero (see Figure
6E.6).
Example 6E.6. Suppose X = [0, 1], and )(r) =
1
3
r
3
1
4
r (as in Figure 6E.8A).
The minimal point of ) is r =
1
2
, where )
t
(
1
2
) = 0 and )(
1
2
) =
1
12
. The
maximal point of ) is r = 1, where )(1) =
1
12
. Thus, [)(r)[ takes a maximum
value of
1
12
at either point, so that |)|
= sup
0x1
1
3
r
3
1
4
r
=
1
12
.
Lemma 6E.7. ||
is a norm. That is:

(a) For any ) : X 1 and : 1, |: )|
= [:[ |)|
.
(b) (Triangle Inequality) For any ), p : X 1, |) +p|
|)|
+|p|
.
(c) For any ) : X 1, |)|
= 0 if and only if ) 0.
f(x)
g(x)
Figure 6E.7: If |) p|
< c, this means that p(r) is conned within an c-tube

around ) for all r.
The uniform distance between two functions ) and p is then given by:
|) p|
= sup
xX
)(x) p(x)
.
One way to interpret this is portrayed in Figure 6E.7. Dene a tube of width c
around the function ). If |) p|
< c, this means that p(r) is conned within

this tube for all r.
Example 6E.8. Let X = [0, 1], and suppose )(r) = r(r + 1) and p(r) = 2r (as
in Figure 6E.8B). For any r [0, 1],
[)(r) p(r)[ =

r
2
+r 2r
r
2
r
= r r
2
.
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
0.1
0 0.2 0.4 0.6 0.8 1
f
1/12
-1/12
0
0.5
1
1.5
2
0 0.2 0.4 0.6 0.8 1
g
(
x
)
f
(
x
)
|f(x)-g(x)|
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
0.5
0 0.2 0.4 0.6 0.8 1
1
2
3
4
(A) (B) (C)
Figure 6E.8: (A) The uniform norm of )(r) =
1
3
r
3
1
4
r (Example 6E.6). (B)
The uniform distance between )(r) = r(r + 1) and p(r) = 2r (Example 6E.8).
(C) p
n
(r) =

r
1
2
n
, for n = 1, 2, 3, 4, 5 (Example (7b))
(because it is nonnegative). This expression takes its maximum at r =
1
2
(to
see this, solve for )
t
(r) = 0), and its value at r =
1
2
is
1
4
. Thus, |) p|
=
sup
xX
r(r 1)
=
1
4
.
Let p
1
, p
2
, p
3
, . . . be functions from X to 1, and let ) : X 1 be some
other function. The sequence p
1
, p
2
, p
3
, . . . converges uniformly to ) if
lim
n
|p
n
)|
= 0. We then write ) = uniflim

n
p
n
. This means not
only that lim
n
p
n
(x) = )(x) for every x X, but furthermore, that the func-
tions p
n
converge to ) everywhere at the same speed. This is portrayed in
Figure 6E.9. For any c 0, we can dene a tube of width c around ), and,
no matter how small we make this tube, the sequence p
1
, p
2
, p
3
, . . . will even-
tually enter this tube and remain there. To be precise: there is some such
that, for all n , the function p
n
is conned within the c-tube around ) i.e.
|) p
n
|
< c.
(a) Suppose, as in Example 6E.5(a) on page 122, and Figure 6E.2B on page
119, that
p
n
(r) =
_
1 if
1
n
< r <
2
n
;
0 otherwise.
Then the sequence p
1
, p
2
, . . . converges pointwise to the constant zero
function, but does not converge to zero uniformly on [0, 1]. (Exercise 6E.12
Verify these claims.). E _
g
1
(x)
f(x)
g
2
(x)
g
3
(x)
g
oo
(x) = f(x)
f(x)
Figure 6E.9: The sequence p
1
, p
2
, p
3
, . . . converges uniformly to ).
(b) If p
n
(r) =

r
1
2
n
(see Figure 6E.8C), then |p
n
|
=
1
2
n
(Exercise 6E.13
). Thus, the sequence p
1
, p
2
, . . . converges to zero uniformly on [0, 1], be- E _
cause lim
n
|p
n
|
= lim
n
1
2
n
= 0.
(c) If p
n
(r) = 1,n for all r [0, 1], then the sequence p
1
, p
2
, . . . converges to
zero uniformly on [0, 1] (Exercise 6E.14 ). E _
(d) Recall the functions p
n
(r) =
1
1+n[x
1
2n
[
from Example 6E.5(e) (Figure 6E.5
on page 122). The sequence p
1
, p
2
, . . . converges pointwise to the con-
stant zero function, but does not converge to zero uniformly on [0, 1].
(Exercise 6E.15 Verify these claims.). E _
n
= ) )
n
for all n N, then
_
)
n

n
) uniformly
_

_
p
n

n
0 uniformly
_
Hence, to understand uniform convergence in general, it is sucient to under-
stand uniform convergence to the constant 0 function.
Uniform convergence is the best kind of convergence. It has the most useful
consequences, but it is also the most dicult to achieve. (In many cases, we
must settle for pointwise or 1
2
convergence instead.) For example, the following
consequences of uniform convergence are extremely useful.
Proposition 6E.10. Let X 1
D
be some domain. Let )
1
, )
2
, )
3
, . . . be
functions from X to 1, and let ) : X 1 be some other function. Suppose
)
n

n
) uniformly.
(a) If )
n
n=1
are all continuous on X, then ) is also continuous on X.
(b) If X is compact (that is, closed and bounded), then lim
n
_
X
)
n
(r) dr =
_
X
)(r) dr.
(c) Suppose the functions )
n
n=1
are all dierentiable on X, and suppose
)
t
n

n
1 uniformly. Then ) is also dierentiable, and )
t
= 1.
Proof. (a) Exercise 6E.16 (Slightly challenging; for students with some analysis E _
background).
For (b,c) see e.g. [Asm05, Theorems 4 and 5, p.91-92 of 2.9]. 2
Note that Proposition 6E.10(a,c) are false if we replace uniformly with
pointwise or in 1
2
. (Proposition 6E.10(b) is sometimes true under these con-
ditions, but only if we also add additional hypotheses.) Indeed, the next result
says that uniform convergence is logically stronger than either pointwise or 1
2
convergence.
Corollary 6E.11. Let )
1
, )
2
, )
3
, . . . be functions from X to 1, and let
) : X 1 be some other function.
(a) If )
n

n
) uniformly, then )
n

n
) pointwise.
(b) Suppose X is compact (that is, closed and bounded). If )
n

n
)
uniformly, then:
[i] )
n

n
) in 1
2
.
[ii] For any p L
2
(X), we have lim
n
)
n
, p) = ), p).
Proof. Exercise 6E.17 (a) is easy. For (b), use Proposition 6E.10(b). 2 E _
Sometimes, uniform convergence is a little too much to ask for, and we must
settle for a slightly weaker form of convergence. Let X 1
D
be some domain.
Let p
1
, p
2
, p
3
, . . . be functions from X to 1, and let ) : X 1 be some other
function. The sequence p
1
, p
2
, p
3
, . . . converges semiuniformly to ) if:
(a) p
1
, p
2
, p
3
, . . . converges pointwise to ) on X; i.e. )(r) = lim
n
p
n
(r) for
all r X.
(b) p
1
, p
2
, p
3
, . . . converges uniformly to ) on any closed subset of int (X). In
other words, if Y int (X) is any closed set, then
lim
n
_
sup
yY
[)(j) p
n
(j)[
_
= 0.
Heuristically speaking, this means that the sequence p
n
n=1
is trying to con-
verge to ) uniformly on X, but it is maybe getting stuck at some of the boundary
points of X.
Example 6E.12. Let X := (0, 1). Recall the functions p
n
(r) =
1
1+n[x
1
2n
[
from Figure 6E.5 on page 122. By Example 6E.9(d) on page 126, we know
that this sequence doesnt converge uniformly to 0 on (0, 1). However, it does
converge semiuniformly to 0. First, we know it converges pointwise on (0, 1),
by Example 6E.5(e) on page 123. Second, if 0 < o < / < 1, it is easy
to check that p
n
n=1
converges to ) uniformly on the closed interval [o, /]
(Exercise 6E.18). It follows that p
n
n=1
converges to ) uniformly on any E _
closed subset of (0, 1).
Summary. The various forms of convergence are logically related as follows:
_
Uniform convergence
_

_
Semiuniform convergence
_

_
Pointwise convergence
_
.
Also, if X is compact, then
_
Uniform convergence
_
=
_
Convergence in 1
2
_
.
Finally, if the sequence of functions is uniformly bounded and X is compact, then
_
_
=
_
Convergence in 1
2
_
.
However, the opposite implications are not true. In general:
_
Convergence in 1
2
_
,=
_
_
,=
_
Uniform convergence
_
6E(iv) Convergence of function series
Let )
1
, )
2
, )
3
, . . . be functions from X to 1. The function series
n=1
)
n
is the
formal innite summation of these functions; we would like to think of this series
as dening another function from X to 1. . Intuitively, the symbol
n=1
)
n
should represent the function which arises as the limit lim

N
1
N
, where, for each
N, 1
N
(r) :=
N
n=1
)
n
(r) = )
1
(r) + )
2
(r) + + )
N
(r) is the th partial
sum. To make this precise, we must specify the sense in which the partial sums
1
1
, 1
2
, . . . converge. If 1 : X 1 is this putative limit function, then we say
that the series
n=1
)
n
....
...converges in 1
2
to 1 if 1 = L
2
lim
N
N
n=1
)
n
. We then write 1

L2
n=1
)
n
.
...converges pointwise to 1 if, for each r X, 1(r) = lim
N
N
n=1
)
n
(r).
We then write 1
n=1
)
n
.
...converges uniformly to 1 if 1 = uniflim
N
N
n=1
)
n
. We then write
1
unif
n=1
)
n
.
The next result provides a useful condition for the uniform convergence of an
innite summation of functions; we will use this result often in our study of
Fourier series and other eigenfunction expansions in Chapters 7 to 9:
Proposition 6E.13. Weierstrass `-test
Let )
1
, )
2
, )
3
, . . . be functions from X to 1. For every n N, let `
n
:= |)
n
|
.
If
n=1
`
n
< , then the series
n=1
)
n
converges uniformly on X.
Proof. Exercise 6E.19 (a) Show that the series converges pointwise to some limit E _
function ) : X 1.
(b) For any N, show that
_
_
_
_
_
1
N
n=1
)
n
_
_
_
_
_
n=N+1
`
n
.
(c) Show that lim
N
n=N+1
`
n
= 0. 2
The next three sucient conditions for convergence are also sometimes useful
(but they are not used later in this book).
Proposition 6E.14. Dirichlet Test
Let )
1
, )
2
, )
3
, . . . be functions from X to 1. Let c
k
k=1
be a sequence of (Optional)
positive real numbers. Then the series
n=1
c
n
)
n
converges uniformly on X if:
lim
n
c
k
= 0; and
There is some ` 0 such that, for all N, we have
_
_
_
_
_
N
n=1
)
n
_
_
_
_
_
< `.
Proof. See [Asm05, Appendix to 2.10, p.99] 2
Proposition 6E.15. Cauchys Criterion
Let )
1
, )
2
, )
3
, . . . be functions from X to 1. For every N, let C
N
:= (Optional)
sup
M>N
_
_
_
_
_
M
n=N
)
n
_
_
_
_
_
.
Then
_
The series
n=1
)
n
converges uniformly on X
_

_
lim
N
C
N
= 0
_
.
Proof. See [CB87, 88]. 2
Proposition 6E.16. Abels Test
Let X 1
N
and Y 1
M
be two domains. Let )
1
, )
2
, )
3
, . . . be a sequence (Optional)
of functions from X to 1, such that the series
n=1
)
n
converges uniformly on X.
Let p
1
, p
2
, p
3
, . . . be another sequence of functions from Y to 1, and consider
the sequence /
1
, /
2
, . . . of functions from X Y to 1, dened by /
n
(r, j) :=
)
n
(r)p
n
(j). Suppose:
(a) The sequence p
n
n=1
is uniformly bounded; i.e. there is some ` 0 such
that [p
n
(j)[ < ` for all n N and j Y.
6F. Orthogonal and orthonormal Bases 131
(b) The sequence p
n
n=1
is monotonic; i.e. either p
1
(j) p
2
(j) p
3
(j)
for all j Y, or p
1
(j) p
2
(j) p
3
(j) for all j Y.
Then the series
n=1
/
n
converges uniformly on X Y.
Proof. See [CB87, 88]. 2
6F Orthogonal and orthonormal Bases
Prerequisites: 6A, 6E(i). Recommended: 6E(iv).
An orthogonal set in L
2
(X) is a (nite or innite) collection of functions
b
1
, b
2
, b
3
, . . . such that b
k
, b
j
) = 0 whenever / ,= ,. Intuitively, the vectors
b
1
, b
2
, b
3
, . . . are all perpendicular to one another in the innite-dimensional
geometry of L
2
(X). One consequence is an 1
2
-version of the Pythagorean
Formula: For any N and any real numbers :
1
, :
2
, . . . , :
N
1, we have
|:
1
b
1
+:
2
b
2
+ +:
N
b
n
|
2
2
= :
2
1
|b
1
|
2
2
+:
2
2
|b
N
|
2
2
+ +:
2
N
|b
N
|
2
2
.
(6F.1)
(Exercise 6F.1 Verify the 1
2
Pythagorean formula). E _
An orthogonal basis for L
2
(X) is an innite collection of functions b
1
, b
2
, b
3
, . . .
such that:
b
1
, b
2
, b
3
, . . . form an orthogonal set (i.e. b
k
, b
j
) = 0 whenever / ,= ,.)
For any g L
2
(X), if we dene
n
=
g, b
n
)
|b
n
|
2
2
, for all n N, then
g

L2
n=1
n
b
n
.
Recall that this means that lim
N
_
_
_
_
_
g
N
n=1
n
b
n
_
_
_
_
_
2
= 0. In other words, we
can approximate g as closely as we want in 1
2
norm with a partial sum
N
n=1
n
b
n
,
if we make large enough.
An orthonormal basis for L
2
(X) is an innite collection of functions b
1
, b
2
, b
3
, . . .
such that:
|b
k
|
2
= 1 for every /.
b
1
, b
2
, b
3
, . . . is an orthogonal basis for L
2
(X). In other words, b
k
, b
j
) =
0 whenever / ,= ,, and, for any g L
2
(X), if we dene
n
= g, b
n
) for all
n N, then g

L2
n=1
n
b
n
.
One consequence of this is
Theorem 6F.1. Parsevals Equality
Let b
1
, b
2
, b
3
, . . . be an orthonormal basis for L
2
(X), and let f , g L
2
(X).
Let
n
:= f , b
n
) and
n
:= g, b
n
) for all n N. Then
(a) f , g) =
n=1
n
.
(b) |g|
2
2
=
n=1
[
n
[
2
.
Proof. Exercise 6F.2 Hint: For all N, let F
N
:=

N
n=1
n
b
n
and G
N
:= E _

N
n=1
n
b
n
.
(i) Show that F
N
, G
N
) =

N
n=1
n
(Hint: the functions b
1
, . . . , b
N
are or-
thonormal).
(ii) To prove (a), show that f , g) = lim
N
F
N
, G
N
) (Hint: Use Lemma 6E.3).
(iii) To prove (b), set f = g in (a). 2
The idea of Fourier analysis is to nd an orthogonal basis for an 1
2
-space,
using familiar trigonometric functions. We will return to this in Chapter 7.
Further reading:
Most of the mathematically rigorous texts on partial dierential equations (such
as [CB87], [Asm05] or [Eva91, Appendix D]) contain detailed and thorough dis-
cussions of 1
2
space, orthogonal basis, and the various convergence concepts
discussed in this chapter. This is because almost all solutions to partial dif-
ferential equations arise through some sort of innite series or approximating
sequence; hence it is essential to properly understand the various forms of func-
tion convergence and their relationships.
The convergence of sequences of functions is part of a subject called real anal-
ysis, and any advanced textbook on real analysis will contain extensive material
on convergence. There are many other forms of function convergence we havent
even mentioned in this chapter, including L
p
convergence (for any value of j
between 1 and ), convergence in measure, convergence almost everywhere, and
weak* convergence. Dierent convergence modes are useful in dierent contexts,
6G. Practice problems 133
1/5 2/5 3/5 4/5 1
5
1/2
1 2 3 4 5
5
1/2
5
1
1
1
(A)
(B)
(C)
f
5
f
5
Figure 6G.1: Problems for Chapter 6
and the logical relationships between them are fairly subtle. See [Fol84, 2.4]
for a good summary. Other standard references are [WZ77, Chap.8], [KF75,
28.4-28.5; 37], [Rud87] or [Roy88].
The geometry of innite-dimensional vector spaces is called functional analy-
sis, and is logically distinct from the convergence theory for functions (although
of course, most of the important innite dimensional spaces are spaces of func-
tions). Innite-dimensional vector spaces fall into several broad classes, depend-
ing upon the richness of the geometric and topological structure, which include
Hilbert spaces [such as L
2
(X)], Banach Spaces [such as ((X) or L
1
(X)] and locally
convex spaces. An excellent introduction to functional analysis is [Con90]. Other
standard references are [Fol84, Chap.5] and [KF75, Chap.4]. Hilbert spaces are
the mathematical foundation of quantum mechanics; see [Pru81, BEH94].
6G Practice problems
1. Let X = (0, 1]. For any n N, dene the function )
n
: (0, 1] 1 by
)
n
(r) = exp(nr). (Fig. 6G.1A)
(a) Compute |)
n
|
2
for all n N.
(b) Does the sequence )
n
n=1
converge to the constant 0 function in
L
2
(0, 1]? Explain.
(c) Compute |)
n
|
for all n N.
(d) Does the sequence )
n
n=1
converge to the constant 0 function uni-
formly on (0, 1]? Explain.
(e) Does the sequence )
n
n=1
converge to the constant 0 function point-
wise on (0, 1]? Explain.
2. Let X = [0, 1]. For any n N, dene )
n
: [0, 1] 1 by )
n
(r) =
_
n if
1
n
r <
2
n
0 otherwise
. (Fig. 6G.1B)
(a) Does the sequence )
n
n=1
wise on [0, 1]? Explain.
(b) Compute |)
n
|
2
for all n N.
(A) (B)
1/2 1
f
n
1/2 1
f
n
1
Figure 6G.2: Problems for Chapter 6
(c) Does the sequence )
n
n=1
L
2
[0, 1]? Explain.
(d) Compute |)
n
|
for all n N.
n
n=1
formly on [0, 1]? Explain.
3. Let X = 1. For any n N, dene )
n
: 1 1 by )
n
(r) =
_
1
n
if 0 r < n
0 otherwise
.
(Fig. 6G.1C)
(a) Compute |)
n
|
for all n N.
(b) Does the sequence )
n
n=1
formly on 1? Explain.
n
n=1
wise on 1? Explain.
(d) Compute |)
n
|
2
for all n N.
n
n=1
L
2
(1)? Explain.
4. Let X = (0, 1]. For all n N, dene )
n
: (0, 1] 1 by )
n
(r) =
1
3
nr
(for all r (0, 1]). (Figure 6G.2A)
n
n=1
wise on (0, 1]? Why or why not?
(b) Compute |)
n
|
2
for all n N.
n
n=1
L
2
(0, 1]? Why or why not?
6G. Practice problems 135
(d) Compute |)
n
|
for all n N.
n
n=1
formly on (0, 1]? Explain.
5. Let X = [0, 1]. For all n N, dene )
n
: [0, 1] 1 by )
n
(r) =
1
(nr + 1)
2
(for all r [0, 1]). (Figure 6G.2B)
n
n=1
wise on [0, 1]? Explain.
(b) Compute |)
n
|
2
for all n N.
n
n=1
L
2
[0, 1]? Explain.
(d) Compute |)
n
|
for all n N.
Hint: Look at the picture. Where is the value of )
n
(r) largest?
n
n=1
formly on [0, 1]? Explain.
6. In each of the following cases, you are given two functions ), p : [0, ] 1.
Compute the inner product ), p).
(a) )(r) = sin(3r), p(r) = sin(2r).
(b) )(r) = sin(nr), p(r) = sin(:r), with n ,= :.
(c) )(r) = sin(nr) = p(r) for some n N. Question: What is |)|
2
?
(d) )(r) = cos(3r), p(r) = cos(2r).
(e) )(r) = cos(nr), p(r) = cos(:r), with n ,= :.
(f) )(r) = sin(3r), p(r) = cos(2r).
7. In each of the following cases, you are given two functions ), p : [, ]
1. Compute the inner product ), p).
(a) )(r) = sin(nr), p(r) = sin(:r), with n ,= :.
(b) )(r) = sin(nr) = p(r) for some n N. Question: What is |)|
2
?
(c) )(r) = cos(nr), p(r) = cos(:r), with n ,= :.
(d) )(r) = sin(3r), p(r) = cos(2r).
8. Determine if )
n
converges to ) pointwise, in 1
2
(X), or uniformly.
(a) )
n
(r) = c
nx
2
, )(r) = 0, X = [1, 1].
(b) )
n
(r) = nsin(r,n), )(r) = r, X = [, ].
137
Chapter 7
Fourier sine series and cosine
series
The art of doing mathematics consists in nding that special case which contains all the
germs of generality. David Hilbert
7A Fourier (co)sine series on [0, ]
Prerequisites: 6E(iv), 6F.
Throughout this section, for all n N, we dene the functions S
n
: [0, ]
1 and C
n
: [0, ] 1 by S
n
(r) := sin(nr) and C
n
(r) := cos(nr), for all
r [0, ] (see Figure 6D.1 on page 113).
7A(i) Sine series on [0, ]
Recommended: 5C(i).
Suppose ) L
2
[0, ] (i.e. ) : [0, ] 1 is a function with |)|
2
< ). We
dene the Fourier sine coecients of ):
1
n
:=
), S
n
)
|S
n
|
2
2
=
2
_

0
)(r) sin(nr) dr, for all n 1. (7A.1)
The Fourier sine series of ) is then the innite summation of functions:
n=1
1
n
S
n
(r). (7A.2)
A function ) : [0, ] 1 is continuously dierentiable on [0, ] if )
is continuous on [0, ] (hence, bounded), )
t
(r) exists for all r (0, ), and
furthermore, the function )
t
: (0, ) 1 is itself bounded and continuous on
(0, ). Let (
1
[0, ] be the space of all continuously dierentiable functions.
138 DRAFT Chapter 7. Fourier sine series and cosine series
We say ) is piecewise continuously dierentiable (or piecewise (
1
, or
sectionally smooth) if there exist points 0 = ,
0
< ,
1
< ,
2
< < ,
M+1
=
(for some ` N) such that ) is bounded and continuously dierentiable on
each of the open intervals (,
m
, ,
m+1
); these are called (
1
intervals for ). In
particular, any continuously dierentiable function on [0, ] is piecewise contin-
uously dierentiable (in this case, ` = 0 and the set ,
1
, . . . , ,
M
is empty, so
all of (0, ) is a (
1
interval).
Exercise 7A.1. (a) Show that any continuously dierentiable function has nite E _
1
2
-norm. In other words, (
1
[0, ] L
2
[0, ].
(b) Show that any piecewise (
1
function on [0, ] is in L
2
[0, ].
Theorem 7A.1. Fourier Sine Series Convergence on [0, ]
(a) The set S
1
, S
2
, S
3
2
[0, ]. Thus, if )
L
2
[0, ], then the sine series (7A.2) converges to ) in 1
2
-norm, i.e. )

L2
n=1
1
n
S
n
.
Furthermore, the coecient sequence 1
n
n=1
is the unique sequence of
coecients with this property. In other words, if 1
t
n
n=1
is some other
sequence of coecients such that )

L2
n=1
1
t
n
S
n
, then we must have 1
t
n
=
1
n
for all n N.
(b) If ) (
1
[0, ], then the sine series (7A.2) converges pointwise on (0, ).
More generally, if ) is piecewise (
1
, then the sine series (7A.2) converges
to ) pointwise on each (
1
interval for ). In other words, if ,
1
, . . . , ,
m
is
the set of discontinuity points of ) and/or )
t
, and ,
m
< r < ,
m+1
, then
)(r) = lim
N
N
n=1
1
n
sin(nr).
(c) If
n=1
[1
n
[ < , then the sine series (7A.2) converges to ) uniformly on
[0, ].
(d) [i] If ) is continuous and piecewise dierentiable on [0, ], and )
t
L
2
[0, ],
and ) satises homogeneous Dirichlet boundary conditions (i.e. )(0) =
)() = 0), then the sine series (7A.2) converges to ) uniformly on [0, ].
[ii] Conversely, if the sine series (7A.2) converges to ) uniformly on [0, ],
then ) is continuous on [0, ], and satises homogeneous Dirichlet boundary
conditions.
7A. Fourier (co)sine series on [0, ] 139
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
= 1 = 3 = 5 = 7 = 9
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
1.2
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
= 11 = 21 = 41 = 101 = 2001
Figure 7A.1:
4
n=1
n odd
1
n
sin(nr), for = 1, 3, 5, 7, 9, 11, 21, 41, and 2001. Notice
the Gibbs phenomenon in the plots for large .
(e) If ) is piecewise (
1
, and K (,
m
, ,
m+1
) is any closed subset of a (
1
interval
of ), then the series (7A.2) converges uniformly to ) on K.
(f ) Suppose 1
n
n=1
is a nonnegative sequence decreasing to zero. (That is,
1
1
1
2
0 and lim
n
1
n
= 0). If 0 < o < / < , then the series
(7A.2) converges uniformly to ) on [o, /].
Proof. (c) is Exercise 7A.2 (Hint: Use the Weierstrass `-test, Proposition 6E.13 E _
on page 129.)
(a,b,e) and (d)[i] are Exercise 7A.3 (Hint: use Theorem 8A.1(a,b,d,e) on page E _
162, and Proposition 8C.5(a) and Lemma 8C.6(a) on page 171).
(d)[ii] is Exercise 7A.4 . (f ) is [Asm05, Thm.2, p.97 of 2.10]. 2 E _
Example 7A.2.
(a) If )(r) = sin(5r) 2 sin(3r), then the Fourier sine series of ) is just
sin(5r) 2 sin(3r). In other words, the Fourier coecients 1
n
are all
zero, except that 1
3
= 2 and 1
5
= 1.
(b) Suppose )(r) 1. For all n N,
1
n
=
2
_

0
sin(nr) dr =
2
n
cos(nr)
x=
x=0
=
2
n
_
1 (1)
n
_
=
_
4
n
if n is odd
0 if n is even
.
Thus, the Fourier sine series is:
4
n=1
n odd
1
n
sin(nr) =
4
_
sin(r) +
sin(3r)
3
+
sin(5r)
5
+
_
(7A.3)
Theorem 7A.1(a) says that 1

L2
4
n=1
n odd
1
n
sin(nr). Figure 7A.1 displays
some partial sums of the series (7A.3). The function ) 1 is clearly
continuously dierentiable, so, by Theorem 7A.1(b), the Fourier sine series
converges pointwise to 1 on the interior of the interval [0, ]. However, the
series does not converge to ) at the points 0 or . This is betrayed by the
violent oscillations of the partial sums near these points; this is an example
of the Gibbs phenomenon.
Since the Fourier sine series does not converge at the endpoints 0 and
, we know automatically that it does not converge to ) uniformly on
[0, ]. However, we could have also deduced this fact by noticing that )
does not have homogeneous Dirichlet boundary conditions (because )(0) =
1 = )()), whereas every nite sum of sin(nr)-type functions does have
homogeneous Dirichlet BC. Thus, the series (7A.3) is trying to converge
to ), but it is stuck at the endpoints 0 and . (This is the idea behind
Theorem 7A.1(d)).
(c) If )(r) = cos (:r), then the Fourier sine series of ) is:
4
n=1
n+m odd
n
n
2
:
2
sin(nr).
(Exercise 7A.5 Hint: Use Theorem 6D.3 on page 113). E _
Example 7A.3: sinh(r)
If 0, and )(r) = sinh(r), then its Fourier sine series is given by:
sinh(r)

L2
2 sinh()
n=1
n(1)
n+1
2
+n
2
sin(nr)
To prove this, we must show that, for all n 0,
1
n
=
2
_

0
sinh(r) sin(nr) dr =
2 sinh()
n(1)
n+1
2
+n
2
.
To begin with, let 1 :=
_

0
sinh(r) sin(nr) dr. Then, applying integration
by parts:
1 =
1
n
_
sinh(r) cos(nr)
x=
x=0

_

0
cosh(r) cos(nr) dr
_
=
1
n
_
sinh() (1)
n

n

_
cosh(r) sin(nr)
x=
x=0

_

0
sinh(r) sin(nr) dr
__
=
1
n
_
sinh() (1)
n

n
(0 1)
_
=
sinh() (1)
n
n

2
n
2
1.
Hence: 1 =
sinh() (1)
n
n

2
n
2
1;
thus
_
1 +

2
n
2
_
1 =
sinh() (1)
n
n
;
i.e.
_
n
2
+
2
n
2
_
1 =
sinh() (1)
n+1
n
;
so that 1 =
n sinh() (1)
n+1
n
2
+
2
.
Thus, 1
n
=
2
1 =
2
n sinh() (1)
n+1
n
2
+
2
.
The function sinh is clearly continuously dierentiable, so Theorem 7A.1(b)
implies that the Fourier sine series converges to sinh(r) pointwise on the
open interval (0, ). However, the series does not converge uniformly on [0, ]
(Exercise 7A.6 Hint: What is sinh()?). E _
7A(ii) Cosine series on [0, ]
Recommended: 5C(ii).
If ) L
2
[0, ], we dene the Fourier cosine coecients of ):
0
:= ), 11) =
1
_

0
)(r) dr, and
n
:=
), C
n
)
|C
n
|
2
2
=
2
_

0
)(r) cos(nr) dr, for all n N. (7A.4)
The Fourier cosine series of ) is then the innite summation of functions:
n=0
n
C
n
(r). (7A.5)
Theorem 7A.4. Fourier Cosine Series Convergence on [0, ]
(a) The set C
0
, C
1
, C
2
2
[0, ]. Thus, if )
L
2
[0, ], then the cosine series (7A.5) converges to ) in 1
2
-norm, i.e.
)

L2
n=0
n
C
n
.
Furthermore, the coecient sequence
n
n=0
is the unique sequence of
coecients with this property. In other words, if
t
n
n=1
is some other
sequence of coecients such that )

L2
n=0
t
n
C
n
, then we must have
t
n
=
n
for all n N.
(b) If ) (
1
[0, ], then the cosine series (7A.5) converges pointwise on (0, ).
If ) is piecewise (
1
on [0, ], then the cosine series (7A.5) converges to
) pointwise on each (
1
interval for ). In other words, if ,
1
, . . . , ,
m
is
the set of discontinuity points of ) and/or )
t
, and ,
m
< r < ,
m+1
, then
)(r) = lim
N
N
n=0
n
cos(nr).
(c) If
n=0
[
n
[ < , then the cosine series (7A.5) converges to ) uniformly on
[0, ].
(d) [i] If ) is continuous and piecewise dierentiable on [0, ], and )
t
L
2
[0, ],
then the cosine series (7A.5) converges to ) uniformly on [0, ].
[ii] Conversely, if
n=0
n[
n
[ < , then ) (
1
[0, ] and ) satises
homogeneous Neumann boundary conditions (i.e. )
t
(0) = )
t
() = 0).
1
, and K (,
m
, ,
m+1
1
interval
(f ) Suppose
n
n=0
is a nonnegative sequence decreasing to zero. (That is,
0

1

2
0 and lim
n
n
= 0). If 0 < o < / < , then the
series (7A.5) converges uniformly to ) on [o, /].
Proof. (c) is Exercise 7A.7 (Hint: Use the Weierstrass `-test, Proposition 6E.13 E _
on page 129.)
(a,b,e) and (d)[i] are Exercise 7A.8 (Hint: use Theorem 8A.1(a,b,d,e) on page E _
162, and Proposition 8C.5(b) and Lemma 8C.6(b) on page 171).
(d)[ii] is Exercise 7A.9 (Hint: Use Theorem 7C.10(b) on page 158). E _
(f ) is [Asm05, Thm.2, p.97 of 2.10]. 2
Example 7A.5.
(a) If )(r) = cos (13r), then the Fourier cosine series of ) is just cos(13r). In
other words, the Fourier coecients
n
are all zero, except that
13
= 1.
(b) Suppose )(r) 1. Then ) = C
0
, so the Fourier cosine coecients are:
0
= 1, while
1
=
2
=
3
= . . . 0.
(c) Let )(r) = sin(:r). If : is even, then the Fourier cosine series of ) is:
4
n=1
n odd
n
n
2
:
2
cos(nr).
If :is odd, then the Fourier cosine series of ) is:
2
:
+
4
n=2
n even
n
n
2
:
2
cos(nr).
(Exercise 7A.10 Hint: Use Theorem 6D.3 on page 113). E _
Example 7A.6: cosh(r)
Suppose )(r) = cosh(r). Then the Fourier cosine series of ) is given by:
cosh(r)

L2
sinh()
+
2 sinh()
n=1
(1)
n
cos(nr)
n
2
+ 1
.
To see this, rst note that
0
=
1
_

0
cosh(r) dr =
1
sinh(r)
x=
x=0
=
sinh()
(because sinh(0) = 0).

Next, let 1 :=
_

0
cosh(r) cos(nr) dr. Then applying integration by parts,
we get:
1 =
1
n
_
cosh(r) sin(nr)
x=
x=0
_

0
sinh(r) sin(nr) dr
_
=
1
n
_

0
sinh(r) sin(nr) dr
=
1
n
2
_
sinh(r) cos(nr)
x=
x=0
_

0
cosh(r) cos(nr) dr
_
=
1
n
2
(sinh() cos(n) 1) =
1
n
2
((1)
n
sinh() 1) .
Thus, 1 =
1
n
2
_
(1)
n
sinh() 1
_
. Hence, (n
2
+ 1)1 = (1)
n
sinh(). Hence, 1 =
(1)
n
sinh()
n
2
+ 1
. Thus,
n
=
2
1 =
2
(1)
n
sinh()
n
2
+ 1
.
Remark. (a) Almost any introduction to the theory of partial dierential equa-
tions will contain a discussion of the Fourier convergence theorems. For example,
see [Pow99, 1.3-1.7, pp.59-85], [dZ86, Thm.6.1, p.72] or [Hab87, 3.2, p.91].
(b) Please see Remark 8D.3 on page 174 for further technical remarks about
the (non)convergence of Fourier (co)sine series, in situations where the hypothe-
ses of Theorems 7A.1 and 7A.4 are not satised.
7B Fourier (co)sine series on [0, L]
Prerequisites: 6E, 6F. Recommended: 7A.
Throughout this section, let 1 0 be some positive real number. For all
n N, we dene the functions S
n
: [0, 1] 1 and C
n
: [0, 1] 1 by
S
n
(r) := sin
_
nr
1
_
and C
n
(r) := cos
_
nr
1
_
, for all r [0, 1] (see Figure 6D.1
on page 113). Notice that, if 1 = , then S
n
(r) = sin(nr) and C
n
(r) = cos(nr),
as in 7A. The results in this section exactly parallel those in 7A, except that
we replace with 1 to obtain slightly greater generality. In principle, every
statement in this section is equivalent to the corresponding statement in 7A,
through the change of variables j = r, (it is a useful exercise to reect on this
as you read this section).
7B(i) Sine series on [0, L]
Recommended: 5C(i), 7A(i).
Fix 1 0, and let [0, 1] be an interval of length 1. If ) L
2
[0, 1], we dene
the Fourier sine coecients of ):
1
n
:=
), S
n
)
|S
n
|
2
2
=
2
1
_
L
0
)(r) sin
_
nr
1
_
dr, for all n 1.
The Fourier sine series of ) is then the innite summation of functions:
n=1
1
n
S
n
(r). (7B.1)
A function ) : [0, 1] 1 is continuously dierentiable on [0, 1] if ) is
continuous on [0, 1] (hence, bounded), and )
t
(r) exists for all r (0, 1), and
7B. Fourier (co)sine series on [0, 1] 145
furthermore, the function )
t
: (0, 1) 1 is itself bounded and continuous on
(0, 1). Let (
1
[0, 1] be the space of all continuously dierentiable functions.
We say ) : [0, 1] 1 is piecewise continuously dierentiable (or
piecewise (
1
, or sectionally smooth) if there exist points 0 = ,
0
< ,
1
< ,
2
<
< ,
M+1
= 1 such that ) is bounded and continuously dierentiable on each
of the open intervals (,
m
, ,
m+1
); these are called (
1
intervals for ). In partic-
ular, any continuously dierentiable function on [0, 1] is piecewise continuously
dierentiable (in this case, all of (0, 1) is a (
1
interval).
Theorem 7B.1. Fourier Sine Series Convergence on [0, 1]
Parts (a-f ) of Theorem 7A.1 on page 138 are all still true if you replace
with 1 everywhere.
Proof. Exercise 7B.1 Hint: Use the change-of-variables j =

L
r to pass from E _
j [0, 1] to r [0, ]. 2
Example 7B.2.
(a) If )(r) = sin
_
5
L
r
_
, then the Fourier sine series of ) is just sin
_
5
L
r
_
. In
other words, the Fourier coecients 1
n
are all zero, except that 1
5
= 1.
(b) Suppose )(r) 1. For all n N,
1
n
=
2
1
_
L
0
sin
_
nr
1
_
dr =
2
n
cos
_
nr
1
_
x=L
x=0
=
2
n
_
1 (1)
n
_
=
_
4
n
if n is odd
0 if n is even
.
Thus, the Fourier sine series is given:
4
n=1
n odd
1
n
sin
_
n
1
r
_
. Figure 7A.1
displays some partial sums of this series (in the case 1 = ). The Gibbs
phenomenon is clearly evident just as in Example 7A.2(b) on page 139.
(c) If )(r) = cos
_
m
L
r
_
, then the Fourier sine series of ) is:
4
n=1
n+m odd
n
n
2
:
2
sin
_
n
1
r
_
.
(Exercise 7B.2 Hint: Use Theorem 6D.3 on page 113). E _
(d) If 0, and )(r) = sinh
_
x
L
_
, then its Fourier sine coecients are
computed:
1
n
=
2
1
_
L
0
sinh
_
r
1
_
sin
_
nr
1
_
dr =
2 sinh()
n(1)
n+1
2
+n
2
.
(Exercise 7B.3 ). E _
7B(ii) Cosine series on [0, L]
Recommended: 5C(ii), 7A(ii).
If ) L
2
[0, 1], we dene the Fourier cosine coecients of ):
0
:= ), 11) =
1
1
_
L
0
)(r) dr,
and
n
:=
), C
n
)
|C
n
|
2
2
=
2
1
_
L
0
)(r) cos
_
nr
1
_
dr, for all n 0.
The Fourier cosine series of ) is then the innite summation of functions:
n=0
n
C
n
(r). (7B.2)
Theorem 7B.3. Fourier Cosine Series Convergence on [0, 1]
Parts (a-f ) of Theorem 7A.4 on page 142 are all still true if you replace
with 1 everywhere.
Proof. Exercise 7B.4 Hint: Use the change-of-variables j :=

L
r to pass from E _
r [0, 1] to j [0, ]. 2
Example 7B.4.
(a) If )(r) = cos
_
13
L
r
_
, then the Fourier cosine series of ) is just cos
_
13
L
r
_
.
In other words, the Fourier coecients
n
are all zero, except that
13
= 1.
(b) Suppose )(r) 1. Then ) = C
0
, so the Fourier cosine coecients are:
0
= 1, while
1
=
2
=
3
= . . . 0.
(c) Let )(r) = sin
_
m
L
r
_
. If : is even, then the Fourier cosine series of ) is:
4
n=1
n odd
n
n
2
:
2
cos
_
n
1
r
_
.
If :is odd, then the Fourier cosine series of ) is:
2
:
+
4
n=2
n even
n
n
2
:
2
cos(nr).
(Exercise 7B.5 Hint: Use Theorem 6D.3 on page 113). E _
7C. Computing Fourier (co)sine coecients 147
7C Computing Fourier (co)sine coecients
Prerequisites: 7B.
When computing the Fourier sine coecient 1
n
=
2
1
_
L
0
)(r)sin
_
n
1
r
_
dr,
it is simpler to rst compute the integral
_
L
0
)(r) sin
_
n
1
r
_
dr, and then
multiply the result by
2
1
. Likewise, to compute a Fourier cosine coecients, rst
compute the integral
_
L
0
)(r) cos
_
n
1
r
_
dr, and then multiply the result by
2
1
. In this section, we review some useful techniques to compute these integrals.
7C(i) Integration by parts
Computing Fourier coecients almost always involves integration by parts. Gen-
erally, if you cant compute it with integration by parts, you cant compute it.
When evaluating a Fourier integral by parts, one almost always ends up with
boundary terms of the form cos(n) or sin
_
n
2
_
, etc. The following formulae
are useful in this regard:
sin(n) = 0 for any n Z. (7C.3)
For example, sin() = sin(0) = sin() = sin(2) = sin(3) = 0.
cos(n) = (1)
n
for any n Z. (7C.4)
For example, cos() = 1, cos(0) = 1, cos() = 1, cos(2) = 1, cos(3) =
1, etc.
sin
_
n
2
_
=
_
0 if n is even
(1)
k
if n is odd, and n = 2/ + 1
(7C.5)
For example, sin(0) = 0, sin
_
1
2
_
= 1, sin() = 0, sin
_
3
2
_
= 1, etc.
cos
_
n
2
_
=
_
0 if n is odd
(1)
k
if n is even, and n = 2/
(7C.6)
For example, cos(0) = 1, cos
_
1
2
_
= 0, cos() = 1, cos
_
3
2
_
= 0, etc.
Exercise 7C.1. Verify equations (7C.3), (7C.4), (7C.5), and (7C.6). E _
7C(ii) Polynomials
Theorem 7C.1. Let n N. Then
(a)
_
L
0
sin
_
n
1
r
_
dr =
_
_
21
n
if n is odd;
0 if n is even.
(7C.7)
(b)
_
L
0
cos
_
n
1
r
_
dr =
_
1 if n = 0
0 if n 0.
(7C.8)
For any / 1, 2, 3, . . ., we have the following recurrence relations:
(c)
_
L
0
r
k
sin
_
n
1
r
_
dr =
(1)
n+1
n

1
k+1
+
/
n

1
_
L
0
r
k1
cos
_
n
1
r
_
, (7C.9)
(d)
_
L
0
r
k
cos
_
n
1
r
_
dr =
/
n

1
_
L
0
r
k1
sin
_
n
1
r
_
. (7C.10)
Proof. Exercise 7C.2 Hint: for (c) and (d), use integration by parts. 2 E _
Example 7C.2. In all of the following examples, let 1 = .
(a)
2
_

0
sin(nr) dr =
2
1 (1)
n
n
.
(b)
2
_

0
r sin(nr) dr = (1)
n+1
2
n
.
(c)
2
_

0
r
2
sin(nr) dr = (1)
n+1
2
n
+
4
n
3
_
(1)
n
1
_
.
(d)
2
_

0
r
3
sin(nr) dr = (1)
n
_
12
n
3

2
2
n
_
.
(e)
2
_

0
cos(nr) dr =
_
2 if n = 0
0 if n 0.
(f)
2
_

0
r cos(nr) dr =
2
n
2
_
(1)
n
1
_
, if n 0.
(g)
2
_

0
r
2
cos(nr) dr = (1)
n
4
n
2
, if n 0.
(h)
2
_

0
r
3
cos(nr) dr = (1)
n
6
n
2

12
n
4
_
(1)
n
1
_
, if n 0.
Proof. (b): We will show this in two ways. First, by direct computation:
_

0
r sin(nr) dr =
1
n
_
r cos(nr)
x=
x=0
_

0
cos(nr) dr
_
=
1
n
_
cos(n)
1
n
sin(nr)
x=
x=0
_
=
1
n
(1)
n
=
(1)
n+1
n
.
Thus,
2
_

0
r sin(nr) dr =
2(1)
n+1
n
, as desired.
Next, we verify (b) using Theorem 7C.1. Setting 1 = and / = 1 in (7C.9),
we have:
_

0
r sin(nr) dr =
(1)
n+1
n

1+1
+
1
n

_

0
r
k1
cos (nr) dr
=
(1)
n+1
n
+
1
n
_

0
cos (nr) dr =
(1)
n+1
n
,
because
_

0
cos (nr) dr = 0 by (7C.8). Thus,
2
_

0
r sin(nr) dr =
2(1)
n+1
n
, as desired.
Proof of (c):
_

0
r
2
sin(nr) dr =
1
n
_
r
2
cos(nr)
x=
x=0
2
_

0
rcos(nr) dr
_
=
1
n
_
2
cos(n)
2
n
_
r sin(nr)
x=
x=0
_

0
sin(nr) dr
__
=
1
n
_
2
(1)
n
+
2
n
_
1
n
cos(nr)
x=
x=0
__
=
1
n
_
2
(1)
n
2
n
2
_
(1)
n
1
_
_
=
2
n
3
_
(1)
n
1
_
+
(1)
n+1
2
n
.
The result follows.
Exercise 7C.3 Verify (c) using Theorem 7C.1. E _
(g) We will show this in two ways. First, by direct computation:
_

0
r
2
cos(nr) dr =
1
n
_
r
2
sin(nr)
x=
x=0
2
_

0
r sin(nr) dr
_
=
2
n
_

0
r sin(nr) dr (because sin(nr) = sin(0) = 0)
=
2
n
2
_
r cos(nr)
x=
x=0
_

0
cos(nr) dr
_
=
2
n
2
_
(1)
n
1
n
sin(nr)
x=
x=0
_
=
2 (1)
n
n
2
.
Thus,
2
_

0
r
2
cos(nr) dr =
4 (1)
n
n
2
, as desired.
Next, we verify (g) using Theorem 7C.1. Setting 1 = and / = 2 in (7C.10),
we have:
_

0
r
2
cos(nr) dr =
/
n

1
_
p
0
ir
k1
sin (nr) =
2
n

_

0
rsin (nr) .
(7C.11)
Next, applying (7C.9) with / = 1, we get:
_

0
rsin (nr) =
(1)
n+1
n

+
1
n
_

0
cos (nr) =
(1)
n+1
n
+
1
n
_

0
cos (nr) .
Substituting this into (7C.11), we get
_

0
r
2
cos(nr) dr =
2
n

_
(1)
n+1
n
+
1
n
_

0
cos (nr)
_
. (7C.12)
Were assuming n 0. But then (7C.8) says
_

0
cos (nr) = 0. Thus, we can
simplify (7C.12) to conclude:
2
_

0
r
2
cos(nr) dr =
2

2
n

(1)
n+1
n
=
4(1)
n
n
2
,
as desired. 2
Exercise 7C.4. Verify all of the other parts of Example 7C.2, both using Theorem E _
7C.1, and through direct integration.
To compute the Fourier series of an arbitrary polynomial, we integrate one
term at a time.
Example 7C.3. Let 1 = and let )(r) = r
2
r. Then the Fourier sine
series of ) is:
8
n=1
n odd
1
n
3
sin(nr) =
8
_
sin(r) +
sin(3r)
27
+
sin(5r)
125
+
sin(7r)
343
+
_
/4 /2 3/4
1
Figure 7C.1: Example 7C.4.
To see this, rst, note that, by Example 7C.2(b)
_

0
r sin(nr) dr =
1
n
(1)
n
=
(1)
n+1
n
.
Next, by Example 7C.2(c),
_

0
r
2
sin(nr) dr =
2
n
3
_
(1)
n
1
_
+
(1)
n+1
2
n
.
Thus,
_

0
_
r
2
r
_
sin(nr) dr =
_

0
r
2
sin(nr) dr
_

0
r sin(nr) dr
=
2
n
3
_
(1)
n
1
_
+
(1)
n+1
2
n

(1)
n+1
n
=
2
n
3
_
(1)
n
1
_
.
Thus,
1
n
=
2
_

0
_
r
2
r
_
sin(nr) dr =
4
n
3
_
(1)
n
1
_
=
_
8,n
3
if n is odd;
0 if n is even.
7C(iii) Step functions

Example 7C.4. Let 1 = , and suppose )(r) =
_
1 if

4
r
3
4
0 otherwise
(see
Figure 7C.1). Then the Fourier sine coecients of ) are given:
1
n
=
_
0 if n is even;
2
2(1)
k
n
if n is odd, and n = 4/ 1 for some / N.
0
0.2
0.4
0.6
0.8
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x 0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
= 1 = 3 = 5 = 7
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
0
0.2
0.4
0.6
0.8
1
0.5 1 1.5 2 2.5 3
x
= 9 = 11 = 21 = 201
Figure 7C.2: Partial Fourier sine series for Example 7C.4, for =
0, 1, 2, 3, 4, 5, 10 and 100. Notice the Gibbs phenomenon in the plots for large .
To see this, observe that
_

0
)(r) sin(nr) dr =
_ 3
4
4
sin(nr) dr =
1
n
cos(nr)
x=
3
4
x=
4
=
1
n
_
cos
_
3n
4
_
cos
_
n
4
_
_
=
_
0 if n is even;
2(1)
k+1
n
if n is odd, and n = 4/ 1 for some / N.
(Exercise 7C.5). Thus, the Fourier sine series for ) is: E _
2
_
_
sin(r) +
N
k=1
(1)
k
_
_
sin
_
(4/ 1)r
_
4/ 1
+
sin
_
(4/ + 1)r
_
4/ + 1
_
_
_
_
(Exercise 7C.6). E _
Figure 7C.2 shows some of the partial sums of this series. The series converges
pointwise to )(r) in the interior of the intervals
_
0,

4
_
,
_
4
,
3
4
_
, and
_
3
4
,
.
However, it does not converge to ) at the discontinuity points

4
and
3
4
. In
the plots, this is betrayed by the violent oscillations of the partial sums near
these discontinuity points this is an example of the Gibbs phenomenon.
0
x
1
x
2
x=L
3
0
x
1
x
2
x=L
3
b=b
1
(A)
(B)
a
2
a
3
a
3
a
2
a
1
b
2
b
3
a
1
Figure 7C.3: (A) A step function. (B) A piecewise linear function.
Example 7C.4 is an example of a step function. A function 1 : [0, 1] 1
is a step function (see Figure 7C.3(A)) if there are numbers 0 = r
0
< r
1
<
r
2
< r
3
< . . . < r
M1
< r
M
= 1 and constants o
1
, o
2
, . . . , o
M
1 such that
1(r) =
_
_
o
1
if 0 r r
1
;
o
2
if r
1
< r r
2
;
.
.
.
.
.
.
o
m
if r
m1
< r r
m
;
.
.
.
.
.
.
o
M
if r
M1
< r 1.
(7C.13)
For instance, in Example 7C.4, ` = 3; r
0
= 0, r
1
=

4
, r
2
=
3
4
, and r
3
= ;
o
1
= 0 = o
3
, and o
2
= 1.
To compute the Fourier coecients of a step function, we simply break the
integral into pieces, as in Example 7C.4. The general formula is given by the
following theorem, but it is really not worth memorizing the formula. Instead,
understand the idea.
Theorem 7C.5. Suppose 1 : [0, 1] 1 is a step function like (7C.13).
Then the Fourier coecients of 1 are given:
1
1
_
L
0
1(r) =
1
1
M
m=1
o
m
(r
m
r
m1
) ;
2
1
_
L
0
1(r) cos
_
n
1
r
_
dr =
2
n
M1
m=1
sin
_
n
1
r
m
_
_
o
m+1
o
m
_
;
2
1
_
L
0
1(r) sin
_
n
1
r
_
dr =
2
n
_
o
1
+ (1)
n+1
o
M
_
+
2
n
M1
m=1
cos
_
n
1
r
m
_
_
o
m+1
o
m
_
.
Proof. Exercise 7C.7 Hint: Integrate the function piecewise. 2 E _
Remark. Note that the Fourier series of a step function ) will converge uni-
formly to ) on the interior of each step, but will not converge to ) at any of
the step boundaries, because ) is not continuous at these points.
/2
1
/2
0
/2
(A) (B)
Figure 7C.4: (A) The step function p(r) in Example 7C.6. (B) The tent
function )(r) in Example 7C.7.
Example 7C.6. Suppose 1 = , and p(r) =
_
1 if 0 r <

2
0 if

2
r
(see
Figure 7C.4A). Then the Fourier cosine series of p(r) is:
1
2
+
2
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
In other words,
0
=
1
2
and, for all n 0,
n
=
_
2
(1)
k
2k+1
if n is odd and n = 2/ + 1;
0 if n is even.
Exercise 7C.8 Show this in two ways: rst by direct integration, and then by E _
applying the formula from Theorem 7C.5.
7C(iv) Piecewise linear functions
Example 7C.7: (The Tent Function)
Let X = [0, ] and let )(r) =
_
_
_
r if 0 r

2
;
r if

2
< r .
(see Figure
7C.4B)
The Fourier sine series of ) is:
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sin(nr).
To prove this, we must show that, for all n 0,
1
n
=
2
_

0
)(r) sin(nr) dr =
_
_
4
n
2
(1)
k
if n is odd, n = 2/ + 1;
0 if n is even.
To verify this, we observe that
_

0
)(r) sin(nr) dr =
_
/2
0
rsin(nr) dr +
_

/2
( r) sin(nr) dr.
Exercise 7C.9 Complete the computation of 1
n
. E _
The tent function in Example 7C.7 is piecewise linear. A function 1 :
[0, 1] 1 is piecewise linear (see Figure 7C.3(B) on page 153) if there
are numbers 0 = r
0
< r
1
< r
1
< r
2
< . . . < r
M1
< r
M
= 1 and constants
o
1
, o
2
, . . . , o
M
1 and / 1 such that
1(r) =
_
_
o
1
(r 1) +/
1
if 0 r r
1
;
o
2
(r r
1
) +/
2
if r
1
< r r
2
;
.
.
.
.
.
.
o
m
(r r
m
) +/
m+1
if r
m
< r r
m+1
;
.
.
.
.
.
.
o
M
(r r
M1
) +/
M
if r
M1
< r 1.
(7C.14)
where /
1
= /, and, for all : 1, /
m
= o
m
(r
m
r
m1
) +/
m1
.
For instance, in Example 7C.7, ` = 2, r
1
=

2
and r
2
= ; o
1
= 1 and
o
2
= 1.
To compute the Fourier coecients of a piecewise linear function, we can
break the integral into pieces, as in Example 7C.7. The general formula is
given by the following theorem, but it is really not worth memorizing the formula.
Instead, understand the idea.
Theorem 7C.8. Suppose 1 : [0, 1] 1 is a piecewise-linear function like
(7C.14). Then the Fourier coecients of 1 are given:
1
1
_
L
0
1(r) =
1
1
M
m=1
o
m
2
(r
m
r
m1
)
2
+ /
m
(r
m
r
m1
) .
2
1
_
L
0
1(r) cos
_
n
1
r
_
dr =
21
(n)
2
M
m=1
cos
_
n
1
r
m
_
_
o
m
o
m+1
_
2
1
_
L
0
1(r) sin
_
n
1
r
_
dr =
21
(n)
2
M1
m=1
sin
_
n
1
r
m
_
_
o
m
o
m+1
_
(where we dene o
M+1
:= o
1
for convenience).
Proof. Exercise 7C.10 Hint: invoke Theorem 7C.5 and integration by parts. 2 E _
Note that the summands in this theorem read o
m
o
m+1
, not the other
way around.
Example 7C.9: (Cosine series of the tent function)
Let Let X = [0, ] and let )(r) =
_
_
_
r if 0 r

2
;
r if

2
< r .
as in Example
7C.7. The Fourier cosine series of ) is:
4

8
n=1
n=4j+2,
for some j
1
n
2
cos(nr).
In other words,
)(r) =

4

8
_
cos(2r)
4
+
cos(6r)
36
+
cos(10r)
100
+
cos(14r)
196
+
cos(18r)
324
+ . . .
_
To see this, rst observe that
0
=
1
_

0
)(r) dr =
1
_
_
/2
0
r dr +
_

/2
( r) dr
_
=
1
_
_
r
2
2
_
/2
0
+

2
2

_
r
2
2
_
/2
_
=
1
2
8
+

2
2

_
2
2

2
8
__
=

2
4
=

4
.
Now lets compute
n
for n 0.
First,
_
/2
0
rcos(nr) dr =
1
n
_
rsin(nr)
/2
0
_
/2
0
sin(nr) dr
_
=
1
n
_
2
sin
_
n
2
_
+
1
n
cos(nr)
/2
0
_
=

2n
sin
_
n
2
_
+
1
n
2
cos
_
n
2
_

1
n
2
.
Next,
_

/2
rcos(nr) dr =
1
n
_
rsin(nr)
/2
_

/2
sin(nr) dr
_
=
1
n
_
2
sin
_
n
2
_
+
1
n
cos(nr)
/2
_
=

2n
sin
_
n
2
_
+
(1)
n
n
2

1
n
2
cos
_
n
2
_
.
Finally,
_

/2
cos(nr) dr =

n
sin(nr)
/2
=

n
sin
_
n
2
_
.
Putting it all together, we have:
_

0
)(r) cos(nr) dr =
_
/2
0
rcos(nr) dr +
_

/2
cos(nr) dr
_

/2
rcos(nr) dr
=

2n
sin
_
n
2
_
+
1
n
2
cos
_
n
2
_

1
n
2

n
sin
_
n
2
_
+

2n
sin
_
n
2
_

(1)
n
n
2
+
1
n
2
cos
_
n
2
_
=
2
n
2
cos
_
n
2
_

1 + (1)
n
n
2
.
Now,
cos
_
n
2
_
=
_
(1)
k
if n is even and n = 2/;
0 if n is odd.
while 1 + (1)
n
=
_
2 if n is even;
0 if n is odd.
Thus,
2 cos
_
n
2
_

_
1 + (1)
n
_
=
_
4 if n is even, n = 2/ and / = 2, + 1 for some ,;
0 otherwise.
=
_
4 if n = 4, + 2 for some ,;
0 otherwise.
(for example, n = 2, 6, 10, 14, 18, . . .). Thus
n
=
2
_

0
)(r) cos(nr) dr
=
_
_
8
n
2
if n = 4, + 2 for some ,;
0 otherwise.
7C(v) Dierentiating Fourier (co)sine series

Prerequisites: 7B, 0F.
Suppose )(r) = 3 sin(r) 5 sin(2r) + 7 sin(3r). Then )
t
(r) = 3 cos(r)
10 cos(2r) +21 cos(3r). Likewise, if )(r) = 3 +2 cos(r) 6 cos(2r) +11 cos(3r),
then )
t
(r) = 2 sin(r)+12 sin(2r)33 sin(3r). This illustrates a general pattern.
Theorem 7C.10. Suppose ) (
[0, 1]
(a) Suppose ) has Fourier sine series
n=1
1
n
S
n
(r). If
n=1
n[1
n
[ < , then
)
t
has Fourier cosine series: )
t
(r) =

1
n=1
n1
n
C
n
(r), and this series
converges uniformly.
(b) Suppose ) has Fourier cosine series
n=0
n
C
n
(r). If
n=1
n[
n
[ < ,
then )
t
has Fourier sine series: )
t
(r) =

1
n=1
n
n
S
n
(r), and this
series converges uniformly.
Proof. Exercise 7C.11 Hint: Apply Proposition 0F.1 on page 565. 2 E _
Consequence: If )(r) = cos
_
nr
1
_
+ 1sin
_
nr
1
_
for some , 1 1,
then )
tt
(r) =
_
n
1
_
2
)(r). In other words, ) is an eigenfunction
1
for the
dierentation operator
2
x
, with eigenvalue =
_
n
L
_
2
. More generally, for any
/ N, we have
2k
x
) =
k
).
In all of these problems, the domain is X = [0, ].
1
See 4B(iv) on page 63
1. Let 0 be a constant. Compute the Fourier sine series of )(r) =
exp( r). At which points does the series converge pointwise? Why?
Does the series converge uniformly? Why or why not?
2. Compute the Fourier cosine series of )(r) = sinh(r). At which points does
the series converge pointwise? Why? Does the series converge uniformly?
Why or why not?
3. Let 0 be a constant. Compute the Fourier sine series of )(r) =
cosh(r). At which points does the series converge pointwise? Why?
Does the series converge uniformly? Why or why not?
4. Compute the Fourier cosine series of )(r) = r. At which points does the
series converge pointwise? Why? Does the series converge uniformly? Why
or why not?
5. Let p(r) =
_
1 if 0 r <

2
0 if

2
r
(Fig. 7C.4A on p. 154)
(a) Compute the Fourier cosine series of p(r). At which points does the
series converge pointwise? Why? Does the series converge uniformly?
Why or why not?
(b) Compute the Fourier sine series of p(r). At which points does the
series converge pointwise? Why? Does the series converge uniformly?
Why or why not?
6. Compute the Fourier cosine series of p(r) =
_
3 if 0 r <

2
1 if

2
r
At which points does the series converge pointwise? Why? Does the series
converge uniformly? Why or why not?
7. Compute the Fourier sine series of )(r) =
_
_
_
r if 0 r

2
r if

2
< r .
(Fig. 7C.4B on p.154) At which points does the series converge pointwise?
Why? Does the series converge uniformly? Why or why not?
Hint: Note that
_

0
)(r) sin(nr) dr =
_
/2
0
rsin(nr) dr +
_

/2
( r) sin(nr) dr.
8. Let ) : [0, ] 1 be dened: )(r) =
_
r if 0 r

2
0 if

2
< r
.
Compute the Fourier sine series for )(r). At which points does the series
converge pointwise? Why? Does the series converge uniformly? Why or
why not?
161
Chapter 8
Real Fourier series and
complex Fourier series
Ordinary language is totally unsuited for expressing what physics really asserts, since the
words of everyday life are not suciently abstract. Only mathematics and mathematical logic
can say as little as the physicist means to say. Bertrand Russell
8A Real Fourier series on [, ]
Prerequisites: 6E, 6F. Recommended: 7A, 5C(iv).
Throughout this section, for all n N, we dene the functions S
n
: [, ]
1 and C
n
: [, ] 1 by S
n
n
(r) := cos(nr), for all
r [, ] (see Figure 6D.1 on page 113). If ) : [, ] 1 is any function
with |)|
2
< , we dene the (real) Fourier coecients:
0
:= ), C
0
) = ), 11) =
1
2
_

)(r) dr,
n
:=
), C
n
)
|C
n
|
2
2
=
1
)(r) cos (nr) dr,

and 1
n
:=
), S
n
)
|S
n
|
2
2
=
1
)(r) sin (nr) dr, for all n 1.

The (real) Fourier series of ) is then the innite summation of functions:
0
+
n=1
n
C
n
(r) + +
n=1
1
n
S
n
(r). (8A.1)
We dene continuously dierentiable and piecewise continuously dierentiable
functions on [, ] in a manner exactly analogous to the denitions on [0, ]
162 DRAFT Chapter 8. Real Fourier series and complex Fourier series
(page 138). Let (
1
[, ] be the set of all continuously dierentiable functions
) : [, ] 1.
Exercise 8A.1. (a) Show that any continuously dierentiable function has nite E _
1
2
-norm. In other words, (
1
[, ] L
2
[, ].
(b) Show that any piecewise (
1
function on [, ] is in L
2
[, ].
Theorem 8A.1. Fourier Convergence on [, ]
(a) The set 11, S
1
, C
1
, S
2
, C
2
2
[, ]. Thus,
if ) L
2
[, ], then the Fourier series (8A.1) converges to ) in 1
2
-norm.
Furthermore, the coecient sequences
n
n=0
and 1
n
n=1
are the unique
sequences of coecients with this property. In other words, if
t
n
n=0
and
1
t
n
n=1
are two other sequences of coecients such that )

L2
n=0
t
n
C
n
+
n=1
1
t
n
S
n
, then we must have
t
n
=
n
and 1
t
n
= 1
n
for all n N.
(b) If ) (
1
[, ] then the Fourier series (8A.1) converges pointwise on
(, ).
1
, then the real Fourier series (8A.1)
converges to ) pointwise on each (
1
interval for ). In other words, if
,
1
, . . . , ,
m
is the set of discontinuity points of ) and/or )
t
in [, ], and
,
m
< r < ,
m+1
, then )(r) =
0
+ lim
N
N
n=1
_
n
cos (nr) +1
n
sin (nr)
_
.
(c) If
n=0
[
n
[ +
n=1
[1
n
[ < , then the series (8A.1) converges to ) uniformly
on [, ].
(d) Suppose ) : [, ] 1 is continuous and piecewise dierentiable, )
t

L
2
[, ], and )() = )(). Then the series (8A.1) converges to )
uniformly on [, ].
1
, and K (,
m
, ,
m+1
1
interval
Proof. For a proof of (a) see 10D on page 207. For a proof of (b), see 10B
on page 197. (Alternately, (b) follows immediately from (e).) For a proof of
(d) see 10C on page 204.
(c) is Exercise 8A.2 (Hint: Use the Weierstrass `-test, Proposition 6E.13 on E _
8B. Computing real Fourier coecients 163
page 129.)
(e) is Exercise 8A.3 (Hint: use Theorem 8D.1(e) and Proposition 8D.2 on page E _
173). 2
There is nothing special about the interval [, ]. Real Fourier series can
be dened for functions on an interval [1, 1] for any 1 0. We chose 1 =
because it makes the computations simpler. If 1 ,= , then we can dene a
Fourier series analogous to (8A.1) using the functions S
n
(r) = sin
_
nr
1
_
and
C
n
(r) = cos
_
nr
1
_
.
Exercise 8A.4. Let 1 0, and let ) : [1, 1] 1. Generalize all parts of E _
Theorem 8A.1 to characterize the convergence of the real Fourier series of ).
Remark. Please see Remark 8D.3 on page 174 for further technical remarks
about the (non)convergence of real Fourier series, in situations where the hy-
potheses of Theorem 8A.1 are not satised.
8B Computing real Fourier coecients
Prerequisites: 8A. Recommended: 7C.
When computing the real Fourier coecient
n
=
1
)(r) cos (nr) dr

(or 1
n
=
1
)(r) sin (nr) dr), it is simpler to rst compute the integral

_

)(r) cos (nr) dr (or

_

)(r) sin (nr) dr), and then multiply the result by

1
. In this section, we review some useful techniques to compute this integral.

8B(i) Polynomials
Recommended: 7C(ii).
Theorem 8B.1.
_

sin(nr) dr = 0 =
_

cos(nr) dr.
For any / 1, 2, 3, . . ., we have the following recurrence relations:
If / is even, then:
_

r
k
sin(nr) dr = 0 and
_

r
k
cos(nr) dr =
/
n
_

r
k1
sin(nr) dr.
If / 0 is odd, then:
_

r
k
sin(nr) dr =
2(1)
n+1
k
n
+
/
n
_

r
k1
cos(nr) dr
and
_

r
k
cos(nr) dr = 0.
Proof. Exercise 8B.1 Hint: use integration by parts. 2 E _
Example 8B.2.
(a) j(r) = r. Since / = 1 is odd, we have
1
r cos(nr) dr = 0,
and
1
r sin(nr) dr =
2(1)
n+1
0
n
+
1
n
_

cos(nr) dr
()
2(1)
n+1
n
.
where equality () follows from case / = 0 in Theorem 8B.1.
(b) j(r) = r
2
. Since / = 2 is even, we have, for all n,
1
r
2
sin(nr) dr = 0,
1
r
2
cos(nr) dr =
2
n
_

r
1
sin(nr) dr
()
2
n
_
2(1)
n+1
n
_
=
4(1)
n
n
2
.
where equality () follows from the previous example.
8B(ii) Step functions
Recommended: 7C(iii).
A function 1 : [, ] 1 is a step function (see Figure 8B.1(A)) if there
are numbers = r
0
< r
1
< r
2
< r
3
< < r
M1
< r
M
= and constants
o
1
, o
2
, . . . , o
M
1 such that
1(r) =
_
_
o
1
if r r
1
;
o
2
if r
1
< r r
2
;
.
.
.
.
.
.
o
m
if r
m1
< r r
m
;
.
.
.
.
.
.
o
M
if r
M1
< r .
(8B.1)
x=
0
x
1
x
2
x=
3
x=
0
x
1
x
2
x=
3
b=b
1
(A)
(B)
a
2
a
3
a
3
a
2
a
1
b
2
b
3
a
1
Figure 8B.1: (A) A step function. (B) A piecewise linear function.

To compute the Fourier coecients of a step function, we break the integral
into pieces. The general formula is given by the following theorem, but it is
really not worth memorizing the formula. Instead, understand the idea.
Theorem 8B.3. Suppose 1 : [, ] 1 is a step function like (8B.1).
Then the Fourier coecients of 1 are given:
1
2
_

1(r) dr =
1
2
M
m=1
o
m
(r
m
r
m1
) ;
1
1(r) cos(nr) dr =
1
n
M1
m=1
sin(n r
m
)
_
o
m+1
o
m
_
;
1
1(r) sin(nr) dr =
(1)
n
n
_
o
1
o
M
_
+
1
n
M1
m=1
cos(n r
m
)
_
o
m+1
o
m
_
.
Proof. Exercise 8B.2 Hint: Integrate the function piecewise. Use the fact that E _
_
xm
xm1
)(r) sin(nr) =
o
m
n
_
cos(n r
m1
) cos(n r
m
)
_
and
_
xm
xm1
)(r) cos(nr) =
o
m
n
_
cos(n r
m
) cos(n r
m1
)
_
.
2
Remark. Note that the Fourier series of a step function ) will converge uni-
formly to ) on the interior of each step, but will not converge to ) at any of
the step boundaries, because ) is not continuous at these points.
-3 /2
/2
2
5
Figure 8B.2: The step function in Example 8B.4.
Example 8B.4. Suppose )(r) =
_
_
_
3 if r <

2
;
5 if

2
r <

2
;
2 if

2
r .
(see Figure
8B.2).
In the notation of Theorem 8B.3, we have ` = 3, and
r
0
= ; r
1
=

2
; r
2
=

2
; r
3
= ;
o
1
= 3; o
2
= 5; o
3
= 2.
Thus,
n
=
1
n
_
8 sin
_
n

2
_
3 sin
_
n

2
_
_
=
_
0 if n is even;
(1)
k
11
n
if n = 2/ + 1 is odd.
and 1
n
=
1
n
_
8 cos
_
n

2
_
3 cos
_
n

2
_
5 cos (n )
_
=
_
_
5
n
if n is odd;
5
n
_
(1)
k
1
_
if n = 2/ is even.
=
_
_
5
n
if n is odd;
0 if n is divisible by 4;
10
n
if n is even but not divisible by 4.
.
8B(iii) Piecewise linear functions
Recommended: 7C(iv).
A continuous function 1 : [, ] 1 is piecewise linear (see Figure
8B.1(B)) if there are numbers = r
0
< r
1
< r
1
< r
2
< . . . < r
M1
< r
M
=
and constants o
1
, o
2
, . . . , o
M
1 and / 1 such that
1(r) =
_
_
o
1
(r ) +/
1
if < r < r
1
;
o
2
(r r
1
) +/
2
if r
1
< r < r
2
;
.
.
.
.
.
.
o
m
(r r
m
) +/
m+1
if r
m
< r < r
m+1
;
.
.
.
.
.
.
o
M
(r r
M1
) +/
M
if r
M1
< r < .
(8B.2)
where /
1
= /, and, for all : 1, /
m
= o
m
(r
m
r
m1
) +/
m1
.
Example 8B.5. If )(r) = [r[, then ) is piecewise linear, with: r
0
= , r
1
= 0,
and r
2
= ; o
1
= 1 and o
2
= 1; /
1
= , and /
2
= 0.
To compute the Fourier coecients of a piecewise linear function, we break
the integral into pieces. The general formula is given by the following theorem,
but it is really not worth memorizing the formula. Instead, understand the idea.
Theorem 8B.6. Suppose 1 : [, ] 1 is a piecewise-linear function like
(8B.2). Then the Fourier coecients of 1 are given:
1
2
_

1(r) dr =
1
2
M
m=1
o
m
2
(r
m
r
m1
)
2
+ /
m
(r
m
r
m1
) ;
1
1(r) cos(nr) dr =
1
n
2
M
m=1
cos(nr
m
)
_
o
m
o
m+1
_
;
1
1(r) sin(nr) dr =
1
n
2
M1
m=1
sin(nr
m
)
_
o
m
o
m+1
_
.
(Here, we dene o
M+1
:= o
1
for convenience.)
Proof. Exercise 8B.3 Hint: invoke Theorem 8B.3 and integration by parts. 2 E _
Note that the summands in this theorem read o
m
o
m+1
, not the other
way around.
Example 8B.7. Recall )(r) = [r[, from Example 8B.5. Applying Theorem
8B.6, we have
0
=
1
2
_
1
2
(0 +)
2
+ (0 +) +
1
2
( 0)
2
+ 0 ( 0)
_
=

2
.
n
=

n
2
[(1 1) cos (n0) (1 + 1) cos (n)]
=
1
n
2
[2 + 2(1)
n
] =
2
n
2
[1 (1)
n
] ,
while 1
n
= 0 for all n N, because ) is an even function.
8B(iv) Dierentiating real Fourier series
Prerequisites: 8A, 0F.
Suppose )(r) = 3 + 2 cos(r) 6 cos(2r) + 11 cos(3r) + 3 sin(r) 5 sin(2r) +
7 sin(3r). Then )
t
(r) = 2 sin(r)+12 sin(2r)33 sin(3r)+3 cos(r)10 cos(2r)+
21 cos(3r). This illustrates a general pattern.
Theorem 8B.8. Let ) (
[, ], and suppose ) has Fourier series
n=0
n
C
n
+
n=1
1
n
S
n
. If
n=1
n[
n
[ < and
n=1
n[1
n
[ < , then )
t
has Fourier Series:
n=1
n
_
1
n
C
n

n
S
n
_
.
Proof. Exercise 8B.4 Hint: Apply Proposition 0F.1 on page 565. 2 E _
Consequence: If )(r) = cos (nr) + 1sin (nr) for some , 1 1, then
)
tt
(r) = n
2
)(r). In other words, ) is an eigenfunction for the dierentation
operator
2
x
, with eigenvalue n
2
. Hence, for any / N, we have
2k
x
) =
(n)
k
).
8C Relation between (co)sine series and real series
Prerequisites: 7A, 8A.
We have seen in 8A how the collection C
n
n=0
S
n
n=1
forms an orthog-
onal basis for L
2
[, ]. However, if we conne our attention to half this interval
that is, to L
2
[0, ] then the results of 7A imply that we only need half as
many basis elements; either the collection C
n
n=0
or the collection S
n
n=1
will
suce. Why is this? And what is the relationship between the Fourier (co)sine
series of 7A and the Fourier series of 8A?
A function ) : [1, 1] 1 is even if )(r) = )(r) for all r [0, 1]. For
example, the following functions are even:
)(r) = 1.
)(r) = [r[.
)(r) = r
2
.
)(r) = r
k
for any even / N.
)(r) = cos(r).
8C. Relation between (co)sine series and real series 169
A function ) : [1, 1] 1 is odd if )(r) = )(r) for all r [0, 1]. For
example, the following functions are odd:
)(r) = r.
)(r) = r
3
.
)(r) = r
k
for any odd / N.
)(r) = sin(r).
Every function can be split into an even part and an odd part.
Proposition 8C.1. (a) For any ) : [1, 1] 1, there is a unique even
function

) and a unique odd function

) such that ) =

) +

). To be specic:
)(r) =
)(r) +)(r)
2
and

)(r) =
)(r) )(r)
2
(b) If ) is even, then ) =

), and

) = 0.
(c) If ) is odd, then

) = 0, and ) =

).
The equation ) =

) +

) is called the even-odd decomposition of ). Next,
we dene the vector spaces:
L
2
even
[, ] := all even elements in L
2
[, ].
and L
2
odd
[, ] := all odd elements in L
2
[, ].
Proposition 8C.1 implies that any ) L
2
[, ] can be written (in a unique way)
as ) =

) +

) for some

) L
2
even
[, ] and

) L
2
odd
[, ]. (This is sometimes
indicated by writing: L
2
[, ] = L
2
even
[, ] L
2
odd
[, ].)
Lemma 8C.2. Let n N.
(a) The function C
n
(r) = cos(nr) is even.
(b) The function S
n
(r) = sin(nr) is odd.
Let ) : [, ] 1 be any function.
(c) If )(r) =
n=0
n
C
n
(r), then ) is even.
(d) If )(r) =
n=1
1
n
S
n
(r), then ) is odd.
In other words, cosine series are even, and sine series are odd. The converse
is also true. To be precise:
Proposition 8C.3. Let ) : [, ] 1 be any function, and suppose ) has
real Fourier series )(r) =
0
+
n=1
n
C
n
(r) +
n=1
1
n
S
n
(r). Then:
(a) If ) is odd, then
n
= 0 for every n N.
(b) If ) is even, then 1
n
= 0 for every n N.
From this, it follows immediately:
Proposition 8C.4.
(a) The set C
0
, C
1
, C
2
2
even
[, ] (where C
0
=
11).
(b) The set S
1
, S
2
, S
3
2
odd
[, ].
(c) Suppose ) has even-odd decomposition ) =

) +

), and ) has real Fourier
series )(r) =
0
+
n=1
n
C
n
(r) +
n=1
1
n
S
n
(r). Then

)(r) =
n=0
n
C
n
(r)
and

)(r) =
n=1
1
n
S
n
(r).
If ) : [0, ] 1, then we can extend ) to a function on [, ] in two ways:
The even extension of ) is dened: )
even
(r) = ) ([r[) for all r [, ].
The odd extension of ) is dened: )
odd
(r) =
_
_
_
)(r) if r 0
0 if r = 0
)(r) if r < 0
Exercise 8C.5. (a) Show that )
even
is even and )
odd
is odd. E _
(b) For all r [0, ], show that )
even
(r) = )(r) = )
odd
(r).
8C. Relation between (co)sine series and real series 171
Proposition 8C.5. Let ) : [0, ] 1 have even extension )
even
: [, ]
1 and odd extension )
odd
: [, ] 1.
(a) The Fourier sine series for ) is the same as the real Fourier series for
)
odd
. In other words, the nth Fourier sine coecient is given: 1
n
=
1
)
odd
(r)S
n
(r) dr.
(b) The Fourier cosine series for ) is the same as the real Fourier series for
)
even
. In other words, the nth Fourier cosine coecient is given:
n
=
1
)
even
(r)C
n
(r) dr.
Let ) (
1
[0, ]. Recall that Theorem 7A.1(d) (on page 138) says that the
Fourier sine series of ) converges to ) uniformly on [0, ] if and only if ) satises
homogeneous Dirichlet boundary conditions on [0, ] (i.e. )(0) = )() = 0). On
the other hand, Theorem 7A.4(d) (on page 142) says that the Fourier cosine series
of ) always converges to ) uniformly on [0, ] if ) (
1
[0, ]; furthermore, if the
formal derivative of this cosine series converges to )
t
uniformly on [0, ], then
) satises homogeneous Neumann boundary conditions on [0, ] (i.e. )
t
(0) =
)
t
() = 0). Meanwhile, if 1 (
1
[, ], then Theorem 8A.1(d) (on page 162)
says that the (real) Fourier series of 1 converges to 1 uniformly on [, ] if 1
satises periodic boundary conditions on [, ] (i.e. 1() = 1()). The next
result explains the logical relationship between these three statements.
Lemma 8C.6. Let ) : [0, ] 1 have even extension )
even
: [, ] 1
and odd extension )
odd
: [, ] 1. Suppose ) is right-continuous at 0 and
left-continuous at .
(a) )
odd
is continuous at zero and satises periodic boundary conditions on
[, ], if and only if ) satises homogeneous Dirichlet boundary conditions
on [0, ].
(b) )
even
is always continuous at zero and always satises periodic boundary
conditions on [, ].
However, the derivative )
t
even
is continuous at zero and satises periodic
boundary conditions on [, ] if and only if ) satises homogeneous Neu-
mann boundary conditions on [0, ].
8D Complex Fourier series
Prerequisites: 6C(i), 6E, 6F, 0C. Recommended: 8A.
Let ), p : X C be complex-valued functions. Recall from 6C(i) that we
dene their inner product:
), p) :=
1
`
_
X
)(x) p(x) dx,
where ` is the length/area/volume of domain X. Once again,
|)|
2
:= ), ))
1/2
=
_
1
`
_
X
)(x))(x) dx
_
1/2
=
_
1
`
_
X
[)(x)[
2
dx
_
1/2
.
The concepts of orthogonality, 1
2
distance, and 1
2
convergence are exactly the
same as before. Let L
2
([1, 1]; C) be the set of all complex-valued functions
) : [1, 1] C with |)|
2
< . For all n Z, let
E
n
(r) := exp
_
inr
1
_
.
(thus, E
0
= 11 is the constant unit function). For all n 0, notice that Eulers
Formula (see page 551) implies:
E
n
(r) = C
n
(r) + i S
n
(r)
and E
n
(r) = C
n
(r) i S
n
(r)
(8D.1)
Also, note that E
n
, E
m
) = 0 if n ,= :, and |E
n
|
2
= 1 (Exercise 8D.1 ), so E _
these functions form an orthonormal set.
If ) : [1, 1] C is any function with |)|
2
< , then we dene the
(complex) Fourier coecients of ):
)
n
:= ), E
n
) =
1
21
_
L
L
)(r) exp
_
inr
1
_
dr. (8D.2)
The (complex) Fourier Series of ) is then the innite summation of functions:
n=
)
n
E
n
. (8D.3)
(note that in this sum, n ranges from to ).
8D. Complex Fourier series 173
Theorem 8D.1. Complex Fourier Convergence
(a) The set . . . , E
1
, E
0
, E
1
, . . . is an orthonormal basis for L
2
([1, 1]; C).
Thus, if ) L
2
([1, 1]; C), then the complex Fourier series (8D.3) con-
verges to ) in 1
2
-norm.
Furthermore,
)
n
n=
is the unique sequence of coecients with this
property.
(b) If ) is continuously dierentiable
1
on [, ], then the Fourier series (8D.3)
converges pointwise on (, ).
1
, then the complex Fourier series (8D.3)
converges to ) pointwise on each (
1
interval for ). In other words, if
,
1
, . . . , ,
m
is the set of discontinuity points of ) and/or )
t
in [1, 1],
and ,
m
< r < ,
m+1
, then )(r) = lim
N
N
n=N
)
n
E
n
(r).
(c) If
n=
)
n
< , then the series (8D.3) converges to ) uniformly on

[, ].
(d) Suppose ) : [, ] 1 is continuous and piecewise dierentiable, )
t

L
2
[, ], and )() = )(). Then the series (8D.3) converges to )
uniformly on [, ].
1
, and K (,
m
, ,
m+1
1
interval
of ), then the series (8D.3) converges uniformly to ) on K.
Proof. For (a) is Exercise 8D.2 (Hint: Use Theorem 8A.1(a) on page 162 and E _
Proposition 8D.2 below).
For a direct proof of (a), see [Kat76, I.5.5, p.29-30].
(b) is Exercise 8D.3 (Hint: (i) use Theorem 8A.1(b) on page 162 and Proposition E _
8D.2 below. (ii) For a second proof, derive (b) from from (e).)
(c) is Exercise 8D.4 (Hint: Use the Weierstrass `-test, Proposition 6E.13 on E _
page 129.)
(d) is Exercise 8D.5 (Hint: use Theorem 8A.1(d) on page 162 and Proposition E _
8D.2 below).
For a direct proof of (d) see [WZ77, Theorem 12.20, p.219].
For (e) see [Fol84, Theorem 8.43, p.256] or [Kat76, Corollary on p.53 of II.2.2].
2
1
This means that f(x) = fr(x) +ifi(x), where fr : [L, L] 1 and fi : [L, L] 1 are
both continuously dierentiable, real-valued functions.
Proposition 8D.2. Relation between Real and Complex Fourier Series
Let ) : [, ] 1 be a real-valued function, and let
n
n=0
and 1
n
n=1
be its real Fourier coecients, as dened on page 161. We can also regard ) as
a complex-valued function; let
)
n
n=
be the complex Fourier coecients of
), as dened by equation (8D.2) on page 172. Let n N
+
. Then
(a)

)
n
=
1
2
(
n
i1
n
), and

)
n
=

)
n
=
1
2
(
n
+i1
n
).
(b) Thus,
n
=

)
n
+

)
n
, and 1
n
= i(
)
n

)
n
).
(c)

)
0
=
0
.
Proof. Exercise 8D.6 Hint: use the equations (8D.1). 2 E _
Exercise 8D.7. Show that Theorem 8D.1(a) and Theorem 8A.1(a) are equivalent, E _
using the Proposition 8D.2.
Remark 8D.3: Further remarks on Fourier convergence
(a) In Theorems 7A.1(b), 7A.4(b), 8A.1(b) and 8D.1(b), if r is a discontinuity
point of ), then the Fourier (co)sine series converges to the average of the
left-hand and right-hand limits of ) at r, namely:
)(r) +)(r+)
2
, where )(r) := lim
y,x
)(j) and )(r+) := lim
yx
)(j).
(b) If the hypothesis of Theorems 7A.1(c), 7A.4(c), 8A.1(c) or 8D.1(c) is sat-
ised, then we say that the Fourier series (real, complex, sine or cosine)
converges absolutely. (In fact, Theorems 7A.1(d)[i], 7A.4(d)[i], 8A.1(d)
or 8D.1(d) can be strengthened to yield absolute convergence). Absolute
convergence is stronger than uniform convergence, and functions with ab-
solutely convergent Fourier series form a special class; see [Kat76, I.6,
p.31-33] for more information.
(c) In Theorems 7A.1(e), 7A.4(e), 8A.1(e) and 8D.1(e), we dont quite need )
to be dierentiable to guarantee uniform convergence of the Fourier (co)sine
series. Let 0 be a constant; we say that ) is -Holder continuous on
[, ] if there is some ` < such that,
For all r, j [0, ],
[)(r) )(j)[
[r j[
`.
Bernsteins Theorem says: If ) is -Holder continuous for some
1
2
, then
the Fourier series (real, complex, sine or cosine) of ) will converge uniformly
(indeed, absolutely) to ); see [Fol84, Theorem 8.39] or [Kat76, Thm 6.3 on
p.32]. (If ) was dierentiable, then ) would be -Holder continuous with
= 1, so Bernsteins Theorem immediately implies Theorems 7A.1(e) and
7A.4(e).)
(d) The total variation of ) is dened
var()) := sup
NN
sup
x
0
<<x
N
n=1
)(r
n
) )(r
n1
)
()
_

)
t
(r)
dr.
Here, the supremum is taken over all nite increasing sequences
r
0
< r
1
< < r
N
(for any N), and equality () is true if
and only if ) is continuously dierentiable. Zygmunds Theorem says: if
var()) < (i.e. ) has bounded variation) and ) is -Holder continuous for
some 0, then the Fourier series of ) will converge uniformly (indeed,
absolutely) to ) on [, ]; see [Kat76, Thm 6.4 on p.33].
(e) However, merely being continuous is not sucient for uniform Fourier con-
vergence, or even pointwise convergence. There exists a continuous func-
tion ) : [0, ] 1 whose Fourier series does not converge pointwise on
(0, ) i.e. the series diverges at some points in (0, ); see [WZ77, Theo-
rem 12.35, p.227] or [Kat76, Theorem 2.1, p.51]. Thus, Theorems 7A.1(b),
7A.4(b), 8A.1(b) and 8D.1(b) are false if we replace dierentiable with
continuous.
(f ) Fix j [1, ). For any ) : [, ] C, we dene the 1
p
-norm of ):
|)|
p
=
__

[)(r)[
p
dr
_
1/p
.
(Thus, if j = 2, we get the familiar 1
2
-norm |)|
2
). Let L
p
[, ] be
the set of all integrable functions ) : [, ] C such that |)|
p
< .
Theorem 8D.1(a) say that, if ) L
2
[, ], then the complex Fourier series
of ) converges to ) in 1
2
-norm. The Fourier series of ) also converges in
1
p
-norm for any other j (1, ). That is, for any j (1, ) and any
) L
p
[, ], we have
lim
N
_
_
_
_
_
)
N
n=N
)
n
E
n
_
_
_
_
_
p
= 0.
See [Kat76, Theorem 1.5, p.50]. If ) L
p
[, ] is purely real-valued, then
the same statement holds for the real Fourier series:
lim
N
_
_
_
_
_
)
_
0
+
N
n=1
n
C
n
+
N
n=1
1
n
S
n
__
_
_
_
_
p
= 0.
To understand the signicance of 1
p
-convergence, we remark that if j is
very large, then 1
p
convergence is almost the same as uniform conver-
gence. Also:
If j , then L
p
[, ] L
q
[, ]. (Exercise 8D.8). E _
For example, if ) L
3
[, ], then it follows that ) L
2
[, ] (but
not vice versa). If ) L
2
[, ], then it follows that ) L
3/2
[, ]
(but not vice versa).
If j , and the Fourier series of ) converges to ) in 1
p
-norm, then
it also converges to ) in 1
q
-norm; see e.g. [Fol84, Proposition 6.12,
p.178].
For example, if ) L
2
[, ], then Theorem 8D.1(a) implies that the
Fourier series of ) converges to ) in 1
q
-norm for all [1, 2]. (How-
ever, if < 2, then there are functions in L
q
[, ] to which Theorem
8D.1(a) does not apply).
Finally, similar 1
p
-convergence statements hold for the Fourier (co)sine
series of real-valued functions in L
p
[0, ].
(g) The pointwise convergence of a Fourier series is a somewhat subtle and
complicated business, once you depart from the realm of (
1
functions. In
particular, the Fourier series of continuous (but non-dierentiable) func-
tions can be badly behaved. This is perplexing, because we know that
Fourier series converge in 1
2
norm for any function in L
2
[, ] (which in-
cludes all sorts of strange functions which are not dierentiable anywhere).
To bridge the gap between 1
2
and pointwise convergence, a variety of other
summation schemes have been introduced for Fourier coecients. These
include:
The Cesaro mean lim
N
1
n=1
o
N
()), where o
N
()) :=
N
n=N
)
n
E
n
is
the th partial sum of the complex Fourier series (8D.3).
The Abel mean lim
r,1
n=
:
[n[
)
n
E
n
.
These sums have somewhat nicer convergence properties than the stan-
dard Fourier series (8D.3). (See 18F on page 461 for further discussion
of the Abel mean.)
(h) There is a close relationship between the Fourier series of complex-valued
functions on [, ], and the Laurent series of complex-analytic functions
dened near the unit circle; see 18E on page 454.
(i) Remark (h) and the periodic boundary conditions required for Theorem
8D.1(d) both suggest that the Fourier series wants us to identify the
interval (, ] with the unit circle S in the complex plane, via the bijection
: (, ] S dened by (r) = c
ix
. Now, S is an abelian group under
the complex multiplication operator. That is: if :, t S, then their product
: t is also in S, the multiplicative inverse :
1
is in S, and the identity
element 1 is an element of S. Furthermore, S is a compact subset of C, and
the multiplication operation is continuous with respect to the topology of
S. In summary, S is a compact abelian topological group. The functions
E
n
n=
are then continuous homomorphisms from S into S (these are
called the characters of the group).
The existence of the Fourier series (8D.3) and the convergence properties
enumerated in Theorem 8D.1 are actually a consequence of these facts. In
fact, if G is any compact abelian topological group, then one can develop
a version of Fourier analysis on G. The characters of G are the continu-
ous homomorphisms from G into the unit circle group S. The set of all
characters of G forms an orthonormal basis for L
2
(G), so that almost any
reasonable function ) : G C can be expressed as a complex-linear
combination of these characters.
The study of Fourier series, their summability, and their generalizations to
other compact abelian groups is called harmonic analysis, and is a crucial
tool in many areas of mathematics, including the ergodic theory of dy-
namical systems and the representation theory of Lie groups. See [Fol84,
Ch.8], [WZ77, Ch.12] or the book [Kat76] to learn more about this vast
and fascinating area of mathematics.
179
Chapter 9
Multidimensional Fourier
series
The scientist does not study nature because it is useful; he studies it because he delights
in it, and he delights in it because it is beautiful. If nature were not beautiful, it would not
be worth knowing, and if nature were not worth knowing, life would not be worth living.
Henri Poincare
9A ...in two dimensions
Prerequisites: 6E, 6F. Recommended: 7B.
Let A, Y 0, and let X := [0, A] [0, Y ] be an A Y rectangle in the
plane. Suppose ) : X 1 is a real-valued function of two variables. For
all n, : N
+
:= 1, 2, 3, . . ., we dene the two-dimensional Fourier sine
coecients:
1
n,m
:=
4
AY
_
X
0
_
Y
0
)(r, j) sin
_
nr
A
_
sin
_
:j
Y
_
drdj.
The two-dimensional Fourier sine series of ) is the doubly innite summa-
tion:
n,m=1
1
n,m
sin
_
nr
A
_
sin
_
:j
Y
_
. (9A.1)
Notice that we are now summing over two independent indices, n and :.
Example 9A.1. Let A = = Y , so that X = [0, ] [0, ], and let
)(r, j) = r j. Then ) has two-dimensional Fourier sine series:
4
n,m=1
(1)
n+m
n:
sin(nr) sin(:j).
180 DRAFT
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
cos (3x) cos (3x) cos (y) cos (3x) cos (2y) cos (3x) cos (3y)
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
cos (2x) cos (2x) cos (y) cos (2x) cos (2y) cos (2x) cos (3y)
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
cos (x) cos (x) cos (y) cos (x) cos (2y) cos (x) cos (3y)
Figure 9A.1: C
n,m
for n = 1...3 and : = 0...3 (rotate page).
181
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
sin (3r) sin (j) sin (3r) sin (2j) sin (3r) sin (3j)
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
sin (2r) sin (j) sin (2r) sin (2j) sin (2r) sin (3j)
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
0
0
.
5
1
1
.
5
2
2
.
5
3
x
0
0
.
5
1
1
.
5
2
2
.
5
3
y
0
.
5 0
0
.
5
1
sin (r) sin (j) sin (r) sin (2j) sin (r) sin (3j)
Figure 9A.2: S
n,m
for n = 1...3 and : = 1...3 (rotate page).
182 DRAFT Chapter 9. Multidimensional Fourier series
To see this, recall from By Example 7C.2(c) on page 148, we know that
2
_

0
rsin(r) dr =
2(1)
n+1
n
.
Thus, 1
n,m
=
4
2
_

0
_

0
rj sin(nr) sin(:j) dr dj
=
_
2
_

0
rsin(nr) dr
_
_
2
_

0
j sin(:j) dj
_
=
_
2(1)
n+1
n
_
_
2(1)
m+1
:
_
=
4(1)
m+n
n:
.
Example 9A.2.
Let A = = Y , so that X = [0, ] [0, ], and let )(r, j) = 1 be the constant
1 function. Then ) has two-dimensional Fourier sine series:
4
n,m=1
[1 (1)
n
]
n
[1 (1)
m
]
:
sin(nr) sin(:j) =
16
n,m=1
both odd
1
n :
sin(nr) sin(:j)
Exercise 9A.1 Verify this. E _
For all n, : N := 0, 1, 2, 3, . . ., we dene the two-dimensional Fourier
cosine coecients of ):
0
:=
1
AY
_
X
0
_
Y
0
)(r, j) drdj,
n,0
:=
2
AY
_
X
0
_
Y
0
)(r, j) cos
_
nr
A
_
drdj for n 0;
0,m
:=
2
AY
_
X
0
_
Y
0
)(r, j) cos
_
:j
A
_
drdj for : 0; and
n,m
:=
4
AY
_
X
0
_
Y
0
)(r, j) cos
_
nr
A
_
cos
_
:j
Y
_
drdj for n, : 0.
The two-dimensional Fourier cosine series of ) is the doubly innite sum-
mation:
n,m=0
n,m
cos
_
nr
A
_
cos
_
:j
Y
_
. (9A.2)
9A. ...in two dimensions 183
In what sense do these series converge to )? For any n, : N, dene the
functions C
n,m
, S
n,m
: [0, A] [0, Y ] 1 by
C
n,m
(r, j) := cos
_
nr
A
_
cos
_
:j
Y
_
,
and S
n,m
(r, j) := sin
_
nr
A
_
sin
_
:j
Y
_
,
for all (r, j) [0, A] [0, Y ] (see Figures 9A.1 and 9A.2).
Theorem 9A.3. Two-dimensional Co/Sine Series Convergence
Let A, Y 0, and let X := [0, A][0, Y ].
(a) [i] The set S
n,m
; n, : N
+
is an orthogonal basis for L
2
(X).
[ii] The set C
n,m
; n, : N is also an orthogonal basis for L
2
(X).
[iii] Thus, if ) L
2
(X), then the series (9A.1) and (9A.2) both converge
to ) in 1
2
-norm. Furthermore, the coecient sequences
n,m
n,m=0
and
1
n,m
n,m=1
are the unique sequences of coecients with this property.
(b) If ) (
1
(X) (i.e. ) is continuously dierentiable on X), then the series
(9A.1) and (9A.2) both converge to ) pointwise on (0, A) (0, Y ).
(c) [i] If
n,m=1
[1
n,m
[ < , then the two-dimensional Fourier sine series (9A.1)
converges to ) uniformly on X.
[ii] If
n,m=0
[
n,m
[ < , then the two-dimensional Fourier cosine series
(9A.2) converges to ) uniformly on X.
(d) [i] If ) (
1
(X), and the derivative functions
x
) and
y
) are both in
L
2
(X), and ) satises homogeneous Dirichlet boundary conditions
1
on X,
then the two-dimensional Fourier sine series (9A.1) converges to ) uniformly
on X.
[ii] Conversely, if the series (9A.1) converges to ) uniformly on X, then )
is continuous and satises homogeneous Dirichlet boundary conditions.
(e) [i] If ) (
1
(X), the derivative functions
x
) and
y
) are both in L
2
(X),
then the two-dimensional Fourier cosine series (9A.2) converges to ) uni-
formly on X.
1
That is, f(0, y) = 0 = f(X, y) for all y [0, Y ], and f(x, 0) = 0 = f(x, Y ) for all x [0, X].
/2
1
/2
Figure 9A.3: The box function )(r, j) in Example 9A.4.

[ii] Conversely, if
n,m=1
n[
nm
[ < and
n,m=1
:[
nm
[ < , then )
(
1
(X), and ) satises homogeneous Neumann boundary conditions
2
on X.
Proof. This is just the case 1 = 2 of Theorem 9B.1 on page 187. 2
Example 9A.4. Suppose A = = Y , and )(r, j) =
_
1 if 0 r <

2
and 0 j <

2
;
0 if

2
r or

2
j.
(See Figure 9A.3). Then the two-dimensional Fourier cosine series of ) is:
1
4
+
1
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
+
1
j=0
(1)
j
2, + 1
cos
_
(2, + 1)j
_
+
4
k,j=0
(1)
k+j
(2/ + 1)(2, + 1)
cos
_
(2/ + 1)r
_
cos
_
(2, + 1)j
_
To see this, note that )(r, j) = p(r)p(j), where p(r) =
_
1 if 0 r <

2
0 if

2
r
.
Recall from Example 7C.6 on page 154 that the (one-dimensional) Fourier co-
sine series of p(r) is
p(r)

L2
1
2
+
2
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
Thus, the cosine series for )(r, j) is given:
)(r, j) = p(r) p(j)
2
That is, x f(0, y) = 0 = x f(X, y) for all y [0, Y ], and y f(x, 0) = 0 = y f(x, Y ) for
all x [0, X].
9A. ...in two dimensions 185
L2
_
1
2
+
2
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
_
_
_
1
2
+
2
j=0
(1)
j
2, + 1
cos
_
(2, + 1)j
_
_
_
.
Mixed Fourier series. (Optional)

We can also dene the mixed Fourier sine/cosine coecients:
C
[sc]
n,0
:=
2
AY
_
X
0
_
Y
0
)(r, j) sin
_
nr
A
_
drdj, for n 0.
C
[sc]
n,m
:=
4
AY
_
X
0
_
Y
0
)(r, j) sin
_
nr
A
_
cos
_
:j
Y
_
drdj, for n, : 0.
C
[cs]
0,m
:=
2
AY
_
X
0
_
Y
0
)(r, j) sin
_
:j
Y
_
drdj, for : 0.
C
[cs]
n,m
:=
4
AY
_
X
0
_
Y
0
)(r, j) cos
_
nr
A
_
sin
_
:j
Y
_
drdj, for n, : 0.
The mixed Fourier sine/cosine series of ) are then:
n=1,m=0
C
[sc]
n,m
sin
_
nr
A
_
cos
_
:j
Y
_
(9A.3)
and
n=0,m=1
C
[cs]
n,m
cos
_
nr
A
_
sin
_
:j
Y
_
For any n, : N, dene the functions M
[sc]
n,m
, M
[cs]
n,m
: [0, A] [0, Y ] 1 by
M
[sc]
n,m
(r, j) := sin
_
n
1
r
A
_
cos
_
n
2
j
Y
_
and M
[cs]
n,m
(r, j) := cos
_
n
1
r
A
_
sin
_
n
2
j
Y
_
.
for all (r, j) [0, A] [0, Y ].
Proposition 9A.5. Two-dimensional Mixed Co/Sine Series Convergence
Let X := [0, A][0, Y ]. The sets of mixed functions,
_
M
[sc]
n,m
; n N
+
, : N
_
and
_
M
[cs]
n,m
; n N, : N
+
_
are both orthogonal basis for L
2
(X). In other
words, if ) L
2
(X), then the series (9A.3) both converge to ) in 1
2
. 2
Exercise 9A.2. Formulate conditions for pointwise and uniform convergence of the E _
mixed series.
9B ...in many dimensions
Prerequisites: 6E, 6F. Recommended: 9A.
Let A
1
, . . . , A
D
0, and let X := [0, A
1
] [0, A
D
] be an A
1
A
D
box in 1-dimensional space. For any n N
D
, dene the functions C
n
: X 1
and S
n
: X 1 by
C
n
(r
1
, . . . , r
D
) := cos
_
n
1
r
1
A
1
_
cos
_
n
2
r
2
A
2
_
cos
_
n
D
r
D
A
D
_
, (9B.1)
S
n
(r
1
, . . . , r
D
) := sin
_
n
1
r
1
A
1
_
sin
_
n
2
r
2
A
2
_
sin
_
n
D
r
D
A
D
_
, (9B.2)
for any x = (r
1
, r
2
, . . . , r
D
) X. Also, for any sequence = (
1
, . . . ,
D
) of 1
symbols : and c, we can dene the mixed functions, M
n
: X 1. For
example, if 1 = 3, then dene
M
[scs]
n
(r, j, .) := sin
_
n
1
r
A
x
_
cos
_
n
2
j
A
y
_
sin
_
n
3
.
A
z
_
.
If ) : X 1 is any function with |)|
2
< , then, for all n N
D
+
, we dene
the multiple Fourier sine coecients:
1
n
:=
), S
n
)
|S
n
|
2
2
=
2
D
A
1
A
D
_
X
)(x) S
n
(x) dx.
The multiple Fourier sine series of ) is then:
nN
D
+
1
n
S
n
. (9B.3)
For all n N
D
, we dene the multiple Fourier cosine coecients:
0
:= ), 11) =
1
A
1
A
D
_
X
)(x) dx,
and
n
:=
), C
n
)
|C
n
|
2
2
=
2
dn
A
1
A
D
_
X
)(x) C
n
(x) dx.
where, for each n N
D
, the number d
n
is the number of nonzero entries in
n = (n
1
, n
2
, . . . , n
D
). The multiple Fourier cosine series of ) is then:
nN
D
n
C
n
, where N := 0, 1, 2, 3, . . .. (9B.4)
9B. ...in many dimensions 187
Finally, we dene the mixed Fourier Sine/Cosine coecients:
C
n
:=
), M
n
)
|M
n
|
2
2
=
2
dn
A
1
A
D
_
X
)(x) M
n
(x) dx,
where, for each n N
D
, the number d
n
is the number of nonzero entries n
i
in
n = (n
1
, . . . , n
D
). The mixed Fourier Sine/Cosine series of ) is then:
nN
D
C
n
M
n
. (9B.5)
Theorem 9B.1. Multidimensional Co/Sine Series Convergence on X
Let X := [0, A
1
] [0, A
D
] be a 1-dimensional box.
(a) [i] The set
_
S
n
; n N
D
+
_
2
(X).
[ii] The set
_
C
n
; n N
D
_
2
(X).
[iii] For any sequence of 1 symbols : and c, the set of mixed
functions,
_
M
n
; n N
D
_
2
(X).
[iv] In other words, if ) L
2
(X), then the series (9B.3), (9B.4), and
(9B.5) all converge to ) in 1
2
-norm. Furthermore, the coecient sequences
nN
D, 1
n
nN
D
+
, and C
nN
D are the unique sequences of coe-
cients with these properties.
(b) If ) (
1
(X) (i.e. ) is continuously dierentiable on X), then the series
(9B.3), (9B.4), and (9B.5) converge pointwise on the interior of X.
(c) [i] If
nN
D
+
[1
n
[ < , then the multidimensional Fourier sine series (9B.3)
[ii] If
nN
D
[
n
[ < , then the multidimensional Fourier cosine series (9B.4)
(d) [i] If ) (
1
k
) are themselves in L
2
(X)
for all / [1...1], and ) satises homogeneous Dirichlet boundary condi-
tions on X, then the multidimensional Fourier sine series (9B.3) converges
to ) uniformly on X.
[ii] Conversely, if the series (9B.3) converges to ) uniformly on X, then )
is continuous and satises homogeneous Dirichlet boundary conditions.
(e) [i] If ) (
1
k
) are themselves in L
2
(X)
for all / [1...1], then the multidimensional Fourier cosine series (9B.4)
[ii] Conversely, if
nN
D
(n
1
+ + n
D
) [
n
[ < , then ) (
1
(X), and )
satises homogeneous Neumann boundary conditions.
Proof. The proof of (c) is Exercise 9B.1 (Hint: Use the Weierstrass `-test, E _
Proposition 6E.13 on page 129.)
The proofs of (d,e)[ii] are Exercise 9B.2 (Hint: Generalize the solutions to E _
Exercises 7A.4 and 7A.9 on pages 139 and 142).
The proof of (a) is Exercise 9B.3 (Hint: Prove this by induction on the dimen- E _
sion 1. The base case (1 = 1) is Theorems 7A.1(a) and 7A.4(a) on pages 138 and
142. Use Lemma 15C.2(f) (on page 330) to handle the induction step.)
We will prove (b), (d)[i] and (e)[i] by induction on the dimension 1. The
base cases (1 = 1) are Theorems 7A.1(b,d[i]) and 7A.4(b,d[i]) on pages 138
and 142.
For the induction step, suppose the theorem is true for 1, and consider 1+1.
Let X := [0, A
0
] [0, A
1
] [0, A
D
] be a (1 + 1)-dimensional box. Note
that X := [0, A
0
] X
, where X
:= [0, A
1
] [0, A
D
] is a 1-dimensional
box. If ) : X 1, then for all j [0, A
0
], let )
y
: X
1 be the function
dened by )
y
(x) := )(j, x) for all x X
.
Claim 1: (a) If ) (
1
(X), then )
y
(
1
(X
) for all j [0, A

0
].
(b) Furthermore, if
k
) L
2
(X) for all / [1...1], then
k
()
y
) L
2
(X
) for
all / [1...1] and all j [0, A
0
].
(c) If ) satises homogenous Dirichlet BC on X, then )
y
satises homogenous
Dirichlet BC on X
, for all j [0, A

0
].
Proof. Exercise 9B.4
Claim 1
E _
For all n N
D
, dene C
n
, S
n
: X
1 as in equations (9B.1) and (9B.2).

