Applied Mathematical Sciences

Valery Serov

Fourier Series,
Fourier Transform and
Their Applications to
Mathematical Physics
Applied Mathematical Sciences

Volume 197

Editors
S.S. Antman, Institute for Physical Science and Technology, University of Maryland, College Park,
MD, USA
[email protected]
Leslie Greengard, Courant Institute of Mathematical Sciences, New York University, New York,
NY, USA
[email protected]
P.J. Holmes, Department of Mechanical and Aerospace Engineering, Princeton University, Princeton,
NJ, USA
[email protected]

Advisors
J. Bell, Lawrence Berkeley National Lab, Center for Computational Sciences and Engineering, Berkeley,
CA, USA
P. Constantin, Department of Mathematics, Princeton University, Princeton, NJ, USA
R. Durrett, Department of Mathematics, Duke University, Durham, NC, USA
R. Kohn, Courant Institute of Mathematical Sciences, New York University, New York, NY, USA
R. Pego, Department of Mathematical Sciences, Carnegie Mellon University, Pittsburgh, PA, USA
L. Ryzhik, Department of Mathematics, Stanford University, Stanford, CA, USA
A. Singer, Department of Mathematics, Princeton University, Princeton, NJ, USA
A. Stevens, Department of Applied Mathematics, University of Münster, Münster, Germany
A. Stuart, California Institute of Technology, Pasadena, CA, USA
S. Wright, Computer Sciences Department, University of Wisconsin, Madison, WI, USA

Founding Editors
Fritz John, Joseph P. LaSalle, Lawrence Sirovich
More information about this series at http://www.springer.com/series/34
Valery Serov

Fourier Series, Fourier

Transform and Their
Applications to Mathematical
Physics
Valery Serov
Department of Mathematical Sciences
University of Oulu
Oulu
Finland

ISSN 0066-5452 ISSN 2196-968X (electronic)

Applied Mathematical Sciences
ISBN 978-3-319-65261-0 ISBN 978-3-319-65262-7 (eBook)
DOI: 10.1007/978-3-319-65262-7

Library of Congress Control Number: 2017950249

Mathematics Subject Classification (2010): 26A16, 26A45, 35A08, 35F50, 35J05, 35J08, 35J10, 35J15,
35K05, 35K08, 35L05, 35P25, 35R30

c Springer International Publishing AG 2017

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Printed on acid-free paper
This Springer imprint is published by Springer Nature
The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface

The modern theory of analysis and differential equations in general certainly in-
cludes the Fourier transform, Fourier series, integral operators, spectral theory of
differential operators, harmonic analysis and much more. This book combines all
these subjects based on a unified approach that uses modern view on all these
themes. The book consists of four parts: Fourier series and the discrete Fourier
transform, Fourier transform and distributions, Operator theory and integral equa-
tions and Introduction to partial differential equations and it outgrew from the half-
semester courses of the same name given by the author at University of Oulu, Fin-
land during 2005–2015.
Each part forms a self-contained text (although they are linked by a common
approach) and can be read independently. The book is designed to be a modern
introduction to qualitative methods used in harmonic analysis and partial differential
equations (PDEs). It can be noted that a survey of the state of the art for all parts of
this book can be found in a very recent and fundamental work of B. Simon [35].
This book contains about 250 exercises that are an integral part of the text. Each
part contains its own collection of exercises with own numeration. They are not only
an integral part of the book, but also indispensable for the understanding of all parts
whose collection is the content of this book. It can be expected that a careful reader
will complete all these exercises.
This book is intended for graduate level students majoring in pure and applied
mathematics but even an advanced researcher can find here very useful information
which previously could only be detected in scientific articles or monographs.
Each part of the book begins with its own introduction which contains the facts
(mostly) from functional analysis used thereinafter. Some of them are proved while
the others are not.
The first part, Fourier series and the discrete Fourier transform, is devoted to
the classical one-dimensional trigonometric Fourier series with some applications
to PDEs and signal processing. This part provides a self-contained treatment of all
well known results (but not only) at the beginning graduate level. Compared with
some known texts (see [12, 18, 29, 35, 38, 44, 45]) this part uses many function
spaces such as Sobolev, Besov, Nikol’skii and Hölder spaces. All these spaces are

v
vi Preface

introduced by special manner via the Fourier coefficients and they are used in the
proofs of main results. Same definition of Sobolev spaces can be found in [35]. The
advantage of such approach is that we are able to prove quite easily the precise em-
beddings for these spaces that are the same as in classical function theory (see [1, 3,
26, 42]). In the frame of this part some very delicate properties of the trigonometric
Fourier series (Chapter 10) are considered using quite elementary proofs (see also
[46]). The unified approach allows us also to consider naturally the discrete Fourier
transform and establish its deep connections with the continuous Fourier transform.
As a consequence we prove the famous Whittaker-Shannon-Boas theorem about the
reconstruction of band-limited signal via the trigonometric Fourier series (see Chap-
ter 13). Many applications of the trigonometric Fourier series to the one-dimensional
heat, wave and Laplace equation are presented in Chapter 14. It is accompanied by a
large number of very useful exercises and examples with applications in PDEs (see
also [10, 17]).
The second part, Fourier transform and distributions, probably takes a central role
in this book and it is concerned with distribution theory of L. Schwartz and its ap-
plications to the Schrödinger and magnetic Schrödinger operators (see Chapter 32).
The estimates for Laplacian and Hamiltonian that generalize well known Agmon’s
estimates on the continuous spectrum are presented in this part (see Chapter 23).
This part can be considered as one of the most important because of numerous ap-
plications in the scattering theory and inverse problems. Here we have considered
for the first time some classical direct scattering problems for the Schrödinger op-
erator and for the magnetic Schrödinger operator with singular (locally unbounded)
coefficients including the mathematical foundations of the classical approximation
of M. Born. Also, the properties of Riesz transform and Riesz potentials (see Chap-
ter 21) are investigated very carefully in this part. Before this material could only be
found in scientific journals or monographs but not in textbooks. There is a good con-
nection of this part with Operator theory and integral equations. The main technique
applied here is the Fourier transform.
The third part, Operator theory and integral equations, is devoted mostly to the
self-adjoint but unbounded operators in Hilbert spaces and their applications to in-
tegral equations in such spaces. The advantage of this part is that many important
results of J. von Neumann’s theory of symmetric operators are collected together.
J. von Neumann’s spectral theorem allows us, for example, to introduce the heat
kernel without solving the heat equation. Moreover, we show applications of the
spectral theorem of J. von Neumann (for these operators) to the spectral theory of
elliptic differential operators. In particular, the existence of Friedrichs extension for
these operators with discrete spectrum is provided. Special attention is devoted to
the Schrödinger and the magnetic Schrödinger operators. The famous diamagnetic
inequality is proved here. We follow in this consideration B. Simon [35] (slightly
different approach can be found in [28]). We recommend (in addition to this part)
the reader get acquainted with the books [4, 13, 15, 24, 41]. As a consequence of
the spectral theory of elliptic differential operators the integral equations with weak
singularities are considered in quite simple manner not only in Hilbert spaces but
also in some Banach spaces, e.g. in the space of continuous functions on closed
Preface vii

manifolds. The central point of this consideration is the Riesz theory of compact
(not necessarily self-adjoint) operators in Hilbert and Banach spaces. In order to
keep this part short, some proofs will not be given, nor will all theorems be proved
in complete generality. For many details of these integral equations we recommend
[22]. We are able to investigate in quite simple manner one-dimensional Volterra in-
tegral equations with weak singularities in L∞ (a, b) and singular integral equations
in the periodic Hölder spaces Cα [−a, a]. Concerning approximation methods our
considerations use the general theory of bounded or compact operators in Hilbert
spaces and we follow mostly the monograph of Kress [22].
The fourth part, Introduction to partial differential equations, serves as an in-
troduction to modern methods for classical theory of partial differential equations.
Fourier series and Fourier transform play crucial role here too. An important (and
quite independent) segment of this part is the self-contained theory of quasi-linear
partial differential equations of order one. The main attention in this part is devoted
to elliptic boundary value problems in Sobolev and Hölder spaces. In particular, the
unique solvability of direct scattering problem for Helmholtz equation is provided.
We investigate very carefully the mapping and discontinuity properties of double
and single layer potentials with continuous densities. We also refer to similar prop-
erties of double and single layer potentials with densities in Sobolev spaces H 1/2 (S)
and H −1/2 (S), respectively, but will not prove any of these results, referring for their
proofs to monographs [22] and [25]. Here (and elsewhere in the book) S denotes the
boundary of a bounded domain in Rn and if the smoothness of S is not specified
explicitly then it is assumed to be such that Sobolev embedding theorem holds.
Compared with well known texts on partial differential equations some direct and
inverse scattering problems for Helmholtz, Schrödinger and magnetic Schrödinger
operators are considered in this part. As it was mentioned earlier this type of mater-
ial could not be found in textbooks. The presentation in many places of this part has
been strongly influenced by the monographs [6, 7, 11] (see also [8, 16, 24, 36, 40]).
In closing we note that this book is not as comprehensive as the fundamental
work of B. Simon [35]. But the book can be considered as a good introduction to
modern theory of analysis and differential equations and might be useful not only
to students and PhD students but also to all researchers who have applications in
mathematical physics and engineering sciences. This book could not have appeared
without the strong participation, both in content and typesetting, of my colleague
Adj. Prof. Markus Harju. Finally, a special thanks to professor David Colton from
University of Delaware (USA) who encouraged the writing of this book and who
has supported the author very much over the years.

Oulu, Finland Valery Serov

June 2017
Contents

Part I Fourier Series and the Discrete Fourier Transform

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Formulation of Fourier Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Fourier Coefficients and Their Properties . . . . . . . . . . . . . . . . . . . . . . . . . 17
4 Convolution and Parseval’s Equality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5 Fejér Means of Fourier Series. Uniqueness of the Fourier Series. . . . . 27
6 The Riemann–Lebesgue Lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
7 The Fourier Series of a Square-Integrable Function. The
Riesz–Fischer Theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
8 Besov and Hölder Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
9 Absolute convergence. Bernstein and Peetre Theorems. . . . . . . . . . . . . . 53
10 Dirichlet Kernel. Pointwise and Uniform Convergence. . . . . . . . . . . . . . 59
11 Formulation of the Discrete Fourier Transform and Its Properties. . . . 77
12 Connection Between the Discrete Fourier Transform
and the Fourier Transform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
13 Some Applications of the Discrete Fourier Transform. . . . . . . . . . . . . . . 93
14 Applications to Solving Some Model Equations . . . . . . . . . . . . . . . . . . . . 99
14.1 The One-Dimensional Heat Equation . . . . . . . . . . . . . . . . . . . . . . . . . 99
14.2 The One-Dimensional Wave Equation . . . . . . . . . . . . . . . . . . . . . . . . 113
14.3 The Laplace Equation in a Rectangle and in a Disk . . . . . . . . . . . . . 121

ix
x Contents

Part II Fourier Transform and Distributions

15 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
16 The Fourier Transform in Schwartz Space . . . . . . . . . . . . . . . . . . . . . . . . 133
17 The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2 . . . . . . . . . . . . . . . . . . . . . . . 143
18 Tempered Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
19 Convolutions in S and S . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
20 Sobolev spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
20.1 Sobolev spaces on bounded domains . . . . . . . . . . . . . . . . . . . . . . . . . 188
21 Homogeneous Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
22 Fundamental Solution of the Helmholtz Operator . . . . . . . . . . . . . . . . . . 207
23 Estimates for the Laplacian and Hamiltonian . . . . . . . . . . . . . . . . . . . . . 217

Part III Operator Theory and Integral Equations

24 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
25 Inner Product Spaces and Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . 249
26 Symmetric Operators in Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
27 John von Neumann’s spectral theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
28 Spectra of Self-Adjoint Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
29 Quadratic Forms. Friedrichs Extension. . . . . . . . . . . . . . . . . . . . . . . . . . . 313
30 Elliptic Differential Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
31 Spectral Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
32 The Schrödinger Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
33 The Magnetic Schrödinger Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
34 Integral Operators with Weak Singularities. Integral Equations
of the First and Second Kinds. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
35 Volterra and Singular Integral Equations . . . . . . . . . . . . . . . . . . . . . . . . . 371
36 Approximate Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379

Part IV Partial Differential Equations

37 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393
38 Local Existence Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
Contents xi

39 The Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421

40 The Dirichlet and Neumann Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
41 Layer Potentials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
42 Elliptic Boundary Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
43 The Direct Scattering Problem for the Helmholtz Equation . . . . . . . . . 485
44 Some Inverse Scattering Problems for the Schrödinger Operator . . . . 493
45 The Heat Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 507
46 The Wave Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 529
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
 Part I
Fourier Series and the Discrete Fourier
Transform
Chapter 1
Introduction

Definition 1.1. A function f (x) of one variable x is said to be periodic with period
T > 0 if the domain D( f ) of f contains x + T whenever it contains x and if for every
x ∈ D( f ), one has
f (x + T ) = f (x). (1.1)

Remark 1.2. If also x − T ∈ D( f ), then

f (x − T ) = f (x).

It follows that if T is a period of f , then mT is also a period for every integer m > 0.
The smallest value of T > 0 for which (1.1) holds is called the fundamental period
of f .

For example, the functions

mπ x mπ x mπ x
sin , cos , ei L , m = 1, 2, . . .
L L

are periodic with fundamental period T = 2L m . Note also that they are periodic with
common period 2L.
If some function f is defined on the interval [a, a + T ], with T > 0 and f (a) =
f (a + T ), then f can be extended periodically with period T to the whole line as

f (x) := f (x − mT ), x ∈ [a + mT, a + (m + 1)T ], m = 0, ±1, ±2, . . . .

Therefore, we may assume from now on that every periodic function is defined on
the whole line.

c Springer International Publishing AG 2017 3

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 1
4 Part I: Fourier Series and the Discrete Fourier Transform

We say that f is p-integrable, 1 ≤ p < ∞, on the interval [a, b] if

 b
| f (x)| p dx < ∞.
a

The set of all such functions is denoted by L p (a, b). When p = 1, we say that f is
integrable.
The following “continuity” in the sense of L p space, 1 ≤ p < ∞, holds: for every
f ∈ L p (a, b) and ε > 0, there is a continuous function g on [a, b] such that
 b
1/p
| f (x) − g(x)| dx p
<ε
a

(see e.g., Corollary 5.3). If f is p-integrable and g is p -integrable on [a, b], where

1 1
+ = 1, 1 < p < ∞, 1 < p < ∞ ,
p p

then their product is integrable on [a, b] and

 b  b
1/p  b
1/p

| f (x)g(x)|dx ≤ | f (x)| p dx |g(x)| p dx .
a a a

This inequality is called Hölder’s inequality for integrals. Fubini’s theorem states
that
 b  d   d  b   b d
F(x, y)dy dx = F(x, y)dx dy = F(x, y)dxdy,
a c c a a c

where F(x, y) ∈ L1 ((a, b) × (c, d)).

If f1 , f2 , . . . , fn are p-integrable on [a, b] for 1 ≤ p < ∞, then so is their sum
∑nj=1 f j , and
   p 1/p  1/p
b n  n
  b
 ∑ f j (x) dx ≤ ∑ | f j (x)| dx
p
. (1.2)

a j=1  j=1 a

This inequality is called Minkowski’s inequality. As a consequence of Hölder’s

inequality we obtain the generalized Minkowski inequality
 b  d  p 1/p  d  b 1/p
 
 F(x, y)dy dx ≤ |F(x, y)| dx
p
dy. (1.3)
a
 c  c a
1 Introduction 5

Exercise 1.1. Prove Hölder’s inequality for integrals for every 1 ≤ p < ∞.

Hint. Prove first Hölder’s inequality for sums, i.e.,

   1/p  1/p
 n  n n
  p
 ∑ a j b j  ≤ ∑ |a j | p
∑ |b j | ,
 j=1  j=1 j=1

where 1 < p < ∞, 1/p + 1/p = 1, and where for p = ∞ (or p = ∞) we consider
max1≤ j≤n |a j | (or max1≤ j≤n |b j |) instead of the corresponding sums.

Exercise 1.2. Prove (1.2) and (1.3).

Lemma 1.3. If f is periodic with period T > 0 and if it is integrable on every finite
interval, then  
a+T T
f (x)dx = f (x)dx (1.4)
a 0

for every a ∈ R.

Proof. Let first a > 0. Then

 a+T  a+T  a
f (x)dx = f (x)dx − f (x)dx
a 0 0
 T  a+T  a
= f (x)dx + f (x)dx − f (x)dx .
0 T 0

The difference in the square brackets is equal to zero due to periodicity of f . Thus,
(1.4) holds for a > 0.
If a < 0, then we proceed similarly, obtaining
 a+T  0  a+T
f (x)dx = f (x)dx + f (x)dx
a a 0
 0  T  T
= f (x)dx + f (x)dx − f (x)dx
a 0 a+T
 T  0  T
= f (x)dx + f (x)dx − f (x)dx .
0 a a+T

Again, the periodicity of f implies that the difference in brackets is zero. Thus the
lemma is proved. 

Definition 1.4. Let us assume that the domain of f is symmetric with respect to
{0}, i.e., if x ∈ D( f ), then −x ∈ D( f ). A function f is called even if

f (−x) = f (x), x ∈ D( f ),
6 Part I: Fourier Series and the Discrete Fourier Transform

and odd if
f (−x) = − f (x), x ∈ D( f ).

Lemma 1.5. If f is integrable on every finite interval and if it is even, then

 a  a
f (x)dx = 2 f (x)dx
−a 0

for every a > 0. Similarly, if f is odd, then

 a
f (x)dx = 0
−a

for every a > 0.

Proof. Since
 a  a  0
f (x)dx = f (x)dx + f (x)dx,
−a 0 −a

then on changing variables in the second integral we obtain

 a  a  a
f (x)dx = f (x)dx + f (−x)dx.
−a 0 0

The assertion of the lemma now follows from Definition 1.4. 

Definition 1.6. The notation f (c ± 0) is used to denote the right and left limits

f (c ± 0) := lim f (x).
x→c±

Definition 1.7. A function f is said to be piecewise continuous (piecewise con-

stant) on an interval [a, b] if there are x0 , x1 , . . . , xn such that a = x0 < x1 < · · · <
xn = b and
(1) f is continuous (constant) on each subinterval (x j−1 , x j ), j = 1, 2, . . . , n,
(2) f (x0 + 0), f (xn − 0), and f (x j ± 0), j = 1, 2, . . . , n − 1, exist.

Definition 1.8. A function f is said to be of bounded variation on an interval [a, b]

if there is c0 ≥ 0 such that for every {x0 , x1 , . . . , xn } with a = x0 < x1 < · · · < xn = b,
one has
n
∑ | f (x j ) − f (x j−1 )| ≤ c0 .
j=1

The number
n
Vab ( f ) := sup
x ,x ,...,xn
∑ | f (x j ) − f (x j−1 )| (1.5)
0 1 j=1
1 Introduction 7

is called the total variation of f on the interval [a, b]. For every x ∈ [a, b] we can
also define Vax ( f ) by (1.5). The class of functions of bounded variation is denoted
by BV [a, b].

Exercise 1.3. (1) Show that the bounded function

x sin 1x , x ∈ (0, 1],

f (x) =
0, x = 0,

is continuous on the interval [0, 1] but is not of bounded variation on [0, 1].
(2) Show that every piecewise constant function on [a, b] is of bounded variation.

Remark 1.9. This exercise shows that C[a, b] and BV [a, b] are not included in each
other, i.e., they represent two different scales of functions.

Exercise 1.4. Prove that

(1) Vax ( f ) is monotone increasing in x,
(2) for every c ∈ (a, b), we have Vab ( f ) = Vac ( f ) +Vcb ( f ).

If f is real-valued, then Exercise 1.4 implies that Vax ( f ) − f (x) is monotone increas-
ing in x. Indeed, for h > 0 we have that
   
Vax+h ( f ) − f (x + h) − (Vax ( f ) − f (x)) = Vax+h ( f ) −Vax ( f ) − ( f (x + h) − f (x))
= Vxx+h ( f ) − ( f (x + h) − f (x))
≥ Vxx+h ( f ) − | f (x + h) − f (x)| ≥ 0.

As an immediate consequence we obtain that every real-valued function f ∈ BV [a, b]

can be represented as the difference of two monotone increasing functions as

f (x) = Vax ( f ) − (Vax ( f ) − f (x)) .

This fact allows us to define the Stieltjes integral

 b
g(x)d f (x), (1.6)
a

where f ∈ BV [a, b] and g is an arbitrary continuous function. The integral (1.6) is

defined as  b n

a
g(x)d f (x) = lim
Δ →0
∑ g(ξ j )( f (x j ) − f (x j−1 )),
j=1

where a = x0 < x1 < · · · < xn = b, ξ j ∈ [x j−1 , x j ], and Δ = max1≤ j≤n (x j − x j−1 ).

8 Part I: Fourier Series and the Discrete Fourier Transform

Let us introduce the modulus of continuity of f by

ωh ( f ) := sup | f (x + h) − f (x)|, h > 0. (1.7)

{x∈[a,b]:x+h∈[a,b]}

Definition 1.10. A bounded function f is said to belong to Hölder space

Cα [a, b], 0 < α ≤ 1, if
ωh ( f ) ≤ Chα

with some constant C > 0. This inequality is called the Hölder condition with expo-
nent α .

Definition 1.11. A function f is said to belong to Sobolev space Wp1 (a, b),
1 ≤ p < ∞, if f ∈ L p (a, b) and there is g ∈ L p (a, b) such that
 x
f (x) = g(t)dt +C (1.8)
a

with some constant C.

Definition 1.12. A function f is said to belong to Sobolev space W∞1 (a, b) if there
is a bounded integrable function g such that
 x
f (x) = g(t)dt +C (1.9)
a

with some constant C.

Remark 1.13. Using Hölder’s inequality we may conclude that

Wp11 (a, b) ⊂ Wp12 (a, b)

for every 1 ≤ p2 < p1 ≤ ∞.

Lemma 1.14. Suppose that f ∈ Wp1 (a, b), 1 ≤ p ≤ ∞. Then f is of bounded vari-
ation. Moreover, if p = 1, then f is also continuous, and if 1 < p ≤ ∞, then
f ∈ C1−1/p [a, b].

Proof. Let first p = 1. Then there is an integrable function g such that (1.8) holds
with some constant C. Hence for fixed x ∈ [a, b] with x + h ∈ [a, b] we have
 x+h
f (x + h) − f (x) = g(t)dt.
x

It follows that
 x+h 
 

| f (x + h) − f (x)| =  g(t)dt  → 0, h → 0,
x
1 Introduction 9

since g is integrable. This proves the continuity of f . At the same time, for every
{x0 , x1 , . . . , xn } such that a = x0 < x1 < · · · < xn = b, we have
 x   xj  b
n  j
n  n

∑ | f (x j ) − f (x j−1 )| = ∑  x g(t)dt  ≤ ∑ |g(t)|dt = |g(t)|dt.
j=1 j=1 j−1 j=1 x j−1 a

Hence, Definition 1.8 is satisfied with constant c0 = ab |g(t)|dt, and f is of bounded
variation.
If 1 < p < ∞, then using Hölder’s inequality for integrals we obtain for h > 0 that

 x+h  x+h
1/p  x+h
1/p
| f (x + h) − f (x)| ≤ |g(t)|dt ≤ dt |g(t)| dt
p
x x x
 b
1/p
≤h 1−1/p
|g(t)| dt
p
,
a

where 1/p + 1/p = 1. If p = ∞, then | f (x + h) − f (x)| ≤ h sup |g|. By Definition

1.10, this means that f ∈ C1−1/p [a, b]. The lemma is proved. 

Remark 1.15. Since every f ∈ Wp1 (a, b), 1 ≤ p ≤ ∞, is continuous, it follows that
the constant C in (1.8)–(1.9) is equal to f (a).

Definition 1.16. Two functions u and v are said to be orthogonal on [a, b] if the
product uv is integrable and
 b
u(x)v(x)dx = 0,
a

where overline indicates the complex conjugation. A set of functions is said to be

mutually orthogonal if each distinct pair in the set is orthogonal on [a, b].

Lemma 1.17. The functions

mπ x mπ x
1, sin , cos , m = 1, 2, . . . ,
L L

form a mutually orthogonal set on the interval [−L, L] as well as on every interval
[a, a + 2L]. In fact,

 L

mπ x nπ x 0, m = n,
cos cos dx = (1.10)
−L L L L, m = n,
 L
mπ x nπ x
cos sin dx = 0, (1.11)
−L L L
10 Part I: Fourier Series and the Discrete Fourier Transform

 L

mπ x nπ x 0, m = n,
sin sin dx = (1.12)
−L L L L, m = n,

and  L  L
mπ x mπ x
sin dx = cos dx = 0. (1.13)
−L L −L L

Proof. By Lemma 1.3, it is enough to prove the equalities (1.10), (1.11), (1.12), and
(1.13) only for integrals over [−L, L]. Let us derive, for example, (1.12). Using the
equality
1
sin α sin β = (cos(α − β ) − cos(α + β )) ,
2
we have for m = n that
 L  L 
mπ x nπ x 1 (m − n)π x 1 L (m + n)π x
sin sin dx = cos dx − cos dx
−L L L 2 −L L 2 −L L
 
(m−n)π x L
 
(m+n)π x L
1 sin L  1 sin L 
=  −  = 0.
2 (m−n)π  2 (m+n)π 
L −L L −L

If m = n, we have
 L  L  L
mπ x nπ x 1 1 2mπ x
sin sin dx = 1dx − cos dx = L.
−L L L 2 −L 2 −L L

The other identities can be proved in a similar manner and are left to the reader. The
lemma is proved. 
nπ x
Remark 1.18. This lemma holds also for the functions ei L , n = 0, ±1, ±2, . . ., in
the form
 L

nπ x m π x 0, n = m,
ei L e−i L dx =
−L 2L, n = m.
Chapter 2
Formulation of Fourier Series

Let us consider a series of the form

∞ 
a0 mπ x mπ x 
+ ∑ am cos + bm sin . (2.1)
2 m=1 L L

This series consists of 2L-periodic functions. Thus, if the series (2.1) converges for
all x, then the function to which it converges will also be 2L-periodic. Let us denote
this limiting function by f (x), i.e.,
∞ 
a0 mπ x mπ x 
f (x) := + ∑ am cos + bm sin . (2.2)
2 m=1 L L

To determine am and bm we proceed as follows: assuming that the integration can be

legitimately carried out term by term (it will be, for example, if ∑∞
m=1 (|am | + |bm |) <
∞), we obtain
 L  L ∞  L
nπ x a0 nπ x mπ x nπ x
f (x) cos dx = cos dx + ∑ am cos cos dx
−L L 2 −L L m=1 −L L L
∞  L
mπ x nπ x
+ ∑ bm −L
sin
L
cos
L
dx
m=1

for each fixed n = 1, 2, . . .. It follows from the orthogonality relations (1.10), (1.11),
and (1.13) that the only nonzero term on the right-hand side is the one for which
m = n in the first summation. Hence
 L
1 nπ x
an = f (x) cos dx, n = 1, 2, . . . . (2.3)
L −L L

c Springer International Publishing AG 2017 11

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 2
12 Part I: Fourier Series and the Discrete Fourier Transform

A similar expression for bn is obtained by multiplying (2.2) by sin nπL x and integrat-
ing termwise from −L to L. The result is
 L
1 nπ x
bn = f (x) sin dx, n = 1, 2, . . . . (2.4)
L −L L

Using (1.13) we can easily obtain that

 L
1
a0 = f (x)dx. (2.5)
L −L

Definition 2.1. Let f be integrable (not necessarily periodic) on the interval [−L, L].
The Fourier series of f is the trigonometric series (2.1), where the coefficients
a0 , am and bm are given by (2.5), (2.3), and (2.4), respectively. In that case, we write
∞ 
a0 mπ x mπ x 
f (x) ∼ + ∑ am cos + bm sin . (2.6)
2 m=1 L L

Remark 2.2. This definition does not imply that the series (2.6) converges to f or
that f is periodic.

Definition 2.1 and Lemma 1.5 imply that if f is even on [−L, L], then the Fourier
series of f has the form
∞
a0 mπ x
f (x) ∼ + ∑ am cos , (2.7)
2 m=1 L

and if f is odd, then

∞
mπ x
f (x) ∼ ∑ bm sin L
. (2.8)
m=1

The series (2.7) and (2.8) are called the Fourier cosine series and Fourier sine series,
respectively.
If L = π , then the Fourier series (2.6) ((2.7) and (2.8)) transforms to
∞
a0
f (x) ∼ + ∑ (am cos mx + bm sin mx) , (2.9)
2 m=1

where the coefficients a0 , am , and bm are given by (2.5), (2.3), and (2.4) with L = π .
There are different approaches if the function f is defined on an asymmetric
interval [0, L] with arbitrary L > 0.
2 Formulation of Fourier Series 13

(1) Even extension. Define a function g(x) on the interval [−L, L] as


f (x), 0 ≤ x ≤ L,
g(x) =
f (−x), −L ≤ x < 0.

Then g(x) is even and its Fourier (cosine) series (2.7) represents f on [0, L].
(2) Odd extension. Define a function h(x) on the interval [−L, L] as

f (x), 0 ≤ x ≤ L,
h(x) =
− f (−x), −L ≤ x < 0.

Then h(x) is odd, and its Fourier (sine) series (2.8) represents f on [0, L].
(3) Define a function f(t) on the interval [−π , π ] as
 
tL L
f(t) = f + .
2π 2

If f (0) = f (L), then we may extend f to be periodic with period L. Then

    
1 π  1 π tL L 1 2π L
a0 ( f) = f (t)dt = f + dt = f (x) dx
π −π π −π 2π 2 π L 0

2 L
= f (x) dx := a0 ( f ),
L 0
 L
2 2mπ x
am ( f) = (−1)m f (x) cos dx = (−1)m am ( f ),
L 0 L

and  L
2 2mπ x
bm ( f) = (−1)m f (x) sin dx = (−1)m bm ( f ).
L 0 L

Hence,
∞
a0
f(t) ∼ + ∑ (am cos mt + bm sin mt) ,
2 m=1

and at the same time,

∞  
a0 2mπ x 2mπ x
f (x) ∼ + ∑ (−1)m am cos + bm sin ,
2 m=1 L L

tL L
where a0 , am , and bm are the same and x = + .
2π 2
These three alternatives allow us to consider (for simplicity) only the case of a sym-
metric interval [−π , π ] such that the Fourier series will be of the form (2.9) i.e.
14 Part I: Fourier Series and the Discrete Fourier Transform

∞
a0
f (x) ∼ + ∑ (am cos mx + bm sin mx) .
2 m=1

Using Euler’s formula, we will rewrite this series in the complex form
∞
f (x) ∼ ∑ cn einx , (2.10)
n=−∞

where the coefficients cn = cn ( f ) are equal to

⎧
⎪
⎪
an bn
+ , n = 1, 2, . . . ,
⎪
⎪
⎪
⎪ 2 2i an = cn + c−n , n = 1, 2, . . . ,
⎨a
0
cn = , n = 0, or a0 = 2c0 , (2.11)
⎪
⎪ 2
⎪
⎪ bn = i(cn − c−n ), n = 1, 2, . . . .
⎪
⎪
⎩ a−n − b−n , n = −1, −2, . . . .
2 2i
The formulas (2.3), (2.4), (2.5), and (2.11) imply that
 π
1
cn ( f ) = f (x)e−inx dx (2.12)
2π −π

for n = 0, ±1, ±2, . . .. We call cn ( f ) the nth Fourier coefficient of f . It can be

checked that
cn ( f ) = c−n ( f ). (2.13)

Exercise 2.1. Prove formulas (2.10), (2.11), (2.12), and (2.13).

Exercise 2.2. Find the Fourier series of

⎧
⎪
⎨−1, −π ≤ x < 0,
(1) sgn(x) = 0, x = 0,
⎪
⎩
1, 0 < x ≤ π.
(2) |x|, −1 ≤ x ≤ 1.
(3) x, −1 ≤x ≤ 1.
0, −L ≤ x ≤ 0,
(4) f (x) =
L, 0 < x ≤ L.
(5) f (x) = sin x, |x| ≤ 2.

Exercise 2.3. Prove, using Part (2) of Exercise 2.2, that

∞ ∞
π2 1 π2 1
=∑ and = ∑ 2.
8 k=1 (2k − 1)2 6 k=1 k
2 Formulation of Fourier Series 15

Exercise 2.4. Suppose that


1 − x, 0 ≤ x ≤ 1,
f (x) =
0, 1 < x ≤ 2.

Find the Fourier cosine and sine series of f (x).

Exercise 2.5. Find the Fourier series of f (x) = cos(x/2), |x| ≤ π . Using this series,
show that
∞
(−1)k+1
(1) π = 2 + ∑ ;
k=1 k − 1/4
2
∞
π (−1)k+1
(2) =∑ ;
4 k=1 2k − 1
∞
1 1
(3) =∑ 2 .
2 k=1 4k − 1

Exercise 2.6. Show that if N is odd, then sinN x can be written as a finite sum of the
form
N
∑ ak sin kx,
k=1

which means that this finite sum is the Fourier series of sinN x and the coefficients
ak (which are real) are the Fourier coefficients of sinN x.

Exercise 2.7. Show that if N is odd, then cosN x can be written as a finite sum of
the form
N
∑ ak cos kx.
k=1
Chapter 3
Fourier Coefficients and Their Properties

Definition 3.1. A trigonometric series

∞
∑ cn einx
n=−∞

is said to
(1) converge pointwise if for each x ∈ [−π , π ] the limit

lim
N→∞
∑ cn einx
|n|≤N

exists,
(2) converge uniformly in x ∈ [−π , π ] if the limit

lim
N→∞
∑ cn einx
|n|≤N

exists uniformly,
(3) converge absolutely if the limit

lim
N→∞
∑ |cn |
|n|≤N

exists, or equivalently, if
∞
∑ |cn | < ∞.
n=−∞

These three different types of convergence appear frequently in the sequel, and they
are presented above from the weakest to the strongest. In other words, absolute

c Springer International Publishing AG 2017 17

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 3
18 Part I: Fourier Series and the Discrete Fourier Transform

convergence implies uniform convergence, which in turn implies pointwise conver-

gence.
If f is integrable on the interval [−π , π ], then the Fourier coefficients cn ( f ) are
uniformly bounded with respect to n = 0, ±1, ±2, . . ., i.e.,
  
1  π −inx 
 1 π
|cn ( f )| = f (x)e dx ≤ | f (x)|dx, (3.1)
2π  −π 2π −π

where the upper bound does not depend on n. Let us assume that a sequence
{cn }∞
n=−∞ is such that
∞
∑ |cn | < ∞.
n=−∞

Then the series

∞
∑ cn einx
n=−∞

converges uniformly in x ∈ [−π , π ] and defines a continuous and periodic function

∞
f (x) := ∑ cn einx , (3.2)
n=−∞

whose Fourier coefficients are {cn }∞ ∞

n=−∞ = {cn ( f )}n=−∞ . More generally, suppose
that
∞
∑ |n|k |cn | < ∞
n=−∞

for some integer k > 0. Then the series (3.2) defines a function that is k times differ-
entiable, with
∞
f (k) (x) = ∑ (in)k cn einx (3.3)
n=−∞

a continuous function. This follows from the fact that the series (3.3) converges
uniformly with respect to x ∈ [−π , π ].
Let us consider a useful example in which Fourier coefficients are applied. If
0 ≤ r < 1, then the geometric series gives
∞
1
= ∑
1 − reix n=0
rn einx , (3.4)

and this series converges absolutely. Using the definition of the Fourier coefficients,
we obtain 
1 π e−inx
r =
n
dx, n = 0, 1, 2, . . . ,
2π −π 1 − reix
3 Fourier Coefficients and Their Properties 19

and  π
1 einx
0= dx, n = 1, 2, . . . .
2π −π 1 − reix

From the representation (3.4), we may conclude also that

  ∞
1 − r cos x 1 1 1 ∞ |n| inx
1 − 2r cos x + r2
= Re
1 − reix
= ∑ r n
cos nx = + ∑ r e
2 2 n=−∞
(3.5)
n=0

and
  ∞
r sin x 1 i ∞
= Im = ∑ rn sin nx = − ∑ r|n| sgn(n)einx . (3.6)
1 − 2r cos x + r 2 1 − reix
n=1 2 n=−∞

Exercise 3.1. Verify formulas (3.5) and (3.6).

Formulas (3.5) and (3.6) can be rewritten as

∞
1 − r2
∑ r|n| einx =
1 − 2r cos x + r2
=: Pr (x) (3.7)
n=−∞

and
∞
2r sin x
−i ∑ sgn(n)r|n| einx =
1 − 2r cos x + r2
=: Qr (x). (3.8)
n=−∞

Definition 3.2. The function Pr (x) is called the Poisson kernel, while Qr (x) is
called the conjugate Poisson kernel.

Since the series (3.7) and (3.8) converge absolutely, we have

 π  π
1 1
r|n| = Pr (x)e−inx dx and − i sgn(n)r|n| = Qr (x)e−inx dx,
2π −π 2π −π

where n = 0, ±1, ±2, . . .. In particular,

 π  π
1 1
Pr (x)dx = 1 and Qr (x)dx = 0.
2π −π 2π −π

Exercise 3.2. Prove that both Pr (x) and Qr (x) are solutions of the Laplace equation

ux1 x1 + ux2 x2 = 0

in the disk x12 + x22 < 1, where x1 + ix2 = reix with 0 ≤ r < 1 and x ∈ [−π , π ].
20 Part I: Fourier Series and the Discrete Fourier Transform

Given a sequence an , n = 0, ±1, ±2, . . ., we define

Δn a = an − an−1 .

Then for any two sequences an and bn and integers M < N, the formula

N N
∑ ak Δk b = aN bN − aM bM − ∑ bk−1 Δk a (3.9)
k=M+1 k=M+1

holds. Formula (3.9) is called summation by parts.

Exercise 3.3. Prove (3.9).
Summation by parts allows us to investigate the convergence of a special type of
trigonometric series.
Theorem 3.3. Suppose that cn > 0, n = 0, 1, 2, . . ., cn ≥ cn+1 , and limn→∞ cn = 0.
Then the trigonometric series
∞
∑ cn einx (3.10)
n=0

converges for every x ∈ [−π , π ] \ {0}.

Proof. Let bn = ∑nk=0 eikx . Since

1 − eix(n+1)
bn = , x = 0,
1 − eix

it follows that
2 1
|bn | ≤ =
|1 − e | | sin 2x |
ix

for x ∈ [−π , π ] \ {0}. Applying (3.9) shows that for M < N we have
 
N N k k−1 N
∑ ikx
ck e = ∑ ck ∑e − ∑e
ilx ilx
= ∑ ck Δk b
k=M+1 k=M+1 l=0 l=0 k=M+1
N
= cN bN − cM bM − ∑ bk−1 Δk c.
k=M+1

Thus
3 Fourier Coefficients and Their Properties 21
 
 N  N
 ikx 
 ∑ ck e  ≤ cN |bN | + cM |bM | + ∑ |bk−1 ||Δk c|
k=M+1  k=M+1
 
N
1
≤ cN + cM + ∑ |ck − ck−1 |
| sin 2x | k=M+1
1 2cM
= x (cN + cM + cM − cN ) = →0
| sin 2 | | sin 2x |

as N > M → ∞. This proves the theorem. 

Corollary 3.4. Under the same assumptions as in Theorem 3.3, the trigonometric
series (3.10) converges uniformly for all π ≥ |x| ≥ δ > 0.
 
Proof. If π ≥ |x| ≥ δ > 0, then sin 2x  ≥ π2 |x| δ
2 ≥ π. 
Theorem 3.3 implies that, for example, the series
∞ ∞
cos nx sin nx
∑ log(2 + n) and ∑ log(2 + n)
n=1 n=1

converge for all x ∈ [−π , π ] \ {0}.

Modulus of continuity and tail sum
For the trigonometric series (2.10) with
∞
∑ |cn | < ∞
n=−∞

we introduce the tail sum by

En := ∑ |ck |, n = 0, 1, 2, . . . . (3.11)
|k|>n

There is a good connection between the modulus of continuity (1.7) and (3.11).
Indeed, if f (x) denotes the series (2.10) and h > 0, we have
∞
| f (x + h) − f (x)| ≤ ∑ |cn ||einh − 1| = ∑ |cn ||einh − 1| + ∑ |cn ||einh − 1|
n=−∞ |n|h≤1 |n|h>1

≤h ∑ |n||cn | + 2 ∑ |cn | =: I1 + I2 ,
|n|≤[1/h] |n|>[1/h]

where [x] denotes the entire part of x. If we denote [1/h] by Nh , then I2 = 2ENh and

 Nh
I1 = h ∑ |n| E|n|−1 − E|n| = −h ∑ l (El − El−1 ) .
|n|≤Nh l=1
22 Part I: Fourier Series and the Discrete Fourier Transform

Using (3.9) in the latter sum, we obtain

 
Nh Nh
I1 = −h Nh ENh − 0 · E0 − ∑ En−1 (n − (n − 1)) = −hNh ENh + h ∑ En−1 .
n=1 n=1

Since hNh = h[1/h] ≤ h · 1h = 1, these formulas for I1 and I2 imply that

Nh
1 Nh −1
ωh ( f ) ≤ 2ENh − hNh ENh + h ∑ En−1 ≤ 2ENh + ∑ En . (3.12)
n=1 Nh n=0

Since En → 0 as n → ∞, the inequality (3.12) implies that ωh ( f ) → 0 as h → 0.

Moreover, if En = O(n−α ), n = 0 for some 0 < α ≤ 1, then

O(hα ), 0 < α < 1,

ωh ( f ) = (3.13)
O(h log 1h ), α = 1.

Here and throughout, the notation A = O(B) on a set X means that |A| ≤ C|B| on X
with some constant C > 0. Similarly, A = o(B) means that A/B → 0.
Exercise 3.4. Prove the second relation in (3.13).
We summarize (3.13) as follows: if the tail of the trigonometric series (2.10) behaves
as O(n−α ) for some 0 < α < 1, then the function f to which it converges belongs
to the Hölder space Cα [−π , π ].
Chapter 4
Convolution and Parseval’s Equality

Let the trigonometric series (2.10) be such that

∞
∑ |cn | < ∞.
n=−∞

Then the function f to which it converges is continuous and periodic. If g(x) is any
continuous function, then the product f g is also continuous and hence integrable on
[−π , π ] and
 π ∞  π ∞
1 1
2π −π
f (x)g(x)dx =
2π ∑ cn ( f )
−π
g(x)einx dx = ∑ cn ( f )c−n (g), (4.1)
n=−∞ n=−∞

where integration of the series term by term is justified by the uniform convergence
of the Fourier series. Putting g = f in (4.1) yields
 π ∞ ∞
1
2π −π
| f (x)|2 dx = ∑ cn ( f )c−n ( f ) = ∑ |cn ( f )|2 (4.2)
n=−∞ n=−∞

by (2.13).
Definition 4.1. Equality (4.2) is called the Parseval’s equality for the trigonometric
Fourier series.
The formula (4.1) can be generalized as follows.
Exercise 4.1. Let a periodic function f be defined by the absolutely convergent
Fourier series (2.10) and let g be integrable and periodic. Prove that
 π ∞
1
2π −π
f (x)g(y − x)dx = ∑ cn ( f )cn (g)einy
n=−∞

c Springer International Publishing AG 2017 23

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 4
24 Part I: Fourier Series and the Discrete Fourier Transform

and that this series converges absolutely.

A generalization of Exercise 4.1 is given by the following theorem.
Theorem 4.2. If f1 and f2 are two periodic L1 functions, then

cn ( f1 ∗ f2 ) = cn ( f1 )cn ( f2 ),

where f1 ∗ f2 denotes the convolution

 π
1
( f1 ∗ f2 )(x) = f1 (y) f2 (x − y)dy (4.3)
2π −π

and where the integral converges for almost every x.

Proof. We note first that by Fubini’s theorem the convolution (4.3) is well defined
as an L1 function. Indeed,
 π  π   π  π −y 
| f1 (y)| · | f2 (x − y)|dy dx = | f1 (y)| | f2 (z)|dz dy
−π −π −π −π −y
 π  π 
= | f1 (y)| | f2 (z)|dz dy
−π −π

by Lemma 1.3. The Fourier coefficients of the convolution (4.3) are equal to

1 π
cn ( f1 ∗ f2 ) = ( f1 ∗ f2 )(x)e−inx dx
2π −π
 π  π 
1
= f 1 (y) f 2 (x − y)dy e−inx dx
(2π )2 −π −π
 π  π 
1 −inx
= f 1 (y) f 2 (x − y)e dx dy
(2π )2 −π −π
 π  π −y 
1 −in(y+z)
= f 1 (y) f 2 (z)e dz dy
(2π )2 −π −π −y
 π  π 
1 −iny −inz
= f 1 (y)e f 2 (z)e dz dy = cn ( f1 )cn ( f2 )
(2π )2 −π −π

by Lemma 1.3. Thus, the theorem is proved. 

Exercise 4.2. Prove that if f1 and f2 are integrable and periodic, then their convo-
lution is symmetric and periodic.

Exercise 4.3. Let f be a periodic L1 function. Prove that

∞
( f ∗ Pr )(x) = (Pr ∗ f )(x) = ∑ r|n| cn ( f )einx
n=−∞
4 Convolution and Parseval’s Equality 25

and that Pr ∗ f satisfies the Laplace equation, i.e., (Pr ∗ f )x1 x1 + (Pr ∗ f )x2 x2 = 0,
where x12 + x22 = r2 < 1 with x1 + ix2 = reix .

Remark 4.3. We are going to prove in Chapter 10 that for every periodic continuous
function f , the limit limr→1− ( f ∗ Pr )(x) = f (x) exists uniformly in x.
Chapter 5
Fejér Means of Fourier Series. Uniqueness
of the Fourier Series.

Let us denote the partial sum of the Fourier series of f ∈ L1 (−π , π ) (not necessarily
periodic) by
SN ( f ) := ∑ cn ( f )einx
|n|≤N

for each N = 0, 1, 2, . . .. The Fejér means are defined by

S0 ( f ) + · · · + SN ( f )
σN ( f ) := .
N +1

Writing this out in detail, we see that

N N
(N + 1)σN ( f ) = ∑ ∑ ck ( f )eikx = ∑ ∑ ck ( f )eikx
n=0 |k|≤n |k|≤N n=|k|

= ∑ (N + 1 − |k|)ck ( f )eikx ,
|k|≤N

which gives the useful representation

 
|k|
σN ( f ) = ∑ 1 − ck ( f )eikx . (5.1)
|k|≤N
N +1

The Fejér kernel is  

|k|
KN (x) := ∑ 1 −
N + 1
eikx . (5.2)
|k|≤N

c Springer International Publishing AG 2017 27

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 5
28 Part I: Fourier Series and the Discrete Fourier Transform

The sum (5.2) can be calculated precisely as

 2
1 sin N+1
2 x
KN (x) = . (5.3)
N +1 sin 2x

Exercise 5.1. Prove the identity (5.3).

Exercise 5.2. Prove that  π

1
KN (x)dx = 1. (5.4)
2π −π

We can rewrite σN ( f ) from (5.1) also as

∞  
|k|
σN ( f ) = ∑ 1[−N,N] (k) 1 − ck ( f )eikx ,
k=−∞ N + 1

where 
1, |k| ≤ N,
1[−N,N] (k) =
0, |k| > N.

Let us assume now that f is periodic. Then Exercise 4.1 and Theorem 4.2 lead to
 π ∞  
1 |k|
f (y)KN (x − y)dy = ∑ 1[−N,N] (k) 1 − ck ( f )eikx . (5.5)
2π −π k=−∞ N +1

Exercise 5.3. Prove (5.5).

Hence, the Fejér means can be represented as

σN ( f )(x) = ( f ∗ KN )(x) = (KN ∗ f )(x). (5.6)

The properties (5.3), (5.4), and (5.6) allow us to prove the following result.
Theorem 5.1. Let f ∈ L p (−π , π ) be periodic with 1 ≤ p < ∞. Then
 π
1/p
lim |σN ( f )(x) − f (x)| p dx = 0. (5.7)
N→∞ −π

If, in addition, f has right and left limits f (x0 ± 0) at a point x0 ∈ [−π , π ], then

1
lim σN ( f )(x0 ) = ( f (x0 + 0) + f (x0 − 0)) . (5.8)
N→∞ 2
5 Fejér Means of Fourier Series. Uniqueness of the Fourier Series. 29

Proof. Let us first prove (5.7). Indeed, (5.4) and (5.6) give
 π
1/p  π
1/p
|σN ( f )(x) − f (x)| p dx = |( f ∗ KN )(x) − f (x)| p dx
−π −π
 
π  1  π   p 1/p
1 π 
=  KN (y) f (x − y)dy − KN (y) f (x)dy dx

−π 2π −π 2π −π
 π  π  p 1/p
1  
=  KN (y)( f (x − y) − f (x))dy dx
2π 
−π −π
 π    p 1/p
1  
≤  K (y)( f (x − y) − f (x))dy  dx
2π 
−π |y|<δ
N 
 π    p 1/p
1  
+  K (y)( f (x − y) − f (x))dy  dx =: I1 + I2 .
2π 
−π π ≥|y|>δ
N 

Using the generalized Minkowski’s inequality, we obtain that

  π
1/p
1 p
I1 ≤ KN (y) | f (x − y) − f (x)| dx dy
2π |y|<δ −π
 π
1/p  π
1
≤ sup | f (x − y) − f (x)| p dx KN (y)dy
|y|<δ −π 2π −π
 π
1/p
p
= sup | f (x − y) − f (x)| dx →0 (5.9)
|y|<δ −π

as δ → 0, since f ∈ L p (−π , π ). Quite similarly,

  π 1/p
1 p
I2 ≤ KN (y) | f (x − y) − f (x)| dx dy
2π π ≥|y|>δ −π
 π 1/p 
p 1
≤2 | f (x)| dx KN (y)dy, (5.10)
−π 2π π ≥|y|>δ

since f is periodic. The next step is to note that

 2  2
y 2 |y| δ
sin ≥
2
≥
2 π 2 π

for π ≥ |y| > δ . That is why (5.3) leads to

1 1 1 π2
KN (y) ≤ · 2y ≤ · 2
N + 1 sin 2 N +1 δ
30 Part I: Fourier Series and the Discrete Fourier Transform

and

1 π 2 2(π − δ ) π2 π2
KN (y)dy ≤ · < 2 < √ →0
2π π ≥|y|>δ 2πδ 2 N +1 δ (N + 1) N

as N → ∞ if we choose δ = N −1/4 . From (5.9) and (5.10) we may conclude (5.7).

In order to prove (5.8), we use (5.4) to consider the difference

1
σN ( f )(x0 ) − ( f (x0 + 0) + f (x0 − 0))
2


1 π 1
= KN (y) f (x0 − y) − ( f (x0 + 0) + f (x0 − 0)) dy
2π −π 2
 π
1
= KN (y)g(y)dy, (5.11)
2π −π

where
1
g(y) = f (x0 − y) − ( f (x0 + 0) + f (x0 − 0)) .
2

Since the Fejér kernel KN (y) is even, we can rewrite the right-hand side of (5.11) as
 π
1
KN (y)h(y)dy,
2π 0

where h(y) = g(y) + g(−y). It is clear that h(y) is an L1 function. But we have more,
namely,
lim h(y) = 0. (5.12)
y→0+

Our task now is to prove that

 π
1
lim KN (y)h(y)dy = 0.
N→∞ 2π 0

We will proceed as in the proof of (5.7), i.e., we split the integral as

 π  δ  π
1 1 1
KN (y)h(y)dy = KN (y)h(y)dy + KN (y)h(y)dy =: I1 + I2 .
2π 0 2π 0 2π δ

The first term can be estimated as

 π
1 1
|I1 | ≤ sup |h(y)| KN (y)dy = sup |h(y)| → 0
2π 0≤y≤δ 0 2 |y|≤δ

as δ → 0+ due to (5.12). For I2 we have

5 Fejér Means of Fourier Series. Uniqueness of the Fourier Series. 31
 π
1  π 2 1
|I2 | ≤ |h(y)|dy → 0
2π δ N + 1 0

as N → ∞ if we choose, for example, δ = N −1/4 . Thus, the theorem is completely

proved. 

Corollary 5.2. If f is periodic and continuous on the interval [−π , π ], then

lim σN ( f )(x) = f (x)

N→∞

uniformly in x ∈ [−π , π ].

Corollary 5.3. Every periodic L p function, 1 ≤ p < ∞, can be approximated in

the sense of L p space by the trigonometric polynomials ∑|k|≤N bk eikx (which are
infinitely differentiable, i.e., C∞ functions).

Theorem 5.4 (Uniqueness of Fourier series). If f ∈ L1 (−π , π ) is periodic and if

its Fourier coefficients are identically zero, then f = 0 almost everywhere.

Proof. Since  π
lim |σN ( f )(x) − f (x)| dx = 0
N→∞ −π

by (5.7), it follows that if all Fourier coefficients are zero, we have

 π
| f (x)| dx = 0,
−π

which means that f = 0 almost everywhere. This proves the theorem. 

Chapter 6
The Riemann–Lebesgue Lemma

Theorem 6.1 (Riemann–Lebesgue lemma). If f is periodic with period 2π and

belongs to L1 (−π , π ), then
 π
lim f (x + z)e−inz dz = 0 (6.1)
n→∞ −π

uniformly in x ∈ R. In particular, cn ( f ) → 0 as n → ∞.

Proof. Since f is periodic with period 2π , it follows that

 π  π +x  π
f (x + z)e−inz dz = f (y)e−in(y−x) dy = einx f (y)e−iny dy (6.2)
−π −π +x −π

by Lemma 1.3. Formula (6.2) shows that to prove (6.1) it is enough to show that the
Fourier coefficients cn ( f ) tend to zero as n → ∞. Indeed,
 π  π +π /n  π
2π cn ( f ) = f (y)e−iny dy = f (y)e−iny dy = f (t + π /n)e−int e−iπ dt
−π −π +π /n −π

by Lemma 1.3. Hence

 π
− 4π cn ( f ) = ( f (t + π /n) − f (t)) e−int dt. (6.3)
−π

If f is continuous on the interval [−π , π ], then

sup | f (t + π /n) − f (t)| → 0, n → ∞.

t∈[−π ,π ]

Hence cn ( f ) → 0 as n → ∞. If f is an arbitrary L1 function, we let ε > 0. Then we

can define a continuous function g (see Corollary 5.3) such that
c Springer International Publishing AG 2017 33
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 6
34 Part I: Fourier Series and the Discrete Fourier Transform
 π
| f (x) − g(x)| dx < ε .
−π

Write
cn ( f ) = cn (g) + cn ( f − g).

The first term tends to zero as n → ∞, since g is continuous, whereas the second
term is less than ε /(2π ). This implies that

ε
sup lim |cn ( f )| ≤ .
n→∞ 2π

Since ε is arbitrary, we have

lim |cn ( f )| = 0.
n→∞

This fact together with (6.2) gives (6.1). The theorem is thus proved. 
Corollary 6.2. Let f be as in Theorem 6.1. If a periodic function g is continuous
on [−π , π ], then  π
lim f (x + z)g(z)e−inz dz = 0
n→∞ −π

and
 π  π
lim f (x + z)g(z) sin(nz)dz = lim f (x + z)g(z) cos(nz)dz = 0
n→∞ −π n→∞ −π

uniformly in x ∈ [−π , π ].
Exercise 6.1. Prove this corollary.
Exercise 6.2. Show that if f satisfies the Hölder condition with exponent α ∈ (0, 1],
then cn ( f ) = O(|n|−α ) as n → ∞.
Exercise 6.3. Suppose that f satisfies the Hölder condition with exponent α > 1.
Prove that f ≡ constant.
Exercise 6.4. Let f (x) = |x|α , where −π ≤ x ≤ π and 0 < α < 1. Prove that cn ( f ) 
|n|−1−α as n → ∞.
Remark 6.3. The notation a  b means that there exist 0 < c1 < c2 such that

c1 |a| < |b| < c2 |a|.

Let us introduce for all 1 ≤ p < ∞ and periodic functions f ∈ L p (−π , π ) the L p -
modulus of continuity of f by
 π
1/p
p
ω p,δ ( f ) := sup | f (x + h) − f (x)| dx .
|h|≤δ −π
6 The Riemann–Lebesgue Lemma 35

The equality (6.3) leads to

 π
1
|cn ( f )| ≤ | f (x + π /n) − f (x)| dx
4π −π
 π
1/p
(2π )1−1/p 1
≤ | f (x + π /n) − f (x)| dx p
≤ (2π )−1/p ω p,π /n ( f ),
4π −π 2

where we have used Hölder’s inequality in the penultimate step.

Exercise 6.5. Suppose that ω p,δ ( f ) ≤ Cδ α for some C > 0 and α > 1. Prove that
f is constant almost everywhere.
Hint. First show that ω p,2δ ( f ) ≤ 2ω p,δ ( f ); then iterate this to obtain a contradiction.

Suppose that f ∈ L1 (−π , π ) but f is not necessarily periodic. We can consider the
Fourier series corresponding to f , i.e.,
∞
f (x) ∼ ∑ cn einx ,
n=−∞

where the cn are the Fourier coefficients cn ( f ). The series on the right-hand side
is considered formally in the sense that we know nothing about its convergence.
However, the limit
 π  π
lim
N→∞ −π
∑ cn ( f )einx dx =
−π
f (x)dx (6.4)
|n|≤N

exists. Indeed,
 π  π  π
∑
−π |n|≤N
cn ( f )einx dx = c0 ( f )
−π
dx + ∑ cn ( f )
−π
einx dx = 2π c0 ( f )
0<|n|≤N
 π
= f (x)dx.
−π

Remark 6.4 (Important properties of the Fourier series). The existence of the limit
(6.4) shows us that we can always integrate the Fourier series of an L1 function term
by term.
Chapter 7
The Fourier Series of a Square-Integrable
Function. The Riesz–Fischer Theorem.

The set of square-integrable functions L2 (−π , π ) is an inner product space (linear

Euclidean space) equipped with the inner product
 π
1
( f , g)L2 (−π ,π ) = f (x)g(x)dx.
2π −π

We can measure the degree of approximation by the (square of) mean square dis-
tance 
1 π
| f (x) − g(x)|2 dx = ( f − g, f − g)L2 (−π ,π ) .
2π −π

In particular, if g(x) = ∑|n|≤N bn einx is a trigonometric polynomial, then this distance

can be written as
 π  π  π
1 1 1
| f (x)|2 dx + |g(x)|2 dx − 2Re f (x)g(x)dx,
2π −π 2π −π 2π −π

or
 π
1
| f (x)|2 dx
2π −π
 π  π
1 1
+
2π ∑
−π |n|≤N
bn einx ∑ bk e−ikx dx −
2π
2Re
−π
f (x) ∑ bn e−inx dx
|k|≤N |n|≤N
  π
1 π 1
=
2π −π
| f (x)|2 dx +
2π ∑ |bn |2
−π
dx − 2Re ∑ bn cn ( f )
|n|≤N |n|≤N
 π
1
=
2π −π
| f (x)|2 dx + ∑ |bn |2 − 2Re ∑ bn cn ( f ) + ∑ |cn ( f )|2 − ∑ |cn ( f )|2 .
|n|≤N |n|≤N |n|≤N |n|≤N

c Springer International Publishing AG 2017 37

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 7
38 Part I: Fourier Series and the Discrete Fourier Transform

So
 π  π
1 1
2π −π
| f (x) − g(x)|2 dx =
2π −π
| f (x)|2 dx − ∑ |cn ( f )|2 + ∑ |bn − cn ( f )|2 .
|n|≤N |n|≤N

This equality has the following consequences:

(1) The minimum error is
 π
1
min | f (x) − g(x)|2 dx
g(x)=∑|n|≤N bn e inx 2π −π
 π
1
=
2π −π
| f (x)|2 dx − ∑ |cn ( f )|2 , (7.1)
|n|≤N

and it is attained when bn = cn ( f ).

(2) For N = 1, 2, . . ., it is true that
 π
1
∑ |cn ( f )|2 ≤
2π −π
| f (x)|2 dx,
|n|≤N

and in particular,
∞  π
1
∑ |cn ( f )|2 ≤
2π −π
| f (x)|2 dx. (7.2)
n=−∞

This inequality is called Bessel’s inequality.

It turns out that (7.2) holds with equality. This is Parseval’s equality for f ∈
L2 (−π , π ), which we state as the following theorem.