For every j [0, A
0
], let
y
n
:=
)
y
, C
n
)
|C
n
|
2
2
and 1
y
n
:=
)
y
, S
n
)
|S
n
|
2
2
be the 1-dimensional Fourier (co)sine coecients for )
y
, so that )
y
has 1-
dimensional Fourier (co)sine series:
nN
D
+
1
y
n
S
n
L2
)
y
L2
nN
D
y
n
C
n
. (9B.6)
Claim 2: For all j [0, A
0
], the two series in eqn.(9B.6) converge to )
y
in
the desired fashion (i.e. pointwise or uniform) on X
.
Proof. Exercise 9B.5 (Hint: Use the induction hypothesis and Claim 1). E _
Claim 2
Fix n N
D
. Dene
n
,
n
: [0, A
0
] 1 by
n
(j) :=
y
n
and
n
(j) := 1
y
n
for all j [0, A
0
].
Claim 3: For all n N
D
,
n
L
2
[0, A
0
] and
n
L
2
[0, A
0
].
Proof. We have
|
n
|
2
2
=
1
A
0
_
X
0
0
[
n
(j)[
2
dj =
1
A
0
_
X
0
0
)
y
, C
n
)
|C
n
|
2
2
2
dj
=
1
A
0
|C
n
|
4
2
_
X
0
0
[)
y
, C
n
)[
2
dj
()
1
A
0
|C
n
|
4
2
_
X
0
0
|)
y
|
2
2
|C
n
|
2
2
dj
=
1
A
0
|C
n
|
2
2
_
X
0
0
|)
y
|
2
2
dj
=
1
A
0
|C
n
|
2
2
__
X
0
0
1
A
1
A
D
_
X
[)
y
(x)[
2
dx
_
dj
=
1
A
0
A
D
|C
n
|
2
2
_
X
[)(j, x)[
2
d(j; x) =
1
|C
n
|
2
2
|)|
2
2
.
Here, () is the Cauchy-Bunyakowski-Schwarz Inequality (Theorem 6B.5 on
page 108).
Thus, |
n
|
2
2
< because |)|
2
2
< because ) L
2
(X) by hypothesis.
Thus,
n
L
2
[0, A
0
]. The proof that
n
L
2
[0, A
0
] is similar.
Claim 3
For all : N, dene S
m
, C
m
: [0, A
0
] 1 by S
m
(j) := sin(:j,A
0
) and
C
m
(j) := cos(:j,A
0
), for all j [0, A
0
]. For all : N, let
n
m
:=

n
, C
m
)
|C
m
|
2
2
and 1
n
m
:=

n
, S
m
)
|S
m
|
2
2
be the one-dimensional Fourier (co)sine coecients for the functions
n
and
n
, so that we get one-dimensional Fourier (co)sine series
n
L2
m=0
n
m
C
m
and
n
L2
m=1
1
n
m
S
m
. (9B.7)
For all n N
D
and all : N, dene S
m;n
, C
m;n
: X 1 by S
m;n
(j; x) :=
S
m
(j) S
n
(x) and C
m;n
(j; x) := C
m
(j) C
n
(x), for all j [0, A
0
] and x X
.
Then let
m;n
:=
), C
m;n
)
|C
m;n
|
2
2
and 1
m;n
:=
), S
m;n
)
|S
m;n
|
2
2
be the (1+1)-dimensional Fourier (co)sine coecients for the function ), so
that we get (1+1)dimensional Fourier (co)sine series
m=0
nN
D
m;n
C
m;n
L2
)

L2
m=1
nN
D
+
1
m;n
S
m;n
. (9B.8)
Claim 4: For all n N
D
and all : N,
n
m
=
m;n
and 1
n
m
= 1
m;n
.
Proof. Exercise 9B.6
Claim 4
E _
Let
0
) be the derivative of ) in the 0th (or j) coordinate, which we regard
as a function
0
) : X 1.
Claim 5: (a) If ) (
1
(X), then for all n N
D
,
n
(
1
[0, A
0
] and
n
(
1
[0, A
0
].
(b) Furthermore, if
0
) L
2
(X), then for all n N
D
,
t
n
L
2
[0, A
0
] and
t
n
L
2
[0, A
0
].
(c) If ) satises homogeneous Dirichlet BC on X, then
n
satises homogeneous
Dirichlet BC on [0, A
0
], for all n N
D
+
.
Proof. To prove (a), we proceed as follows.
Exercise 9B.7 (a) Show that ) is uniformly continuous on X. (Hint: ) is E _
continuous on X, and X is compact.)
(b) Show: for any j
0
[0, A
0
], the functions )
y
converge uniformly to )
y0
as
j j
0
.
(c) For any xed n N, deduce that lim
yy0
y
n
=
y0
n
and lim
yy0
1
y
n
= 1
y0
n
. (Hint:
Use Corollary 6E.11(b)[ii] on page 127.)
(d) Conclude that the functions
n
and
n
are continuous at j
0
.
The conclusion of Exercise 9B.7(d) holds for all j
0
[0, A
0
] and all n N.
Thus, the functions
n
and
n
are continuous on [0, A
0
], for all n N.
For all j [0, A
0
], let (
0
))
y
: X
1 be the function dened by

(
0
))
y
(x) :=
0
)(j, x) for all x X
.
Exercise 9B.8 Suppose ) (
1
(X). Use Proposition 0G.1 on page 567 to E _
show, for all n N
D
, that the functions
n
and
n
are dierentiable on [0, A
0
];
furthermore, for all j [0, A
0
],
t
n
(j) =
(
0
))
y
, C
n
)
|C
n
|
2
2
and
t
n
(j) =
(
0
))
y
, S
n
)
|S
n
|
2
2
. (9B.9)
Exercise 9B.9 Using the same technique as Exercise 9B.7, use eqn.(9B.9) to E _
prove that the functions
n
and
n
are continuous on [0, A
0
].
Thus,
n
(
1
[0, A
0
] and
n
(
1
[0, A
0
]; this proves part (a) of the Claim.
The proof of (b) is Exercise 9B.10 (Hint. Imitate the proof of Claim 3). E _
The proof of (c) is Exercise 9B.11 .
Claim 5
E _
Claim 6: The one-dimensional Fourier cosine series in eqn.(9B.7) converges
to
n
, and the one-dimensional Fourier sine series in eqn.(9B.7) converges to
n
in the desired fashion (i.e. pointwise or uniform), for all n N
D
.
Proof. Exercise 9B.12 (Hint: Use Theorems 7A.1(b,d[i]) and 7A.4(b,d[i]) and E _
Claim 5).
Now, Claim 4 implies that the (1+1)dimensional Fourier (co)sine series in
(9B.8) can be rewritten as
m=0
nN
D
n
m
C
m
C
n
L2
)

L2
m=1
nN
D
+
1
n
m
S
m
S
n
. (9B.10)
Exercise 9B.13 Suppose ) (
1
(X). Use the pointwise versions of Claims 2 and E _
6 to show that the two series in eqn.(9B.10) converges to ) pointwise on the interior
of X.
This proves part (b) of the Theorem.
Exercise 9B.14 (hard) Suppose ) (
1
(X), the derivative functions
0
),
1
), . . . ,
0
) E _
are all in L
2
(X), and (for the sine series) ) satises homogeneous Dirichlet boundary
conditions. Use the uniform versions of Claims 2 and 6 to show that the two series
in eqn.(9B.10) converges to ) uniformly if ) (
1
(X).
This proves parts (d)[i] and (e)[i] of the Theorem. 2
Remarks. (a) If ) is a piecewise (
1
function on the interval [0, ], then The-
orems 7A.1 and 7A.4 also yield pointwise convergence and local uniform con-
vergence of one-dimensional Fourier (co)sine to ) inside the (
1
intervals of ).
Likewise, if ) is a piecewise (
1
function on the 1-dimensional domain X, then
one can extend Theorem 9B.1 to get pointwise convergence and local uniform
convergence of 1-dimensional Fourier (co)sine to ) inside the (
1
regions of );
however, it is too technically complicated to formally state this here.
(b) Remark 8D.3 on page 174 provided some technical remarks about the
(non)convergence of one-dimensional Fourier (co)sine series, when the hypotheses
of Theorems 7A.1 and 7A.4 are further weakened. Similar remarks apply to 1-
dimensional Fourier series.
(c) It is also possible to dene 1-dimensional complex Fourier series on the
1-dimensional box [, ]
D
, in a manner analogous to the results of Section
8D, and then state and prove a theorem analogous to Theorem 9B.1 for such
1-dimensional complex Fourier series. (Exercise 9B.15 (Challenging) Do this.) E _
In Chapters 11-14, we will often propose a multiple Fourier series (or similar
object) as the solution to some PDE, perhaps with certain boundary conditions.
To verify that the Fourier series really satises the PDE, we must be able to
compute its Laplacian. If we also require the Fourier series solution to satisfy
some Neumann boundary conditions, then we must be able to compute its normal
derivatives on the boundary of the domain. For these purposes, the next result
is crucial.
Proposition 9B.2. The Derivatives of a Multiple Fourier (co)sine series
Let X := [0, A
1
] [0, A
D
]. Let ) : X 1 be have uniformly convergent
Fourier series
)
unif

0
+
nN
D
n
C
n
+
nN
D
+
1
n
S
n
.
(a) Fix i [1...1]. Suppose that
nN
D
n
i
[
n
[ +
nN
D
+
n
i
[1
n
[ < . Then the
function
i
) exists, and
i
)

L2
nN
D
_
n
i
A
i
_
_
1
n
S
t
n

n
C
t
n
_
.
Here, for all n N
D
, and all x X, we dene
C
t
n
(x) := sin
_
n
i
r
i
A
i
_
C
n
(x), cos
_
n
i
r
i
A
i
_
,
and S
t
n
(x) := cos
_
n
i
r
i
A
i
_
S
n
(x), sin
_
n
i
r
i
A
i
_
.
(b) Fix i [1...1]. Suppose that
nN
D
n
2
i
[
n
[ +
nN
D
+
n
2
i
[1
n
[ < . Then
the function
2
i
) exists, and
2
i
)

L2
nN
D
_
n
i
A
i
_
2
n
C
n
+1
n
S
n
_
.
(c) Suppose that
nN
D
[n[
2
[
n
[ +
nN
D
[n[
2
[1
n
[ < (where we dene
[n[
2
:= n
2
1
+.... +n
2
D
). Then ) is twice-dierentiable, and
)

L2
nN
D
_
_
n
1
A
1
_
2
+ +
_
n
D
A
D
_
2
_
n
C
n
+1
n
S
n
_
.
Proof. Exercise 9B.16 Hint: Apply Proposition 0F.1 on page 565 2 E _
9C. Practice problems 193
Example 9B.3. Fix n N
D
. If ) = C
n
+1 S
n
, then
) =
2
_
_
n
1
A
1
_
2
+ +
_
n
D
A
D
_
2
_
).
In particular, if A
1
= = A
D
= , then this simplies to: ) = [n[
2
).
In other words, ) is an eigenfunction of the Laplacian operator, with eigenvalue
= [n[
2
.
9C Practice problems
Compute the two-dimensional Fourier sine transforms of the following functions.
For each question, also determine: at which points does the series converge
pointwise? Why? Does the series converge uniformly? Why or why not?
1. )(r, j) = r
2
j.
2. p(r, j) = r +j.
3. )(r, j) = cos(r) cos(`j), for some integers `, 0.
4. )(r, j) = sin(r) sinh(j), for some integer 0.
195
Chapter 10
Proofs of the Fourier
convergence theorems
The profound study of nature is the most fertile source of mathematical discoveries.
Jean Joseph Fourier
In this section, we will prove Theorem 8A.1(a,b,d) on page 162 (and thus,
indirectly prove 7A.1(a,b,d) and 7A.4(a,b,d) on pages 138 and 142). Along the
way, we will introduce some ideas which are of independent interest: Bessels
inequality, the Riemann-Lebesgue lemma, the Dirichlet kernel, convolutions and
molliers, and the relationship between the smoothness of a function and the
asymptotic decay of its Fourier coecients. This chapter assumes no prior knowl-
edge of analysis, beyond some background from Chapter 6. However, the pre-
sentation is slightly more abstract than most of the book, and is intended for
more theoretically inclined students.
10A Bessel, Riemann and Lebesgue
Prerequisites: 6D. Recommended: 7A, 8A.
We begin with a general result which is true for any orthonormal set in any
1
2
space.
Theorem 10A.1. (Bessels Inequality)
Let X 1
D
be any bounded domain. Let
n
n=1
be any orthonormal set of
functions in L
2
(X). Let ) L
2
(X), and for all n N, let c
n
:= ),
n
). Then for
all N,
N
n=1
[c
n
[
2
|)|
2
2
.
In particular, lim
n
c
n
= 0.
196 DRAFT Chapter 10. Proofs of the Fourier convergence theorems
Proof. Without loss of generality, suppose [X[ = 1, so that ), p) =
_
X
)(r) p(r) dr
for any ), p L
2
(X). First note that
_
)(r)
N
n=1
c
n
n
(r)
_
2
= )(r)
2
2)(r)
N
n=1
c
n
n
(r) +
_
N
n=1
c
n
n
(r)
_
_
N
m=1
c
m
m
(r)
_
= )(r)
2
2
N
n=1
c
n
)(r)
n
(r) +
N
n,m=1
c
n
c
m
n
(r)
m
(r). (10A.1)
Thus,
0
_
_
_
_
_
)
N
n=1
c
n
n
_
_
_
_
_
2
2
=
_
X
_
)(r)
N
n=1
c
n
n
(r)
_
2
dr
()
_
X
_
_
)(r)
2
2
N
n=1
c
n
)(r)
n
(r) +
N
n,m=1
c
n
c
m
n
(r)
m
(r)
_
_
dr
=
_
X
)(r)
2
dr 2
N
n=1
c
n
_
X
)(r)
n
(r) dr +
N
n,m=1
c
n
c
m
_
X
n
(r)
m
(r) dr
= ), ))
. .
|f|
2
2
2
N
n=1
c
n
),
n
)
. .
cn
+
N
n,m=1
c
n
c
m
n
,
m
)
. .
=1 if n=m
=0 if n=m
= |)|
2
2
2
N
n=1
c
2
n
+
N
n=1
c
2
n
= |)|
2
2

N
n=1
c
2
n
.
Here, () is by eqn.(10A.1). Thus, 0 |)|
2
2

N
n=1
c
2
n
. Thus
N
n=1
c
2
n
|)|
2
2
,
as desired. 2
Example 10A.2. Suppose ) L
2
[, ] has real Fourier coecients
n
n=0
and 1
n
n=1
, as dened on page 161. Then for all N,
2
0
+
N
n=1
[
n
[
2
2
+
N
n=1
[1
n
[
2
2
|)|
2
2
.
Exercise 10A.1 Prove this. (Hint: Let X = [, ] and let
k
k=1
=
_
2C
n
_
n=0
. E _
_
2S
n
_
n=1
. Show that
k
k=1
is an orthonormal set of functions (Use Proposi-
tion 6D.2 on page 112). Now apply Bessels Inequality).
10B. Pointwise convergence 197
Corollary 10A.3. (Riemann-Lebesgue Lemma)
(a) Suppose ) L
2
[, ] has real Fourier coecients
n
n=0
and 1
n
n=1
,
as dened on page 161. Then lim
n
n
= 0 and lim
n
1
n
= 0.
(b) Suppose ) L
2
[0, ] has Fourier cosine coecients
n
n=0
, as dened by
eqn.(7A.4) on page 141, and Fourier sine coecients 1
n
n=1
, as dened
by eqn.(7A.1) on page 137. Then lim
n
n
= 0 and lim
n
1
n
= 0.
Proof. Exercise 10A.2 Hint: Use Example 10A.2. 2 E _
10B Pointwise convergence
Prerequisites: 8A, 10A. Recommended: 17B.
In this section we will prove Theorem 8A.1(b), through a common strategy
in harmonic analysis: the use of a summation kernel. For all N, the th
Dirichlet kernel is the function D
N
: [2, 2] 1 dened by
D
N
(r) := 1 + 2
N
n=1
cos(nr) (see Figure 10B.1).
Note that D
N
is 2-periodic (i.e. D
N
(r + 2) = D
N
(r) for all r [2, 0]).
Thus, we could represent D
N
as a function from [, ] into 1. However, it is
sometimes convenient to extend D
N
to [2, 2]. For example, for any function
) : [, ] 1, the convolution of D
N
and ) is the function D
N
) :
[, ] 1 dened by
D
N
)(r) :=
1
2
_

)(j) D
N
(r j) dj, for all r [, ].
(Note that, to dene D
N
), we must evaluate D
N
(.) for all . [2, 2]).
The connection between Dirichlet kernels and Fourier series is given by the next
lemma:
Lemma 10B.1. Let ) L
2
[, ], and for all n N, let
n
:=
1
cos(nj))(j) dj and 1
n
:=
1
sin(nj))(j) dj
be the real Fourier coecients of ). Then for any N, and every r [, ],
we have
0
+
N
n=1
n
C
n
(r) +
N
n=1
1
n
S
n
(r) = D
N
)(r).
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
D
1
D
2
D
3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
D
4
D
5
D
6
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
-2
0
2
4
6
8
10
-3 -2 -1 0 1 2 3
D
7
D
8
D
9
Figure 10B.1: The Dirichlet kernels D1, D2, . . . , D9 plotted on interval [, ]. Note the
increasing concentration of the function near x = 0. (In the terminology of Section 10D(ii) and
17B, the sequence |D1, D2, . . . is like an approximation of the identity.)
Proof. For any r [, ], we have
0
+
N
n=1
n
C
n
(r) +
N
n=1
1
n
S
n
(r)
=
0
+
N
n=1
cos(nr)
_
1
cos(nj))(j) dj
_
+
N
n=1
sin(nr)
_
1
sin(nj))(j) dj
_
=
0
+
N
n=1
1
__

cos(nr) cos(nj))(j) dj +
_

sin(nr) sin(nj))(j) dj
_
=
0
+
N
n=1
1
_
cos(nr) cos(nj) + sin(nr) sin(nj)
_
)(j) dj
()
1
2
_

)(j) dj +
N
n=1
1
cos
_
n(r j)
_
)(j) dj
=
1
2
_

_
)(j) + 2
N
n=1
cos
_
n(r j)
_
)(j)
_
dj
=
1
2
_

)(j) D
N
(r j) dj = D
N
)(r).
Here, () uses the fact that
0
:=
1
2
_
)(j) dj, and also the well-known

trigonometric identity cos(n ) = cos(n) cos() + sin(n) sin() (with n = nr
and = nj). 2
Remark. See Exercise 18F.7 on page 464 for another proof of Lemma 10B.1
for complex Fourier series.
Figure 10B.1 shows how the mass of the Dirichlet kernel D
N
becomes in-
creasingly concentrated near r = 0 as . In the terminology of Sections
10D and 17B (pages 207 and 379), the sequence D
1
, D
2
, . . . is like an approxi-
mation of the identity. Thus, our strategy is to show that D
N
)(r) )(r) as
, whenever ) is continuous at r. Indeed, we will go further: when ) is
discontinuous at r, we will show that D
N
)(r) converges to the average of the
left-hand and right-hand limits of ) at r. First we need some technical results.
Lemma 10B.2.
(a) For any N, we have
_

0
D
N
(r) dr = .
(b) For any N and r (, 0).(0, ), we have D
N
(r) =
sin((2 + 1)r,2)
sin(r,2)
.
(c) Let p : [0, ] 1 be a piecewise continuous function. Then
lim
N
_

0
p(r) sin
_
(2 + 1)r
2
_
dr = 0.
Proof. The proof of (b) is Exercise 10B.1 (Hint: Use trigonometric identities). E _
To prove (a), note that
_

0
D
N
(r) dr =
_

0
1 + 2
N
n=1
cos(nr) dr =
_

0
1 dr + 2
N
n=1
_

0
cos(nr) dr
= + 2
N
n=1
0 = .
To prove (c), rst observe that
sin
_
(2 + 1)r
2
_
= sin
_
r +
r
2
_
= sin(r) cos(r,2) + cos(r) sin(r,2), (10B.1)
where the last step uses the well-known trigonometric identity sin(n + ) =
sin(n) cos() + cos(n) sin() (with n := r and := r,2). Thus,
_

0
p(r) sin
_
(2 + 1)r
2
_
dr
()
_

0
p(r)
_
sin(r) cos(r,2) + cos(r) sin(r,2)
_
dr
=
_

0
p(r) cos(r,2)
. .
G
1
(x)
sin(r)
. .
S
N
(x)
dr +
_

0
p(r) sin(r,2)
. .
G
2
(x)
cos(r)
. .
C
N
(x)
dr
()
2
2
_

0
G
1
(r) S
N
(r) dr +
2
2
_

0
G
2
(r) C
N
(r) dr
()
2
G
1
, S
N
) +

2
G
2
, C
N
)
N
0 + 0, by Corollary 10A.3(b) (the Riemann-Lebesgue Lemma).
Here () is by eqn.(10B.1) and () is by denition of the inner product on
L
2
[0, ]. In (), we dene the functions G
1
(r) := p(r) cos(r,2) G
2
(r) :=
p(r) sin(r,2); these functions are piecewise continuous because p is piecewise
continuous; thus they are in L
2
[0, ], so the Riemann-Lebesgue Lemma is
applicable. 2
L
R
x
f
(A)
x
f
(B)
x
f
(C)
f

(
(
x
)
f

)
(
x
)
Figure 10B.2: (A) Left-hand and right-hand limits. Here, L := lim
y,x
f(x) and R := lim
y_x
f(x).
(B) The right-hand derivative f
(x). (C) The left-hand derivative f

)
(x).
Let ) : [, ] 1 be a function. For any r [, ), the right-hand
limit of ) at r is dened
lim
yx
)(j) := lim
0
)(r +[c[) (if this limit exists).
Likewise, for any r (, ], the left-hand limit of ) at r is dened
lim
y,x
)(j) := lim
0
)(r [c[) (if this limit exists).
See Figure 10B.2(A). Clearly, if ) is continuous at r, then the left-hand and
right-hand limits both exist, and lim
yx
)(j) = )(r) = lim
y,x
)(j). However, the
left-hand and right-hand limits may exist even when ) is not continuous.
For any r [, ), let )(r
+
) := lim
yx
)(j). The right-hand derivative of
) at r is dened
)
(r) := lim
yx
)(j) )(r
+
)
j r
= lim
0
)(r +[c[) )(r
+
)
[c[
(if this limit exists).
See Figure 10B.2(B). Likewise, for any r (, ], let )(r
) := lim
y,x
)(j). The
left-hand derivative of ) at r is dened
)
)
(r) := lim
y,x
)(j) )(r
)
j r
= lim
0
)(r [c[) )(r
)
[c[
(if this limit exists).
See Figure 10B.2(C). If )
(r) and )
)
(r) both exist, then we say ) is semidif-
ferentiable at r. Clearly, ) is dierentiable at r if and only if ) is continuous
at r (so that )(r
) = )(r
+
)), and ) semidierentiable at r, and )
(r) = )
)
(r).
In this case, )
t
(r) = )
(r) = )
)
(r). However, ) can be semidierentiable at r
even when ) is not dierentiable (or even continuous) at r.
Lemma 10B.3. Let

) : [, ] 1 be a piecewise continuous function
which is semidierentiable at 0. Then
lim
N
_

)(r) D
N
(r) dr =
_
lim
x,0
)(r) + lim
x0
)(r)
_
.
Proof. It suces to show that
lim
N
_
0
)(r) D
N
(r) dr = lim
x,0
)(r) (10B.2)
and lim
N
_

0
)(r) D
N
(r) dr = lim
x0
)(r). (10B.3)
We will prove eqn.(10B.3). Let

)(0
+
) := lim
x0
)(r), and consider the function

p : [0, ] 1 dened by p(r) :=

)(r)

)(0
+
)
sin(r,2)
if r 0, while p(0) :=
2

)
(0).
Claim 1: p is piecewise continuous on [0, ].
Proof. Clearly, p is piecewise continuous on (0, ] because

) is piecewise
continuous, while sin(r,2) is nonzero on (0, ]. The only potential location
of an unbounded discontinuity is at 0. But
lim
x0
p(r) = lim
x0
)(r)

)(0
+
)
sin(r,2)
= lim
x0
_

)(r)

)(0
+
)
r
_
_
r
sin(r,2)
_
=
_
lim
x0
)(r)

)(0
+
)
r 0
_
. .
(0)
2
_
lim
x0
r,2
sin(r,2)
_
. .
=1
= 2

)
(0) =: p(0).
Thus, p is (right-)continuous at 0, as desired.
Claim 1
Now,
lim
N
_

0
)(r) D
N
(r) dr
= lim
N
_

0
_
)(0
+
) +

)(r)

)(0
+
)
_
D
N
(r) dr
= lim
N
_

0
)(0
+
) D
N
(r) dr +
_

0
_
)(r)

)(0
+
)
_
D
N
(r) dr
(a)

)(0
+
) +
_

0
_
)(r)

)(0
+
)
_
D
N
(r) dr
(b)

)(0
+
) + lim
N
_

0
)(r)

)(0
+
)
sin(r,2)
sin
_
(2 + 1)r
2
_
dr
=

)(0
+
) + lim
N
_

0
p(r) sin
_
(2 + 1)r
2
_
dr
(c)

)(0
+
) + 0 =

)(0
+
),
as desired. Here, (a) is by Lemma 10B.2(a), (b) is by Lemma 10B.2(b), and (c)
is by Lemma 10B.2(c), which is applicable because p is piecewise continuous
by Claim 1.
This proves eqn.(10B.3). The proof of eqn.(10B.2) is Exercise 10B.2. Adding E _
together equations (10B.2) and (10B.3) proves the lemma. 2
Lemma 10B.4. Let ) : [, ] 1 be piecewise continuous, and suppose
that ) is semidierentiable at r (i.e. )
(r) and )
)
(r) exist). Then
lim
N
D
N
)(r) =
1
2
_
lim
yx
)(j) + lim
y,x
)(j)
_
. (10B.4)
In particular, if ) is continuous and semidierentiable at r, then lim
N
D
N

)(r) = )(r).

f (y)
x x
f (y+x)
x
x
f (y)
(A)
(B) (C)
0 0 0
-x
x
Figure 10B.3: The 2-periodic phase-shift of a function. (A) A function f : [, ] 1.
(B) The function y f(x +y). (C) The function
f : [, ] 1.
Proof. Suppose r [0, ] (the case r [, 0] is handled similarly). Dene
) : [, ] 1 by
) (j) :=
_
)(j +r) if j [, r];
)(j +r 2) if j [ r, ].
(Eectively, we are treating ) as a 2-periodic function, and phase-shifting
) by r; see Figure 10B.3). Then
2 D
N
)(r) =
_

)(j)D
N
(r j) dj
()
_

)(j)D
N
(j r) dj
(c)
_
x
x
)(. +r)D
N
(.) d.
=
_

x
)(. +r)D
N
(.) d. +
_
x
)(. +r)D
N
(.) d.
(@)
_

x
)(n +r 2)D
N
(n 2) dn +
_
x
)(. +r)D
N
(.) d.
()
_

x
) (n)D
N
(n) dn +
_
x
) (.)D
N
(.) d.
=
_

) (.)D
N
(.) d.
()

_
lim
z0
) (.) + lim
z,0
) (.)
_
()

_
lim
z0
)(. +r) + lim
z,0
)(. +r)
_
(c)

_
lim
yx
)(j) + lim
y,x
)(j)
_
.
Now divide both sides by 2 to get equation (10B.4).
Here, () is because D
N
is even (i.e. D
N
(:) = D
N
(:) for all : 1).
Both equalities marked (c) are the change of variables . := j r (so that
j = . + r). Likewise, equality (@) is the change of variables n := . + 2 (so
that . = n 2). Both () and () use the denition of

) , and () also uses
the fact that D
N
is 2-periodic, so that D
N
(n 2) = D
N
(n) for all n
[r, ]. Finally, () is by Lemma 10B.3 applied to

) (which is continuous
and semidierentiable at 0 because ) is continuous and semidierentiable at
r). 2
Proof of Theorem 8A.1(b). Let r [, ], and suppose ) is continuous and
dierentiable at r. Then
lim
N
0
+
N
n=1
n
C
n
(r) +
N
n=1
1
n
S
n
(r)
()
lim
N
D
N
)(r)
()
)(r),
as desired. Here, () is by Lemma 10B.1 and () is by Lemma 10B.4. 2
Remarks. (a) Note that we have actually proved a slightly stronger result
than Theorem 8A.1(b). If ) is discontinuous, but semidierentiable at r, then
Lemmas 10B.1 and 10B.4 together imply that
lim
N
0
+
N
n=1
n
C
n
(r) +
N
n=1
1
n
S
n
(r) =
1
2
_
lim
yx
)(j) + lim
y,x
)(j)
_
.
This is how the Pointwise Fourier Convergence Theorem is stated in some texts.
(b) For other good expositions of this material, see [CB87, 30-31, pp.87-92].
[Asm05, Thm. 1, p.30 of 2.2], [Pow99, 1.7, p.79], or [Bro89, Corollary 1.4.5,
p.16].
10C Uniform convergence
Prerequisites: 8A, 10A.
In this section, we will prove Theorem 8A.1(d). First we state a discrete ver-
sion of the Cauchy-Bunyakowski-Schwarz Inequality (Theorem 6B.5 on page 108).
Lemma 10C.1. Cauchy-Bunyakowski-Schwarz Inequality in /
2
(N)
Let (o
n
)
n=1
and (/
n
)
n=1
be two innite sequences of real numbers. Then
_

n=1
o
n
/
n
_
2
n=1
o
2
n
_
n=1
/
2
n
_
,
whenever these sums are nite.
Proof. Exercise 10C.1 Hint: imitate the proof of Theorem 6B.5 on page 108 2 E _
10C. Uniform convergence 205
Remark. For any innite sequences of real numbers a := (o
n
)
n=1
and b :=
(/
n
)
n=1
, we can dene a, b) :=
n=1
o
n
/
n
and |a|
2
:=
_
a, a) =
_
n=1
o
2
n
.
The set of all sequences a such that |a|
2
< is denoted /
2
(N). Lemma 10C.1
can then be reformulated as the statement: For all a, b /
2
(N), [a, b)[
|a|
2
|b|
2
.
Next, we will prove a result which relates the smoothness of the function )
to the asymptotic decay rate of its Fourier coecients.
Lemma 10C.2. Let ) : [, ] 1 be continuous, with )() = )(). Let
n=0
and 1
n
n=1
be the real Fourier coecients of ), as dened on page 161.
If ) is piecewise dierentiable on [, ], and )
t
L
2
[, ], then the sequences
n=0
and 1
n
n=1
converge to zero fast enough that
n=1
[
n
[ < and
n=1
[1
n
[ < .
Proof. If )
t
L
2
[, ], then we can compute its real Fourier coecients
t
n
n=0
and 1
t
n
n=1
.
Claim 1: For all n N,
n
= 1
t
n
,n and 1
n
=
t
n
,n.
Proof. By denition,
t
n
:=
1
)
t
(r) cos(nr) dr
(p)
1
)(r) cos(nr)
x=
x=
+
1
)(r) nsin(nr) dr
(c)
(1)
n
_
)() )()
_
+ n1
n
()
n1
n
.
Likewise, 1
t
n
:=
1
)
t
(r) sin(nr) dr
(p)
1
)(r) sin(nr)
x=
x=
)(r) ncos(nr) dr
(s)
_
0 0
_
n
n
= n
n
.
Here, (p) is integration by parts, (c) is because cos(n) = (1)
n
=
cos(n), () is because )() = )(), and (s) is because sin(n) = 0 =
sin(n).
Thus, 1
t
n
= n
n
and
t
n
= n1
n
; hence
n
= 1
t
n
,n and 1
n
=
t
n
,n.
Claim 1
Let 1 :=
n=1
1
n
2
(a nite value). Then
_

n=1
[
n
[
_
2
()
_

n=1
1
n
[1
t
n
[
_
2
(CBS)
_

n=1
1
n
2
_
n=1
[1
t
n
[
2
_
= 1
n=1
[1
t
n
[
2
(B)
1
_
_
)
t
_
_
2
2
<
()
.
Here, () is by Claim 1, (CBS) is by Lemma 10C.1, and (B) is by Bessels
inequality (Theorem 10A.1 on page 195). Finally, () is because )
t
L
2
[, ]
by hypothesis. It follows that
n=1
[
n
[ < . The proof that
n=1
[1
n
[ < is
similar. 2
Proof of Theorem 8A.1(d). If ) : [, ] 1 is continuous and piecewise
dierentiable, )
t
L
2
[, ], and )() = )(), then Lemma 10C.2 implies
that
n=1
[
n
[ +
n=1
[1
n
[ < . But then Theorem 8A.1(c) says that the
Fourier series of ) converges uniformly. (Theorem 8A.1(c), in turn, is a direct
consequence of the Weierstrass ` test, Proposition 6E.13 on page 129.). 2
Remarks. (a) For other treatments of the material in this section, see [CB87,
34-35, pp.105-109] or [Asm05, Thm. 3, p.90 of 2.9].
(b) The connection between smoothness of ) and the asymptotic decay of
its Fourier coecients is a recurring theme in harmonic analysis. In general, the
smoother a function is, the faster its Fourier coecients decay to zero. The
weakest statement of this kind is the Riemann-Lebesgue Lemma (Corollary 10A.3
on page 197), which says that if ) is merely in 1
2
, then its Fourier coecients
must converge to zero although perhaps very slowly. (In the context of Fourier
transforms of functions on 1, the corresponding statement is Theorem 19B.1 on
page 492). If ) is slightly smoother specically, if ) is absolutely continuous
or if ) has bounded variation then its Fourier coecients decay to zero with
speed comparable to the sequence 1,
1
2
,
1
3
, . . . ,
1
n
, . . .; see [Kat76, Thm.4.3 and
4.5, pp.24-25]. If ) is dierentiable, then Lemma 10C.2 says that its Fourier
coecients must decay fast enough that the sums

n=1
[
n
[ and

n=1
[1
n
[
converge. (For Fourier transforms of functions on 1, the corresponding result
is Theorem 19B.7 on page 496.) More generally, if ) is / times dierentiable
on [, ], then its Fourier coecients must decay fast enough that the sums
10D. 1
2
convergence 207
n=1
n
k1
[
n
[ and
n=1
n
k1
[1
n
[ converge; see [Kat76, Thm.4.3, p.24]. Finally, if
) is analytic
1
, on [, ], then its Fourier coecients must decay exponentially
quickly to zero; that is, for small enough : 0, we have lim
n
:
n
[
n
[ = 0 and
lim
n
:
n
[1
n
[ = 0 (see Proposition 18E.3 on page 458).
At the other extreme, what about a sequence of Fourier coecients which
does not satisfy the Riemann-Lebesgue lemma that is, which does not converge
to zero? This corresponds to the Fourier series of an object which is more
singular than any function can be: a Laurent distribution or a measure on
[, ], which can have innitely dense concentrations of mass at some points.
See [Kat76, 1.7, pp.34-46] or [Fol84, 8.5 and 8.8 on p.258 and p.281].
10D L
2
convergence
Prerequisites: 6B.
In this section, we will prove Theorem 8A.1(a) (concerning the 1
2
conver-
gence of Fourier series). For any / N, let (
k
per
[, ] be the set of func-
tions ) which are / times continuously dierentiable on [, ], and such that
)() = )(), )
t
() = )
t
(), )
tt
() = )
tt
(), . . ., and )
(k)
() = )
(k)
().
If ) (
1
per
[, ], then Theorem 8A.1(d) (which we just proved in 10C) says
the Fourier series of ) converges uniformly. Then Corollary 6E.11(b)[i] (on page
127) immediately implies that the Fourier series of ) converges in 1
2
norm. Un-
fortunately, this argument does not work for most functions in L
2
[, ], which
are not in (
1
per
[, ]. Our strategy will be to show that (
1
per
[, ] is dense
in L
2
[, ]; thus, the 1
2
convergence of Fourier series in (
1
per
[, ] can be
leveraged to obtain 1
2
convergence for all functions in L
2
[, ].
A subset ( L
2
[, ] is dense in L
2
[, ] if, for any ) L
2
[, ], and
any c 0, we can nd some p ( such that |) p|
2
< c. In other words,
any element of L
2
[, ] can be approximated arbitrarily closely
2
by elements
of (. Aside from Theorem 8A.1(a), the major goal of this section is to prove the
following result:
Theorem 10D.1. For all / N, the subset (
k
per
[, ] is dense in L
2
[, ].
To achieve this goal, we must rst examine the structure of integrable func-
tions, and develop some useful machinery involving convolutions and molli-
ers. Then we will prove Theorem 10D.1. Once Theorem 10D.1 is established,
1
See page 570 of Appendix 0H
2
In the same way, the set of rational numbers is dense in the set 1 of real numbers:
any real number can be approximated arbitrarily closely by rational numbers. Indeed, we
exploit this fact every time we approximate a real number using a decimal expansion e.g.
3.141592653 =
3141592653
100000000
.
we will prove Theorem 8A.1(a) by using Theorem 8A.1(d) and the triangle in-
equality.
10D(i) Integrable functions and step functions in L
2
[, ]
Prerequisites: 6B, 6E(i).
We have dened L
2
[, ] to be the set of integrable functions ) : [, ]
1 such that
_
[)(r)[
2
dr < . But what exactly does integrable mean? To ex-
plain this, let Step[, ] be the set of all step functions on [, ] (see 8B(ii)
on page 164 for the denition of step functions). If ) : [, ] 1 is any
bounded function, then we can approximate ) using step functions in a natural
way. First, let := = j
0
< j
1
< j
2
< j
3
< < j
M1
< j
M
= be
some nite mesh of points in [, ]. For all n N, let o
n
:= inf
y
n1
xyn
)(r)
and o
n
:= sup
y
n1
xyn
)(r). Then dene step functions o
: [, ] 1 and
o
: [, ] 1 by
o
(r) :=
_
_
o
1
if r j
1
;
o
2
if j
1
< r j
2
;
.
.
.
o
m
if j
m1
< r j
m
;
.
.
.
o
M
if j
M1
< r .
and o
(r) :=
_
_
o
1
if r j
1
;
o
2
if j
1
< r j
2
;
.
.
.
o
m
if j
m1
< r j
m
;
.
.
.
o
M
if j
M1
< r .
It is easy to compute the integrals of o
and o
:
_

(r) dr =
N
n=1
o
n
[j
n
j
n1
[ and
_

(r) dr =
N
n=1
o
n
[j
n
j
n1
[ .
(You may recognize these as upper and lower Riemann sums of )). If the mesh
j
0
, j
1
, j
2
,. . ., j
M
is dense enough in [, ], so that o
and o
are good
approximations of ), then we might expect
_
(r) dr and
_
(r) dr
to be good approximations of
_
)(r) dr (if the integral of ) is well-dened).

Furthermore, it is clear from their denitions that o
(r) )(r) o
(r) for all

10D. 1
2
convergence 209
r [, ]; thus we would expect
_

(r) dr
_

)(r) dr
_

(r) dr,
whenever the integral
_
)(r) dr exists. Let Y be the set of all nite meshes

of points in [, ]. Formally:
Y :=
_
[, ] ; = j
0
, j
1
, j
2
, . . . , j
M
, for some ` N
and = j
0
< j
1
< j
2
< < j
M
=
_
.
We dene the lower and upper semi-integrals of ):
1()) := sup
Y
_

(r) dr (10D.1)
= sup
_
N
n=1
_
[j
n
j
n1
[ inf
y
n1
xyn
)(r)
_
; ` N, = j
0
< j
1
< < j
M
=
_
.
1()) := inf
Y
_

(r) dr (10D.2)
= inf
_
N
n=1
_
[j
n
j
n1
[ sup
y
n1
xyn
)(r)
_
; ` N, = j
0
< j
1
< < j
M
=
_
.
It is easy to see that 1()) 1()) (Exercise 10D.1 Check this.). Indeed, if E _
) is a suciently pathological function, then we may have 1()) < 1()). If
1()) = 1()), then we say that ) is (Riemann) integrable, and we dene the
(Riemann) integral of ):
_

)(r) dr := 1()) = 1()).

For example:
Any bounded, piecewise continuous function on [, ] is Riemann-integrable.
Any continuous function on [, ] is Riemann-integrable.
Any step function on [, ] is Riemann-integrable.
If ) : [, ] 1 is not bounded, then the denitions of o
and/or o
make
no sense (because at least one of them is dened as or on some
interval). Thus, at least one of the expressions (10D.1) and (10D.2) is not well-
dened if ) is unbounded. In this case, for any N, we dene the truncated
functions )
+
N
: [, ] [0, ] and )
N
: [, ] [, 0] as follows
)
+
N
(r) :=
_
_
_
0 if )(r) 0;
)(r) if 0 )(r) ;
if )(r).
and )
N
(r) :=
_
_
_
if )(r) ;
)(r) if )(r) 0;
0 if 0 )(r).
The functions )
+
N
and )
N
are clearly bounded, so their Riemann integrals are
potentially well-dened. If )
+
N
and )
N
are integrable for all N, then we say
that ) is (Riemann) measurable. We then dene
_

)(r) dr := lim
N
_

)
+
N
(r) dr + lim
N
_

N
(r) dr.
If both these limits are nite, then
_
)(r) dr is well-dened, and we say that the

unbounded function ) is (Riemann) integrable. The set of all integrable func-
tions (bounded or unbounded) is denoted L
1
[, ], and for any ) L
1
[, ],
we dene
|)|
1
=
_

[)(r)[ dr.
We can now dene L
2
[, ]:
L
2
[, ] :=
_
all measurable functions ) : [, ] 1 such that
)
2
: [, ] 1 is integrable i.e.
_
[)(r)[
2
dr <
_
.
Proposition 10D.2. Step[, ] is dense in L
2
[, ].
Proof. Let ) L
2
[, ] and let c 0. We want to nd some o Step[, ]
such that |) o|
2
< c.
First suppose that ) is bounded. Since )
2
is integrable, we know that 1()
2
) =
_
)
2
(r) dr, where 1()
2
) is dened by expression (10D.1). Thus, we can
nd some step function o
0
Step[, ] such that 0 o
0
(r) )
2
(r) for all
r [, ], and such that
0
_

)(r)
2
dr
_

o
0
(r) dr < c. (10D.3)
Dene the step function o Step[, ] by o(r) := sign[)(r)]
_
o
0
(r). Thus,
o
2
(r) = o
0
(r), and the sign of o agrees with that of ) everywhere. Observe
that
() o)
2
=
) o
) +o
() o)() +o) =
) o
) +o
()
2
o
2
)
()
)
2
o
2
= )
2
o
0
. (10D.4)
10D. 1
2
convergence 211
Here, () is because 0 <
fS
f+S
< 1 (because for all r, either )(r) o(r) 0,
or 0 o(r) )(r)), while )
2
o
2
0 (because )
2
o
0
= o
2
). Thus,
0 |) o|
2
2
=
_

()(r) o(r))
2
dr
()
_

)(r)
2
o
0
(r) dr
=
_

)(r)
2
dr
_

o
0
(r) dr <
()
c,
where () is by eqn.(10D.4) and () is by eqn.(10D.3).
This works for any c 0; thus the set Step[, ] is dense in the space of
bounded elements of L
2
[, ].
The case when ) is unbounded is Exercise 10D.2 (Hint: approximate ) with E _
bounded functions). 2
Remark 10D.3: To avoid developing a considerable amount of technical back-
ground, we have dened L
2
[, ] using the Riemann integral. The true deni-
tion of L
2
[, ] involves the more powerful and versatile Lebesgue integral. (See
6C(ii) on page 110 for an earlier discussion of Lebesgue integration). The de-
nition of the Lebesgue integral is similar to the Riemann integral, but instead of
approximating ) using step functions, we use simple functions. A simple function
is a piecewise-constant function, like a step function, but instead of open inter-
vals, the pieces of a simple function are Borel-measurable subsets of [, ]. A
Borel measurable subset is a countable union of countable intersections of count-
able unions of countable intersections of .... of countable unions/intersections
of open and/or closed subsets of [, ]. In particular, any interval is Borel
measurable (so any step function is a simple function), but Borel measurable
subsets can be very complicated indeed. Thus, simple functions are capable of
approximating even pathological, wildly discontinuous functions on [, ], so
that the Lebesgue integral can be evaluated even on such crazy functions. The
set of Lebesgue-integrable functions is thus much larger than the set of Riemann-
integrable functions. Every Riemann-integrable function is Lebesgue integrable
(and its Lebesgue integral is the same as its Riemann integral), but not vice
versa.
The analogy of Proposition 10D.2 is still true if we dene L
2
[, ] using
Lebesgue-integrable functions, and if we replace Step[, ] with the set of all
simple functions. The other results in this section can also be extended to the
Lebesgue version of L
2
[, ], but at the cost of considerable technical complex-
ity.
Let ) : [, ] 1. Let )
: [2, 2] 1 be the 2-periodic exten-


f
2
2
f
0
Figure 10D.1: f
: [2, 2] is the 2-periodic extension of f : [, ] 1.

sion of ), dened:
)
(r) :=
_
_
_
)(r + 2) if 2 r < ;
)(r) if r ;
)(r 2) if < r 2.
(See Figure 10D.1)
(Observe that )
is continuous if and only if ) is continous and )() = )()).

For any t 1, dene the function )
t
: [, ] 1 by )
t
(r) = )
(r t).
(For example, the function

) dened on page 203 could be written:

) = )
x
; see
Figure 10B.3 on page 203).
Lemma 10D.4. Let ) L
2
[, ]. Then ) = L
2
lim
t0
)
t
.
Proof. We will employ a classic strategy in real analysis: rst prove the
result for some nice class of functions, and then prove it for all functions by
approximating them with these nice functions. In this case, the nice functions
are the step functions.
Claim 1: Let o Step[, ]. Then o = L
2
lim
t0
o
t
.
Proof. Exercise 10D.3
Claim 1
E _
Now, let ) L
2
[, ], and let c 0. Proposition 10D.2 says there is some
o Step[, ] such that
|o )|
2
<
c
3
. (10D.5)
Claim 2: For all t 1,
_
_
_o
t
)
t
_
_
_
2
= |o )|
2
.
Claim 2
E _
Now, using Claim 1, nd 0 such that, if [t[ < , then
_
_
_
_
o o
t
_
_
_
_
2
<
c
3
. (10D.6)
10D. 1
2
convergence 213
Then
_
_
_
_
) )
t
_
_
_
_
2
=
_
_
_
_
) o + o o
t
+ o
t
)
t
_
_
_
_
2
(z)
|) o|
2
+
_
_
_
_
o o
t
_
_
_
_
2
+
_
_
_
_
o
t
)
t
_
_
_
_
2
()
c
3
+
c
3
+
c
3
= c.
Here () is the triangle inequality, and () is by equations (10D.5) and (10D.6)
and Claim 2.
This works for all c 0; thus, ) = L
2
lim
t0
)
t
. 2
There is one nal technical result we will need about L
2
[, ]. If )
1
, )
2
, . . . , )
N

L
2
[, ] and :
1
, :
2
, . . . , :
N
1 are real numbers, then the triangle inequality
implies that
|:
1
)
1
+:
2
)
2
+ +:
N
)
N
|
2
[:
1
[ |)
1
|
2
+[:
2
[ |)
2
|
2
+ +[:
N
[ |)
N
|
2
.
This is a special case of Minkowskis inequality. The next result says that the
same inequality holds if we sum together a continuum of functions.
Theorem 10D.5. (Minkowskis inequality for integrals)
Let o < /, and for all t [o, /], let )
t
L
2
[, ]. Dene 1 : [o, /][, ] 1
by 1(t, r) = )
t
(r) for all (t, r) [o, /] [, ], and suppose that the family
)
t
t[a,b]
is such that the function 1 is integrable on [o, /] [, ]. Let 1 :
[o, /] 1 be some other integrable function, and dene G : [, ] 1 by
G(r) :=
_
b
a
1(t))
t
(r) dt, for all r [, ].
Then G L
2
[, ], and
|G|
2

_
b
a
[1(t)[ |)
t
|
2
dt,
In particular, if |)
t
|
2
< ` for all t [o, /], then |G|
2
` |1|
1
, where
|1|
1
:=
_
b
a
[1(t)[ dt.
Proof. See [Fol84, Thm 6.18, p.186]. 2
10D(ii) Convolutions and molliers
Prerequisites: 10D(i). Recommended: 17B.
Let ), p : [, ] 1 be two integrable functions. Let p
: [2, 2] 1
be the 2-periodic extension of p (see Figure 10D.1 on page 212). The (2-
periodic) convolution of ) and p is the function ) p : [, ] 1 dened
by
) p(r) :=
1
2
_

)(j) p

Convolution is an important and versatile mathematical operation, which ap-
pears frequently in harmonic analysis, probability theory, and the study of partial
dierential equations. We will encounter it again in Chapter 17, in the context
of impulse-response solutions to boundary value problems. In this subsection,
we will develop the theory of convolutions on [, ]. We will actually develop
slightly more than we need in order to prove Theorems 10D.1 and 8A.1(a). Re-
sults which are not logically required for the proofs of Theorems 10D.1 and
8A.1(a) are marked with the margin symbol (Optional) and can be skipped
on a rst reading; however, we feel that these results are interesting enough in
themselves to be worth including in the exposition.
Lemma 10D.6. (Properties of convolutions)
Let ), p : [, ] 1 be integrable functions. The convolution of ) and p has
the following properties:
(a) (Commutativity) ) p = p ).
(b) (Linearity) If / : [, ] 1 is another integrable function, then )
(p +/) = ) p +) / and () +p) / = ) / +p /.
(c) If ), p L
2
[, ], then ) p is bounded: for all r [, ], we have
[) p(r)[ |)|
2
|p|
2
. (In other words, |) p|
|)|
2
|p|
2
.) (Optional)
Proof. (a) is Exercise 10D.5 . To prove (b), let r [, ]. Then E _
) (p +/)(r) =
1
2
_

)(j)
_
p
(r j) +/
(r j)
_
dj
=
1
2
_

)(j)p
(r j) dj +
1
2
_

)(j)/
(r j) dj
= ) p(r) +) /(r).
(c) Let r [, ]. Dene / L
2
[, ] by /(j) := p
(r j). Then (Optional)

) p(r) =
1
2
_

)(j)p
(r j) dj =
1
2
_

)(j)/(j) dj = ), /) .
Thus, [) p(r)[ = [), /)[
(CBS)
|)|
2
|/|
2
, (10D.7)
10D. 1
2
convergence 215
where (CBS) is the Cauchy-Bunyakowski-Schwarz inequality (Theorem 6B.5
on page 108). But
|/|
2
2
=
1
2
_

/(j)
2
dj =
1
2
_

(r j)
2
dj
()
1
2
_
x
x+
p
(.)
2
d.
=
1
2
_
x+
x
p
(.)
2
d.
()
1
2
_

p(.)
2
d. = |p|
2
2
. (10D.8)
Here, () is the change of variables . = rj (so that d. = dj) and () is by
denition of the periodic extension p
of p.
Combining equations (10D.7) and (10D.8) we conclude that [) p(r)[ < |)|
2
|p|
2
, as claimed. 2
Remarks. (a) Proposition 17G.1 on page 409 provides an analog to Lemma
10D.6 for convolutions on 1
D
.
(b) There is also an interesting relationship between convolution and complex
Fourier coecients; see Lemma 18F.3 on page 463.
Elements of L
1
[, ] and L
2
[, ] need not be dierentiable, or even con-
tinuous (indeed, some of these functions are discontinuous almost everywhere).
But the convolution of even two highly discontinuous elements of L
2
[, ] will
be a continuous function. Furthermore, convolution with a smooth function has
a powerful smoothing eect on even the nastiest elements of L
1
[, ].
Lemma 10D.7. Let ), p L
1
[, ].
(a) ) p() = ) p().
(b) If ) L
1
[, ] and p is continuous with p() = p(), then ) p is
continuous.
(c) If ), p L
2
[, ], then ) p is continuous. (Optional)
(d) If p is dierentiable on [, ], then ) p is also dierentiable on [, ],
and () p)
t
= ) (p
t
).
(e) If p (
1
[, ], then ) p (
1
per
[, ].
(f ) For any / N, if p (
k
[, ], then
3
) p (
k
per
[, ]. Furthermore, (Optional)
() p)
t
= ) p
t
, () p)
tt
= ) p
tt
, . . ., and () p)
(k)
= ) p
(k)
.
3
See page 207 for the denition of (
k
per
[, ].
Proof. (a) ) p() =
1
2
_

)(j)p
( j) dj
()
1
2
_

)(j)p
( j
2) dj =
1
2
_

)(j)p
( j) dj = ) p(). Here, () is because p
is
2-periodic.
(b) Fix r [, ] and let c 0. We must nd some 0 such that, for any
r
1
[, ], if [r r
1
[ < then [) p(r) ) p(r
1
)[ < c. But
) p(r) ) p(r
1
) =
1
2
_

)(j) p
(r j) dj
1
2
_

)(j) p
(r
1
j) dj
=
1
2
_

)(j) p
(r j) )(j)p
(r
1
j) dj
=
1
2
_

)(j)
_
p
(r j) p
(r
1
j)
_
dj. (10D.9)
Since p is continuous on [, ] and p() = p(), it follows that p
is contin-
uous on [2, 2]; since [2, 2] is a closed and bounded set, it then follows
that p
is uniformly continuous on [2, 2]. That is, there is some 0 such

that, for any ., .
1
[2, 2],
if [. .
1
[ < , then [p
(.) p
(.
1
)[ <
2c
|)|
1
. (10D.10)
Now, suppose [r r
1
[ < . Then
[) p(r) ) p(r
1
)[
()
1
2
_

)(j)
_
p
(r j) p
(r
1
j)
_
dj
(z)
1
2
_

[)(j)[
(r j) p
(r
1
j)
dj
()
1
2
_

[)(j)[
2c
|)|
1
dj =
c
|)|
1
[)(j)[ dj
=
c
|)|
1
|)|
1
= c.
Here, () is by eqn.(10D.9), () is the triangle inequality for integrals, and ()
is by eqn.(10D.10), because [(rj) (r
1
j)[ < for all j [, ], because
[r r
1
[ < .
Thus, if [r r
1
[ < then [) p(r) ) p(r
1
)[ < c. This argument works for
any c 0 and r [, ]. Thus, ) p is continuous, as desired.
(c) Fix r [, ] and let c 0. We must nd some 0 such that, for any (Optional)
t [, ], if [t[ < then [) p(r) ) p(r t)[ < c. But
) p(r t) =
1
2
_

)(j) p
(r t j) dj
10D. 1
2
convergence 217
=
1
2
_

)(j) (p
t
)
(r j) dj = ) p
t
(r).
Thus, ) p(r) ) p(r t) = ) p(r) ) p
t
(r)
()
) (p p
t
)(r),
so
) p(r) ) p(r t)
) (p p
t
)(r)
()
|)|
2

_
_
_
_
p p
t
_
_
_
_
2
. (10D.11)
Here, () is by Lemma 10D.6(b) and () is by Lemma 10D.6(c).
However, Lemma 10D.4 on page 212 says that p = L
2
lim
t0
p
t
. Thus, there
exists some 0 such that, if [t[ < , then
_
_
_p p
t
_
_
_
2
< c, |)|
2
. Thus, if
[t[ < , then
) p(r) ) p(r t)

()
|)|
2

_
_
_
_
p p
t
_
_
_
_
2
|)|
2

c
|)|
2
= c.
where () is by eqn.(10D.11). This argument works for any c 0 and r
[, ]. Thus, ) p is continuous, as desired.
(d) We have
2 () p)
t
(r) = 2
x
() p)(r) =
x
_

)(j) p(r j) dj
()
_

)(j)
x
p(r j) dj =
_

)(j) p
t
(r j) dj
= 2 ) (p
t
)(r).
Here, () is by Proposition 0G.1 on page 567.
(e) Follows immediately from (a), (b) and (d).
(f ) is Exercise 10D.6 Hint: Use proof by induction, along with parts (b) and E _
(d). 2
Remarks. (a) Proposition 17G.2 on page 410 provides an analog to Lemma
10D.7 for convolutions on 1
D
.
(b) (for algebraists) Let (
per
[, ] be the set of all continuous functions
) : [, ] 1 such that )() = )(). Then Lemmas 10D.6(a,b) and
10D.7(a,b) imply that (
per
[, ] is a commutative ring, where functions are
added pointwise, and where the convolution operator plays the role of mul-
tiplication. Furthermore, Lemma 10D.7(f) says that, for all / N, the set
(
k
per
[, ] is an ideal of the ring (
per
[, ]. Note that this ring does not have

3
4
Area/2 = 0.2
Area/2 = 0.7 Area/2 = 0.9
Area/2 = 0.5
Figure 10D.2: An approximation of identity on [, ]. Here, > 0 is xed, and
lim
n
1
2
_

n(x) dx = 1.
a multiplicative identity element. However, it does have approximations of
identity, as we shall now see.
For all n N, let
n
: [, ] 1 be a nonnegative function. The se-
quence
n
n=1
is called an approximation of identity if it has the following
properties:
(AI1)
1
2
_

n
(j) dj = 1 for all n N.
(AI2) For any c 0, lim
n
1
2
_

n
(r) dr = 1. (See Figure 10D.2).
Example 10D.8. Let : [, ] 1 be any nonnegative function with
1
2
_

(r) dr = 1. For all n N, dene

n
: [, ] 1 by
n
(r) :=
_
_
_
0 if r < ,n;
n(nr) if ,n r ,n;
0 if ,n < r.
(see Figure 10D.3).
Then
n
n=1
is a 2-periodic approximation of identity (Exercise 10D.7). E _
The term approximation of identity is due to the following result:

10D. 1
2
convergence 219
1
2
=
1
0

1/3
6
3
0
1
/
4
8
4
0
1/2
4
2
0
Figure 10D.3: Example 10D.8.
Proposition 10D.9. Let
n
n=1
be an approximation of identity. Let ) :
[, ] 1 be some integrable function.
(a) If ) L
2
[, ] then ) = L
2
lim
n
n
).
(b) If r (, ) and ) is continuous at r, then )(r) = lim
n
n
)(r). (Optional)
Proof. (a) Fix c 0. We must nd N such that, for all n ,
|)
n
)|
2
< c. First, nd some 0 which is small enough that
(2 |)|
2
+ 1) < c. (10D.12)
Now, Lemma 10D.4 on page 212 says that there is some 0 such that,
For any t (, )
_
_
_
_
) )
t
_
_
_
_
2
< . (10D.13)
Next, property (AI2) says there is some N such that,
For all n , 1 <
1
2
_

n
(j) dj 1. (10D.14)
Now, for any r [, ] and n N, observe that
)(r) = )(r) 1
()
)(r)
1
2
_

n
(j) dj =
1
2
_

)(r)
n
(j) dj
(10D.15)
where () is by property (AI1). Thus, for all r [, ] and n N,
)(r)
n
)(r)
()
1
2
_

)(r)
n
(t) dt
1
2
_

(r t)
n
(t) dt
=
1
2
_

_
)(r) )
(r t)
_

n
(t) dt
=
1
2
_

_
)(r) )
t
(r)
_

n
(t) dt
=
1
2
_

n
(t) 1
t
(r) dt,
where, () is by eqn.(10D.15), and where, for all t [, ] we dene the
function 1
t
: [, ] 1 by 1
t
(r) := )(r) )
t
(r) for all r [, ]. Thus
|)
n
)|
2
=
_
_
_
_
1
2
_

n
(t) 1
t
dt
_
_
_
_
2
(M)
1
2
_

[
n
(t)[ |1
t
|
2
dt
=
1
2
_

n
(t) |1
t
|
2
dt +
1
2
_

n
(t) |1
t
|
2
dt +
1
2
_

n
(t) |1
t
|
2
dt
()
|)|
2
n
(t) dt +
1
2
_

n
(t) |1
t
|
2
dt +
|)|
2
n
(t) dt
()
|)|
2
__

n
(t) dt +
_

n
(t) dt
_
+

2
_

n
(t) dt
<
()
2 |)|
2
+ 1 = (2 |)|
2
+ 1)
()
c.
Here, (M) is Minkowskis inequality for integrals (Theorem 10D.5 on page 213).
Next, () is because
|1
t
|
2
=
_
_
_
_
) )
t
_
_
_
_
2
(z)
|)|
2
+
_
_
_
_
)
t
_
_
_
_
2
= |)|
2
+|)|
2
= 2 |)|
2
.
Next, () is because |1
t
|
2
< for all t (, ) by equation (10D.13). In-
equality () is because equation (10D.14) says 1 <
1
2
_

n
(t) dt 1;
thus, we must have
1
2
_

n
(t) dt +
1
2
_

n
(t) dt < . Finally, () is by
eqn.(10D.12).
This argument works for any c 0. We conclude that ) = L
2
lim
n
n
).
(b) is Exercise 10D.8 . 2 E _
For any / N, a (
k
-mollier is an approximation of identity
n
n=1
such
that
n
(
k
per
[, ] for all n N. Lemma 10D.7(f) says that you can mollify
some initially pathological function ) into a nice smooth approximation by con-
volving it with
n
. Our last task in this section is to show how to construct such
a (
k
-mollier.
10D. 1
2
convergence 221
Lemma 10D.10. Let (
k
[, ] be such that
1
2
_

(r) dr = 1, and
such that
() =
t
() =
tt
() = =
(k)
() = 0
and () =
t
() =
tt
() = =
(k)
() = 0.
Dene
n
n=1
as in Example 10D.8. Then
n
n=1
is a (
k
-mollier.
Proof. Exercise 10D.9 2 E _
Example 10D.11. Let p(r) = (r+)
k+1
(r)
k+1
, let G =
1
2
_
p(r) dr, and

then let (r) := p(r),G. Then satises the hypotheses of Lemma 10D.10
(Exercise 10D.10). E _
Remark. For more information about convolutions and molliers, see [Fol84,
8.2, pp.230-237] or [WZ77, Chap.9, pp.145-160].
10D(iii) Proof of Theorems 8A.1(a) and 10D.1.
Prerequisites: 8A, 10A, 10D(ii).
Proof of Theorem 10D.1. Let
n
n=1
(
k
per
[, ] be the (
k
-mollier from
Lemma 10D.10. Then Proposition 10D.9(a) says that ) = L
2
lim
n
n
).
Thus, for any c 0, we can nd some n N such that |)
n
)|
2
< c.
Furthermore,
n
(
k
[, ], so Lemma 10D.7(f) says that
n
) (
k
per
[, ],
for all n N. 2
The proof of Theorem 8A.1(a) now follows a standard strategy in analysis:
approximate the function ) with a nice function

), establish convergence for
the Fourier series of

) rst, and then use the triangle inequality to leverage this
into convergence for the Fourier series of ).
Proof of Theorem 8A.1(a). Let ) L
2
[, ]. Fix c 0. Theorem 10D.1
says there exists some

) (
1
per
[, ] such that
_
_
_)

)
_
_
_
2
<
c
3
. (10D.16)
Let
n
n=0
and 1
n
n=1
be the real Fourier coecients for ), and let
n=0
and
1
n
n=1
be the real Fourier coecients for

). Let ) := )

),
and let
n
n=0
and 1
n
n=1
be the real Fourier coecients for ). Then for
all n N, we have
n
=
n
+

n
and 1
n
= 1
n
+

1
n
. (10D.17)
Also, for any N, we have
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
_
_
_
_
_
2
2
(z)
2
0
+
n=0
[
n
[
2
|C
n
|
2
2
+
n=1
[1
n
[
2
|C
n
|
2
2
()
2
0
+
n=0
[
n
[
2
2
+
n=1
[1
n
[
2
2
(B)
_
_
)
_
_
2
2
=
_
_
_)

)
_
_
_
2
2
<
()
_
c
3
_
2
. (10D.18)
Here, () is by the triangle inequality
4
, and () is because |C
n
|
2
2
=
1
2
= |S
n
|
2
2
for all n N (by Proposition 6D.2 on page 112). (B) is Bessels Inequality
(Example 10A.2 on page 196), and () is by eqn.(10D.16).
Now,

) (
1
per
[, ], so Theorem 8A.1(d) (which we proved in Section 10C)
says that
uniflim
N
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
_
=

).
Thus Corollary 6E.11(b)[i] on page 127 implies that
L
2
lim
N
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
_
=

).
Thus, there exists some N such that
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n

)
_
_
_
_
_
2
<
c
3
. (10D.19)
Thus,
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
)
_
_
_
_
_
2
()
_
_
_
_
_
(
0
+

0
) +
N
n=0
(
n
+

n
)C
n
+
N
n=1
(1
n
+

1
n
)S
n

) +

) )
_
_
_
_
_
2
4
Actually, this is an equality, because of the L
2
Pythagorean formula (equation (6F.1) on
page 131)
10D. 1
2
convergence 223
=
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
+

0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n

) +

) )
_
_
_
_
_
2
(z)
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n
_
_
_
_
_
2
+
_
_
_
_
_
0
+
N
n=0
n
C
n
+
N
n=1
1
n
S
n

)
_
_
_
_
_
2
+
_
_
_
) )
_
_
_
2
()
c
3
+
c
3
+
c
3
= c.
Here, () is by eqn.(10D.17), () is the Triangle inequality, and () is by
inequalities (10D.16), (10D.18) and (10D.19).
This argument works for any c 0. We conclude that
0
+
n=0
n
C
n
+
n=1
1
n
S
n
L2
). 2
Recall that a function ) : [, ] 1 is analytic if ) is innitely dier-
entiable, and the Taylor expansion of ) around any r [, ] has a nonzero
radius of convergence.
5
Let (
per
[, ] be the set of all analytic functions ) on
[, ] such that )() = )(), and )
(k)
() = )
(k)
() for all / N. For
example, the functions sin and cos are in (
per
[, ]. Elements of (
per
[, ] are
some of the nicest possible functions on [, ]. On the other hand, arbitrary
elements of L
2
[, ] can by quite nasty (i.e. nondierentiable, discontinuous).
Thus, the following result is quite striking.
Corollary 10D.12. (
per
[, ] is dense in L
2
[, ].
Proof. Theorem 8A.1(a) says that any function in L
2
[, ] can be ap-
proximated arbitrarily closely by a trigonometric polynomial of the form
0
+

N
n=1
n
C
n
+

N
n=1
1
n
S
n
. But all trigonometric polynomials are
in (
per
[, ] (because they are nite linear combinations of the functions
S
n
n
(r) := cos(nr), which are all in (
per
[, ]). Thus,
any function in L
2
[, ] can be approximated arbitrarily closely by an ele-
ment of (
per
[, ] in other words, (
per
[, ] is dense in L
2
[, ]. 2
Remarks. (a) Proposition 17G.3 on page 411 provides a pointwise version of
the Theorem 10D.1 for convolutional smoothing on 1
D
.
(b) For another proof of the 1
2
-convergence of real Fourier series, see [Bro89,
Theorems 1.5.4 (p.20) and 2.3.10 (p.35)]. For a proof of the 1
2
-convergence of
complex Fourier series (which is very similar), see [Kat76, I.5.5, p.29-30].
5
See Appendix 0H(i) on page 568.
224
IV BVP solutions via
eigenfunction expansions
A powerful and general method for solving linear PDEs is to represent the
solutions using eigenfunction expansions. Rather than rst deploying this idea
in full abstract generality, we will start by illustrating it in a variety of special
cases. We will gradually escalate the level of abstraction, so that the general
theory is almost obvious when it is nally stated explicitly.
The orthogonal trigonometric functions S
n
and C
n
in a Fourier series are
eigenfunctions of the Laplacian operator . Furthermore, the eigenfunctions S
n
and C
n
are particularly well-adapted to domains like the interval [0, ], the
square [0, ]
2
, or the cube [0, ]
3
, for two reasons:
The functions S
n
and C
n
and the domain [0, ]
k
are both easily expressed
in a Cartesian coordinate system.
The functions S
n
and C
n
satisfy desirable boundary conditions (e.g. ho-
mogeneous Dirichlet/Neumann) on the boundaries of the domain [0, ]
k
.
Thus, we can use S
n
and C
n
as building blocks to construct a solution to
a given partial dierential equation a solution which also satises specied
initial conditions and/or boundary conditions on [0, ]
k
. In particular, we will
use Fourier sine series to obtain homogeneous Dirichlet boundary conditions [by
Theorems 7A.1(d), 9A.3(d) and 9B.1(d)] , and Fourier cosine series to obtain
homogeneous Neumann boundary conditions [by Theorems 7A.4(d), 9A.3(e) and
9B.1(e)]. This basic strategy underlies all the solution methods developed in
Chapters 11 to 13.
When we consider other domains (e.g. disks, annuli, balls, etc.), the func-
tions C
n
and S
n
are no longer so well-adapted. In Chapter 14, we discover
that, in polar coordinates, the well-adapted eigenfunctions are combinations of
trigonometric functions (C
n
and S
n
) with another class of transcendental func-
tions called Bessel functions. This yields another orthogonal system of eigen-
functions. We can then represent most functions on the disks and annuli using
Fourier-Bessel expansions (analogous to Fourier series), and we can then mimic
the solution methods of Chapters 11 to 13.
225
Chapter 11
Boundary value problems on a
line segment
Mathematics is the music of reason. James Joseph Sylvester
Prerequisites: 7A, 5C.
This chapter concerns boundary value problems on the line segment [0, 1],
and provides solutions in the form of innite series involving the functions
S
n
(r) := sin
_
n
L
r
_
and C
n
(r) := cos
_
n
L
r
_
. For simplicity, we will assume
throughout the chapter that 1 = . Thus S
n
(r) = sin (nr) and C
n
(r) =
cos (nr). We will also assume that (through an appropriate choice of time units)
the physical constants in the various equations are all equal to one. Thus, the
heat equation becomes
t
n = n, the wave equation is
2
t
n = n, etc.
This does not limit the generality of our results. For example, faced with a
general heat equation of the form
t
n(r, t) = n for r [0, 1], (with ,= 1
and 1 ,= ) you can simply replace the coordinate r with a new space coordinate
j =

L
r and replace t with a new time coordinate : = t, to reformulate the
problem in a way compatible with the following methods.
11A The heat equation on a line segment
Prerequisites: 7B, 5B, 5C, 1B(i),0F. Recommended: 7C(v).
Proposition 11A.1. (Heat equation; homogeneous Dirichlet boundary)
Let X = [0, ], and let ) L
2
[0, ] be some function describing an initial heat
distribution. Suppose ) has Fourier Sine Series )(r)

L2
n=1
1
n
sin(nr), and
226 DRAFT Chapter 11. Boundary value problems on a line segment
dene the function n : X 1
,
1 by
n(r; t)

L2
n=1
1
n
sin(nr) exp
_
n
2
t
_
, for all r [0, ] and t 0.
Then n is the unique solution to the one-dimensional heat equation
t
n =
2
x
n,
with homogeneous Dirichlet boundary conditions
n(0; t) = n(; t) = 0, for all t 0.
and initial conditions: n(r; 0) = )(r), for all r [0, ].
Furthermore, the series dening n converges semiuniformly on X 1
+
.
Proof. Exercise 11A.1 Hint: E _
(a) Show that, when t = 0, the Fourier series of n(r; 0) agrees with that of )(r); hence
n(r; 0) = )(r).
(b) Show that, for all t 0,
n=1
n
2
1
n
c
n
2
t
< .
(c) For any T 0, apply Proposition 0F.1 on page 565 to conclude that
t
n(r; t)
unif
n=1
n
2
1
n
sin(nr) exp
_
n
2
t
_
unif
n(r; t) on [T, ).
(d) Observe that for any xed t 0,
n=1
1
n
c
n
2
t
< .
(e) Apply part (c) of Theorem 7A.1 on page 138 to show that the Fourier series of
n(r; t) converges uniformly for all t 0.
(f) Apply part (d) of Theorem 7A.1 on page 138 to conclude that n(0; t) = 0 = n(, t)
for all t 0.
(g) Apply Theorem 5D.8 on page 91 to show that this solution is unique. 2
Example 11A.2. Consider a metal rod of length , with initial tempera-
ture distribution )(r) = sinh(r) (where , 0 are constants), and
homogeneous Dirichlet boundary condition. Proposition 11A.1 tells us to get
the Fourier sine series for )(r). In Example 7A.3 on page 140, we computed
this to be
2 sinh()
n=1
n(1)
n+1
2
+n
2
sin(nr). The evolving temperature
distribution is therefore given:
n(r; t) =
2 sinh()
n=1
n(1)
n+1
2
+n
2
sin(nr) c
n
2
t
.
11A. The heat equation on a line segment 227
Proposition 11A.3. (Heat equation; homogeneous Neumann boundary)
Let X = [0, ], and let ) L
2
[0, ] be some function describing an initial heat
distribution. Suppose ) has Fourier Cosine Series )(r)

L2
n=0
n
cos(nr),
and dene the function n : X 1
,
1 by
n(r; t)

L2
n=0
n
cos(nr) exp
_
n
2
t
_
, for all r [0, ] and t 0.
Then n is the unique solution to the one-dimensional heat equation
t
n =
2
x
n,
with homogeneous Neumann boundary conditions
x
n(0; t) =
x
n(; t) = 0, for all t 0.
and initial conditions: n(r; 0) = )(r), for all r [0, ].
+
.
Proof. Setting t = 0, we get:
n(r; 0) =
n=1
n
cos(nr) exp
_
n
2
0
_
=
n=1
n
cos(nr) exp (0)
=
n=1
n
cos(nr) 1 =
n=1
n
cos(nr) = )(r),
so we have the desired initial conditions.
Let ` := max
nN
[
n
[. Then ` < (because ) L
2
).
Claim 1: For all t 0,
n=0
n
2

n
c
n
2
t
< .
Proof. Since ` = max
nN
[
n
[, we know that [
n
[ < ` for all n N. Thus,
n=0
n
2

n
c
n
2
t
n=0
n
2
c
n
2
t
= `
n=0
n
2
c
n
2
t
Hence,it suces to show that
n=0
n
2
c
n
2
t
< . To see this, let 1 = c
t
.
Then 1 1 (because t 0). Also, n
2
c
n
2
t
=
n
2
1
n
2
, for each n N. Thus,
n=1
n
2
c
n
2
t
=
n=1
n
2
1
n
2

m=1
:
1
m
(11A.1)
We must show that right-hand series in (11A.1) converges. We apply the
Ratio Test:
lim
m
m+1
E
m+1
m
E
m
= lim
m
:+ 1
:
1
m
1
m+1
= lim
m
1
1
< 1.
Hence the right-hand series in (11A.1) converges.
Claim 1
Claim 2: For any T 0, we have
x
n(r; t)
unif

n=1
n
n
sin(nr)
exp
_
n
2
t
_
on X [T, ), and also
2
x
n(r; t)
unif

n=1
n
2
n
cos(nr)
exp
_
n
2
t
_
on X [T, ).
Proof. This follows from Claim 1 and two applications of Proposition 0F.1
on page 565.
Claim 2
Claim 3: For any T 0, we have
t
n(r; t)
unif

n=1
n
2
n
cos(nr)
exp
_
n
2
t
_
on [T, ).
Proof.
t
n(r; t) =
t
n=1
n
cos(nr) exp
_
n
2
t
_
()
n=1
n
cos(nr)
t
exp
_
n
2
t
_
=
n=1
n
cos(nr) (n
2
) exp
_
n
2
t
_
,
where () is by Claim 1 and Proposition 0F.1 on page 565.
Claim 3
Combining Claims 2 and 3, we conclude that
t
n(r; t) = n(r; t).
Finally Claim 1 also implies that, for any t 0,
n=0
n
n
c
n
2
t
<
n=0
n
2

n
c
n
2
t
< .
Hence, Theorem 7A.4(d)[ii] on p.142 implies that n(r; t) satises homogeneous
Neumann boundary conditions for any t 0. (This can also be seen directly
via Claim 2).
Finally, Theorem 5D.8 on page 91 implies that this solution is unique. 2
11B. The wave equation on a line (the vibrating string) 229
Example 11A.4. Consider a metal rod of length , with initial temper-
ature distribution )(r) = cosh(r) and homogeneous Neumann boundary
condition. Proposition 11A.3 tells us to get the Fourier cosine series for
)(r). In Example 7A.6 on page 143, we computed this to be
sinh()
+
2 sinh()
n=1
(1)
n
cos(nr)
n
2
+ 1
. The evolving temperature distribution is there-
fore given:
n(r; t)

L2
sinh()
+
2 sinh()
n=1
(1)
n
cos(nr)
n
2
+ 1
c
n
2
t
.
Exercise 11A.2. Let 1 0 and let X := [0, 1]. Let 0 be a diusion constant, E _
and consider the general one-dimensional heat equation
t
n =
2
x
n. (11A.2)
(a) Generalize Proposition 11A.1 to nd the solution to eqn.(11A.2) on X satisfying
prescribed initial conditions and homogeneous Dirichlet boundary conditions.
(b) Generalize Proposition 11A.3 to nd the solution to eqn.(11A.2) on X satisfying
prescribed initial conditions and homogeneous Neumann boundary conditions.
In both cases, prove that your solution converges, satises the desired initial conditions
and boundary conditions, and satises eqn.(11A.2) (Hint: imitate the strategy suggested
in Exercise 11A.1)
Exercise 11A.3 Let X = [0, ], and let ) L
2
(X) be a function whose Fourier E _
sine series satises
n=1
n
2
[1
n
[ < . Imitate Proposition 11A.1, to nd a Fourier
series solution to the initial value problem for the one-dimensional free Schrodinger
equation
i
t
=
1
2

2
x
, (11A.3)
on X, with initial conditions
0
= ), and satisfying homogeneous Dirichlet
boundary conditions. Prove that your solution converges, satises the desired
initial conditions and boundary conditions, and satises eqn.(11A.3). (Hint:
imitate the strategy suggested in Exercise 11A.1).
11B The wave equation on a line (the vibrating string)
Prerequisites: 7B(i), 5B, 5C, 2B(i). Recommended: 17D(ii).
Imagine a piano string stretched tightly between two points. At equilibrium,
the string is perfectly at, but if we pluck or strike the string, it will vibrate,
meaning there will be a vertical displacement from equilibrium. Let X = [0, ]
represent the string, and for any point r X on the string and time t 0,
let n(r; t) be the vertical displacement of the string. Then n will obey the one-
dimensional wave equation:
2
t
n(r; t) = n(r; t). (11B.1)
However, since the string is xed at its endpoints, the function n will also exhibit
homogeneous Dirichlet boundary conditions
n(0; t) = n(; t) = 0 (for all t 0). (11B.2)
Proposition 11B.1. (Initial Position Problem for Vibrating String with xed
endpoints)
)
0
: X 1 be a function describing the initial displacement of the string.
Suppose )
0
has Fourier Sine Series )
0
(r)

L2
n=1
1
n
sin(nr), and dene the
function n : X 1
,
1 by
n(r; t)

L2
n=1
1
n
sin(nr) cos (nt) , for all r [0, ] and t 0. (11B.3)
Then n is the unique solution to the wave equation (11B.1), satisfying the
Dirichlet boundary conditions (11B.2), as well as
Initial Position: n(r, 0) = )
0
(r),
Initial Velocity:
t
n(r, 0) = 0,
_
for all r [0, ].
Proof. Exercise 11B.1 Hint: E _
(a) Prove the trigonometric identity sin(nr) cos(nt) =
1
2
_
sin (n(r t)) + sin (n(r +t))
_
.
(b) Use this identity to show that the Fourier sine series (11B.3) converges in 1
2
to
the dAlembert solution from Theorem 17D.8(a) on page 401.
(c) Apply Theorem 5D.11 on page 94 to show that this solution is unique. 2
Example 11B.2. Let )
0
(r) = sin(5r). Thus, 1
5
= 1 and 1
n
= 0 for all
n ,= 5. Proposition 11B.1 tells us that the corresponding solution to the wave
equation is n(r, t) = cos(5t) sin(5r). To see that n satises the wave equation,
note that, for any r [0, ] and t 0,
t
n(r, t) = 5 sin(5t) sin(5r) and 5 cos(5t) cos(5r) =
x
n(r, t);
Thus
2
t
n(r, t) = 25 cos(5t) sin(5r) = 25 cos(5t) cos(5r) =
2
x
n(r, t).
Figure 11B.1: (A) A harpstring at rest. (B) A harpstring being plucked. (C) The harp-
string vibrating. (D) A big hammer striking a xylophone. (E) The initial velocity of the
xylophone when struck.
Also n has the desired initial position because, for any r [0, ], we have
n(r; 0) = cos(0) sin(5r) = sin(5r) = )
0
(r), because cos(0) = 1.
Next, n has the desired initial velocity because for any r [0, ], we have
t
n(r; 0) = 5 sin(0) sin(5r) = 0, because sin(0) = 0.
Finally n satises homogeneous Dirichlet BC because, for any t 0, we
have n(0, t) = cos(5t) sin(0) = 0 and n(, t) = cos(5t) sin(5) = 0, because
sin(0) = 0 = sin(5).
Example 11B.3: (The plucked harp string)
A harpist places her ngers at the midpoint of a harp string and plucks it.
What is the formula describing the vibration of the string?
Solution: For simplicity, we imagine the string has length . The tight string
forms a straight line when at rest (Figure 11B.1A); the harpist plucks the
string by pulling it away from this resting position and then releasing it. At
the moment she releases it, the strings initial velocity is zero, and its initial
position is described by a tent function like the one in Example 7C.7 on
page 155
)
0
(r) =
_
r if 0 r

2
( r) if

2
< r .
(Figure 11B.1B)
where 0 is a constant describing the force with which she plucks the string
(and its resulting amplitude).
t = 0 t = 0.4 t = 0.8 t = 1.2 t = 1.6 /2
t = 2.0 t = 2.4 t = 2.8 t = 3.2 t = 3.6
Figure 11B.2: The plucked harpstring of Example 11B.3. From t = 0 to t = /2, the initially
triangular shape is blunted; at t = /2 it is totally at. From t = /2 to t = , the process
happens in reverse, only the triangle grows back upside down. At t = , the original triangle
reappears, upside down. Then the entire process happens in reverse, until the original triangle
reappears at t = 2.
The endpoints of the harp string are xed, so it vibrates with homogeneous
Dirichlet boundary conditions. Thus, Proposition 11B.1 tells us to nd the
Fourier sine series for )
0
. In Example 7C.7, we computed this to be:
)
0
L2
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sin(nr).
Thus, the resulting solution is: n(r; t)

L2
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sin(nr) cos(nt);
(See Figure 11B.2). This is not a very accurate model because we have not
accounted for energy loss due to friction. In a real harpstring, these per-
fectly triangular waveforms rapidly decay into gently curving waves depicted
in Figure 11B.1(C); these slowly settle down to a stationary state.
Proposition 11B.4. (Initial Velocity Problem for Vibrating String with xed
endpoints)
)
1
: X 1 be a function describing the initial velocity of the string.
Suppose )
1
1
(r)

L2
n=1
1
n
sin(nr), and dene the
function : X 1
,
1 by
(r; t)

L2
n=1
1
n
n
sin(nr) sin (nt) , for all r [0, ] and t 0. (11B.4)
Then is the unique solution to the wave equation (11B.1), satisfying the
Dirichlet boundary conditions (11B.2), as well as
Initial Position: (r, 0) = 0;
Initial Velocity:
t
(r, 0) = )
1
(r),
_
for all r [0, ].
(a) Prove the trigonometric identity sin(nr) sin(nt) =
1
2
_
cos (n(r +t)) cos (n(r t))
_
.
(b) Use this identity to show that the Fourier sine series (11B.4) converges in 1
2
to
the dAlembert solution from Theorem 17D.8(b) on page 401.
(c) Apply Theorem 5D.11 on page 94 to show that this solution is unique. 2
Example 11B.5. Let )
1
(r) = 3 sin(8r). Thus, 1
8
= 3 and 1
n
= 0 for all n ,= 8.
Proposition 11B.4 tells us that the corresponding solution to the wave equation
is n(r, t) =
3
8
sin(8t) sin(8r). To see that n satises the wave equation, note
that, for any r [0, ] and t 0,
t
n(r, t) = 3 sin(8t) cos(8r) and 3 cos(8t) sin(8r) =
x
n(r, t);
Thus
2
t
n(r, t) = 24 cos(8t) cos(8r) = 24 cos(8t) cos(8r) =
2
x
n(r, t).
Also n has the desired initial position because, for any r [0, ], we have
n(r; 0) =
3
8
sin(0) sin(8r) = 0, because sin(0) = 0.
Next, n has the desired initial velocity because for any r [0, ], we have
t
n(r; 0) =
3
8
8 cos(0) sin(8r) = 3 sin(8r) = )
1
(r), because cos(0) = 1.
Finally n satises homogeneous Dirichlet BC because, for any t 0, we have
n(0, t) =
3
8
sin(8t) sin(0) = 0 and n(, t) =
3
8
sin(8t) sin(8) = 0, because
sin(0) = 0 = sin(8).
Example 11B.6: (The Xylophone)
A musician strikes the midpoint of a xylophone bar with a broad, at hammer.
What is the formula describing the vibration of the string?
Solution: For simplicity, we imagine the bar has length and is xed at
its endpoints (actually most xylophones satisfy neither requirement). At the
moment when the hammer strikes it, the strings initial position is zero, and
its initial velocity is determined by the distribution of force imparted by the
hammer head. For simplicity, we will assume the hammer head has width ,2,
and hits the bar squarely at its midpoint (Figure 11B.1D). Thus, the initial
velocity is given by the function:
)
1
(r) =
_
if

4
r
3
4
0 otherwise
(Figure 11B.1E)
where 0 is a constant describing the force of the impact. Proposition
11B.4 tells us to nd the Fourier sine series for )
1
(r). From Example 7C.4 on
page 151, we know this to be
)
1
(r)

L2
2
_
_
sin(r) +
k=1
(1)
k
sin
_
(4/ 1)r
_
4/ 1
+
k=1
(1)
k
sin
_
(4/ + 1)r
_
4/ + 1
_
_
.
The resulting vibrational motion is therefore described by:
(r, t)

L2
2
_
_
sin(r) sin(t) +
k=1
(1)
k
sin
_
(4/ 1)r
_
sin
_
(4/ 1)t
_
(4/ 1)
2
+
k=1
(1)
k
sin
_
(4/ + 1)r
_
sin
_
(4/ + 1)t
_
(4/ + 1)
2
_
_
.
Exercise 11B.3 Let 1 0 and let X := [0, 1]. Let 0 be a parameter E _

describing wave velocity (determined by the strings tension, elasticity, density,
etc.), and consider the general one-dimensional wave equation
2
t
n =
2
2
x
n. (11B.5)
(a) Generalize Proposition 11B.1 to nd the solution to eqn.(11B.5) on X hav-
ing zero initial velocity and a prescribed initial position, and homogeneous
Dirichlet boundary conditions.
(b) Generalize Proposition 11B.4 to nd the solution to eqn.(11B.5) on X hav-
ing zero initial position and a prescribed initial velocity, and homogeneous
Dirichlet boundary conditions.
In both cases, prove that your solution converges, satises the desired initial
conditions and boundary conditions, and satises eqn.(11B.5) (Hint: imitate the
strategy suggested in Exercises 11B.1 and 11B.2.)
11C. The Poisson problem on a line segment 235
11C The Poisson problem on a line segment
Prerequisites: 7B, 5C, 1D. Recommended: 7C(v).
We can also use Fourier series to solve the one-dimensional Poisson problem
on a line segment. This is not usually a practical solution method, because we
already have a simple, complete solution to this problem using a double integral
(see Example 1D.1 on page 13). However, we include this result anyways, as a
simple illustration of Fourier techniques.
Proposition 11C.1. Let X = [0, ], and let : X 1 be some func-
tion, with semiuniformly convergent Fourier sine series: (r)

L2
n=1
Q
n
sin(nr).
Dene the function n : X 1 by
n(r)
unif
n=1
Q
n
n
2
sin(nr), for all r [0, ].
Then n is the unique solution to the Poisson equation n(r) = (r) satisfying
homogeneous Dirichlet boundary conditions: n(0) = n() = 0.
Proof. Exercise 11C.1 Hint: (a) Apply Proposition 0F.1 on page 565 twice to E _
show that n(r)
unif
n=1
Q
n
sin(nr) = (r), for all r int (X). (Hint: The Fourier
series of is semiuniformly convergent).
(b) Observe that
n=1
Q
n
n
2
< .
(c) Apply Theorem 7A.1(c) (p.138) to show that the given Fourier sine series for n(r)
converges uniformly.
(d) Apply Theorem 7A.1(d)[ii] (p.138) to conclude that n(0) = 0 = n().
(e) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique. 2
Proposition 11C.2. Let X = [0, ], and let : X 1 be some some func-
tion, with semiuniformly convergent Fourier cosine series: (r)

L2
n=1
Q
n
cos(nr),
and suppose that Q
0
= 0. Fix any constant 1 1, and dene the function
n : X 1 by
n(r)
unif
n=1
Q
n
n
2
cos(nr) + 1, for all r [0, ]. (11C.1)
Then n is a solution to the Poisson equation n(r) = (r), satisfying homo-
geneous Neumann boundary conditions n
t
(0) = n
t
() = 0.
Furthermore, all solutions to this Poisson equation with these boundary con-
ditions have the form (11C.1) for some choice of 1.
If Q
0
,= 0, however, the problem has no solution.
Proof. Exercise 11C.2 Hint: (a) Apply Proposition 0F.1 on page 565 twice to E _
show that n(r)
unif
n=1
Q
n
cos(nr) = (r), for all r int (X). (Hint: The Fourier
series of is semiuniformly convergent).
(b) Observe that
n=1
Q
n
n
< .
(c) Apply Theorem 7A.4(d)[ii] (p.142) to conclude that n
(0) = 0 = n
().
(d) Apply Theorem 5D.5(c) on page 88 to conclude that this solution is unique up to
addition of a constant. 2
Exercise 11C.3. Mathematically, it is clear that the solution of Proposition 11C.2 E _
cannot be well-dened if Q
0
,= 0. Provide a physical explanation for why this is to be
expected.
1. Let p(r) =
_
1 if 0 r <

2
0 if

2
r
. (see problem #5 of 7D)
(a) Find the solution to the one-dimensional heat equation
t
n(r, t) =
n(r, t) on the interval [0, ], with initial conditions n(r, 0) = p(r)
and homogeneous Dirichlet Boundary conditions.
(b) Find the solution to the one-dimensional heat equation
t
n(r, t) =
n(r, t) on the interval [0, ], with initial conditions n(r, 0) = p(r)
and homogeneous Neumann Boundary conditions.
(c) Find the solution to the one-dimensional wave equation
2
t
n(r, t) =
n(r, t) on the interval [0, ], satisfying homogeneous Dirichlet Bound-
ary conditions, with initial position n(r, 0) = 0 and initial velocity
t
n(r, 0) = p(r).
2. Let )(r) = sin(3r), for r [0, ].
(a) Compute the Fourier Sine Series of )(r) as an element of L
2
[0, ].
(b) Compute the Fourier Cosine Series of )(r) as an element of L
2
[0, ].
(c) Solve the one-dimensional heat equation (
t
n = n) on the do-
main X = [0, ], with initial conditions n(r; 0) = )(r), and the
following boundary conditions:
i. Homogeneous Dirichlet boundary conditions.
ii. Homogeneous Neumann boundary conditions.
(d) Solve the the one-dimensional wave equation (
2
t
= ) on the
domain X = [0, ], with homogeneous Dirichlet boundary conditions,
and with
Initial position: (r; 0) = 0,
Initial velocity:
t
(r; 0) = )(r).
3. Let ) : [0, ] 1, and suppose ) has
Fourier cosine series: )(r) =
n=0
1
2
n
cos(nr)
Fourier sine series: )(r) =
n=1
1
n!
sin(nr)
t
n = n,
with homogeneous Neumann boundary conditions, and initial con-
ditions n(r; 0) = )(r) for all r [0, ].
(b) Verify your solution in part (a). Check the heat equation, the initial
conditions, and boundary conditions. [Hint: Use Proposition 0F.1 on
page 565]
(c) Find the solution to the one-dimensional wave equation
2
t
n(r; t) =
n(r; t) with homogeneous Dirichlet boundary conditions, and
Initial position n(r; 0) = )(r), for all r [0, ].
Initial velocity
t
n(r; 0) = 0, for all r [0, ].
4. Let ) : [0, ] 1 be dened by )(r) = r.
(a) Compute the Fourier sine series for ).
(b) Does the Fourier sine series converge pointwise to ) on (0, )? Justify
your answer.
(c) Does the Fourier sine series converge uniformly to ) on [0, ]? Justify
your answer in two dierent ways.
(d) Compute the Fourier cosine series for ).
(e) Solve the one-dimensional heat equation (
t
n = n) on the do-
main X := [0, ], with initial conditions n(r, 0) := )(r), and with the
following boundary conditions:
[i] Homogeneous Dirichlet boundary conditions.
[ii] Homogeneous Neumann boundary conditions.
(f) Verify your solution to question (e) part [i]. That is: check that your
solution satises the heat equation, the desired initial conditions, and
homogeneous Dirichlet BC. [You may assume that the relevent series converge
uniformly, if necessary. You may dierentiate Fourier series termwise, if necessary.]
(g) Find the solution to the one-dimensional wave equation on the do-
main X := [0, ], with homogeneous Dirichlet boundary conditions,
and with
Initial position n(r; 0) = )(r), for all r [0, ].
Initial velocity
t
n(r; 0) = 0, for all r [0, ].
239
Chapter 12
square
Each problem that I solved became a rule which served afterwards to solve other problems.
Rene Descartes
Prerequisites: 9A, 5C. Recommended: 11.
Multiple Fourier series can be used to nd solutions to boundary value prob-
lems on a box [0, A] [0, Y ]. The key idea is that the functions S
n,m
(r, j) :=
sin
_
n
X
r
_
sin
_
m
Y
j
_
and C
n,m
(r, j) := cos
_
n
X
r
_
cos
_
m
Y
j
_
are eigenfunctions of
the Laplacian operator. Furthermore, S
n,m
satises Dirichlet boundary condi-
tions, so any (uniformly convergent) Fourier sine series will also do so. Likewise,
C
n,m
satises Neumann boundary conditions, so any (suciently convergent)
Fourier cosine series will also do so.
For simplicity, we will assume throughout that A = Y = . Thus S
n,m
(r) =
sin (nr) sin (:j) and C
n,m
(r) = cos (nr) cos (:j). We will also assume that
(through an appropriate choice of time units) the physical constants in the vari-
ous equations are all equal to one. Thus, the heat equation becomes
t
n = n,
the wave equation is
2
t
n = n, etc. This will allow us to develop the solution
methods in the simplest possible scenario, without a lot of distracting technical-
ities.
The extension of these solution methods to equations with arbitrary physical
constants on an arbitrary rectangular domain [0, A] [0, Y ] (for some A, Y
0) are left as exercises. These exercises are quite straightforward, but are an
eective test of your understanding of the solution techniques.
Remark on Notation: Throughout this chapter (and the following ones) we
will often write a function n(r, j; t) in the form n
t
(r, j). This emphasizes the
distinguished role of the time coordinate t, and makes it natural to think of
xing t at some value and applying the 2-dimensional Laplacian =
2
x
+
2
y
to
the resulting 2-dimensional function n
t
.
240 DRAFT Chapter 12. Boundary value problems on a square
u(x,0)=0
u
(
0
,
y
)
=
0
u
(
,
y
)
=
R
u(x,0)=B
u(x,)=T
u
(
0
,
y
)
=
L
(B) (A)
u(x,)=1
u
(
,
y
)
=
0
Figure 12A.1: The Dirichlet problem on a square. (A) Proposition 12A.1; (B) Proposi-
tions 12A.2 and 12A.4.
Some authors use the subscript notation n
t
to denote the partial derivative
t
n. We never use this notation. In this book, partial derivatives are always
denoted by
t
n, etc.
12A The Dirichlet problem on a square
Prerequisites: 9A, 5C(i), 1C, 0F. Recommended: 7C(v).
In this section we will learn to solve the Dirichlet problem on a square
domain X: that is, to nd a function which is harmonic on the interior of X
and which satises specied Dirichlet boundary conditions on the boundary X.
Solutions to the Dirichlet problem have several physical interpretations.
Heat: Imagine that the boundaries of X are perfect heat conductors, which are
in contact with external heat reservoirs with xed temperatures. For ex-
ample, one boundary might be in contact with a heat source, and another,
in contact with a coolant liquid. The solution to the Dirichlet problem is
then the equilibrium temperature distribution on the interior of the box,
given these constraints.
Electrostatic: Imagine that the boundaries of X are electrical conductors which
are held at some xed voltage by the application of an external electric
potential (dierent boundaries, or dierent parts of the same boundary,
may be held at dierent voltages). The solution to the Dirichlet problem
is then the electric potential eld on the interior of the box, given these
constraints.
Minimal surface: Imagine a squarish frame of wire, which we have bent in the
vertical direction to have some shape. If we dip this wire frame in a soap
12A. The Dirichlet problem on a square 241
y
2.5
0.5
3
0
1.5
1.5
3 0.5 1
2
0 2
1
2.5
x
y
2.5
0.5
3
1.5
1
2
1 2
x
1.5
0.5
0
2.5 0 3
y
2.5
0.5
3
x
0.5 3 2.5
1.5
0 2
1
2
1.5
0
1
Proposition 12A.1 Example 12A.3 Example 12A.6
T = 1, R = L = B = 0 T = 3, L = 5, R = B = 0 T = tent function , R = L = B = 0
sin(x) sinh(y) sin(2x) sinh(2y) sin(3x) sinh(3y)
T(x) = sin(x) , R = L = B = 0 T(x) = sin(2x) , R = L = B = 0 T(x) = sin(3x) , R = L = B = 0
Figure 12A.2: The Dirichlet problem on a box. The curves represent isothermal contours (of
a temperature distribution) or equipotential lines (of an electric voltage eld).
solution, we can form a soap bubble (i.e. minimal-energy surface) which
must obey the boundary conditions imposed by the shape of the wire.
The dierential equation describing a minimal surface is not exactly the
same as the Laplace equation; however, when the surface is not too steeply
slanted (i.e. when the wire frame is not too bent), the Laplace equation
is a good approximation; hence the solution to the Dirichlet problem is a
good approximation of the shape of the soap bubble.
We will begin with the simplest problem: a constant, nonzero Dirichlet
boundary condition on one side of the box, and zero boundary conditions on
the other three sides.
Proposition 12A.1. (Dirichlet problem; one constant nonhomog. BC)
Let X = [0, ] [0, ], and consider the Laplace equation n = 0, with
3 0
0 2.5
0.2
0.5
2
0.4
1
1.5
0.6
1.5 y
0.8
1 x 2
0.5 2.5
3 0
3
0 2.5
-2
-1
0.5
2
0
1
1
1.5
2
1.5 y
3
1 x
4
2
0.5 2.5
3 0
Proposition 12A.1 Example 12A.3 Example 12A.6
T = 1, R = L = B = 0 T = 3, L = 5, R = B = 0 T = tent function , R = L = B = 0
3
-1
0 2.5
-0.5
0.5
2
1
0
1.5
1.5 y
0.5
1 x 2
1
0.5 2.5
3 0
sin(x) sinh(y) sin(2x) sinh(2y) sin(3x) sinh(3y)
T(x) = sin(x) , R = L = B = 0 T(x) = sin(2x) , R = L = B = 0 T(x) = sin(3x) , R = L = B = 0
Figure 12A.3: The Dirichlet problem on a box: 3-dimensional plots. You can imagine these
as soap lms.
nonhomogeneous Dirichlet boundary conditions [see Figure 12A.1(A)]:
n(0, j) = n(, j) = 0, for all j [0, ). (12A.1)
n(r, 0) = 0 and n(r, ) = 1, for all r [0, ]. (12A.2)
The unique solution to this problem is the function n : X 1 dened:
n(r, j)

L2
4
n=1
n odd
1
nsinh(n)
sin(nr) sinh(nj), for all (r, j) X.
[See Figures 12A.2(a) and 12A.3(a).] Furthermore, this series converges semiu-
niformly on int (X).
Proof. Exercise 12A.1 E _
(a) Check that, for all n N, the function n
n
(r, j) = sin(nr) sinh(nj) satises the
Laplace equation and the rst boundary condition (12A.1). See Figures 12A.2(d,e,f)
and 12A.3(d,e,f).
(b) Show that
n=1
n odd
n
2
sinh(nj)
nsinh(n)
< , for any xed j < . (Hint. If j < ,

then sinh(nj), sinh(n) decays like exp(n(j )) as n .)
(c) Apply Proposition 0F.1 on page 565 to conclude that n(r, j) = 0.
(d) Observe that
n=1
n odd
sinh(nj)
nsinh(n)
< , for any xed j < .