Theorem 7.1. For every periodic function f ∈ L2 (−π , π ) with period 2π , its
Fourier series converges in L2 (−π , π ), i.e.,
 2
 π  
1  
lim  f (x) − ∑ cn ( f )einx  dx = 0,
N→∞ 2π −π  |n|≤N


and Parseval’s equality

 π ∞
1
2π −π
| f (x)|2 dx = ∑ |cn ( f )|2 (7.3)
n=−∞

holds.
7 The Fourier Series of a Square-Integrable Function 39

Proof. By Bessel’s inequality (7.2), we have for every f ∈ L2 (−π , π ) that

 2
 π  
1  inx 
2π
 ∑
−π |n|≤N
cn ( f )e inx
− ∑ cn ( f )e  dx =
 ∑ |cn ( f )|2 → 0
|n|≤M M+1≤|n|≤N

as N > M → ∞. Due to the completeness of the trigonometric polynomials in

L2 (−π , π ) (see Corollary 5.3), we may now conclude that there exists F ∈ L2 (−π , π )
such that  2

1 π  
inx 
lim F(x) − ∑ cn ( f )e  dx = 0.
N→∞ 2π −π  
|n|≤N

It remains to show that F(x) = f (x) almost everywhere. To do this, we compute the
Fourier coefficients cn (F) by writing

 π  π
 
2π cn (F) =
−π
F(x)e −inx
dx =
−π
F(x) − ∑ ck ( f )e
ikx
e−inx dx
|k|≤N
 π
+ ∑ ck ( f )
−π
ei(k−n)x dx.
|k|≤N

If N > |n|, then the last sum is equal to 2π cn ( f ). Thus, by Hölder’s inequality,
 
 π  
 
2π |cn (F) − cn ( f )| ≤ F(x) − ∑ ck ( f )eikx  dx
−π  |k|≤N

⎛  2 ⎞1/2
√  π  
 
≤ 2π ⎝ F(x) − ∑ ck ( f )eikx  dx⎠ → 0
−π  |k|≤N


as N → ∞, i.e., cn (F) = cn ( f ) for all n = 0, ±1, ±2, . . .. Theorem 5.4 (uniqueness

of Fourier series) implies now that F = f almost everywhere. Parseval’s equality
follows from (7.1) if we let N → ∞. 

Corollary 7.2 (Riesz–Fischer theorem). Suppose {bn }∞ n=−∞ is a sequence of

complex numbers with ∑∞ n=−∞ |bn |2 < ∞. Then there is a unique periodic function

f ∈ L2 (−π , π ) such that bn = cn ( f ).

Proof The proof is identical to that of Theorem 7.1. 

Theorem 7.3. Suppose that f ∈ L2 (−π , π ) is periodic with period 2π and that its
Fourier coefficients satisfy
∞
∑ |n|2 |cn ( f )|2 < ∞. (7.4)
n=−∞
40 Part I: Fourier Series and the Discrete Fourier Transform

Then f ∈ W21 (−π , π ) with the Fourier series for f  (x) given by
∞
f  (x) ∼ ∑ incn ( f )einx .
n=−∞

Proof. Since (7.4) holds, it follows that by the Riesz–Fischer theorem there is a
unique function g(x) ∈ L2 (−π , π ) such that
∞
g(x) ∼ ∑ incn ( f )einx .
n=−∞

Integrating term by term, we obtain

 π
g(x)dx = 0.
−π
 x
Let F(x) := g(t)dt. Then for n = 0, we have
−π
 π
 x  
 π 
1 1 π
cn (F) = g(t)dt e−inx dx = g(t) e−inx dx dt
2π
−π −π 2π −π t
 π
−inπ −int
  π
1 e e 1 1 1
= g(t) + dt = g(t)e−int dt = cn (g).
2π −π −in in in 2π −π in

On the other hand, cn (g) = incn ( f ). Thus, by the uniqueness of Fourier series, we
obtain that F(x)− f (x) = constant almost everywhere, or f  (x) = g(x) almost every-
where. This means that
 x
f (x) = g(t)dt + constant,
−π

where g ∈ L2 (−π , π ). Therefore, f ∈ W21 (−π , π ) and

∞
f  (x) = g(x) ∼ ∑ incn ( f )einx .
n=−∞

This completes the proof. 

Corollary 7.4. Under the conditions of Theorem 7.3, it is true that

∞
∑ |cn ( f )| < ∞.
n=−∞
7 The Fourier Series of a Square-Integrable Function 41

Proof. Due to (7.4) we have

∞
∑ |cn ( f )| = |c0 ( f )| + ∑ |cn ( f )|
n=−∞ n=0
1 1 1
2 n∑
≤ |c0 ( f )| + |n|2 |cn ( f )|2 + ∑ 2 < ∞,
=0 2 n=0 |n|

where we have used the basic inequality 2ab ≤ a2 + b2 for real numbers a and b. 

Using Parseval’s equality (7.3), we can obtain for every periodic function f ∈
L2 (−π , π ) and N = 1, 2, . . . that
 2
 π  
1  
2π
 ∑
−π |n|≤N
cn ( f )e
inx
− f (x) dx = ∑ |cn ( f )|2 .

(7.5)
|n|>N

Exercise 7.1. Prove (7.5).

Using Parseval’s equality again, we have

 π ∞
1
2π −π
| f (x + h) − f (x)|2 dx = ∑ |cn ( f (x + h) − f (x))|2
n=−∞
∞ (7.6)
= ∑ |eihn − 1|2 |cn ( f )|2 .
n=−∞

Theorem 7.5. Suppose f ∈ L2 (−π , π ) is periodic with period 2π . Then

∑ |cn ( f )|2 = O(N −2α ), N = 1, 2, . . . , (7.7)

|n|>N

with 0 < α < 1 if and only if

∞
∑ |einh − 1|2 |cn ( f )|2 = O(|h|2α ) (7.8)
n=−∞

for |h| sufficiently small.

Proof. From (7.6) we have for every integer M > 0 that

 π
1
2π −π
| f (x + h) − f (x)|2 dx ≤ ∑ n2 h2 |cn ( f )|2 + 4 ∑ |cn ( f )|2 (7.9)
|n|≤M |n|>M

if M|h| ≤ 1. If (7.7) holds, then the second sum is O(M −2α ). To estimate the first
sum we use summation by parts. Writing
42 Part I: Fourier Series and the Discrete Fourier Transform

In := ∑ |ck ( f )|2 ,
|k|≤n

we have
M
∑ n2 |cn ( f )|2 = M 2 IM − 0 · I0 − ∑ In−1 (n2 − (n − 1)2 )
1≤|n|≤M n=1
M
= M 2 IM − ∑ (2n − 1)In−1 − I0 .
n=2

By hypothesis,
I∞ − In = O(n−2α ), n = 1, 2, . . . .

Thus,

M 
M 2 IM − ∑ (2n − 1) I∞ + O((n − 1)−2α )
n=2
 M M
= M 2 I∞ + O(M −2α ) − I∞ ∑ (2n − 1) − ∑ (2n − 1)O((n − 1)−2α )
n=2 n=2
 
M M
= O(M 2−2α ) + I∞ M 2 − ∑ (2n − 1) − ∑ (2n − 1)O((n − 1)−2α )
n=2 n=2
  
=1
= O(M 2−2α ) + O(M 2−2α ) = O(M 2−2α ).

Exercise 7.2. Prove that

n=1 (2n − 1) = M ;
(1) ∑M 2

(2) ∑M
n=2 O((2n − 1)(n − 1)−2α ) = O(M 2−2α ) for 0 < α < 1.

Combining these two estimates, we may conclude from (7.9) that there exists C > 0
such that 
1 π 
| f (x + h) − f (x)|2 dx ≤ C h2 M 2−2α + M −2α .
2π −π

Since 0 < α < 1, choosing M = [1/|h|] we obtain

 π 
1
| f (x + h) − f (x)|2 dx ≤ C h2 [1/|h|]2−2α + [1/|h|]−2α
2π −π
 
≤ C h2 (1/|h|)2−2α + (1/|h| − {1/|h|})−2α
 
≤ C |h|2α + (1/|h| − 1)−2α

= C |h|2α + |h|2α /(1 − |h|)2α ≤ C|h|2α
7 The Fourier Series of a Square-Integrable Function 43

if |h| ≤ 1/2. Here {1/|h|} denotes the fractional part of 1/|h|, i.e., {1/|h|} = 1/|h| −
[1/|h|] ∈ [0, 1).
Conversely, if (7.8) holds, then
∞
∑ (1 − cos(nh))|cn ( f )|2 ≤ Ch2α
n=−∞

with some C > 0. Integrating this inequality with respect to h over the interval [0, l],
l > 0, we have
∞  l  l
∑ |cn ( f )|2
0
(1 − cos(nh))dh ≤ C
0
h2α dh,
n=−∞

or
∞

sin(nl)
∑ |cn ( f )| 2
l−
n
≤ Cl 2α +1 ,
n=−∞

or
∞

sin(nl)
∑ |cn ( f )| 2
1−
nl
≤ Cl 2α .
n=−∞

It follows that


sin(nl) 1
Cl 2α
≥ ∑ |cn ( f )| 2
1−
nl
≥ ∑ |cn ( f )|2 .
2 |n|l≥2
(7.10)
|n|l≥2

Taking l = 2/N for the integer N > 0 in (7.10) yields

∑ |cn ( f )|2 ≤ CN −2α .

|n|≥N

This completes the proof. 

Remark 7.6. If α = 1, then (7.8) implies (7.7) but not conversely.

Exercise 7.3. Suppose that a periodic function f ∈ L2 (−π , π ) satisfies the condi-
tion  π
| f (x + h) − f (x)|2 dx ≤ Ch2
−π

with some C > 0. Prove that

∞
∑ |n|2 |cn ( f )|2 < ∞
n=−∞

and therefore f ∈ W21 (−π , π ).

44 Part I: Fourier Series and the Discrete Fourier Transform

For an integrable function f periodic on the interval [−π , π ] let us introduce the
mapping
f
→ {cn ( f )}∞
n=−∞ ,

where cn ( f ) are the Fourier coefficients of f . This mapping is a linear transforma-

tion. Formula (3.1) says that this mapping is bounded from L1 (−π , π ) to l ∞ (Z).
Here Z denotes all integers, and the sequence space l p (Z) consists of sequences
{bn }∞
n=−∞ for which
∞
∑ |bn | p < ∞
n=−∞

if 1 ≤ p < ∞ and supn∈Z |bn | < ∞ if p = ∞.

Parseval’s equality (7.3) shows that it is also bounded from L2 (−π , π ) to l 2 (Z).
By the Riesz–Thorin interpolation theorem (Theorem 17.7), we may conclude that

this mapping is bounded from L p (−π , π ) to l p (Z) for every 1 < p < 2, 1/p+1/p =
1, and
∞
 π  p /p
p
∑ |cn ( f )| ≤ c p | f (x)| dx
−π
p
< ∞.
n=−∞
Chapter 8
Besov and Hölder Spaces

In this chapter we will consider integrable 2π -periodic functions f defined via

trigonometric Fourier series in L2 (−π , π ) as
∞
f (x) ∼ ∑ cn ( f )einx , (8.1)
n=−∞

where the Fourier coefficients cn ( f ) satisfy Parseval’s equality

 π ∞
1
2π −π
| f (x)|2 dx = ∑ |cn ( f )|2 ;
n=−∞

that is, (8.1) can be understood in the sense of L2 (−π , π ) as

 2
 π  
 
lim  f (x) − ∑ cn ( f )einx  dx = 0.
N→∞ −π  
|n|≤N

We will introduce new spaces of functions (as subspaces of L2 (−π , π )) in terms of

Fourier coefficients. The motivation of such an approach is the following: we proved
(see Theorem 7.3 and Exercise 7.3) that a periodic function f belongs to W21 (−π , π )
if and only if
∞
∑ |n|2 |cn ( f )|2 < ∞.
n=−∞

This fact and Parseval’s equality justify the following definitions.

c Springer International Publishing AG 2017 45

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 8
46 Part I: Fourier Series and the Discrete Fourier Transform

Definition 8.1. A 2π -periodic function f is said to belong to the Sobolev space

W2α (−π , π )

for some α ≥ 0 if
∞
∑ |n|2α |cn ( f )|2 < ∞, 00 := 1.
n=−∞

Definition 8.2. A 2π -periodic function f is said to belong to the Besov space

Bα2,θ (−π , π )

for some α ≥ 0 and some 1 ≤ θ < ∞ if

⎛ ⎞θ /2
∞
∑⎝ ∑ |n|2α |cn ( f )|2 ⎠ < ∞.
j=0 2 j ≤|n|<2 j+1

Definition 8.3. A 2π -periodic function f is said to belong to the Nikol’skii space

H2α (−π , π )

for some α ≥ 0 if
sup ∑
j=0,1,2,... 2 j ≤|n|<2 j+1
|n|2α |cn ( f )|2 < ∞.

Definition 8.4 (See also Definition 1.10).

1) A 2π -periodic function f is said to belong to the Hölder space Cα [−π , π ] for
some noninteger α > 0 if f is continuous on the interval [−π , π ], there is a
continuous derivative f (k) of order k = [α ] on the interval [−π , π ], and for all
h = 0 small enough, we have

sup | f (k) (x + h) − f (k) (x)| ≤ C|h|α −k ,

x∈[−π ,π ]

where the constant C > 0 does not depend on h.

2) By the space Ck [−π , π ] for integer k > 0 we mean the set of 2π -periodic func-
tions f that have continuous derivatives f (k) of order k on the interval [−π , π ].

Remark 8.5. We shall use later the following sufficient condition (see (3.13)): if
there is a constant C > 0 such that for each n = 1, 2, . . . we have

∑ |m|k |cm ( f )| ≤ Cn−α (8.2)

|m|≥n
8 Besov and Hölder Spaces 47

with some integer k ≥ 0 and some 0 < α < 1, then f belongs to the Hölder space
Ck+α [−π , π ].

The definitions 8.1–8.3 imply the following equalities and embeddings:

(1) Bα2,2 (−π , π ) = W2α (−π , π ), α ≥ 0.
(2) W20 (−π , π ) = L2 (−π , π ).
(3) Bα2,1 (−π , π ) ⊂ Bα2,θ (−π , π ) ⊂ H2α (−π , π ), α ≥ 0, 1 ≤ θ < ∞.
(4) B02,θ (−π , π ) ⊂ L2 (−π , π ), 1 ≤ θ ≤ 2 and L2 (−π , π ) ⊂ B02,θ (−π , π ), 2 ≤ θ < ∞.
(5) L2 (−π , π ) ⊂ H20 (−π , π ).

Exercise 8.1. Prove embeddings (3), (4), and (5).

More embeddings are formulated in the following theorems.

Theorem 8.6. If α ≥ 0, then

Cα [−π , π ] ⊂ W2α (−π , π )

and
Cα [−π , π ] ⊂ H2α (−π , π ).

Proof. Let us prove the first claim for integer α ≥ 0. If α = 0 then

C[−π , π ] ⊂ L2 (−π , π ) = W20 (−π , π ).

If α = k > 0 is an integer, then Definition 8.4 implies that

 π
1
| f (k−1) (x + h) − f (k−1) (x)|2 dx ≤ Ch2 .
2π −π

Using Parseval’s equality, we obtain

∞
∑ |n|2k−2 |cn ( f )|2 |einh − 1|2 ≤ Ch2
n=−∞

or
∞
4 ∑ |n|2k−2 |cn ( f )|2 sin2 (nh/2) ≤ Ch2 .
n=−∞

It follows that
∑ |n|2k−2 |cn ( f )|2 n2 h2 ≤ Ch2
|nh|≤2

or
∑ |n|2k |cn ( f )|2 ≤ C.
|n|≤2/|h|
48 Part I: Fourier Series and the Discrete Fourier Transform

Letting h → 0 yields
∞
∑ |n|2k |cn ( f )|2 ≤ C
n=−∞

i.e., f ∈ W2k (−π , π ). For noninteger α one needs to interpolate between the spaces
C[α ] [−π , π ] and C[α ]+1 [−π , π ].
Now let us consider the second claim. As above, for f ∈ Cα [−π , π ], we have
 π
1
| f (k) (x + h) − f (k) (x)|2 dx ≤ C|h|2(α −k) ,
2π −π

where k = [α ] if α is not an integer and k = α − 1 if α is an integer. By Parseval’s

equality,
∞
∑ |n|2k |cn ( f )|2 |einh − 1|2 ≤ C|h|2(α −k) ,
n=−∞

or
∞
2 ∑ |n|2k |cn ( f )|2 (1 − cos(nh)) ≤ C|h|2(α −k) .
n=−∞

It suffices to consider h > 0. If we integrate the last inequality with respect to h > 0
from 0 to l, then
∞ 
sin(nl)
∑ |n|2k
|cn ( f )|2
1 −
nl
≤ Cl 2(α −k) .
n=−∞

It follows that
∑ |n|2k |cn ( f )|2 ≤ Cl 2(α −k)
2≤|n|l≤4

or equivalently,
∑ |n|2k l 2(k−α ) |cn ( f )|2 ≤ C,
2≤|n|l≤4

where the constant C > 0 does not depend on l. Since 2(k − α ) < 0, it follows that

∑ |n|2α |cn ( f )|2 ≤ C

2/l≤|n|≤4/l

for every l > 0. Choosing l = 2− j+1 , we obtain

∑ |n|2α |cn ( f )|2 ≤ C

2 j ≤|n|≤2 j+1

i.e., f ∈ H2α (−π , π ). This completes the proof. 

8 Besov and Hölder Spaces 49

Theorem 8.7. Assume that α > 1/2 and that α − 1/2 is not an integer. Then

H2α (−π , π ) ⊂ Cα −1/2 [−π , π ].

Proof. Let k = [α ], so that α = k + {α }, where {α } denotes the fractional part of α .

Note that in general, 0 ≤ {α } < 1 and in this theorem {α } = 1/2. We will assume
first that {α } = 0, i.e., α = k is an integer and k ≥ 1. If f ∈ H2k (−π , π ), then there
is a constant C > 0 such that

∑ |m|2k |cm ( f )|2 ≤ C (8.3)

2 j ≤|m|<2 j+1

for each j = 0, 1, 2, . . .. Let us estimate the tail (8.2). Indeed, by the Cauchy–
Bunyakovsky–Schwarz inequality and (8.3),
∞
∑ |m|k−1 |cm ( f )| ≤ ∑ ∑ |m|k−1 |cm ( f )|
|m|≥n j= j0 2 j ≤|m|<2 j+1
2 j0 ∼n
⎛ ⎞1/2 ⎛ ⎞1/2
∞
⎝ 1 ⎠
≤ ∑ ∑ |m|2k |cm ( f )|2 ⎠ ⎝ ∑ 2
j= j0 2 j ≤|m|<2 j+1 2 j ≤|m|<2 j+1
m
2 j0 ∼n
⎛ ⎞1/2
√ ∞
1 √ ∞ j j0
≤ C ∑ ⎝ ∑ m2 ⎠ ≤ C ∑ 2− 2 ≤ C2− 2 .
j= j0 2 j ≤|m|<2 j+1 j= j0
2 j0 ∼n 2 j0 ∼n

Here, 2 j0 ∼ n means that 2 j0 ≤ n < 2 j0 +1 . Therefore, we obtain

∑ |m|k−1 |cm ( f )| ≤ Cn−1/2 ,

|m|≥n

where the constant C > 0 is independent of n. This means that (see (8.2)) f belongs
to Ck−1+1/2 [−π , π ] = Ck−1/2 [−π , π ].
If α > 1/2 is not an integer and α −1/2 is not an integer, then for f ∈ H2α (−π , π )
we have instead of (8.3) the estimate

∑ |m|2k+2{α } |cm ( f )|2 ≤ C,

2 j ≤|m|<2 j+1

where k = [α ], 0 < {α } < 1 and {α } = 1/2. If k = 0, then 1/2 < α = {α } < 1.

Repeating now the above procedure, we obtain easily
50 Part I: Fourier Series and the Discrete Fourier Transform

⎛ ⎞1 ⎛ ⎞1
2 2
∞
∑ |cm ( f )| ≤ ∑ ⎝ ∑ 2α
|m| |cm ( f )| 2⎠ ⎝ ∑ −2α ⎠
|m|
|m|≥n j= j0 2 j ≤|m|<2 j+1 2 j ≤|m|<2 j+1
2 j0 ∼n
⎛ ⎞1/2
∞ ∞
≤C ∑ ⎝ ∑ |m|−2α ⎠ ≤C ∑ 2−(α −1/2) j ≤ Cn−(α −1/2) ,
j= j0 2 j ≤|m|<2 j+1 j= j0
2 j0 ∼n 2 j0 ∼n

i.e., we have again that f ∈ Cα −1/2 [−π , π ].

For the case [α ] = k ≥ 1, α is not an integer, and α − 1/2 is not an integer, we
consider two cases: 0 < {α } < 1/2 and 1/2 < {α } < 1. In the first case we have

∑ |m|k−1 |cm ( f )| ≤
|m|≥n
⎛ ⎞1/2 ⎛ ⎞1/2
∞
∑ ⎝ ∑ |m| 2k+2{α }
|cm ( f )|2⎠ ⎝ ∑ |m| −2−2{α } ⎠
j= j0 2 j ≤|m|<2 j+1 2 j ≤|m|<2 j+1
2 j0 ∼n
∞
≤C ∑ 2− j− j{α } 2 j/2 ≤ Cn−1/2−{α } .
j= j0
2 j0 ∼n

This means again that f ∈ Ck−1+1/2+{α } [−π , π ] = Cα −1/2 [−π , π ]. In the second
case, 1/2 < {α } < 1, we proceed as follows:

∑ |m|k |cm ( f )| ≤
|m|≥n
⎛ ⎞1/2 ⎛ ⎞1/2
∞
∑ ⎝ ∑ |m|2k+2{α } |cm ( f )|2 ⎠ ⎝ ∑ |m|−2{α } ⎠
j= j0 2 j ≤|m|<2 j+1 2 j ≤|m|<2 j+1
2 j0 ∼n
∞
≤C ∑ 2−({α }−1/2) j ≤ Cn−{α }+1/2 .
j= j0
2 j0 ∼n

This means that f ∈ Ck+{α }−1/2 [−π , π ] = Cα −1/2 [−π , π ]. Hence, the theorem is
completely proved. 

Corollary 8.8. Assume that α = k + 1/2 for some integer k ≥ 1. Then

H2α (−π , π ) ⊂ Cβ −1/2 [−π , π ]

for every 1/2 < β < α .

 8 Besov and Hölder Spaces 51

Corollary 8.9. Assume that α > 1/2 and α − 1/2 is not an integer. Then

Bα2,θ (−π , π ) ⊂ Cα −1/2 [−π , π ]

for every 1 ≤ θ < ∞.

Exercise 8.2. Prove Corollaries 8.8 and 8.9.
Exercise 8.3. Prove that the Fourier series (8.1) with coefficients

1
cn ( f ) = , n = 0, c0 ( f ) = 1
|n| log(1 + |n|)

1/2
defines a function from the Besov space B2,θ (−π , π ) for every 1 < θ < ∞, but not
for θ = 1.
Exercise 8.4. Prove that the Fourier series (8.1) with coefficients

1
cn ( f ) = , n = 0, c0 ( f ) = 1
|n|3/2 log(1 + |n|)

defines a function from the Besov space B12,θ (−π , π ) for every 1 < θ < ∞, but not
for θ = 1.
Exercise 8.5. Consider the Fourier series (8.1) with coefficients

1
cn ( f ) = , n = 0, c0 ( f ) = 1.
|n|2 logβ (1 + |n|)

Prove that
3/2
(1) f ∈ H2 (−π , π ) if β ≥ 0
3/2
(2) f ∈ W2 (−π , π ) if β > 1/2
(3) f ∈ C1 [−π , π ] if β > 1 but f ∈
/ C1 [−π , π ] if β ≤ 1.
So the embeddings W2α (−π , π ) ⊂ Cα −1/2 [−π , π ] and H2α (−π , π ) ⊂ Cα −1/2 [−π , π ]
are not valid for α − 1/2 an integer (see Theorem 8.6).
Exercise 8.6. Let
∞
∑ ak cos(bk x)
k=0

be a trigonometric series, where b = 2, 3, . . . and 0 < a < 1. Prove that the series
defines a function from C1 [−π , π ] if 0 < ab < 1 and a function from the Hölder
space Cγ [−π , π ], γ < 1 if ab = 1.
Exercise 8.7. Assume that a = 1/b2 in Exercise 8.6. Is it true that this function
belongs to C1 [−π , π ]?
Chapter 9
Absolute Convergence. Bernstein and Peetre
Theorems.

We begin by proving the equivalence between a 2π -periodic function f belonging

to the Nikol’skii space H2α (−π , π ) for some 0 < α < 1 in the sense of Definition 8.3
and the L2 Hölder condition of order α , i.e.,
 π
1
| f (x + h) − f (x)|2 dx ≤ K|h|2α (9.1)
2π −π

for some constant K > 0 and for all h = 0 sufficiently small. Indeed, due to Parseval’s
equality we have
 π ∞
1
2π −π
| f (x + h) − f (x)|2 dx = ∑ |cn ( f )|2 |einh − 1|2
n=−∞
≤ ∑j |n|2 |h|2 |cn ( f )|2 + 4 ∑j |cn ( f )|2 , (9.2)
|n|≤2 0 |n|≥2 0

where j0 is chosen so that 2 j0 ≤ 1

|h| < 2 j0 +1 . The first sum on the right-hand side of
(9.2) is estimated from above as

j0
|h|2 ∑j |n|2 |cn ( f )|2 ≤ |h|2 ∑ ∑ |n|2 |cn ( f )|2
|n|≤2 0 j=0 2 j ≤|n|<2 j+1
j0  2−2α
≤ |h|2 ∑ 2 j+1 ∑ |n|2α |cn ( f )|2
j=0 2 j ≤|n|<2 j+1

 j +2
j0   22−2α 0 − 22−2α
≤ C|h| 2
∑ 2 2−2α j+1
= C|h| 2
22−2α − 1
j=0
 2−2α
 2−2α 1
≤ C|h|2 2 j0 ≤ C|h|2
= C|h|2α
|h|

c Springer International Publishing AG 2017 53

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 9
54 Part I: Fourier Series and the Discrete Fourier Transform

if 0 < α < 1. We used this condition for α because we considered a geometric sum
with common ratio 22−2α = 1. The second sum on the right-hand side of (9.2) is
estimated from above as
∞ ∞
4 ∑ ∑ |n|2α |n|−2α |cn ( f )|2 ≤ 4 ∑ 2−2α j ∑ |n|2α |cn ( f )|2
j= j0 2 j ≤|n|<2 j+1 j= j0 2 j ≤|n|<2 j+1
∞
≤C ∑ 2−2α j ≤ C2−2 j0 α ≤ C|h|2α ,
j= j0

since 1
|h| ≤ 2 j0 +1 and the criterion of Definition 8.3 is satisfied. Thus, (9.1) is proved.
Conversely, if the L2 Hölder condition (9.1) is fulfilled, then Theorem 7.5 implies
for each N = 1, 2, . . . the inequality

∑ |cn ( f )|2 ≤ CN −2α

|n|≥N

with the same α as in (9.1). But this leads to the inequality

N 2α ∑ |cn ( f )|2 ≤ C,
N≤|n|<2N

where the constant C is independent of N. Thus, we obtain for every integer N > 0
that
∑ |n|2α |cn ( f )|2 ≤ C.
N≤|n|<2N

Since N is arbitrary, we may conclude that f ∈ H2α (−π , π ) for 0 < α ≤ 1 in the
sense of Definition 8.3. Therefore, the L2 Hölder condition (9.1) can be considered
as an equivalent definition of the Nikol’skii space H2α (−π , π ) for 0 < α < 1.
Exercise 9.1. Prove that f belongs to the Nikol’skii space H2α (−π , π ) for every
noninteger α > 0 in the sense of Definition 8.3 if and only if the following L2 Hölder
condition holds:
 π
1
| f (k) (x + h) − f (k) (x)|2 dx ≤ K|h|2α −2k
2π −π

with some constant K > 0 and k = [α ].

Exercise 9.2. Prove that f ∈ W2k (−π , π ), k = 1, 2, . . ., if and only if

 π
1
| f (k−1) (x + h) − f (k−1) (x)|2 dx ≤ C|h|2 .
2π −π

Theorem 9.1 (Bernstein, 1914). Assume that a 2π -periodic function f satisfies

the L2 Hölder condition with 1/2 < α ≤ 1. Then its trigonometric Fourier series
converges absolutely, i.e.,
9 Absolute Convergence. Bernstein and Peetre Theorems. 55

∞
∑ |cn ( f )| < ∞.
n=−∞

Proof. Since the L2 Hölder condition (9.1) is equivalent to f ∈ H2α (−π , π ) for
0 < α < 1, there is a constant C > 0 such that

∑ |n|2α |cn ( f )|2 ≤ C

2 j ≤|n|<2 j+1

for each j = 0, 1, 2, . . .. Hence we have

∞ ∞
∑ |cn ( f )| = |c0 ( f )| + ∑ ∑ |n|α |cn ( f )||n|−α
n=−∞ j=0 2 j ≤|n|<2 j+1
⎛ ⎞1/2 ⎛ ⎞1/2
∞
≤ |c0 ( f )| + ∑ ⎝ ∑ |n|2α |cn ( f )|2 ⎠ ⎝ ∑ |n|−2α ⎠
j=0 2 j ≤|n|<2 j+1 2 j ≤|n|<2 j+1
√ ∞ √ ∞  j
≤ |c0 ( f )| + C ∑ 2−α j 2 j/2 = |c0 ( f )| + C ∑ 2−(α −1/2) < ∞,
j=0 j=0

since α > 1/2. Thus, the theorem is proved. 

Corollary 9.2. Theorem 9.1 holds for Cα [−π , π ], Bα2,θ (−π , π ), and H2α (−π , π ) for
every α > 1/2 and 1 ≤ θ < ∞.

Exercise 9.3. Prove this Corollary.

Theorem 9.3 (Peetre, 1967). Assume that a 2π -periodic function f belongs to

1/2
the Besov space B2,1 (−π , π ). Then its trigonometric Fourier series converges
absolutely.
1/2
Proof. If f ∈ B2,1 (−π , π ), then

⎛ ⎞1/2
∞
∑⎝ ∑ |n||cn ( f )|2 ⎠ < ∞. (9.3)
j=0 2 j ≤|n|<2 j+1

Hence we have
∞ ∞
∑ |cn ( f )| = |c0 ( f )| + ∑ ∑ |n|1/2 |cn ( f )||n|−1/2
n=−∞ j=0 2 j ≤|n|<2 j+1
⎛ ⎞1/2 ⎛ ⎞1/2
∞
≤ |c0 ( f )| + ∑ ⎝ ∑ |n||cn ( f )|2 ⎠ ⎝ ∑ |n|−1 ⎠
j=0 2 j ≤|n|<2 j+1 2 j ≤|n|<2 j+1
56 Part I: Fourier Series and the Discrete Fourier Transform

⎛ ⎞1/2
∞  1/2
≤ |c0 ( f )| + ∑ ⎝ ∑ |n||cn ( f )|2 ⎠ 2− j 2 j
j=0 2 j ≤|n|<2 j+1
⎛ ⎞1/2
∞
= |c0 ( f )| + ∑ ⎝ ∑ |n||cn ( f )|2 ⎠ <∞
j=0 2 j ≤|n|<2 j+1

due to (9.3). This proves the theorem. 

1/2
Corollary 9.4. It is true that B2,1 (−π , π ) ⊂ C[−π , π ].
1/2
Exercise 9.4. Prove that the embedding B2,θ (−π , π ) ⊂ C[−π , π ] does not hold for
1 < θ < ∞ by considering the function from Exercise 8.3.

Exercise 9.5. Prove that H2α (−π , π ) ⊂ B2,1 (−π , π ) if α > 1/2.
1/2

Theorem 9.5. Assume that a 2π -periodic function f belongs to the Sobolev space
Wp1 (−π , π ) with some 1 < p < ∞. Then its trigonometric Fourier series converges
absolutely.
Proof. Since Wp11 (−π , π ) ⊂ Wp12 (−π , π ) for 1 ≤ p2 < p1 , we may assume without
loss of generality that f ∈ Wp1 (−π , π ) with 1 < p ≤ 2. Then there is a function
g ∈ L p (−π , π ) with 1 < p ≤ 2 such that
 x  π
f (x) = g(t)dt + f (−π ), g(t)dt = 0. (9.4)
−π −π

As we know from the proof of Theorem 7.3, (9.4) leads to

1
cn ( f ) = cn (g), n = 0.
in

Since g ∈ L p (−π , π ) with 1 < p ≤ 2, the results of Chapter 7 give

1/p
∞
p
∑ |cn (g)| < ∞, (9.5)
n=−∞

where 1
p + 1
p = 1. The facts (9.4), (9.5) and Hölder’s inequality imply that
∞
1
∑ |cn ( f )| = |c0 ( f )|+ ∑ |cn (g)|
n=−∞ n=0 |n|
1/p 1/p
1 p
≤ |c0 ( f )| + ∑ p ∑ |cn (g)| < ∞,
n=0 |n| n=0

since 1 < p ≤ 2. This completes the proof. 

9 Absolute Convergence. Bernstein and Peetre Theorems. 57

Remark 9.6. For the Sobolev space W11 (−π , π ), this theorem is not valid, i.e., there
is a function f from W11 (−π , π ) with absolutely divergent trigonometric Fourier
series. More precisely, we will prove in the next chapter that the function
∞
sin nx
f (x) := ∑ n log(1 + n) (9.6)
n=1

belongs to the Sobolev space W11 (−π , π ) and is continuous on the interval [−π , π ],
but its trigonometric Fourier series (9.6) diverges absolutely.

The next theorem is due to Zigmund (1958–1959).

Theorem 9.7. Suppose that f ∈ W11 (−π , π ) ∩ Cα [−π , π ] with some 0 < α < 1.
Then its trigonometric Fourier series converges absolutely.

Proof. Since f ∈ W11 (−π , π ), it follows that f is of bounded variation. The period-
icity of f implies that
 π  π N
1 1
2π −π
| f (x+h)− f (x)|2 dx =
2π N −π k=1
∑ | f (x+kh)− f (x+(k −1)h)|2 dx, (9.7)

where the integer N is chosen so that N|h| ≤ 1 for h = 0 sufficiently small. We

choose N = [1/|h|]. Since f ∈ Cα [−π , π ], the right-hand side of (9.7) can be esti-
mated as
 π N
1
2π N ∑ | f (x + kh) − f (x + (k − 1)h)|2dx
−π k=1

C|h|α π N
≤
2π N −π k=1∑ | f (x + kh) − f (x + (k − 1)h)|dx
C|h|α  −π  C|h|α
≤ V−π −1 ( f ) +V−ππ ( f ) +Vππ +1 ( f ) 2π ≤
2π N N
C|h|α C|h|α C|h|α C|h|α +1
= = ≤ = ≤ C|h|α +1
[1/|h|] 1/|h| − {1/|h|} 1/|h| − 1 1 − |h|
α +1
if |h| ≤ 1/2. This inequality means (see (9.7)) that f ∈ H2 2 (−π , π ) with α +1
2 >
1/2 for α > 0. An application of Bernstein’s theorem completes the proof of the
theorem. 

Exercise 9.6. Let a (periodic) function f be defined by

∞
eikx
f (x) := ∑ .
k=1 k
 58 Part I: Fourier Series and the Discrete Fourier Transform

1/2
Prove that the function f belongs to the Nikol’skii space H2 (−π , π ) but its
trigonometric Fourier series is not absolutely convergent.

Exercise 9.7. Let a (periodic) function f be defined by the absolutely convergent

Fourier series
∞
eikx
f (x) := ∑ 3/2 .
k=1 k

(1) Show that f belongs to the Nikol’skii space H21 (−π , π ) but
 π
1 4h2 π
| f (x + h) − f (x)|2 dx ≥ log
2π −π π2 |h|

for 0 < |h| < 1, that is, (9.1) does not hold for α = 1.
(2) Show that f does not belong to the Besov space B12,θ (−π , π ) for any 1 ≤ θ < ∞.
Chapter 10
Dirichlet Kernel. Pointwise and Uniform
Convergence.

The material of this chapter forms a central part of the theory of trigonometric
Fourier series. In this chapter we will answer the following question: to what value
does a trigonometric Fourier series converge?
The Dirichlet kernel DN (x), which is defined by the symmetric finite trigonomet-
ric sum
DN (x) := ∑ einx , (10.1)
|n|≤N

plays a key role in this chapter. If x ∈ [−π , π ] \ {0}, then DN (x) from (10.1) can be
recalculated as follows. Using Euler’s formula, we have

N N 2N
DN (x) = ∑ einx = e−iNx ∑ ei(n+N)x = e−iNx ∑ eikx
n=−N n=−N k=0
1 − ei(2N+1)x e−iNx − ei(N+1)x
= e−iNx =
1 − eix 1 − eix
ei(N+1/2)x −e −i(N+1/2)x sin(N + 1/2)x
= = .
eix/2 − e−ix/2 sin x/2

Thus, the Dirichlet kernel equals

sin(N + 1/2)x
DN (x) = , x = 0. (10.2)
sin x/2

For x = 0 we have
DN (0) = 2N + 1 = lim DN (x),
x→0

so that (10.2) holds for all x ∈ [−π , π ].

c Springer International Publishing AG 2017 59

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 10
60 Part I: Fourier Series and the Discrete Fourier Transform

Exercise 10.1. Prove that


1 π
(1) DN (x)dx = 1, N = 0, 1, 2, . . .;
2π −π
(2) KN (x) = N+1
1
∑Nj=0 D j (x), where KN (x) is the Fejér kernel (5.2)
Recall that the trigonometric Fourier partial sum is given by

SN f (x) = ∑ cn ( f )einx . (10.3)

|n|≤N

The Fourier coefficients of DN (x) are equal to

 π

1 0, |n| > N,
cn (DN ) =
2π −π
e −inx
∑ e dx = 1, |n| ≤ N.
ikx

|k|≤N

Hence, if f is periodic and integrable, then the partial sum (10.3) can be rewritten
as (see Exercise 4.1)
∞  π
1
SN f (x) = ∑ cn (DN )cn ( f )einx = ( f ∗ DN )(x) =
2π −π
DN (x − y) f (y)dy
n=−∞
 π  π
1 1 sin(N + 1/2)y
= DN (y) f (x + y)dy = f (x + y) dy. (10.4)
2π −π 2π −π sin y/2

Exercise 10.2. Let f be the function


− 2xπ ,
1
0 < x ≤ π,
f (x) = 2
− 2 − 2π , −π ≤ x < 0.
1 x

Show that
(1) (SN f ) (x) = 21π (DN (x) − 1);
(2) limN→∞ SN f (x) = f (x), x = 0;
(3) limN→∞ SN f (0) = 0.
Exercise 10.3. Prove that as N → ∞,
 π
1 4 log N
|DN (x)|dx = + O(1).
2π −π π2

Since the Dirichlet kernel is an even function (see (10.2)), we can rewrite (10.4) as
 π
1 sin(N + 1/2)y
SN f (x) = ( f (x + y) + f (x − y)) dy.
2π 0 sin y/2

Using the normalization of the Dirichlet kernel (see Exercise 10.1), we have for
every function S(x) that
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 61
 π
1 sin(N + 1/2)y
SN f (x) − S(x) = ( f (x + y) + f (x − y) − 2S(x)) dy. (10.5)
2π 0 sin y/2

Our aim is to define S(x) so that the limit of the right-hand side of (10.5) is equal to
zero. We will simplify the problem by splitting it into two steps. The first simplifi-
cation is connected with the following technical lemma.

Lemma 10.1. For all z ∈ [−π , π ], it is true that

 
 1 2  π 2

 sin z/2 − z  ≤ 24 .

Proof. First we show that

 z  |z|3

sin z/2 −  ≤ (10.6)
2 48

for all z ∈ [−π , π ]. In order to prove this inequality, it is enough to show that

x3
0 < x − sin x <
6

for all 0 < x < π /2. The left inequality is well known. To prove the right inequality
we introduce h(x) as
h(x) = x − sin x − x3 /6.

Then its derivative satisfies

h (x) = 1 − cos x − x2 /2 = 2(sin2 x/2 − x2 /4) < 0

for all 0 < x < π /2. Thus, h(x) is monotonically decreasing on the interval [0, π /2],
which implies that
0 = h(0) > h(x) = x − sin x − x3 /6

for all 0 < x < π /2. This proves (10.6), which in turn yields
 
 1 2  2 |z/2 − sin(z/2)| |z|3 /24 |z|3 /24 π |z| π 2

 sin z/2 z  = |z|| sin(z/2)| ≤ |z|| sin(z/2)| ≤ |z||z|/π ≤ 24 ≤ 24 ,
−

since | sin z/2| ≥ |z|/π for all z ∈ [−π , π ]. This finishes the proof. 

As an immediate corollary of Lemma 10.1, we obtain for every periodic and inte-
grable function f that the function
 
1 2
y → ( f (x + y) + f (x − y) − 2S(x)) −
sin y/2 y
62 Part I: Fourier Series and the Discrete Fourier Transform

is integrable on the interval [0, π ] uniformly in x ∈ [−π , π ] if S(x) is bounded, i.e.,

 π  
 1 2 
| f (x + y) + f (x − y) − 2S(x)|  − dy
0 sin y/2 y 
  π  π 
π2
≤ | f (x + y)|dy + | f (x − y)|dy + 2π |S(x)|
24 0 0
 
π2 π
≤ | f (y)|dy + 2π sup |S(x)| .
24 −π x∈[−π ,π ]

Application of the Riemann–Lebesgue lemma (Theorem 6.1) gives us that

 π  
1 1 2
lim ( f (x + y) + f (x − y) − 2S(x)) − sin(N + 1/2)y = 0
N→∞ 2π 0 sin y/2 y

pointwise in x ∈ [−π , π ] and even uniformly in x ∈ [−π , π ] if S(x) is bounded on the

interval [−π , π ].
Thus, we have reduced the question of pointwise or uniform convergence in
(10.5) to proving that
 π
sin(N + 1/2)y
lim ( f (x + y) + f (x − y) − 2S(x)) dy = 0 (10.7)
N→∞ 0 y

pointwise or uniformly in x ∈ [−π , π ].

For the second simplification we consider the contribution to (10.7) from the
interval 0 < δ ≤ y ≤ π . Note that the function

f (x + y) + f (x − y) − 2S(x)
y

is integrable in y on the interval [δ , π ] uniformly in x ∈ [−π , π ] if S(x) is bounded.

Hence, by the Riemann–Lebesgue lemma this contribution tends to zero as N → ∞.
We summarize these two simplifications as

SN f (x) − S(x)
 δ
1 sin(N + 1/2)y
= ( f (x + y) + f (x − y) − 2S(x)) dy + o(1) (10.8)
π 0 y

as N → ∞ pointwise or uniformly in x ∈ [−π , π ].

Let us assume now that f is a piecewise continuous periodic function. We wish
to know the values of S(x) in (10.8) to which the trigonometric Fourier partial sum
can converge. The second part of Theorem 5.1 shows that the Fejér means converge
to
1
lim σN f (x) = ( f (x + 0) + f (x − 0))
N→∞ 2
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 63

for every x ∈ [−π , π ] pointwise. But since

∑Nj=0 S j f (x)
σN f (x) = ,
N +1

SN f (x) can converge (if it converges) only to the value

1
( f (x + 0) + f (x − 0)) .
2
We can obtain some sufficient conditions when the limit in (10.8) exists.

Theorem 10.2. Suppose that S(x) is chosen so that

 δ
| f (x + y) + f (x − y) − 2S(x)|
dy < ∞ (10.9)
0 y

pointwise or uniformly in x ∈ [−π , π ]. Then

lim SN f (x) = S(x) (10.10)

N→∞

pointwise or uniformly in x ∈ [−π , π ].

Proof. The result follows immediately from (10.8) and the Riemann–Lebesgue
lemma. 

Remark 10.3. If in (10.10) we have uniform convergence, then S(x) must necessar-
ily be periodic (S(−π ) = S(π )) and continuous on the interval [−π , π ].

Corollary 10.4. Suppose a periodic function f belongs to the Hölder space

Cα [−π , π ] for some 0 < α ≤ 1. Then

lim SN f (x) = f (x)

N→∞

uniformly in x ∈ [−π , π ].

Proof. Since f ∈ Cα [−π , π ], it follows that

| f (x + y) + f (x − y) − 2 f (x)| ≤ | f (x + y) − f (x)| + | f (x − y) − f (x)| ≤ Cyα

for 0 < y < δ . This means that the condition (10.9) holds with S(x) = f (x) uniformly
in x ∈ [−π , π ], from which the statement of the corollary follows. 
64 Part I: Fourier Series and the Discrete Fourier Transform

Theorem 10.5 (Dirichlet, Jordan). Suppose that a periodic function f is of

bounded variation on the interval [x − δ , x + δ ] for some δ > 0 and some fixed x.
Then
1
lim SN f (x) = ( f (x + 0) + f (x − 0)) .
N→∞ 2

Proof. Since f is of bounded variation, the limit

1 1
lim ( f (x + y) + f (x − y)) = ( f (x + 0) + f (x − 0)) =: S(x) (10.11)
y→0+ 2 2

exists. For 0 < y < δ we define

F(y) := f (x + y) + f (x − y) − 2S(x),

where S(x) is defined by (10.11). Note that F(0) = 0. Let us also define
 y
sin(N + 1/2)t
GN (y) := dt, 0 < y ≤ δ. (10.12)
0 t

It is easy to check that

 (N+1/2)y
sin ρ
GN (y) = dρ , 0 < y ≤ δ.
0 ρ

This representation implies that

 ∞
sin ρ π
lim GN (y) = dρ = . (10.13)
N→∞ 0 ρ 2

For fixed x we have from (10.8) and (10.12) that

 δ
1
SN f (x) − S(x) = F(y)GN (y)dy + o(1), N → ∞.
π 0

Here integration by parts gives

  δ 
1
SN f (x) − S(x) = F(y)GN (y)|δ0 − GN (y)dF(y) + o(1)
π 0
  δ 
1
= F(δ )GN (δ ) − GN (y)dF(y) + o(1), N → ∞, (10.14)
π 0

where the last integral is well defined as the Stieltjes integral of the continuous
function GN (y) with respect to the function of bounded variation F(y). Since the
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 65

limit (10.13) holds and GN (y) is continuous, we can consider the limit in (10.14) as
N → ∞. Hence, we obtain
  
1 π π δ
lim (SN f (x) − S(x)) = F(δ ) − dF(y)
N→∞ π 2 2 0
1
= (F(δ ) − F(δ ) + F(0)) = 0.
2
This completes the proof. 

Corollary 10.6. If f is periodic and if f and f  are piecewise continuous, then the
Fourier series of f converges to 12 ( f (x + 0) + f (x − 0)) at all points. If in addition
f is continuous on (−∞, ∞), then its Fourier series converges to f (x) uniformly on
(−∞, ∞).

Corollary 10.7. If f is periodic and belongs to the Sobolev space W11 (−π , π ), then
its trigonometric Fourier series converges pointwise to f (x) everywhere.

Proof. Since f ∈ W11 (−π , π ), it is of bounded variation and continuous on the in-
terval [−π , π ]. In this case, the value S(x) from (10.11) equals f (x) at every point
x ∈ [−π , π ]. Thus,
lim SN f (x) = f (x)
N→∞

pointwise in x ∈ [−π , π ]. 

Remark 10.8. The above proof does not allow us to conclude uniform convergence
of the trigonometric Fourier series of functions from the Sobolev space W11 (−π , π ).
However, uniform convergence is in fact the case, as we will prove later in this
chapter.

Exercise 10.4. Show that

1
f (x) = 1
, |x| < 1/2
log |x|

is of bounded variation but this function does not satisfy condition (10.9) at x = 0.
Hint.  δ 
 1 
 
 0 y log y dy = +∞.

Exercise 10.5. Show that

1
f (x) = x sin
x
satisfies condition (10.9) at x = 0 but this function is not of bounded variation, see
Exercise 1.3.
66 Part I: Fourier Series and the Discrete Fourier Transform

We can return now to the question of term-by-term integration of trigonometric

Fourier series.
Theorem 10.9. Suppose f belongs to L1 (−π , π ). Then
 b  b
lim SN f (x)dx = f (x)dx
N→∞ a a

for every interval (a, b) ⊂ [−π , π ].

Proof. For a given L1 function f (not necessarily periodic) we introduce a new
function F as  x
F(x) := ( f (t) − c0 ( f ))dt. (10.15)
−π

It is clear that F(x) belongs to the Sobolev space W11 (−π , π ) with F(−π ) = F(π ) =
0 (periodicity) and
F  (x) = f (x) − c0 ( f ).

This implies

cn (F  ) = incn (F) = cn ( f ), n = 0, c0 (F  ) = 0.

Corollary 10.7 gives us that F(x) has everywhere convergent trigonometric Fourier
series
cn ( f ) inx
F(x) = c0 (F) + ∑ e . (10.16)
n=0 in

In particular, for every −π ≤ a < b ≤ π we have from (10.16) that

 b
cn ( f )
F(b) − F(a) = ∑ in (einb − eina ) = ∑ cn ( f ) a
einx dx,
n=0 n=0

or equivalently (see (10.15)),

 b  a  b
( f (x) − c0 ( f ))dx − ( f (x) − c0 ( f ))dx = f (x)dx − (b − a)c0 ( f )
−π −π a
 b
= ∑ cn ( f ) a
einx dx.
n=0

Thus, we obtain finally

 b ∞  b  b

a
f (x)dx = ∑ cn ( f )
a
einx dx = lim
N→∞
∑ cn ( f )
a
einx dx
n=−∞ |n|≤N
 b
= lim SN f (x)dx.
N→∞ a
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 67

This proves the theorem. 

Exercise 10.6. Calculate c0 (F) for F defined by (10.15).

Corollary 10.10. If f ∈ L1 (−π , π ), then the series

cn ( f ) cn ( f )(−1)n
∑ n and ∑ n
n=0 n=0

converge.

Proof. The result follows from (10.16). 

Corollary 10.11. The series

∞
sin(nx)
∑ log(1 + n)
n=1

is not the Fourier series of an L1 function.

Proof. Let us assume to the contrary that there is a function f ∈ L1 (−π , π ) such
that
∞
sin(nx) 1 ∞ einx 1 ∞ e−inx
f (x) ∼ ∑ = ∑ log(1 + n) 2i ∑ log(1 + n)
−
n=1 log(1 + n) 2i n=1 n=1
1 ∞ einx 1 −∞ sgn(n)einx 1 sgn(n)einx
= ∑ + ∑ = ∑
2i n=1 log(1 + n) 2i n=−1 log(1 + |n|) 2i n=0 log(1 + |n|)
,

i.e., we have
1 sgn(n)
cn ( f ) = , n = 0, c0 ( f ) = 0.
2i log(1 + |n|)

Since cn ( f ) = −c−n ( f ), this trigonometric Fourier series can be interpreted as the

Fourier series of some odd L1 function. Then Corollary 10.10 implies that

sgn(n) 1 ∞ 1
∑ 2in log(1 + |n|) i ∑ n log(1 + n)
=
n=0 n=1

must be convergent. But this is not true. This contradiction proves this corollary. 

Remark 10.12. If we define the function f by the series in Corollary 10.11, then it
turns out that  π  π
f (x)dx = 0, | f (x)|dx = +∞.
−π −π

Recall that the Poisson kernel is equal to

68 Part I: Fourier Series and the Discrete Fourier Transform

1 − r2
Pr (x) = , 0≤r<1
1 − 2r cos x + r2

and its trigonometric Fourier series is

∞
Pr (x) = ∑ r|n| einx .
n=−∞

Corollary 10.4 shows us that this series converges to Pr (x) uniformly in x ∈ [−π , π ].
Theorem 10.13. Suppose that f ∈ C[−π , π ] is periodic. Then

lim (Pr ∗ f )(x) = f (x)

r→1−

uniformly in x ∈ [−π , π ] or
∞
lim
r→1− n=−∞
∑ r|n| cn ( f )einx = f (x) (10.17)

uniformly in x ∈ [−π , π ] even if f has no convergent trigonometric Fourier series.

Proof. Using the normalization
 π
1
Pr (x)dx = 1,
2π −π

we have

1 π
(Pr ∗ f )(x) − f (x) = Pr (y)( f (x − y) − f (x))dy
2π −π

1
= Pr (y)( f (x − y) − f (x))dy
2π |y|≤δ

1
+ Pr (y)( f (x − y) − f (x))dy =: I1 + I2 .
2π δ ≤|y|≤π

Since f is continuous on [−π , π ], it follows that I1 can be estimated as

 π
1
|I1 | ≤ sup | f (x − y) − f (x)| Pr (y)dy
x∈[−π ,π ],|y|≤δ 2π −π

= sup | f (x − y) − f (x)| → 0
x∈[−π ,π ],|y|≤δ

as δ → 0. At the same time, I2 can be estimated as


1
|I2 | ≤ 2 max | f (x)| Pr (y)dy.
|x|≤π 2π δ ≤|y|≤π
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 69

For δ ≤ |y| ≤ π the Poisson kernel can be estimated as

1 − r2 1 − r2
Pr (y) = =
1 − 2r cos y + r2 4r sin2 y/2 + (1 − r)2
1−r 1−r π2 1 − r
≤ ≤ = .
2r sin y/2 2rδ /π
2 2 2 2 rδ 2

If we choose δ 4 = 1 − r, then I2 is estimated as

√
1 π2 1 − r π 1−r
|I2 | ≤ max | f (x)| √ = max | f (x)| →0
π |x|≤π 2 r 1−r 2 |x|≤π r

as r → 1−. Hence, the estimates for I1 and I2 show that

lim ((Pr ∗ f )(x) − f (x)) = 0

r→1−

uniformly in x ∈ [−π , π ]. The equality (10.17) follows from this fact and
Exercise 4.1. This completes the proof. 

We will prove now the well known Hardy’s theorem and then apply it to the uniform
convergence of the trigonometric sum SN f (x).

Theorem 10.14 (Hardy, 1949). Let {ak }∞

k=0 be a sequence of complex numbers
such that
k|ak | ≤ M, k = 0, 1, 2, . . . , (10.18)

where the constant M is independent of k. If the limit

n  
j
lim σn := lim
n→∞ n→∞
∑ 1−
n+1
aj = a (10.19)
j=0

exists, then
lim (σn − sn ) = 0,
n→∞

where sn = ∑nj=0 a j , i.e. also

lim sn = a. (10.20)
n→∞

If ak depends on x and (10.18) holds uniformly in x and convergence in (10.19) is

uniform, then convergence in (10.20) is also uniform.

Proof. For n < m it is true that

m
(m + 1)σm − (n + 1)σn − ∑ (m + 1 − j)a j = (m − n)sn .
j=n+1
70 Part I: Fourier Series and the Discrete Fourier Transform

Indeed,
m
(m + 1)σm − (n + 1)σn − ∑ (m + 1 − j)a j
j=n+1
m n m
= ∑ (m + 1 − j)a j − ∑ (n + 1 − j)a j − ∑ (m + 1 − j)a j
j=0 j=0 j=n+1
n n n
= ∑ (m + 1 − j)a j − ∑ (n + 1 − j)a j = ∑ (m − n)a j = (m − n)sn .
j=0 j=0 j=0

Therefore,
m
(m + 1)σm − (n + 1)σn − ∑ (m + 1 − j)a j − (m − n)σn = (m − n)sn − (m − n)σn
j=n+1

or equivalently,
m
m(σm − σn ) + (σm − σn ) − ∑ (m + 1 − j)a j = (m − n)(sn − σn ),
j=n+1

i.e.,  
m+1 m+1 m
j
m−n
(σm − σn ) −
m−n ∑ 1−
m+1
a j = sn − σn .
j=n+1

Let m > n → ∞ be such that m/n ∼ 1 + δ (i.e., limm,n→∞ m/n = 1 + δ ) with some
positive δ to be chosen. Since σm is a Cauchy sequence by (10.19), it follows that

m+1
(σm − σn ) → 0, m > n → ∞.
m−n

At the same time, the condition (10.18) implies that

     
m+1
 m+1
m m
 j j M

m−n ∑ 1−
m+1
a j ≤
 m−n ∑ 1− m+1 j
j=n+1 j=n+1
m  
1 1
∼ M(1 + 1/δ ) ∑ −
j=n+1 j m+1
 
m
1 m−n
= M(1 + 1/δ ) ∑ − m+1
j=n+1 j
 m 
1 m−n
≤ M(1 + 1/δ ) dξ −
n ξ m+1
 
m m−n
= M(1 + 1/δ ) log −
n m+1
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 71
 
δ
∼ M(1 + 1/δ ) log(1 + δ ) −
1 + δ + 1/n
∼ M(1 + 1/δ )

× δ − δ 2 /2 + o(δ 2 ) − δ (1 − δ + O(δ 2 ))

= M(1 + 1/δ ) δ 2 /2 + o(δ 2 ) ≤ 2M δ

if δ is chosen small enough. Since δ is arbitrary, we may conclude that also the
second term converges to zero. This proves the theorem. 

Corollary 10.15. Suppose that f ∈ C[−π , π ] is periodic and that its Fourier coeffi-
cients satisfy
M
|cn ( f )| ≤ , n = 0
|n|

with positive constant M that does not depend on n. Then the trigonometric Fourier
series of f converges to f uniformly in x ∈ [−π , π ].

Proof. Since f ∈ C[−π , π ] is periodic, Theorem 5.1 gives the convergence of Fejér
means
lim σN f (x) = f (x)
N→∞

uniformly in x ∈ [−π , π ]. Let us define a0 = c0 ( f ) and

ak (x) = ck ( f )eikx + c−k ( f )e−ikx , k = 1, 2, . . . .

Then
n
sn ≡ ∑ ak (x) = Sn f (x)
k=0

and
1 n 1 n
σn f (x) = ∑
n + 1 k=0
Sk f (x) = ∑ sk .
n + 1 k=0

Thus, we are in the setting of Hardy’s theorem, because

2M
|ak (x)| ≤ , k = 1, 2, . . . .
k

Since this inequality is uniform in x ∈ [−π , π ], on applying Hardy’s theorem we

obtain that
lim sN ≡ lim SN f (x) = f (x)
N→∞ N→∞

uniformly in x ∈ [−π , π ]. 
72 Part I: Fourier Series and the Discrete Fourier Transform

Corollary 10.16. If f ∈ W11 (−π , π ) is periodic, then

lim SN f (x) = f (x)

N→∞

uniformly in x ∈ [−π , π ].

Proof. Since f ∈ W11 (−π , π ) is periodic, there is g ∈ L1 (−π , π ) such that

 x  π
f (x) = g(t)dt + f (−π ), g(t)dt = 0.
−π −π

Thus, f  = g and
cn (g) = incn ( f ),

or equivalently,  
 cn (g)  M
|cn ( f )| =  ≤ , n = 0.
in  |n|

Due to embedding (see Lemma 1.14), the function f is continuous on the interval
[−π , π ]. Using again Hardy’s theorem, we obtain

lim SN f (x) = f (x)

N→∞

uniformly in x ∈ [−π , π ]. This completes the proof. 

Exercise 10.7. Using Theorem 10.9, prove the embedding

W2α (−π , π ) ⊂ W11 (−π , π ), α > 1/2.

Remark 10.17. Corollary 10.16 and Exercise 10.7 show that for every function f
from the spaces H2α and Bα2,θ with 1 ≤ θ < ∞ and α > 1/2, its trigonometric Fourier
series converges to this function uniformly. Here one must take into account that f
might be changed on a set of measure zero.