(e) Apply part (c) of Theorem 7A.1 on page 138 to show that the series given for
n(r, j) converges uniformly for any xed j < .
(f) Apply part (d) of Theorem 7A.1 on page 138 to conclude that n(0, j) = 0 = n(, j)
for all j < .
(g) Observe that sin(nr) sinh(n 0) = 0 for all n N and all r [0, ]. Conclude
that n(r, 0) = 0 for all r [0, ].
(h) To check that the solution also satises the boundary condition (12A.2), subsiti-
tute j = to get:
n(r, ) =
4
n=1
n odd
1
nsinh(n)
sin(nr)sinh(n) =
4
n=1
n odd
1
n
sin(nr)

L2
1.
because
4
n=1
n odd
1
n
sin(nr) is the (one-dimensional) Fourier sine series for the function
/(r) = 1 (see Example 7A.2(b) on page 139).
(i) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique.
2
Proposition 12A.2. (Dirichlet Problem; four constant nonhomog. BC)
nonhomogeneous Dirichlet boundary conditions [see Figure 12A.1(B)]:
n(0, j) = 1 and n(, j) = 1, for all j (0, );
n(r, ) = T and n(r, 0) = 1, for all r (0, ).
where 1, 1, T, and 1 are four constants. The unique solution to this problem
is the function n : X 1 dened:
n(r, j) := |(r, j) +:(r, j) +t(r, j) +/(r, j), for all (r, j) X.
where, for all (r, j) X,
|(r, j)

L2
1
n=1
n odd
c
n
sinh
_
n( r)
_
sin(nj), :(r, j)

L2
1
n=1
n odd
c
n
sinh(nr) sin(nj),
t(r, j)

L2
T
n=1
n odd
c
n
sin(nr) sinh(nj), /(r, j)

L2
1
n=1
n odd
c
n
sin(nr) sinh
_
n( j)
_
.
where c
n
:=
4
n sinh(n)
, for all n N.
Furthermore, these four series converge semiuniformly on int (X).
Proof. Exercise 12A.2 E _
(a) Apply Proposition 12A.1 to show that each of the functions |(r, j), :(r, j), t(r, j),
/(r, j) satises a Dirichlet problem where one side has nonzero temperature and the
other three sides have zero temperature.
(b) Add these four together to get a solution to the original problem.
(c) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique. 2
Exercise 12A.3. What happens to the solution at the four corners (0, 0), (0, ), E _
(, 0) and (, )?
Example 12A.3. Suppose 1 = 0 = 1, T = 3, and 1 = 5. Then the solution
is:
n(r, j)

L2
1
n=1
n odd
c
n
sinh
_
n( r)
_
sin(nj) + T
n=1
n odd
c
n
sin(nr) sinh(nj)
=
20
n=1
n odd
sinh
_
n( r)
_
sin(nj)
nsinh(n)

12
n=1
n odd
sin(nr) sinh(nj)
nsinh(n)
.
See Figures 12A.2(b) and 12A.3(b).
Proposition 12A.4. (Dirichlet Problem; arbitrary nonhomogeneous boundaries)
nonhomogeneous Dirichlet boundary conditions [see Figure 12A.1(B)]:
n(0, j) = 1(j) and n(, j) = 1(j), for all j (0, );
n(r, ) = T(r) and n(r, 0) = 1(r), for all r (0, ).
where 1, 1, T, 1 : [0, ] 1 are four arbitrary functions. Suppose these func-
tions have (one-dimensional) Fourier sine series:
1(j)

L2
n=1
1
n
sin(nj), 1(j)

L2
n=1
1
n
sin(nj), for all j [0, ];
T(r)

L2
n=1
T
n
sin(nr), and 1(r)

L2
n=1
1
n
sin(nr), for all r [0, ].
The unique solution to this problem is the function n : X 1 dened:
n(r, j) := |(r, j) + :(r, j) + t(r, j) + /(r, j), for all (r, j) X.
where, for all (r, j) X,
|(r, j)

L2
n=1
1
n
sinh(n)
sinh
_
n( r)
_
sin(nj),
:(r, j)

L2
n=1
1
n
sinh(n)
sinh(nr) sin(nj),
t(r, j)

L2
n=1
T
n
sinh(n)
sin(nr) sinh(nj),
and /(r, j)

L2
n=1
1
n
sinh(n)
sin(nr) sinh
_
n( j)
_
.
Furthermore, these four series converge semiuniformly on int (X).
Proof. Exercise 12A.4 First we consider the function t(r, j). E _
(a) Same as Exercise 12A.1(a)
(b) For any xed j < , show that
n=1
n
2
T
n
sinh(nj)
sinh(n)
< . (Hint. If j < , then

sinh(nj), sinh(n) decays like exp(n(j )) as n .)
(c) Combine part (b) and Proposition 0F.1 on page 565 to conclude that t(r, j) is
harmonic i.e. t(r, j) = 0.
Through symmetric reasoning, conclude that the functions /(r, j), :(r, j) and /(r, j)
are also harmonic.
(d) Same as Exercise 12A.1(d)
(e) Apply part (c) of Theorem 7A.1 on page 138 to show that the series given for
t(r, j) converges uniformly for any xed j < .
(f) Apply part (d) of Theorem 7A.1 on page 138 to conclude that t(0, j) = 0 = t(, j)
for all j < .
(g) Observe that sin(nr) sinh(n 0) = 0 for all n N and all r [0, ]. Conclude
that t(r, 0) = 0 for all r [0, ].
(h) To check that the solution also satises the boundary condition (12A.2), subsiti-
tute j = to get:
t(r, ) =
n=1
T
n
sinh(n)
sin(nr) sinh(n) =
4
n=1
T
n
sin(nr) = T(r).
(j) At this point, we know that t(r, ) = T(r) for all r [0, ], and t 0 on the other
three sides of the square. Through symmetric reasoning, show that:
/(0, j) = 1(j) for all j [0, ], and / 0 on the other three sides of the square.
:(, j) = 1(j) for all j [0, ], and : 0 on the other three sides of the square.
/(r, 0) = 1(r) for all r [0, ], and / 0 on the other three sides of the square.
(k) Conclude that n = t + / + : + / is harmonic and satises the desired boundary
conditions.
(l) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique. 2
Example 12A.5. If T(r) = sin(3r), and 1 1 1 0, then n(r, j) =
sin(3r) sinh(3j)
sinh(3)
.
Example 12A.6. Let X = [0, ] [0, ]. Solve the 2-dimensional Laplace
Equation on X, with inhomogeneous Dirichlet boundary conditions:
n(0, j) = 0; n(, j) = 0; n(r, 0) = 0;
n(r, ) = T(r) =
_
r if 0 r

2
r if

2
< r
(see Figure 7C.4(B) on page 154)
Solution: Recall from Example 7C.7 on page 155 that T(r) has Fourier series:
T(r)

L2
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sin(nr).
Thus, the solution is n(r, j)

L2
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sinh(n)
sin(nr) sinh(nj).
See Figures 12A.2(c) and 12A.3(c).
Exercise 12A.5. Let A, Y 0 and let X := [0, A] [0, Y ]. Generalize Proposition E _
12A.4 to nd the solution to the Laplace equation on X, satisfying arbitrary nonhomo-
geneous Dirichlet boundary conditions on the four sides of X.
12B The heat equation on a square
12B(i) Homogeneous boundary conditions
Prerequisites: 9A, 5B, 5C, 1B(ii), 0F. Recommended: 11A, 7C(v).
12B. The heat equation on a square 247
Proposition 12B.1. (Heat equation; homogeneous Dirichlet boundary)
Consider the box X = [0, ] [0, ], and let ) : X 1 be some function
describing an initial heat distribution. Suppose ) has Fourier Sine Series
)(r, j)

L2
n,m=1
1
n,m
sin(nr) sin(:j)
,
1 by
n
t
(r, j)

L2
n,m=1
1
n,m
sin(nr) sin(:j) exp
_
(n
2
+:
2
) t
_
,
for all (r, j) X and t 0. Then n is the unique solution to the heat equation
t
n = n, with homogeneous Dirichlet boundary conditions
n
t
(r, 0) = n
t
(0, j) = n
t
(, j) = n
t
(r, ) = 0, for all r, j [0, ] and t 0.
and initial conditions: n
0
(r, j) = )(r, j), for all (r, j) X.
+
.
(a) Show that, when t = 0, the two-dimensional Fourier series of n
0
(r, j) agrees with
that of )(r, j); hence n
0
(r, j) = )(r, j).
n,m=1
(n
2
+:
2
) 1
n,m
c
(n
2
+m
2
)t
< .
(c) For any T 0, apply Proposition 0F.1 on page 565 to conclude that
t
n
t
(r, j)
unif
n,m=1
(n
2
+:
2
)1
n,m
sin(nr)sin(:j)exp
_
(n
2
+:
2
) t
_
unif
n
t
(r, j),
for all (r, j; t) X [T, ).
(d) Observe that for all t 0,
n,m=1
1
n,m
c
(n
2
+m
2
)t
< .
(e) Apply part (c)[i] of Theorem 9A.3 on page 183 to show that the two-dimensional
Fourier series of n
t
converges uniformly for any xed t 0.
(f) Apply part (d)[ii] of Theorem 9A.3 on page 183 to conclude that n
t
satises
homogeneous Dirichlet boundary conditions, for all t 0.
(g) Apply Theorem 5D.8 on page 91 to show that this solution is unique. 2
Figure 12B.1: (A) A hot metal rod quenched in a cold bucket. (B) A cross
section of the rod in the bucket.
Example 12B.2: (The quenched rod)
On a cold January day, a blacksmith is tempering an iron rod. He pulls it out
of the forge and plunges it, red-hot, into ice-cold water (Figure 12B.1A). The
rod is very long and narrow, with a square cross section. We want to compute
how the rod cooled.
Answer: The rod is immersed in freezing cold water, and is a good conduc-
tor, so we can assume that its outer surface takes the the surrounding water
temperature of 0 degrees. Hence, we assume homogeneous Dirichlet boundary
conditions.
Endow the rod with coordinate system (r, j, .), where . runs along the length
of the rod. Since the rod is extremely long relative to its cross-section, we
can neglect the . coordinate, and reduce to a 2-dimensional equation (Figure
12B.1B). Assume the rod was initially uniformly heated to a temperature of T.
The initial temperature distribution is thus a constant function: )(r, j) = T.
From Example 9A.2 on page 182, we know that the constant function 1 has
two-dimensional Fourier sine series:
1

L2
16
n,m=1
both odd
1
n :
sin(nr) sin(:j).
Thus, )(r, j)

L2
16T
n,m=1
both odd
1
n :
sin(nr) sin(:j). Thus, the time-varying
thermal prole of the rod is given:
n
t
(r, j)

L2
16T
n,m=1
both odd
1
n :
sin(nr) sin(:j) exp
_
(n
2
+:
2
) t
_
.
Proposition 12B.3. (Heat equation; homogeneous Neumann boundary)
Consider the box X = [0, ] [0, ], and let ) : X 1 be some function
describing an initial heat distribution. Suppose ) has Fourier Cosine Series
)(r, j)

L2
n,m=0
n,m
cos(nr) cos(:j)
,
1 by:
n
t
(r, j)

L2
n,m=0
n,m
cos(nr) cos(:j) exp
_
(n
2
+:
2
) t
_
,
for all (r, j) X and t 0. Then n is the unique solution to the heat equation
t
n = n, with homogeneous Neumann boundary conditions
y
n
t
(r, 0) =
y
n
t
(r, ) =
x
n
t
(0, j) =
x
n
t
(, j) = 0, for all r, j [0, ] and t 0.
and initial conditions: n
0
(r, j) = )(r, j), for all (r, j) X. 2
+
.
(a) Show that, when t = 0, the two-dimensional Fourier cosine series of n
0
(r, j) agrees
with that of )(r, j); hence n
0
(r, j) = )(r, j).
n,m=0
(n
2
+:
2
)
n,m
c
(n
2
+m
2
)t
< .
(c) Apply Proposition 0F.1 on page 565 to conclude that
t
n
t
(r, j)
unif
n,m=0
(n
2
+:
2
)
n,m
cos(nr)cos(:j)exp
_
(n
2
+:
2
) t
_
unif
n
t
(r, j),
for all (r, j) X and t 0.
(d) Observe that for all t 0,
n,m=0
n
n,m
c
(n
2
+m
2
)t
< and
n,m=0
:
n,m
c
(n
2
+m
2
)t
< .
(e) Apply part (e)[ii] of Theorem 9A.3 on page 183 to conclude that n
t
satises
homogeneous Neumann boundary conditions, for any xed t 0.
(f) Apply Theorem 5D.8 on page 91 to show that this solution is unique. 2
Example 12B.4. Suppose X = [0, ] [0, ]
(a) Let )(r, j) = cos(3r) cos(4j) + 2 cos(5r) cos(6j). Then
3,4
= 1 and
5,6
= 2, and all other Fourier coecients are zero. Thus, n(r, j; t) =
cos(3r) cos(4j) c
25t
+ cos(5r) cos(6j) c
59t
.
(b) Suppose )(r, j) =
_
1 if 0 r <

2
and 0 j <

2
;
0 if

2
r or

2
j.
We know
from Example 9A.4 on page 184 that the two-dimensional Fourier cosine
series of ) is:
)(r, j)

L2
1
4
+
1
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
+
1
j=0
(1)
j
2, + 1
cos
_
(2, + 1)j
_
+
4
k,j=1
(1)
k+j
(2/ + 1)(2, + 1)
cos
_
(2/ + 1)r
_
cos
_
(2, + 1)j
_
Thus, the solution to the heat equation, with initial conditions n
0
(r, j) =
)(r, j) and homogeneous Neumann boundary conditions is given:
n
t
(r, j)

L2
1
4
+
1
k=0
(1)
k
2/ + 1
cos
_
(2/ + 1)r
_
c
(2k+1)
2
t
+
1
j=0
(1)
j
2, + 1
cos
_
(2, + 1)j
_
c
(2j+1)
2
t
+
4
k,j=1
(1)
k+j
(2/ + 1)(2, + 1)
cos
_
(2/ + 1)r
_
cos
_
(2, + 1)j
_
c
[(2k+1)
2
+(2j+1)
2
]t
Exercise 12B.3. Let A, Y 0 and let X := [0, A] [0, Y ]. Let 0 be a diusion E _

constant, and consider the general two-dimensional heat equation
t
n = n. (12B.1)
(a) Generalize Proposition 12B.1 to nd the solution to eqn.(12B.1) on X satisfying
prescribed initial conditions and homogeneous Dirichlet boundary conditions.
(b) Generalize Proposition 12B.3 to nd the solution to eqn.(12B.1) on X satisfying
prescribed initial conditions and homogeneous Neumann boundary conditions.
In both cases, prove that your solution converges, satises the desired initial conditions
and boundary conditions, and satises eqn.(12B.1) (Hint: imitate the strategy suggested
in Exercises 12B.1 and 12B.2)
Exercise 12B.4. Let ) : X 1 and suppose the Fourier sine series of ) satises E _
the constraint
n,m=1
(n
2
+ :
2
)[1
nm
[ < . Imitate Proposition 12B.1 to nd a Fourier
series solution to the initial value problem for the two-dimensional free Schrodinger
equation
i
t
=
1
2
(12B.2)
on the box X = [0, ]
2
, with homogeneous Dirichlet boundary conditions. Prove that
your solution converges, satises the desired initial conditions and boundary conditions,
and satises eqn.(12B.2). (Hint: imitate the strategy suggested in Exercise 12B.1, and
also Exercise 12D.1 on page 260).
12B(ii) Nonhomogeneous boundary conditions
Prerequisites: 12B(i), 12A. Recommended: 12C(ii).
Proposition 12B.5. (Heat equation on box; nonhomogeneous Dirichlet BC)
Let X = [0, ] [0, ]. Let ) : X 1 and let 1, 1, T, 1 : [0, ] 1 be
functions. Consider the Heat equation
t
n(r, j; t) = n(r, j; t),
with initial conditions
n(r, j; 0) = )(r, j), for all (r, j) X, (12B.3)
and nonhomogeneous Dirichlet boundary conditions:
n(r, ; t) = T(r) and n(r, 0; t) = 1(r), for all r [0, ]
n(0, j; t) = 1(j) and n(, j; t) = 1(j), for all j [0, ]
_
for all t 0.
(12B.4)
This problem is solved as follows:
1. Let n(r, j) be the solution
1
to the Laplace Equation n(r, j) = 0, with
the nonhomogeneous Dirichlet BC (12B.4).
2. Dene p(r, j) := )(r, j) n(r, j). Let (r, j; t) be the solution
2
to
the heat equation
t
(r, j; t) = (r, j; t) with initial conditions
(r, j; 0) = p(r, j), and homogeneous Dirichlet BC.
3. Dene n(r, j; t) := (r, j; t) +n(r, j). Then n(r, j; t) is a solution to the
heat equation with initial conditions (12B.3) and nonhomogeneous Dirich-
let BC (12B.4).
Interpretation: In Proposition 12B.5, the function n(r, j) represents the
long-term thermal equilibrium that the system is trying to attain. The function
p(r, j) = )(r, j) n(r, j) thus measures the deviation between the current
state and this equilibrium, and the function (r, j; t) thus represents how this
transient deviation decays to zero over time.
Example 12B.6. Suppose T(r) = sin(2r) and 1 1 0 and 1 0. Then
Proposition 12A.4 on page 244 says
n(r, j) =
sin(2r) sinh(2j)
sinh(2)
.
Suppose )(r, j) := sin(2r) sin(j). Then
p(r, j) = )(r, j) n(r, j) = sin(2r) sin(j)
sin(2r) sinh(2j)
sinh(2)
()
sin(2r) sin(j)
_
sin(2r)
sinh(2)
__
2 sinh(2)
m=1
:(1)
m+1
2
2
+:
2
sin (:j)
= sin(2r) sin(j)
2 sin(2r)
m=1
:(1)
m+1
4 +:
2
sin (:j) .
Here () is because Example 7A.3 on page 140 says sinh(2j) =
2 sinh(2)
m=1
:(1)
m+1
2
2
+:
2

sin (:j). Thus, Proposition 12B.1 on page 247 says that
(r, j; t) = sin(2r) sin(j)c
5t
2 sin(2r)
m=1
:(1)
m+1
4 +:
2
sin (:r) exp((4+:
2
)t).
Finally, Proposition 12B.5 says the solution is n(r, j; t) := (r, j; t) +
sin(2r) sinh(2j)
sinh(2)
.
Example 12B.7. A freshly baked baguette is removed from the oven and left on
a wooden plank to cool near the window. The baguette is initially at a uniform
temperature of 90
o
C; the air temperature is 20
o
C, and the temperature of the
wooden plank (which was sitting in the sunlight) is 30
o
C.
Mathematically model the cooling process near the center of the baguette. How
long will it be before the baguette is cool enough to eat? (assuming cool enough
is below 40
o
C.)
1
Obtained from Proposition 12A.4 on page 244, for example.
2
Obtained from Proposition 12B.1 on page 247, for example.
Figure 12B.2: The temperature distribution of a baguette
Answer: For simplicity, we will assume the baguette has a square cross-
section (and dimensions , of course). If we conne our attention to the
middle of the baguette, we are far from the endpoints, so that we can neglect
the longitudinal dimension and treat this as a two-dimensional problem.
Suppose the temperature distribution along a cross section through the center
of the baguette is given by the function n(r, j; t). To simplify the problem, we
will subtract 20
o
C o all temperatures. Thus, in the notation of Proposition
12B.5 the boundary conditions are:
1(j) = 1(j) = T(r) = 0 (the air)
and 1(r) = 10. (the wooden plank)
and our initial temperature distribution is )(r, j) = 70 (see Figure 12B.2).
From Proposition 12A.1 on page 241, we know that the long-term equilibrium
for these boundary conditions is given by:
n(r, j)

L2
40
n=1
n odd
1
nsinh(n)
sin(nr) sinh(n( j)).
We want to represent this as a two-dimensional Fourier sine series. To do this,
we need the (one-dimensional) Fourier sine series for sinh(nr). We set = n
in Example 7A.3 on page 140, and get:
sinh(nr)

L2
2 sinh(n)
m=1
:(1)
m+1
n
2
+:
2
sin (:r) . (12B.5)
Thus,
sinh
_
n( j)
_

L2
2 sinh(n)
m=1
:(1)
m+1
n
2
+:
2
sin (: :j)
=
2 sinh(n)
m=1
:
n
2
+:
2
sin (:j) ,
because sin (: nj) = sin(:) cos(nj)cos(:) sin(nj) = (1)
m+1
sin (nj).
Substituting this into (12B.5) yields:
n(r, j)

L2
80
n=1
n odd
m=1
: sinh(n)
n sinh(n)(n
2
+:
2
)
sin(nr) sin (:j)
=
80
n=1
n odd
m=1
: sin(nr) sin(:j)
n (n
2
+:
2
)
(12B.6)
Now, the initial temperature distribution is the constant function with value
70. Take the two-dimensional sine series from Example 9A.2 on page 182, and
multiply it by 70, to obtain:
)(r, j) = 70

L2
1120
n,m=1
both odd
1
n :
sin (nr) sin (:j)
Thus,
p(r, j) = )(r, j) n(r, j)
L2
1120
n,m=1
both odd
sin (nr) sin (:j)
n :

80
n=1
n odd
m=1
: sin(nr) sin(:j)
n (n
2
+:
2
)
Thus,
(r, j; t)

L2
1120
n,m=1
both odd
sin (nr) sin (:j)
n :
exp
_
(n
2
+:
2
)t
_
80
n=1
n odd
m=1
: sin(nr) sin(:j)
n (n
2
+:
2
)
exp
_
(n
2
+:
2
)t
_
If we combine the second term in this expression with (12B.6), we get the nal
answer:
n(r, j; t) = (r, j; t) +n(r, j)
L2
1120
n,m=1
both odd
sin (nr) sin (:j)
n :
exp
_
(n
2
+:
2
)t
_
+
80
n=1
n odd
m=1
: sin(nr) sin(:j)
n (n
2
+:
2
)
_
1 exp
_
(n
2
+:
2
)t
_
_
12C. The Poisson problem on a square 255
12C The Poisson problem on a square
12C(i) Homogeneous boundary conditions
Prerequisites: 9A, 5C, 1D. Recommended: 11C, 7C(v).
Proposition 12C.1. Let X = [0, ] [0, ], and let : X 1 be some
function with semiuniformly convergent Fourier sine series:
(r, j)

L2
n,m=1
Q
n,m
sin(nr) sin(:j).
Dene the function n : X 1 by n(r, j)
unif
n,m=1
Q
n,m
n
2
+:
2
sin(nr) sin(:j),
for all (r, j) X.
Then n is the unique solution to the Poisson equation n(r, j) = (r, j),
satisfying homogeneous Dirichlet boundary conditions n(r, 0) = n(0, j) = n(r, ) =
n(, j) = 0.
Proof. Exercise 12C.1 (a) Use Proposition 0F.1 on page 565 to show that n E _
satises the Poisson equation on int (X).
(b) Use Proposition 9A.3(e) on page 183 to show that n satises homogeneous Dirich-
let BC.
(c) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique. 2
Example 12C.2. A nuclear submarine beneath the Arctic Ocean has jettisoned
a fuel rod from its reactor core (Figure 12C.1). The fuel rod is a very long,
narrow, enriched uranium bar with square cross section. The radioactivity
causes the fuel rod to be uniformly heated from within at a rate of Q, but the
rod is immersed in freezing Arctic water. We want to compute its internal
temperature distribution.
Answer: The rod is immersed in freezing cold water, and is a good conduc-
tor, so we can assume that its outer surface takes the the surrounding water
temperature of 0 degrees. Hence, we assume homogeneous Dirichlet boundary
conditions.
Endow the rod with coordinate system (r, j, .), where . runs along the length
of the rod. Since the rod is extremely long relative to its cross-section, we can
neglect the . coordinate, and reduce to a 2-dimensional equation. The uniform
heating is described by a constant function: (r, j) = Q. From Example 9A.2
Figure 12C.1: A jettisoned fuel rod in the Arctic Ocean
on page 182, know that the constant function 1 has two-dimensional Fourier
sine series:
1

L2
16
n,m=1
both odd
1
n :
sin(nr) sin(:j)
Thus, (r, j)

L2
16Q
n,m=1
both odd
1
n :
sin(nr) sin(:j). The temperature dis-
tribution must satisfy Poissons equation. Thus, the temperature distribution
is:
n(r, j)
unif
16Q
n,m=1
both odd
1
n : (n
2
+:
2
)
sin(nr) sin(:j).
Example 12C.3. Suppose (r, j) = r j. Then the solution to the Poisson
equation n = on the square, with homogeneous Dirichlet boundary
conditions, is given by:
n(r, j)
unif
4
n,m=1
(1)
n+m+1
n: (n
2
+:
2
)
sin(nr) sin(:j)
To see this, recall from Example 9A.1 on page 179 that the two-dimensional
Fourier sine series for (r, j) is:
rj

L2
4
n,m=1
(1)
n+m
n:
sin(nr) sin(:j).
Now apply Proposition 12C.1.
12C. The Poisson problem on a square 257
Proposition 12C.4. Let X = [0, ] [0, ], and let : X 1 be some
function with semiuniformly convergent Fourier cosine series:
(r, j)

L2
n,m=0
Q
n,m
cos(nr) cos(:j).
Suppose that Q
0,0
= 0. Fix some constant 1 1, and dene the function
n : X 1 by
n(r, j)
unif
n,m=0
not both zero
Q
n,m
n
2
+:
2
cos(nr) cos(:j) + 1, (12C.1)
for all (r, j) X. Then n is a solution to the Poisson equation n(r, j) =
(r, j), satisfying homogeneous Neumann boundary conditions
y
n(r, 0) =
x
n(0, j) =
y
n(r, ) =
x
n(, j) = 0.
ditions have the form (12C.1).
If Q
0,0
Proof. Exercise 12C.2 (a) Use Proposition 0F.1 on page 565 to show that n E _
satises the Poisson equation on int (X).
(b) Use Proposition 9A.3 on page 183 to show that n satises homogeneous Neumann
BC.
(c) Apply Theorem 5D.5(c) on page 88 to conclude that this solution is unique up to
addition of a constant. 2
Exercise 12C.3. Mathematically, it is clear that the solution of Proposition 12C.4 E _
cannot be well-dened if Q
0,0
,= 0. Provide a physical explanation for why this is to be
expected.
Example 12C.5. Suppose (r, j) = cos(2r) cos(3j). Then the solution to
the Poisson equation n = on the square, with homogeneous Neumann
boundary conditions, is given by:
n(r, j) =
cos(2r) cos(3j)
13
.
To see this, note that the two-dimensional Fourier Cosine series of (r, j) is
just cos(2r) cos(3j). In other words,
2,3
= 1, and
n,m
= 0 for all other n
and :. In particular,
0,0
= 0, so we can apply Proposition 12C.4 to conclude:
n(r, j) =
cos(2x)cos(3y)
2
2
+3
2
=
cos(2x)cos(3y)
13
.
12C(ii) Nonhomogeneous boundary conditions
Prerequisites: 12C(i), 12A. Recommended: 12B(ii).
Proposition 12C.6. (Poisson equation on box; nonhomogeneous Dirichlet BC)
Let X = [0, ] [0, ]. Let : X 1 and 1, 1, T, 1 : [0, ] 1 be functions.
Consider the Poisson equation
n(r, j) = (r, j), (12C.2)
with nonhomogeneous Dirichlet boundary conditions:
n(r, ) = T(r) and n(r, 0) = 1(r), for all r [0, ]
n(0, j) = 1(j) and n(, j) = 1(j), for all j [0, ]
(12C.3)
(see Figure 12A.1(B) on page 240). This problem is solved as follows:
1. Let (r, j) be the solution
3
to the Poisson equation (12C.2) with homoge-
neous Dirichlet BC: (r, 0) = (0, j) = (r, ) = (, j) = 0.
2. Let n(r, j) be the solution
4
to the Laplace Eqation n(r, j) = 0, with
the nonhomogeneous Dirichlet BC (12C.3).
3. Dene n(r, j) := (r, j)+n(r, j); then n(r, j) is a solution to the Poisson
problem with the nonhomogeneous Dirichlet BC (12C.3).
Example 12C.7. Suppose (r, j) = r j. Find the solution to the Poisson
equation n = on the square, with nonhomogeneous Dirichlet boundary
conditions:
n(0, j) = 0; n(, j) = 0; n(r, 0) = 0; (12C.4)
n(r, ) = T(r) =
_
r if 0 r

2
2
r if

2
< r
(see Figure 7C.4(B) on page 154)
(12C.5)
Solution: In Example 12C.3, we found the solution to the Poisson equation
= , with homogeneous Dirichlet boundary conditions; it was:
(r, j)
unif
4
n,m=1
(1)
n+m+1
n: (n
2
+:
2
)
sin(nr) sin(:j).
3
Obtained from Proposition 12C.1 on page 255, for example.
4
Obtained from Proposition 12A.4 on page 244, for example.
12D. The wave equation on a square (the square drum) 259
In Example 12A.6 on page 246, we found the solution to the Laplace equation
n = 0, with nonhomogeneous Dirichlet boundary conditions (12C.4) and
(12C.5); it was:
n(r, j)

L2
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sinh(n)
sin(nr) sinh(nj).
Thus, according to Proposition 12C.6 on the facing page, the solution to the
nonhomogeneous Poisson problem is:
n(r, j) = (r, j) +n(r, j)
L2
4
n,m=1
(1)
n+m+1
n: (n
2
+:
2
)
sin(nr) sin(:j) +
4
n=1
n odd;
n=2k+1
(1)
k
n
2
sinh(n)
sin(nr) sinh(nj).
12D The wave equation on a square (the square drum)

Prerequisites: 9A, 5B, 5C, 2B(ii), 0F. Recommended: 11B, 7C(v).
Imagine a drumskin stretched tightly over a square frame. At equilibrium,
the drumskin is perfectly at, but if we strike the skin, it will vibrate, meaning
that the membrane will experience vertical displacements from equilibrium. Let
X = [0, ] [0, ] represent the square skin, and for any point (r, j) X on the
drumskin and time t 0, let n(r, j; t) be the vertical displacement of the drum.
Then n will obey the two-dimensional wave equation:
2
t
n(r, j; t) = n(r, j; t). (12D.1)
However, since the skin is held down along the edges of the box, the function n
will also exhibit homogeneous Dirichlet boundary conditions
n(r, ; t) = 0 and n(r, 0; t) = 0, for all r [0, ]
n(0, j; t) = 0 and n(, j; t) = 0, for all j [0, ]
_
for all t 0.
(12D.2)
Proposition 12D.1. (Initial Position for Square Drumskin)
Let X = [0, ] [0, ], and let )
0
: X 1 be a function describing
the initial displacement of the drumskin. Suppose )
0
has Fourier Sine Series
)
0
(r, j)
unif
n,m=1
1
n,m
sin(nr) sin(:j), such that:
n,m=1
(n
2
+:
2
)[1
n,m
[ < . (12D.3)
Dene the function n : X 1
,
1 by
n(r, j; t)
unif
n,m=1
1
n,m
sin(nr) sin(:j) cos
_
_
n
2
+:
2
t
_
, (12D.4)
for all (r, j) X and t 0. Then series (12D.4) converges uniformly, and
n(r, j; t) is the unique solution to the wave equation (12D.1), satisfying the
Dirichlet boundary conditions (12D.2), as well as
Initial Position: n(r, j, 0) = )
0
(r, j),
Initial Velocity:
t
n(r, j, 0) = 0,
_
for all (r, j) X.
Proof. Exercise 12D.1 (a) Use the hypothesis (12D.3) and Proposition 0F.1 on E _
page 565 to conclude that
2
t
n(r, j; t)
unif

n,m=1
(n
2
+:
2
)1
n,m
sin(nr)sin(:j)cos
_
_
n
2
+:
2
t
_
unif
n(r, j; t)
(b) Check that the Fourier series (12D.4) converges uniformly.
(c) Use Theorem 9A.3(d)[ii] on page 183 to conclude that n satises Dirichlet bound-
ary conditions.
(d) Set t = 0 to check the initial position.
(e) Set t = 0 and use Proposition 0F.1 on page 565 to check initial velocity.
Example 12D.2. Suppose )
0
(r, j) = sin(2r) sin(3j). Then the solution to
the wave equation on the square, with initial position )
0
, and homogeneous
Dirichlet boundary conditions, is given by:
n(r, j; t) = sin(2r) sin(3j) cos(
13 t).
To see this, note that the two-dimensional Fourier sine series of )
0
(r, j) is
just sin(2r) sin(3j). In other words, 1
2,3
= 1, and 1
n,m
= 0 for all other n
and :. Apply Proposition 12D.1 to conclude: n(r, j; t) = sin(2r) sin(3j)
cos
_
2
2
+ 3
2
t
_
= sin(2r) sin(3j) cos(
13 t).
12D. The wave equation on a square (the square drum) 261
Proposition 12D.3. (Initial Velocity for Square Drumskin)
Let X = [0, ] [0, ], and let )
1
: X 1 be a function describing the
initial velocity of the drumskin. Suppose )
1
1
(r, j)
unif
n,m=1
1
n,m
sin(nr) sin(:j), such that
n,m=1
_
n
2
+:
2
[1
n,m
[ < . (12D.5)
Dene the function : X 1
,
1 by:
(r, j; t)
unif
n,m=1
1
n,m
n
2
+:
2
sin(nr) sin(:j) sin
_
_
n
2
+:
2
t
_
, (12D.6)
for all (r, j) X and t 0. Then the series (12D.6) converges uniformly,
and (r, j; t) is the unique solution to the wave equation (12D.1), satisfying the
Dirichlet boundary conditions (12D.2), as well as
Initial Position: (r, j, 0) = 0;
Initial Velocity:
t
(r, j, 0) = )
1
(r, j).
_
for all (r, j) X.
Proof. Exercise 12D.2 (a) Use the hypothesis (12D.5) and Proposition 0F.1 on E _
page 565 to conclude that
2
t
(r, j; t)
unif

n,m=1
_
n
2
+:
2
1
n,m
sin(nr) sin(:j) cos
_
_
n
2
+:
2
t
_
unif
(r, j; t),
(b) Check that the Fourier series (12D.6) converges uniformly.
(c) Use Theorem 9A.3(d)[ii] on page 183 to conclude that (r, j; t) satises Dirichlet
(d) Set t = 0 to check the initial position.
(e) Set t = 0 and use Proposition 0F.1 on page 565 to check initial velocity.
Remark. Note that it is important in these theorems not only for the Fourier
series (12D.4) and (12D.6) to converge uniformly, but also for their formal second
derivative series to converge uniformly. This is not guaranteed. This is the reason
for imposing the hypotheses (12D.3) and (12D.5).
Example 12D.4. Suppose
)
1
(r, j) =
16
2
99
n,m=1
both odd
1
n :
sin(nr) sin(:j)
(This is a partial sum of the two-dimensional Fourier sine series for the constant
function

)
1
(r, j) 1, from Example 9A.2 on page 182). Then the solution
to the two-dimensional wave equation, with homogeneous Dirichlet boundary
conditions and initial velocity )
1
, is given:
n(r, j; t)

L2
16
2
99
n,m=1
both odd
1
n :
n
2
+:
2
sin(nr) sin(:j) sin
_
_
n
2
+:
2
t
_
.
Question: Why cant we apply Theorem 12D.3 to the full Fourier series for
the function )
1
= 1? (Hint: Is (12D.5) satised?)
Question: For the solutions of the heat equation and Poisson equation,
in Propositions 12B.1, 12B.3, and 12C.1, we did not need to impose explicit
hypotheses guaranteeing the uniform convergence of the given series (and its
derivatives). But we do need explicit hypotheses to get convergence for the wave
equation. Why is this?
12E Practice problems
1. Let )(j) = 4 sin(5j) for all j [0, ].
(a) Solve the two-dimensional Laplace Equation (n = 0) on the
square domain X = [0, ] [0, ], with nonhomogeneous Dirichlet
boundary conditions:
n(r, 0) = 0 and n(r, ) = 0, for all r [0, ]
n(0, j) = 0 and n(, j) = )(j), for all j [0, ].
(b) Verify your solution to part (a) (i.e. check boundary conditions,
Laplacian, etc.).
2. Let )
1
(r, j) = sin(3r) sin(4j).
12E. Practice problems 263
(a) Solve the two-dimensional wave equation (
2
t
n = n) on the
square domain X = [0, ] [0, ], with on the square domain X =
[0, ] [0, ], with homogeneous Dirichlet boundary conditions,
and initial conditions:
Initial position: n(r, j, 0) = 0 for all (r, j) X
Initial velocity:
t
n(r, j, 0) = )
1
(r, j) for all (r, j) X
(b) Verify your that solution in part (a) satises the required initial con-
ditions (dont worry about boundary conditions or checking the wave
equation).
3. Solve the two-dimensional Laplace Equation / = 0 on the square do-
main X = [0, ]
2
, with inhomogeneous Dirichlet boundary conditions:
(a) /(, j) = sin(2j) and /(0, j) = 0, for all j [0, ];
/(r, 0) = 0 = /(r, ) for all r [0, ].
(b) /(, j) = 0 and /(0, j) = sin(4j), for all j [0, ];
/(r, ) = sin(3r); /(r, 0) = 0, for all r [0, ].
4. Let X = [0, ]
2
and let (r, j) = sin(r) sin(3j) +7 sin(4r) sin(2j). Solve
the Poisson Equation n(r, j) = (r, j). with homogeneous Dirichlet
5. Let X = [0, ]
2
. Solve the heat equation
t
n(r, j; t) = n(r, j; t) on X,
with initial conditions n(r, j; 0) = cos(5r) cos(j). and homogeneous
Neumann boundary conditions.
6. Let )(r, j) = cos(2r) cos(3j). Solve the following boundary value prob-
lems on the square domain X = [0, ]
2
(Hint: see problem #3 of 9C).
(a) Solve the two-dimensional heat equation
t
n = n, with homoge-
neous Neumann boundary conditions, and initial conditions n(r, j; 0) =
)(r, j).
(b) Solve the two-dimensional wave equation
2
t
n = n, with homoge-
neous Dirichlet boundary conditions, initial position n(r, j; 0) =
)(r, j) and initial velocity
t
n(r, j; 0) = 0.
(c) Solve the two-dimensional Poisson Equation n = ) with homo-
geneous Neumann boundary conditions.
(d) Solve the two-dimensional Poisson Equation n = ) with homo-
geneous Dirichlet boundary conditions.
(e) Solve the two-dimensional Poisson Equation = ) with inho-
mogeneous Dirichlet boundary conditions:
(, j) = sin(2j); (0, j) = 0 for all j [0, ].
(r, 0) = 0 = (r, ) for all r [0, ].
7. X = [0, ]
2
be the box of sidelength . Let )(r, j) = sin(3r) sinh(3j).
(Hint: see problem #4 of 9C).
(a) Solve the heat equation on X, with initial conditions n(r, j; 0) =
)(r, j), and homogeneous Dirichlet boundary conditions.
(b) Let T(r) = sin(3r). Solve the Laplace Equation n(r, j) = 0
on the box, with inhomogeneous Dirichlet boundary conditions:
n(r, ) = T(r) and n(r, 0) = 0 for r [0, ]; n(0, j) = 0 =
n(, j), for j [0, ].
(c) Solve the heat equation on the box with initial conditions on the
box X, with initial conditions n(r, j; 0) = 0, and the same inho-
mogeneous Dirichlet boundary conditions as in part (b).
265
Chapter 13
cube
Mathematical Analysis is as extensive as nature herself. Jean Joseph Fourier
The Fourier series technique used to solve BVPs on a square box extends
readily to 3-dimensional cubes, and indeed, to rectilinear domains in any number
of dimensions. As in Chapter 12, we will conne our exposition to the cube
[0, ]
3
, and assume that the physical constants in the various equations are all
set to one. Thus, the heat equation becomes
t
n = n, the wave equation is
2
t
n = n, etc. This allows us to develop the solution methods with minimum
technicalities. The extension of each solution method to equations with arbitrary
physical constants on an arbitrary box [0, A] [0, Y ] [0, 7] (for some A, Y, 7
0) is left as a straightforward (but important!) exercise.
We will use the following notation:
The cube of dimensions is denoted X = [0, ] [0, ] [0, ] =
[0, ]
3
.
A point in the cube will be indicated by a vector x = (r
1
, r
2
, r
3
), where
0 r
1
, r
2
, r
3
.
If ) : X 1 is a function on the cube, then
)(x) =
2
1
)(x) +
2
2
)(x) +
2
3
)(x).
A triple of natural numbers will be denoted by n = (n
1
, n
2
, n
3
), where
n
1
, n
2
, n
3
N := 0, 1, 2, 3, 4, . . .. Let N
3
be the set of all triples n =
(n
1
, n
2
, n
3
), where n
1
, n
2
, n
3
N. Thus, an expression of the form
nN
3
(something about n)
266 DRAFT Chapter 13. Boundary value problems on a cube
should be read as:
n
1
=0
n
2
=0
n
3
=0
(something about (n
1
, n
2
, n
3
)).
Let N
+
:= 1, 2, 3, 4, . . . be the set of nonzero natural numbers, and let
N
3
+
be the set of all such triples. Thus, an expression of the form
nN
3
+
(something about n)
should be read as:
n
1
=1
n
2
=1
n
3
=1
(something about (n
1
, n
2
, n
3
)).
For any n N
3
+
, S
n
(x) = sin(n
1
r
1
) sin(n
2
r
2
) sin(n
3
r
3
). The Fourier
sine series of a function )(x) thus has the form: )(x) =
nN
3
+
1
n
S
n
(x)
For any n N
3
, C
n
(x) = cos(n
1
r
1
) cos(n
2
r
2
) cos(n
3
r
3
). The Fourier
cosine series of a function )(x) thus has the form: )(x) =
nN
3
n
C
n
(x)
For any n N
3
, let |n| =
_
n
2
1
+n
2
2
+n
2
3
. In particular, note that:
S
n
= |n|
2
S
n
, and C
n
= |n|
2
C
n
(Exercise 13.1) E _
13A The heat equation on a cube
Prerequisites: 9B, 5B, 5C, 1B(ii). Recommended: 11A, 12B(i), 7C(v).
Proposition 13A.1. (Heat equation; homogeneous Dirichlet BC)
Consider the cube X = [0, ]
3
, and let ) : X 1 be some function describ-
ing an initial heat distribution. Suppose ) has Fourier sine series )(x)

L2
nN
3
+
1
n
S
n
(x).
Dene the function n : X 1
,
1 by:
n(x; t)

L2
nN
3
+
1
n
S
n
(x) exp
_
|n|
2
t
_
.
Then n is the unique solution to the heat equation
t
n = n, with homoge-
neous Dirichlet boundary conditions
n(r
1
, r
2
, 0; t) = n(r
1
, r
2
, ; t) = n(r
1
, 0, r
3
; t) (see Figure 13A.1A)
= n(r
1
, , r
3
; t) = n(0, r
2
, r
3
; t) = n(, r
2
, r
3
, ; t) = 0,
13A. The heat equation on a cube 267
(x
1
,,x
3
)
(B)
(A)
(x
1
,x
2
,)
x
1
x
3
x
2
1
u
3
u
1
u
3
u
2
u
2
u
(x
1
,0,x
3
)
(x
1
,x
2
,0)
(
x
1
,
,
x
3
)
(
x
1
,
0
,
x
3
)
Figure 13A.1: Boundary conditions on a cube: (A) Dirichlet. (B) Neumann.
and initial conditions: n(x; 0) = )(x).
+
.
Proof. Exercise 13A.1 2 E _
Example: An ice cube of dimensions is removed from a freezer
(ambient temperature 10
o
C) and dropped into a pitcher of freshly brewed tea
(initial temperature +90
o
C). We want to compute how long it takes the ice cube
to melt.
Answer: We will assume that the cube has an initially uniform temperature of
10
o
C and is completely immersed in the tea
1
. We will also assume that the
pitcher is large enough that its temperature doesnt change during the experi-
ment.
We assume the outer surface of the cube takes the temperature of the sur-
rounding tea. By subtracting 90 from the temperature of the cube and the
water, we can set the water to have temperature 0 and the cube, 100. Hence,
we assume homogeneous Dirichlet boundary conditions; the initial temperature
distribution is a constant function: )(x) = 100. The constant function 100
has Fourier sine series:
100

L2
6400
nN
3
+
n
1
,n
2
,n
3
all odd
1
n
1
n
2
n
3
S
n
(x).
(Exercise 13A.2 Verify this Fourier series). Let be the thermal conductivity E _
1
Unrealistic, since actually the cube oats just at the surface.
of the ice. Thus, the time-varying thermal prole of the cube is given
2
n(x; t)

L2
6400
nN
3
+
n
1
,n
2
,n
3
all odd
1
n
1
n
2
n
3
S
n
(x) exp
_
|n|
2
t
_
.
Thus, to determine how long it takes the cube to melt, we must solve for the min-
imum value of t such that n(x, t) 90 everywhere (recall than 90 corresponds
to 0
o
C.). The coldest point in the cube is always at its center (Exercise 13A.3), E _
which has coordinates
_
2
,

2
,

2
_
, so we need to solve for t in the inequality
n
__
2
,

2
,

2
_
; t
_
90, which is equivalent to
90
3
6400

nN
3
+
n
1
,n
2
,n
3
all odd
1
n
1
n
2
n
3
S
n
_
2
,

2
,

2
_
exp
_
|n|
2
t
_
=
nN
3
+
n
1
,n
2
,n
3
all odd
1
n
1
n
2
n
3
sin
_
n
1
2
_
sin
_
n
2
2
_
sin
_
n
3
2
_
exp
_
|n|
2
t
_
(7C.5)
k
1
,k
2
,k
3
N
+
(1)
k
1
+k
2
+k
3
exp
_

_
(2/
1
+ 1)
2
+ (2/
2
+ 1)
2
+ (2/
3
+ 1)
2
t
_
(2/
1
+ 1) (2/
2
+ 1) (2/
3
+ 1)
.
where (7C.5) is by eqn. (7C.5) on p. 147. The solution of this inequality is
Exercise 13A.4 . E _
Exercise 13A.5. Imitating Proposition 13A.1, nd a Fourier series solution to the E _
initial value problem for the free Schrodinger equation
i
t
=
1
2
,
on the cube X = [0, ]
3
, with homogeneous Dirichlet boundary conditions. Prove that
your solution converges, satises the desired initial conditions and boundary conditions,
and satises the Schrodinger equation.
Proposition 13A.2. (Heat equation; homogeneous Neumann BC)
2
Actually, this is physically unrealistic for two reasons. First, as the ice melts, additional
thermal energy is absorbed in the phase transition from solid to liquid. Second, once part of
the ice cube has melted, its thermal properties change; liquid water has a dierent thermal
conductivity, and in addition, transports heat through convection.
13B. The Dirichlet problem on a cube 269
T
B
S
N
E
W
(B)
East
West
U
p
D
o
w
n
N
o
r
t
h
S
o
u
t
h
x
1
x
3
x
2
1
0
0
0
0
0
(A)
Figure 13B.1: Dirichlet boundary conditions on a cube (A) Constant; Nonho-
mogeneous on one side only. (B) Arbitrary nonhomogeneous on all sides.
Consider the cube X = [0, ]
3
, and let ) : X 1 be some function
describing an initial heat distribution. Suppose ) has Fourier Cosine Series
)(x)

L2
nN
3
n
C
n
(x). Dene the function n : X 1
,
1 by:
n(x; t)

L2
nN
3
n
C
n
(x) exp
_
|n|
2
t
_
.
Then n is the unique solution to the heat equation
t
n = n, with homoge-
neous Neumann boundary conditions
3
n(r
1
, r
2
, 0; t) =
3
n(r
1
, r
2
, ; t) =
2
n(r
1
, 0, r
3
; t) =
2
n(r
1
, , r
3
; t) =
1
n(0, r
2
, r
3
; t) =
1
n(, r
2
, r
3
, ; t) = 0. (see Figure 13A.1B)
and initial conditions: n(x; 0) = )(x).
+
.
13B The Dirichlet problem on a cube
Prerequisites: 9B, 5C(i), 1C. Recommended: 7C(v), 12A.
Proposition 13B.1. (Laplace Equation; one constant nonhomog. Dirichlet BC)
Let X = [0, ]
3
, and consider the Laplace equation n = 0, with nonho-
mogeneous Dirichlet boundary conditions (see Figure 13B.1A):
n(r
1
, 0, r
3
) = n(r
1
, , r
3
) = n(0, r
2
, r
3
) = n(, r
2
, r
3
, ) = 0; (13B.1)
n(r
1
, r
2
, 0) = 0;
n(r
1
, r
2
, ) = 1. (13B.2)
The unique solution to this problem is the function n : X 1 dened
n(r
1
, r
2
, r
3
)

L2
n,m=1
n,m both odd
16
n: sinh(
n
2
+:
2
)
sin(nr) sin(:j)sinh(
_
n
2
+:
2
r
3
).
for all (r
1
, r
2
, r
3
) X. Furthermore, this series converges semiuniformly on
int (X).
Proof. Exercise 13B.1 (a) Check that the series and its formal Laplacian both E _
converge semiuniformly on int (X). (b) Check that each of the functions n
n,m
(x) =
sin(nr) sin(:j) sinh(
n
2
+:
2
r
3
) satises the Laplace equation and the rst bound-
ary condition (13B.1). (c) To check that the solution also satises the boundary
condition (13B.2), subsititute r
2
= to get:
n(r
1
, r
2
, ) =
n,m=1
n,m both odd
16
n: sinh(
n
2
+:
2
)
sin(nr) sin(:j) sinh(
_
n
2
+:
2
)
=
n,m=1
n,m both odd
16
n:
sin(nr) sin(:j)

L2
1,
because this is the Fourier sine series for the function /(r
1
, r
2
) = 1, by Example 9A.2
on page 182.
(d) Apply Theorem 5D.5(a) on page 88 to conclude that this solution is unique. 2
Proposition 13B.2. (Laplace Equation; arbitrary nonhomogeneous Dirichlet BC)
Let X = [0, ]
3
, and consider the Laplace equation / = 0, with nonho-
mogeneous Dirichlet boundary conditions (see Figure 13B.1B):
/(r
1
, r
2
, 0) = 1(r
1
, r
2
) /(r
1
, r
2
, ) = l(r
1
, r
2
)
/(r
1
, 0, r
3
) = o(r
1
, r
3
) /(r
1
, , r
3
) = (r
1
, r
3
)
/(0, r
2
, r
3
) = \(r
2
, r
3
) /(, r
2
, r
3
, ) = 1(r
2
, r
3
)
13B. The Dirichlet problem on a cube 271
where 1(r
1
, r
2
), l(r
1
, r
2
), o(r
1
, r
3
), (r
1
, r
3
), \(r
2
, r
3
), and 1(r
2
, r
3
)
are six functions. Suppose that these functions have two-dimensional Fourier
sine series:
1(r
1
, r
2
)

L2
n
1
,n
2
=1
1
n
1
,n
2
sin(n
1
r
1
) sin(n
2
r
2
);
l(r
1
, r
2
)

L2
n
1
,n
2
=1
l
n
1
,n
2
sin(n
1
r
1
) sin(n
2
r
2
);
o(r
1
, r
3
)

L2
n
1
,n
3
=1
o
n
1
,n
3
sin(n
1
r
1
) sin(n
3
r
3
);
(r
1
, r
3
)

L2
n
1
,n
3
=1
n
1
,n
3
sin(n
1
r
1
) sin(n
3
r
3
);
\(r
2
, r
3
)

L2
n
2
,n
3
=1
\
n
2
,n
3
sin(n
2
r
2
) sin(n
3
r
3
);
1(r
2
, r
3
)

L2
n
2
,n
3
=1
1
n
2
,n
3
sin(n
2
r
2
) sin(n
3
r
3
).
Then the unique solution to this problem is the function:
/(x) = d(x) +n(x) +:(x) +n(x) +n(x) +c(x)
d(r
1
, r
2
, r
3
)

L2
n
1
,n
2
=1
1
n
1
,n
2
sinh
_
_
n
2
1
+n
2
2
_ sin(n
1
r
1
) sin(n
2
r
2
) sinh
_
_
n
2
1
+n
2
2
r
3
_
;
n(r
1
, r
2
, r
3
)

L2
n
1
,n
2
=1
l
n
1
,n
2
sinh
_
_
n
2
1
+n
2
2
_ sin(n
1
r
1
) sin(n
2
r
2
) sinh
_
_
n
2
1
+n
2
2
( r
3
)
_
;
:(r
1
, r
2
, r
3
)

L2
n
1
,n
3
=1
o
n
1
,n
3
sinh
_
_
n
2
1
+n
2
3
_ sin(n
1
r
1
) sin(n
3
r
3
) sinh
_
_
n
2
1
+n
2
3
r
2
_
;
n(r
1
, r
2
, r
3
)

L2
n
1
,n
3
=1
n
1
,n
3
sinh
_
_
n
2
1
+n
2
3
_ sin(n
1
r
1
) sin(n
3
r
3
) sinh
_
_
n
2
1
+n
2
3
( r
2
)
_
;
n(r
1
, r
2
, r
3
)

L2
n
2
,n
3
=1
\
n
2
,n
3
sinh
_
_
n
2
2
+n
2
3
_ sin(n
2
r
2
) sin(n
3
r
3
) sinh
_
_
n
2
2
+n
2
3
r
1
_
;
c(r
1
, r
2
, r
3
)

L2
n
2
,n
3
=1
1
n
2
,n
3
sinh
_
_
n
2
2
+n
2
3
_ sin(n
2
r
2
) sin(n
3
r
3
) sinh
_
_
n
2
2
+n
2
3
( r
1
)
_
.
Furthermore, these six series converge semiuniformly on int (X).
13C The Poisson problem on a cube
Prerequisites: 9B, 5C, 1D. Recommended: 11C, 12C, 7C(v).
Proposition 13C.1. Poisson Problem on Cube; homogeneous Dirichlet BC
Let X = [0, ]
3
, and let : X 1 be some function with semiuniformly
convergent Fourier sine series: (x)

L2
nN
3
+
Q
n
S
n
(x). Dene the function
n : X 1 by
n(x)
unif
nN
3
+
Q
n
|n|
2
S
n
(x), for all x X.
Then n is the unique solution to the Poisson equation n(x) = (x), satis-
fying homogeneous Dirichlet boundary conditions n(r
1
, r
2
, 0) = n(r
1
, r
2
, ) =
n(r
1
, 0, r
3
) = n(r
1
, , r
3
) = n(0, r
2
, r
3
) = n(, r
2
, r
3
, ) = 0.
Proposition 13C.2. Poisson Problem on Cube; homogeneous Neumann BC
Let X = [0, ]
3
, and let : X 1 be some function with semiuniformly
convergent Fourier cosine series: (x)

L2
nN
3
+
Q
n
C
n
(x).
Suppose Q
0,0,0
= 0. Fix some constant 1 1, and dene the function
n : X 1 by
n(x)
unif
nN
3
n
1
,n
2
,n
3
not all zero
Q
n
|n|
2
C
n
(x) + 1, for all x X. (13C.1)
Then n is a solution to the Poisson equation n(x) = (x), satisfying homoge-
neous Neumann boundary conditions
3
n(r
1
, r
2
, 0) =
3
n(r
1
, r
2
, ) =
2
n(r
1
, 0, r
3
) =
2
n(r
1
, , r
3
) =
1
n(0, r
2
, r
3
) =
1
n(, r
2
, r
3
, ) = 0.
ditions have the form (13C.1).
If Q
0,0,0
273
Chapter 14
Boundary value problems in
polar coordinates
The source of all great mathematics is the special case, the concrete example. It is frequent
in mathematics that every instance of a concept of seemingly great generality is in essence the
same as a small and concrete special case. Paul Halmos
14A Introduction
Prerequisites: 0D(ii).
When solving a boundary value problem, the shape of the domain dictates
the choice of coordinate system. Seek the coordinate system yielding the simplest
description of the boundary. For rectangular domains, Cartesian coordinates are
the most convenient. For disks and annuli in the plane, polar coordinates are
a better choice. Recall that polar coordinates (:, ) on 1
2
are dened by the
transformation:
r = : cos() and j = : sin(). (Figure 14A.1A)
with reverse transformation:
: =
_
r
2
+j
2
and = arctan
_
j
r
_
.
Here, the coordinate : ranges over 1
,
, while the variable ranges over [, ).
(Clearly, we could let range over any interval of length 2; we just nd [, )
the most convenient).
The three domains we will examine are:
| = (:, ) ; : 1, the disk of radius 1; see Figure 14A.1B. For simplic-
ity we will usually assume 1 = 1.
274 DRAFT Chapter 14. Boundary value problems in polar coordinates
R R
R
min
R
max
(B)
(C)
(D)
(A)
r
x
y
Figure 14A.1: (A) Polar coordinates; (B) The disk |; (C) The codisk |
;
(D) The annulus A.
|
= (:, ) ; 1 :, the codisk or punctured plane of radius 1; see

Figure 14A.1C. For simplicity we will usually assume 1 = 1.
A = (:, ) ; 1
min
: 1
max
, the annulus, of inner radius 1
min
and
outer radius 1
max
; see Figure 14A.1D.
The boundaries of these domains are circles. For example, the boundary of
the disk | of radius 1 is the circle:
| = S = (:, ) ; : = 1 .
The circle can be parameterized by a single angular coordinate [, ).
Thus, the boundary conditions will be specied by a function / : [, ) 1.
Note that, if /() is to be continuous as a function on the circle, then it must be
2-periodic as a function on [, ).
In polar coordinates, the Laplacian is written:
n =
2
r
n +
1
:

r
n +
1
:
2

2
n. (14A.1)
(Exercise 14A.1) E _
14B The Laplace equation in polar coordinates
14B(i) Polar harmonic functions
Prerequisites: 0D(ii), 1C.
275
1
0
1
2
1
0
1
2
2 0 2 4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
2
(r, )
3
(r, )
4
(r, )
5
(r, )
5
(r, )
= r
2
cos (2) = r
3
cos (3) = r
4
cos (4) = r
5
cos (5) = r
6
cos (6)
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
8
0
.
6
0
.
4
0
.
2
0
0
.
2
0
.
4
0
.
6
0
.
8
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
2
(r, )
3
(r, )
4
(r, )
5
(r, )
5
(r, )
= r
2
sin (2) = r
3
sin (3) = r
4
sin (4) = r
5
sin (5) = r
6
sin (6)
Figure 14B.1:
n
and
n
for n = 2...6 (rotate page).
276 DRAFT
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2 3
1
(:, )
2
(:, )
3
(:, )
4
(:, )
= :
1
cos (2) = :
2
cos (3) = :
3
cos (4) = :
4
cos (5)
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2 3
1
0
1
2
1
0
1
2
1 0
1
2
1
(:, )
2
(:, )
3
(:, )
4
(:, )
= :
1
sin (2) = :
2
sin (3) = :
3
sin (4) = :
4
sin (5)
Figure 14B.2:
n
and
n
for n = 1...4 (rotate page). Note that these plots have
been truncated to have vertical bounds 3, because these functions explode to
at zero.
14B. The Laplace equation in polar coordinates 277
2
1
0
1
2
2
1
0
1
2
2
1.5
1
0.5
0
0.5
Figure 14B.3:
0
(:, ) = log [:[ (vertically truncated near zero).
0
1
2
3
4
5
6
7
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
PSfrag replacements
r
r
2
r
3
r
4
-2
-1
0
1
2
3
4
5
0 0.5 1 1.5 2 2.5 3
PSfrag replacements
l
o
g
(
r
)
1
,
r
1
,
r
2
1
,
r
3
(A): r
n
, for n = 1, 2, 3, 4; (B): log(r) and
1
x
n
, for n = 1, 2, 3
(these plots are vertically truncated).
Figure 14B.4: Radial growth/decay of polar-separated harmonic functions.
The following important harmonic functions separate in polar coordinates:
n
(:, ) = cos(n) :
n
;
n
(:, ) = sin(n) :
n
; for n N
+
(Fig.14B.1)
n
(:, ) =
cos(n)
:
n
;
n
(:, ) =
sin(n)
:
n
; for n N
+
(Fig.14B.2)
0
(:, ) = 1 and
0
(:, ) = log(:) (Fig.14B.3)
Proposition 14B.1. The functions
n
,
n
,
n
, and
n
are harmonic, for
all n N.
Proof. See practice problems #1 to #5 in 14I. 2
Exercise 14B.1. (a) Show that
1
(:, ) = r and
1
(:, ) = j in Cartesian E _
coordinates.
(b) Show that
2
(:, ) = r
2
j
2
and
2
(:, ) = 2rj in Cartesian coordinates.
(c) Dene 1
n
: C C by 1
n
(.) := .
n
. Show that
n
(r, j) = Re [1
n
(r +ji)] and
n
(r, j) = Im[1
n
(r +ji)].
(d) (Hard) Show that
n
can be written as a homogeneous polynomial of degree n
in r and j.
(e) Show that, if (r, j) | (i.e. if r
2
+j
2
= 1), then
N
(r, j) =
N
(r), where
N
(r) := 2
(N1)
r
N
+
N
2
|
n=1
(1)
n
2
(N12n)
n
_
n 1
n 1
_
r
(N2n)
.
is the th Chebyshev polynomial. (For more information, see [Bro89, 3.4]).
We will solve the Laplace equation in polar coordinates by representing so-
lutions as sums of these simple functions. Note that
n
and
n
are bounded
at zero, but unbounded at innity (Figure 14B.4(A) shows the radial growth
of
n
and
n
). Conversely,
n
and
n
are unbounded at zero, but bounded at
innity) (Figure 14B.4(B) shows the radial decay of
n
and
n
). Finally,
0
being constant, is bounded everywhere, while
0
is unbounded at both 0 and
(see Figure 14B.4B). Hence, when solving BVPs in a neighbourhood around
zero (e.g. the disk), it is preferable to use
0
,
n
and
n
. When solving BVPs
on an unbounded domain (i.e. one containing innity) it is preferable to use
0
,
n
and
n
. When solving BVPs on a domain containing neither zero nor
innity (e.g. the annulus), we use all of
n
,
n
,
n
,
n
,
0
, and
0
.
14B(ii) Boundary value problems on a disk
Prerequisites: 5C, 14A, 14B(i), 8A, 0F.
Proposition 14B.2. (Laplace Equation, Unit Disk, nonhomog. Dirichlet BC)
Let | = (:, ) ; : 1 be the unit disk, and let / L
2
[, ) be some func-
tion. Consider the Laplace equation n = 0, with nonhomogeneous Dirichlet
n(1, ) = /(), for all [, ). (14B.1)
Suppose / has real Fourier series: /()

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n).
Then the unique solution to this problem is the function n : | 1 dened:
n(:, )

L2
0
+
n=1
n
(:, ) +
n=1
1
n
n
(:, )
1
0.5
0
0.5
1
1
0
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
1
0.5
0
0.5
1
(A): A bent circular wire frame: (B): A bubble in the frame:
/() = sin(3). n(:, ) = :
3
sin(3).
Figure 14B.5: A soap bubble in a bent wire frame.
=
0
+
n=1
n
cos(n) :
n
+
n=1
1
n
sin(n) :
n
(14B.2)
Furthermore, the series (14B.2) converges semiuniformly to n on int (|).
Proof. Exercise 14B.2 (a) Fix 1 < 1 and let |(1) := (:, ) ; : < 1. Show E _
that on the domain |(1), the conditions of Proposition 0F.1 on page 565 are satised;
use this to show that
n(:, )
unif
n=1
n

n
(:, ) +
n=1
1
n

n
(:, )
for all (:, ) |(1). Now use Proposition 14B.1 on page 277 to deduce that
n(:, ) = 0 for all : 1. Since this works for any 1 < 1, conclude that n 0
on |.
(b) To check that n also satises the boundary condition (14B.1), substitute : = 1
into (14B.2) to get: n(1, )

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n) = /().
(c) Use Proposition 5D.5(a) on page 88 to conclude that this solution is unique. 2
Example 14B.3. Take a circular wire frame of radius 1, and warp it so that its
vertical distortion is described by the function /() = sin(3), shown in Figure
14B.5(A). Dip the frame into a soap solution to obtain a bubble with the bent
wire as its boundary. What is the shape of the bubble?
Solution: A soap bubble suspended from the wire is a minimal surface,
and minimal surfaces of low curvature are well-approximated by harmonic
functions. Let n(:, ) be a function describing the bubble surface. As long as
the distortion /() is relatively small, n(:, ) will be a solution to Laplaces
equation, with boundary conditions n(1, ) = /(). Thus, as shown in Figure
14B.5(B), n(:, ) = :
3
sin(3).
Exercise 14B.3. Let n(r, ) be a solution to the Dirichlet problem with boundary E _
conditions n(1, ) = /(). Let 0 be the center of the disk (i.e. the point with radius 0).
Use Proposition 14B.2 to prove that n(0) =
1
2
_

/() d.
Proposition 14B.4. (Laplace Equation, Unit Disk, nonhomog. Neumann BC)
Let | = (:, ) ; : 1 be the unit disk, and let / L
2
[, ). Consider
the Laplace equation n = 0, with nonhomogeneous Neumann boundary con-
ditions:
r
n(1, ) = /(), for all [, ). (14B.3)

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n).
If
0
= 0, then the solutions to this problem are all functions n : | 1 of
the form
n(:, )

L2
C +
n=1
n
n

n
(:, ) +
n=1
1
n
n

n
(:, )
= C +
n=1
n
n
cos(n) :
n
+
n=1
1
n
n
sin(n) :
n
(14B.4)
where C is any constant. Furthermore, the series (14B.4) converges semiuni-
formly to n on int (|).
However, if
0
,= 0, then there is no solution.
Proof.
Claim 1: For any : < 1,
n=1
n
2
[
n
[
n
:
n
+
n=1
n
2
[1
n
[
n
:
n
< .
Proof. Let ` = max
_
max[
n
[
n=1
, max[1
n
[
n=1
_
. Then
n=1
n
2
[
n
[
n
:
n
+
n=1
n
2
[1
n
[
n
:
n
n=1
n
2
`
n
:
n
+
n=1
n
2
`
n
:
n
= 2`
n=1
n:
n
. (14B.5)
Let )(:) =
1
1 :
. Then )
t
(:) =
1
(1 :)
2
. Recall that, for [:[ < 1,
)(:) =
n=0
:
n
. Thus, )
t
(:) =
n=1
n:
n1
=
1
:
n=1
n:
n
. Hence, the right
hand side of eqn.(14B.5) is equal to
2`
n=1
n:
n
= 2`: )
t
(:) = 2`:
1
(1 :)
2
< ,
for any : < 1.
Claim 1
Let 1 < 1 and let |(1) = (:, ) ; : 1 be the disk of radius 1. If n(:, ) =
C +
n=1
n
n

n
(:, ) +
n=1
1
n
n

n
(:, ), then for all (:, ) |(1),
n(:, )
unif
n=1
n
n

n
(:, ) +
n=1
1
n
n

n
(:, )
()
n=1
n
n
(0) +
n=1
1
n
n
(0) = 0,
on |(1). Here,
unif
is by Proposition 0F.1 on page 565 and Claim 1, while
() is by Proposition 14B.1 on page 277.
To check boundary conditions, observe that, for all 1 < 1 and all (:, ) |(1),
r
n(:, )
unif
n=1
n
n

r

n
(:, ) +
n=1
1
n
n

r

n
(:, )
=
n=1
n
n
n:
n1
cos(n) +
n=1
1
n
n
n:
n1
sin(n)
=
n=1
n
:
n1
cos(n) +
n=1
1
n
:
n1
sin(n).
Here
unif
is by Proposition 0F.1 on page 565. Hence, letting 1 1, we get
n(1, ) =
r
n(1, ) =
n=1
n
(1)
n1
cos(n) +
n=1
1
n
(1)
n1
sin(n)
=
n=1
n
cos(n) +
n=1
1
n
sin(n)

L2
/(),
as desired. Here,

L2
is because this is the Fourier Series for /(), assuming
0
= 0. (If
0
,= 0, then this solution doesnt work.)
Finally, Proposition 5D.5(c) on page 88 implies that this solution is unique up
to addition of a constant. 2
0.5
1 0.5
-1
1
-0.5 -1
-0.5
0
0
-1
-0.5
0
0.5
1-1
-0.5
0
0.5
1
-0.8
-0.4
0
0.4
0.8
(A) Contour plot (B) Surface plot
Figure 14B.6: The electric potential deduced from Scullys voltage measurements in Example
14B.5.
Remark. Physically speaking, why must
0
= 0?
If n(:, ) is an electric potential, then
r
n is the radial component of the
electric eld. The requirement that
0
= 0 is equivalent to requiring that the
net electric ux entering the disk is zero, which is equivalent (via Gausss law)
to the assertion that the net electric charge contained in the disk is zero. If
0
,= 0, then the net electric charge within the disk must be nonzero. Thus,
if : | 1 is the charge density eld, then we must have , 0. However,
= n (see Example 1D.2 on page 14), so this means n ,= 0, which means n
is not harmonic.
Example 14B.5. While covertly investigating mysterious electrical phenomena
on a top-secret military installation in the Nevada desert, Mulder and Scully
are trapped in a cylindrical concrete silo by the Cancer Stick Man. Scully
happens to have a voltimeter, and she notices an electric eld in the silo.
Walking around the (circular) perimeter of the silo, Scully estimates the radial
component of the electric eld to be the function /() = 3 sin(7) cos(2).
Estimate the electric potential eld inside the silo.
Solution: The electric potential will be a solution to Laplaces equation,
with boundary conditions
r
n(1, ) = 3 sin(7) cos(2). Thus,
n(:, ) = C +
3
7
sin(7) :
7
1
2
cos(2) :
2
. (see Figure 14B.6)
Question: Moments later, Mulder repeats Scullys experiment, and nds that
the perimeter eld has changed to /() = 3 sin(7)cos(2)+6. He immediately
suspects that an Alien Presence has entered the silo. Why?
14B(iii) Boundary value problems on a codisk
Prerequisites: 5C, 14A, 14B(i), 8A, 0F. Recommended: 14B(ii).
We will now solve the Dirichlet problem on an unbounded domain: the codisk
|
:= (:, ) ; 1 :, [, ) .
Physical Interpretations:
Chemical Concentration: Suppose there is an unknown source of some chem-
ical hidden inside the disk, and that this chemical diuses into the sur-
rounding medium. Then the solution function n(:, ) represents the equi-
librium concentration of the chemical. In this case, it is reasonable to
expect n(:, ) to be bounded at innity, by which we mean:
lim
r
[n(:, )[ ,= , for all [, ). (14B.6)
Otherwise the chemical concentration would become very large far away
from the center, which is not realistic.
Electric Potential: Suppose there is an unknown charge distribution inside
the disk. Then the solution function n(:, ) represents the electric potential
eld generated by this charge. Even though we dont know the exact charge
distribution, we can use the boundary conditions to extrapolate the shape
of the potential eld outside the disk.
If the net charge within the disk is zero, then the electric potental far
away from the disk should be bounded (because from far away, the charge
distribution inside the disk looks neutral); hence, the solution n(:, ) will
again satisfy the Boundedness Condition (14B.6).
However, if there is a nonzero net charge within the the disk, then the
electric potential will not be bounded (because, even from far away, the
disk still looks charged). Nevertheless, the electric eld generated by this
potential should still decay to zero (because the inuence of the charge
should be weak at large distances). This means that, while the potential is
unbounded, the gradient of the potential must decay to zero near innity.
In other words, we must impose the decaying gradient condition:
lim
r
n(:, ) = 0, for all [, ). (14B.7)
Proposition 14B.6. (Laplace equation, Codisk, nonhomog. Dirichlet BC)
Let |
= (:, ) ; 1 : be the codisk, and let / L

2
[, ). Consider the
Laplace equation n = 0, with nonhomogeneous Dirichlet boundary condi-
tions:
n(1, ) = /(), for all [, ). (14B.8)

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n).
Then the unique solution to this problem which is bounded at innity as in
(14B.6) is the function n : |
1 dened:
n(:, )

L2
0
+
n=1
n
cos(n)
:
n
+
n=1
1
n
sin(n)
:
n
(14B.9)
Furthermore, the series (14B.9) converges semiuniformly to n on int
_
|
_
.
Proof. Exercise 14B.4 (a) To show that n is harmonic, apply eqn.(14A.1) on E _
page 274 to get
.u(r, ) =
2
r
_

n=1
An
cos(n)
r
n
+
n=1
Bn
sin(n)
r
n
_
+
1
r
r
_

n=1
An
cos(n)
r
n
+
n=1
Bn
sin(n)
r
n
_
+
1
r
2
n=1
An
cos(n)
r
n
+
n=1
Bn
sin(n)
r
n
_
. (14B.10)
Now let 1 1. Check that, on the domain |
(1) = (:, ) ; : 1, the conditions

of Proposition 0F.1 on page 565 are satised; use this to simplify the expression
(14B.10). Finally, apply Proposition 14B.1 on page 277 to deduce that n(:, ) = 0
for all : 1. Since this works for any 1 1, conclude that n 0 on |
.
(b) To check that the solution also satises the boundary condition (14B.8), subsititute
: = 1 into (14B.9) to get: n(1, )

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n) = /().
(c) Use Proposition 5D.5(a) on page 88 to conclude that this solution is unique. 2
Example 14B.7. An unknown distribution of electric charges lies inside the
unit disk in the plane. Using a voltimeter, the electric potential is measured
along the perimeter of the circle, and is approximated by the function /() =
sin(2) +4 cos(5). Far away from the origin, the potential is found to be close
to zero. Estimate the electric potential eld.
with boundary conditions n(1, ) = sin(2)+4 cos(5). Far away, the potential
apparently remains bounded. Thus, as shown in Figure 14B.7,
n(:, ) =
sin(2)
:
2
+
4 cos(5)
:
5
.
-2
-1
-1
-3
0 2
2
-2
3
-3 3
1
0
1
-3
-2
-1
0
1
2
3-3
-2
-1
0
1
2
3
-4
-2
0
2
4
(unit disk occulted) (unit disk deleted)
Figure 14B.7: The electric potential deduced from voltage measurements in Example 14B.7.
Remark. Note that, for any constant C 1, another solution to the Dirichlet
problem with boundary conditions (14B.8) is given by the function
n(:, ) =
0
+ C log(:) +
n=1
n
cos(n)
:
n
+
n=1
1
n
sin(n)
:
n
.
(Exercise 14B.5). However, unless C = 0, this will not be bounded at innity. E _
Proposition 14B.8. (Laplace equation, Codisk, nonhomog. Neumann BC)
Let |
= (:, ) ; 1 : be the codisk, and let / L

2
[, ). Consider the
Laplace equation n = 0, with nonhomogeneous Neumann boundary condi-
tions:
n(1, ) =
r
n(1, ) = /(), for all [, ). (14B.11)

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n).
Fix a constant C 1, and dene n : |
1 by:
n(:, )

L2
C +
0
log(:) +
n=1
n
n
cos(n)
:
n
+
n=1
1
n
n
sin(n)
:
n
(14B.12)
Then n is a solution to the Laplace equation, with nonhomogeneous Neumann
boundary conditions (14B.11), and furthermore, obeys the Decaying Gradient
Condition (14B.7) on p.283. Furthermore, all harmonic functions satisfying equa-
tions (14B.11) and (14B.7) must be of the form (14B.12). However, the solution
(14B.12) is bounded at innity as in (14B.6) if and only if
0
= 0.
Finally, the series (14B.12) converges semiuniformly to n on int
_
|
_
.
Proof. Exercise 14B.6 (a) To show that n is harmonic, apply eqn.(14A.1) on E _
page 274 to get
n(:, ) =
2
r
_
0
log(:)
n=1
n
n
cos(n)
:
n

n=1
1
n
n
sin(n)
:
n
_
+
1
:
r
_
0
log(:)
n=1
n
n
cos(n)
:
n

n=1
1
n
n
sin(n)
:
n
_
+
1
:
2
0
log(:)
n=1
n
n
cos(n)
:
n

n=1
1
n
n
sin(n)
:
n
_
. (14B.13)
Now let 1 1. Check that, on the domain |
(1) = (:, ) ; : 1, the conditions

of Proposition 0F.1 on page 565 are satised; use this to simplify the expression
(14B.13). Finally, apply Proposition 14B.1 on page 277 to deduce that n(:, ) = 0
for all : 1. Since this works for any 1 1, conclude that n 0 on |
.
(b) To check that the solution also satises the boundary condition (14B.11), subsi-
titute : = 1 into (14B.12) and compute the radial derivative (using Proposition 0F.1
on page 565) to get:
r
n(1, ) =
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n)

L2
/().
(c) Use Proposition 5D.5(c) on page 88 to show that this solution is unique up to
addition of a constant.
(d) What is the physical interpretation of
0
= 0? 2
Example 14B.9. An unknown distribution of electric charges lies inside the
unit disk in the plane. The radial component of the electric eld is measured
along the perimeter of the circle, and is approximated by the function /() =
0.9 + sin(2) + 4 cos(5). Estimate the electric potential potential (up to a
constant).
with boundary conditions
r
n(1, ) = 0.9 + sin(2) + 4 cos(5). Thus, as
shown in Figure 14B.8,
n(:, ) = C + 0.9 log(:) +
sin(2)
2 :
2
+
4 cos(5)
5 :
5
.
14B(iv) Boundary value problems on an annulus
Prerequisites: 5C, 14A, 14B(i), 8A, 0F. Recommended: 14B(ii), 14B(iii).
4
0
2
4 -4
-2
-2 0 2
-4
-4
-2
0
2
4
-4
-2
0
2
-2
4
-1.5
-1
-0.5
0
0.5
1
1.5
(unit disk occulted) (unit disk deleted)
Figure 14B.8: The electric potential deduced from eld measurements in Example 14B.9.
Proposition 14B.10. (Laplace Equation, Annulus, nonhomog. Dirichlet BC)
Let A = (:, ) ; 1
min
: 1
max
be an annulus, and let /, 1 : [, )
1 be two functions. Consider the Laplace equation n = 0, with nonhomo-
geneous Dirichlet boundary conditions:
n(1
min
, ) = /() and n(1
max
, ) = 1(), for all [, ). (14B.14)
Suppose / and 1 have real Fourier series:
/()

L2
o
0
+
n=1
o
n
cos(n) +
n=1
/
n
sin(n)
and 1()

L2
0
+
n=1
n
cos(n) +
n=1
1
n
sin(n).
Then the unique solution to this problem is the function n : A 1 dened
n(:, ) =

L2
l
0
+ n
0
log(:) +
n=1
_
l
n
:
n
+
n
n
:
n
_
cos(n)
+
n=1
_
\
n
:
n
+

n
:
n
_
sin(n). (14B.15)
where the coecients n
n
, l
n
,
n
, \
N
n=1
are the unique solutions to the equa-
2
1
0
1
2
2
1
0
1
1
2
3
4
5
Figure 14B.9: A bubble between two concentric circular wires
tions:
l
0
+n
0
log(1
min
) = o
0
; l
0
+n
0
log(1
max
) =
0
;
l
n
1
n
min
+
n
n
1
n
min
= o
n
; l
n
1
n
max
+
n
n
1
n
max
=
n
;
\
n
1
n
min
+

n
1
n
min
= /
n
; \
n
1
n
max
+

n
1
n
max
= 1
n
.
Furthermore, the series (14B.15) converges semiuniformly to n on int (A).
Proof. Exercise 14B.7 (a) To check that n is harmonic, generalize the strategies E _
used to prove Proposition 14B.2 on page 278 and Proposition 14B.6 on page 284.
(b) To check that the solution also satises the boundary condition (14B.14), sub-
sititute : = 1
min
and : = 1
max
into (14B.15) to get the Fourier series for / and
1.
(c) Use Proposition 5D.5(a) on page 88 to show that this solution is unique. 2
Example: Consider an annular bubble spanning two concentric circular wire
frames. The inner wire has radius 1
min
= 1, and is unwarped, but is elevated
to a height of 4c:, while the outer wire has radius 1
max
= 2, and is twisted to
have shape 1() = cos(3) 2 sin(). Estimate the shape of the bubble between
the two wires.
Solution: We have /() = 4, and 1() = cos(3) 2 sin(). Thus:
o
0
= 4;
3
= 1; and 1
1
= 2
and all other coecients of the boundary conditions are zero. Thus, our solution
will have the form:
n(:, ) = l
0
+ n
0
log(:) +
_
l
3
:
3
+
n
3
:
3
_
cos(3) +
_
\
1
: +

1
:
_
sin(),
0
x
s
|x-s|
p
r
o
p
o
r
t
i
o
n
a
l
t
o

1
-
|
x
|
2
0
x
s
r
(A) (B)
Figure 14B.10: The Poisson kernel (see also Figure 17F.1 on page 407)
where l
0
, n
0
, l
3
, n
3
, \
1
, and
1
are chosen to solve the equations:
l
0
+n
0
log(1) = 4; l
0
+n
0
log(2) = 0;
l
3
+n
3
= 0; 8 l
3
+
n
3
8
= 1;
\
1
+
1
= 0; 2\
1
+

1
2
= 2.
which is equivalent to:
l
0
= 4; n
0
=
l
0
log(2)
=
4
log(2)
;
n
3
= l
3
;
_
8
1
8
_
l
3
= 1, and thus l
3
=
8
63
;
1
= \
1
;
_
2
1
2
_
\
1
= 2, and thus \
1
=
4
3
.
so that n(:, ) = 4
4 log(:)
log(2)
+
8
63
_
:
3
1
:
3
_
cos(3)
4
3
_
:
1
:
_
sin().
14B(v) Poissons solution to Dirichlet problem on the disk
Prerequisites: 14B(ii). Recommended: 17F.
1
Let | = (:, ) ; : 1 be the disk of radius 1, and let | = S = (:, ) ; : = 1
be its boundary, the circle of radius 1. Recall the Dirichlet problem on the disk
1
See 17F on page 406 for a dierent development of the material in this section, us-
ing impulse-response functions. For yet another approach, using complex analysis, see Corol-
lary 18C.13 on page 445.
from 14B(ii). We will now construct an integral representation formula for the
solution to this problem. The Poisson kernel is the function T : | S 1
dened:
T(x, s) :=
1
2
|x|
2
|x s|
2
for any x | and s S.
In polar coordinates (Figure 14B.10B), we can parameterize s S with a single
angular coordinate [, ), and assign x the coordinates (:, ). Poissons
kernel then takes the form:
T(x, s) = T(:, ; ) =
1
2
:
2
1
2
2:1cos( ) +:
2
.
(Exercise 14B.8) E _
Proposition 14B.11. Poissons Integral Formula
Let | = (:, ) ; : 1 be the disk of radius 1, and let / L
2
[, ). Con-
sider the Laplace equation n = 0, with nonhomogeneous Dirichlet boundary
conditions n(1, ) = /(). The unique solution to this problem satises:
For any : [0, 1) and [, ), n(:, ) =
1
2
_

T(:, ; ) /() d.
(14B.16)
or, more abstractly, n(x) =
1
2
_
S
T(x, s) /(s) ds, for any x int (|).
Proof. For simplicity, assume 1 = 1 (the general case can be obtained by
rescaling). From Proposition 14B.2 on page 278, we know that
n(:, )

L2
0
+
n=1
n
cos(n) :
n
+
n=1
1
n
sin(n) :
n
,
where
n
and 1
n
are the (real) Fourier coecients for the function /. Substi-
tuting in the denition of these coecients (see 8A on page 161), we get:
n(:, ) =
1
2
_

/() d +
n=1
cos(n) :
n
_
1
/() cos(n) d
_
+
n=1
sin(n) :
n
_
1
/() sin(n) d
_
=
1
2
_

/()
_
1 + 2
n=1
:
n
cos(n) cos(n) + 2
n=1
:
n
sin(n) sin(n)
_
d
()
1
2
_

/()
_
1 + 2
n=1
:
n
cos
_
n( )
_
_
(14B.17)
14C. Bessel functions 291
where () is because cos(n) cos(n) + sin(n) sin(n) = cos
_
n( )
_
.
It now suces to prove:
Claim 1: 1 + 2
n=1
:
n
cos
_
n( )
_
= T(:, ; ).
Proof. By Eulers Formula (see page 551), 2 cos
_
n( )
_
= c
in()
+
c
in()
. Hence,
1 + 2
n=1
:
n
cos
_
n( )
_
= 1 +
n=1
:
n
_
c
in()
+c
in()
_
.
(14B.18)
Now dene complex number . = : c
i()
; then observe that :
n
c
in()
=
.
n
and :
n
c
in()
= .
n
. Thus, we can rewrite the right hand side of
(14B.18) as:
1 +
n=1
:
n
c
in()
+
n=1
:
n
c
in()
= 1 +
n=1
.
n
+
n=1
.
n
(a)
1 +
.
1 .
+
.
1 .
= 1 +
. .. +. ..
1 . . +..
(b)
1 +
2Re [.] 2[.[
2
1 2Re [.] +[.[
2
=
1 2Re [.] +[.[
2
1 2Re [.] +[.[
2
+
2Re [.] 2[.[
2
1 2Re [.] +[.[
2
=
1 [.[
2
1 2Re [.] +[.[
2 (c)
1 :
2
1 2: cos ( ) + :
2
= T(:, ; ).
(a) is because
n=1
x
n
=
x
1 x
for any x C with [x[ < 1. (b) is because z + z =
2Re [z] and zz = [z[
2
for any z C. (c) is because [z[ = r and Re [z] = cos( ) by
denition of z.
Claim 1
Now, use Claim 1 to substitute T(:, ; ) into (14B.17); this yields the Poisson
integral formula (14B.16). 2
14C Bessel functions
14C(i) Bessels equation; Eigenfunctions of in Polar Coordi-
nates
Prerequisites: 4B, 14A. Recommended: 16C.
J_0
J_1
J_2
J_3
Legend
0.4
0.2
0
0.2
0.4
0.6
0.8
1
2 4 6 8 10 12
x
Y_0
Y_1
Y_2
Y_3
Legend
2
1.5
1
0.5
0
0.5
2 4 6 8 10 12
x
(A) (B)
0
(r),
1
(r),
2
(r), and
3
(r),
0
(r),
1
(r),
2
(r), and
3
(r) for r [0, 12]
for r [0, 12]; (these four plots are vertically truncated).
Figure 14C.1: Bessel functions near zero.
Fix n N. The (2-dimensional) Bessels Equation (of order n) is the
ordinary dierential equation
r
2
1
tt
(r) + r1
t
(r) + (r
2
n
2
) 1(r) = 0, (14C.1)
where 1 : [0, ] 1 is an unknown function. In 16C, we will explain how
this equation was rst derived. In the present section, we will investigate its
mathematical consequences.
The Bessel equation has two solutions:
1(r) =
n
(r) the nth order Bessel function of the rst kind.
[See Figures 14C.1(A) and 14C.2(A)]
1(r) =
n
(r) the nth order Bessel function of the second kind, or
Neumann function. [See Figures 14C.1(B) and 14C.2(B)]
Bessel functions are like trigonometric or logarithmic functions; the simplest
expression for them is in terms of a power series. Hence, you should treat the
functions
n
and
n
the same way you treat elementary functions like sin,
tan or log. In 14G we will derive an explicit power series for Bessels func-
tions, and in 14H, we will derive some of their important properties. However,
for now, we will simply take for granted that some solution functions
n
ex-
ists, and discuss how we can use these functions to build eigenfunctions for the
Laplacian which separate in polar coordinates.
Proposition 14C.1.
0.4
0.2
0
0.2
0.4
0.6
0.8
1
10 20 30 40 50 60 70 80
x
(A):
0
(r), for r [0, 100].
The r-intercepts of this graph are the roots
01
,
02
,
03
,
04
, . . .
1.5
1
0.5
0
0.5
10 20 30 40 50 60 70 80
x
(B):
0
(r), for r [0, 80].
Figure 14C.2: Bessel functions are asymptotically periodic.
Fix 0. For any n N, dene the functions
n,
,
n,
,
n,
,
n,
: 1
2
1 by
n,
(:, ) =
n
( :) cos(n);
n,
(:, ) =
n
( :) sin(n);
n,
(:, ) =
n
( :) cos(n); and
n,
(:, ) =
n
( :) sin(n).
(see Figures 14C.3 and 14C.4). Then
n,
,
n,
,
n,
, and
n,
are all eigen-
functions of the Laplacian with eigenvalue
2
:
n,
=
2
n,
;
n,
=
2
n,
;
n,
=
2
n,
; and
n,
=
2
n,
.
Proof. See practice problems #12 to #15 of 14I. 2
We can now use these eigenfunctions to solve PDEs in polar coordinates.
Notice that
n
and thus, eigenfunctions
n,
and
n,
are bounded around
zero (see Figure 14C.1A). On the other hand,
n
and thus, eigenfunctions
n,
and
n,
are unbounded at zero (see Figure 14C.1B). Hence, when solving
BVPs in a neighbourhood around zero (e.g. the disk), we should use
n
,
n,
and
n,
. When solving BVPs on a domain away from zero (e.g. the annulus),
we can also use
n
,
n,
and
n,
.
294 DRAFT
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
21
(r, ) =
22
(r, ) =
23
(r, ) =
24
(r, ) =
25
(r, ) =
2
(
21
r) cos (2)
2
(
22
r) cos (2)
2
(
23
r) cos (2)
2
(
24
r) cos (2)
2
(
25
r) cos (2)
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
6
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
6
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
6
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
11
(r, ) =
12
(r, ) =
13
(r, ) =
14
(r, ) =
15
(r, ) =
1
(
11
r) cos ()
1
(
12
r) cos ()
1
(
13
r) cos ()
1
(
14
r) cos ()
1
(
15
r) cos ()
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
0
.
2
0
.
4
0
.
6
0
.
8
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
8
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
8
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
8
1
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
6
0
.
8
1
01
(r, ) =
02
(r, ) =
03
(r, ) =
04
(r, ) =
05
(r, ) =
0
(
01
r)
0
(
02
r)
0
(
03
r)
0
(
04
r)
0
(
05
r)
Figure 14C.3:
n,m
for n = 0, 1, 2 and for : = 1, 2, 3, 4, 5 (rotate page).
295
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
3
0
.
2
0
.
1 0
0
.
1
0
.
2
0
.
3
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
3
0
.
2
0
.
1 0
0
.
1
0
.
2
0
.
3
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
3
0
.
2
0
.
1 0
0
.
1
0
.
2
0
.
3
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
3
0
.
2
0
.
1 0
0
.
1
0
.
2
0
.
3
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
3
0
.
2
0
.
1 0
0
.
1
0
.
2
0
.
3
51
(r, ) =
52
(r, ) =
53
(r, ) =
54
(r, ) =
55
(r, ) =
5
(
51
r) cos (5)
5
(
52
r) cos (5)
5
(
53
r) cos (5)
5
(
54
r) cos (5)
5
(
55
r) cos (5)
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
41
(r, ) =
42
(r, ) =
43
(r, ) =
44
(r, ) =
45
(r, ) =
4
(
41
r) cos (4)
4
(
42
r) cos (4)
4
(
43
r) cos (4)
4
(
44
r) cos (4)
4
(
45
r) cos (4)
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
0
.
5
0
0
.
5
1
0
.
5
0
0
.
5
1
0
.
4
0
.
2 0
0
.
2
0
.
4
31
(r, ) =
32
(r, ) =
33
(r, ) =
34
(r, ) =
35
(r, ) =
3
(
31
r) cos (3)
3
(
32
r) cos (3)
3
(
33
r) cos (3)
3
(
34
r) cos (3)
3
(
35
r) cos (3)
Figure 14C.4:
n,m
for n = 3, 4, 5 and for : = 1, 2, 3, 4, 5 (rotate page).
14C(ii) Boundary conditions; the roots of the Bessel function
Prerequisites: 5C, 14C(i).
To obtain homogeneous Dirichlet boundary conditions on a disk of radius 1,
we need an eigenfunction of the form
n,
(or
n,
) such that
n,
(1, ) = 0 for
all [, ). Hence, we need:
n
( 1) = 0. (14C.2)
The roots of the Bessel function
n
are the values 1
,
such that
n
() = 0.
These roots form an increasing sequence
0
n1
<
n2
<
n3
<
n4
< . . . (14C.3)
of irrational values
2
. Thus, to satisfy the homogeneous Dirichlet boundary condi-
tion (14C.2), we must set :=
nm
,1 for some : N. This yields an increasing
sequence of eigenvalues:
2
n1
=
_
n1
1
_
2
<
2
n2
=
_
n2
1
_
2
<
2
n3
=
_
n3
1
_
2
<
2
n4
=
_
n4
1
_
2
< . . .
(14C.4)
which are the eigenvalues which we can expect to see in this problem. The
corresponding eigenfunctions will then have the form:
n,m
(:, ) =
n
(
n,m
:) cos(n)
n,m
(:, ) =
n
(
n,m
:) sin(n)
(14C.5)
(see Figures 14C.3 and 14C.4).
14C(iii) Initial conditions; Fourier-Bessel expansions
Prerequisites: 5B, 6F, 14C(ii).
To solve an initial value problem, while satisfying the desired boundary con-
ditions, we express our initial conditions as a sum of the eigenfunctions from
expression (14C.5). This is called a Fourier-Bessel Expansion:
)(:, ) =
n=0
m=1
nm

nm
(:, ) +
n=1
m=1
1
nm

nm
(:, )
+
n=0
m=1
o
nm

nm
(:, ) +
n=1
m=1
/
nm

nm
(:, ), (14C.6)
where
nm
, 1
nm
, o
nm
, and /
nm
are all real-valued coecients. Suppose we
are considering boundary value problems on the unit disk |. Then we want
2
Computing these roots is dicult; tables of nm can be found in most standard references
on PDEs.
this expansion to be bounded at 0, so we dont want the second two types of
eigenfunctions. Thus, expression (14C.6) simplies to:
n=0
m=1
nm

nm
(:, ) +
n=1
m=1
1
nm

nm
(:, ). (14C.7)
If we substitute the explicit expressions from (14C.5) for
nm
(:, ) and
nm
(:, )
into expression (14C.7), we get:
n=0
m=1
nm
n
_
nm
:
1
_
cos(n) +
n=1
m=1
1
nm
n
_
nm
:
1
_
sin(n).
(14C.8)
Now, if ) : | 1 is some function describing initial conditions, is it always
possible to express ) using an expansion like (14C.8)? If so, how do we com-
pute the coecients
nm
and 1
nm
in expression (14C.8)? The answer to these
questions lies in the following result:
Theorem 14C.2. The collection
n,m
,
,m
; n = 0..., / N, : N is
an orthogonal basis for L
2
(|). Thus, suppose ) L
2
(|), and for all n, : N,
we dene
nm
:=
),
nm
)
|
n
:|
2
2
=
2
1
2

2
n+1
(
nm
)

_

_
R
0
)(:, )
n
_
nm
:
1
_
cos(n) : d: d,
and 1
nm
:=
),
nm
)
|
n
:|
2
2
=
2
1
2

2
n+1
(
nm
)

_

_
R
0
)(:, )
n
_
nm
:
1
_
sin(n) : d: d.
Then the Fourier-Bessel series (14C.8) converges to ) in 1
2
-norm.
Proof. (sketch) The fact that the collection
n,m
,
,m
; n = 0..., / N, : N
is an orthogonal set will be veried in Proposition 14H.4 on page 313 of 14H.
The fact that this orthogonal set is actually a basis of L
2
(|) is too com-
plicated for us to prove here. Given that this is true, if we dene
nm
:=
),
nm
),|
n
:|
2
2
and 1
nm
:= ),
nm
),|
n
:|
2
2
, then the Fourier-Bessel
series (14C.8) converges to ) in 1
2
-norm, by denition of orthogonal basis
(see 6F on page 131).
It remains to verify the integral expressions given for the two inner products.
To do this, recall that
),
nm
) =
1
Area[|]
_
|
)(x)
nm
(x) dx
=
1
1
2
_
2
0
_
R
0
)(:, )
n
_
nm
:
1
_
cos(n) : d: d
and |
nm
|
2
2
=
nm
,
nm
) =
1
1
2
_

_
R
0
2
n
_
nm
:
1
_
cos
2
(n) : d: d
=
_
1
1
2
_
R
0
2
n
_
nm
:
1
_
: d:
_
_
1
cos
2
(n) d
_
()
1
1
2
_
R
0
2
n
_
nm
:
1
_
: d:.
()
_
1
0
2
n
(
nm
:) : d:.
()
1
2
2
n+1
(
nm
)
here, () is by Proposition 6D.2 on page 112, () is the change of variables
: :=
r
R
, so that d: = 1d:), and () is by Lemma 14H.3(b) on page 310. 2
To compute the integrals in Theorem 14C.2, one generally uses integration
by parts techniques similar to those used to compute trigonometic Fourier coef-
cients (see e.g. 7C on page 147). However, instead of the convenient trigono-
metric facts that sin
t
= cos and cos
t
= sin, one must make use of slightly more
complicated recurrence relations of Proposition 14H.1 on page 309 of 14H. See
Remark 14H.2 on page 310.
We will do not have space in this book to properly develop integration tech-
niques for computing Fourier-Bessel coecients. Instead, in the remaining dis-
cussion, we will simply assume that ) is given to us in the form (14C.8).
14D The Poisson equation in polar coordinates
Prerequisites: 1D, 14C(ii), 0F. Recommended: 11C, 12C, 13C, 14B .
Proposition 14D.1. (Poisson Equation on Disk; homogeneous Dirichlet BC)
Let | = (:, ) ; : 1 be a disk, and let L
2
(|) be some function.
Consider the Posson equation n(:, ) = (:, ), with homogeneous Dirich-
let boundary conditions. Suppose has semiuniformly convergent Fourier-Bessel
series:
(:, )

L2
n=0
m=1
nm
n
_
nm
:
1
_
cos(n) +
n=1
m=1
1
nm
n
_
nm
:
1
_
sin(n)
14D. The Poisson equation in polar coordinates 299
Then the unique solution to this problem is the function n : | 1 dened
n(:, )
unif

n=0
m=1
1
2

nm
2
nm

n
_
nm
:
1
_
cos(n)
n=1
m=1
1
2
1
nm
2
nm

n
_
nm
:
1
_
sin(n)
Remark. If 1 = 1, then the expression for simplies to:
(:, )

L2
n=0
m=1
nm
n
(
nm
:)cos(n) +
n=1
m=1
1
nm
n
(
nm
:)sin(n)
and the solution simplies to
n(:, )
unif

n=0
m=1
nm
2
nm
n
(
nm
:)cos(n)
n=1
m=1
1
nm
2
nm
n
(
nm
:)sin(n)
Example 14D.2. Suppose 1 = 1, and (:, ) =
0
(
0,3
:) +
5
(
2,5
:)
sin(2). Then
n(:, ) =

0
(
0,3
:)
2
0,3

5
(
2,5
:) sin(2)
2
2,5
.
Proposition 14D.3. (Poisson Equation on Disk; nonhomogeneous Dirichlet BC)
Let | = (:, ) ; : 1 be a disk. Let / L
2
[, ) and L
2
(|). Con-
sider the Poisson equation n(:, ) = (:, ), with nonhomogeneous Dirichlet
n(1, ) = /(), for all [, ). (14D.1)
1. Let n : | 1 be the solution
3
to the Laplace Equation; n = 0, with
the nonhomogeneous Dirichlet BC (14D.1).
2. Let : | 1 be the solution
4
to the Poisson Equation; = , with
the homogeneous Dirichlet BC.
3. Dene n(:, ) := (:, ; t) + n(:, ). Then n(:, ) is a solution to the
Poisson Equation with inhomogeneous Dirichlet BC (14D.1).
Example 14D.4. Suppose 1 = 1, and (:, ) =
0
(
0,3
:) +
2
(
2,5
:)
sin(2). Let /() = sin(3).
From Example 14B.3 on page 279, we know that the (bounded) solution to
the Laplace equation with Dirichlet BC n(1, ) = /() is:
n(:, ) = :
3
sin(3).
From Example 14D.2, we know that the solution to the Poisson equation
= , with homogeneous Dirichlet BC is:
(:, ) =

0
(
0,3
:)
2
0,3
+

2
(
2,5
:) sin(2)
2
2,5
.
Thus, by Proposition 14D.3, the the solution to the Poisson equation n =
, with Dirichlet BC n(1, ) = /(), is given:
n(:, ) = (:, ) + n(:, ) =

0
(
0,3
:)
2
0,3
+

2
(
2,5
:) sin(2)
2
2,5
+
:
3
sin(3).
14E The heat equation in polar coordinates
Prerequisites: 1B, 14C(iii), 0F. Recommended: 11A, 12B, 13A, 14B .
Proposition 14E.1. (Heat equation on Disk; homogeneous Dirichlet BC)
Let | = (:, ) ; : 1 be a disk, and consider the heat equation
t
n =
n, with homogeneous Dirichlet boundary conditions, and initial conditions
n(:, ; 0) = )(:, ). Suppose ) has Fourier-Bessel series:
)(:, )

L2
n=0
m=1
nm
n
_
nm
:
1
_
cos(n) +
n=1
m=1
1
nm
n
_
nm
:
1
_
sin(n)
3
4
Obtained from Proposition 14D.1, for example.
14E. The heat equation in polar coordinates 301
Then the unique solution to this problem is the function n : | 1
,
1
dened:
n(:, ; t)

L2
n=0
m=1
nm

n
_
nm
:
1
_
cos(n) exp
_
2
nm
1
2
t
_
+
n=1
m=1
1
nm

n
_
nm
:
1
_
sin(n) exp
_
2
nm
1
2
t
_
Furthermore, the series dening n converges semiuniformly on | 1
+
.
Remark. If 1 = 1, then the initial conditions simplify to:
)(:, )

L2
n=0
m=1
nm
n
(
nm
:)cos(n) +
n=1
m=1
1
nm
n
(
nm
:)sin(n)
and the solution simplies to:
n(:, ; t)

L2
n=0
m=1
nm

n
(
nm
:) cos(n) c
2
nm
t
+
n=1
m=1
1
nm

n
(
nm
:) sin(n) c
2
nm
t
.
Example 14E.2. Suppose 1 = 1, and )(:, ) =
0
(
0,7
:) 4
3
(
3,2
:)
cos(3). Then
n(:, ; t) =
0
(
0,7
:) c
2
0,7
t
4
3
(
3,2
:) cos(3) c
2
32
t
.
Proposition 14E.3. (Heat equation on Disk; nonhomogeneous Dirichlet BC)
Let | = (:, ) ; : 1 be a disk, and let ) : | 1 and / : [, ) 1
be given functions. Consider the Heat equation
t
n = n, with initial
conditions n(:, ) = )(:, ), and nonhomogeneous Dirichlet boundary conditions:
n(1, ) = /(), for all [, ). (14E.1)
1. Let n : | 1 be the solution
5
to the Laplace Equation; n = 0, with
the nonhomogeneous Dirichlet BC (14E.1).
5
2. Dene p(:, ) := )(:, ) n(:, ). Let : | 1
,
1 be the solution
6
to the heat equation
t
= with initial conditions (:, ) = p(:, ),
and homogeneous Dirichlet BC.
3. Dene n(:, ; t) := (:, ; t) + n(:, ). Then n(:, ; t) is a solution to the
heat equation with initial conditions n(:, ) = )(:, ), and inhomogeneous
Dirichlet BC (14E.1).
14F The wave equation in polar coordinates
Prerequisites: 2B, 14C(ii), 14C(iii), 0F. Recommended: 11B, 12D, 14E.
Imagine a drumskin stretched tightly over a circular frame. At equilibrium,
the drumskin is perfectly at, but if we strike the skin, it will vibrate, meaning
that the membrane will experience vertical displacements from equilibrium. Let
| = (:, ) ; : 1 represent the round skin, and for any point (:, ) | on
the drumskin and time t 0, let n(:, ; t) be the vertical displacement of the
drum. Then n will obey the two-dimensional wave equation:
2
t
n(:, ; t) = n(:, ; t). (14F.1)
However, since the skin is held down along the edges of the circle, the function
n will also exhibit homogeneous Dirichlet boundary conditions:
n(1, ; t) = 0, for all [, ) and t 0. (14F.2)
Proposition 14F.1. (Wave equation on Disk; homogeneous Dirichlet BC)
Let | = (:, ) ; : 1 be a disk, and consider the wave equation
2
t
n =
n, with homogeneous Dirichlet boundary conditions, and
Initial position: n(:, ; 0) = )
0
(:, );
Initial velocity:
t
n(:, ; 0) = )
1
(:, )
Suppose )
0
and )
1
have Fourier-Bessel series:
)
0
(:, )

L2
n=0
m=1
nm

n
_
nm
:
1
_
cos(n)
+
n=1
m=1
1
nm

n
_
nm
:
1
_
sin(n);
6
Obtained from Proposition 14E.1, for example.
14F. The wave equation in polar coordinates 303
and )
1
(:, )

L2
n=0
m=1
t
nm

n
_
nm
:
1
_
cos(n)
+
n=1
m=1
1
t
nm

n
_
nm
:
1
_
sin(n).
Assume that
n=0
m=1
2
nm
[
nm
[ +
n=1
m=1
2
nm
[1
nm
[ < ,
and
n=0
m=1
nm
[
t
nm
[ +
n=1
m=1
nm
[1
t
nm
[ < .
Then the unique solution to this problem is the function n : | 1
,
1
dened:
n(:, ; t)

L2
n=0
m=1
nm

n
_
nm
:
1
_
cos(n) cos
_
nm
1
t
_
+
n=1
m=1
1
nm

n
_
nm
:
1
_
sin(n) cos
_
nm
1
t
_
+
n=0
m=1
1
t
nm
nm

n
_
nm
:
1
_
cos(n) sin
_
nm
1
t
_
+
n=1
m=1
1 1
t
nm
nm

n
_
nm
:
1
_
sin(n) sin
_
nm
1
t
_
.
Proof. Exercise 14F.1 2 E _
Remark. If 1 = 1, then the initial conditions would be:
)
0
(:, )

L2
n=0
m=1
nm

n
(
nm
:) cos(n)
+
n=1
m=1
1
nm

n
(
nm
:) sin(n),
and )
1
(:, )

L2
n=0
m=1
t
nm

n
(
nm
:) cos(n)
+
n=1
m=1
1
t
nm

n
(
nm
:) sin(n).
and the solution simplies to:
n(:, ; t)

L2
n=0
m=1
nm

n
(
nm
:) cos(n) cos (
nm
t)
+
n=1
m=1
1
nm

n
(
nm
:) sin(n) cos (
nm
t)
+
n=0
m=1
t
nm
nm

n
(
nm
:) cos(n) sin (
nm
t)
+
n=1
m=1
1
t
nm
nm

n
(
nm
:) sin(n) sin (
nm
t) .
Acoustic Interpretation: The vibration of the drumskin is a superposition
of distinct modes of the form
nm
(:, ) =
n
_
nm
:
1
_
cos(n) and
nm
(:, ) =
n
_
nm
:
1
_
sin(n),
for all :, n N. For xed : and n, the modes
nm
and and
nm
vibrate at
(temporal) frequency
nm
=

nm
1
. In the case of the vibrating string, all the
dierent modes vibrated at frequences that were integer multiples of the funda-
mental frequency; musically speaking, this means that they are in harmony. In
the case of a drum, however, the frequencies are all irrational multiples (because
the roots
nm
are all irrationally related). Acoustically speaking, this means we
expect a drum to sound somewhat more discordant than a string.
Notice also that, as the radius 1 gets larger, the frequency
nm
=

nm
1
gets smaller. This means that larger drums vibrate at lower frequencies, which
matches our experience.
Example 14F.2. A circular membrane of radius 1 = 1 is emitting a pure pitch
at frequency
35
. Roughly describe the space-time prole of the solution (as a
pattern of distortions of the membrane).
Answer: The spatial distortion of the membrane must be a combination of
modes vibrating at this frequency. Thus, we expect it to be a function of the
form:
n(:, ; t) =
3
(
35
:)
__
cos(3) + 1 sin(3)
_
cos (
35
t) +
_

t
35
cos(3) +
1
t
35
sin(3)
_
sin (
35
t)
_
By introducing some constant angular phase-shifts and
t
, as well as new
constants C and C
t
, we can rewrite this (Exercise 14F.2 ) as: E _
14G. The power series for a Bessel function 305
n(:, ; t) =
3
(
35
:)
_
C cos(3( +)) cos (
35
t) +
C
t
35
cos(3( +
t
)) sin (
35
t)
_
.
Example 14F.3. An initially silent circular drum of radius 1 = 1 is struck in its

exact center with a drumstick having a spherical head. Describe the resulting
pattern of vibrations.
Solution: This is a problem with nonzero initial velocity and zero initial
position. Since the initial velocity (the impact of the drumstick) is rotationally
symmetric (dead centre, spherical head), we can write it as a Fourer-Bessel
expansion with no angular dependence:
)
1
(:, ) = )(:)

L2
m=1
t
m
0
(
0m
:) (
t
1
,
t
2
,
t
3
, . . . some constants)
(all the higher-order Bessel functions disappear, since
n
is always associated
with terms of the form sin(n) and cos(n), which depend on .) Thus, the
solution must have the form:
n(:, ; t) = n(:, t)

L2
m=1
t
m
0m

0
(
0m
:) sin (
0m
t).
14G The power series for a Bessel function
Prerequisites: 0H(iii). Recommended: 14C(i).
In 14C-14F, we claimed that Bessels equation had certain solutions called
Bessel functions, and showed how to use these Bessel functions to solve dier-
ential equations in polar coordinates. Now we will derive an an explicit formula
for these Bessel functions in terms of their power series.
Proposition 14G.1. Set := 1. For any xed : N there is a solution
m
: 1
,
1 to the Bessel Equation
r
2
tt
(r) +r
t
(r) + (r
2
:
2
) (r) = 0, for all r 0. (14G.1)
with a power series expansion:
m
(r) =
_
r
2
_
m
k=0
(1)
k
2
2k
/! (:+/)!
r
2k
(14G.2)
(
m
is called the :th order Bessel function of the rst kind.)
Proof. The ODE (14G.1) satises the hypotheses of the Frobenius Theorem
(see Example 0H.5 on page 574 of Appendix 0H(iii)). Thus, we can apply the
Method of Frobenius to solve (14G.1). Suppose that is a solution, with an
(unknown) power series (r) = r
M
k=0
o
k
r
k
, where o
0
, o
1
, . . . are unknown
coecients, and ` 0. We assume that o
0
,= 0. We substitute this power
series into eqn.(14G.1) to get equations relating the coecients. The details
of this computation are shown in Table 14.1.
Claim 1: ` = :.
Proof. If the Bessel equation is to be satised, the power series in the bottom
row of Table 14.1 must be identically zero. In particular, this means that
the coecient labeled (a) must be zero; in other words o
0
(`
2
:
2
) = 0.
Since we know that o
0
,= 0, this means (`
2
:
2
) = 0 i.e. `
2
= :
2
.
But ` 0, so this means ` = :.
Claim 1
Claim 2: o
1
= 0.
Proof. If the Bessel equation is to be satised, the power series in the bottom
row of Table 14.1 must be identically zero. In particular, this means that the
coecient labeled (b) must be zero; in other words, o
1
_
(` + 1)
2
:
2
=
0.
Claim 1 says that ` = :; hence this is equivalent to o
1
_
(:+ 1)
2
:
2
=
0. Clearly,
_
(:+ 1)
2
:
2
,= 0; hence we conclude that o

1
= 0.
Claim 2
Claim 3: For all / 2, the coecients o
2
, o
3
, o
4
, . . . must satisfy the
following recurrence relation:
o
k
=
1
(:+/)
2
:
2
o
k2
, for all even / N with / 2. (14G.3)
On the other hand, o
k
= 0 for all odd / N.
Proof. If the Bessel equation is to be satised, the power series in the
bottom row of Table 14.1 must be identically zero. In particular, this
means that all the coecients /
k
must be zero. In other words, o
k2
+
_
(` +/)
2
:
2
_
o
k
= 0.
From Claim 1, we know that ` = :; hence this is equivalent to o
k2
+
_
(:+/)
2
:
2
_
o
k
= 0. In other words, o
k
= o
k2
,
_
(:+/)
2
:
2
_
o
k
.
From Claim 2, we know that o
1
= 0. It follows from this equation that
o
3
= 0; hence o
5
= 0, etc. Inductively, o
n
= 0 for all odd n.
Claim 3
Claim 4: Assume we have xed a value for o
0
. Dene
o
2j
:=
(1)
j
o
0
2
2j
,!(:+ 1)(:+ 2) (:+,)
, for all , N.
307
(
r
)
=
o
0
r
M
+
o
1
r
M
+
1
+
o
2
r
M
+
2
+
+
o
k
r
M
+
k
+
T
h
e
n
:
2
(
r
)
=
:
2
o
0
r
M
:
2
o
1
r
M
+
1
+
:
2
o
2
r
M
+
2
+
:
2
o
k
r
M
+
k
+
r
2
(
r
)
=
o
0
r
M
+
2
+
+
o
k
2
r
M
+
k
+
t
(
r
)
=
`
o
0
r
M
+
(
`
+
1
)
o
1
r
M
+
1
+
(
`
+
2
)
o
2
r
M
+
2
+
+
(
`
+
/
)
o
k
r
M
+
k
+
r
2
t
t
(
r
)
=
`
(
`
1
)
o
0
r
M
+
(
`
+
1
)
`
o
1
r
M
+
1
+
(
`
+
2
)
(
`
+
1
)
o
2
r
M
+
2
+
+
(
`
+
/
)
(
`
+
/
1
)
o
k
r
M
+
k
+
T
h
u
s
0
=
r
2
t
t
(
r
)
+
r
t
(
r
)
+
(
r
2
:
2
)
(
r
)
=
(
`
2
:
2
)
.
.
(
a
)
o
0
r
M
+
(
(
`
+
1
)
2
:
2
)
.
.
(
b
)
o
1
r
M
+
1
+
_
a
0
+
(
(
M
+
2
)
2
m
2
)
a
2
_
r
M
+
2
+
+
/
k
r
M
+
k
+
w
h
e
r
e
/
k
:
=
o
k
2
+
(
`
+
/
)
o
k
+
(
`
+
/
)
(
`
+
/
1
)
o
k
:
2
o
k
=
o
k
2
+
(
`
+
/
)
(
1
+
`
+
/
1
)
o
k
:
2
o
k
=
o
k
2
+
_
(
`
+
/
)
2
:
2
_
o
k
.
Table 14.1: The method of Frobenius to solve Bessels equation in the proof of 14G.1.
Then the sequence o
0
, o
2
, o
4
, . . . satises the recurrence relation (14G.3).
Proof. Set / = 2, in eqn.(14G.3). For any , 2, we must show that
o
2j
=
o
2j2
(:+ 2,)
2
:
2
. Now, by denition,
o
2j2
= o
2(j1)
:=
(1)
j1
o
0
2
2j2
(, 1)!(:+ 1)(:+ 2) (:+, 1)
,
Also,
(:+2,)
2
:
2
= :
2
+4,:+4,
2
:
2
= 4,:+4,
2
= 2
2
,(:+,).
Hence
o
2j2
(:+ 2,)
2
:
2
=
o
2j2
2
2
,(:+,)
=
(1)(1)
j1
o
0
2
2
,(:+,) 2
2j2
(, 1)!(:+ 1)(:+ 2) (:+, 1)
=
(1)
j
o
0
2
2j2+2
,(, 1)! (:+ 1)(:+ 2) (:+, 1)(:+,)
=
(1)
j
o
0
2
2j
,!(:+ 1)(:+ 2) (:+, 1)(:+,)
= o
2j
,
as desired.
Claim 4
By convention we dene o
0
:=
1
2
m
1
:!
. We claim that that the resulting
coecients yield the Bessel function
m
(r) dened by (14G.2) To see this, let
/
2k
be the 2/th coecient of the Bessel series. By denition,
/
2k
:=
1
2
m

(1)
k
2
2k
/! (:+/)!
=
1
2
m

(1)
k
2
2k
/! :!(:+ 1)(:+ 2) (:+/ 1)(:+/)
=
1
2
m
:!

(1)
k
2
2k
/! (:+ 1)(:+ 2) (:+/ 1)(:+/)
= o
0

_
(1)
k+1
2
2k
/!(:+ 1)(:+ 2) (:+/ 1)(:+/)
_
= o
2k
,
as desired. 2
Corollary 14G.2. Fix : N. For any 0, the Bessel Equation (16C.12)
has solution 1(:) :=
m
(:).
Proof. Exercise 14G.1 . 2 E _
14H. Properties of Bessel functions 309
Remarks: (a) We can generalize the Bessel Equation be replacing : with
an arbitrary real number j 1 with j 0. The solution to this equation is the
Bessel function
(r) =
_
r
2
_
k=0
(1)
k
2
2k
/! (j +/ + 1)
r
2k
Here, is the Gamma function; if j = : N, then (:+/ +1) = (:+/)!, so
this expression agrees with (14G.2).
(b) There is a second solution to (14G.1); a function
m
(r) which is un-
bounded at zero. This is called a Neumann function (or a Bessel function of
the second kind or a Weber-Bessel function). Its derivation is too complicated
to discuss here. See [Bro89, 6.8, p.115] or [CB87, 68, p.233].
14H Properties of Bessel functions
Prerequisites: 14G. Recommended: 14C(i).
Let
n
(r) be the Bessel function dened by eqn.(14G.2) on page 305 of 14G.
In this section, we will develop some computational tools to work with these
functions. First, we will dene Bessel functions with negative order as follows:
for any n N, we dene
n
(r) := (1)
n
n
(r). (14H.4)
We can now state the following useful recurrence relations
Proposition 14H.1. For any : Z,
(a)
2:
r

m
(r) =
m1
(r) +
m+1
(r).
(b) 2
t
m
(r) =
m1
(r)
m+1
(r).
(c)
t
0
(r) =
1
(r).
(d)
x
_
r
m

m
(r)
_
= r
m

m1
(r).
(e)
x
_
1
r
m

m
(r)
_
=
1
r
m

m+1
(r).
(f )
t
m
(r) =
m1
(r)
:
r
m
(r).
(g)
t
m
(r) =
m+1
(r) +
:
r
m
(r).
Proof. Exercise 14H.1 (i) Prove (d) for : 1 by substituting in the power E _
series (14G.2) and dierentiating.
(ii) Prove (e) for : 0 by substituting in the power series (14G.2) and dierentiating.
(iii) Use the denition (14H.4) and (i) and (ii) to prove (d) for : 0 and (e) for
: 1.
(iv) Set : = 0 in (e) to obtain (c).
(v) Deduce (f) and (g) from (d) and (e).
(vi) Compute the sum and dierence of (f) and (g) to get (a) and (b). 2
Remark 14H.2: (Integration with Bessel functions)
The recurrence relations of Proposition 14H.1 can be used to simplify inte-
grals involving Bessel functions. For example, parts (d) and (e) immediately
imply that
_
r
m

m1
(r) dr = r
m

m
(r) +C
and
_
1
r
m

m+1
(r) dr =
1
r
m

m
(r) +C.
The other relations are sometimes useful in an integration by parts strategy.
For any n N, let 0
n,1
<
n,2
<
n,3
< be the zeros of the nth Bessel
function
n
(i.e.
n
(
n,m
) = 0 for all : N). Proposition 14C.1 on page 292 of
14C(i) says we can use Bessel functions to dene a sequence of polar-separated
eigenfunctions of the Laplacian:
n,m
(:, ) :=
n
(
n,m
:) cos(n);
n,m
(:, ) :=
n
(
n,m
:) sin(n).
In the proof of Theorem 14C.2 on page 297 of 14C(iii), we claimed that these
eigenfunctions were orthogonal as elements of L
2
(|). We will now verify this
claim. First we must prove a technical lemma.
Lemma 14H.3. Fix n N.
(a) If : ,= `, then
_
1
0
n
(
n,m
:)
n
(
n,M
:) : d: = 0.
(b)
_
1
0
n
(
n,m
:)
2
: d: =
1
2
n+1
(
n,m
)
2
.
Proof. (a) Let =
n,m
and =
n,M
. Dene )(r) :=
m
(r) and
p(r) :=
m
(r). Hence we want to show that
_
1
0
)(r)p(r)r dr = 0.
Dene /(r) = r
_
)(r)p
t
(r) p(r))
t
(r)
_
.
Claim 1: /
t
(r) = (
2
2
))(r)p(r)r.
Proof. First observe that
/
t
(r) = r
x
_
)(r)p
t
(r) p(r))
t
(r)
_
+
_
)(r)p
t
(r) p(r))
t
(r)
_
= r
_
)(r)p
tt
(r) +)
t
(r)p
t
(r) p
t
(r))
t
(r) p(r))
tt
(r)
_
+
_
)(r)p
t
(r) p(r))
t
(r)
_
= r
_
)(r)p
tt
(r) p(r))
tt
(r)
_
+
_
)(r)p
t
(r) p(r))
t
(r)
_
.
By setting 1 = ) or 1 = p in Corollary 14G.2, we obtain:
r
2
)
tt
(r) +r)
t
(r) + (
2
r
2
n
2
))(r) = 0,
and r
2
p
tt
(r) +rp
t
(r) + (
2
r
2
n
2
)p(r) = 0.
We multiply the rst equation by p(r) and the second by )(r) to get
r
2
)
tt
(r)p(r) +r)
t
(r)p(r) +
2
r
2
)(r)p(r) n
2
)(r)p(r) = 0,
and r
2
p
tt
(r))(r) +rp
t
(r))(r) +
2
r
2
p(r))(r) n
2
p(r))(r) = 0.
We then subtract these two equations to get
r
2
_
)
tt
(r)p(r) p
tt
(r))(r)
_
+r
_
)
t
(r)p(r) p
t
(r))(r)
_
+
_
2
_
)(r)p(r)r
2
= 0.
Divide by r to get
r
_
)
tt
(r)p(r) p
tt
(r))(r)
_
+
_
)
t
(r)p(r) p
t
(r))(r)
_
+
_
2
_
)(r)p(r)r = 0.
Hence we conclude
_
2
_
)(r)p(r)r = r
_
p
tt
(r))(r) )
tt
(r)p(r)
_
+
_
p
t
(r))(r) )
t
(r)p(r)
_
= /
t
(r),
as desired
Claim 1
It follows from Claim 1 that
(
2
2
)
_
1
0
)(r)p(r)r dr =
_
1
0
/
t
(r) dr = /(1)/(0)
()
00 = 0.
To see (), observe that /(0) = 0
_
)(0)p
t
(0) p(0))
t
(0)
_
= 0. Also,
/(1) = (1)
_
)(1)p
t
(1) p(1))
t
(1)
_
= 0,
because )(1) =
n
(
n,m
) = 0 and p(1) =
n
(
n,N
) = 0.
(b) Let =
n,m
and )(r) :=
m
(r). Hence we want to evaluate
_
1
0
)(r)
2
r dr.
Dene /(r) := r
2
()
t
(r))
2
+ (
2
r
2
n
2
))
2
(r).
Claim 2: /
t
(r) = 2
2
)(r)
2
r.
Proof. By setting 1 = ) in Corollary 14G.2, we obtain:
0 = r
2
)
tt
(r) +r)
t
(r) + (
2
r
2
n
2
))(r).
We multiply by )
t
(r) to get
0 = r
2
)
t
(r))
tt
(r) +r()
t
(r))
2
+ (
2
r
2
n
2
))(r))
t
(r)
= r
2
)
t
(r))
tt
(r) +r()
t
(r))
2
+ (
2
r
2
n
2
))(r))
t
(r) +
2
r)
2
(r)
2
r)
2
(r)
=
1
2
x
_
r
2
()
t
(r))
2
+ (
2
r
2
n
2
))
2
(r)
_
2
r)
2
(r)
=
1
2
/
t
(r)
2
r)
2
(r).
Claim 2
It follows from Claim 2 that
2
2
_
1
0
)(r)
2
r dr
=
_
1
0
/
t
(r) dr = /(1) /(0)
= 1
2
()
t
(1))
2
+ (
2
1
2
n
2
) )
2
(1)
. .
2
n
(n,m)
=0
0
2
()
t
(0))
2
. .
0
+(
2
0
2
n
2
)
. .
[0 if n = 0]
)
2
(0)
. .
[0 if n ,= 0]
= )
t
(1)
2
=
_
t
n
()
_
2
=
2
t
n
()
2
.
Hence
_
1
0
)(r)
2
r dr =
1
2
t
n
()
2
()
1
2
_
n
n
()
n+1
()
_
2
()
1
2
_
_
n
n,m
n
(
n,m
)
. .
=0
n+1
(
n,m
)
_
_
2
=
1
2
n+1
(
n,m
)
2
,
where () is by Proposition 14H.1(g) and () is because :=
n,m
. 2
Proposition 14H.4. Let | = (:, ) ; : 1 be the unit disk. Then the
collection
n,m
,
,m
; n = 0..., / N, : N
is an orthogonal set for L
2
(|). In other words, for any n, :, , ` N,
(a)
n,m
,
N,M
) =
1
_
1
0
_

n,m
(:, )
N,M
(:, ) d : d: = 0.
Furthermore, if (n, :) ,= (, `), then
(b)
n,m
,
N,M
) =
1
_
1
0
_

n,m
(:, )
N,M
(:, ) d : d: = 0.
(c)
n,m
,
N,M
) =
1
_
1
0
_

n,m
(:, )
N,M
(:, ) d : d: = 0.
Finally, for any (n, :),
(d) |
n,m
|
2
=
1
_
1
0
_

n,m
(:, )
2
d : d: =
1
2
n+1
(
n,m
)
2
.
(e) |
n,m
|
2
=
1
_
1
0
_

n,m
(:, )
2
d : d: =
1
2
n+1
(
n,m
)
2
.
Proof. (a)
n,m
and
N,M
separate in the coordinates (:, ), so the integral
splits in two:
_
1
0
_

n,m
(:, )
N,M
(:, ) d : d:
=
_
1
0
_

n
(
n,m
:) cos(n)
N
(
N,M
:) sin() d : d:
=
_
1
0
n
(
n,m
:)
N
(
N,M
:) : d:
_

cos(n) sin() d
. .
= 0 by Prop. 6D.2(c), p.112
= 0.
(b) or (c) (Case n ,= ). Likewise, if n ,= , then
_
1
0
_

n,m
(:, )
N,M
(:, ) d : d:
=
_
1
0
_

n
(
n,m
:) cos(n)
N
(
N,M
:) cos() d : d:
=
_
1
0
n
(
n,m
:)
N
(
N,M
:) : d:
_

cos(n) cos() d
. .
= 0 by Prop. 6D.2(a), p.112
= 0.
the case (c) is proved similarly.
(b) or (c) (Case n = but : ,= `). If n = , then
_
1
0
_

n,m
(:, )
n,M
(:, ) d : d:
=
_
1
0
_

n
(
n,m
:) cos(n)
n
(
n,M
:) cos(n) d : d:
=
_
1
0
n
(
n,m
:)
n
(
n,M
:) : d:
. .
= 0 by Lemma 14H.3(a)
cos(n)
2
d
. .
= by 6D.2(d),
on p.112.
= 0 = 0.
(d): If n = and : = ` then
_
1
0
_

n,m
(:, )
2
d : d: =
_
1
0
_

n
(
n,m
:)
2
cos(n)
2
d : d:
=
_
1
0
n
(
n,m
:)
2
: d:
. .
=
1
2
n+1(n,m)
2
by Lemma 14H.3(b)
cos(n)
2
d
. .
= by Prop 6D.2(d)
on p.112.
=

2
n+1
(
n,m
)
2
.
The proof of (e) is Exercise 14H.2 . 2 E _
Exercise 14H.3. (a) Use a separation of variables argument (similar to Proposi- E _
tion 16C.2) to prove:
Proposition: Let ) : 1
2
1 be a harmonic function in other words suppose
) = 0.
Suppose ) separates in polar coordinates, meaning that there is a function :
[, ] 1 (satisfying periodic boundary conditions) and a function 1 : 1
1
such that
)(:, ) = 1(:) (), for all : 0 and [, ].
Then there is some : N such that
() = cos(:) +1sin(:), (for constants , 1 1.)
and 1 is a solution to the Cauchy-Euler Equation:
:
2
1
(:) +: 1
(:) :
2
1(:) = 0, for all : 0. (14H.5)
(b) Let 1(:) = :
where = :. Show that 1(:) is a solution to the Cauchy-Euler

equation (14H.5).
(c) Deduce that
m
(:, ) = :
m
sin(:);
m
(:, ) = :
m
cos(:);
m
(:, ) =
:
m
sin(:); and
m
(:, ) = :
m
cos(:) are harmonic functions in 1
2
.
14I. Practice problems 315
14I Practice problems
1. For all (:, ), let
n
(:, ) = :
n
cos(n). Show that
n
is harmonic.
2. For all (:, ), let
n
(:, ) = :
n
sin(n). Show that
n
is harmonic.
3. For all (:, ) with : 0, let
n
(:, ) = :
n
cos(n). Show that
n
is
harmonic.
n
(:, ) = :
n
sin(n). Show that
n
is
harmonic.
0
(:, ) = log [:[. Show that
0
is harmonic.
6. Let /() = cos(3) + 2 sin(5) for [, ).
(a) Find the bounded solution(s) to the Laplace equation on |, with
nonhomogeneous Dirichlet boundary conditions n(1, ) = /(). Is
the solution unique?
(b) Find the bounded solution(s) to the Laplace equation on |
, with
nonhomogeneous Dirichlet boundary conditions n(1, ) = /(). Is
the solution unique?
(c) Find the decaying gradient solution(s) to the Laplace equation on
|
, with nonhomogeneous Neumann boundary conditions

r
n(1, ) =
/(). Is the solution unique?
7. Let /() = 2 cos() 6 sin(2), for [, ).
(a) Find the bounded solution(s) to the Laplace equation on |, with
nonhomogeneous Dirichlet boundary conditions: n(1, ) = /() for
all [, ). Is the solution unique?
(b) Find the bounded solution(s) to the Laplace equation on |, with
nonhomogeneous Neumann boundary conditions:
r
n(1, ) = /()
for all [, ). Is the solution unique?
8. Let /() = 4 cos(5) for [, ).
(a) Find the bounded solution(s) to the Laplace equation on the disk
| = (:, ) ; : 1, with nonhomogeneous Dirichlet boundary con-
ditions n(1, ) = /(). Is the solution unique?
(b) Verify your answer in part (a) (i.e. check that the solution is harmonic
and satises the prescribed boundary conditions.)
(Hint: Recall that =
2
r
+
1
r
r
+
1
r
2
.)
9. Let /() = 5 + 4 sin(3) for [, ).
(a) Find the decaying gradient solution(s) to the Laplace equation on
the codisk |
= (:, ) ; : 1, with nonhomogeneous Neumann

boundary conditions
r
n(1, ) = /(). Is the solution unique?
(b) Verify that your answer in part (a) satises the prescribed boundary
conditions. (Forget about the Laplacian).
10. Let /() = 2 cos(5) + sin(3), for [, ).
(a) Find the solution(s) (if any) to the Laplace equation on the disk
| = (:, ) ; : 1, with nonhomogeneous Neumann boundary
conditions:
n(1, ) = /(), for all [, ).

Is the solution unique? Why or why not?
(b) Find the bounded solution(s) (if any) to the Laplace equation on
the codisk |
= (:, ) ; : 1, with nonhomogeneous Dirichlet

boundary conditions: n(1, ) = /(), for all [, ).
Is the solution unique? Why or why not?
11. Let | be the unit disk. Let / : | 1 be some function, and let
n : | 1 be the solution to the corresponding Dirichlet problem with
boundary conditions /(). Prove that
n(0, 0) =
1
2
_

/() d.
Remark: This is a special case of the Mean Value Theorem for Harmonic
Functions (Theorem 1E.1 on page 16), but do not simply quote Theorem
1E.1 to solve this problem. Instead, apply Proposition 14B.11 on page 290.
12. Let
n,
(:, ) :=
n
( :) cos(n). Show that
n,
=
2
n,
.
13. Let
n,
(:, ) :=
n
( :) sin(n). Show that
n,
=
2
n,
.
14. Let
n,
(:, ) :=
n
( :) cos(n). Show that
n,
=
2
n,
.
15.
n,
(:, ) :=
n
( :) sin(n). Show that
n,
=
2
n,
.
317
Chapter 15
Eigenfunction methods on
arbitrary domains
Science is built up with facts, as a house is with stones. But a collection of facts is no
more a science than a heap of stones is a house. Henri Poincare
The methods given in Chapters 11-14 are all special cases of a single, general
technique: the solution of initial/boundary value problems using eigenfunction
expansions. The time has come to explicate this technique in full generality. The
exposition in this chapter is somewhat more abstract than in previous chapters,
but that is because the concepts we introduce are of such broad applicability.
Technically, this chapter can be read without having read Chapters 11-14; how-
ever, this chapter will be easier to understand if you have have already read
Chapters 11-14.
15A General solution to Poisson, heat and wave equa-
tion BVPs
Prerequisites: 4B(iv), 5D, 6F, 0D. Recommended: Chapters 11, 12, 13, 14 .
Throughout this section:
Let X 1
D
be any bounded domain (e.g. a line segment, box, disk,
sphere, etc. see 0D). When we refer to Neumann boundary conditions,
we will also assume that X has a piecewise smooth boundary (so the normal
derivative is well-dened).
Let o
k
k=1
L
2
(X) be a Dirichlet eigenbasis that is, o
k
k=1
is an
orthogonal basis of L
2
(X), such that every o
k
is an eigenfunction of the
Laplacian, and satises homogeneous Dirichlet boundary conditions on X
(i.e. o
k
(x) = 0 for all x X). For every / N, let
k
< 0 be the
eigenvalue associated with o
k
(i.e. o
k
=
k
o
k
). We can assume
without loss of generality that
k
,= 0 for all / N (Exercise 15A.1 E _
318 DRAFT Chapter 15. Eigenfunction methods on arbitrary domains
Why? Hint: Lemma 5D.3(a)).
Let (
k
k=0
L
2
(X) be a Neumann eigenbasis that is, (
k
k=0
is
an orthogonal basis L
2
(X), such that every (
k
is an eigenfunction of the
Laplacian, and satises homogeneous Neumann boundary conditions on X
(i.e.
(
k
(x) = 0 for all x X). For every / N, let j
k
0 be
the eigenvalue associated with (
k
(i.e. (
k
= j
k
(
k
). We can assume
without loss of generality that (
0
is a constant function (so that j
0
= 0),
while j
k
,= 0 for all / 1 (Exercise 15A.2 Why? Hint: Lemma 5D.3(b)). E _
Theorem 15E.12 (page 347 below) will guarantee that we will be able to nd a
Dirichlet eigenbasis for any domain X 1
D
, and a Neumann eigenbasis for many
domains. If ) L
2
(X) is some other function (describing, for example, an initial
condition), then we can express ) as a combination of these basis elements, as
described in 6F:
)

L2
k=0
k
(
k
, where
k
:=
), (
k
)
|(
k
|
2
2
, for all / N; (15A.1)
and )

L2
k=1
1
k
o
k
, where 1
k
:=
), o
k
)
|o
k
|
2
2
, for all / N. (15A.2)
These expressions are called eigenfunction expansions for ).
Example 15A.1. (a) If X = [0, ] 1, then we could use the eigenbases
o
k
k=1
= S
n
n=1
and (
k
k=0
= C
n
n=0
, where S
n
(r) := sin(nr) and
C
n
(r) := cos(nr) for all n N. In this case,
n
= n
2
= j
n
for all n N.
Also the eigenfunction expansions (15A.1) and (15A.2) are, respectively, the
Fourier Cosine Series and Fourier Sine Series for ), from 7A.
(b) If X = [0, ]
2
1
2
, then we could use the eigenbases o
k
k=1
= S
n,m
n,m=1
and (
k
k=0
= C
n,m
n,m=0
, where S
n,n
(r, j) := sin(nr) sin(:j) and C
n,m
(r) :=
cos(nr) cos(:j) for all n, : N. In this case,
n,m
= n
2
+ :
2
= j
n,m
for
all (n, :) N. Also, the eigenfunction expansions (15A.1) and (15A.2) are,
respectively, the two-dimensional Fourier Cosine Series and Fourier Sine Series
for ), from 9A.
(c) If X = | 1
2
, then we could use the Dirichlet eigenbasis o
n
k=1
=
n,m
n=0,m=1
.
n,m
n,m=1
, where
n,m
and
n,m
are the type-1 Fourier-
Bessel eigenfunctions dened by eqn.(14C.5) on page 296 of 14C(ii). In this
case, we have eigenvalues
n,m
=
2
n,m
, as dened in equation (14C.4) on
page 296. Then the eigenfunction expansion in (15A.2) is the Fourier-Bessel
expansion for ), from 14C(iii).
15A. General solution to Poisson, heat and wave equation BVPs 319
Theorem 15A.2. General Solution of the Poisson Equation
Let X 1
D
be a bounded domain. Let ) L
2
(X), and let / : X 1 be
some other function. Let n : X 1 be a solution to the Poisson equation
n = ).
(a) Suppose o
k
,
k
k=1
is a Dirichlet eigenbasis, and 1
n
n=1
are as in equa-
tion (15A.2). Assume that [
k
[ 1 for all but nitely many / N. If n
satises homogeneous Dirichlet BC (i.e. n(x) = 0 for all x X), then
n

L2
n=1
1
n
n
o
n
.
(b) Let / : X 1 be a solution to the Laplace equation / = 0 satisfying
the nonhomogeneous Dirichlet BC /(x) = /(x) for all x X. If n is as in
part (a), then n := n+/ is a solution to the Poisson equation n = )
and also satises Dirichlet BC n(x) = /(x) for all x X.
(c) Suppose (
k
, j
k
k=1
is a Neumann eigenbasis, and suppose that [j
k
[ 1
for all but nitely many / N. Let
n
n=1
be as in equation (15A.1),
and suppose
0
= 0. For any , [1...1], let |
j
(
k
|
be the supremum of
the ,-derivative of (
k
on X, and suppose that
k=1
k
=0
[
k
[
[j
k
[
|
j
(
k
|
< . (15A.3)
If n satises homogeneous Neumann BC (i.e.
n(x) = 0 for all x X),

then n

L2
C
k=1
k
=0
k
j
k
(
k
, where C 1 is an arbitrary constant.
However, if
0
,= 0, then there is no solution to this problem with homo-
geneous Neumann BC.
(d) Let / : X 1 be a solution to the Laplace equation / = 0 satisfying
the nonhomogeneous Neumann BC
/(x) = /(x) for all x X. If n

is as in part (c), then n := n + / is a solution to the Poisson equation
n = ) and also satises Neumann BC
n(x) = /(x) for all x X.

Proof. Exercise 15A.3 Hint: To show solution uniqueness, use Theorem 5D.5. E _
For (a), imitate the proofs of Propositions 11C.1, 12C.1, 13C.1, and 14D.1.
For (b,d), imitate the proofs of Propositions 12C.6 and 14D.3.
For (c), imitate the proofs of Propositions 11C.2, 12C.4 and 13C.2. Note that you
need hypothesis (15A.3) to apply Proposition 0F.1. 2
Exercise 15A.4. Show how Propositions 11C.1, 11C.2, 12C.1, 12C.4,12C.6 13C.1, E _
13C.2, 14D.1, and 14D.3 are all special cases of Theorem 15A.2. For the results involving
Neumann BC, dont forget to check that (15A.3) is satised.
Theorem 15A.3. General Solution of the heat equation
Let X 1
D
be a bounded domain. Let ) L
2
(X), and let / : X 1 be
some other function. Let n : X 1
,
1 be a solution to the heat equation
t
n = n, with initial conditions n(x, 0) = )(x) for all x X.
(a) Suppose o
k
,
k
k=1
n
n=1
are as in equa-
tion (15A.2). If n satises homogeneous Dirichlet BC (i.e. n(x, t) = 0 for
all x X and t 1
,
), then n

L2
n=1
1
n
exp(
n
t) o
n
.
(b) Let / : X 1 be a solution to the Laplace equation / = 0 satisfying
the nonhomogeneous Dirichlet BC /(x) = /(x) for all x X. If n is as in
part (a), then n := n +/ is a solution to the heat equation
t
n = n,
with initial conditions n(x, 0) = )(x)+/(x) for all x X, and also satises
Dirichlet BC n(x, t) = /(x) for all (x, t) X 1
+
.
(c) Suppose (
k
, j
k
k=0
is a Neumann eigenbasis, and
n
n=0
are as in equa-
tion (15A.1). Suppose the sequence j
k
k=0
grows fast enough that
lim
k
log(/)
j
k
= 0, and, for all , [1...1], lim
k
log |
j
(
k
|
j
k
= 0.
(15A.4)
n(x, t) = 0 for all x X

and t 1
,
), then n

L2
n=0
n
exp(j
n
t) (
n
.
(d) Let / : X 1 be a solution to the Laplace equation / = 0 satisfying
the nonhomogeneous Neumann BC
/(x) = /(x) for all x X. If n is as

in part (c), then n := n+/ is a solution to the heat equation
t
n = n
with initial conditions n(x, 0) = )(x)+/(x) for all x X, and also satises
Neumann BC
n(x, t) = /(x) for all (x, t) X 1

+
.
Furthermore, in parts (a) and (c), the series dening n converges semiuniformly
on X 1
+
.
For part (a), imitate the proofs of Propositions 11A.1, 12B.1, 13A.1, and 14E.1.
For (b,d) imitate the proofs of Propositions 12B.5 and 14E.3.
For (c), imitate the proofs of Propositions 11A.3, 12B.3, and 13A.2. First use hy-
pothesis (15A.4) to show that the sequence c
nt
|
j
(
n
|
n=0
is square-summable
for any t 0. Use Parsevals equality (Theorem 6F.1) to show that the sequence
[
k
[
k=0
is also square-summable. Use the Cauchy-Bunyakowski-Schwarz inequal-
ity to conclude that the sequence c
nt
[
k
[ |
j
(
n
|
n=0
is absolutely summable,
which means the formal derivative
j
n is absolutely convergent. Now apply Proposi-
tion 0F.1. 2
Exercise 15A.6. Show how Propositions 11A.1, 11A.3, 12B.1, 12B.3, 12B.5, 13A.1, E _
13A.2, 14E.1, and 14E.3. are all special cases of Theorem 15A.3. For the results involving
Neumann BC, dont forget to check that (15A.4) is satised.
Theorem 15A.4. General Solution of the wave equation
Let X 1
D
be a bounded domain and let ) L
2
(X). Suppose n : X1
,
1
is a solution to the wave equation
2
t
n = n, and has initial position n(x; 0) =
)(x) for all x X.
(a) Suppose o
k
,
k
k=1
n
n=1
are as in equa-
tion (15A.2). Suppose
n=1
[
n
1
n
[ < . If n satises homogeneous Dirich-
let BC (i.e. n(x, t) = 0 for all x X and t 1
,
), then n

L2
n=1
1
n
cos(
_
n
t) o
n
.
(b) Suppose (
k
, j
k
k=0
is a Neumann eigenbasis, and
n
n=0
are as in equa-
tion (15A.1). Suppose the sequence
n
n=0
decays quickly enough that
n=0
[j
n
n
[ < , and, for all , [1...1],
n=0
[
n
[ |
j
(
n
|
< .
(15A.5)
n(x, t) = 0 for all x X

and t 1
,
), then n

L2
n=0
n
cos(
j
n
t) (
n
.
Now suppose n : X1
,
1 is a solution to the wave equation
2
t
n = n,
and has initial velocity
t
n(x; 0) = )(x) for all x X.
(c) Suppose
n=1
_
n
[1
n
[ < . If n satises homogeneous Dirichlet BC, then
n

L2
n=1
1
n
n
sin(
_
n
t) o
n
.
(d) Suppose the sequence
n
n=0
decays quickly enough that
n=1
j
n
[
n
[ < , and, for all , [1...1],
n=1
[
n
[
j
n
|
j
(
n
|
< .
(15A.6)
If n satises homogeneous Neumann BC, then there is some constant C 1
such that, for all x X, we have n(x; 0) = C, and for all t 1, we have
n(x; t)

L2
0
t +
n=1
j
n
sin(
j
n
t) (
n
(x) + C.
(e) To obtain a solution with both a specied initial position and a specied
initial velocity, add the solutions from (a) and (c) for homogeneous Dirich-
let BC. Add the solutions from (b) and (d) for homogeneous Neumann BC
(setting C = 0 in part (d)).
For (a), imitate Propositions 12D.1 and 14F.1. For (c) imitate the proof of Proposi-
tions 12D.3 and 14F.1. For (b) and (d), use hypotheses (15A.5) and (15A.6) to apply
Proposition 0F.1. 2
Exercise 15A.8. Show how Propositions 11B.1, 11B.4 12D.1, 12D.3, and 14F.1 are E _
all special cases of Theorem 15A.4(a,c).
Exercise 15A.9. What is the physical meaning of a nonzero value of
0
in Theorem E _
15A.4(d)?
Theorems 15A.2, 15A.3, and 15A.4 allow us to solve I/BVPs on any domain,
once we have a suitable eigenbasis. We illustrate with a simple example.
Proposition 15A.5. Eigenbases for a Triangle
Let X :=
_
(r, j) [0, ]
2
; j r
_
be a lled right-angle triangle (Figure
15A.1).
(a) For any two-element subset n, : N (i.e. n ,= :), let o
n,m
:=
sin(nr) sin(:j) sin(:r) sin(nj), and let
n,m
:= n
2
+:
2
. Then:
[i] o
n,m
is an eigenfunction of the Laplacian: o
n,m
=
n,m
o
n,m
.
[ii] o
n,m
n,mN
is a Dirichlet eigenbasis for L
2
(X).
(b) Let C
0,0
= 1, and for any two-element subset n, : N, let (
n,m
:=
cos(nr) cos(:j) + cos(:r) cos(nj), and let
n,m
:= n
2
+:
2
. Then:

X
Figure 15A.1: Right-angle triangular domain of Proposition 15A.5
[i] (
n,m
is an eigenfunction of the Laplacian: (
n,m
=
n,m
(
n,m
.
[ii] (
n,m
n,mN
is a Neumann eigenbasis for L
2
(X).
Proof. Exercise 15A.10 Hint: Part [i] is a straightforward computation, as is E _
the verication of the homogeneous boundary conditions (Hint: on the hypotenuse,
=
2

1
). To verify that the specied sets are orthogonal bases, use Theorem
9A.3. 2
Exercise 15A.11. (a) Combine Proposition 15A.5 with Theorems 15A.2, 15A.3, E _
and 15A.4 to provide a general solution method for solving the Poisson equation, heat
equation, and wave equation on a right-angle triangle domain, with either Dirichlet or
Neumann boundary conditions.
(b) Set up and solve some simple initial/boundary value problems using your method.
Remark 15A.6. There is nothing special about the role of the Laplacian
in Theorems 15A.2, 15A.3, and 15A.4. If L is any linear dierential operator,
for which we have solution uniqueness results analogous to the results of 5D,
then Theorems 15A.2, 15A.3, and 15A.4 are still true if you replace with
L everywhere (Exercise 15A.12 Verify this). In particular, if L is an elliptic E _
dierential operator (see 5E), then:
Theorem 15A.2 becomes the general solution to the boundary value prob-
lem for the nonhomogeneous elliptic PDE Ln = ).
Theorem 15A.3 becomes the general solution to the the initial/boundary
value problem for the homogeneous parabolic PDE
t
n = Ln.
Theorem 15A.4 becomes the general solution to the initial value problem
for the homogeneous hyperbolic PDE
2
t
n = Ln.
Theorem 15E.17 on page 349 (below) discusses the existence of Dirichlet eigen-
bases for other elliptic dierential operators.
Exercise 15A.13. Let X 1
3
be a bounded domain, and consider a quantum E _
particle conned to the domain X by an innite potential well \ : 1
3
1 ,
where \ (x) = 0 for all x X, and \ (x) = for all x , X (see Examples 3C.4 and 3C.5
on pages 49-50 for discussion of the physical meaning of this model). Modify Theorem
15A.3 to state and prove a theorem describing the general solution to the initial value
problem for the Schrodinger equation with the potential \ .
Hint. If : 1
3
1 C is a solution to the corresponding Schrodinger equation,
then we can assume
t
(x) = 0 for all x , X. If is also continuous, then we can model
the particle using a function : X 1 C, which satises homogeneous Dirichlet
boundary conditions on X.
15B General solution to Laplace equation BVPs
Prerequisites: 4B(iv), 5B, 5C, 6F, 0D.
Recommended: 5D, 12A, 13B, 14B, 15A.
Theorems 15A.2(b,d) and 15A.3(b,d) both used the same strategy to solve a
PDE with nonhomogeneous boundary conditions:
Solve the original PDE with homogeneous boundary conditions.
Solve the Laplace equation with the specied nonhomogeneous BC.
Add these two solutions together to get a solution to the original problem.
However, we do not yet have a general method for solving the Laplace equation.
That is the goal of this section. Throughout this section, we make the following
assumptions.
Let X 1
D
be a bounded domain, whose boundary X is piecewise
smooth. This has two consequences: (1) The normal derivative on the
boundary is well-dened (so we can meaningfully impose Neumann bound-
ary conditions); and (2) We can meaningfully speak of integrating func-
tions over X. For example, if X 1
2
, then X should be a nite union
of smooth curves. If X 1
3
, then X should be a nite union of smooth
surfaces, etc. If /, c : X 1 are functions, then dene
/, c) :=
_
X
/(x)c(x) dr and |/|
2
:=
_
/, /) :=
__
X
[/(x)[
2
dr
_
1/2
,
where these are computed as contour integrals (or surface integrals, etc.)
over X. As usual, let L
2
(X) be the set of all integrable functions / :
X 1 such that |/|
2
< (see 6B for further discussion).
15B. General solution to Laplace equation BVPs 325
Let
n
n=1
be an orthogonal basis for L
2
(X). Thus, for any / L
2
(X),
we can write
/

L2
n=1
1
n
n
where 1
n
:=
/,
n
)
|
n
|
2
2
, for all n N.
(15B.1)
For all n N, let H
n
: X 1 be a harmonic function (i.e. H
n
=
0) satisfying the nonhomogeneous Dirichlet boundary condition H
n
(x) =
n
(x) for all x X. The system H := H
n
n=1
is called a Dirichlet
harmonic basis for X.
Suppose
1
1 is the constant function. Then
_
X

n
(x) dr =
n
, 1) =
0, for all n 2 (by orthogonality). For all n 2, let (
n
: X 1
be a harmonic function (i.e. (
n
= 0) satisfying the nonhomogeneous
Neumann boundary condition
(
n
(x) =
n
(x) for all x X. The
system G := 1 .(
n
n=2
is called a Neumann harmonic basis for X.
(Note that
1 = 0, not
1
).
Note. Although they are called harmonic bases for X, H and
(
n
n=2
are
actually orthogonal bases for L
2
(X), not for L
2
(X).
Exercise 15B.1. Show that there is no harmonic function (
1
on X satisfying the E _
Neumann boundary condition
(
1
(x) = 1 for all x X. Hint: Use Corollary
5D.4(b)[i] on page 87.
Example 15B.1. If X = [0, ]
2
1
2
, then X = L R T B, where
L := 0[0, ], R := [0, ], B := [0, ]0, and T := [0, ].
(See Figure 12A.1(B) on page 240).
(a) Let
k
k=1
:= L
n
n=1
.1
n
n=1
.B
n
n=1
.T
n
n=1
, where, for all
n N, the functions L
n
, 1
n
, B
n
, T
n
: X 1 are dened by
L
n
(r, j) :=
_
sin(nj) if (r, j) L;
0 otherwise.
1
n
(r, j) :=
_
sin(nj) if (r, j) R;
0 otherwise.
B
n
(r, j) :=
_
sin(nr) if (r, j) B;
0 otherwise.
and T
n
(r, j) :=
_
sin(nr) if (r, j) T;
0 otherwise.
Now, L
n
n=1
2
(L) (by Theorem 7A.1). Likewise,
1
n
n=1
2
(R), B
n
n=1
is an orthogonal basis for
L
2
(B), and T
n
n=1
2
(T). Thus,
k
k=1
is an
orthogonal basis for L
2
(X).
Let H := H
L
n
n=1
.H
R
n
n=1
.H
T
n
n=1
.H
B
n
n=1
, where for all n N, and
all (r, j) [0, ]
2
, we dene
H
L
n
(r, j) :=
sinh (n( r)) sin(nj)
sinh(n)
;
H
R
n
(r, j) :=
sinh(nr) sin(nj)
sinh(n)
;
H
B
n
(r, j) :=
sin(nr) sinh(n( j))
sinh(n)
;
and H
T
n
(r, j) :=
sin(nr) sinh(nj)
sinh(n)
.
(See Figures 12A.2 and 12A.3 on pages 241-242). Then H is a Dirichlet har-
monic basis for X. This was the key fact employed by Proposition 12A.4 on
page 244 to solve the Laplace Equation on [0, ]
2
with arbitrary nonhomoge-
neous Dirichlet boundary conditions.
(b) Let
k
k=1
:=
1
,
=
,
[[
,
. L
n
n=1
. 1
n
n=1
. B
n
n=1
.
T
n
n=1
. Here, for all (r, j) [0, ]
2
, we dene
1
(r, j) := 1;
[[
(r, j) :=
_
_
_
1 if (r, j) R;
1 if (r, j) L;
0 if (r, j) B. T.
=
(r, j) :=
_
_
_
1 if (r, j) T;
1 if (r, j) B;
0 if (r, j) L . R.
and
(r, j) :=
_
1 if (r, j) L . R;
1 if (r, j) T. B.
Meanwhile, for all n N, the functions L
n
, 1
n
, B
n
, T
n
: X 1 are dened
by
L
n
(r, j) :=
_
cos(nj) if (r, j) L;
0 otherwise.
1
n
(r, j) :=
_
cos(nj) if (r, j) R;
0 otherwise.
B
n
(r, j) :=
_
cos(nr) if (r, j) B;
0 otherwise.
and T
n
(r, j) :=
_
cos(nr) if (r, j) T;
0 otherwise.
Now, 1 . L
n
n=1
2
(L) (by Theorem 7A.1).
Likewise, 1.1
n
n=1
2
(R), 1.B
n
n=1
is an
2
(B), and 1.T
n
n=1
2
(T).
It follows that
k
k=1
2
(X) (Exercise 15B.2). E _
Let G := 1, (
=
, (
[[
, (
. (
L
n
n=1
. (
R
n
n=1
. (
B
n
n=1
. (
T
n
n=1
, where,
for all (r, j) [0, ]
2
,
(
[[
(r, j) := r;
(
=
(r, j) := j;
and (
(r, j) :=
1
_
_
r

2
_
2
_
j

2
_
2
_
.
The graphs of (
[[
(r, j) and (
=
(r, j) are inclined planes at 45
o
in the r and
j directions respectively. The graph of (
is a saddle shape very similar to

Figure 1C.1(B) on page 10. Meanwhile, for all n 1, and all (r, j) [0, ]
2
,
we dene
(
L
n
(r, j) :=
cosh (n( r)) cos(nj)
n sinh(n)
;
(
R
n
(r, j) :=
cosh(nr) cos(nj)
n sinh(n)
;
(
B
n
(r, j) :=
cos(nr) cosh (n( j))
n sinh(n)
;
and (
T
n
(r, j) :=
cos(nr) cosh(nj)
n sinh(n)
.
Then G is a Neumann harmonic basis for X (Exercise 15B.3). E _
Example 15B.2. (a) If X = | = (:, ) ; : 1 (the unit disk in polar coor-
dinates), then X = S = (:, ) ; : = 1 (the unit circle). In this case, let
k=1
:= (
n
n=0
. o
n
n=1
, where, for all n N and [, ),
(
n
(, 1) := cos(n) and o
n
(, 1) := sin(n).
Then
k
k=1
is a basis of L
2
(S), by Theorem 8A.1. Let H :=
n
n=0
.
n=1
, where
0
1, and where, for all n 1 and (:, ) |, we dene
n
(:, ) := cos(n) :
n
and
n
(:, ) := sin(n) :
n
.
(See Figure 14B.1 on page 275). Then H is a Dirichlet harmonic basis for |;
this was the key fact employed by Proposition 14B.2 on page 278, to solve the
Laplace Equation on | with arbitrary nonhomogeneous Dirichlet boundary
conditions.
Suppose
1
= (
0
(i.e.
1
1). Let G := 1 .
n
,n
n=1
.
n
,n
n=1
,
where, for all n N and (:, ) |, we have
n
(:, ),n :=
cos(n) :
n
n
and
n
(:, ),n :=
sin(n) :
n
n
.
Then G is a Neumann harmonic basis for |; this was the key fact employed
by Proposition 14B.4 on page 280, to solve the Laplace Equation on | with
arbitrary nonhomogeneous Neumann boundary conditions.
(b) If X = |
= (:, ) ; : 1 (in polar coordinates)

1
, then X = S =
(:, ) ; : = 1. In this case, let
k
k=1
:= (
n
n=0
. o
n
n=1
, just as in
Example (a). However, this time, let H :=
0
.
n
n=1
.
n
n=1
, where
0
1, and where, for all n 1 and (:, ) |, we dene
n
(:, ) := cos(n),:
n
and
n
(:, ) := sin(n),:
n
.
(See Figure 14B.2 on page 276). Then H is a Dirichlet harmonic basis for |
;
this was the key fact employed by Proposition 14B.6 on page 284, to solve the
Laplace Equation on |
with arbitrary nonhomogeneous Neumann boundary

conditions.
Recall
1
= (
0
1. Let G := 1.
n
,n
n=1
.
n
,n
n=1
, where
n
and
n
are as dened above, for all n 1. Then G is a Neumann harmonic basis
for |
; this was the key fact employed by Proposition 14B.8 on page 285, to
solve the Laplace Equation on |
with arbitrary nonhomogeneous Neumann

2
Theorem 15B.3. General solution to Laplace equation

Let / L
2
(X) have orthogonal expansion (15B.1). Let H := H
k
k=1
be
a Dirichlet harmonic basis for X and let G := 1 . (
k
k=2
be a Neumann
harmonic basis for X.
1
Technically, we are here developing a theory for bounded domains, and |
is obviously
not bounded. But it is interesting to note that many our techniques still apply to |
. This is
because |
is conformally isomorphic to a bounded domain, once we regard |
as a subset of the
Riemann sphere by including the point at innity. See 18B on page 422 for an introduction
to conformal isomorphism. See Remark 18G.4 on page 469 for a discussion of the Riemann
sphere.
2
Note that our Neumann harmonic basis does not include the element 0(r, ) := log(r).
This is because
0 = 1. Of course, the domain |
is not bounded, so Corollary 5D.4(b)[i]

does not apply, and indeed 0 is a continuous harmonic function on |
. However, unlike the

elements of G, the function 0 is not bounded, and thus does not extend to a continuous real-
valued harmonic function when we embed |
in the Riemann sphere by adding the point at

innity.
(a) Let n

L2
k=1
1
k
H
k
. If this series converges uniformly to n on the interior
of X, then n is the unique continuous harmonic function with nonhomoge-
neous Dirichlet BC n(x) = /(x) for all x X.
(b) Suppose
1
1. If 1
1
,= 0, then there is no continuous harmonic function
on X with nonhomogeneous Neumann BC
n(x) = /(x) for all x X.

Suppose 1
1
= 0. Let n

L2
k=2
1
k
(
k
+ C, where C 1 is any con-
stant. If this series converges uniformly to n on the interior of X, then
it is a continuous harmonic function with nonhomogeneous Neumann BC
n(x) = /(x) for all x X. Furthermore, all solutions to this BVP have
this form, for some value of C 1.
Proof. Exercise 15B.4 Hint: The boundary conditions follow from expansion E _
(15B.1). To verify that n is harmonic, use the Mean Value Theorem (Theorem 1E.1
on page 16). (Use Proposition 6E.10(b) on page 127 to guarantee that the integral of
the sum is the sum of the integrals.) Finally, use Corollary 5D.4 on page 87 to show
solution uniqueness. 2
Exercise 15B.5. Show how Propositions 12A.4, 13B.2, 14B.2, 14B.4 14B.6, 14B.8 E _
and 14B.10 are all special cases of Theorem 15B.3.
Remark. There is nothing special about the role of the Laplacian in Theo-
rem 15B.3. If L is any linear dierential operator, then something like Theorem
15B.3 is still true if you replace with L everywhere. In particular, if L is
an elliptic dierential operator (see 5E), then Theorem 15B.3 becomes the gen-
eral solution to the boundary value problem for the homogeneous elliptic PDE
L n 0.
However, if L is an arbitrary dierential operator, then there is no guarantee
that you will nd a harmonic basis H
k
k=1
of functions such that L H
k
0
for all / N, and such that the collection H
k
k=1
(or
H
k
k=1
) provides
an orthonormal basis for L
2
(X). (Even for the Laplacian, this is a nontrivial
problem; see e.g. Corollary 15C.8 on page 333 below.)
Furthermore, once you dene n

L2
k=1
1
k
H
k
as in Theorem 15B.3, you might
not be able to use something like the Mean Value Theorem to guarantee that
L n = 0. Instead you must formally dierentiate the series

k=1
1
k
H
k
and
k=1
1
k
(
k
using Proposition 0F.1 on page 565. For this to work, you need some
convergence conditions on the formal derivatives of these series. For example,
if L was an th order dierential operator, it would be sucient to require
that
k=1
[1
k
[
_
_
N
j
H
k
_
_
< and
k=1
[1
k
[
_
_
N
j
(
k
_
_
< for all , [1...1]

(Exercise 15B.6 Verify this). E _
Finally, for an arbitrary dierential operator, there may not be a result like
Corollary 5D.4 on page 87, which guarantees a unique solution to a Dirich-
let/Neumann BVP. It may be necessary to impose further constraints to get a
unique solution.
15C Eigenbases on Cartesian products
Prerequisites: 4B(iv), 5B, 5C, 6F, 0D.
If X
1
1
D
1
and X
2
1
D
2
are two domains, then their Cartesian product
is the set
X
1
X
2
:= (x
1
, x
2
) ; x
1
X
1
and x
2
X
2
1
D
1
+D
2
.
Example 15C.1. (a) if X
1
= [0, ] 1 and X
2
= [0, ]
2
1
2
then X
1
X
2
=
[0, ]
3
1
3
.
(b) If X
1
= | 1
2
and X
2
= [0, ] 1, then X
1
X
2
= (:, , .) ; (:, )
| and 0 . 1
3
is the cylinder of height .
To apply the solution methods from Sections 15A and 15B, we must rst
construct eigenbases and/or harmonic bases on the domain X; that is the goal of
this section. We begin with some technical results which are useful and straight-
forward to prove.
Lemma 15C.2. Let X
1
1
D
1
and X
2
1
D
2
. Let X := X
1
X
2
1
D
1
+D
2
.
(a) X = [(X
1
) X
2
] [X
1
(X
2
)].
Let
1
: X
1
1 and
2
: X
2
1, and dene =
1

2
: X 1 by
(x
1
, x
2
) :=
1
(x
1
)
2
(x
2
) for all (x
1
, x
2
) X.
(b) If
1
satises homogeneous Dirichlet BC on X
1
and
2
satises homoge-
neous Dirichlet BC on X
2
, then satises homogeneous Dirichlet BC on
X.
(c) If
1
satises homogeneous Neumann BC on X
1
and
2
satises homoge-
neous Neumann BC on X
2
, then satises homogeneous Neumann BC on
X.
15C. Eigenbases on Cartesian products 331
(d) ||
2
= |
1
|
2
|
2
|
2
. Thus, if
1
L
2
(X
1
) and
2
L
2
(X
2
) then
L
2
(X).
(e) If
1
L
2
(X
1
) and
2
L
2
(X
2
) and =
1

2
then , ) =
1
,
1
)
2
,
2
).
(f ) Let
(1)
n

n=1
2
(X
1
) and let
(2)
m

m=1
be an
2
(X
2
). For all (n, :) N, let
n,m
:=
(1)
n

(2)
m
.
Then
n,m
n,m=1
2
(X).
Let
1
be the Laplacian operator on 1
D
1
, let
2
be the Laplacian operator on
1
D
2
, and let be the Laplacian operator on 1
D
1
+D
2
.
(g) (x
1
, x
2
) =
_
1

1
(x
1
)
_

2
(x
2
) +
1
(x
1
)
_
2

2
(x
2
)
_
.
(h) Thus, if
1
is an eigenfunction of
1
with eigenvalue
1
, and
2
is an
eigenfunction of
2
with eigenvalue
2
, then is an eigenfunction of
with eigenvalue (
1
+
2
).
Proof. Exercise 15C.1 (Remark: For part (f), just show that
n,m
n,m=1
is an E _
orthogonal collection of functions. Showing that
n,m
n,m=1
is actually a basis for
L
2
(X) requires methods beyond the scope of this course.) 2
Corollary 15C.3. Eigenbases for Cartesian Products
Let X
1
1
D
1
and X
2
1
D
2
. Let X := X
1
X
2
1
D
1
+D
2
. Let
(1)
n

n=1
be a Dirichlet (or Neumann) eigenbasis for L
2
(X
1
), and let
(2)
m

m=1
be a
Dirichlet (respectively Neumann) eigenbasis for L
2
(X
2
). For all (n, :) N,
dene
n,m
=
(1)
n

(2)
m
. Then
n,m
n,m=1
Dirichlet (respectively Neumann)
eigenbasis for L
2
(X).
Proof. Exercise 15C.2 Just combine Lemma 15C.2(b,c,f,h). 2 E _
Example 15C.4. Let X
1
= [0, ] and X
2
= [0, ]
2
, so X
1
X
2
= [0, ]
3
.
Note that ([0, ]
3
) =
_
0, [0, ]
2
_

_
[0, ] [0, ]
2
_
. For all / N,
dene C
and S
L
2
[0, ] by C
(r) := cos(/r) and S
(r) := sin(/r). For all

:, n N, dene C
m,n
and S
m,n
L
2
([0, ]
2
) by C
m,n
(j, .) := cos(:j) cos(n.)
and S
m,n
(j, .) := sin(:j) sin(n.).
For any /, :, n N, dene C
,m,n
and S
,m,n
L
2
(X) by C
,m,n
(r, j, .) :=
C
(r) C
m,n
(j, .) = cos(/r) cos(:j) cos(n.) and S
,m,n
(r, j, .) := S
(r)
S
m,n
(j, .) = sin(/r) sin(:j) sin(n.).
Now, S
=1
is a Dirichlet eigenbasis for [0, ] (by Theorem 7A.1), and
S
m,n
m,n=1
is a Dirichlet eigenbasis for [0, ]
2
(by Theorem 9A.3(a)); thus,
Corollary 15C.3 says that S
,m,n
,m,n=1
is a Dirichlet eigenbasis for [0, ]
3
(as earlier noted by Theorem 9B.1).
Likewise, C
=0
is a Neumann eigenbasis for [0, ] (by Theorem 7A.4), and
C
m,n
m,n=0
is a Neumann eigenbasis for [0, ]
2
; (by Theorem 9A.3(b)); thus,
Corollary 15C.3 says that C
,m,n
,m,n=0
is a Neumann eigenbasis for [0, ]
3
(as earlier noted by Theorem 9B.1).
1
= | and X
2
= [0, ], so that X
1
X
2
is the cylinder
of height and radius 1. Let S := | (the unit circle). Note that X =
_
S [0, ]
_

_
| 0,
_
. For all n N, dene S
n
L
2
[0, ] as in Example
15C.4. For all /, : N, let
,m
and
,m
be the type-1 Fourier-Bessel eigen-
functions dened by eqn.(14C.5) on page 296 of 14C(ii). For any /, :, n N,
dene
,m,n
and
,m,n
L
2
(X) by
,m,n
(:, , .) :=
,m
(:, ) S
n
(.) and
,m,n
(:, , .) :=
,m
(:, ) S
n
(.).
Now
m,n
,
m,n
m,n=1
is a Dirichlet eigenbasis for the disk | (by Theorem
14C.2) and S
n
n=1
is a Dirichlet eigenbasis for the line [0, ] (by Theorem
7A.1); thus, Corollary 15C.3 says that
,m,n
,
,m,n
,m,n=1
is a Dirichlet
eigenbasis for the cylinder X.
Exercise 15C.3. (a) Combine Example 15C.5 with Theorems 15A.2, 15A.3, and E _
15A.4 to provide a general solution method for solving the Poisson equation, heat equa-
tion, and wave equation on a nite cylinder with Dirichlet boundary conditions.
(b) Set up and solve some simple initial/boundary value problems using your method.
Exercise 15C.4. In cylindrical coordinates on 1

3
, let X = (:, , .); 1 :, 0 E _
. , and < be the punctured slab of thickness , having a cylindrical
hole of radius 1.
(a) Express X as a Cartesian product of the punctured plane and a line segment.
(b) Use Corollary 15C.3 to obtain a Dirichlet eigenbasis for X.
(c) Apply Theorems 15A.2, 15A.3, and 15A.4 to provide a general solution method
for solving the Poisson equation, heat equation, and wave equation on the punctured
slab with Dirichlet boundary conditions.
(d) Set up and solve some simple initial/boundary value problems using your method.
Exercise 15C.5. Let X

1
=
_
(r, j) [0, ]
2
; j r
_
be the right angle triangle E _
from Proposition 15A.5 on page 322, and let X
2
= [0, ] 1. Then X = X
1
X
2
is a
right-angle triangular prism.
(a) Use Proposition 15A.5 and Corollary 15C.3 to obtain Dirichlet and Neumann
eigenbases for the prism X.
(b) Apply Theorems 15A.2, 15A.3, and 15A.4 to provide a general solution method
for solving the Poisson equation, heat equation, and wave equation on the prism with
Dirichlet or Neumann boundary conditions.
(c) Set up and solve some simple initial/boundary value problems using your method.
We now move on to the problem of constructing harmonic bases on a Carte-

sian product. We will need two technical lemmas.
Lemma 15C.6. Harmonic functions on Cartesian products
Let X
1
1
D
1
and X
2
1
D
2
. Let X := X
1
X
2
1
D
1
+D
2
.
Let c
1
: X
1
1 be an eigenfunction of
1
with eigenvalue , and let
c
2
: X
2
1 be an eigenfunction of
2
with eigenvalue . If we dene
H := c
1
c
2
: X 1, as in Lemma 15C.2, then H is a harmonic function
that is, H = 0.
Proof. Exercise 15C.6 Hint: Use Lemma 15C.2(h). 2 E _
Lemma 15C.7. Orthogonal bases on almost-disjoint unions
Let Y
1
, Y
2
1
D
be two (11)-dimensional subsets (e.g. two curves in 1
2
,
two surfaces in 1
3
, etc.). Suppose that Y
1
Y
2
has dimension (12) (e.g. it
is a discrete set of points in 1
2
, or a curve in 1
3
, etc.). Let
(1)
n

n=1
be an
2
(Y
1
), such that
(1)
n
(y) = 0 for all y Y
2
and n N.
Likewise, let
(2)
n

n=1
2
(Y
2
), such that
(2)
n
(y) = 0
for all y Y
1
and n N. Then
(1)
n

n=1
.
(2)
n

n=1
is an orthogonal basis for
L
2
(Y
1
Y
2
).
Proof. Exercise 15C.7 Hint: the (11)-dimensional integral of any function on E _
Y
1
Y
2
must be zero. 2
For the rest of this section we adopt the following notational convention: if
) : X 1 is a function, then let

) denote the restriction of ) to a function
) : X 1 (that is,

) := )
[
X
).
Corollary 15C.8. Harmonic bases on Cartesian products
Let X
1
1
D
1
and X
2
1
D
2
. Let X := X
1
X
2
1
D
1
+D
2
.
Let
2
m
mM
2
2
(X
2
) (here, M
2
is some indexing
set, either nite or innite; e.g. M
2
= N). Let c
1
n
n=1
be a Dirichlet eigenbasis
for X
1
. For all n N, suppose
1
c
1
n
=
(1)
n
c
1
n
, and for all : M
2
, let
T
2
n,m
L
2
(X
2
) be an eigenfunction of
2
with eigenvalue +
(1)
n
, such that
T
2
n,m
=
2
m
. Let H
1
n,m
:= c
1
n
T
2
n,m
: X 1, for all n N and : M
2
.
Likewise, let
1
m
mM
1
2
(X
1
) (where M
1
is
some indexing set), and let c
2
n
n=1
be a Dirichlet eigenbasis for X
2
. For all
n N, suppose
2
c
2
n
=
(2)
n
c
2
n
, and for all : M
1
, let T
1
n,m
L
2
(X
1
) be
an eigenfunction of
1
with eigenvalue +
(2)
n
, such that

T
1
n,m
=
1
m
. Dene
H
2
n,m
:= T
1
n,m
c
2
n
: X 1, for all n N and : M
1
.
Then H := H
1
n,m
; n N, : M
2
. H
2
n,m
; n N, : M
1
is a
Dirichlet harmonic basis for L
2
(X).
Proof. Exercise 15C.8 (a) Use Lemma 15C.6 to verify that all the functions E _
H
1
n,m
and H
2
n,m
are harmonic on X.
(b) Show that
H
1
n,m
nN,mM2
2
_
X
1
(X
2
)
_
, while
H
2
n,m
nN,mM1
2
_
(X
1
) X
2
_
. Use Lemma 15C.2(f).
(c) Show that H is an orthogonal basis for L
2
(X). Use Lemma 15C.2(a) and Lemma
15C.7. 2
1
= [0, ] = X
2
, so that X = [0, ]
2
. Observe that
_
[0, ]
2
_
=
_
0, [0, ]
_

_
[0, ] 0,
_
.
Observe that X
1
= 0, = X
2
(a two-element set), and L
2
0, is 2-
dimensional vector space (isomorphic to 1
2
). Let M
1
:= 1, 2 =: M
2
. Let
1
1
=
2
1
=
1
and
1
2
=
2
2
=
2
, where
1
,
2
: 0, 1 are dened:
2
(0) := 1 =:
1
(), and
2
() := 0 =:
1
(0).
Then
1
,
2
2
0, . For all n N, let
c
1
n
(r) = c
2
n
(r) = c
n
(r) := sin(nr).
T
1
n,1
(r) = T
2
n,1
(r) = T
n,1
(r) := sinh(nr), sinh(n)
T
1
n,2
(r) = T
2
n,2
(r) = T
n,2
(r) := sinh (n( r)) , sinh(n).
Then c
n
n=1
is a Dirichlet eigenbasis for [0, ] (by Theorem 7A.1), while
T
n,1
=
1
and

T
n,2
=
2
for all n N.
For each n N, we have eigenvalue
n
:= n
2
. That is c
n
(r) = n
2
c
n
(r)
while T
n,m
(r) = n
2
T
n,m
(r). Thus, the functions H
n
(r, j) := c
n
(r)T
n,m
(j)
are harmonic, by Lemma 15C.6. Thus, if we dene
H
1
n,1
(r, j) := c
1
n
(r) T
2
n,1
(j) =
sin(nr) sinh(nj)
sinh(n)
,
H
1
n,2
(r, j) := c
1
n
(r) T
2
n,2
(j) =
sin(nr) sinh(n( j))
sinh(n)
,
H
2
n,1
(r, j) := T
1
n,1
(r) c
2
n
(j) =
sinh(nr) sin(nj)
sinh(n)
, and
H
2
n,2
(r, j) := T
1
n,2
(r) c
2
n
(j) =
sinh (n( r)) sin(nj)
sinh(n)
,
then Corollary 15C.8 says that the collection H
1
n,1
nN
.H
1
n,2
nN
.H
2
n,1
nN
.
H
2
n,2
nN
is a Dirichlet harmonic basis for [0, ]
2
a fact we already observed
in Example 15B.1(a), and exploited earlier in Proposition 12A.4.
1
= | 1
2
and X
2
= [0, ], so that X
1
X
2
1
3
is the cylinder of height and radius 1. Note that X =
_
S [0, ]
_

_
| 0,
_
.
For all n N and / N, let c
1
,n
:=
,n
, and c
1
,n
:=
,n
, where
,n
and
,n
are the type-1 Fourier-Bessel eigenfunctions dened by eqn.(14C.5) on
page 296 of 14C(ii). Then c
1
,n
; / N and n Z is a Dirichlet eigenbasis
for |, by Theorem 14C.2.
As in Example 15C.9, [0, ] = 0, . Let M
2
:= 0, 1 and let
2
1
: 0,
1 and
2
2
: 0, 1 be as in Example 15C.9. Let
,n
,n=1
be the roots
of the Bessel function
n
, as described in equation (14C.3) on page 296. For
every (/, n) N Z, dene T
2
,n;1
and T
2
,n;2
L
2
[0, ] by
T
2
,n;1
(.) :=
sinh
_
,[n[
.
_
sinh(
,[n[
)
and T
2
,n;2
(.) :=
sinh
_
,[n[
( .)
_
sinh(
,[n[
)
,
for all . [0, ]. Then clearly

T
2
,n;1
=
2
1
and

T
2
,n;2
=
2
2
.
For each (/, n) NZ, we have eigenvalue
2
,[n[
by equation (14C.4) on page
296. That is
,n
(:, ) =
2
,n

,n
(:, ) and
,n
(:, ) =
2
,n

,n
(:, );
thus, c
1
,n
(.) =
2
,[n[
c
1
,n
(.) for all (/, n) NZ. Meanwhile, T
1
,n;m
(.) =
2
,[n[
T
1
,n;m
(.), for all (/, n; :) N Z 1, 2. Thus, the functions
H
1
,n,1
(:, , .) := c
1
,n
(:, ) T
2
,n;1
(.) =

,n
(:, ) sinh(
,n
.)
sinh(
,n
)
,
H
1
,n,2
(:, , .) := c
1
,n
(:, ) T
2
,n;2
(.) =

,n
(:, ) sinh (
,n
( .))
sinh(
,n
)
,
H
1
,n,1
(:, , .) := c
1
,n
(:, ) T
2
,n;1
(.) =

,n
(:, ) sinh(
,n
.)
sinh(
,n
)
, and
H
1
,n,2
(:, , .) := c
1
,n
(:, ) T
2
,n;2
(.) =

,n
(:, ) sinh (
,n
( .))
sinh(
,n
)
are all harmonic, by Lemma 15C.6.
Recall that | = S. Let M
1
:= Z, and for all : Z, dene
1
m
L
2
(S) by
1
m
(1, ) := sin(:) (if : 0) and
1
m
(1, ) := cos(:) (if : 0), for all
[, ]; then
1
m
mZ
is an orthogonal bass for L
2
(S), by Theorem 8A.1.
For all n N and . [0, ], dene c
2
n
(.) := sin(n.) as in Example 15C.9.
Then c
2
n
n=1
is a Dirichlet eigenbasis for [0, ], by Theorem 7A.1. For all
n N, the eigenfunction c
2
n
has eigenvalue
(2)
n
:= n
2
. For all : Z, let
T
1
n,m
: | 1 be an eigenfunction of the Laplacian with eigenvalue n
2
, and
with boundary condition T
1
n,m
(1, ) =
1
m
() for all [, ] (see Exercise
15C.9(a) below). The function H
2
n,m
(:, , .) := T
1
n,m
(:, ) c
2
n
(.) is harmonic,
by Lemma 15C.6. Thus, Corollary 15C.8 says that the collection
_
H
1
,n,m
; / N, n Z, : = 1, 2
_
.
_
H
2
n,m
; n N, : Z
_
is a Dirichlet harmonic basis for the cylinder X.
Exercise 15C.9. (a) Example 15C.10 posits the existence of eigenfunctions T
1
n,m
: E _
| 1 of the Laplacian with eigenvalue n
2
and with boundary condition T
1
n,m
(1, ) =
1
m
() for all [, ]. Assume T
1
n,m
separates in polar coordinates that is,
T
1
n,m
(:, ) = 1(:)
m
(), where 1 : [0, 1] 1 is some unknown function with
1(1) = 1. Show that 1 must satisfy the ordinary dierential equation :
2
1
(:) +
:1
(:) (:
2
+1)n
2
1(:) = 0. Use the Method of Frobenius (0H(iii)) to solve this ODE
and get an expression for T
1
n,m
.
(b) Combine Theorem 15B.3 with Example 15C.10 to obtain a general solution
to the Laplace equation on a nite cylinder with nonhomogeneous Dirichlet boundary
conditions.
(c) Set up and solve a few simple Dirichlet problems using your method.
Exercise 15C.10. Let X = (:, , .) ; 1 : and 0 . be the punctured E _
slab from Exercise 15C.4.
(a) Use Corollary 15C.8 to obtain a Dirichlet harmonic basis for X.
(b) Apply Theorem 15B.3 to obtain a general solution to the Laplace equation on
the punctured slab with nonhomogeneous Dirichlet boundary conditions.
15D. General method for solving I/BVPs 337
Exercise 15C.11. Let X be the right-angle triangular prism from Exercise 15C.5. E _
(a) Use Proposition 15A.5 and Corollary 15C.8 to obtain a Dirichlet harmonic basis
for X.
(b) Apply Theorem 15B.3 to obtain a general solution to the Laplace equation on
the prism with nonhomogeneous Dirichlet boundary conditions.
Exercise 15C.12. State and prove a theorem analogous to Corollary 15C.8 for E _
Neumann harmonic bases.
15D General method for solving I/BVPs
Prerequisites: 15A, 15B. Recommended: 15C.
We now provide a general method for solving initial/boundary value prob-
lems. Throughout this section, let X 1
D
be a domain. Let L be a linear
dierential operator on X (e.g. L = ).
1. Pick a suitable coordinate system. Find the coordinate system where
your problem can be expressed in simplest form. Generally, this is a coordinate
system where the domain X can be described using a few simple inequalities.
For example, if X = [0, 1]
D
, then probably the Cartesian coordinate system is
best. If X = | or |
or A, then probably polar coordinates on 1

2
are the most
suitable. If X = B or X = B, then probably spherical polar coordinates on 1
3
are best.
If the dierential operator L has nonconstant coecients, then you should
also seek a coordinate system where these coecients can be expressed using the
simplest formulae. (If L = , then it has constant coecients, so this is not an
issue).
Finally, if several coordinate systems are equally suitable for describing X and
L, then nd the coordinate system where the initial conditions and/or boundary
conditions can be expressed most easily. For example, if X = 1
2
and L = ,
then either Cartesian or polar coordinates might be appropriate. However, if the
initial conditions are rotationally symmetric around zero, then polar coordinates
would be more appropriate. If the initial conditions are invariant under transla-
tion in some direction, then Cartesian coordinates would be more appropriate.
Note. Dont forget to nd the correct expression for L in the new coordinate
system. For example, in Cartesian coordinates on 1
2
, we have n(r, j) =
2
y
n(r, j) +
2
y
n(r, j). However, in polar coordinates, n(:, ) =
2
r
n(:, ) +
1
r

r
n(:, ) +
1
r
2

2
n(:, ). If you apply the Cartesian Laplacian to a function

expressed in polar coordinates, the result will be nonsense.
2. Eliminate irrelevant coordinates. A coordinate r is irrelevant if:
(a) membership in the domain X does not depend on this coordinate; and
(b) the coecients of L do not depend on this coordinate; and
(c) the initial and/or boundary conditions do not depend on this coordinate.
In this case, we can eliminate the r coordinate from all equations, by ex-
pressing the domain X, the operator L and the initial/boundary conditions as
functions of only the non-r coordinates. This reduces the dimension of the prob-
lem, thereby simplifying it.
To illustrate (a), suppose X = | or |
or A, and we use the polar coordinate

system (:, ); then the angle coordinate is irrelevant to membership in X. On
the other hand, suppose X = 1
2
[0, 1] is the slab of thickness 1 in 1
3
, and we
use Cartesian coordinates (r, j, .). Then the coordinates r and j are irrelevant
to membership in X.
If L = or any other dierential operator with constant coecients, then
(b) is automatically satised.
To illustrate (c), suppose X = | and we use polar coordinates. Let ) : |
1 be some initial condition. If )(:, ) is a function only of :, and doesnt depend
on , then is a redundant coordinate and can be eliminated, thereby reducing
the BVP to a one-dimensional problem, as in Example 14F.3 on page 305.
On the other hand, let / : S 1 be a boundary condition. Then is only
irrelevant if / is a constant function (otherwise / has nontrivial dependence on
).
Now, suppose X = 1
2
[0, 1] is the slab of thickness 1 in 1
3
. If the
boundary condition / : X 1 is constant on the top and bottom faces of the
slab, then the r and j coordinates can be eliminated, thereby reducing the BVP
to a one-dimensional problem: a BVP on the line segment [0, 1], which can be
solved using the methods of Chapter 11.
In some cases, a certain coordinate can be eliminated if it is approximately
irrelevant. For example, if the domain X is particularly long in the r dimen-
sion relative to its other dimensions, and the boundary conditions are roughly
constant in the r dimension, then we can approximate long with innite
and roughly constant with exactly constant, and eliminate the r dimension
from the problem. This method was used in Example 12B.2 on page 248 (the
quenched rod), Example 12B.7 on page 252 (the baguette), and Example
12C.2 on page 255 (the nuclear fuel rod).
3. Find an eigenbasis for L
2
(X). If X is one of the standard domains we
have studied in this book, then use the eigenbases we have introduced in Chap-
ters 7-9, Section 14C, or Section 15C. Otherwise, you must construct a suitable
eigenbasis. Theorem 15E.12 (page 347) guarantees that such an eigenbasis exists,
15D. General method for solving I/BVPs 339
but it doesnt tell you how to construct it. The actual construction of eigenbases
is usually done using Separation of Variables, discussed in Chapter 16. The sep-
aration of the time variable is really just a consequence of the fact that we
have an eigenfunction. The separation of the space variables is not necessary
to get an eigenfunction, but it is very convenient, for two reasons:
1. Separation of variables is a powerful strategy for nding the eigenfunctions;
it reduces the problem to set of independent ODEs which can each be solved
using classical ODE methods.
2. If an eigenfunction c
n
appears in separated form, then it is often easier to
compute the inner product c
n
, )), where ) is some other function. This is
important when the eigenfunctions form an orthogonal basis, and we want
to compute the coecients of ) in this basis.
4. Find a harmonic basis for L
2
(X) (if there are nonhomogeneous bound-
ary conditions). The same remarks apply as in Step 3.
5. Solve the problem Express any initial conditions in terms of the eigenbasis
from step #3, as described in 15A Express any boundary conditions in terms
of the harmonic basis from step #4, as described in 15B.
If L = , then use Theorems 15A.2, 15A.3, 15A.4, and/or 15B.3. If L is
some other linear dierential operator, then use the appropriate analogues of
these theorems (see Remark 15A.6).
6. Verify convergence. Note that Theorems 15A.2, 15A.3, 15A.4, and/or
15B.3 require the eigenvalue sequences
n
n=1
and/or j
n
n=1
to grow at a
certain speed, or require the coecient sequences
n
n=0
and 1
n
n=1
to decay
at a certain speed, so as to guarantee that the solution series and its formal
derivatives are absolutely convergent. These conditions are important, and must
be checked. Typically, if L = , the growth conditions on
n
n=1
and j
n
n=1
are easily satised. However, if you try to extend these theorems to some other
linear dierential operator, the conditions on
n
n=1
and j
n
n=1
must be
checked.
7. Check the uniqueness of the solution. Section 5D describes condi-
tions under which boundary value problems for the Poisson, Laplace, Heat, and
wave equations will have a unique solution. Check that these conditions are
satised. If L ,= , then you will need to establish solution uniqueness using
theorems analogous to those found in Section 5D. (General theorems for the ex-
istence/uniqueness of solutions to I/BVPs can be found in most advanced texts
on PDE theory, such as [Eva91]).
If the solution is not unique, then it is important to enumerate all solutions to
the problem. Remember that your ultimate goal here is to predict the behaviour
of some physical system in response to some initial or boundary condition. If the
solution to the I/BVP is not unique, then you cannot make a precise prediction;
instead, your prediction must take the form of a precisely specied range of
possible outcomes.
15E Eigenfunctions of self-adjoint operators
Prerequisites: 4B(iv), 5C, 6F. Recommended: 7A, 8A, 9B, 15A.
The solution methods of Section 15A are only relevant if we know that a
suitable eigenbasis for the Laplacian exists on the domain of interest. If we
want to develop similar methods for some other linear dierential operator L
(as described in Remark 15A.6 on page 323), then we must rst know that
suitable eigenbases exists for L. In this section, we will discuss the eigenfunctions
and eigenvalues of an important class of linear operators: self-adjoint operators.
This class includes the Laplacian and all other symmetric elliptic dierential
operators.
A linear operator 1 : 1
D
1
D
is self-adjoint if, for any vectors x, y 1
D
,
1(x), y) = x, 1(y)) .
Example 15E.1. The matrix
_
1 2
2 1
_
denes a self-adjoint operator on 1
2
,
because for any x =
_
x
1
x
2
_
and y =
_
y
1
y
2
_
in 1
2
, we have
1(x), y) =
__
x
1
2x
2
x
2
2x
1
_
,
_
y
1
y
2
__
= j
1
_
r
1
2r
2
_
+j
2
_
r
2
2r
1
_
= r
1
_
j
1
2j
2
_
+r
2
_
j
2
2j
1
_
=
__
x
1
x
2
_
,
_
y
1
2y
2
y
2
2y
1
__
= x, 1(y)) .
Theorem 15E.2. Let 1 : 1
D
1
D
be a linear operator with matrix A.
Then 1 is self-adjoint if and only if A is symmetric (i.e. o
ij
= o
ji
for all ,, i)
Proof. Exercise 15E.1 . 2 E _
A linear operator L : (
is self-adjoint if, for any two functions

), p (
,
L[)], p) = ), L[p])
whenever both sides are well-dened
3
.
3
This is an important point. Often, one of these inner products (say, the left one) will
not be well-dened, because the integral
_
X
L[f] g dx does not converge, in which case self-
adjointness is meaningless.
15E. Eigenfunctions of self-adjoint operators 341
Example 15E.3: Multiplication Operators are Self-Adjoint.
Let X 1
D
be any bounded domain. Let (
:= (
(X; 1). Fix (
(X),
and dene the operator Q : (
by Q()) := ) for any ) (
. Then
Q is self-adjoint. To see this, let ), p (
. Then
), p) =
_
X
( )) p dr =
_
X
) ( p) dr = ), p) .
(These integrals are all well-dened because , ) and p are all continuous and
hence bounded on X.)
Let 1 0, and consider the interval [0, 1]. Recall that (
[0, 1] is the set of

all smooth functions from [0, 1] into 1, and that:
(
0
[0, 1] is the space of all ) (
[0, 1] satisfying homogeneous Dirichlet bound-

ary conditions: )(0) = 0 = )(1) (see 5C(i)).
(
[0, 1] satisfying ) : [0, 1] 1 satisfy-

ing homogeneous Neumann boundary conditions: )
t
(0) = 0 = )
t
(1) (see
5C(ii)).
(
per
[0, 1] satisfying ) : [0, 1] 1 satisfying

periodic boundary conditions: )(0) = )(1) and )
t
(0) = )
t
(1) (see 5C(iv)).
(
h,h
[0, 1] satisfying homogeneous mixed bound-

ary conditions, for any xed real numbers /(0), /
(0), /(1) and /
(1)
(see 5C(iii)).
When restricted to these function spaces, the one-dimensional Laplacian op-
erator
2
x
is self-adjoint.
Proposition 15E.4. Let 1 0, and consider the operator
2
x
on (
[0, 1].
(a)
2
x
is self-adjoint when restricted to (
0
[0, 1].
(b)
2
x
[0, 1].
(c)
2
x
per
[0, 1].
(d)
2
x
h,h
[0, 1], for any /(0), /
(0), /(1)
and /
(1) in 1.
Proof. Let ), p : [0, 1] 1 be smooth functions. We apply integration by
parts to get:
2
x
), p
_
=
_
L
0
)
tt
(r) p(r) dr = )
t
(r) p(r)
x=L
x=0
_
L
0
)
t
(r) p
t
(r) dr.
(15E.1)
But if we apply Dirichlet, Neumann, or Periodic boundary conditions, we get:
)
t
(r) p(r)
x=L
x=0
= )
t
(1) p(1) )
t
(0) p(0)
=
_
_
_
)
t
(1) 0 )
t
(0) 0 = 0 (if homog. Dirichlet BC)
0 p(1) 0 p(0) = 0 (if homog. Neumann BC)
)
t
(0) p(0) )
t
(0) p(0) = 0 (if Periodic BC)
= 0 in all cases.
Thus, the rst term in (15E.1) is zero, so

2
x
), p
_
=
_
L
0
)
t
(r) p
t
(r) dr.
But by the same reasoning, with ) and p interchanged,
_
L
0
)
t
(r) p
t
(r) dr =
),
2
x
p
_
.
Thus, weve proved parts (a), (b), and (c). To prove part (d), rst note that
)
t
(r) p(r)
x=L
x=0
= )
t
(1) p(1) )
t
(0) p(0)
= )(1)
/(1)
/
(1)
p(1) + )(0)
/(0)
/
(0)
p(0)
= )(1) p
t
(1) )(0) p
t
(0) = )(r) p
t
(r)
x=L
x=0
.
Hence, substituting )(r) p
t
(r)
x=L
x=0
for )
t
(r) p(r)
x=L
x=0
in (15E.1), we get:
2
x
), p
_
=
_
L
0
)
tt
(r) p(r) dr =
_
L
0
)(r) p
tt
(r) dr =

),
2
x
p
_
. 2
Proposition 15E.4 generalizes to higher-dimensional Laplacians in the obvious
way:
Theorem 15E.5. Let 1 0.
(a) The Laplacian operator is self-adjoint on any of the spaces: (
0
[0, 1]
D
,
(
[0, 1]
D
, (
h,h
[0, 1]
D
or (
per
[0, 1]
D
.
(b) More generally, if X 1
D
is any bounded domain with a smooth bound-
ary
4
, then the Laplacian operator is self-adjoint on any of the spaces:
(
0
(X), (
(X), or (
h,h
(X).
In other words, the Laplacian is self-adjoint whenever we impose homogeneous
Dirichlet, Neumann, or mixed boundary conditions, or (when meaningful) peri-
odic boundary conditions.
Proof. (a) Exercise 15E.2 Hint: The argument is similar to Proposition 15E.4. E _
Apply integration by parts in each dimension, and cancel the boundary terms using
the boundary conditions.
(b) Exercise 15E.3 Hint: Use Greens Formulae (Corollary 0E.5(c) on page 564) E _
to set up an integration by parts argument similar to Proposition 15E.4. 2
If L
1
and L
2
are two self-adjoint operators, then their sum L
1
+ L
2
is also
self-adjoint (Exercise 15E.4). E _
Example 15E.6. Let X 1
D
be some domain (e.g. a cube), and let \ : X 1
be a potential describing the force acting on a quantum particle (e.g. an
electron) conned to the region X by an innite potential barrier along X.
Consider the Hamiltonian operator H dened in Section 3B on page 41:
H(x) =
/
2
2 :
(x) + \ (x) (x), for all x X.
(Here, / is Planks constant, : is the mass of the particle, and (
0
X is
its wavefunction.) The operator H is self-adjoint on (
0
(X). To see this, note
that H[] =
/
2
2 m
+ V[], where V[] = \ . Now, is self-adjoint
by Theorem 15E.5(b), and V is self-adjoint from Example 15E.3; thus, their
sum H is also self-adjoint. The stationary Schrodinger equation
5
H =
simply says that is an eigenfunction of H with eigenvalue .
Example 15E.7. Let :, : [0, 1] 1 be dierentiable. The Sturm-
Liouville operator
SL
s,q
[)] := : )
tt
+:
t
)
t
+ )
is self-adjoint on any of the spaces (
0
[0, 1], (
[0, 1], (
h,h
[0, 1] or (
per
[0, 1].
To see this, notice that
SL
s,q
[)] = (: )
t
)
t
+ ( )) = S[)] + Q[)], (15E.2)
where Q[)] = ) is just a multiplication operator, and S[)] = (: )
t
)
t
. We
know that Q is self-adjoint from Example 15E.3. We claim that S is also
self-adjoint. To see this, note that:
S[)], p) =
_
L
0
(: )
t
)
t
(r) p(r) dr
()
:(r) )
t
(r) p(r)
x=L
x=0
_
L
0
:(r) )
t
(r) p
t
(r) dr
4
See page 85 of 5D.
5
See 3C.
()
:(r) )
t
(r) p(r)
x=L
x=0
:(r) )(r) p
t
(r)
x=L
x=0
+
_
L
0
)(r) (: p
t
)
t
(r) dr
()
_
L
0
)(r) (: p
t
)
t
(r) dr = ), S[p]) .
Here, each () is integration by parts, and () follows from any of the cited
boundary conditions as in Proposition 15E.4 on page 341 (Exercise 15E.5). E _
Thus, S is self-adjoint, so SL
s,q
= S + Q is self-adjoint.
If SL
s,q
is a Sturm-Liouville operator, then the corresponding Sturm-Liouville
equation is the linear ordinary dierential equation
SL
s,q
[)] = ). (15E.3)
where ) : [0, 1] C and C are unknown. Clearly, equation (15E.3) simply
asserts that ) is an eigenfunction of SL
s,q
, with eigenvalue . Sturm-Liouville
equations appear frequently in the study of ordinary and partial dierential
equations.
Example 15E.8. (a) The one-dimensional Helmholtz equation )
tt
(r) = )(r)
is a Sturm-Liouville equation, with : 1 (constant) and 0.
(b) The one-dimensional stationary Schrodinger equation
/
2
2 :
)
tt
(r) + \ (r) )(r) = )(r), for all r [0, 1].
is a Sturm-Liouville equation, with :
/
2
2 m
(constant) and (r) := \ (r).
(c) The Cauchy-Euler equation
6
r
2
)
tt
(r) +2r)
t
(r) )(r) = 0 is a Sturm-
Liouville equation: let :(r) := r
2
and 0
(d) The Legendre equation
7
(1 r
2
) )
tt
(r) 2r)
t
(r) +j)(r) = 0 is a Sturm-
Liouville equation: let :(r) := (1 r
2
), 0, and let := j.
For more information about Sturm-Liouville problems, see [Bro89, 2.6, pp.39-
44], [Pow99, 2.7, pp.146-150], [Con90, II.6, pp.49-53], and especially [CB87,
Chap.6, pp.159-203].
Examples 15E.6 and 15E.8, Theorem 15E.5, and the solution methods of 15A
all illustrate the importance of the eigenfunctions of self-adjoint operators. One
nice property of self-adjoint operators is that their eigenfunctions are orthogonal.
6
See equation (14H.5) on page 314, and equation (16D.20) on page 361.
7
See equation (16D.19) on page 361.
Proposition 15E.9. Suppose L is a self-adjoint operator. If )
1
and )
2
are
eigenfunctions of L with eigenvalues
1
,=
2
, then )
1
and )
2
are orthogonal.
Proof. By hypothesis, L[)
k
] =
k
)
k
, for / = 1, 2. Thus,
1
)
1
, )
2
) =
1
)
1
, )
2
) = L[)
1
], )
2
)
()
)
1
, L[)
2
]) = )
1
,
2
)
2
) =
2
)
1
, )
2
) ,
where () follows from self-adjointness. Since
1
,=
2
, this can only happen
if )
1
, )
2
) = 0. 2
Example 15E.10. Eigenfunctions of
2
x
(a) Let
2
x
act on (
[0, 1]. Then all real numbers 1 are eigenvalues of

2
x
.
For any j 1,
If = j
2
0, the eigenfunctions are of the form (r) = sinh(j
r) +1cosh(j r) for any constants , 1 1.
If = 0, the eigenfunctions are of the form (r) = r + 1 for any
constants , 1 1.
If = j
2
< 0, the eigenfunctions are of the form (r) = sin(j
r) +1cos(j r) for any constants , 1 1.
Note: Because we have not imposed any boundary conditions, Proposition
15E.4 does not apply; indeed
2
x
is not a self-adjoint operator on (
[0, 1].
(b) Let
2
x
act on (
([0, 1]; C). Then all complex numbers C are eigenval-

ues of
2
x
. For any j C, with = j
2
, the eigenvalue has eigenfunctions
of the form (r) = exp(j r) +1exp(j r) for any constants , 1 C.
(Note that the three cases of the previous example arise by taking 1.)
Again, Proposition 15E.4 does not apply in this case, because
2
x
is not a
self-adjoint operator on (
([0, 1]; C).

(c) Now let
2
x
act on (
0
[0, 1]. Then the eigenvalues of
2
x
are
n
=
_
n
1
_
2
for every n N, each of multiplicity 1; the corresponding eigenfunctions
are all scalar multiples of S
n
(r) := sin
_
nx
L
_
.
(d) If
2
x
acts on (
[0, 1], then the eigenvalues of

2
x
are again
n
=
_
n
1
_
2
for every n N, each of multiplicity 1, but the corresponding eigenfunctions
are now all scalar multiples of C
n
(r) := cos
_
nx
L
_
. Also, 0 is an eigenvalue,
with eigenfunction C
0
= 11.
(e) Let / 0, and let
2
x
act on ( = ) (
[0, 1] ; )(0) = 0 and / )(1) +)

t
(1) = 0.
Then the eigenfunctions of
2
x
are all scalar multiples of
n
(r) := sin(j
n
r) ,
with eigenvalue
n
= j
2
n
, where j
n
0 is any real number such that
tan(1 j
n
) =
j
n
/
This is a transcendental equation in the unknown j
n
. Thus, although
there is an innite sequence of solutions j
0
< j
1
< j
2
< . . ., there is
no closed-form algebraic expression for j
n
. At best, we can estimate j
n
through numerical methods.
(f) Let /(0), /
(0), /(1), and /
(1) be real numbers, and let

2
x
act on
(
h,h
[0, 1]. Then the eigenfunctions of

2
x
are all scalar multiples of
n
(r) := sin
_
n
+j
n
r
_
,
with eigenvalue
n
= j
2
n
, where
n
[0, 2] and j
n
0 are constants
satisfying the transcendental equations:
tan (
n
) = j
n

/
(0)
/(0)
and tan(j
n
1 +
n
) = j
n

/
(1)
/(1)
.
(Exercise 15E.6). In particular, if /
(0) = 0, then we must have = 0. E _

If /(1) = / and /
(1) = 1, then we return to Example (e).

(g) Let
2
x
act on (
per
[1, 1]. Then the eigenvalues of
2
x
are again
n
=
_
n
1
_
2
, for every n N, each having multiplicity 2. The corresponding
eigenfunctions are of the form S
n
+1C
n
for any , 1 1. In particular,
0 is an eigenvalue, with eigenfunction C
0
= 11.
(h) Let
2
x
act on (
per
([1, 1]; C). Then the eigenvalues of
2
x
are again
n
=
_
n
1
_
2
, for every n N, each having multiplicity 2. The corresponding
eigenfunctions are of the form E
n
+ 1 E
n
for any , 1 1, where
E
n
(r) := exp
_
inr
1
_
. In particular 0 is an eigenvalue, with eigenfunction
E
0
= 11.
Example 15E.11. Eigenfunctions of
(a) Let act on (
0
[0, 1]
D
. Then the eigenvalues of are the numbers
m
:=
_
1
_
2
|m|
2
for all m N
D
+
. (Here, if m = (:
1
, . . . , :
D
), then
|m|
2
:= :
2
1
+. . . +:
2
d
). The eigenspace of
m
is spanned by all functions
S
n
(r
1
, ..., r
D
) := sin
_
n
1
r
1
1
_
sin
_
n
2
r
2
1
_
sin
_
n
D
r
D
1
_
,
for all n = (n
1
, . . . , n
D
) N
D
+
such that |n| = |m|.
(b) Now let act on (
[0, 1]
D
. Then the eigenvalues of are
m
for all
m N
D
. The eigenspace of
m
C
n
(r
1
, ..., r
D
) := cos
_
n
1
r
1
1
_
cos
_
n
2
r
2
1
_
cos
_
n
D
r
D
1
_
,
for all n N
D
such that |n| = |m|. In particular, 0 is an eigenvalue whose
eigenspace is the set of constant functions i.e. multiples of C
0
= 11.
(c) Let act on (
per
[1, 1]
D
. Then the eigenvalues of are again
m
for all
m N
D
. The eigenspace of
m
contains C
n
and S
n
for all n N
D
such
that |n| = |m|.
(d) Let act on (
per
_
[1, 1]
D
; C
_
. Then the eigenvalues of are again
m
for all m N
D
. The eigenspace of
m
E
n
(r
1
, . . . , r
D
) := exp
_
in
1
r
1
1
_
exp
_
in
D
r
D
1
_
,
for all n Z
D
such that |n| = |m|.
The alert reader will notice that, in each of the above scenarios (except Exam-
ples 15E.10(a) and 15E.10(b), where
2
x
is not self-adjoint), the eigenfunctions
are not only orthogonal, but actually form an orthogonal basis for the corre-
sponding 1
2
-space. This is not a coincidence. If ( is a subspace of L
2
(X), and
L : ( ( is a linear operator, then a set
n
n=1
( is an L-eigenbasis for
L
2
(X) if
n
n=1
2
(X), and for every n N,
n
is
an eigenfunction for L.
Theorem 15E.12. Eigenbases of the Laplacian
(a) Let 1 0. Let ( be any one of (
0
[0, 1]
D
, (
[0, 1]
D
, or (
per
[0, 1]
D
,
and treat as a linear operator on (. Then there is a -eigenbasis for
L
2
[0, 1]
D
consisting of elements of (. The corresponding eigenvalues of
are the values
m
dened in Example 15E.11(a), for all m N
D
.
(b) More generally, if X 1
D
is any bounded open domain, then there is a
-eigenbasis for L
2
[X] consisting of elements of (
0
[X]. The corresponding
eigenvalues of on ( form a decreasing sequence 0
1

2

3

with lim
n
n
= .
In both (a) and (b), some of the eigenspaces may be many-dimensional.
Proof. (a) we have already established. The eigenfunctions of the Lapla-
cian in these contexts are
_
C
n
; n N
D
_
and/or
_
S
n
; n N
D
+
_
. Theorem
8A.1(a) on page 162 and Theorem 9B.1(a) on page 187 tell us that these form
orthogonal bases for L
2
[0, 1]
D
.
(b) follows from Theorem 15E.17 on the next page. Alternately, see [War83],
Chapter 6, p. 255; exercise 16(g), or [Cha93], Theorem 3.21, p. 156. 2
Example 15E.13. (a) Let B =
_
x 1
D
; |x| < 1
_
be the ball of radius 1.
Then there is a -eigenbasis for L
2
(B) consisting of functions which are zero
on the spherical boundary of B.
(b) Let A =
_
(r, j) 1
2
; :
2
< r
2
+j
2
< 1
2
_
be the annulus of inner radius
: and outer radius 1 in the plane. Then there is a -eigenbasis for L
2
(A)
consisting of functions which are zero on the inner and outer boundary circles
of A.
Theorem 15E.14. Eigenbases for Sturm-Liouville operators
Let 1 0, let :, : [0, 1] 1 be dierentiable functions, and let SL
s,q
be the
Sturm-Liouville operator dened by : and on (
0
[0, 1]. Then there exists an
SL
s,q
-eigenbasis for L
2
[0, 1] consisting of elements of (
0
[0, 1]. The corresponding
eigenvalues of SL
s,q
form an innite increasing sequence 0 <
0
<
1
<
2
< . . .,
with lim
n
n
= . Each eigenspace is one-dimensional.
Proof. See [Tit62, Theorem 1.9]. For a proof in the special case when : 1,
see [Con90, Theorem 6.12, p.52]. 2
Symmetric Elliptic Operators. The rest of this section concerns the eigen-
functions of symmetric elliptic operators. (Please see 5E for the denition of
an elliptic operator.)
Lemma 15E.15. Let X 1
D
. If L is an elliptic dierential operator on
(
(X), then there are functions

cd
: X 1 for all c, d [1...1], and functions
,
1
, . . . ,
D
: X 1 such that L can be written in divergence form:
L[n] =
D
c,d=1
c
(
cd

d
n) +
D
d=1
d

d
n + n,
= div [ ] + , ) + n,
where =
_
_
_
1
.
.
.
D
_
_, and =
_
_
_
11
. . .
1D
.
.
.
.
.
.
.
.
.
D1
. . .
DD
_
_ is a symmetric, positive-denite
matrix.
Proof. Exercise 15E.7 Hint. Use the same strategy as as equation (15E.2). 2 E _
L is called symmetric if, in the divergence form, 0. For example, in the
case when L = , we have = Id and = 0, so is symmetric.
Theorem 15E.16. If X 1
D
is an open bounded domain, then any symmetric
elliptic dierential operator on (
0
(X) is self-adjoint.
Proof. This is a generalization of the integration-by-parts argument used to
prove Proposition 15E.4 on page 341 and Theorem 15E.5 on page 342. See
[Eva91, 6.5, p.334]. 2
Theorem 15E.17. Let X 1
D
be an open, bounded domain, and let L
be any symmetric, elliptic dierential operator on (
0
(X). Then there exists
an L-eigenbasis for L
2
(X) consisting of elements of (
0
(X). The corresponding
eigenvalues of L form an innite decreasing series 0
0

1

2
. . ., with
lim
n
n
= .
Proof. See of [Eva91, Theorem 1, 6.5.1, p.335]. 2
Remark. Theorems 15E.12, 15E.14, and 15E.17 are all manifestations of a far
more general result, the Spectral Theorem for Unbounded Self-Adjoint Opera-
tors. Unfortunately, it would take us too far aeld to even set up the necessary
background to precisely state this theorem. See [Con90, X.4 p. 319] for a good
exposition.
Further reading
The study of eigenfunctions and eigenvalues is sometimes called spectral theory.
For a good introduction to the spectral theory of linear operators on function
spaces, see [Con90]. An analogy of the Laplacian can be dened on any Rieman-
nian manifold; it is often called the Laplace-Beltrami operator, and its eigenfunc-
tions reveal much about the geometry of the manifold; see [War83, Chap.6] or
[Cha93, 3.9]. In particular, the eigenfunctions of the Laplacian on spheres have
been extensively studied. These are called spherical harmonics, and a sort of
Fourier theory can be developed on spheres, analogous to multivariate Fourier
theory on the cube [0, 1]
D
, but with the spherical harmonics forming the or-
thonormal basis [Tak94, M ul66]. Much of this theory generalizes to a broader
family of manifolds called symmetric spaces [Ter85, Hel81]. The eigenfunctions
of the Laplacian on symmetric spaces are closely related to the theory of Lie
groups and their representations [CW68, Sug75], a subject which is sometimes
called noncommutative harmonic analysis [Tay86].
351
V Miscellaneous solution
methods
In Chapters 11 to 15, we saw how initial/boundary value problems for linear
partial dierential equations could be solved by rst identifying an orthogonal
basis of eigenfunctions for the relevant dierential operator (usually the Lapla-
cian), and then representing the desired initial conditions or boundary conditions
as an innite summation of these eigenfunctions. For each bounded domain, each
boundary condition, and each coordinate system we considered, we found a sys-
tem of eigenfunctions that was adapted to that domain, boundary conditions,
and coordinate system.
This method is extremely powerful, but it raises several questions:
1. What if you are confronted with a new domain or coordinate system, where
none of the known eigenfunction bases is applicable? Theorem 15E.12 on
page 347 says that a suitable eigenfunction basis for this domain always
exists, in principle. But how do you go about discovering such a basis in
practice? For that matter, how were eigenfunctions bases like the Fourier-
Bessel functions discovered in the rst place? Where did Bessels equation
come from?
2. What if you are dealing with an unbounded domain, such as diusion in 1
3
?
In this case, Theorem 15E.12 is not applicable, and in general, it may not be
possible (or at least, not feasible) to represent initial/boundary conditions
in terms of eigenfunctions. What alternative methods are available?
3. The eigenfunction method is dicult to connect to our physical intuitions.
For example, intuitively, heat seaps slowly through space, and temper-
ature distributions gradually and irreversibly decay towards uniformity.
It is thus impossible to send a long-distance signal using heat. On the
other hand, waves maintain their shape and propagate across great dis-
tances with a constant velocity; hence they can be used to send signals
through space. These familiar intiutions are not explained or justied by
the eigenfunction method. Is there an alternative solution method where
these intuitions have a clear mathematical expression?
Part V provides answers to these questions. In Chapter 16, we introduce
a powerful and versatile technique called separation of variables, to construct
eigenfunctions adapted to any coordinate system. In Chapter 17, we develop
the entirely dierent solution technology of impulse-response functions, which
allows you to solve dierential equations on unbounded domains, and which has
an an appealing intuitive interpretation. Finally, in Chapter 18, we explore some
352 DRAFT
surprising and beautiful applications of complex analysis to harmonic functions
and Fourier theory.
353
Chapter 16
Separation of variables
Before creation God did just pure mathematics. Then He thought it would be a pleasant
change to do some applied. J. E. Littlewood
16A Separation of variables in Cartesian coordinates
on 1
2
Prerequisites: 1B, 1C.
A function n : 1
2
1 is said to separate if we can write n(r, j) =
A(r) Y (j) for some functions A, Y : 1 1. If n is a solution to some partial
dierential equation, we say n is a separated solution.
Example 16A.1. The heat equation on 1
We wish to nd n : 1 1
,
1 such that
t
n =
2
x
n. Suppose n(r; t) =
A(r) T(t), where
A(r) = exp(ijr) and T(t) = exp(j
2
t),
for some constant j 1. Then n(r; t) = exp(jir j
2
t), so that
2
x
n =
j
2
n =
t
n. Thus, n is a separated solution to the heat equation.
Separation of variables is a strategy for for solving partial dierential equa-
tions by specically looking for separated solutions. At rst, it seem like we are
making our lives harder by insisting on a solution in separated form. However,
often, we can use the hypothesis of separation to actually simplify the problem.
Suppose we are given some PDE for a function n : 1
2
1 of two variables.
Separation of variables is the following strategy:
1. Hypothesize that n can be written as a product of two functions, A(r) and
Y (j), each depending on only one coordinate; in other words, assume that
n(r, j) = A(r) Y (j) (16A.1)
354 DRAFT Chapter 16. Separation of variables
2. When we evaluate the PDE on a function of type (16A.1), we may nd
that the PDE decomposes into two separate, ordinary dierential equations
for each of the two functions A and Y . Thus, we can solve these ODEs
independently, and combine the resulting solutions to get a solution for n.
Example 16A.2. Laplaces Equation in 1
2
Suppose we want to nd a function n : 1
2
1 such that n 0. If
n(r, j) = A(r) Y (j), then
n =
2
x
(A Y ) +
2
y
(A Y ) =
_
2
x
A
_
Y + A
_
2
y
Y
_
= A
tt
Y +AY
tt
,
where we denote A
tt
=
2
x
A and Y
tt
=
2
y
Y . Thus,
n(r, j) = A
tt
(r) Y (j) + A(r) Y
tt
(j)
=
_
A
tt
(r) Y (j) + A(r) Y
tt
(j)
_
A(r)Y (j)
A(r)Y (j)
=
_
A
tt
(r)
A(r)
+
Y
tt
(j)
Y (j)
_
n(r, j).
Thus, dividing by n(r, j), Laplaces equation is equivalent to:
0 =
n(r, j)
n(r, j)
=
A
tt
(r)
A(r)
+
Y
tt
(j)
Y (j)
.
This is a sum of two functions which depend on dierent variables. The only
way the sum can be identically zero is if each of the component functions is
constant:
A
tt
A
,
Y
tt
Y

So, pick some separation constant 1, and then solve the two ordinary
dierential equations:
A
tt
(r) = A(r) and Y
tt
(j) = Y (j) (16A.2)
The (real-valued) solutions to (16A.2) depends on the sign of . Let j =
_
[[.
Then the solutions of (16A.2) have the form:
A(r) =
_
_
_
sinh(jr) +1cosh(jr) if 0;
r +1 if = 0;
sin(jr) +1cos(jr) if < 0;
where and 1 are arbitrary constants. Assuming < 0, and j =
_
[[, we
get:
A(r) = sin(jr)+1cos(jr) and Y (j) = C sinh(jr)+1cosh(jr).
16B. ...in Cartesian coordinates on 1
D
355
This yields the following separated solution to Laplaces equation:
n(r, j) = A(r)Y (j) =
_
sin(jr) +1cos(jr)
_
_
C sinh(jr) +1cosh(jr)
_
(16A.3)
Alternately, we could consider the general complex solution to (16A.2), given
by:
A(r) = exp
_
r
_
,
where

C is some complex number. For example, if < 0 and j =
_
[[,
then

= ji are imaginary, and
A
1
(r) = exp(ijr) = cos(jr) +i sin(jr)
and A
2
(r) = exp(ijr) = cos(jr) i sin(jr)
are two linearly independent solutions to (16A.2). The general solution is then
given by:
A(r) = o A
1
(r) +/ A
2
(r) = (o +/) cos(jr) + i (o /) sin(jr).
Meanwhile, the general form for Y (j) is
Y (j) = c exp(jj) +d exp(jj) = (c +d) cosh(jj) + (c d) sinh(jj).
The corresponding separated solution to Laplaces equation is:
n(r, j) = A(r)Y (j) =
_
sin(jr) +1i cos(jr)
_
_
C sinh(jr) +1cosh(jr)
_
,
(16A.4)
where = (o +/), 1 = (o /), C = (c +d), and 1 = (c d). In this case,
we just recover solution (16A.3). However, we could also construct separated
solutions where C is an arbitrary complex number, and

is one of its
square roots.
16B Separation of variables in Cartesian coordinates
on 1
1
Recommended: 16A.
Given some PDE for a function n : 1
D
1, we apply the strategy of
separation of variables as follows:
1. Hypothesize that n can be written as a product of 1 functions, each de-
pending on only one coordinate; in other words, assume that
n(r
1
, . . . , r
D
) = n
1
(r
1
) n
2
(r
2
) . . . n
D
(r
D
) (16B.5)
2. When we evaluate the PDE on a function of type (16B.5), we may nd
that the PDE decomposes into 1 separate, ordinary dierential equations
for each of the 1 functions n
1
, . . . , n
D
. Thus, we can solve these ODEs
independently, and combine the resulting solutions to get a solution for n.
Example 16B.1. Laplaces Equation in 1
D
:
Suppose we want to nd a function n : 1
D
1 such that n 0. As in
the two-dimensional case (Example 16A.2), we reason:
If n(x) = A
1
(r
1
)A
2
(r
2
) . . . A
D
(r
D
), then n =
_
A
tt
1
A
1
+
A
tt
2
A
2
+. . . +
A
tt
D
A
D
_
n.
Thus, Laplaces equation is equivalent to:
0 =
n
n
(x) =
A
tt
1
A
1
(r
1
) +
A
tt
2
A
2
(r
2
) + . . . +
A
tt
D
r
D
(r
D
).
This is a sum of 1 distinct functions, each of which depends on a dierent vari-
able. The only way the sum can be identically zero is if each of the component
functions is constant:
A
tt
1
A
1

1
,
A
tt
2
A
2

2
, . . . ,
A
tt
D
A
D

D
, (16B.6)
such that
1
+
2
+. . . +
D
= 0. (16B.7)
So, pick some separation constant = (
1
,
2
, . . . ,
D
) 1
D
satisfying
(16B.7), and then solve the ODEs:
A
tt
d
=
d
A
d
for d=1,2,. . . ,D (16B.8)
The (real-valued) solutions to (16B.8) depends on the sign of (and clearly,
if (16B.7) is going to be true, either all
d
are zero, or some are negative and
some are positive). Let j =
_
[[. Then the solutions of (16B.8) have the
form:
A(r) =
_
_
_
exp(jr) +1exp(jr) if 0;
r +1 if = 0;
sin(jr) +1cos(jr) if < 0;
where and 1 are arbitrary constants. We then combine these as in Example
16A.2.
16C. ...in polar coordinates: Bessels equation 357
16C Separation in polar coordinates: Bessels equa-
tion
Prerequisites: 0D(ii), 1C. Recommended: 14C, 16A.
In 14C-14F, we explained how to use solutions of Bessels equation to solve
the heat equation or wave equation in polar coordinates. In this section, we will
see how Bessel derived his equation in the rst place: it arises naturally when
one uses separation of variables to nd eigenfunctions of the Laplacian in polar
coordinates. First, a technical lemma from the theory of ordinary dierential
equations:
Lemma 16C.1. Let : [, ] 1 be a function satisfying periodic
boundary conditions [i.e. () = () and
t
() =
t
()]. Let j 0 be
some constant, and suppose satises the linear ordinary dierential equation:
tt
() = j (), for all [, ]. (16C.9)
Then j = :
2
for some : N, and must be a function of the form:
() = cos(:) +1sin(:), (for constants , 1 C.)
Proof. Eqn.(16C.9) is a second-order linear ODE, so the set of all solutions to
eqn.(16C.9) is a two-dimensional vector space. This vector space is spanned
by functions of the form () = c
r
, where : is any root of the characteristic
polynomial j(r) = r
2
+ j. The two roots of this polynomial are of course
: =
ji. Let : =

j (it will turn out that : is an integer, although we
dont know this yet). Hence the general solution to (16C.9) is
() = C
1
c
mi
+ C
2
c
mi
,
where C
1
and C
2
are any two constants. The periodic boundary conditions
mean that
() = () and
t
() =
t
(),
which means
C
1
c
mi
+ C
2
c
mi
= C
1
c
mi
+ C
2
c
mi
, (16C.10)
and :iC
1
c
mi
:iC
2
c
mi
= :iC
1
c
mi
:iC
2
c
mi
. (16C.11)
If we divide both sides of the eqn.(16C.11) by :i, we get
C
1
c
mi
C
2
c
mi
= C
1
c
mi
C
2
c
mi
.
If we add this to eqn.(16C.10), we get
2C
1
c
mi
= 2C
1
c
mi
,
which is equivalent to c
2mi
= 1. Hence, : must be some integer, and
j = :
2
.
Now, let := C
1
+ C
2
and 1
t
:= C
1
C
2
. Then C
1
=
1
2
( + 1
t
) and
C
2
=
1
2
(1
t
). Thus,
() = C
1
c
mi
+ C
2
c
mi
= (+1
t
)c
mi
+ (1
t
)c
mi
=

2
_
c
mi
+c
mi
_
+
1
t
i
2i
_
c
mi
c
mi
_
= cos(:) +1
t
i sin(:)
because of the Euler formulas: cos(r) =
1
2
(c
ix
+ c
ix
) and sin(r) =
1
2i
(c
ix
c
ix
).
Now let 1 = 1
t
i; then () = cos(:) +1sin(:), as desired. 2
Proposition 16C.2. Let ) : 1
2
1 be an eigenfunction of the Laplacian
[i.e. ) =
2
) for some constant 1]. Suppose ) separates in polar
coordinates, meaning that there is a function : [, ] 1 (satisfying
periodic boundary conditions) and a function 1 : 1
,
1 such that
)(:, ) = 1(:) (), for all : 0 and [, ].
Then there is some : N such that
() = cos(:) +1sin(:), (for constants , 1 1.)
and 1 is a solution to the (:th order) Bessel Equation:
:
2
1
tt
(:) +: 1
t
(:) + (
2
:
2
:
2
) 1(:) = 0, for all : 0. (16C.12)
Proof. Recall that, in polar coordinates, ) =
2
r
) +
1
:
r
) +
1
:
2
).
Thus, if )(:, ) = 1(:) (), then the eigenvector equation ) =
2
)
becomes
2
1(:) () = 1(:) ()
=
2
r
1(:) () +
1
:
r
1(:) () +
1
:
2
1(:) ()
= 1
tt
(:)() +
1
:
1
t
(:)() +
1
:
2
1(:)
tt
(),
16D. ...in spherical coordinates: Legendres equation 359
which is equivalent to
2
=
1
tt
(:)() +
1
r
1
t
(:)() +
1
r
2
1(:)
tt
()
1(:) ()
=
1
tt
(:)
1(:)
+
1
t
(:)
:1(:)
+

tt
()
:
2
()
, (16C.13)
If we multiply both sides of (16C.13) by :
2
and isolate the
tt
term, we get:
2
:
2
:
2
1
tt
(:)
1(:)
+
:1
t
(:)
1(:)
=

tt
()
()
. (16C.14)
Abstractly, equation (16C.14) has the form: 1(:) = G(), where 1 is a func-
tion depending only on : and G is a function depending only on . The only
way this can be true is if there is some constant j 1 such that 1(:) = j
for all : 0 and G() = j for all [, ). In other words,
tt
()
()
= j, for all [, ), (16C.15)
and
2
:
2
+
:
2
1
tt
(:)
1(:)
+
:1
t
(:)
1(:)
= j, for all : 0. (16C.16)
Multiply both sides of equation (16C.15) by () to get:
tt
() = j (), for all [, ). (16C.17)
Multiply both sides of equation (16C.16) by 1(:) to get:
:
2
1
tt
(:) +: 1
t
(:) +
2
:
2
1(:) = j1(:), for all : 0. (16C.18)
Apply Lemma 16C.1 to to eqn.(16C.17) to deduce that j = :
2
for some
: N, and that () = cos(:) + 1sin(:). Substitute j = :
2
into
eqn.(16C.18) to get
:
2
1
tt
(:) +: 1
t
(:) +
2
:
2
1(:) = :
2
1(:),
Now subtract :
2
1(:) from both sides to get Bessels equation (16C.12). 2
16D Separation in spherical coordinates: Legendres
equation
Prerequisites: 0D(iv), 1C, 5C(i), 6F, 0H(iii). Recommended: 16C.
z
0
x
y
c
o
s
(
)
z
x
y
(A) (B)
Figure 16D.1: (A) Spherical coordinates. (B) Zonal functions.
Recall that spherical coordinates (:, , ) on 1
3
are dened by the transfor-
mation:
r = :sin()cos(), j = :sin()sin() and . = :cos().
where : 1
,
, [, ), and [0, ]. The reverse transformation is dened:
: =
_
r
2
+j
2
+.
2
, = arctan
_
j
r
_
and = arctan
_
_
r
2
+j
2
.
_
.
[See Figure 16D.1(A)]. Geometrically, : is the radial distance from the origin. If
we x : = 1 , then we get a sphere of radius 1. On the surface of this sphere,
is longitude and is latitude. In terms of these coordinates, the Laplacian is
written:
)(:, , ) =
2
r
) +
2
:
r
) +
1
:
2
sin()
) +
cot()
:
2

) +
1
:
2
sin()
2
).
(Exercise 16D.1) E _
A function ) : 1
3
1 is called zonal if )(:, , ) depends only on on :
and in other words, )(:, , ) = 1(:, ), where 1 : 1
,
[0, ] 1 is some
other function. If we restrict ) to the aforementioned sphere of radius 1, then )
is invariant under rotations around the north-south axis of the sphere. Thus,
) is constant along lines of equal latitude around the sphere, so it divides the
sphere into zones from north to south [Figure 16D.1(B)].
Proposition 16D.1. Let ) : 1
3
1 be zonal. Suppose ) is a harmonic
function (i.e. ) = 0). Suppose ) separates in spherical coordinates, meaning
that there are (bounded) functions : [0, ] 1 and 1 : 1
,
1 such that
)(:, , ) = 1(:) (), for all : 0, [0, ], and [, ].
Then there is some j 1 such that () = L[cos()], where L : [1, 1] 1
is a (bounded) solution of the Legendre Equation:
(1 r
2
)L
tt
(r) 2rL
t
(r) +jL(r) = 0, (16D.19)
and 1 is a (bounded) solution to the Cauchy-Euler Equation:
:
2
1
tt
(:) + 2: 1
t
(:) j 1(:) = 0, for all : 0. (16D.20)
Proof. By hypothesis
0 = )(:, , )
=
2
r
) +
2
:
r
) +
1
:
2
sin()
) +
cot()
:
2

) +
1
:
2
sin()
2
)
()
1
tt
(:) () +
2
:
1
t
(:) ()
+
1
:
2
sin()
1(:)
tt
() +
cot()
:
2
1(:)
t
() + 0.
[where () is because )(:, , ) = 1(:) ().] Hence, multiplying both sides
by
:
2
1(:) ()
, we obtain
0 =
:
2
1
tt
(:)
1(:)
+
2:1
t
(:)
1(:)
+
1
sin()
tt
()
()
+
cot()
t
()
()
,
Or, equivalently,
:
2
1
tt
(:)
1(:)
+
2:1
t
(:)
1(:)
=
1
sin()
tt
()
()

cot()
t
()
()
. (16D.21)
Now, the left hand side of (16D.21) depends only on the variable :, whereas the
right hand side depends only on . The only way that these two expressions
can be equal for all values of : and is if both expressions are constants. In
other words, there is some constant j 1 (called a separation constant) such
that
:
2
1
tt
(:)
1(:)
+
2:1
t
(:)
1(:)
= j, for all : 0,
and
1
sin()
tt
()
()
+
cot()
t
()
()
= j, for all [0, ].
Or, equivalently,
:
2
1
tt
(:) + 2:1
t
(:) = j1(:), for all : 0, (16D.22)
and

tt
()
sin()
+ cot()
t
() = j(), for all [0, ]. (16D.23)
If we make the change of variables r = cos() (so that = arccos(r), where
r [1, 1]), then () = L(cos()) = L(r), where L is some other (unknown)
function.
Claim 1: The function satises the ODE (16D.23) if and only if L satises
the Legendre equation (16D.19).
Proof. Exercise 16D.2 (Hint: This is a straightforward application of the E _
Chain Rule.)
Claim 1
Finally, observe that the ODE (16D.22) is equivalent to the Cauchy-Euler
equation (16D.20). 2
For all n N, we dene the nth Legendre Polynomial by
T
n
(r) :=
1
n! 2
n
n
x
_
(r
2
1)
_
n
. (16D.24)
For example:
T
0
(r) = 1 T
3
(r) =
1
2
(5r
3
3r)
T
1
(r) = r T
4
(r) =
1
8
(35r
4
30r
2
+ 3)
T
2
(r) =
1
2
(3r
2
1) T
5
(r) =
1
8
(63r
5
70r
3
+ 15r).
(see Figure 16D.2(A))
Lemma 16D.2. Let n N. Then the Legendre Polynomial T
n
is a solution
to the Legendre Equation (16D.19) with j = n(n + 1).
Proof. Exercise 16D.3 (Direct computation) 2 E _
Is T
n
the only solution to the Legendre Equation (16D.19)? No, because the
Legendre Equation is an order-two linear ODE, so the set of solutions forms
a two-dimensional vector space 1. The scalar multiples of T
n
form a one-
dimensional subspace of 1. However, to be physically meaningful, we need the
solutions to be bounded at r = 1. So instead we ask: is T
n
the only bounded
solution to the Legendre Equation (16D.19)? Also, what happens if j ,= n(n+1)
for any n N?
363
0.5
-0.5
1
0
-1
x
1 0.5 -0.5 0 -1
1
0.8
0.6
-0.2
0.4
0.2
0
-0.4
x
1 0.5 0 -0.5 -1
T
1
(r) T
2
(r)
0
-0.5
-1
1
0.5
x
1 -1 0.5 -0.5 0
1
0.6
-0.2
0.8
0.4
0
0.2
-0.4
x
1 0.5 0 -1 -0.5
T
3
(r) T
4
(r)
1
0
0.5
-0.5
x
1 0.5 0 -1 -0.5
-1
1
0.6
-0.2
0.8
0.4
0
0.2
-0.4
x
0.5 1 0 -1 -0.5
T
5
(r) T
6
(r)
Above. The Legendre polynomials 11(x) to 16(x),
plotted for x [1, 1].
Right. Substitution of the power series
n=0
anx
n
into
the Legendre Equation (16D.19), in the proof of Claim
1 of Lemma 16D.3.
I
f
/
(
x
)
=
a
0
+
a
1
x
+
a
2
x
2
+
+
a
k
x
k
+
t
h
e
n
/
(
x
)
=
a
0
+
a
1
x
+
a
2
x
2
+
a
k
x
k
+
a
n
d
2
x
/
/
(
x
)
=
2
a
1
x
+
4
a
2
x
2
+
2
k
a
k
x
k
+
a
n
d
/
/
/
(
x
)
=
2
a
2
+
6
a
3
x
+
1
2
a
4
x
2
+
+
(
k
+
2
)
(
k
+
1
)
a
k
+
2
x
k
+
a
n
d
x
2
/
/
/
(
x
)
=
2
a
2
x
2
+
(
k
1
)
k
a
k
x
k
+
T
h
u
s
,
0
=
(
1
x
2
)
/
/
/
(
x
)
2
x
/
/
(
x
)
+
/
(
x
)
=
(
a
0
2
a
2
)
+
(
(
2
)
a
1
+
6
a
3
)
x
+
(
(
6
)
a
2
+
1
2
a
4
)
x
2
+
+
b
k
x
k
+
w
h
e
r
e
b
k
:
=
(
k
+
2
)
(
k
+
1
)
a
k
+
2
+
_
k
(
k
+
1
)
_
a
k
f
o
r
a
l
l
k
N
.
Figure 16D.2:
Lemma 16D.3.
(a) If j = n(n+1) for some n N, then (up to multiplication by a scalar), the
Legendre polynomial T
n
(r) is the unique solution to the Legendre Equation
(16D.19) which is bounded on [1, 1].
(b) If j ,= n(n +1) for any n N, then all solutions to the Legendre Equation
(16D.19) are innite power series which diverge at r = 1 (and thus, are
unsuitable for Proposition 16D.1).
Proof. We apply the Power series method (see 0H(iii) on page 571). Sup-
pose L(r) =
n=0
o
n
r
n
is some analytic function dened on [1, 1] (where the
coecients o
n
n=1
are as yet unknown).
Claim 1: L(r) satises the Legendre Equation (16D.19) if and only if the
coecients o
0
, o
1
, o
2
, . . . satisfy the recurrence relation
o
k+2
=
/(/ + 1) j
(/ + 2)(/ + 1)
o
k
, for all / N. (16D.25)
In particular, o
2
=
j
2
o
0
and o
3
=
2 j
6
o
1
.
Proof. We will substitute the power series
n=0
o
n
r
n
into the Legendre
Equation (16D.19). The details of the computation are shown on the right
side of Figure 16D.2. The computation yields the equation 0 =
k=0
/
k
r
k
,
where /
k
:= (/ + 2)(/ + 1)o
k+2
+
_
j /(/ + 1)
_
o
k
for all / N. It follows
that /
k
= 0 for all / N; in other words, that
(/ + 2)(/ + 1)o
k+2
+
_
j /(/ + 1)
_
o
k
= 0, for all / N.
Rearranging this equation produces the desired recurrence relation (16D.25).
Claim 1
The space of all solutions to the Legendre Equation (16D.19) is a two-dimensional
vector space, because the Legendre equation is a linear dierential equation
of order 2. We will now nd a basis for this space. Recall that L is even
if L(r) = L(r) for all r [1, 1], and L is odd if L(r) = L(r) for all
r [1, 1].
Claim 2: There is a unique even analytic function c(r) and a unique odd
analytic function O(r) which satisfy the Legendre Equation (16D.19), such
that c(1) = 1 = O(1), and such that any other solution L(r) can be written
as a linear combination L(r) = o c(r) +/ O(r), for some constants o, / 1.
Proof. Claim 1 implies that the power series L(r) =
n=0
o
n
r
n
is entirely
determined by the coecients o
0
and o
1
. To be precise, L(r) = c(r) +
O(r), where c(r) =
n=0
o
2n
r
2n
and O(r) =
n=0
o
2n+1
r
2n+1
both satisfy
the recurrence relation (16D.25), and thus, are solutions to the Legendre
Equation (16D.19).
Claim 2
Claim 3: Suppose j = n(n + 1) for some n N. Then the Legendre
equation (16D.19) has a degree-n polynomial as one of its solutions. To be
precise:
(a) If n is even, then o
k
= 0 for all even / n. Hence, c(r) is a degree-n
polynomial.
(b) If n is odd, then o
k
= 0 for all odd / n. Hence, O(r) is a degree-n
polynomial.
Claim 3
E _
Thus, there is a one-dimensional space of polynomial solutions to the Legendre
equation namely all scalar multiples of c(r) (if n is even) or O(r) (if n is
odd).
Claim 4: If j ,= n(n + 1) for any n N, the series c(r) and O(r) both
diverge at r = 1.
Proof. Exercise 16D.5 (a) First note that an innite number of coecients E _
o
n
n=0
are nonzero.
(b) Show that lim
n
[o
n
[ = 1.
(c) Conclude that the series c(r) and O(r) diverge when r = 1.
Claim 4
So, there exist solutions to the Legendre equation (16D.19) that are bounded
on [1, 1] if and only if j = n(n + 1) for some n N, and in this case, the
bounded solutions are all scalar multiples of a polynomial of degree n [either
c(r) or O(r)]. But Lemma 16D.2 says that the Legendre polynomial T
n
(r) is
a solution to the Legendre equation (16D.19). Thus, (up to multiplication by
a constant), T
n
(r) must be equal to c(r) (if n is even) or O(r) (if n is odd).
2
Remark: Sometimes the Legendre polynomials are dened as the (unique)
polynomial solutions to Legendres equation; the denition we have given in
eqn.(16D.24) is then derived from this denition, and is called Rodrigues Formula.
Lemma 16D.4. Let 1 : 1
,
1 be a solution to the Cauchy-Euler equation
:
2
1
tt
(:) + 2: 1
t
(:) n(n + 1) 1(:) = 0, for all : 0. (16D.26)
Then 1(:) = :
n
+
B
r
n+1
for some constants and 1.
If 1 is bounded at zero, then 1 = 0, so 1(:) = :
n
.
Proof. Check that )(:) = :
n
and p(:) = :
n1
are solutions to eqn.(16D.26).
But (16D.26) is a second-order linear ODE, so the solutions form a 2-dimensional
vector space. Since ) and p are linearly independent, they span this vector
space. 2
Corollary 16D.5. Let ) : 1
3
1 be a zonal harmonic function that
separates in spherical coordinates (as in Proposition 16D.1). Then there is some
: N such that )(:, , ) = C:
n
T
n
[cos()], where T
n
is the nth Legendre
Polynomial, and C 1 is some constant. (See Figure 16D.3.)
Proof. Combine Proposition 16D.1 with Lemmas 16D.3 and 16D.4 2
Thus, the Legendre polynomials are important when solving the Laplace
equation on spherical domains. We now describe some of their important prop-
erties
Proposition 16D.6. Legendre polynomials satisfy the following recurrence
relations:
(a) (2n + 1)T
n
(r) = T
t
n+1
(r) T
t
n1
(r).
(b) (2n + 1)rT
n
(r) = (n + 1)T
n+1
(r) +nT
t
n1
(r).
Proposition 16D.7. The Legendre polynomials form an orthogonal set for
L
2
[1, 1]. That is:
(a) For any n ,= :, T
n
, T
m
) =
1
2
_
1
1
T
n
(r)T
m
(r) dr = 0.
(b) For any n N, |T
n
|
2
2
=
1
2
_
1
1
T
2
n
(r)dr =
1
2n + 1
.
Proof. (a) Exercise 16D.7 (Hint: Start with the Rodrigues formula (16D.24). E _
Apply integration by parts n times.)
(b) Exercise 16D.8 (Hint: Use Proposition 16D.6(b).) 2 E _
367
6
2
-6
4
6 0 2 -4
-2
4 -2
-4
0
-6
-6
6 -4
4
0
-2
4
-4
2
2
-6 -2
0
6
:T
1
(cos()) :
2
T
2
(cos())
6
2
-6
4
6
-2
2 -4
-4
4 -2
0
-6 0
-6
-2
4
2
-4
6
-2
0 2
0
-4 6 4 -6
:
3
T
3
(cos()) :
4
T
4
(cos())
6
2
-6
4
6 2 -4
-2
4
0
-6 -2 0
-4
6
2
-6
4
6 0 2 -4 4 -2
0
-2
-6
-4
:
5
T
5
(cos()) :
6
T
6
(cos())
Figure 16D.3: Planar cross-sections of the zonal harmonic functions r11(cos()) to
r
6
16(cos()), plotted for r [0...6]; see Corollary 16D.5 on the preceding page. Remember
that these are functions in 1
3
. To visualize these functions in three dimensions, take the above
contour plots and mentally rotate them around the vertical axis.
Because of Proposition 16D.7, we can try to represent an arbitrary function
) L
2
[1, 1] in terms of Legendre polynomials, to obtain a Legendre Series:
)(r)