Let us return to some special trigonometric Fourier series. Namely, we consider

functions f1 (x) and f2 (x) that are defined by the Fourier series
∞
sin(nx)
f1 (x) = ∑ n log(1 + n) (10.21)
n=1

and
∞
cos(nx)
f2 (x) = ∑ n log(1 + n) . (10.22)
n=1
10 Dirichlet Kernel. Pointwise and Uniform Convergence. 73

These functions are well defined for all x ∈ [−π , π ] \ {0}; see Theorem 3.3. In
addition, f1 (0) = 0, whereas f2 (0) is not defined, since the series (10.22) diverges
at zero. We will show that f2 (x) does not belong to W11 (−π , π ) but f1 (x) does.
If we assume to the contrary that f2 ∈ W11 (−π , π ), then its derivative f2 has the
Fourier series
∞
sin(nx)
f2 (x) ∼ − ∑ .
n=1 log(1 + n)

But due to Corollary 10.11 this is not a Fourier series of an L1 function. This con-
tradiction proves that f2 ∈ / W11 (−π , π ).
Concerning the series (10.21), let us prove first that it converges uniformly in
x ∈ [−π , π ], i.e., f1 (x) is (at least) continuous on the interval [−π , π ]. Indeed, by
summation by parts we obtain for 0 ≤ M < N that
 
N N n n−1
sin(nx) 1
∑ n log(1 + n) = ∑ n log(1 + n) ∑ sin(kx) − ∑ sin(kx)
n=M+1 n=M+1 k=1 k=1
sin(Nx) sin(Mx)
= −
N log(1 + N) M log(1 + M)
  
N+1 n
1 1
− ∑ ∑ sin(kx) (n + 1) log(2 + n) − n log(1 + n) . (10.23)
n=M+2 k=1

Using the calculation from Exercise 5.1, we have

n
cos x/2 − cos(n + 1/2)x sin(nx/2) sin((n + 1)x/2)
∑ sin(kx) = 2 sin x/2
=
sin x/2
. (10.24)
k=1

The first two terms on the right-hand side of (10.23) converge to zero as N > M → ∞
uniformly in x ∈ [−π , π ]. The sum on the right-hand side of (10.23) becomes, using
(10.24),

sin(nx/2) sin((n + 1)x/2) 1+n + log(2 + n)

n log 2+n
N+1
− ∑ sin x/2 n(n + 1) log(1 + n) log(2 + n)
n=M+2
N+1
sin(nx/2) sin((n + 1)x/2) log 2+n
=− ∑ sin x/2 (n + 1) log(1 +
1+n
n) log(2 + n)
n=M+2
N+1
sin(nx/2) sin((n + 1)x/2) 1
− ∑ sin x/2 n(n + 1) log(1 + n)
=: I1 + I2 .
n=M+2

Let us consider two cases: n|x| < 1 and n|x| > 1. In the first case,
 
 sin(nx/2) sin((n + 1)x/2)  n|x|/2 · 1 π n
 ≤ = .
 sin x/2  |x|/π 2
74 Part I: Fourier Series and the Discrete Fourier Transform

Then
   
1 1
|x| |x|
π n log (1 + 1/(n + 1)) π n 1n
|I1 | ≤ ∑
2 n=M+2 (n + 1) log(1 + n) log(2 + n)
< ∑
2 n=M+2 n log2 n
π ∞ 1
< ∑
2 n=M+2 n log2 n
→0 (10.25)

as M → ∞ uniformly in x. In the first case, for I2 we have, by integration by parts,

that
   
1 1
|x| |x|
π 2 |x|
n n+1 π 1
|I2 | ≤ ∑ n(n + 1) log(1 + n) 4 ∑ log(1 + n)
2 n=M+2
= |x|
n=M+2
 1/|x|
π dt
< |x|
4 M+2 logt
   1/|x|

π t 1/|x| dt
= |x| +
4 logt M+2 M+2 log t
2
  1/|x| 
π 1/|x| M+2 dt
= |x| − +
4 log(1/|x|) log(M + 2) M+2 log t
2
 
π 1 |x|(M + 2) |x|
≤ + + (1/|x| + M + 2)
4 log(1/|x|) log(M + 2) log2 (M + 2)
 
π 1 1 1 1
< + + +
4 log(M + 2) log(M + 2) log2 (M + 2) log2 (M + 2)
π
≤ →0
log(M + 2)

uniformly in x as M → ∞. In the second case,

 
 sin(nx/2) sin((n + 1)x/2)  1 π
 ≤ ≤ ≤ π n.
 sin x/2  | sin x/2| |x|

Then
N+1
n log (1 + 1/(n + 1)) N+1 n1
|I1 | ≤ π ∑
  (n + 1) log(1 + n) log(2 + n)
<π ∑ n
  n log2 n
1
n= |x| 1
n= |x|
∞
1
 ∑
<π → 0, M→∞
1
n log2 n
n= |x|
≥M

uniformly in x. For I2 we have, by integration by parts,

10 Dirichlet Kernel. Pointwise and Uniform Convergence. 75

N+1  ∞
1 1 1 dt
∑
|I2 | ≤ ≤π  
| sin x/2| 1 +1
2
n log n |x| 1
|x|
≥M t 2 logt
n= |x|
⎛ ⎞
  ∞
1 ⎝ 1 ∞ dt ⎠
=π − + 1
t logt  1 ≥M
 
|x| |x|
2
≥M t 2 log t
|x|
 

1/|x| 1 ∞ dt
=π +  
[1/|x|] log[1/|x|] |x| |x|1 ≥M t 2 log2 t
  
[1/|x|] + 1 [1/|x|] + 1 ∞ dt
<π +
[1/|x|] log[1/|x|] [1/|x|] 2
M t log t
 
1 + 1/M 1 + 1/M
≤π + → 0, M → ∞ (10.26)
log M log M

uniformly in x. Finally, we may conclude that the trigonometric Fourier series

(10.21) converges uniformly on the interval [−π , π ], and therefore it defines a con-
tinuous function f1 (x). This series, as well as (10.22), can be differentiated term by
term for π ≥ |x| ≥ δ > 0, because the series
∞
cos(nx)
∑ log(1 + n)
n=1

converges uniformly (see Corollary 3.4) for π ≥ |x| ≥ δ > 0. This means that for
this interval, f1 (x) belongs to C1 . Thus, it remains to investigate the behavior of
the series (10.21) as x → 0+. But the estimates (10.25)–(10.26) show us that (if
we choose M  1/x, x → 0+) the function f1 (x) from (10.21) has the asymptotic
behavior
C
f1 (x) ∼ . (10.27)
log x

It is possible to prove (see [46, Chapter V, formula (2.19)]) that the asymptotic
(10.27) can be differentiated, and we obtain

C
f1 (x) ∼ − .
x log2 x

This singularity is integrable at zero. Therefore, the function f1 (x) belongs to

W11 (−π , π ).
Exercise 10.8. Prove that the series (10.21) and (10.22) do not converge absolutely.
Chapter 11
Formulation of the Discrete Fourier
Transform and Its Properties.

Let x(t) be a 2π -periodic continuous signal. Assume that x(t) can be represented by
an absolutely convergent trigonometric Fourier series
∞
x(t) = ∑ cm eimt , t ∈ [−π , π ], (11.1)
m=−∞

where the cm are the Fourier coefficients of x(t).

Let now N be an even positive integer and

2π k N N
tk = , k = − , . . . , − 1.
N 2 2

Then x(tk ) is a response at tk , i.e.,

∞ 2π km
x(tk ) = ∑ cm ei N . (11.2)
m=−∞

Since ei2π kl = 1 for integers k and l, the series (11.2) can be rewritten as
∞ 2π k (m−lN)
∞ 2π k (m−lN)
∑ cm ei N = ∑ ∑ cm ei N
m=−∞ l=−∞ −N/2≤m−lN≤N/2−1
∞ N/2−1
2π k n
= ∑ ∑ cn+lN ei N

l=−∞ n=−N/2
N/2−1
2π k n
∞ N/2−1
2π k n
= ∑ ei N
∑ cn+lN = ∑ ei N Xn , (11.3)
n=−N/2 l=−∞ n=−N/2

where Xn , n = −N/2, . . . , N/2 − 1 is given by

c Springer International Publishing AG 2017 77

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 11
78 Part I: Fourier Series and the Discrete Fourier Transform

∞
Xn = ∑ cn+lN . (11.4)
l=−∞

In the calculation (11.3) we have used the fact that the series (11.1) converges ab-
solutely. Combining (11.2) and (11.3), we obtain

N/2−1
2π kn
xk := x(tk ) = ∑ Xn ei N . (11.5)
n=−N/2

The formula (11.5) can be viewed as an inverse discrete Fourier transform, and it
appeared quite naturally in the discretization of a continuous periodic signal. More-
over, the formula (11.4) becomes the main property of this approach. Since

N/2−1
i(k−m) 2Nπ n 0, k − m = 0, ±N, ±2N, . . .
∑ e =
N, k − m = 0, ±N, ±2N, . . .
(11.6)
n=−N/2

we solve the linear system (11.5) with respect to Xn , n = −N/2, . . . , N/2 − 1, and
obtain
1 N/2−1 2π kn
Xn = ∑ xk e−i N .
N k=−N/2
(11.7)

Exercise 11.1. Prove (11.6) and (11.7).

In fact, the formulas (11.5) and (11.7) give us the inverse and direct discrete Fourier
transforms, respectively.
N/2−1
Definition 11.1. The sequence {Xn }n=−N/2 of complex numbers is called the dis-
N/2−1
crete Fourier transform (DFT) of the sequence {Yk }k=−N/2 if for each n = −N/2, . . . ,
N/2 − 1 we have

1 N/2−1 2π kn
Xn = ∑ Yk e−i N .
N k=−N/2
(11.8)

We use the symbol F for the DFT and write

Xn = F (Yk )n

or simply X = F (Y ).
N/2−1
Definition 11.2. The sequence {Zk }k=−N/2 of complex numbers is said to be the
N/2−1
inverse discrete Fourier transform (IDFT) of the sequence {Xn }n=−N/2 if for each
k = −N/2, . . . , N/2 − 1 we have
11 Formulation of the Discrete Fourier Transform and Its Properties. 79

N/2−1
2π kn
Zk = ∑ Xn ei N . (11.9)
n=−N/2

We use the symbol F −1 for the IDFT and write

Zk = F −1 (Xn )k

or simply Z = F −1 (X).

The properties of the DFT and IDFT are collected in the following lemmas.
Lemma 11.3. The following equalities hold:
(1) F −1 (F (Y )) = Y ;
(2) F (F −1 (X)) = X;
(3)
N/2−1
1 N/2−1
∑ F (Xn )k F (Yn )k = ∑ XnYn .
N n=−N/2
k=−N/2

Proof. Using (11.6), (11.8), and (11.9), we have

 
N/2−1
i 2πNkn 1 N/2−1 N/2−1
−i 2πNln 2π kn
F −1
(F (Y ))k = ∑ F (Yl )n e = ∑
N n=−N/2 ∑ Yl e ei N

n=−N/2 l=−N/2
 
N/2−1 N/2−1
1 2π n(k−l) 1
= ∑ Yl
N l=−N/2 ∑ ei N =
N
Yk N = Yk .
n=−N/2

This proves part (1). Part (2) can be proved in the same manner. 

Exercise 11.2. Prove part (3) of Lemma 11.3.

Corollary 11.4 (Parseval’s equality).

1 N/2−1 N/2−1
∑
N n=−N/2
|Xn |2 = ∑ |F (Xn )k |2 .
k=−N/2

Remark 11.5. Due to the periodicity of the complex exponential, we may extend
the values of Xm , m = −N/2, . . . , N/2 − 1 periodically to any integer by

Xm+lN = Xm , l = 0, ±1, ±2, . . . . (11.10)

N/2−1
Corollary 11.6. For a sequence X = {Xn }n=−N/2 we define

N/2−1
Xrev = {XN−n }n=−N/2 .
80 Part I: Fourier Series and the Discrete Fourier Transform

Then
F −1 (X) = NF (Xrev ).

Proof. By Definition 11.2 and the periodicity condition (11.10) we have

N/2−1 N/2−1 −N/2+1

2π kn 2π kn 2π kn
NF (Xrev )k = ∑ XN−n e−i N = ∑ X−n e−i N = ∑ Xn ei N

n=−N/2 n=−N/2 n=N/2

N/2 N/2−1
2π kn 2π kn
= ∑ Xn ei N = ∑ Xn ei N + XN/2 eiπ k − X−N/2 e−iπ k
n=−N/2+1 n=−N/2
N/2−1
2π kn
= ∑ Xn ei N = F −1 (X)k .
n=−N/2

This concludes the proof. 

N/2−1
Definition 11.7. The convolution of discrete sequences X = {Xn }n=−N/2 and Y =
N/2−1
{Yn }n=−N/2 is defined as the sequence whose elements are given by

N/2−1
(X ∗Y )k = ∑ Xl Yk−l , (11.11)
l=−N/2

where Xn and Yn satisfy the periodicity condition (11.10).

Proposition 11.8. For every integer l, it is true that

N/2−l−1 N/2−1
2π nm 2π nm
∑ Ym e−i N = ∑ Ym e−i N .
m=−N/2−l m=−N/2

Proof. The claim is trivial for l = 0. If l > 0, then

N/2−l−1 N/2−1 −N/2−1 N/2−1

2π nm 2π nm 2π nm 2π nm
∑ Ym e−i N = ∑ Ym e−i N + ∑ Ym e−i N − ∑ Ym e−i N

m=−N/2−l m=−N/2 m=−N/2−l m=N/2−l

N/2−1 N/2−1
2π nm 2π n (m−N)
= ∑ Ym e−i N + ∑ Ym−N e−i N

m=−N/2 m=N/2−l
N/2−1 N/2−1
2π nm 2π nm
− ∑ Ym e−i N = ∑ Ym e−i N

m=N/2−l m=−N/2
11 Formulation of the Discrete Fourier Transform and Its Properties. 81

due to the periodicity condition (11.10). If l < 0, then

N/2−l−1 N/2−1 N/2−l−1 −N/2−l−1

2π nm 2π nm 2π nm 2π nm
∑ Ym e−i N = ∑ Ym e−i N + ∑ Ym e−i N − ∑ Ym e−i N

m=−N/2−l m=−N/2 m=N/2 m=−N/2

N/2−1 N/2−l−1
2π nm 2π nm
= ∑ Ym e−i N + ∑ Ym e−i N

m=−N/2 m=N/2
N/2−l−1 N/2−1
2π n (m−N) 2π nm
− ∑ Ym−N e−i N = ∑ Ym e−i N

m=N/2 m=−N/2

due to periodicity condition (11.10). This proves the proposition. 

Corollary 11.9. For every integer l it is true that

N/2−l−1 N/2−1
∑ Ym = ∑ Ym .
m=−N/2−l m=−N/2

Lemma 11.10. The convolution (11.11) is symmetric, i.e.,

(X ∗Y )k = (Y ∗ X)k

for every k = −N/2, . . . , N/2 − 1.

Proof. We have

N/2−1 k+1−N/2
(X ∗Y )k = ∑ Xl Yk−l = ∑ Y j Xk− j
l=−N/2 j=k+N/2
N/2−1+(k+1) N/2−1
= ∑ Y j Xk− j = ∑ Y j Xk− j = (Y ∗ X)k
j=−N/2+(k+1) j=−N/2

by Corollary 11.9. 

Lemma 11.11. For each n = −N/2, . . . , N/2 − 1 it is true that

(1) F (X ∗Y )n = NF (X)n F (Y )n ;
(2) F −1 (X ∗Y )n = F −1 (X)n F −1 (Y )n .

Proof. Using (11.11), we have

1 N/2−1 2π kn 1 N/2−1 N/2−1

2π kn
F (X ∗Y )n = ∑
N k=−N/2
(X ∗Y )k e−i N = ∑ Xl ∑ Yk−l e−i N
N l=−N/2 k=−N/2
82 Part I: Fourier Series and the Discrete Fourier Transform

1 N/2−1 N/2−l−1
2π n
= ∑ Xl ∑
N l=−N/2 m=−N/2−l
Ym e−i N (m+l)

1 N/2−1 2π nl
N/2−l−1
2π nm
= ∑
N l=−N/2
Xl e−i N ∑ Ym e−i N . (11.12)
m=−N/2−l

Proposition 11.8 allows us to rewrite (11.12) as

1 N/2−1 2π nl
N/2−1
2π nm
F (X ∗Y )n = ∑
N l=−N/2
Xl e−i N ∑ Ym e−i N = NF (X)n F (Y )n .
m=−N/2

Part (2) is proved in a similar manner. 

Corollary 11.12. For each n = −N/2, . . . , N/2 − 1 it is true that

1  −1 
F −1 (X ·Y )n = F (X) ∗ F −1 (Y ) n ,
N
F (X ·Y )n = (F (X) ∗ F (Y ))n ,

N/2−1
where X ·Y denotes the sequence {Xk ·Yk }k=−N/2 .

Proof. Lemmas 11.3 and 11.11 imply that

   
X ∗ Y = F −1 F X ∗ Y  · F (Y ) .
= NF −1 F (X)
n n n

Setting X := F −1 (X) and Y := F −1 (Y ), we obtain easily from the latter equality

that  
NF −1 (X ·Y )n = F −1 (X) ∗ F −1 (Y ) n .

The second part is proved in a similar manner. 

Let us return to the continuous signal x(t), t ∈ [−π , π ], which is represented by an

absolutely convergent trigonometric Fourier series (11.1). Formula (11.4) allows us
to obtain

N/2−1 N/2−1
∞
N/2−1

∑ |Xn − cn | = ∑

∑ cn+lN − cn

= ∑

∑ cn+lN

n=−N/2 n=−N/2 l=−∞ n=−N/2 l=0

N/2−1
≤ ∑ ∑ |cn+lN | ≤ ∑ |cν |. (11.13)
n=−N/2 l=0 |ν |≥N/2

Similarly, we have
11 Formulation of the Discrete Fourier Transform and Its Properties. 83

N/2−1

N/2−1

−1 i 2πNkn

i 2πNkn

F (Xn )k − ∑ cn e
=
∑ (Xn − cn )e

n=−N/2


n=−N/2

N/2−1
≤ ∑ |Xn − cn | ≤ ∑ |cν |. (11.14)
n=−N/2 |ν |≥N/2

In the formulas (11.13) and (11.14) the numbers cn are the Fourier coefficients of
N/2−1 N/2−1
the signal x(t), and {Xn }n=−N/2 is the DFT of {x(tk )}k=−N/2 with tk = 2π k/N.

Theorem 11.13. If x(t) is periodic and belongs to the Sobolev space W2m (−π , π )
for some m = 1, 2, . . ., then
 
1
Xn = cn + o , N→∞ (11.15)
N m−1/2

and
N/2−1  
2π kn 1
F −1 (Xn )k = ∑ cn ei N +o
N m−1/2
, N→∞ (11.16)
n=−N/2

uniformly in n and k from the set {−N/2, . . . , N/2 − 1}.

Proof. Using Hölder’s inequality, we have

 1/2  1/2
∑ |cν | ≤ ∑ |ν |2m |cν |2 ∑ |ν |−2m .
|ν |≥N/2 |ν |≥N/2 |ν |≥N/2

The first sum on the right-hand side tends to zero as N → ∞ due to Parseval’s equality
for a function from the Sobolev space W2m (−π , π ). The second sum can be estimated
precisely. Namely, since for every m = 1, 2, . . . we have
 1/2  1/2
∞
∑ |ν | −2m

N/2
t −2m
dt  N −m+1/2 ,
|ν |≥N/2

we conclude that (11.15) and (11.16) follow from the last estimate and (11.13) and
(11.14), respectively. 

Corollary 11.14. An unknown periodic function x(t) ∈ W2m (−π , π ), m = 1, 2, . . .,

can be recovered from its IDFT as
   
2π k −1 1
x =F (Xn )k + o , N→∞
N N m−1/2

uniformly in k from the set {−N/2, . . . , N/2 − 1}.

84 Part I: Fourier Series and the Discrete Fourier Transform

Exercise 11.3. (1) Show that

⎧ 2π n
 ⎨1, a = ei N ,
k N/2−1
F {a }k=−N/2 = 1 −N/2 −aN/2 2π n
n ⎩ N (−1)n a −i 2π n , a= ei N .
1−ae N

N/2−1
(2) If a sequence Y = {Yk }k=−N/2 is real, then show that

F (Y )n = F (Y )N−n .
Chapter 12
Connection Between the Discrete Fourier
Transform and the Fourier Transform.

If a function f (x) is integrable over the whole line, i.e.,

 ∞
| f (x)|dx < ∞,
−∞

then its Fourier transform is defined as

 ∞
1
F f (ξ ) = f(ξ ) := √ f (x)e−ixξ dx.
2π −∞

Similarly, the inverse Fourier transform of an integrable function g(ξ ) is defined as

 ∞
1
F −1 g(x) := √ g(ξ )eixξ dξ .
2π −∞

Theorem 12.1 (Riemann–Lebesgue lemma). For an integrable function f (x), its

Fourier transform F f (ξ ) is continuous, and

lim F f (ξ ) = 0.
ξ →±∞

Proof. Since eiπ = −1, we have

 ∞  ∞
1 1
f(ξ ) = − √ f (x)e−ixξ +iπ dx = − √ f (y + π /ξ )e−iξ y dy.
2π −∞ 2π −∞

This fact implies that

 ∞
1
−2 f(ξ ) = √ ( f (x + π /ξ ) − f (x))e−iξ x dx.
2π −∞

c Springer International Publishing AG 2017 85

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 12
86 Part I: Fourier Series and the Discrete Fourier Transform

Hence  ∞
1
2| f(ξ )| ≤ √ | f (x + π /ξ ) − f (x)|dx → 0
2π −∞

as |ξ | → ∞, since f is integrable on the whole line. This is a well known property of

integrable functions. Continuity of f(ξ ) follows from the representation
 ∞
1
f(ξ + h) − f(ξ ) = √ f (x)e−ixξ (e−ixh − 1)dx
2π −∞

and its consequence

| f(ξ + h) − f(ξ )|
 
1 2
≤√ | f (x)||e−ixh − 1|dx + √ | f (x)|dx =: I1 + I2 .
2π |xh|<δ 2π |xh|>δ

For the first term I1 we have the estimate

  ∞
1 δ
I1 ≤ √ | f (x)||xh|dx < √ | f (x)|dx → 0, δ → 0.
2π |xh|<δ 2π −∞

For the second term I2 we have


2
I2 = √ | f (x)|dx → 0
2π |x|>δ /|h|

as |h| → 0. If we choose δ = |h|1/2 , then both I1 and I2 tend to zero as |h| → 0. This
completes the proof. 

If a function f (x) has integrable derivatives f (k) (x) of order k = 0, 1, 2, . . . , m, then

we say that f belongs to the Sobolev space W1m (R).
Exercise 12.1. Prove that if f ∈ W11 (R), then limx→±∞ f (x) = 0.

Theorem 12.2 (Fourier inversion formula). Suppose that f belongs to W11 (R).
Then
F −1 (F f )(x) = f (x)

at every point x ∈ R.

Proof. First we prove that

 ∞  ∞
f (x)
g(x)dx = f(ξ )g(ξ )dξ
−∞ −∞

for every pair of integrable functions f and g. Indeed,

12 Connection Between the Discrete Fourier Transform and the Fourier Transform. 87
 ∞ ∞  ∞ 
1
g(x)dx = √
f (x) f (x) g(ξ )e−ixξ dξ dx
−∞ 2π −∞ −∞
 ∞  ∞  ∞
1
=√ g(ξ ) f (x)e−ixξ dxdξ = f(ξ )g(ξ )dξ
2π −∞ −∞ −∞

by Fubini’s theorem. Suppose now that g(ξ ) is given by


1, |ξ | < n,
g(ξ ) =
0, |ξ | > n.

Its Fourier transform is equal to

 n   
1 −ixξ 1 e−ixn eixn 2 sin(nx)
g(x) = √ e dξ = √ − = .
2π −n 2π −ix −ix π x

Thus, we have the equality

   n
2 ∞ sin(nx)
f (x) dx = f(ξ )dξ ,
π −∞ x −n

where f ∈ W11 (R). Letting n → ∞, we obtain

 ∞  
2 ∞ sin(nx)
p. v. f(ξ )dξ = lim f (x) dx. (12.1)
−∞ n→∞ π −∞ x
√
We will prove that the limit in (12.1) is actually equal to 2π f (0). Since
 ∞
sin(nx)
dx = π ,
−∞ x

the limit in (12.1) can be rewritten as

 
2 ∞ sin(nx)
lim f (x) dx
n→∞ π −∞ x
 
√ 2 ∞ sin(nx)
= 2π f (0) + lim ( f (x) − f (0)) dx
n→∞ π −∞ x
 
√ 2 ∞ sin(t)
= 2π f (0) + lim ( f (t/n) − f (0)) dt.
n→∞ π −∞ t

It remains to show that the latter limit is equal to zero. In order to prove this fact we
split the above integral as follows:
88 Part I: Fourier Series and the Discrete Fourier Transform
 ∞
sin(t)
( f (t/n) − f (0)) dt
−∞ t
 
sin(t) sin(t)
= ( f (t/n) − f (0)) dt + ( f (t/n) − f (0)) dt =: I1 + I2 .
|t|<1 t |t|>1 t

Since f ∈ W11 (R), it is continuous with respect to the definition of Sobolev spaces
in this chapter. Therefore,
  
 sin(t) 
|I1 | ≤ sup | f (t/n) − f (0)|  
|t|<1
 t  dt → 0, n → ∞.
|t|<1

For the second term, I2 , we first change variables to obtain


sin(nz)
I2 = ( f (z) − f (0)) dz
|z|>1/n z
 ∞  
z
sin(ny)
= 1 ( f (z) − f (0)) dy dz
n 0 y
 −1  z 
n sin(ny)
+ ( f (z) − f (0)) dy dz.
−∞ 0 y

Integration by parts in these integrals leads to

 z ∞  ∞  z

sin(ny) sin(ny)
I2 = ( f (z) − f (0)) dy − f  (z) dydz
0 y 1/n 1/n 0 y
 z   −1/n  z
sin(ny) −1/n sin(ny)
+ ( f (z) − f (0)) dy − f  (z) dydz
0 y −∞ −∞ 0 y
  ∞  1/n
sin(ny) sin(ny)
= lim f (z) − f (0) dy − ( f (1/n) − f (0)) dy
z→∞ 0 y 0 y
 ∞  nz −1/n sin(ny) 
sin(t)
− f  (z) dtdz + ( f (−1/n) − f (0)) dy
1/n 0 t 0 y
   −∞  −1/n  nz
sin(ny) sin(t)
− lim f (z) − f (0) dy − f  (z) dtdz.
z→−∞ 0 y −∞ 0 t

Since limz→±∞ f (z) = 0 (see Exercise 12.1) and since f is continuous, we obtain (as
n → ∞)
 ∞  nz
π sin(t)
I2 → − f (0) − lim f  (z) dtdz
2 n→∞ 1/n 0 t
 −1/n  nz
π sin(t)
− f (0) − lim f  (z) dtdz
2 n→∞ −∞ 0 t
12 Connection Between the Discrete Fourier Transform and the Fourier Transform. 89
 ∞  nz
sin(t)
= −π f (0) − lim f  (z) dtdz
n→∞ 1/n 0 t
 −1/n  nz
sin(t)
− lim f  (z) dtdz
n→∞ −∞ 0 t
 ∞ 
π 0 π
= −π f (0) − f  (z) dz + f  (z) dz
0 2 −∞ 2
π π
= −π f (0) + f (0) + f (0) = 0.
2 2

Here we used again the fact that limz→±∞ f (z) = 0 and Lebesgue’s dominated con-
vergence theorem. Thus, (12.1) transforms to
 ∞ √
p. v. f(ξ )dξ = 2π f (0),
−∞

or equivalently,
F −1 (F f )(0) = f (0).

In order to prove the Fourier inversion formula for every x ∈ R, let us note that
∞ 
 1
fx (y)(ξ ) = f 
(x + y)(ξ ) = √ f (x + y)e−iyξ dy
2π −∞
 ∞
1
=√ f (z)e−izξ eixξ dz = eixξ f(ξ ).
2π −∞

Since fx (0) = f (x), it follows that

f (x) = fx (0) = F −1 (F fx )(0)

 ∞  ∞
1 1
=√ (F fx )(ξ )dξ = √ eixξ f(ξ )dξ = F −1 (F f )(x).
2π −∞ 2π −∞

This completes the proof. 

Remark 12.3. As a by-product of the above proof we record the limit
 ∞
1 sin(nx)
lim f (x) dx = f (0)
n→∞ π −∞ x

for every f ∈ W11 (R).

Lemma 12.4. If f belongs to the Sobolev space W1m (R) for some m = 1, 2, . . ., then
 
1
f(ξ ) = o (12.2)
|ξ |m

as |ξ | → ∞.
90 Part I: Fourier Series and the Discrete Fourier Transform

Proof. Since f ∈ W1m (R), it follows that f  , f  , . . . , f (m−1) ∈ W11 (R). By Exercise
12.1, we have
lim f (k) (x) = 0
x→±∞

for each k = 0, 1, . . . , m − 1. This fact and repeated integration by parts give us

∞
 ∞   
−ixξ e−ixξ  1 ∞  1 ∞ 
f (x)e dx = f (x) + f (x)e−ixξ dx = f (x)e−ixξ dx
−∞ −iξ  iξ −∞ iξ −∞
−∞
∞
−ix ξ   ∞
e  1
=− f 
(x)  + f  (x)e−ixξ dx
(iξ )2  (iξ )2 −∞
−∞
 ∞  ∞  
1  −ixξ 1 (m) −ixξ 1
= f (x)e dx = · · · = f (x)e dx = o
(iξ )2 −∞ (iξ )m −∞ |ξ |m

due to the Riemann–Lebesgue lemma. 

The equality (12.2) allows us to consider (with respect to the accuracy of calcu-
lations) the Fourier transform only on the interval (−R, R) with R > 0 sufficiently
large, i.e., we may neglect the values of F f (ξ ) for |ξ | > R. This simplification
justifies the following approximation of the inverse Fourier transform:
 R
1
f ∗ (x) := √ F f (ξ )eixξ dξ .
2π −R

At the same time and without loss of generality we may assume that the function
f (x) is equal to zero outside some finite interval. In that case it can be proved that
F f (ξ ) is a smooth function for which (12.2) holds.
◦
1 (−R, R) if f ∈ W1 (R) and f ≡ 0 for x ∈
Definition 12.5. We say that f ∈ W m m /
(−R, R).
◦
Theorem 12.6. Suppose that f ∈ W m 1 (−R, R) is supported in a fixed interval [a, b] ⊂
(−R, R) with R > 0 sufficiently large for some m = 2, 3, . . .. Then
 N/2−1   x(2n+1)
2 1 2n + 1 i
f (x) = ∑ Ff
π N m/(m+2) n=−N/2 N m/(m+2)
e N m/(m+2)
 
1
+O (12.3)
N (2m−2)/(m+2)

uniformly in x ∈ (−R, R) for R = N 2/(m+2) and even N → ∞.

12 Connection Between the Discrete Fourier Transform and the Fourier Transform. 91

Proof. Since f (x) = 0 for x ∈

/ (−R, R), using the Fourier inversion formula (see
Theorem 12.2) we have
∞  R
1 1
f (x) − f ∗ (x) = √ F f (ξ )eixξ dξ − √ F f (ξ )eixξ dξ
2π −∞ 2π −R

1
=√ F f (ξ )eixξ dξ .
2π |ξ |>R

Lemma 12.4 implies then that

 
∗ 1
f (x) − f (x) = o . (12.4)
Rm−1

Let us divide the interval [−R, R] into N + 1 subintervals [ξn , ξn+1 ], where n =
−N/2, . . . , N/2 − 1 such that

−R = ξ−N/2 < ξ−N/2+1 < · · · < ξN/2 = R,

where
2Rn 2R
ξn = , ξn+1 − ξn = .
N N
Let us also set
ξn + ξn+1 R
ξn∗ = = (2n + 1).
2 N
Then we obtain
 R
1
f ∗ (x) = √ F f (ξ )eixξ dξ
2π −R
N/2−1
1 ∗ 2R
=√
2π
∑ F f (ξn∗ )eixξn
N
n=−N/2

1 N/2−1 ξn+1 ∗

+√ ∑
2π n=−N/2 ξn
F f (ξ )eixξ − F f (ξn∗ )eixξn dξ

  3
2 R N/2−1 R
= ∑ F f (R(2n + 1)/N)e
π N n=−N/2
ixR(2n+1)/N
+O
N2
. (12.5)
92 Part I: Fourier Series and the Discrete Fourier Transform

Exercise 12.2. Prove that

N/2−1  ξn+1

1 ∗
√
2π
∑ F f (ξ )eixξ − F f (ξn∗ )eixξn dξ
n=−N/2 ξn
⎧ 3

⎨O R 2 , supp f = [a, b]
= N

⎩O R52 , supp f ⊂ (−R, R)

N

uniformly in x ∈ supp f .

Hint. Use the Taylor expansion for the smooth function F f (ξ )eixξ at the point ξn∗ .
If we combine (12.5) with (12.4) and choose R = N 2/(m+2) , then we obtain (12.3).
This completes the proof. 

Remark 12.7. The main part of (12.3) represents some kind of inverse discrete
Fourier transform. In order to reconstruct f at a point x ∈ [−R, R] we need to know
only the Fourier transform of this unknown function at the points

2n + 1
, n = −N/2, . . . , N/2 − 1,
N m/(m+2)
where m is the smoothness index of f . What is more, the formula (12.3) shows us
that it is effective if
2m − 2 m
>
m+2 m+2
◦
or m > 2. This means that f must belong to the Sobolev space W m
1 (−R, R) with
some m ≥ 3.
Chapter 13
Some Applications of the Discrete Fourier
Transform.

First we prove the Poisson summation formula.

Definition 13.1. Let f be a function such that

lim
N→∞
∑ f (x + 2π n)
|n|≤N

exists pointwise in x ∈ R. Then

∞
fp (x) := ∑ f (x + 2π n) (13.1)
n=−∞

is called the periodization of f .

Remark 13.2. It is clear that fp (x) is periodic with period 2π . Hence we will con-
sider it only on the interval [−π , π ].

Theorem 13.3 (Poisson summation formula). Suppose that f ∈ L1 (R). Then

fp (x) from (13.1) is finite almost everywhere, satisfies fp (x + 2π ) = fp (x) almost
everywhere, and is integrable on the interval [−π , π ]. The Fourier coefficients of
fp (x) are given by
 π
1 1
fp (x)e−imx dx = √ F f (m).
2π −π 2π

If in addition
∞
∑ |F f (m)| < ∞,
m=−∞

c Springer International Publishing AG 2017 93

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 13
94 Part I: Fourier Series and the Discrete Fourier Transform

then
∞ ∞
1
∑ f (x + 2π n) = √2π ∑ F f (m)eimx . (13.2)
n=−∞ m=−∞

In particular, fp (x) is continuous, and we have the Poisson identity

∞ ∞
1
∑ f (2π n) = √
2π
∑ F f (m). (13.3)
n=−∞ m=−∞

Proof. Since f ∈ L1 (R), we have

 π  π ∞ ∞  π

−π
| fp (x)|dx ≤ ∑
−π n=−∞
| f (x + 2π n)|dx = ∑ | f (x + 2π n)|dx
n=−∞ −π
∞  π +2π n  ∞
= ∑ | f (t)|dt = | f (t)|dt < ∞.
n=−∞ −π +2π n −∞

This shows that fp is finite almost everywhere and integrable on [−π , π ]. Applying
the same calculation to fp (x)e−imx allows us to integrate term by term to obtain
 π ∞  π
1 1
cm ( fp ) =
2π −π
fp (x)e−imx dx = ∑ 2π −π
f (x + 2π n)e−imx dx
n=−∞
∞  π +2π n
1
= ∑ f (t)e−im(t−2π n) dt
n=−∞ 2π −π +2π n
∞ π +2π n 
1 1
=√ ∑ √
2π n=−∞ 2π −π +2π n
f (t)e−imt dt
 ∞
1 1 1
= √ √ f (t)e−imt dt = √ F f (m).
2π 2π −∞ 2π

Now, if the series

∞
∑ |F f (m)|
m=−∞

converges, then that convergence is equivalent to the fact that

∞
∑ |cm ( fp )| < ∞.
m=−∞

Thus, fp (x) can be represented by its Fourier series at least almost everywhere (and
we can redefine fp (x) so that this representation holds pointwise), i.e.,

∞
fp (x) = ∑ cm ( fp )eimx .
m=−∞
13 Some Applications of the Discrete Fourier Transform. 95

This means that (see (13.1))

∞ ∞
1
∑ f (x + 2π n) = √
2π
∑ F f (m)eimx .
n=−∞ m=−∞

Finally, set x = 0 to obtain the Poisson identity (13.3). 

Example 13.4. If
1 − x2
f (x) = √ e 4t , x ∈ R,
4π t

where t > 0 is a parameter, then it is very well known (see, e.g., Example 16.7 and
Exercise 16.4) that
1
F f (ξ ) = √ e−t ξ .
2

2π

Formula (13.2) transforms in this case to

∞ (x+2π n)2 ∞
1 1
∑ e− 4t = ∑ e−tm eimx ,
2
√
4π t n=−∞ 2π m=−∞

and the Poisson identity transforms to

 ∞ ∞
π
∑ e−π ∑
2 n2 /t
e−tm .
2
=
t n=−∞ m=−∞

As an application of the Poisson summation formula we consider the problem of

reconstructing a band-limited signal from its values on the integers.
Definition 13.5. A signal f (t) is called band-limited if it has a representation
 2πλ
1
f (t) = √ F(ξ )eit ξ dξ , (13.4)
2π −2πλ

where λ is a positive parameter and F is some integrable function.

Remark 13.6. If we set F(ξ ) = 0 for |ξ | > 2πλ , then (13.4) is the inverse Fourier
transform of F ∈ L1 (R). In that case f is bounded and continuous.

Theorem 13.7 (Whittaker, Shannon, Boas). Suppose that F ∈ L1 (R) and F(ξ ) =
0 for |ξ | > 2πλ . If λ ≤ 1/2, then for every t ∈ R we have
∞
sin π (t − n)
f (t) = ∑ f (n)
π (t − n)
, (13.5)
n=−∞
96 Part I: Fourier Series and the Discrete Fourier Transform

where the fraction is equal to 1 when t = n. If λ > 1/2, we have

  
 

∞
sin π (t − n)  2


f (t) − ∑ f (n)
π (t − n) 
 ≤
π π <|ξ |<2πλ
|F(ξ )|dξ .
n=−∞

Proof. Let
∞
Fp (ξ ) = ∑ F(ξ + 2π n)
n=−∞

be the periodization of F. Formula (13.4) shows that

F (F)(t) = f (−t),

where F (F) denotes the Fourier transform of F(ξ ). Hence, by the Poisson summa-
tion formula, we have (see (13.2))
∞ ∞ ∞
1 1
∑ F(ξ + 2π n) = √
2π
∑ f (−m)eimξ = √
2π
∑ f (m)e−imξ .
n=−∞ m=−∞ m=−∞

Since every trigonometric Fourier series can be integrated term by term, we obtain
 π  π ∞

−π
Fp (ξ )eit ξ dξ = ∑
−π n=−∞
F(ξ + 2π n)eit ξ dξ
∞  π
1
=√
2π
∑ f (m)
−π
ei(t−m)ξ dξ
m=−∞
∞
1 ei(t−m)π − e−i(t−m)π
=√
2π
∑ f (m)
i(t − m)
m=−∞
∞
2π sin π (t − m)
=√
2π
∑ f (m)
π (t − m)
.
m=−∞

Now, if λ ≤ 1/2, then F(ξ ) for |ξ | ≤ π is equal to its periodization Fp (ξ ) (see

Definition 13.1) and  π √
Fp (ξ )eit ξ dξ = 2π f (t).
−π

These equalities imply immediately that

∞
sin π (t − n)
f (t) = ∑ f (n)
π (t − n)
,
n=−∞

so that (13.5) is proved. If λ > 1/2, then we cannot expect that F(ξ ) = Fp (ξ ) for
|ξ | > π , but using Definition 13.1 we have
13 Some Applications of the Discrete Fourier Transform. 97

 π  
1 2
it ξ
f (t) = √ F(ξ )e dξ + F(ξ )eit ξ dξ
2π
−π π π <|ξ |<2πλ
 
∞
sin π (t − n) 2
= ∑ f(n) + F(ξ )eit ξ dξ ,
n=−∞ π (t − n) π π <| ξ |<2 πλ

where  π
1
f(n) = √ F(ξ )einξ dξ .
2π −π

Therefore,
∞
sin π (t − n)
f (t) = ∑ f (n)
π (t − n)
n=−∞
∞   sin π (t − n) 
1
+ ∑ f(n) − f (n) +√ F(ξ )eit ξ dξ .
n=−∞ π (t − n) 2π π <|ξ |<2πλ

Here the middle series is equal to

∞ 
1 sin π (t − n)
−√
2π
∑ π <|ξ |<2πλ
F(ξ )e inξ
dξ
π (t − n)
,
n=−∞

or



∞
1 sin π (t − n) inξ
−√ F(ξ ) ∑ e dξ .
2π π <|ξ |<2πλ n=−∞ π (t − n)

Exercise 13.1. Prove that

∞
sin π (t − n) inξ
∑ e = eit ξ , ξ ∈ [−π , π ].
n=−∞ π (t − n)

Hint. Show that   sin π (t − n)

cn eit ξ = .
π (t − n)

Using Exercise 13.1 and the periodicity of einξ , we have

  
 

∞
sin π (t − n)  2
 f (t) − ∑ f (n) ≤ |F(ξ )|dξ .
 n=−∞ π (t − n)  π π <|ξ |<2πλ

This proves the theorem. 

98 Part I: Fourier Series and the Discrete Fourier Transform

Theorem 13.7 shows that in order to reconstruct a band-limited signal f (t), it is

enough to know the values of the signal at integers n. In turn, to evaluate f (n) it is
enough to use the IDFT of F(ξ ); see (13.4). Indeed, let us assume without loss of
generality that λ = 1/2. Then
 π
1
f (n) = √ F(ξ )einξ dξ .
2π −π

If F is smooth enough (say F ∈ C2 [−π , π ]), then formula (12.5) gives (R = π and
N  1)
√ N/2−1 
2π π n(2k+1) 1
f (n) =
N ∑ Fk ei N +O
N2
k=−N/2
√ 
2π i π n N/2−1 2π kn 1
= e N ∑ Fk ei N + O ,
N k=−N/2
N2

where Fk denotes the value of F(ξ ) at the point π (2k + 1)/N. Therefore, up to the
accuracy of calculations,
√
2π i π n −1
f (n) ≈ e N F (Fk )n ,
N
i.e., for N a sufficiently large even integer,
√ ∞
2π i π n −1 sin π (t − n)
f (t) ≈ ∑ e N F (Fk )n .
n=−∞ N π (t − n)
Chapter 14
Applications to Solving Some Model
Equations

14.1 The One-Dimensional Heat Equation

Let us consider a heat conduction problem for a straight bar of uniform cross sec-
tion and homogeneous material. Let x = 0 and x = L denote the ends of the bar
(the x-axis is chosen to lie along the axis of the bar). Suppose that no heat passes
through the sides of the bar. We also assume that the cross-sectional dimensions
are so small that the temperature u can be considered constant on any given cross
section (Figure 14.1).

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

Fig. 14.1 Geometry of the heat conduction problem for a bar.

Then u is a function only of the coordinate x and the time t. The variation of
temperature in the bar is governed by the partial differential equation

α 2 uxx (x,t) = ut (x,t), 0 < x < L,t > 0, (14.1)

where α 2 is a constant known as the thermal diffusivity. This equation is called the
heat conduction equation or heat equation.
In addition, we assume that the initial temperature distribution in the bar is given
by
u(x, 0) = f (x), 0 ≤ x ≤ L, (14.2)

where f is a given function. Finally, we assume that the temperature at each end of
the bar is given by
c Springer International Publishing AG 2017 99
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 14
100 Part I: Fourier Series and the Discrete Fourier Transform

u(0,t) = g0 (t), u(L,t) = g1 (t), t > 0, (14.3)

where g0 and g1 are given functions. The problem (14.1), (14.2), (14.3) is an ini-
tial value problem in the time variable t. With respect to the space variable x it is a
boundary value problem, and the conditions of (14.3) are called the boundary con-
ditions. Alternatively, this problem can be considered a boundary value problem in
the xt-plane (Figure 14.2):

u(0,t) = g0 (t) α 2 uxx = ut u(L,t) = g1 (t)

x
u(x, 0) = f (x)

x=0 x=L

Fig. 14.2 Geometric illustration of the heat equation as a boundary value problem.

We begin by considering the homogeneous boundary conditions when the func-

tions g0 (t) and g1 (t) in (14.3) are identically zero:
⎧
⎪
⎨α uxx = ut ,
2 0 < x < L,t > 0,
u(0,t) = u(L,t) = 0, t > 0, (14.4)
⎪
⎩
u(x, 0) = f (x), 0 ≤ x ≤ L.

We look for a solution to the problem (14.4) in the form

u(x,t) = X(x)T (t). (14.5)

Such a method is called a separation of variables (or Fourier’s method). Substituting

(14.5) into (14.1) yields

α 2 X  (x)T (t) = X(x)T  (t),

or
X  (x) 1 T  (t)
= 2 ,
X(x) α T (t)
14 Applications to Solving Some Model Equations 101

in which the variables are separated, that is, the left-hand side depends only on x,
and the right-hand side only on t. This is possible only when both sides are equal to
the same constant:
X  1 T
= 2 = −λ .
X α T

Hence, we obtain two ordinary differential equations for X(x) and T (t):

X  + λ X = 0,

T  + α 2 λ T = 0. (14.6)

The boundary condition for u(x,t) at x = 0 leads to

u(0,t) = X(0)T (t) = 0.

It follows that
X(0) = 0,

since otherwise, T ≡ 0, and so u ≡ 0, which we do not accept. Similarly, the bound-

ary condition at x = L requires that

X(L) = 0.

So, for the function X(x) we obtain the homogeneous boundary value problem

X  + λ X = 0, 0 < x < L,
(14.7)
X(0) = X(L) = 0.

The values of λ for which nontrivial solutions of (14.7) exist are called eigenvalues,
and the corresponding nontrivial solutions are called eigenfunctions. The problem
(14.7) is called an eigenvalue problem.
Lemma 14.1. The problem (14.7) has an infinite sequence of positive eigenvalues

n2 π 2
λn = , n = 1, 2, . . . ,
L2
with the corresponding eigenfunctions

nπ x
Xn (x) = c sin ,
L
where c is an arbitrary nonzero constant.
Proof. Suppose first that λ > 0, i.e., λ = μ 2 . The characteristic equation for (14.7)
is r2 + μ 2 = 0 with roots r = ±iμ , so the general solution is
102 Part I: Fourier Series and the Discrete Fourier Transform

X(x) = c1 cos μ x + c2 sin μ x.

Note that μ is nonzero, and there is no loss of generality if we assume that μ > 0.
The first boundary condition in (14.7) implies

X(0) = c1 = 0,

and the second reduces to

c2 sin μ L = 0,

or
sin μ L = 0,

since we do not allow c2 = 0 either. It follows that

μ L = nπ , n = 1, 2, . . . ,

or
n2 π 2
λn = , n = 1, 2, . . . .
L2
Hence the corresponding eigenfunctions are

nπ x
Xn (x) = c sin .
L

If λ = −μ 2 < 0, μ > 0, then the characteristic equation for (14.7) is r2 − μ 2 = 0

with roots r = ±μ . Hence the general solution is

X(x) = c1 cosh μ x + c2 sinh μ x.

Since
eμ x + e−μ x eμ x − e−μ x
cosh μ x = and sinh μ x = ,
2 2
this is equivalent to
X(x) = c1 eμ x + c2 e−μ x .

The first boundary condition requires again that c1 = 0, while the second gives

c2 sinh μ L = 0.

Since μ = 0 (μ > 0), it follows that sinh μ L = 0, and therefore we must have c2 = 0.
Consequently, X ≡ 0, i.e., there are no nontrivial solutions for λ < 0.
14 Applications to Solving Some Model Equations 103

If λ = 0, the general solution is

X(x) = c1 x + c2 .

The boundary conditions can be satisfied only if c1 = c2 = 0, so there is only the

trivial solution in this case as well. 

n2 π 2
Turning now to (14.6) for T (t) and substituting L2
for λ , we have

T (t) = ce−( ) .
nπα 2 t
L

Hence the functions

un (x,t) = e−( ) sin nπ x
nπα 2 t
L (14.8)
L
satisfy (14.1) and the homogeneous boundary conditions from (14.4) for each n =
1, 2, . . .. The linear superposition principle gives that every linear combination

) sin nπ x
N
∑ cn e−(
nπα 2 t
u(x,t) = L
n=1 L

is also a solution of the same problem. In order to take into account infinitely many
functions (14.8), we assume that
∞
) sin nπ x ,
∑ cn e−(
nπα 2 t
u(x,t) = L (14.9)
n=1 L

where the coefficients cn are still undetermined, and the series converges in some
sense. To satisfy the initial condition from (14.4) we must have
∞
nπ x
u(x, 0) = ∑ cn sin L
= f (x), 0 ≤ x ≤ L. (14.10)
n=1

In other words, we need to choose the coefficients cn so that the series (14.10) con-
verges to the initial temperature distribution f (x). We extend f from [0, L] to [−L, L]
as an odd function and then obtain that
 L
2 nπ x
cn = f (x) sin dx.
L 0 L

It is not difficult to prove that for t > 0, 0 < x < L, the series (14.9) converges
(with any derivative with respect to x and t) and solves (14.1) with boundary condi-
tions (14.4). Only one question remains: can every function f (x) be represented by
a Fourier sine series (14.10)? Some sufficient conditions for such a representation
are given in Chapter 10.
104 Part I: Fourier Series and the Discrete Fourier Transform

Remark 14.2. We can consider the boundary value problem for a linear differential
equation
y + p(x)y + q(x)y = g(x) (14.11)

of order two on the interval (a, b) with the boundary conditions

y(a) = y0 , y(b) = y1 , (14.12)

where y0 and y1 are given constants. Let us assume that we have found a fundamen-
tal set of solutions y1 (x) and y2 (x) to the corresponding homogeneous equation

y + p(x)y + q(x)y = 0.

Then the general solution to (14.11) is

y(x) = c1 y1 (x) + c2 y2 (x) + yp (x),

where yp (x) is a particular solution to (14.11) and c1 and c2 are arbitrary constants.
To satisfy the boundary conditions (14.12) we have the linear inhomogeneous
algebraic system 
c1 y1 (a) + c2 y2 (a) = y0 − yp (a),
(14.13)
c1 y1 (b) + c2 y2 (b) = y1 − yp (b).

If the determinant  
 y1 (a) y2 (a) 
 
 y1 (b) y2 (b) 

is nonzero, then the constants c1 and c2 can be determined uniquely, and therefore
the boundary value problem (14.11)–(14.12) has a unique solution. If
 
 y1 (a) y2 (a) 
 
 y1 (b) y2 (b)  = 0,

then (14.11)–(14.12) has either no solutions or infinitely many solutions.

Example 14.3. Let us consider the boundary value problem


y + μ 2 y = 1, 0 < x < 1,
y(0) = y0 , y(1) = y1 ,

where μ > 0 is fixed. This differential equation has a particular solution yp (x) = 1
μ2
.
Hence, the system (14.13) becomes
14 Applications to Solving Some Model Equations 105

c1 sin 0 + c2 cos 0 = y0 − μ12 ,
c1 sin μ + c2 cos μ = y1 − μ12 ,

or ⎧
⎨c2 = y0 − 12 ,
μ

⎩c1 sin μ = y1 − 12 − y0 − 12 cos μ .

μ μ

If  
 0 1 

 sin μ cos μ  = 0,

i.e., sin μ = 0, then c1 is uniquely determined, and the boundary value problem in
question has a unique solution. If sin μ = 0, then the problem has solutions (in fact,
infinitely many) if and only if
 
1 1
y1 − 2 = y0 − 2 cos μ .
μ μ

If μ = 2π k, then sin μ = 0 and cos μ = 1 and the following equation must hold:

1 1
y1 − = y0 − 2 ,
μ2 μ

i.e., y1 = y0 . If μ = π + 2π k, then sin μ = 0 and cos μ = −1, and we must have

2
y1 + y0 = .
μ2

Suppose now that the ends of the bar are held at constant temperatures T1 and T2 .
The corresponding boundary value problem is then
⎧
⎪
⎨α uxx = ut ,
2 0 < x < L,t > 0,
u(0,t) = T1 , u(L,t) = T2 , t > 0, (14.14)
⎪
⎩
u(x, 0) = f (x).

After a long time (t → ∞) we anticipate that a steady temperature distribution v(x)

will be reached that is independent of time and the initial condition. Since the solu-
tion of (14.14) with T1 = T2 = 0 tends to zero as t → ∞, see (14.9), we look for the
solution to (14.14) in the form

u(x,t) = v(x) + w(x,t). (14.15)

106 Part I: Fourier Series and the Discrete Fourier Transform

Substituting (14.15) into (14.14) leads to

⎧
⎪
⎨α (vxx + wxx ) = wt ,
2

v(0) + w(0,t) = T1 , v(L) + w(L,t) = T2 ,

⎪
⎩
v(x) + w(x, 0) = f (x).

Let us assume that v(x) satisfies the steady-state problem


v (x) = 0, 0 < x < L,
(14.16)
v(0) = T1 , v(L) = T2 .

Then w(x,t) satisfies the homogeneous boundary value problem for the heat equa-
tion: ⎧
⎪
⎨α wxx = wt ,
2 0 < x < L,t > 0,
w(0,t) = w(L,t) = 0, (14.17)
⎪
⎩
w(x, 0) = f˜(x),

where f˜(x) = f (x) − v(x). Since the solution of (14.16) is

T2 − T1
v(x) = x + T1 , (14.18)
L
the solution of (14.17) is
∞
) sin nπ x ,
∑ cn e−(
nπα 2 t
w(x,t) = L (14.19)
n=1 L

where the coefficients cn are given by

 L 
2 T2 − T1 nπ x
cn = f (x) − x − T1 sin dx.
L 0 L L

Combining (14.18) and (14.19), we obtain

∞
T2 − T1 nπ x
x + T1 + ∑ cn e−( L ) t sin
nπα 2
u(x,t) = .
L n=1 L

Let us slightly complicate the problem (14.14), namely assume that

⎧
⎪
⎨α uxx = ut + p(x),
2 0 < x < L,t > 0,
u(0,t) = T1 , u(L,t) = T2 , t > 0, (14.20)
⎪
⎩
u(x, 0) = f (x).
14 Applications to Solving Some Model Equations 107

We begin by assuming that the solution to (14.20) consists of a steady-state solution

v(x) and a transient solution w(x,t) that tends to zero as t → ∞:

u(x,t) = v(x) + w(x,t).

Then for v(x) we will have the problem


v (x) = α12 p(x), 0 < x < L,
(14.21)
v(0) = T1 , v(L) = T2 .

To solve this we integrate twice to get

 x  y
1
v(x) = 2 dy p(s)ds + c1 x + c2 .
α 0 0

The boundary conditions yield c2 = T1 and

   y 
1 1 L
c1 = T2 − T1 − 2 dy p(s)ds .
L α 0 0

Therefore, the solution of (14.21) has the form

 L  y  x  y
T2 − T1 x 1
v(x) = x− dy p(s)ds + dy p(s)ds + T1 .
L Lα 2 0 0 α2 0 0

For w(x,t) we will have the homogeneous problem

⎧
⎪
⎨α wxx = wt , 0 < x < L,t > 0,
2

w(0,t) = w(L,t) = 0,t > 0,

⎪
⎩
w(x, 0) = f˜(x) := f (x) − v(x).

A different problem occurs if the ends of the bar are insulated so that there is no
passage of heat through them. Thus, in the case of no heat flow, the boundary value
problem is ⎧
⎪
⎨α uxx = ut , 0 < x < L,t > 0,
2

ux (0,t) = ux (L,t) = 0,t > 0, (14.22)

⎪
⎩
u(x, 0) = f (x).

This problem can also be solved by the method of separation of variables. If we let
u(x,t) = X(x)T (t), it follows that

X  + λ X = 0, T  + α 2 λ T = 0. (14.23)

The boundary conditions now yield

108 Part I: Fourier Series and the Discrete Fourier Transform

X  (0) = X  (L) = 0. (14.24)

If λ = −μ 2 < 0, μ > 0, then (14.23) for X(x) becomes X  − μ 2 X = 0 with general

solution
X(x) = c1 sinh μ x + c2 cosh μ x.

Therefore, the conditions (14.24) give c1 = 0 and c2 = 0, which is unacceptable.

Hence λ cannot be negative.
If λ = 0, then
X(x) = c1 x + c2 .

Thus X  (0) = c1 = 0 and X  (L) = 0 for every c2 , leaving c2 undetermined. Therefore

λ = 0 is an eigenvalue, corresponding to the eigenfunction X0 (x) = 1. It follows
from (14.23) that T (t) is also a constant. Hence, for λ = 0 we obtain the constant
solution u0 (x,t) = c2 .
If λ = μ 2 > 0, then X  + μ 2 X = 0 and consequently

X(x) = c1 sin μ x + c2 cos μ x.

nπ
The boundary conditions imply c1 = 0 and μ = L , n = 1, 2, . . ., leaving c2 arbi-
n2 π 2
trary. Thus we have an infinite sequence of positive eigenvalues λn = L2
with the
corresponding eigenfunctions

nπ x
Xn (x) = cos , n = 1, 2, . . . .
L

If we combine these eigenvalues and eigenfunctions with zero eigenvalue λ0 = 0

and X0 (x) = 1, we may conclude that we have the infinite sequences

n2 π 2 nπ x
λn = , Xn (x) = cos , n = 0, 1, 2, . . . ,
L2 L
and
nπ x −( nπα )2 t
un (x,t) = cos e L , n = 0, 1, 2, . . . .
L
Each of these functions satisfies the equation and boundary conditions from (14.22).
It remains to satisfy the initial condition. In order to do so, we assume that u(x,t)
has the form
∞
c0 nπ x −( nπα )2 t
u(x,t) = + ∑ cn cos e L , (14.25)
2 n=1 L
14 Applications to Solving Some Model Equations 109

where the coefficients cn are determined by the requirement that

∞
c0 nπ x
u(x, 0) = + ∑ cn cos = f (x), 0 ≤ x ≤ L.
2 n=1 L

Thus the unknown coefficients in (14.25) must be the Fourier coefficients in the
Fourier cosine series of period 2L for the even extension of f . Hence
 L
2 nπ x
cn = f (x) cos dx, n = 0, 1, 2, . . . ,
L 0 L

and the series (14.25) provides the solution to the heat conduction problem (14.22)
for a bar with insulated ends. The physical interpretation of the term
 L
c0 1
= f (x)dx
2 L 0

is that it is the mean value of the original temperature distribution and a steady-state
solution in this case.

Exercise 14.1. Let v(x) be a solution of the problem


v (x) = 0, 0 < x < L,
v (0) = T1 , v (L) = T2 .

Show that the problem

⎧
⎪
⎨α uxx = ut , 0 < x < L,t > 0,
2

ux (0,t) = T1 , ux (L,t) = T2 ,t > 0,

⎪
⎩
u(x, 0) = f (x),

has a solution of the form u(x,t) = v(x) + w(x,t) if and only if T1 = T2 .

Example 14.4. ⎧
⎪
⎨uxx = ut , 0 < x < 1,t > 0,
u(0,t) = u(1,t) = 0,
⎪
⎩
u(x, 0) = ∑∞
n=1 n2 sin(nπ x) := f (x).
1

As we know, the solution of this problem is given by

∞ 2
u(x,t) = ∑ cn sin(nπ x)e−(nπ ) t .
n=1
110 Part I: Fourier Series and the Discrete Fourier Transform

Since
∞ ∞
1
u(x, 0) = ∑ cn sin(nπ x) = ∑ n2 sin(nπ x),
n=1 n=1

we conclude that necessarily cn = 1

n2
(since the Fourier series is unique). Hence the
solution is
∞
1 2
u(x,t) = ∑ n2 sin(nπ x)e−(nπ ) t .
n=1

Exercise 14.2. Find a solution of the problem

⎧
⎪
⎨uxx = ut , 0 < x < π ,t > 0,
ux (0,t) = ux (π ,t) = 0,t > 0,
⎪
⎩
u(x, 0) = 1 − sin x,

using the method of separation of variables.

Let us consider a bar with mixed boundary conditions at the ends. Assume that
the temperature at x = 0 is zero, while the end x = L is insulated so that no heat
passes through it: ⎧
⎪
⎨α uxx = ut , 0 < x < L,t > 0,
2

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

⎪
⎩
u(x, 0) = f (x).

Separation of variables leads to


X  + λ X = 0, 0 < x < L,
(14.26)
X(0) = X  (L) = 0,

and
T  + α 2 λ T = 0, t > 0.

As above, one can show that (14.26) has nontrivial solutions only for λ > 0, namely

(2m − 1)2 π 2 (2m − 1)π x

λm = , Xm (x) = sin , m = 1, 2, 3, . . . .
4L2 2L
The solution to the mixed boundary value problem is

∞
2
(2m − 1)π x − (2m−1)πα
∑
t
u(x,t) = cm sin e 2L

m=1 2L
14 Applications to Solving Some Model Equations 111

with arbitrary constants cm . To satisfy the initial condition we have

∞
(2m − 1)π x
f (x) = ∑ cm sin 2L
, 0 ≤ x ≤ L.
m=1

This is a Fourier sine series but in some specific form. We show that the coefficients
cm can be calculated as
 L
2 (2m − 1)π x
cm = f (x) sin dx,
L 0 2L

and such a representation is possible.

In order to prove this, let us first extend f (x) to the interval 0 ≤ x ≤ 2L so that
it is symmetric about x = L, i.e., f (2L − x) = f (x) for 0 ≤ x ≤ L. We then extend
the resulting function to the interval (−2L, 0) as an odd function and elsewhere as a
periodic function f of period 4L. In this procedure we need to define

f(0) = f(2L) = f(−2L) = 0.

Then the Fourier series contains only sines:

∞
nπ x
f(x) = ∑ cn sin ,
n=1 2L

with the Fourier coefficients

 2L
2 nπ x
cn = f(x) sin dx.
2L 0 2L

Let us show that cn = 0 for even n = 2m. Indeed,


1 2L  mπ x
c2m = f (x) sin dx
L 0 L
 
1 L mπ x 1 2L mπ x
= f (x) sin dx + f (2L − x) sin dx
L 0 L L L L
 L  0
1 mπ x 1 mπ (2L − y)
= f (x) sin dx − f (y) sin dy
L 0 L L L L
 
1 L mπ x 1 0 mπ y
= f (x) sin dx + f (y) sin dy = 0,
L 0 L L L L

which is why
∞
(2m − 1)π x
f(x) = ∑ c2m−1 sin ,
m=1 2L
112 Part I: Fourier Series and the Discrete Fourier Transform

where

1 2L  (2m − 1)π x
c2m−1 = f (x) sin dx
L 0 2L
 
1 L (2m − 1)π x 1 2L (2m − 1)π x
= f (x) sin dx + f (2L − x) sin dx
L 0 2L L L 2L
 L
2 (2m − 1)π x
= f (x) sin dx,
L 0 2L

as claimed. Let us remark that the series

∞
(2m − 1)π x
∑ cm sin 2L
m=1

represents f (x) on (0, L].

Remark 14.5. For the boundary conditions

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

the function f (x) must be extended to the interval 0 ≤ x ≤ 2L as f (x) = − f (2L − x)

with f (L) = 0. Furthermore, f is an even extension to the interval (−2L, 0). Then
the corresponding Fourier series represents f (x) on the interval [0, L).

Exercise 14.3. (1) Let u(x,t) satisfy

⎧
⎪
⎨uxx = ut , 0 < x < 1,t > 0,
u(0,t) = u(1,t) = 0, t ≥ 0,
⎪
⎩
u(x, 0) = f (x), 0 ≤ x ≤ 1,

where f ∈ C[0, 1]. Show that for every T ≥ 0 we have

 1  1
|u(x, T )|2 dx ≤ | f (x)|2 dx.
0 0

Hint: Use the identity 2u(ut − uxx ) = ∂t u2 − ∂x (u · ux ) + 2(ux )2 .

(2) Use Fourier’s method to solve
⎧
⎪
⎨uxx = ut , 0 < x < 1,t > 0,
u(0,t) = u(1,t), ux (0,t) = ux (1,t) t ≥ 0,
⎪
⎩
u(x, 0) = f (x), 0 ≤ x ≤ 1,

where f ∈ C[0, 1] with piecewise continuous derivative.

14 Applications to Solving Some Model Equations 113

(3) Use Fourier’s method to solve

⎧
⎪
⎨uxx + u = ut , 0 < x < 1,t > 0,
u(0,t) = ux (0,t) = 0 t ≥ 0,
⎪
⎩
u(x, 0) = x(1 − x), 0 ≤ x ≤ 1.

(4) Show that  ∞

1
ϕ (ξ )e−(x−ξ +t)
2 /(4t)
u(x,t) = √ dξ
4π t −∞

solves the problem

⎧
⎪
⎨uxx + ux = ut , −∞ < x < ∞,t > 0,
u(x, 0) = ϕ (x), −∞ < x < ∞,
⎪
⎩
u(x,t) is bounded for − ∞ < x < ∞,t ≥ 0.

14.2 The One-Dimensional Wave Equation

Another situation in which the separation of variables applies occurs in the study
of a vibrating string. Suppose that an elastic string of length L is tightly stretched
between two supports, so that the x-axis lies along the string. Let u(x,t) denote the
vertical displacement experienced by the string at the point x at time t. It turns out
that if damping effects are neglected, and if the amplitude of the motion is not too
large, then u(x,t) satisfies the partial differential equation

a2 uxx = utt , 0 < x < L,t > 0. (14.27)

Equation (14.27) is known as the one-dimensional wave equation. The constant a2

is equal to T /ρ , where T is the force in the string and ρ is the mass per unit length
of the string material (Figure 14.3).

x
x=0 x=L

u(x,t)

Fig. 14.3 Geometry of the one-dimensional wave equation.

To describe the motion completely it is necessary also to specify suitable initial

and boundary conditions for the displacement u(x,t). The ends of the string are
assumed to remain fixed:
114 Part I: Fourier Series and the Discrete Fourier Transform

u(0,t) = u(L,t) = 0, t ≥ 0. (14.28)

The initial conditions are (since (14.27) is of second order with respect to t)

u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L, (14.29)

where f and g are given functions. In order for (14.28) and (14.29) to be consistent,
it is also necessary to require that

f (0) = f (L) = g(0) = g(L) = 0. (14.30)

Equations (14.27)–(14.30) can be interpreted as the following boundary value prob-

lem for the wave equation (Figure 14.4):

u(0,t) = 0 a2 uxx = utt u(L,t) = 0

x
u(x, 0) = f (x)
ut (x, 0) = g(x)
x=0 x=L

Fig. 14.4 Geometric illustration of the wave equation as a boundary value problem.

Let us apply the method of separation of variables to this homogeneous boundary

value problem. Assuming that u(x,t) = X(x)T (t), we obtain

X  + λ X = 0, T  + a2 λ T = 0.

The boundary conditions (14.28) imply that


X  + λ X = 0, 0 < x < L,
X(0) = X(L) = 0.
14 Applications to Solving Some Model Equations 115

This is the same boundary value problem that we have considered previously.
Hence,
n2 π 2 nπ x
λn = 2 , Xn (x) = sin , n = 1, 2, . . . .
L L

Taking λ = λn in the equation for T (t), we have

nπ a
2
T  (t) + T (t) = 0.
L
The general solution to this equation is

nπ at nπ at
T (t) = k1 cos + k2 sin ,
L L
where k1 and k2 are arbitrary constants. Using the linear superposition principle, we
consider the series
∞
nπ x nπ at nπ at

u(x,t) = ∑ sin L
an cos
L
+ bn sin
L
, (14.31)
n=1

where the coefficients an and bn are to be determined. It is clear that u(x,t)

from (14.31) satisfies (14.27) and (14.28) (at least formally). The initial conditions
(14.29) imply
∞
nπ x
f (x) = ∑ an sin , 0 ≤ x ≤ L,
n=1 L
∞ (14.32)
nπ a nπ x
g(x) = ∑ bn sin , 0 ≤ x ≤ L.
n=1 L L

Since (14.30) are fulfilled it follows that equations (14.32) are the Fourier sine series
for f and g, respectively. Therefore,
 L
2 nπ x
an = f (x) sin
dx,
L
0 L
 L (14.33)
2 nπ x
bn = g(x) sin dx.
nπ a 0 L

Finally, we may conclude that the series (14.31) with the coefficients (14.33) solves
(at least formally) the boundary value problem (14.27)–(14.30).
Each displacement pattern

nπ x nπ at nπ at

un (x,t) = sin an cos + bn sin

L L L
116 Part I: Fourier Series and the Discrete Fourier Transform

is called a natural mode of vibration and is periodic in both the space variable x
and time variable t. The spatial period 2L
n in x is called the wavelength, while the
numbers nπLa are called the natural frequencies.

Exercise 14.4. Find a solution of the problem

⎧
⎪
⎨uxx = utt , 0 < x < 1,t > 0,
u(0,t) = u(1,t) = 0,t ≥ 0,
⎪
⎩
u(x, 0) = x(1 − x), ut (x, 0) = sin(7π x),

using the method of separation of variables.

If we compare the two series

∞
nπ x nπ at nπ at

u(x,t) = ∑ sin L
an cos
L
+ bn sin
L
,
n=1
∞
nπ x −( nπα )2 t
u(x,t) = ∑ cn sin L
e L
n=1

for the wave and heat equations, we see that the second series has an exponential
factor that decays rapidly with n for every t > 0. This guarantees convergence of the
series as well as the smoothness of the sum. This is no longer true for the first series,
because it contains only oscillatory terms that do not decay with increasing n. This
means that the solution of the heat equation is a C∞ function in the corresponding
domain, but the solution of the wave equation is not necessarily smooth.
The boundary value problem for the wave equation with free ends of the string
can be formulated as follows:
⎧
⎪
⎨a uxx = utt , 0 < x < L,t > 0,
2

ux (0,t) = ux (L,t) = 0,t ≥ 0,

⎪
⎩
u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L.

Let us first note that the boundary conditions imply that f (x) and g(x) must satisfy

f  (0) = f  (L) = g (0) = g (L) = 0.

The method of separation of variables gives that the eigenvalues are

nπ
2
λn = , n = 0, 1, 2, . . . ,
L

and the formal solution u(x,t) is

b0t + a0 ∞
nπ x nπ at nπ at

u(x,t) = + ∑ cos an cos + bn sin .

2 n=1 L L L
14 Applications to Solving Some Model Equations 117

The initial conditions are satisfied when

∞
a0 nπ x
f (x) = + ∑ an cos
2 n=1 L

and ∞
b0 nπ a nπ x
g(x) = + ∑ bn cos ,
2 n=1 L L

where

2 L nπ x
an = f (x) cos dx, n = 0, 1, 2, . . . ,
L 0 L

2 L
b0 = g(x)dx,
L 0

and  L
2 nπ x
bn = g(x) cos dx, n = 1, 2, . . . .
nπ a 0 L

Exercise 14.5. (1) Show that there is no uniqueness for the problem
⎧
⎪
⎨uxx = utt , 0 < x < L,t > 0,
u(0,t) = u(L,t), ux (0,t) = ux (L,t), t ≥ 0,
⎪
⎩
u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L,

i.e., this problem is ill posed.

(2) Find a solution of the problem
⎧
⎪
⎨uxx = utt , 0 < x < L,t > 0,
ux (0,t) = u(L,t) = 0, t ≥ 0,
⎪
⎩
u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L.

Let us consider the wave equation on the whole line. It corresponds, so to speak,
to an infinite string. In that case we no longer have the boundary conditions, but we
have the initial conditions

a2 uxx = utt , −∞ < x < ∞,t > 0,
(14.34)
u(x, 0) = f (x), ut (x, 0) = g(x).

Proposition 14.6. The solution u(x,t) of the wave equation is of the form

u(x,t) = ϕ (x − at) + ψ (x + at),

where ϕ and ψ are two arbitrary C2 functions of one variable.

118 Part I: Fourier Series and the Discrete Fourier Transform

Proof. By the chain rule,

∂tt u − a2 ∂xx u = 0

if and only if
∂ξ ∂η u = 0,

where ξ = x + at and η = x − at (and so ∂x = ∂ξ + ∂η , 1a ∂t = ∂ξ − ∂η ). It follows

that
∂ξ u = Θ (ξ ),

or
u = ψ (ξ ) + ϕ (η ),

where ψ  = Θ . 

To satisfy the initial conditions, we have

f (x) = ϕ (x) + ψ (x), g(x) = −aϕ  (x) + aψ  (x).

It follows that
1  1 1  1
ϕ  (x) = f (x) − g(x), ψ  (x) = f (x) + g(x).
2 2a 2 2a
Integrating, we obtain
 x  x
1 1 1 1
ϕ (x) = f (x) − g(s)ds + c1 , ψ (x) = f (x) + g(s)ds + c2 ,
2 2a 0 2 2a 0

where c1 and c2 are arbitrary constants. But ϕ (x) + ψ (x) = f (x) implies c1 + c2 = 0.
Therefore, the solution of the initial value problem is
 x+at
1 1
u(x,t) = ( f (x − at) + f (x + at)) + g(s)ds. (14.35)
2 2a x−at

This formula is called d’Alembert’s formula.

Exercise 14.6. Prove that if f is a C2 function and g is a C1 function, then u from
(14.35) is a C2 function that satisfies (14.34) in the classical sense (pointwise).
Exercise 14.7. Prove that if f and g are merely locally integrable, then u from
(14.35) is a solution of (14.34) in the sense of integral equalities and the initial
conditions are satisfied pointwise.
Example 14.7. The solution of

uxx = utt , −∞ < x < ∞,t > 0,
u(x, 0) = f (x), ut (x, 0) = 0,
14 Applications to Solving Some Model Equations 119

where 
1, |x| ≤ 1,
f (x) =
0, |x| > 1,

is given by d’Alembert’s formula

1
u(x,t) = ( f (x − t) + f (x + t)) .
2
Some solutions are graphed below (Figure 14.5).

u(x, 0)

x
−1 1
u(x, 1/2)

x
−1 1
u(x, 2)

1/2

x
−1 1

Fig. 14.5 Some solutions of Example 14.7.

120 Part I: Fourier Series and the Discrete Fourier Transform

We can also apply d’Alembert’s formula to the finite string. Consider again the
boundary value problem with homogeneous boundary conditions with fixed ends of
the string: ⎧ 2
⎪a uxx = utt , 0 < x < L,t > 0,
⎪
⎪
⎨u(0,t) = u(L,t) = 0,t ≥ 0,
⎪
⎪ u(x, 0) = f (x), ut (x, 0) = g(x), 0 ≤ x ≤ L,
⎪
⎩
f (0) = f (L) = g(0) = g(L) = 0.

Let h(x) be the function defined for all x ∈ R such that


f (x), 0 ≤ x ≤ L,
h(x) =
− f (−x), −L ≤ x ≤ 0,

and h(x) is 2L-periodic, and let k(x) be the function defined for all x ∈ R such that

g(x), 0 ≤ x ≤ L,
k(x) =
−g(−x), −L ≤ x ≤ 0,

and k(x) is 2L-periodic. Let us also assume that f and g are C2 functions on the inter-
val [0, L]. Then the solution to the boundary value problem is given by d’Alembert’s
formula 
1 1 x+at
u(x,t) = (h(x − at) + h(x + at)) + k(s)ds.
2 2a x−at

Remark 14.8. It can be checked that this solution is equivalent to the solution given
by the Fourier series.

Exercise 14.8. Prove that

 x+at
1 1
u(x,t) = ( f (x − at) + f (x + at)) + g(s)ds
2 2a x−at
 
1 t x+a(t−τ )
− F(s, τ )dsdτ
2a 0 x−a(t−τ )

solves 
a2 uxx = utt + F(x,t), −∞ < x < ∞,t > 0,
u(x, 0) = f (x), ut (x, 0) = g(x), −∞ < x < ∞.

Example 14.9. (Linearized system of the equation of gas dynamics) The isen-
tropic (the entropy is assumed to be constant) flow of an inviscid gas in the one-
dimensional case satisfies the nonlinear equations
14 Applications to Solving Some Model Equations 121
⎧ 2
⎨u + u · u + c ρ = 0,
t x x
ρ
⎩
ρt + u · ρx + ρ ux = 0, −∞ < x < ∞,t > 0,

where u(x,t) is the velocity of the gas at x and time t, ρ (x,t) is the density, and
c = c(ρ ) is the known local speed of sound.
As a first step in understanding the general nature of solutions to this system we
assume that u(x,t) and ρ (x,t) are not very different from their values at time t = 0
and that these values and their derivatives are “small.” Neglecting products of terms
of “small” order, we arrive at the linear system (which is the linearization of the
original system)
⎧ 2 ⎧
⎨u + c0 ρ = 0, ⎨c0 ρxx = ρtt ,
t x
ρ0 so c2
⎩ ⎩ 0 uxx = utt ,
ρt + ρ0 ux = 0, ρ0

where ρ0 is the density of the fluid at rest and c0 = c(ρ0 ). This is just a wave equation
for ρ and u. Thus we have that
c0
u(x,t) = √ ( f (x − c0t) + g(x + c0t))
ρ0
√
ρ (x,t) = ρ0 ( f (x − c0t) − g(x + c0t)) ,

where f and g are arbitrary C2 functions, and they are the same here due to the
linearized system. If in addition

u(x, 0) = ϕ1 (x), ρ (x, 0) = ϕ2 (x),

then
1 c0
u(x,t) = (ϕ1 (x − c0t) + ϕ1 (x + c0t)) + (ϕ2 (x − c0t) − ϕ2 (x + c0t)) ,
2 2ρ0
1 ρ0
ρ (x,t) = (ϕ2 (x − c0t) + ϕ2 (x + c0t)) + (ϕ1 (x − c0t) − ϕ1 (x + c0t)) .
2 2c0

14.3 The Laplace Equation in a Rectangle and in a Disk

One of the most important of all partial differential equations in applied mathematics
is the Laplace equation:

uxx + uyy = 0 2D equation,

(14.36)
uxx + uyy + uzz = 0 3D equation.
122 Part I: Fourier Series and the Discrete Fourier Transform

The Laplace equation appears quite naturally in many applications. For example, a
steady-state solution of the heat equation in two space dimensions,

α 2 (uxx + uyy ) = ut ,

satisfies the 2D Laplace equation (14.36). When electrostatic fields are considered,
the electric potential function must satisfy either the 2D or the 3D equation (14.36).
A typical boundary value problem for the Laplace equation is (in dimension two)

uxx + uyy = 0, (x, y) ∈ Ω ⊂ R2 ,
(14.37)
u(x, y) = f (x, y), (x, y) ∈ ∂ Ω ,

where f is a given function on the boundary ∂ Ω of the domain Ω . The problem

(14.37) is called the Dirichlet problem (Dirichlet boundary conditions). The problem

uxx + uyy = 0, (x, y) ∈ Ω ,
∂u
∂ν (x, y) = g(x, y), (x, y) ∈ ∂ Ω ,

where g is given and ∂∂ νu is the outward normal derivative, is called the Neumann
problem (Neumann boundary conditions) (Figure 14.6).

y ν |ν| = 1

Ω ∂Ω

Fig. 14.6 Domain Ω and the outward unit normal vector ν on the boundary ∂ Ω .

Dirichlet problem for a rectangle

Consider the boundary value problem in most general form:
⎧
⎪
⎨wxx + wyy = 0, 0 < x < a, 0 < y < b,
w(x, 0) = g1 (x), w(x, b) = f1 (x), 0 < x < a,
⎪
⎩
w(0, y) = g2 (y), w(a, y) = f2 (y), 0 ≤ y ≤ b,
14 Applications to Solving Some Model Equations 123

for fixed a > 0 and b > 0. The solution of this problem can be reduced to the solu-
tions of
⎧
⎪
⎨uxx + uyy = 0, 0 < x < a, 0 < y < b,
u(x, 0) = u(x, b) = 0, 0 < x < a, (14.38)
⎪
⎩
u(0, y) = g2 (y), u(a, y) = f2 (y), 0 ≤ y ≤ b,

and ⎧
⎪
⎨uxx + uyy = 0, 0 < x < a, 0 < y < b,
u(x, 0) = g1 (x), u(x, b) = f1 (x), 0 < x < a,
⎪
⎩
u(0, y) = 0, u(a, y) = 0, 0 ≤ y ≤ b.

Due to symmetry in x and y, we consider (14.38) only (Figure 14.7).

y
u(x, b) = 0
b

u(0, y) = g(y) Ω u(a, y) = f (y)

x
u(x, 0) = 0 a

Fig. 14.7 Geometric illustration of the boundary value problem (14.38).

The method of separation of variables gives for u(x, y) = X(x)Y (y),


Y  + λ Y = 0, 0 < y < b,
(14.39)
Y (0) = Y (b) = 0,

and
X  − λ X = 0, 0 < x < a. (14.40)

From (14.39) one obtains the eigenvalues and eigenfunctions

nπ
2 nπ y
λn = , Yn (y) = sin , n = 1, 2, . . . .
b b
124 Part I: Fourier Series and the Discrete Fourier Transform

We substitute λn into (14.40) to get the general solution

nπ x nπ x
X(x) = c1 cosh + c2 sinh .
b b
As above, we represent the solution to (14.38) in the form
∞
nπ y nπ x nπ x

u(x, y) = ∑ sin b
an cosh
b
+ bn sinh
b
. (14.41)
n=1

The boundary condition at x = 0 gives

∞
nπ y
g(y) = ∑ an sin b
,
n=1

with  b
2 nπ y
an = g(y) sin dy.
b 0 b

At x = a we obtain
∞
nπ y nπ a nπ a

f (y) = ∑ sin b
an cosh
b
+ bn sinh
b
.
n=1

It is a Fourier sine series for f (y). Hence,

 b
nπ a nπ a 2 nπ y
an cosh + bn sinh = f (y) sin dy =: 
bn .
b b b 0 b

This implies

bn − an cosh nπb a
bn = . (14.42)
sinh nπb a

Substituting (14.42) into (14.41) gives

 
∞
nπ y nπ x bn − an cosh nπb a nπ x
u(x, y) = ∑ sin an cosh + sinh
n=1 b b sinh nπb a b
∞
nπ y  sinh nπb x
= ∑ sinb
bn
sinh nπb a
n=1
 
∞
nπ y cosh nπb x sinh nπb a − cosh nπb a sinh nπb x
+ ∑ sin an
n=1 b sinh nπb a
nπ (a−x)
∞
nπ y  sinh nπb x ∞
nπ y sinh b
= ∑ sin b
bn + ∑
sinh nπb a n=1
sin
b
an
sinh nπb a
,
n=1

because cosh α sinh β − sinh α cosh β = sinh(β − α ). Using the properties of sinh α
and cosh α for large α , we may conclude that inside of the rectangle, i.e., for
14 Applications to Solving Some Model Equations 125

0 < x < a, 0 < y < b, we may differentiate this series term by term as often as we
wish, and so u is a C∞ function there.
Exercise 14.9. Find a solution of the problem
⎧
⎪
⎨uxx + uyy = 0, 0 < x < 2, 0 < y < 1,
u(x, 0) = u(x, 1) = 0, 0 < x < 2,
⎪
⎩
u(0, y) = 0, u(2, y) = y(1 − y), 0 ≤ y ≤ 1,

using the method of separation of variables.

Exercise 14.10. Find a solution of the problem
⎧
⎪
⎨uxx + uyy = 0, 0 < x < a, 0 < y < b,
uy (x, 0) = uy (x, b) = 0, 0 < x < a,
⎪
⎩
ux (0, y) = f (y), ux (a, y) = g(y), 0 < y < b,

using the method of separation of variables.

Dirichlet problem for a disk
 
Consider the problem of solving the Laplace equation in a disk x ∈ R2 : |x| < a
subject to the boundary condition

u(a, θ ) = f (θ ), (14.43)

where f is a given function on 0 ≤ θ ≤ 2π . In polar coordinates x = r cos θ , y =

r sin θ , the Laplace equation takes the form

1 1
urr + ur + 2 uθ θ = 0. (14.44)
r r
We apply again the method of separation of variables and assume that

u(r, θ ) = R(r)T (θ ). (14.45)

Substitution for u in (14.44) yields

1 1
R T + R T + 2 RT  = 0,
r r
or 
r2 R + rR − λ R = 0,
T  + λ T = 0.

There are no homogeneous boundary conditions; however, we require T (θ ) to be

2π -periodic and also bounded. This fact in particular leads to ( f (0) = f (2π ) and
126 Part I: Fourier Series and the Discrete Fourier Transform

f  (0) = f  (2π ))
T (0) = T (2π ), T  (0) = T  (2π ). (14.46)

It is possible to show that equations (14.46) require λ to be real. In what follows we

will consider the three possible cases.
If λ = −μ 2 < 0, μ > 0, then the equation for T becomes T  − μ 2 T = 0, and
consequently
T (θ ) = c1 eμθ + c2 e−μθ .

It follows from (14.46) that


c1 + c2 = c1 e2π μ + c2 e−2π μ ,
c1 − c2 = c1 e2π μ − c2 e−2π μ ,

so that c1 = c2 = 0.

If λ = 0, then T  = 0 and T (θ ) = c1 + c2 θ . The first condition in (14.46) implies

then that c2 = 0 and therefore T (θ ) ≡ constant.
If λ = μ 2 > 0, μ > 0, then

T (θ ) = c1 cos(μθ ) + c2 sin(μθ ).

Now the conditions (14.46) imply that


c1 = c1 cos(2π μ ) + c2 sin(2π μ ),
c2 = −c1 sin(2π μ ) + c2 cos(2π μ ),

or 
c1 sin2 (μπ ) = c2 sin(μπ ) cos(μπ ),
c2 sin2 (μπ ) = −c1 sin(μπ ) cos(μπ ).

If sin(μπ ) = 0, then 
c1 = c2 cot(μπ ),
c2 = −c1 cot(μπ ).

Hence c21 + c22 = 0, i.e., c1 = c2 = 0. Thus we must have sin(μπ ) = 0, and so

λn = n2 , Tn (θ ) = c1 cos(nθ ) + c2 sin(nθ ), n = 0, 1, 2, . . . . (14.47)

Turning now to R, for λ = 0 we have r2 R + rR = 0, i.e., R(r) = k1 + k2 log r. Since
log r → −∞ as r → 0, we must choose k2 = 0 in order for R (and u) to be bounded.
That is why
R0 (r) ≡ constant. (14.48)
14 Applications to Solving Some Model Equations 127

For λ = μ 2 = n2 the equation for R becomes

r2 R + rR − n2 R = 0.

Hence
R(r) = k1 rn + k2 r−n .

Again, we must choose k2 = 0, and therefore

Rn (r) = k1 rn , n = 1, 2, . . . . (14.49)

Combining (14.45), (14.47), (14.48), and (14.49), we obtain

∞
a0
u(r, θ ) = + ∑ rn (an cos(nθ ) + bn sin(nθ )). (14.50)
2 n=1

The boundary condition (14.43) then requires

∞
a0
u(a, θ ) = + ∑ an (an cos(nθ ) + bn sin(nθ )) = f (θ ).
2 n=1

Hence the coefficients are given by


1 2π
a0 = f (θ )dθ ,
π 0
 2π
1
an = n f (θ ) cos(nθ )dθ ,
πa 0

and  2π
1
bn = f (θ ) sin(nθ )dθ .
π an 0

This procedure can be used also to study the Neumann problem, i.e., the problem
in the disk with the boundary condition

∂u
(a, θ ) = f (θ ). (14.51)
∂r

Also in this case the solution u(r, θ ) has the form (14.50). The boundary condition
(14.51) implies that

∂u  ∞
(r, θ ) = ∑ nan−1 (an cos(nθ ) + bn sin(nθ )) = f (θ ).
∂r r=a n=1
128 Part I: Fourier Series and the Discrete Fourier Transform

Hence  2π
1
an = f (θ ) cos(nθ )dθ
π nan−1 0

and  2π
1
bn = f (θ ) sin(nθ )dθ .
π nan−1 0

Remark 14.10. For the Neumann problem a solution is defined up to an arbitrary

constant a20 . Moreover, f must satisfy the consistency condition
 2π
f (θ )dθ = 0,
0

since integrating
∞
f (θ ) = ∑ nan−1 (an cos(nθ ) + bn sin(nθ ))
n=1

termwise gives us zero.

Remark 14.11. The solution u(r, θ ) of the Laplace equation in a disk {x ∈ R2 :

|x| < a} subject to the boundary condition u(a, θ ) = f (θ ) without the assumption
that f  (0) = f  (2π ) (but with the assumption f (0) = f (2π )) can be obtained as
∞ ∞
a0
u(r, θ ) = + ∑ rn (an cos(nθ ) + bn sin(nθ )) + ∑ rn−1/2 cn sin(n − 1/2)θ .
2 n=1 n=1
(14.52)

Exercise 14.11. Prove (14.52) and then show that we have no uniqueness in this
boundary value problem. Hint: Use the fact that if cn = 0, n = 1, 2, . . ., then we may
uniquely determine a0 , an , bn satisfying the boundary condition u(a, θ ) = f (θ ), and
if an = bn = 0, n = 1, 2, . . ., we may uniquely define cn , n = 1, 2, . . . that depend
parametrically on an arbitrary constant a0 .

Remark 14.12. The solution u(r, θ ) of the Laplace equation in the disk {x ∈ R2 :
|x| < a} subject to the boundary condition uθ (a, θ ) = f (θ ) with (possibly non-
smooth) periodic function f (θ ) can be obtained as
∞ ∞
a0 θ + b0
u(r, θ ) = + ∑ rn (an cos(nθ ) + bn sin(nθ )) + ∑ rn−1/2 cn cos(n − 1/2)θ .
2 n=1 n=1
(14.53)

Exercise 14.12. Prove (14.53) and then show that we have no uniqueness in this
boundary value problem (see the previous exercise).
 Part II
Fourier Transform and Distributions
Chapter 15
Introduction

In this part we assume that the reader is familiar with the following concepts:
(1) Metric spaces and their completeness.
(2) The Lebesgue integral in a bounded domain Ω ⊂ Rn and in Rn .
(3) The Banach spaces (L p , 1 ≤ p ≤ ∞, Ck ) and Hilbert spaces (L2 ): If 1 ≤ p < ∞,
then we set
 1/p
L (Ω ) := { f : Ω → C measurable :  f L p (Ω ) :=
p
| f (x)| dx
p
< ∞},
Ω

while

L∞ (Ω ) := { f : Ω → C measurable :  f L∞ (Ω ) := ess sup | f (x)| < ∞}.

x∈Ω

Moreover,

Ck (Ω ) := { f : Ω → C :  f Ck (Ω ) := max ∑
x∈Ω |α |≤k
|∂ α f (x)| < ∞},

where Ω is the closure of Ω . We say that f ∈ C∞ (Ω ) if f ∈ Ck (Ω1 ) for all k ∈ N

and for all bounded subsets Ω1 ⊂ Ω . The space C∞ (Ω ) is not a normed space.
The inner product in L2 (Ω ) is denoted by

( f , g)L2 (Ω ) = f (x)g(x)dx.
Ω

Also in L2 (Ω ), the duality pairing is given by


 f , gL2 (Ω ) = f (x)g(x)dx.
Ω

c Springer International Publishing AG 2017 131

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 15
132 Part II: Fourier Transform and Distributions

(4) Hölder’s inequality: Let 1 ≤ p ≤ ∞, u ∈ L p , and v ∈ L p with

1 1
+ = 1.
p p

Then uv ∈ L1 and
   1   1
p
p
p
|u(x)v(x)|dx ≤ |u(x)| dx
p
|v(x)| dx ,
Ω Ω Ω

where the Hölder conjugate exponent p of p is obtained via

p
p =
p−1

with the understanding that p = ∞ if p = 1 and p = 1 if p = ∞.

(5) Lebesgue’s dominated convergence theorem:
Let A ⊂ Rn be measurable and let { fk }∞ k=1 be a sequence of measurable func-
tions converging to f (x) pointwise in A. If there exists a function g ∈ L1 (A) such
that | fk (x)| ≤ g(x) in A, then f ∈ L1 (A) and
 
lim fk (x)dx = f (x)dx.
k→∞ A A

(6) Fubini’s theorem on the interchange of the order of integration:

      
f (x, y)dxdy = dx f (x, y)dy = dy f (x, y)dx
X×Y X Y Y X

if f ∈ L1 (X ×Y ).
(7) The divergence theorem: Let Ω ⊂ Rn be a bounded domain with C1 boundary
∂ Ω , and let F be a C1 vector-valued function on Ω . Then
 
div F(x)dx = ν · Fdσ (x),
Ω ∂Ω

where ν is the outward normal vector to ∂ Ω .

Chapter 16
The Fourier Transform in Schwartz Space

Consider
 the Euclidean space Rn , n ≥ 1, with x = (x1 , . . . , xn ) ∈ Rn and with |x| =
x12 + · · · + xn2 and scalar product (x, y) = ∑nj=1 x j y j . The open ball of radius δ > 0
centered at x ∈ Rn is denoted by

Uδ (x) := {y ∈ Rn : |x − y| < δ }.

Recall the Cauchy–Bunyakovsky–Schwarz inequality

|(x, y)| ≤ |x||y|.

Following Laurent Schwartz, we call an n-tuple α = (α1 , . . . , αn ), α j ∈ N∪{0} ≡ N0

an n-dimensional multi-index. Define

|α | = α1 + · · · + αn , α ! = α1 ! · · · αn !

and
xα = x1α1 · · · xnαn , 00 = 1, 0! = 1.

Moreover, multi-indices α and β can be ordered according to

α ≤β

if α j ≤ β j for all j = 1, 2, . . . , n. Let us also introduce a shorthand notation

∂
∂ α = ∂1α1 · · · ∂nαn , ∂j = .
∂xj

c Springer International Publishing AG 2017 133

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 16
134 Part II: Fourier Transform and Distributions

Definition 16.1. The Schwartz space S(Rn ) of rapidly decaying functions is

defined as
 
 
S(Rn ) = { f ∈ C∞ (Rn ) : | f |α ,β := sup xα ∂ β f (x) < ∞ for any α , β ∈ Nn0 }.
x∈Rn

The following properties of S = S(Rn ) are readily verified.

(1) S is a linear space.
(2) ∂ α : S → S for every α ≥ 0.
(3) xβ · : S → S for every β ≥ 0.
(4) If f ∈ S(Rn ), then | f (x)| ≤ cm (1 + |x|)−m for every m ∈ N. The converse is not
true (see part (3) of Example 16.2).
(5) It follows from part (4) that S(Rn ) ⊂ L p (Rn ) for every 1 ≤ p ≤ ∞.

Example 16.2.
f (x) = e−a|x| ∈ S for every a > 0.
2
(1)
f (x) = e−a(1+|x| ) ∈ S for every a > 0.
2 a
(2)
(3) f (x) = e−|x| ∈
/ S.
(4) C0∞ (Rn ) ⊂ S(Rn ), where

C0∞ (Rn ) = { f ∈ C∞ (Rn ) : supp f compact in Rn }

and supp f = {x ∈ Rn : f (x) = 0}.

The space S(Rn ) is generated by a countable family of seminorms because | f |α ,β is

only a seminorm for α ≥ 0 and β > 0, i.e., the condition

| f |α ,β = 0 if and only if f =0

fails to hold for, e.g., a constant function f . The space (S, ρ ) is not normable but it
is a metric space if the metric ρ is defined by

| f − g|α ,β
ρ ( f , g) = ∑ 2−|α |−|β | ·
1 + | f − g|α ,β
.
α ,β ≥0

Exercise 16.1. Prove that ρ is a metric, that is,

(1) ρ ( f , g) ≥ 0 and ρ ( f , g) = 0 if and only if f = g.
(2) ρ ( f , g) = ρ (g, f ).
(3) ρ (g, h) ≤ ρ (g, f ) + ρ ( f , h).
Prove also that |ρ ( f , h) − ρ (g, h)| ≤ ρ ( f , g).

Theorem 16.3 (Completeness). The space (S, ρ ) is a complete metric space, i.e.,
every Cauchy sequence converges.
16 The Fourier Transform in Schwartz Space 135

Proof. Let { fk }∞k=1 , f k ∈ S, be a Cauchy sequence, that is, for every ε > 0 there
exists n0 (ε ) ∈ N such that

ρ ( fk , fm ) < ε , k, m ≥ n0 (ε ).

It follows that  
 
sup ∂ β ( fk − fm ) < ε
x∈K

for every β ≥ 0 and for every compact set K in Rn . This means that { fk }∞
k=1 is a
Cauchy sequence in the Banach space C|β | (K). Hence there exists a function f ∈
C|β | (K) such that
C|β | (K)
lim fk = f.
k→∞

Thus we may conclude that our function f is in C∞ (Rn ). It remains only to prove
that f ∈ S. It is clear that

sup |xα ∂ β f | ≤ sup |xα ∂ β ( fk − f )| + sup |xα ∂ β fk |

x∈K x∈K x∈K

≤ Cα (K) sup |∂ β ( fk − f )| + sup |xα ∂ β fk |.

x∈K x∈K

Taking k → ∞, we obtain

sup |xα ∂ β f | ≤ lim sup | fk |α ,β < ∞.

x∈K k→∞

The last inequality is valid, since { fk }∞

k=1 is a Cauchy sequence, so that | f k |α ,β
is bounded. The last inequality doesn’t depend on K, and we may conclude that
| f |α ,β < ∞, or f ∈ S. 
S
Definition 16.4. We say that fk → f as k → ∞ if

| fk − f |α ,β → 0, k→∞

for all α , β ≥ 0.
Exercise 16.2. Prove that C0∞ (Rn ) = S, that is, for every f ∈ S there exists { fk }∞
k=1 ,
S
fk ∈ C0∞ (Rn ), such that fk → f , k → ∞.
Now we are in position to define the Fourier transform in S(Rn ).
Definition 16.5. The Fourier transform F f (ξ ) or f(ξ ) of the function f (x) ∈ S is
defined by

F f (ξ ) ≡ f(ξ ) := (2π )−n/2 e−i(x,ξ ) f (x)dx, ξ ∈ Rn .
Rn
136 Part II: Fourier Transform and Distributions

Remark 16.6. This integral is well defined, since

  
 
 f (ξ ) ≤ cm (2π )−n/2 (1 + |x|)−m dx < ∞,
Rn

for m > n.

Next we prove the following properties of the Fourier transform:

(1) F is a linear continuous map from S into S.
β 
(2) ξ α ∂ξ f(ξ ) = (−i)|α |+|β | ∂xα (x β f (x)).

Indeed, we have

β
∂ξ f(ξ ) = (2π )−n/2 (−ix)β e−i(x,ξ ) f (x)dx
Rn

and hence 
  |x||β |
 β 
∂ξ f (ξ ) ≤ cm (2π )−n/2 dx < ∞
L∞ (Rn ) Rn (1 + |x|)m

if we choose m > n + |β |. At the same time we have obtained the formula

β 
∂ξ f(ξ ) = (−ix)β f (x). (16.1)

Further, integration by parts gives us


ξ f(ξ ) = (−i)|α | (2π )−n/2
α
e−i(x,ξ ) ∂xα f (x)dx,
Rn

from which we have the estimate

  
 α 
ξ f  ∞ ≤c |∂xα f (x)| dx < ∞,
L (Rn ) Rn

since ∂xα f (x) ∈ S for every α ≥ 0 if f (x) ∈ S. And also we have the formula


ξ α f = (−i)|α | ∂ α f . (16.2)
x

If we combine these last two estimates, we may conclude that F : S → S and F

is a continuous map (in the sense of the metric space (S, ρ )), since F maps every
bounded set from S again to a bounded set from S.
The formulas (16.1) and (16.2) show us that it is more convenient to use the
following notation:

∂
D j = −i∂ j = −i , Dα = Dα1 1 · · · Dαn n .
∂xj
16 The Fourier Transform in Schwartz Space 137

For this new derivative the formulas (16.1) and (16.2) can be rewritten as

Dξα f = (−1)|α | x
α f, ξ α f = D
α f.

Example 16.7. It is true that

F (e− 2 |x| )(ξ ) = e− 2 |ξ | .

1 2 1 2

Proof. The definition gives us directly


e−i(x,ξ )− 2 |x| dx
1 1
F (e− 2 |x| )(ξ ) = (2π )−n/2
2 2

Rn

= (2π )−n/2 e− 2 |ξ |
1 1 2 +2i(x,ξ )−|ξ |2 )
e− 2 (|x|
2
dx
Rn
n  ∞
= (2π )−n/2 e− 2 |ξ | ∏ e− 2 (t+iξ j ) dt.
1 2 1 2

j=1 −∞

z2
In order to calculate the last integral, we consider the function f (z) = e− 2 of the
complex variable z and the domain DR depicted in Figure 16.1.

DR
t
−R R

Fig. 16.1 Domain DR .

We consider the positive direction of going around the boundary ∂ DR . It is clear

that f (z) is a holomorphic function in this domain, and by Cauchy’s theorem we
have 
z2
e− 2 dz = 0.
∂ DR

But
  R  ξj
z2 t2
e− 2 (R+iτ ) dτ
1
e− 2 dz = e− 2 dt + i
2

∂ DR −R 0
 −R  0
− 12 (t+iξ j )2
e− 2 (−R+iτ ) dτ .
1 2
+ e dt + i
R ξj
138 Part II: Fourier Transform and Distributions

If R → ∞, then
 ξj
e− 2 (±R+iτ ) dτ → 0.
1 2

Hence  ∞  ∞
t2
e− 2 (t+iξ j ) dt =
1
e− 2 dt,
2
j = 1, . . . , n.
−∞ −∞

Using Fubini’s theorem and polar coordinates, we can evaluate the last integral as
 ∞

2   2π  ∞
t2 1 r2
e− 2 dt e− 2 (t
2 +s2 )
= dtds = dθ e− 2 rdr
−∞ R2 0 0
 ∞
−m
= 2π e dm = 2π .
0

Thus  ∞ √
e− 2 (t+iξ j ) dt =
1 2
2π
−∞

and
|x|2 n √
)(ξ ) = (2π )− 2 e− 2 |ξ | ∏ 2π = e− 2 |ξ | .
n 1 1
F(e−
2 2
2
j=1

This completes the proof. 

Exercise 16.3. Let P(D) be a differential operator,

P(D) = ∑ aα Dα ,
|α |≤m

 = P(ξ )
with constant coefficients. Prove that P(D)u u.
Definition 16.8. We adopt the following notation for translation and dilation of a
function:
(τh f )(x) := f (x − h), (σλ f )(x) := f (λ x), λ = 0.

Exercise 16.4. Let f ∈ S(Rn ), h ∈ Rn , and λ ∈ R, λ = 0. Prove that

(1) σ −n   −n 
λ f (ξ ) = |λ | σ 1 f (ξ ) and σλ f (ξ ) = |λ | σ 1 f (ξ );
λ λ

(2) τ −i(h,ξ ) f(ξ ) and τ f(ξ ) = e

h f (ξ ) = e
i(h,·) f (ξ ).
h

Exercise 16.5. Let A be a real-valued n × n matrix such that A−1 exists. Define
fA (x) := f (A−1 x). Prove that

fA (ξ ) = ( f)A (ξ )

if and only if A is an orthogonal matrix (a rotation), that is, AT = A−1 .

16 The Fourier Transform in Schwartz Space 139

Let us now consider f and g from S(Rn ). Then

  

(F f , g)L2 = 
f (ξ )g(ξ )dξ = (2π )−n/2
g(ξ ) e−i(x,ξ )
f (x)dx dξ
Rn Rn Rn
 

= (2π )−n/2 f (x) ei(x,ξ ) g(ξ )dξ dx = ( f , F ∗ g)L2 ,

Rn Rn

where F ∗ g(x) := F g(−x).

Remark 16.9. Here F ∗ is the adjoint operator (in the sense of L2 ), which maps S
into S since F : S → S. The inverse Fourier transform F −1 is defined as F −1 :=
F ∗.

In order to justify this definition we will prove the following theorem.

Theorem 16.10 (Fourier inversion formula). Let f be a function from S(Rn ). Then

F ∗F f = f .

To this end we will prove first the following (somewhat technical) lemma.

Lemma 16.11. Let f0 (x) be a function from L1 (Rn ) with Rn f0 (x)dx = 1 and let
f (x) be a function from L∞ (Rn ) that is continuous at {0}. Then
 x
lim ε −n f0 f (x)dx = f (0).
ε →0+ Rn ε

Proof. Since
 x  x
ε −n f0 f (x)dx − f (0) = ε −n f0 ( f (x) − f (0))dx,
Rn ε Rn ε

we may assume without loss of generality that f (0) = 0. Since f is continuous at

{0}, there exists δ > 0 for every η > 0 such that

η
| f (x)| <
 f 0 L 1

whenever |x| < δ . Note that

 
 
 
 Rn f0 (x)dx ≤  f0 L1 .

We may therefore conclude that

140 Part II: Fourier Transform and Distributions
 x    x 
  η  
 ε −n
f f (x)dx  ≤ · ε −n
 f0  dx
 Rn 0
ε   f0  1 |x|< δ ε
L
  x 
 
+  f L∞ ε −n  f0  dx
|x|>δ ε

η
≤ ·  f0 L1 +  f L∞ | f0 (y)| dy
 f0 L1 |y|> δε

= η +  f L∞ Iε .

But Iε → 0 as ε → 0+. This proves the lemma.

|x|2
Proof (Proof of

theorem 16.10). Let us consider v(x) = e− 2 . We know from Exam-
ple 16.7 that Rn v(x)dx = (2π )n/2 and Fv = v. If we apply Lemma 16.11 with
f0 = (2π )−n/2 v(x) and f ∈ S(Rn ), then by Exercise 16.4,
 x
n
(2π ) 2 f (0) = lim ε −n v
f (x)dx = lim f , ε −n σ 1 vL2
ε →0+ Rn ε ε →0+ ε
−n
= lim f , ε σ 1 F vL2 = lim f , F (σε v)L2
ε →0+ ε ε →0+

= lim F f , σε vL2 = F f , 1 = F f (ξ )ei(0,ξ ) dξ ,
ε →0+ Rn

where we have used Lebesgue’s dominated convergence theorem in the last step.
This proves that

f (0) = (2π )−n/2 F f (ξ )ei(0,ξ ) dξ = (F ∗ F f )(0).
Rn

The proof is now finished by


f (x) = (τ−x f )(0) = (F ∗ F (τ−x f ))(0) = (2π )−n/2 F (τ−x f )(ξ )ei(0,ξ ) dξ
Rn

= (2π )−n/2 ei(x,ξ ) F f (ξ )dξ = F ∗ F f (x),
Rn

where we have used Exercise 16.4. 

Corollary 16.12. The Fourier transform is an isometry (in the sense of L2 ).

Proof. The fact that the Fourier transform preserves the norm of f ∈ S follows from

F f 2L2 = (F f , F f )L2 = ( f , F ∗ F f )L2 = ( f , f )L2 =  f 2L2 .

This is called Parseval’s equality. 

Note that
(F f , g)L2 = ( f , F ∗ g)L2
16 The Fourier Transform in Schwartz Space 141

means that
  
f(ξ )g(ξ )dξ = f (x)F ∗ g(x)dx = f (x)F (g)(x)dx.
Rn Rn Rn

This implies that  

f(ξ )g(ξ )dξ = f (x)
g(x)dx,
Rn Rn

or
F f , gL2 (Rn ) = f , F gL2 (Rn ) .
Chapter 17
The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2

Let us begin with a preliminary proposition.

Proposition 17.1. Let X be a linear normed space and E ⊂ X a subspace of X such
that E = X; that is, the closure E of E in the sense of the norm in X is equal to X.
Let Y be a Banach space. If T : E → Y is a continuous linear map, i.e., there exists
M > 0 such that
TuY ≤ M uX , u ∈ E,

then there exists a unique linear continuous map Tex : X → Y such that Tex |E = T
and
Tex uY ≤ M uX , u ∈ X.

Exercise 17.1. Prove the previous proposition.

Lemma 17.2. Let 1 ≤ p < ∞. Then

Lp
C0∞ (Rn ) = L p (Rn ),

that is, C0∞ (Rn ) is dense in L p (Rn ) in the sense of L p -norm.

Proof. We will use the fact that the set of finite linear combinations of characteristic
functions of bounded measurable sets in Rn is dense in L p (Rn ), 1 ≤ p < ∞. This is
a well known fact from functional analysis.
Let now A ⊂ Rn be a bounded measurable set and let ε > 0. Then there exist a
closed set F and an open set Q such that F ⊂ A ⊂ Q and μ (Q \ F) < ε p (or only
μ (Q) < ε p if there is no closed set F ⊂ A). Here μ is the Lebesgue measure in Rn .
Let now ϕ be a function from C0∞ (Rn ) such that supp ϕ ⊂ Q, ϕ |F ≡ 1 and 0 ≤ ϕ ≤ 1;
see [19]. Then
 
ϕ − χA Lp p (Rn ) = |ϕ (x) − χA (x)| p dx ≤ 1dx = μ (Q \ F) < ε p ,
Rn Q\F

c Springer International Publishing AG 2017 143

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 17
144 Part II: Fourier Transform and Distributions

or
ϕ − χA L p (Rn ) < ε ,

where χA denotes the characteristic function of A, i.e.,


1, x ∈ A,
χA (x) =
0, x ∈
/ A.

Thus, we may conclude that C0∞ (Rn ) = L p (Rn ) for 1 ≤ p < ∞. 

Remark 17.3. Lemma 17.2 does not hold for p = ∞. Indeed, for a function f ≡ c0 =
0 and for every function ϕ ∈ C0∞ (Rn ), we have that

 f − ϕ L∞ (Rn ) ≥ |c0 | > 0.

Hence we cannot approximate a function from L∞ (Rn ) by functions from C0∞ (Rn ).
This means that ∞ L
C0∞ (Rn ) = L∞ (Rn ).

But the following result holds:

L∞
Exercise 17.2. Prove that S(Rn ) = Ċ(Rn ), where

Ċ(Rn ) := { f ∈ C(Rn ) : lim f (x) = 0}.

|x|→∞

Now we are in a position to extend F from S ⊂ L1 to L1 .

Theorem 17.4 (Riemann–Lebesgue lemma). Let F : S → S be the Fourier trans-

form in Schwartz space S(Rn ). Then there exists a unique extension Fex as a map
from L1 (Rn ) to Ċ(Rn ) with norm Fex L1 →L∞ = (2π )−n/2 .

Proof. We know that F f L∞ ≤ (2π )−n/2  f L1 for f ∈ S. Now we apply the pre-
L1
liminary proposition to E = S, X = L1 , and Y = L∞ . Since S = L1 (which follows
L1
from C0∞ ⊂ S and C0∞ = L1 ) for every f ∈ L1 (Rn ), there exists { fk } ⊂ S such that
 fk − f L1 → 0 as k → ∞. In that case, we can define

L∞
Fex f := lim F fk .
k→∞

L∞
Since S = Ċ (see Exercise 17.2), it follows that Fex f ∈ Ċ and Fex L1 →L∞ ≤
(2π )−n/2 . On the other hand,

F f L∞ ≥ | f(0)| = (2π )−n/2  f L1

17 The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2 145

for f ∈ L1 and f ≥ 0. Hence Fex L1 →L∞ = (2π )−n/2 . 

Proof (Alternative proof). If f ∈ L1 (Rn ), then we can define the Fourier transform
F f (ξ ) directly by

F f (ξ ) := (2π )−n/2 e−i(x,ξ ) f (x)dx,
Rn

since   
 
 e−i(x,ξ )
f (x)dx ≤ | f (x)|dx =  f L1 .
 Rn  Rn

Also we have
   
   
(2π )n/2  f(ξ + h) − f(ξ ) = sup  e−i(ξ ,x) −i(h,x)
(e − 1) f (x)dx 
L∞ (Rn )  Rn

n ξ ∈R
 
≤2 ε
| f (x)|dx + ε | f (x)|dx =: I1 + I2 .
|x|> |h| |x||h|≤ε

Here we have used the fact that |eiy − 1| ≤ |y| for y ∈ R with |y| ≤ 1. It is easily seen
that I1 → 0 for |h| → 0 and I2 → 0 for ε → 0, since f ∈ L1 (Rn ).
This means that the Fourier transform f(ξ ) is continuous (even uniformly con-
tinuous) on Rn . Moreover, we have
   
ξπ
2 f(ξ ) = (2π )−n/2 e −i(x,ξ )
f (x) − f x + 2 dx.
Rn |ξ |

This equality follows from


f(ξ ) = −(2π )−n/2 eiπ e−i(x,ξ ) f (x)dx
Rn
 
 i ξ , ξ2 π
−n/2 −i(x,ξ )
= −(2π ) e |ξ |
e f (x)dx
Rn

and Exercise 16.4. Thus,

  
 ξπ 
2| f(ξ )| ≤ (2π )−n/2 
 f (x) − f x +  →0
|ξ |2 L1 (Rn )

for |ξ | → ∞. 

Theorem 17.5 (Plancherel). Let F : S → S be the Fourier transform in S with

F f L2 =  f L2 . Then there exists a unique extension Fex of F to L2 -space such
onto
that Fex : L2 → L2 and Fex L2 →L2 = 1. Also Parseval’s equality remains valid.
146 Part II: Fourier Transform and Distributions

L2 L2
Proof. We know that S = L2 , since C0∞ = L2 . Thus for every f ∈ L2 (Rn ) there exists
{ fk }∞
k=1 ⊂ S(R ) such that  f k − f L2 (Rn ) → 0 as k → ∞. By Parseval’s equality in
n

S we get
F fk − F fl L2 =  fk − fl L2 → 0, k, l → ∞.

L2
Thus {F fk }∞
k=1 is a Cauchy sequence in L (R ), and therefore, F f k → g, where
2 n

g ∈ L . Therefore, we may put Fex f := g. Also we have Parseval’s equality

2

Fex f L2 = lim F fk L2 = lim  fk L2 =  f L2 ,

k→∞ k→∞

which proves the statement about the operator norm. 

∗F f =
Remark 17.6. In L2 -space we also have the Fourier inversion formula Fex ex
−1
f , or Fex Fex f = f .

Exercise 17.3. Prove that if f ∈ L2 (Rn ), then

L2
(1) Fex f (ξ ) = lim F fR (ξ ), where fR (x) = χ{x:|x|≤R} (x) f (x).
R→+∞
L2
(2) Fex f (ξ ) = lim F (e−ε |x| f ).
ε →0+

Exercise 17.4. Let us assume that f ∈ L1 (Rn ) and F f (ξ ) ∈ L1 (Rn ). Prove that

f (x) = (2π )−n/2 ei(x,ξ ) F f (ξ )dξ = F −1 F f (x).
Rn

This means that the Fourier inversion formula is valid.

Exercise 17.5. Let f1 and f2 belong to L2 (Rn ). Prove that

( f1 , f2 )L2 = (F f1 , F f2 )L2 .

Theorem 17.7 (Riesz–Thorin interpolation theorem). Let T be a linear contin-

uous map from L p1 (Rn ) to Lq1 (Rn ) with norm estimate M1 and from L p2 (Rn ) to
Lq2 (Rn ) with norm estimate M2 . Then T is a linear continuous map from L p (Rn ) to
Lq (Rn ) with p and q such that

1 θ 1−θ 1 θ 1−θ
= + , = + , 0 ≤ θ ≤ 1,
p p1 p2 q q1 q2

with norm estimate M1θ M21−θ .

Proof. Let F and G be two functions with the following properties:

(1) F, G ≥ 0,
(2) FL1 = GL1 = 1.
17 The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2 147

Let us consider now the function Φ (z) of a complex variable z ∈ C given by

 z
z + 1−z  + 1−z

Φ (z) := M1−z M2−1+z T ( f0 F p1 p 2 )(x)g0 G q1 q2
(x)dx,
Rn

where 1
q1 + q1 = 1, 1
q2 + q1 = 1, | f0 | ≤ 1, and |g0 | ≤ 1. The two functions f0 and g0
1 2
will be selected later. We assume also that 0 ≤ Re(z) ≤ 1.
Our aim is to prove the inequality

|T f , g L2 | ≤ M1θ M21−θ  f L p gLq ,

where
1 θ 1−θ 1 θ 1−θ 1 1
= + , = + , + = 1.
p p1 p2 q q1 q2 q q

z z + 1−z
+ 1−z  
Since T is a linear continuous map and F p1 p2 , G q1 q2 are holomorphic func-
tions with respect to z (consider az = ez log a , a > 0), we may conclude that Φ (z) is a
holomorphic function also.
(1) Let us assume now that Re(z) = 0, i.e., z = iy. Then we have
iy
iy
+ 1−iy + 1−iy
Φ (iy) = M1−iy M2−1+iy T ( f0 F p1
 
p 2 ), g0 G q1 q2
L2 .

Since |aix | = 1 for a, x ∈ R, a > 0, it follows from Hölder’s inequality and the
assumptions on T that
   
 iy 1−iy   iy
+ 1−iy 
p + p
|Φ (iy)| ≤ M2−1 M2 
 f0 F 1 2


g0 G q1 q2 
  
q
L p2 L 2
   
 1   1  1 1
q
= p2  |g0 |G q2  p
|
 0f |F    
≤ FL12 GL12 = 1.
q
L p2 L 2

(2) Let us assume now that Re(z) = 1, i.e., z = 1 + iy. Then we have similarly that
   
 1+iy −iy   1+iy −iy 
+
p + p2 
|Φ (1 + iy)| ≤ M1−1 M1 
 f0 F 1  p
g0 G q1 q2 
  
q
L 1 L 1
   
 1   1  1 1
q1
= p1  |g0 |G q1  p
|
 0f |F    ≤ FL11 G L1
= 1.
q
L p1 L 1

If we apply now the Phragmén–Lindelöf theorem for the domain 0 < Re(z) < 1, we
obtain that |Φ (z)| ≤ 1 for every z such that 0 < Re(z) < 1. Then |Φ (θ )| ≤ 1 also for
0 < θ < 1. But this is equivalent to the estimate
148 Part II: Fourier Transform and Distributions

1 1
|T ( f0 F p ), g0 G q L2 | ≤ M1θ M21−θ , (17.1)

θ θ 1 θ θ
where 1
p = p1 + 1−
p2 , q = q1 + 1−
q2 and
1
q + q1 = 1. In order to finish the proof of

this theorem let us choose (for arbitrary functions f ∈ L p and g ∈ Lq with p and q
as above) the functions F, G, f0 and g0 as follows:

F = | f1 | p , G = |g1 |q , f0 = sgn f1 , g0 = sgn g1 ,

f g
where f1 =  f L p , g1 = g q , and
L

⎧
⎪
⎨1, f > 0,
sgn f = 0, f = 0,
⎪
⎩
−1, f < 0.

1 1
In that case, f1 = f0 F p and g1 = g0 G q . Applying the estimate (17.1), we obtain
    

 f g 
 T ,  ≤ M1θ M21−θ ,
  f L p gLq L2 

which is equivalent to

|T f , g L2 | ≤ M1θ M21−θ  f L p gLq .

This implies the desired final estimate

T f Lq ≤ M1θ M21−θ  f L p

and finishes the proof. 

Theorem 17.8 (Hausdorff–Young). Let F : S → S be the Fourier transform in

Schwartz space. Then there exists a unique extension Fex as a linear continuous
map

Fex : L p (Rn ) → L p (Rn ),

where 1 ≤ p ≤ 2 and 1
p + 1
p = 1. What is more, we have the norm estimate
 
−n 1p − 12
Fex L p →L p ≤ (2π ) .

This is called the Hausdorff–Young inequality.

Proof. We know from Theorems 17.4 and 17.5 that there exists a unique extension
Fex of the Fourier transform from S to S for spaces:
17 The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2 149

n
(1) Fex : L1 (Rn ) → L∞ (Rn ) with norm estimate M1 = (2π )− 2 ;
(2) Fex : L2 (Rn ) → L2 (Rn ) with norm estimate M2 = 1.
Applying now Theorem 17.7, we obtain that Fex : L p → Lq , where

1 θ 1−θ 1 θ 1 θ 1−θ 1 θ
= + = + , = + = − .
p 1 2 2 2 q ∞ 2 2 2

It follows that
1 1
+ = 1,
p q

i.e., q = p and θ = 2p − 1. For θ to satisfy the condition 0 ≤ θ ≤ 1 we get 1 ≤ p ≤ 2.

We may also conclude that
 
− 2n θ 1−θ −n 1p − 21
Fex L p →L p ≤ ((2π ) ) 1 = (2π )

to obtain the desired norm estimate. 

Remark 17.9. In order to obtain Fex in L p (Rn ), 1 ≤ p ≤ 2, constructively we can

apply the following procedure. Let us assume that f ∈ L p (Rn ), 1 ≤ p ≤ 2, and
{ fk }∞
k=1 ⊂ S(R ) such that
n

 fk − f L p (Rn ) → 0, k → ∞.

It follows from the Hausdorff–Young inequality that

F fk − F fl L p (Rn ) ≤ Cn  fk − fl L p (Rn ) .


This means that {F fk }∞
k=1 is a Cauchy sequence in L (R ). We can therefore define
p n


Lp
Fex f := lim F fk .
k→∞

And we also have the Hausdorff–Young inequality

Fex f L p = lim F fk L p ≤ lim Cn  fk L p = Cn  f L p .

k→∞ k→∞

Example 17.10 (Fourier transform on the line). Let f2 (x) = 1

(x−iε )2
, where ε > 0 is
fixed. It is clear that f2 ∈ L1 (R) and
 +∞ −ixξ
1 e dx
f2 (ξ ) = √ .
2π −∞ (x − iε )2
150 Part II: Fourier Transform and Distributions

e −izξ
In order to calculate this integral we consider the function F(z) := (z−i ε )2
of a com-
plex variable z ∈ C. It is easily seen that z = iε , ε > 0, is a pole of order 2. We
consider the cases ξ > 0 and ξ < 0 separately; see Figure 17.1.

Im z Im z

i i
−R D+
R
R
Re z Re z
−R R
D−
R

Fig. 17.1 Domains D− +

R and DR of integration.

(1) Let ξ > 0. By Cauchy’s theorem we have

  R 
−
F(z)dz = 0 = F(z)dz + F(z)dz =: I1 + I2 .
∂ DR −R
|z|=R
Imz<0

It follows that  +∞ −ixξ

e dx
I1 → , R → ∞,
−∞ (x − iε )2

and
I2 → 0, R→∞

due to Jordan’s lemma, since ξ Imz < 0. We therefore may conclude that
 +∞ −ixξ
e dx
=0
−∞ (x − iε )2

for ξ > 0.
(2) Let ξ < 0. In this case again ξ Imz < 0. So we may apply Jordan’s lemma again
and obtain
  R −ixξ 
e dx
F(z)dz = + F(z)dz = 2π iRes F(z).
∂ D+
R −R (x − iε )2 z=iε
|z|=R
Imz>0
17 The Fourier Transform in L p (Rn ), 1 ≤ p ≤ 2 151

Hence  ∞ −ixξ
e dx
= 2π i((z − iε )2 F(z)) |z=iε = 2πξ eεξ .
−∞ (x − iε )2

If we combine these two cases, we obtain

 1 √
(ξ ) = 2πξ H(−ξ )eεξ ,
(x − iε )2

where 
1, t ≥ 0,
H(t) =
0, t < 0,

is the Heaviside function. Similarly we obtain

 1 √
(ξ ) = − 2πξ H(ξ )e−εξ .
(x + iε )2

Example 17.11. Let f1 (x) = x−i ε , where ε > 0 is fixed. It is clear that f 1 ∈
/ L1 (R),
1

but f1 ∈ L (R), 1 < p ≤ 2. Analogously to Example 17.10 we obtain

p

 √
1 −i 2π H(ξ )e−εξ , ξ = 0,
(ξ ) = 
x + iε −i π2 , ξ = 0,

and √
1 i 2π H(−ξ )eεξ , ξ = 0,
(ξ ) = 
x − iε i π2 , ξ = 0.