L2
n=0
o
n
T
n
(r), (16D.27)
where o
n
:=
), T
n
)
|T
n
|
2
2
=
2n + 1
2
_
1
1
)(r)T
n
(r) dr is the nth Legendre coef-
cient of ).
Theorem 16D.8. The Legendre polynomials form an orthogonal basis for
L
2
[1, 1]. Thus, if ) L
2
[1, 1], then the Legendre series (16D.27) converges
to ) in 1
2
.
Proof. See [Bro89, Thm 3.2.4, p.50]. 2
Let B = (:, , ) ; : 1, [, ], [0, ] be the unit ball in spherical
coordinates. Thus, B = (1, , ) ; [, ], [0, ] is the unit sphere.
Recall that a zonal function on B is a function which depends only on the
latitude coordinate , and not on the longitude coordinate .
Theorem 16D.9. Dirichlet problem on a ball
Let ) : B 1 be some function describing a heat distribution on the
surface of the ball. Suppose ) is zonal i.e. )(1, , ) = 1 (cos()), where
1 L
2
[1, 1], and 1 has Legendre series
1(r)

L2
n=0
o
n
T
n
(r).
Dene n : B 1 by n(:, , ) =
n=0
o
n
:
n
T
n
(cos()). Then n is the uniqe solu-
tion to the Laplace equation, satisfying the nonhomogeneous Dirichlet boundary
conditions
n(1, , )

L2
)(, ), for all (1, , ) B.
16E. Separated vs. quasiseparated 369
16E Separated vs. quasiseparated
Prerequisites: 16B.
If we use complex-valued functions like (16A.4) as the components of the
separated solution (16B.5) on page 355, then we will still get mathematically
valid solutions to Laplaces equation (as long as (16B.7) is true). However,
these solutions are not physically meaningful what does a complex-valued heat
distribution feel like? This is not a problem, because we can extract real-valued
solutions from the complex solution as follows.
Proposition 16E.1. Suppose L is a linear dierential operator with real-
valued coecients, and p : 1
D
1, and consider the nonhomogeneous PDE
L n = p.
If n : 1
D
C is a (complex-valued) solution to this PDE, and we dene
n
R
(x) = Re [n(x)] and n
I
(x) = Im[n(x)], then L n
R
= p and L n
I
= 0.
In this case, the solutions n
R
and n
I
are not themselves generally going to
be in separated form. Since they arise as the real and imaginary components of
a complex separated solution, we call n
R
and n
I
quasiseparated solutions.
Example 16E.2. Recall the separated solutions to the two-dimensional
Laplace equation from Example 16A.2 on page 354. Here, L = and p 0,
and, for any xed j 1, the function
n(r, j) = A(r) Y (j) = exp(jj) exp(jij)
is a complex solution to Laplaces equation. Thus,
n
R
(r, j) = exp(jr) cos(jj) and n
I
(r, j) = exp(jr) sin(jj)
are real-valued solutions of the form obtained earlier.
16F The polynomial formalism
Prerequisites: 16B, 4B.
Separation of variables seems like a bit of a miracle. Just how generally
applicable is it? To answer this, it is convenient to adopt a polynomial for-
malism for dierential operators. If L is a dierential operator with constant
1
1
This is important.
coecients, we will formally represent L as a polynomial in the variables
1
,
2
, . . . ,
D
. For example, we can write the Laplacian:
=
2
1
+
2
2
+. . . +
2
D
= T(
1
,
2
, . . . ,
D
),
where T(r
1
, r
2
, . . . , r
D
) = r
2
1
+r
2
2
+. . . +r
2
D
.
In another example, the general second-order linear PDE
2
x
n + 1
x
y
n + C
2
y
n + 1
x
n + 1
y
n + 1n = G
(where , 1, C, . . . , 1 are constants) can be rewritten:
T(
x
,
y
)n = p
where T(r, j) = r
2
+1rj +Cj
2
+1r +1j +1.
The polynomial T is called the polynomial symbol of L, and provides a
convenient method for generating separated solutions
Proposition 16F.1. Suppose that L is a linear dierential operator on 1
D
with polynomial symbol T. Regard T : C
D
C as a function.
Fix z = (.
1
, . . . , .
D
) C
D
, and dene n
z
: 1
D
1 by
n
z
(r
1
, . . . , r
D
) = exp(.
1
r
1
) exp(.
2
r
2
) . . . exp(.
D
r
D
) = exp(z x).
Then Ln
z
(x) = T(z) n
z
(x) for all x 1
D
.
In particular, if z is a root of T (that is, T(.
1
, . . . , .
D
) = 0), then Ln = 0.
Proof. Exercise 16F.1 Hint: First, use formula (0C.1) on page 551 to show that E _
d
n
z
= .
d
n
z
, and, more generally,
n
d
n
z
= .
n
d
n
z
. 2
Thus, many
2
separated solutions of the dierential equation Ln = 0 are
dened by the the complex-valued solutions of the algebraic equation T(z) = 0.
Example 16F.2. Consider again the two-dimensional Laplace equation
2
x
n +
2
y
n = 0
The corresponding polynomial is T(r, j) = r
2
+ j
2
. Thus, if .
1
, .
2
C are
any complex numbers such that .
2
1
+.
2
2
= 0, then
n(r, j) = exp(.
1
r +.
2
j) = exp(.
1
r) exp(.
2
j)
is a solution to Laplaces equation. In particular, if .
1
= 1, then we must have
.
2
= i. Say we pick .
2
= i; then the solution becomes
n(r, j) = exp(r) exp(ij) = c
x
_
cos(j) +i sin(j)
_
.
2
But not all.
16F. The polynomial formalism 371
More generally, if we choose .
1
= j 1 to be a real number, then we must
choose .
2
= ji to be purely imaginary, and the solution becomes
n(r, j) = exp(jr) exp(jij) = c
x
_
cos(jj) +i sin(jj)
_
.
Compare this with the separated solutions obtained from Example 16A.2 on
page 354.
Example 16F.3. Consider the one-dimensional telegraph equation:
2
t
n + 2
t
n +n = n. (16F.28)
We can rewrite this as
2
t
n + 2
t
n +n
2
x
n = 0,
which is equivalent to L n = 0, where L is the linear dierential operator
L =
2
t
+ 2
t
+n
2
x
,
with polynomial symbol
T(r, t) = t
2
+ 2t + 1 r
2
= (t + 1 +r)(t + 1 r).
Thus, the equation T(, ) = 0 has solutions:
= ( + 1)
So, if we dene n(r, t) = exp( r) exp( t), then n is a separated solution to
equation (16F.28). (Exercise 16F.2 Check this.). In particular, suppose we E _
choose = 1. Then the separated solution is n(r, t) = exp((t r) r).
If =
R
+
I
i is a complex number, then the quasiseparated solutions are:
n
R
= exp (
R
(r +t) r) cos (
I
(r +t))
n
I
= exp (
R
(r +t) r) sin (
I
(r +t)) .
Remark 16F.4: The polynomial formalism provides part of the motivation
for the classication of PDEs as elliptic, hyperbolic
3
, etc. Notice that, if L is an
elliptic dierential operator on 1
2
, then the real-valued solutions to T(.
1
, .
2
) = 0
(if any) form an ellipse in 1
2
. In 1
D
, the solutions form an ellipsoid.
Similarly, if we consider the parabolic PDE
t
n = Ln, the the correspond-
ing dierential operator L
t
has polynomial symbol Q(x; t) = T(x) t. The
real-valued solutions to Q(x; t) = 0 form a paraboloid in 1
D
1. For example,
the 1-dimensional heat equation
2
x
n
t
n = 0 yields the classic equation
t = r
2
for a parabola in the (r, t)-plane. Similarly, with a hyperbolic PDE,
the dierential operator L
2
t
has polynomial symbol Q(x; t) = T(x) t
2
, and
the roots form a hyperboloid.
3
See 5E on page 95.
16G Constraints
Prerequisites: 16F.
Normally, we are not interested in just any solution to a PDE; we want a
solution which satises certain constraints. The most common constraints are:
Boundary Conditions: If the PDE is dened on some bounded domain
X 1
D
, then we may want the solution function n (or its derivatives) to
have certain values on the boundary of this domain.
Boundedness: If the domain X is unbounded (e.g. X = 1
D
), then we
may want the solution n to be bounded; in other words, we want some nite
` 0 such that [n(x)[ < ` for all values of some coordinate r
d
.
16G(i) Boundedness
The solution obtained through Proposition 16F.1 is not generally going to be
bounded, because the exponential function )(r) = exp(r) is not bounded as a
function of r, unless is a purely imaginary number. More generally:
Proposition 16G.1. Fix z = (.
1
, . . . , .
D
) C
D
, and suppose n
z
: 1
D
1
is dened as in Proposition 16F.1:
n
z
(r
1
, . . . , r
D
) = exp(.
1
r
1
) exp(.
2
r
2
) . . . exp(.
D
r
D
) = exp(z x).
Then:
1. n(x) is bounded for all values of the variable r
d
1 if and only if .
d
= i
for some 1.
2. n(x) is bounded for all r
d
0 if and only if .
d
= +i for some 0.
3. n(x) is bounded for all r
d
< 0 if and only if .
d
= +i for some 0.
Proof. Exercise 16G.1 2 E _
Example 16G.2. Recall the one-dimensional telegraph equation of Example
16F.3:
2
t
n + 2
t
n +n = n
We constructed a separated solution of the form: n(r, t) = exp(r + t),
where = ( + 1). This solution will be bounded in time if and only if
is a purely imaginary number; i.e. =
I
i. Then = (
I
i + 1), so that
16G. Constraints 373
n(r, t) = exp(r) exp
_
I
(t r) i
_
; thus, the quasiseparated solutions
are:
n
R
= exp (r)cos
_
I
(t r)
_
and n
I
= exp (r)sin
_
I
(t r)
_
.
Unfortunately, this solution is unbounded in space, which is probably not what
we want. An alternative is to set =
I
i 1, and then set = + 1 =
I
i.
Then the solution becomes n(r, t) = exp(
I
i(r +t) t) = c
t
exp(
I
i(r +t)),
and the quasiseparated solutions are:
n
R
= c
t
cos (
I
(r +t)) and n
I
= c
t
sin (
I
(r +t)) .
These solutions are exponentially decaying as t , and thus, bounded in
forward time. For any xed time t, they are also bounded (and actually
periodic) functions of the space variable r.
16G(ii) Boundary conditions
Prerequisites: 5C.
There is no cureall like Proposition 16G.1 for satisfying boundary conditions,
since generally they are dierent in each problem. Generally, a single separated
solution (say, from Proposition 16F.3) will not be able to satisfy the conditions;
we must sum together several solutions, so that they cancel out in suitable
ways along the boundaries. For these purposes, the following Euler identities
are often useful:
sin(r) =
c
xi
c
xi
2i
; cos(r) =
c
xi
+c
xi
2i
;
sinh(r) =
c
x
c
x
2
; cosh(r) =
c
x
+c
x
2
.
which we can utilize along with the following boundary information:
cos
t
(n) = sin(n) = 0, for all n Z;
sin
t
__
n +
1
2
_
_
= cos
__
n +
1
2
_
_
= 0, for all n Z;
cosh
t
(0) = sinh(0) = 0.
For rectangular domains, the boundaries are obtained by xing a partic-
ular coordinate at a particular value; i.e. they are each of the form form
_
x 1
D
; r
d
= 1
_
for some constant 1 and some dimension d. The conve-
nient thing about a separated solution is that it is a product of 1 functions, and
only one of them is involved in satisfying this boundary condition.
For example, recall Example 16F.2 on page 370, which gave the separated so-
lution n(r, j) = c
x
_
cos(jj) +i sin(jj)
_
for the two-dimensional Laplace
equation, where j 1. Suppose we want the solution to satisfy homogeneous
Dirichlet boundary conditions:
n(r, j) = 0 if r = 0, or j = 0, or j = .
To satisfy these three conditions, we proceed as follows:
First, let n
1
(r, j) = c
x
_
cos(jj) +i sin(jj)
_
,
and n
2
(r, j) = c
x
_
cos(jj) +i sin(jj)
_
= c
x
_
cos(jj) i sin(jj)
_
.
If we dene (r, j) = n
1
(r, j) n
2
(r, j), then
(r, j) = 2c
x
i sin(jj).
At this point, (r, j) already satises the boundary conditions for j = 0 and
j = . To satisfy the remaining condition:
Let
1
(r, j) = 2c
x
i sin(jj),
and
1
(r, j) = 2c
x
i sin(jj).
If we dene n(r, j) =
1
(r, j)
2
(r, j), then
n(r, j) = 4 sinh(jr) i sin(jj)
also satises the boundary condition at r = 0.
375
Chapter 17
Impulse-response methods
Nature laughs at the diculties of integration. Pierre-Simon Laplace
17A Introduction
A fundamental concept in science is causality: an initial event (an impulse)
at some location y causes a later event (a response) at another location x (Figure
17A.1A). In an evolving, spatially distributed system (e.g. a temperature dis-
tribution, a rippling pond, etc.), the system state at each location results from
a combination of the responses to the impulses from all other locations (as in
Figure 17A.1B).
If the system is described by a linear PDE, then we expect some sort of
superposition principle to apply (Theorem 4C.3 on page 65). Hence, we can
replace the word combination with sum, and say:
The state of the system at x is a sum of the responses to the
impulses from all other locations.
(17A.1)
(See Figure 17A.1B). However, there are an innite number indeed, a contin-
uum of other locations, so we are summing over a continuum of responses.
But a sum over a continuum is just an integral. Hence, statement (17A.1)
becomes:
In a linear PDE, the solution at x is an integral of the re-
sponses to the impulses from all other locations.
(17A.2)
The relation between impulse and response (i.e. between cause and eect) is
described by impulse-response function, (y x), which measures the degree
of inuence which point y has on point x. In other words, (y x) measures
the strength of the response at x to an impulse at y. In a system which evolves
in time, may also depend on time (since it takes time for the eect from y to
propagate to x), so also depends on time, and is written
t
(y x).
Intuitively, should have four properties:
376 DRAFT Chapter 17. Impulse-response methods
y
x
x
y
1
y
2
y
3
y
4
y
5
(B) (A)
(

)
y
-
>
x
(
y
1

-
>

x
)
(y
2 ->
x)
(
y
3

-
>

x
)
(
y
4

-
>

x
)
(
y
5
-
>

x
)
Figure 17A.1: (A) (y x) describes the response at x to an impulse at y.
(B) The state at x is a sum of its responses to the impulses at y
1
, y
2
, . . . , y
5
.
(i) Inuence should decay with distance. In other words, if y and x are close
together, then (y x) should be large; if y and x are far apart, then
(y x) should be small (Figure 17A.2).
(ii) In a spatially homogeneous or translation invariant system (Figure 17A.3(A)),
should only depend on the displacement from y to x, so that we can write
(y x) = (x y), where is some other function.
(iii) In an isotropic or rotation invariant system system (Figure 17A.3(B)),
should only depend on the distance between y and x, so that we can write
(y x) =
_
[x y[
_
, where is a function of one real variable, and
lim
r
(:) = 0.
(iv) In a time-evolving system, the value of
t
(y x) should rst grow as t
increases (as the eect propagates from y to x), reach a maximum value,
and then decrease to zero as t grows large (as the eect dissipates through
space) (see Figure 17A.4).
Thus, if there is an impulse of magnitude 1 at y, and 1(x) is the response
at x, then
1(x) = 1 (y x) (see Figure 17A.5A)
What if there is an impulse 1(y
1
) at y
1
, an impulse 1(y
2
) at y
2
, and an impulse
1(y
3
) at y
3
? Then statement (17A.1) implies:
1(x) = 1(y
1
) (y
1
x) + 1(y
2
) (y
2
x) + 1(y
3
) (y
3
x).
(see Figure 17A.5B). If X is the domain of the PDE, then suppose, for every y
377
y
x
(
y

x
)
Figure 17A.2: The inuence of y on x becomes small as the distance from y to
x grows large.
(y
1
x
1
)
x
1
y
1
(y
2
x
2
)
x
2
= x
1
+ v
y
2
= y
1
+ v
v
v
(A)
x
y
1
y
2
(y2
x
)
(
y
1

x
)
r
(B)
Figure 17A.3: (A) Translation invariance: If y
2
= y
1
+v and x
2
= x
1
+v, then
(y
2
x
2
) = (y
1
x
1
). (B) Rotation invariance: If y
1
and y
2
are both
the same distance from x (i.e. they lie on the circle of radius : around x), then
(y
2
x) = (y
1
x).
t
t

(
y
-
>
x
)
Figure 17A.4: The time-dependent impulse-response function rst grows large,
and then decays to zero.
(y x)
y y
y y y
1 2 3
y y y
1 2 3
3
2
1
x
x
(A)
(B)
Figure 17A.5: (A) An impulse of magnitude 1 at y triggers a response of
magnitude 1 (y x) at x. (B) Multiple impulses of magnitude 1
1
, 1
2
and 1
3
at y
1
, y
2
and y
3
, respectively, triggers a response at x of magnitude
1
1
(y
1
x) + 1
2
(y
2
x) + 1
3
(y
3
x).
in X, that 1(y) is the impulse at y. Then statement (17A.1) takes the form:
1(x) =
yX
1(y) (y x). (17A.3)
But now we are summing over all y in X, and usually, X = 1
D
or some subset, so
the summation in (17A.3) doesnt make mathematical sense. We must replace
the sum with an integral, as in statement (17A.2), to obtain:
1(x) =
_
X
1(y) (y x) dy. (17A.4)
If the system is spatially homogeneous, then according to (ii), this becomes
1(x) =
_
1(y) (x y) dy.
This integral is called a convolution, and is usually written as 1 . In other
words,
1(x) = 1 (x), where 1 (x) :=
_
1(y) (x y) dy.
(17A.5)
Note that 1 is a function of x. The variable y appears on the right hand side,
but as only an integration variable.
In a time-dependent system, (17A.4) becomes:
1(x; t) =
_
X
1(y)
t
(y x) dy.
while (17A.5) becomes:
1(x; t) = 1
t
(x), where 1
t
(x) =
_
1(y)
t
(x y) dy.
(17A.6)
The following surprising property is often useful:
Proposition 17A.1. If ), p : 1
D
1 are integrable functions, then p ) =
) p.
Proof. (Case 1 = 1) Fix r 1. Then
(p ))(r) =
_

p(j) )(r j) dj
(s)
_

p(r .) )(.) (1) d.

=
_

)(.) p(r .) d. = () p)(r).

Here, step (s) was the substitution . = rj, so that j = r. and dj = d..
Exercise 17A.1 Generalize this proof to the case 1 2. 2 E _
17B. Approximations of identity 379
Remarks: (a) Depending on the context, impulse-response functions are some-
times called solution kernels, or Greens functions or impulse functions.
(b) If ) and p are analytic functions, then there is an ecient way to compute
) p using complex analysis; see Corollary 18H.3 on page 474.
17B Approximations of identity
17B(i) ...in one dimension
Prerequisites: 17A.
Suppose : 11
+
1 was a one-dimensional impulse response function,
as in equation (17A.6). Thus, if 1 : 1 1 is a function describing the initial
impulse, then for any time t 0, the response is given by the function 1
t
dened:
1
t
(r) := 1
t
(r) =
_

1(j)
t
(r j) dj. (17B.1)
Intuitively, if t is close to zero, then the response 1
t
should be concentrated near
the locations where the impulse 1 is concentrated (because the energy has not
yet been able to propagate very far). By inspecting eqn.(17B.1), we see that this
means that the mass of
t
should be concentrated near zero. Formally, we say
that is an approximation of the identity if it has the following properties
(Figure 17B.1):
(AI1)
t
(r) 0 everywhere, and
_

t
(r) dr = 1 for any xed t 0.
t0
_

t
(r) dr = 1.
Property (AI1) says that
t
is a probability density. (AI2) says that
t
concentrates all of its mass at zero as t 0. (Heuristically speaking, the
function
t
is converging to the Dirac delta function
0
as t 0.)
Example 17B.1.
(a) Let
t
(r) =
_
1
t
if 0 r t;
0 if r < 0 or t < r.
(Figure 17B.2)
Thus, for any t 0, the graph of
t
is a box of width t and height 1,t.
Then is an approximation of identity. (See Practice Problem # 11 on
page 413.)
380 DRAFT

(AI1)
(AI2) t=10
(AI2) t=1
(AI2) t=0.1
(AI2) t=0.01
F
G
H
I
J
1
0.2
0.5
0.8
0.95
Figure 17B.1: is an approximation of the identity.
2
1/2
1
1
1/2
2
1/3
3
1/4
4
1/2
1/3
1/4
Figure 17B.2: Example 17B.1(a)
(b) Let
t
(r) =
_
1
2t
if [r[ t
0 if t < [r[
.
Thus, for any t 0, the graph of
t
is a box of width 2t and height 1,2t.
Then is an approximation of identity. (See Practice Problem # 12 on
page 413.)
A function satisfying properties (AI1) and (AI2) is called an approximation
of the identity because of the following theorem:
Proposition 17B.2. Let : 1 1
+
1 be an approximation of identity.
(a) Let 1 : 1 1 be a bounded continuous function. Then for all r 1,
lim
t0
1
t
(r) = 1(r).
(b) Let 1 : 1 1 be any bounded integrable function. If r 1 is any
continuity-point of 1, then lim
t0
1
t
(r) = 1(r).
Proof. (a) Fix r 1. Given any c 0, nd 0 such that,
For all j 1,
_
[j r[ <
_
=
_
1(j) 1(r)
<

3
_
.
(Such an c exists because 1 is continuous). Thus,
1(r)
_
x+
x
t
(r j) dj
_
x+
x
1(j)
t
(r j) dj
_
x+
x
_
1(r) 1(j)
_

t
(r j) dj

_
x+
x
1(r) 1(j)

t
(r j) dj
<
c
3
_
x+
x
t
(r j) dj <
(AI1)
c
3
. (17B.2)
(Here (AI1) is by property (AI1) of
t
.)
Recall that 1 is bounded. Suppose [1(j)[ < ` for all j 1; using (AI2), nd
some small 0 such that, if t < , then
_
x+
x
t
(j) dj 1
c
3`
; hence
_
x
t
(j) dj +
_

x+
t
(j) dj =
_

t
(j) dj
_
x+
x
t
(j) dj
<
(AI1)
1
_
1
c
3`
_
=
c
3`
. (17B.3)
(Here (AI1) is by property (AI1) of
t
.) Thus,
1
t
(r)
_
x+
x
1(j)
t
(r j) dj
1(j)
t
(r j) dj
_
x+
x
1(j)
t
(r j) dj
_
x
1(j)
t
(r j) dj +
_

x+
1(j)
t
(r j) dj
_
x
1(j)
t
(r j)
dj +
_

x+
1(j)
t
(r j)
dj
_
x
`
t
(r j) dj +
_

x+
`
t
(r j) dj
`
__
x
t
(r j) dj +
_

x+
t
(r j) dj
_
()
`
c
3`
=
c
3
. (17B.4)
(Here, () is by eqn.(17B.3).) Combining equations (17B.2) and (17B.4) we
have:
1(r)
_
x+
x
t
(r j) dj 1
t
(r)
1(r)
_
x+
x
t
(r j) dj
_
x+
x
1(j)
t
(r j) dj
_
x+
x
1(j)
t
(r j) dj 1
t
(r)
c
3
+
c
3
=
2c
3
. (17B.5)
But if t < ,then
1
_
x+
x
t
(r j) dj
<
c
3`
. Thus,
1(r) 1(r)
_
x+
x
t
(r j) dj
[1(r)[
1
_
x+
x
t
(r j) dj
< [1(r)[
c
3`
`
c
3`
=
c
3
. (17B.6)
Combining equations (17B.5) and (17B.6) we have:
[1(r) 1
t
(r)[
1(r) 1(r)
_
x+
x
t
(r j) dj
1(r)
_
x+
x
t
(r j) dj 1
t
(r)
c
3
+
2c
3
. = c.
Since c can be made arbitrarily small, were done.
(b) Exercise 17B.1 (Hint: imitate part (a)). 2 E _
In other words, as t 0, the convolution 1
t
resembles 1 with arbitrarily
high accuracy. Similar convergence results can be proved in other norms (e.g.
1
2
convergence, uniform convergence).
Example 17B.3. Let
t
(r) =
_
1
t
if 0 r t
0 if r < 0 or t < r
, as in Example
17B.1(a). Suppose 1 : 1 1 is a continuous function. Then for any r 1,
1
t
(r) =
_

1(j)
t
(rj) dj =
1
t
_
x
xt
1(j) dj =
1
t
_
(r) (r t)
_
,
where is an antiderivative of 1. Thus, as implied by Proposition 17B.2,
lim
t0
1
t
(r) = lim
t0
(r) (r t)
t
()
t
(r)
()
1(r).
(Here () is just the denition of dierentiation, and () is because is an antiderivative of
1.)
17B(ii) ...in many dimensions
Prerequisites: 17B(i). Recommended: 17C(i).
A nonnegative function : 1
D
1
+
1
,
is called an approximation of
the identity if it has the following two properties:
(AI1)
_
1
D
t
(x) dx = 1 for all t [0, ].
t0
_
B(0;)
t
(x) dx = 1.
Property (AI1) says that
t
is a probability density. (AI2) says that
t
con-
centrates all of its mass at zero as t 0.
Example 17B.4. Dene : 1
2
1
+
1 by
t
(r, j) =
_
1
4t
2
if [r[ t and [j[ t;
0 otherwise.
.
Then is an approximation of the identity on 1
2
. (Exercise 17B.2 ) E _
Proposition 17B.5. Let : 1
D
1
+
1 be an approximation of the
identity.
(a) Let 1 : 1
D
1 be a bounded continuous function. Then for every
x 1
D
, we have lim
t0
1
t
(x) = 1(x).
(b) Let 1 : 1
D
1 be any bounded integrable function. If x 1
D
is any
continuity-point of 1, then lim
t0
1
t
(x) = 1(x).
Proof. Exercise 17B.3 Hint: the argument is basically identical to that of E _
Proposition 17B.2; just replace the interval (c, c) with a ball of radius c. 2
In other words, as t 0, the convolution 1
t
resembles 1 with arbitrarily
high accuracy. Similar convergence results can be proved in other norms (e.g.
1
2
convergence, uniform convergence).
When solving partial dierential equations, approximations of identity are
invariably used in conjunction with the following result:
Proposition 17B.6. Let L be a linear dierential operator on (
(1
D
; 1).
(a) If : 1
D
1 is a solution to the homogeneous equation L = 0, then
for any function 1 : 1
D
1, the function n = 1 satises: L n = 0.
(b) If : 1
D
1
+
1 satises the evolution equation
n
t
= L , and
we dene
t
(x) := (x; t), then for any function 1 : 1
D
1, the function
n
t
= 1
t
satises:
n
t
n = L n.
Proof. Exercise 17B.4 Hint: Generalize the proof of Proposition 17C.1 on E _
the facing page, by replacing the one-dimensional convolution integral with a 1-
dimensional convolution integral, and by replacing the Laplacian with an arbitrary
linear operator L. 2
Corollary 17B.7. Suppose is an approximation of the identity and satises
the evolution equation
n
t
= L . For any 1 : 1
D
1, dene n : 1
D
1
,
1 by:
n(x; 0) = 1(x).
n
t
= 1
t
, for all t 0.
Then n is a solution to the equation
n
t
n = L n, and n satises the initial
conditions n(x, 0) = 1(x) for all x 1
D
.
Proof. Combine Propositions 17B.5 and 17B.6. 2
17C. The Gaussian convolution solution (heat equation) 385
We say that is the fundamental solution (or solution kernel, or Greens
function or impulse function) for the PDE. For example, the 1-dimensional
Gauss-Weierstrass kernel is a fundamental solution for the 1-dimensional heat
equation.
17C The Gaussian convolution solution (heat equation)
17C(i) ...in one dimension
Prerequisites: 1B(i), 17B(i), 0G. Recommended: 17A, 20A(ii) .
Given two functions 1, ( : 1 1, recall (from 17A) that their convolu-
tion is the function 1 ( : 1 1 dened:
1 ((r) :=
_

1(j) ((r j) dj, for all r 1.

Recall the Gauss-Weierstrass kernel from Example 1B.1 on page 6:
(
t
(r) :=
1
2
t
exp
_
r
2
4t
_
, for all r 1 and t 0.
We will use (
t
(r) as an impulse-response function to solve the one-dimensional
heat equation.
Proposition 17C.1. Let 1 : 1 1 be a bounded integrable function.
Dene n : 1 1
+
1 by n(r; t) := 1 (
t
(r) for all r 1 and t 0. Then
n is a solution to the one-dimensional heat equation.
Proof. For any xed j 1, dene n
y
(r; t) = 1(j) (
t
(r j).
Claim 1: n
y
(r; t) is a solution of the one-dimensional heat equation.
Proof. First note that
t
(
t
(r j) =
2
x
(
t
(r j) (Exercise 17C.1 ). E _
Now, j is a constant, so we treat 1(j) as a constant when dierentiating by
r or by t. Thus,
t
n
y
(r, t) = 1(j)
t
(
t
(r j) = 1(j)
2
x
(
t
(r j)
=
2
x
n
y
(r, t) = n
y
(r, t),
as desired.
Claim 1
Now, n(r, t) = 1 (
t
=
_

1(j) (
t
(r j) dj =
_

n
y
(r; t) dj. Thus,
t
n(r, t)
()
_

t
n
y
(r; t) dj
()
_

n
y
(r; t) dj
()
n(r, t).
386 DRAFT
(A)
(B)
(C)
4 3 2 0 2 3 4 5 6 7 5
0
4
4 3 2 0 2 3 4 5 6 7 5
-5
-4
-3
-2
-1
0
1
2
3
4
5
4 3 2 0 2 3 4 5 6 7 5
6
7
(D)
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
(
t
(r)
(
(5)
t
(r)
1(r)
1
1
1
1
1
1
1
1
1
1
1
1
n=
1
n
(
(n)
t
(r)
Figure 17C.1: Discrete convolution: a superposition of Gaussians
(A)
(B)
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
(
t
(
r
)
(
(
5
)
t
(
r
)
1(r)
1
n
=
1
n
(
(
n
)
t
(
r
)
1 (
t
(r) =
_
1
1(j) (
t
(r j) dj = lim
0
c
n=
1(nc) (
t
(r nc)
Figure 17C.2: Convolution as a limit of discrete convolutions.
Here, () is by Claim 1, and () is by Proposition 0G.1 on page 567.
(Exercise 17C.2 Verify that the conditions of Proposition 0G.1 are satised.) 2 E _
Remark. One way to visualize the Gaussian convolution n(r; t) = 1 (
t
(r)
is as follows. Consider a nely spaced c-mesh of points on the real line,
c Z = nc ; n Z .
For every n Z, dene the function (
(n)
t
(r) = (
t
(r nc). For example,
(
(5)
t
(r) = (
t
(r 5c) looks like a copy of the Gauss-Weierstrass kernel, but
centered at 5c (see Figure 17C.1A).
For each n Z, let 1
n
= 1(n c) (see Figure 17C.1C). Now consider the
innite linear combination of Gauss-Weierstrass kernels (see Figure 17C.1D):
n
(r; t) = c
n=
1
n
(
(n)
t
(r).
Now imagine that the c-mesh become innitely dense, by letting c 0. Dene
n(r; t) = lim
0
n
(r; t). I claim that n(r; t) = 1 (

t
(r). To see this, note that
n(r; t) = lim
0
c
n=
1
n
(
(n)
t
(r) = lim
0
c
n=
1(nc) (
t
(r nc)
()
_

1(j) (
t
(r j) dj = 1 (
t
(j),
as shown in Figure 17C.2.
Exercise 17C.3. Rigorously justify step () in the previous computation. (Hint. E _
Use a Riemann sum.)
Proposition 17C.2. The Gauss-Weierstrass kernel is an approximation of
identity (see 17B(i)), meaning that it satises the following two properties:
(AI1) (
t
(r) 0 everywhere, and
_

(
t
(r) dr = 1 for any xed t 0.
t0
_

(
t
(r) dr = 1.
0
Figure 17C.3: The Heaviside step function H(r).
Corollary 17C.3. Let 1 : 1 1 be a bounded integrable function. Dene
the function n : 1 1
,
1 by
n
0
(r) := 1(r) for all r 1 (initial conditions).
n
t
:= 1 (
t
, for all t 0.
Then n is a solution to the one-dimensional heat equation. Furthermore:
(a) If 1 is continuous on 1, then n is continuous on 1 1
,
.
(b) Even if 1 is not continuous, the function n is still continuous on 1 1
+
,
and n is also continuous at (r, 0) for any r 1 where 1 is continuous.
Proof. Propositions 17C.1 says that n is a solution to the heat equation.
Combine Proposition 17C.2 with Proposition 17B.2 on page 381 to verify the
continuity assertions (a) and (b). 2
The continuity part of Corollary 17C.3 means that n is the solution to the
initial value problem for the heat equation with initial conditions 1. Because
of Corollary 17C.3, we say that ( is the fundamental solution (or solution
kernel, or Greens function or impulse function) for the heat equation.
Example 17C.4: The Heaviside Step function
Consider the Heaviside step function H(r) =
_
1 if r 0
0 if r < 0
(see Fig-
ure 17C.3). The solution to the one-dimensional heat equation with initial
conditions n(r, 0) = H(r) is given by:
n(r, t)
()
H (
t
(r)
()
(
t
H(r) =
_

(
t
(j) H(r j) dj
=
1
2
t
_

exp
_
j
2
4t
_
H(r j) dj
()
1
2
t
_
x
exp
_
j
2
4t
_
dj
()
1
2
_
x/
2t
exp
_
.
2
2
_
d. =
_
r
2t
_
.
389
0
Figure 17C.4: n
t
(r) = (H (
t
)(r) evaluated at several r 1.
t=1
t=3
t=5
t=7
t=9
t=1
t=3
t=5
t=7
t=9
P
S
f
r
a
g
r
e
p
l
a
c
e
m
e
n
t
s
(
t
(r)
n
t
(r)
Figure 17C.5: n
t
(r) = (H (
t
)(r) for several t 0.
Here, () is by Prop. 17C.1 on page 385; () is by Prop. 17A.1 on page 378;
() is because H(rj) =
_
1 if j r
0 if j r
, and () is where we make the
substitution . =
y
2t
; thus, dj =

2t d..
Here, (r) is the cumulative distribution function of the standard normal
probability measure
1
, dened:
(r) :=
1
2
_
x
exp
_
.
2
2
_
d..
(see Figure 17C.4). At time zero, n(r, 0) = H(r) is a step function. For t 0,
n(r, t) looks like a compressed version of (r): a steep sigmoid function. As
t increases, this sigmoid becomes broader and atter. (see Figure 17C.5).
When computing convolutions, you can often avoid a lot of messy integrals
by exploiting the following properties:
Proposition 17C.5. Let ), p : 1 1 be integrable functions. Then:
(a) If / : 1 1 is another integrable function, then ) (p +/) = () p) +
() /).
(b) If : 1 is a constant, then ) (: p) = : () p).
(c) Suppose d 1 is some displacement, and we dene )
d
(r) = )(r d).
Then ()
d
p)(r) = () p)(r d). (i.e. ()
d
) p = () p)
d
.)
Proof. See Practice Problems #2 and # 3 on page 411 of 17H. 2
Example 17C.6: A staircase function
Suppose 1(r) =
_
_
0 if r < 0
1 if 0 r < 1
2 if 1 r < 2
0 if 2 r
(see Figure 17C.6A). Let (r) be
the sigmoid function from Example 17C.4. Then
n(r, t) =
_
r
2t
_
+
_
r 1
2t
_
2
_
r 2
2t
_
(see Figure 17C.6B)
1
This is sometimes called the error function or sigmoid function. Unfortunately, no
simple formula exists for (x). It can be computed with arbitrary accuracy using a Taylor
series, and tables of values for (x) can be found in most statistics texts.
2
1
1 2 x
t
0.5
0
0.5
1
1.5
2
2.5
x 0
0.05
0.1
t
0
1
2
3
4
(A) (B)
Figure 17C.6: (A) A staircase function. (B) The resulting solution to the
heat equation.
To see this, observe that we can write:
1(r) = H(r) + H(r 1) 2 H(r 2) (17C.1)
= H + H
1
(r) 2H
2
(r), (17C.2)
where eqn. (17C.2) uses the notation of Proposition 17C.5(c). Thus,
n(r; t)
()
1 (
t
(r)
()
_
H + H
1
2H
2
_
(
t
(r)
()
H (
t
(r) + H
1
(
t
(r) 2H
2
(
t
(r)
()
H (
t
(r) + H (
t
(r 1) 2H (
t
(r 2)
()
_
r
2t
_
+
_
r 1
2t
_
2
_
r 2
2t
_
. (17C.3)
Here, () is by Proposition 17C.1 on page 385; () is by eqn. (17C.2); ()
is by Proposition 17C.5(a) and (b); () is by Proposition 17C.5(c); and
() is by Example 17C.4.
Another approach. Begin with eqn. (17C.1), and, rather than using Propo-
sition 17C.5, use instead the linearity of the heat equation, along with Theo-
rem 4C.3 on page 65, to deduce that the solution must have the form:
n(r, t) = n
0
(r, t) + n
1
(r, t) 2 n
2
(r, t), (17C.4)
where
n
0
(r, t) is the solution with initial conditions n
0
(r, 0) = H(r),
n
1
1
(r, 0) = H(r 1),
n
2
2
(r, 0) = H(r 2),
But then we know, from Example 17C.4 that
n
0
(r, t) =
_
r
2t
_
; n
1
(r, t) =
_
r 1
2t
_
; and n
2
(r, t) =
_
r 2
2t
_
;
(17C.5)
Now combine (17C.4) with (17C.5) to again obtain the solution (17C.3).
Remark. The Gaussian convolution solution to the heat equation is revisited
in 20A(ii) on page 530, using the methods of Fourier transforms.
17C(ii) ...in many dimensions
Prerequisites: 1B(ii), 17B(ii). Recommended: 17A, 17C(i).
Given two functions 1, ( : 1
D
1, their convolution is the function
1 ( : 1
D
1 dened:
1 ((x) :=
_
1
D
1(y) ((x y) dy.
Note that 1 ( is a function of x. The variable y appears on the right hand
side, but as an integration variable.
Consider the the 1-dimensional Gauss-Weierstrass kernel:
(
t
(x) :=
1
(4t)
D/2
exp
_
|x|
2
4t
_
, for all x 1
D
and t 0.
(See Examples 1B.2(b,c) on page 8). We will use (
t
(r) as an impulse-response
function to solve the 1-dimensional heat equation.
Theorem 17C.7.
Suppose 1 : 1
D
1 is a bounded continuous function. Dene the function
n : 1
D
1
,
1 by:
n
0
(x) := 1(x) for all x 1
D
(initial conditions).
n
t
:= 1 (
t
, for all t 0.
Then n is a continuous solution to the heat equation on 1
D
with initial conditions
1.
Proof.
Claim 1: n(x; t) is a solution to the 1-dimensional heat equation.
Proof. Exercise 17C.5 Hint: Combine Example 1B.2(c) on page 8 with E _
Proposition 17B.6(b) on page 384.
Claim 1
17D. dAlemberts solution (one-dimensional wave equation) 393
Claim 2: ( is an approximation of the identity on 1
D
.
Proof. Exercise 17C.6
Claim 2
E _
Now apply Corollary 17B.7 on page 384 2
Because of Theorem 17C.7, we say that ( is the fundamental solution for
the heat equation.
Exercise 17C.7. In Theorem 17C.7, suppose the initial condition 1 had some E _
points of discontinuity in 1
D
. What can you say about the continuity of the function
n? In what sense is n still a solution to the initial value problem with initial conditions
n
0
= 1?
17D dAlemberts solution (one-dimensional wave equa-
tion)
Algebra is generous; she often gives more than is asked of her. Jean le Rond dAlembert
dAlemberts method provides a solution to the one-dimensional wave equa-
tion
2
t
n =
2
x
n (17D.1)
with any initial conditions, using combinations of travelling waves and ripples.
First well discuss this in the innite domain X = 1, then well consider a nite
domain like X = [o, /].
17D(i) Unbounded domain
Prerequisites: 2B(i). Recommended: 17A.
Lemma 17D.1. (Travelling Wave Solution)
Let )
0
: 1 1 be any twice-dierentiable function. Dene the functions
n
L
, n
R
: 1 1
,
1 by n
L
(r, t) := )
0
(r + t) and n
R
(r, t) := )
0
(r t), for
any r 1 and any t 0 (see Figure 17D.1). Then n
L
and n
R
are solutions to
the wave equation (17D.1), with
Initial Position: n
L
(r, 0) = )
0
(r) = n
R
(r, 0),
Initial Velocities:
t
n
L
(r, 0) = )
t
0
(r);
t
n
R
(r, 0) = )
t
0
(r).
Dene n : 11
,
1 by n(r, t) :=
1
2
_
n
L
(r, t) +n
R
(r, t)
_
, for all r 1 and
t 0. Then n is a solution to the wave equation (17D.1), with Initial Position
n(r, 0) = )
0
(r) and Initial Velocity
t
n(r, 0) = 0.
Proof. See Practice Problem # 5 on page 412. 2
Physically, n
L
represents a leftwards-travelling wave: take a copy of the
function )
0
and just rigidly translate it to the left. Similarly, n
R
represents a
rightwards-travelling wave. (Navely, it seems that n
L
(r, t) = )
0
(r + t) should
be a rightwards travelling wave, while n
R
should be leftwards travelling wave.
Yet the opposite is true. Think about this until you understand it. It may
be helpful to do the following: Let )
0
(r) = r
2
. Plot )
0
(r), and then plot
n
L
(r, 5) = )(r + 5) = (r + 5)
2
. Observe the motion of the parabola.)
f (x)
0
w (x,t)
R
w(x,t) =
w (x,t)
L
1
2
w (x,t)
R
+
w (x,t)
L
Figure 17D.1: The dAlembert travelling wave solution; )
0
(r) =
1
x
2
+1
from Ex-
ample 17D.2.
Example 17D.2. (a) If )
0
(r) =
1
r
2
+ 1
, then n(r) =
1
2
_
1
(r +t)
2
+ 1
+
1
(r t)
2
+ 1
_
(Figure 17D.1)
(b) If )
0
(r) = sin(r), then
n(r; t) =
1
2
_
sin(r +t) + sin(r t)
_
=
1
2
_
sin(r) cos(t) + cos(r) sin(t) + sin(r) cos(t) cos(r) sin(t)
_
=
1
2
_
2 sin(r) cos(t)
_
= cos(t) sin(r),
In other words, two sinusoidal waves, traveling in opposite directions, when
superposed, result in a sinusoidal standing wave.
x
x+t
f(x)
w
L
(x,t)
0
t
t
t
-1 1 -1-t 1-t
Figure 17D.2: The travelling box wave n
L
(r, t) = )
0
(r + t) from Example
17D.2(c).
(c) (see Figure 17D.2) Suppose )
0
(r) =
_
1 if 1 < r < 1
0 otherwise
. Then:
n
L
(r, t) = )
0
(r+t) =
_
1 if 1 < r +t < 1
0 otherwise
=
_
1 if 1 t < r < 1 t;
0 otherwise.
(Notice that the solutions n
L
and n
R
are continuous (or dierentiable) only
when )
0
is continuous (or dierentiable). But the formulae of Lemma 17D.1
make sense even when the original wave equation itself ceases to make sense,
as in Example (c). This is an example of a generalized solution of the wave
equation.)
Lemma 17D.3. (Ripple Solution)
Let )
1
: 1 1 be a dierentiable function. Dene the function : 1
1
,
1 by (r, t) :=
1
2
_
x+t
xt
)
1
(j) dj, for any r 1 and any t 0. Then
is a solution to the wave equation (17D.1), with
Initial Position: (r, 0) = 0; Initial Velocity:
t
(r, 0) = )
1
(r).
Proof. See Practice Problem #6 in 17H. 2
Physically, represents a ripple. You can imagine that )
1
describes the
energy prole of an impulse which is imparted into the vibrating medium at
time zero; this energy propagates outwards, leaving a disturbance in its wake
(see Figure 17D.5).
100
50
0
50
100
x
0
20
40
60
80
100
t
0
1
2
3
t
x
Figure 17D.3: The ripple solution with initial velocity )
1
(r) =
1
1+x
2
(see Example
17D.4(a)).
Example 17D.4. (a) If )
1
(r) =
1
1 +r
2
, then the dAlembert solution to the
initial velocity problem is
(r, t) =
1
2
_
x+t
xt
)
1
(j) dj =
1
2
_
x+t
xt
1
1 +j
2
dj
=
1
2
arctan(j)
y=x+t
y=xt
=
1
2
_
arctan(r +t) arctan(r t)
_
.
(see Figure 17D.3).
(b) If )
1
(r) = cos(r), then
(r, t) =
1
2
_
x+t
xt
cos(j) dj =
1
2
_
sin(r +t) sin(r t)
_
=
1
2
_
sin(r) cos(t) + cos(r) sin(t) sin(r) cos(t) + cos(r) sin(t)
_
=
1
2
_
2 cos(r) sin(t)
_
= sin(t) cos(r).
(c) Let )
1
(r) =
_
2 if 1 < r < 1
0 otherwise
(Figures 17D.4 and 17D.5). If
t 2, then
(r, t) =
_
_
0 if r +t < 1;
r +t + 1 if 1 r +t < 1;
2 if r t 1 < 1 r +t;
t + 1 r if 1 r t < 1;
0 if 1 r t.
397
x x-t x+t
1
-1
x x-t x+t
x
x-t
x+t 1 -1
1 -1
-1 1
-1 1
x
x+t
x-t
x
x+t x-t
-1 1 x x+t
x-t
-1 1
-1-t
1-t t-1 t+1
x <-1-t
-1-t< x < 1-t
1-t < x < t-1
t-1 < x < t+1
t+1 < x
-1-t
1-t
t-1 t+1
u(x,t)
f (x)
1
Figure 17D.4: The dAlembert ripple solution from Example 17D.4(c), evaluated
for various r 1, assuming t 2.
f(x)
v(x,0.2)
v(x,0.7)
v(x,1.0)
v(x,1.5)
v(x,2.2)
1
T
i
m
e
Figure 17D.5: The dAlembert ripple solution from Example 17D.4(c), evolving
in time.
f(x)
1
v(x,t) = g(x+t) - g(x-t)
g(x)
Figure 17D.6: The ripple solution with initial velocity: )
1
(r) =
2x
(x
2
+1)
2
(Example
17D.4(d)).
=
_
_
0 if r < 1 t;
r +t + 1 if 1 t r < 1 t;
2 if 1 t r < t 1;
t + 1 r if t 1 r < t + 1;
0 if t + 1 r.
Exercise 17D.1 Verify this formula. Find a similar formula for when t < 2. E _
Notice that, in this example, the wave of displacement propagates outwards
through the medium, and the medium remains displaced. The model contains
no restoring force which would cause the displacement to return to zero.
(d) If )
1
(r) =
2r
(r
2
+ 1)
2
, then p(r) =
1
r
2
+ 1
, and (r) =
1
2
_
1
(r +t)
2
+ 1

1
(r t)
2
+ 1
_
(see Figure 17D.6)
Remark. If p : 1 1 is an antiderivative of )
1
(i.e. p
t
(r) = )
1
(r), then
(r, t) = p(r+t)p(rt). Thus, the dAlembert ripple solution looks like the
dAlembert travelling wave solution, but with the rightward travelling wave
being vertically inverted.
Exercise 17D.2. (a) Express the dAlembert ripple solution as a convolution, as E _
described in 17A on page 375. Hint: Find an impulse-response function
t
(r), such
that )
1

t
(r) =
1
2
_
x+t
xt
)
1
(j) dj.
(b) Is
t
an approximation of identity? Why or why not?
Proposition 17D.5. (dAlembert Solution on an innite wire)
0 L 2L 3L -L
-2L -3L
Figure 17D.7: The odd 21-periodic extension.
Let )
0
: 1 1 be twice-dierentiable, and )
1
: 1 1 be dierentiable.
Dene the function n : 1 1
,
1 by
n(r, t) :=
1
2
_
n
L
(r, t) + n
R
(r, t)
_
+ (r, t), for all r 1 and t 0,
where n
L
, n
R
, and are as in Lemmas 17D.1 and 17D.3. Then n satises the
wave equation, with
Initial Position: (r, 0) = )
0
(r); Initial Velocity:
t
(r, 0) = )
1
(r).
Furthermore, all solutions to the wave equation with these initial conditions are
of this type.
Proof. This follows from Lemmas 17D.1 and 17D.3. 2
Remark. There is no nice extension of the dAlembert solution in higher di-
mensions. The closest analogy is Poissons spherical mean solution to the
three-dimensional wave equation in free space, which is discussed in 20B(ii) on
page 534.
17D(ii) Bounded domain
Prerequisites: 17D(i), 5C(i).
The dAlembert solution in 17D(i) works ne if X = 1, but what if X =
[0, 1)? We must extend the initial conditions in some way. If ) : [0, 1) 1
is any function, then an extension of ) is any function ) : 1 1 such that
)(r) = )(r) whenever 0 r 1. If ) is continuous and dierentiable, then we
normally require its extension to also be continuous and dierentiable.
The extension we want is the odd, 21-periodic extension, which is dened
as the unique function ) : 1 1 with the following three properties (see Figure
17D.7):
1
1
(A)
(B)
(C)
(D)
(E)
(F)
2 3 - -2 -3 0
0
0
0
1 2 3 -1 -2 -3
1 2 3 -1 -2 -3
0
0
Figure 17D.8: The odd, 21-periodic extension.
1. )(r) = )(r) whenever 0 r 1.
2. ) is an odd function,
2
meaning: )(r) = )(r) for all r 1.
3. ) is 21-periodic, meaning )(r + 21) = )(r) for all r 1.
Example 17D.6.
(a) Suppose 1 = 1, and )(r) = 1 for all r [0, 1) (Figure 17D.8A). Then the
odd, 2-periodic extension is dened:
)(r) =
_
1 if r . . . [2, 1) [0, 1) [2, 3) . . .
1 if r . . . [1, 0) [1, 2) [3, 4) . . .
(Figure 17D.8B)
(b) Suppose 1 = 1, and )(r) =
_
1 if r
_
0,
1
2
_
0 if r
_
1
2
, 1
_
(Figure 17D.8C).
Then the odd, 2-periodic extension is dened:
)(r) =
_
_
_
1 if r . . .
_
2, 1
1
2
_
_
0,
1
2
_
_
2, 2
1
2
_
. . .
1 if r . . .
_
1
2
, 0
_
_
1
1
2
, 2
_
_
3
1
2
, 4
_
. . .
0 otherwise
(Figure 17D.8D)
2
See 8C on page 168 for more information about odd functions.
(c) Suppose 1 = , and )(r) = sin(r) for all r [0, ) (Figure 17D.8E)
Then the odd, 2-periodic extension is given by )(r) = sin(r) for all r 1
(Figure 17D.8E).
Exercise 17D.3 Verify this. E _
We will now provide a general formula for the odd periodic extension, and
characterize its continuity and/or dierentiability. First some terminology. If
) : [0, 1) 1 is a function, then we say that ) is right-dierentiable at 0
if the right-hand derivative )
(0) is well-dened (see page 201). We can usually

extend ) to a function ) : [0, 1] 1 by dening )(1
) := lim
x,L
)(r), where
this denotes the left-hand limit of ) at 1, if this limit exists (see page 201 for
denition). We then say that ) is left-dierentiable at 1 if the left-hand
derivative )
)
(1) exists.
Proposition 17D.7. Let ) : [0, 1) 1 be any function
(a) The odd, 21-periodic extension of ) is given:
)(r) =
_
_
)(r) if 0 r < 1
)(r) if 1 r < 0
)(r 2n1) if 2n1 r (2n + 1)1, for some n Z
)(2n1 r) if (2n 1)1 r 2n1, for some n Z
(b) ) is continuous at 0, 1, 21 etc. if and only if )(0) = )(1
) = 0.
(c) ) is dierentiable at 0, 1, 21, etc. if and only if it is continuous, ) is
right-dierentiable at 0, and ) and left-dierentiable at 1.
Proposition 17D.8. (dAlembert solution on a nite string)
Let )
0
: [0, 1) 1 and )
1
: [0, 1) 1 be dierentiable functions, and let
their odd periodic extensions be )
0
: 1 1 and )
1
: 1 1.
(a) Dene n : [0, 1] 1
,
1 by
n(r, t) :=
1
2
_
)
0
(r t) +)
0
(r +t)
_
, for all r [0, 1] and t 0.
Then n is a solution to the wave equation (17D.1) with initial conditions:
n(r, 0) = )
0
(r) and
t
n(r, 0) = 0, for all r [0, 1],
and homogeneous Dirichlet boundary conditions:
n(0, t) = 0 = n(1, t), for all t 0.
The function n is continuous if and only if )
0
satises homogeneous Dirich-
let boundary conditions (i.e. )(0) = )(1
) = 0). In addition, n is dieren-

tiable if and only if )
0
is also right-dierentiable at 0 and left-dierentiable
at 1.
(b) Dene : [0, 1] 1
,
1 by
(r, t) :=
1
2
_
x+t
xt
)
1
(j) dj, for all r [0, 1] and t 0.
Then is a solution to the wave equation (17D.1) with initial conditions:
(r, 0) = 0 and
t
(r, 0) = )
1
(r), for all r [0, 1],
(0, t) = 0 = (1, t), for all t 0.
The function is always continuous. However, is dierentiable if and
only if )
1
satises homogeneous Dirichlet boundary conditions.
(c) Dene n : [0, 1] 1
,
1 by n(r, t) := n(r, t) +(r, t), for all r [0, 1]
and t 0. Then n(r, t) is a solution to the wave equation (17D.1) with
initial conditions:
n(r, 0) = )
0
(r) and
t
n(r, 0) = )
1
(r), for all r [0, 1],
n(0, t) = 0 = n(1, t), for all t 0.
Clearly, n is continuous (respectively, dierentiable) whenever and n are
continuous (respectively, dierentiable).
Proof. The fact that n, n, and are solutions to their respective initial
value problems follows from Proposition 17D.5 on page 398. The verication
of homogeneous Dirichlet conditions is Exercise 17D.5 . The conditions for E _
continuity/dierentiability are Exercise 17D.6 . 2 E _
17E. Poissons solution (Dirichlet problem on half-plane) 403
(A) (B)
u(x,y)
b(x)
={(x,y); y > 0}
H
H = R
Figure 17E.1: The Dirichlet problem on a half-plane.
17E Poissons solution (Dirichlet problem on half-plane)
Prerequisites: 1C, 5C, 0G, 17B(i). Recommended: 17A.
Consider the half-plane domain H :=
_
(r, j) 1
2
; j 0
_
. The boundary
of this domain is just the r axis: H = (r, 0) ; r 1. Thus, we impose
boundary conditions by choosing some function / : 1 1. Figure 17E.1
illustrates the corresponding Dirichlet problem: nd a continuous function
n : H 1 such that
1. n is harmonic i.e. n satises the Laplace equation: n(r, j) = 0 for all
r 1 and j 0.
2. n satises the nonhomogeneous Dirichlet boundary condition: n(r, 0) =
/(r), for all r 1.
Physical Interpretation: Imagine that H is an innite ocean, so that H
is the beach. Imagine that /(r) is the concentration of some chemical which has
soaked into the sand of the beach. The harmonic function n(r, j) on H describes
the equilibrium concentration of this chemical, as it seeps from the sandy beach
and diuses into the water
3
. The boundary condition n(r, 0) = /(r) represents
the chemical content of the sand. Note that /(r) is constant in time; this
represents the assumption that the chemical content of the sand is large compared
to the amount seeping into the water; hence, we can assume the sands chemical
content remains eectively constant over time, as small amounts diuse into the
water.
3
Of course this an unrealistic model: in a real ocean, currents, wave action, and weather
transport chemicals far more quickly than mere diusion alone.
4
0
3 2
0.5
2.5
1
0 2
1.5
x
2
1.5
-2
2.5
y 1
0.5 -4
0
0
0
0.5
0.5
4
1
1
2
1.5
1.5
y 0
2
2 x
2.5
-2
2.5
-4
3
Figure 17E.2: Two views of the Poisson kernel /
y
(r).
We will solve the half-plane Dirichlet problem using the impulse-response
method. For any j 0, dene the Poisson kernel /
y
: 1 1 by:
/
y
(r) :=
j
(r
2
+j
2
)
. (Figure 17E.2) (17E.1)
Observe that:
/
y
(r) is smooth for all j 0 and r 1.
/
y
(r) has a singularity at (0, 0). That is: lim
(x,y)(0,0)
/
y
(r) = ,
/
y
(r) decays near innity. That is, for any xed j 0, lim
x
/
y
(r) =
0, and also, for any xed r 1, lim
y
/
y
(r) = 0.
Thus, /
y
(r) has the prole of an impulse-response function as described in 17A
on page 375. Heuristically speaking, you can think of /
y
(r) as the solution to
the Dirichlet problem on H, with boundary condition /(r) =
0
(r), where
0
is
the infamous Dirac delta function. In other words, /
y
(r) is the equilibrium
concentration of a chemical diusing into the water from an innite concentra-
tion of chemical localized at a single point on the beach (say, a leaking barrel of
toxic waste).
Proposition 17E.1. Poisson Kernel Solution to Half-Plane Dirichlet problem
Let / : 1 1 be a bounded, continuous, integrable function. Dene n : H
1 as follows:
n(r, j) := / /
y
(r) =
j
/(.)
(r .)
2
+j
2
d.,
17E. Poissons solution (Dirichlet problem on half-plane) 405
(A)
(x,y)
A B
y
x
x-A B-x
B
A
(B)
(x,y)
A B
x
A-x
B-x
B
A
Figure 17E.3: Example 17E.2.
for all r 1 and j 0, while for all r 1, we dene n(r, 0) := /(r). Then n is
the solution to the Laplace equation (n = 0) which is bounded at innity and
which satises the nonhomogeneous Dirichlet boundary condition n(r, 0) = /(r),
for all r 1.
Proof. (sketch)
Claim 1: Dene /(r, j) = /
y
(r) for all (r, j) H, except (0, 0). Then the
function / : H 1 is harmonic on the interior of H.
Proof. See Practice Problem # 14 on page 414 of 17H.
Claim 1
Claim 2: Thus, the function n : H 1 is harmonic on the interior of H.
Proof. Exercise 17E.1 Hint: Combine Claim 1 with Proposition 0G.1 on E _
page 567
Claim 2
Recall that we dened n on the boundary of H by n(r, 0) = /(r). It remains
to show that n is continuous when dened in this way.
Claim 3: For any r 1, lim
y0
n(r, j) = /(r).
Proof. Exercise 17E.2 Show that the kernel /
y
is an approximation of the E _
identity as j 0. Then apply Proposition 17B.2 on page 381 to conclude that
lim
y0
(/ /
y
)(r) = /(r) for all r 1.
Claim 3
Finally, this solution is unique by Theorem 5D.5(a) on page 88. 2
Example 17E.2. Let < 1 be real numbers. Let /(r) :=
_
1 if < r < 1;
0 otherwise.
.
Then Proposition 20C.3 yields solution:
l(r, j)
()
/ /
y
(r)
()
j
_
B
A
1
(r .)
2
+j
2
d.
(S)
j
2
_ Bx
y
Ax
y
1
j
2
n
2
+j
2
dn
=
1
_ Bx
y
Ax
y
1
n
2
+ 1
dn =
1
arctan(n)
w=
Bx
y
w=
Ax
y
=
1
arctan
_
1 r
j
_
arctan
_
r
j
_
(T)
1
B

A
_
,
where
B
and
A
are as in Figure 17E.3. Here, () is Proposition 20C.3; ()
is eqn.(17E.1); (S) is the substitution n =
zx
y
, so that dn =
1
y
d. and
d. = j dn; and (T) follows from elementary trigonometry.
Note that, if < r (as in Fig. 17E.3A), then r < 0, so
A
is negative,
so that l(r, j) =
1
B
+[
A
[
_
. If r, then we have the situation in Fig.
17E.3B. In either case, the interpretation is the same:
l(r, j) =
1
B

A
_
=
1
_
the angle subtended by interval [, 1], as
seen by an observer at the point (r, j)
_
.
This is reasonable, because if this observer moves far away from the interval
[, 1], or views it at an acute angle, then the subtended angle
_
B

A
_
will
become small hence, the value of l(r, j) will also become small.
Remark. We will revisit the Poisson kernel solution to the half-plane Dirichlet
problem in 20C(ii) on page 539, where we will prove Proposition 17E.1 using
Fourier transform methods.
17F Poissons solution (Dirichlet problem on the disk)
Prerequisites: 1C, 0D(ii), 5C, 0G. Recommended: 17A, 14B(v).
4
Let | :=
_
(r, j) 1
2
;
_
r
2
+j
2
1
_
be the disk of radius 1 in 1
2
.
Thus, | has boundary | = S :=
_
(r, j) 1
2
;
_
r
2
+j
2
= 1
_
(the circle
of radius 1). Suppose / : | 1 is some function on the boundary. The
Dirichlet problem on | asks for a continuous function n : | 1 such that:
n is harmonici.e. n satises the Laplace equation n 0.
4
See 14B(v) on page 289 for a dierent development of the material in this section, using
the methods of polar-separated harmonic functions. For yet another approach, using complex
analysis, see Corollary 18C.13 on page 445.
17F. Poissons solution (Dirichlet problem on the disk) 407
0
x
s
|x-s|
p
r
o
p
o
r
t
i
o
n
a
l
t
o

1
-
|
x
|
2
0
x
s
r
(A) (B)
Figure 17F.1: The Poisson kernel
n satises the nonhomogeneous Dirichlet Boundary Condition n(r, j) =
/(r, j) for all (r, j) |.
If n represents the concentration of some chemical diusing into | from the
boundary, then the value of n(r, j) at any point (r, j) in the interior of the disk
should represent some sort of average of the chemical reaching (r, j) from all
points on the boundary. This is the inspiration of Poissons Solution. We dene
the Poisson kernel T : | S 1 as follows:
T(x, s) :=
1
2
|x|
2
|x s|
2
, for all x | and s S.
As shown in Figure 17F.1(A), the denominator, |x s|
2
, is just the squared-
distance from x to s. The numerator, 1
2
|x|
2
, roughly measures the distance
from x to the boundary S; if x is close to S, then 1
2
|x|
2
becomes very small.
Intuitively speaking, T(x, s) measures the inuence of the boundary condition
at the point s on the value of n at x; see Figure 17F.2.
In polar coordinates (Figure 17F.1B), we can parameterize s S with a
single angular coordinate [, ), so that s =
_
1cos(), 1sin()
_
. If x
has coordinates (r, j), then Poissons kernel takes the form:
T(x, s) = T
(r, j) =
1
2
r
2
j
2
(r 1cos())
2
+ (j 1sin())
2
.
Proposition 17F.1. Poissons Integral Formula
1
0
0.5
1
0.5
0
0.5
1
0
2
4
6
8
10
Figure 17F.2: The Poisson kernel T(x, s) as a function of x. (for some xed
value of s). This surface illustrates the inuence of the boundary condition at
the point s on the point x. (The point s is located at the peak of the surface.)
Let | =
_
(r, j) ; r
2
+j
2
1
2
_
be the disk of radius 1, and let / : | 1
be continuous. The unique solution to the corresponding Dirichlet problem is
the function n : | 1 dened as follows:
For any (r, j) on the interior of | n(r, j) :=
1
2
_

/() T
(r, j) d,
while, for (r, j) |, we dene n(r, j) := /(r, j).
That is, for any x |, n(x) :=
_
_
1
2
_
S
/(s) T(x, s) ds if |r| < 1;
/(x) if |r| = 1.
Proof. (sketch) For simplicity, assume 1 = 1 (the proof for 1 ,= 1 is similar).
Thus,
T
(r, j) =
1 r
2
j
2
(r cos())
2
+ (j sin())
2
.
Claim 1: Fix [, ). The function T
: | 1 is harmonic on the
interior of |.
Proof. Exercise 17F.1
Claim 1
E _
Claim 2: Thus, the function n is harmonic on the interior of |.
Proof. Exercise 17F.2 Hint: Combine Claim 1 with Proposition 0G.1 on E _
page 567.
Claim 2
17G.
Properties of convolution 409

Recall that we dened n on the boundary S of | by n(s) = /(s). It remains
to show that n is continuous when dened in this way.
Claim 3: For any s S, lim
(x,y)s
n(r, j) = /(s).
Proof. Exercise 17F.3 (Hard) E _
Hint: Write (r, j) in polar coordinates as (:, ). Thus, our claim becomes lim
lim
r1
n(:, ) =
/().
(a) Show that T
(r, j) = T
r
( ), where, for any : [0, 1), we dene
T
r
() =
1 :
2
1 2: cos() +:
2
, for all [, ).
(b) Thus, n(:, ) =
1
2
_

/() T
r
( ) d is a sort of convolution on a circle.
We can write this: n(:, ) = (/ T
r
)().
(c) Show that the function T
r
is an approximation of the identity as : 1,
meaning that, for any continuous function / : S 1, lim
r1
(/ T
r
)() = /().
For your proof, borrow from the proof of Proposition 17B.2 on page 381
Claim 3
Finally, this solution is unique by Theorem 5D.5(a) on page 88. 2
17G
Properties of convolution
Prerequisites: 17A. Recommended: 17C.
We have introduced the convolution operator to solve the Heat Equation,
but it is actually ubiquitous, not only in the theory of PDEs, but in other
areas of mathematics, especially probability theory, harmonic analysis, and group
representation theory. We can dene an algebra of functions using the operations
of convolution and addition; this algebra is as natural as the one you would form
using normal multiplication and addition.
5
Proposition 17G.1. Algebraic Properties of Convolution
Let ), p, / : 1
D
1 be integrable functions. Then the convolutions of ), p,
and / have the following relations:
Commutativity: ) p = p ).
Associativity: ) (p /) = () p) /.
Distribution: ) (p +/) = () p) + () /).
5
Indeed, in a sense, it is the same algebra, seen through the prism of the Fourier transform;
see Theorem 19B.2 on page 494.
Linearity: ) (: p) = : () p) for any constant : 1.
Proof. Commutativity is just Proposition 17A.1. In the case 1 = 1, the proofs
of the other three properties are Practice Problems #1 and #2 in 17H. The
proofs for 1 2 are Exercise 17G.1 . 2 E _
Remark. Let L
1
(1
D
) be the set of all integrable functions on 1
D
. The proper-
ties of Commutativity, Associativity, and Distribution mean that the set L
1
(1
D
),
together with the operations + (pointwise addition) and (convolution), is a
ring (in the language of abstract algebra). This, together with Linearity, makes
L
1
(1
D
) an algebra over 1.
Example 17C.4 on page 388 exemplies the convenient smoothing proper-
ties of convolution. If we convolve a rough function with a smooth function,
then this smooths out the rough function.
Proposition 17G.2. Regularity Properties of Convolution
Let ), p : 1
D
1 be integrable functions.
(a) If ) is continuous, then so is ) p (regardless of whether p is.)
(b) If ) is dierentiable, then so is ) p. Furthermore,
d
() p) = (
d
)) p.
(c) If ) is times dierentiable, then so is ) p, and
n
1
1

n
2
2
. . .
n
D
D
() p) =
_
n
1
1

n
2
2
. . .
n
D
D
)
_
p,
for any n
1
, n
2
, . . . , n
D
such that n
1
+. . . +n
D
.
(d) More generally, if L is any linear dierential operator of degree or less,
with constant coecients, then L () p) = (L )) p.
(e) Thus, if ) is a solution to the homogeneous linear equation L ) = 0, then
so is ) p.
(f ) If ) is innitely dierentiable, then so is ) p.
Proof. Exercise 17G.2 2 E _
This has a convenient consequence: any function, no matter how rough,
can be approximated arbitrarily closely by smooth functions.
1
1
1
1
-1
-1
(B) (C)
1 2 -1 -2
(A)
Figure 17H.1: Problems #1(a), #1(b), #1(c) and #2(a).
Proposition 17G.3. Suppose ) : 1
D
1 is integrable. Then there
is a sequence )
1
, )
2
, )
3
, . . . of innitely dierentiable functions which converges
pointwise to ). In other words, for every x 1
D
, lim
n
)
n
(x) = )(x).
Proof. Exercise 17G.3 Hint: Use the fact that the Gauss-Weierstrass kernel is E _
innitely dierentiable, and is also an approximation of identity. Then use Part 6 of
the previous theorem. 2
Remarks. (a) We have formulated Proposition 17G.3 in terms of pointwise
convergence, but similar results hold for 1
2
convergence, 1
1
convergence, uniform
convergence, etc. Were neglecting these to avoid technicalities.
(b) In 10D(ii) on page 214, we discuss the convolution of periodic functions
on the interval [, ], and develop a theory quite similar to the theory developed
here. In particular, Lemma 10D.6 on page 214 is analogous to Proposition 17G.1,
Lemma 10D.7 on page 215 is analogous to Proposition 17G.2, and Theorem 10D.1
on page 207 is analogous to Proposition 17G.3, except that the convergence is
in 1
2
norm.
17H Practice problems
1. Let ), p, / : 1 1 be integrable functions. Show that ) (p /) =
() p) /.
2. Let ), p, / : 1 1 be integrable functions, and let : 1 be a constant.
Prove that ) (: p +/) = : () p) + () /).
3. Let ), p : 1 1 be integrable functions. Let d 1 be some displace-
ment and dene )
d
(r) = )(r d). Prove that ()
d
) p = () p)
d
.
4. In each of the following, use the method of Gaussian convolutions to nd
the solution to the one-dimensional heat equation
t
n(r; t) =
2
x
n(r; t)
with initial conditions n(r, 0) = 1(r).
(a) 1(r) =
_
1 if 1 r 1
0 if r < 1 or 1 < r
. (see Figure 17H.1A).
(In this case, sketch your solution evolving in time.)
(b) 1(r) =
_
1 if 0 r 1
0 otherwise
(see Figure 17H.1B).
(c) 1(r) =
_
_
_
1 if 1 r 0
1 if 0 r 1
0 otherwise
(see Figure 17H.1C).
5. Let ) : 1 1 be some dierentiable function. Dene (r; t) =
1
2
_
)(r +t) +)(r t)
_
.
(a) Show that (r; t) satises the one-dimensional wave equation
2
t
(r; t) =
2
x
(r; t)
(b) Compute the initial position (r; 0).
(c) Compute the initial velocity
t
(r; 0).
6. Let )
1
: 1 1 be a dierentiable function. For any r 1 and any t 0,
dene (r, t) =
1
2
_
x+t
xt
)
1
(j) dj.
(a) Show that (r; t) satises the one-dimensional wave equation
2
t
(r; t) =
2
x
(r; t)
(b) Compute the initial position (r; 0).
(c) Compute the initial velocity
t
(r; 0).
7. In each of the following, use the dAlembert method to nd the solution to
the one-dimensional wave equation
2
t
n(r; t) =
2
x
n(r; t) with initial
position n(r, 0) = )
0
(r) and initial velocity
t
n(r, 0) = )
1
(r).
In each case, identify whether the solution satises homogeneous Dirich-
let boundary conditions when treated as a function on the interval [0, ].
Justify your answer.
(a) )
0
(r) =
_
1 if 0 r < 1
0 otherwise
; and )
1
(r) = 0 (see Figure
17H.1B).
(b) )
0
(r) = sin(3r) and )
1
(r) = 0.
(c) )
0
(r) = 0 and )
1
(r) = sin(5r).
(d) )
0
(r) = cos(2r) and )
1
(r) = 0.
(e) )
0
(r) = 0 and )
1
(r) = cos(4r).
(f) )
0
(r) = r
1/3
and )
1
(r) = 0.
(g) )
0
(r) = 0 and )
1
(r) = r
1/3
.
(h) )
0
(r) = 0 and )
1
(r) = tanh(r) =
sinh(x)
cosh(x)
.
8. Let (
t
(r) =
1
2
t
exp
_
x
2
4t
_
be the Gauss-Weierstrass Kernel. Fix :, t 0;
we claim that (
s
(
t
= (
s+t
. (For example, if : = 3 and t = 5, this means
that (
3
(
5
= (
8
).
(a) Prove that (
s
(
t
= (
s+t
by directly computing the convolution
integral.
(b) Use Corollary 17C.3 on page 388 to nd a short and elegant proof
that (
s
(
t
= (
s+t
without computing any convolution integrals.
Remark. Because of this result, probabilists say that the set (
t
t1
+
forms a stable family of probability distributions on 1. Analysts say that
(
t
t1
+
is a one-parameter semigroup under convolution.
9. Let (
t
(r, j) =
1
4t
exp
_
(r
2
+j
2
)
4t
_
be the 2-dimensional Gauss-Weierstrass
Kernel. Suppose / : 1
2
1 is a harmonic function. Show that /(
t
= /
for all t 0.
10. Let | be the unit disk. Let / : | 1 be some function, and let
n : | 1 be the solution to the corresponding Dirichlet problem with
boundary conditions /(). Prove that
n(0, 0) =
1
2
_

/() d.
Remark. This is a special case of the Mean Value Theorem for Harmonic
Functions (Theorem 1E.1 on page 16), but do not simply quote Theorem
1E.1 to solve this problem. Instead, apply Proposition 17F.1 on page 407.
11. Let
t
(r) =
_
1
t
if 0 r t;
0 if r < 0 or t < r.
(Figure 17B.2). Show that
is an approximation of identity.
12. Let
t
(r) =
_
1
2t
if [r[ t
0 if t < [r[
. Show that is an approximation of
identity.
13. Let | =
_
x 1
2
; [x[ 1
_
be the unit disk.
(a) Let n : | 1 be the unique solution to the Laplace equation
(n = 0) satisfying the nonhomogeneous Dirichlet boundary con-
ditions n(s) = 1, for all s S. Show that n must be constant: n(x) = 1
for all x |.
(b) Recall that the Poisson Kernel T : |S 1 is dened by T(x, s) =
1|x|
2
|xs|
2
, for any x | and s S. Show that, for any xed x |,
1
2
_
S
T(x, s) ds = 1.
(c) Let / : S 1 be any function, and n = 0) satisfying the nonho-
mogeneous Dirichlet boundary conditions n(s) = /(s), for all s S.
Let : := min
sS
/(s), and ` := max
sS
/(s). Show that:
For all x |, : n(x) `.
[ In other words, the harmonic function u must take its maximal and minimal
values on the boundary of the domain |. This is a special case of the Maximum
Principle for harmonic functions; see Corollary 1E.2 on page 17]
14. Let H :=
_
(r, j) 1
2
; j 0
_
be the half-plane. Recall that the half-plane
Poisson kernel is the function / : H 1 dened /(r, j) :=
j
(r
2
+j
2
)
for all (r, j) H except (0, 0) (where it is not dened). Show that / is
harmonic on the interior of H.
415
Chapter 18
Applications of complex
analysis
The shortest path between two truths in the real domain passes through the complex
domain. Jacques Hadamard
Complex analysis is one of the most surprising and beautiful areas of math-
ematics. It also has some unexpected applications to PDEs and Fourier theory,
which we will briey survey in this chapter. Our survey is far from comprehen-
sive that would require another entire book. Instead, our goal in this chapter
is to merely to sketch the possibilities. If you are interested in further explor-
ing the interactions between complex analysis and PDEs, we suggest [Asm02]
and [CB03], as well as [Asm05, Chapter 12], [Fis99, Chapters 4 and 5], [Lan85,
Chapter VIII], or the innovative and lavishly illustrated [Nee97, Chapter 12].
This chapter assumes no prior knowledge of complex analysis. However, the
presentation is slightly more abstract than most of the book, and is intended for
more theoretically inclined students. Nevertheless, someone who only wants
the computational machinery of residue calculus can skip Sections 18B, 18E and
18F, and skim the proofs in Sections 18C, 18D, and 18G, proceeding rapidly to
Section 18H.
18A Holomorphic functions
Prerequisites: 0C, 1C.
Let U C be a open set, and let ) : U C be a complex-valued function.
If n U, then the (complex) derivative of ) at n is dened:
)
t
(n) := lim
cu
cC
)(c) )(n)
c n
, (18A.1)
where all terms in this formula are understood as complex numbers. We say that
) is complex-dierentiable at n if )
t
(n) exists.
416 DRAFT Chapter 18. Applications of complex analysis
If we identify C with 1
2
in the obvious way, then we might imagine ) as a
function from a domain U 1
2
into 1
2
, and assume that the complex derivate )
t
was just another way of expressing the (real-valued) Jacobian matrix of ). But
this is not the case. Not all (real-)dierentiable functions on 1
2
can be regarded
as complex-dierentiable functions on C. To see this, let )
r
:= Re [)] : U 1
and )
i
:= Im[)] : U 1 be the real and imaginary parts of ), so that we can
write )(n) = )
r
(n) + )
i
(n)i for any n U. For any n U, let n
r
:= Re [n] and
n
i
:= Im[n], so that n = n
r
+ n
i
i. Then the (real-valued) Jacobian matrix of )
has the form
_

r
)
r

r
)
i
i
)
r

i
)
i
_
. (18A.2)
The relationship between the complex derivative (18A.1) and the Jacobian (18A.2)
is the subject of the following fundamental result:
Theorem 18A.1. (Cauchy-Riemann)
Let ) : U C and let n U. Then ) is complex-dierentiable at n if and
only if the partial derivatives
r
)
r
(n),
r
)
i
(n),
i
)
r
(n) and
i
)
i
(n) all exist, and
furthermore, satisfy the Cauchy-Riemann dierential equations (CRDEs)
r
)
r
(n) =
i
)
i
(n) and
i
)
r
(n) =
r
)
i
(n). (18A.3)
In this case, )
t
(n) =
r
)
r
(n) i
i
)
r
(n) =
i
)
i
(n) +i
r
)
i
(n).
Proof. Exercise 18A.1 (a) Compute the limit (18A.1) along the real axis E _
that is, let c = n + c where c 1, and show that lim
R0
)(n +c) )(n)
c
=
r
)
r
(n) +i
r
)
i
(n).
(b) Compute the limit (18A.1) along the imaginary axis that is, let c = n + ci
where c 1, and show that lim
R0
)(n +ci) )(n)
ci
=
i
)
i
(n) i
i
)
r
(n).
(c) If the limit (18A.1) is well-dened, then it must be the same no matter the direction
from which c approaches n. Conclude that the results of (a) and (b) must be equal.
Derive equation (18A.3). 2
Thus, the complex-dierentiable functions are actually a very special subclass
of the set of all (real-)dierentiable functions on the plane. The function ) is
called holomorphic on U if ) is complex-dierentiable at all n U. This is ac-
tually a much stronger requirement than merely requiring a real-valued function
to be (real-)dierentiable everywhere in some open subset of 1
2
. For example,
later we will show that every holomorphic function is analytic (Theorem 18D.1
on page 450). But one immediate indication of the special nature of holomorphic
functions is their close relationship to two-dimensional harmonic functions.
18A. Holomorphic functions 417
Proposition 18A.2. Let U C be an open set, and also regard U as a
subset of 1
2
in the obvious way. If ) : U C is any holomorphic function,
then )
r
: U 1 and )
i
: U 1 are both harmonic functions.
Proof. Exercise 18A.2 Hint: apply the Cauchy-Riemann dierential equations E _
(18A.3) twice to get Laplaces equation. 2
So, we can convert any holomorphic map into a pair of harmonic functions.
Conversely, we can convert any harmonic function into a holomorphic map. To
see this, suppose / : U 1 is a harmonic function. A harmonic conjugate
for / is a function p : U 1 which satises the dierential equation:
2
p(n) =
1
/(n) and
1
p(n) =
2
/(n), for all n U. (18A.4)
Proposition 18A.3. Let U 1
2
be a convex open set (e.g. a disk or a
rectangle). Let / : U 1 be any harmonic function.
(a) There exist harmonic conjugates for / on U that is, the equations (18A.4)
have solutions.
(b) Any two harmonic conjugates for / dier by a constant.
(c) If p is a harmonic conjugate to /, and we dene ) : U C by )(n) =
/(n) +p(n)i, then ) is holomorphic.
Proof. Exercise 18A.3 Hint: (a) Dene p(0) arbitrarily, and then for any E _
n = (n
1
, n
2
) U, dene p(n) =
_
u1
0
2
/(0, r) dr +
_
u2
0
1
/(n
1
, j) dj. Show that
p is dierentiable and satises eqn.(18A.4).
For (b), suppose p
1
and p
2
both satisfy eqn.(18A.4); show that p
1
p
2
is a constant
by showing that
1
(p
1
p
2
) = 0 =
2
(p
1
p
2
).
For (c), derive the CRDEs (18A.3) from the harmonic conjugacy equation (18A.4).
2
Remark. (a) If / satises a Dirichlet boundary condition on U, then its
harmonic conjugate satises an associated Neumann boundary condition on U,
and vice versa; see Exercise 18A.7 on page 421. Thus, harmonic conjugation can
be used to convert a Dirichlet BVP into a Neumann BVP, and vice versa.
(b) The convexity requirement in Proposition 18A.2 can be weakened to
simply connected. However, Proposition 18A.2 is not true if the domain U is
not simply connected (i.e. has a hole); see Exercise 18C.16(e) on page 448.
Holomorphic functions have a rich and beautiful geometric structure, with
many surprising properties. The study of such functions is called complex anal-
ysis. Propositions 18A.2 and 18A.3 imply that every fact about harmonic func-
tions in 1
2
is also a fact about complex analysis, and vice versa.
Complex analysis also has important applications to uid dynamics and elec-
trostatics, because any holomorphic function can be interpreted as sourceless,
irrotational ow, as we now explain. Let U 1
2
and let

V : U 1
2
be a
two-dimensional vector eld. Recall that the divergence of

V is the scalar eld
div

V : U 1 dened by div

V(u) :=
1
\
1
(u) +
2
\
2
(u) for all u U (see
0E(ii) on page 558). We say

V is locally sourceless if div

V 0. If

V
represents the two-dimensional ow of an incompressible uid (e.g. water) in
U, then div

V 0 means there are no sources or sinks in U. If

V represents a
two-dimensional electric (or gravitational) eld, then div

V 0 means there are
no charges (or masses) inside U.
The curl of

V is the scalar eld curl

V : U 1 dened by curl

V(u) :=
1
\
2
(u)
2
\
1
(u) for all u U. We say

V is locally irrotational if curl

V
0. If

V represents a force eld, then curl

V 0 means that the net energy
absorbed by a particle moving around a closed path in

V is zero (i.e. the eld is
conservative). If

V represents the ow of a uid, then curl

V 0 means there
are no vortices in U. (Note that this does not mean the uid must move in
straight lines without turning. It simply means that the uid turns in a uniform
manner, without turbulence).
Regard U as a subset of C, and let ) : U C be some function, with real
and imaginary parts )
r
: U 1 and )
i
: U 1. The complex conjugate
of ) is the function ) : U C dened by )(n) = )
r
(n) i)
i
(n). We can treat
) as vector eld

V : U 1
2
, where \
1
)
r
and \
2
)
i
.
Proposition 18A.4. (Holomorphic sourceless irrotational ow)
The function ) is holomorphic on U if and only if

V is locally sourceless and
irrotational on U.
In 18B, we shall see that Proposition 18A.4 yields a powerful technique for
studying uids (or electric elds) conned to a subset of the plane (see Proposi-
tion 18B.6 on page 430). In 18C, we shall see that Proposition 18A.4 is also the
key to understanding complex contour integration, through its role in the proof
of Cauchys Theorem 18C.5 on page 438.
To begin our study of complex analysis, we will verify that all the standard
facts about the dierentiation of real-valued functions carry over to complex
dierentiation, pretty much verbatim.
Proposition 18A.5. (Closure properties of holomorphic functions)
Let U C be an open set. Let ), p : U C be holomorphic functions.
(a) The function /(n) := )(n) + p(n) is also holomorphic on U, and /
t
(n) =
)
t
(n) +p
t
(n) for all n U.
(b) (Leibniz rule) The function /(n) := )(n) p(n) is also holomorphic on U,
and /
t
(n) = )
t
(n)p(n) +p
t
(n))(n) for all n U.
(c) (Quotient rule) Let U
:= n U ; p(n) ,= 0. The function /(n) :=

)(n),p(n) is also holomorphic on U
, and /
t
(n) = [p(n))
t
(n))(n)p
t
(n)],p(n)
2
for all n U
.
(d) For any n N, the function /(n) := )
n
(n) is holomorphic on U, and
/
t
(n) = n)
n1
(n) )
t
(n) for all n U.
(e) Thus, for any c
0
, c
1
, . . . , c
n
C, the polynomial function /(.) := c
n
.
n
+
+c
1
. +c
0
is holomorphic on C.
(f ) For any n N, the function /(n) := 1,p
n
(n) is holomorphic on U
:=
n U ; p(n) ,= 0 and /
t
(n) = np
t
(n),p
n+1
(n) for all n U
.
(g) For all n N, let )
n
: U C be a holomorphic function. Let ), 1 : U
C be two other functions. If uniflim
n
)
n
= ) and uniflim
n
)
t
n
= 1, then
) is holomorphic on U, and )
t
= 1.
(h) Let c
n
n=0
be any sequence of complex numbers, and consider the power
series
n=0
c
n
.
n
= c
0
+c
1
. +c
2
.
2
+c
3
.
3
+c
4
.
4
+
Suppose this series converges on U to dene a function ) : U C. Then
) is holomorphic on U. Furthermore, )
t
is given by the formal derivative
of the power series. That is:
)
t
(n) =
n=1
nc
n
.
n1
= c
1
+ 2c
2
. + 3c
3
.
2
+ 4c
4
.
3
+
(i) Let X 1 be open, let ) : X 1, and suppose ) is analytic at r X,
with a Taylor series
1
T
x
) which converges in the interval (r 1, r +1)
for some 1 0. Let | := c C ; [c r[ < 1 be the open disk of radius
1 around r in the complex plane. Then the Taylor series T
x
) converges
uniformly on |, and denes a holomorphic function 1 : | C which
extends ) (i.e. 1(:) = )(:) for all : (r 1, r +1) 1).
1
See 0H(ii) on page 569.
(j) (Chain rule) Let U, V C be open sets. Let p : U V and ) : V C
be holomorphic functions. Then the function /(n) = ) p(n) = )[p(n)] is
holomorphic on U, and /
t
(n) = )
t
[p(n)] p
t
(n) for all n U.
(k) (Inverse function rule) Let U, V C be open sets. Let p : U V be a
holomorphic function. Let ) : V U be be an inverse for p that is,
)[p(n)] = n for all n U. Let n U and = p(n) V. If p
t
(n) ,= 0, then
) is holomorphic in a neighbourhood of , and )
t
() = 1,p
t
(n).
Proof. Exercise 18A.5 Hint: For each part, the proof from single-variable (real) E _
dierential calculus generally translates verbatim to complex numbers. 2
Theorem 18A.5(i) implies that all the standard real-analytic functions have
natural extensions to the complex plane, obtained by evaluating their Taylor
series on C.
Example 18A.6. (a) We dene exp : C C by exp(.) =
n=0
.
n
n!
for all . C.
The function dened by this power series is the same as the exponential func-
tion dened by Eulers formula (0C) on page 551 in Appendix 0C. It satis-
es the same properties as the real exponential function that is, exp
t
(.) =
exp(.), exp(.
1
+.
2
) = exp(.
1
) exp(.
2
), etc. (See Exercise 18A.8 on the next
page.)
(b) We dene sin : C C by sin(.) =
n=0
(1)
n
.
2n+1
(2n + 1)!
for all . C.
(c) We dene cos : C C by cos(.) =
n=0
(1)
n
.
2n
(2n)!
for all . C.
(d) We dene sinh : C C by sinh(.) =
n=0
.
2n+1
(2n + 1)!
for all . C.
(e) We dene cosh : C C by cosh(.) =
n=0
.
2n
(2n)!
for all . C.
The complex trigonometric functions satisfy the same algebraic relations and
dierentiation rules as the real trigonometric functions (see Exercise 18A.9 on
page 422). We will later show that any analytic function on 1 has a unique ex-
tension to a holomorphic function on some open subset of C (see Corollary 18D.4
on page 453).
Exercise 18A.6. Proposition 18A.2 says that the real and imaginary parts of any E _
holomorphic function will be harmonic functions.
(a) Let :
0
, :
1
, . . . , :
n
1, and consider the real-valued polynomial )(r) = :
n
r
n
+
+ :
1
r + :
0
. Proposition 18A.5(e) says that ) extends to a holomorphic function
) : C C. Express the real and imaginary parts of ) in terms of the polar harmonic
functions
n
n=0
and
n
n=0
introduced in 14B on page 274.
(b) Express the real and imaginary parts of each of the holomorphic functions sin,
cos, sinh and cosh (from Example 18A.6) in terms of the harmonic functions introduced
in 12A on page 240.
Exercise 18A.7. (Harmonic conjugation of boundary conditions) E _
Let U 1
2
be an open subset whose boundary U is a smooth curve. Let :
[0, o] U be a clockwise, arc-length parameterization of U. That is: is a dier-
entiable bijection from [0, o) into U with (0) = (o), and [ (:)[ = 1 for all : [0, o].
Let / : U 1 be a continuous function describing a Dirichlet boundary condition on
U, and dene 1 := / : [0, o] 1. Suppose 1 is dierentiable; let 1
: [0, o] 1
be its derivative, and then dene the function /
: U 1 by /
((:)) = 1
(:) for all

: [0, o) (this denes /
on U because is a bijection). Thus, we can regard /
as the
derivative of / along the boundary of U.
Let / : U 1 be a harmonic function, and let p : U 1 be a harmonic conjugate
for /. Show that / satises the Dirichlet boundary condition
2
/(r) = /(r) + C for all
r U (where C is some constant) if and only if p satises the Neumann boundary
condition
p(r) = /
(r) for all r U.

Hint: For all : [0, o], let

N(:) denote the outward unit normal vector of U at
(:). Let R =
_
0 1
1 0
(thus, left-multiplying the matrix R rotates a vector clockwise by

90
o
).
(a) Show that (p) R = /. (Here we regard / and p as 2 1 row matrices).
(b) Show that R (:) =

N(:) for all : [0, o] (Here we regard and

N as a 1 2
column matrices. Hint: recall that is a clockwise parameterization).
(c) Show that (/ )
(:) = /[(:)] (:), for all : [0, o]. (To make sense of this,
recall that / is 2 1 matrix, while is a 1 2 matrix. Hint: use the chain rule).
(d) Show that (
p)[(:)] = (/ )
(:) for all : [0, o]. (Hint: Recall that

(
p)[(:)] = (p)[(:)]

N(:)).
(e) Conclude that
p[(:)] = /
[(:)] for all : [0, o] if and only if /[(:)] = /(:)+C

for all : [0, o] (where C is some constant).
Exercise 18A.8. (a) Show that exp
(.) = exp(.) for all . C. E _

(b) Fix r 1, and consider the smooth path : 1 1
2
dened by
(t) := [exp
r
(r +it), exp
i
(r +it)],
where exp
r
(.) and exp
i
(.) denote the real and imaginary parts of exp(.). Let 1 := c
x
;
note that (0) = (1, 0). Use (a) to show that satises the ordinary dierential equation
_

1
(t)

2
(t)
_
=
_

2
(t)
1
(t)
_
2
See 5C(i) on page 73 and 5C(ii) on page 76.
Conclude that (t) = [1cos(t), 1sin(t)] for all t 1.
(c) For any r, j 1, use (b) to show that exp(r +ij) = c
x
(cos(j) +i sin(j)).
(d) Deduce that exp(c
1
+c
2
) = exp(c
1
) exp(c
2
) for all c
1
, c
2
C.
Exercise 18A.9. (a) Show that sin
(.) = cos(.), cos
(.) = sin(.), sinh
(.) = E _
cos(.), and cosh
(.) = sinh(.), for all . C.

(b) For all . C, verify the Euler Identities:
sin(.) =
exp(.i) exp(.i)
2i
cos(.) =
exp(.i) + exp(.i)
2
sinh(.) =
exp(.) exp(.)
2
cosh(.) =
exp(.) + exp(.)
2
(c) Deduce that sinh(.) = i sin(i.) and cosh(.) = cos(i.).
(d) For all r, j 1, prove the following identities:
cos(r +ji) = cos(r) cosh(j) i sin(r) sinh(j);
sin(r +ji) = sin(r) cosh(j) +i cos(r) sinh(j).
(e) For all . C, verify the Pythagorean Identities:
cos(.)
2
+ sin(.)
2
= 1 and cosh(.)
2
sinh(.)
2
= 1.
(Later we will show that pretty much every trigonometric identity which is true on
1 will also be true over all of C; see Exercise 18D.4 on page 454.)
18B Conformal maps
Prerequisites: 1B, 5C, 18A.
A linear map ) : 1
D
1
D
is called conformal if it preserves the angles
between vectors. Thus, for example, rotations, reections, and dilations are all
conformal maps.
Let U, V 1
D
be open subsets of 1
D
. A dierentiable map ) : U V
is called conformal if its derivative D)(x) is a conformal linear map, for every
x U. One way to interpret this is depicted in Figure 18B.1. Suppose two
smooth paths
1
and
2
cross at x, and their velocity vectors
1
and
2
form an
angle at x. Let
1
= )
1
and
2
= )
2
, and let y = )(x). Then
1
and
2
are smooth paths, and cross at y, forming an angle . The map ) is conformal
if, for every x,
1
, and
2
, the angles and are equal.
Complex analysis could be redened as the study of two-dimensional con-
formal maps, because of the next result.
18B. Conformal maps 423
x
y
1
1
2
2
f
1
1
2
2
Figure 18B.1: A conformal map preserves the angle of intersection between two paths.
Proposition 18B.1. (Holomorphic conformal)
Let U 1
2
be an open subset, and let ) : U 1
2
be a dierentiable function,
with )(u) = ()
1
(u), )
2
(u)) for all u U. Identify U with a subset

U of the plane
C in the obvious way, and dene

) :

U C by

)(r +ji) = )
1
(r, j) +)
2
(r, j)i
that is,

) is just the representation of ) as a complex-valued function on C.
Then
_
) is conformal
_

_

) is holomorphic
_
.
Proof. Exercise 18B.1 (Hint: The derivative D) is a linear map on 1
2
. Show that E _
D) is conformal if and only if

) satises the Cauchy-Riemann dierential equations
(18A.3) on page 416.). 2
If U C is open, then Proposition 18B.1 means that every holomorphic map
) : U C can be treated as a conformal transformation of U. In particular
we can often conformally identify U with some other domain in the complex
plane via a suitable holomorphic map. A function ) : U V is a conformal
isomorphism if ) is conformal, invertible, and )
1
: V U is also conformal.
Proposition 18B.1 says that this is equivalent to requiring ) and )
1
to be
holomorphic.
Example 18B.2. (a) In Figure 18B.2, U = r +ji ; r 1, 0 < j < is a bi-
innite horizontal strip, and C
+
= r +ji ; r 1, j 0 is the open upper
half-plane, and )(.) = exp(.). Then ) : U C
+
is a conformal isomorphism
from U to C
+
.
(b) In Figure 18B.3, U = r +ji ; r 0, j 1 is the open right half of the
complex plane, and
o
|
=
_
r +ji ; r
2
+j
2
1
_
is the complement of the
0
i
i
i
2
0 1 2 3 -3 -2 -1
z exp(z)
Figure 18B.2: Example 18B.2(a). The map f(z) = exp(z) conformally identies a bi-innite
horizontal strip with the upper half-plane.
0 1 2 3 4 5
2
4
2
0
i
i
i
i
i
i
i
z exp(z)
-1
Figure 18B.3: Example 18B.2(b). The map f(z) = exp(z) conformally projects the right
half-plane onto the complement of the unit disk.
closed unit disk, and )(.) = exp(.). Then ) : U
o
|
is not a conformal
isomorphism (because it is many-to-one). However, ) is a conformal covering
map. This means that ) is locally one-to-one: for any point n U, with
= )(n)
o
|
, there is a neighbourhood 1
o
|
of and a neighbourhood
| U of n such that )
[
: | 1 is one-to-one. (Note that ) is not globally
one-to-one because it is periodic in the imaginary coordinate).
(c) In Figure 18B.4, U = r +ji ; r < 0, 0 < j < is a left half-innite
rectangle, and V =
_
r +ji ; j 1, r
2
+j
2
< 1
_
is the open half-disk, and
)(.) = exp(.). Then ) : U V is a conformal isomorphism from U to V.
(d) In Figure 18B.5, U = r +ji ; r 0, 0 < j < is a right half-innite
rectangle, and V =
_
r +ji ; j 1, r
2
+j
2
1
_
is the amphitheatre, and
)(.) = exp(.). Then ) : U V is a conformal isomorphism from U to V.
0
i
i
i
2
0 -1 -2 -3 -4 -5
z exp(z)
1 2
1 -1 0
i
Figure 18B.4: Example 18B.2(c). The map f(z) = exp(z) conformally identies a left
half-innite rectangle with the half-disk.
(e) In Figure 18B.6(A,B), U = r +ji ; r, j 0 is the open upper right
quarter-plane, and C
+
= r +ji ; j 0 is the open upper half-plane, and
)(.) = .
2
. Then ) is a conformal isomorphism from U to C
+
.
(f) Let C
+
:= r +ji ; j 0 be the upper half-plane, and U := C
+

ji ; 0 < j < 1; that is, U is is the upper half-plane with a vertical line-
segment of length 1 removed above the origin. Let )(.) = (.
2
+ 1)
1/2
; then )
is a conformal isomorphism from U to C
+
, as shown in Figure 18B.7(a).
(g) Let U := r+ji ; either j ,= 0 or 1 < r < 1, and let V :=
_
r +ji ;

2
< j <

2
_
be a bi-innite horizontal strip of width . Let )(.) := i arcsin(.); then ) is
a conformal isomorphism from U to V, as shown in Figure 18B.7(b).
Exercise 18B.2 Verify each of examples (a)-(g). E _
Conformal maps are very useful for solving boundary value problems, because
of the following result:
Proposition 18B.3. Let X, Y 1
2
be open domains with closures X and Y.
Let ) : X Y be a continuous surjection which conformally maps X into Y.
Let / : Y 1 be some smooth function, and dene H = / ) : X 1 .
(a) / is harmonic on X if and only if H is harmonic on Y.
(b) Let / : Y 1 be some function on the boundary of Y. Then 1 = / ) :
X 1 is a function on the boundary of X. The function / satises
the nonhomogeneous Dirichlet boundary condition
3
/(y) = /(y) for all
3
See 5C(i) on page 73.
0
i
i
i
2
0 1 2 3 4 5 6
z exp(z)
-1 -2
1 -1
i
0
Figure 18B.5: Example 18B.2(d). The map f(z) = exp(z) conformally identies a right
half-innite rectangle with the amphitheatre
y Y if and only if H satises the nonhomogeneous Dirichlet boundary
condition H(x) = 1(x) for all x X.
(c) For all x X, let

N
X
(x) be the outward unit normal vector to X at
x, let

N
Y
(x) be the outward unit normal vector to Y at )(x), and let
D)(x) be the derivative of ) at x (a linear transformation of 1
D
). Then
D)(x)[
N
X
(x)] = (x)

N
Y
(x) for some scalar (x) 0.
(d) Let / : Y 1 be some function on the boundary of Y, and dene
1 : X 1 by 1(x) := (x) /[)(x)] for all x X. Then / satises
the nonhomogeneous Neumann boundary condition
4
/(y) = /(y) for all

y Y if and only if H satises the nonhomogeneous Neumann boundary
condition
H(x) = 1(x) for all x X.