Exercise 17.6. Find the Fourier transforms of the following functions on the line.

e−x , x > 0,
(1) f (x) =
0, x ≤ 0,
(2) f (x) = e−|x| and f (x) = 1
1+x2
,
(3) f3 (x) = (x±i1ε )3 , ε > 0.

Exercise 17.7. Define the Laplace transform by

 ∞
L(p) := f (x)e−px dx,
0

where | f (x)| ≤ Meax , x > 0, f (x) = 0, x < 0, and p = p1 + ip2 , p1 > a. Prove that
152 Part II: Fourier Transform and Distributions

√

(1) L(p) = 2π f (x)e −p1 x (p ).
2
(2) Apply the Fourier inversion formula to prove the Mellin formula
 p1 +i∞
1
f (x) = L(p)e px dp, p1 > a.
2π i p1 −i∞
Chapter 18
Tempered Distributions

In this chapter we will consider two types of distributions: Schwartz distributions

and tempered distributions. To that end we consider the space D := C0∞ (Rn ) of test
functions. It is clear that D is a linear space and D ⊂ S. A notion of convergence is
given in the following definition.
Definition 18.1. A sequence {ϕk }∞
k=1 is a null sequence in D if

(1) there exists a compact set K ⊂ Rn such that supp ϕk ⊂ K for every k and
(2) for every α ≥ 0 we have

sup |Dα ϕk (x)| → 0, k → ∞.

x∈K

D D D
We denote this fact by ϕk → 0. As usual, ϕk → ϕ ∈ D means that ϕk − ϕ → 0.
Now we are in a position to define the Schwartz distribution space.
Definition 18.2. A functional T : D → C is a Schwartz distribution if it is linear
and continuous, that is,
(1) T (α1 ϕ1 + α2 ϕ2 ) = α1 T (ϕ1 ) + α2 T (ϕ2 ) for every ϕ1 , ϕ2 ∈ D and α1 , α2 ∈ C
(2) for every null sequence ϕk in D, one has T (ϕk ) → 0 in C as k → ∞.
The linear space of Schwartz distributions is denoted by D  . The action of T on ϕ
is denoted by T (ϕ ) = T, ϕ .

Example 18.3. Every locally integrable function f , that is, f ∈ Lloc

1 (Rn ), defines a

Schwartz distribution by the formula


T f , ϕ  := f (x)ϕ (x)dx.
Rn

c Springer International Publishing AG 2017 153

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 18
154 Part II: Fourier Transform and Distributions

It is clear that T f is a linear map. It remains only to prove that T f is a continuous

map on D. Let {ϕk }∞ k=1 be a null sequence in D. Then

|T f , ϕk | ≤ sup |ϕk (x)| | f (x)|dx → 0, k→∞
x∈K K

by the definition of null sequence.

D
Example 18.4. If T, ϕ  := ϕ (0), then T ∈ D  . Indeed, T is linear, and if ϕk → 0,
then T, ϕk  = ϕk (0) → 0 for k → ∞. This distribution is called the δ -function and
is denoted by δ , i.e.,
δ , ϕ  = ϕ (0), ϕ ∈ D.

Remark 18.5. A distribution T is regular if it can be written in the form


T, ϕ  = f (x)ϕ (x)dx
Rn

for some locally integrable function f . All other distributions are singular.

Exercise 18.1. Prove that δ is a singular distribution.

Definition 18.6. The functional T defined by

  ∞
ϕ (x) ϕ (x)
T, ϕ  := lim dx ≡ p. v. dx
ε →0+ |x|>ε x −∞ x

on D(R) is called the principal value of 1x . We denote it by T = p. v. 1x .

Remark 18.7. Note that 1

x ∈
/ Lloc
1 (R), but we have the following result.

Exercise 18.2. Prove that

 ∞  ∞
1 ϕ (x) − ϕ (−x) ϕ (x) − ϕ (0)
p. v. , ϕ  = dx = p. v. dx.
x 0 x −∞ x

Example 18.8. Let σ be a hypersurface of dimension n − 1 in Rn and let dσ stand

for an element of surface area on σ . Consider the functional

T, ϕ  = a(x)ϕ (x)dσ
σ

on D, where a(x) is a locally integrable function on σ . We can interpret T in terms

of surface source. Indeed,
  
 a(ξ )δ (x − ξ )dσξ , ϕ  := a(ξ )δ (x − ξ ), ϕ (x)dσξ = a(ξ )ϕ (ξ )dσξ .
σ σ σ
18 Tempered Distributions 155

It is easy to see that T is a singular distribution. This distribution is known as the

simple layer.

Definition 18.9. If T ∈ D  and g ∈ C∞ (Rn ), then we may define the product gT by

gT, ϕ  := T, gϕ , ϕ ∈ D.

This product is well defined, because gϕ ∈ D.

If f is a locally integrable function whose derivative ∂∂xfj is also locally integrable,

then
     
∂f ∂f ∂ϕ ∂ϕ
,ϕ = ϕ (x)dx = − f dx = − f , , ϕ ∈D
∂xj Rn ∂ x j Rn ∂ x j ∂xj

by integration by parts. This property is used to define the derivative of any distrib-
ution.
Definition 18.10. Let T be a distribution from D  . For a multi-index α we define
the derivative ∂ α T by

∂ α T, ϕ  := T, (−1)|α | ∂ α ϕ , ϕ ∈ D.

It is easily seen that ∂ α T ∈ D  .

Example 18.11. Consider the Heaviside function H(x). Since H ∈ Lloc

1 (R),

it follows that
 ∞
H  , ϕ  = −H, ϕ   = − ϕ  (x)dx = ϕ (0) = δ , ϕ .
0

Hence H  = δ .

Example 18.12. Let us prove that (log |x|) = p. v. 1x in the sense of Schwartz distri-
butions. Indeed,
 ∞
(log |x|) , ϕ  = −log |x|, ϕ   = − log(|x|)ϕ  (x)dx
−∞
 ∞  0
=− log(x)ϕ  (x)dx − log(−x)ϕ  (x)dx
0 −∞
 ∞  ∞
=− log(x)(ϕ  (x) + ϕ  (−x))dx = − log(x)(ϕ (x) − ϕ (−x)) dx
0 0
 ∞
ϕ (x) − ϕ (−x) 1
= − log(x) [ϕ (x) − ϕ (−x)]∞
0 + dx = p. v. , ϕ 
0 x x

by integration by parts and Exercise 18.2.

156 Part II: Fourier Transform and Distributions

Exercise 18.3. Prove that

 
1  1
p. v. = − p. v. 2 ,
x x

where p. v. x12 is defined as


1 ϕ (x) − ϕ (0)
p. v. , ϕ  = lim dx.
x2 ε →0+ |x|>ε x2

The following characterization of D  is given without proof: T ∈ D  if and only if

for every compact K ⊂ Rn there exists n0 (K) ∈ N0 such that

|T, ϕ | ≤ C0 ∑ sup |Dα ϕ |

|α |≤n0 x∈K

for every ϕ ∈ D with supp ϕ ⊂ K.

Definition 18.13. A functional T : S → C is a tempered distribution if
(1) T is linear, i.e., T, αϕ + β ψ  = α T, ϕ  + β T, ψ  for all α , β ∈ C and
ϕ , ψ ∈ S.
(2) T is continuous on S, i.e., there exist n0 ∈ N0 and a constant c0 > 0 such that

|T, ϕ | ≤ c0 ∑ |ϕ |α ,β
|α |,|β |≤n0

for every ϕ ∈ S.
The space of tempered distributions is denoted by S . In addition, for Tk , T ∈ S the
S C
convergence Tk → T means that Tk , ϕ  → T, ϕ  for all ϕ ∈ S.

Remark 18.14. Since D ⊂ S, the space of tempered distributions is narrower than

the space of Schwartz distributions, S ⊂ D  . Later we will consider the even nar-
rower distribution space E  , which consists of continuous linear functionals on the
(widest test function) space E := C∞ (Rn ). In short, D ⊂ S ⊂ E implies that

E  ⊂ S ⊂ D  .

It turns out that members of E  have compact support, and they are therefore called
distributions with compact support. But more on that later.

Example 18.15. Let us consider R1 .

(1) It is clear that f (x) = e|x| is a Schwartz distribution but not a tempered distrib-
2

ution, because part (2) of the previous definition is not satisfied.

m
(2) If f (x) = ∑ ak xk is a polynomial, then f (x) ∈ S , since
k=0
18 Tempered Distributions 157
 
 m 
 
|T f , ϕ | =  ∑ ak x ϕ (x)dx
k
 R k=0 
m 
≤ ∑ |ak | R
(1 + |x|)−1−δ (1 + |x|)1+δ |x|k |ϕ (x)|dx
k=0
m 
≤ C ∑ |ak ||ϕ |0,k+1+δ (1 + |x|)−1−δ dx,
k=0 R

so condition (2) is satisfied, e.g., for δ = 1, n0 = m + 2. This polynomial is a

regular distribution, since T f , ϕ  = R f (x)ϕ (x)dx is well defined.
Definition 18.16. Let T be a distribution from D  . Then the support of T is defined
by
supp T := Rn \ A,

where A = {x ∈ Rn : T, ϕ  = 0 for all ϕ ∈ C∞ with supp ϕ ⊂ Uδ (x)}.

Exercise 18.4. Prove that
(1) if f is continuous, then supp T f = supp f ;
(2) supp(∂ α T ) ⊂ supp T ;
(3) supp δ = {0}.
Example 18.17. (1) The weighted Lebesgue spaces are defined as
 1
p
Lσp (Rn ) := { f ∈ Lloc
p
(Rn ) :  f Lσp := (1 + |x|)σ p | f (x)| p dx < ∞}
Rn

for 1 ≤ p < ∞ and

Lσ∞ (Rn ) := { f ∈ Lloc

∞
(Rn ) :  f Lσ∞ := ess sup(1 + |x|)σ | f (x)| < ∞}.
x∈Rn

1 (Rn ) for some δ > 0, then T ∈ S . In fact,

If f ∈ L−δ f

 
 

|T f , ϕ | =  f ϕ dx ≤  f L1 ϕ L∞ .
Rn −δ δ


This means that R f ϕ dx is well defined in this case and

T f , ϕ  := f ϕ dx.
Rn

(2) If f ∈ L p , 1 ≤ p ≤ ∞, then f ∈ S . Indeed,

n
L p (Rn ) ⊂ L− δ (R ) for δ > ,
1 n
p
158 Part II: Fourier Transform and Distributions

where 1
p + 1
p = 1. This follows from Hölder’s inequality

   1
 p
(1 + |x|)−δ | f (x)|dx ≤ (1 + |x|)−δ p dx  f L p .
R R

(3) Let T ∈ S and ϕ0 (x) ∈ C0∞ (Rn ) with ϕ0 (0) = 1. The product ϕ0 k T is well
defined in S by  x   x 
ϕ0 T, ϕ := T, ϕ0 ϕ .
k k

If we consider the sequence Tk := ϕ0 kx T , then

k→∞ S S
(a) Tk , ϕ  ≡ T, ϕ0 ( kx )ϕ  → T, ϕ  (since ϕ0 ( kx )ϕ → ϕ ), so that Tk → T .
(b) Tk has compact support as a tempered distribution. This fact follows from
the compactness of support of ϕk = ϕ0 ( kx ).

Now we are ready to prove a more serious and more useful fact.
Theorem 18.18. Let T ∈ S . Then there exists Tk ∈ S such that

Tk , ϕ  = Tk (x)ϕ (x)dx → T, ϕ , k → ∞,
Rn

S
where ϕ ∈ S. In short, S = S .

Proof. Let j(x) be a function from D ≡ C0∞ (Rn ) with Rn j(x)dx = 1 and j(−x) =
j(x). Let jk (x) := kn j(kx). By Lemma 16.11 we have

lim  jk , ϕ  = lim jk (x)ϕ (x)dx = ϕ (0)
k→∞ k→∞ Rn

S
for every ϕ ∈ S. That is, jk (x) → δ (x).
The convolution of two integrable functions g and ϕ is defined by

(g ∗ ϕ )(x) := g(x − y)ϕ (y)dy.
Rn

If h and g are integrable functions and ϕ ∈ S, then it follows from Fubini’s theorem
that
   
h ∗ g, ϕ  = ϕ (x)dx h(x − y)g(y)dy = g(y)dy h(x − y)ϕ (x)dx
R R Rn Rn
n n

= g(y)dy Rh(y − x)ϕ (x)dx = g, Rh ∗ ϕ ,

Rn Rn

where Rh(z) := h(−z) is the reflection of h.

18 Tempered Distributions 159

Let now ϕ0 (x)  

x
∈ D with ϕ0 (0) = 1. For every T ∈ S let us put Tk := jk ∗ Tk ,
where Tk = ϕ0 k T . From the above considerations we know that  jk ∗ Tk , ϕ  =
Tk , R jk ∗ ϕ .
Let us prove that this Tk meets the requirements of the theorem. First of all,
x
Tk , ϕ  ≡  jk ∗ Tk , ϕ  = Tk , R jk ∗ ϕ  = ϕ0 T, jk ∗ ϕ 
x k
= T, ϕ0 ( jk ∗ ϕ ) → T, ϕ , k → ∞,
k
because

S
(a) ϕ0 kx → 1 pointwise for k → ∞, since ϕ0 (0) = 1 and ϕ0 ( kx )ϕ → ϕ ,
S
(b) jk ∗ ϕ → ϕ for k → ∞ by Lemma 16.11:
 
jk (x − y)ϕ (y)dy = jk (z)ϕ (x − z)dz → ϕ (x).
Rn Rn

Finally, jk (x) ∈ C0∞ (Rn ) implies that Tk ∈ C0∞ (Rn ) ⊂ S also. 

Definition 18.19. Let us assume that L : S → S is a continuous linear map. The

adjoint map L : S → S is defined by

L T, ϕ  := T, Lϕ , T ∈ S .

Clearly, L is also a continuous linear map.

Corollary 18.20. Every continuous linear map (operator) L : S → S admits a con-

 : S → S .
tinuous linear extension L
S
Proof. If T ∈ S , then by Theorem 18.18 there exists Tk ∈ S such that Tk → T . Then

 ϕ
LTk , ϕ  = Tk , L ϕ  → T, L ϕ  =: LT,

as k → ∞. 

Now we are in a position to formulate the following theorem.

Theorem 18.21 (Properties of tempered distributions). The following linear
continuous operators from S into S admit unique continuous linear extensions as
maps from S into S :
(1) uT, ϕ  := T, uϕ , u ∈ S;
(2) ∂ α T, ϕ  := T, (−1)|α | ∂ α ϕ ;
(3) τh T, ϕ  := T, τ−h ϕ ;
160 Part II: Fourier Transform and Distributions

(4) σλ T, ϕ  := T, |λ |−n σ 1 ϕ , λ = 0;

λ
(5) F T, ϕ  := T, F ϕ .

Proof. See Theorem 18.18, Definition 18.19, and Corollary 18.20. 

Remark 18.22. Since F −1 F T, ϕ  = F T, F −1 ϕ  = T, F F −1 ϕ  = T, ϕ , we

have that F −1 F = F F −1 = I in S .

Example 18.23. (1) Since

 
ei(0,ξ ) F ϕ dξ
n n
F 1, ϕ  ≡ 1, F ϕ  = (F ϕ )(ξ )dξ = (2π ) 2 (2π )− 2
Rn Rn
n n n
= (2π ) 2 F −1 F ϕ (0) = (2π ) 2 ϕ (0) = (2π ) 2 δ , ϕ 

for every ϕ ∈ S, we have that

 n
1 = (2π ) 2 δ

in S .
(2) δ = (2π )− 2 · 1, since for ϕ ∈ S we have
n


δ, ϕ  = δ , F ϕ  = F ϕ (0) = (2π )− 2
n n
e−i(0,x) ϕ (x)dx = (2π )− 2 1, ϕ .
Rn

n
Moreover, F −1 δ = (2π )− 2 · 1 in S .
 x2 ξ2
(3) e−a 2 = a− 2 e− 2a , Re a ≥ 0, a = 0. Indeed, for a > 0 we know that
n

√
x2 ( ax)2 n ξ2
F (e−a 2 ) = F (e− 2 ) = a− 2 e− 2a .

If a is such that Re a ≥ 0, a = 0, then we can use analytic continuation of these

formulas.
(4) Consider (1 − Δ )u = f , where Δ = ∂∂x2 + · · · + ∂∂x2 is the Laplacian in Rn and
2 2

1 n
and u, f ∈ S . This equation can be solved in S using the Fourier transform.
Indeed, we get
u = f,
(1 + |ξ |2 )

or
u = (1 + |ξ |2 )−1 f,

or
u = F −1 ((1 + |ξ |2 )−1 F f ).
18 Tempered Distributions 161

If f ∈ S, then F f ∈ S and (1 + |ξ |2 )−1 F f ∈ S also, and then u ∈ S exists. If

S
f ∈ S , then by Theorem 18.18 there exists fk ∈ S such that fk → f . We conclude
that
S
u = lim uk ,
k→∞

where uk = F −1 ((1 + |ξ |2 )−1 F fk ).

Exercise 18.5. Let P(D) be an elliptic partial differential operator

P(D) = ∑ aα Dα
|α |≤m

with constant coefficients and P(ξ ) = 0 for ξ = 0. Prove that if u ∈ S and Pu = 0,

then u is a polynomial.

Corollary 18.24. If Δ u = 0 in S and |u| is less than or equal to some constant, then
u is constant.

Exercise 18.6. Prove that


(1) F (p. v. 1x ) = −i π2 sgn ξ ;

(2) F (p. v. x12 ) = − π2 |ξ |.

Definition 18.25. Let us introduce the tempered distributions

1 1
:= lim
x ± i0 ε →0+ x ± iε

(if they exist), i.e.,

   
1 1
,ϕ = lim ,ϕ , ϕ ∈ S.
x ± i0 ε →0+ x ± iε

In a similar fashion,
1 1
:= lim
(x ± i0)2 ε →0+ (x ± iε )2

in S (if they exist).

Example 18.26. We know from Example 17.11 that

 √
1 −i 2π H(ξ )e−εξ , ξ = 0,
(ξ ) = 
x + iε −i π2 , ξ = 0,
162 Part II: Fourier Transform and Distributions

and √
1 i 2π H(−ξ )eεξ , ξ = 0,
(ξ ) = 
x − iε i π2 , ξ = 0.

Hence  √
1 1 −i 2π H(ξ ), ξ = 0,
= lim = 
x + i0 ε →0+ x + iε −i π2 , ξ = 0,

and √
1 1 i 2π H(−ξ ), ξ = 0,
= lim = 
x − i0 ε →0+ x − iε i π2 , ξ = 0.

It follows from Exercise 18.6 that

  
1 1 √ π  1
+ = −i 2π sgn ξ = 2 −i sgn ξ = 2p. v.
x + i0 x − i0 2 x

and thus
1 1 1
+ = 2 p. v. .
x + i0 x − i0 x

In a similar fashion,

1 1 √ √ √
− = i 2π · 1 = i 2π 2π δ = 2π iδ,
x − i0 x + i0

and so
1 1
− = 2π iδ .
x − i0 x + i0

We add and subtract to get finally

1 1 1 1
= p. v. − iπδ and = p. v. + iπδ .
x + i0 x x − i0 x

Exercise 18.7. Prove that

(1)
 1 √  1 √
= − 2πξ H(ξ ) and = 2πξ H(−ξ );
(x + i0)2 (x − i0)2

(2)

1 1 1 1 1
+ = 2 p. v. 2 and − = −2π iδ  ;
(x + i0)2 (x − i0)2 x (x − i0)2 (x + i0)2
18 Tempered Distributions 163

(3)
1 1 1 1
= p. v. 2 + π iδ  and = p. v. 2 − π iδ  ;
(x + i0)2 x (x − i0)2 x

(4) 
 π 1
log |x| = − p. v. ;
2 |ξ |

(5)
xβ = (2π )n/2 i|β | ∂ β δ .

Exercise 18.8. Prove that

(1)
 ξ ) = − √i · 1 ;
H(
2π ξ − i0

(2)
i 1
 ξ ) = −  π p. v. .
sgn(
2
ξ

Example 18.27. Since

∂  = (−1)|α | δ , ∂ α ϕ = δ , i|α | ξ

α δ , ϕ  = ∂ α δ , ϕ α ϕ

= δ, (iξ )α ϕ  = (2π )− 2 , (iξ )α ϕ  = (2π )− 2 (iξ )α , ϕ ,

n n

we get
∂
α δ = (2π )− 2 (iξ )α .
n

In particular, in dimension one,

1 k k √ k (k)
δ
(k) = √ iξ , xk = 2π i δ .
2π

Let us consider the Cauchy–Riemann operator

 
1 ∂ ∂
∂ := +i
2 ∂x ∂y

and  
1 ∂ ∂
∂ := −i
2 ∂x ∂y

in R2 .
164 Part II: Fourier Transform and Distributions

Let us prove the following facts about these operators:

(1)
1
∂ · ∂¯ = ∂¯ · ∂ = Δ ,
4
(2)  
1¯ 1
∂ =δ in R2 .
π z

The last fact means that

1 1
π x + iy

is the fundamental solution (see Chapter 22) of ∂¯ . Taking the Fourier transform of
(2) gives us
1
∂ (ξ ) = π δ(ξ ),
z

which is equivalent to

1 
1 1
(iξ1 − ξ2 ) · (ξ ) = π · (2π )−1 · 1 = ,
2 z 2
or

1 1 1
(ξ ) = = −i , ξ = 0.
z iξ1 − ξ2 ξ1 + iξ2

Let us check that this is indeed the case. We have, by Example 17.11,

    ∞ −iξ1 x
1 1 e−i(ξ1 x+ξ2 y) 1 ∞ −iξ2 y e
(ξ ) = dxdy = e dy dx
z 2π R2 x + iy 2π −∞ −∞ x + iy
 
∞ e−iξ1 x  ∞ e−iξ1 x 
1 ∞ −iξ2 y 1 0 −iξ2 y
= e dy dx + e dy dx
2π 0 −∞ x + iy 2π −∞ −∞ x + iy
 ∞ √ √
1
= e−iξ2 y 2π (−i 2π H(ξ1 )e−yξ1 )dy
2π 0

1 0 −iξ2 y √ √
+ e 2π (i 2π H(−ξ1 )e−yξ1 )dy
2π −∞
  ∞  0 
−y(ξ1 +iξ2 ) −y(ξ1 +iξ2 )
= −i H(ξ1 ) e dy − H(−ξ1 ) e dy .
0 −∞

For ξ1 > 0 we have

∞
 ∞
−y(ξ1 +iξ2 ) e−y(ξ1 +iξ2 )  1
−i e dy = i  = −i .
0 ξ1 + iξ2  ξ1 + iξ2
0
18 Tempered Distributions 165

For ξ1 < 0 we have

0
 0
−y(ξ1 +iξ2 ) e−y(ξ1 +iξ2 )  1
i e dy = −i  = −i .
−∞ ξ1 + iξ2  ξ1 + iξ2
−∞

Hence

1 i
(ξ ) = − ,
z ξ1 + iξ2

which proves (2). Part (1) is established with a simple calculation:

    2  2 
1 ∂ ∂ ∂ ∂ 1 ∂ ∂ 1
∂ ·∂ = −i +i = + = Δ = ∂ ·∂.
4 ∂x ∂y ∂x ∂y 4 ∂x ∂y 4
Chapter 19
Convolutions in S and S

Let us consider first the direct product of distributions. Let us assume that T1 , . . . , Tn
are one-dimensional tempered distributions, T j ∈ S (R), j = 1, 2, . . . , n. The product
T1 (x1 ) · · · Tn (xn ) can be formally defined by

T1 (x1 ) · · · Tn (xn ), ϕ (x1 , . . . , xn ) = T1 (x1 ) · · · Tn−1 (xn−1 ), ϕ1 (x1 , . . . , xn−1 )
= T1 (x1 ) · · · Tn−2 (xn−2 ), ϕ2 (x1 , . . . , xn−2 )
= · · · = T1 (x1 ), ϕn−1 (x1 ),

where

ϕ1 (x1 , . . . , xn−1 ) := Tn (xn ), ϕ (x1 , . . . , xn ) ∈ S(Rn−1 ),

ϕ j (x1 , . . . , xn− j ) := Tn− j+1 , ϕ j−1 (x1 , . . . , xn− j+1 ) ∈ S(Rn− j ).

In this sense, it is clear that

δ (x1 , . . . , xn ) = δ (x1 ) · · · δ (xn ).

But the product T1 (x)T2 (x), where the x are the same, in general case does not exist,
that is, it is impossible to define such a product. We remedy this by recalling the
following definition.

Definition 19.1. The convolution ϕ ∗ ψ of the functions ϕ ∈ S and ψ ∈ S is defined

as 
(ϕ ∗ ψ )(x) := ϕ (x − y)ψ (y)dy.
Rn

We can observe the following immediately.

c Springer International Publishing AG 2017 167

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 19
168 Part II: Fourier Transform and Distributions

(1) The convolution is commutative for every n ≥ 1. If n ≥ 2, then

 
(ϕ ∗ ψ )(x) = ϕ (x − y)ψ (y)dy = ϕ (z)ψ (x − z)dz = (ψ ∗ ϕ )(x).
Rn Rn

If n = 1, then
 ∞  −∞
(ϕ ∗ ψ )(x) = ϕ (x − y)ψ (y)dy = − ϕ (z)ψ (x − z)dz
−∞ ∞
 ∞
= ψ (x − z)ϕ (z)dz = (ψ ∗ ϕ )(x).
−∞

(2) It is also clear that the convolution is well defined for ϕ and ψ from S, and
moreover, for every α ≥ 0,

∂xα (ϕ ∗ ψ )(x) = (∂ α ϕ ∗ ψ )(x) = ∂xα ϕ (x − y)ψ (y)dy
Rn

= (−1)|α | ∂yα ϕ (x − y)ψ (y)dy
Rn

= (−1)2|α | ϕ (x − y)∂yα ψ (y)dy = (ϕ ∗ ∂ α ψ )(x),
Rn

where we integrated by parts and used the fact that ∂x j ϕ (x − y) = −∂y j ϕ (x − y).

We would like to prove that for ϕ and ψ from S it follows that ϕ ∗ ψ from S also. In
fact,
(1) ϕ ∗ ψ ∈ C∞ (Rn ) since ∂ α (ϕ ∗ ψ ) = ϕ ∗ ∂ α ψ and ∂ α : S → S.
(2) ϕ ∗ ψ decreases at infinity faster than any inverse power:
   
  1
 ϕ (x − y)ψ (y)dy ≤ c1 |ψ (y)|dy + c |ψ (y)|dy
 Rn  |y|≤ 2 |x − y|
|x| m 2 |x|
|y|> 2
 
c1
≤ |ψ (y)|dy + c2 |y|−m |y|m |ψ (y)|dy
|x|m |y|≤ 2
|x| |x|
|y|> 2
c1 c2
≤ + = c|x|−m , m ∈ N.
|x|m |x|m

Next we collect some important inequalities involving the convolution.

(1) Hölder’s inequality implies that

ϕ ∗ ψ
L∞ (Rn ) ≤
ϕ
L p (Rn ) ·
ψ
L p (Rn ) , (19.1)

where 1
p + 1
p = 1, 1 ≤ p ≤ ∞. This means that the convolution is well defined

even for ϕ ∈ L p (Rn ) and ψ ∈ L p (Rn ). In particular,
19 Convolutions in S and S 169

ϕ ∗ ψ
L∞ (Rn ) ≤
ϕ
L1 (Rn ) ·
ψ
L∞ (Rn ) . (19.2)

(2) It follows from Fubini’s theorem that

 

ϕ ∗ ψ
L 1 ≤ dx |ϕ (x − y)||ψ (y)|dy
R Rn
n

= |ψ (y)|dy |ϕ (x − y)|dx =
ϕ
L1
ψ
L1 . (19.3)
Rn Rn

(3) Interpolating (19.2) and (19.3) leads us to (see the Riesz–Thorin theorem, The-
orem 17.7)

ϕ ∗ ψ
L p ≤
ψ
L1 ·
ϕ
L p . (19.4)

(4) Interpolating (19.1) and (19.4) leads us to (again by the Riesz–Thorin theorem)

ϕ ∗ ψ
Ls ≤
ψ
Lr ·
ϕ
L p ,

where
1 1 1
1+ = + .
s r p

Indeed, the linear operator T ψ = ϕ ∗ ψ with ϕ ∈ L p (Rn ) maps as

 1 1
T : L p (Rn ) → L∞ (Rn ), + =1
p p

and
T : L1 (Rn ) → L p (Rn );

see (19.1) and (19.4), respectively. Thus

T : Lr (Rn ) → Ls (Rn ),

where
1 θ 1−θ θ
= + = 1−
r p 1 p

and
1 θ 1−θ 1 θ
= + = − .
s ∞ p p p

This gives
1 1 1
− = 1− .
r s p

Now we are in a position to consider the Fourier transform of a convolution.

170 Part II: Fourier Transform and Distributions

(1) Let ϕ , ψ ∈ S. Then ϕ ∗ ψ ∈ S and F (ϕ ∗ ψ ) ∈ S. Moreover,

 
e−i(x,ξ ) dx
n
F (ϕ ∗ ψ ) = (2π )− 2 ϕ (x − y)ψ (y)dy
R  R
n n

e−i(x,ξ ) ϕ (x − y)dx
n
= (2π )− 2 ψ (y)dy
Rn Rn
 
ψ (y)e−i(y,ξ ) dy ϕ (z)e−i(z,ξ ) dz = (2π ) 2 F ϕ · F ψ ,
n n
= (2π )− 2
Rn Rn

i.e., n
∗ ψ = (2π ) 2 ϕ · ψ
ϕ .

Similarly, n
F −1 (ϕ ∗ ψ ) = (2π ) 2 F −1 ϕ · F −1 ψ .

Hence n
ϕ ∗ ψ = (2π ) 2 F (F −1 ϕ · F −1 ψ ),

which implies that

n
F ϕ1 ∗ F ψ1 = (2π ) 2 F (ϕ1 · ψ1 ),

or n

ϕ · ψ = (2π )− 2 ϕ ∗ ψ
.

(2) Let us assume that ϕ ∈ L1 and ψ ∈ L p , 1 ≤ p ≤ 2. Then (19.4) implies that


ϕ ∗ ψ ∈ L p , 1 ≤ p ≤ 2. Further, F (ϕ ∗ ψ ) belongs to L p by the Hausdorff–
Young inequality. Thus,
n 
∗ ψ = (2π ) 2 ϕ · ψ
ϕ  ∈ Lp .

Lemma 19.2. Let ϕ (x) be a function from L1 (Rn ) with  Rn ϕ (x)dx = 1 and let ψ (x)
be a function from L2 (Rn ). Let us set ϕε (x) := ε −n ϕ εx , ε > 0. Then

L2 (Rn )
lim ϕε ∗ ψ = ψ.
ε →0+

Proof. By (19.4) we have that ϕε ∗ ψ ∈ L2 (Rn ). Then

n
ϕ ε · ψ
ε ∗ ψ = (2π ) 2 ϕ 

in L2 . But
x
 −n
−n  −n  −n 1 L∞
ε = ε
ϕ ϕ = ε σ 1 ϕ (ξ ) = ε ϕ(εξ ) = ϕ(εξ ) → ϕ(0)
ε ε ε
19 Convolutions in S and S 171

as ε → 0+. Note also that


n n
ϕ(0) = (2π )− 2 e−i(0,x) ϕ (x)dx = (2π )− 2 .
Rn

Hence
n L2
ϕ (εξ ) · ψ
ε ∗ ψ = (2π ) 2 ϕ  (ξ ) → ψ
 (ξ ), ε → 0+.

By the Fourier inversion formula it follows that

L2
ϕε ∗ ψ → ψ

as ε → 0+. 

Theorem 19.3. For every fixed function ϕ from S(Rn ) the map ϕ ∗ T has, as a
continuous linear map from S to S (with respect to T ), a unique continuous linear
extension as a map from S to S (with respect to T ) as follows:

ϕ ∗ T, ψ  := T, Rϕ ∗ ψ ,

where Rϕ (x) := ϕ (−x). Moreover, this extension has the properties

∗ T = (2π ) 2 ϕ · T,
(1) ϕ
n

(2) ∂ (ϕ ∗ T ) = ∂ α ϕ ∗ T = ϕ ∗ ∂ α T .
α

Proof. Let us assume that ϕ , ψ , and T belong to the Schwartz space S(Rn ). Then
we have checked already the properties (1) and (2) above. But we can easily check
that for such functions the definition is also true. In fact,
  
ϕ ∗ T, ψ  = (ϕ ∗ T )(x)ψ (x)dx = ϕ (x − y)T (y)dyψ (x)dx
R Rn Rn
n

= T (y) ϕ (x − y)ψ (x)dxdy
R Rn
n

= T (y)dy Rϕ (y − x)ψ (x)dx = T, Rϕ ∗ ψ .
Rn Rn

S
For the case T ∈ S the statement of this theorem follows from the fact that S = S
(see Theorem 18.18). 

Corollary 19.4. Since ϕ ∗ T = T ∗ ϕ for ϕ and T from S, we may define T ∗ ϕ as

follows (for T ∈ S ):
T ∗ ϕ , ψ  := T, Rϕ ∗ ψ .
172 Part II: Fourier Transform and Distributions

Example 19.5.
(1) It is true that δ ∗ ϕ = ϕ . Indeed,

δ ∗ ϕ , ψ  = δ , Rϕ ∗ ψ  = (Rϕ ∗ ψ )(0) = ϕ (y)ψ (y)dy = ϕ , ψ .
Rn

Alternatively, we note that

∗ ϕ = (2π ) 2 δ · ϕ = 1 · ϕ = ϕ
δ
n

is equivalent to
δ ∗ϕ = ϕ

in S .
(2) Property (2) of Theorem 19.3 and part (1) of this example imply that

∂ α (δ ∗ ϕ ) = δ ∗ ∂ α ϕ = ∂ α ϕ .

(3) Let us consider again the equation (1 − Δ )u = f for u and f ∈ L2 (or even from
u = f is still valid in L2 and u = (1 + |ξ |2 )−1 f or
S ). Then (1 + |ξ |2 )

−1 1 
u(x) = F f
1 + |ξ |2
 
− 2n −1 1
= (2π ) F ∗f= K(x − y) f (y)dy,
1 + |ξ |2 Rn

where 
1 ei(x−y,ξ )
K(x − y) := dξ .
(2π )n Rn 1 + |ξ |2

This is the inverse Fourier transform of a locally integrable function. This func-
tion K is the free space Green’s function of the operator 1 − Δ in Rn . We will
calculate this integral precisely in Chapter 22.

 x 19.6. Let j(x) be a function from L (R ) with j(x)dx = 1. Set jε (x) =
Lemma 1 n
Rn
−n
ε j ε , ε > 0. Then

jε ∗ f − f
L p → 0, ε → 0+,

for every function f ∈ L p (Rn ), 1 ≤ p < ∞. In the case p = ∞ we can state only the
fact  
( jε ∗ f )gdx → f · gdx, ε → 0+
Rn Rn

for every g ∈ L1 (Rn ).

19 Convolutions in S and S 173

Exercise 19.1. Prove Lemma 19.6 and find a counterexample showing that the first
part fails for p = ∞.

Remark 19.7. If j ∈ C0∞ (Rn ) or S(Rn ), then jε ∗ f ∈ C0∞ (Rn ) or S(Rn ) also for every
f ∈ L p (Rn ), 1 ≤ p < ∞.
Chapter 20
Sobolev Spaces

Lemma 20.1. For every function f ∈ L2 (Rn ) the following statements are
equivalent:
∂f
∂ x j (x) ∈ L (R ),
(1) 2 n

(2) ξ j f(ξ ) ∈ L2 (Rn ),

Δ t f (x)
(3) lim j t exists in L2 (Rn ). Here Δ tj f (x) := f (x + te j ) − f (x) with t ∈ R and
t→0
e j = (0, . . . , 1, 0, . . . , 0).
L2 ∂ fk
(4) There exists { fk }∞
k=1 , f k ∈ S, such that f k → f and ∂xj has a limit in L2 (Rn ).

Proof. (1) ⇔ (2): Since


D 
j f = ξj f ,

we have    
 
ξ j f  = D j f L2
L2

by Parseval’s equality.

(2) ⇒ (3): Let ξ j f be a function from L2 (Rn ). Then the equality


1 t 1 eit ξ j − 1
Δ j f (ξ ) = (eit ξ j − 1) f(ξ ) = · ξ j f(ξ )
t t tξ j

holds. But
eit ξ j − 1
→i
tξ j

c Springer International Publishing AG 2017 175

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 20
176 Part II: Fourier Transform and Distributions

pointwise as t → 0. Hence


1 t L2
Δ f → iξ j f, t →0
t j
i.e. (again due to Parseval’s equality),

1 t L2 ∂ f
Δ f→ , t → 0.
t j ∂xj

The same arguments lead us to the statement that (3) ⇒ (1).

L2 S ∂ f k S
(4) ⇒ (1): Let fk be a sequence from S such that fk → f . Then fk → f and ∂xj →
∂f ∂ f k L2
∂xj also. By condition (4) we have that the limit lim =g exists. We
k→∞ ∂ x j
S
may conclude that ∂∂ xfkj → g. This means that g = ∂∂xfj in S .
(2) ⇒ (4): Let us write f(ξ ) as the sum of two functions f(ξ ) = g(ξ ) + h(ξ ),
where

g(ξ ) = f(ξ )χ{|ξ j |<1} , h(ξ ) = f(ξ )χ{|ξ j |>1} .

L2
Let {gk } be a sequence in S such that gk → g and supp gk ⊂ {|ξ j | < 2}.
L2
 ξ j h and
Let {hk } be a sequence in S such that hk →  supp hk ⊂ {|ξ j | > 2 }.
1

If we define the sequence fk (x) = F −1 gk + hξk (x), then

j

 hk L 2
fk (ξ ) = gk + → g(ξ ) + h(ξ ) = f(ξ ).
ξj

But
∂fk L2
= iξ j gk + ihk → iξ j (g + h) = iξ j f.
∂xj

This means that (by the Fourier inversion formula or Parseval’s equality)

∂ fk L2 −1 ∂f
→ F (iξ j f) = .
∂xj ∂xj

This completes the proof. 

We have also the following generalization of Lemma 20.1 to a multi-index α .

20 Sobolev Spaces 177

Lemma 20.2. Let f be a function from L2 (Rn ) and let s ∈ N. Then the following
statements are equivalent:
(1) Dα f ∈ L2 (Rn ), |α | ≤ s;
(2) ξ α f ∈ L2 (Rn ), |α | ≤ s;
Δhα f α α1 αn
α exists in L (R ), |α | ≤ s. Here Δ h f := (Δ h · · · Δ h ) f and h ∈ R with
(3) lim 2 n n
h→0 h 1 n
h j = 0 for all j = 1, 2, . . . , n.
L2
(4) There exists fk ∈ S such that fk → f and Dα fk has a limit in L2 (Rn ) for |α | ≤ s.

Proof. The result follows from Lemma 20.1 by induction on |α |. 

Definition 20.3. Let s > 0 be an integer. Then

H s (Rn ) := { f ∈ L2 (Rn ) : ∑
Dα f
L2 < ∞}
|α |≤s

is called the (L2 -based) Sobolev space of order s with norm

 1/2

f
H s (Rn ) := ∑
D α
f
2L2 (Rn ) .
|α |≤s

Remark 20.4. It is easy to check that H s (Rn ), s ∈ N, can be characterized by

H s (Rn )={ f ∈ L2 (Rn ) : (1 + |ξ |2 )s | f(ξ )|2 dξ < ∞}.
Rn

Proof. It follows from Parseval’s equality that

 2  
   α 2
∑
Dα f
2L2 = ∑ Dα f  2 =
L
∑ ξ f  2
L
|α |≤s |α |≤s |α |≤s

= ∑ |ξ α |2 | f(ξ )|2 dξ = ∑ |ξ α |2 | f(ξ )|2 dξ .
|α |≤s R Rn |α |≤s
n

But it is easily seen that there are positive constants c1 and c2 such that

c1 (1 + |ξ |2 )s ≤ ∑ |ξ α |2 ≤ c2 (1 + |ξ |2 )s ,
|α |≤s

or
∑ |ξ α |2  (1 + |ξ |2 )s .
|α |≤s
178 Part II: Fourier Transform and Distributions

Therefore we may conclude that

∑
Dα f
L2 < ∞ ⇔ ∑
Dα f
2L2 < ∞ ⇔ (1 + |ξ |2 )s | f(ξ )|2 dξ < ∞.
|α |≤s |α |≤s Rn

This establishes the characterization. 

This property of an integer s justifies the following definition.

Definition 20.5. Let s be a real number. Then we define

H s (Rn ) := { f ∈ S : (1 + |ξ |2 ) 2 f ∈ L2 (Rn )}
s

with the norm

1
2

f
H s (Rn ) := (1 + |ξ |2 )s | f(ξ )|2 dξ .
Rn

Definition 20.6. Let s > 0 be an integer and 1 ≤ p ≤ ∞. Then

Wps (Rn ) := { f ∈ L p (Rn ) : ∑

Dα f
L p (Rn ) < ∞}
|α |≤s

is called the Sobolev space with norm

 1/p

f
Wps (Rn ) := ∑
Dα f
Lp p (Rn ) .
|α |≤s

Exercise 20.1. Let s > 0 be an even integer and 1 ≤ p ≤ ∞. Prove that

1
p
|F −1 ((1 + |ξ |2 ) 2 f)| p dx
s
| f |Wps (Rn ) :=
Rn

is an equivalent norm in Wps (Rn ).

Definition 20.7. Let s > 0 be a real number and 1 ≤ p ≤ ∞. Then

1
p
Wps (Rn ) := { f ∈S :
|F −1
((1 + |ξ | ) f)| p dx
2 2s
< ∞}
Rn

with the norm

1
p

f
Wps (Rn ) := |F −1
((1 + |ξ | ) f)| p dx
2 2s
.
Rn
20 Sobolev Spaces 179

Exercise 20.2. Let s ∈ R. Prove that

f ∈ H s (Rn )

if and only if
f ∈ Ls2 (Rn ).

Proposition 20.8. Let us assume that 0 < s < 1. Then

| f (x) − f (y)|2
(1 + |ξ |2s )| f(ξ )|2 dξ = | f (x)|2 dx + As dxdy, (20.1)
Rn Rn Rn Rn |x − y|n+2s

where As is a positive constant depending on s and n.

Remark 20.9. Since 1 + |ξ |2s  (1 + |ξ |2 )s , 0 < s < 1, the right-hand side of (20.1)
is an equivalent norm in H s (Rn ).

Proof. Denote by I the double integral appearing on the right-hand side of (20.1).
Then

I= | f (y + z) − f (y)|2 |z|−n−2s dydz
Rn Rn

|ei(z,ξ ) − 1|2
= |z|−n−2s dz |ei(z,ξ ) − 1|2 | f(ξ )|2 dξ = | f(ξ )|2 dξ dz
Rn Rn Rn Rn |z|n+2s

by Parseval’s equality and Exercise 16.4. We claim that

|ei(z,ξ ) − 1|2
dz = |ξ |2s A−1
s .
Rn |z|n+2s

Indeed, if we consider the Householder reflection matrix

2vvT
A := I − , v = ξ − |ξ |e1 , ξ ∈ Rn ,
|v|2

then AT = A−1 = A and Aξ = |ξ |e1 = (|ξ |, 0, . . . , 0). It follows that

|ei(z,ξ ) − 1|2 |ei(Az,Aξ ) − 1|2
|ξ |−2s dz = |ξ |−2s dz
Rn |z|n+2s Rn |z|n+2s

|ei(y,Aξ ) − 1|2 |eiy1 |ξ | − 1|2
= |ξ |−2s dy = |ξ |−2s dy
Rn |y|n+2s Rn |y|n+2s

|e − 1|2
iz 1
= dz =: A−1
s .
Rn |z|n+2s
180 Part II: Fourier Transform and Distributions

Therefore,

| f (x)|2 dx + As | f (x) − f (y)|2 |x − y|−n−2s dxdy
Rn Rn Rn

= | f(ξ )|2 dξ + |ξ |2s | f(ξ )|2 dξ .
Rn Rn

This completes the proof. 

Remark 20.10. Note that As exists only for 0 < s < 1.

Exercise 20.3. Prove that

 2
 

f
2H k+s (Rn )  (1 + |ξ |2k+2s )| f(ξ )|2 dξ   f 2 + ∑ 
|ξ |2s |Dα f |2 dξ
Rn L
|α |=k R
n

=
f
2L2 + As ∑ |Dα f (x) − Dα f (y)|2 |x − y|−n−2s dxdy.
|α |=k R R
n n

Example 20.11. 1 ∈ / H s (Rn ) for all s. Indeed, assume that 1 ∈ H s0 (Rn ) for some s0
s0
(it is clear that s0 > 0). This means that (1 + |ξ |2 ) 2  1 ∈ L2 (Rn ). It follows from this
fact that  2 (Rn ), and further,  1 (Rn ). But  n
1 ∈ Lloc 1 ∈ Lloc 1 = (2π ) 2 δ , and we know that
δ is not a regular distribution.

Next we list some properties of H s (Rn ).

(1) Since f ∈ H s (Rn ) if and only if f(ξ ) ∈ Ls2 (Rn ) and Ls2 (Rn ) is a separable Hilbert
space with the scalar product

( f1 , g1 )Ls2 (Rn ) = (1 + |ξ |2 )s f1 · g1 dξ ,
Rn

it follows that H s (Rn ) is also a separable Hilbert space, and the scalar product
can be defined by

( f , g)H s (Rn ) = (1 + |ξ |2 )s f· gdξ .
Rn

We may prove the following property:

H −s (Rn ) = (H s (Rn ))∗ , s∈R

in the sense that

|( f , g)L2 (Rn ) |

f
H −s (Rn ) := sup .
0 =g∈H s (Rn )
g
H s (Rn )
20 Sobolev Spaces 181

This means that H −s (Rn ) is the dual space to H s (Rn ) with respect to the Hilbert
space L2 (Rn ). Indeed, since by Parseval’s equality

( f , g)L2 (Rn ) = ( f, g)L2 (Rn ) = ((1 + |ξ |2 )−s/2 f, (1 + |ξ |2 )s/2 g)L2 (Rn ) ,

it follows that

|( f , g)L2 (Rn ) | |( f, g)L2 (Rn ) |  

 
sup = sup =  f =
f
H −s (Rn ) .
0 =g∈H s (Rn )
g
H s (Rn ) 0 =g∈Ls2 (Rn )

g
Ls2 (Rn ) 2 (Rn )
L−s

We have used here the fact that the space L−s2 (Rn ) is dual to L2 (Rn ) for every
s
s ∈ R.
(2) For −∞ < s < t < ∞, it follows that S ⊂ H t (Rn ) ⊂ H s (Rn ) ⊂ S .

Example 20.12. δ ∈ H s (Rn ) if and only if s < − n2 . Indeed, if we define

1
ξ  := (1 + |ξ |2 ) 2 ,
n
then δ ∈ H s (Rn ) is equivalent to (2π )− 2 ξ s ∈ L2 (Rn ), which in turn is equivalent
to s < − n2 .

(3) Let ϕ be a function from H s (Rn ), and ψ a function from H −s (Rn ). Then ϕ ∈
 ∈ L−s
Ls2 (Rn ) and ψ ·ψ
2 (Rn ), so that ϕ  ∈ L1 (Rn ) by Hölder’s inequality. We
may therefore define (temporarily, and with slight abuse of notation)

ϕ , ψ L2 (Rn ) := ϕ · ψ
 dξ
Rn

and obtain
| ϕ , ψ L2 (Rn ) | ≤
ϕ
H s (Rn ) ·
ψ
H −s (Rn ) .

∂δ
For example, if ϕ is a function from H 2 +1+ε (Rn ), ε > 0, and ψ =
n
∂xj , then

 
∂δ ∂δ n
,ϕ = · ϕdξ = i(2π )− 2 ξ j ϕ(ξ )dξ
∂xj L2 (Rn ) Rn ∂xj Rn

is well defined, since ϕ ∈ L2n +1+ε (Rn ) and ξ j ∈ L−

2
n −1−ε (R ).
n
2 2
(4) Consider a differential operator with constant coefficients

P(D) = ∑ aα Dα .
|α |≤m

Then P(D) : H s (Rn ) → H s−m (Rn ) for all real s.

182 Part II: Fourier Transform and Distributions

Proof. By the properties of the Fourier transform we have

P(D) f
2H s−m (Rn ) = f |2 dξ
(1 + |ξ |2 )s−m |P(D)
Rn

= (1 + |ξ |2 )s−m |P(ξ )|2 · | f(ξ )|2 dξ
Rn

≤c (1 + |ξ |2 )s−m (1 + |ξ |2 )m | f(ξ )|2 dξ = c
f
2H s (Rn )
Rn

for all real s. 

There is a generalization of this result. Let P(x, D) be a differential operator

P(x, D) = ∑ aα (x)Dα
|α |≤m

with variable coefficients such that |aα (x)| ≤ c0 for all x ∈ Rn and |α | ≤ m. Then

P(x, D) : H m (Rn ) → L2 (Rn ).

Indeed,
 
 α 

P(x, D) f
L2 ≤ c0 ∑
Dα f
L2 = c0 ∑ ξ f 
L2
|α |≤m |α |≤m
 
 
f
m
≤ c0 (1 + |ξ |2 ) 2 = c0
f
H m .
L2

(5) We have the following lemma.

Lemma 20.13. Let ϕ be a function from S, and f a function from H s (Rn ) for s ∈ R.
Then ϕ · f ∈ H s (Rn ) and
 
 |s| 

ϕ f
H s ≤ c (1 + |ξ |2 ) 2 ϕ 1 ·
f
H s .
L

Proof. We know that

ϕ ϕ(ξ − η ) f(η )dη .
n
· f (ξ ) = (2π )− 2
Rn

Hence
ξ s
ξ  ϕ ϕ(ξ − η ) η s f(η )dη .
n
s
· f = (2π )− 2
Rn η s

Let us prove that

|s|
ξ s ≤ 2 2 η s · ξ − η |s|
20 Sobolev Spaces 183

for all s ∈ R. Indeed,

1 1
ξ  = (1 + |ξ |2 ) 2 ≤ (1 + |η |2 ) 2 + |ξ − η | = η  + |ξ − η | ≤ η (1 + |ξ − η |).
√
Since 1 + |ξ − η | ≤ 2 ξ − η , we have
s
ξ s ≤ 2 2 η s · ξ − η s

for s ≥ 0. Moreover, for s < 0 we have

ξ s η |s| |s|
= ≤ 2 2 η − ξ |s| .
η s ξ |s|

It now follows from (19.4) that

       
       

ϕ f
H s =  ξ s ϕf  2 ≤ c | ξ |s| ϕ| ∗ | η s f| ≤ c  ξ |s| ϕ 1  η s f
L L2 L L2

for all s ∈ R. 

Exercise 20.4. Suppose s > n/2. Show that if u, v ∈ L2 (Rn ) and

u(ξ − η )v(η )dη
w(ξ ) = ,
Rn (1 + |η |2 )s/2

then w ∈ L2 (Rn ) and

w
L2 ≤ C
u
L2
v
L2 .

(6) Let us now consider distributions with compact support in greater detail than
what we saw in Chapter 18.

Definition 20.14. Set E = C∞ (Rn ). We say that T ∈ E  if T is a linear functional

on E that is also continuous, i.e., ϕk → 0 in E implies that T, ϕk  → 0 in C. Here
ϕk → 0 in E means that
sup |∂ α ϕk | → 0, k → ∞
K

for every compact subset K ⊂ Rn and multi-index α .

It can be proved that T ∈ E  if and only if there exist c0 > 0, R0 > 0, and n0 ∈ N0
such that
| T, ϕ | ≤ c0 ∑ sup |Dα ϕ (x)|
|α |≤n0 |x|≤R0

for all ϕ ∈ C∞ (Rn ). Moreover, members of E  have compact support.

184 Part II: Fourier Transform and Distributions

Assume that T ∈ E  . Since ϕ (x) = e−i(x,ξ ) ∈ C∞ (Rn ), it follows that T, e−i(x,ξ ) 

is well defined and that there exist c0 > 0, R0 > 0, and n0 ∈ N0 such that
n0
| T, e−i(x,ξ ) | ≤ c0 ∑ sup |Dαx e−i(x,ξ ) | ≤ c0 ∑ |ξ α |  (1 + |ξ |2 ) 2 .
|α |≤n0 |x|≤R0 |α |≤n0

If we now set
T(ξ ) := (2π )−n/2 T, e−i(x,ξ ) ,

then T is a usual function of ξ . The same is true for

∂ α T(ξ ) = (2π )−n/2 (−1)|α | T, ∂ α e−i(x,ξ ) 

and hence T ∈ C∞ (Rn ). On the other hand, | T, e−i(x,ξ ) | ≤ c0 ξ n0 implies that
|T(ξ )| ≤ c0 ξ n0 and hence T ∈ Ls2 (Rn ) for s < −n0 − n2 . So, by Exercise 20.2, we
may conclude that every T ∈ E  belongs to H s (Rn ) for s < −n0 − n2 .
(7) We have the following lemma.
Lemma 20.15. The closure of C0∞ (Rn ) in the norm of H s (Rn ) is H s (Rn ) for all
Hs
s ∈ R. In short, C0∞ (Rn ) = H s (Rn ).
Proof. Let f be an arbitrary function from H s (Rn ) and let fR be a new function such
that 
  f(ξ ), |ξ | < R,
fR (ξ ) = χR (ξ ) f (ξ ) =
0, |ξ | > R.

Then fR (x) = F −1 (χR f)(x) = (2π )− 2 (F −1 χR ∗ f )(x). It follows from the above
n

considerations that F −1 χR ∈ C∞ (Rn ) as the inverse Fourier transform of a com-

/ C0∞ (Rn )) and
pactly supported function (but ∈


f − fR
2H s = | f(ξ ) − fR (ξ )|2 ξ 2s dξ = | f(ξ )|2 ξ 2s dξ → 0
Rn |ξ |>R

as R → ∞, since f ∈ H s (Rn ). This completes the first step. 

The second step is as follows. Let j(ξ ) ∈ C0∞ (|ξ | < 1) with Rn j(ξ )dξ = 1. Let us
Lp
set jk (ξ ) := kn j(kξ ). We recall from Lemma 19.6 that jk ∗ g → g,
1 ≤ p < ∞. Define the sequence vk := F −1 ( jk ∗ fR ). Since vk = jk ∗ fR , it follows
that supp vk ⊂ UR+1 (0), and so vk ∈ C0∞ (Rn ). Hence vk ∈ S. Therefore, vk ∈ H s (Rn )
and


vk − fR
2H s (Rn ) = ξ 2s | jk ∗ fR − fR |2 dξ
|ξ |<R+1

≤ CR | jk ∗ fR − fR |2 dξ → 0, k → ∞.
|ξ |<R+1
20 Sobolev Spaces 185

/ C0∞ (Rn ), we take a function κ ∈ C0∞ (Rn ) with κ (0) = 1. Then

Since vk ∈
x
S
κ vk → vk , A → ∞.
A
x Hs x
This fact implies that κ A vk → vk as A → ∞. Setting fk (x) := κ A vk (x) ∈
C0∞ (Rn ), we get finally
 x 
 

f − fk
H s ≤
f − fR
H s +
fR − vk
H s + vk − κ vk  → 0
A Hs

if A, k, and R are sufficiently large. 

Now we are in a position to formulate the main result concerning H s (Rn ).

Theorem 20.16 (Sobolev embedding theorem). Let f be a function from H s (Rn )

for s > k + n2 , where k ∈ N0 . Then Dα f ∈ Ċ(Rn ) for all α such that |α | ≤ k. In short,

n
H s ⊂ C˙k (Rn ), s > k+ .
2

Proof. Let f ∈ H s (Rn ) ⊂ S . Then

Dα f = F −1 F (Dα f ) = F −1 (ξ α f(ξ )).

What is more,

|ξ ||α |
|ξ f(ξ )|dξ ≤ c
α
|ξ | | f(ξ )|dξ = c
|α |
ξ s | f(ξ )|dξ
Rn ξ 
s
Rn Rn
 1/2
1/2
|ξ |2|α | 2s 
≤c dξ ξ  | f (ξ )| dξ2
≤ c
f
H s (Rn )
Rn ξ 2s Rn

if and only if 2s − 2|α | > n, or s > |α | + n/2.

This means that for such s and α the function Dα f is the Fourier transform of
some function from L1 (Rn ). By the Riemann–Lebesgue lemma we have that Dα f
from Ċ(Rn ). 

 Ls (R ) ⊂ L (R ) if and only if q = 2 and s ≥ 0 or 1 ≤ q < 2 and

Lemma 2 n q n
 20.17.
s > n 1q − 12 .

Exercise 20.5. Prove Lemma 20.17.

186 Part II: Fourier Transform and Distributions

Lemma 20.18 (Hörmander).

 
(1) F : H s (Rn ) → Lq (Rn ) for 1 ≤ q < 2 and s > n − 12 . 1
q
 
(2) F : L p (Rn ) → H −s (Rn ) for 2 < p ≤ ∞ and s > n 12 − 1p .
(3) F : L2 (Rn ) → L2 (Rn ).

Proof. (1) See Lemma 20.17.

(2) Let f be a function from L p (Rn ) for 2 < p ≤ ∞. Then f ∈ S and | f, ϕ L2 (Rn ) | =
| f , ϕL2 (Rn ) | ≤
f
p ·
ϕ
p , where 1 ≤ p < 2. But if ϕ ∈ H s (Rn ) for s >
 
n p1 − 12 , then
ϕ
L p ≤ c
ϕ
H s . So

| f, ϕ L2 (Rn ) | ≤ c

f
p ·
ϕ
H s .

Therefore (by duality),  

 
f ≤ c
f
L p
H −s
   
for s > n 1
p − 12 = n 12 − 1p .
 
 
(3) This is simply Parseval’s equality  f 2 =
f
L2 .
L
This completes the proof. 

Exercise 20.6. Prove that

(1) χ[0,1] ∈ H s (R) if and only if s < 1/2.
(2) χ[0,1]×[0,1] ∈ H s (R2 ) if and only if s < 1/2.
 
(3) K(x) := F −1 1+|1ξ |2 ∈ H s (Rn ) if and only if s < 2 − n/2.
(4) Let f (x) = χ (x) log log |x|−1 in R2 , where χ (x) ∈ C0∞ (|x| < 1/3). Prove that
/ L∞ (R2 ).
f ∈ H 1 (R2 ) but f ∈

Remark 20.19. This counterexample shows us that the Sobolev embedding theorem
is sharp.

Lemma 20.20. Let us assume that ϕ and f from H s (Rn ) for s > n2 . Then F (ϕ f ) ∈
L1 (Rn ).

Proof. Since f , ϕ ∈ H s (Rn ), it follows that f, ϕ ∈ Ls2 (Rn ) for s > n2 . But this implies
(see Lemma 20.17) that f and ϕ ∈ L1 (Rn ) and

F (ϕ f ) = (2π )− 2 ϕ ∗ f
n

also belongs to L1 (Rn ). 

20 Sobolev Spaces 187

Remark 20.21. It is possible to prove that if ϕ , f ∈ H s (Rn ) for s > n2 , then ϕ f ∈

H s (Rn ) with the same s.

Exercise 20.7. Prove that Wp1 (Rn ) ·Wp1 (Rn ) ⊂ Wp1 (Rn ) if p > n.

Next, we consider the trace map τ , defined initially on S(Rn ) by τ u = f , where

f (x ) = u(0, x ) if x = (x1 , . . . , xn ) and x = (x2 , . . . , xn ).
Proposition 20.22. The map τ extends uniquely to a continuous linear map

τ : H s (Rn ) → H s−1/2 (Rn−1 )

for all s > 1/2, and this map is surjective.

Proof. If f = τ u, then for all u ∈ S(Rn ) we may define

∞
f(ξ  ) = u(ξ )dξ1 .
−∞

Hence, using the Cauchy-Bunyakovsky-Schwarz inequality, we have

∞ ∞
| f(ξ  )|2 ≤ u(ξ )| (1 + |ξ | ) dξ1
| 2 2 s
(1 + |ξ |2 )−s dξ1 .
−∞ −∞

Since
∞ ∞
1 1 1
dξ1 =
 2 s dξ1
−∞ (1 + |ξ  |2 + ξ1 )s
2 (1 + |ξ  |2 )s −∞ ξ1
1+ (1+|ξ  |2 )1/2
∞
dρ
= (1 + |ξ  |2 )−s+1/2 , s > 1/2,
−∞ (1 + ρ 2 )s

we have
∞
(1 + |ξ  |2 )−s+1/2 | f(ξ  )|2 ≤ Cs u(ξ )|2 (1 + |ξ |2 )s dξ1 ,
|
−∞

where Cs denotes the latter convergent integral with respect to ρ . Integrating with
respect to ξ  in the latter inequality leads to

f
2H s−1/2 (Rn−1 ) ≤ Cs
u
2H s (Rn ) .
188 Part II: Fourier Transform and Distributions

This means that the first part of this proposition is proved, since S(Rn ) is dense in
H s (Rn ). Surjectivity follows also. If g ∈ H s−1/2 (Rn−1 ), s > 1/2, we can define

(1 + |ξ  |2 )s/2−1/4
u(ξ ) := g(ξ  ) .
(1 + |ξ |2 )s/2

Then u := F −1 (
u(ξ )) defines u ∈ H s (Rn ) and u(0, x ) = Cg(x ) with some nonzero
constant C. 

20.1 Sobolev Spaces on Bounded Domains

Let Ω be a bounded domain in Rn with smooth boundary ∂ Ω . We define for integers

k ≥ 0 the Sobolev space H k (Ω ) = W2k (Ω ) as the set of all f ∈ L2 (Ω ) for which
there exist (by analogy with Lemma 20.1 if we extend f by zero outside of Ω ) the
generalized derivatives ∂ α f in L2 (Ω ) for all |α | ≤ k. The norm in this space is
defined then by

1/2

f
H k (Ω ) = ∑ |∂ α f |2 dx . (20.2)
|α |≤k Ω

We define the space H0k (Ω ) as the completion of C0∞ (Ω ) with respect to the norm of
H k (Ω ).
Theorem 20.23 (Poincaré’s inequality). Suppose f ∈ H0k (Ω ), k ≥ 1. Then there is
a constant M > 0 such that
 
 

f
H k−1 (Ω ) ≤ M ∑ ∂ β f  2 . (20.3)
L (Ω )
|β |=k

Proof. We apply induction with respect to k. Since Ω is bounded, it can be enclosed

in a cube
Qn := {x ∈ Rn : |x j | ≤ A, j = 1, . . . , n},

and f ∈ C0∞ (Ω ) will continue to be identically zero outside of Ω . Then for all x ∈ Qn
we have x1
f (x) = ∂x1 f (ξ1 , x )dξ1 , x = (x1 , x ). (20.4)
−A

We have used here the fact f = 0 on ∂ Ω . Using the Cauchy-Bunyakovsky-Schwarz

inequality, we obtain that
A A x1 A
| f (x)|2 dx1 ≤ 2A dx1 |∂x1 f (ξ1 , x )|2 dξ1 ≤ 4A2 |∂x1 f (ξ1 , x )|2 dξ1 .
−A −A −A −A
20 Sobolev Spaces 189

Integrating now with respect to x , we obtain

Qn
| f (x)|2 dx ≤ 4A2
Qn
|∂x1 f |2 dx ≤ 4A2 ∑ |∂ β f |2 dx.
|β |=1 Qn

Since C0∞ (Ω ) is dense in H01 (Ω ), the case k = 1 is established. Let us assume that
for all f ∈ H0k (Ω ), k ≥ 1, we have
 
 β 

f
H k−1 (Ω ) ≤ M ∑ ∂ f 
L 2 (Ω )
.
|β |=k

Then for all f ∈ H k+1 (Ω ) we have that ∂ j f ∈ H0k (Ω ), j = 1, . . . , n, and by induction

we have
   
∂ j f  k−1  
H (Ω )
≤ M ∑ ∂ β (∂ j f ) 2 ≤ M ∑
∂ γ f
L2 (Ω ) .
L (Ω )
|β |=k |γ |=k+1

Thus
n  
∑ ∂ j f H k−1 (Ω ) ≤ M ∑
∂ γ f
L2 (Ω ) ,
j=1 |γ |=k+1

or

f
H k (Ω ) ≤ M  ∑
∂ γ f
L2 (Ω ) .
|γ |=k+1

Hence the theorem is completely proved. 

For all real s > 0 the space H s (Ω ) with fractional s can be obtained as the inter-
polation space between L2 (Ω ) and H k (Ω ) with some integer k ≥ 1 (see, e.g., [39,
p. 286] for details).

Corollary 20.24. Suppose f ∈ H0k (Ω ), k ≥ 1. Then ∂ α f |∂ Ω = 0 for all |α | ≤ k − 1.

Proof. This fact can be proved also by induction on k ≥ 1 using (20.4). Since C0∞ (Ω )
is complete in H0k (Ω ), it follows that for k = 1 and x = (x1 , x ) ∈ ∂ Ω we have from
(20.4) that x1
f (x) = ∂x1 f (ξ1 , x )dξ1 = 0.
−A

Using induction on k and the representation

x1
α
∂ f (x) = ∂x1 (∂ α f (ξ1 , x ))dξ1
−A

for x ∈ ∂ Ω and |α | ≤ k − 1, we obtain the desired result. 

190 Part II: Fourier Transform and Distributions

Also, the converse of this corollary holds: if f ∈ H k (Ω ) and ∂ α f |∂ Ω = 0 for |α | ≤

k − 1, then f ∈ H0k (Ω ). For details, see [11].

Exercise 20.8. Prove Rellich’s theorem: if Ω is bounded and s > t ≥ 0, then the
inclusion map H0s (Ω ) → H0t (Ω ) is compact. Hint: Use the Ascoli–Arzelà theorem
(Theorem 34.7).

We define for every integer k ≥ 0 the space

H k (Rn+ ) = { f ∈ L2 (Rn+ ) : ∂ α f ∈ L2 (Rn+ ), |α | ≤ k}

with the norm

 ∞
1/2

f
H k (Rn+ ) = ∑ dx  α  2
|∂ f (x1 , x )| dx1 . (20.5)
|α |≤k R
n−1 0

We may say that an element f ∈ H k (Rn+ ) is the restriction on Rn+ of some element
from H k (Rn ). More precisely, the following proposition holds.

Proposition 20.25. For every integer k ≥ 0 there is an extension linear operator E

defined on H k (Rn+ ) with image in H k (Rn ) such that

E f
H k (Rn ) ≤ C
f
H k (Rn+ ) , (20.6)

where the constant C > 0 is independent of f .

Proof. Since S(Rn+ ) is dense in H k (Rn+ ) (see, for example, Lemma 20.15), it follows
that for every f ∈ S(Rn+ ) and integer k ≥ 0 we may define E as

f (x), x1 ≥ 0,
E f (x) = 
j=1 a j f (− jx1 , x ),
∑k+1 x1 < 0,

where a j , j = 1, . . . , k + 1, are to be determined. Let us prove that this operator E

satisfies this proposition. It is clear that E is linear and that for k = 0 (a1 = 1),

f (x), x1 ≥ 0,
E f (x) = 
f (−x1 , x ), x1 < 0,

belongs to L2 (Rn ) and (20.6) holds. If k ≥ 1 is an integer, then for 0 ≤ l ≤ k we have

formally 
∂ l1 f (x), x1 > 0,
∂xl1 (E f (x)) = xk+1 l ∂ l f (− jx , x ), x < 0.
∑ j=1 j a (− j) x1 1 1

Let us choose a j as the solution of the linear algebraic system

20 Sobolev Spaces 191

k+1
∑ a j (− j)l = 1, l = 0, 1, . . . , k.
j=1

The determinant of this system is the well known Vandermonde polynomial, and it
is not equal to zero. Hence, this system has a unique solution with respect to the
coefficients a j , j = 1, . . . , k − 1. After these coefficients have been determined, the
inequality (20.6) follows immediately. 
Remark 20.26. For arbitrary s ≥ 0 the result of Proposition 20.25 can be obtained
by interpolation of Sobolev spaces H s ; see [39, p. 285].
Let Ω ⊂ Rn be a bounded domain with a C∞ boundary ∂ Ω . Since ∂ Ω is a compact
set, it can be covered by finitely many open sets U j , j = 1, . . . , m, such that


m
∂Ω ⊂ Uj
j=1

and there is a partition of unity {ϕ j }mj=1 with ϕ j ∈ C0∞ (U j ) and ∑mj=1 ϕ j ≡ 1 on ∂ Ω .

Thus for every function u(x) defined on ∂ Ω we have u = ∑mj=1 ϕ j u. The functions
ϕ j u can be written in local coordinates y as ϕ j u = (ϕ j u)(x(y )), y ∈ Rn−1 , and the
U j are mapped into the unit ball {y ∈ Rn : |y| < 1} so that (see Figure 20.1)

U j ∩ Ω → {y ∈ Rn : |y| < 1, y1 > 0}

and
U j ∩ ∂ Ω → {y ∈ Rn : |y| < 1, y1 = 0}.

For these purposes we may use the extension operator E from Proposition 20.25
such that (s ≥ 0)

m
E : H s (Ω ) → H s (U), U= Uj.
j=1

y1

∂Ω 1
Uj

y

Fig. 20.1 Representation of the boundary in local coordinates.

192 Part II: Fourier Transform and Distributions

Definition 20.27. A function u is said to belong to H s (∂ Ω ) if ϕ j u, j = 1, 2, . . . , m,

belong to H s (Rn−1 ) as a function of y . The norm in H s (∂ Ω ), s ≥ 0, is defined as
m  

u
H s (∂ Ω ) := ∑ ϕ j uH s (Rn−1 ) .
j=1

Remark 20.28. It can be shown that H s (∂ Ω ) is independent of the partition of unity

and that the norms are equivalent with respect to the different partitions of unity.

Let τ be a trace map (linear):

τ u := u|∂ Ω

for all u ∈ C∞ (Ω ) defined on Ω .

Proposition 20.29. For s > 1/2 the map τ extends uniquely to a continuous linear
map
τ : H s (Ω ) → H s−1/2 (∂ Ω )

with the norm estimate

τ u
H s−1/2 (∂ Ω ) ≤ C
u
H s (Ω ) .

Moreover, this map is surjective.

Proof. The proof is based on Proposition 20.22, the definition of H s (∂ Ω ), and the
following diagram:
τ
H s (Ω) H s−1/2 (∂ Ω)

E in local coordinates
τ
H s (U) H s−1/2 (Rn−1 )
See [39, p. 287] for details. 
Chapter 21
Homogeneous Distributions

We begin this chapter with the Fourier transform of a radially symmetric function.
Lemma 21.1. Let f (x) be a radially symmetric function in Rn , i.e., f (x) = f1 (|x|).
Let us assume also that f (x) ∈ L1 (Rn ). Then the Fourier transform f(ξ ) is also
radial and  ∞
f(ξ ) = |ξ |1− 2
n n
f1 (r)r 2 J n−2 (r|ξ |)dr,
0 2

where Jν (·) is the Bessel function of order ν .

Proof. Let us take the Fourier transform


f(ξ ) = (2π )− 2 e−i(x,ξ ) f1 (|x|)dx
n

R∞
n

e−i|ξ |r(ϕ ,θ ) dθ ,
n
−2
= (2π ) f1 (r)rn−1 dr
0 Sn−1

where x = rθ , ξ = |ξ |ϕ , and θ , ϕ ∈ Sn−1 := {x ∈ Rn : |x| = 1}. It is known [43] that

 n−1  π
2π 2
e−i|ξ |r(ϕ ,θ ) dθ = e−i|ξ |r cos ψ (sin ψ )n−2 dψ ,
Sn−1 Γ ( n−1
2 ) 0

where Γ is the gamma function. This fact implies that f(ξ ) is a radial function,
since the last integral depends only on |ξ |. A property of Bessel functions [23] is
that
 π   J n−2 (r|ξ |)
√ n−1
e−i|ξ |r cos ψ (sin ψ )n−2 dψ = 2 2 −1 πΓ
n
2
n−2 . (21.1)
0 2 (r|ξ |) 2

c Springer International Publishing AG 2017 193

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 21
194 Part II: Fourier Transform and Distributions

Collecting these things, we obtain

 ∞
f(ξ ) = |ξ | 2
2−n n
f1 (r)r 2 J n−2 (r|ξ |)dr,
0 2

and the claim follows. 

Remark 21.2. If we put the variable u = cos ψ in the integral I appearing in (21.1),
then we obtain
 π  1 
I= e−i|ξ |r cos ψ (sin ψ )n−2 dψ = e−i|ξ |ru ( 1 − u2 )n−3 du.
0 −1

In particular, if n = 3, then (21.1) implies that

 1
sin(|ξ |r) √ J 12 (r|ξ |)
I= e−i|ξ |ru du = 2 = 2π 1 ,
−1 |ξ |r (r|ξ |) 2

i.e., 
2 sin(|ξ |r)
J 1 (r|ξ |) = .
2 π (|ξ |r) 21

If n = 2, then
 1 −i|ξ |ru
e
I= √ du = π J0 (r|ξ |),
−1 1 − u2

i.e.,
 1 −i|ξ |ru
1 e
J0 (r|ξ |) = √ du.
π −1 1 − u2

Remark 21.3. For later considerations we state the small- and large-argument
asymptotics of Jν for ν > −1 as

cν |x|ν , |x| → 0+,
Jν (|x|) ≈  √1
cν cos(Aν |x| + Bν ), |x| → +∞
|x|

(see [23]).

Exercise 21.1. Prove that f(Aξ ) = f(ξ ) if A is a linear transformation in Rn with

rotation A = A−1 and f is radially symmetric.

Let us return again to the distribution (cf. Example 19.5)

 
1 −1 1
K1 (x) := n F (x).
(2π ) 2 1 + |ξ |2
21 Homogeneous Distributions 195

Let us assume now that n = 1, 2, 3, 4. Then the last integral can be understood in the
classical sense. It follows from Lemma 21.1 that
 ∞ r 2n J n−2 (r|x|)dr
1 (|x|) = (2π )
K1 (x) = K − 2n
|x|1− 2n 2
0 1 + r2
 ∞ ρ n2 J n−2 (ρ )dρ
− 2n
= (2π ) |x|
2−n 2
.
0 ρ 2 + |x|2

It is not too difficult to prove that for |x| < 1 we have

⎧
⎪
⎨1, n = 1,
|K1 (x)| ≤ c log |x|
1
, n = 2,
⎪
⎩ 2−n
|x| , n = 3, 4.

Exercise 21.2. Prove this fact.

Remark 21.4. A little later we will prove estimates for K1 (x) for every dimension
and for all x ∈ Rn .
There is one more important example. If we have the equation
n
 (−1 − Δ )u = f in
2 n −
L (R ) (or even in S), then formally u = (2π ) 2 F −1 1
∗ f = K−1 ∗ f , where
|ξ |2 −1

 ∞ ρ n2 J n−2 (ρ )dρ
− 2n
K−1 (|x|) = (2π ) |x| 2−n 2
.
0 ρ 2 − |x|2

But there is a problem with the convergence of this integral near ρ = |x|. Therefore,
this integral must be regularized as

 ∞ ρ n2 J n−2 (ρ )dρ
lim 2
.
ε →0+ 0 ρ 2 − |x|2 − iε

Recall that
(1) σλ f (x) := f (λ x), λ = 0 and
(2)
σλ T, ϕ  := λ −n
T, σ 1 ϕ , λ > 0.
λ

Definition 21.5. A tempered distribution T is said to be a homogeneous distribution

of degree m ∈ C if
σλ T = λ m T

for every λ > 0. In other words,

σλ T, ϕ  = λ m
T, ϕ ,
196 Part II: Fourier Transform and Distributions

or

T, ϕ  = λ −n−m
T, σ 1 ϕ ,
λ

for ϕ ∈ S. The space of all such distributions is denoted by Hm (Rn ).

Lemma 21.6. F : Hm (Rn ) → H−m−n (Rn ).
Proof. Let T ∈ Hm (Rn ). Then

σλ T, ϕ  = λ −n
T, σ 1 ϕ  = λ −n
T, σ
1 ϕ = λ
−n

T, λ n σλ ϕ
λ λ

=
T, σλ ϕ = λ −n
σ 1 T, ϕ = λ −n λ −m
T, ϕ = λ −n−m
T, ϕ 
λ

for all ϕ ∈ S. 
Definition 21.7. We set Hm∗ (Rn ) := {T ∈ Hm (Rn ) : T ∈ C∞ (Rn \ {0})}.
Exercise 21.3. Prove that
(1) if T ∈ Hm∗ , then Dα T ∈ Hm−|
∗ α ∗
α | and x T ∈ Hm+|α | ;
∗ ∗
(2) F : Hm → H−m−n .

Exercise 21.4. Let ρ (x) be a function from C∞ (Rn ) with |Dα ρ (x)| ≤ c
xm−|α | for
all α ≥ 0 and m ∈ R. Prove that ρ(ξ ) ∈ C∞ (Rn \ {0}) and (1 − ϕ )ρ ∈ S, where
ϕ ∈ C0∞ (Rn ) and ϕ ≡ 1 in Uδ (0).
∗ (Rn ). Indeed,
Example 21.8. (1) δ ∈ H−n

σλ δ , ϕ  = λ −n
δ , σ 1 ϕ  = λ −n σ 1 ϕ (0) = λ −n ϕ (0) = λ −n
δ , ϕ .
λ λ

But supp δ = {0}. This means that δ ∈ C∞ (Rn \ {0}). Alternatively, one could
note that
δ = (2π )− 2 · 1 ∈ H0∗ (Rn )
n

and use Exercise 21.3 to conclude that

n
δ = F −1 ((2π )− 2 · 1) ∈ H−n
∗
(Rn ).
 
(2) Let us assume that ω ∈ C∞ (Sn−1 ) and m > −n. Set Tm (x) := |x|m ω x
|x| for
x ∈ Rn \ {0}. Then Tm (x) ∈ Lloc
1 (Rn ) and Tm ∈ Hm∗ (Rn ). Indeed,
   
λx

σλ Tm , ϕ  = σλ Tm (x)ϕ (x)dx = |λ x|m ω ϕ (x)dx = λ m
Tm , ϕ .
Rn Rn |λ x|
 
Since |x|m and ω |x| x
are from C∞ (Rn \{0}), we have Tm ∈ Hm∗ (Rn ). Moreover,
Dα Tm ∈ Hm−|
∗ n α ∗
α | (R ) and x Tm ∈ Hm+|α | (R ) by Exercise 21.3.
n
21 Homogeneous Distributions 197


(3) Let now m = −n in part (2) and in addition assume that Sn−1 ω (θ )dθ = 0. Note
that T−n (x) ∈
/ Lloc
1 (Rn ). But we can define T 
−n as a distribution from S by


p. v. T−n , ϕ  := T−n (x)[ϕ (x) − ϕ (0)ψ (|x|)]dx,
Rn

where ϕ ∈ S(Rn ) and ψ ∈ S(R) with ψ (0) = 1. We assume that  ψ is fixed. But
it is clear that this definition does not depend on ψ , because Sn−1 ω (θ )dθ = 0.

Exercise 21.5. Prove that



p. v. T−n , ϕ  = lim T−n (x)ϕ (x)dx,
ε →0+ |x|≥ε

  
where T−n = |x|−n ω |x| , Sn−1 ω (θ )dθ = 0.
x

Let us prove the following:

(1) p. v. T−n ∈ H−n∗ ;

(2) p.v. T−n ∈ H0∗ (Rn ), and moreover, it is bounded.

(3) p. v. T−n ∗ : L2 (Rn ) → L2 (Rn ).

Proof. Part (1) is clear. Part (2) follows from

 
 
|
p. v. T−n , ϕ | = |
p. v. T−n , ϕ| = 
 T−n (x)[ϕ(x) − ϕ(0)ψ (|x|)]dx
R n
  
 
≤ (2π ) −2 n

|ϕ (ξ )|dξ  T−n (x)[e−i(x,ξ )
− ψ (|x|)]dx
R n R n
   ∞ 
 1 −ir(θ ,ξ ) 
= (2π )
n
−2
|ϕ (ξ )|dξ  ω (θ )dθ [e − ψ (r)]dr
R n Sn−1 0 r
≤ c ϕ L1 (Rn ) .

Hence p.
v. T−n ∈ L∞ (Rn ) by duality.
Finally, if f ∈ L2 (Rn ), then

v. T−n · f,
F (p. v. T−n ∗ f ) = (2π ) 2 p.
n

which implies that


n 


p. v. T−n ∗ f L2 (Rn ) ≤ (2π ) 2 p.
v. T−n  ·  f L2 .
L∞

This proves part (3). 

198 Part II: Fourier Transform and Distributions

Remark 21.9. In fact, it follows from Calderón–Zigmund theory that

p. v. T−n ∗ : L p (Rn ) → L p (Rn ), 1 < p < ∞.

Next we want to consider a more difficult case than the previous one. Define

1

p. v. , ϕ  := |x|−n [ϕ (x) − ϕ (0)ψ (|x|)]dx, (21.2)
|x|n Rn

where ϕ ∈ S and ψ ∈ S with ψ (0) = 1. But now we don’t have the condition
Sn−1 ω (θ )dθ = 0 as above. Therefore, (21.2) must depend on the function ψ (|x|).
We will try to choose an appropriate function ψ . Applying the operator σλ , we get
 
1 1

σλ p. v. n , ϕ  =
p. v. n , λ −n σ 1 ϕ 
|x| |x| λ
  x 
= |x|−n λ −n ϕ − ϕ (0)ψ (|x|) dx
Rn λ
= λ −n |y|−n [ϕ (y) − ϕ (0)ψ (λ |y|)]dy
Rn

= λ −n |y|−n [ϕ (y) − ϕ (0)ψ (|y|)]dy
Rn

− λ −n |y|−n ϕ (0)[ψ (λ |y|) − ψ (|y|)]dy
Rn
1
=
λ −n p. v. , ϕ  + Rest,
|x|n

where

Rest = −λ −n ϕ (0) |y|−n [ψ (λ |y|) − ψ (|y|)]dy
Rn
 ∞ 
ψ (λ r) − ψ (r)
= −λ −n
δ , ϕ  dr dθ
r
0 Sn−1
 ∞
ψ (λ r) − ψ (r)
= −ωn λ −n
δ , ϕ  dr,
0 r
n
and ωn = Γ2π( n2) is the area of the unit sphere Sn−1 . Let us denote the last integral by
2
G(λ ), λ > 0. Then
 ∞  ∞
1 1 1
G (λ ) = ψ  (λ r)dr = ψ  (t)dt = − ψ (0) = − .
0 λ 0 λ λ

We also have that G(1) = 0. We may therefore conclude that G(λ ) = − log λ , which
implies that
Rest = ωn λ −n log λ
δ , ϕ ,
21 Homogeneous Distributions 199

and so  
1 1
σλ p. v. n = λ −n p. v. n + ωn λ −n log λ · δ (x).
|x| |x|

Taking the Fourier transform, we get

    
1 −n 1 n
F σλ p. v. n =λ F p. v. n + (2π )− 2 ωn λ −n log λ ,
|x| |x|

or    
1 1 n
λ −n σ 1 F p. v. n = λ −n F p. v. n + (2π )− 2 ωn λ −n log λ ,
λ |x| |x|

or     
1 ξ 1 n
F p. v. n = F p. v. n (ξ ) + (2π )− 2 ωn log λ .
|x| λ |x|

Let us put now λ = |ξ |. Then

    
1 − 2n 1 ξ
F p. v. n (ξ ) = −(2π ) ωn log |ξ | + F p. v. n .
|x| |x| |ξ |

Since p. v. |x|1n for such ψ is a radial homogeneous distribution, we must have that
 
1
F p. v. n
|x|
  
ξ
is also a radial homogeneous distribution. Therefore, F p. v. |x|1n |ξ |
depends
 
 
only on  |ξξ |  = 1. So this term is a constant that depends on the choice of ψ . We
will choose our function ψ (|x|) so that this constant is zero. Then finally,
 
1 n
F p. v. n (ξ ) = −(2π )− 2 ωn log |ξ |.
|x|

Now let us consider T−m = |x|−m , 0 < m < n. It is clear that |x|−m ∈ Lloc
1 (Rn ). Thus

the situation here is simpler. We have




|x|−m , ϕ  =
|x|−m , ϕ
 = |x|−m ϕ(x)dx.
Rn

Lemma 21.1 implies that

 ∞ r n2 J n−2 (r|ξ |)  ∞

|x|−m = |ξ | 1− 2n 2
dr = |ξ |−n+m
n
ρ 2 −m J n−2 (ρ )dρ .
0 rm 0 2
200 Part II: Fourier Transform and Distributions

The last integral converges if n−1

2 < m < n. We may therefore write that

 −m = C
n−1
|x| n,m |ξ | , < m < n.
m−n
2

In fact, this is true even for m such that 0 < Re(m) < n, which follows by analytic
continuation on m. In order to calculate the constant Cn,m , let us apply this distribu-
|x|2
tion to ϕ = e− 2 . Since ϕ = ϕ , we get

|x|2 |ξ |2

|x|−m , e− 2  =
Cn,m |ξ |m−n , e− 2 .

The left-hand side is

  ∞
|x|2 r2
|x|−m e− 2 dx = ωn rn−m−1 e− 2 dr
Rn 0
 ∞  
n−m−2 n−m −1 n−m−2 n−m
=2 2 ωn t 2 e−t dt = 2 2 ωnΓ .
0 2

Using this, the right-hand side becomes

|ξ |2 n−(n−m)−2
m
Cn,m
|ξ |m−n , e− 2  = Cn,m 2 2 ωnΓ .
2
Therefore,  
m−2
m n−m−2 n−m
Cn,m 2 2 ωnΓ =2 2 ωnΓ ,
2 2

which gives us  
n −m Γ n−m
Cn,m = 2 2   .
2
Γ m2

Finally, we have  
 n
−m = 2 2 −m
Γ n−m
|x|  m2  · |ξ |m−n . (21.3)
Γ 2

Definition 21.10. The Hilbert transform H f of f ∈ S is defined by

 
1 1
H f := p. v. ∗ f ,
π x

i.e., 
1 f (t)dt
H f (x) = lim , x ∈ R.
π ε →0+ |x−t|≥ε x−t
21 Homogeneous Distributions 201

Exercise 21.6. Prove that

(1) H f L2 (R) =  f L2 (R) .
(2) Hilbert transform has an extension to functions from L2 (R).
(3) H 2 = −I, i.e., H −1 = −H.

(4) (H f1 , H f2 )L2 = ( f1 , f2 )L2 for f1 ∈ L p and f2 ∈ L p , where 1p + p1 = 1, 1 < p < ∞.
(5) H : L p (R) → L p (R), 1 < p < ∞, i.e.,
  
1 f (t)dt 
 
 π |x−t|≥ε x − t  p ≤ c  f L p
L

for ell ε > 0, where c does not depend on ε .

The multidimensional analogue of the Hilbert transform is developed in the follow-

ing definition.
Definition 21.11. The functions
xj
R j (x) := , x = 0, j = 1, 2, . . . , n,
|x|n+1

are called the Riesz kernels.

Remark 21.12. We can rewrite R j (x) in the form R j (x) = |x|−n ω j (x), where ω j (x) =
xj
|x| and conclude that

(1) Sn−1 ω j (θ )dθ = 0;
(2) R j (λ x) = λ −n R j (x), λ > 0.

These properties imply that we may define the Riesz transform by

R j ∗ f = p. v. R j ∗ f ,

because in our previous notation, R j (x) = T−n ∈ H−n∗ (Rn ) is a homogeneous distri-

bution. Let us calculate the Fourier transform of the Riesz kernels. By homogeneity,
it suffices to consider |ξ | = 1. We have

e−i(x,ξ ) x j
Rj (ξ ) = p.

n
v. R j (ξ ) = (2π )− 2 dx
Rn |x|n+1

n e−i(x,ξ ) x j
= lim (2π )− 2 dx.
ε →0+ ε <|x|<μ |x|n+1
μ →+∞

We split
  
e−i(x,ξ ) x j e−i(x,ξ ) x j e−i(x,ξ ) x j
dx = dx + dx =: I1 + I2 .
ε <|x|<μ |x|n+1 ε <|x|<1 |x|
n+1
1<|x|<μ |x|n+1
202 Part II: Fourier Transform and Distributions

For I1 we will use integration by parts:


1 ∂
I1 = e−i(x,ξ ) (|x|1−n )dx
1−n ε <|x|<1 ∂xj
  
e−i(x,ξ ) e−i(x,ξ ) x j e−i(x,ξ ) x j
= cn iξ j dx + dσ − dσ
ε <|x|<1 |x| |x|n |x|n
n−1
|x|=1 |x|=ε
 
e−i(x,ξ )
→ cn iξ j dx + e−i(x,ξ ) x j dσ − 0, ε → 0+.
|x|<1 |x|n−1 |x|=1

But
  
−i(x,ξ ) ∂ ∂
x je dσ = i e−i(x,ξ ) dσ = i cos(|ξ | · x1 )dσ
|x|=1 ∂ ξ j |x|=1 ∂ ξ j |x|=1

iξ j
=− x1 · sin(|ξ | · x1 )dσ = −iξ j ·C1 , |ξ | = 1,
|ξ | |x|=1

where we have used the fact that a rotation maps ξ to (|ξ |, 0, . . . , 0). Similarly, we
may conclude that
 
e−i(x,ξ )
dx = cos(|ξ |x1 )|x|1−n dx = C2 , |ξ | = 1.
|x|<1 |x|
n−1
|x|<1

If we collect all of these things, we obtain

I1 → Cn iξ j , |ξ | = 1

as ε → 0+. For I2 we will use the following technique:

 
e−i(x,ξ ) x j ∂ e−i(x,ξ )
I2 → dx = i dx
|x|>1 |x|
n+1 ∂ ξ j |x|>1 |x|n+1

∂
=i |x|−n−1 cos(|ξ | · x1 )dx
∂ ξ j |x|>1

iξ j x1 sin(|ξ |x1 )
=− dx = −iξ j · const, |ξ | = 1, μ → +∞.
|ξ | |x|>1 |x|n+1

Exercise 21.7. Prove the convergence of the last integral.

Collecting these integrals, we obtain that


p. v. R j = iξ j ·Cn
21 Homogeneous Distributions 203


for |ξ | = 1. But we know from Exercise 21.3 that p. v. R j ∈ H0∗ (Rn ). We conclude
ξ

that p. v. R j (ξ ) = iCn j . Moreover, we have
|ξ |


R
n
j · f = iCn ξ j f,
j ∗ f = (2π ) 2 R
|ξ |

or  
ξj 
Rj ∗ f = iCn F −1 f .
|ξ |

It is easy to see that √

π
Cn = − n/2 .
2 Γ ((n + 1)/2)

Corollary 21.13. It is true that

n
∑ R j ∗ R j ∗ = −C n δ ∗ .
j=1

Proof. By the above results we have

n n
ξj ξj
∑ F (R j ∗ R j ∗ f ) = ∑ iCn |ξ | · iCn |ξ | f(ξ ) = −(Cn )2 f(ξ ).
j=1 j=1

Taking the inverse Fourier transform, we obtain the claim. 

Remark 21.14. By Parseval’s equality we have
     
   ξj 
R j ∗ f  2 = 
R 
j ∗ f =C  ≤ C 
f

 f 2 = C  f L2 ,
L L2  |ξ |  2 L
L

i.e.,
R j ∗ : L2 (Rn ) → L2 (Rn ),

and it follows from Calderón–Zigmund theory that

R j ∗ : L p (Rn ) → L p (Rn ), 1 < p < ∞.

Let us now introduce the Riesz potential by

   
1  n 1
I −1 f := F −1 f (ξ ) = (2π )− 2 F −1 ∗ f = I1 ∗ f ,
|ξ | |ξ |

where by (21.3),
1
I1 (x) = cn .
|x|n−1
204 Part II: Fourier Transform and Distributions

Therefore, we have 
f (y)dy
I −1 f (x) = cn ,
Rn |x − y|n−1

where
1 Γ ((n − 1)/2)
cn = .
2 π (n+1)/2
∂
It is straightforward to verify that ∂ x j I1 = cn R j and hence

∂ −1
I f = cn R j ∗ f .
∂xj

We would like to prove that

I −1 : Lσs (Rn ) → Ws1 (Rn )

for some s and σ . Since R j ∗ is a bounded map from L2 (Rn ) to L2 (Rn ), we may
conclude that
∂ −1 2 n
I : L (R ) → L2 (Rn ). (21.4)
∂xj

Now let us assume for simplicity that n ≥ 3. Let us try to prove that

I −1 : Lσ2 (Rn ) → L2 (Rn ). (21.5)

Indeed, for f ∈ L2 (Rn ),

I −1 f ∈ L2 (Rn )

if and only if
1 
f ∈ L2 (Rn ).
|ξ |

Let us assume now that σ > 1.  

Lemma 20.17 implies that Lσ2 (Rn ) ⊂ Lr (Rn ) for all 1 ≤ r < 2 and σ > n 1r − 12 .
But for σ > 1 we may find appropriate r such that r < n+2 2n
. We conclude that for
a function f ∈ Lσ (R ) with σ > 1 it follows from the Hausdorff–Young inequality
2 n
 r
that f ∈ Lr (Rn ) for some r > n−2
2n
or | f|2 ∈ L 2 (Rn ). This fact implies that for
|ξ | < 1 we have
|ξ |−1 f(ξ ) ∈ Lloc
2
.

Indeed,

  2/r      1

 −2 r2 r 
|ξ | | f(ξ )|2 dξ ≤
−2
| f(ξ )|r dξ |ξ | dξ 2 < ∞,
|ξ |<1 |ξ |<1 |ξ |<1
21 Homogeneous Distributions 205
  
r
since 2 > n−2
n
and < n2 . For |ξ | > 1 the function |ξ |−1 f(ξ ) belongs to L2 (Rn ).
r
2
This fact follows from the inequality |ξ |−1 | f(ξ )| < | f(ξ )| and from the positivity
of σ (see Lemma 20.17). This proves (21.5) for σ > 1.
If we combine (21.4) and (21.5), we obtain that

I −1 : Lσ2 (Rn ) → W21 (Rn ), σ > 1.

Let us consider now Lσ∞ (Rn ) for σ > 1. If f ∈ Lσ∞ (Rn ), then | f (x)| ≤ C(1 + |x|)−σ
and thus 
(1 + |y|)−σ dy
|I −1 f (x)| ≤ C < ∞.
Rn |x − y|n−1

This means that

I −1 : Lσ∞ (Rn ) → L∞ (Rn ).

Interpolating this with (21.5), we can obtain the following result:

I −1 : Lσs (Rn ) → Ls (Rn ), 2 ≤ s ≤ ∞, σ > 1.

If we recall the fact that R j ∗ : Ls (Rn ) → Ls (Rn ) for all 1 < s < ∞, then we have

I −1 : Lσs (Rn ) → Ws1 (Rn ), 2 ≤ s < ∞, σ > 1.

Chapter 22
Fundamental Solution of the Helmholtz
Operator

Let us consider a linear partial differential operator of order m in the form

L(x, D) = ∑ aα (x)Dα , x ∈ Rn ,
|α |≤m

where α = (α1 , . . . , αn ) is a multi-index, Dα = Dα1 1 · · · Dαn n , and D j = 1 ∂

i ∂xj .
In this chapter, Ω is a bounded domain in Rn , or Ω = Rn .
Definition 22.1. A fundamental solution for L in Ω is a distribution E in x that
satisfies
Lx E(x|y) = δ (x − y)

in D  (Ω ) with parameter y ∈ Ω , i.e., Lx E, ϕ  = ϕ (y) for ϕ ∈ C0∞ (Ω ).

We understand that LE, ϕ  is defined in distributional form

LE, ϕ  = E, L ϕ ,

where L is the formal adjoint operator of L given by

L f = ∑ (−1)|α | Dα (aα (x) f (x)).

|α |≤m

Here, L ϕ must be in D(Ω ) for ϕ from D(Ω ). This will be the case, for example,
for aα (x) ∈ C∞ (Ω ).
Two fundamental solutions for L with the same parameter y differ by a solution
of the homogeneous equation Lu = 0. Unless boundary conditions are imposed,
the homogeneous equation will have many solutions, and the fundamental solution
will not be uniquely determined. In most problems there are grounds of symmetry
or causality for selecting the particular fundamental solution for the appropriate
physical behavior.
c Springer International Publishing AG 2017 207
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 22
208 Part II: Fourier Transform and Distributions

We also observe that if L has constant coefficients, we can find the fundamental
solution in the form E(x|y) = E(x − y|0) := E(x − y). This fact follows from the
properties of the Fourier transform:

  
Lx E(x − y) = ∑ aα ξ α E(x − y) = ∑ aα ξ α e−i(ξ ,y) E(x)
|α |≤m |α |≤m

= e−i(ξ ,y) δ
(x) = δ
(x − y),

i.e.,
Lx E(x − y) = δ (x − y).

Exercise 22.1. Let L be a differential operator with constant coefficients. Prove that
u = q ∗ E = E ∗ q solves the inhomogeneous equation

Lu = q

in D .

Remark 22.2. In many cases the fundamental solution is a function. We can there-
fore write u as an integral

u(x) = E(x − y)q(y)dy.
Ω

Remark 22.3. In order for the convolution product E ∗q (or q∗E) to be well defined,
we have to assume that, for example, q vanishes outside a finite sphere.

Remark 22.4. If L does not have constant coefficients, we can no longer appeal to
convolution products; instead, one can often show that

u(x) = E(x|y)q(y)dy.
Ω

Definition 22.5. We denote by a0 (x, ξ ) the main (or principal) symbol of L(x, D)

a0 (x, ξ ) = ∑ aα (x)ξ α , ξ ∈ Rn .
|α |=m

Assume that the aα (x) are “smooth.” An operator L(x, D) is said to be elliptic in Ω
if for every x ∈ Ω and ξ ∈ Rn \{0} it follows that

a0 (x, ξ ) = 0.

Exercise 22.2. Let aα (x) be real for |α | = m. Prove that the previous definition is
equivalent to
22 Fundamental Solution of the Helmholtz Operator 209

(1) m is even,
(2) a0 (x, ξ ) ≥ CK |ξ |m (or −a0 (x, ξ ) ≥ CK |ξ |m ), CK > 0, for every compact set K ⊂
Ω and for all ξ ∈ Rn and x ∈ K.
Let us consider the heat equation
⎧
⎨∂u
= Δ u, t > 0, x ∈ Rn ,
∂t
⎩u(x, 0) = f (x), x ∈ Rn

in S (Rn ). Take the Fourier transform with respect to x to obtain

⎧
⎨∂
u(ξ ,t) = −ξ 2 u(ξ ,t), t > 0,
∂t
⎩u(ξ , 0) = f(ξ ).

This initial value problem for an ordinary differential equation has the solution

u(ξ ,t) = e−t|ξ | f(ξ ).

2

Hence

u(x,t) = F −1 (e−t|ξ | f(ξ )) = (2π )− 2 F −1 (e−t|ξ | ) ∗ f = P(·,t) ∗ f ,

2 n 2

where 
1 |x|2
P(x,t) = (2π )−n e−t|ξ | ei(x,ξ ) dξ = − 4t
2
n e .
Rn (4π t) 2

This formula implies that


1 |x−y|2
u(x,t) = n e− 4t f (y)dy.
(4π t) 2 Rn

Definition 22.6. The function P(x,t) is the fundamental solution of the heat equa-
tion and satisfies ⎧

⎪ ∂
⎨ − Δ P(x,t) = 0, t > 0,
∂t
⎪
⎩ lim P(x,t) = S
δ (x).
t→0+

It is also called the Gaussian kernel or heat kernel.

We can generalize this situation as follows. Let us consider an elliptic differential
operator
L(D) = ∑ aα Dα
|α |≤m
210 Part II: Fourier Transform and Distributions

with constant coefficients. Assume that L(ξ ) = ∑|α |≤m aα ξ α > 0 for all ξ ∈ Rn \{0}.
If we consider PL (x,t) as a solution of
⎧

⎪ ∂
⎨ + L(D) PL (x,t) = 0, t > 0,
∂t
⎪ S
⎩ lim PL (x,t) = δ (x),
t→0+

∂
then PL (x,t) is the fundamental solution of ∂t + L(D) and can be calculated by

PL (x,t) = (2π )−n e−tL(ξ ) ei(x,ξ ) dξ .
Rn

Lemma 22.7. Let PL (x,t) be as above. Then the function

 ∞
S
F(x, λ ) := lim e−λ t PL (x,t)dt
ε →0+ ε

is a fundamental solution of the operator L(D) + λ I, λ > 0.

Proof. By the definitions of F and PL we have

 ∞  ∞
F(x, λ ), ϕ  = lim  e−λ t PL (x,t)dt, ϕ  = lim e−λ t PL , ϕ dt.
ε →0+ ε ε →0+ ε

Therefore,
 ∞
(L(D) + λ )F, ϕ  = lim e−λ t (L(D) + λ )PL , ϕ dt
ε →0+ ε
 ∞  ∞
= lim e−λ t L(D)PL , ϕ dt + λ e−λ t PL , ϕ dt
ε →0+ ε 0
 ∞
∂
= lim e−λ t − PL , ϕ dt + λ F, ϕ 
ε →0+ ε ∂t
  ∞

= lim −e−λ t PL , ϕ |∞
ε − λ e−λ t PL , ϕ dt + λ F, ϕ 
ε →0+ ε
−λ ε
= lim e PL (·, ε ), ϕ  = δ , ϕ 
ε →0+

for all ϕ ∈ S. 
∂
Exercise 22.3. Let us define a fundamental solution Γ (x,t) of ∂t + L(D) as a solu-
tion of
( ∂∂t + L)Γ (x,t) = δ (x)δ (t),
Γ (x, 0) = 0.
22 Fundamental Solution of the Helmholtz Operator 211

Prove that  ∞
F(x, λ ) := e−λ t Γ (x,t)dt
0

is a fundamental solution of the operator L(D) + λ I, λ > 0.

 2
Example 22.8. Let us consider L(D) = ∑nj=1 1i ∂∂x j = −Δ . Then L(ξ ) = |ξ |2 ,
and the fundamental solution F(x, λ ) of the operator L(D) + λ = −Δ + λ has the
form
 ∞  ∞
1 2
−λ t − x4t 1 x2
e−λ t− 4t · t − 2 dt
n
F(x, λ ) = n e · e dt = n
0 (4π t) 2 (4π ) 2 0
 ∞ √  ∞
1 ( λ |x|)2 1 r2
−τ − 4τ
e−τ − 4τ τ − 2 dτ ,
n −1 − n2 n −1 n
= n λ e τ dτ = n λ
2 2
(4π ) 2 0 (4π ) 2 0

√
where r = λ |x|. From our previous considerations we know that

1
F(x, λ ) = (2π )−n/2 F −1 (x),
|ξ |2 + λ

where F −1 is the inverse Fourier transform. The function

 ∞
1  r ν r2
Kν (r) = e−t− 4t · t −1−ν dt
2 2 0

is called the Macdonald function of order ν . So we have

1− n √
n |x| 2
F(x, λ ) = (2π )− 2 √ K n2 −1 ( λ |x|).
λ

It is known that
π i iπ ν (1)
Kν (r) = e 2 Hν (ir), r > 0,
2
(1)
where Hν is the Hankel function of first kind of order ν .
Next we want to obtain estimates for F(x, λ ) for x ∈ Rn , λ > 0, and n ≥ 1. Let
 ∞ −τ − r2 − n 1 ∞
4τ τ 2 dτ in two parts I1 + I2 =
us consider the integral 0 e 0 + 1 .

(1) If 0 < r < 1, then

 1  1  ∞
r2 r2
e−y− 4y y− 2 dy ≤ e− 4y y− 2 dy = cn r2−n
n n n
I1 = r2
e−z z 2 −2 dz = cn r2−n I1 .
0 0 4
212 Part II: Fourier Transform and Distributions

Since ⎧
⎪ −1
⎨r , n = 1,

I1 ∼ c log 1r , n = 2,
⎪
⎩
1, n ≥ 3,

as r → 0+, we have ⎧
⎪
⎨1, n = 1,
|I1 | ≤ cn log r , n = 2,
1
⎪
⎩ 2−n
r , n ≥ 3.

For I2 we can simply argue that

 ∞  ∞
r2 r2 r2
e−y− 4y y− 2 dy ≤ e− 4
n
I2 = e−y dy ≤ e− 4 ≤ 1, r → 0+.
1 1

(2) If r > 1, then

 1  ∞
r2
r2
r−2 e− 4 , n = 1, 2, 3, 4
e− 4y y− 2 dy
n −z n2 −2
I1 ≤ = cn r 2−n
e z dz ≤ cn
0 r2 2−n
r e − δ r 2
, n ≥ 5,
4

n
where 0 < δ < 14 . The last inequality follows from the fact that z 2 −2 ≤ cε eε z for
2 − 2 > 0 and all ε > 0 (z > 1).
n

Since  ∞
r2
I2 ≤ e−y− 4y dy,
1

2
we perform the change of variable z := y + 4y
r
. Then z ≥ r and z → +∞. Thus
 ∞  ∞

r2 z
e−y− 4y dy = c e−z 1 + √ dz
1 r z2 − r 2
 ∞  ∞
zdz
=c e−z dz + c e−z √
r r z2 − r 2
 ∞  ∞ 


= ce−r + c e−z z2 − r2  + e−z z2 − r2 dz
r r
 ∞ 

= c e−r + e−z z2 − r2 dz ≤ ce−δ r

r

for all 0 < δ < 1.

22 Fundamental Solution of the Helmholtz Operator 213

If we collect all these estimates, we obtain the following:

√
(1) If λ |x| < 1, then
⎧
⎪
⎨1, n = 1,
n −1
|F(x, λ )| ≤ cn λ 2 log √ 1
, n = 2,
⎪
⎩√
λ |x|
( λ |x|) , n ≥ 3,
2−n
⎧
√
⎪
⎨1, n = 1,
 n2 −1 −δ λ |x| log √ 1 , n = 2,
≤ cn λ e
⎪
⎩√
λ |x|
( λ |x|) , n ≥ 3.
2−n

√
(2) If λ |x| > 1, then
√
|F(x, λ )| ≤ cn e−δ λ |x|
, n ≥ 1.

We will rewrite these estimates in a more appropriate form for all λ > 0 and x ∈
Rn as ⎧
⎪
⎨ λ, n = 1,
⎪ √1
√
−δ λ |x|
|F(x, λ )| ≤ cn e 1 + | log √ |, n = 2,
1
⎪
⎪ λ |x|
⎩|x|2−n , n ≥ 3.

Remark 22.9. It is not too difficult to observe that F(x, λ ) is positive.

Example 22.10. Recall from Chapter 21 that the solution of the equation
(−1 − Δ )u = f can be written in the form

−1 1
u(x) = K−1 ∗ f = F ∗ f,
|ξ |2 − 1

where
 ∞ ρ n2 J n−2 (ρ )dρ
K−1 (|x|) = cn |x|2−n
lim 2
.
ε →0+ 0 ρ 2 − |x|2 − iε

In fact, K−1 is a fundamental solution of the operator −1 − Δ . Let us consider the

more general operator −Δ − λ for λ > 0 or even for λ ∈ C. The operator −Δ − λ
is called the Helmholtz operator. Its fundamental solution En (x, λ ) satisfies

−Δ En − λ En = δ (x).
√ √
We define λ with nonnegative imaginary part, i.e., λ = α + iβ , where β ≥ 0
and β = 0 if and only if λ ∈ [0, +∞). We require that En is radially symmetric. Then
for x = 0, En must solve the equation
214 Part II: Fourier Transform and Distributions

(rn−1 u ) + λ rn−1 u = 0.

This equation can be reduced to one of Bessel type by making the substitution u =
n
wr1− 2 . A straightforward calculation shows that
 n 2 w
(rw ) − 1 − + λ rw = 0,
2 r
or 

w n 2 1
w + + λ − 1− w = 0,
r 2 r2

or
√ 
√

√ v (r λ ) n 2 1 √
v (r λ ) + √ + 1− 1− v(r λ ) = 0, w(r) = v(r λ ).
r λ 2 λ r2

This is the Bessel equation of order n2 − 1. Its two linearly independent solutions
are the Bessel functions J n2 −1 and Y 2n −1 of the first and second kinds, respectively.
Therefore the general solution is of the form
√ √
w(r) = c0 J n2 −1 ( λ r) + c1Y n2 −1 ( λ r).

For us it is convenient to write it in terms of Hankel functions of the first and second
kinds as √ √
(1) (2)
w(r) = c0 H n −1 ( λ r) + c1 H n −1 ( λ r),
2 2

where
(1) (2)
Hν (z) = Jν (z) + iYν (z), Hν (z) = Jν (z) − iYν (z).

The corresponding general solution u is

 √ √ 
n (1) (2)
u(r) = r1− 2 c0 H n −1 ( λ r) + c1 H n −1 ( λ r) .
2 2

√ (2) √
If λ ∈
/ [0, +∞), then λ has positive imaginary part, and the solution H n −1 ( λ r) is
(1) √
2
exponentially large at z = +∞, whereas H n −1 ( λ r) is exponentially small. Hence
2
we take √
1− 2n (1)
En (x, λ ) = c0 r H n −1 ( λ r).
2

Exercise 22.4. Prove that


∂ En
lim dσ (x) = 1
ε →0+ |x|=ε ∂r
22 Fundamental Solution of the Helmholtz Operator 215

or
∂ En
lim rn−1 ωn = 1,
r→0 ∂r

where ωn = |Sn−1 | is the area (measure) of the unit sphere Sn−1 .

For small values of r, we have the asymptotic expansions [23]

n−2
(1) i2 2 Γ ( n−2
2 ) − n−2
H n−2 (r) ∼ − r 2 , n = 2
2 π

and
(1) 2i
H0 (r) ∼ log r.
π
It can be proved using Exercise 22.4 that
 √  n−2
λ
2
i
c0 = .
4 2π

Thus for n ≥ 2 and λ ∈

/ [0, +∞) we obtain
 √  n−2
i λ
2
(1)
√
En (x, λ ) = H n−2 ( λ |x|). (22.1)
4 2π |x| 2

A direct calculation shows that for n = 1 we have

√
i
E1 (x, λ ) = √ ei λ |x|
2 λ

for all λ = 0. The formula (22.1) is valid also for λ ∈ (0, +∞). This fact follows
from the definition:
√  n−2
i λ + iε
2
(1)

En (x, λ ) = lim En (x, λ + iε ) = lim H n−2 ( λ + iε |x|)
ε →0+ 4 ε →0+ 2π |x| 2

 √  n−2
i λ
2
(1)
√
= H n−2 ( λ |x|).
4 2π |x| 2
216 Part II: Fourier Transform and Distributions

Remark 22.11. We conclude that

 √  n−2
1 i λ
2
(1)
√
(2π )−n/2 F −1 = H n−2 ( λ |x|),
|ξ | − λ − i0
2 4 2π |x| 2

for λ > 0. A direct calculation shows that

⎧
⎪ |x|
⎨− 2 ,
⎪ n = 1,
En (x, 0) = 2π log |x| , n = 2,
1 1
⎪
⎪
⎩ |x|2−n , n ≥ 3.
(n−2)ωn
Chapter 23
Estimates for the Laplacian and Hamiltonian

Let us recall Agmon’s (2, 2)-estimate for the Laplacian [2]:

c
(−Δ − k2 − i0)−1 L2 →L2 ≤ , (23.1)
δ −δ |k|

where (−Δ − k2 − i0)−1 is an integral operator with kernel En (x, k) from the previ-
ous chapter and δ > 12 . In fact, this estimate allows us to consider the Hamiltonian
∞ -potentials only (if we want to preserve (2, 2)-estimates). But we would
with Lloc
p
like to consider the Hamiltonian with Lloc -potentials. We therefore need to prove
(p, q)-estimates.
1 1
We proved in Example 18.26 that the limit lim x−i ε := x−i0 exists in the sense
ε →0+
of tempered distributions and

1 1
= p. v. + iπδ (x),
x − i0 x

i.e., 
1 ϕ (x)
 , ϕ  = lim dx + iπϕ (0).
x − i0 δ →0+ |x|>δ x

In Example 18.8 we considered the simple layer


T, ϕ  := a(ξ )ϕ (ξ )dσξ ,
σ

where σ is a hypersurface of dimension n − 1 in Rn and a(ξ ) is a density. These

examples can be extended as follows. If H : Rn → R and |∇H| = 0 at every point
where H(ξ ) = 0, then we can define the distribution

1
(H(ξ ) − i0)−1 := lim
ε →0+ H(ξ ) − iε
c Springer International Publishing AG 2017 217
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 23
218 Part II: Fourier Transform and Distributions

in S (R), and we can also prove that

1
(H(ξ ) − i0)−1 = p. v. + iπδ (H(ξ ) = 0),
H(ξ )

where δ (H(ξ ) = 0) is defined as follows:


δ (H), ϕ = ϕ (ξ )dσξ , ϕ ∈ S(Rn ).
H(ξ )=0

The equality H(ξ ) = 0 defines an (n − 1)-dimensional hypersurface, and σξ is any

(n − 1)-form such that d σξ ∧ |∇H|
dH
= dξ (in local coordinates).

Exercise 23.1. Prove that

1
δ (α H) = δ (H)
α
for every positive differentiable function α .
 
Due to Exercise 23.1 we may conclude that δ (H) = |∇H| 1
δ |∇H|
H
if |∇H| = 0
for H = 0.
Let us consider now H(ξ ) := −|ξ |2 + k2 , k > 0. Then H(ξ ) = 0 or |ξ | = k is a
sphere and ∇H(ξ ) = −2ξ and |∇H(ξ )| = 2k at every point on this sphere. If we
change variables, we then obtain
 
1
δ (H), ϕ = ϕ (ξ )dσξ = ϕ (kθ )dθ .
H(ξ )=0 2k Sn−1

We know that (−Δ − k2 − i0)−1 f can be represented as


(−Δ − k2 − i0)−1 f = G+
k (|x − y|) f (y)dy,
Rn

  n−2
|k| (1)
where G+
2
k (|x|) = H n−2 (|k||x|). On the other hand, we can write
i
4 2π |x|
2


n f(ξ )ei(x,ξ ) dξ
(−Δ − k2 − i0)−1 f = F −1 (F [(−Δ − k2 − i0)−1 f ]) = (2π )− 2
Rn |ξ |2 − k2 − i0
 n 
1 iπ (2π )− 2
f(ξ )ei(x,ξ ) dξ + δ (H) f(ξ )ei(x,ξ ) dξ
n
= (2π )− 2 p. v.
Rn |ξ |2 − k2 2k Rn
  
f (ξ )ei(x,ξ ) dξ iπ
f(kθ )eik(x,θ ) dθ
n
= (2π )− 2 p. v. +
|ξ |2 − k2
n
Rn 2k(2π ) 2 Sn−1
   
f (ξ )ei(x,ξ ) dξ iπ
eik(θ ,x−y) dθ .
n
= (2π )− 2 p. v. + f (y)dy
Rn |ξ |2 − k2 2k(2π )n Rn Sn−1
23 Estimates for the Laplacian and Hamiltonian 219

Our aim is to prove the following result.

Theorem 23.1. Let k > 0 and 2n ≥ 1p − p1 ≥ n+1 2
for n ≥ 3 and 1 > 1p − p1 ≥ 23 for
n = 2, where 1p + p1 = 1. Then there exists a constant C independent of k and f such
that  
n 1p − p1 −2
(−Δ − k2 − i0)−1 f L p (Rn ) ≤ Ck  f L p (Rn ) .

k instead of (−Δ − k2 −
Remark 23.2. In what follows we will use the notation G
−1
i0) .

Proof. First we prove that if the claim holds for k = 1, then it holds for every k > 0.
So let us assume that
G 1 f  p n ≤ C f L p (Rn ) .
L (R )

n
Set Tδ f := f (δ x), δ > 0. It is clear that Tδ f L p (Rn ) = δ − p  f L p (Rn ) . It is not dif-
ficult to show that Gk = k−2 Tk G1 T 1 . Indeed, since
k

 
k f = (2π )−n ei(y,ξ ) f (x − y)dξ dy
G ,
Rn Rn |ξ |2 − k2 − i0

we get
 
ei(y,ξ ) f ( x−y
k )dξ dy
1 T 1 f = (2π )−n
G .
k Rn Rn |ξ | − 1 − i0
2

It follows that
 
ei(y,ξ ) f (x − ky )dξ dy
1 T 1 f = (2π )−n
Tk G
k Rn Rn |ξ |2 − 1 − i0
 
ei(z,η ) k−n f (x − z)kn dzdη
= (2π )−n
|η |2
Rn Rn
k2
− 1 − i0
  η )
ei(z, f (x − z)dzdη
= (2π )−n k2 .
Rn Rn |η |2 − k2 − i0

This proves that

k f ≡ Tk G
k2 G 1 T 1 f ,
k

which we use to get

k f  p = k−2 Tk G
1 T 1 f  p = k−2 k − pn 1 T 1 f  p
G L L
G L
k k
 − n  
−2− pn −2− pn 1 p
n 1 − 1 −2
≤ Ck T 1 f L p = Ck  f L p = Ck p p  f L p .
k k
220 Part II: Fourier Transform and Distributions

It therefore suffices to prove this theorem for k = 1.

The rest of the proof makes use of the following lemmas.

 
ξ
Lemma 23.3. Let ω (x) ∈ S(Rn ), 0 < ε < 1, and σε ω (ξ ) = ε −n ω ε . Let us set
 
1
Pε (ξ ) := p. v. ∗ σε ω (ξ ).
|η |2 − 1

Then
c
|Pε (ξ )| ≤ .
ε
Proof. For Pε we have the following representation:
   
σε ω (ξ − η )
Pε = p. v. + + dη =: I1 + I2 + I3 .
1−ε ≤|η |≤1+ε |η |<1−ε |η |>1+ε |η |2 − 1

−1
The integrals I2 and  I3 can be easily bounded by ε ω L1 , because  |η | < 1 − ε
 1   1 
implies that  |η |2 −1  = 1−|η |2 < ε and |η | > 1+ ε implies that  |η |2 −1  = |η |21−1 < ε1 .
1 1

By the definition of p. v. we have


σε ω (ξ − η )
I1 = lim dη
|η |2 − 1
δ →0+ δ <|1−|η ||<ε
 1−δ  1+ε  
rn−1
= lim + σε ω (ξ − rθ ) 2 dθ dr.
δ →0+ 1−ε 1+δ Sn−1 r −1

Replacing r with 2 − r in the latter integral, we obtain

 1−δ
F(r, ξ )
I1 = lim dr,
δ →0+ 1−ε r−1

where


rn−1 (2 − r)n−1
F(r, ξ ) = σε ω (ξ − rθ ) − σε ω (ξ − (2 − r)θ ) dθ .
Sn−1 r+1 3−r

If we observe that F(1, ξ ) = 0, then we get by the mean value theorem (Lagrange
formulas) that
 1−δ     
 F(r, ξ )   1−δ F(r, ξ ) − F(1, ξ )  ∂F 
 dr ≤ (ε − δ ) sup   (r, ξ )
 1−ε r − 1 dr =  1−ε r−1 1−ε <r<1 ∂ r
 
∂F 
≤ ε sup  .
∂r 
1−ε <r<1
23 Estimates for the Laplacian and Hamiltonian 221

But
   
∂F rn−1 rn−1
= σε ω (ξ − rθ )dθ − θ · ∇(σε ω (ξ − rθ ))dθ
∂r r+1 Sn−1 r + 1 Sn−1
  
(2 − r)n−1
− σε ω (ξ − (2 − r)θ )dθ
3−r Sn−1

(2 − r)n−1
− θ · ∇(σε ω (ξ − (2 − r)θ ))dθ =: θ1 + θ2 + θ3 + θ4 .
3−r Sn−1

By the proof of Lemma 23.4 below we get |θ1 | ≤ c1 ε −1 and |θ3 | ≤ c3 ε −1 , where
the constants c1 and c3 depend on ω . The second integral, θ2 , can be estimated as
(see Lemma 23.4)
  
n
rn−1 ∂
ε −1 ∑ θ j σε ω (ξ − rθ )dθ ≤ c2 ε −2 .
j=1 r + 1 Sn−1 ∂xj

The same estimate holds for θ4 . Thus, Lemma 23.3 is proved. 