Proof. Exercise 18B.3 Hint: (a) Combine Propositions 18A.2, 18A.3, and 18B.1. E _
For (c), use the fact that ) is a conformal map, so D)(x) is a conformal linear
transformation; thus, if

N
X
(x) is normal to X, then D)(x)[
N
X
(x)] must be normal
to Y. To prove (d), use (c) and the chain rule. 2
We can apply Proposition 18B.3 as follows: given a boundary value problem
on some nasty domain X, nd a nice domain Y (e.g. a box, a disk, or a
half-plane), and a conformal isomorphism ) : X Y. Solve the boundary
value problem in Y (e.g. using the methods from Chapters 12-17), to get a
solution function / : Y 1. Finally, pull back this solution to get a solution
H = / ) : X 1 to the original BVP on X. We can obtain a suitable
4
See 5C(ii) on page 76.
427
0 0.5 1 1.5 2 2.5 3
0
0.5
1
1.5
2
2.5
3
-10 -5 0 5 10
0
2
4
6
8
10
(A) (B)
-4
-2
0
2
4
6
8
10
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
-4 -2 02468 10
-4
-2
0
2
4
6
8
10
-20 -15 -10 -5 0 5 10 15 20
0
5
10
15
20
-4 -2 02468 10
(C) (D)
Figure 18B.6: (A,B): Example 18B.2(e). The map f(z) = z
2
conformally identies the
quarter-plane (A) and the half-plane (B). The mesh of curves in (A) is the preimage of the
Cartesian grid in (B). Note that these curves always intersect at right angles; this is because
f is a conformal map. The solid curves are the streamlines: the preimages of horizontal grid
lines. The streamlines describe a sourceless, irrotational ow conned to the quarter-plane
(see Proposition 18B.6 on page 430). The dashed curves are the equipotential contours: the
preimages of vertical grid lines. The streamlines and equipotentials can be interpreted as the
level curves of two harmonic functions (by Proposition 18A.2). They can also be interpreted
as the voltage contours and eld lines of an electric eld in a quarter-plane bounded by perfect
conductors on the x and y axes. (C,D): Example 18B.4 on the following
page. The map f(z) = z
2
can be used to pull back solutions to BVPs from the half-plane to
the quarter-plane. Figure (C) shows a greyscale plot of the harmonic function H dened on
the quarter-plane by eqn.(18B.2). Figure (D) shows a greyscale plot of the harmonic function
h dened on the half-plane by eqn.(18B.1); the two functions are related by H = h f.
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
i
-3 -2 -1 0 1 2 3
-1
-0.5
0
0.5
1
-1 +1
(A) (B)
Figure 18B.7: (A) Example 18B.2(f). The map f(z) = (z
2
+1)
1/2
from the set C+ \ |yi ; 0 < y < 1 to the upper half-plane C+. (B) Example 18B.2(g).
The map f(z) := i arcsin(z) is a conformal isomorphism from the domain U := |x + yi ;
either y ,= 0 or 1 < x < 1 to a bi-innite horizontal strip. In these gures, as in Figure
18B.6(A,B), the mesh is the preimage of a Cartesian grid on the image domain; the solid
lines are streamlines, and the dashed lines are equipotential contours. In Figure (A), we can
interpret these streamlines as the ow of uid over an obstacle; in Figure (B); they represent
the ow of uid through a narrow aperature between two compartments. Alternately, we can
interpret these curves as the voltage contours and eld lines of an electric eld, where the
domain boundaries are perfect conductors.
conformal isomorphism fromX to Y using holomorphic mappings, by Proposition
18B.1.
Example 18B.4. Let X =
_
(r
1
, r
2
) 1
2
; r
1
, r
2
0
_
be the open upper right
quarter-plane. Suppose we want to nd a harmonic function H : X 1
satisfying the nonhomogeneous Dirichlet boundary conditions H(x) = 1(x)
for all x X, where 1 : X 1 is dened:
1(r
1
, 0) =
_
3 if 1 r
1
2;
0 otherwise.
and 1(0, r
2
) =
_
1 if 3 r
2
4;
0 otherwise.
Identify X with the complex right quarter-plane U from Example 18B.2(e). Let
)(.) := .
2
; then ) is a conformal isomorphism from U to the upper half-plane
C
+
. If we identify C
+
with the real upper half-plane Y :=
_
(j
1
, j
2
) 1
2
; j
2
0
_
,
then we can treat ) as a function ) : X Y, given by the formula )(r
1
, r
2
) =
(r
2
1
r
2
2
, 2r
1
r
2
).
Since ) is bijective, the inverse )
1
: Y X is well-dened. Thus, we can
dene a function / := 1 )
1
: Y 1. To be concrete:
/(j
1
, 0) =
_
_
_
3 if 1 j
1
4;
1 if 16 j
1
9;
0 otherwise.
Now we must nd a harmonic function / : Y 1 satisfying the Dirichlet
boundary conditions /(j
1
, 0) = /(j
1
, 0) for all j
1
1. By adding together two
copies of the solution from Example 17E.2 on page 405, we deduce that
/(j
1
, j
2
) =
3
_
arcsin
_
4 j
1
j
2
_
arcsin
_
1 j
1
j
2
__
(18B.1)
_
arcsin
_
9 j
1
j
2
_
arcsin
_
16 j
1
j
2
__
,
for all (j
1
, j
2
) Y; see Figure 18B.6(D). Finally, dene H := / ) : X 1.
That is,
H(r
1
, r
2
) =
3
_
arcsin
_
4 r
2
1
+r
2
2
2r
1
r
2
_
arcsin
_
1 r
2
1
+r
2
2
2r
1
r
2
__
(18B.2)
_
arcsin
_
9 r
2
1
+r
2
2
2r
1
r
2
_
arcsin
_
16 r
2
1
+r
2
2
2r
1
r
2
__
,
for all (r
1
, r
2
) X; see Figure 18B.6(C). Proposition 18B.3(a) says that H is
harmonic on X, because / is harmonic on Y. Finally, / satises the Dirichlet
boundary conditions specied by /, and 1 = / ); thus Proposition 18B.3(b)
says that H satises the Dirichlet boundary conditions specied by 1, as
desired.
For Proposition 18B.3 to be useful, we must nd a conformal map from our
original domain X to some nice domain Y where we are able to easily solve
BVPs. For example, ideally, Y should be a disk or a half-plane, so that we can
apply the Fourier techniques of Section 14B, or the Poisson kernel methods from
Sections 14B(v), 17F and 17E. If X is a simply connected open subset of the
plane, then a deep result in complex analysis says that it is always possible to
nd such a conformal map. An open subset U C is simply connected if any
closed loop in U can be continuously shrunk down to a point without ever leaving
U. Heuristically speaking, this means that U has no holes. (For example, the
open disk is simply connected, and so is the upper half-plane. However, the open
annulus is not simply connected.)
Theorem 18B.5. Riemann Mapping Theorem
Let U, V C be two open, simply connected regions of the complex plane.
Then there is always a holomorphic bijection ) : U V.
Proof. See [Lan85, Chapter XIV, pp.340-358]. 2
In particular, this means that any simply connected open subset of C is
conformally isomorphic to the disk, and also conformally isomorphic to the upper
half-plane. Thus, in theory, a technique like Example 18B.4 can be applied to
any such region.
Unfortunately, the Riemann Mapping Theorem does not tell you how to con-
struct the conformal isomorphism it merely tells you that such an isomorphism
exists. This is not very useful when we want to solve a specic boundary value
problem on a specic domain. If V is a region bounded by a polygon, and U
is the upper half-plane, then it is possible to construct an explicit conformal
isomorphism from U to V using Schwarz-Christoel transformations; see [Fis99,
3.5, p.227] or [Asm05, 12.6]. For further information about conformal maps in
general, see [Fis99, 3.4], [Lan85, Chapter VII], or the innovatively visual [Nee97,
Chapter 12]. Older, but still highly respected references are [Neh75], [Bie53] and
[Sch79].
Application to uid dynamics. Let U C be an open connected set, and
let

V : U 1
2
be a two-dimensional vector eld (describing a ow). De-
ne ) : U C by )(n) = \
1
(n) i\
2
(n). Recall that Proposition 18A.4 on
page 418 says that

V is sourceless and irrotational (e.g. describing a nontur-
bulent, incompressible uid) if and only if ) is holomorphic. Suppose 1 is a
complex antiderivative
5
of ) on U that is 1 : U C is a holomorphic map
such that 1
t
). Then 1 is called a complex potential for

V. The function
(n) = Re [1(n)] is called the (real) potential of the ow. An equipotential
contour of 1 is a level curve of . That is, it is a set c
x
= n U; Re [1(n)] = r
for some xed r 1. For example, in Figures 18B.6(A) and 18B.7(A,B), the
equipotential contours are the dashed curves. A streamline of 1 is a level curve
of the imaginary part of 1. That is, it is a set o
y
= n U; Im[1(n)] = j
for some xed j 1. For example, in Figures 18B.6(A) and 18B.7(A,B), the
streamlines are the solid curves.
A trajectory of

V is the path followed by a particle carried in the ow
that is, it is a smooth path : (T, T) U (for some T (0, ]) such that
(t) =

V[(t)] for all t (T, T). The ow

V is conned to U if no trajec-
tories of

V ever pass through the boundary U. (Physically, U represents an
impermeable barrier). The equipotentials and streamlines of 1 are important
for understanding the ow dened by

V, because the following result:
Proposition 18B.6. Let

V : U 1
2
be a sourceless, irrotational ow, and
let 1 : U C be a complex potential for

V.
5
We will discuss how to construct complex antiderivatives in Exercise 18C.15 on page 447;
for now, just assume that F exists.
(a) If = Re [1], then =

V. Thus, particles in the ow can be thought
of as descending the potential energy landscape determined by . In
particular, every trajectory of the ow cuts orthogonally through every
equipotential contour of 1.
(b) Every streamline of 1 also cuts orthogonally through every equipotential
contour.
(c) Every trajectory of

V parameterizes a streamline of 1, and every stream-
line can be parameterized by a trajectory. (Thus, by plotting the stream-
lines of 1, we can visualize the ow

V).
(d)

V is conned to U if and only if 1 conformally maps U to a bi-innite
horizontal strip V C, and maps each connected component of U to a
horizontal line in V.
Proof. Exercise 18B.4 Hint: (a) Follows from the denitions of 1 and

V. To E _
prove (b) use the fact that 1 is a conformal map. (c) follows by combining (a) and
(b), and then (d) follows from (c). 2
Thus, the set of conformal mappings from U onto such horizontal strips
describes all possible sourceless, irrotational ows conned to U. If U is simply
connected, then we can assume 1 maps U to the upper half-plane and maps U
to 1 (as in Example 18B.2(e)). Or, if we are willing to allow one point source
(or sink) j in U, we can nd a mapping from U to a bi-innite horizontal strip,
which maps the half of the boundary on one side of j to the top edge of strip,
maps the other half of the boundary to the bottom edge, and maps j itself to
(as in Example 18B.2(a); in this case, the point source is at 0).
Application to electrostatics. Proposition 18B.6 has another important
physical interpretation. The function = Im[)] is harmonic (by Proposi-
tion 18A.2 on page 417). Thus, it can be interpreted as an electrostatic potential
(see Example 1D.2 on page 14). In this case, we can regard the streamlines of
1 as the voltage contours of the resulting electric eld; then the equipotentials
1 of are the eld lines (note the reversal of roles here). If U is a perfect con-
ductor (e.g. a metal), then the eld lines must always intersect U orthogonally,
and the voltage contours (i.e. the streamlines) can never intersect U thus,
in terms of our uid dynamical model, the ow is conned to U. Thus, the
streamlines and equipotentials in Figures 18B.6(A) and 18B.7(A,B) portray the
(two-dimensional) electric eld generated by charged metal plates.
For more about the applications of complex analysis to uid dynamics and
electrostatics, see [Fis99, 4.2, pp.261-278] or [Nee97, 12.V, pp.527-540].
1
V
-1+ i
-1- i
-1 0 0
V
U
(a)
(b)
(c)
1
U
(e)
-1
U
(d)
1 0 -1 1
i
-i
c
r
b a
Figure 18B.8: Exercise 18B.5.
Exercise 18B.5. (a) Let (0, 2], and consider the pie-wedge domain V := E _
: cis ; 0 < : < 1, 0 < < (in polar coordinates); see Figure 18B.8(a). Find a con-
formal isomorphism fromV to the left half-innite rectangle U = r +ji ; r 0, 0 < j < .
(b) Let U := r +ji ; < j < be a bi-innite horizontal strip of width 2, and
let V := r + ji ; either j ,= i or r 1, as shown in Figure 18B.8(b). Show that
)(.) := . + exp(.) is a conformal isomorphism from U to V.
(c) Let C
+
:= r +ji ; j 0 be the upper half-plane. Let U := r + ji; either
j 0 or 1 < r < 1. That is, U is the complex plane with two lower quarter-planes
removed, leaving a narrow chasm in between them, as shown in Figure 18B.8(c). Show
that )(.) =
2
.
2
1 + arcsin(1,.)
_
is a conformal isomorphism from C
+
to U.
(d) Let C
+
:= r +ji ; j 0 be the upper half-plane. Let U := r+ji; either j 0
or r < 1 or 1 < r. That is, U is the complex plane with a vertical half-innite rectangle
removed, as shown in Figure 18B.8(d). Show that )(.) =
2
_
.(1 .
2
)
1/2
+ arcsin(.)
_
is a conformal isomorphism from C
+
to U.
(e) Let c 0, let 0 < : < 1, and let U := r+ji ; r
2
+j
2
< 1 and (rc)
2
+j
2
< :
2
.
That is, U is the o-centre annulus, obtained by removing from the unit disk a smaller
smaller disk of radius : centered at (c, 0), as shown in Figure 18B.8(e). Let o := c :
and / := c +:, and dene
:=
1 +o/
_
(1 o
2
)(1 /
2
)
o +/
and 1 :=
1 o/
_
(1 o
2
)(1 /
2
)
/ o
.
Let A := r+ji ; 1 < r
2
+j
2
< 1 be an annulus with inner radius 1 and outer radius
1, and let )(.) :=
.
1 .
. Show that ) is a conformal isomorphism from U into A.
(f) Let U be the upper half-disk shown on the right side of Figure 18B.4, and let | be
the unit disk. Show that the function )(.) = i
.
2
+ 2i. + 1
.
2
2i. + 1
from U into |.
(g) Let | =
_
r +ji ; r
2
+j
2
< 1
_
be the open unit disk and let C
+
= r +ji ; j 0
be the open upper half-plane. Dene ) : | C
+
by )(.) = i
1+z
1z
. Show that ) is a
conformal isomorphism from | into C
+
.
Exercise 18B.6. (a) Combine Example 18B.2(a) with Proposition 17E.1 on E _
page 404 (or Proposition 20C.1 on page 538) to propose a general method for solving
the Dirichlet problem on the bi-innite strip U = r +ji ; r 1, 0 < j < .
(b) Now combine Exercise (a) with Exercise 18B.5(b) to propose a general method
for solving the Dirichlet problem on the domain portrayed in Figure 18B.8(b). (Note:
despite the fact that the horizontal barriers are lines of zero thickness, your method
allows you to assign dierent boundary conditions to the two sides of these barriers.)
Use your method to nd the equilibrium heat distribution when the two barriers are each
a dierent constant temperature. Reinterpret this solution as the electric eld between
two charged electrodes.
(c) Combine Exercise 18B.5(c) with Proposition 17E.1 on page 404 (or Proposi-
tion 20C.1 on page 538) to propose a general method for solving the Dirichlet problem
on the chasm domain portrayed in Figure 18B.8(c). Use your method to nd the
equilibrium heat distribution when the boundaries on either side of the chasm are two
dierent constant temperatures. Reinterpret this solution as the electric eld near the
edge of a narrow gap between two large, oppositely charged parallel plates.
(d) Combine Exercise 18B.5(d) with Proposition 17E.1 on page 404 (or Proposi-
on the domain portrayed in Figure 18B.8(d). Use your method to nd the equilibrium
heat distribution when the left side of the rectangle has temperature 1, the right side
has temperature +1, and the top has temperature 0.
(e) Combine Exercise 18B.5(e) with Proposition 14B.10 on page 287 to propose a
general method for solving the Dirichlet problem on the o-centre annulus portrayed in
Figure 18B.8(e). Use your method to nd the equilibrium heat distribution when the
inner and outer circles are two dierent constant temperatures. Reinterpret this solution
as an electric eld between two concentric, oppositely charged cylinders.
(f) Combine Exercise 18B.5(f) with the methods of Sections 14B, 14B(v), and/or 17F
to propose a general method for solving the Dirichlet and Neumann problems on the half-
disk portrayed in Figure 18B.4. Use your method to nd the equilibrium temperature
distribution when the semicircular top of the half-disk is one constant temperature, and
the base is another constant temperature.
(g) Combine Exercise 18B.5(g) with the Poisson Integral Formula on a disk (Propo-
sition 14B.11 on page 290 or Proposition 17F.1 on page 407) to obtain another solution
to the Dirichlet problem on a half-plane. Show that this is actually equivalent to the
Poisson Integral Formula on a half-plane (Proposition 17E.1 on page 404).
(h) Combine Example 18B.2(f) with Proposition 17E.1 on page 404 (or Proposi-
on the domain portrayed in Figure 18B.7(a). (Note: despite the fact that the vertical
obstacle is a line of zero thickness, your method allows you to assign dierent boundary
conditions to the two sides of this line.) Use your method to nd the equilibrium tem-
perature distribution when the obstacle has one constant temperature and the the real
line has another constant temperature. Reinterpret this as the electric eld generated
by a charged electrode protruding but insulated from a horizontal, neutrally charged
conducting barrier.
(i) Combine Exercise (a) with Example 18B.2(g) to propose a general method for
solving the Dirichlet problem on the domain portrayed in Figure 18B.7(b). (Note: de-
spite the fact that the horizontal barriers are lines of zero thickness, your method allows
you to assign dierent boundary conditions to the two sides of these barriers.) Use
your method to nd the equilibrium temperature distribution when the two horizon-
tal barriers have dierent constant temperatures. Reinterpret this as the electric eld
between two charged electrodes.
Exercise 18B.7. (a) Figure 18B.6(A) portrays the map )(.) = .
2
from Example E _
18B.2(e). Show that in this case, the equipotential contours are all curves of the form
r +ij; j =

r
2
c for some xed c 0. Show that the streamlines are all curves of
the form r +ij; j = c,r for some xed c 0.
(b) Figure 18B.7(B) portrays the map )(.) = i arcsin(.) from Example 18B.2(f).
Show that in this case, the equipotential contours are all ellipses of the form
_
r +ij ;
r
2
cosh(:)
2
+
j
2
sinh(:)
2
= 1
_
,
for some xed : 1. Likewise, show that the streamlines are all hyperbolas
_
r +ij ;
r
2
sin(:)
2

j
2
cos(:)
2
= 1
_
,
for some xed : 1. Hint: Use Exercises 18A.9(d,e) on page 422.
(c) Find an equation describing all streamlines and equipotentials of the conformal
map in Example 18B.2(a). Sketch the streamlines. (This describes a ow into a large
body of water, from a point source on the boundary).
(d) Fix (, ), and consider the innite wedge-shaped region U = : cis ;
: 0, 0 < < 2 . Find a conformal isomorphism from U to the upper half-plane.
Sketch the streamlines of this map. (This describes the ow near the bank of a wide
river, at a corner where the river bends by angle of ).
(e) Suppose = 2,3. Find an exact equation to describe the streamlines and
equipotentials from question (d) (analogous to the equations j =

r
2
c and j =
c,r from question (a)).
(f) Sketch the streamlines and equipotentials dened by the conformal map in Ex-
ercise 18B.5(b). (This describes the ow out of a long pipe or channel into a large body
of water).
(g) Sketch the streamlines and equipotentials dened by the inverse of the confor-
mal map ) in Exercise 18B.5(c). (In other words, sketch the )-images of vertical and
horizontal lines in C
+
). This describes the ow over a deep chasm in the streambed.
(h) Sketch the streamlines and equipotentials dened by the inverse of the confor-
mal map ) in Exercise 18B.5(d). (In other words, sketch the )-images of vertical and
horizontal lines in C
+
). This describes the ow around a long rectangular peninsula in
an ocean.
18C Contour integrals and Cauchys Theorem
Prerequisites: 18A.
A contour in C is a continuous function : [0, o] C (for some o 0)
such that (0) = (o), and such that does not self-intersect that is, :
[0, o) C is injective.
6
. Let
r
,
i
: [0, o] 1 be the real and imaginary
parts of (so (:) =
r
(:) +
i
(:)i, for all : 1). For any : (0, o), we
6
What we are calling a contour is sometimes called a simple, closed curve.
18C. Contour integrals and Cauchys Theorem 435
-i
1
-1
0
R i
R -R 0
(A)
(B)
(
s
)
=
e
x
p
(
i
s
)
(
s
)

=

i

e
x
p
(
i

s
)
i
Figure 18C.1: (A) The counterclockwise unit circle contour from Example 18C.1. (B)
The D contour from Example 18C.3
dene the (complex) velocity vector if at : by (:) :=
t
r
(:) +
t
i
(:)i (if these
derivatives exist). We say that is smooth if (:) exists for all : (0, o).
Example 18C.1. Dene : [0, 2] C by (:) = exp(i:); then is a
counterclockwise parameterization of the unit circle in the complex plane, as
shown in Figure 18C.1(A). For any : [0, 2], we have (:) = cos(:)+i sin(:),
so that (:) = cos
t
(:) +i sin
t
(:) = sin(:) +i cos(:) = i(:).
Let U C be an open subset, let ) : U C be a complex function, and
let : [0, o] U be a smooth contour. The contour integral of ) along is
dened:
_
) :=
_
S
0
)[(:)] (:) d:.
(Recall that (:) is a complex number, so )[(:)] (:) is a product of two
complex numbers). Another notation we will sometimes use is
_
)(.) d..
Example 18C.2. Let : [0, 2] C be the unit circle contour from Example
18C.1.
(a) Let U := C and let )(.) := 1, a constant function. Then
_
) =
_
2
0
1 (:) d: =
_
2
0
sin(:) +i cos(:) d:
=
_
2
0
sin(:) d: +i
_
2
0
cos(:) d: = 0 +i0 = 0.
(b) Let U := C and let )(.) := .
2
. Then
_
) =
_
2
0
(:)
2
(:) d: =
_
2
0
exp(i:)
2
i exp(i:) d:
= i
_
2
0
exp(i:)
3
d: = i
_
2
0
exp(3i:) d:
= i
_
2
0
cos(3:) +i sin(3:) d: = i
_
2
0
cos(3:) d:
_
2
0
sin(3:) d:
= i0 0 = 0.
(c) More generally, for any n Z except n = 1, we have
_
.
n
d. = 0
(Exercise 18C.1). E _
(What happens if n = 1? See Example 18C.6 below).
(d) It follows that, if c
n
, . . . , c
2
, c
1
, c
0
C, and )(.) = c
n
.
n
+ +c
2
.
2
+c
1
.+c
0
is a complex polynomial function, then
_
) = 0.
A contour : [0, o] U C is piecewise smooth if (:) exists for all :
[0, o], except for perhaps nitely many points 0 = :
0
:
1
:
2
. . . :
N
= o.
If ) : U C is a complex function, we dene the contour integral
_
) :=
N
n=1
_
sn
s
n1
)[(:)] (:) d:.
Example 18C.3. Fix 1 0, and dene
R
: [0, + 21] C as follows:
R
(:) :=
_
1 exp(i:) if 0 : ;
: 1 if : + 21.
(18C.1)
This contour looks like a D turned on its side; see Figure 18C.1(B). The
rst half of the contour parameterizes the upper half of the circle from 1 to
1. The second half parameterizes a straight horizontal line segment from
1 back to 1. It follows that

R
(:) :=
_
1i exp(i:) if 0 : ;
1 if : + 21.
(18C.2)
(a) Let U := C and let )(.) := .. Then
_
R
) =
_

0
(:) (:) d: +
_
+2R
(:) (:) d:
()
_

0
1 exp(i:) 1i exp(i:) d: +
_
+2R
(: 1) d:
= 1
2
i
_

0
exp(i:)
2
d: +
_
R
R
t dt
= 1
2
i
_

0
cos(2:) +i sin(2:) d: +
t
2
2
t=R
t=R
=
1
2
i
2
_
sin(2:) i cos(2:)
_
s=
s=0
+
1
2
_
1
2
(1)
2
_
=
1
2
i
2
_
(0 0) i(1 1)
_
+ 0
= 0 + 0 = 0.
Here, () is by equations (18C.1) and (18C.2).
(b) For generally, for any n Z, if n ,= 1, then
_
R
.
n
d. = 0 (Exercise 18C.2). E _
Thus, if ) is any complex polynomial, then
_
R
) = 0.
Any contour : [0, o] C cuts the complex plane into exactly two pieces.
Formally the set C [0, o] has exactly two connected components, and ex-
actly one of these components (the one inside ) is bounded.
7
The bounded
component is called the purview of ; see Figure 18C.2(A). For example, the
purview of the unit circle is the unit disk. If G is the purview of , then clearly
G = [0, o]. We say that is counterclockwise if the outward normal vector
of G is always on the righthand side of the vector . We say is clockwise if
the outward normal vector of G is always on the lefthand side of the vector ;
see Figure 18C.2(C).
The contour is called nullhomotopic in U if the purview of is entirely
contained in U; see Figure 18C.2(B). Equivalently: it is possible to continuously
shrink down to a point without the any part of the contour leaving U; this is
called a nullhomotopy of , and is portrayed in see Figure 18C.2(D). Heuris-
tically speaking, is nullhomotopic in U if and only if does not encircle any
holes in the domain U.
Example 18C.4. (a) The unit circle from Examples 18C.1 and the D contour
from Example 18C.3 are both counterclockwise, and both are nullhomotopic
in the domain U = C.
(b) The unit circle is not nullhomotopic on the domain C
:= C 0. The
purview of (the unit disk) is not entirely contained in C
, because the point 0

is missing. Equivalently, it is not possible to shrink down to a point without
passing the curve through 0 at some moment; at this moment the curve would
not be contained in U.
The zero outcomes of Examples 18C.2 and 18C.3 not accidents; they are
consequences of one of the fundamental results of complex analysis.
7
This seemingly innocent statement is actually the content of the Jordan Curve Theorem,
which is a surprisingly dicult and deep result in planar topology.
(B)
(D)
U
U
(A)
(C)
Figure 18C.2: (A) Three contours and their purviews. (B) Contour is nullhomotopic
in U, but contours and are not nullhomotopic in U. (C) Contour is clockwise;
contours and are counterclockwise. (D) A nullhomotopy of .
Theorem 18C.5. (Cauchys Theorem)
Let U C be an open subset, and let ) : U C be holomorphic on U. If
: [0, o] U is a contour which is nullhomotopic in U, then
_
) = 0.
Proof. Let G be the purview of . If is nullhomotopic in U, then G U and
parameterizes the boundary G. Treat U as a subset of 1
2
. Let )
r
: U 1
and )
i
: U 1 be the real and imaginary parts of ). The function ) can
be expressed as a vector eld

V : U 1
2
dened by \
1
(n) := )
r
(n) and
\
2
(n) := )
i
(n). For any b G, let

N[b] denote the outward unit normal
vector to G at b. We dene
Flux(
V, ) :=
_
S
0
V[(:)]
N[(:)] d:, and Work(
V, ) :=
_
S
0
V[(:)] (:)] d:.

The rst integral is the ux of

V across the boundary of G; this is just a
reformulation of equation (0E.1) on page 562 (see Figure 0E.1(B) on page
561). The second integral is the work of

V along the contour .
Claim 1: (a) Re
__
)
_
= Work(
V, ) and Im
__
)
_
= Flux(
V, ).
(b) If div (
V) 0, then Flux(
V, ) = 0.
(c) If curl (
V) 0, then Work(
V, ) = 0.
Proof. (a) is Exercise 18C.3. (b) is Greens Theorem (Theorem 0E.3 on E _
page 562). (c) is Exercise 18C.4 (Hint: its a variant of Greens Theorem). E _
Claim 1
Now, if ) is holomorphic on U, then Proposition 18A.4 on page 418 says that
div (
V) 0 and curl (
V) 0. Then Claim 1 implies

_
) = 0.
(For other proofs, see [Fis99, Theorem 1, 2.3, p.107], [Lan85, IV.3, p.137],
or [Nee97, 8.X, p.410]). 2
At this point you are wondering: what are complex contour integrals good
for, if they are always equal to zero? The answer is that
_
) is only zero if the

function ) is holomorphic in the purview of . If ) has a singularity inside this
purview (i.e. a point where ) is not complex-dierentiable, or perhaps not even
dened), then
_
) might be nonzero.
Example 18C.6. Let : [0, 2] C be the unit circle contour from Example
18C.1. Let C
:= C 0, and dene ) : C
C by )(.) := 1,.. Then

_
) =
_
2
0
(:)
(:)
d: =
_
2
0
i exp(i:)
exp(i:)
d: =
_
2
0
i d: = 2i.
Notice that is not nullhomotopic on C
. Of course, we could extend ) to all

of C by dening )(0) in some arbitrary way. But no matter how we do this, )
will never be complex-dierentiable at zero in other words, 0 is a singularity
of ).
If the purview of contains one or more singularities of ), then the value of
_
) reveals important information about these singularities. Indeed, the value

of
_
) depends only on the singularities within the purview of , and not on the
shape of itself. This is a consequence of the homotopy-invariance of contour
integration.
Let U C be an open subset, and let
0
,
1
: [0, o] U be two contours.
We say that
0
is homotopic to
1
in U if
0
can be continuously deformed
into
1
without ever moving outside of U; see Figure 18C.3. (In particular, is
0
is homotopic to
1
, but not to or .
1
is homotopic to
0
, but not to or .
is not homotopic to
0
,
1
, or . Likewise, is not homotopic to
0
,
1
, or .
1
A homotopy from
0
to
1
.
(A) (B)
U U
Figure 18C.3: Homotopy
nullhomotopic if is homotopic to a constant path in U.) Formally, this means
there is a continuous function : [0, 1] [0, o] U such that:
For all : [0, o], (0, :) =
0
(:).
For all : [0, o], (1, :) =
1
(:).
For all t [0, 1], if we x t and dene
t
: [0, o] U by
t
(:) := (t, :)
for all : [0, o], then
t
is a contour in U.
The function is called a homotopy of
0
into
1
. See Figure 18C.4(A).
Proposition 18C.7. (Homotopy invariance of contour integration)
Let U C be an open subset, and let ) : U C be a holomorphic function.
Let
0
,
1
: [0, o] U be two contours. If
0
is homotopic to
1
in U, then
_
0
) =
_
1
).
Before proving this result, it will be useful to somewhat extend our deni-
tion of contour integration. A chain is a piecewise-continuous, piecewise dif-
ferentiable function : [0, o] C (for some o 0). (Thus, a chain is
a contour if is continuous, (o) = (0), and is not self-intersecting). If
: [0, o] U C is a chain, and ) : U C is a complex-valued function,
then the integral of ) along is dened
) =
_
S
0
)[(:)] (:) d:. (18C.3)
(0,0) (0,S)
(1,S)
(1,0)
0
(0)
1
(0)
(A)
(D)
(B)

(C)
Figure 18C.4: (A) is a homotopy from 0 to 1. (B) The reversal
of . (C)
The linking . (D) The contour
dened by the boundary of the homotopy map

from Figure (A).
Here we dene (:) = 0 whenever : is one of the (nitely many) points where
is nondierentiable or discontinuous. (Thus, if is a contour, then (18C.3) is
just the contour integral
_
)).
The reversal of chain is the chain

: [0, o] C dened by

(:) :=
(o :); see Figure 18C.4(B). If : [0, o] C and : [0, T] C are two
chains, then the linking of and is the chain : [0, o +T] C dened
(:) :=
_
(:) if 0 : o;
(: o) if o : o +T.
(Figure 18C.4(C))
Lemma 18C.8. Let U C be an open set and let ) : U C be a complex-
valued function. Let : [0, o] U be a chain.
(a)
_
) =
_
).
(b) If : [0, T] C is another chain, then
_

) =
_
) +
_
).
(c) is continuous if and only if and are both continuous, and (0) =
(o).
(d) The linking operation is associative: that is, if is another chain, then
( ) = ( ).
(Thus, we normally drop the brackets and just write ).
Proof of Proposition 18C.7. Dene the continuous path : [0, 1] C by
(t) := (0, t) = (o, t), for all t [0, 1].
Figure 18C.4(A) shows how traces the path dened by the homotopy
from
0
(o) (=
0
(0)) to
1
(o) (=
1
(0)). We assert (without proof) that the
homotopy can always be chosen such that is piecewise smooth; thus we
regard as a chain. Figure 18C.4(D) portrays the contour
:=
0

, which traces the -image of the four sides of the rectangle [0, 1] [0, 2].
Claim 1:
is nullhomotopic in U.
Proof. The purview of
is simply the image of the open rectangle (0, 1)

(0, 2) under the function . But by denition, maps (0, 1) (0, 2) into
U; thus the purview of
is contained in U, so
is nullhomotopic in U.
Claim 1
Thus, 0
(C)
_
)
()
_
)
()
_
0
) +
_
)
_
1
)
_
) =
_
0
)
_
1
).
Here (C) is by Cauchys Theorem and Claim 1, () is by denition of , and
() is by Lemma 18C.8(a,b).
Thus, we have
_
0
)
_
1
) = 0, which means
_
0
) =
_
1
), as claimed.
2
Example 18C.6 is a special case of the following important result:
Theorem 18C.9. (Cauchys Integral Formula)
Let U C be an open subset, let ) : U C be holomorphic, let n U, and
let : [0, o] U be any counterclockwise contour whose purview contains n
and is contained in U. Then )(n) =
1
2i
_
)(.)
. n
d..
In other words: if U
:= U n, and we dene 1
u
: U
C by 1
u
(.) :=
)(.)
. n
for all . U
, then )(n) =
1
2i
_
1
u
.
Proof. For simplicity, we will prove this in the case n = 0. We must show that
1
2i
_
)(.)
.
d. = )(0).
Let G U be the purview of . For any : 0, let |
r
be the disk of radius
: around 0, and let
r
be a counterclockwise parameterization of |
r
(e.g.
r
(:) := :c
is
for all : [0, 2]). Let U
:= U 0.
Claim 1: If : 0 is small enough, then |
r
G. In this case, is homotopic
to
r
in U
.
Claim 1
E _
Now, dene : U C as follows.
(n) :=
)(.) )(0)
.
for all . U
, and (0) := )
t
(0).
Then is holomorphic on U
. Observe that
1
()
1
2i
_
r
1
.
d.
()
1
2i
_
1
.
d.. (18C.4)
where () is by Example 18C.6 on page 439, and () is by Claim 1 and Propo-
sition 18C.7. Thus,
)(0)
()
)(0)
2i
_
1
.
d. =
1
2i
_
)(0)
.
d..
Thus,
1
2i
_
)(.)
.
d. )(0) =
1
2i
_
)(.)
.
d.
1
2i
_
)(0)
.
d.
=
1
2i
_
)(.)
.

)(0)
.
d.
=
1
2i
_
)(.) )(0)
.
d. =
1
2i
_
()
1
2i
_
r
, for any : 0.
Here, () is by eqn.(18C.4), and () is again by Claim 1 and Proposition 18C.7.
Thus, we have
1
2i
_
)(.)
.
d. )(0) = lim
r0
1
2i
_
r
. (18C.5)
Thus, it suces to show that lim
r0
_
r
= 0. To see this, rst note that
is continuous at 0 (because lim
z0
(.) = )
t
(0) by denition of the derivative),
and is also continuous on the rest of U (where is just another holomorphic
function). Thus, is bounded on G (because G is a bounded set whose closure
is inside U). Thus, if ` := sup
zG
[(.)[, then ` < . But then
_
r

()
` length(
r
) = ` 2:
r0
0,
where () is by Lemma 18C.10 (below).
Thus, lim
r0
_
r
= 0, so eqn. (18C.5) implies that
1
2i
_
)(.)
.
d. = )(0),
as desired. 2
If : [0, o] C is a chain, then we dene length() :=
_
S
0
[ (:)[ d:. The
proof of Theorem 18C.9 invoked the following useful lemma.
Lemma 18C.10. Let ) : U C and let be a chain in U. If ` :=
sup
uU
[)(n)[, then
` length().
Exercise 18C.8. Prove the general case of Theorem 18C.9, for an arbitrary n C. E _
Corollary 18C.11. (Mean Value Theorem for holomorphic functions)

Let U C be an open set and let ) : U C be holomorphic. Let : 0 be
small enough that the circle S(:) of radius : around n is contained in U. Then
)(n) =
1
2
_
S(r)
)(:) d: =
1
2
_
2
0
)(n +: c
i
) d.
Proof. Dene : [0, 2] U by (:) := n+:c
is
for all : [0, 2]; thus, is a
counterclockwise parameterization of S(:), and (:) = i:c
is
for all : [0, 2].
Then
)(n)
()
1
2i
_
)(.)
. n
d. =
1
2i
_
2
0
)[()]
() n
() d
=
1
2i
_
2
0
)(n +: c
i
)
:c
i
i:c
i
d =
1
2
_
2
0
)(n +: c
i
) d,
as desired. Here () is by Cauchys Integral Formula. 2
Exercise 18C.9. Using Proposition 18C.11, derive another proof of the Mean Value E _
Theorem for harmonic functions on U (Theorem 1E.1 on page 16). (Hint: Use Proposi-
tion 18A.3 on page 417).
Corollary 18C.12. (Maximum Modulus Principle)
Let U C be an open set and let ) : U C be holomorphic. Then the
function :(.) := [)(.)[ has no local maxima inside U.
Proof. Exercise 18C.10 Hint: Use the Mean Value Theorem. 2 E _
Exercise 18C.11. Let ) : U C be holomorphic. Show that the functions E _
1(.) := Re [)(.)] and 1(.) := Im[)(.)] have no local maxima or minima inside U.
Let
o
| := . C ; [.[ < 1 be the open unit disk in the complex plane, and
let S := | be the unit circle. The Poisson kernel for | is the function
T : S
o
| 1 dened by
T(:, n) :=
1 [n[
2
[: n[
2
, for all : S and n
o
|.
Corollary 18C.13. (Poisson Integral Formula for holomorphic functions)
Let U C be an open subset containing the unit disk
o
|, and let ) : U C
be holomorphic. Then for all n
o
|,
)(n) =
1
2
_
S
)(:) T(:, n) d: =
1
2
_
2
0
)(c
i
) T(c
i
, n) d.
Proof. If n
o
|, then n
1
is outside
o
| (because [n
1
[ = [n[
1
= [n[
1
1 if
[n[ < 1). Thus, the set C
u
:= C n
1
contains
o
|. Fix n
o
| and dene
the function p
u
: C
u
C by
p
u
(.) :=
)(.) n
1 n.
.
Claim 1: p
u
is holomorphic on C
u
.
Claim 1
E _
Now, dene 1
u
: U C by 1
u
(.) :=
)(.)
. n
, and let : [0, 2] S be the
unit circle contour from Example 18C.1 (i.e. (:) = c
is
for all : [0, 2]).
Then
)(n) =
1
2i
_
1
u
by Cauchys Integral Formula (Theorem 18C.9),
and 0 =
1
2i
_
p
u
by Cauchys Theorem (Theorem 18C.5),
Thus, )(n) =
1
2i
_
(1
u
+p
u
) =
1
2i
_
2
0
_
1
u
[()] +p
u
[()]
_
() d
=
1
2i
_
2
0
_
)(c
i
)
c
i
n
+
)(c
i
) n
1 nc
i
_
ic
i
d
=
1
2
_
2
0
)(c
i
)
_
c
i
c
i
n
+
c
i
n
1 nc
i
_
d
()
1
2
_
2
0
)(c
i
)
1 [n[
2
[c
i
n[
2
d =
1
2
_
2
0
)(c
i
)T(c
i
, n) d,
as desired. Here, () uses the fact that, for any : S and n C,
:
: n
+
:n
1 n:
=
:
: n
+
::n
: n::
=
:
: n
+
[:[
2
n
: n[:[
2
()
:
: n
+
n
: n
=
: (: n) +n (: n)
(: n) (: n)
=
[:[
2
:n +n: [n[
2
(: n) (: n)
=
[:[
2
[n[
2
[: n[
2 ()
1 [n[
2
[: n[
2
.
where both () are because [:[ = 1. 2
Exercise 18C.13. Using Corollary 18C.13, derive yet another proof of the Poisson E _
Integral Formula for harmonic functions on |. (See Proposition 14B.11 on page 290,
and also Proposition 17F.1 on page 407.) Hint: Use Proposition 18A.3 on page 417.
At this point, we have proved the Poisson Integral Formula three entirely
dierent ways: using Fourier series (Proposition 14B.11), using impulse-response
methods (Proposition 17F.1), and now, using complex analysis (Corollary 18C.13).
In 18F on page 461 below, we will encounter the Poisson Integral Formula yet
again, while studying the Abel mean of a Fourier series.
An equation which expresses the solution to a boundary value problem in
terms of an integral over the boundary of the domain is called an integral repre-
sentation formula. For example, Poisson Integral Formula is such a formula, as
is Poissons solution to the Dirichlet problem on a half-space (Proposition 17E.1
on page 404). Cauchys Integral Formula provides an integral representation
formula for any holomorphic function on any domain in C which is bounded by
a contour. Our proof of Corollary 18C.13 shows how this can be used to obtain
integral representation formulae for harmonic functions on planar domains.
Exercise 18C.14. (Liouvilles Theorem) E _
Suppose ) : C C is holomorphic and bounded i.e. there is some ` 0 such
that [)(.)[ < ` for all . C. Show that ) must be a constant function.
Hint. Dene p(.) :=
f(z)f(0)
z
.
(a) Show that p is holomorphic on C.
(b) Show that [p(.)[ < 2`,[.[ for all . C.
(c) Let . C. Let be a circle of radius 1 0 around 0, where 1 is large enough
that . is in the purview of . Use Cauchys Integral Formula and Lemma 18C.10 on
page 444 (below) to show that [p(.)[ <
1
2
2`
1
21
1 [.[
. Now let 1 .
Exercise 18C.15. (Complex antiderivatives) E _
Let U C be an open connected set. We say that U is simply connected if
every contour in U is nullhomotopic. Heuristically speaking, this means U doesnt have
any holes. For any n
0
, n
1
U, a path in U from n
0
to n
1
is a continuous function
: [0, o] U such that (0) = n
0
and (o) = n
1
.
Let ) : U C be holomorphic. Pick a basepoint / U, and dene a function
1 : U C as follows.
For all n U, 1(n) :=
_
), where is any path in U from / to n. (18C.6)

(a) Show that 1(n) is well-dened by expression (18C.6), independent of the path
you use to get from / to n.
(Hint. If
1
and
2
are two paths from / to n, show that
1
2
is a contour.
Then apply Cauchys Theorem).
(b) For any n
1
, n
2
U, show that 1(n
2
) 1(n
1
) =
_
), where is any path in U

from n
1
to n
2
.
(c) Show that 1 is a holomorphic function, and 1
(n) = )(n) for all n U.

(Hint. Write 1
(n) as the limit (18A.1) on page 415. For any c close to n let
: [0, 1] U be the straight-line path linking n to c (i.e. (:) = :c +(1 :)n.).
Deduce from part (b) that
1(c) 1(n)
c n
=
1
c n

_
). Now take the limit as

c n.)
The function 1 is called a complex antiderivative of ), based at /. Part (c) is
the complex version of the Fundamental Theorem of Calculus.
(d) Let U = C, and let / U, and let )(n) = exp(n). Let 1 be the complex
antiderivative of ) based at /. Show that 1(n) = exp(n) exp(/) for all n C.
(e) Let U = C, and let / U, and let )(n) = n
n
for some n N. Let 1 be the
complex antiderivative of ) based at /. Show that 1(n) =
1
n+1
(n
n+1
/
n+1
), for
all n C.
We already encountered one application of complex antiderivatives in Proposition 18B.6
on page 430. The next two exercises describe another important application.
Exercise 18C.16. Complex logarithms (follows Exercise 18C.15). E _
(a) Let U C be an open, simply connected set which does not contain 0. Dene a
complex logarithm function log : U C as the complex antiderivative of 1,.
based at 1. That is, log(n) :=
_
1,. d., where is any path in U from 1 to .. Show

that log is a right-inverse of the exponential function that is, exp(log(n)) = n
for all n U.
(b) What goes wrong with part (a) if 0 U? What goes wrong if 0 , U, but U
contains an annulus which encircles 0? (Hint. Consider Example 18C.6)
Remark: This is the reason why we required U to be simply connected in Exercise
18C.15.
(c) Suppose our denition of complex logarithm is any right-inverse of the complex
exponential function that is, any holomorphic function 1 : U C such that
exp(1(n)) = n for all n U. Suppose 1
0
: U C is one such logarithm function
(dened as in part (a), for example). Dene 1
1
: U C by 1
1
(n) = 1
0
(n)+2i.
Show that 1
1
is also a logarithm. Relate this to the problem you found in part
(b).
(d) Indeed, for any n Z, dene 1
n
: U C by 1
n
(n) = 1
0
(n) + 2ni. Show that
1
n
is also a logarithm in the sense of part (c). Make a sketch of the surface
described by the functions Im[1
n
] : C 1, for all n Z at once.
(e) Proposition 18A.2 on page 417 asserted that any harmonic function on a convex
domain U 1
2
can be represented as the real part of a holomorphic function on
U, treated as a subset of C. The Remark following Proposition 18A.2 said that U
actually doesnt need to be convex, but it does need to be simply connected. We
will not prove that simple-connectedness is sucient, but we can now show that
it is necessary.
18D. Analyticity of holomorphic maps 449
Consider the harmonic function /(r, j) = log(r
2
+j
2
) dened on 1
2
0. Show
that, on any simply connected subset U C
, there is a holomorphic function

1 : U C with / = Re [1]. However, show that there is no holomorphic function
1 : C
C with / = Re [1].
The functions 1
n
(for n Z) are called the branches of the complex logarithm
. This exercise shows that the complex logarithm is a much more complicated object
than the real logarithm indeed, the complex log is best understood as a holomorphic
multifunction which takes countably many distinct values at each point in C
. The
surface in part (d) is an abstract representation of the graph of this multifunction it
is called a Riemann surface .
Exercise 18C.17. Complex root functions (follows Exercise 18C.16). E _
(a) Let U C be an open, simply connected set which does not contain 0, and let
log : U C be any complex logarithm function, as dened in Exercise 18C.16.
Fix n N. Show that exp(n log(n)) = n
n
for all n N.
(b) Fix n N and now dene
n
: U C by
n
n = exp(log(n),n) for all n N.

Show that
n
is a complex nth root function. That is, (

n
n)
n
= n for all n U.
Dierent branches of logarithm dene dierent branches of the nth root function.
However, while there are innitely many distinct branches of logarithm, there are exactly
n distinct branches of the nth root function.
(c) Fix n N, and consider the equation .
n
= 1. Show that the set of all solutions to
this equation is Z
n
:= 1, c
2i/n
, c
4i/n
, c
6i/n
, . . . c
2(n1)i/n
. (These numbers
are called the nth roots of unity). For example, Z
2
= 1 and Z
4
= 1, i.
(d) Suppose :
1
: U C and :
2
: U C are two branches of the square root
function (dened by applying the denition in part (b) to dierent branches of
the logarithm). Show that :
1
(n) = :
2
(n) for all n U.
Sketch the Riemann surface for the complex square root function.
(e) More generally, let n 2, and suppose :
1
: U C and :
2
: U C are two
branches of the nth root function (dened by applying the denition in part (b)
to dierent branches of the logarithm). Show that there is some Z
n
(the set
of nth roots of unity from part (c)) such that :
1
(n) = :
2
(n) for all n U.
Bonus: Sketch the Riemann surface for the complex nth root function. (Note that it is
not possible to embed this surface in three dimensions without some self-intersection).
18D Analyticity of holomorphic maps

Prerequisites: 18C, 0H(ii).
In 18A, we said that the holomorphic functions formed a very special sub-
class within the set of all (real)-dierentiable functions on the plane. One in-
dication of this was Proposition 18A.2 on page 417. Another indication is the
following surprising and important result.
Theorem 18D.1. (Holomorphic Analytic)
Let U C be an open subset. If ) : U C is holomorphic on U, then
) is innitely (complex-)dierentiable everywhere in U. Thus, the functions
)
t
, )
tt
, )
ttt
, . . . are also holomorphic on U. Finally, for all n U, the (complex)
Taylor series of ) at n converges uniformly to ) in open disk around n.
Proof. Since any analytic function is (
, it suces to prove the last sentence,

and the rest of the theorem follows. Suppose 0 U; we will prove that ) is
analytic at n = 0 (the general case n ,= 0 is similar).
Let be a counterclockwise circular contour in U centered at 0 (e.g. dene
: [0, 2] U by (:) = :c
is
for some : 0). Let W U be the purview of
(an open disk centered at 0). For all n N, let
c
n
:=
1
2i
_
)(.)
.
n+1
d..
We will show that the power series
n=0
c
n
n
n
converges to ) for all n W. For
any n W, we have
)(n)
()
1
2i
_
)(.)
. n
d.
()
1
2i
_
)(.)
.

n=0
_
n
.
_
n
d.
=
1
2i
_
n=0
)(.)
.
n+1
n
n
d.
()
n=0
_
1
2i
_
)(.)
.
n+1
d.
_
n
n
()
n=0
c
n
n
n
,
as desired. Here, () is Cauchys Integral Formula (Theorem 18C.9 on page 443),
and () is by the denition of c
n
. Step () is because
1
. n
=
_
1
.
_
_
1
1
w
z
_
=
1
.

n=0
_
n
.
_
n
. (18D.1)
Here, the last step is the geometric series expansion
1
1 r
=
n=1
r
n
(with
r := n,.), which is valid because [n,.[ < 1 because [n[ < [.[ because n is
inside the disk W and . is a point on the boundary of W.
It remains to justify step (). For any N, observe that
1
2i
_
n=0
)(.)
.
n+1
n
n
d. =
N
n=0
_
1
2i
_
)(.)
.
n+1
d.
_
n
n
+
1
2i
_
n=N+1
)(.)
.
n+1
n
n
d.
(18D.2)
Thus, to justify (), it suces to show that the second term on the right hand
side of (18D.2) tends to zero as . Let 1 be the length of (i.e. 1 = 2:
if describes a circle of radius :). The function . )(.),. is continuous on
the boundary of W, so it is bounded. Let ` := sup
zW
)(.)
.
. Fix c 0 and
nd some N such that
n=N
_
n
.
_
n
<
c
1`
. (Such an exists because
the geometric series (18D.1) converges because [n,.[ < 1.) It follows that:
For all . W,
)(.)
.

n=N
_
n
.
_
n
< `
c
1`
=
c
1
. (18D.3)
Thus
n=N+1
)(.)
.
n+1
n
n
d.
)(.)
.

n=N+1
_
n
.
_
n
d.

()
c
1
1 = c.
Here, () is by equation (18D.3) above and Lemma 18C.10 on page 444. This
works for any c 0, so we conclude that the second term on the right side of
(18D.2) tends to zero as . This justies step (), which completes the
proof. 2
Corollary 18D.2. (Case 1 = 2 of Proposition 1E.5 on page 18)
Let U 1
2
be open. If / : U 1 is a harmonic function, then / is analytic
on U.
Proof. Exercise 18D.1 Hint. Combine Theorem 18D.1 with Proposition 18A.3 E _
on page 417. Note that this is not quite as trivial as it sounds: you must show how to
translate the (complex) Taylor series of a holomorphic function on C into the (real)
Taylor series of a real-valued function on 1
2
. 2
Because of Theorem 18D.1 and Proposition 18A.5(h), holomorphic functions
are also called complex-analytic functions (or even simply analytic functions)
in some books. Analytic functions are extremely rigid: for any n U, the
behaviour of ) in a tiny neighbourhood around n determines the structure of )
everywhere on U, as we now explain. Recall that a subset U C is connected if
it is not possible to write U as a union of two nonempty disjoint open subsets. A
subset X C is perfect if, for every r X, every open neighbourhood around r
contains other points in X besides r. (Equivalently: every point in X is a cluster
point of X). In particular, any open subset of C is perfect. Also, 1 and are a
perfect subsets of C. Any disk, annulus, line segment, or unbroken curve in C is
both connected and perfect.
Theorem 18D.3. (Identity Theorem)
Let U C be a connected open set, and let ) : U C and p : U C be
two holomorphic functions.
(a) Suppose there is some o U such that )(o) = p(o), )
t
(o) = p
t
(o), )
tt
(o) =
p
tt
(o), and in general, )
(n)
(o) = p
(n)
(o) for all n N. Then )(n) = p(n)
for all n U.
(b) Suppose there is a perfect subset X U such that )(r) = p(r) for all
r X. Then )(n) = p(n) for all n U.
Proof. (a) Let / := ) p. It suces to show that / 0. Let
W :=
_
n U ; /(n) = 0, and /
(n)
(n) = 0 for all n N
_
.
The set W is nonempty, because o W by hypothesis. We will show that
W= U; it follows that / 0.
Claim 1: W is an open subset of U.
Proof. Let n W; we must show that there is a nonempty open disk around
n that is also in W. Now, / is analytic at n because ) and p are analytic at
n. Thus, there is some nonempty open disk | centered at n such that the
Taylor expansion of / converges to /(.) for all . |. The Taylor expansion
of / at n is c
0
+c
1
(.n)+c
2
(.n)
2
+c
3
(.n)
3
+ , where c
n
:= /
(n)
(n),n!,
for all n N. But for all n N, c
n
= 0 because /
(n)
(n) = 0 because n W.
Thus, the Taylor expansion is 0+0(.n)+0(.n)
2
+ ; hence it converges
to zero. Thus, / is equal to the constant zero function on |. Thus, | W.
This holds for any n W; hence W is an open subset of C.
Claim 1
Claim 2: W is a closed subset of U.
Proof. For all n N, the function /
(n)
: U C is continuous (because
)
(n)
and p
(n)
are continuous, since they are dierentiable). Thus, the set
W
n
:=
_
n U ; /
(n)
(n) = 0
_
is a closed subset of U (because 0 is a closed
subset of C). But clearly W= W
0
W
1
W
2
. Thus, W is also closed
(because it is an intersection of closed sets).
Claim 2
Thus, Wnonempty, and is both open and closed in U. Thus, the set V := UW
is also open and closed in U, and U = V . W. If V ,= , then we have
expressed U as a union of two nonempty disjoint open sets, which contradicts
the hypothesis that U is connected. Thus, V = , which means W= U. Thus,
/ 0. Thus, ) p.
(b) Fix r X, and let r
n
n=1
X be a sequence converging to r (which
exists because X is perfect).
Claim 3: )
t
(r) = p
t
(r).
Proof. Exercise 18D.2 Hint: Compute the limit (18A.1) on page 415 along E _
the sequence r
n
n=1
.
Claim 3
This argument works for any r X; thus )
t
(r) = p
t
(r) for all r X. Repeat-
ing the same argument, we get )
tt
(r) = p
tt
(r) for all r X. By induction,
)
(n)
(r) = p
(n)
(r) for all n N and all r X. But then part (a) implies that
) p. 2
Remark. The Identity Theorem is true (with pretty much the same proof)
for any 1
M
-valued, analytic function on any connected open subset U 1
N
,
for any , ` N. (Exercise 18D.3 Verify this.) In particular, the Identity E _
Theorem holds for any harmonic functions dened on a connected open subset of
1
N
(for any N). This result nicely complements Corollary 5D.4 on page 87,
which established the uniqueness of the harmonic function which satises spec-
ied boundary conditions. (Note that neither Corollary 5D.4 nor the Identity
Theorem for harmonic functions is a special case of the other; they apply to
distinct situations.)
In Proposition 18A.5(i) and Example 18A.6 on pages 419 and 420, we showed
how the standard real-analytic functions on 1 can be extended to holomorphic
functions on C in a natural way. We now show that these are the only holo-
morphic extensions of these functions. In other words, there is a one-to-one
relationship between real-analytic functions and their holomorphic extensions.
Corollary 18D.4. (Analytic extension)
Let X 1 be an open subset, and let ) : X 1 be an analytic function.
There exists some open subset U C containing X, and a unique holomorphic
function 1 : U C such that 1(r) = )(r) for all r X.
Proof. For any r X, Proposition 18A.5(i) says that the (real) Taylor series of
) around r can be extended to dene a holomorphic function 1
x
: |
x
C,
where |
x
C is an open disk centered at r. Let U :=
_
xX
|
x
; then U is an
open subset of C containing X. We would like to dene 1 : U C as follows:
1(n) := 1
x
(n), for any r X and n |
x
. (18D.4)
But theres a problem: what if n |
x
and also n |
y
for some r, j X. We
must make sure that 1
x
(n) = 1
y
(n) otherwise 1 will not be well-dened by
equation (18D.4).
So, let r, j X, and suppose the disks |
x
and |
y
overlap. Then P := X
|
x
|
y
is a nonempty open subset of 1 (hence perfect). The functions 1
x
and 1
y
both agree with ) on P; thus, they agree with each other on P. Thus,
the Identity Theorem says that 1
x
and 1
y
agree everywhere on |
x
|
y
.
Thus, 1 is well-dened by equation (18D.4). By construction, 1 is a holomor-
phic function on U which extends ). Furthermore, 1 is the only holomorphic
extension of ), by the Identity Theorem. 2
Exercise 18D.4. Let 1 : 1 1 1 be any real-analytic function, and sup- E _
pose 1(sin(:), cos(:)) = 0 for all : 1. Use the Identity Theorem to show that
1(sin(c), cos(c)) = 0 for all c C.
Conclude that any algebraic relation between sin and cos (i.e. any trigonometric
identity) which is true on 1 will also be true over all of C.
18E Fourier series as Laurent series
Prerequisites: 18D, 8D. Recommended: 10D(ii).
For any : 0, let
o
|(:) := . C ; [.[ < : be the open disk of radius :
around 0, and let
o
|
(:) := . C ; [.[ : be the open codisk of coradius

:. Let S(:) := . C ; [.[ = : be the circle of radius :; then
o
|(:) = S(:) =
o
|
(:). Finally, for any 1 : 0, let

o
A(:, 1) := . C ; : < [.[ < 1 be the
open annulus of inner radius : and outer radius 1.
Let c
0
, c
1
, c
2
, . . . be complex numbers, and consider the complex-valued power
series:
n=0
c
n
.
n
= c
0
+ c
1
. + c
2
.
2
+ c
3
.
3
+ c
4
.
4
+ (18E.1)
For any coecients c
n
n=0
, there is some radius of convergence 1 0
(possibly zero) such that the power series (18E.1) converges uniformly on the
open disk
o
|(1) and diverges for all .
o
|
(1). (The series (18E.1) may or

may not converge on the boundary circle S(1)). The series (18E.1) then denes
a holomorphic function on
o
|(1). (Exercise 18E.1 Prove the preceding three E _
sentences.) Conversely, if U C is any open set containing 0, and ) : U C
is holomorphic, then Theorem 18D.1 on page 450 says that ) has a power series
expansion like (18E.1) which converges to ) in a disk of nonzero radius around
0.
Next, let c
0
, c
1
, c
2
, c
3
. . . be complex numbers, and consider the complex-
valued inverse power series
0
n=
c
n
.
n
= c
0
+
c
1
.
+
c
2
.
2
+
c
3
.
3
+
c
4
.
4
+ (18E.2)
18E. Fourier series as Laurent series 455
For any coecients c
n
0
n=
, there is some coradius of convergence : 0
(possibly innity) such that the inverse power series (18E.2) converges uniformly
on the open codisk
o
|
(:) and diverges for all .

o
|(:). (The series (18E.2) may
or may not converge on the boundary circle S(:)). The series (18E.2) then denes
a holomorphic function on
o
|
(:). Conversely, if U C is any open set, then

we say that U is a neighbourhood of innity if U contains
o
|
(:) for some

: < . If ) : U C is holomorphic, and lim
z
)(.) exists and is nite, then
) has a inverse power series expansion like (18E.2) which converges to ) in a
codisk of nite coradius (i.e. a nonempty open disk around innity).
8
Exercise 18E.2. Prove all statements in the paragraph above. Hint: Consider the E _
change of variables n := 1,.. Now use the results about power series from the paragraph
between equations (18E.1) and (18E.2).
Finally, let . . . , c
2
, c
1
, c
0
, c
1
, c
2
, . . . be complex numbers, and consider the
complex-valued Laurent series:
n=
c
n
.
n
= +
c
2
.
2
+
c
1
.
+ c
0
+ c
1
. + c
2
.
2
+ c
3
.
3
+ (18E.3)
For any coecients c
n
n=
, there exist 0 : 1 such that the
Laurent series (18E.3) converges uniformly on the open annulus
9 o
A(:, 1) and
diverges for all .
o
|(:) and all .
o
|
(1). (The series (18E.3) may or may

not converge on the boundary circles S(:) and S(1).) The series (18E.3) then
denes a holomorphic function on
o
A(:, 1).
Exercise 18E.3. Prove all statements in the paragraph above. Hint: Combine E _
the results about power series and inverse power series from the from the paragraphs
between equations (18E.1) and (18E.3).
Proposition 18E.1. Let 0 : < 1 , and suppose the Laurent series
(18E.3) converges on
o
A(:, 1) to dene the function ) :
o
A(:, 1) C. Let be
a counterclockwise contour in
o
A(:, 1) which encircles the origin (for example,
could be a counterclockwise circle of radius :
0
, where : < :
0
< 1). Then for all
n Z,
c
n
=
1
2i
_
)(.)
.
n+1
d..
8
This is not merely fanciful terminology; see Remark 18G.4 on page 469.
9
Note that
o
A(r, R) = if r = R.
Proof. For all .
o
A(:, 1), we have )(.) =
k=
c
k
.
k
. Thus, for any n Z,
)(.)
.
n+1
=
1
.
n+1
k=
c
k
.
k
=
k=
c
k
.
kn1
()
m=
c
m+n+1
.
m
,
where () is the change of variables : := / n 1, so that / = :+n +1. In
other words,
)(.)
.
n+1
= +
c
n1
.
2
+
c
n
.
+ c
n+1
+ c
n+2
. + c
n+3
.
2
+ (18E.4)
Thus,
_
)(.)
.
n+1 ()
+
_
c
n1
.
2
+
_
c
n
.
+
_
c
n+1
+
_
c
n+2
. +
_
c
n+3
.
2
+
()
+ 0 + 2i c
n
+ 0 + 0 + 0 +
= 2i c
n
, as desired.
Here, () is because the series (18E.4) converges uniformly on
o
A(:, 1) (be-
cause the Laurent series (18E.3) converges uniformly on
o
A(:, 1)); thus, Propo-
sition 6E.10(b) on page 127 implies we can compute the contour integral of
series (18E.4) term-by-term. Next, () is by Examples 18C.2(c) and 18C.6
(pages 435 and 439). 2
Laurent series are closely related to Fourier series.
Proposition 18E.2. Suppose 0 : < 1 < 1 and suppose the Laurent series
(18E.3) converges to the function ) :
o
A(:, 1) C. Dene p : [, ] C by
p(r) := )(c
ix
) for all r [, ]. For all n Z, let
p
n
:=
1
2
_

p(r) exp(nir) dr (18E.5)

be the nth complex Fourier coecient of p (see 8D on page 172). Then:
(a) p
n
= c
n
for all n Z.
(b) For any r [, ], if . = c
ix
S(1), then for all N, the th partial
Fourier sum of p(r) equals the th partial Laurent sum of )(.); that is:
N
n=N
p
n
exp(nir) =
N
n=N
c
n
.
n
.
Thus, the Fourier Series for p converges on [, ] in exactly the same
ways (i.e. uniformly, in 1
2
, etc.), and at exactly the same speed, as the
Laurent series for ) converges on S(1).
Proof. (a) As in Example 18C.1 on page 435, dene the unit circle contour
: [0, 2] C by (:) := exp(i:) for all : [0, 2]. Then
c
n
()
1
2i
_
)(.)
.
n+1
=
1
2i
_
2
0
)[(:)])
(:)
n+1
(:) d:
=
1
2i
_
2
0
)(c
is
)
c
is(n+1)
ic
is
d: =
1
2
_
2
0
)(c
is
) c
nis
d:
()
1
2
_

)(c
ix
) c
nix
dr =
1
2
_

p(r) c
nix
dr
()
p
n
.
Here, () is by Proposition 18E.1, and () is because the function : c
is
is
2-periodic. Finally, () is by equation (18E.5).
(b) follows immediately from (a), because if . = c
ix
, then for all n Z we
have .
n
= exp(nir). 2
We can also reverse this logic: given the Fourier series for a function p :
[, ] C, we can interpret it as the Laurent series of some hypothetical
function ) dened on an open annulus in C (which may or may not contain
S(1)); then by studying ) and its Laurent series, we can draw conclusions about
p and its Fourier series.
Let p : [, ] C be some function, let p
n
n=
be its Fourier co-
ecients, as dened by equation (18E.5), and consider the complex Fourier
series
10
n=
p
n
E
n
. If p L
2
[, ], then the Riemann-Lebesgue Lemma
(Corollary 10A.3 on page 197) says that lim
n
p
n
= 0; however, the sequence
p
n
n=
might converge to zero very slowly, if p is nondierentiable and/or
discontinuous. We would like the sequence p
n
n=
to converges to zero as
quickly as possible, for two reasons:
1. The faster the sequence p
n
n=
converges to zero, the easier it will be
to approximate the function p using a partial sum of the form
N
n=N
p
n
E
n
,
for some N. (This is important for numerical computations.)
2. The faster the sequence p
n
n=
converges to zero, the more computa-
tions we can perform with p by formally manipulating its Fourier series.
For example, if p
n
n=
converges to zero faster than
1
n
k
, then we can
compute the derivatives p
t
, p
tt
, p
ttt
, . . . , p
(k1)
by formally dierentiating
the Fourier series for p (see 8B(iv) on page 168). This is necessary to
verify the Fourier series solutions to I/BVPs which we constructed in
Chapters 11- 14.
10
For all n Z, recall that En(x) := exp(nix) for all x [, ].
We say the sequence p
n
n=
has exponential decay if there is some o 1
such that
lim
n
o
n
[ p
n
[ = 0 and lim
n
o
n
[ p
n
[ = 0.
This is an extremely rapid decay rate, which causes the partial sum
N
n=N
p
n
E
n
to uniformly converge to p very quickly as . This means we can formally
dierentiate the Fourier series of p as many times as we want. In particular, any
formal solution to an I/BVP which we obtain through such formal dierentiation
will converge to the correct solution.
Suppose p : [, ] C has real and imaginary parts p
r
, p
i
: [, ] 1
(so that p(r) = p
r
(r) + p
i
(r)i for all r [, ]). We say that p is analytic
and periodic if the functions p
r
and p
i
are (real)-analytic everywhere on [, ],
and if we have p() = p(), p
t
() = p
t
(), p
tt
() = p
tt
(), etc. (where
p
t
(r) = p
t
r
(r) +p
t
i
(r)i, etc.).
Proposition 18E.3. Let p : [, ] C have complex Fourier coecients
p
n
n=
. Then
_
p is analytic and periodic
_

_
The sequence p
n
n=
decays exponentially
_
.
Proof. = Dene the function ) : S(1) C by )(c
ix
) := p(r) for all
r [, ].
Claim 1: ) can be extended to a holomorphic function 1 : A(:, 1) C,
for some 0 : < 1 < 1 .
Proof. Let p : 1 C be the 2-periodic extension of p (i.e. p(r+2n) :=
p(r) for all r [, ] and n Z). Then p is analytic on 1, so Corol-
lary 18D.4 on page 453 says that there is an open subset U C containing
1 and a holomorphic function G : U C which agrees with p on 1. With-
out loss of generality, we can assume that U is a horizontal strip of width 2\
around 1, for some \ 0 that is, U = r +ji ; r 1, j (\, \).
Claim 1.1: G is horizontally 2-periodic (i.e. G(n + 2) = 1(n) for all
n U).
Proof. Exercise 18E.4 Hint: Use the Identity Theorem 18D.3 on page 452, E _
and the fact that p is 2-periodic.
Claim 1.1
Dene 1 : C C by 1(.) := exp(i.); thus, 1 maps 1 to the unit circle
S(1). Let : := c
W
and 1 := c
W
; then : < 1 < 1. Then 1 maps the
strip U to the open annulus A(:, 1) C. Note that 1 is horizontally 2-
periodic (i.e. 1(n + 2) = 1(n) for all n U). Dene 1 : A C by
1(1(n)) := G(n) for all n U.
Claim 1.2: 1 is well-dened on A(:, 1).
Proof. Exercise 18E.5 Hint: use the fact that both G and 1 are 2-periodic. E _
Claim 1
Claim 1.3: 1 is holomorphic on A(:, 1).
Proof. Let o A(:, 1); we must show that 1 is dierentiable at o.
Suppose o = G(n) for some n U. There are open sets V U (containing
n) and B A(:, 1) (containing o) such that 1 : V B is bijective. Let
1 : B V be a local inverse of 1. Then 1 is holomorphic on V by
Proposition 18A.5(k) on page 419 (because 1
t
() ,= 0 for all V). But
by denition, 1(/) = G(1(/)) for all / B; Thus, 1 is holomorphic on B
by Proposition 18A.5(j) (the chain rule).
This argument works for any o A(:, 1); thus, 1 is holomorphic on
A(:, 1).
Claim 1.3
It remains to show that 1 is an extension of ). But by denition, )(1(r)) =
p(r) for all r [, ]. Since G is an extension of p, and 1 1 = G, it
follows that 1 is an extension of ).
Claim 1
Let c
n
n=
be the Laurent coecients of 1. Then Proposition 18E.2 on
page 456 says that c
n
= p
n
for all n Z. However, the Laurent series (18E.3)
of 1 (on page 455) converges on A(:, 1). Thus, if [.[ < 1, then the power
series (18E.1) on page 454 converges absolutely at .. This means that, if
1 < o < 1, then
n=0
o
n
[ p
n
[ is nite. Thus, lim
n
o
n
[ p
n
[ = 0.
Likewise, if : < [.[, then the inverse power series (18E.2) on page 454 converges
absolutely at .. This means that, if 1 < o < 1,:, then
n=0
o
n
[ p
n
[ is nite.
Thus, lim
n
o
n
[ p
n
[ = 0.
= Dene c
n
:= p
n
for all n Z, and consider the resulting Laurent series
(18E.3). Suppose there is some o 1 such that
lim
n
o
n
[c
n
[ = 0 and lim
n
o
n
[c
n
[ = 0. (18E.6)
Claim 2: Let : := 1,o and 1 := o. For all . A(:, 1), the Laurent series
(18E.3) converges absolutely at ..
Proof. Let .
+
:= .,o; then [.
+
[ < 1 because [.[ < 1 := o. Also, let
.
:= 1,o.; then [.
[ < 1, because [.[ : := 1,o. Thus,
n=1
[.
[
n
< and
n=0
[.
+
[
n
< . (18E.7)
Evaluating the Laurent series (18E.3) at ., we see that
n=
c
n
.
n
=
n=1
c
n
.
n
+
n=0
c
n
.
n
=
n=1
o
n
c
n
(o.)
n
+
n=0
o
n
c
n
(.,o)
n
=
n=1
(o
n
c
n
) .
n
n=0
o
n
c
n
.
n
+
.
Thus,
n=
[c
n
.
n
[
n=1
o
n
[c
n
[ [.
[
n
+
n=0
o
n
[c
n
[ [.
+
[
n
<
()
,
where () is by equations (18E.6) and (18E.7).
Claim 2
Claim 2 implies that the Laurent series (18E.3) converges to some holomorphic
function ) : A(:, 1) C. But p(r) = )(c
ix
) for all r [, ]; thus, p
is (real)-analytic on [, ], because ) is (complex-)analytic on A(:, 1), by
Theorem 18D.1 on page 450. 2
Exercise 18E.6. Let ) : [, ] 1, and consider the real Fourier series for ) E _
(see 8A on page 161). Show that the real Fourier coecients
n
n=0
and 1
n
n=0
have exponential decay if and only if ) is analytic and periodic on [, ]. (Hint: Use
Proposition 8D.2 on page 174.)
Exercise 18E.7. Let ) : [0, ] 1, and consider the Fourier sine series and cosine E _
series for ) (see 7A(i) on page 137 and 7A(ii) on page 141).
(a) Show that the Fourier cosine coecients
n
n=0
have exponential decay if and
only if ) is analytic on [0, ] and )
(0) = 0 = )
(), and )
(n)
(0) = 0 = )
(n)
() for all
odd n N.
(b) Show that the Fourier sine coecients 1
n
n=1
have exponential decay if and
only if ) is analytic on [0, ] and )(0) = 0 = )(), and )
(n)
(0) = 0 = )
(n)
() for all
even n N.
(c) Conclude that if both the sine and cosine series have exponential decay, then
) 0.
(Hint. Use the previous exercise and Proposition 8C.5 on page 171.)
Exercise 18E.8. Let X = [0, 1] be an interval, let ) L
2
(X) be some initial tem- E _
perature distribution, and let 1 : X 1
1 be the solution to the one-dimensional

heat equation (
t
1 =
2
x
1) with initial conditions 1(r; 0) = )(r) for all r X, and
satisfying either homogeneous Dirichlet boundary conditions, or homogeneous Neumann
boundary conditions, or periodic boundary conditions on X, for all t 0. Show that, for
any xed t 0, the function 1
t
(r) := 1(r, t) is analytic on X. (Hint: Apply Propositions
11A.1 and 11A.3 on pages 225 and 227.)
This shows how the action of the heat equation can rapidly smooth even a highly
irregular initial condition.
18F.
Abel means and Poisson kernels 461

Exercise 18E.9. Compute the complex Fourier series of ) : [, ] C when ) E _
is dened as follows:
(a) )(r) = sin(c
ix
).
(b) )(r) = cos(c
3ix
).
(c) )(r) = c
2ix
cos(c
3ix
).
(d) )(r) = (5 +c
2ix
) cos(c
3ix
).
(e) )(r) =
1
c
2ix
4
.
(f) )(r) =
c
ix
c
2ix
4
.
Exercise 18E.10. (a) Show that the Laurent series (18E.3) can be written in the E _
form 1
+
(.) +1
(1,.), where 1
+
and 1
are both power series.

(b) Suppose 1
+
has radius of convergence 1
+
, and 1
has radius of convergence

1
. Let 1 := 1
+
and : := 1,1
, and show that the Laurent series (18E.3) converges

on A(:, 1).
18F
Abel means and Poisson kernels

Prerequisites: 18E. Prerequisites (for proofs): 10D(ii).
Theorem 18E.3 showed that, if p : [, ] C is analytic, then its Fourier
series

n=
p
n
E
n
will converge uniformly and extremely quickly to p. At
the opposite extreme, if p is not even dierentiable, then

n=
p
n
E
n
might
not converge uniformly, or even pointwise, to p. To address this problem, we
introduce the Abel mean. For any : < 1, the :th Abel mean of the Fourier
series
n=
p
n
E
n
is dened:
A
r
[p] :=
n=
:
[n[
p
n
E
n
.
As : 1, each summand :
[n[
p
n
E
n
in the Abel mean converges to the corre-
sponding summand p
n
E
n
in the Fourier series for p. Thus, we expect that A
r
[p]
should converge to p as : 1. The goal of this section is to verify this intuition.
For any : [0, 1), we dene the Poisson kernel P
r
: [2, 2] 1 by
P
r
(r) :=
1 :
2
1 2: cos(r) +:
2
, for all r [2, 2].
(See Figure 18F.1). Note that P
r
is 2-periodic (i.e. P
r
(r +2) = P
r
(r) for all
r [2, 0]). For any function p : [, ] C, the convolution of P
r
and p
0
5
10
15
20
-3 -2 -1 0 1 2 3
f(0.7,x)
0
5
10
15
20
-3 -2 -1 0 1 2 3
f(0.8,x)
0
5
10
15
20
-3 -2 -1 0 1 2 3
f(0.9,x)
P
0.7
P
0.8
P
0.9
Figure 18F.1: The Poisson kernels P0.7, P0.8, and P0.9, plotted on interval [, ]. Note
the increasing concentration of Pr near x = 0 as r , 1. (In the terminology of Section 10D(ii),
the system |Pr0<r<1 is like an approximation of the identity.)
is the function P
r
p : [, ] C dened by
P
r
p(r) :=
1
2
_

p(j) P
r
The next result tells us that lim
r,1
A
r
[p](r) = p(r), whenever the function p is
continuous at r. Furthermore, for all : < 1, the functions A
r
[p] : [, ] C
are extremely smooth, and two-variable function G(r, :) := A
r
[p](r) is also
extremely smooth.
Proposition 18F.1. Let p L
2
[, ].
(a) For any : [0, 1) and r [, ], P
r
p(r) = A
r
[p](r).
(b) For any r (, ), if p is continuous at r, then lim
r,1
A
r
[p](r) = p(r).
(c) Let | be the closed unit disk, and dene ) : | C by
)(: c
i
) :=
_
P
r
p() if : < 1,
p() if : = 1,
for all [, ] and : [0, 1].
Then ) is holomorphic on
o
|.
(d) Thus, for any xed : < 1, the function A
r
[p] : [, ] C is analytic.
(e) Let (, ) and let : = c
i
S. If p is continuous in a neighbourhood
of , then ) is continuous at : i.e. lim
us
uD
)(n) = )(:).
(f ) If p is continuous on [, ] and p() = p(), then ) is continuous on |.
18F.
Abel means and Poisson kernels 463

Proof. (a) Exercise 18F.1 (Hint: Use Lemmas 18F.2 and 18F.3 below). E _
(b) is Exercise 18F.2 (Hint: Use Lemma 18F.4 below, and Proposition 10D.9(b) E _
on page 219).
(e) and (f ) follow immediately from (b), while (d) follows from (c).
(c) is Exercise 18F.3 Hint: (i) Let T : S
o
| 1 be the Poisson kernel E _
dened on page 445. For any : S and n
o
|, if : = c
iy
and n = : c
ix
, show that
T(:, n) = P
r
(r j).
(ii) Use this to show that P
r
p(r) =
1
2
_
2
0
p(j) T(c
iy
, n) dj.
(ii) Now apply the Poisson Integral Formula for holomorphic functions (Corollary
18C.13). 2
To prove parts (a) and (b) of Proposition 18F.1, we require the following
three lemmas.
Lemma 18F.2. Fix : [0, 1). Then
P
r
(r) = 1 + 2
n=1
:
n
cos(nr) =
n=
:
[n[
exp(inr).
Thus, if
P
n
r
n=
are the complex Fourier coecients of the Poisson kernel
P
r
, then

P
n
r
= :
[n[
for all n Z.
For any ), p : [, ] C, recall the denition of the convolution ) p from
10D(ii) on page 214. The passage from a function to its Fourier coecients
converts the convolution operator into multiplication, as follows:
11
Lemma 18F.3. Let ), p L
2
[, ] and suppose / = ) p L
2
[, ] also.
Then for all n Z, we have

/
n
=

)
n
p
n
.
11
For the corresponding result for Fourier transforms of functions on 1, see Theorem 19B.2(b)
on page 494.
Lemma 18F.4. The set of Poisson kernels P
r
0r<1
is an approximation
of identity, in the following sense:
(AI1)
1
2
_

P
r
(r) dr = 1 for all : [0, 1).
r,1
1
2
_

P
r
(r) dr = 1. (See Figure 10D.2 on
page 218).
Exercise 18F.7. For all N, the th Dirichlet kernel is the function D
N
: E _
[, ] 1 dened by
D
N
(r) := 1 + 2
N
n=1
cos(nr) (see Figure 10B.1 on page 198).
(a) Show that D
N
(r) =
N
n=N
exp(nri).
(b) Let p : [, ] C have complex Fourier series
n=
p
n
E
n
. Use Lemma 18F.3
to show that D
N
p =
N
n=N
p
n
E
n
. (Compare this with Lemma 10B.1).
18G Poles and the residue theorem
Prerequisites: 18D.
Let U C be an open subset, let j U, and let U
:= U j. Let ) :
U
C be a holomorphic function. We say that j is an isolated singularity

of ), because ) is well-dened and holomorphic for all points near j, but not at
j itself.
Now, it might be possible to extend ) to a holomorphic function ) : U C
by dening )(j) in some suitable way. In this case, we say that j is a removable
singularity of ); it is merely a point we forgot when dening ) on U
. However,
sometimes there is no way to dene )(j) such that the resulting function ) :
U C is complex-dierentiable (or even continuous) at j; in this case, we say
that j is an indelible singularity. In this section, we will be concerned with a
particularly nice class of indelible singularities, called poles.
Dene 1
1
: U
C by 1
1
(n) = (j n) )(n). We say that j is a simple
pole of ) if j is a removable singularity of 1
1
i.e. if 1
1
(j) can be dened
18G. Poles and the residue theorem 465
such that 1
1
is complex-dierentiable at j. Now, 1
1
is already holomorphic on
U
(because it is a product of two holomorphic functions ) and . (. n)).

Thus, if 1
1
is dierentiable at j, then 1
1
is holomorphic on all of U. Then
Theorem 18D.1 on page 450 says that 1
1
is analytic at j i.e. 1
1
has a Taylor
expansion near j:
1
1
(.) = o
0
+ o
1
(. j) + o
2
(. j)
2
+ o
3
(. j)
3
+ o
4
(. j)
4
+
Thus,
)(.) =
1
1
(.)
. j
=
o
0
. j
+ o
1
+ o
2
(. j) + o
3
(. j)
2
+ o
4
(. j)
3
+
This expression is called a Laurent expansion (of order 1) for ) at the pole j.
The coecient o
0
is called the residue of ) at the pole j, and denoted res
p
()).
But suppose j is not a simple pole (i.e. it is not a removable singularity for
1
1
). Let n N, and dene 1
n
: U
C by 1
n
(n) = (j n)
n
)(n). We say
that j is a pole if there is some n N such that j is a removable singularity
of 1
n
i.e. if 1
n
(j) can be dened such that 1
n
is complex-dierentiable at j.
The smallest value of n for which this is true is called the order of the pole j.
Now, 1
n
is already holomorphic on U
. Thus, if 1
n
is dierentiable at j, then
1
n
is holomorphic on all of U. Again, Theorem 18D.1 says that 1
n
is analytic
at j, with Taylor expansion
1
n
(.) =
o
0
+ o
1
(. j) + + o
n1
(. j)
n1
+ o
n
(. j)
n
+ o
n+1
(. j)
n+1
+
Thus,
)(.) =
1
n
(.)
(. j)
n
=
o
0
(. j)
n
+
o
1
(. j)
n1
+ +
o
n1
(. j)
+ o
n
+ o
n+1
(. j) +
This expression is called a Laurent expansion (of order n) for ) at the pole
j. The coecient o
n1
is called the residue of ) at the pole j, and denoted
res
p
()).
Let

C := C . , where the symbol represents a point at innity.
If ) : U
C has a pole at j, then it is easy to check that lim

zp
[)(.)[ =
(Exercise 18G.1). Thus, it is natural and convenient to extend ) to a function E _
) : U

C by dening )(j) = . (Later, in Remark 18G.4 on page 469, we
will explain why this is not merely a cute notational device, but is actually the
correct thing to do). The extended function ) : U

C is called a meromorphic
function.
Formally, if U C is an open set, then a function ) : U

C is meromor-
phic if there is a discrete subset P U (possibly empty) such that, if U
:= UP,
then:
1. ) : U
C is holomorphic.
2. Every j P is a pole of ) (hence lim
zp
[)(.)[ = ).
3. )(j) = for all j P.
Example 18G.1. (a) Any holomorphic function is meromorphic, since it has no
poles.
(b) Let ) : U C be holomorphic, let j U, and dene 1 : U

C by
1(.) = )(.),(. j). Then 1 is meromorphic on U, with a single pole at j,
and res
p
(1) = )(j).
(c) Fix j 0, and dene /
y
: C

C by
/
y
(.) =
j
(.
2
+j
2
)
=
j
(. +ji)(. ji)
, for all . C.
Then /
y
is meromorphic on C, with simple poles at . = ji. Observe that
/
y
(.) =
)
+
(.)
(. ji)
, where )
+
(.) :=
j
(. +ji)
, for all . C.
Note that )
+
is holomorphic near ji, so Example (b) says that res
yi
(/
y
) =
)
+
(ji) =
j
(2ji)
=
1
2i
. Likewise,
/
y
(.) =
)
(.)
(. +ji)
, where )
(.) :=
j
(. ji)
, for all . C.
Note that )
is holomorphic near ji, so Example (b) says that res

yi
(/
y
) =
)
(ji) =
j
(2ji)
=
1
2i
.
(d) More generally, let 1(.) = (. j
1
)
n
1
(. j
2
)
n
2
(. j
J
)
n
J
be any com-
plex polynomial with roots j
1
, . . . , j
n
C. Let ) : C C be any other
holomorphic function (e.g. another polynomial), and dene 1 : C

C by
1(.) = )(.),1(.) for all . C. Then 1 is a meromorphic function, whose
poles are located at j
1
, j
2
, . . . , j
J
. For any , [1...J], dene 1
j
(.) :=
)(.),(. j
1
)
n
1
(. j
j1
)
n
j1
(. j
j+1
)
n
j+1
(. j
J
)
n
J
. Then
res
p
j
(1) =
1
(n
j
1)
j
(j
j
)
(n
j
1)!
(i.e. the (n
j
1)th term in the Taylor expansion of 1
j
at j
j
). In particular, if
n
j
= 1 (i.e. j
j
is a simple pole), then res
p
j
(1) = 1
j
(j
j
).
(e) Let p : U

C be meromorphic and let j U. Suppose p has a simple
pole at j. If ) : U C is holomorphic, and )(j) ,= 0, then the function ) p
is meromorphic, with a pole at j, and res
p
() p) = )(j) res
p
(p).
Exercise 18G.2 Verify Examples (d) and (e). E _
We now come to one of the most important results in complex analysis.
U
p
0
p
3
p
1
p
2
U
p
0
0
p
3
3
p
1
1
p
2
U
p
0
0
x
0
y
0
p
3
3
x
3
p
1
1
x
1
y
1
p
2
2
x
2 y
2
U
p
0
p
3
p
1
p
2
(A) (B)
(C) (D)
3
y
3
Figure 18G.1: (A) The hypotheses of the Residue Theorem. (B) For all j [0...J], j is
a small, counterclockwise circular contour around the pole pj. (C) The paths 0, . . . , J,
0, . . . , J, and 1, . . . , J. (D) The chain is a contour homotopic to
Theorem 18G.2. (Residue Theorem)
Let U C be an open, simply-connected subset of the plane. Let ) : U

C
be meromorphic on U. Let : [0, o] U be a counterclockwise contour
which is nullhomotopic in U, and suppose the purview of contains the poles
j
0
, j
1
, . . . , j
J
U, and no other poles, as shown in Figure 18G.1(A). Then
_
) = 2i
J
j=0
res
p
j
()).
Proof. For all , [0...J], let
j
: [0, 2] U be a small, counterclockwise
circular contour around the pole j
j
, as shown in Figure 18G.1(B).
Claim 1: For all , [0...J],
_
j
) = res
p
j
()).
Proof. Suppose ) has the following Laurent expansion around j
j
:
f(z) =
an
(z pj)
n
+
a1n
(z pj)
n1
+ +
a1
(z pj)
+ a0 + a1(z pj) +
This series converges uniformly, so
_
j
) can integrated term-by-term to get:
_
j
an
(z pj)
n
+
_
j
a1n
(z pj)
n1
+ +
_
j
a1
(z pj)
+
_
j
a0 +
_
j
a1(z pj) +
()
0 + 0 + + a1 2i + 0 + 0 +
= 2i o
1
()
2i res
p
j
()).
Here, () is by Examples 18C.2(c) and 18C.6 on pages 435 and 439. Mean-
while, () is because o
1
= res
p
j
()) by denition.
Claim 1
Figure 18G.1(C) portrays the smooth paths
j
: [0, ] U and
j
: [0, ]
U dened by
j
(:) :=
j
(:) and
j
(:) :=
j
(: +), for all : [0, ].
That is:
j
and
j
parameterize the rst half and the second half of
j
,
respectively, so that
j
=
j

j
. (18G.1)
For all , [0...J], let r
j
:=
j
(0) =
j
(). and let j
j
:=
j
() =
j
(0). For
all , [1...J], let
j
: [0, 1] U be a smooth path from j
j1
to r
j
. For all
i [0...J], we can assume that
j
is drawn so as not to intersect
i
or
i
, and
for all i [1...J], i ,= ,, we can likewise assume that
j
does not intersect
i
.
Figure 18G.1(D) portrays the chain
:=
0

1

1

2

J

J

J

J1
J1

3

2
2

1
1

0
.
(18G.2)
The chain is actually a contour, by Lemma 18C.8(c) on page 441.
Claim 2: is homotopic to in U.
Proof. Exercise 18G.3 (Not as easy as it looks)
Claim 2
E _
Thus,
_
)
()
_
)
()
0
) +
_
1
) +
_
1
) +
_
2
) + +
_
J
) +
_
J
)
_
J
) +
_
J1
)

_
J1
) +
_
3
) +
_
2
)
_
2
) +
_
1
)
_
1
) +
_
0
).
=
_
0
) +
_
1
) + +
_
J1
) +
_
J
) +
_
J1
) + +
_
1
) +
_
0
).
(@)
0

0
) +
_
1

1
) + +
_
J1

J1
) +
_
J
)
()
_
0
) +
_
1
) + +
_
J1
) +
_
J
)
()
2i res
p
0
()) + 2i res
p
1
()) + + 2i res
p
J1
()) + 2i res
p
J
()),
as desired. Here, () is by Claim 2 and Proposition 18C.7 on page 440. ()
is by eqn.(18G.2) and Lemma 18C.8(a,b) on page 441, and (@) is again by
Lemma 18C.8(b). Finally, () is by eqn.(18G.1), and () is by Claim 1. 2
Example 18G.3. (a) Suppose ) is holomorphic inside the purview of . Then
it has no poles, so the residue-sum in the Residue Theorem is zero. Thus,
we get
_
) = 0, in agreement with Cauchys Theorem (Theorem 18C.5 on

page 438).
(b) Suppose )(.) = 1,., and encircles 0. Then ) has exactly one pole in the
purview of (namely, at 0), and res
0
()) = 1 (because the Laurent expansion
of ) is just 1,.). Thus, we get
_
) = 2i, in agreement with Example 18C.6

on page 439.
(c) Suppose ) is holomorphic inside the purview of . Let j be in the purview
of and dene 1(.) :=
f(z)
zp
. Then 1 has exactly one pole in the purview
of (namely, at j), and res
p
(1) = )(.), by Example 18G.1(b). Thus, we
get
_
1 = 2i )(.), in agreement with Cauchys Integral Formula (Theo-

rem 18C.9 on page 443).
Remark 18G.4: (The Riemann Sphere) Earlier we introduced the notational
convention of dening )(j) = whenever j was a pole of a holomorphic function
0
1 2 3
3i
2i
i
4i
5i
6i
4 5 6
-6 -5 -4 -3 -2 -1
-5i
-6i
-4i
-3i
-2i
-i
0
o
-i
-2 i
-3 i
-4 i
6
1
2
3
4
5
7
8
9
10
11
12
13
14
15
16 17
18
-1
-2
-3
-4
-5
-7
-8
9
-10
-11
-12
-13
-14
-15
-16-17
18
-6
C
C
o
Figure 18G.2: The identication of the complex plane C with the Riemann sphere

C.
) : U j C, thereby extending ) to a meromorphic function ) : U

C,
where

C = C . . We will now explain how this cute notation is actually
quite sensible. The Riemann sphere is the topological space

C constructed by
taking the complex plane C and adding a point at innity, denoted by .
An open set U C is considered a neighbourhood of if there is some : 0
such that U contains the codisk |
(:) := c C ; [c[ :. See Figure 18G.2.

Now, let U
:= U j and suppose ) : U
C is a continuous function
with a singularity at j. Suppose we dene )(j) = , thereby extending ) to
a function ) : U

C. If lim
zp
[)(j)[ = (e.g. if j is a pole of )), then this
extended function will be continuous at j, with respect to the topology of the
Riemann sphere. In particular, any meromorphic function ) : U

C is is a
continuous mapping from U into

C.
If ) : C

C is meromorphic, and 1 := lim
c
)(c) is well-dened, then
we can extend ) to a continuous function ) :

C

C by dening )() :=
1. We can then even dene the complex derivatives of ) at ; ) eectively
becomes a complex-dierentiable transformation of the entire Riemann sphere.
Many of ideas in complex analysis are best understood by regarding meromorphic
functions in this way.
Remark. Not all indelible singularities are poles. Suppose j is a singularity of
), and the Laurent expansion of ) at j has an innite number of negative-power
terms (i.e. it looks like the Laurent series (18E.3) on page 455). Then j is called
a essential singularity of ). The Casorati-Weierstrass Theorem says that, if
B is any open neighbourhood of j, however tiny, then the image )[B] is dense in
C. In other words, the value of )(.) wildly oscillates all over the complex plane
innitely often as . j. This is a much more pathological behaviour than a
pole, where we simply have )(.) as . j.
Exercise 18G.4. Let )(.) = exp(1,.). E _
(a) Show that ) has an essential singularity at 0.
(b) Verify the conclusion of the Casorati-Weierstrass Theorem for this function. In
fact, show that, for any c 0, if |(c) is the disk of radius c around 0, then )[|(c)] =
C 0.
Exercise 18G.5. For each of the following functions, nd all poles and compute E _
the residue at each pole. Then use the Residue Theorem to compute the contour integral
along a counterclockwise circle of radius 1.8 around the origin.
(a) )(.) =
1
.
4
+ 1
. (Hint: .
4
+1 = (. c
i/4
)(. c
3i/4
)(. c
5i/4
)(. c
7i/4
).)
(b) )(.) =
.
3
1
.
4
+ 5.
2
+ 4
. (Hint: .
4
+ 5.
2
+ 4 = (.
2
+ 1)(.
2
+ 4).)
(c) )(.) =
.
4
.
6
+ 14.
4
+ 49.
2
+ 36
. (Hint: .
6
+ 14.
4
+ 49.
2
+ 36 = (.
2
+ 1)(.
2
+
4)(.
2
+ 9).)
(d) )(.) =
. +i
.
4
+ 5.
2
+ 4
. (Careful!)
(e) )(.) = tan(.) = sin(.), cos(.).
(f) )(.) = tanh(.) = sinh(.), cosh(.).
Exercise 18G.6. (For algebraists) E _
(a) Let H be the set of all holomorphic functions ) : C C. (These are sometimes
called entire functions). Show that H is an integral domain under the operations of
pointwise addition and multiplication. That is: if ), p H, then the functions () + p)
and () p) are in H. (Hint: Use Proposition 18A.5(a,b) on page 18A.5). Also, if
) ,= 0 ,= p, then ) p ,= 0. (Hint: Use the Identity Theorem 18D.3 on page 452).
(b) Let M be the set of all meromorphic functions ) : C

C. Show that M is a
eld under the operations of pointwise addition and multiplication. That is: if ), p M,
then the functions () +p) and () p) are in M, and if p , 0, then the function (),p) is
also in M.
(c) Suppose ) Mhas only a nite number of poles. Show that ) can be expressed
in the form ) = p,/, where p, / H. (Hint: you can make / a polynomial).
(d) (hard) Show that any function ) M can be expressed in the form ) = p,/,
where p, / H. (Thus, M is related to H the same way the eld of rational functions is
related to the ring of polynomials, and the same way that the eld of rational numbers
is related to the ring of integers. Technically, M is the eld of fractions of H.).
18H Improper integrals and Fourier transforms
Prerequisites: 18G. Recommended: 17A, 19A.
The Residue Theorem is a powerful tool for evaluating contour integrals in
the complex plane. We shall now see that it is also useful for computing improper
integrals over the real line, such as convolutions and Fourier transforms. First
some notation. Let C
+
:= c C ; Im[c] 0 and C
:= c C ; Im[c] < 0
be the upper and lower halves of the complex plane. If 1 : C

C is some
meromorphic function, then we say that 1 uniformly decays at innity on
C
+
with order o(1,.) if,
12
for any c 0, there is some : 0 such that:
For all . C
+
,
_
[.[ :
_
=
_
[.[ [1(.)[ < c
_
. (18H.1)
In other words, lim
C
+
z
[.[ [1(.)[ = 0, and this convergence is uniform as
. in any direction in C
+
. We dene uniform decay on C
in the same
fashion.
Example 18H.1. (a) The function )(.) = 1,.
2
uniformly decays at innity on
both C
+
and C
with order o(1,.).

(b) However, the function )(.) = 1,. does not uniformly decays at innity
with order o(1,.) (it decays just a little bit too slowly).
(c) The function )(.) = exp(i.),.
2
uniformly decays at innity with order
o(1,.) on C
+
, but does not decay on C
.
(d) If 1
1
, 1
2
: C C are two complex polynomials of degree
1
and
2
respectively, and
2

1
+ 2, then the rational function )(.) = 1
1
(.),1
2
(.)
uniformly decays with order o(1,.) on both C
+
and C
.
Exercise 18H.1. Verify Examples 18H.1(a-d). E _
12
This is pronounced, small oh of 1/z.
18H. Improper integrals and Fourier transforms 473
Proposition 18H.2. (Improper integrals of analytic functions)
Let ) : 1 C be an analytic function, and let 1 : C

C be an extension
of ) to a meromorphic function on C.
(a) Suppose that 1 uniformly decays with order o(1,.) on C
+
. If j
1
, j
2
, . . . , j
J

C
+
are all the poles of 1 in C
+
, then
_

)(r) dr = 2i
J
j=1
res
p
j
(1).
(b) Suppose that 1 uniformly decays with order o(1,.) on C
. If j
1
, j
2
, . . . , j
J

C
are all the poles of 1 in C
, then
_

)(r) dr = 2i
J
j=1
res
p
j
(1).
Proof. (a) Note that 1 has no poles on the real line 1, because ) is analytic
on 1. For any 1 0, let
R
be the D-shaped contour of radius 1 from
Example 18C.3 on page 436. If 1 is made large enough, then
R
encircles all
of j
1
, j
2
, . . . , j
J
. Thus, the Residue Theorem 18G.2 on page 467 says that
_
R
1 = 2i
J
j=1
res
p
j
(1). (18H.2)
But by denition,
_
R
1 =
_
+R
0
1[
R
(:)]
R
(:)
()
_

0
1(1c
is
) 1i c
is
d: +
_
R
R
)(r) dr,
where () is by equations (18C.1) and (18C.2) on page 436. Thus,
lim
R
_
R
1 = lim
R
_

0
1(1c
is
) 1ic
is
d: + lim
R
_
R
R
)(r) dr
()
_

)(r) dr. (18H.3)

Now combine equations (18H.2) and (18H.3) to prove part (a).
In equation (18H.3), step () is because
lim
R
_
R
R
)(r) dr =
_

)(r) dr, (18H.4)

while lim
R
_

0
1(1c
is
) 1ic
is
d:
= 0 (18H.5)
Equation (18H.4) is just the denition of an improper integral. To see equation
(18H.5), note that
_

0
1(1c
is
) 1ic
is
d:

(z)
_

0
1(1c
is
) 1ic
is
d: =
_

0
1
1(1c
is
)
d:,
(18H.6)
where () is just the triangle inequality for integrals. But for any c 0, we
can nd some : 0 satisfying equation (18H.1). Then for all 1 :, and all
: [0, ], we have 1
1(1c
is
)
< c, which means

_

0
1
1(1c
is
)
d:
_

0
c = c. (18H.7)
Since c 0 can be made arbitrarily small, equations (18H.6) and (18H.7)
imply (18H.5).
Exercise 18H.2 Prove part (b) of the theorem. 2 E _
If ), p : 1 C are integrable functions, recall that their convolution is
the function ) p : 1 1 dened by ) p(:) :=
_
)(r) p(: r) dr,

for any : 1. Chapter 17 showed how to solve I/BVPs by convolving with
impulse-response functions like the Poisson kernel.
Corollary 18H.3. (Convolutions of analytic functions)
Let ), p : 1 C be analytic functions, with meromorphic extensions 1, G :
C

C. Suppose the function . 1(.) G(.) uniformly decays with order
o(1,.) on C
+
. Suppose 1 has simple poles j
1
, j
2
, . . . , j
J
C
+
, and no other
poles in C
+
. Suppose G has simple poles
1
,
2
, . . . ,
K
C
, and no other poles

in C
. Then for all : 1,

) p(:) = 2i
J
j=1
G(: j
j
) res
p
j
(1) 2i
K
k=1
1(:
k
) res
q
j
(G).
Proof. Fix : 1, and consider the function H(.) := 1(.)G(: .). For all
, [1...J], Example 18G.1(e) on page 466 says that H has a simple pole at
j
j
C
+
, with residue G(: j
j
) res
p
j
(1). For all / [1...1], the function
. G(: .) has a simple pole at :
k
, with residue res
q
k
(G). Thus,
Example 18G.1(e) says that H has a simple pole at :
k
, with residue 1(:
k
) res
q
k
(G). Note that (:
k
) C
+
, because
k
C
and : 1. Now
apply Proposition 18H.2. 2
Example 18H.4. For any j 0, recall the half-plane Poisson kernel /
y
: 1
1 from 17E, dened by
/
y
(r) :=
j
(r
2
+j
2
)
, for all r 1.
Let H :=
_
(r, j) 1
2
; j 0
_
(the upper half-plane). If / : 1 1 is
bounded and continuous, then Proposition 17E.1 on page 404 says that the
function /(r, j) := /
y
/(r) is the unique continuous harmonic function on H
which satises the Dirichlet boundary condition /(r, 0) = /(r) for all r 1.
Suppose / : 1 1 is analytic, with a meromorphic extension 1 : C

C
which is asymptotically bounded near innity in C
that is, there exist

1, 1 0 such that [1(.)[ < 1 for all . C
with [.[ 1. Then the

function /
y
1 asymptotically decays near innity with order o(1,.) on C
+
,
so Corollary 18H.3 is applicable.
In Example 18G.1(c) on page 466, we saw that /
y
has a simple pole at ji,
with res
yi
(/
y
) = 1,2i, and no other poles in C
+
. Suppose 1 has simple poles
1
,
2
, . . . ,
K
C
, and no other poles in C
. Then setting ) := /
y
, p := /,
J := 1 and j
1
:= ji in Corollary 18H.3, we get for any (r, j) H,
/(r, j) = 2i 1(r ji) res
yi
(/
y
)
. .
=1/2i
2i
K
k=1
/
y
(r
k
) res
q
k
(1)
= 1(r ji) 2ji
k
k=1
res
q
k
(1)
(r
k
)
2
+j
2
.
Exercise 18H.3. (a) Show that, in fact, /(r, j) = Re [1(r ji)]. (Thus, if we E _
could compute 1, then the BVP would already be solved, and we actually wouldnt
need to apply Proposition 17E.1).
(b) Deduce that Im[1(r ji)] = 2j
K
k=1
res
q
k
(1)
(r
k
)
2
+j
2
.
Exercise 18H.4. Let ) : 1 1 be a bounded analytic function, with meromor- E _
phic extension 1 : C

C. Let n : 1 1
1 be the unique solution to the

one-dimensional heat equation (
t
n =
2
x
n) with initial conditions n(r; 0) = )(r) for
all r 1. Combine Proposition 18H.3 with Proposition 17C.1 on page 385 to nd a
formula for n(r; t) in terms of the residues of 1.
Exercise 18H.5. For any t 0, let
t
: 1 1 be the dAlembert kernel: E _
t
(r) =
_
1
2
if t < r < t;
0 otherwise.
Let )
1
: 1 1 be an analytic function. Lemma 17D.3 on page 395 says that we
can solve the initial velocity problem for the one-dimensional wave equation by dening
(r; t) :=
t
)
1
(r). Explain why Proposition 18H.3 is not suitable for computing
t
)
1
.
If ) : 1 C is an integrable function, then its Fourier transform is the

function

) : 1 C dened by
)(j) :=
1
2
_

exp(jri))(r) dr, for all j 1.

(See 19A for more information). Proposition 18H.2 can also be used to compute
Fourier transforms, but it is not quite the strongest result for this purpose. If
1 : C

C is some meromorphic function, then we say that 1 uniformly
decays at innity with order O(1,.) if
13
there exists some ` 0 and some
: 0 such that:
For all . C,
_
[.[ :
_
=
_
[1(.)[ < `,[.[
_
. (18H.8)
In other words, the function [. 1(.)[ is uniformly bounded (by `) as .
in any direction in C.
Example 18H.5. (a) The function )(.) = 1,. uniformly decays at innity
with order O(1,.).
(b) If ) uniformly decays at innity on C
with order o(1,.), then it also

uniformly decays at innity with order O(1,.). (Exercise 18H.6 Verify E _
this).
(c) In particular, if 1
1
, 1
2
: C C are two complex polynomials of degree
1
and
2
respectively, and
2

1
+ 1, then the rational function )(.) =
1
1
(.),1
2
(.) uniformly decays at innity with order O(1,.).
Thus, decay with order O(1,.) is a slightly weaker requirement than decay
with order o(1,.).
Proposition 18H.6. (Fourier transforms of analytic functions)
Let ) : 1 C be an analytic function. Let 1 : C

C be an exten-
sion of ) to a meromorphic function on C which uniformly decays with order
O(1,.). Let j
K
, . . . , j
2
, j
1
, j
0
, j
1
, . . . , j
J
be all the poles of 1 in C, where
j
K
, . . . , j
2
, j
1
C
and j
0
, j
1
, . . . , j
J
C
+
. Then:
)(j) = i
J
j=0
res
p
j
(c
1), for all j < 0, (18H.9)

and

)(j) = i
K
k=1
res
p
k
(c
1), for all j 0,

13
This is pronounced, big oh of 1/z.
R i
R -R 0
(A) (B)
R
R+2R i
-R+2R i
-R
p
4
p
0
p
2
p
3
p
1
p
4
p
5
p
2
p
3
p
1
r
2
r
1
R
Figure 18H.1: (A) The square contour in the proof of Proposition 18H.6. (B) The
contour in Exercise 18H.12 on page 480
where c
: C C is the holomorphic function dened c
(.) := exp(j . i)
for all . C. In particular, if all the poles of 1 are simple, then
)(j) = i
J
j=0
exp(jj
j
i) res
p
j
(1), for all j < 0, (18H.10)
and

)(j) = i
K
k=1
exp(jj
k
i) res
p
k
(1), for all j 0.
Proof. We will prove the theorem for j < 0. Fix j < 0 and dene G : C C
by G(.) := exp(j. i) 1(.). For any 1 0, dene the chains
R
,
R
,
r
,
and
R
as shown in Figure 18H.1(A):
For all : [1, 1],
R
(:) := : so that

R
(:) = 1.
For all : [0, 21],
R
(:) := 1 +:i so that
R
(:) = i.
For all : [1, 1],
R
(:) := : + 21i so that
R
(:) = 1.
For all : [0, 21],
R
(:) := 1 +:i so that

R
(:) = i.
(18H.11)
(Mnemonic: ottom, ight, op, eft.) Thus, if =

, then
R
traces a square in C
+
of sidelength 21. If 1 is made large enough, then
R
encloses all of j
0
, j
2
, . . . , j
J
. Thus, for any large enough 1 0,
2i
J
j=0
res
p
j
(G)
()
_
R
G
()
R
G +
_
R
G
_
R
G
_
R
G
(18H.12)
where () is by the Residue Theorem 18G.2 on page 467, and where () is by
Lemma 18C.8(a,b) on page 441.
Claim 1: lim
R
R
G = 2

)(j).
Proof. Exercise 18H.7
Claim 1
E _
Claim 2: (a) lim
R
R
G = 0 and (b) lim
R
R
G = 0.
Proof. (a) By hypothesis, ) decays with order O(1,.). Thus, we can nd
some : 0 and ` 0 satisfying eqn.(18H.8). If 1 :, then for all
: [0, 21],
[G(1 +:i)[ = [exp[ji(1 +:i)][ [1(1 +:i)[
()
[exp(j: j1i)[
`
[1 +:i[
c
s
`
1
, (18H.13)
where () is by equation (18H.8). Thus,
R
G
()
_
2R
0
G(1 +:i) i d:

_
2R
0
[G(1 +:i)[ d:
()
`
1
_
2R
0
c
s
d:
=
`
j1
c
s
s=2R
s=0
=
`
j1
(1 c
2R
)
()
`
j1

R
0,
as desired. Here, () is by eqn.(18H.11), () is by eqn.(18H.13), and () is
because j < 0. This proves (a). The proof of (b) is similar.
Claim 2
Claim 3: lim
R
R
G = 0.
Proof. Again nd : 0 and ` 0 satisfying eqn.(18H.8). If 1 :, then
for all : [1, 1],
[G(: + 21i)[ = [exp[ji(: + 21i)][ [1(: + 21i)[
()
[exp(21j :ji)[
`
[: + 21i[
c
2R
`
21
, (18H.14)
where () is by equation (18H.8). Thus,
R
G

()
c
2R
`
21
length(
R
) = c
2R
`
21
21 = `c
2R
()
R
0,
as desired. Here, () is by eqn.(18H.14) and Lemma 18C.10 on page 444,
while () is because j < 0.
Claim 3
Now we put it all together:
2i
J
j=0
res
p
j
(G)
()
lim
R
_
R
G +
_
R
G
_
R
G
_
R
G
_
()
2

)(j) + 0 + 0 + 0 = 2

)(j).
Now divide both sides by 2 to get eqn.(18H.9). Here, () is by eqn.(18H.12),
and () is by Claims 1-3.
Finally, to see eqn.(18H.10), suppose all the poles j
0
, . . . , j
J
are simple. Then
res
p
j
(G) = exp(ijj
j
) res
p
j
(1) for all , [0...J], by Example 18G.1(e) on
page 466. Thus,
i
J
j=0
res
p
j
(G) = i
J
j=0
exp(jj
j
i) res
p
j
(1),
so eqn.(18H.10) follows from (18H.9). 2
Exercise 18H.8. Prove Proposition 18H.6 in the case j 0. E _
Example 18H.7. Dene ) : 1 1 by )(r) := 1,(r
2
+ 1) for all r 1. The
meromorphic extension of ) is simply the complex polynomial 1 : C

C
dened
1(.) :=
1
.
2
+ 1
=
1
(. +i)(. i)
, for all . C.
Clearly 1 has simple poles at i, with res
i
(1) =
1
2i
and res
i
(1) =
1
2i
.
Thus, Proposition 18H.6 says
If j < 0, then

)(j) = i exp(ji i)
1
2i
=
c
2
=
c
[[
2
.
If j 0, then

)(j) = i exp(ji (i))
1
2i
=
c
2
=
c
[[
2
.
We conclude that p(j) =
c
[[
2
for all j 1.
Exercise 18H.9. Compute the Fourier transforms of the following rational func- E _
tions
(a) )(r) =
1
r
4
+ 1
. (Hint: r
4
+1 = (rc
i/4
)(rc
3i/4
)(rc
5i/4
)(rc
7i/4
).)
(b) )(r) =
r
3
1
r
4
+ 5r
2
+ 4
. (Hint: r
4
+ 5r
2
+ 4 = (r
2
+ 1)(r
2
+ 4).)
(c) )(r) =
r
4
r
6
+ 14r
4
+ 49r
2
+ 36
. (Hint: r
6
+14r
4
+49r
2
+36 = (r
2
+1)(r
2
+
4)(r
2
+ 9).)
(d) )(r) =
r +i
r
4
+ 5r
2
+ 4
. (Careful!)
Exercise 18H.10. Why is Proposition 18H.6 not suitable to compute the Fourier E _
transforms of the following functions?
(a) )(r) =
1
r
3
+ 1
.
(b) )(r) =
sin(r)
r
4
+ 1
.
(c) )(r) =
1
[r[
3
+ 1
.
(d) )(r) =
3
r
r
3
+ 1
.
Exercise 18H.11. Let ) : 1 1 be an analytic function whose meromorphic E _
extension 1 : C

C decays with order O(1,.).
(a) State and prove a general formula for trigonometric integrals of the form
_

cos(nr) )(r) dr and

_

sin(nr) )(r) dr
(Hint: Use Proposition 18H.6 and the formula exp(jr) = cos(jr) +i sin(jr)).
Use your method to compute the following integrals:
(b)
_

sin(r)
r
2
+ 1
dr.
(c)
_

cos(r)
r
4
+ 1
dr.
(d)
_

sin(r)
2
r
2
+ 1
dr. (Hint: 2 sin(r)
2
= 1 cos(2r).)
Exercise 18H.12. Proposition 18H.2 requires the function ) to have no poles on E _
the real line 1. However, this is not really necessary.
(a) Let 1 := :
1
, . . . , :
N
1. Let ) : 1 1 1 be an analytic function whose
meromorphic extension 1 : C

C decays with order o(1,.) on C
+
, and has poles
j
1
, . . . , j
J
C
+
, and also has simple poles at :
1
, . . . , :
N
. Show that
_

)(r) dr = 2i
J
j=1
res
pj
(1) + i
N
n=1
res
rn
(1).
18I.
Homological extension of Cauchys theorem 481

Hint: For all c 0 and 1 0, let
R,
be the contour shown in Figure 18H.1(B) on
page 477. This is like the D contour in the proof of Propositions 18H.2, except that it
makes a little semicircular detour of radius c around each of the poles :
1
, . . . , :
N
1.
Show that the integral along each of these c-detours tends to i res
rn
(1) as c 0,
while the integral over the remainder of the real line tends to
_
)(r) dr as c 0 and
1 .
(b) Use your method to compute
_

exp(ijr)
r
2
dr.
Exercise 18H.13. (a) Let 1 := :
1
, . . . , :
N
1. Let ) : 1 1 1 be an E _
analytic function whose meromorphic extension 1 : C

C decays with order O(1,.)
and has poles j
1
, . . . , j
J
C
+
and also has a simple poles at :
1
, . . . , :
N
. Show that, for
any j < 0,
)(j) = i
J
j=1
exp(jj
j
i) res
pj
(1) +
i
2
N
n=1
exp(j:
j
i) res
rn
(1).
(If j 0, its a similar formula, only summing over the residues in C
and multiplying
by 1). Hint. Combine the method from Exercise 18H.12 with the proof technique from
Proposition 18H.6.
(b) Use your method to compute

)(j) when )(r) =
1
r(r
2
+ 1)
.
Exercise 18H.14. The Laplace inversion integral is dened by equation (19H.3) E _
on page 518. State and prove a formula similar to Theorem 18H.6 for the computation
of Laplace inversion integrals.
18I
Homological extension of Cauchys theorem

Prerequisites: 18C.
We have dened contours to be non-self-intersecting curves only so as to
simplify the exposition in Section 18C.
14
All of the results of Section 18C are
true for any piecewise smooth closed curve in C . Indeed, the results of Section
18C can be even extended to integrals on chains, as we now discuss.
Let U C be a connected open set, and let G
1
, G
2
, . . . , G
N
be the connected
components of the boundary U. Suppose each G
n
can be parameterized by
a piecewise smooth contour
n
, such that the outward normal vector eld of
G
n
is always on the right-hand side of
n
. The chain :=
1

2

n
is called the positive boundary of U. Its reversal

is called the negative

boundary of U. Both the negative and positive boundaries of a set are called
14
To be precise, it made it simpler for us to dene the purview of the contour, by invoking the
Jordan Curve Theorem. It also made it simpler to dene clockwise versus counterclockwise
contours.
oriented boundaries. For example, any contour is the oriented boundary
of the purview of . Theorem 18C.5 on page 438 now extends to the following
theorem.
Theorem 18I.1. Cauchys Theorem on oriented boundaries
Let U C be any open set, and let ) : U C be holomorphic on U. If is
an oriented boundary of U, then
_
) = 0. 2
Let
1
,
2
, . . . ,
N
: [0, 1] C be continuous, piecewise smooth curves in C
(not necessarily closed), and consider the chain =
1

2

N
(note that
any chain can be expressed in this way). Lets refer to the paths
1
, . . . ,
N
as
the the links of the chain . We say that is a cycle if the endpoint of each
link is the starting point of exactly one other link, and the starting point of each
link is the endpoint of exactly one other link. In other words, for all : [1..],
there exists a unique /, n [1...C] such that
(1) =
m
(0) and
m
(1) =
n
(0).
Example 18I.2. (a) Any contour is a cycle.
(b) If and are two cycles, then is also a cycle.
(c) Thus, if
1
, . . . ,
N
are contours, then
1

N
is a cycle.
(d) In particular, the oriented boundary of an open set is a cycle.
(e) If is a cycle, then
is a cycle.
Not all cycles are oriented boundaries. For example, let
1
and
2
be two
concentric counterclockwise circles around the origin; then
1

2
is not an
oriented boundary. (Although
1
2
is.)
Let U C be an open set. Let and be two cycles in U. We say that is
homologous to in U if the cycle
is the oriented boundary of some open

subset V U. We then write
U

Example 18I.3. (a) Let be a clockwise circle of radius 1 around the origin,
and let be a clockwise circle of radius 2 around the origin. Then is
homologous to in C
, because
is the positive boundary of the annulus

A := c C ; 1 < [c[ < 2 C
.
(b) If
0
and
1
are contours, then they are cycles. If
0
is homotopic to
1
in
U, then
0
is also homologous to
1
in U. To see this, let : [0, 1] [0, o] C
be a homotopy from
0
to
1
, and let V := ((0, 1) [0, o]). Then V is an
open subset of U, and
1
2
is an oriented boundary of V.
Thus, homology can be seen as a generalization of homotopy. Proposi-
tion 18C.7 on page 440 can be extended as follows:
18I.
Homological extension of Cauchys theorem 483

Proposition 18I.4. (Homology invariance of chain integrals)
Let U C be any open set, and let ) : U C be holomorphic on U. If and
are two cycles which are homologous in U, then
_
) =
_
). 2
Proof. Exercise 18I.1 Hint: Use Theorem 18I.1. 2 E _
The relation
U
is an equivalence relation. That is, for all cycles , , and
,

U
;.
If
U
, then
U
;
If
U
, and
U
, then
U
.
(Exercise 18I.2 Verify these three properties.) E _
For any cycle , let []
U
denote its equivalence class under
U
(this is
called a homology class). Let H
1
(U) denote the set of all homology classes
of cycles. In particular, let []
U
denote the homology class of the empty cycle
then []
U
contains all cycles which are oriented boundaries of subsets of U.
Corollary 18I.5. Let U C be any open set.
(a) If
1
U

2
and
1
U

2
, then (
1

1
)
U
(
2

2
). Thus, we can dene
an operation on H
1
(U) by []
U
[]
U
:= [ ]
U
.
(b) H
1
(U) is an abelian group under the operation .
(c) If ) : U C is holomorphic, then the function []
U

_
) is a group
homomorphism from (H
1
(U), ) to the group (C, +) of complex numbers
under addition.
Proof. (a) is Exercise 18I.3 . Verify the following: E _
(i) The operation is commutative. That is, for any cycles and , we have

U
; thus, []
U
[]
U
= []
U
[]
U
.
(ii) The operation is associative. That is, for any cycles , , and , we have
( )
U
( ) ; thus, []
U
([]
U
[]
U
) = ([]
U
[]
U
)[]
U
.
(iii) The cycle []
U
is an identity element. For any cycle , we have []
U

[]
U
= []
U
.
(iv) For any cycle , the class [
]
U
is an additive inverse for []
U
. That is:
[]
U
[
]
U
= []
U
.
(b) follows immediately from (a). (c) follows from Proposition 18I.4 and
Lemma 18C.8(a,b). 2
The group H
1
(U) is called the rst homology group of U. In general,
H
1
(U) is a free abelian group of rank 1, where 1 is the number of holes in
U. One can similarly dene homology groups for any subset U 1
N
for any
N (e.g. a surface or a manifold), or even for more abstract spaces. The alge-
braic properties of the homology groups of U encode the large-scale topological
properties of U (e.g. the presence of holes or twists). The study of homology
groups is one aspect of a vast and beautiful area in mathematics called algebraic
topology. Surprisingly, the algebraic topology of a dierentiable manifold indi-
rectly inuences the behaviour of partial dierential equations dened on this
manifold; this is content of deep results such as the Atiyah-Singer Index Theo-
rem. For an elementary introduction to algebraic topology, see [Hen94]. For a
comprehensive text, see the beautiful book [Hat02].
485
VI Fourier transforms on
unbounded domains
In Part III, we saw that trigonometric functions like sin and cos formed
orthogonal bases of L
2
(X), where X was one of several bounded subsets of 1
D
.
Thus, any function in L
2
(X) could be expressed using a Fourier series. In Part
IV we used these Fourier series to solve initial/boundary value problems on X.
A Fourier transform is similar to a Fourier series, except that now X is an
unbounded set (e.g. X = 1 or 1
D
). This introduces considerable technical
complications. Nevertheless, the underlying philosophy is the same; we will
construct something analogous to an orthogonal basis for L
2
(X), and use this to
solve partial dierential equations on X.
It is technically convenient (although not strictly necessary) to replace sin
and cos with the complex exponential functions like exp(ri) = cos(r) +i sin(r).
The material on Fourier series in Part III could have also been developed using
these complex exponentials, but in that context, this would have been a needless
complication. In the context of Fourier transforms, however, it is actually a
simplication.
486 DRAFT
487
Chapter 19
Fourier transforms
There is no branch of mathematics, however abstract, which may not someday be applied
to the phenomena of the real world. Nicolai Lobachevsky
19A One-dimensional Fourier transforms
Prerequisites: 0C. Recommended: 6C(i), 8D.
Fourier series help us to represent functions on a bounded domain, like
X = [0, 1] or X = [0, 1] [0, 1]. But what if the domain is unbounded, like
X = 1? Now, instead of using a discrete collection of Fourier coecients like
0
,
1
, 1
1
,
2
, 1
2
, . . . or
)
1
,

)
0
,

)
1
,

)
2
, . . ., we must use a continuously pa-
rameterized family.
For every j 1, we dene the function c
: 1 C by c
(r) := exp(jir).
You can visualize this function as a helix which spirals with frequency j around
the unit circle in the complex plane (see Figure 19A.1). Indeed, using Eulers
Formula (see page 551), it is not hard to check that c
(r) = cos(jr) +i sin(jr)

(Exercise 19A.1). In other words, the real and imaginary parts of c
(r) act E _
like a cosine wave and a sine wave, respectively, both of frequency j.
Heuristically speaking, the (continuously parameterized) family of functions
c
1
acts as a kind of orthogonal basis for a certain space of functions
Figure 19A.1: c
(r) := exp(j r i) as a function of r.