Lemma 23.4. Let us assume that f ∈ L∞ (Sn−1 ) and ω ∈ S(Rn ). Then

 
 
  ≤ Cε −1 .
 Sn−1 σε ω (ξ − θ ) f (θ )dθ 
L∞ (Rn )

S
Proof. We can reduce the proof to compactly supported ω , since C0∞ = S. Let 
us con-

sider a C0∞ partition of unity in Rn such that ∑∞j=0 ψ j (ξ ) = 1 or even ∑∞j=0 ψ j 1
ξ =
1, where ψ0 is supported in |ξ | < 1 and ψ j = ψ (2− j ξ )
for j = 1, 2, 3, . . . with ψ
supported in the annulus 1/2 < |ξ | < 2. We may therefore write
 ∞     
ξ −θ ξ −θ
Sn−1
σε ω (ξ − θ ) f (θ )dθ = ∑ ε −n ψ j
ε
ω
ε
f (θ )dθ .
j=0 S
n−1

   
ξ −θ
For j = 1, 2, 3, . . . , the function ψ j ε ω ξ −ε θ is supported in the annulus
2 j−1 ≤ | · | ≤ 2 j+1 . Since ω is rapidly decreasing, we have that in this annulus,
  
 
ω ξ − θ  ≤ CM
,
 ε  (1 + 2 j )M

for all M ∈ N. Hence

   
 ξ −θ  (2 j+1 ε )n−1 −n  −1 (2 )
j n−1
 −n
f (θ )dθ  ≤ CM
 Sn−1 ε (ψ j ω ) ε (1 + 2 j )M
ε ≤ CM ε
(1 + 2 j )M
.
222 Part II: Fourier Transform and Distributions

Taking M large enough, we see that the sum in j converges to Cε −1 . To end the
proof of Lemma 23.4, notice that the term for j = 0 satisfies this inequality trivially.


Exercise 23.2. Prove that (−Δ )−1 : Lδ2 (R3 ) → L−
2 (R3 ) for δ > 1.
δ

1 f in the form
Let us return to the proof of Theorem 23.1. We can rewrite G

1 f = C p. v. f(ξ )ei(x,ξ ) dξ
G + I1 f ,
Rn |ξ |2 − 1

where 
I1 f = C f(θ )ei(θ ,x) dθ .
Sn−1

Let us take a partition of unity ∑∞j=0 ψ j (x) = 1 such that supp ψ0 ⊂ {|x| < 1} and
supp ψ j ⊂ {2 j−1 < |x| < 2 j+1 }, where ψ j = ψ (2− j x) with a fixed function ψ ∈ S.
We set Ψj := ψ j G+ +
1 and K j f := Ψj ∗ f , where G1 is the kernel of the integral operator
1 . Using the estimates of the Hankel function H (1)
G n−2 (|x|) for |x| < 2, we obtain
2

|Ψ0 | ≤ C|x|2−n , n ≥ 3,

and
|Ψ0 | ≤ C(|log |x|| + 1), n = 2.

Exercise 23.3 (Sobolev inequality). Let 0 < α < n, 1 < p < q < ∞, and 1q = 1p − αn .
Prove that  
 f (y)dy 
 
 Rn |x − y|n−α  q ≤ C f L p .
L

Hint: For K := |x|−n+α use the representation K = K1 + K2 , where

 
K, |x| < μ , 0, |x| < μ ,
K1 = and K2 =
0, |x| > μ , K, |x| > μ .

From Sobolev’s inequality for α = 2 we may conclude that the operator K0 is


bounded from L p (Rn ) → L p (Rn ) for the range 2n ≥ 1p − p1 ≥ 0 if n ≥ 3, and for
the range 1 > 1p − p1 ≥ 0 if n = 2. From Lemmas 23.3 and 23.4 with ε = 21j we can
obtain that
F (Ψj )∞ = (|ξ |2 − 1 − i0)−1 ∗ ψ j ∞ ≤ C · 2 j .

This inequality leads to

K j L2 →L2 ≤ C · 2 j ,
23 Estimates for the Laplacian and Hamiltonian 223

because

j · f 2 ≤ Ψ
K j f L2 = F (Ψj ∗ f )L2 = CΨ j L∞  f 2 ≤ C · 2 j  f  2 .
L L L

On the other hand, due to the estimate of the fundamental solution at infinity we can
n−1
obtain that |Ψj (x)| ≤ C · 2− j· 2 and

n−1
K j L1 →L∞ ≤ C · 2− j· 2 .

We have used here two facts:

 
 (1) 
H n−2 (|x|) ≤ C , |x| > 1,
 2  1
|x| 2

and supp Ψj (x) ⊂ {x : 2 j−1 < |x| < 2 j+1 }. Interpolating these estimates, we obtain
the self-dual estimates
   
2 1− 1p − n−1 2 −1
K j L p →L p ≤ C(2 j ) 2 p
.

For convergence of this series we need the condition 2(1 − 1p ) − n−1

2 ( p − 1) < 0, or
2

p − p > n+1 . If we want to get the sharper inequality p − p ≥ n+1 , we have to use
1 1 2 1 1 2

Stein’s theorem on interpolation [37]. Thus, Theorem 23.1 is proved. 


It follows from Theorem 23.1 that if we consider the values of p from the interval

2n 2n + 2
≤ p≤ , n ≥ 3,
n+2 n+3

1 < p ≤ 6/5, n = 2,

then we have the self-dual estimate

k  p p ≤ C
G L →L
 .
2−n 1p − p1
|k|

But we would like to extend the estimates for G k for 2n ≤ p ≤ 2, n ≥ 3, and

n+2
1 < p ≤ 2, n = 2. In order to do so, we use interpolation of Agmon’s estimate and
the latter estimate for p = 2n+2
n+3 . This process leads to the estimate

k  C
G p p ≤  ,
Lδ →L−δ 1−(n−1) 1p − 21
|k|
 
where 2n+2
n+3 < p ≤ 2, n ≥ 2, and δ > 12 − (n + 1) 1
2p − 14 .
224 Part II: Fourier Transform and Distributions

Theorem 23.5. Assume that the potential q(x) belongs to Lσp (Rn ), n ≥ 2, with n2 <
p ≤ ∞ and σ = 0 for n2 < p ≤ n+1
2 and σ > 1 − 2p for 2 < p ≤ +∞. Then for
n+1 n+1

all k = 0, the limit

q := lim (H − k2 − iε )−1
G
ε →0+

2p 2p
exists in the uniform operator topology from L σp+1 (Rn ) to L−p−1
σ (R ) with the norm
n
2 2
estimate
q f  2p ≤ C|k|−γ  f  2p
G
p−1 p+1
L−σ /2 Lσ /2

for large k with p and σ as above and with γ = 2 − n2 for n

2 <p≤ n+1
2 and γ =
2p for 2 < p ≤ ∞.
1 − n−1 n+1

 with kernel
Proof. Let us prove first that the integral operator K
1
K(x, y) := |q| 2 (x)G+
k (|x − y|)q 1 (y), 2

1
where q 1 (y) = |q(y)| 2 sgn q(y) maps from L2 (Rn ) to L2 (Rn ) with the same norm
2
1
estimate as in Theorem 23.5. Indeed, if f ∈ L2 (Rn ) and q ∈ Lσp (Rn ), then |q| 2 ∈
2p
1
L2p
σ (R ), and therefore, f |q| 2 ∈ L
n p+1
σ (Rn ). Applying Theorem 23.1, we obtain
2 2

k (|q| 2 f )
G
1
2p ≤ C|k|−γ  f  2p ,
p−1 p+1
L−σ /2 Lσ /2

k (q 1 f ) ∈
where γ is as in Theorem 23.5. Then by Hölder’s inequality we have |q| 2 G
1

2
L2 (Rn ) as asserted.
q . This operator satisfies the resolvent equa-
Let us consider now the operator G
tion
q = G
G k − Gk qG
q ,

which follows easily from (H − k2 )G q = I. We denote by G l and Gr the integral op-
+ 1 +
erators having kernels Gk (|x − y|)q 1 (y) and |q(x)| 2 Gk (|x − y|), respectively. Then
2
one can show that
Gq = Gk − G l (1 + K)
 −1 Gr

2p 2p
r : L σp+1 → L2 , and G
 : L2 → L2 , G
for large k. Since K l : L2 → L p−1 , Theorem 23.5
2 − σ2
is proved. 

The fundamental solution of the Helmholtz operator that was considered in the pre-
vious chapter can be effectively used for the following scattering problem: find
23 Estimates for the Laplacian and Hamiltonian 225

u ∈ Hloc
2 (Rn ), n ≥ 2 that satisfies

− Δ u + qu = k2 u, x ∈ Rn , k>0 (23.2)
u = u0 + usc , u0 = eik(x,θ ) , θ ∈ Sn−1
 
∂ usc
lim r(n−1)/2 − ikusc = 0, r = |x|.
r→∞ ∂r

The latter condition is called the Sommerfeld radiation condition at infinity. The
problem (23.2) is called the scattering problem.

Theorem 23.6. Assume that q ∈ L1 (Rn ) ∩ Lσp (Rn ), n/2 < p ≤ ∞, σ > max{0, 1 −
(n + 1)/(2p)}, is real-valued. Then there exists a unique solution u of (23.2) such
that usc ∈ L∞ (Rn ), and this solution u necessarily satisfies the Lippmann–Schwinger
equation 
u = u0 − G+
k (|x − y|)q(y)u(y)dy. (23.3)
Rn

Proof. Let us show first that there is a constant C > 0 such that

lim |usc (y)|2 dσ (y) ≤ C. (23.4)
R→∞ |y|=R

Indeed, the Sommerfeld radiation condition at infinity and Green’s identity imply
that
 
2ik |usc (y)|2 dσ (y) = [usc (y)ikusc (y) − usc (y)(−ikusc (y))]dσ (y)
|y|=R |y|=R

∂ ∂
= [usc (y) usc (y) − usc (y) usc (y)]dσ (y)
|y|=R ∂r ∂r

+ o(1/R(n−1)/2 ) (usc (y) + usc (y))dσ (y)
|y|=R

= [usc (y)Δ usc (y) − usc (y)Δ usc (y)]dy
|y|≤R
 1/2
+ o(1) |usc (y)| dσ (y)
2
|y|=R

= q(y)[usc (y)u0 (y) − usc (y)u0 (y)]dy
|y|≤R
 1/2
+ o(1) |usc (y)|2 dσ (y) , R → ∞.
|y|=R

This equality leads to the inequality

226 Part II: Fourier Transform and Distributions

2k |usc (y)|2 dσ (y)
|y|=R
 1/2
≤ 2 qL1 (Rn ) usc L∞ (Rn ) + o(1) |usc (y)|2 dσ (y) , R → ∞.
|y|=R

This inequality clearly implies (23.4). The next observation is that G+ k clearly satis-
fies the Sommerfeld radiation condition at infinity. Fixing now x ∈ Rn and R > 0 suf-
ficiently large that x ∈ BR = {y : |y| < R} and applying Green’s identity to usc (y) and
G+ ∂
k (|x − y|), we obtain (using the fact that on the sphere |y| = r we have ∂νy = ∂ r )


∂ + ∂
[usc (y) Gk (|x − y|) − G+ (|x − y|) usc (y)]dσ (y)
|y|=R ∂r k ∂r

= [usc (y)(Δy + k2 )G+ +
k (|x − y|) − Gk (|x − y|)(Δ + k )usc (y)]dy.
2
|y|≤R

The usual procedure of allocation of the singularity of G+

k allows us to obtain

usc (x) = − G+k (|x − y|)q(y)u(y)dy
|y|≤R
      
∂ ∂
− usc (y) − ik G+ (|x − y|) − G+
(|x − y|) − ik usc (y) dσ (y).
|y|=R ∂r k k ∂r

The integral over the sphere |y| = R can be estimated from above by

 1/2    2 1/2
 ∂ 
|usc (y)| dσ (y)  + 
 ∂ r − ik Gk (|x − y|) dσ (y)
2
|y|=R |y|=R
 1/2    2 1/2
 ∂ 
|G+  
+ k (|x − y|)| dσ (y)  ∂ r − ik usc (y) dσ (y) .
2
|y|=R |y|=R

Since for fixed x and R → ∞ we have |G+ k (|x − y|)| ≤ C/R

(n−1)/2 and since G+ and
k
usc both satisfy the radiation condition, we may estimate the latter sum from above
using (23.4) as
 1/2
(n−1)/2
Co(1/R ) dσ (y)
|y|=R
 1/2  1/2
C
+ dσ (y) o(1/R(n−1)/2 ) dσ (y) .
R(n−1)/2 |y|=R |y|=R

But this sum tends to zero as R → ∞. Thus we have

23 Estimates for the Laplacian and Hamiltonian 227

usc (x) = − G+
k (|x − y|)q(y)u(y)dy.
Rn

So u from (23.2) necessarily satisfies the Lippmann–Schwinger equation (23.3). For

k > 0 sufficiently large we can prove the unique solvability of (23.3) (and (23.2) as
well) as follows. Theorem 23.5 allows us to rewrite equation (23.3) in the form


v = v0 − Kv, (23.5)

 is as in Theorem 23.5. Since the conditions

where v = |q|1/2 u, v0 = |q|1/2 u0 , and K
on q and u imply that v, v0 ∈ L (R ) and
2 n

 
  C0
K  ≤ ,
L2 (Rn )→L2 (Rn ) kγ

where C0 as in Theorem 23.5 and γ = 2 − n/2 for n/2 < p ≤ (n + 1)/2 and γ =
1/γ
1 − (n − 1)/(2p) for (n + 1)/2 < p ≤ ∞, we obtain for k > C0 that there is a
unique solution v of (23.5), namely
∞
v= ∑ K j v0 .
j=0

Moreover, the estimate

2C0
v − v0 L2 (Rn ) ≤ v0 L2 (Rn )
kγ

holds uniformly in k ≥ (2C0 )1/γ . This is equivalent to the estimate

 
 1/2  2C0 1/2
|q| usc  ≤ qL1 (Rn ) . (23.6)
L2 (Rn ) kγ

For the values of k from the interval 0 < k ≤ (C0 )1/γ we proceed as follows.
Exercise 23.4. Show that the integral operator G k ◦ q for all k > 0 is a compact
operator in L∞ (Rn ), where q satisfies the conditions of Theorem 23.6.
This exercise implies that the integral operators K  and G q ◦ q are also compact
in L2 (Rn ) and L∞ (Rn ), respectively. Next, using Agmon’s estimate and Theorems
q = (−Δ − k2 + q − i0)−1
23.1 and 23.5, we conclude that for all k > 0 the operator G
exists in the appropriate operator topology (see Theorem 23.5), and therefore for the
solution u of (23.2) (or equivalently, (23.3)) the representation

q ◦ q)u0
u = (I − G (23.7)

holds. The Lippmann–Schwinger equation can be rewritten in the operator form

228 Part II: Fourier Transform and Distributions

usc − Tusc = u0 , (23.8)

where u0 = −(G k ◦ q)u0 ∈ L∞ (Rn ) and T is a compact operator in L∞ (Rn ). By Riesz
theory (see Chapter 34) we shall obtain the unique solvability of (23.8) if we are able
to show that I − T is injective. But injectivity follows immediately from (23.7). The
theorem is therefore completely proved. 

Remark 23.7. For k > 0 large enough, the unique solvability in Theorem 23.6 holds
for a complex-valued potential q.
Corollary 23.8. Let v be the outgoing solution of the inhomogeneous Schrödinger
equation
(H − k2 )v = f ,

i.e.,
v = (H − k2 − i0)−1 f ,

where f ∈ S(Rn ). Then the following representation holds:

k ( f − qG
v(x) = G q ( f ))(x).

Moreover, for |x| → ∞ and fixed positive k,

n−3
 
eik|x| k 2
 1
v(x, k) = Cn n−1 A f (k, θ ) + o n−1 ,
|x| 2 |x| 2

where θ  = x
|x| and the function A f , called the scattering amplitude, is defined by


A f (k, θ  ) := q ( f ))dy.
e−ik(θ ,y) ( f (y) − q(y)G
Rn

Proof. The first representation follows immediately from the definition of G q .

   
Indeed, since v = Gq f , we must have Gk f = v + Gk qv, or v = Gk f − Gk qv =
Gk ( f − qG
q f ).
In order to prove the asymptotic behavior for v let us assume that q and f have
compact support, say in the ball {x : |x| ≤ R}. We will use the following asymptotic
behavior of G+ k (|x|):
(1) k|x| < 1 :
(a) G+
k (|x|) ∼ C|x|
2−n , n ≥ 3,
+
(b) Gk (|x|) ∼ C log(k|x|), n = 2.
(2) k|x| > 1:
n−3
k 2
G+
k (|x|) ∼ C n−1 eik|x| , n ≥ 2.
|x| 2
23 Estimates for the Laplacian and Hamiltonian 229

Since k is fixed, |y| ≤ R, and |x| → +∞, we may assume that k|x − y| > 1 for x large
enough. Therefore, as |x| → ∞, we have
n−3 
eik|x| k 2
q f )dy
v(x) = C n−1 eik(|x−y|−|x|) ( f − qG
|x| 2 |y|≤R

 
1 q f )dy =: I1 + I2 .
+ o n−1 ( f − qG
|y|≤R |x − y| 2

It is clear that for I2 the following is true:

 
  
1  1
I2 = o n−1 ( f (y) − q(y)Gq f (y))dy = o n−1 , |x| → ∞,
|x| 2 |y|≤R |x| 2

q f is an integrable function. Next, let us note that

because f − qG
   
|x − y|2 − |x|2 y2 − 2(x, y) x 1
|x − y| − |x| = = =− ,y +O
|x − y| + |x| |x − y| + |x| |x| |x|

as |x| → +∞. We can therefore rewrite the integral appearing in I1 as follows:

    
x ,y +O 1
−ik |x|
e |x| ( f q f )dy
− qG
|y|≤R
  

q f )dy + O 1 
q f )dy
= e−ik(θ ,y) ( f − qG e−ik(θ ,y) ( f − qG
|y|≤R |x| |y|≤R
  
=

q f )dy + O 1 , |x| → ∞,
e−ik(θ ,y) ( f − qG
|y|≤R |x|

where θ  = |x|
x
∈ Sn−1 . Thus, Corollary 23.8 is proved when q and f have com-
pact support. The proof in the general case is much more difficult and is therefore
omitted. 

Remark 23.9. Hint for the general case: The integral over Rn might be divided into
two parts: |y| < |x|ε and |y| > |x|ε , where ε > 0 is chosen appropriately.
Lemma 23.10 (Optical lemma). For the function A f (k, θ  ) the following equality
holds:  
1
|A f (k, θ  )|2 dθ  = − 2 n−2 Im( f v)dx,
Sn−1 C k Rn

where C is the constant from the asymptotic representation of v = (H − k2 − i0)−1 f .

Proof Let ρ be a smooth real-valued function on [0, +∞) such that  0 ≤ ρ ≤ 1 and
ρ (r) = 1 for 0 ≤ r < 1 and ρ (r) = 0 for r ≥ 2. We set ρm (r) = ρ mr . Multiplying
f by vρm (|x|), integrating over Rn , and taking the imaginary parts leads to
230 Part II: Fourier Transform and Distributions
 
Im f (x)ρm (|x|)v(x)dx = Im (−Δ v)ρm (|x|)v(x)dx.
Rn Rn

As m tends to infinity, the left-hand side converges to Im Rn f (x)v(x)dx. To get the
desired limit for the right-hand side, we integrate by parts and obtain
 
x
Im (−Δ v)ρm (|x|)v(x)dx = Im · ∇vρm (|x|)v(x)dx
Rn Rn |x|
   
= Im (θ · ∇v − ikv)v(x)ρm (|x|) + ikρm (|x|)|v|2 dx
Rn
 
= Im (θ · ∇v − ikv)v(x)ρm (x)dx + k

ρm (|x|)|v(x)|2 dx =: I1 + I2 .
Rn Rn

Since v = (H − k2 − i0)−1 f , using the asymptotic representation we may conclude

that v satisfies the Sommerfeld radiation condition
 
∂v 1
− ikv = o n−1 , r = |x|,
∂r r 2
at infinity. Hence I1 → 0 as m → ∞. By Corollary 23.8, the second term I2 is equal
to
 
kn−3
k ρm (|x|)|v(x)|2 dx = k ρm (|x|) ·C2 n−1 |A f (k, θ  )|2 dx
Rn Rn |x|
  
1
+k ρm (|x|)o dx
Rn |x|n−1
 
2m r n−1
= C2 kn−2 |A f (k, θ  )|2 dθ  ρ  (r)dr
n−1 m
Sn−1 m r
  2m  
1
+k dθ rn−1 o n−1 ρm (r)dr
Sn−1 m r
  2m
= C2 kn−2 |A f (k, θ  )|2 dθ  ρm (r)dr + o(1)
Sn−1 m

= C2 kn−2 |A f (k, θ  )|2 dθ  [ρ (2) − ρ (1)] + o(1)
Sn−1

= −C2 kn−2 |A f (k, θ  )|2 dθ  + o(1), m → ∞.
Sn−1

Letting m → ∞, we obtain
 
Im f (x)v(x)dx = −C2 kn−2 |A f (k, θ  )|2 dθ  .
Rn Sn−1

Thus, Lemma 23.10 is proved. 


23 Estimates for the Laplacian and Hamiltonian 231

Exercise 23.5. Let n = 2 or n = 3. Assume that q ∈ L p (Rn )∩L1 (Rn ) with 1 < p ≤ ∞
if n = 2 and 3 < p ≤ ∞ if n = 3. Prove that the generalized eigenfunctions u(x,k),
that is, the solutions of the problem (23.2) with (k,k) = k2 , are uniformly bounded
with respect to x ∈ Rn and |k| sufficiently large.

We will obtain very important corollaries from the optical lemma. Let Aq (k) denote
the linear mapping that takes the inhomogeneity f to the corresponding scattering
amplitude
Aq (k) : f (x) → A f (k, θ  ).

Lemma 23.11. Let the potential q(x) satisfy the conditions from Theorem 23.5.
2p

/2 (R ) to L (S
Then Aq is a well defined bounded operator from Lσp+1 n 2 n−1 ) with the

operator norm estimate

C
Aq  2p ≤ γ n−2 ,
p+1
Lσ /2 →L2 |k| 2+ 2

where p, σ , and γ are as in Theorem 23.5.

Proof. By Lemma 23.10 and the definition of Aq f we have that

 
1
Aq f 2L2 (Sn−1 ) = |A f (k, θ  )|2 dθ  = − Im( f · v)dx
Sn−1 C2 |k|n−2 Rn
1
≤ v 2p f 2p .
C2 |k|n−2 p−1
L−σ /2 (Rn )
p+1
Lσ /2 (Rn )

q f , we obtain from Theorem 23.5 that

Further, since v = G

C
Aq f 2L2 (Sn−1 ) ≤ · |k|−γ  f 2 2p .
|k|n−2 p+1
Lσ /2 (Rn )

Thus, Lemma 23.11 is proved. 



Let us denote by A0 (k) the operator Aq (k) that corresponds to the potential q ≡ 0,
i.e., 

A0 f (θ  ) = e−ik(θ ,y) f (y)dy.
Rn

It is not difficult to see that


Aq f (θ  ) = A f (k, θ  ) = f (y)u(y, k, θ  )dy,
Rn
232 Part II: Fourier Transform and Distributions

where u(·, k, θ  ) is the solution of the Lippmann–Schwinger equation. Indeed, by

Corollary 23.8 we have


A f (k, θ  ) = e−ik(θ ,y) ( f (y) − q(y)G q ) f , eik(θ  ,y) ) 2 n
q ( f ))dy = ((I − qG
L (R )
Rn


q (q))eik(θ ,·) ) 2 n =
= ( f , (I − G q (q))(eik(θ ,·) )(y)dy
f (y)(I − G

L (R )
Rn

= f (y)u(y, k, θ  )dy,
Rn

q is a self-adjoint operator.
since G
Let us prove now that

q (q))(eik(θ  ,·) )(y)

u(y, k, θ  ) := (I − G

is the solution of the Lippmann–Schwinger equation. Indeed,

 
q (q) · (eik(θ ,·) )(y)
(H − k2 )u = (H − k2 )(eik(θ ,y) ) − (H − k2 )G
  
= (−Δ − k2 )eik(θ ,y) + qeik(θ ,y) − qeik(θ ,y) = 0,

q = I. This means that this u(y, k, θ  ) is
since (−Δ − k2 )eik(θ ,y) = 0 and (H − k2 )G
the solution of the equation (H − k )u = 0.
2

Remark 23.12. Let us consider the Lippmann–Schwinger equation


u(x, k, θ ) = eik(x,θ ) − G+
k (|x − y|)q(y)u(y, k, θ )dy.
Rn

Then for fixed k > 0 and |x| → ∞, the solution u(x, k, θ ) admits the asymptotic
representation
n−3
 
ik(x,θ ) eik|x| k 2
 1
u(x, k, θ ) = e +Cn n−1 A(k, θ , θ ) + o n−1 ,
|x| 2 |x| 2

where θ  = x
|x| and the function A(k, θ  , θ ) is called the scattering amplitude and has
the form 

A(k, θ  , θ ) = e−ik(θ ,y) q(y)u(y, k, θ )dy.
Rn

For k < 0 we set

A(k, θ  , θ ) = A(−k, θ  , θ ), u(x, k, θ ) = u(x, −k, θ ).

Proof. If (H − k2 )u = 0 and u = eik(θ ,x) + usc (x, k, θ ), then usc (x, k, θ ) satisfies the
equation
23 Estimates for the Laplacian and Hamiltonian 233

(H − k2 )usc = −qeik(θ ,x) .

We may therefore apply Corollary 23.8 with v := usc and f := −qeik(θ ,x) to obtain
n−3
 
eik|x| k 2
 1
usc (x, k, θ ) = Cn n−1 A f (k, θ ) + o n−1 ,
|x| 2 |x| 2

where


A f (k, θ  ) = q (qeik(θ ,·) )(y))dy
e−ik(θ ,y) (−qeik(θ ,y) + qG
Rn


=− q (qeik(θ ,·) )dy.
e−ik(θ ,y) q(y)(eik(θ ,y) − G
Rn

q (qeik(θ ,·) )(y) is a solution of the equation (H −

But we have proved that eik(θ ,y) − G
k )u = 0. We conclude that
2



A f (k, θ  ) = − e−ik(θ ,y) q(y)u(y, k, θ )dy =: −A(k, θ  , θ ).
Rn

This proves the remark. 



Now let Φ0 (k) and Φ (k) be the operators defined for f ∈ L2 (Sn−1 ) as

eik(x,θ ) f (θ )dθ
1
(Φ0 (k) f )(x) := |q(x)| 2 (23.9)
Sn−1

and 
1
(Φ (k) f )(x) := |q(x)| 2 u(x, k, θ ) f (θ )dθ . (23.10)
Sn−1

Lemma 23.13. The operators Φ0 (k) and Φ (k) are bounded from L2 (Sn−1 ) to
L2 (Rn ) with the norm estimates

C
Φ0 (k), Φ (k) ≤ γ n−2 , k > 0,
k 2+ 2

where γ is as in Theorem 23.5.

Proof. Let us prove that

1
(Φ0 (k) f )(x) = |q(x)| 2 (A∗0 f )(x) (23.11)

and 1
(Φ (k) f )(x) = |q(x)| 2 (A∗q f )(x), (23.12)
234 Part II: Fourier Transform and Distributions

where A∗0 and A∗q are the adjoint operators for A0 and Aq , respectively. Indeed, if
f ∈ L2 (Sn−1 ) and g ∈ L2 (Rn ), then
  
f (θ )(A0 g)(θ )dθ = f (θ )dθ eik(θ ,y) g(y)dy
Sn−1 Sn−1 Rn
 
= g(y)dy eik(θ ,y) f (θ )dθ
Rn Sn−1
  
= eik(θ ,y) f (θ )dθ g(y)dy.
Rn Sn−1

This means that 

A∗0 f (y) = eik(θ ,y) f (θ )dθ ,
Sn−1

and (23.11) is immediate. Similarly one proves (23.12). Since (see Lemma 23.11)

C
A0 , Aq  2p ≤ γ n−2 ,
p+1
Lσ /2 →L2 (Sn−1 ) k 2+ 2

we have that
C
A∗0 , A∗q  2p ≤ γ n−2 .
p−1
L2 (Sn−1 )→L−σ /2 (Rn ) k 2+ 2

The proof is finished by

1 1
Φ0 (k) f L2 (Rn ) = |q| 2 (A∗0 f )L2 (Rn ) ≤ qL2 p (Rn ) A∗0 f  2p
σ p−1
L−σ /2 (Rn )

C 1
≤ γ n−2 qL2 p (Rn )  f L2 (Sn−1 ) ,
k 2+ 2 σ

where we have made use of Hölder’s inequality in the first estimate. It is clear that
the same is true for Φ (k). 


Agmon’s estimate (23.1) can be applied to the magnetic Schrödinger operator. In

fact, in the work [2], Agmon proved a more general estimate than (23.1). Namely, it
was proved that for all g ∈ H−2 δ (Rn ) and |k| ≥ 1,

1  
gH 2 (Rn ) + gH 1 (Rn ) + |k| gL2 (Rn ) ≤ C (Δ + k2 )gL2 (Rn ) ,
|k| −δ −δ −δ δ

where δ > 1/2 and H−s δ (Rn ), s = 0, 1, 2, denotes the weighted Sobolev space (see
below for a precise definition). As a consequence of this estimate, for all f ∈ Lδ2 (Rn ),
δ > 1/2, one has the estimates
23 Estimates for the Laplacian and Hamiltonian 235

 
(−Δ − k2 − i0)−1 f  2 n ≤ β  f  2 n
L−δ (R ) Lδ (R )
|k| (23.13)
 
(−Δ − k2 − i0)−1 f  1 n ≤ β  f  2 n .
H (R ) L (R ) δ
−δ

Here (−Δ − k2 − i0)−1 is the integral operator with kernel G+ k (|x − y|), see (22.1),
and the weighted Sobolev spaces Wp,1 σ (Rn ) (or Hσ1 (Rn ) if p = 2) are understood so
that f belongs to Wp,1 σ (Rn ) if and only if f and ∇ f belong to the weighted Lebesgue
space Lσp (Rn ) (see Example 18.17).
Since the integral operator (−Δ − k2 − i0)−1 is of convolution type, using duality
we can conclude that it maps Hδ−1 (Rn ) to L− 2 (Rn ) with the norm estimate
δ
 
(−Δ − k2 − i0)−1 f  2 ≤ β  f H −1 (Rn ) , |k| ≥ 1, (23.14)
L −δ
(Rn ) δ

where Hδ−1 (Rn ) denotes the dual space of the Sobolev space H−1 δ (Rn ) and the con-
stant β is the same as in (23.13).
We will consider now the scattering problem for the magnetic Schrödinger oper-
ator in Rn , n ≥ 2, of the form

 (x))2 · +V (x)·,
Hm := −(∇ + iW (23.15)

 (x) and V (x) are assumed to be real and are from the spaces
where the coefficients W

 ∈ Wp,1 σ (Rn ),
W V ∈ Lσp (Rn ), n < p ≤ ∞, σ > n/p ,
1/p + 1/p = 1.
(23.16)
We are looking for the solutions to the equation Hm u = k2 u, k = 0, with Hm from
(23.15) in the form

u(x) = u0 (x) + usc (x), u0 (x) = eik(x,θ ) , θ ∈ Sn−1 ,
∂ usc (x) (23.17)
limr→∞ r(n−1)/2 ∂r − ikusc (x) = 0, r = |x|.

Using the same procedure as for the Schrödinger operator (see Theorem 23.6), we
conclude that the solution (23.17) necessarily satisfies the Lippmann–Schwinger
integral equation

u(x) = u0 (x) + G+ 
k (|x − y|)(2i∇(W (y)u(y)) − q(y)u(y))dy,
Rn

 + |W
where q = i∇W  |2 +V . This equation can be rewritten as the following integral
equation:
usc (x) = u0 (x) + Lk (usc )(x), u0 (x) = Lk (u0 )(x), (23.18)

with the integral operator Lk defined as

236 Part II: Fourier Transform and Distributions

Lk f (x) = G+ 
k (|x − y|)(2i∇(W (y) f (y)) − q(y) f (y))dy. (23.19)
Rn

Lemma 23.14. Suppose that the conditions (23.16) are fulfilled. Then u0 belongs
σ /2 (R ), and Lk from (23.19) maps L−σ /2 (R ) into itself with σ as (23.16).
2
to L− n 2 n

Moreover, for |k| ≥ 1 uniformly,

 
 
 
u0 L2 (Rn ) ≤β 2 W  + qL2 (Rn ) ,
−σ /2 Lσ2 /2 (Rn ) σ /2
    (23.20)
 
Lk f L2 (Rn ) ≤ β 2 W  +Cp qLσp (Rn )  f L2 (Rn ) ,
−σ /2 Lσ∞ (Rn ) −σ /2

where β is the same as in (23.13) and the constant Cp is equal to

 1/p
1 Γ ((p − n)/2)
Cp = √ n .
(2 π ) Γ (p/2)

Proof. Conditions (23.16) imply that σ /2 > 1/2 and

Lσp (Rn ) → Lσ2 /2 (Rn ).

 and |W
It is therefore true that under these conditions the functions V, ∇W  |2 belong
to Lσ /2 (R ) and W
2 n  ∈ Lσ (R ). Using the first Agmon’s estimate (23.13), one can
∞ n

easily obtain
 
β  
 
u0 L2 (Rn ) ≤ 2|k| W  + qL2 (Rn ) .
−σ /2 |k| Lσ2 /2 (Rn ) σ /2

Hence the first inequality in (23.20) is proved. Next, applying now (23.14), we ob-
tain that
 
 
  
Lk f L2 (Rn ) ≤ β 2 ∇(W f ) −1 n + q f H −1 (Rn )
−σ /2 Hσ /2 (R ) σ /2
 
 
 
≤β 2 W f + q f H −1 (Rn )
Lσ2 /2 (Rn ) σ /2
   
 
≤ β 2 W   f L2 (Rn ) + q f  H (Rn ) .
−1
Lσ∞ (Rn ) −σ /2 σ /2

To estimate the second term q f H −1 (Rn ) we proceed using Hölder’s inequality and
σ /2
the Hausdorff–Young inequalities as follows:
23 Estimates for the Laplacian and Hamiltonian 237
     
   )  )
q f H −1 (Rn ) = q f = F (q f  ≤ C0 F (
q f  2p/(p−2) n
σ /2 H −1 (Rn ) 2 (Rn )
L−1 L (R )
   
   
≤ C0 (2π )−n/p q f 2p/(p+2) n ≤ C0 (2π )−n/p qL p (Rn )  f 2 ,
L (R ) L (Rn )

where p > n, q(x) = (1 + |x|2 )σ /2 q(x), f(x) = (1 + |x|2 )−σ /4 f (x), and C0 is equal to
 1/p  √ n ∞ 1/p
dx ( π) (n−2)/2 −p/2
C0 = = r (1 + r) dr .
Rn (1 + |x|2 ) p/2 Γ (n/2) 0

Combining this constant C0 with the latter inequality, we obtain Cp from this Lemma
and (23.20). 

We denote by α and γ the following constants:
   
   
α = 2 W  +Cp qLσp (Rn ) , γ = 2 W  + qL2 (Rn ) . (23.21)
Lσ∞ (Rn ) Lσ2 /2 (Rn ) σ /2

Theorem 23.15. Assume that the conditions (23.16) are satisfied and assume that
β α < 1 with β and α from (23.13) and (23.21), respectively. Then the integral
σ /2 (R ), and uniformly
2
equation (23.18) has a unique solution usc from the space L− n

in k, |k| ≥ 1, the following estimate holds:

βγ
usc L2 (Rn ) ≤ . (23.22)
−σ /2 1−βα

σ /2 (R ) and
2
Proof. Lemma 23.14 says that Lk maps in L− n

Lk L2 (Rn )→L−

2 (Rn ) ≤ β α < 1.
−σ /2 σ /2

Since u0 belongs to L− σ /2 (R ) with the norm estimate β γ , the integral equation
2 n

σ /2 (R ) that can be obtained by the itera-

(23.18) has a unique solution usc from L−2 n

tions
∞
usc = (I − Lk )−1 (u0 ) = ∑ Lkj+1 (u0 ).
j=0

The estimate (23.22) follows now from Lemma 23.14 and from the latter represen-
tation for usc . 

Corollary 23.16. If the constant α from (23.21) is small enough, then for fixed k,
|k| ≥ 1, usc (x, k, θ ) belongs to L∞ (Rn ) in x ∈ Rn and uniformly in θ ∈ Sn−1 .
Proof. For α small enough, u0 ∈ L∞ (Rn ) and Lk maps in L∞ (Rn ) with the norm
estimate
Lk L∞ (Rn )→L∞ (Rn ) ≤ c(k)α . (23.23)
238 Part II: Fourier Transform and Distributions

These two facts yield the proof. 



Lemma 23.17. Under the assumptions of Theorem 23.15, for fixed k ≥ 1 and for
f ∈ L∞ (Rn ) the following asymptotic representation holds:
  
eik|x| k(n−3)/2 
 + q) f (y)dy + o 1
Lk f (x) = C e−ik(θ ,y) (2kθ W ,
|x|(n−1)/2 Rn |x|(n−1)/2
(23.24)
as |x| → ∞, where θ  = x/|x| and
π
1 e−i 4 (n+1)
C= .
2 (2π )(n−1)/2


Proof. In this proof we assume (for simplicity) that n ≥ 3. Since f ∈ L∞ (Rn ) and W
vanishes at infinity, integration by parts leads to
 
Lk f (x) = −2i ∇y G+  G+
k (|x − y|)W (y) f (y)dy − k (|x − y|)q(y) f (y)dy.
Rn Rn

In view of this, one must study the behavior as |x| → ∞ of the functions
 (n−2)/2
i k (1)
G+
k (|x − y|) = H(n−2)/2 (k|x − y|)
4 2π |x − y|

and  (n−2)/2
x−y i k (1)
∇y G+
k (|x − y|) = k Hn/2 (k|x − y|),
|x − y| 4 2π |x − y|

(1)
where Hν denotes the Hankel function of the first kind of order ν . The behavior
of the latter integrals can be studied by dividing them into two cases: |y| ≤ |x|a and
|y| > |x|a , where a > 0 is a parameter that we can adjust to our liking. In the first
case we have for a < 1/2 that

|x − y| = |x| − (θ  , y) + O(|x|2a−1 )

and (as a consequence of it) k|x − y| → ∞ for |x| → ∞. Thus, we use the behavior of
(1)
Hν for large argument (see [23])
 
(1) eiz 1
H(n−2)/2 (z) = Cn √ ,+O
z z3/2
  (23.25)
(1) eiz 1
Hn/2 (z) = −iCn √ + O 3/2 ,
z z
23 Estimates for the Laplacian and Hamiltonian 239

2 −i π4 (n−1)
as |z| → ∞, where Cn = πe , n ≥ 2. Hence we obtain in this case that

 
iCn eik|x−y| 1
G+
k (|x − y|) = k (n−3)/2
+ O
4(2π )(n−2)/2 |x − y|(n−1)/2 |x|(n+1)/2
ik|x−y|
 
+ Cn  e (n−3)/2 1
∇y Gk (|x − y|) = θk k +O .
4(2π )(n−2)/2 |x − y|(n−1)/2 |x|(n+1)/2

Since for |y| ≤ |x|a we have in addition that

x−y x
|x − y|−(n−1)/2 = |x|−(n−1)/2 + O(|x|−(n−1)/2+a−1 ), = + O(|x|a−1 ),
|x − y| |x|

and  
ik|x| −ik(θ  ,y) 1
eik|x−y|
=e e +O ,
|x|1−2a

it follows that the first part (|y| ≤ |x|a ) of Lk f (x) is equal to

  
n eik|x| −ik(θ  ,y)  1
C e (2kθ W (y) + q(y)) f (y)dy + O ,
|x|(n−1)/2 |y|≤|x|a |x|(n−1)/2+1−2a

where Cn = − iCn

. We have used here the fact that the conditions (23.16)
4(2π )(n−2)/2
 and V belong to L1 (Rn ). This means that Lk f is of the desired
guarantee that W
form as |x| → ∞ in this case.
Turning now to Lk f , where |y| > |x|a , we have two more possibilities: |x|a < |y| ≤
|x|/2 and |y| > |x|/2. In the first case we have that |x − y| ≥ |x|/2. Using again the
asymptotic (23.25) we may estimate Lk f from above by
  
C  (y)| + |V (y)|)dy  f L∞ (Rn ) = o 1
(|W ,
|x|(n−1)/2 |x|<|y|≤|x|/2 |x|(n−1)/2

 and V belong to L1 (Rn ). For the case |y| > |x|/2 we have two subcases:
since W
(1)
k|x − y| < 1 and k|x − y| > 1. For the first subcase we use the behavior of Hν for
small argument, see [23],

(1)
Hν (z) = cν z−ν + o(z−ν ), z → 0+,

and obtain that

G+
k (|x − y|) = c(k)|x − y|
2−n
+ o(|x − y|2−n ),
∇y G+
k (|x − y|) = c(k)|x − y|
1−n
+ o(|x − y|1−n ).
240 Part II: Fourier Transform and Distributions

Hence, Lk f (x) in this subcase can be estimated from above by

  (y)|| f (y)|dy 
|W |V (y)|| f (y)|dy
C +C
k|x−y|<1,|y|>|x|/2 |x − y| n−1
k|x−y|<1,|y|>|x|/2 |x − y|n−2
  
  dy
≤ C  f L∞ (Rn ) W  ∞ n
Lσ (R ) |y|>|x|/2 |x − y|n−1 |y|σ
 1/p 
dy
+ V Lσp (Rn )  
|y|>|x|/2 |x − y|(n−2)p |y|σ p
⎛ 
 ⎞
W  ∞ n  
⎜ Lσ (R ) V Lσp (Rn ) ⎟ 1
≤ C  f L∞ (Rn ) ⎝ n−1+σ −n + n−2+σ −n/p ⎠ = o ,
|x| |x| |x|(n−1)/2

since σ > n/p , n < p ≤ ∞, and n ≥ 3. We have used here the estimates for the
convolution of the weak singularities (see, for example, Lemma 34.3).
For the second subcase we can use (23.25) and estimate this part of Lk f (x) from
above by
  | + |V |)(1 + |y|)σ
(|W
C  f L∞ (Rn ) dy
k|x−y|>1,|y|>|x|/2 |x − y|(n−1)/2 (1 + |y|)σ
    1/p
  dy
≤ C  f L∞ (Rn ) W  p n + V Lσp (Rn )  
Lσ (R ) |y|>|x|/2 |x − y|(n−1)p /2 |y|σ p
 
1
=o
|x|(n−1)/2

as |x| → ∞, using the estimates for convolution of weak singularities and conditions
(23.16). 


Since usc (x, k, θ ) for fixed k ≥ 1 is an L∞ -function in x, Lemma 23.17 yields the
asymptotic representation for u as
π  
ik(x,θ ) 1 e−i 4 (n+1) eik|x| k(n−3)/2  1
u(x, k, θ ) = e + A(k, θ , θ ) + o
2 (2π )(n−1)/2 |x|(n−1)/2 |x|(n−1)/2

as |x| → ∞, where the function A(k, θ  , θ ) is called the scattering amplitude for the
magnetic Schrödinger operator and it is defined as


A(k, θ  , θ ) =  (y) + q(y))u(y, k, θ )dy.
e−ik(θ ,y) (2kθ W (23.26)
Rn

Substituting u = u0 + usc into (23.26) implies

23 Estimates for the Laplacian and Hamiltonian 241


A(k, θ  , θ ) =  (y) + q(y))dy
e−ik(θ ,y)+ik(θ ,y) (2kθ W
Rn


+  (y) + q(y))usc (y, k, θ )dy
e−ik(θ ,y) (2kθ W
Rn
=: AB (k, θ  , θ ) + R(k, θ  , θ ). (23.27)

The function AB (k, θ  , θ ) is called the direct Born approximation. It can be easily
checked that

 )(k(θ − θ  )) + F (q)(k(θ − θ  ))
AB (k, θ  , θ ) = 2kθ  F (W
 )(k(θ − θ  )) + F (|W
= k(θ + θ  )F (W  |2 +V )(k(θ − θ  )), (23.28)

where F denotes the usual n-dimensional Fourier transform.

The direct Born approximation allows us to obtain the approximation uB (x, k, θ )
for the solution u(x, k, θ ) of the equation Hm u = k2 u as
π
1 e−i 4 (n+1) eik|x| k(n−3)/2
uB (x, k, θ ) = eik(x,θ ) + AB (k, θ  , θ ) (23.29)
2 (2π )(n−1)/2 |x|(n−1)/2

and secondly, to prove the following very practical statement.

 |2 +V and W
Proposition 23.18. The Fourier transforms of |W  can be evaluated as

F (|W  |2 +V )(ξ ) = 1 (AB (k, θ  , θ ) + AB (k, −θ , −θ  ))

2
 1
 ))(ξ ) = (AB (k, θ  , θ ) − AB (k, −θ , −θ  )),
4k2 − ξ 2 (ξ̂⊥ , F (W
2

where ξ = 0, ξ̂⊥ is any unit vector that is orthogonal to ξ , and k, θ  , θ are defined
by
ξ ξ̂⊥  2 ξ ξ̂⊥  2
θ= + 4k − ξ 2 , θ  = − + 4k − ξ 2
2k 2k 2k 2k

so that ξ = k(θ − θ  ) and k2 ≤ ξ 2 /4.

Proof. The result follows straightforwardly from (23.28). 



All these results, in particular the direct Born approximation, are valid also for the
Schrödinger operator (W  = 0) as well as the approximation for the backscattering
amplitude (see results below).
One may have interest in the particular case θ  = −θ . This case leads to the
so-called direct backscattering Born approximation, i.e.,

 |2 +V )(2kθ ).
A(k, −θ , θ ) ≈ AbB (k, −θ , θ ) := F (|W (23.30)
242 Part II: Fourier Transform and Distributions

But the approximation for the backscattering amplitude admits more terms than just
the Born backscattering approximation. Namely, the following theorem holds.

Theorem 23.19. Under the conditions of Theorem 23.15 the backscattering ampli-
tude A(k, −θ , θ ) admits the following representation:

 |2 +V )(2kθ ) − F (q)(kθ + η )F (q)(kθ − η )
1
A(k, −θ , θ ) = F (|W dη
(2π )n η 2 − k2 − i0
Rn
  )(kθ + η )η F (W
 )(kθ − η )
4k θ F (W
+ dη + hrest (kθ ),
(2π )n Rn η − k − i0
2 2
(23.31)

 + |W
where q denotes the complex conjugate of q = i∇W  |2 + V and where hrest
∞
belongs to L (R ) and
n

β 2 αγ 2
hrest L∞ (Rn ) ≤ 3 . (23.32)
1−βα

Proof. The formulas (23.27) and (23.28) for the case θ  = −θ show that we need
to investigate only
 
R(k, −θ , θ ) = −2kθ  (y)usc (y, k, θ )dy +
eik(θ ,y)W eik(θ ,y) q(y)usc (y, k, θ )dy.
Rn Rn
(23.33)

But since usc = ∑∞j=1 Lkj (u0 ), we see that (23.33) can be rewritten as

R(k, −θ , θ ) = −2kθ  (y)Lk u0 (y, k, θ )dy
eik(θ ,y)W
Rn

+ eik(θ ,y) q(y)Lk u0 (y, k, θ )dy
Rn
 ∞
− 2kθ  (y) ∑ L j u0 (y, k, θ )dy
eik(θ ,y)W k
Rn j=2
 ∞
+ eik(θ ,y) q(y) ∑ Lkj u0 (y, k, θ )dy =: R1 + R2 .
Rn j=2

The definition of Lk u0 allows us to obtain

23 Estimates for the Laplacian and Hamiltonian 243
 
eik(θ ,y+z) G+  
R1 = 4k2 k (|y − z|)θ W (y)θ W (z)dydz
Rn Rn
 
eik(θ ,y+z) G+ 
+ 2k k (|y − z|)θ W (y)(q(z) − q(z))dydz
Rn Rn
 
− eik(θ ,y+z) G+
k (|y − z|)q(y)q(z)dydz =: I1 + I2 + I3 .
Rn Rn

Using now the facts

1
F (G+
k (·))(η ) = η 2 − k2 − i0

and F (ϕψ ) = (2π )−n F (ϕ ) ∗ F (ψ ), we obtain that I j , j = 1, 2, 3, can be written

as
  )(kθ + η )θ F (W )(kθ − η )
4k2 θ F (W
I1 = dη ,
(2π )nRn η 2 − k2 − i0
  )(kθ + η )F (∇W  )(kθ − η )
4ik θ F (W
I2 = − dη ,
(2π ) R
n n η − k − i0
2 2
  )(kθ + η )η F (W )(kθ − η )
4k θ F (W
= −I1 + dη ,
(2π ) Rn
n η − k − i0
2 2

1 F (q)(kθ + η )F (q)(kθ − η )
I3 = − dη .
(2π ) Rn
n η 2 − k2 − i0

Thus, the second terms in (23.31) are proved. It remains to estimate hrest (or R2 ).
Indeed, the definition of R2 allows us to obtain (using integration by parts) that
 ∞
R2 (k, −θ , θ ) = − eik(θ ,y) q(y) ∑ Lkj u0 (y, k, θ )dy
Rn j=2

 
∞
ik(θ ,y) 
+ 2i
Rn
e W (y) · ∇ ∑ Lkj u0 (y, k, θ ) dy.
j=2

Since
Lk H 1 (Rn )→L−
2 (Rn ) ≤ β α,
−σ /2 σ /2

it follows using duality that R2 (k, −θ , θ ) can be estimated as

|R2 (k, −θ , θ )|
    
∞     ∞ 
     
≤ qL2 (Rn )  ∑ Lkj u0  + 2 W  ∇ ∑ L j
u0 
σ /2  j=2  Hσ /2 (R )
1 n  j=2
k  −1
2
L− σ /2
(Rn ) H− σ /2
(Rn )

β αβ γ   β αβ γ β 2 αγ 2
 
≤ qL2 (Rn ) + 2 W  1 ≤ 3 .
σ /2 1−βα Hσ /2 (Rn ) 1 − β α 1−βα
244 Part II: Fourier Transform and Distributions

Thus, Theorem 23.19 is completely proved. 



Remark 23.20. This theorem (as well as Theorem 23.15) is a generalization of the
corresponding results for the Schrödinger operator. But the difference is that com-
pared with the Schrödinger operator, the magnetic Schrödinger operator is not a
“small” perturbation of the Laplacian. This is a reason for the smallness of norms in
Theorem 23.15. For the Schrödinger operator we do not need this requirement.

Remark 23.21 (One-dimensional case). There is one interesting remark that should
be made here. The asymptotic representations (see Theorem 23.15 of formulas
(23.24)) and (23.26) coincide with well known formulas in the one-dimensional
case. Moreover, the definition (23.26) of the scattering amplitude defines the reflec-
tion and transmission coefficients for the one-dimensional “magnetic” Schrödinger
operator.
 Part III
Operator Theory and Integral Equations
Chapter 24
Introduction

Despite the fact that this part is devoted to Hilbert spaces, it is assumed that the fol-
lowing concepts are known (they are necessary mainly for examples and exercises):
(1) the Lebesgue integral in a bounded domain Ω ⊂ Rn and in Rn ;
(2) functions of bounded variation BV [a, b] on an interval [a, b] (see Part I for
details);
(3) the Stieltjes integral of continuous functions on [a, b];
(4) a complete normed space Ck (Ω ), k = 0, 1, 2, . . ., on a closed bounded domain
Ω ⊂ Rn defined by

Ck (Ω ) := { f : Ω → C :  f Ck (Ω ) := max ∑
Ω |α |≤k
|∂ α f (x)| < ∞},

where α is an n-dimensional multi-index, i.e., α = (α1 , . . . , αn ), α j ∈ N ∪ {0},

∂ |α | f
j = 1, 2, . . . , n with |α | = α1 + α2 + · · · + αn and ∂ α f = α ;
∂ x1 1 ···∂ xnαn
(5) a complete normed space L∞ (Ω ) defined by

L∞ (Ω ) := { f : Ω → C :  f L∞ (Ω ) := ess sup | f (x)| < ∞};

Ω

(6) a complete normed space L1 (Ω ) for an open set Ω ⊂ Rn defined by


L1 (Ω ) := { f : Ω → C :  f L1 (Ω ) := | f (x)|dx < ∞};
Ω

(7) the generalized (in the L2 sense) derivatives ∂ α f (x), α = (α1 , . . . , αn ) (see Part
II for details);

c Springer International Publishing AG 2017 247

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 24
248 Part III: Operator Theory and Integral Equations

(8) Lebesgue’s dominated convergence theorem: let Ω ⊂ Rn be measurable and let

{ fk (x)}∞
k=1 be a sequence of measurable functions converging to f (x) point-
wise in Ω ; if there exists a function g(x) ∈ L1 (Ω ) such that | fk (x)| ≤ g(x),
k = 1, 2, . . ., then f (x) ∈ L1 (Ω ) and
 
lim fk (x)dx = f (x)dx;
k→∞ Ω Ω

(9) Fubini’s theorem on interchanging the order of integration: if f (x, y) is inte-

grable on X ×Y , then
      
dx f (x, y)dy = dy f (x, y)dx = f (x, y)dxdy;
X Y Y X X×Y

(10) the uniform boundedness principle in Hilbert space (Banach–Steinhaus the-

orem): let H be a Hilbert space; suppose that F is a collection of bounded
(continuous) linear operators in H; if for all x ∈ H, then one has

sup AxH < ∞,

A∈F

whence
sup AxH = sup AH→H < ∞,
A∈F,x=1 A∈F

see Theorem 26.3.

Chapter 25
Inner Product Spaces and Hilbert Spaces

A collection of elements H is called a complex (real) vector space (linear space) if

the following axioms are satisfied:

(1) To every pair x, y ∈ H there corresponds a vector x + y, called the sum, with the
following properties:
(a) x + y = y + x;
(b) x + (y + z) = (x + y) + z ≡ x + y + z;
(c) there exists a unique element 0 ∈ H such that x + 0 = x;
(d) for every x ∈ H there exists a unique element y1 ∈ H such that x + y1 = 0.
We set y1 := −x.
(2) For every x ∈ H and every λ , μ ∈ C there corresponds a vector λ · x such that
(a) λ (μ x) = (λ μ )x ≡ λ μ x;
(b) (λ + μ )x = λ x + μ x;
(c) λ (x + y) = λ x + λ y;
(d) 1 · x = x.

Definition 25.1. For a linear space H, a mapping (·, ·) : H × H → C is called an

inner product or a scalar product if for every x, y, z ∈ H and λ ∈ C the following
conditions are satisfied:
(1) (x, x) ≥ 0 and (x, x) = 0 if and only if x = 0;
(2) (x, y + z) = (x, y) + (x, z);
(3) (λ x, y) = λ (x, y);
(4) (x, y) = (y, x).
A linear space equipped with an inner product is called an inner product space.

c Springer International Publishing AG 2017 249

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 25
250 Part III: Operator Theory and Integral Equations

An immediate consequence of this definition is that

(λ x + μ y, z) = λ (x, z) + μ (y, z),

(x, λ y) = λ (x, y),

for every x, y, z ∈ H and λ , μ ∈ C.

Example 25.2. In the complex Euclidean space H = Cn the standard inner product
is
n
(x, y) = ∑ x jy j,
j=1

where x = (x1 , . . . , xn ) ∈ Cn and y = (y1 , . . . , yn ) ∈ Cn .

Example 25.3. In the linear space C[a, b] of continuous complex-valued functions,

the formula  b
( f , g) = f (x)g(x)dx
a

defines an inner product.

Definition 25.4. Suppose H is an inner product space. Then

(1) x ∈ H is orthogonal to y ∈ H if (x, y) = 0. 
1, α = β ,
(2) A system {xα }α ∈I ⊂ H is orthonormal if (xα , xβ ) = δα ,β =
0, α =
 β,
where I
is some index set.
(3) x := (x, x) is called the length of x ∈ H.

Exercise 25.1. Prove the Pythagorean theorem: If {x j }kj=1 , k ∈ N, is an orthonor-

mal system in an inner product space H, then
 2
k  k 
 
x2 = ∑ |(x, x j )|2 + x − ∑ (x, x j )x j 
j=1
 j=1


for every x ∈ H.

Exercise 25.2. Prove Bessel’s inequality: if {x j }kj=1 , k ≤ ∞, is an orthonormal sys-

tem, then
k
∑ |(x, x j )|2 ≤ x2 ,
j=1

for every x ∈ H.
25 Inner Product Spaces and Hilbert Spaces 251

Exercise 25.3. Prove the Cauchy–Bunyakovsky–Schwarz inequality:

|(x, y)| ≤ x y , x, y ∈ H.

Prove also that (·, ·) is continuous as a map from H × H to C.

If H is an inner product space, then


x := (x, x)

has the following properties:

(1) x ≥ 0 for every x ∈ H and x = 0 if and only if x = 0.
(2) λ x = |λ | x for every x ∈ H and λ ∈ C.
(3) x + y ≤ x + y for every x, y ∈ H. This is the triangle inequality.

The function · = (·, ·) is thus a norm on H. It is called the norm induced by the
inner product.
Every inner product space H is a normed space under the induced norm. The
neighborhood of x ∈ H is the open ball Br (x) = {y ∈ H : x − y < r}. This system
of neighborhoods defines the norm topology on H such that the following conditions
are satisfied:
(1) Addition x + y is a continuous map H × H → H.
(2) Scalar multiplication λ · x is a continuous map C × H → H.
(3) The inner product (x, y) : H × H → C is continuous.

Definition 25.5. (1) A sequence {x j }∞j=1 ⊂ H is called a Cauchy sequence if for

 
every ε > 0 there exists n0 ∈ N such that xk − x j  < ε for k, j ≥ n0 .
(2) A sequence {x j }∞j=1 ⊂ H is said to be convergent if there exists x ∈ H such that
 
for every ε > 0 there exists n0 ∈ N such that x − x j  < ε whenever j ≥ n0 .
(3) An inner product space H is a complete space if every Cauchy sequence in H
converges.

Corollary 25.6. (1) Every convergent sequence is a Cauchy sequence.

(2) If {x j }∞j=1 converges to x ∈ H, then
 
lim x j  = x .
j→∞

Definition 25.7 (J. von Neumann, 1925). A Hilbert space is an inner product
space that is complete (with respect to its norm topology).

Exercise 25.4. Prove that in an inner product space the norm induced by this inner
product satisfies the parallelogram law

x + y2 + x − y2 = 2 x2 + 2 y2 .

252 Part III: Operator Theory and Integral Equations

Exercise 25.5. Prove that if in a normed space H the parallelogram law holds, then
there is an inner product on H such that x2 = (x, x) and that this inner product is
defined by the polarization identity

1 
(x, y) := x + y2 − x − y2 + i x + iy2 − i x − iy2 .
4

Exercise 25.6. Prove that on C[a, b] the norm

 f  = max | f (x)|
x∈[a,b]

is not induced by an inner product.

Exercise 25.7. Give an example of an inner product space that is not complete.
Next we list some examples of Hilbert spaces.
(1) The Euclidean spaces Rn and Cn .
(2) The matrix space Mn (C) consisting of n × n matrices whose elements are com-
plex numbers. For A, B ∈ Mn (C) the inner product is given by
n
(A, B) = ∑ ak j bk j = Tr(AB∗ ),
k, j=1

T
where B∗ = B .
(3) The sequence space l 2 (C) defined by
 
∞
l 2 (C) := {x j }∞j=1 , x j ∈ C : ∑ |x j |2 < ∞ .
j=1

The estimates

|x j + y j |2 ≤ 2 |x j |2 + |y j |2 , |λ x j |2 = |λ |2 |x j |2

and
1 2

|x j y j | ≤ |x j | + |y j |2
2

imply that l 2 (C) is a linear space. Let us define the inner product by
∞
(x, y) := ∑ x jy j
j=1

and prove that l 2 (C) is complete. Suppose that {x(k) }∞

k=1 ∈ l (C) is a Cauchy
2

sequence. Then for every ε > 0 there exists n0 ∈ N such that

25 Inner Product Spaces and Hilbert Spaces 253

 2 ∞
 (k)  (k) (m)
x − x(m)  = ∑ |x j − x j |2 < ε 2
j=1

for k, m ≥ n0 . This implies that

(k) (m)
|x j − x j | < ε , j = 1, 2, . . . ,

(k)
or that {x j }∞
k=1 is a Cauchy sequence in C for every j = 1, 2, . . .. Since C is a
(k)
complete space, it follows that {x j }∞
k=1 converges for every fixed j = 1, 2, . . .,
i.e., there exists x j ∈ C such that

(k)
x j = lim x j .
k→∞

This fact and

l
(k) (m)
∑ |x j − x j |2 < ε 2 , l ∈ N,
j=1

imply that
l l
(k) (m) (k)
lim
m→∞
∑ |x j − x j |2 = ∑ |x j − x j |2 ≤ ε 2
j=1 j=1

for all k ≥ n0 and l ∈ N. Therefore, the sequence

l
(k)
sl := ∑ |x j − x j |2 , k ≥ n0 ,
j=1

is a monotonically increasing sequence that is bounded from above by ε 2 .