488 DRAFT Chapter 19. Fourier transforms
from 1 into C (although making this rigorous is very complicated). This is the
motivating idea behind the Fourier transform.
Let ) : 1 C be some function. The Fourier transform of ) is the
function

) : 1 C dened:
)(j) :=
1
2
_

)(r)c
(r) dr =
1
2
_

)(r) exp(j r i) dr,

for any j 1. (In other words,

)(j) :=
1
2
), c
), in the notation of 6C(i)

on page 109). Notice that this integral may not converge, in general. We need
)(r) to decay fast enough as r goes to . To be precise, we need ) to be
an absolutely integrable function, meaning that
_

[)(r)[ dr < .
We indicate this by writing: ) L
1
(1).
The Fourier transform

)(j) plays the same role that the complex Fourier
coecients . . .

)
1
,

)
0
,

)
1
,

)
2
, . . . play for a function on an interval (see 8D
on page 172). In particular, we can express )(r) as a sort of generalized Fourier
series. We would like to write something like:
)(r) =
)(j)c
(r).
However, this expression makes no mathematical sense, because you cant sum
over all real numbers (there are too many). Instead of summing over all Fourier
coecients, we must integrate. For this to work, we need a technical condition.
We say that ) is piecewise smooth if there is a nite set of points :
1
< :
2
<
< :
N
in 1 such that ) is continuously dierentiable on the open intervals
(, :
1
), (:
1
, :
2
), (:
1
, :
2
), . . . , (:
N1
, :
N
), and (:
N
, ), and furthermore, the
left-hand and right-hand limits
1
of ) and )
t
exist at each of the points :
1
, . . . , :
N
.
Theorem 19A.1. Fourier Inversion Formula
Suppose that ) L
1
(1) is piecewise smooth. For any r 1, if ) is continuous
at r, then
)(r) = lim
M
_
M
M
)(j) c
(r) dj = lim
M
_
M
M
)(j) exp(j r i) dj.

(19A.1)
If ) is discontinuous at r, then we have
lim
M
_
M
M
)(j) c
(r) dj =
1
2
_
lim
yx
)(j) + lim
y,x
)(j)
_
.
1
See page 201 for denition.
19A. One-dimensional Fourier transforms 489
Proof. See [Wal88, Theorem 5.17, p.244], [Kor88, Theorem 61.1, p.300], or
[Fis99, 5.2, p.335-342]. 2
It follows that, under mild conditions, a function can be uniquely identied
from its Fourier transform:
Proposition 19A.2. Suppose ), p ((1) L
1
(1) are continuous and inte-
grable. Then
_

) = p
_

_
) = p
_
.
Proof. = is obvious. The proof of = is Exercise 19A.2 (Hint. (a) If E _
) and p are piecewise smooth, then show that this follows immediately from Theorem
19A.1.
(b) In the general case (where ) and p might not be piecewise smooth), proceed as
follows. Let / ((1) L
1
(1). Suppose

/ 0; show that we must have / 0. Now
let / := ) p; then

/ =

) p 0 (because

) = p). Thus / = 0; thus, ) = p.) 2
1
0
2
-2 -1
0.3
0.1
0.2
0
u
40 20 0 -40 -20
(A) (B)
Figure 19A.2: (A) Example 19A.3. (B) The Fourier transform

)(r) =
sin()
from Example 19A.3.

Example 19A.3. Suppose )(r) =
_
1 if 1 < r < 1;
0 otherwise
[see Figure
19A.2(A)]. Then
For all j 1,

)(j) =
1
2
_

)(r) exp(j r i) dr =
1
2
_
1
1
exp(j r i) dr
=
1
2ji
exp
_
j r i
_
x=1
x=1
=
1
2ji
_
c
i
c
i
_
=
1
j
_
c
i
c
i
2i
_
(Eu)
1
j
sin(j) [see Fig.19A.2(B)]
where (Eu) is Eulers Formula (see page 551).
Thus, the Fourier Inversion Formula says, that, if 1 < r < 1, then
lim
M
_
M
M
sin(j)
j
exp(j r i) dj = 1,
while, if r < 1 or r 1, then lim
M
_
M
M
sin(j)
j
exp(j r i) dj = 0. If
r = 1, then the Fourier inversion integral will converge to
1
2
.
1
0
2
-1
y
0.35
-0.25
x
20 -20
(A) (B)
Figure 19A.3: (A) Example 19A.4. (B) The real and imaginary parts of the
Fourier transform

)(r) =
1e
i
2i
from Example 19A.4.
_
1 if 0 < r < 1;
0 otherwise
[see Figure 19A.3(A)].
Then

)(j) =
1e
i
2i
[see Figure 19A.3(B)]; the verication of this is practice
problem # 1 on page 523 of 19I. Thus, the Fourier inversion formula says,
that, if 0 < r < 1, then
lim
M
_
M
M
1 c
i
2ji
exp(j r i) dj = 1,
while, if r < 0 or r 1, then lim
M
_
M
M
1 c
i
2ji
exp(j r i) dj = 0. If
r = 0 or r = 1, then the Fourier inversion integral will converge to
1
2
.
19A. One-dimensional Fourier transforms 491
1
0.5
x
10 -10
1
0.5
x
10 -10
(A) (B)
Figure 19A.4: (A) The symmetric exponential tail function )(r) = c
[x[
from Example 19A.7. (B) The Fourier transform

)(r) =
a
(x
2
+a
2
)
of the
symmetric exponential tail function from Example 19A.7.
In the Fourier Inversion Formula, it is important that the positive and neg-
ative bounds of the integral go to innity at the same rate in the limit (19A.1).
In particular, it is not the case that )(r) = lim
N,M
_
M
N
)(j) exp(j r i) dj;

in general, this integral may not converge. The reason is this: even if ) is ab-
solutely integrable, its Fourier transform

) may not be. If we assume that

) is
also absolutely integrable, then things become easier.
Theorem 19A.5. Strong Fourier Inversion Formula
Suppose that ) L
1
(1), and that

) is also in L
1
(1). If r 1, and ) is
continuous at r, then )(r) =
_

)(j) exp(j r i) dj.

Proof. See [Kor88, Theorem 60.1, p.296], [Wal88, Theorem 4.11, p.236], [Fol84,
Theorem 8.26, p. 243], or [Kat76, VI.1.12, p.126]. 2
Corollary 19A.6. Suppose ) L
1
(1), and there exists some p L
1
(1) such
that ) = p. Then

)(j) =
1
2
p(j) for all j 1.
Example 19A.7. Let 0 be a constant, and suppose )(r) = c
[x[
. [see
Figure 19A.4(A)]. Then
2
)(j) =
_

c
[x[
exp(jri) dr
=
_

0
c
x
exp(jri) dr +
_
0
c
x
exp(jri) dr
=
_

0
exp(r jri) dr +
_
0
exp(r jri) dr
=
1
( +ji)
exp
_
( +ji) r
_
x=
x=0
+
1
ji
exp
_
( ji) r
_
x=0
x=
()
1
+ji
(0 1) +
1
ji
(1 0) =
1
+ji
+
1
ji
=
ji + +ji
( +ji)( ji)
=
2
2
+j
2
.
Thus, we conclude:

)(j) =

(
2
+j
2
)
. [see Figure 19A.4(B)].
To see equality (), recall that
exp
_
( +ji) r
_
= c
x
. Thus,
lim
exp
_
( +ji) r
_
= lim
c
x
= 0.
Likewise, lim
exp
_
( ji) r
_
= lim
c
x
= 0.
Example 19A.8. Conversely, suppose 0, and p(r) =
1
(
2
+r
2
)
. Then
p(j) =
1
2
c
[[
; the verication of this is practice problem # 6 on page 523
of 19I.
Remark. Proposition 18H.6 on page 476 provides a powerful technique for
computing the Fourier transform of any analytic function ) : 1 C, using
residue calculus.
19B Properties of the (one-dimensional) Fourier trans-
form
Prerequisites: 19A, 0G.
Theorem 19B.1. Riemann-Lebesgue Lemma
Let ) L
1
(1).
(a) The function

) is continuous and bounded. To be precise: If 1 :=
_

[)(r)[ dr,
then, for all j 1, we have
)(j)
< 1.
(b)

) asymptotically decays near innity. That is, lim
)(j)
= 0.
19B. Properties of the (one-dimensional) Fourier transform 493
Proof. (a) Exercise 19B.1 Hint: Boundedness follows from applying the triangle E _
inequality to the integral dening

)(j). For continuity, x j
1
, j
2
1, and dene
1 : 1 1 by 1(r) := exp(j
1
ri) exp(j
2
ri). For any A 0, we can write
)(j
1
)
)(j
2
) =
1
2
_
_
X
)(r) 1(r) dr
. .
(A)
+
_
X
X
)(r) 1(r) dr
. .
(B)
+
_

X
)(r) 1(r) dr
. .
(C)
_
.
(i) Show that, if A is large enough, then the integrals (A) and (C) can be made
arbitrarily small, independent of the values of j
1
and j
2
. (Hint. Recall that )
L
1
(1). Observe that [1(r)[ 2 for all r 1.)
(ii) Fix A 0. Show that, if j
1
and j
2
are close enough, then integral (B) can also
be made arbitrarily small (Hint: if j
1
and j
2
are close, then [1(r)[ is small for all
r 1.)
(iii) Show that, if j
1
and j
2
are close enough, then
)(j
1
)

)(j
2
)
can be made
arbitrarily small. (Hint: Combine (i) and (ii), using the triangle inequality). Hence,
) is continuous.
(b) (if ) is continuous) Exercise 19B.2 Hint. For any A 0, we can write E _
)(j) =
1
2
_
_
X
)(r) c
(r) dr
. .
(A)
+
_
X
X
)(r) c
(r) dr
. .
(B)
+
_

X
)(r) c
(r) dr
. .
(C)
_
.
(i) Show that, if A is large enough, then the integrals (A) and (C) can be made
arbitrarily small, independent of the value of j. (Hint. Recall that ) L
1
(1).
Observe that [c
(r)[ = 1 for all r 1.)

(ii) Fix A 0. Show that, if j is large enough, then integral (B) can also be made
arbitrarily small (Hint: ) is uniformly continuous on the interval [A, A] (why?).
Thus, for any c 0, there is some ` such that, for all j `, and all r [A, A],
we have [)(r) )(r +,j)[ < c. But c
(r +,j) = c
(r)).
(iii) Show that, if j is large enough, then
)(j)
can be made arbitrarily small. (Hint:

Combine (i) and (ii), using the triangle inequality).
For the proof of (b) when ) is an arbitrary (discontinuous) element of L
1
(1),
see [Fol84, Theorem 8.22(f), p.241] or [Fis99, Exercise 15, 5.2, p.343] or
[Kat76, Theorem 1.7, p.123]. 2
Recall that, if ), p : 1 1 are two functions, then their convolution is
the function () p) : 1 1 dened:
() p)(r) :=
_

)(j) p(r j) dj.

(see 17A on page 375 for a discussion of convolutions). Similarly, if ) has
Fourier transform

) and p has Fourier transform p, we can convolve

) and p to
get a function (
) p) : 1 1 dened:
(
) p)(j) :=
_

)() p(j ) d.
Theorem 19B.2. Algebraic Properties of the Fourier Transform
Suppose ), p L
1
(1) are two functions.
(a) If / := ) +p, then for all j 1, we have

/(j) =

)(j) + p(j).
(b) If / := ) p, then for all j 1, we have

/(j) = 2

)(j) p(j).
(c) Conversely, suppose / := ) p. If

), p and

/ are in L
1
(1), then for all
j 1, we have

/(j) = (
) p)(j).
Proof. See practice problems #11 to # 13 on page 524. 2
This theorem allows us to compute the Fourier transform of a complicated
function by breaking it into a sum/product of simpler pieces.
Theorem 19B.3. Translation and Phase Shift
Suppose ) L
1
(1).
(a) If 1 is xed, and p L
1
(1) is dened by: p(r) := )(r + ), then for
all j 1, we have p(j) = c
i

)(j).
(b) Conversely, if 1 is xed, and p L
1
(1) is dened: p(r) := c
xi
)(r),
then for all j 1, we have p(j) =

)(j ).
Proof. See practice problems #14 and # 15 on page 524. 2
Thus, translating a function by in physical space corresponds to phase-
shifting its Fourier transform by c
i
, and vice versa. This means that, via a
suitable translation, we can put the center of our coordinate system wherever
it is most convenient to do so.
Example 19B.4. Suppose p(r) =
_
1 if 1 < r < 1 ;
0 otherwise
. Thus,
p(r) = )(r + ), where )(r) is as in Example 19A.3 on page 489. We know
that

)(j) =
sin(j)
j
; thus, it follows from Theorem 19B.3 that p(j) =
c
i
sin(j)
j
.
y
x
f^
g^
Figure 19B.1: Plot of

) (black) and p (grey) in Example 19B.6, where p(r) =
)(r,3).
Theorem 19B.5. Rescaling Relation
Suppose ) L
1
(1). If 0 is xed, and p is dened by: p(r) = )
_
r
_
, then
for all j 1, p(j) =

)( j).
Proof. See practice problem # 16 on page 525. 2
In Theorem 19B.5, the function p is the same as function ), but expressed
in a coordinate system rescaled by a factor of .
Example 19B.6. Suppose p(r) =
_
1 if 3 < r < 3;
0 otherwise
. Thus, p(r) =
)(r,3), where )(r) is as in Example 19A.3 on page 489. We know that

)(j) =
sin(j)
j
; thus, it follows from Theorem 19B.5 that p(j) = 3
sin(3j)
3j
=
sin(3j)
j
. See Figure 19B.1.
A function ) : 1 C is continuously dierentiable if )
t
(r) exists for
all r 1, and the function )
t
: 1 C is itself continuous. Let (
1
(1) be the
set of all continuously dierentiable functions from 1 to C. For any n N let
)
(n)
(r) :=
d
n
dx
n
)(r). The function ) is n times continuously dierentiable if
)
(n)
(r) exists for all r 1, and the function )
(n)
: 1 C is itself continuous.
Let (
n
(1) be the set of all n-times continuously dierentiable functions from 1
to C.
F
F
Jaggy Smooth
Slow decay Rapid decay
Figure 19B.2: Smoothness vs. asymptotic decay in the Fourier Transform.
Theorem 19B.7. Dierentiation and Multiplication
Suppose ) L
1
(1).
(a) Suppose ) (
1
(1) and lim
x
[)(r)[ = 0. Let p(r) := )
t
(r). If p L
1
(1),
then for all j 1, we have p(j) = ij

)(j).
(b) More generally, suppose ) (
n
(1) and lim
x
[)
(n1)
(r)[ = 0. Let p(r) :=
)
(n)
(r). If p L
1
(1), then for all j 1, we have p(j) = (ij)
n

)(j).
Thus,

)(j) asymptotically decays faster than
1
j
n
as j . That is,
lim
j
n
)(j) = 0.
(c) Conversely, let p(r) := r
n
)(r), and suppose that ) decays quickly
enough that p is also in L
1
(1) [for example, this happens if lim
x
r
n+1
)(r) =
0]. Then the function

) is n times dierentiable, and, for all j 1,
p(j) = i
n
d
n
dj
n
)(j).
Proof. (a) is practice problem # 17 on page 525 of 19I.
(b) is just the result of iterating (a) n times.
(c) is Exercise 19B.3 (Hint: either reverse the result of (a) using the Fourier E _
Inversion Formula (Theorem 19A.1 on page 488), or use Proposition 0G.1 on page 567
to directly dierentiate the integral dening

)(j).) 2
This theorem says that the Fourier transform converts dierentiation-by-
r into multiplication-by-ji. This implies that the smoothness of a function )
is closely related to the asymptotic decay rate of its Fourier transform. The
smoother ) is (i.e. the more times we can dierentiate it), the more rapidly
)(j) decays as j (see Figure 19B.2).

Physically, we can interpret this as follows. If we think of ) as a signal,
then

)(j) is the amount of energy at the frequency j in the spectral decom-
position of this signal. Thus, the magnitude of

)(j) for extremely large j is the
amount of very high frequency energy in ), which corresponds to very nely
featured, jaggy structure in the shape of ). If ) is smooth, then we expect
there will be very little of this jagginess; hence the high frequency part of the
energy spectrum will be very small.
Conversely, the asymptotic decay rate of ) determines the smoothness of its
Fourier transform. This makes sense, because the Fourier inversion formula can
be (loosely) intepreted as saying that ) is itself a sort of backwards Fourier
transform of

).
One very important Fourier transform is the following:
Theorem 19B.8. Fourier Transform of a Gaussian
(a) If )(r) = exp
_
r
2
_
, then

)(j) =
1
2
)
_
j
2
_
=
1
2
exp
_
j
2
4
_
.
(b) Fix 0. If )(r) =
1
2
exp
_
r
2
2
2
_
is a Gaussian probability distri-
bution with mean 0 and variance
2
, then
)(j) =
1
2
exp
_
2
j
2
2
_
.
(c) Fix 0 and 1. If )(r) =
1
2
exp
_
[r [
2
2
2
_
is a Gaussian
probability distribution with mean and variance
2
, then
)(j) =
c
i
2
exp
_
2
j
2
2
_
.
Proof. Well start with part (a). Let p(r) = )
t
(r). Then by Theorem
19B.7(a),
p(j) = ij

)(j). (19B.1)
However direct computation says p(r) = 2r )(r), so
1
2
p(r) = r )(r), so
Theorem 19B.7(c) implies
i
2
p(j) = (
))
t
(j). (19B.2)
Combining (19B.2) with (19B.1), we conclude:
(
))
t
(j)
(19B.2)
i
2
p(j)
(19B.1)
i
2
ij

)(j) =
j
2
)(j). (19B.3)
Dene /(j) =

)(j) exp
_
j
2
4
_
. If we dierentiate /(j), we get:
/
t
(j)
(dL)
)(j)
j
2
exp
_
j
2
4
_

j
2
)(j)
. .
()
exp
_
j
2
4
_
= 0.
Here, (dL) is dierentiating using the Leibniz rule, and () is by eqn.(19B.3).
In other words, /(j) = H is a constant. Thus,
)(j) =
/(j)
exp (j
2
,4)
= H exp
_
j
2
4
_
= H )
_
j
2
_
.
To evaluate H, set j = 0, to get
H = H exp
_
0
2
4
_
=

)(0) =
1
2
_

)(r) dr =
1
2
_

exp
_
r
2
_
=
1
2
.
(where the last step is Exercise 19B.4 ). Thus, we conclude:

)(j) =
1
2
E _
)
_
j
2
_
.
Part (b) follows by setting :=

2 in Theorem 19B.5 on page 495.
Part (c) is Exercise 19B.5 (Hint: Apply Theorem 19B.3 on page 494). 2 E _
Loosely speaking, Theorem 19B.8 says, The Fourier transform of a Gaussian
is another Gaussian
2
. However, notice that, in Part (b) of the theorem, as
the variance of the Gaussian (that is,
2
) gets bigger, the variance of its
Fourier transform (which is eectively
1
2
) gets smaller (see Figure 19B.3). If
we think of the Gaussian as the probability distribution of some unknown piece
of information, then the variance measures the degree of uncertainty. Hence,
we conclude: the greater the uncertainty embodied in the Gaussian ), the less
the uncertainty embodied in

), and vice versa. This is a manifestation of the
so-called Heisenberg Uncertainty Principle (see Theorem 19G.2 on 513).
2
This is only loosely speaking, however, because a proper Gaussian contains the multiplier
2
to make it a probability distribution, whereas the Fourier transform does not.
F
3
F
2
F
1
F
3
F
2
1
F
small big
small big
1/
1/
Figure 19B.3: The Uncertainty Principle.
Proposition 19B.9. Inversion and Conjugation
For any . C, let . denote the complex conjugate of .. Let ) L
1
(1).
(a) For all j 1, we have

)(j) =

)(j). In particular,
_
) purely real-valued
_

_
for all j 1, we have

)(j) =

)(j)
_
.
(b) Suppose p(r) = )(r) for all r 1. Then for all j 1, p(j) =

)(j).
In particular, if ) purely real-valued, then p(j) =

)(j). for all j 1.
(c) If ) is real-valued and even (i.e. )(r) = )(r)), then

) is purely real-
valued.
(d) If ) is real-valued and odd (i.e. )(r) = )(r)), then

) is purely imaginary-
valued.
Example 19B.10: Autocorrelation and Power Spectrum
If ) : 1 1, then the autocorrelation function of ) is dened by
A)(r) :=
_

)(j) )(r +j) dj.

Heuristically, if we think of )(r) as a random signal, then A)(r) measures
the degree of correlation in the signal across time intervals of length r i.e.
it provides a crude measure of how well you can predict the value of )(j +r)
given information about )(r). In particular, if ) has some sort of T-periodic
component, then we expect A)(r) to be large when r = nT for any n Z.
If we dene p(r) = )(r), then we can see that A)(r) = () p)(r)
(Exercise 19B.7). Thus, E _
A)(j)
()
) p (j)
()
)(j) p(j)
()
)(j)

)(j) =

)(j)

)(j) =
)(j)
2
.
Here, both () are by Proposition 19B.9(b), while () is by Theorem 19B.2(b).
The function
)(j)
2
measures the absolute magnitude of the Fourier transform
of

), and is sometimes called the power spectrum of

).
Evil twins of the Fourier transform. Unfortunately, the mathematics lit-
erature contains at least four dierent denitions of the Fourier transform. In
this book, the Fourier transform and its inversion are dened with the integrals
)(j) :=
1
2
_

)(r) exp(i rj) dj and )(r) =

_

)(r) exp(i rj) dj.

Some books (e.g. [Kat76, Kor88, Fis99, Hab87]) instead use what we will call
the opposite Fourier transform:
)(j) :=
_

)(r) exp(i rj) dr,

with inverse transform
)(r) =
1
2
_

)(j) exp(i rj) dj.

Other books (e.g. [Asm05]) instead use what we will call the symmetric Fourier
transform:
)(j) :=
1
2
_

)(r) exp(i rj) dr,

)(r) =
1
2
_

)(j) exp(i rj) dj.

Finally, some books (e.g. [Fol84, Wal88]) use what we will call the canonical
Fourier transform:
)(j) :=
_

)(r) exp(2i rj) dr,

)(r) =
_

)(j) exp(2i rj) dj.

All books use the symbol
) to denote the Fourier transform of ) we are

using four dierent accents simply to avoid confusing the four denitions. It is
easy to translate the Fourier transform in this book into its evil twins. For any
) L
1
(1) and any j 1, we have:
)(j) = 2

)(j) and

)(j) =
1
2
)(j);
)(j) =

2

)(j) and

)(j) =
1
)(j);
)(j) = 2

)(2 j) and

)(j) =
1
2
)
_
j
2
_
.
(19B.4)
(Exercise 19B.8 Check this.) All of the formulae and theorems we have derived E _
in this section are still true under these alternate denitions, except that one must
multiply or divide by 2 or

2 at certain key points, and replace c
i
with c
2i
(or vice versa) at others.
Exercise 19B.9. (Annoying) Use the identities (19B.4) to reformulate all the for- E _
mulae and theorems in this chapter in terms of (a) The opposite Fourier transform

);
or (b) The symmetric Fourier transform

); or (c) The canonical Fourier transform

).
Each of the four denitions has advantages and disadvantages; some formu-
lae become simpler, others become more complex. Clearly, both the symmetric
and canonical versions of the Fourier transform have some appeal because the
Fourier transform and its inverse have symmetrical formulae using these de-
nitions. Furthermore, in both of these versions, the 2 factor disappears from
Parsevals and Plancherels Theorems (see 19C below) in other words, the
Fourier transform becomes an isometry of L
2
(1). The symmetric Fourier trans-
form has the added advantage that it maps a Gaussian distribution into another
Gaussian (no scalar multiplication required). The canonical Fourier transform
has the added advantage that

) p =

) p (without the 2 factor required in
Theorem 19B.2(a)), while simultaneously,

) p =

) p (unlike the symmetric
Fourier transform).
The denition used in this book (and also in [Pin98, CB87, Pow99, Bro89,
McW72], among others) has none of these advantages. Its major advantage is
that it will yield simpler expressions for the abstract solutions to partial dier-
ential equations in Chapter 20. If one uses the symmetric Fourier transform,
then every one of the solution formulae in Chapter 20 must be multiplied by
some power of
1
2
. If one uses the canonical Fourier transform, then every
spacetime variable (i.e. r, j, ., t) in every formula must by multiplied by 2 or
sometimes by 4
2
, which makes all the formulae look much more complicated.
3
3
Of course, when you actually apply these formulae to solve specic problem, you will end
We end with a warning. When comparing or combining formulae from two or
more books, make sure to rst compare their denitions of the Fourier transform,
and make the appropriate conversions using formulae (19B.4), if necessary.
Further reading. Almost any book on PDEs contains a discussion of Fourier
transforms, but for greater depth (and rigour) it is better to seek a text dedi-
cated to Fourier analysis. Good introductions to Fourier transforms and their
applications can be found in [Wal88, Chapter 6-7] and [Kor88, Part IV]. (In addi-
tion to a lot of serious mathematical content, Korners book contains interesting
and wide-ranging discussions about the history of Fourier theory and its many
scientic applications, and is written in a delightfully informal style).
19C
Parseval and Plancherel

Prerequisites: 19A. Recommended: 6C(i), 6F.
Let L
2
(1) be the set of all square-integrable complex-valued functions on 1
that is, all integrable functions ) : 1 C such that |)|
2
< , where we
dene
|)|
2
:=
__

[)(r)[
2
dr
_
1/2
(see 6C(i) for more information).
Note that L
2
(1) is neither a subset nor a superset of L
1
(1); however, the two
spaces do overlap. If ), p L
2
(1), then we dene
), p) :=
_

)(r) p(r) dr.

The following identity is useful in many applications of Fourier theory, especially
quantum mechanics. It can be seen as the continuum analog of Parsevals
equality for an orthonormal basis (Theorem 6F.1 on page 132).
Theorem 19C.1. Parsevals Equality for Fourier Transforms
If ), p L
1
(1) L
2
(1), then ), p) = 2
_
), p
_
.
Proof. Dene / : 1 1 by /(r) := )(r)p(r). Then / L
1
(1) because
), p L
2
(1). We have
/(0) =
1
2
_

)(r) p(r) exp(i0r) dr

()
1
2
_

)(r) p(r) dr =
), p)
2
, (19C.1)
up with exactly the same solution no matter which version of the Fourier transform you use
why?
19C.
Parseval and Plancherel 503

where () is because exp(i0r) = exp(0) = 1 for all r 1. But we also have
/(0)
()
p(0) =
_

)()
p() d
()
_

)() p() d =
_
), p
_
. (19C.2)
Here, () is by Theorem 19B.2(c), because / = ) p so

/ =

)
p. Meanwhile,
() is by Proposition 19B.9(a).
Combining (19C.1) and (19C.2) yields
_
), p
_
=

/(0) = ), p) ,2. The result
follows. 2
Corollary 19C.2. Plancherels Theorem
Suppose ) L
1
(1) L
2
(1). Then

) L
1
(1) L
2
(1) also, and |)|
2
=
2
_
_
_
)
_
_
_
2
.
Proof. Set ) = p in the Parseval equality. Recall that |)|
2
=
_
), )). 2
In fact, the Plancherel Theorem says much more than this. Dene the linear
operator F
1
: L
1
(1) L
1
(1) by F
1
()) :=

2

) for all ) L
1
(1); then the
full Plancherel Theorem says that F
1
extends uniquely to a unitary isomorphism
F
2
: L
2
(1) L
2
(1) that is, a bijective linear transformation from L
2
(1) to
itself such that |F
2
())|
2
= |)|
2
for all ) L
2
(1). For any j [1, ), let L
p
(1)
be the set of all integrable functions ) : 1 C such that |)|
p
< , where
|)|
p
:=
__

[)(r)[
p
dr
_
1/p
.
For any j [1, 2], let j [2, ] be the unique number such that
1
j
+
1
j
= 1 (for
example, if j = 3,2, then j = 3). Then, through a process called Riesz-Thorin
interpolation, it is possible to extend the Fourier transform even further, to get
a linear transformation F
p
: L
p
(1) L
p
(1). For example, one can dene a
Fourier transform F
3/2
: L
3/2
(1) L
3
(1). All these transformations agree on
the overlaps of their domains, and satisfy the Hausdor-Young inequality:
|F
p
())|
p
|)|
p
, for any j [1, 2] and ) L
p
(1).
However, the details are well beyond the scope of this text. For more information,
see [Fol84, Chapter 8] or [Kat76, Chapter VI].
19D Two-dimensional Fourier transforms
Prerequisites: 19A. Recommended: 9A.
Let L
1
(1
2
) be the set of all functions ) : 1
2
C which are absolutely
integrable on 1
2
, meaning that
_
1
2
[)(r, j)[ dr dj < .
If ) L
1
(1
2
), then the Fourier transform of ) is the function

) : 1
2
C
dened:
)(j, ) :=
1
4
2
_
1
2
)(r, j) exp
_
(jr +j) i
_
dr dj,
for all (j, ) 1
2
.
Theorem 19D.1. Strong Fourier Inversion Formula
Suppose that ) L
1
(1
2
), and that

) is also in L
1
(1
2
). For any (r, j) 1
2
, if
) is continuous at (r, j), then
)(r, j) =
_
1
2
)(j, ) exp
_
(jr +j) i
_
dj d.
Proof. [Fol84, Theorem 8.26, p. 243] or [Kat76, VI.1.12, p.126]. 2
Unfortunately, not all the functions one encounters have the property that
their Fourier transform is in L
1
(1
2
). In particular,

) L
1
(1
2
) only if ) agrees
almost everywhere with a continuous function (thus, Theorem 19D.1 is inap-
plicable to step functions, for example). We want a result analogous to the
weak Fourier Inversion Theorem 19A.1 on page 488. It is surprisingly di-
cult to nd clean, simple inversion theorems of this nature for multidimen-
sional Fourier transforms. The result given here is far from the most gen-
eral one in this category, but it has the advantage of being easy to state and
prove. First, we must dene an appropriate class of functions. Let

L
1
(1
2
) :=
_
) L
1
(1
2
) ;

) L
1
(1
2
)
_
; this is the class considered by Theorem 19D.1. Let
L
1
(1) be the set of all piecewise smooth functions in L
1
(1) (the class considered
by Theorem 19A.1). Let T(1
2
) be the set of all functions ) L
1
(1
2
) such that
there exist )
1
, )
2

L
1
(1) with )(r
1
, r
2
) = )
1
(r
1
) )
2
(r
2
) for all (r
1
, r
2
) 1
2
.
Let H(1
2
) denote the set of all functions in L
1
(1
2
) which can be written as a
nite sum of elements in T(1
2
). Finally, we dene
L
1
(1
2
) :=
_
) L
1
(1
2
) ; ) = p +/ for some p

L
1
(1
2
) and / H(1
2
)
_
.
19D. Two-dimensional Fourier transforms 505
1
Y
-
Y
-X X
Figure 19D.1: Example 19D.4
Theorem 19D.2. 2-dimensional Fourier Inversion Formula
Suppose )

L
1
(1
2
). If (r, j) 1
2
and ) is is continuous at (r, j), then
)(r, j) = lim
M
_
M
M
_
M
M
)(j, ) exp
_
(jr +j) i
_
dj d. (19D.1)
Proof. Exercise 19D.1 (a) First show that eqn.(19D.1) holds for any element of E _
T(1
2
). (Hint. If )
1
, )
2

L
1
(1), and )(r
1
, r
2
) = )
1
(r
1
) )
2
(r
2
) for all (r
1
, r
2
) 1
2
,
then show that

)(j
1
, j
2
) =

)
1
(j
1
)

)
2
(j
2
). Substitute this expression into the right
hand side of eqn.(19D.1); factor the integral into two one-dimensional Fourier inversion
integrals, and then apply Theorem 19A.1 on page 488.)
(b) Deduce that eqn.(19D.1) holds for any element of H(1
2
). (Hint. The Fourier
transform is linear.)
(c) Now combine (b) with Theorem 19D.1 to conclude that eqn.(19D.1) holds for any
element of

L
1
(1
2
). 2
Proposition 19D.3. If ), p ((1
2
) L
1
(1
2
) are continuous, integrable
functions, then
_

) = p
_

_
) = p
_
. 2
Example 19D.4. Let A, Y 0, and let )(r, j) =
_
_
_
1 if A r A
and Y j Y ;
0 otherwise.
(Figure 19D.1) Then:
)(j, ) =
1
4
2
_

)(r, j) exp
_
(jr +j) i
_
dr dj
=
1
4
2
_
X
X
_
Y
Y
exp(jri) exp(ji) dr dj
=
1
4
2
__
X
X
exp(jri) dr
_
__
Y
Y
exp(ji)dj
_
=
1
4
2

_
1
ji
exp
_
jri
_
x=X
x=X
_
_
1
i
exp
_
ji
_
y=Y
y=Y
_
=
1
4
2
_
c
Xi
c
Xi
ji
__
c
Y i
c
Y i
i
_
=
1
2
j
_
c
Xi
c
Xi
2i
__
c
Y i
c
Y i
2i
_
(Eu)
1
2
j
sin(jA) sin(Y ),
where (Eu) is by double application of Eulers formula (see page 551). Note
that ) is in T(1
2
) (why?), and thus, in

L
1
(1
2
). Thus, Theorem 19D.2 says,
that, if A < r < A and Y < j < Y , then
lim
M
_
M
M
_
M
M
sin(jA) sin(Y )
2
j
exp
_
(jr +j) i
_
dj d = 1,
while, if (r, j) , [A, A] [Y, Y ], then
lim
M
_
M
M
_
M
M
sin(jA) sin(Y )
2
j
exp
_
(jr +j) i
_
dj d = 0.
At points on the boundary of the box [0, A] [0, Y ], however, the Fourier
inversion integral will converge to neither of these values.
Example 19D.5. If )(r, j) =
1
2
2
exp
_
r
2
j
2
2
2
_
is a two-dimensional
Gaussian distribution, then

)(j, ) =
1
4
2
exp
_
2
2
_
j
2
+
2
_
_
.
(Exercise 19D.2 ) E _
Exercise 19D.3. State and prove 2-dimensional versions of all results in 19B. E _
19E Three-dimensional Fourier transforms
Prerequisites: 19A. Recommended: 9B, 19D.
In three or more dimensions, it is cumbersome to write vectors as an explicit
list of coordinates. We will adopt a more compact notation. Bold-face letters
will indicate vectors, and italic letters, their components. For example:
x = (r
1
, r
2
, r
3
), y = (j
1
, j
2
, j
3
), = (j
1
, j
2
, j
3
), and = (
1
,
2
,
3
)
19E. Three-dimensional Fourier transforms 507
We dene the inner product x y := r
1
j
1
+ r
2
j
2
+ r
3
j
3
. Let L
1
(1
3
) be
the set of all functions ) : 1
3
C which are absolutely integrable on 1
3
,
meaning that
_
1
3
[)(x)[ dx < .
If ) L
1
(1
3
), then we can dene
_
1
3
)(x) dx :=
_

)(r
1
, r
2
, r
3
) dr
1
dr
2
dr
3
,
where this integral is understood to be absolutely convergent. In particular if
) L
1
(1
3
), then the Fourier transform of ) is the function

) : 1
3
C
dened:
)() :=
1
8
3
_
1
3
)(x) exp (x i) dx,
for all 1
3
. Dene

L
1
(1
3
) in a manner analogous to the denition of

L
1
(1
2
)
on page 505.
Theorem 19E.1. 3-dimensional Fourier Inversion Formula
(a) Suppose )

L
1
(1
3
). For any x 1
3
, if ) is continuous at x, then
)(x) = lim
M
_
M
M
_
M
M
_
M
M
)() exp ( x i) d.
(b) Suppose ) L
1
(1
3
), and

) is also in L
1
(1
3
). For any x 1
3
, if ) is
continuous at x, then )(x) =
_
1
3
)() exp( x i) dj.

Proof. (a) Exercise 19E.1 (Hint: Generalize the proof of Theorem 19D.2 on E _
page 505. You may assume (b) is true.)
(b) See [Fol84, Thm 8.26, p. 243] or [Kat76, VI.1.12, p.126]. 2
Proposition 19E.2. If ), p ((1
3
) L
1
(1
3
) are continuous, integrable
functions, then
_

) = p
_

_
) = p
_
. 2
Example 19E.3: A Ball
For any x 1
3
, let )(x) =
_
1 if |x| 1;
0 otherwise.
. Thus, )(x) is nonzero
on a ball of radius 1 around zero. Then
)() =
1
2
2
_
sin(j1)
j
3

1cos(j1)
j
2
_
,
where j := ||.
Exercise 19E.2. Verify Example 19E.3. Hint: Argue that, by spherical symme- E _
try, we can rotate without changing the integral, so we can assume that = (j, 0, 0).
Switch to the spherical coordinate system (r
1
, r
2
, r
3
) = (: cos(), : sin() sin(), : sin() cos()),
to express the Fourier integral as
1
8
3
_
R
0
_

0
_

exp (j : cos() i) : sin() d d d:.

Use Claim 1 from Theorem 20B.6 on page 534 to simplify this to
1
2
2
j
_
R
0
:sin (j :) d:.
Now apply integration by parts.
Exercise 19E.3 The Fourier transform of Example 19E.3 contains the terms E _
sin(j1)
j
3
and
cos(j1)
j
2
, both of which go to innity as j 0. However, these
two innities cancel out. Use lHopitals rule to show that lim
0
)() =
1
24
3
.
Example 19E.4: A spherically symmetric function
Suppose ) : 1
3
1 was a spherically symmetric function; in other words,
)(x) = (|x|) for some function : 1
,
1. Then for any 1
3
,
)() =
1
2
2
_

0
(:) : sin (|| :) d:.
(Exercise 19E.4 ) E _
1-dimensional Fourier transforms. Fourier transforms can be dened in
an analogous way in higher dimensions. Let L
1
(1
D
) be the set of all functions
) : 1
D
C such that
_
1
D
[)(x)[ dx < . If ) L
1
(1
D
), then the Fourier
transform of ) is the function

) : 1
D
C dened:
)() :=
1
(2)
D
_
1
D
)(x) exp (x i) dx,
for all 1
D
. Dene

L
1
(1
D
) in a manner analogous to the denition of

L
1
(1
2
)
on page 505.
Theorem 19E.5. 1-dimensional Fourier Inversion Formula
(a) Suppose )

L
1
(1
D
). For any x 1
D
, if ) is continuous at x, then
)(x) = lim
M
_
M
M
_
M
M

_
M
M
)() exp ( x i) d.
19E. Three-dimensional Fourier transforms 509
(b) Suppose ) L
1
(1
D
), and

) is also in L
1
(1
D
). For any x 1
D
, if ) is
continuous at x, then )(x) =
_
1
D
)() exp( x i) dj.

Proof. (a) Exercise 19E.5 E _
(b) See [Fol84, Thm 8.26, p. 243] or [Kat76, VI.1.12, p.126]. 2
Exercise 19E.6. State and prove 1-dimensional versions of all results in 19B. E _
Evil twins of multidimensional Fourier transform. Just as with the one-
dimensional Fourier transform, the mathematics literature contains at least four
dierent denitions of multidimensional Fourier transform. Instead of the trans-
form we have dened here, some books use what we will call the opposite Fourier
transform:
)() :=
_
1
D
)(x) exp(i x ) dx,
)(x) =
1
(2)
D
_
1
D
)() exp(i x ) d.
Other books instead use the symmetric Fourier transform:
)() :=
1
(2)
D/2
_
1
D
)(x) exp (i x ) dx,
)(x) =
1
(2)
D/2
_
1
D
)() exp (i x ) d,
Finally, some books use the canonical Fourier transform:
)() :=
_
1
D
)(x) exp (2i x ) dx,
)(x) =
_
1
D
)() exp (2i x ) d,

For any ) L
1
(1
D
) and any 1
D
, we have:
)() = (2)
D

)() and

)() =
1
(2)
D
)();
)() = (2)
D/2

)() and

)() =
1
(2)
D/2
)();
)() = (2)
D

)(2 ) and

)() =
1
(2)
D
)
_

2
_
.
(19E.1)
(Exercise 19E.7 Check this.) When comparing or combining formulae from E _
two or more books, always compare their denitions of the Fourier transform,
and make the appropriate conversions using formulae (19E.1), if necessary.
19F Fourier (co)sine Transforms on the half-line
Prerequisites: 19A. Recommended: 7A, 8A.
In 8A, to represent a function on the symmetric interval [, ], we used a
real Fourier series (with both sine and cosine terms). However, to represent
a function on the interval [0, ], we found in 7A that it was only necessary to
employ half as many terms, using either the Fourier sine series or the Fourier
cosine series. A similar phenomenon occurs when we go from functions on the
whole real line to functions on the positive half-line.
Let 1
,
:= r 1 ; r 0 be the half-line: the set of all nonnegative real
numbers. Let
L
1
(1
,
) :=
_
) : 1
,
1 ;
_

0
[)(r)[ dr <
_
be the set of absolutely integrable functions on the half-line.
The boundary of the half-line is just the point 0. Thus, we will say that
a function ) satises homogeneous Dirichlet boundary conditions if )(0) = 0.
Likewise, ) satises homogeneous Neumann boundary conditions if )
t
(0) = 0.
If ) L
1
(1
,
), then the Fourier Cosine Transform of ) is the function
)
cos
: 1
,
1 dened:
)
cos
(j) :=
2
_

0
)(r) cos(jr) dr, for all j 1
,
.
The Fourier Sine Transform of ) is the function

)
sin
: 1
,
1 dened:
)
sin
(j) :=
2
_

0
)(r) sin(jr) dr, for all j 1
,
.
In both cases, for the transform to be well-dened, we require ) L
1
(1
,
).
Theorem 19F.1. Fourier (co)sine Inversion Formula
Suppose that ) L
1
(1
,
) be peicewise smooth. Then for any r 1
+
such that
) is continuous at r,
)(r) = lim
M
_
M
0
)
cos
(j) cos(j r) dj,
and )(r) = lim
M
_
M
0
)
sin
(j) sin(j r) dj,
19G.
Momentum representation & Heisenberg uncertainty 511

The Fourier cosine series also converges at 0. If )(0) = 0, then the Fourier sine
series converges at 0.
Proof. Exercise 19F.1 Hint: Imitate the methods of 8C. 2 E _
19G
Momentum representation & Heisenberg uncer-

tainty
Anyone who is not shocked by quantum theory has not understood it. Niels Bohr
Prerequisites: 3B, 6B, 19C.
Let : 1
3
1 C be the wavefunction of a quantum particle (e.g. an
electron). Fix t 1, and dene the instantaneous wavefunction
t
: 1
3
C
by
t
(x) = (x; t) for all x 1
3
. Recall from3B that
t
encodes the probability
distribution for the classical position of the particle at time t. However,
t
seems
to say nothing about the classical momentum of the particle. In Example 3B.2
on page 42, we stated (without proof) the wavefunction of a particle with a
particular known velocity. Now we make a more general assertion:
Suppose a particle has instantaneous wavefunction
t
: 1
3
C.
Let
t
: 1
3
C be the (3-dimensional) Fourier transform of
t
,
and dene
t
:=
t
_
p
/
_
for all p 1
3
. Then
t
is the wavefunction
for the particles classical momentum at time t. That is: if we dene

t
(p) := [
t
[
2
(p), |
t
|
2
2
for all p 1
3
, then
t
is the probability
distribution for the particles classical momentum at time t.
Recall that we can reconstruct
t
from
t
via the inverse Fourier transform.
Hence, the (positional) wavefunction
t
implicitly encodes the (momentum)
wavefunction
t
, and conversely the (momentum) wavefunction
t
implicitly
encodes the (positional) wavefunction
t
. This answers the question we posed
on page 38 of 3A. The same applies to multi-particle quantum systems:
Suppose an -particle quantum system has instantaneous (posi-
tion) wavefunction
t
: 1
3N
C. Let
t
: 1
3N
C be the
(3-dimensional) Fourier transform of
t
, and dene
t
:=
t
_
p
/
_
for all p 1
3N
. Then
t
is the joint wavefunction for the clas-
sical momenta of all the particles at time t. That is: if we dene

t
(p) := [
t
[
2
(p), |
t
|
2
2
for all p 1
3N
, then
t
is the joint prob-
ability distribution for the classical momenta of all the particles at
time t.
Because the momentum wavefunction contains exactly the same information
as the positional wavefunction, we can reformulate the Schrodinger equation in
momentum terms. For simplicity, we will only do this in the case of a single
particle. Suppose the particle is subjected to a potential energy function \ :
1
3
1. Let

\ be the Fourier transform of \ , and dene

\ :=
1
/
3
\
_
p
/
_
for all p 1
3
. Then the momentum wavefunction evolves according to the
momentum Schrodinger Equation:
i
t
(p; t) =
/
2
2:
[p[
2

t
(p) + (
\
t
)(p). (19G.1)
(here, if p = (j
1
, j
2
, j
3
), then [p[
2
:= j
2
1
+j
2
2
+j
2
3
). In particular, if the potential
eld is trivial, we get the free momentum Schrodinger equation:
i
t
(j
1
, j
2
, j
3
; t) =
/
2
2:
(j
2
1
+j
2
2
+j
2
3
) (j
1
, j
2
, j
3
; t).
Exercise 19G.1. Verify eqn.(19G.1) by applying the Fourier transform to the (po- E _
sitional) Schrodinger equation eqn.(3B.3) on page 41. Hint: Use Theorem 19B.7 on
page 496 to show that

t
(p) = [p[
2

t
(p). Use Theorem 19B.2(c) to show that
(
\
t
)(p,/) =

\
t
(p).
Exercise 19G.2. Formulate the momentum Schrodinger equation for an single par- E _
ticle conned to a 1-dimensional or 2-dimensional environment. Be careful how you
dene

\ .
Exercise 19G.3. Formulate the momentum Schrodinger equation for an -particle E _
quantum system. Be careful how you dene

\ .
Recall that Theorem 19B.8 said: if ) is a Gaussian distribution, then

)
is also a Gaussian (after multiplying by a scalar), but the variance of

) is
inversely proportional to the variance of ). This is an example of a general
phenomenon, called Heisenbergs Inequality. To state this formally, we need
some notation. Recall from 6B that L
2
(1) is the set of all square-integrable
complex-valued functions on 1 that is, all integrable functions ) : 1 C
such that |)|
2
< , where
|)|
2
:=
__

[)(r)[
2
dr
_
1/2
.
If ) L
2
(1), and r 1, then dene the uncertainty of ) around r to be

x
()) :=
1
|)|
2
2
_

[)(j)[
2
[j r[
2
dj.
19G.
Momentum representation & Heisenberg uncertainty 513

(In most physics texts, the uncertainty of ) is denoted by
x
); however, we will
not use this symbol because it looks too much like the Laplacian operator.)
Example 19G.1. (a) If ) L
2
(1), then (r) := )(r)
2
, |)|
2
2
is a probability
density function on 1 (why?). If r is the mean of the distribution (i.e.
r =
_
r(r) dr), then

x
()) =
_

(j) [j r[
2
dj
is the variance of the distribution. Thus, if describes the probability density
of a random variable A 1, then r is the expected value of A, and
x
()
measures the degee of uncertainty we have about the value of A. If
x
()
is small, then the distribution is tightly concentrated around r, so we can be
fairly condent that A is close to r. If
x
() is large, then the distribution
is broadly dispersed around r, so we really have only a vague idea where A
might be.
(b) In particular, suppose )(r) = exp
_
x
2
4
2
_
. Then )
2
,|)|
2
2
=
1
2
exp
_
x
2
2
2
_
is a Gaussian distribution with mean 0 and variance
2
. It follows that

0
()) =
2
.
(c) Suppose : 11 C is a one-dimensional wavefunction, and x t 1;
thus, the function
t
(r) = [
t
[
2
(r), |
t
|
2
is the probability density for the
classical position of the particle at time t in a one-dimensional environment
(e.g. an electron in a thin wire). If r is the mean of this distribution, then

x
(
t
) is the variance of the distribution; this reects our degree of uncertainty
about the particles classical position at time t.
Why the subscript r in
x
())? Why not just measure the uncertainty
around the mean of the distribution as in Example 19G.1? Three reasons. First,
because the distribution might not have a well-dened mean (i.e. the integral
_
r(r) dr might not converge). Second, because it is sometimes useful to

measure the uncertainty around other points in 1 besides the mean value. Third,
because we do not need to use the mean value to state the next result.
Theorem 19G.2. Heisenbergs Inequality
Let ) L
2
(1) be a nonzero function, and let

) be its Fourier transform. Then
for any r, j 1, we have
x
())
))
1
4
.
Example 19G.3. (a) If )(r) = exp
_
r
2
4
2
_
, then

)(j) =

exp
_
2
j
2
_
(Exercise 19G.4).
4
Thus,

)(j)
2
,
_
_
_
)
_
_
_
2
2
=
2
2
exp(2
2
j
2
) is a Gaussian E _
distribution with mean 0 and variance 1,4
2
. Thus, Example 19G.1(b) says
that
0
()) =
2
and
0
(
)) = 1,4
2
. Thus,

0
())
0
(
)) =

2
4
2
=
1
4
.
(b) Suppose
t
L
2
(1) is the instantaneous wavefunction for the position of a
particle at time t, so that
t
(p) =
t
(p,/) is the instantaneous wavefunction
for the momentum of the particle at time t. Then Heisenbergs Inequality
becomes Heisenbergs Uncertainty Principle: For any r, j 1,

x
(
t
)
/
2
4
p
(
t
)
and
p
(
t
)
/
2
4
x
(
t
)
(Exercise 19G.5). In other words: if our uncertainty
(
t
) about the E _
particles momentum is small, then our uncertainty
x
(
t
) about its position
must be big. Conversely, if our uncertainty
x
(
t
) about the particles position
is small, then our uncertainty
(
t
) about its momentum must be big.
In physics popularizations, the Uncertainty Principle is usually explained as
a practical problem of measurement precision: any attempt to measure an
electrons position (e.g. by deecting photons o of it) will impart some un-
predictable momentum into the particle. Conversely, any attempt to measure
its momentum disturbs its position. However, as you can see, Heisenbergs Un-
certainty Principle is actually an abstract mathematical theorem about Fourier
transforms it has nothing to do with the limitations of experimental equip-
ment or the unpredictable consequences of photon bombardment.
Proof of Heisenbergs Inequality. For simplicity, assume lim
x
r[)(r)[
2
= 0.
Case r = j = 0. Dene : 1 1 by (r) := r for all r 1. Thus,
t
(r) := 1 for all r 1. Observe that
|) |
2
2
=
_

[) [
2
(r) dr =
|)|
2
2
|)|
2
2
_

[)(r)[
2
[r[
2
dr
= |)|
2
2

0
()). (19G.2)
Also, Theorem 19B.7 implies that
()
t
) = i

). (19G.3)
Now,
|)|
2
2
:=
_

[)[
2
(r) dr
()
(r) [)(r)[
2
x=
x=
(r) ([)[
2
)
t
(r) dr
4
Hint: set := 2 in Theorem 19B.5 on page 495, and then apply it to Theorem 19B.8(a)
on page 497.
19H.
Laplace transforms 515

()
r ([)[
2
)
t
(r) dr
()
r 2 Re
_
)
t
(r))(r)
dr
= 2 Re
__

r)(r))
t
(r) dr
_
(19G.4)
Here, () is integration by parts, because
t
(r) = 1. Next, () is because
lim
x
r[)(r)[
2
= 0. Meanwhile, () is because [)[
2
(r) = )(r))(r), so that
([)[
2
)
t
(r) = )
t
(r))(r) + )(r))
t
(r) = 2 Re
_
)
t
(r))(r)
, where the last step

uses the identity . +. = 2 Re [.], with . = )
t
(r))(r). Thus,
1
4
|)|
4
2
()
2
2
4
Re
__

r)(r))
t
(r) dr
_
2
r)(r))
t
(r) dr
2
=

), )
t
_
(CBS)
_
_
)
_
_
2
2

_
_
)
t
_
_
2
2
= | )|
2
2

_
_
)
t
_
_
2
2
()

0
()) |)|
2
2

_
_
)
t
_
_
2
2 (Pl)

0
()) |)|
2
2
(2)
_
_
_
()
t
)
_
_
_
2
2
()
2
0
()) |)|
2
2

_
_
_i
)
_
_
_
2
2
= 2
0
()) |)|
2
2

_
_
_
)
_
_
_
2
2
()
2
0
()) |)|
2
2

0
(
))
_
_
_
)
_
_
_
2
2
(Pl)

0
()) |)|
2
2

0
(
)) |)|
2
2
=
0
())
0
(
)) |)|
4
2
.
Cancelling |)|
4
2
from both sides of this equation, we get
1
4

0
())
0
(
)),
as desired.
Here, () is by eqn.(19G.4), while (CBS) is the Cauchy-Bunyakowski-Schwarz
inequality (Theorem 6B.5 on page 108). Both () are by eqn.(19G.2). Both
(Pl) are by Plancharels theorem (Corollary 19C.2 on page 503) . Finally, ()
is by eqn.(19G.3).
Case r ,= 0 and/or j ,= 0. Exercise 19G.6 (Hint: Combine the case r = j = 0 E _
with Theorem 19B.3). 2
Exercise 19G.7. State and prove a form of Heisenbergs Inequality for a function E _
) L
1
(1
D
) for 1 2. Hint: You must compute the uncertainty in one coordinate at
a time. Integrate out all the other dimensions to reduce the 1-dimensional problem to
a one-dimensional problem, and then apply Theorem 19G.2.
19H
Laplace transforms
Recommended: 19A, 19B.
The Fourier transform

) is only well-dened if ) L
1
(1), which implies
that lim
t
[)(t)[ = 0 relatively quickly.
5
This is often inconvenient in physical
models where the function )(t) is bounded away from zero, or even grows without
bound as t . In some cases, we can handle this problem using a Laplace
transform, which can be thought of as a Fourier transform rotated by 90
o
in the
complex plane. The price we pay is that we must work on the half-innite line
1
,
:= [0, ), instead of the entire real line.
Let ) : 1
,
C. We say that ) has exponential growth if there are
constants 1 and 1 0 such that
[)(t)[ 1 c
t
, for all t 1
,
. (19H.1)
If 0, then inequality (19H.1) even allows lim
t
)(t) = , as long as )(t)
doesnt grow too quickly. (However, if < 0, then inequality (19H.1) requires
lim
t
)(t) = 0 exponentially fast). The exponential order of ) is the inmum
of all satisfying inequality (19H.1). Thus, if ) has exponential order
0
, then
(19H.1) is true for all
0
(but may or may not be true for =
0
).
Example 19H.1. (a) Fix : 0. If )(t) = t
r
for all t 1
,
, then ) has
exponential order 0.
(b) Fix : < 0 and t
0
0. If )(t) = (t + t
0
)
r
for all t 1
,
, then ) has
exponential order 0. (However, if t
0
0, then )(t) does not have exponential
growth, because in this case lim
tt
0
)(t) = , so inequality (19H.1) is always
false near t
0
).
(c) Fix 1. If )(t) = c
t
for all t 1
,
, then ) has exponential order .
(d) Fix j 1. If )(t) = sin(jt) or )(t) = cos(jt), then ) has exponential
order 0.
(e) If ) : 1
,
C has exponential order , and : 1 is any constant, then
: ) also has exponential order . If p : 1
,
C has exponential order ,
then ) +p has exponential order at most max, , and ) p has exponential
order +.
(f) Combining (a) and (e): any polynomial )(t) = c
n
t
n
+ c
2
t
2
+ c
1
t + c
0
has exponential order 0. Likewise, combining (d) and (e): any trigonometric
polynomial has exponential order 0.
Exercise 19H.1 Verify examples (a-f). E _
If c = r + ji is a complex number, recall that Re [c] := r. Let H
:=
c C ; Re [c] be the half of the complex plane to the right of the vertical
5
Actually, we can dene

f if f L
p
(1) for any p [1, 2], as discussed in 19C. However,
this isnt really that much of an improvement; we still need lim
t
[f(t)[ = 0 quickly.
19H.

line c C ; Re [c] = . In particular, H
0
:= c C ; Re [c] 0. If ) has
exponential order , then the Laplace transform of ) is the function L[)] :
H
C dened as follows:
For all : H
, L[)](:) :=
_

0
)(t) c
ts
dt. (19H.2)
Lemma 19H.2. If ) has exponential order , then the integral (19H.2) con-
verges for all : H
. Thus, L[)] is well-dened on H
.
Proof. Exercise 19H.2 2 E _
Example 19H.3. (a) If )(t) = 1, then ) has exponential order 0. For all : H
0
,
L[)](:) =
_

0
c
ts
dt =
c
ts
:
t=
t=0
()
(0 1)
:
=
1
:
.
Here () is because Re [:] 0.
(b) If 1 and )(t) = c
t
, then ) has exponential order . For all : H
,
L[)](:) =
_

0
c
t
c
ts
dt =
_

0
c
t(s)
dt
= =
c
t(s)
:
t=
t=0
()
(0 1)
:
=
1
:
.
Here () is because Re [ :] < 0 because Re [:] because : H
.
(c) If )(t) = t, then ) has exponential order 0. For all : H
0
,
L[)](:) =
_

0
t c
ts
dt
(p)
t c
ts
:
t=
t=0
_

0
c
ts
:
dt
=
(0 0)
:

c
ts
:
2
t=
t=0
=
(0 1)
:
2
=
1
:
2
,
where (p) is integration by parts.
(d) Suppose ) : 1
,
C has exponential order < 0. Extend ) to a
function ) : 1 C by dening )(t) = 0 for all t < 0. Then the Fourier
transform

) of ) is well-dened, and for all j 1, we have 2

)(j) = L[)](ji)
(Exercise 19H.3). E _
(e) Fix j 1. If )(t) = cos(jt), then L[)] =
:
:
2
+j
2
. If )(t) = sin(jt), then
L[)] =
j
:
2
+j
2
.
Exercise 19H.4 Verify (e). Hint: recall that exp(ijt) = cos(jt) +i sin(jt). E _
Example 19H.3(d) suggests that most properties of the Fourier transform
should translate into properties of the Laplace transform, and vice versa. First,
like Fourier, the Laplace transform is invertible.
Theorem 19H.4. Laplace Inversion Formula
Suppose ) : 1
,
C has exponential order , and let 1 := L[)]. Then for
any xed :
r
, and any t 1
,
, we have
)(t) =
1
2i
_

1(:
r
+:
i
i) exp(t:
r
+t:
i
i) d:
i
. (19H.3)
In particular, if p : 1
,
C, and L[p] = L[)] on any innite vertical strip in
C, then we must have p = ).
Proof. Exercise 19H.5 (Hint: Use an argument similar to Example 19H.3(d) to E _
represent 1 as a Fourier transform. Then apply the Fourier Inversion Formula.) 2
The integral (19H.3) is called the Laplace inversion integral, and is de-
noted by L
1
[1]. The integral (19H.3) is sometimes written
)(t) =
1
2i
_
sr+i
sri
1(:) exp(t:) d:.
The integral (19H.3) is can be treated as a contour integral along the vertical
line c C ; Re [c] = :
r
in the complex plane, and evaluated using residue cal-
culus
6
. However, in many situations, it is neither easy nor particularly necessary
to explicitly compute (19H.3); instead, we can determine the inverse Laplace
transform by inspection, by simply writing 1 as a sum of Laplace transforms
of functions we recognize. Most books on ordinary dierential equations contain
an extensive table of Laplace transforms and their inverses, which is useful for
this purpose.
Example 19H.5. Suppose 1(:) =
3
:
+
5
: 2
+
7
:
2
. Then by inspecting
Example 19H.3(a,b,c), we deduce that )(t) = L
1
[1](t) = 3 + 5c
2t
+ 7t.
Most of the results about Fourier transforms from Section 19B have equiva-
lent formulations for Laplace transforms.
Theorem 19H.6. Properties of the Laplace transform
Let ) : 1
,
C have exponential order .
(a) (Linearity) Let p : 1
,
C have exponential order , and let =
max, . Let /, c C. Then /) + cp has exponential order at most ,
and for all : H
, L[/) +cp](:) = /L[)](:) +cL[p](:).

6
See 18H on page 472 for a discussion of residue calculus and its application to improper
integrals. See [Fis99, 5.3] for some examples of computing Laplace inversion integrals using
this method.
19H.

(b) (Transform of a derivative)
(i) Suppose ) (
1
(1
,
) (i.e. ) is continuously dierentiable on 1
,
) and
)
t
has exponential order . Let = max, . Then for all : H
,
L[)
t
](:) = :L[)](:) )(0).
(ii) Suppose ) (
2
(1
,
), )
t
has exponential order
1
and )
tt
has expo-
nential order
2
. Let = max,
1
,
2
. Then for all : H
,
L[)
tt
](:) = :
2
L[)](:) )(0): )
t
(0).
(iii) Let N, and suppose ) (
N
(1
,
). Suppose )
(n)
has exponential
order
n
for all n [1...]. Let = max,
1
, . . . ,
N
. Then for
all : H
,
L[)
(N)
](:) = :
N
L[)](:))(0):
N1
)
t
(0):
N2
)
tt
(0):
N3
)
(N2)
(0) :)
(N1)
(0).
(c) (Derivative of a transform) For all n N, the function p
n
(t) = t
n
)(t) also
has exponential order .
If ) is piecewise continuous, then the function 1 = L[)] : H
C is
(complex)-dierentiable,
7
and for all : H
, 1
t
(:) = L[p
1
](:), 1
tt
(:) =
L[p
2
](:), and in general 1
(n)
(:) = (1)
n
L[p
n
](:).
(d) (Translation) Fix T 1
,
, and dene p : 1
,
C by p(t) = )(t T) for
t T and p(t) = 0 for t [0, T). Then p also has exponential order . For
all : H
, L[p](:) = c
Ts
L[)](:).
(e) (Dual translation) For all 1, the function p(t) = c
t
)(t) has exponen-
tial order +. For all : H
+
, L[p](:) = L[)](: ).
Proof. Exercise 19H.6 (Hint: Imitate the proofs of Theorems 19B.2(a), 19B.7 E _
and 19B.3.) 2
Exercise 19H.7. Show by a counterexample that Theorem 19H.6(d) is false if E _
T < 0.
Corollary 19H.7. Fix n N. If )(t) = t
n
for all t 1
,
, then L[)](:) =
n!
:
n+1
for all : H
0
.
Proof. Exercise 19H.8 (Hint: Combine Theorem 19H.6(c) with Example E _
19H.3(a).) 2
Exercise 19H.9. Combine Corollary 19H.7 with Theorem 19H.6(b) to get a for- E _
mula for the Laplace transform of any polynomial.
Exercise 19H.10. Combine Theorem 19H.6(c,d) with Example 19H.3(a) to get a E _
formula for the Laplace transform of )(t) =
1
(1 +t)
n
for all n N.
Corollary 19H.7 does not help us to compute the Laplace transform of )(t) =
t
r
when : is not an integer. To do this, we must introduce the gamma function
: 1
,
C, which is dened
(:) :=
_

0
t
r1
c
t
dt, for all : 1
,
.
This is regarded as a generalized factorial because of the following properties.
Lemma 19H.8.
(a) (1) = 1.
(b) For any : 1
,
, (: + 1) = : (:).
(c) Thus, for any n N, (n + 1) = n!. (For example, (5) = 4! = 24.)
Exercise 19H.12. (a) Show that (1,2) =

. E _
(b) Deduce that (3,2) =

,2 and (5,2) =
3
4
.
The gamma function is useful when computing Laplace transforms because
of the next result.
Proposition 19H.9. Laplace transform of )(t) = t
r
Fix : 1, and let )(t) := t
r
for all t 1
,
. Then L[)](:) =
(: + 1)
:
r+1
for all
: H
0
.
Remark. If : < 0, then technically, )(t) = t
r
does not have exponential
growth, as noted in Example 19H.1(b). Hence Lemma 19H.2 does not apply.
However, the Laplace transform integral (19H.2) converges in this case anyways,
7
See 18A on page 415 for more about complex dierentiation.
19H.

because although lim
t0
t
r
= , it goes to innity slowly, so that the integral
_
1
0
t
r
dt is still nite.
Example 19H.10. (a) If : N and )(t) = t
r
, then Proposition 19H.9 and
Lemma 19H.8(c) together imply L[)](:) =
:!
:
r+1
, in agreement with Corollary
19H.7.
(b) Let : = 1,2. Then )(t) =

t, and Proposition 19H.9 says L[)](:) =
(3,2)
:
3/2
=
2:
3/2
, where the last step is by Exercise 19H.12(b).
(c) Let : = 1,2. Then )(t) =
1
t
, and Proposition 19H.9 says L[)](:) =
(1,2)
:
1/2
=
_
:
, where the last step is by Exercise 19H.12(a).
Theorems 19B.2(c) showed how the Fourier transform converts function con-
volution into multiplication. A similar property holds for the Laplace transform.
Let ), p : 1
,
C be two functions. The convolution of ), p is the function
) p : 1
,
C dened
) p(T) :=
_
T
0
)(T t)p(t) dt, for all T 1
,
.
Note that ) p(T) is an integral over a nite interval [0, T]; thus it is well-dened
no matter how fast )(t) and p(t) grow as t .
Theorem 19H.11. Let ), p : 1
,
C have exponential order. Then L[)
p] = L[)] L[p] wherever all these functions are dened.
Theorem 19H.6(b) makes the Laplace transform a powerful tool for solving
linear ordinary dierential equations.
Proposition 19H.12. Laplace solution to linear ODE
Let ), p : 1
,
C have exponential order and respectively, and let
= max, . Let 1 := L[)] and G := L[p]. Let c
0
, c
1
, . . . , c
n
C be
constants. Then ) and p satises the linear ODE
p = c
0
) +c
1
)
t
+c
2
)
tt
+ +c
n
)
(n)
(19H.4)
if and only if 1 and G satisfy the algebraic equation
G(:) =
_
c
0
+c
1
: +c
2
:
2
+c
3
:
3
+ + c
n
:
n
_
1(:)
_
c
1
+c
2
:
1
+c
3
:
2
+ +c
n
:
n1
_
)(0)
_
c
2
+ c
3
: + +c
n
:
n2
_
)
t
(0)
.
.
.
.
.
.
.
.
.
.
.
.
(c
n1
+c
n
:) )
(n2)
(0)
c
n
)
(n1)
(0).
for all : H
. In particular, ) satises ODE (19H.4) and homogeneous boundary

conditions )(0) = )
t
(0) = )
tt
(0) = = )
(n1)
(0) = 0 if and only if
1(:) =
G(:)
c
0
+c
1
: +c
2
:
2
+ +c
3
:
3
+ +c
n
:
n
for all : H
Proof. Exercise 19H.15 (Hint: Apply Theorem 19H.6(b), then reorder terms.) E _
2
Laplace transforms can also be used to solve partial dierential equations.
Let X 1 be some one-dimensional domain, let ) : X 1
,
C, and write
)(r; t) as )
x
(t) for all r X. Fix 1. For all r X suppose that )
x
has
exponential order , so that L[)
x
] is a function H
C. Then we dene
L[)] : X H
C by L[)](r; :) = L[)
x
](:) for all r X and : H
.
Proposition 19H.13. Suppose
x
)(r, t) is dened for all (r, t) int (X)1
+
.
Then
x
L[)](r, :) is dened for all (r, :) int (X) H
, and
x
L[)](r, :) =
L[
x
)](r, :).
Proof. Exercise 19H.16 (Hint: Apply Theorem 0G.1 (on page 567) to the Laplace E _
transform integral (19H.2).) 2
By iterating Proposition 19H.13, we have
n
x
L[)](r, :) = L[
n
x
)](r, :) for
any n N. Through Proposition 19H.13 and Theorem 19H.6(b), we can convert
a PDE about ) into an ODE involving only the r-derivatives of L[)].
Example 19H.14. Let ) : X 1
,
C and let 1 := L[)] : X H
C.
(a) (heat equation) Dene )
0
(r) := )(r, 0) for all r X (the initial tem-
perature distribution). Then ) satises the one-dimensional heat equation
t
)(r, t) =
2
x
)(r, t) if and only if, for every : H
, the function 1
s
(r) =
1(r, :) satises the second-order linear ODE
2
x
1
s
(r) = : 1
s
(r) )
0
(r), for all r X.
(b) (wave equation) Dene )
0
(r) := )(r, 0) and )
1
(r) :=
t
)(r, 0) for all r X
(the initial position and initial velocity, respectively). Then ) satises the
one-dimensional wave equation
2
t
)(r, t) =
2
x
)(r, t) if and only if, for every
: H
, the function 1
s
(r) = 1(r, :) satises the second-order linear ODE
2
x
1
s
(r) = :
2
1
s
(r) : )
0
(r) )
1
(r), for all r X.
(Exercise 19H.17 Verify examples (a) and (b).) E _
We can then use solution methods for ordinary dierential equations to solve
for 1
s
for all : H
, obtain an expression for the function 1, and then apply

the Laplace Inversion Theorem 19H.4 to obtain an expression for ). We will
not pursue this approach further here; however, we will develop a very similar
approach in the Chapter 20 using Fourier transforms. For more information, see
[Asm05, Chapt.8], [Fis99, 5.3-5.4], [Hab87, Chapt.13], or [Bro89, Chapt.5]
19I Practice problems
1. Suppose )(r) =
_
1 if 0 < r < 1;
0 otherwise
, as in Example 19A.4 on page 490.
Check that

)(j) =
1 c
i
2ji
2. Compute the one-dimensional Fourier transforms of p(r), when:
(a) p(r) =
_
1 if < r < 1 ;
0 otherwise
,
(b) p(r) =
_
1 if 0 < r < ;
0 otherwise
.
3. Let A, Y 0, and let )(r, j) =
_
1 if 0 r A and 0 j Y ;
0 otherwise.
.
Compute the two-dimensional Fourier transform of )(r, j). What does the
Fourier Inversion formula tell us?
4. Let ) : 1 1 be the function dened: )(r) =
_
r if 0 r 1
0 otherwise
(Fig.19I.1) Compute the Fourier Transform of ).
5. Let )(r) = r exp
_
r
2
2
_
. Compute the Fourier transform of ).
6. Let 0, and let p(r) =
1
2
+r
2
. Example 19A.8 claims that p(j) =
1
2
c
[[
. Verify this. Hint: Use the Fourier Inversion Theorem.
1 2 3 -1
1
Figure 19I.1: Problem #4
7. Fix j 0, and let /
y
(r) =
j
(r
2
+j
2
)
(this is the half-space Poisson
Kernel from 17E and 20C(ii)). Compute the one-dimensional Fourier
transform

/
y
(j) =
1
2
_

/
y
(r) exp
_
jir
_
dj.
8. Let )(r) =
2r
(1 +r
2
)
2
. Compute the Fourier transform of ).
9. Let )(r) =
_
1 if 4 < r < 5;
0 otherwise.
Compute the Fourier transform
)(j).
10. Let )(r) =
rcos(r) sin(r)
r
2
. Compute the Fourier transform

)(j).
11. Let ), p L
1
(1), and let /(r) = )(r) + p(r). Show that, for all j 1,
/(j) =

)(j) + p(j).
12. Let ), p L
1
(1), and let / = ) p. Show that for all j 1,

/(j) =
2

)(j) p(j).
Hint: exp (ijr) = exp (ijj) exp
_
ij(r j)
_
.
13. Let ), p L
1
(1), and let /(r) = )(r) p(r). Suppose

/ is also in L
1
(1).
Show that, for all j 1,

/(j) = (
) p)(j).
Hint: Combine problem #12 with the Strong Fourier Inversion Formula
(Theorem 19A.5 on page 491).
14. Let ) L
1
(1). Fix 1, and dene p : 1 C by: p(r) = )(r + ).
Show that, for all j 1, p(j) = c
i

)(j).
15. Let ) L
1
(1). Fix 1 and dene p : 1 C by p(r) = c
xi
)(r).
Show that, for all j 1, p(j) =

)(j ).
16. Suppose ) L
1
(1). Fix 0, and dene p : 1 C by: p(r) = )
_
r
_
.
Show that, for all j 1, p(j) =

)( j).
17. Suppose ) : 1 1 is dierentiable, and that ) L
1
(1) and p := )
t

L
1
(1). Assume that lim
x
)(r) = 0. Show that p(j) = ij

)(j).
18. Let (
t
(r) =
1
2
t
exp
_
x
2
4t
_
be the Gauss-Weierstrass kernel. Recall that
(
t
(j) =
1
2
c
2
t
. Use this to construct a simple proof that, for any
:, t 0, (
t
(
s
= (
t+s
.
(Hint: Use problem #12. Do not compute any convolution integrals, and
do not use the solution to the heat equation argument from Problem # 8
on page 413.)
Remark. Because of this result, probabilists say that the set (
t
t1
+
forms a stable family of probability distributions on 1. Analysts say that
(
t
t1
+
is a one-parameter semigroup under convolution.
527
Chapter 20
Fourier transform solutions to
PDEs
Mathematics compares the most diverse phenomena and discovers the secret analogies
that unite them. Jean Joseph Fourier
We will now see that the Fourier series solutions to the PDEs on a bounded
domain (Chapters 11-14) generalize to Fourier transform solutions on the un-
bounded domain in a natural way.
20A The heat equation
20A(i) Fourier transform solution
Prerequisites: 1B, 19A, 5B, 0G. Recommended: 11A, 12B, 13A, 19D, 19E.
Proposition 20A.1. Heat equation on an innite rod
Let 1 : 1 1 be a bounded function (of j 1).
(a) Dene n : 1 1
+
1 by
n(r; t) :=
_

1(j) exp(jri) c
2
t
dj, (20A.1)
for all t 0 and all r 1. Then n is a smooth function and satises the
heat equation.
(b) In particular, suppose ) L
1
(1), and

)(j) = 1(j). If we dene n(r; 0) :=
)(r) for all r 1, and dene n(r; t) by eqn.(20A.1), when t 0, then n is
continuous on 1 1
,
, and is a solution to the heat equation with initial
conditions n(r; 0) = )(r).
Proof. Exercise 20A.1 (Hint: Use Proposition 0G.1 on page 567.) 2 E _
528 DRAFT Chapter 20. Fourier transform solutions to PDEs
_
1 if 1 < r < 1;
0 otherwise.
We know
from Example 19A.3 on page 489 that

)(j) =
sin(j)
j
. Thus,
n(r, t) =
_

)(j) exp(jri) c
2
t
dj =
_

sin(j)
j
exp(jri) c
2
t
dj.
(Exercise 20A.2 Verify that n satises the one-dimensional heat equation and the E _
specied initial conditions.)
Example 20A.3: The Gauss-Weierstrass kernel
For all r 1 and t 0, dene the Gauss-Weierstrass Kernel: (
t
(r) :=
1
2
t
exp
_
r
2
4t
_
(see Example 1B.1(c) on page 6). Fix t 0; then setting
=

2t in Theorem 19B.8(b), we get
(
t
(j) =
1
2
exp
_
(
2t)
2
j
2
2
_
=
1
2
exp
_
2tj
2
2
_
=
1
2
c
2
t
.
Thus, applying the Fourier Inversion formula (Theorem 19A.1 on page 488),
we have:
((r, t) =
_

(
t
(j) exp(jri) dj =
1
2
_

2
t
exp(jri) dj,
which, according to Proposition 20A.1, is a smooth solution of the heat equa-
tion, where we take 1(j) to be the constant function: 1(j) = 1,2. Thus, 1
is not the Fourier transform of any function ). Hence, the Gauss-Weierstrass
kernel solves the heat equation, but the initial conditions (
0
do not cor-
respond to a function, but instead a dene more singular object, rather like
an innitely dense concentration of mass at a single point. Sometimes (
0
is
called the Dirac delta function, but this is a misnomer, since it isnt really
a function. Instead, (
0
is an example of a more general class of objects called
distributions.
Proposition 20A.4. Heat equation on an innite plane
Let 1 : 1
2
C be some bounded function (of (j, ) 1
2
).
(a) Dene n : 1
2
1
+
1 by
n(r, j; t) :=
_
1
2
1(j, ) exp
_
(jr +j) i
_
c
(
2
+
2
)t
dj d, (20A.2)
for all t 0 and all (r, j) 1
2
. Then n is continuous on 1
3
1
+
and
satises the two-dimensional heat equation.
20A. The heat equation 529
1
(1
2
), and

)(j, ) = 1(j, ). If we dene
n(r, j, 0) := )(r, j) for all (r, j) 1
2
, and dene n(r, j, t) by eqn.(20A.2)
when t 0, then n is continuous on 1
2
1
,
, and is a solution to the heat
equation with initial conditions n(r, j, 0) = )(r, j).
Example 20A.5. Let A, Y 0 be constants, and consider the initial conditions:
)(r, j) =
_
1 if A r A and Y j Y ;
0 otherwise.
From Example 19D.4 on page 505, the Fourier transform of )(r, j) is given:
)(j, ) =
sin(jA) sin(Y )
2
j
.
Thus, the corresponding solution to the two-dimensional heat equation is:
n(r, j, t) =
_
1
2
)(j, ) exp
_
(jr +j) i
_
c
(
2
+
2
)t
dj d
=
_
1
2
sin(jA) sin(Y )
2
j
exp
_
(jr +j) i
_
c
(
2
+
2
)t
dj d .
Proposition 20A.6. Heat equation in innite space
Let 1 : 1
3
C be some bounded function (of 1
3
).
(a) Dene n : 1
3
1
+
1 by
n(x; t) :=
_
1
3
1() exp
_
x i
_
c
||
2
t
d, (20A.3)
for all t 0 and all x 1
3
, (where ||
2
:= j
2
1
+ j
2
2
+ j
2
3
). Then n is
continuous on 1
3
1
+
and satises the three-dimensional heat equation.
1
(1
3
), and

)() = 1(). If we dene
n(x, 0) := )(x) for all x 1
3
, and dene n(x, t) by eqn.(20A.3) when
t 0, then n is continuous on 1
3
1
,
, and is a solution to the heat
equation with initial conditions n(x, 0) = )(x).
Example 20A.7: A ball of heat
Suppose the initial conditions are: )(x) =
_
1 if |x| 1;
0 otherwise.
Setting 1 = 1 in Example 19E.3 (p.507) yields the three-dimensional Fourier
transform of ):
)() =
1
2
2
_
sin ||
||
3

cos ||
||
2
_
.
The resulting solution to the heat equation is:
n(x; t) =
_
1
3
)() exp
_
x i
_
c
||
2
t
d
=
1
2
2
_
1
3
_
sin ||
||
3

cos ||
||
2
_
exp
_
x i
_
c
||
2
t
d.
20A(ii) The Gaussian convolution formula, revisited
Prerequisites: 17C(i), 19B, 20A(i).
Recall from 17C(i) on page 385 that the Gaussian Convolution formula
solved the initial value problem for the heat equation by locally averaging the
initial conditions. Fourier methods provide another proof that this is a solution
to the heat equation.
Theorem 20A.8. Gaussian convolutions and the heat equation
Let ) L
1
(1), and let (
t
(r) be the Gauss-Weierstrass kernel from Example
20A.3. For all t 0, dene l
t
:= ) (
t
; in other words, for all r 1,
l
t
(r) :=
_

)(j) (
t
(r j) dj.
Also, for all r 1, dene l
0
(r) := )(r). Then l is continuous on 11
,
, and
is a solution to the Heat Equation with initial conditions l(r, 0) = )(r).
Proof. l(r, 0) = )(r) by denition. To show that l satises the heat equation,
we will show that it is in fact equal to the Fourier solution n described in
Theorem 20A.1 on page 527. Fix t 0, and let n
t
(r) = n(r, t); recall that, by
denition
n
t
(r) =
_

)(j) exp(jri) c
2
t
dj =
_

)(j)c
2
t
exp(jri) dj
Thus, Proposition 19A.2 on page 489 says that
n
t
(j) =

)(j) c
t
2
()
2

)(j)

(
t
(j). (20A.4)
Here, () is because Example 20A.3 on page 528 says that c
t
2
= 2

(
t
(j).
But remember that l
t
= ) (
t
, so, Theorem 19B.2(b) on page 494 says
l
t
(j) = 2

)(j)

(
t
(j). (20A.5)
Thus (20A.4) and (20A.5) mean that

l
t
= n
t
. But then Proposition 19A.2 on
page 489 implies that n
t
(r) = l
t
(r). 2
For more discussion and examples of the Gaussian convolution approach to
the heat equation, see 17C(i) on page 385.
Exercise 20A.5. State and prove a generalization of Theorem 20A.8 to solving the E _
1-dimensional heat equation, for 1 2.
20B The wave equation
20B(i) Fourier transform solution
Prerequisites: 2B, 19A, 5B, 0G. Recommended: 11B, 12D, 19D, 19E, 20A(i).
Proposition 20B.1. Wave equation on an innite wire
Let )
0
, )
1
L
1
(1) be twice-dierentiable, and suppose )
0
and )
1
have Fourier
transforms

)
0
and

)
1
, respectively. Dene n : 1 1
,
1 by
n(r, t) =
_

)
0
(j) cos(jt) +

)
1
(j)
j
sin(jt)
_
exp(jri) dj.
Then n is a solution to the one-dimensional wave equation with initial position
n(r, 0) = )
0
(r), and initial velocity
t
n(r, 0) = )
1
(r), for all r 1.
Proof. Exercise 20B.1 (Hint: Show that this solution is equivalant to the E _
dAlembert solution of Proposition 17D.5 on page 398.) 2
Example 20B.2. Fix 0, and suppose we have initial position )
0
(r) := c
[x[
for all r 1, and initial velocity )
1
0. From Example 19A.7 on page 491,
we know that

)
0
(j) =
2
(
2
+j
2
)
. Thus, Proposition 20B.1 says:
n(r, t) =
_

)
0
(j) exp(jri) cos(jt) dj
=
_

2
(
2
+j
2
)
exp(jri) cos(jt) dj.
(Exercise 20B.2 Verify that n satises the one-dimensional wave equation and E _
the specied initial conditions.)
Proposition 20B.3. Wave equation on an innite plane
Let )
0
, )
1
L
1
(1
2
) be twice dierentiable functions, whose Fourier transforms
)
0
and

)
1
decay fast enough that
_
1
2
(j
2
+
2
)
)
0
(j, )
dj d <
and
_
1
2
_
j
2
+
2
)
1
(j, )
dj d < .
(20B.6)
Dene n : 1
2
1
,
1 by
n(r, j, t) =
_
1
2
)
0
(j, ) cos
_
_
j
2
+
2
t
_
exp
_
(jr +j) i
_
dj d.
+
_
1
2
)
1
(j, )
_
j
2
+
2
sin
_
_
j
2
+
2
t
_
exp
_
(jr +j) i
_
dj d.
Then n is a solution to the two-dimensional wave equation with initial position
n(r, j, 0) = )
0
(r, j), and initial velocity
t
n(r, j, 0) = )
1
(r, j), for all (r, j)
1
2
.
Proof. Exercise 20B.3 (Hint: Equation (20B.6) makes the integral absolutely E _
convergent, and also enables you to apply Proposition 0G.1 on page 567 to compute
the relevant derivatives of n.) 2
Example 20B.4. Let , 0 be constants, and suppose we have initial
position )
0
0, and initial velocity )
1
(r, j) =
1
(
2
+r
2
)(
2
+j
2
)
for all
(r, j) 1
2
. By adapting Example 19A.8 on page 492, one can check that
)
1
(j, ) =
1
4
exp
_
[j[ [[
_
.
Thus, Proposition 20B.3 says
n(r, j, t) =
_
1
2
)
1
(j, )
_
j
2
+
2
sin
_
_
j
2
+
2
t
_
exp
_
(jr +j) i
_
dj d
=
_
1
2
sin
_
_
j
2
+
2
t
_
exp
_
(jr +j) i [j[ [[
_
4
_
j
2
+
2
dj d.
(Exercise 20B.4 Verify that n satises the two-dimensional wave equation and E _
the specied initial conditions.)
Proposition 20B.5. Wave equation in innite space
Let )
0
, )
1
L
1
(1
3
) be twice dierentiable functions whose Fourier transforms
)
0
and

)
1
decay fast enough that
_
1
3
||
2
)
0
()
d <
and
_
1
3
||
)
1
()
d < .
(20B.7)
Dene n : 1
3
1
,
1 by
n(x, t) :=
_
1
3
_
)
0
() cos (|| t) +

)
1
()
||
sin (|| t)
_
exp ( xi) . d
Then n is a solution to the three-dimensional wave equation with initial position
n(x, 0) = )
0
(x) and initial velocity
t
n(x, 0) = )
1
(x), for all x 1
3
.
Proof. Exercise 20B.5 (Hint: Equation (20B.7) makes the integral absolutely E _
convergent, and also enables you to apply Proposition 0G.1 on page 567 to compute
the relevant derivatives of n.) 2
Exercise 20B.6. Show that the decay conditions (20B.6) or (20B.7) are satised if E _
)
0
and )
1
are asymptotically at in the sense that lim
|x|
[)(x)[ = 0 and lim
|x|
[)(x)[ =
0, while (
i
j
)) L
1
(1
2
) for all i, , 1, ..., 1 (where 1 = 2 or 3).
Hint. Apply Theorem 19B.7 on page 496 to compute the Fourier transforms of the
derivative functions
i
j
); conclude that the function

) must itself decay at a certain
speed.
20B(ii) Poissons spherical mean solution; Huygens principle
Prerequisites: 17A, 19E, 20B(i). Recommended: 17D, 20A(ii).
The Gaussian Convolution formula of 20A(ii) solves the initial value problem
for the heat equation in terms of a kind of local averaging of the initial con-
ditions. Similarly, dAlemberts formula (17D) solves the initial value problem
for the one-dimensional wave equation in terms of a local average.
There is an analogous result for higher-dimensional wave equations. To ex-
plain it, we must introduce the concept of spherical averages. Let ) : 1
3
1
be some integrable function. If x 1
3
is a point in space, and 1 0, then the
spherical average of ) at x, of radius 1, is dened:
M
R
)(x) :=
1
41
2
_
S(R)
)(x +s) ds.
Here, S(1) :=
_
s 1
3
; |s| = 1
_
is the sphere around 0 of radius 1. The
total surface area of the sphere is 41
2
; we divide by this quantity to obtain
an average. We adopt the notational convention that M
0
)(x) := )(x). This is
justied by the next exercise.
Exercise 20B.7. Suppose ) is continuous at x. Show that lim
R0
M
R
)(x) = )(x). E _
Theorem 20B.6. Poissons Spherical Mean Solution to wave equation

(a) Let )
1
L
1
(1
3
) be twice-dierentiable, and such that

)
1
satises eqn.(20B.7)
on page 533. Dene : 1
3
1
,
1 by
(x; t) := t M
t
)
1
(x), for all x 1
3
and t 0.
Then is a solution to the wave equation with
Initial Position: (x, 0) = 0; Initial Velocity:
t
(x, 0) = )
1
(x).
(b) Let )
0
L
1
(1
3
) be twice-dierentiable and such that

)
0
satises eqn.(20B.7)
on page 533. For all x 1
3
and t 0, dene \(x; t) := t M
t
)
0
(x), and
then dene n : 1
3
1
,
1 by
n(x; t) :=
t
\(x; t), for all x 1
3
and t 0.
Then n is a solution to the wave equation with
Initial Position: n(x, 0) = )
0
(x); Initial Velocity:
t
n(x, 0) = 0.
(c) Let )
0
, )
1
L
1
(1
3
) be as in (a) and (b), and dene n : 1
3
1
,
1 by
n(x; t) := n(x; t) + (x; t), for all x 1
3
and t 0,
where n is as in Part (b) and is as in Part (a). Then n is the unique
solution to the wave equation with
Initial Position: n(x, 0) = )
0
(x); Initial Velocity:
t
n(x, 0) = )
1
(x).
Proof. We will prove Part (a). First we will need a certain calculation.
Claim 1: For any 1 0, and any 1
3
,
_
S(R)
exp
_
si
_
ds =
41 sin (|| 1)
||
.
Proof. By spherical symmetry, we can rotate the vector without aecting
the value of the integral, so rotate until it becomes = (j, 0, 0), with
j 0. Thus, || = j, and, if a point s S(1) has coordinates (:
1
, :
2
, :
3
)
in 1
3
, then s = j :
1
. Thus, the integral simplies to:
_
S(R)
exp ( si) ds =
_
S(R)
exp (j :
1
i) ds
We will integrate using a spherical coordinate system (, ) on the sphere,
where 0 < < and < < , and where
(:
1
, :
2
, :
3
) = 1 (cos(), sin() sin(), sin() cos()) .
On the sphere of radius 1, the surface area element is ds = 1
2
sin() d d.
Thus,
_
S(R)
exp (j :
1
i) ds =
_

0
_

exp (j 1 cos() i) 1
2
sin() d d
()
2
_

0
exp (j 1 cos() i) 1
2
sin() d
()
2
_
R
R
exp (j :
1
i) 1 d:
1
=
21
ji
exp
_
j :
1
i
_
s
1
=R
s
1
=R
= 2
21
j

_
c
Ri
c
Ri
2i
_
()
41
j
sin (j1) .
() The integrand is constant in the coordinate. () Making substitution
:
1
= 1cos(), so d:
1
= 1sin() d. () By Eulers Formula (see page
551).
Claim 1
Now, by Proposition 20B.5 on page 533, a solution to the three-dimensional
wave equation with zero initial position and initial velocity )
1
is given by:
n(x, t) =
_
1
3
)
1
()
sin (|| t)
||
exp ( xi) d. (20B.8)
However, if we set 1 = t in Claim 1, we have:
sin (|| t)
||
=
1
4t
_
S(t)
exp ( si) ds.
Thus,
sin (|| t)
||
exp ( xi) = exp ( xi)
1
4t
_
S(t)
exp ( si) ds
=
1
4t
_
S(t)
exp
_
xi + si
_
ds
=
1
4t
_
S(t)
exp
_
(x +s) i
_
ds.
Substituting this into (20B.8), we get:
n(x, t) =
_
1
3
)
1
()
4t

_
_
S(t)
exp
_
(x +s) i
_
ds
_
d
()
1
4t
_
S(t)
_
1
3
)
1
() exp
_
(x +s) i
_
d ds
()
1
4t
_
S(t)
)
1
(x +s) ds = t
1
4t
2
_
S(t)
)
1
(x +s) ds
= t M
t
)
1
(x).
() We simply interchange the two integrals
1
. () This is just the Fourier
Inversion Theorem 19E.1 on page 507.
Part (b) is Exercise 20B.8 . Part (c) follows by combining Part (a) and E _
Part (b). 2
Corollary 20B.7. Huygens principle
Let )
0
, )
1
L
1
(1
3
), and suppose there is some bounded region K 1
3
such
that )
0
and )
1
are zero outside of K that is: )
0
(y) = 0 and )
1
(y) = 0 for all
y , K (see Figure 20B.1A). Let n : 1
3
1
,
1 be the solution to the wave
equation with initial position )
0
and initial velocity )
1
, and let x 1
3
.
1
This actually involves some subtlety, which we will gloss over.
20C. The Dirichlet problem on a half-plane 537
t R
x
K K K K
x x
x
R
(1) Wave originates
inside K at time t=0
(2) If t<R, wave
hasnt yet reached x
(3) Wave reaches
x around t R.
(4) For t >>R, wave
has completely
passed by x.
t<R
t>>R
t=0
Figure 20B.1: Huygens principle.
(a) Let 1 be the distance from K to x. If t < 1 then n(x; t) = 0 (Figure
20B.1B).
(b) If t is large enough that K is entirely contained in a ball of radius t around
x, then n(x; t) = 0 (Figure 20B.1D).
Part (a) of Huygens Principle says that, if a sound wave originates in the
region K at time 0, and x is of distance 1 then it does not reach the point x
before time 1. This is not surprising; it takes time for sound to travel through
space. Part (b) says that the soundwave propagates through the point x in a
nite amount of time, and leaves no wake behind it. This is somewhat more
surprising, but corresponds to our experience; sounds travelling through open
spaces do not reverberate (except due to echo eects). It turns out, however,
that Part (b) of the theorem is not true for waves travelling in two dimensions
(e.g. ripples on the surface of a pond).
20C The Dirichlet problem on a half-plane
Prerequisites: 1C, 19A, 5C, 0G. Recommended: 12A, 13B, 19D, 19E.
In 12A and 13B, we saw how to solve Laplaces equation on a bounded
domain such as a rectangle or a cube, in the context of Dirichlet boundary con-
ditions. Now consider the half-plane domain H :=
_
(r, j) 1
2
; j 0
_
. The
boundary of this domain is just the r axis: H = (r, 0) ; r 1; thus, bound-
ary conditions are imposed by choosing some function / : 1 1. Figure 17E.1
on page 403 illustrates the corresponding Dirichlet problem: nd a continuous
function n : H 1 such that
1. n satises the Laplace equation: n(r, j) = 0 for all r 1 and j 0.
2. n satises the nonhomogeneous Dirichlet boundary condition: n(r, 0) =
/(r) for all r 1.
20C(i) Fourier solution
Heuristically speaking, we will solve the problem by dening n(r, j) as a con-
tinuous sequence of horizontal bres, parallel to the r axis, and ranging over
all values of j 0. Each bre is a function only of r, and thus, has a one-
dimensional Fourier transform. The problem then becomes determining these
Fourier transforms from the Fourier transform of the boundary function /.
Proposition 20C.1. Fourier Solution to Half-Plane Dirichlet problem
Let / L
1
(1). Suppose that / has Fourier transform

/, and dene n : H 1
by
n(r, j) :=
_

/(j) c
[[y
exp
_
jir
_
dj, for all r 1 and j 0.
Then n is the solution to the Laplace equation (n = 0) which is bounded at
innity and which satises the nonhomogeneous Dirichlet boundary condition
n(r, 0) = /(r), for all r 1.
Proof. For any xed j 1, the function )
(r, j) = exp
_
[j[ j
_
exp
_
jir
_
is harmonic (see practice problem # 10 on page 543). Thus, Proposition 0G.1
on page 567 implies that the function n(r, j) is also harmonic. Finally, no-
tice that, when j = 0, the expression for n(r, 0) is just the Fourier inversion
integral for /(r). 2
Example 20C.2. Suppose /(r) =
_
1 if 1 < r < 1;
0 otherwise.
We already
know from Example 19A.3 on page 489 that

/(j) =
sin(j)
j
.
Thus, n(r, j) =
1
sin(j)
j
c
[[y
exp
_
jir
_
dj.
Exercise 20C.1. Note the boundedness condition in Proposition 20C.1. Find E _
another solution to the Dirichlet problem on H which is unbounded at innity.
20C. The Dirichlet problem on a half-plane 539
20C(ii) Impulse-response solution
Prerequisites: 20C(i). Recommended: 17E.
For any j 0, dene the Poisson kernel /
y
: 1 1 by:
/
y
(r) :=
j
(r
2
+j
2
)
. (see Figure 17E.2 on page 404) (20C.9)
In 17E, we used the Poisson kernel to solve the half-plane Dirichlet problem
using impulse-response methods (Proposition 17E.1 on page 404). We can now
use the Fourier solution to provide another proof Proposition 17E.1.
Proposition 20C.3. Poisson Kernel Solution to Half-Plane Dirichlet problem
Let / L
1
(1). Dene n : H 1 by
l(r, j) = / /
y
(r) =
j
/(.)
(r .)
2
+j
2
d., (20C.10)
for all j 0 and r 1. Then l is the solution to the Laplace equation (l = 0)
which is bounded at innity and which can be continuously extended to satisfy
the nonhomogeneous Dirichlet boundary condition l(r, 0) = /(r) for all r 1.
Proof. Well show that the solution l in eqn. (20C.10) is actually equal to
the Fourier solution n from Proposition 20C.1.
Fix j 0, and dene l
y
(r) = l(r, j) for all r 1. Equation (20C.10) says
l
y
= / /
y
; hence Theorem 19B.2(b) (p.494) says:
l
y
= 2
/

/
y
. (20C.11)
Now, by practice problem # 7 on page 524 of 19I, we have:
/
y
(j) =
c
y[[
2
, (20C.12)
Combine (20C.11) and (20C.12) to get:
l
y
(j) = c
y[[

/(j). (20C.13)
Now apply the Fourier inversion formula (Theorem 19A.1 on page 488) to eqn
(20C.13) to obtain:
l
y
(r) =
_

l(j) exp(j r i) dj =
_

c
y[[

/(j) crj(j r i) dj
= n(r, j),
where n(r, j) is the solution from Proposition 20C.1. 2
20D PDEs on the half-line
Prerequisites: 1B(i), 19F, 5C, 0G.
Using the Fourier (co)sine transform, we can solve PDEs on the half-line.
Theorem 20D.1. The heat equation; Dirichlet boundary conditions
Let ) L
1
(1
,
) have Fourier sine transform

)
sin
, and dene n : 1
,
1
,
1
by:
n(r, t) :=
_

0
)
sin
(j) sin(j r) c
2
t
dj
Then n(r, t) is a solution to the heat equation, with initial conditions n(r, 0) =
)(r), and satises the homogeneous Dirichlet boundary condition: n(0, t) = 0.
Proof. Exercise 20D.1 (Hint: Use Proposition 0G.1 on page 567.) 2 E _
Theorem 20D.2. The heat equation; Neumann boundary conditions
Let ) L
1
(1
,
) have Fourier cosine transform

)
cos
, and dene n : 1
,
1
,
1
by:
n(r, t) :=
_

0
)
cos
(j) cos(j r) c
2
t
dj
Then n(r, t) is a solution to the heat equation, with initial conditions n(r, 0) =
)(r), and the homogeneous Neumann boundary condition:
x
n(0, t) = 0.
Proof. Exercise 20D.2 (Hint: Use Proposition 0G.1 on page 567.) 2 E _
20E General solution to PDEs using Fourier transforms
Prerequisites: 16F, 18A, 19A, 19D, 19E.
Recommended: 20A(i), 20B(i), 20C, 20D.
Most of the results of this chapter can be subsumed into a single abstraction,
which makes use of the polynomial formalism developed in 16F on page 369.
Theorem 20E.1. Fix 1 N, and let L be a linear dierential operator on
(
(1
D
; 1) with constant coecients. Suppose L has polynomial symbol T.
(a) If ) L
1
(1
D
) has Fourier Transform

) : 1
D
1, and p = L ), then p
has Fourier transform: p() = T(i )

)(), for all 1
D
.
20E. General solution to PDEs using Fourier transforms 541
(b) If L
1
(1
D
) has Fourier transform L
1
(1
D
), and ) L
1
(1
D
) has
Fourier transform
)() =
()
T(i )
, for all 1
D
,
then ) is a solution to the Poisson-type nonhomogeneous equation L ) =
.
Let n : 1
D
1
,
1 be another function, and, for all t 0, dene
n
t
: 1
D
1 by n
t
(x) := n(x, t) for all x 1
D
. Suppose n
t
L
1
(1
D
), and let
n
t
have Fourier transform n
t
.
(c) Let ) L
1
(1
D
), and suppose n
t
() = exp(T(i ) t)

)(), for all 1
D
and t 0. Then n is a solution to the rst-order evolution equation
t
n(x, t) = L n(x, t), for all x 1
D
and t 0,
with initial conditions n(x, 0) = )(x), for all x 1
D
.
(d) Suppose ) L
1
(1
D
) has Fourier transform

) which decays fast enough
that
_
1
D
T(i ) cos
_
_
T(i ) t
_

)()
d < , for all t 0.

2
Suppose n
t
() = cos
_
_
T(i ) t
_

)(), for all 1
D
and t 0.
Then n is a solution to the second-order evolution equation
2
t
D
and t 0,
with initial position n(x, 0) = )(x) and initial velocity
t
n(x, 0) = 0, for all
x 1
D
.
(e) Suppose ) L
1
(1
D
) has Fourier transform

) which decays fast enough
that
_
1
D
_
T(i ) sin
_
_
T(i ) t
_

)()
d < , for all t 0.

Suppose n
t
() =
sin
_
_
T(i ) t
_
_
T(i )

)(), for all 1

D
and t 0. Then
the function n(x, t) is a solution to the second-order evolution equation
2
t
D
and t 0,
with initial position n(x, 0) = 0 and initial velocity
t
n(x, 0) = )(x), for all
x 1
D
.
Proof. Exercise 20E.1 (Hint: Use Proposition 0G.1 on page 567. In each case, E _
be sure to verify that convergence conditions of Proposition 0G.1 are satised.) 2
2
See Example 18A.6(b,c) on page 420 for the denitions of complex sine and cosine functions.
See Exercise 18C.17 on page 449 for a discussion of complex square roots.
Exercise 20E.2. Go back through this chapter and see how all of the dierent E _
solution theorems for the heat equation (20A(i)) wave equation (20B(i)), and Poisson
equation (20C) are special cases of this result. What about the solution for the Dirichlet
problem on a half-space in 20D? How does it t into this formalism?
Exercise 20E.3. State and prove a theorem analogous to Theorem 20E.1 for solving E _
a 1-dimensional Schrodinger equation using Fourier transforms.
20F Practice problems
1. Let )(r) =
_
1 if 0 < r < 1;
0 otherwise.
, as in Example 19A.4 on page 490
(a) Use the Fourier method to solve the Dirichlet problem on a half-space,
with boundary condition n(r, 0) = )(r).
(b) Use the Fourier method to solve the heat equation on a line, with
initial condition n
0
(r) = )(r).
2. Solve the two-dimensional heat equation, with initial conditions
)(r, j) =
_
1 if 0 r A and 0 j Y ;
0 otherwise.
where A, Y 0 are constants. (Hint: See problem # 3 on page 523 of
19I)
3. Solve the two-dimensional wave equation, with
Initial Position: n(r, j, 0) = 0,
Initial Velocity:
t
n(r, j, 0) = )
1
(r, j) =
_
1 if 0 r 1 and 0 j 1;
0 otherwise.
.
(Hint: See problem # 3 on page 523 of 19I)
4. Let ) : 1 1 be the function dened: )(r) =
_
r if 0 r 1
0 otherwise
(see Figure 19I.1 on page 524). Solve the heat equation on the real
line, with initial conditions n(r; 0) = )(r). (Use the Fourier method; see
problem # 4 on page 523 of 19I)
5. Let )(r) = r exp
_
r
2
2
_
. (See problem # 5 on page 523 of 19I.)
(a) Solve the heat equation on the real line, with initial conditions
n(r; 0) = )(r). (Use the Fourier method.)
(b) Solve the wave equation on the real line, with initial position n(r; 0) =
)(r) and initial velocity
t
n(r, 0) = 0. (Use the Fourier method.)
6. Let )(r) =
2r
(1 +r
2
)
2
. (See problem # 8 on page 524 of 19I.)
(a) Solve the heat equation on the real line, with initial conditions
n(r; 0) = )(r). (Use the Fourier method.)
(b) Solve the wave equation on the real line, with initial position n(r, 0) =
0 and initial velocity
t
n(r, 0) = )(r). (Use the Fourier method.)
7. Let )(r) =
_
1 if 4 < r < 5;
0 otherwise.
(See problem # 9 on page 524 of
19I.) Use the Fourier Method to solve the one-dimensional heat equation
(
t
n(r; t) = n(r; t)) on the domain X = 1, with initial conditions
n(r; 0) = )(r).
8. Let )(r) =
rcos(r) sin(r)
r
2
. (See problem # 10 on page 524 of
19I.) Use the Fourier Method to solve the one-dimensional heat equa-
tion (
t
n(r; t) = n(r; t)) on the domain X = 1, with initial conditions
n(r; 0) = )(r).
9. Suppose ) : 1 1 had Fourier transform

)(j) =
j
j
4
+ 1
.
t
n = n,
with initial conditions n(r; 0) = )(r) for all r 1.
(b) Find the solution to the one-dimensional wave equation
2
t
n = n,
with
Initial position n(r; 0) = 0, for all r 1.
Initial velocity
t
n(r; 0) = )(r), for all r 1.
(c) Find the solution to the two-dimensional Laplace Equation n(r, j) =
0 on the half-space H = (r, j) ; r 1, j 0, with boundary con-
dition: n(r, 0) = )(r) for all r 1.
(d) Verify your solution to question (c). That is: check that your solution
satises the Laplace equation and the desired boundary conditions.
10. Fix j 1, and dene )
: 1
2
C by )
(r, j) := exp
_
[j[ j
_
exp
_
jir
_
.
Show that ) is harmonic on 1
2
.
(This function appears in the Fourier solution to the half-plane Dirichlet
problem; see Proposition 20C.1 on page 538.)
545
Chapter 0
Appendices
0A Sets and functions
Sets: A set is a collection of objects. If o is a set, then the objects in o are
called elements of o; if : is an element of o, we write : o. A subset of
o is a smaller set 1 such that every element of 1 is also an element of o. We
indicate this by writing 1 o.
Sometimes we can explicitly list the elements in a set; we write o = :
1
, :
2
, :
3
, . . ..
Example 0A.1.
(a) In Figure 0A.1(A), o is the set of all cities in the world, so Toronto o. We
might write o = Toronto, Beijing, London, Kuala Lampur, Nairobi, Santiago,
Pisa, Sidney, . . .. If 1 is the set of all cities in Canada, then 1 o.
(b) In Figure 0A.1(B), the set of natural numbers is N := 0, 1, 2, 3, 4, . . ..
The set of positive natural numbers is N
+
:= 1, 2, 3, 4, . . ..
Thus, 5 N, but , N and 2 , N.
(c) In Figure 0A.1(B), the set of integers is Z := . . . , 3, 2, 1, 0, 1, 2, 3, 4, . . ..
Thus, 5 Z and 2 Z, but , Z and
1
2
, Z. Observe that N
+
N Z.
(d) In Figure 0A.1(B), the set of real numbers is denoted by 1. It is best
visualised as an innite line. Thus, 5 1, 2 1, 1 and
1
2
1.
Observe that N Z 1.
(e) In Figure 0A.1(B), the set of nonnegative real numbers is denoted by
[0, ) or 1
,
. It is best visualised as a half-innite line, including zero.
Observe that [0, ) 1.
(f) In Figure 0A.1(B), the set of positive real numbers is denoted by (0, )
or 1
+
. It is best visualised as a half-innite line, excluding zero. Observe
that 1
+
1
,
1.
546 DRAFT Chapter 0. Appendices
Nairobi
Cairo
Paris
Kuala Lampur
Copenhagen
Beijing
Pisa
Marakesh
Buenos Aires
St. Petersburg
Kyoto
Santiago New York
Toronto
Peterborough
Montreal
Barcelona
Edmonton
Vancouver
Halifax
Bombay
Sidney
Vladivostok
Berlin
(A)
1 2 3 4 5......
0 1 2 3 4 5...... -1 -2 -3 .....-4
0
R
:= [0,oo)
:= (0,oo)
(B)
(x,y)
x
y
R
(C)
(D)
(x , x , x )
1 2 3
R
3
2
R
+
R
0 1 2 3 4 5......
+
Figure 0A.1: (A) 1 is a subset of o (B) Important Sets: N
+
, N, Z, 1, 1
,
:=
[0, ) and 1
+
:= (0, ). (C) 1
2
is two-dimensional space. (D) 1
3
is
three-dimensional space.
(g) Figure 0A.1(C) depicts two-dimensional space: the set of all coordinate
pairs (r, j), where r and j are real numbers. This set is denoted by 1
2
,
and is best visualised as an innite plane.
(h) Figure 0A.1(D) depicts three-dimensional space: the set of all coordi-
nate triples (r
1
, r
2
, r
3
), where r
1
, r
2
, and r
3
are real numbers. This set is
denoted by 1
3
, and is best visualised as an innite void.
(i) If 1 is any natural number, then 1-dimensional space is the set of all
coordinate triples (r
1
, r
2
, . . . , r
D
), where r
1
, . . . , r
D
are all real numbers.
This set is denoted by 1
D
. It is hard to visualize when 1 is bigger than
3.
Cartesian Products: If o and T are two sets, then their Cartesian product
is the set of all pairs (:, t), where : is an element of o, and t is an element of T .
We denote this set by o T .
Example 0A.2.
(a) 11 is the set of all pairs (r, j), where r and j are real numbers. In other
words, 1 1 = 1
2
.
0A. Sets and functions 547
Nairobi
Cairo
Paris
Beijing
Pisa
Marakesh
Santiago
Montreal
Bombay
Sidney
C
K
S
B F
D P
Z
H
V
N
E M
Halifax
Vancouver
Kyoto
Vladivostok
St. Petersburg
Copenhagen
Peterborough
T
New York
Edmonton
Toronto
Buenos Aires
Barcelona
Kuala Lampur
Berlin
A
T
R
Q
L
(A) (B)
R
R
3
Figure 0A.2: (A) )(C) is the rst letter of city C. (B) p(t) is the position of
the y at time t.
(b) 1
2
1 is the set of all pairs (w, .), where w 1
2
and . 1. But if w is
an element of 1
2
, then w = (r, j) for some r 1 and j 1. Thus, any
element of 1
2
1 is an object
_
(r, j), .
_
. By suppressing the inner pair
of brackets, we can write this as (r, j, .). In this way, we see that 1
2
1
is the same as 1
3
.
(c) In the same way, 1
3
1 is the same as 1
4
, once we write
_
(r, j, .), t
_
as
(r, j, ., t). More generally, 1
D
1 is mathematically the same as 1
D+1
.
Often, we use the nal coordinate to store a time variable, so it is useful
to distinguish it, by writing
_
(r, j, .), t
_
as (r, j, .; t).
Functions: If o and T are sets, then a function from o to T is a rule which
assigns a specic element of T to every element of o. We indicate this by writing
) : o T .
Example 0A.3.
(a) In Figure 0A.2(A), o is the cities in the world, and T = , 1, C, . . . , 7
is the letters of the alphabet, and ) is the function which is the rst letter
in the name of each city. Thus )(Peterborough) = 1, )(Santiago) = o,
etc.
(b) if 1 is the set of real numbers, then sin : 1 1 is a function: sin(0) = 0,
sin(,2) = 1, etc.
Two important classes of functions are paths and elds.
8
6
4
2
0
2
4
6
8
x
3
2
1
0
1
2
3
y
10
5
0
5
10
(A) (B)
Figure 0A.3: (A) A height function describes a landscape. (B) A density distribution in
1
2
.
Paths: Imagine a y buzzing around a room. Suppose you try to represent
its trajectory as a curve through space. This denes a a function p from 1 into
1
3
, where 1 represents time, and 1
3
represents the (three-dimensional) room,
as shown in Figure 0A.2(B). If t 1 is some moment in time, then p(t) is the
position of the y at time t. Since this p describes the path of the y, we call p
a path.
More generally, a path (or trajectory or curve) is a function p : 1 1
D
,
where 1 is any natural number. It describes the motion of an object through
1-dimensional space. Thus, if t 1, then p(t) is the position of the object at
time t.
Scalar Fields: Imagine a three-dimensional topographic map of Antarctica.
The rugged surface of the map is obtained by assigning an altitude to every
location on the continent. In other words, the map implicitly denes a function
h from 1
2
(the Antarctic continent) to 1 (the set of altitudes, in metres above
sea level). If (r, j) 1
2
is a location in Antarctica, then h(r, j) is the altitude
at this location (and h(r, j) = 0 means (r, j) is at sea level).
This is an example of a scalar eld. A scalar eld is a function n : 1
D
1;
it assigns a numerical quantity to every point in 1-dimensional space.
Example 0A.4.
(a) In Figure 0A.3(A), a landscape is represented by a height function h :
1
2
1.
(b) Figure 0A.3(B) depicts a concentration function on a two-dimensional
plane (e.g. the concentration of bacteria on a petri dish). This is a function
: 1
2
1
,
(where (r, j) = 0 indicates zero bacteria at (r, j)).
0B. Derivatives notation 549
(c) The mass density of a three-dimensional object is a function : 1
3
1
,
(where (r
1
, r
2
, r
3
) = 0 indicates vacuum).
(d) The charge density is a function q : 1
3
1 (where q(r
1
, r
2
, r
3
) = 0
indicates electric neutrality)
(e) The electric potential (or voltage) is a function V : 1
3
1.
(f) The temperature distribution in space is a function n : 1
3
1 (so
n(r
1
, r
2
, r
3
) is the temperature at location (r
1
, r
2
, r
3
))
A time-varying scalar eld is a function n : 1
D
1 1, assigning a
quantity to every point in space at each moment in time. Thus, for example,
n(x; t) is the temperature at location x, at time t
Vector Fields: A vector eld is a function

V : 1
D
1
D
; it assigns a
vector (i.e. an arrow) at every point in space.
Example 0A.5.
(a) The electric eld generated by a charge distribution (denoted by

E).
(b) The ux of some material owing through space (often denoted by

F).
Thus, for example,

F(x) is the ux of material at location x.
0B Derivatives notation
If ) : 1 1, then )
t
is the rst derivative of ); )
tt
is the second deriva-
tive,... )
(n)
the nth derivative, etc. If x : 1 1
D
is a path, then the velocity
of x at time t is the vector
x(t) =
_
r
t
1
(t), r
t
2
(t), . . . , r
t
D
(t)
If n : 1
D
1 is a scalar eld, then the following notations will be used
interchangeably:
for all , [1...1],
j
n :=
n
r
j
If n : 1
2
1 (i.e. n(r, j) is a function of two variables), then we will also write
x
n :=
n
r
and
y
n :=
n
j
.
Multiple derivatives will be indicated by iterating this procedure. For example,
3
x
2
y
n :=

3
r
3
2
n
j
2
A useful notational convention (which we rarely use) is multiexponents. If
1
, . . . ,
D
are positive integers, and = (
1
, . . . ,
D
), then
x
:= r
1
1
r
2
2
. . . , r
D
D
For example, if = (3, 4), and z = (r, j) then z
= r
3
j
4
.
This generalizes to multi-index notation for derivatives. If = (
1
, . . . ,
D
),
then
n :=
1
1

2
2
. . .
D
D
n
For example, if = (1, 2), then
n =

r
2
n
j
2
.
Remark. Many authors use subscripts to indicate partial derivatives. For
example, they would write
n
x
:=
x
n, n
xx
:=
2
x
n, n
xy
:=
x
y
n, etc.
This notation is very compact and intuitive, but it has two major disadvantages:
1. When dealing with an -dimensional function n(r
1
, r
2
, . . . , r
N
) (where
is either large or indeterminate), you have only two options. You can either
either use awkward nested subscript expressions like
n
x
3
:=
3
n, n
x
5
x
5
:=
2
5
n, n
x
2
x
3
:=
2
3
n, etc.,
or you must adopt the numerical subscript convention that
n
3
:=
3
n, n
55
:=
2
5
n, n
23
:=
2
3
n, etc.
But once numerical subscripts are reserved to indicate derivatives in this
fashion, they can no longer be used for other purposes (e.g. indexing
a sequence of functions, or indexing the coordinates of a vector-valued
function). This can create further awkwardness.
2. We will often be considering functions of the form n(r, j; t), where (r, j)
are space coordinates and t is a time coordinate. In this situation, it is
often convenient to x a value of t and consider the two-dimensional scalar
eld n
t
(r, j) := n(r, j; t). Normally, when we use t as a subscript, it will
be indicate a time-frozen scalar eld of this kind.
Thus, in this book, we will never use subscripts to indicate partial derivatives.
Partial derivatives will always be indicated by the notation
x
n or
u
x
(almost
always the rst one). However, when consulting other texts, you should be aware
of the subscript notation for derivatives, because it is used quite frequently.
0C. Complex numbers 551
0C Complex numbers
Complex numbers have the form . = r + ji, where i
2
= 1. We say that r
is the real part of ., and j is the imaginary part; we write: r = Re [.] and
j = Im[.].
If we imagine (r, j) as two real coordinates, then the complex numbers form
a two-dimensional plane. Thus, we can also write a complex number in polar
coordinates (see Figure 0C.1) If : 0 and 0 < 2, then we dene
: cis = : [cos() +i sin()]
Addition: If .
1
= r
1
+ j
1
i, .
2
= r
2
+ j
2
i, are two complex numbers, then
.
1
+.
2
= (r
1
+r
2
) + (j
1
+j
2
)i. (see Figure 0C.2)
Multiplication: If .
1
= r
1
+ j
1
i, .
2
= r
2
+ j
2
i, are two complex numbers,
then .
1
.
2
= (r
1
r
2
j
1
j
2
) + (r
1
j
2
+r
2
j
1
) i.
Multiplication has a nice formulation in polar coordinates; If .
1
= :
1
cis
1
and .
2
= :
2
cis
2
, then .
1
.
2
= (:
1
:
2
) cis (
1
+
2
). In other words,
multiplication by the complex number . = : cis is equivalent to dilating the
complex plane by a factor of :, and rotating the plane by an angle of . (see
Figure 0C.3)
Exponential: If . = r +ji, then exp(.) = c
x
cis j = c
x
[cos(j) +i sin(j)].
(see Figure 0C.4) In particular, if r 1, then
exp(r) = c
x
is the standard real-valued exponential function.
exp(ji) = cos(j) + sin(j)i is a periodic function; as j moves along the
real line, exp(ji) moves around the unit circle. (This is Eulers formula.)
The complex exponential function shares two properties with the real exponential
function:
If .
1
, .
2
C, then exp(.
1
+.
2
) = exp(.
1
) exp(.
2
).
If n C, and we dene the function ) : C C by )(.) = exp(n .),
then )
t
(.) = n )(.).
Consequence: If n
1
, n
2
, . . . , n
D
C, and we dene ) : C
D
C by
)(.
1
, . . . , .
D
) = exp(n
1
.
1
+n
2
.
2
+. . . n
D
.
D
),
then
d
)(z) = n
d
)(z). More generally,
n
1
1

n
2
2
. . .
n
D
D
)(z) = n
n
1
1
n
n
2
2
. . . n
n
D
D
)(z). (0C.1)
552 DRAFT
x= r cos()
y
=

r

s
i
n
(
)
r
z = x + y i
= r [cos() + i sin()]
= r cis
Figure 0C.1: . = r +ji; : =
_
r
2
+j
2
, = tan(j,r).
x
1
x
2
+
y
1
y
2
+
x = 4
y

=

2
z = 4 + 2

i
1
1
1
x = 2
z = 2 + 3

i
2
2
y

=

3
2
6+5
i
2
= 6
=

5
Figure 0C.2: The addition of complex numbers .
1
= r
1
+j
1
i and .
2
= r
2
+j
2
i.
= 30
z = 2 cis 30
1
1
= 75
= 45
2
z =(1.5) cis 45
1
z = 3 cis 75
r
=
3
r

=

2
1
r
=
1
.
5
2
Figure 0C.3: The multiplication of complex numbers .
1
= :
1
cis
1
and .
2
=
:
2
cis
2
.
x=1/2
z = +

i
y
=

/
3
exp
e
exp(z) = e
= e [cos( ) + sin( )]
= e [ 3 /2 + /2]
i
cis
1/2
1/2
1/2
/3
/3 /3
/3
i
1/2
/3
1/2
Figure 0C.4: The exponential of complex number . = r +ji.
0D. Coordinate systems and domains 553
L
(A) Box
L
L
(B) Ball
1
3
2
R
(C) Cube
1
1
1
(D) Slab
L
(E) Rectangular
Column
Figure 0D.1: Some domains in 1
3
.
For example, if )(r, j) = exp(3r + 5ij), then
)
xxy
(r, j) =
2
x
y
)(r, j) = 45 i exp(3r + 5ij).
If w = (n
1
, . . . , n
D
) and z = (.
1
, . . . , .
D
), then we will sometimes write:
exp(n
1
.
1
+n
2
.
2
+. . . n
D
.
D
) = exp w, z) .
Conjugation and Norm: If . = r + ji, then the complex conjugate of .
is . = r ji. In polar coordinates, if . = : cis , then . = : cis ().
The norm of . is [.[ =
_
r
2
+j
2
. We have the formula:
[.[
2
= . ..
0D Coordinate systems and domains
Prerequisites: 0A.
Boundary Value Problems are usually posed on some domain some region
of space. To solve the problem, it helps to have a convenient way of mathemati-
cally representing these domains, which can sometimes be simplied by adopting
a suitable coordinate system. We will rst give a variety of examples of domains
in dierent coordinate systems in 0D(a,b,c,d). Then in 0D(e) we will give a
formal denition of the word domain.
0D(i) Rectangular coordinates
Rectangular coordinates in 1
3
are normally denoted (r, j, .). Three common
domains in rectangular coordinates:
The slab X =
_
(r, j, .) 1
3
; 0 . 1
_
, where 1 is the thickness of the
slab (see Figure 0D.1D).
The unit cube: X = (r, j, .) 1
3
; 0 r 1, 0 j 1, and 0
. 1 (see Figure 0D.1C).
The box: X = (r, j, .) 1
3
; 0 r 1
1
, 0 j 1
2
, and 0 .
1
3
, where 1
1
, 1
2
, and 1
3
are the sidelengths (see Figure 0D.1A).
The rectangular column: X =
_
(r, j, .) 1
3
; 0 r 1
1
and 0 j 1
2
_
(see Figure 0D.1E).
0D(ii) Polar coordinates on 1
2
r
x
y
Figure 0D.2: Polar coordinates

Polar coordinates (:, ) on 1
2
are dened by the transformation:
r = : cos() and j = : sin().
: =
_
r
2
+j
2
and = Arctan(j, r) .
Here, the coordinate : ranges over 1
,
, while the variable ranges over [, ).
Finally we dene
Arctan(j, r) :=
_
_
_
arctan (j,r) if r 0;
arctan (j,r) + if r < 0 and j 0;
arctan (j,r) if r < 0 and j < 0.
Three common domains in polar coordinates are:
| = (:, ) ; : 1 is the disk of radius 1 (see Figure 0D.3A).
0D. Coordinate systems and domains 555
R
L
(A) Disk (B) Annulus

(C) Finite Cylinder (D) Pipe
(E) Infinite
Cylinder
R
min
R
max
R
Figure 0D.3: Some domains in polar and cylindrical coordinates.
|
= (:, ) ; 1 : is the codisk of inner radius 1.

A = (:, ) ; 1
min
: 1
max
is the annulus, of inner radius 1
min
and
outer radius 1
max
(see Figure 0D.3B).
0D(iii) Cylindrical coordinates on 1
3
Cylindrical coordinates (:, , .) on 1
3
, are dened by the transformation:
r = : cos(), j = : sin() and . = .
: =
_
r
2
+j
2
, = Arctan(j, r) and . = ..
Five common domains in cylindrical coordinates are:
X = (:, , .) ; : 1 is the (innite) cylinder of radius 1 (see Figure
0D.3E).
X = (:, , .) ; 1
min
: 1
max
is the (innite) pipe of inner radius
1
min
and outer radius 1
max
(see Figure 0D.3D).
X = (:, , .) ; : 1 is the wellshaft of radius 1.
z
0
x
y
c
o
s
(
)
Figure 0D.4: Spherical coordinates
X = (:, , .) ; : 1 and 0 . 1 is the nite cylinder of radius 1
and length 1 (see Figure 0D.3C).
In cylindrical coordinates on 1
3
, we can write the slab as (:, , .) ; 0 . 1.
0D(iv) Spherical coordinates on 1
3
Spherical coordinates (:, , ) on 1
3
are dened by the transformation:
r = : sin() cos(), j = : sin() sin()
and . = : cos().
: =
_
r
2
+j
2
+.
2
, = Arctan (j, r)
and = Arctan
_
_
r
2
+j
2
, .
_
.
In spherical coordinates, the set B = (:, , ) ; : 1 is the ball of radius
1 (see Figure 0D.1B).
0D(v) What is a domain ?
Formally, a domain is a subset X 1
D
which satises three conditions:
(a) X is closed. That is, X must contain all its boundary points.
(b) X has a dense interior. That is, every point in X is a limit point of a
sequence r
n
n=1
of interior points of X. (A point r X is an interior
point if B(r, c) X for some c 0).
0E. Vector calculus 557
(c) X is connected. That is, we cannot nd two disjoint closed subsets X
1
and
X
2
such that X = X
1
. X
2
.
Observe that all of the examples in 0D(a,b,c,d) satisfy these three conditions.
Why conditions (a), (b), and (c)? We are normally interested in nding a
function ) : X 1 which satises a certain partial dierential equation on
X. However, such a PDE only makes sense on the interior of X (because the
derivatives of ) at r are only well-dened if r is an interior point of X). Thus,
rst X must have a nonempty interior, and second, this interior must ll most
of X. This is the reason for condition (b). We often represent certain physical
constraints by requiring ) to satisfy certain boundary conditions on the boundary
of X. (Thats what a boundary value problem means). But this cannot make
sense unless X satises condition (a). Finally, we dont actually need condition
(c). But if X is disconnected, then we could split X into two or more disconnected
pieces and solve the equations separately on each piece. Thus, we can always
assume without loss of generality that X is connected.
0E Vector calculus
Prerequisites: 0A, 0B.
0E(i) Gradient
....in two dimensions:
Suppose X 1
2
was a two-dimensional region. To dene the topography of a
landscape on this region, it suces
1
to specify the height of the land at each
point. Let n(r, j) be the height of the land at the point (r, j) X. (Technically,
we say: n : X 1 is a two-dimensional scalar eld.)
The gradient of the landscape measures the slope at each point in space. To
be precise, we want the gradient to be an arrow pointing in the direction of most
rapid ascent. The length of this arrow should then measure the rate of ascent.
Mathematically, we dene the two-dimensional gradient of n by:
n(r, j) =
_
n
r
(r, j),
n
j
(r, j)
_
The gradient arrow points in the direction where n is increasing the most rapidly.
If n(r, j) was the height of a mountain at location (r, j), and you were trying
to climb the mountain, then your (naive) strategy would be to always walk in
the direction n(r, j). Notice that, for any (r, j) X, the gradient n(r, j) is
a two-dimensional vector that is, n(r, j) 1
2
. (Technically, we say n :
X 1
2
is a two-dimensional vector eld.)
1
Assuming no overhangs!
....in many dimensions:
This idea generalizes to any dimension. If n : 1
D
1 is a scalar eld, then
the gradient of n is the associated vector eld n : 1
D
1
D
, where, for any
x 1
D
,
n(x) =
_
1
n,
2
n, . . . ,
D
n
_
(x)
Proposition 0E.1. Algebra of gradients
Let X 1
D
be a domain. Let ), p : X 1 be dierentiable scalar elds, and
let () +p) : X 1 and () p) : X 1 denote the sum and product of ) and
p.
(a) (Linearity) For all x X, and any : 1,
(:) +p)(x) = : )(x) + p(x).
(b) (Leibniz rule) For all x X,
() p)(x) = )(x)
_
p(x)
_
+ p(x)
_
)(x)
_
.
0E(ii) Divergence
....in one dimension:
Imagine a current of uid (e.g. air, water, electricity) owing along the real
line 1. For each point r 1, let \ (r) describe the rate at which uid is
owing past this point. Now, in places where the uid slows down, we expect
the derivative \
t
(r) to be negative. We also expect the uid to accumulate
(i.e. become compressed) at such locations (because uid is entering the region
more quickly than it leaves). In places where the uid speeds up, we expect the
derivative \
t
(r) to be positive, and we expect the uid to be depleted (i.e. to
decompress) at such locations (because uid is leaving the region more quickly
than it arrives).
If the uid is incompressible (e.g. water), then we can assume that the quan-
tity of uid at each point is constant. In this case, the uid cannot accumulate
or be depleted. In this case, a negative value of \
t
(r) means that uid is some-
how being absorbed (e.g. being destroyed or leaking out of the system) at r.
Likewise, a positive value of \
t
(r) means that uid is somehow being generated
(e.g. being created, or leaking into the system) at r.
In general, positive \
t
(r) may represent some combination of uid depletion,
decompression, or generation at r, while negative \
t
(r) may represent some
combination of local accumulation, compression or absorption at r. Thus, if
we dene the divergence of the ow to be the rate at which uid is being de-
pleted/decompressed/generated (if positive) or being accumulated/compressed/absorbed
(if positive), then mathematically speaking,
div \ (r) = \
t
(r).
This physical model yields an interesting interpretation of the Fundamental The-
orem of Calculus. Suppose o < / 1, and consider the interval [o, /]. If
\ : 1 1 describes the ow of uid, then \ (o) is the amount of uid owing
into the left end of the interval [o, /] (or owing out, if \ (o) < 0). Likewise,
\ (/) is the amount of uid owing out of the right end of the interval [o, /] (or
owing in, if \ (/) < 0). Thus, \ (/) \ (o) is the net amount of uid owing out
through the endpoints of [o, /] (or owing in, if this quantity is negative). But
the Fundamental Theorem of Calculus asserts that
\ (/) \ (o) =
_
b
a
\
t
(r) dr =
_
b
a
div \ (r) dr.
In other words, the net amount of uid instantaneously leaving/entering through
the endpoints of [o, /] is equal to the integral of the divergence over the interior.
But if div \ (r) is the amount of uid being instantaneously generated at r (or
absorbed if negative) this integral can be interpreted as the saying:
The net amount of uid instantaneously leaving the endpoints of
[o, /] is equal to the net quantity of uid being instantaneously gener-
ated throughout the interior of [o, /].
From a physical point of view, this makes perfect sense; it is simply con-
servation of mass. This is the one-dimensional form of the Divergence Theorem
(Theorem 0E.4 on page 563 below).
....in two dimensions:
Let X 1
2
be some planar region, and consider a uid owing through X. For
each point (r, j) X, let

V(r, j) be a two-dimensional vector describing the
current at that point
2
.
Think of this two-dimensional current as a superposition of a horizontal cur-
rent \
1
and a vertical current \
2
. For each of the two currents, we can reason
as in the one-dimensional case. If
x
\
1
(r, j) 0, then the horizontal current is
accelerating at (r, j), so we expect it to deplete the uid at (r, j) (or, if the uid
2
Technically, we say
V : X 1
2
is a two-dimensional vector eld.
is incompressible, we interpret this to mean that additional uid is being gener-
ated at (r, j)). If
x
\
1
(r, j) < 0, then the horizontal current is decelarating, we
expect it to deposit uid at (r, j) (or, if the uid is incompressible, we interpret
this to mean that uid is being absorbed or destroyed at (r, j)).
The same reasoning applies to
y
\
2
(r, j). The divergence of the two-
dimensional current is thus just the sum of the divergences of the horizontal
and vertical currents
div

V(r, j) =
x
\
1
(r, j) +
y
\
2
(r, j).
Notice that, although

V(r, j) was a vector, the divergence div

V(r, j) is a scalar
3
.
Just as in the one-dimensional case, we interpret div

V(r, j) to be the the instan-
taneous rate at which uid is being depleted/decompressed/generated at (r, j) (if
positive) or being accumulated/compressed/absorbed at (r, j) (if negative).
For example, suppose 1
2
represents the ocean, and

V : 1
2
1
2
is a vector
eld representing ocean currents. If div

V(r, j) 0, this means that there is a
net injection of water into the ocean at the point (r, j) e.g. due to rainfall
or a river outow. If div

V(r, j) < 0, this means that there is a net removal of
water from the ocean at the point (r, j) e.g. due to evaporation or hole in the
bottom of the sea.
....in many dimensions:
We can generalize this idea to any number of dimensions. If

V : 1
D
1
D
is
a vector eld, then the divergence of

V is the associated scalar eld div

V :
1
D
1, where, for any x 1
D
,
div

V(x) =
1
\
1
(x) +
2
\
2
(x) +. . . +
D
\
D
(x).
If

V represents the ow of some uid through 1
D
, then div

V(x) represents the
instantaneous rate at which uid is being depleted/decompressed/generated at x
(if positive) or being accumulated/compressed/absorbed at x (if negative). For
example, if

E is the electric eld, then div

E(x) is the amount of electric eld
being generated at x that is, div

E(x) = q(x) is the charge density at x.
Proposition 0E.2. Algebra of Divergences
Let X 1
D
be a domain. Let

V,

W : X 1
D
be dierentiable vector elds,
and let ) : X 1 be a dierentiable scalar eld, and let ()

V) : X 1
D
represent the product of ) and

V.
(a) (Linearity) For all x X, and any : 1,
div (:
V+

W)(x) = : div

V(x) + div

Wp(x).
3
Technically, we say div

V : X 1
2
is a two-dimensional scalar eld.
N(x)
V(x)
V(x) N(x)
x
L
X X
V
(A) (B)
Figure 0E.1: (A) Line segment ' is tangent to X at x. Vector

N(x) is normal to X
at x. If

V(x) is another vector based at x, then the dot product

V(x)

N(x) measures the
orthogonal projection of

V(x) onto

N(x) that is, the part of

V(x) which is normal to X.
(B) Here

V : 1
2
1
2
is a dierentiable vector eld, and we portray the scalar eld

V

N
along the curve X (although we have visualized it as a vector eld, to help your intuitions).
The ux of

V across the boundary of X is obtained by integrating

V

N along X.
(b) (Leibniz rule) For all x X,
div ()

V)(x) = )(x)
_
div

V(x)
_
+
_
)(x)
_

V(x).
Exercise 0E.3. Let

V,

W : X 1
D
be dierentiable vector elds, and consider E _
their dot product (
V

W) : X 1
D
(a dierentiable scalar eld). State and prove a
Leibniz-like rule for (
V

W).... (a) In the case 1 = 3; ....(b) In the case 1 4.
0E(iii) The Divergence Theorem.
...in two dimensions
Let X 1
2
be some domain in the plane, and let X be the boundary of
X. (For example, if X is the unit disk, then X is the unit circle. If X is a
square domain, then X is the four sides of the square, etc.). Let x X. A
line segment ' through x is tangent to X if ' touches X only at x; that is,
' X = x (see Figure 0E.1(A)). A unit vector

N is normal to X at x if
there is a line segment through x which is orthogonal to

N and which is tangent
to X. We say X is piecewise smooth if there is a unique unit normal vector
N(x) at every x X, except perhaps at nitely many points (the corners

of the boundary). For example, the disk, the square, and any other polygonal
domain have piecewise smooth boundaries. The function

N : X 1
2
is then
called the normal vector eld for X.
If

V = (\
1
, \
2
) and

N = (
1
,
2
) are two vectors, then dene

V

N :=
\
1
1
+\
2
2
. If

V : 1
2
1
2
is a vector eld, and X 1
2
is a domain with a
smooth boundary X, then we can dene the ux of

V across X as the integral:
_
X
V(s)

N(s) ds. (see Figure 0E.1(B)). (0E.1)
Here, by integrating over X, we are assuming that X can be parameterized
as a smooth curve or a union of smooth curves; this integral can then be com-
puted (via this parameterization) as one or more one-dimensional integrals over
intervals in 1. The value of integral (0E.1) is independent of the choice of param-
eterization you use. If

V describes the ow of some uid, then the ux (0E.1)
represents the net quantity of uid owing across the boundary of X.
On the other hand, if div

V(x) represents the instantaneous rate at which
uid is being generated/destroyed at the point x, then the two-dimensional in-
tegral
_
X
div

V(x) dx
is the net rate at which uid is being generated/destroyed throughout the interior
of the region X. The next result then simply says that the total mass of the
uid must be conserved when we combine these two processes:
Theorem 0E.3. (Greens Theorem)
If X 1
2
is an bounded domain with a piecewise smooth boundary, and

V :
X 1
2
is a continuously dierentiable vector eld, then
_
X
V(s)
N(s) ds =
_
X
div

V(x) dx.
Proof. See any introduction to vector calculus; see e.g. [Ste08, 16.5, p.1067]
2
...in many dimensions
Let X 1
D
be some domain, and let X be the boundary of X. (For example,
if X is the unit ball, then X is the unit sphere). If 1 = 2, then X will be
a 1-dimensional curve. If 1 = 3, then X will be a 2-dimensional surface. In
general, if 1 4, then X will be a (11)-dimensional hypersurface.
Let x X. A (hyper)plane segment P through x is tangent to X if P
touches X only at x; that is, P X = x. A unit vector

N is normal to X
at x if there is a (hyper)plane segment through x which is orthogonal to

N and
which is tangent to X. We say X is smooth if there is a unique unit normal
vector

N(x) at each x X.
4
The function

N : X 1
D
is then called the
normal vector eld for X.
If

V = (\
1
, . . . , \
D
) and

N = (
1
, . . . ,
D
) are two vectors, then dene
V

N := \
1
1
+ +\
D
D
. If

V : 1
D
1
D
is a vector eld, and X 1
D
is
a domain with a smooth boundary X, then we can dene the ux of

V across
X as the integral:
_
X
V(s)

N(s) ds. (0E.2)
Here, by integrating over X, we are assuming that X can be parameterized as
a smooth (hyper)surface or a union of smooth (hyper)surfaces; this integral can
then be computed (via this parameterization) as one or more (11)-dimensional
integrals over open subsets of 1
D1
. The value of integral (0E.2) is independent
of the choice of parameterization you use. If

V describes the ow of some uid,
then the ux (0E.2) represents the net quantity of uid owing across the bound-
ary of X.
On the other hand, if div

V(x) represents the instantaneous rate at which
uid is being generated/destroyed at the point x, then the 1-dimensional integral
_
X
div

V(x) dx
is the net rate at which uid is being generated/destroyed throughout the interior
of the region X. The next result then simply says that the total mass of the
uid must be conserved when we combine these two processes:
Theorem 0E.4. (Divergence Theorem)
If X 1
D
is a bounded domain with a piecewise smooth boundary, and
V : X 1
D
is a continuously dierentiable vector eld, then
_
X
V(s)
N(s) ds =
_
X
div

V(x) dx.
Proof. If 1 = 1, this just a restates the Fundamental Theorem of Calculus.
If 1 = 2, this just a restates of Greens Theorem (Theorem 0E.3).
For the case 1 = 3, this result can be found in any introduction to vector
calculus; see e.g. [Ste08, 16.9, p.1099]. This theorem is often called Gausss
4
More generally, X is piecewise smooth if there is a unique unit normal vector

N(x) at
almost every x X, except perhaps for a subset of dimension (D2). For example, a surface
in 1
3
is piecewise smooth if it has a normal vector eld everywhere except at some union of
curves, which represent the edges between the smooth faces the surface. In particular, a
cube, a cylinder, or any other polyhedron is has a piecewise smooth boundary.
Theorem (after C.F. Gauss) or Ostrogradskys Theorem (after Mikhail Ostro-
gradsky).
For the case 1 4, this is a special case of the Generalized Stokes Theorem,
one of the fundamental results of modern dierential geometry, which unies
the classic (2-dimensional) Stokes theorem, Greens theorem, Gauss theorem,
and the Fundamental Theorem of Calculus. A statement and proof can be
found in any introduction to dierential geometry or tensor calculus. See e.g.
[BG80, Theorem 4.9.2, p.196].
Some texts on partial dierential equations also review the Divergence Theo-
rem, usually in an appendix. See for example [Eva91, Appendix C.2, p. 627].
2
Greens formulae. Let n : 1
D
1 be a scalar eld. If X is a domain, and
s X, then the outward normal derivative of n at s is dened
n(s) := n(s)

N(s).
(see 5C(ii) for more information). Meanwhile, the Laplacian of n is dened by
n = div ((n)).
(see 1B(ii) on page 7 for more information). The Divergence Theorem then has
the following useful consequences.
Corollary 0E.5. (Greens Formulae)
Let X 1
D
be a bounded domain, and let n : X 1 be a scalar eld which
is (
2
(i.e. twice continuously dierentiable). Then
(a)
_
X
n(s) ds =
_
X
n(x) dx.
(b)
_
X
n(s)
n(s) ds =
_
X
n(x) n(x) + [n(x)[
2
dx.
(c) For any other (
2
function n : X 1,
_
X
_
n(x) n(x) n(x) n(x)
_
dx =
_
X
_
n(s)
n(s) n(s)
n(s)
_
ds.
Proof. (a) is Exercise 0E.4 . To prove (b), note that E _
2
_
X
n(s)
n(s) ds
()
_
X
(n
2
)(s) ds
()
_
X
(n
2
)(x) dx.
()
2
_
X
n(x) n(x) + [n(x)[
2
dx.
0F. Dierentiation of function series 565
Here, () is because
(n
2
)(s) = 2n(s)
n(s), by the Leibniz rule for nor-

mal derivatives (Exercise 0E.6 below), while () is because (n
2
)(x) =
2 [n(x)[
2
+ 2 n(x) n(x) by the Leibniz rule for Laplacians (Exercise
1B.4 on page 9). Finally, () is by part (a). The result follows.
(c) is Exercise 0E.5 . 2 E _
Exercise 0E.6. Prove the Leibniz rule for normal derivatives: if ), p : X 1 are E _
two scalar elds, and () p) : X 1 is their product, then for all s X,
() p)(s) =
_
)(s)
_
p(s) + )(s)
_
p(s)
_
.
Hint: Use the Leibniz rules for gradients (Propositions 0E.1(b) on page 558) and the
linearity of the dot product.
0F Dierentiation of function series
Recommended: 6E(iii), 6E(iv).
Many of our methods for solving partial dierential equations will involve
expressing the solution function as an innite series of functions (e.g. Taylor
series, Fourier series, etc.). To make sense of such solutions, we must be able to
dierentiate them.
Proposition 0F.1. Dierentiation of Series
Let o < / . For all n N, let )
n
: (o, /) 1 be a dierentiable
function, and dene 1 : (o, /) 1 by
1(r) =
n=0
)
n
(r), for all r (o, /).
(a) Suppose that
n=0
)
n
converges uniformly
5
to 1 on (o, /), and that
n=0
)
t
n
also converges uniformly on (o, /). Then 1 is dierentiable, and 1
t
(r) =
n=0
)
t
n
(r) for all r (o, /).
(b) Suppose there is a sequence 1
n
n=1
of positive real numbers such that
n=1
1
n
< .
5
See 6E(iii) and 6E(iv) for the denition of uniform convergence of a function series.
For all r (o, /), and all n N, [)
n
(r)[ 1
n
and [)
t
n
(r)[ 1
n
.
Then 1 is dierentiable, and 1
t
(r) =
n=0
)
t
n
(r) for all r (o, /).
Proof. (a) follows immediately from Proposition 6E.10(c) on page 127.
(b) follows from (a) and the Weierstras `-test (Proposition 6E.13 on page
129).
For a direct proof, see [Asm05, Theorems 1 and 5, p.87 and p.92 of 2.9] or
[Fol84, Theorem 2.27(b), p.54]. 2
Example 0F.2. Let o = 0 and / = 1. For all n N, let )
n
(r) =
r
n
n!
. Thus,
1(r) =
n=0
)
n
(r) =
n=0
r
n
n!
= exp(r),
(because this is the Taylor series for the exponential function). Now let 1
0
= 1
and let 1
n
=
1
(n1)!
for n 1. Then for all r (0, 1), and all n N,
[)
n
(r)[ =
1
n!
r
n
<
1
n!
<
1
(n1)!
= 1
n
and [)
t
n
(r)[ =
n
n!
r
n1
=
(n1)!
r
n1
<
1
(n1)!
= 1
n
. Also,
n=1
1
n
=
n=1
1
(n 1)!
< .
Hence the conditions of Proposition 0F.1(b) are satised, so we conclude that
1
t
(r) =
n=0
)
t
n
(r) =
n=0
n
n!
r
n1
=
n=1
r
n1
(n 1)!
(c)
m=0
r
m
:!
= exp(r),
where (c) is the change of variables : = n 1. In this case, the conclusion is
a well-known fact. But the same technique can be applied to more mysterious
functions.
Remarks: (a) The series
n=0
)
t
n
(r) is sometimes called the formal derivative
of the series
n=0
)
n
(r). It is formal because it is obtained through a purely
symbolic operation; it is not true in general that the formal derivative is re-
ally the derivative of the series, or indeed, if the formal derivative series even
0G. Dierentiation of integrals 567
converges. Proposition 0F.1 essentially says that, under certain conditions, the
formal derivative equals the true derivative of the series.
(b) Proposition 0F.1 is also true if the functions )
n
involve more than one
variable and/or more than one index. For example, if )
n,m
(r, j, .) is a function
of three variables and two indices, and
1(r, j, .) =
n=0
m=0
)
n,m
(r, j, .), for all (r, j, .) (o, /)
3
.
then under similar hypothesis, we can conclude that
y
1(r, j, .) =
n=0
m=0
y
)
n,m
(r, j, .),
for all (r, j, .) (o, /)
3
.
0G Dierentiation of integrals
Recommended: 0F.
Many of our methods for solving partial dierential equations will involve
expressing the solution function 1(r) as an integral of functions; i.e. 1(r) =
_

)
y
(r) dj, where, for each j 1, )
y
(r) is a dierentiable function of the
variable r. This is a natural generalization of the solution series mentioned
in 0F. Instead of beginning with a discretely paramaterized family of functions
)
n
n=1
, we begin with a continuously paramaterized family, )
y
y1
. Instead
of combining these functions through a summation to get 1(r) =
n=1
)
n
(r), we
combine them through integration, to get 1(r) =
_

)
y
(r) dj. However, to
make sense of such integrals as the solutions of dierential equations, we must
be able to dierentiate them.
Proposition 0G.1. Dierentiation of Integrals
Let o < / . For all j 1, let )
y
: (o, /) 1 be a dierentiable
function, and dene 1 : (o, /) 1 by
1(r) =
_

)
y
(r) dj, for all r (o, /).
Suppose there is a function : 1 1
,
such that
(a)
_

(j) dj < .
(b) For all j 1 and for all r (o, /), [)
y
(r)[ (j) and
)
t
y
(r)
(j).
Then 1 is dierentiable, and, for all r (o, /), 1
t
(r) =
_

)
t
y
(r) dj.
Proof. See [Fol84, Theorem 2.27(b), p.54]. 2
Example 0G.2. Let o = 0 and / = 1. For all j 1 and r (0, 1), let
)
y
(r) =
r
[y[+1
1 +j
4
. Thus,
1(r) =
_

)
y
(r) dj =
_

r
[y[+1
1 +j
4
dj.
Now, let (j) =
1 +[j[
1 +j
4
. Then
(a)
_

(j) dj =
_

1 +[j[
1 +j
4
dj < (check this).
(b) For all j 1 and all r (0, 1), [)
y
(r)[ =
r
[y[+1
1 +j
4
<
1
1 +j
4
<
1 +[j[
1 +j
4
= (j), and

)
t
n
(r)
=
([j[ + 1) r
[y[
1 +j
4
<
1 +[j[
1 +j
4
= (j).
Hence the conditions of Proposition 0G.1 are satised, so we conclude that
1
t
(r) =
_

)
t
n
(r) dj =
_

([j[ + 1) r
[y[
1 +j
4
dj.
Remarks: Proposition 0G.1 is also true if the functions )
y
involve more than
one variable. For example, if )
v,w
(r, j, .) is a function of ve variables, and
1(r, j, .) =
_

)
u,v
(r, j, .) dn d for all (r, j, .) (o, /)
3
.
then under similar hypothesis, we can conclude that
2
y
1(r, j, .) =
_

2
y
)
u,v
(r, j, .) dn d,
for all (r, j, .) (o, /)
3
.
0H Taylor polynomials
0H(i) Taylor polynomials in one dimension
Let X 1 be an open set and let ) : X 1 be an -times dierentiable
function. Fix o X. The Taylor polynomial of order for ) around o is
0H. Taylor polynomials 569
the function
T
N
a
)(r) := )(o) + )
t
(o) (r o) +
)
tt
(o)
2
(r o)
2
+
)
ttt
(o)
6
(r o)
3
+
)
(4)
(o)
4!
(r o)
4
+ +
)
(N)
(o)
!
(r o)
N
. (0H.1)
Here, )
(n)
(o) denotes the nth derivative of ) at o [e.g. )
(3)
(o) = )
ttt
(o)], and n!
(pronounced n factorial) is the product n (n 1) 4 3 2 1. For example,
T
0
a
)(r) = )(o) (a constant);
T
1
a
)(r) = )(o) + )
t
(o) (r o) (a linear function);
T
2
,
o )(r) = )(o) + )
t
(o) (r o) +
f
//
(a)
2
(r o)
2
(a quadratic function);
Note that T
1
a
)(r) parameterizes the tangent line to the graph of )(r) at the point
(o, )(o)) that is, the best linear approximation of ) in a neighbourhood of o.
Likewise, T
2
a
)(r) is the best quadratic approximation of ) in a neighbourhood
of o. In general T
N
a
)(r) is the polynomial of degree which provides the best
approximation of )(r) if r is reasonably close to . The formal statement of
this is Taylors theorem, which states that
)(r) = T
N
a
)(r) + O([r o[
N+1
).
Here O([r o[
N+1
) means some function which is smaller than a constant
multiple of [r o[
N+1
. In other words, there is a constant 1 0 such that
)(r) T
N
a
)(r)
1 [r o[
N+1
.
If [r o[ is large, then [r o[
N+1
is huge, so this inequality isnt particularly
useful. However, as [ro[ becomes small, [ro[
N+1
becomes really, really small.
For example, if [r o[ < 0.1, then [r o[
N+1
< 10
N1
. In this sense, T
N
a
)(r)
is a very good approximation of )(r) if r is close enough to o.
Further reading. More information about Taylor polynomials can be found
in any introduction to single-variable calculus; see e.g. [Ste08, p.253-254].
0H(ii) Taylor series and analytic functions
Prerequisites: 0H(i), 0F.
Let X 1 be an open set, let ) : X 1, let o X, and suppose ) is
innitely dierentiable at o. By letting in equation (0H.1), we obtain
the Taylor series (or power series) for ) at o:
T
a
)(r) := )(o) +)
t
(o)(ro) +
)
tt
(o)
2
(ro)
2
+ =
n=0
)
(n)
(o)
n!
(ro)
n
.
(0H.2)
Taylors Theorem suggests that T
a
)(r) = )(r) if r is close enough to o. Un-
fortunately, this is not true for all innitely dierentiable functions; indeed, the
series (0H.2) might not even converge for any r ,= o. However, we have the
following result:
Proposition 0H.1. Suppose the series (0H.2) converges for some r ,= o.
In that case, there is some 1 0 such that the series (0H.2) converges uni-
formly to )(r) on the interval (o 1, o +1). Thus, T
a
)(r) = )(r) for all
r (o 1, o +1). On the other hand, (0H.2) diverges for all r (, o 1)
and all r (o +1, ). 2
The 1 0 in Proposition 0H.1 is called the radius of convergence of
the power series (0H.2), and the interval (o 1, o +1) is the interval of con-
vergence. (Note that Proposition 0H.1 says nothing about the convergence of
(0H.2) at o 1; this varies from case to case). When the conclusion of Propo-
sition 0H.1 is true, we say that ) is analytic at o.
Example 0H.2. (a) All the basic functions of calculus are analytic every-
where on their domain: all polynomials, all rational functions, all trigonomet-
ric functions, the exponential function, the logarithm, and any sum, product,
or quotient of these functions.
(b) More generally, if ) and p are analytic at o, then () + p) and () p) are
analytic at o. If p(o) ,= 0, then ),p is analytic at o.
(c) If p is analytic at o, and p(o) = /, and ) is analytic at /, then ) p is
analytic at o.
If ) is innitely dierentiable at o = 0, then we can compute the Taylor series
T
0
)(r) := c
0
+ c
1
r +c
2
r
2
+c
3
r
3
+ =
n=0
c
n
r
n
. (0H.3)
where c
n
:=
)
(n)
(0)
n!
, for all n N. This special case of the Taylor series (with
o = 0) is sometimes called a Maclaurin series.
Dierentiating a Maclaurin series. If ) is analytic at o = 0, then there
is some 1 0 such that )(r) = T
0
)(r) for all r (1, 1). It follows that
)
t
(r) = (T
0
))
t
(r), and )
tt
(r) = (T
0
))
tt
(r), and so on, for all r (1, 1).
Proposition 0F.1 says that we can compute (T
0
))
t
(r), (T
0
))
tt
(r) etc. by
formally dierentiating the Maclaurin series (0H.3). Thus, we get:
)(r) = c
0
+ c
1
r + c
2
r
2
+ c
3
r
3
+ c
4
r
4
+ =
n=0
c
n
r
n
;
)
t
(r) = c
1
+ 2c
2
r + 3c
3
r
2
+ 4c
4
r
3
+ =
n=1
nc
n
r
n1
;
)
tt
(r) = 2c
2
+ 6c
3
r + 12c
4
r
2
+ =
n=1
n(n 1)c
n
r
n2
;
)
ttt
(r) = 6c
3
+ 24c
4
r + =
n=1
n(n 1)(n 2)c
n
r
n3
;
(0H.4)
etc.
Further reading. More information about Taylor series can be found in any
introduction to single-variable calculus; see e.g. [Ste08, 11.10, p.734].
0H(iii) Taylor series to solve ordinary dierential equations
Prerequisites: 0H(ii).
Suppose ) is an unknown analytic function (so the coecients c
0
, c
1
, c
2
, . . .
are unknown). An ordinary dierential equation in ) can be reformulated in
terms of the Maclaurin series in (0H.4); this yields a set of equations involving
the coecients c
0
, c
1
, c
2
, . . .. For example, let , 1, C 1 be constants. The
second-order linear ODE
)(r) + 1)
t
(r) + C)
tt
(r) = 0 (0H.5)
can be reformulated as a power-series equation
0 = c
0
+ c
1
r + c
2
r
2
+ c
3
r
3
+ c
4
r
4
+
+ 1c
1
+ 21c
2
r + 31c
3
r
2
+ 41c
4
r
3
+
+ 2Cc
2
+ 6Cc
3
r + 12Cc
4
r
2
+
(0H.6)
When we collect like terms in the r variable, this becomes:
0 = (c
0
+1c
1
+2Cc
2
)+(c
1
+21c
2
+6Cc
3
)r+(c
2
+31c
3
+12Cc
4
)r
2
+
(0H.7)
This yields an (innite) system of linear equations
0 = c
0
+ 1c
1
+ 2Cc
2
;
0 = c
1
+ 21c
2
+ 6Cc
3
;
0 = c
2
+ 31c
3
+ 12Cc
4
;
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
(0H.8)
If we dene c
n
:= n! c
n
for all n N, then the system (0H.8) reduces to the
simple linear recurrence relation:
c
n
+ 1c
n+1
+ Cc
n+2
= 0, for all n N. (0H.9)
[Note the relationship between (0H.9) and (0H.5); this is because, if ) is analytic
and has Maclaurin series (0H.3), then c
n
= )
(n)
(0) for all n N.]
We can then solve the linear recurrence relation (0H.9) using standard meth-
ods (e.g. characteristic polynomials), and obtain the coecients c
0
, c
1
, c
2
, . . ..
If the resulting power series converges, then it is a solution of the ODE (0H.5)
which is analytic in a neighbourhood of zero.
This technique for solving an ordinary dierential equation is called the Power
Series Method. It is not necessary to work in a neighbourhood of zero to apply
this method; we assumed o = 0 only to simplify the exposition. The Power
Series Method can be applied to a Taylor expansion around any point in 1.
We used the constant-coecient linear ODE (0H.5) just to provide a sim-
ple example. In fact, there are much easier ways to solve these sorts of ODEs
(e.g. characteristic polynomials, matrix exponentials). However, the Power Se-
ries Method is also applicable to linear ODEs with nonconstant coecients. For
example, if the coecients , 1, and C in equation (0H.5) were themselves ana-
lytic functions in r, then we would simply substitute the Taylor series expansions
of (r), 1(r) and C(r) into the power series equation (0H.6). This would make
the simplication into equation (0H.7) much more complicated, but we would
still end up with a system of linear equations in c
n
n=0
, like (0H.8). In general,
this will not simplify into a neat linear recurrence relation like (0H.9). But it
can still be solved one term at a time.
Indeed, the Power Series Method is also applicable to nonlinear ODEs. In
this case, we may end up with a system of nonlinear equations in c
n
n=0
instead
of the linear system (0H.8). For example, if the ODE (0H.5) contained a term
like )(r) )
tt
(r), then the system of equations (0H.8) would contain quadratic
terms like c
0
c
2
, c
1
c
3
, c
2
c
4
, etc.
Our analysis is actually incomplete, because we didnt check that the power
series (0H.3) had a nonzero radius of convergence when we obtained the sequence
c
n
n=0
as solutions to (0H.9). If (0H.9) is a linear recurrence relation (as in the
example here), then the sequence c
n
n=0
will grow subexponentially, and it is
easy to show that the radius of convergence for (0H.3) will always be nonzero.
However, in the case of nonconstant coecients or a nonlinear ODE, the power
series (0H.3) may not converge; this needs to be checked. For most of the second-
order linear ODEs we will encounter in this book, convergence is assured by the
following result.
Theorem 0H.3. (Fuchs)
Let o 1, let 1 0, and let " := (o 1, o + 1). Let j, , : : " 1 be
analytic functions whose Taylor series at o all converges everywhere in ". Then
every solution of the ODE
)
tt
(r) +j(r))
t
(r) +(r))(r) = :(r) (0H.10)
is an analytic function, whose Taylor series at o converges on ". The coecients
of this Taylor series can be found using the Power Series Method.
Proof. See [RB69, Chapter 3]. 2
If the conditions for Fuchs theorem are satised (i.e. if j, , : are all analytic
at o), then o is called an ordinary point for the ODE (0H.10). Otherwise, if one
of j, , : is not analytic at o, then o is called a singular point for ODE (0H.10).
In this case, we can sometimes use a modication of the Power Series Method:
the Method of Frobenius. For simplicity, we will discuss this method in the
case o = 0. Consider the homogeneous linear ODE
)
tt
(r) +j(r))
t
(r) +(r))(r) = 0. (0H.11)
Suppose that o = 0 is a singular point i.e. either j or is not analytic at 0.
Indeed, perhaps j and/or are not even dened at zero (e.g. j(r) = 1,r). We
say that 0 is a regular singular point if there are functions 1(r) and Q(r)
which are analytic at 0, such that j(r) = 1(r),r and (r) = Q(r),r
2
for all
r ,= 0. Let j
0
:= 1(0) and
0
:= Q(0) (the zeroth terms in the Maclaurin series
of 1 and Q), and consider the indicial polynomial
r(r 1) +j
0
r +
0
.
The roots :
1
:
2
of the indicial polynomial are called the indicial roots of the
ODE (0H.11).
Theorem 0H.4. (Frobenius)
Suppose r = 0 is a regular singular point of the ODE (0H.11), and let " be
the largest open interval of 0 where the Taylor series of both 1(r) and Q(r)
converge. Let "
:= " 0. Then there are two linearly independent functions

)
1
, )
2
: "
1 which satisfy the ODE (0H.11), and which depend on the

indicial roots :
1
:
2
as follows:
(a) If :
1
:
2
is not an integer, then )
1
(r) = [r[
r
1
n=0
/
n
r
n
and )
2
(r) =
[r[
r
2
n=0
c
n
r
n
.
(b) If :
1
= :
2
= :, then )
1
(r) = [r[
r
n=0
/
n
r
n
and )
2
(r) = )
1
(r) ln[r[ +
[r[
r
n=0
c
n
r
n
.
(c) If :
1
:
2
N, then )
1
(r) = [r[
r
n=0
/
n
r
n
and )
2
(r) = / )
1
(r) ln[r[ +
[r[
r
2
n=0
c
n
r
n
, for some / 1.
In all three cases, to obtain explicit solutions, substitute the expansions for )
1
and )
2
into the ODE (0H.11), along with the power series for 1(r) and Q(r), to
obtain recurrence relations characterizing the coecients /
n
n=0
, c
n
n=0
.
Proof. See [Asm05, Appendex A.6]. 2
Example 0H.5: (Bessels equation)
For any n N, the (2-dimensional) Bessel equation of order n is the
ordinary dierential equation
r
2
)
tt
(r) +r)
t
(r) + (r
2
n
2
))(r) = 0. (0H.12)
To put this in the form of ODE (0H.11), we divide by r
2
, to get
)
tt
(r) +
1
r
)
t
(r) +
_
1
n
2
r
2
_
)(r) = 0.
Thus, we have j(r) =
1
x
and (r) =
_
1
n
2
x
2
_
; hence 0 is a singular point of
ODE (0H.12), because j and are not dened (and hence not analytic) at
zero. However, clearly j(r) = 1(r),r and (r) = Q(r),r
2
, where 1(r) = 1
and Q(r) = (r
2
n
2
) are analytic at zero; thus 0 is a regular singular point
of ODE (0H.12). We have j
0
= 1 and
0
= n
2
, so the indicial polynomial
is r(r 1) + 1r n
2
= r
2
n
2
, which has roots :
1
= n and :
2
= n. Since
:
1
:
2
= 2n N, we apply Case (c) of Frobenius Theorem, and look for
solutions of the form
)
1
(r) = [r[
n
n=0
/
n
r
n
and )
2
(r) = / )
1
(r) ln[r[ +[r[
n
n=0
c
n
r
n
,
(0H.13)
To identify the coecients /
n
n=0
, c
n
n=0
, we substitute the power series
(0H.13) into ODE (0H.12) and simplify. The resulting solutions are called the
Bessel functions of types 1 and 2, respectively. The details can be found in
the proof of Proposition 14G.1 on page 305 of 14G.
Finally, we remark that a multivariate version of the Power Series Method
can be applied to a multivariate Taylor series, to obtain solutions to partial
dierential equations. (However, this book provides many other, much nicer
methods for solving linear PDEs with constant coecients).
Further reading. More information about the power series method and the
method of Frobenius can be found in any introduction to ordinary dierential
equations. See e.g. [Cod89, 3.9,p.138 and 4.6,p.162]. Some books on partial
dierential equations also contain this information (usually in an appendix); see
e.g. [Asm05, Appendix A.5-A.6].
0H(iv) Taylor polynomials in two dimensions
Prerequisites: 0B. Recommended: 0H(i).
Let X 1
2
be an open set and let ) : X 1 be an -times dierentiable
function. Fix a = (o
1
, o
2
) X. The Taylor polynomial of order for )
around a is the function
T
N
a
)(r
1
, r
2
) := )(a) +
1
)(a) (r
1
o
1
) +
2
)(a) (r
2
o
2
)
+
1
2
_
2
1
)(a) (r
1
o
1
)
2
+ 2
1
2
)(a) (r
1
o
1
)(r
2
o
2
) +
2
2
)(a) (r
2
o
2
)
2
_
+
1
6
_
3
1
)(a) (r
1
o
1
)
3
+ 3
2
1
2
)(a) (r
1
o
1
)
2
(r
2
o
2
)
+ 3
1
2
2
)(a) (r
1
o
1
)(r
2
o
2
)
2
+
3
2
)(a) (r
1
o
1
)
3
_
+
1
4!
_
4
1
)(a) (r
1
o
1
)
4
+ 4
3
1
2
)(a) (r
1
o
1
)
3
(r
2
o
2
)
+ 6
2
1
2
2
)(a) (r
1
o
1
)
2
(r
2
o
2
)
2
+ 4
1
3
2
)(a) (r
1
o
1
)(r
2
o
2
)
2
+
4
2
)(a) (r
1
o
1
)
4
_
+
+
1
!
N
n=0
_

n
_
(Nn)
1

n
2
)(a) (r
1
o
1
)
Nn
(r
2
o
2
)
n
.
For example, T
0
a
)(r
1
, r
2
) = )(a) is just a constant, while
T
1
a
)(r
1
, r
2
) = )(a) +
1
)(a) (r
1
o
1
) +
2
)(a) (r
2
o
2
)
is an ane function which parameterizes the tangent plane to the surface graph
of )(r) at the point (a, )(a)) that is, the best linear approximation of ) in a
neighbourhood of a. In general, T
N
a
)(x) is the 2-variable polynomial of degree
which provides the best approximation of )(x) if x is reasonably close to .
The formal statement of this is multivariate Taylors theorem, which states that
)(x) = T
N
a
)(x) + O([x a[
N+1
).
For example, if we set = 1, we get:
)(r
1
, r
2
) = )(a) +
1
)(a) (r
1
o
1
) +
2
)(a) (r
2
o
2
) + O([x a[
2
).
0H(v) Taylor polynomials in many dimensions
Prerequisites: 0E(i). Recommended: 0H(iv).
Let X 1
D
be an open set and let ) : X 1 be an -times dierentiable
function. Fix a = (o
1
, . . . , o
D
) X. The Taylor polynomial of order for
) around a is the function
T
N
a
)(x) := )(a) +)(a)
(x a) +
1
2
(x a)
D
2
)(a) (x a) +
+
1
!
n
1
++n
D
=N
_

n
1
. . . n
D
_
n
1
1

n
2
2

n
D
D
)(a) (r
1
o
1
)
n
1
(r
2
o
2
)
n
2
(r
D
o
D
)
n
D
.
Here, we regard x and a as column vectors, and the transposes x
, a
etc. as row
vectors. )(a)
:= [
1
)(a),
2
)(a), . . . ,
D
)(a)] is the (transposed) gradient
vector of ) at a, and
D
2
) :=
_
2
1
)
1
2
) . . .
1
D
)
1
)
2
2
) . . .
2
D
)
.
.
.
.
.
.
.
.
.
.
.
.
1
)
D
2
) . . .
2
D
)
_
_
is the Hessian derivative matrix of ). For example, T
0
a
)(x) = )(a) is just a
constant, while
T
1
a
)(r) = )(a) +)(a)
(x a)
is an ane function which paramaterizes the tangent hyperplane to the hypersur-
face graph of )(r) at the point (a, )(a)) that is, the best linear approximation
of ) in a neighbourhood of a. In general T
N
a
)(x) is the multivariate polynomial
of degree which provides the best approximation of )(x) if x is reasonably
close to . The formal statement of this is multivariate Taylors theorem, which
states that
)(x) = T
N
a
)(x) + O([x a[
N+1
).
For example, if we set = 2, we get
)(x) = )(a)+)(a)
(xa) +
1
2
(xa)
D
2
)(a) (xa) + O
_
[x a[
3
_
.
577
Bibliography
[Asm02] Nakhle H. Asmar. Applied Complex Analysis with Partial Dierential Equations.
Prentice-Hall, Upper Saddle River, NJ, 1st edition, 2002.
[Asm05] Nakhle H. Asmar. Partial dierential equations, with Fourier series and boundary
value problems. Prentice-Hall, Upper Saddle River, NJ, 2nd edition, 2005.
[BEH94] Jiri Blank, Pavel Exner, and Miloslav Havlicek. Hilbert Space Operators in Quantum
Physics. AIP series in computational and applied mathematical science. American
Institute of Physics, New York, 1994.
[BG80] Richard L. Bishop and Samuel I. Goldberg. Tensor analysis on manifolds. Dover,
Mineola, NY, 1980.
[Bie53] L. Bieberbach. Conformal mapping. Chelsea Publishing Co., New York, 1953. Trans-
lated by F. Steinhardt.
[Boh79] David Bohm. Quantum Theory. Dover, Mineola, NY, 1979.
[Bro89] Arne Broman. Introduction to partial dierential equations. Dover Books on Ad-
vanced Mathematics. Dover Publications Inc., New York, second edition, 1989. From
Fourier series to boundary value problems.
[CB87] Ruel V. Churchill and James Ward Brown. Fourier series and boundary value prob-
lems. McGraw-Hill Book Co., New York, fourth edition, 1987.
[CB03] Ruel V. Churchill and James Ward Brown. Complex Variables and Applications.
McGraw-Hill Book Co., New York, 7th edition, 2003.
[Cha93] Isaac Chavel. Riemannian Geometry: A modern introduction. Cambridge UP, Cam-
bridge, MA, 1993.
[Cod89] Earl A. Coddington. An introduction to ordinary dierential equations. Dover, Mi-
neola, NY, 1989.
[Con90] John B. Conway. A Course in Functional Analysis. Springer-Verlag, New York,
second edition, 1990.
[CW68] R. R. Coifman and G. Wiess. Representations of compact groups and spherical
harmonics. Lenseignment Math., 14:123173, 1968.
[dZ86] Paul duChateau and David W. Zachmann. Partial Dierential Equations. Schaums
Outlines. McGraw-Hill, New York, 1986.
[Eva91] Lawrence C. Evans. Partial Dierential Equations, volume 19 of Graduate Studies
in Mathematics. American Mathematical Society, Providence, Rhode Island, 1991.
[Fis99] Stephen D. Fisher. Complex variables. Dover Publications Inc., Mineola, NY, 1999.
Corrected reprint of the second (1990) edition.
[Fol84] Gerald B. Folland. Real Analysis. John Wiley and Sons, New York, 1984.
[Hab87] Richard Haberman. Elementary applied partial dierential equations. Prentice Hall
Inc., Englewood Clis, NJ, second edition, 1987. With Fourier series and boundary
value problems.
578 DRAFT BIBLIOGRAPHY
[Hat02] Allen Hatcher. Algebraic topology. Cambridge University Press, Cambridge, 2002.
[Hel81] Sigurdur Helgason. Topics in Harmonic Analysis on Homogeneous Spaces.
Birkhauser, Boston, Massachusetts, 1981.
[Hen94] Michael Henle. A combinatorial introduction to topology. Dover, Mineola, NY, 1994.
[Kat76] Yitzhak Katznelson. An Introduction to Harmonic Analysis. Dover, New York,
second edition, 1976.
[KF75] A. N. Kolmogorov and S. V. Fomn. Introductory real analysis. Dover Publications
Inc., New York, 1975. Translated from the second Russian edition and edited by
Richard A. Silverman, Corrected reprinting.
[Kor88] T. W. Korner. Fourier Analysis. Cambridge University Press, Cambridge, UK, 1988.
[Lan85] Serge Lang. Complex Analysis. Springer-Verlag, New York, second edition, 1985.
[McW72] Roy McWeeny. Quantum Mechanics: Principles and Formalism. Dover, Mineola,
NY, 1972.
[M ul66] C. M uller. Spherical Harmonics. Number 17 in Lecture Notes in Mathematics.
Springer-Verlag, New York, 1966.
[Mur93] James D. Murray. Mathematical Biology, volume 19 of Biomathematics. Springer-
Verlag, New York, second edition, 1993.
[Nee97] Tristan Needham. Visual complex analysis. The Clarendon Press Oxford University
Press, New York, 1997.
[Neh75] Zeev Nehari. Conformal mapping. Dover Publications Inc., New York, 1975. Reprint-
ing of the 1952 edition.
[Pin98] Mark A. Pinsky. Partial Dierential Equations and Boundary-Value Problems with
Applications. International Series in Pure and Applied Mathematics. McGraw-Hill,
Boston, third edition, 1998.
[Pow99] David L. Powers. Boundary value problems. Harcourt/Academic Press, San Diego,
CA, fourth edition, 1999.
[Pru81] Eduard Prugovecki. Quantum Mechanicss in Hilbert Space. Academic Press, New
York, second edition, 1981.
[RB69] Gian-Carlo Rota and Garret Birkho. Ordinary Dierential Equations. Wiley, 2nd
edition, 1969.
[Roy88] H. L. Royden. Real analysis. Macmillan Publishing Company, New York, third
edition, 1988.
[Rud87] Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third
edition, 1987.
[Sch79] Hans Schwerdtfeger. Geometry of complex numbers. Dover Publications Inc., New
York, 1979. Circle geometry, Moebius transformation, non-Euclidean geometry, A
corrected reprinting of the 1962 edition, Dover Books on Advanced Mathematics.
[Ste95] Charles F. Stevens. The six core theories of modern physics. A Bradford Book. MIT
Press, Cambridge, MA, 1995.
[Ste08] James Stewart. Calculus, Early Transcendentals. Thomson Brooks/Cole, Belmont,
CA, 6e edition, 2008.
[Str93] Daniel W. Strook. Probability Theory: An analytic view. Cambridge University
Press, Cambridge, UK, revised edition, 1993.
[Sug75] M. Sugiura. Unitary Representations and Harmonic Analysis. Wiley, New York,
1975.
BIBLIOGRAPHY 579
[Tak94] Masaru Takeuchi. Modern Spherical Functions, volume 135 of Translations of Math-
ematical Monographs. American Mathematical Society, Providence, Rhode Island,
1994.
[Tay86] Michael Eugene Taylor. Noncommutative Harmonic Analysis. American Mathemat-
ical Society, Providence, Rhode Island, 1986.
[Ter85] Audrey Terras. Harmonic Analysis on Symmetric Spaces and Applications, volume I.
Springer-Verlag, New York, 1985.
[Tit62] E.C. Titchmarsh. Eigenfunction expansions associated with second-order dierential
equations, Part I. Oxford University Press, London, 2nd edition, 1962.
[Wal88] James S. Walker. Fourier Analysis. Oxford UP, Oxford, UK, 1988.
[War83] Frank M. Warner. Foundations of Dierentiable Manifolds and Lie Groups. Springer-
Verlag, New York, 1983.
[WZ77] Richard L. Wheeden and Antoni Zygmund. Measure and integral. Marcel Dekker Inc.,
New York, 1977. An introduction to real analysis, Pure and Applied Mathematics,
Vol. 43.
580 DRAFT
Index
dAlembert
ripple solution (initial velocity), 395
solution to wave equation, 398, 401, 531
travelling wave solution (initial position),
393
Abel mean, 461
Abel sum, 176
Abels test, 130
Abelian group, 177
Absolute convergence
of Fourier series, 174
Absolutely integrable function
on half-line 1
= [0, ), 510
on the real line 1, 488
on the two-dimensional plane 1
2
, 504
on the two-dimensional plane 1
3
, 507
Airys equation, 67, 99
Algebraic topology, 484
Analytic
harmonic functions are, 18
Analytic extension, 453
Analytic function
denition, 570
Fourier coecient decay rate, 207
Analytic functions
convolutions of, 474
Fourier transform of, 476
improper integral of, 473
Annulus, 274
Antiderivative, complex, 448
Approximation of identity
denition (on 1), 379
denition (on 1
D
), 383
Gauss-Weierstrass Kernel
many-dimensional, 393
one-dimensional, 387
on [, ], 199, 218
Poisson kernel (on disk), 409
Poisson kernel (on half-plane), 405
use for smooth approximation, 411
Atiyah-Singer Index Theorem, 484
Autocorrelation Function, 499
Baguette example, 252
Balmer, J.J, 54
BC, see Boundary conditions
Beam equation, 67, 99
Bernsteins theorem, 175
Bessel functions, 574
and eigenfunctions of Laplacian, 292
denition, 292
roots, 296
Bessels Equation, 292
Bessels equation, 574
Bessels inequality, 195
Big O notation, see Order O(1/z)
Bilinearity, 107
Binary expansion, 115
Borel-measurable set, 110
Borel-measurable subset, 211
Boundary
denition of, 71
examples (for various domains), 71
X, see Boundary
Boundary Conditions
denition, 72
Homogeneous Dirichlet, see Dirichlet Bound-
ary Conditions, Homogeneous
Homogeneous Mixed, see Mixed Bound-
Homogeneous Neumann, see Neumann Bound-
Homogeneous Robin, see Mixed Bound-
Nonhomogeneous Dirichlet, see Dirichlet
Boundary Conditions, Nonhomoge-
neous
Nonhomogeneous Mixed, see Mixed Bound-
ary Conditions, Nonhomogeneous
Nonhomogeneous Neumann, see Neumann
Boundary Conditions, Nonhomoge-
neous
Nonhomogeneous Robin, see Mixed Bound-
ary Conditions, Nonhomogeneous
Periodic, see Periodic Boundary Condi-
tions
Boundary conditions
and harmonic conjugacy, 421
Boundary value problem, 72
Bounded variation, 175
Branches
of complex logarithm, 449
of complex roots, 449
Brownian Motion, 21
Burgers equation, 67, 99
BVP, see Boundary value problem
(
1
[0, L], 145
(
1
[0, ], 137
(
1
interval, 138, 145
Casorati-Weierstrass Theorem, 471
Cauchy problem, see Initial value problem
Cauchy Residue Theorem, see Residue Theo-
rem
Cauchys Criterion, 130
Cauchys integral formula, 443
Cauchys Theorem
on contours, 438
on oriented boundaries, 482
Cauchy-Bunyakowski-Schwarz Inequality
for complex functions, 109
for sequences in l
2
(N), 204
in L
2
, 108
Cauchy-Euler Equation
as Sturm-Liouville equation, 344
polar eigenfunctions of . (2 dimensions),
314
zonal eigenfunctions of . (3 dimensions),
361
Cauchy-Riemann dierential equations, 416
Cauchy-Schwarz inequality, see Cauchy-Bunyakowski-
Schwarz Inequality
CBS inequality, see Cauchy-Bunyakowski-Schwarz
Inequality
Cesaro sum, 176
Character (of a group), 177
Chasm in streambed (ow), 434
Chebyshev polynomial, 278
Codisk, 273
Compact abelian topological group, 177
Complex nth root, 449
Complex antiderivative, 448
as complex potential, 430
Complex derivative, 415
Complex logarithm, 449
Complex numbers
addition, 551
conjugate, 553
exponential, 551
derivative of, 551
multiplication, 551
norm, 553
polar coordinates, 551
Complex potentials, 430
Complex-analytic, see Holomorphic function
Complex-dierentiable, 415
Componentwise addition, 57
Conformal
holomorphic, 423
Conformal isomorphism, 423
Conformal map
denition, 422
Riemann Mapping Theorem, 429
Connected
denition, 451
Conservation of energy
in wave equation, 92
Continuously dierentiable, 137, 144
Contour, 434
piecewise smooth, 436
purview of, 437
smooth, 435
Contour integral, 435, 436
Convergence
as approximation, 117
in L
2
, 118
of complex Fourier series, 173
of Fourier cosine series, 142, 146
of Fourier series; Bernsteins Theorem, 175
of Fourier sine series, 138, 145
of function series, 129
of multidimensional Fourier series, 187
of real Fourier series, 162
of two-dimensional Fourier (co)sine series,
183
of two-dimensional mixed Fourier series,
185
pointwise, 120
pointwise = L
2
, 120
semiuniform, 128
uniform, 125
Uniform = pointwise, 127
Uniform = L
2
, 127
Convolution
= multiplication of Fourier coecients,
463
of 2-periodic functions, 214
of analytic functions, 474
with the Dirichlet kernel, 197, 464
with the Poisson kernel, 462
convolution
continuity of, 410
denition of (f g), 17, 378
dierentiation of, 410
is associative (f (g h) = (f g) h),
409
is commutative (f g = g f), 378, 409
is distributive (f (g + h) = (f g) +
(f h)), 409
use for smooth approximation, 411
Convolution ring, 217
Coordinates
cylindrical, 555
polar, 554
rectangular, 554
spherical, 556
Cosine series, see Fourier series, cosine
Coulomb potential (electrostatics), 14
Coulombs Law (electrostatics), 15
Countour integral
is homotopy-invariant, 440
Curl, 418
Cycle (in homology), 482
Cylindrical coordinates, 555
., see Laplacian
dAlembert
ripple solution (initial velocity), 395
solution to wave equation, 398, 401, 531
travelling wave solution (initial position),
393
Davisson, C.J., 37
de Broglie, Louis
matter wave hypothesis, 37
de Broglie wavelength, 43
Decay at of order O(1/z), see Order O(1/z)
Decay at of order o(1/z), see Order o(1/z)
Decaying gradient condition, 283
Dense subspace of L
2
[, ], 207
X, see Boundary
u, see Outward normal derivative

Dierence operator, 60
Dierentiation as linear operator, 61
Diraction of matter waves, 37
Dirac delta function 0, 379, 404
Dirac delta function 0, 528
Dirichlet Boundary Conditions
Homogeneous
2-dim. Fourier sine series, 183
denition, 73
Fourier sine series, 138
multidim. Fourier sine series, 187
physical interpretation, 73
Nonhomogeneous
denition, 75
Dirichlet kernel, 197, 464
Dirichlet problem
around chasm, 433
denition, 75
on a half-disk, 433
on annulus
Fourier solution, 287
on bi-innite strip, 432
on codisk
on cube
nonconstant, nonhomog. Dirichlet BC,
270
one constant nonhomog. Dirichlet BC,
270
on disk
denition, 406
Poisson (impulse-response) solution, 290,
407
on half-plane
denition, 403, 537
Poisson (impulse-response) solution, 404,
539
on half-plane with vertical obstacle, 433
on interval [0, L], 76
on o-centre annulus, 433
on quarter-plane
conformal mapping solution, 428
on square
four constant nonhomog. Dirichlet BC,
243
nonconstant nonhomog. Dirichlet BC,
244
one constant nonhomog. Dirichlet BC,
241
two compartments separated by aperture,
433
unique solution of, 87
Dirichlet test, 130
Distance
L
2
, 117
L
, 124
uniform, 124
Divergence (div V )
in many dimensions, 560
in one dimension, 558
in two dimensions, 559
Divergence theorem, 563
Dot product, 103
Drumskin
round, 302
square, 259, 261
-tube, 124
Eigenfunction
denition, 63
of dierentiation operator, 158, 168
of Laplacian, 63, 67
polar-separated, 292
polar-separated; homog. Dirichlet BC,
296
Eigenfunctions
of
2
x
, 345
of self-adjoint operators, 345
of the Laplacian, 347
Eigenvalue
denition, 63
of Hamiltonian as energy levels, 46
Eigenvector
denition, 63
of Hamiltonian as stationary quantum states,
46
Eikonal equation, 67, 99
Electric eld, 15
Electric eld lines, 431
Electrostatic potential, 14, 431
Elliptic dierential equation, 98
motivation: polynomial formalism, 371
two-dimensional, 96
Elliptic dierential operator
denition, 96, 98
divergence form, 349
self-adjoint
eigenvalues of, 349
if symmetric, 349
symmetric, 349
Entire function, 471
Equipotential contour, 430
Error function , 390
Essential singularity, 471
Eulers formula, 551
Even extension, 170
Even function, 168
Even-odd decomposition, 169
Evolution equation, 69
Extension
even, see Even extension
odd, see Odd extension
odd periodic, see Odd Periodic Extension
(error function or sigmoid function), 390
Factorial, 569
gamma function, 520
Field of fractions, 472
Flow
along river bank, 434
around peninsula, 434
conned to domain, 430
irrotational, 418
out of pipe, 434
over chasm, 434
sourceless, 418
sourceless and irrotational, 418, 430
Fluid
incompressible and nonturbulent, 430
Fluid dynamics, 430
Flux across boundary
in 1
2
, 562
in 1
D
, 563
Fokker-Plank equation, 19
is homogeneous linear, 64
is parabolic PDE, 98
Forced heat equation
Forced wave equation
Fourier (co)sine transform
denition, 510
inversion, 510
Fourier cosine series, see Fourier series, cosine
Fourier series
absolute convergence, 174
convergence; Bernsteins theorem, 175
failure to converge pointwise, 175
Fourier series, (co)sine
of derivative, 158
of piecewise linear function, 156
of polynomials, 148
of step function, 153
relation to real Fourier series, 171
Fourier series, complex
coecients, 172
convergence, 173
denition, 172
relation to real Fourier series, 174
Fourier series, cosine
coecents
on [0, ], 141
on [0, L], 146
convergence, 142, 146
denition
on [0, ], 141
on [0, L], 146
is even function, 169
of f(x) = cosh(x), 143
of f(x) = sin(mx/L), 146
of f(x) = sin(mx), 143
of f(x) = x, 148
of f(x) = x
2
, 148
of f(x) = x
3
, 148
of f(x) 1, 143, 146
of half-interval, 154
Fourier series, multidimensional
complex, 191
convergence, 187
cosine
coecients, 186
series, 186
Derivatives of, 192
mixed
coecients, 187
series, 187
sine
coecients, 186
series, 186
Fourier series, real
coecents, 161
convergence, 162
denition, 161
of f(x) = x, 164
of f(x) = x
2
, 164
of derivative, 168
of piecewise linear function, 167
of polynomials, 163
of step function, 165
relation to complex Fourier series, 174
relation to Fourier (co)sine series, 171
Fourier series, sine
coecents
on [0, ], 137
on [0, L], 144
convergence, 138, 145
denition
on [0, ], 137
on [0, L], 144
is odd function, 169
of f(x) = cos(mx/L), 145
of f(x) = cos(mx), 140
of f(x) = sinh(x/L), 145
of f(x) = sinh(x), 140
of f(x) = x, 148
of f(x) = x
2
, 148
of f(x) = x
3
, 148
of f(x) 1, 140, 145
of tent function, 155, 159
Fourier series, two-dimensional
convergence, 183
cosine
coecients, 182
denition, 182
sine
coecients, 179
denition, 179
of f(x, y) = x y, 179
of f(x, y) 1, 182
Fourier series, two-dimensional, mixed
coecients, 185
convergence, 185
denition, 185
Fourier sine series, see Fourier series, sine
Fourier transform
D-dimensional
inversion, 508
asymptotic decay, 492
convolution, 494
D-dimensional
denition, 508
derivative of, 496
evil twins of, 500
is continuous, 492
of analytic function, 476
one-dimensional
denition, 488
inversion, 488, 491
of box function, 489
of Gaussian, 497
of Poisson kernel (on half-plane), 539
of rational functions, 479
of symmetric exponential tail function,
491
rescaling, 495
smoothness vs. asymptotic decay, 497
three-dimensional
denition, 507
inversion, 507
of ball, 507
translation vs. phase shift, 494
two-dimensional
denition, 504
inversion, 504, 505
of box function, 505
of Gaussian, 506
Fouriers Law of Heat Flow
many dimensions, 4
one-dimension, 4
Fourier-Bessel series, 296
Frequency spectrum, 53
Frobenius, method of, 573
to solve Bessel equation, 306
Fuchs power series solution to ODE, 573
Fuel rod example, 255
Functions as vectors, 58
Fundamental solution, 385
heat equation (many-dimensional), 393
heat equation (one-dimensional), 388
Fundamental theorem of calculus, 559
as special case of Divergence Theorem,
563
2
, see Laplacian
Gamma function, 520
Gauss theorem, see Divergence theorem
Gausss Law (electrostatics), 15
Gauss-Weierstrass Kernel
convolution with, see Gaussian Convolu-
tion
many-dimensional
denition, 8
is approximation of identity, 393
one-dimensional, 385, 528
denition, 6
two-dimensional, 8
Gaussian
one-dimensional
cumulative distribution function of, 390
integral of, 390
stochastic process, 21
two-dimensional
Gaussian Convolution, 388, 392, 530
General Boundary Conditions, 82
Generation equation, 12
equilibrium of, 13
Generation-diusion equation, 12
Germer, L.H, 37
Gibbs phenomenon, 140, 145, 152
Gradient u
two-dimensional, 557
Gradient vector eld
Gravitational potential, 14
Greens function, 379
Greens Theorem, 562
as special case of Divergence Theorem,
563
Guitar string, 71
Holder continuous, 174
Haar basis, 115
Hamiltonian operator
eigenfunctions of, 46
in Schrodinger equation, 41
is self-adjoint, 343
Harmonic locally holomorphic, 417
Harmonic analysis, 177
noncommutative, 350
Harmonic conjugate, 417
swaps Neumann and Dirichlet BC, 421
Harmonic function
saddle shape, 10
analyticity, 18
convolution against Gauss-Weierstrass, 413
denition, 10
Maximum Principle, 17
Mean value theorem, 16, 316, 413
separated (Cartesian), 356
smoothness properties, 18
two-dimensional
separated (Cartesian), 354
two-dimensional, separated (polar coordi-
nates), 274
Harp string, 231
Hausdor-Young inequality, 503
HDBC, see Dirichlet Boundary Conditions, Ho-
mogeneous
Heat equation
denition, 8
derivation and physical interpretation
many dimensions, 7
one-dimension, 5
equilibrium of, 9
fundamental solution of, 388, 393
Initial conditions: Heaviside step func-
tion, 388
is evolution equation., 69
is parabolic PDE, 96, 98
norm decay, 89
on 2-dim. plane
Fourier transform solution, 528
on 3-dim. space
on cube; Homog. Dirichlet BC
on cube; Homog. Neumann BC
on disk; Homog. Dirichlet BC
Fourier-Bessel solution, 300
on disk; Nonhomog. Dirichlet BC
on interval; Homog. Dirichlet BC
on interval; Homog. Neumann BC
on real line
Gaussian Convolution solution, 388, 530
on square; Homog. Dirichlet BC
on square; Homog. Neumann BC
on square; Nonhomog. Dirichlet BC
on unbounded domain
Gaussian Convolution solution, 392
Heaviside step function, 388
Heisenberg Uncertainty Principle, see Uncer-
tainty Principle
Heisenberg, Werner, 513
Helmholtz equation, 67, 99
is not evolution equation., 70
Hermitian, 109
Hessian derivative, 26
Hessian derivative matrix, 576
HNBC, see Neumann Boundary Conditions,
Homogeneous
Holomorphic
conformal, 423
sourceless irrotational ow, 418
complex-analytic, 450
harmonic, 417
function, 416
Homogeneous Boundary Conditions
Dirichlet, see Dirichlet Boundary Condi-
tions, Homogeneous
Mixed, see Mixed Boundary Conditions,
Homogeneous
Neumann, see Neumann Boundary Con-
ditions, Homogeneous
Robin, see Mixed Boundary Conditions,
Homogeneous
Homogeneous linear dierential equation
denition, 64
superposition principle, 65
Homologous (cycles), 482
Homology group, 484
Homology invariance (of chain integrals), 483
Homotopic (contours), 439
Homotopy invariance
of contour integration, 440
Huygens Principle, 536
Hydrogen atom
Balmer lines, 53
Bohr radius, 52
energy spectrum, 53
frequency spectrum, 53
ionization potential, 52
Schrodinger equation, 42
Stationary Schrodinger equation, 51
Hyperbolic dierential equation, 98
one-dimensional, 96
I/BVP, see Initial/Boundary value problem
Ice cube example, 267
Identity Theorem, 452
Imperfect Conductor (Robin BC), 81
Impermeable barrier (Homog. Neumann BC.,
77
Improper integral
of analytic functions, 473
Impulse function, 379
Impulse-response function
four properties, 375
interpretation, 375
Impulse-response solution
to Dirichlet problem on disk, 290, 407
to half-plane Dirichlet problem, 404, 539
to heat equation, 392
to heat equation (one dimensional), 388
to wave equation (one dimensional), 398
Indelible singularity, 464
indicial polynomial, 573
indicial roots, 573
, see Point at innity
Initial conditions, 70
Initial position problem, 230, 259, 302, 393
Initial value problem, 70
Initial velocity problem, 232, 261, 302, 395
Initial/Boundary value problem, 72
Inner product
of complex functions, 109
of functions, 105, 107
of functions (complex-valued), 172
of vectors, 103
int (X), see Interior
Integral domain, 471
Integral representation formula, 447
Integration as linear operator, 62
Integration by parts, 147
Interior (of a domain), 71
Irrotational ow, 418
IVP, see Initial value problem
Jordan Curve Theorem, 437
Kernel
convolution, see Impulse-response func-
tion
Gauss-Weierstrass, see Gauss-Weierstrass
Kernel
Poisson
on disk, see Poisson Kernel (on disk)
on half-plane, see Poisson Kernel (on
half-plane)
Kernel of linear operator, 63
L
2
norm (|f|
2
), see Norm, L
2
L
2
-convergence, see Convergence in L
2
L
2
-distance, 117
L
2
-norm, 118
L
2
-space, 40, 106, 107
L
-convergence, see Convergence, uniform

L
-distance, 124
L
-norm (|f|
), 123
L
1
(1
), 510
L
1
(1), 488
L
1
(1
2
), 504
L
1
(1
3
), 507
L
1
(1
D
), 508
L
2
(X), 40, 106, 107
L
2
even
[, ], 169
L
2
odd
[, ], 169
Landau big O notation, see Order O(1/z)
Landau small o notation, see Order o(1/z)
Laplace equation
denition, 9
is elliptic PDE, 96, 98
nonhomogeneous Dirichlet BC, see Dirich-
let Problem
on codisk
on codisk; homog. Neumann BC
on disk; homog. Neumann BC
one-dimensional, 10
polynomial formalism, 370
quasiseparated solution, 369
separated solution (Cartesian), 356
separated solution of, 12
three-dimensional, 11
two-dimensional, 10
separated solution (Cartesian), 354, 370
separated solution (polar coordinates),
274
Laplace transform, 515
Laplace-Beltrami operator, 21, 350
Laplacian, 7
eigenfunctions (polar-separated), 292
eigenfunctions (polar-separated) homog.
Dirichlet BC, 296
eigenfunctions of, 347
in polar coordinates, 274
is linear operator, 62
spherical mean formula, 16, 25
Laurent expansion, 465
Lebesgue integral, 110, 211
Lebesgues Dominated Convergence Theorem,
121
Left-hand derivative (f
)
(x)), 201
Left-hand limit (lim
y,x
f(y)), 201
Legendre Equation, 361
Legendre equation
Legendre polynomial, 362
Legendre series, 368
Leibniz rule
for divergence, 561
for gradients, 558
for Laplacians, 9
for normal derivatives, 565
lim
y,x
f(y), see Left-hand limit
lim
y_x
f(y), see Right-hand limit
Linear dierential operator, 62
Linear function, see Linear operator
Linear operator
denition, 60
kernel of, 63
Linear transformation, see Linear operator
Liouvilles equation, 19
Liouvilles theorem, 447
Logarithm, complex, 449
L
p
norm
on [, ], 175
on 1, 503
L
p
(1), 503
L
p
[, ], 175
Maclaurin series, 570
derivatives of, 570
Maximum Principle, 17
Mean Value Theorem
for harmonic functions, 445
for holomorphic functions, 444
Mean value theorem, 316, 413
for harmonic functions, 16
Meromorphic function, 466
Method of Frobenius, see Frobenius, method
of
Minkowskis Inequality, 213
Mixed Boundary Conditions
Homogeneous
denition, 81
Nonhomogeneous
as Dirichlet, 81
as Neumann, 81
denition, 81
Mollier, 220
Monge-Ampère equation, 67, 99
Multiplication operator
continuous, 62
discrete, 61
2
, see Laplacian
Negative denite matrix, 96, 98
Neumann Boundary Conditions
Homogeneous
2-dim. Fourier cosine series, 183
denition, 77
Fourier cosine series, 142
multidim. Fourier cosine series, 187
Nonhomogeneous
denition, 80
Neumann Problem
denition, 80
Neumann problem
Newtons law of cooling, 81
Nonhomogeneous Boundary Conditions
Dirichlet, see Dirichlet Boundary Condi-
tions, Nonhomogeneous
Mixed, see Mixed Boundary Conditions,
Nonhomogeneous
Neumann, see Neumann Boundary Con-
ditions, Nonhomogeneous
Robin, see Mixed Boundary Conditions,
Nonhomogeneous
Nonhomogeneous linear dierential equation
denition, 65
subtraction principle, 66
Norm
L
2
(|f|
2
), 40, 106, 107, 109
of a vector, 104
uniform (|f|
), 123
Norm decay
in heat equation, 89
Normal derivative, see Outward normal deriva-
tive
Normal vector
in 1
D
, 562
Normal vector eld
in 1
2
, 562
Nullhomotopic, 437
O(1/z), see Order O(1/z)
o(1/z), see Order o(1/z)
Ocean pollution, 403
Odd extension, 170
Odd function, 169
Odd periodic extension, 399
One-parameter semigroup, 413, 525
Open source, xiii
Order
of dierential equation, 70
of dierential operator, 70
Order O(1/z), 476
Order o(1/z), 472
ordinary point for ODE, 573
Oriented boundary (of a subset of C), 482
Orthogonal
basis, see Orthogonal basis
eigenfunctions of self-adjoint operators, 345
functions, 112
set, see Orthogonal set
set of functions, 112
trigonometric functions, 112, 113
vectors, 103
Orthogonal basis
eigenfunctions of Laplacian, 347
for L
2
([0, X] [0, Y ]), 183, 185
for L
2
([0, X1] ... [0, XD]), 187
for L
2
(D), using Fourier-Bessel functions,
297
for L
2
[, ]
using (co)sine functions, 162
for L
2
[0, ]
using cosine functions, 142
using sine functions, 138
for even functions Leven[, ], 170
for odd functions L
odd
[, ], 170
of functions, 131
Orthogonal set
of functions, 131
Orthonormal basis
for L
2
[L, L]
using exp(inx) functions, 173
of functions, 131
of vectors, 104
Orthonormal set of functions, 112
Ostrogradskys Theorem, see Divergence the-
orem
Outward normal derivative
denition in special cases, 76
Outward normal derivative (
u)
examples (various domains), 76
Outward normal derivative (
u)
Outward normal derivative
u)
abstract denition, 564
Parabolic dierential equation, 98
one-dimensional, 96
Parsevals equality
for Fourier transforms, 502
for orthonormal bases, 132
for vectors, 104
X, see Boundary
u, see Outward normal derivative

Peninsula (ow), 434
Perfect Conductor (Homog. Dirichlet BC., 73
Perfect Insulator (Homog. Neumann BC., 77
Perfect set
deniton, 451
Periodic Boundary Conditions
complex Fourier series, 173
denition
on cube, 84
on interval, 82
on square, 83
interpretation
on interval, 82
on square, 83
real Fourier series, 162
(error function or sigmoid function), 390
Piano string, 71
Piecewise (
1
, 138, 145
Piecewise continuously dierentiable, 138, 145
Piecewise linear function, 155, 167
Piecewise smooth boundary, 85
in 1
2
, 561
in 1
D
, 563
Pipe into lake (ow), 434
Plancherels theorem, 503
Plucked string problem, 230
Point at innity, 465
Pointwise convergence, see Convergence, point-
wise
Poisson equation
denition, 13
electrostatic potential, 14
is elliptic PDE, 98
is nonhomogeneous, 66
on cube; Homog. Dirichlet BC
on cube; Homog. Neumann BC
on disk; Homog. Dirichlet BC
on disk; nonhomog. Dirichlet BC
on interval; Homog. Dirichlet BC
on interval; Homog. Neumann BC
on square; Homog. Dirichlet BC
on square; Homog. Neumann BC
on square; Nonhomog. Dirichlet BC
one-dimensional, 13
Poisson Integral Formula
for harmonic functions on disk, 290
for holomorphic functions on disk, 445
Poisson kernel
and Abel mean of Fourier series, 462
Fourier series of, 463
Poisson kernel (on disk)
denition, 290, 407
in complex plane, 445
in polar coordinates, 290, 407
is approximation of identity, 409, 464
picture, 407
Poisson kernel (on half-plane)
denition, 404, 539
picture, 404
Poisson solution
to Dirichlet problem on disk, 290, 407
to half-plane Dirichlet problem, 404, 433,
539
to three-dimensional wave equation, 534
Poissons equation
Polar coordinates, 554
Pole, 465
simple, 464
Pollution, oceanic, 403
Polynomial formalism
denition, 369
elliptic, parabolic & hyperbolic, 371
Laplace equation, 370
telegraph equation, 371, 372
Polynomial symbol, 370
Positive denite matrix, 96, 97
Positive-denite
inner product is, 107, 109
Potential, 14
complex, 430
Coulomb, 14
electrostatic, 14, 431
gravitational, 14
of a ow, 430
Potential elds and Poissons equation, 14
Power series, 569
Power series method, 571
to solve Legendre equation, 364
Power spectrum, 500
Punctured plane, 273
Purview (of a contour), 437
Pythagorean formula
in 1
N
, 104
in L
2
, 131
Quantization of energy
hydrogen atom, 53
in nite potential well, 48
in innite potential well, 50
Quantum numbers, 50
Quasiseparated solution, 369
of Laplace equation, 369
Reaction kinetic equation, 19
Reaction-diusion equation, 20, 67, 99
is nonlinear, 66
Rectangular coordinates, 554
regular singular point for ODE, 573
Removable singularity, 464
Residue, 465
Residue Theorem, 467
Riemann integrable function, 209
Riemann integral
of bounded function on [, ], 209
of step function on [, ], 208
of unbounded function on [, ], 210
Riemann Mapping Theorem, 429
Riemann Sphere, 469
Riemann surface, 449
Riemann-Lebesgue Lemma
for Fourier series, 197
for Fourier transforms, 492
Riesz-Thorin interpolation, 503
Right-hand derivative (f
(x)), 201
Right-hand limit (lim
y_x
f(y)), 200
River bank (ow), 434
Robin Boundary Conditions
Homogeneous, see Mixed Boundary Con-
ditions, Homogeneous
Nonhomogeneous, see Mixed Boundary
Conditions, Nonhomogeneous
Rodrigues Formula, 365
Root, complex, 449
Roots of unity, 449
Rydberg, J.R, 54
Scalar conservation law, 67, 99
Schrodinger Equation
abstract, 41
is evolution equation, 99
is linear, 67
momentum representation, 512
positional, 41
Schrodinger Equation, Stationary, 46
Schrodinger Equation
abstract, 64
of electron in Coulomb eld, 41
of free electron, 41
solution, 42
of hydrogen atom, 42
Schrodinger Equation, Stationary, 70
hydrogen atom, 51
of free electron, 46
potential well (one-dimensional)
nite voltage, 47
innite voltage, 49
potential well (three-dimensional), 50
Sectionally smooth, see Piecewise smooth
Self-adjoint
2
x
, 341
multiplication operators, 341
Self-adjoint operator
denition, 340
eigenfunctions are orthogonal, 345
Laplacian, 342
Sturm-Liouville operator, 343
Semidierentiable, 201
separation constant, 354, 356
Separation of variables
boundary conditions, 373
bounded solutions, 372
description
many dimensions, 355
two dimensions, 353
Laplace equation
two-dimensional, 354, 370
telegraph equation, 371, 372
Sesquilinearity, 109
Sigmoid function , 390
Simple closed curve, see Contour
Simple function, 211
Simple pole, 464
Simply connected, 429, 447
Sine series, see Fourier series, sine
singular point for ODE, 573
Singularity
essential, 471
indelible, 464
of holomorphic function, 439
pole, 464
removable, 464
Small o notation, see Order o(1/z)
Smooth approximation (of function), 411
Smooth boundary, 85
in 1
D
, 563
Smooth graph, 85
Smooth hypersurface, 85
Smoothness vs. asymptotic decay
of Fourier coecients, 206
of Fourier transform, 497
Soap bubble example, 279
Solution kernel, 379
Sourceless ow, 418
Spectral signature, 54
Spectral theory, 350
Spherical coordinates, 556
Spherical harmonics, 350
Spherical mean
denition, 24
formula for Laplacian, 16, 25
Poisson soln. to wave equation, 534
solution to 3-dim. wave equation, 534
spherically symmetric, 17
Stable family of probability distributions, 413,
525
Standing wave
one-dimensional, 30
two-dimensional, 32
Stationary Schrodinger equation
Step[, ], see Step function
Step function, 153, 164, 208
Stokes theorem, see Divergence theorem
Streamline, 430
Struck string problem, 232
Sturm-Liouville equation, 344
Sturm-Liouville operator
self-adjoint
eigenvalues of, 348
Subtraction principle for nonhomogeneous lin-
ear PDE, 66
Summation operator, 60
Superposition principle for homogeneous lin-
ear PDE, 65
Tangent (hyper)plane, 562
Tangent line, 561
Taylor polynomial
Taylor series, 569
Taylors theorem
Telegraph equation
denition, 34
polynomial formalism, 371, 372
separated solution, 371, 372
Tent function, 155, 159
Thompson, G.P, 37
Topological group, 177
Torus, 83
Total variation, 175
Trajectory
of ow, 430
Transport equation, 18
Travelling wave
one-dimensional, 31
two-dimensional, 33
Trigonometric orthogonality, 112, 113
Uncertainty Principle
Examples
electron with known velocity, 44
Normal (Gaussian) distribution, 498,
513
Uniform convergence, see Convergence, uni-
form
Abels test, 130
Cauchys Criterion, 130
Dirichlet test, 130
of continuous functions, 127
of derivatives, 127
of integrals, 127
Weierstrass M-test, 129
Uniform distance, 124
Uniform norm (|f|
), 123
Unique solution
of forced heat equation, 91
of forced wave equation, 94
of heat equation, 91
of Laplace equation, 86
of Poisson equation, 88
of wave equation, 94
to Dirichlet Problem, 87
to Neumann Problem, 87
Vector addition, 57
Velocity vector
of a contour in C, 435
Vibrating string
initial position, 230
initial velocity, 232
Violin string, 95
Voltage contour, 431
Wave equation
conservation of energy, 92
denition, 34
derivation and physical interpretation
one dimension, 30
two dimensions, 32
is hyperbolic PDE, 96, 98
on 2-dim. plane
on 3-dim. space
Huygens principle, 536
Poissons (spherical mean) solution, 534
on disk
on interval
dAlembert solution, 401
on interval; Initial position
on interval; Initial velocity
on real line
dAlembert solution, 398, 531
on real line; initial position
dAlembert (travelling wave) solution,
393
on real line; initial velocity
dAlembert (ripple) solution, 395
on square; Initial position
on square; Initial velocity
Wave vector
many dimensions, 34
two dimensions, 33
Wavefunction
phase, 44
probabilistic interpretation, 40
Wavelet basis, 115
convergence in L
2
, 118
pointwise convergence, 123
Weierstrass M-test, 129
Wind instrument, 95
X, see Boundary
Xylophone, 233
Notation
Sets and domains:
A(r, R): The 2-dimensional closed annulus of inner radius r and outer radius R: the set of
all (x, y) 1
2
such that r x
2
+y
2
R.
o
A(r, R): The 2-dimensional open annulus of inner radius r and outer radius R: the set of all
(x, y) 1
2
such that r < x
2
+y
2
< R.
B: A D-dimensional closed ball (often the unit ball centred at the origin).
B(x, ): The D-dimensional closed ball; of radius around the point x; the set of all y 1
D
such that |x y| < .
C: The set of complex numbers of the form x +yi, where x, y 1, and i is the square root
of 1.
C+: The set of complex numbers x +yi with y > 0.
C: The set of complex numbers x +yi with y < 0.
C = C |, the Riemann Sphere (the range of a meromorphic function).

|: A 2-dimensional closed disk (usually the unit disk centred at the origin).
|(R): A 2-dimensional closed disk of radius R, centred at the origin: the set of all (x, y) 1
2
such that x
2
+y
2
R.
o
|(R): A 2-dimensional open disk of radius R, centred at the origin: the set of all (x, y) 1
2
such that x
2
+y
2
< R.
|
(R): A 2-dimensional closed codisk of coradius R, centred at the origin: the set of all
(x, y) 1
2
such that x
2
+y
2
R.
o
|
(R): A 2-dimensional open codisk of coradius R, centred at the origin: the set of all
(x, y) 1
2
such that x
2
+y
2
> R.
H: A half-plane. Usually H =
_
(x, y) 1
2
; y 0
_
(the upper half-plane).
N: = |0, 1, 2, 3, . . ., the set of natural numbers.
N+: = |1, 2, 3, . . ., the set of positive natural numbers.
N
D
: The set of all n = (n1, n2, . . . , nD), where n1, . . . , nD are natural numbers.
: The empty set, also denoted |.
: The rational numbers: the set of all fractions n/m, where n, m Z, and m ,= 0.
1: The set of real numbers (e.g. 2, 3,
7 +, etc.)
1+ := (0, ) = |r 1 ; r 0.
1
:= [0, ) = |r 1 ; r 0.
1
2
: The 2-dimensional innite plane the set of all ordered pairs (x, y), where x, y 1.
1
D
: D-dimensional space the set of all D-tuples (x1, x2, . . . , xD), where x1, x2, . . . , xD 1.
Sometimes we will treat these D-tuples as points (representing locations in physical
space); normally points will be indicated in bold face, eg: x = (x1, . . . , xD). Sometimes
we will treat the D-tuples as vectors (pointing in a particular direction); then they will
be indicated with arrows, eg:

V = (V1, V2, . . . , VD).
1
D
1: The set of all pairs (x; t), where x 1
D
is a vector, and t 1 is a number. (Of
course, mathematically, this is the same as 1
D+1
, but sometimes it is useful to regard
the last dimension as time.)
1 1
: The half-space of all points (x, y) 1

2
, where y 0.
S: The 2-dimensional unit circle; the set of all (x, y) 1
2
such that x
2
+y
2
= 1.
S
D1
(x; R): The D-dimensional sphere; of radius R around the point x; the set of all y 1
D
such that |x y| = R
U, V, W usually denote open subsets of 1
D
or C.
X, Y: usually denote domains closed connected subsets of 1
D
with dense interiors.
Z: The integers |. . . , 2, 1, 0, 1, 2, 3, . . ..
Z
D
: The set of all n = (n1, n2, . . . , nD), where n1, . . . , nD are integers.
[1...D] = |1, 2, 3, . . . , D.
[0, ]: The closed interval of length ; the set of all real numbers x where 0 x .
(0, ): The open interval of length ; the set of all real numbers x where 0 < x < .
[0, ]
2
: The (closed) square; the set of all points (x, y) 1
2
where 0 x, y .
[0, ]
D
: The D-dimensional unit cube; the set of all points (x1, . . . , xD) 1
D
where 0 x
d

1 for all d [1...D].
[L, L]: The interval of all real numbers x with L X L.
[L, L]
D
: The D-dimensional cube of all points (x1, . . . , xD) 1
D
where L x
d
L for
all d [1...D].
Set operations:
int (X) The interior of the set X (i.e. all points in X not on the boundary of X).
Intersection. If X and Y are sets, then X Y := |z ; z X and z Y. If X1, . . . , XN are

sets, then
N
n=1
Xn := X1 X2 XN.
Union. If X and Y are sets, then X Y := |z ; z X or z Y. If X1, . . . , XN are sets,

then
N
_
n=1
Xn := X1 X2 XN.
Disjoint union. If X and Y are sets, then X Y means the same as X Y, but conveys
the added information that X and Y are disjoint i.e. X Y = . Likewise,
N
_
n=1
Xn :=
X1 X2 XN.
\ Dierence. If X and Y are sets, then X \ Y = |x X ; x , Y.
Spaces of Functions:
(
: A vector space of (innitely) dierentiable functions. Some examples:

(
[1
2
; 1]: The space of dierentiable scalar elds on the plane.
(
[1
D
; 1]: The space of dierentiable scalar elds on D-dimensional space.
(
[1
2
; 1
2
]: The space of dierentiable vector elds on the plane.
(
0
[0, 1]
D
: The space of dierentiable scalar elds on the cube [0, 1]
D
satisfying Dirichlet
boundary conditions: f(x) = 0 for all x [0, 1]
D
.
(
[0, 1]
D
D
satisfying Neumann
f(x) = 0 for all x [0, 1]

D
.
(
h
[0, 1]
D
D
satisfying mixed bound-
ary conditions:

f
f
(x) = h(x) for all x [0, 1]
D
.
(
per
[, ]: The space of dierentiable scalar elds on the interval [, ] satisfying periodic
L
1
(1) : The set of all functions f : 1 1 such that
_
[f(x)[ dx < .
L
1
(1
2
) : The set of all functions f : 1
2
1 such that
_
[f(x, y)[ dx dy < .

L
1
(1
3
) : The set of all functions f : 1
3
1 such that
_
R
3
[f(x)[ dx < .
L
2
(X) : The set of all functions f : X 1 such that |f|
2
=
__
X
[f(x)[
2
dx
_
1/2
< .
L
2
(X; C) : The set of all functions f : X C such that |f|
2
=
__
X
[f(x)[
2
dx
_
1/2
< .
Derivatives and Boundaries:
k
f =
df
dx
k
.
f = (1f, 2f, . . . , Df), the gradient of scalar eld f.
div f = 1f1 + 2f2 + . . . + DfD, the divergence of vector eld f.
f is the derivative of f normal to the boundary of some region. Sometimes this is written
as
f
n
or
f
, or as f n.
.f =
2
1
f +
2
2
f +. . . +
2
D
f. Sometimes this is written as
2
f.
L f sometimes means a general linear dierential operator L being applied to the function
f.
SLs,q() = s
2
+ s
/
+ q . Here, s, q : [0, L] 1 are predetermined functions, and
: [0, L] 1 is the function we are operating on by the Sturm-Liouville operator
SLs,q.
= (
/
1
, . . . ,
/
D
) is the velocity vector of the path : 1 1
D
.
X: If X 1
D
is some region in space, then X is the boundary of that region. For example:
[0, 1] = |0, 1.
B
2
(0; 1) = S
2
(0; 1).
B
D
(x; R) = S
D
(x; R).
(1 1
) = 1 |0.
Norms and Inner products:
|x|: If x 1
D
is a vector, then |x| =
_
x
2
1
+x
2
2
+. . . +x
2
D
is the norm (or length) of x.
|f|
2
: Let X 1
D
be a bounded domain, with volume M =
_
X
1dx. If f : X 1 is an
integrable function, then |f|
2
=
1
M
__
X
[f(x)[ dx
_
1/2
is the L
2
-norm of f.
f, g): If f, g : X 1 are integrable functions, then their inner product is given by:
f, g) =
1
M
_
X
f(x) g(x) dx.
|f|
1
: Let X 1
D
be any domain. If f : X 1 is an integrable function, then |f|
=
_
X
[f(x)[ dx is the L
1
-norm of f.
|f|
: Let X 1
D
be any domain. If f : X 1 is a bounded function, then |f|
=
sup
xX
[f(x)[ is the L
-norm of f.
Other Operations on Functions:
An: normally denotes the nth Fourier cosine coecient of a function f on [0, ] or [, ].
That is, An :=
2
0
f(x) cos(nx) dx or An :=
1
f(x) cos(nx) dx.

An,m: normally denotes a 2-dimensional Fourier cosine coecient, while An normally denotes
a D-dimensional Fourier cosine coecient.
Bn: normally denotes the nth Fourier cosine coecient of a function f on [0, ] or [, ].
That is, An :=
2
0
f(x) sin(nx) dx or An :=
1
f(x) sin(nx) dx.

Bn,m: normally denotes a 2-dimensional Fourier sine coecient, while Bn normally denotes a
D-dimensional Fourier sine coecient.
f g: If f, g : 1
D
1, then their convolution is the function f g : 1
D
1 dened by
(f g)(x) =
_
R
D
f(y) g(x y) dy.
f() =
1
(2)
D
_
R
D
f(x) exp(i x) dx is the Fourier transformof the function f : 1
D
C.
It is dened for all 1
D
.
fn =
1
2
_

f(x) exp(i nx) dx is the nth complex Fourier coecient of a function f :

[, ] C (here n Z).
f : =
_
S
0
f[(s)] (s) ds is a chain integral. Here, f : U C is some complex-valued
function and : [0, S] U is a chain (a piecewise-continuous, piecewise dierentiable
path).
_
f : A contour integral. The same denition as the chain integral

_
f, but is a contour.
: If and are two chains, then represents the linking of the two chains.
/[f] =
_

0
f(t) e
ts
dt is the Laplace transform of the function f : 1
C; it is dened
for all s C with Re [s] > , where is the exponential order of f.
M
R
u(x) =
1
4R
2
_
S(R)
f(x + s) ds is the spherical average of f at x, of radius R. Here,
x 1
3
is a point in space, R > 0, and S(R) =
_
s 1
3
; |s| = R
_
.
Special functions.:
Cn(x) = cos(nx) for all n N and x [, ].
Cn,m(x) = cos(nx) cos(my) for all n, m N and (x, y) [, ]
2
.
Cn(x) = cos(n1x1) cos(nDxD) for all n N
D
and x [, ]
D
.
DN(x) = 1 + 2
N
n=1
cos(nx) is the nth Dirichlet kernel, for all n N and x [2, 2].
En(x) = exp(i nx) for all n Z and x [, ].
c(x) = exp(i x) for all 1 and x [, ].
((x; t) =
1
2
t
exp
_
x
2
4t
_
is the (one-dimensional) Gauss-Weierstrass kernel.
((x, y; t) =
1
4t
exp
_
x
2
y
2
4t
_
is the (two-dimensional) Gauss-Weierstrass kernel.
((x; t) =
1
(4t)
D/2
exp
_
jxj
2
4t
_
is the (D-dimensional) Gauss-Weierstrass kernel.
n is the nth Bessel function of the rst kind.
ly(x) =
y
(x
2
+y
2
)
is the half-plane Poisson kernel, for all x 1 and y > 0.
N(x) is the outward unit normal vector to a domain X at a point x X.

1(x, s) =
R
2
|x|
2
|x s|
2
is the Poisson kernel on the disk, for all x | and s S.
Pr(x) =
1 r
2
1 2r cos(x) +r
2
is the Poisson kernel in polar coordinates, for all x [2, 2]
and r < 1.
n, n, n and n refer to the harmonic functions on the unit disk which separate in polar
coordinates. 0(r, ) = 1 and 0(r, ) = log(r), while for all n 1, we have n(r, ) =
cos(n) r
n
, n(r, ) = sin(n) r
n
, n(r, ) =
cos(n)
r
n
, and n(r, ) =
sin(n)
r
n
.
n,
,
n,
,
n,
, and
n,
refer to eigenfunctions of the Laplacian on the unit disk which
separate in polar coordinates. For all n N and > 0,
n,
(r, ) = n( r) cos(n),
n,
(r, ) = n( r) sin(n),
n,
(r, ) = }n( r) cos(n), and
n,
(r, ) =
}n( r) sin(n)
Sn(x) = sin(nx) for all n N and x [, ].
Sn,m(x) = sin(nx) sin(my) for all n, m N and (x, y) [, ]
2
.
Sn(x) = sin(n1x1) sin(nDxD) for all n N
D
and x [, ]
D
.
}n is the nth Bessel function of the second kind.

Cambridge PDE

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Cambridge PDE

Uploaded by

Copyright:

Available Formats

Linear Partial Differential

Properties of convolution . . . . . . . . . . . . . . . . . . . . . . 409

Abel means and Poisson kernels . . . . . . . . . . . . . . . . . . 461

Parseval and Plancherel . . . . . . . . . . . . . . . . . . . . . . 502

Momentum representation & Heisenberg uncertainty . . . . . . 511

Laplace transforms . . . . . . . . . . . . . . . . . . . . . . . . . . 515

By changing to the appropriate time units, we can assume = 1, so the heat

(r) = Q(r) = 1,r

Notice that the electric/gravitational potential eld is not uniquely dened

)(s) ds, for some constant 1 0.

Transport and diusion

Reaction and diusion 19

Reaction and diusion

be the surface area of the sphere. For example:

)(0) is the average value of )(s) over all

n(r), where n(r) is the height of bead

is to remind us that this is just an c-bead

:= c. Thus, by Newtons law, the vertical acceleration of bead r must

, then it experiences acceleration o = 1,:

, and each of these is expressed by a formula

The major mathematical problems of quantum mechanics come down to nd-

. This is the so-called de Broglie wavelength of an electron

Remark. (The meaning of phase) At any point x in space and moment t in

21 1) = 0. (Figure 3C.2) (3C.9)

21 is an integer, which means 21 0, 1, 4, 9, 16, 25, . . ., which means

Exercise 3C.2. (a) Check that eqn.(3C.11) is a solution for eqn.(3C.10). E _

, and let u := v+w =

, then Ax = 3Av. That is:

Remark Notice that the transformation 1 is a left-inverse to the transforma-

Recommended: 1B, 1C, 2B.

is called a linear operator if, for any two

, we have L() +p) = L()) + L(p), and, for any

(1; 1). For example, sin and cos are elements of (

(1; 1), and we

(1; 1), and we have

by: [)] = ). In other words, for all x 1

(1). However, D is not a right-inverse for S, because in general

in terms of multiplication by some

(1; 1). The kernel

, then an eigenvector of L is sometimes called an

is a linear dierential operator, then the equation Ln 0

(X; 1), and L = .

(X 1; 1), and, for any n (

are solutions to the

is some xed nonzero function, then the equation Lj is

is not a linear operator, then a PDE of the form Pn 0

A boundary value problem (BVP) is a problem of the following kind:

Exercise 5C.1. (i) Verify examples (b) to (e) above E _

n(s) to denote the outward

n(s) is the rate of change

n(r, j) is not well-dened).

We say that n satises homogeneous Neumann boundary conditions

n(s) = 0 for all s X. (5C.5)

n(s) is the component of n(s) which is per-

(X). The set of innitely dierentiable

(X). Thus, for example

) 0 on the boundary of the disk, so )

(|). However, ) is not smooth (it is nondierentiable at

) 0 on the boundary of the disk, and ) is smooth

(X). More generally, any nite sum

n(s) = /(s) for all s X. (5C.6)

n(s) is the ux of heat/material across

n(s) = /(r) for all s X. (5C.7)

(X) satisfying this property will be denoted