Hence this sequence has a limit with the same upper bound, i.e.,

∞ l
(k) (k)
∑ |x j − x j |2 = lim
l→∞
∑ |x j − x j |2 ≤ ε 2 ,
j=1 j=1

from which we conclude that

     
     
x ≤ x(k)  + x(k) − x ≤ x(k)  + ε

and x ∈ l 2 (C).
(4) The Lebesgue space L2 (Ω ), where Ω ⊂ Rn is an open set. The space L2 (Ω )
consists of all Lebesgue measurable functions f that are square integrable, i.e.,

| f (x)|2 dx < ∞.
Ω
254 Part III: Operator Theory and Integral Equations

This space is a linear space with the inner product


( f , g) = f (x)g(x)dx
Ω

and the Riesz–Fischer theorem reads as follows: L2 (Ω ) is a Hilbert space.

(5) The Sobolev spaces W2k (Ω ) consisting of functions f ∈ L2 (Ω ) whose weak
or distributional derivatives ∂ α f also belong to L2 (Ω ) up to order |α | ≤ k,
k = 1, 2, . . .. On the space W2k (Ω ) the natural inner product is

( f , g) = ∑ ∂ α f (x)∂ α g(x)dx.
|α |≤k Ω

Definition 25.8. Let H be an inner product space. For a linear subspace M ⊂ H the
orthogonal complement of M is defined as

M ⊥ := {y ∈ H : (y, x) = 0, for all x ∈ M} .

Remark 25.9. It is clear that M ⊥ is a linear subspace of H. Moreover, M ∩ M ⊥ =

{0}, since we always have 0 ∈ M.

Definition 25.10. A closed subspace of a Hilbert space H is a linear subspace of H

that is closed (i.e., M = M) with respect to the induced norm.

Remark 25.11. The subspace M ⊥ is closed if M is a subset of a Hilbert space.

Theorem 25.12 (Projection theorem). Suppose M is a closed subspace of a

Hilbert space H. Then every x ∈ H has a unique representation as

x = u + v,

where u ∈ M and v ∈ M ⊥ , or equivalently,

H = M ⊕ M⊥.

Moreover, one has that

v = inf x − y =: d(x, M).

y∈M

Proof. Let x ∈ H. Then

d := d(x, M) ≡ inf x − y ≤ x − u
y∈M

for all u ∈ M. The definition of infimum implies that there exists a sequence
{u j }∞j=1 ⊂ M such that
25 Inner Product Spaces and Hilbert Spaces 255
 
d = lim x − u j  .
j→∞

The parallelogram law implies that

 2
       
u j − uk 2 = (u j − x) + (x − uk )2 = 2 u j − x2 +2 x − uk 2 −4 x − u j + uk  .
 2 

Since (u j + uk )/2 ∈ M, it follows that

   
u j − uk 2 ≤ 2 u j − x2 + 2 x − uk 2 − 4d 2 → 2d 2 + 2d 2 − 4d 2 = 0

as j, k → ∞. Hence {u j }∞j=1 ⊂ M is a Cauchy sequence in the Hilbert space H. This

means that there exists u ∈ H such that

u = lim u j .
j→∞

But M = M implies that u ∈ M. By construction one has that

 
d = lim x − u j  = x − u .
j→∞

Let us set v := x − u and show that v ∈ M ⊥ . For all y ∈ M, y = 0, we introduce the

number
(v, y)
α =− .
y2

Since u − α y ∈ M, we have

d 2 ≤ x − (u − α y)2 = v + α y2 = v2 + (v, α y) + (α y, v) + |α |2 y2

(v, y)(v, y) (v, y)(y, v) |(v, y)|2 |(v, y)|2
= d2 − − + = d2 − .
y2 y2 y2 y2

This inequality implies that (y, v) = 0, which means that v ∈ M ⊥ . In order to prove
uniqueness, assume that x = u1 + v1 = u2 + v2 , where u1 , u2 ∈ M and v1 , v2 ∈ M ⊥ .
It follows that
u1 − u2 = v2 − v1 ∈ M ∩ M ⊥ .

But M ∩ M ⊥ = {0}, so that u1 = u2 and v1 = v2 . 

Remark 25.13. In the framework of this theorem we have that

x2 = u2 + v2 , v2 = (x, v), u2 = (x, u).

256 Part III: Operator Theory and Integral Equations

Corollary 25.14 (Riesz-Fréchet theorem). If T is a linear continuous functional

on the Hilbert space H, then there exists a unique h ∈ H such that T (x) = (x, h) for
all x ∈ H. Moreover, T H→C = h.
Proof. If T ≡ 0, then h = 0 will do. If T = 0, then there exists v0 ∈ H such that
T (v0 ) = 0. Let
M := {u ∈ H : T (u) = 0} .

Then v0 ∈ M ⊥ , v0 = 0, and T (v0 ) = 0. Since T is linear and continuous, M is a

closed subspace. It follows from Theorem 25.12 that

H = M ⊕ M⊥,

i.e., every x ∈ H has a unique representation as x = u+ v. Therefore, for every x ∈ H,

we can define
T (x)
u := x − v0 .
T (v0 )

Then T (u) = 0, i.e., u ∈ M. It follows that

T (x) T (x)
(x, v0 ) = (u, v0 ) + v0 2 = v0 2 ,
T (v0 ) T (v0 )

or 
T (v0 ) T (v0 )
T (x) = (x, v0 ) = x, v0 ,
v0 2 v0 2

which is of the desired form. The uniqueness of h can be seen as follows. If T (x) =
 2
 
(x, h) = (x,  h) = 0 for all x ∈ H. In particular, h − 
h), then (x, h −  h = (h − h,
h− h) = 0, i.e., h = 
h. It remains to prove the statement about the norm T H→C =
T . Firstly,
T  = sup |T (x)| = sup |(x, h)| ≤ h .
x≤1 x≤1

On the other hand, T (h/ h) = h implies that T  ≥ h. Thus T  = h. This
completes the proof. 
Corollary 25.15. If M is a linear subspace of a Hilbert space H, then
 ⊥
M ⊥⊥ := M ⊥ = M.

⊥
Proof. It is not difficult to check that M ⊥ = M . Therefore,

 ⊥
⊥
M ⊥⊥ = M ,
25 Inner Product Spaces and Hilbert Spaces 257

and Theorem 25.12 implies that

⊥
⊥
H =M⊕ M , H = M ⊕ M ⊥⊥ .

The uniqueness of this representation guarantees that M ⊥⊥ = M. 

Definition 25.16. Let A ⊂ H be a subset of an inner product space. The subset
 
k
span A := x∈H :x= ∑ λ j x j , x j ∈ A, λ j ∈ C
j=1

is called the linear span of A.

Definition 25.17. Let H be a Hilbert space.
(1) A subset B ⊂ H is called a basis of H if B is linearly independent in H and

span B = H,

i.e., for every x ∈ H and every ε > 0 there exist k ∈ N and {c j }kj=1 ⊂ C such
that  
 k 
 
x − ∑ c j x j  < ε , x j ∈ B.
 j=1


(2) H is called separable if it has a countable or finite basis.

(3) An orthonormal system B = {xα }α ∈A in H that is a basis is called an orthonor-
mal basis.
By Gram–Schmidt orthonormalization we may conclude that every separable
Hilbert space has an orthonormal basis.
Theorem 25.18 (Characterization of an orthonormal basis). Let B = {x j }∞j=1
be an orthonormal system in a separable Hilbert space H. Then the following state-
ments are equivalent:
(1) B is maximal, i.e., it is not a proper subset of any other orthonormal system.
(2) For every x ∈ H the condition (x, x j ) = 0, j = 1, 2, . . ., implies that x = 0.
(3) Every x ∈ H has the Fourier expansion
∞
x= ∑ (x, x j )x j ,
j=1

i.e.,  
 k 
 
x − ∑ (x, x j )x j  → 0, k → ∞.
 j=1


This means that B is an orthonormal basis.

258 Part III: Operator Theory and Integral Equations

(4) Every pair x, y ∈ H satisfies the completeness relation

∞
(x, y) = ∑ (x, x j )(y, x j ).
j=1

(5) Every x ∈ H satisfies Parseval’s equality

∞
x2 = ∑ |(x, x j )|2 .
j=1

Proof. (1)⇒(2) Suppose that there is z ∈ H, z = 0 such that (z, x j ) = 0 for all j =
1, 2, . . .. Then  
 z
B := , x1 , x2 , . . .
z

is an orthonormal system in H. This fact implies that B is not maximal, which

contradicts (1) and proves (2).
(2)⇒(3) Given x ∈ H, we introduce the sequence

k
x(k) = ∑ (x, x j )x j .
j=1

The Pythagorean theorem and Bessel’s inequality (Exercises 25.1 and 25.2) im-
ply that
 2 k
 (k) 
x  = ∑ |(x, x j )|2 ≤ x .
2

j=1

It follows that
∞
∑ |(x, x j )|2
j=1

converges. Therefore, for m < k,

 2 k
 (k) 
x − x(m)  = ∑ |(x, x j )|2 → 0
j=m+1

as k, m → ∞. Hence x(k) is a Cauchy sequence in H. Thus there exists y ∈ H such

that
∞
y = lim x(k) =
k→∞
∑ (x, x j )x j .
j=1

Next, since the inner product is continuous, we deduce that

25 Inner Product Spaces and Hilbert Spaces 259

(y, x j ) = lim (x(k) , x j ) = (x, x j )

k→∞

for all j = 1, 2, . . .. Therefore, (y − x, x j ) = 0 for all j = 1, 2, . . .. Part (2) implies

that y = x and part (3) follows.
(3)⇒(4) Let x, y ∈ H. We know from part (3) that
∞ ∞
x= ∑ (x, x j )x j , y= ∑ (y, xk )xk .
j=1 k=1

The continuity of the inner product and the orthonormality of {x j }∞j=1 allow us
to conclude that
∞ ∞ ∞
(x, y) = ∑ ∑ (x, x j )(y, xk )(x j , xk ) = ∑ (x, x j )(y, x j ).
j=1 k=1 j=1

(4)⇒(5) Take y = x in part (4).

(5)⇒(1) Suppose that B is not maximal. Then we can add a unit vector z ∈ H to it
that is orthogonal to B. Parseval’s equality gives then
∞
1 = z2 = ∑ |(z, x j )|2 = 0.
j=1

This contradiction proves the result. 

Exercise 25.8. Let {x j }∞j=1 be an orthonormal system in an inner product space H.

Let x ∈ H, {c j }kj=1 ⊂ C, and k ∈ N. Prove that
   
 k   k 
   
x − ∑ (x, x j )x j  ≤ x − ∑ c j x j  .
 j=1
  j=1

Chapter 26
Symmetric Operators in Hilbert Spaces

Assume that H is a Hilbert space. A linear operator from H to H is a mapping

A : D(A) ⊂ H → H,

where D(A) is a linear subspace of H and A satisfies the condition

A(λ x + μ y) = λ Ax + μ Ay

for all λ , μ ∈ C and x, y ∈ D(A). The space D(A) is called the domain of A. The
space
N(A) := {x ∈ D(A) : Ax = 0}

is called the null space (or kernel) of A. The space

R(A) := {y ∈ H : y = Ax for some x ∈ D(A)}

is called the range of A. Both N(A) and R(A) are linear subspaces of H. We say that
A is bounded if there exists M > 0 such that

Ax ≤ M x , x ∈ D(A).

We say that A is densely defined if D(A) = H. In such a case, A can be extended

to Aex , which will be defined on the whole H with the same norm estimate, and we
may define
AH→H := inf{M : Ax ≤ M x , x ∈ D(A)},

or equivalently,
AH→H = sup Ax .
x=1

c Springer International Publishing AG 2017 261

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 26
262 Part III: Operator Theory and Integral Equations

Example 26.1 (Integral operator in L2 ). Suppose that K(s,t) ∈ L2 (Ω × Ω ), Ω ⊂ Rn .

 defined as
Let us show that the integral operator K

 f (s) =
K K(s,t) f (t)dt, f ∈ L2 (Ω )
Ω

is bounded. Indeed,
 2    2

  
K f  2 =  f (s)|2 ds =
|K K(s,t) f (t)dt  ds

L (Ω ) Ω ΩΩ
  2   2
= (K(s, ·), f ) 2  ds ≤ K(s, ·)2L2  f L2 ds
L
Ω Ω
   
= |K(s,t)| dt | f (t)| dt ds = K2L2 (Ω ×Ω )  f 2L2 (Ω ) ,
2 2
Ω Ω Ω

where we have made use of the Cauchy–Bunyakovsky–Schwarz inequality. We

therefore have  
 
K  2 2 ≤ KL2 (Ω ×Ω ) .
L →L

The norm  
 
K  := KL2 (Ω ×Ω )
HS


is called the Hilbert–Schmidt norm of K.

Example 26.2 (Schur test). Assume that p and q are positive measurable functions
on Ω ⊂ Rn and α and β are positive numbers such that

|K(x, y)|p(y)dy ≤ α q(x), a.e. in Ω
Ω

and 
|K(x, y)|q(x)dx ≤ β p(y), a.e. in Ω .
Ω

 is bounded and
Then K  
  
K  ≤ αβ .
L2 →L2

Proof. For all f ∈ L2 (Ω ) we have

26 Symmetric Operators in Hilbert Spaces 263

  2
|K(x, y)| · | f (y)|dy dx
Ω Ω

2
   |K(x, y)|
= |K(x, y)| p(y) | f (y)|dy dx
Ω Ω p(y)
    
|K(x, y)|
≤ |K(x, y)|p(y)dy | f (y)|2 dy dx
Ω Ω Ω p(y)
   
| f (y)|2
≤α |K(x, y)|q(x)dx dy ≤ αβ | f (y)|2 dy
Ω Ω p(y) Ω

by the Cauchy–Bunyakovsky–Schwarz inequality and Fubini’s theorem. 

Exercise 26.1. Assume that α and β are positive constants such that

|K(x, y)|dy ≤ α , a.e. in Ω
Ω

and 
|K(x, y)|dx ≤ β , a.e. in Ω
Ω

for some measurable function K(x, y) on Ω × Ω , Ω ⊂ Rn . Show that the integral

 with kernel K is bounded in L p (Ω ) for all 1 ≤ p ≤ ∞ and
operator K
 
  1 1 1 1
K  ≤ α p β p , + = 1.
L p →L p p p

The following fundamental result can be used in the theory of bounded linear
operators (see [29]).

Theorem 26.3 (Uniform boundedness principle). Suppose that a sequence An :

H → H of bounded linear operators satisfies the property (pointwise boundedness)

sup An uH ≤ Cu . (26.1)

n

Then there is a constant C > 0 such that

sup An H→H ≤ C.

n

Proof. Let us assume to the contrary that

sup An H→H = +∞. (26.2)

n
264 Part III: Operator Theory and Integral Equations

Then for each k ≥ 1 there exist Ank and uk ∈ H such that

     
uk  = 4−k , An uk  ≥ 2 An  uk  , An uk  ≥ 2(Mk−1 + k), (26.3)
k k k
3

where M0 = 1 and Mk = supn An (u1 + · · · + uk ). Indeed, by (26.2) there exists An1
with An1  ≥ 24. The definition of the norm of a linear operator allows us to choose
u1 such that u1  = 1 and An1 u1  ≥ 23 An1 . Setting u1 = u1 /4 shows that all
conditions (26.3) are satisfied for k = 1 and with M0 = 1.
Assuming that u1 , u2 , . . . , uk−1 , An1 , An2 , . . . , Ank−1 have been defined, choose Ank
such that  
An  ≥ 3 · 4k (Mk−1 + k),
k

which is possible by hypothesis  With this choicek of Ank there exists uk such
 2  (26.2).

that uk  = 1 and Ank uk ≥ 3 Ank . Setting uk = uk /4 , we have again that uk  =
 
4−k and
   
An uk  ≥ 2 · 4−k An  ≥ 2 · 4−k 3 · 4k (Mk−1 + k) = 2(Mk−1 + k).
k k
3 3

To complete the proof we put u := ∑∞

k=1 uk , which is well defined in H. But then we
have
 
 ∞    ∞     ∞ 1 
 
Ank ∑ u j  ≤ Ank  ∑ u j  ≤ Ank  ∑ 4− j = Ank  uk  .
 j=k+1
 j=k+1 j=k+1 3

By the triangle inequality and the definition of Mk we have

   
     k−1   ∞ 

An u ≥ An uk  − An ∑ u j  
 − A ∑ j 
u

k k
 k j=1   nk j=k+1
  1 
≥ Ank uk  − Mk−1 − Ank uk  ≥ k.
2
This contradiction with (26.1) proves the theorem. 
Remark 26.4. The uniform boundedness principle holds not only in Hilbert spaces
but also in Banach spaces (complete normed spaces). This fact follows straightfor-
wardly from the proof.
Corollary 26.5 (Banach–Steinhaus). Under the conditions of Theorem 26.3 it is
true that for pointwise convergence An u → Au, n → ∞, for all u ∈ H it is necessary
and sufficient that supn An H→H ≤ C and that An u → Au, n → ∞ for all u ∈ U,
where U is some dense subset of H.
Corollary 26.6 (Trigonometric Fourier series). There exists a continuous func-
tion whose Fourier partial sums Sn f (x) do not remain uniformly bounded. For every
26 Symmetric Operators in Hilbert Spaces 265

x ∈ [−π , π ] there exists a continuous function whose Fourier partial sums Sn f (x) are
unbounded at x.

Proof. Let us consider on the Banach space C[−π , π ] of continuous and periodic
functions on the interval [−π , π ] the linear operators

f → Sn f (x) = ∑ cn ( f )einx ,
n≤N

where cn ( f ) are the trigonometric Fourier coefficients of f . Since we have the sharp
estimate 
1 π
Sn f L∞ (−π ,π ) ≤ |DN (x)|dx  f L∞ (−π ,π ) ,
2π −π

where DN (x) is the Dirichlet kernel (see Chapter 10), choosing the sequence
fn (x) := σn ( f0 ) defined by Fejér means with f0 (x) = sgn DN (x), we obtain
 π
2π SN fn (0) = fn (x)DN (x)dx
−π
 π  π
8 log N
→ f0 (x)DN (x)dx = |DN (x)|dx = + O(1), N → ∞,
−π −π π

as is stated in Exercise 10.3. Thus, the linear operators f → SN f (x) are bounded (for
each fixed N) with operator norms

4 log N
SN  = + O(1).
π2
Therefore, by the uniform boundedness principle, there exists a continuous function
satisfying the present corollary. 

Exercise 26.2 (Hellinger–Toeplitz). Suppose that D(A) = H and

(Ax, y) = (x, Ay), x, y ∈ H.

Prove that A is bounded.

Exercise 26.3. Suppose that f ∈ L1 (−π , π ) is periodic. Prove that if for some x
there exists
f (x + y) + f (x − y)
lim ,
y→0 2

then SN f (x) = o(log N) as N → ∞.

266 Part III: Operator Theory and Integral Equations

Example 26.7 (Differential operator in L2 ). Consider the differential operator

d
A := i
dt

of order 1 in L2 (0, 1) with domain


D(A) = f ∈ C1 [0, 1] : f (0) = f (1) = 0 .

First of all, we have that D(A) = L2 (0, 1); see, e.g., Lemma 17.2. Moreover, inte-
gration by parts gives
 1  1  1
(A f , g) = i f  (t)g(t)dt = i f g|10 − i f (t)g (t)dt = f (t)ig (t)dt = ( f , Ag)
0 0 0

for all f , g ∈ D(A). Let us now consider the sequence

un (t) := sin(nπ t), n = 1, 2, . . . .

Clearly, un ∈ D(A) and

 1
1
un 2L2 = | sin(nπ t)|2 dt = .
0 2

But
 1  2
  1
 d  dt = (nπ )2 (nπ )2
Aun 2L2 = i
 dt sin(nπ t)  | cos(nπ t)|2 dt = = (nπ )2 un 2L2 .
0 0 2

Therefore, A is unbounded. This shows that D(A) = H is an essential assumption in

Exercise 26.2.

Example 26.8 (Differential operator in L2 ). Consider the differential operator

d2 d
A := p0 + ip1 + p2
dt 2 dt

of order 2 in L2 (0, 1) with domain

D(A) = { f ∈ C2 [0, 1] : f (0) = f (1) = 0}

and with real nonzero constant coefficients p0 , p1 , and p2 . The fact D(A) = L2 and
integration by parts gives
26 Symmetric Operators in Hilbert Spaces 267
 1  1  1
(A f , g) = p0 f  · gdt + ip1 f  · gdt + p2 f · gdt
0 0
  0

1  1  1
= p0 f  g0 −  
f · g dt + ip1 f g|0 − 1 
f · g dt + p2 ( f , g)L2
0 0
 1  1
= −p0 f  · g dt − ip1 f · g dt + ( f , p2 g)L2
0 0
  1 

f g 0 −
1
= −p0 f · g dt + ( f , ip1 g )L2 + ( f , p2 g)L2

0
 1
= p0 f · g dt + ( f , ip1 g )L2 + ( f , p2 g)L2 = ( f , Ag)L2
0

for all f , g ∈ D(A). Moreover, for the sequence un (t) = sin(nπ t) we have (for suffi-
ciently large n) that
 1
Aun 2L2 = |p0 (sin(nπ t)) + ip1 (sin(nπ t)) + p2 sin(nπ t)|2 dt
0
 1 
= (p0 (nπ )2 − p2 )2 sin2 (nπ t) + (nπ )2 p21 cos2 (nπ t) dt
0
 1 
(nπ )4 2 2
≥ p0 sin (nπ t) + (nπ )2 p21 cos2 (nπ t) dt
0 2
 1 
≥ (nπ )2 p21 sin2 (nπ t) + cos2 (nπ t) dt
0
1
= 2(nπ )2 p21 = 2(nπ )2 p21 un 2L2 .
2
So A is unbounded, since
A2L2 →L2 ≥ 2(nπ )2 p21

for n → ∞.

From now on we assume that D(A) = H, i.e., that A is densely defined in any case.

Definition 26.9. The graph Γ (A) of a linear operator A in a Hilbert space H is

defined as
Γ (A) := {(x; y) ∈ H × H : x ∈ D(A) and y = Ax} .

Remark 26.10. The graph Γ (A) is a linear subspace of a Hilbert space H × H. The
inner product in H × H can be defined as

((x1 ; y1 ) , (x2 ; y2 ))H×H := (x1 , x2 )H + (y1 , y2 )H

for all (x1 ; y1 ) , (x2 ; y2 ) ∈ H × H.

268 Part III: Operator Theory and Integral Equations

Definition 26.11. The operator A is said to be closed if Γ (A) = Γ (A). We denote

this fact by A = A.

By definition, the criterion for closedness is that

⎧
⎪ 
⎨xn ∈ D(A), x ∈ D(A),
xn → x, ⇒
⎪
⎩ y = Ax.
Axn → y

The reader is asked to verify that it is also possible to use a seemingly weaker, but
equivalent, criterion:
⎧
⎪ 
⎨xn ∈ D(A), x ∈ D(A),
w
xn → x, ⇒
⎪
⎩ w y = Ax,
Axn → y

w
where xn → x indicates weak convergence in the sense that

(xn , y) → (x, y)

for all y ∈ H.

Remark 26.12. It is important from the point of view of applications (in particular,
for numerical procedures) that the closedness of an operator guarantees the conver-
gence of some process to the “correct” result.

Definition 26.13. Let A and A1 be two linear operators in a Hilbert space H. The
operator A1 is called an extension of A (or A is a restriction of A1 ) if D(A) ⊂ D(A1 )
and Ax = A1 x for all x ∈ D(A). We denote this fact by A ⊂ A1 and A = A1 |D(A) .

Definition 26.14. An operator A is called closable if A has an extension A1 and

A1 = A1 . The closure of A, denoted by A, is the smallest closed extension of A if it
exists, i.e., 
A= A1 .
A⊂A1
A1 =A1

1 we mean the operator whose domain is D(A1 ∩ A

Here, by A1 ∩ A 1 ) := D(A1 ) ∩
1 ) and
D(A
1 )x := A1 x = A
(A1 ∩ A 1 x, x ∈ D(A1 ∩ A
1 ),

whenever A ⊂ A1 = A1 and A ⊂ A 1 .
1 = A
 
If A is closable, then Γ A = Γ (A).
26 Symmetric Operators in Hilbert Spaces 269

Definition 26.15. Consider the subspace

D∗ := {v ∈ H : there exists h ∈ H such that (Ax, v) = (x, h) for all x ∈ D(A)} .

The operator A∗ with domain D(A∗ ) := D∗ and mapping A∗ v = h is called the adjoint
operator of A.

Exercise 26.4. Prove that A∗ exists as a unique linear operator.

Remark 26.16. The adjoint operator is maximal among all linear operators B (in the
sense that B ⊂ A∗ ) that satisfy

(Ax, y) = (x, By)

for all x ∈ D(A) and y ∈ D(B).

Example 26.17. Consider the operator

A f (x) := x−α f (x), α >0

in the Hilbert space H = L2 (0, 1). Let us define


D(A) := f ∈ L2 (0, 1) : f (x) = χn (x)g(x), g ∈ L2 for some n ∈ N ,

where 
0, 0 ≤ x ≤ 1/n,
χn (x) =
1, 1/n < x ≤ 1.

It is clear that D(A) = L2 (0, 1). For v ∈ D(A∗ ) we have

 1  1
−α
(A f , v) = x χn (x)g(x)v(x)dx = f (x)x−α v(x)dx = ( f , A∗ v).
0 0

We conclude that 
D(A∗ ) = v ∈ L2 : x−α v ∈ L2 .

Let us show that A is not closed. To see this, we take the sequence

xα , 1/n < x ≤ 1,
fn (x) =
0, 0 ≤ x ≤ 1/n.

Then fn ∈ D(A) and 

1, 1/n < x ≤ 1,
A fn (x) =
0, 0 ≤ x ≤ 1/n.
270 Part III: Operator Theory and Integral Equations

If we assume that A = A, then

⎧
⎪ 
⎨ fn ∈ D(A), xα ∈ D(A),
f n → xα , ⇒
⎪
⎩ 1 = Axα .
A fn → 1

But xα ∈ / D(A). This contradiction shows that A is not closed. It is not bounded
either, since α > 0.
Theorem 26.18. Let A be a linear and densely defined operator. Then
(1) A∗ = A∗ .
∗ ∗∗ ∗ ∗
 if∗ D(A∗ ) = H. In this case A := (A ) = A.
(2) A is closable if and only
(3) If A is closable, then A = A .
Proof. (1) Let us define in H ×H the linear and bounded operator V as the mapping

V : (u; v) → (v; −u).

It has the property V 2 = −I. The equality (Au, v) = (u, A∗ v) for u ∈ D(A) and
v ∈ D(A∗ ) can be rewritten as

(V (u; Au), (v; A∗ v))H×H = 0.

This implies that Γ (A∗ ) ⊥V Γ (A) and Γ (A∗ ) ⊥V Γ (A), which in turn means that
 ⊥
Γ (A∗ ) ⊂ V Γ (A) . Let us check that the criterion for closedness holds, i.e.,
⎧
⎪ ∗ 
⎨vn ∈ D(A ), v ∈ D(A∗ ),
vn → v, ⇒
⎪
⎩ ∗ y = A∗ v.
A vn → y

Indeed, for all u ∈ D(A) we have

(Au, vn ) → (Au, v).

On the other hand,

(Au, vn ) = (u, A∗ vn ) → (u, y).

Hence (Au, v) = (u, y). Thus v ∈ D(A∗ ) and y = A∗ v. This proves (1).
(2) Assume D(A∗ ) = H. Then A∗∗ exists and due to part (1) we may conclude that

⊥
Γ (A∗ ) ⊂ V Γ (A) .
26 Symmetric Operators in Hilbert Spaces 271

Then
V Γ (A) ⊂ Γ (A∗ )⊥ .

It follows that
Γ (A) ⊂ (−V Γ (A∗ ))⊥ ,

since V 2 = −I. Here

−V Γ (A∗ ) = {(−A∗ u; u), u ∈ D(A∗ )}.

Thus
(−V Γ (A∗ ))⊥ = {(e1 ; e2 )},

so that
(−A∗ u, e1 )H + (u, e2 )H = 0,

or
(A∗ u, e1 )H = (u, e2 )H .

Therefore, e1 ∈ D(A∗∗ ) and A∗∗ e1 = e2 . This shows that (e1 ; e2 ) ∈ Γ (A∗∗ ) and
hence
(−V Γ (A∗ ))⊥ ⊂ Γ (A∗∗ ).

Therefore,
Γ (A) ⊂ Γ (A∗∗ ),

which means that A ⊂ A∗∗ , and since A∗∗ is closed, A is closable and A ⊂ A∗∗ .
Let us show that in this case, in fact, A = A∗∗ . Indeed, if u ∈ D(A∗∗ ), then

(v, A∗∗ u) = (A∗ v, u), v ∈ D(A∗ ),

or
(u, A∗ v) = (A∗∗ u, v), v ∈ D(A∗ ),

or
(Au, v) = (A∗∗ u, v), v ∈ D(A∗ ).

Since D(A∗ ) = H, we obtain Au = A∗∗ u on D(A∗∗ ). It follows that A∗∗ ⊂ A and

furthermore A∗∗ ⊂ A. Hence A = A∗∗ .
This proves (2) in one direction. Let us assume now that A is closable (i.e., A
exists and is minimal among all closed extensions) but D(A∗ ) = H. Then there
exists u0 = 0 such that u0 ⊥D(A∗ ). So u0 ⊥D(A∗ ) also. Then

(u0 ; 0)⊥(v; A∗ v), v ∈ D(A∗ ).

272 Part III: Operator Theory and Integral Equations

It follows that
(u0 , v) = (0, A∗ v),

or
(A∗ v, 0) = (v, u0 ).

In part (1) it is shown that Γ (A)⊥(−V Γ (A∗ )). Then

Γ (A)⊥(−V Γ (A∗ ))

or
Γ (A)⊥(−V Γ (A∗ ))

since A exists. Since also (0; u0 )⊥(−V Γ (A∗ )) then (0; u0 ) ∈ Γ (A) i.e. 0 =
A(0) = u0 = 0. This contradiction proves (2).
(3) Since A is closable, (1) and (2) imply
 ∗
A∗ = A∗ = (A∗ )∗∗ = (A)∗∗∗ = (A∗∗ )∗ = A .

This completes the proof. 

Example 26.19. Consider the Hilbert space H = L2 (R) and the operator

Au(x) = (u, f0 )u0 (x),

where u0 ≡ 0, u0 ∈ L2 (R), is fixed and f0 = 0 is an arbitrary but fixed constant. We

consider A on the domain
  
D(A) = u ∈ L2 (R) : | f0 u(x)|dx < ∞ = L2 (R) ∩ L1 (R).
R

It is known that L2 (R) ∩ L1 (R) = L2 (R). Thus A is densely defined. Let v be an

element of D(A∗ ). Then
 
(Au, v) = ((u, f0 )u0 , v) = (u, f0 )(u0 , v) = u, (u0 , v) f0 = (u, (v, u0 ) f0 ) .

It means that
A∗ v = (v, u0 ) f0 .

But (v, u0 ) f0 must belong to L2 (R). Since (v, u0 ) f0 is a constant and f0 = 0, it fol-
lows that (v, u0 ) must be equal to 0. Thus

u0 ⊥D(A∗ ),

which implies that

26 Symmetric Operators in Hilbert Spaces 273

u0 ⊥D(A∗ ).

Since u0 = 0, we have D(A∗ ) = H. Thus A∗ exists but is not densely defined. So A

is not closable.

Exercise 26.5. Assume that A is closable. Prove that D(A) can be obtained as the
closure of D(A) by the norm
 1/2
Au2 + u2 .

Theorem 26.20 (Closed graph theorem). If A : H → H is a linear operator whose

graph Γ (A) is closed in H × H, then A is bounded.

Proof. As a closed subspace of the Hilbert space H × H, the graph Γ (A) is a Hilbert
space (see Exercise 26.5). Let us define the projection mappings P1 and P2 as fol-
lows:

P1 : Γ (A) → H, P1 (u, v) = u,
P2 : Γ (A) → H, P2 (u, v) = v.

Since A is linear, both P1 and P2 are linear. Moreover, P1 is injective and surjective
and P1 and P2 are continuous, since

P1 (u, v)H = uH ≤ uH + vH ,

P2 (u, v)H = vH ≤ uH + vH .

Hence P1 is a bijective continuous (bounded) linear map of Γ (A) onto H and has a
continuous (bounded) inverse, since it is open; see [5]. But at the same time,

Au = P2 (P1 )−1 (u), u ∈ H,

and therefore as a superposition of two bounded linear operators, A is also

bounded. 

Definition 26.21. An operator A : H → H with D(A) = H is called

(1) symmetric if A ⊂ A∗ ;
(2) self-adjoint if A = A∗ ;  
∗
(3) essentially self-adjoint if A = A.

Remark 26.22. A symmetric operator is always closable, and its closure is also
symmetric. Indeed, if A ⊂ A∗ , then D(A) ⊂ D(A∗ ). Hence

H = D(A) ⊂ D(A∗ ) ⊂ H
274 Part III: Operator Theory and Integral Equations

implies that D(A∗ ) = H. Therefore, A is closable. Since A is the smallest closed

extension of A, we have  ∗
A ⊂ A ⊂ A∗ = A ,

i.e., A is also symmetric.

Some properties of a symmetric operator A are as follows:

(1) A ⊂ A = A∗∗ ⊂ A∗ ,
(2) A = A = A∗∗ ⊂ A∗ if A is closed,
(3) A = A = A∗∗ = A∗ if A is self-adjoint,
(4) A ⊂ A = A∗∗ = A∗ if A is essentially self-adjoint.

Example 26.23. Consider the operator

d2
A :=
dx2

in the Hilbert space H = L2 (0, 1) with domain

D(A) = { f ∈ C2 [0, 1] : f (0) = f (1) = f  (0) = f  (1) = 0}.

It is clear that D(A) = L2 (0, 1) and A is not closed. Moreover, integration by parts
gives
(A f , g)L2 = ( f , Ag)L2

for every f ∈ D(A) and g ∈ W22 (0, 1). That is, A is symmetric such that A ⊂ A∗ and
D(A∗ ) = W22 (0, 1). As we know, A∗ = A∗ always. Now we will show that A is the
◦ ◦
same differential operator of order 2 with D(A) = W 22 (0, 1), where W 22 (0, 1) denotes
the closure of D(A) with respect to the norm of the Sobolev space W22 (0, 1). Indeed,
for every f ∈ D(A) we have

A f 2L2 +  f 2L2 ≤  f W
2
2
2

and
 1
2
 f W 2 2
2 = A f L2 +  f L2 + | f  |2 dx
2 0
 1
3 3
= A f 2L2 +  f 2L2 − f f  dx ≤ A f 2L2 +  f 2L2 .
0 2 2

This means that

A f 2L2 +  f 2L2   f W
2
2.
2

Exercise 26.5 gives now that

26 Symmetric Operators in Hilbert Spaces 275

◦
D(A) = W 22 (0, 1).

So we have finally
◦
D(A)  D(A) = W 22 (0, 1) = D(A∗∗ )  W22 (0, 1).

The closure A is symmetric but not self-adjoint, since

◦ ∗
W 22 (0, 1) = D(A) = D(A ) = D(A∗ ) = W22 (0, 1).

Theorem 26.24 (J. von Neumann). Assume that A ⊂ A∗ .

(1) If D(A) = H, then A = A∗ and A is bounded.
(2) If R(A) = H, then A = A∗ and A−1 exists and is bounded.
∗
(3) If A−1 exists, then A = A∗ if and only if A−1 = A−1 .

Proof. (1) Since A ⊂ A∗ , we have H = D(A) ⊂ D(A∗ ) ⊂ H and hence D(A) =

D(A∗ ) = H. Thus A = A∗ , and the Hellinger–Toeplitz theorem (Exercise 26.2)
says that A is bounded.
(2), (3) Let us assume that u0 ∈ D(A) and Au0 = 0. Then for all v ∈ D(A) we obtain
that
0 = (Au0 , v) = (u0 , Av).

This means that u0 ⊥H and therefore −1

 −1 ∗ u0 = 0. It follows that A ∗exists and
−1
D(A ) = R(A) = H. Hence A exists. Let us prove that (A )−1 exists
 ∗  ∗ 
too and (A∗ )−1 = A−1 . Indeed, if u ∈ D(A) and v ∈ D A−1 , then
 ∗
(u, v) = (A−1 Au, v) = (Au, A−1 v).

This equality implies that

 ∗
A−1 v ∈ D(A∗ )

and  ∗
A∗ A−1 v = v. (26.4)

Similarly, if u ∈ D(A−1 ) and v ∈ D(A∗ ), then

(u, v) = (AA−1 u, v) = (A−1 u, A∗ v)

and therefore  ∗ 
A∗ v ∈ D A−1

and
276 Part III: Operator Theory and Integral Equations
 ∗
A−1 A∗ v = v. (26.5)
 ∗
It follows from (26.4) and (26.5) that (A∗ )−1 exists and (A∗ )−1 = A−1 .
The boundedness of A−1 follows from part (1).

Exercise 26.6. Let A and B be injective operators. Prove that if A ⊂ B, then A−1 ⊂
B−1 .

Since A ⊂ A∗ , we have by Exercise 26.6 that

 ∗
A−1 ⊂ (A∗ )−1 = A−1 ,

i.e., −1 D(A−1 ) 
= H. We conclude that H = D(A−1 ) ⊂
 A−1 ∗is also symmetric. But −1 ∗
D A ⊂ H and hence D(A ) = D A−1 = H. Thus A−1 is self-adjoint
and bounded (Hellinger–Toeplitz theorem; see Exercise 26.2). Finally,
 ∗
A−1 = A−1 = (A∗ )−1

if and only if A = A∗ .
This completes the proof. 
Theorem 26.25 (Basic criterion of self-adjointness). If A ⊂ A∗ , then the following
statements are equivalent:
(1) A = A∗ .
(2) A = A and N(A∗ ± iI) = {0}.
(3) R(A ± iI) = H.

Proof. (1) ⇒ (2) Since A = A∗ , it follows that A is closed. Suppose that u0 ∈

N(A∗ − iI), i.e., u0 ∈ D(A∗ ) = D(A) and Au0 = iu0 . Then

i(u0 , u0 ) = (iu0 , u0 ) = (Au0 , u0 ) = (u0 , Au0 ) = (u0 , iu0 ) = −i(u0 , u0 ).

This implies that u0 = 0, i.e., N(A∗ − iI) = {0}. The proof of N(A∗ + iI) = {0}
is left to the reader.
(2) ⇒ (3) Since A = A and N(A∗ ± iI) = {0}, it follows, for example, that
the equation A∗ u = −iu has only the trivial solution u = 0. This implies that
R(A − iI) = H. For otherwise, there exists u0 = 0 such that u0 ⊥R(A − iI). This
means that for all u ∈ D(A) we have

((A − iI)u, u0 ) = 0

and therefore u0 ∈ D(A∗ + iI) and (A∗ + iI)u0 = 0, or A∗ u0 = −iu0 , u0 = 0. This

contradiction proves that R(A − iI) = H. Next, since A is closed, Γ (A) is also
closed, and due to the fact that A is symmetric, we have
26 Symmetric Operators in Hilbert Spaces 277

(A − iI)u2 = ((A − iI)u, (A − iI)u)

= Au2 − i(u, Au) + i(Au, u) + u2 = Au2 + u2

for u ∈ D(A). It follows that if (A − iI)un → v0 , then Aun and un are conver-
gent, i.e., Aun → v0 , un → u0 , and un ∈ D(A). The closedness of A implies that
u0 ∈ D(A) and v0 = Au0 , i.e., (A − iI)un → Au0 − iu0 = v0 . This means that
R(A − iI) is a closed set, i.e., R(A − iI) = R(A − iI) = H. The proof of R(A + iI) =
H is left to the reader.
(3) ⇒ (1) Assume that R(A ± iI) = H. Since A ⊂ A∗ , it suffices to show that
D(A∗ ) ⊂ D(A). For every u ∈ D(A∗ ) we have (A∗ − iI)u ∈ H. Part (3) implies
that there exists v0 ∈ D(A) such that

(A − iI)v0 = (A∗ − iI)u.

It is clear that u − v0 ∈ D(A∗ ) (since A ⊂ A∗ ) and

(A∗ − iI)(u − v0 ) = (A∗ − iI)u − (A∗ − iI)v0 = (A∗ − iI)u − (A − iI)v0

= (A − iI)v0 − (A − iI)v0 = 0.

Hence u − v0 ∈ N(A∗ − iI).

Exercise 26.7. Let A be a linear and densely defined operator in the Hilbert space
H. Prove that
H = N(A∗ ) ⊕ R(A).

By this exercise we know that

H = N(A∗ − iI) ⊕ R(A + iI).

But in our case R(A + iI) = H. Hence N(A∗ − iI) = {0} and therefore u = v0 . Thus
D(A) = D(A∗ ).
This concludes the proof. 
Example 26.26. Assume that an operator A is symmetric and closed in a Hilbert
space H. Consider the operator A∗ A on the domain

D(A∗ A) = { f ∈ D(A) : A f ∈ D(A∗ )}.

This operator is self-adjoint. Indeed, since (A∗ A)∗ = A∗ A∗∗ = A∗ A = A∗ A, we have

that A∗ A is symmetric. At the same time, for all f ∈ D(A), we have

(A∗ A f , f ) = (A f , A∗∗ f ) = (A f , A f ) = A f 2H .

This fact leads to R(A∗ A ± iI) = H, since A∗ A ± iI is invertible in this case. Thus,
Theorem 26.25 gives us that A∗ A is self-adjoint. The same is true for the operator
278 Part III: Operator Theory and Integral Equations

AA∗ on the domain

D(AA∗ ) = { f ∈ D(A∗ ) : A∗ f ∈ D(A)}.

It is clear that in general,

AA∗ = A∗ A.

If equality holds here, the operator A is said to be normal.

d
Exercise 26.8. Let H = L2 (0, 1) and A := i .
dx
(1) Prove that A is closed and symmetric on the domain
◦
D(A) = { f ∈ L2 (0, 1) : f  ∈ L2 (0, 1), f (0) = f (1) = 0} ≡ W 12 (0, 1).

(2) Prove that A is self-adjoint on the domain

!
Dγ (A) = f ∈ L2 (0, 1) : f  ∈ L2 (0, 1), f (0) = f (1)eiγ , γ ∈ R .
Chapter 27
John von Neumann’s Spectral Theorem

Definition 27.1. A bounded linear operator P on a Hilbert space H that is self-

adjoint and idempotent, i.e., P2 = P, is called an orthogonal projection operator or a
projector.
Proposition 27.2. Let P be a projector. Then
(1) P = 1 if P = 0.
(2) P is a projector if and only if P⊥ := I − P is a projector.
(3) H = R(P) ⊕ R(P⊥ ), P|R(P) = I, and P|R(P⊥ ) = 0.
(4) There is a one-to-one correspondence between projectors on H and closed lin-
ear subspaces of H. More precisely, if M ⊂ H is a closed linear subspace, then
there exists a projector PM : H → M, and conversely, if P : H → H is a projector,
then R(P) is a closed linear subspace.
(5) If {e j }Nj=1 , N ≤ ∞ is an orthonormal system, then

N
PN x := ∑ (x, e j )e j , x ∈ H,
j=1

is a projector.
Proof. (1) Since P = P∗ and P = P2 , we have P = P∗ P. Hence P = P∗ P. But
P∗ P = P2 . Indeed,

P∗ P ≤ P∗  P ≤ P2

and

P2 = sup Px2 = sup (Px, Px)

x=1 x=1

= sup (P∗ Px, x) ≤ sup P∗ Px = P∗ P .

x=1 x=1

c Springer International Publishing AG 2017 279

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 27
280 Part III: Operator Theory and Integral Equations

Therefore, P = P2 , or P = 1, if P = 0.

(2) Since P is linear and bounded, the same is true about I − P. Moreover,

(I − P)∗ = I − P∗ = I − P

and
(I − P)2 = (I − P)(I − P) = I − 2P + P2 = I − P.

(3) It follows immediately from I = P + P⊥ that every x ∈ H is of the form u + v,

 ⊥
where u ∈ R(P) and v ∈ R(P⊥ ). Let us prove that R(P) = R(P⊥ ) . First
 ⊥
assume that w ∈ R(P⊥ ) , i.e., (w, (I − P)x) = 0 for all x ∈ H. This is equiva-
lent to
(w, x) = (w, Px) = (Pw, x), x ∈ H,
 ⊥
or Pw = w. Hence w ∈ R(P), and so we have proved that R(P⊥ ) ⊂ R(P).
For the opposite embedding we let w ∈ R(P). Then there exists xw ∈ H such
that w = Pxw . If z ∈ R(P⊥ ), then z = P⊥ xz = (I − P)xz for some xz ∈ H. Thus

(w, z) = (Pxw , (I − P)xz ) = (Pxw , xz ) − (Pxw , Pxz ) = 0,

 ⊥
since P is a projector. Therefore, w ∈ R(P⊥ ) , and we may conclude that
 ⊥
R(P) = R(P⊥ ) . This fact allows us to conclude that R(P) = R(P) and H =
R(P) ⊕ R(P⊥ ). Moreover, it is easy to check from the definition that P|R(P) = I
and P|R(P⊥ ) = 0.
(4) If M ⊂ H is a closed subspace, then Theorem 25.12 implies that x = u + v ∈ H,
where u ∈ M and v ∈ M ⊥ . In that case, let us define PM : H → M as

PM x = u.

It is clear that PM2 x = PM u = u = PM x, i.e., PM2 = PM . Moreover, if y ∈ H, then

y = u1 + v1 , u1 ∈ M, v1 ∈ M ⊥ and

(PM x, y) = (u, u1 + v1 ) = (u, u1 ) = (u + v, u1 ) = (u + v, PM y) = (x, PM y),

i.e., PM∗ = PM . Hence PM is a projector. If P is a projector, then we know from

part (3) that M := R(P) is a closed subspace of H.
(5) Let us assume that N = ∞. Define M as
 
∞ ∞
M := x∈H :x= ∑ c j e j , ∑ |c j |
2
<∞ .
j=1 j=1
27 John von Neumann’s Spectral Theorem 281

Then M is a closed subspace of H. If we define a linear operator PM as

∞
PM x := ∑ (x, e j )e j , x ∈ H,
j=1

then by Bessel’s inequality we obtain that PM x ∈ M and

PM x ≤ x .

This means that PM is a bounded linear operator into M. But PM e j = e j and thus
PM2 x = PM x for all x ∈ H. Next, for all x, y ∈ H we have
 
∞ ∞ ∞
(PM x, y) = ∑ (x, e j )e j , y = ∑ (x, e j ) (e j , y) = ∑ (x, (y, e j )e j )
j=1 j=1 j=1
 
∞
= x, ∑ (y, e j )e j = (x, PM y),
j=1

i.e., PM∗ = PM . The case of finite N requires no convergence questions and is left
to the reader.
This completes the proof. 

Definition 27.3. A bounded linear operator A on a Hilbert space H is said to be

smaller than or equal to a bounded operator B on H if

(Ax, x) ≤ (Bx, x), x ∈ H.

We denote this fact by A ≤ B. The operator A is nonnegative if A ≥ 0; A is positive,

denoted by A > 0, if A ≥ c0 I for some c0 > 0.

Remark 27.4. In the framework of this definition, (Ax, x) and (Bx, x) must be real
for all x ∈ H.

Proposition 27.5. For two projectors P and Q the following statements are equiv-
alent:
(1) P ≤ Q.
(2) Px ≤ Qx for all x ∈ H.
(3) R(P) ⊂ R(Q).
(4) P = PQ = QP.

Proof. (1) ⇒ (2) Follows directly from (Px, x) = (P2 x, x) = (Px, Px) = Px2 .
(3) ⇒ (4) Assume R(P) ⊂ R(Q). Then QPx = Px or QP = P. Conversely, if QP = P,
then clearly R(P) ⊂ R(Q). Finally, P = QP = P∗ = (QP)∗ = P∗ Q∗ = PQ.
282 Part III: Operator Theory and Integral Equations

(2) ⇒ (4) If (4) holds, then Px = PQx and Px = PQx ≤ Qx for all x ∈ H.
Conversely, if Px ≤ Qx, then Px = QPx + Q⊥ Px implies that
 2
 
Px2 = QPx2 + Q⊥ Px ≤ QPx2 .

Hence  
 ⊥ 2
Q Px = 0,

i.e., Q⊥ Px = 0 for all x ∈ H. Hence P = QP = PQ.

This completes the proof. 

Exercise 27.1. Let {Pj }∞j=1 be a sequence of projectors with Pj ≤ Pj+1 for each
j = 1, 2, . . .. Prove that lim j→∞ Pj := P exists and that P is a projector.

Definition 27.6. A linear map A : H → H with the property

Ax = x , x ∈ H,

is called an isometry.

Exercise 27.2. Prove that

(1) A is an isometry if and only if A∗ A = I.
(2) Every isometry A has an inverse A−1 : R(A) → H and A−1 = A∗ |R(A) .
(3) If A is an isometry, then AA∗ is a projector on R(A).

Definition 27.7. A surjective isometry U : H → H is called a unitary operator.

Remark 27.8. It follows that U is unitary if and only if it is surjective and U ∗U =

UU ∗ = I, i.e., (Ux,Uy) = (x, y) for all x, y ∈ H.

Definition 27.9. Let H be a Hilbert space. The family of operators {Eλ }∞

λ =−∞ is
called a spectral family if the following conditions are satisfied:
(1) Eλ is a projector for all λ ∈ R.
(2) Eλ ≤ Eμ for all λ < μ .
(3) {Eλ } is right continuous with respect to the strong operator topology, i.e.,

lim Es x − Et x = 0
s→t+

for all x ∈ H.
27 John von Neumann’s Spectral Theorem 283

(4) {Eλ } is normalized as follows:

lim Eλ x = 0, lim Eλ x = x

λ →−∞ λ →+∞

for all x ∈ H. The latter condition can also be formulated as

lim Eλ x − x = 0.
λ →+∞

Remark 27.10. It follows from the previous definition and Proposition 27.5 that

Eλ Eμ = Emin{λ ,μ } .

Proposition 27.11. For every fixed x, y ∈ H, (Eλ x, y) is a function of bounded vari-

ation with respect to λ ∈ R.

Proof. Let us define

E(α , β ] := Eβ − Eα , α < β.

Then E(α , β ] is a projector. Indeed,

E(α , β ]∗ = Eβ∗ − Eα∗ = Eβ − Eα = E(α , β ],

i.e., E(α , β ] is self-adjoint. It is also idempotent due to

(E(α , β ])2 = (Eβ − Eα )(Eβ − Eα ) = Eβ2 − Eα Eβ − Eβ Eα + Eα2

= Eβ − Eα − Eα + Eα = E(α , β ].

Another property is that

E(α1 , β1 ]x⊥E(α , β ]y, x, y ∈ H

if β1 ≤ α or β ≤ α1 . To see this for β1 ≤ α we calculate

 
(E(α1 , β1 ]x, E(α , β ]y) = Eβ1 x − Eα1 x, Eβ y − Eα y
     
= Eβ1 x, Eβ y − Eα1 x, Eβ y − Eβ1 x, Eα y + (Eα1 x, Eα y)
   
= x, Eβ1 y − (x, Eα1 y) − x, Eβ1 y + (x, Eα1 y) = 0.

Let now
λ0 < λ1 < · · · < λn .
284 Part III: Operator Theory and Integral Equations

Then
n n  
∑ |(Eλ j x, y) − (Eλ j−1 x, y)| = ∑ E(λ j−1 , λ j ]x, y
j=1 j=1

n 
= ∑ E(λ j−1 , λ j ]x, E(λ j−1 , λ j ]y
j=1
n   
≤ ∑ E(λ j−1 , λ j ]x E(λ j−1 , λ j ]y
j=1
 1/2  1/2
n  2 
n 2
≤ ∑ E(λ j−1 , λ j ]x ∑ E(λ j−1 , λ j ]y
j=1 j=1
  
 n  n 
  
=  ∑ E(λ j−1 , λ j ]x  ∑ E(λ j−1 , λ j ]y
 j=1   j=1 
= E(λ0 , λn ]x E(λ0 , λn ]y ≤ x y .

Here we have made use of orthogonality, normalization, and the

Cauchy–Bunyakovsky–Schwarz inequality. 

By Proposition 27.11 we can define a Stieltjes integral. Indeed, for every continuous
function f (λ ) we may conclude the equality of limits
 
n   n
lim ∑
Δ →0 j=1
f (λ j∗ ) E(λ j−1 , λ j ]x, y = lim
Δ →0
∑ f (λ j∗ )E(λ j−1 , λ j ]x, y ,
j=1

where λ j∗ ∈ [λ j−1 , λ j ], α = λ0 < λ1 < · · · < λn = β , and Δ = max1≤ j≤n |λ j−1 − λ j |

exists, and by definition this limit is

β
f (λ )d(Eλ x, y), x, y ∈ H.
α

It can be shown that this is equivalent to the existence of the limit in H

n
lim
Δ →0
∑ f (λ j∗ )E(λ j−1 , λ j ]x,
j=1

which we denote by

β
f (λ )dEλ x.
α

Thus
β 
β

f (λ )d(Eλ x, y) = f (λ )dEλ x, y , x, y ∈ H.
α α
27 John von Neumann’s Spectral Theorem 285

For the spectral representation of self-adjoint operators one needs integrals not only
over finite intervals but also over the whole line, which is naturally defined as the
limit

∞
β 
∞ 
f (λ )d(Eλ x, y) = lim f (λ )d(Eλ x, y) = f (λ )dEλ x, y
−∞ α →−∞ α −∞
β →∞

if it exists. Deriving first some basic properties of the integral just defined, one can
check that

∞
β
f (λ )d(Eλ Eβ x, y) = f (λ )d(Eλ x, y)
−∞ −∞

β
:= lim f (λ )d(Eλ x, y), x, y ∈ H.
α →−∞ α

Theorem 27.12. Let {Eλ }∞ λ =−∞ be a spectral family on a Hilbert space H and let
f be a real-valued continuous function on the line. Define

∞

D := x ∈ H : | f (λ )| d(Eλ x, x) < ∞
2
−∞

 ∞ 
(or D := x ∈ H : −∞ f (λ )dEλ x exists ). Let us define on this domain an operator
A as
∞
(Ax, y) = f (λ )d(Eλ x, y), x ∈ D(A) := D, y ∈ H
−∞
∞
(or Ax = −∞ f (λ )dEλ x, x ∈ D(A)). Then A is self-adjoint and satisfies

E(α , β ]A ⊂ AE(α , β ], α < β.

Proof. It can be shown that the integral

∞
f (λ )d (Eλ x, y)
−∞

exists for x ∈ D and y ∈ H. Thus (Ax, y) is well defined. Let v be an element of H and
let ε > 0. Then by normalization, there exist α < −R and β > R with R sufficiently
large such that
   
v − E(α , β ]v = v − Eβ v + Eα v ≤ (I − Eβ )v + Eα v < ε .
286 Part III: Operator Theory and Integral Equations

On the other hand,

∞
∞
| f (λ )|2 d(Eλ E(α , β ]v, E(α , β ]v) = | f (λ )|2 d(Eλ E(α , β ]v, v)
−∞ −∞

∞
∞
= | f (λ )|2 d(Eλ Eβ v, v) − | f (λ )|2 d(Eλ Eα v, v)
−∞ −∞

β
α
= | f (λ )|2 d(Eλ v, v) − | f (λ )|2 d(Eλ v, v)
−∞ −∞

β
= | f (λ )|2 d(Eλ v, v) < ∞.
α

These two facts mean that E(α , β ]v ∈ D and D = H. Since f (λ ) = f (λ ), it follows

that A is symmetric. Indeed,

∞
β
(Ax, y) = f (λ )d(Eλ x, y) = lim f (λ )d(Eλ x, y)
−∞ α →−∞ α
β →∞

β 
β

= lim f (λ )d(x, Eλ y) = lim x, f (λ )dEλ y
α →−∞ α α →−∞ α
β →∞ β →∞
⎛ ⎞

β
= ⎝x, lim f (λ )dEλ y⎠ = (x, Ay).
α →−∞ α
β →∞

In order to prove that A = A∗ , it remains to show that D(A∗ ) ⊂ D(A). Let u ∈ D(A∗ ).
Then
β
(E(α , β ]z, A∗ u) = (AE(α , β ]z, u) = f (λ )d(Eλ z, u)
α

for all z ∈ H. This equality implies that

β
∞
(z, A∗ u) = lim f (λ )d(Eλ z, u) = f (λ )d(Eλ z, u)
α →−∞ α −∞
β →∞

∞
∞
= f (λ )d(z, Eλ u) = f (λ )d(Eλ u, z) = (Au, z) = (z, Au),
−∞ −∞

where the integral exists because (z, A∗ u) exists. Hence u ∈ D(A) and A∗ u = Au. For
the second claim we first calculate
27 John von Neumann’s Spectral Theorem 287

   
∞
E(α , β ]Ax = Eβ − Eα Ax = Eβ − Eα f (λ )dEλ x
−∞

∞
∞
= f (λ )dEλ Eβ x − f (λ )dEλ Eα x
−∞ −∞

β
α
= f (λ )dEλ x − f (λ )dEλ x
−∞ −∞

β
∞
 
= f (λ )dEλ x = f (λ )dEλ Eβ − Eα x
α −∞
 
= A Eβ − Eα x = AE(α , β ]x

for all x ∈ D(A). Since the left-hand side is defined on D(A) and the right-hand side
on all of H, the latter is an extension of the former. 
Exercise 27.3. Let A be as in Theorem 27.12. Prove that

∞
Au2 = | f (λ )|2 d(Eλ u, u)
−∞

if u ∈ D(A).
Exercise 27.4. Let H = L2 (R) and Au(t) = tu(t), t ∈ R. Define D(A) on which
A = A∗ and evaluate the spectral family {Eλ }∞
λ =−∞ .

Theorem 27.13 (John von Neumann’s spectral theorem). Every self-adjoint

operator A on a Hilbert space H has a unique spectral representation, i.e., there
is a unique spectral family {Eλ }∞
λ =−∞ such that

∞
Ax = λ dEλ x, x ∈ D(A)
−∞
∞
(i.e., (Ax, y) = −∞ λ d(Eλ x, y), x ∈ D(A), y ∈ H), where D(A) is defined as

∞ 
D(A) = x ∈ H : λ d(Eλ x, x) < ∞ .
2
−∞

Proof. First we assume that this theorem holds when A is bounded, that is, that there
is a unique spectral family {Fμ }∞
μ =−∞ such that

∞
Au = μ dFμ u, u ∈ H,
−∞

since D(A) = H in this case. But Fμ ≡ 0 for μ < m and Fμ ≡ I for μ > M, where

m = inf (Ax, x), M = sup (Ax, x).

x=1 x=1
288 Part III: Operator Theory and Integral Equations

The spectral representation therefore can be written in the form

M
Au = μ dFμ u, u ∈ H.
m

Let us consider now an unbounded operator that is semibounded from below, i.e.,

(Au, u) ≥ m0 (u, u), u ∈ D(A)

with some constant m0 . We assume without loss of generality that (Au,u) ≥ (u, u).
This condition implies that A−1 exists, it is defined over all of H, and A−1  ≤ 1.
Indeed, A−1 exists and is bounded because Au = 0 if and only if u = 0. The norm
estimate follows from
 2
(v, A−1 v) ≥ A−1 v , v ∈ D(A−1 ).

Since A−1 is bounded, D(A−1 ) is a closed subspace in H. The self-adjointness of A

 ∗
means that A−1 = A−1 . Therefore, A−1 is closed and D(A−1 ) = H, i.e., A−1 is
densely defined. Therefore, D(A−1 ) = H and R(A) = H. Since

0 ≤ (A−1 v, v) ≤ v2 , v ∈ H,

we may conclude in this case that m ≥ 0, M ≤ 1, and

1
A−1 v = μ dFμ v, v ∈ H,
0

where {Fμ } is the spectral family of A−1 . Let us note that F1 = I and F0 = 0, which
follows from the spectral theorem for bounded operators and from the fact that
A−1 v = 0 if and only if v = 0. Next, let us define the operator Bε , ε > 0, as

1
1
Bε u := dFμ u, u ∈ D(A).
ε μ

For every v ∈ H we have

1
1 
1 
1 1
Bε A−1 v = dFμ (A−1 v) = dFμ λ dFλ v
ε μ ε μ 0

1 
1 
1 
μ 
1
1 1 1
= d λ d(Fμ Fλ v) = d λ dFλ v = μ dFμ v
ε μ 0 ε μ ε ε μ

1
= dFμ v = F1 v − Fε v = v − Fε v.
ε
27 John von Neumann’s Spectral Theorem 289

Since every spectral family is right continuous, it follows that

lim Bε A−1 v = v
ε →0+

exists. For every u ∈ D(A) we have similarly

1
1 
μ

−1 1
A Bε u = μ dFμ (Bε u) = μd dF u = u − Fε u,
0 ε ε λ λ

and hence
lim A−1 Bε u = u
ε →0+

exists. These two equalities mean that

 −1
lim Bε = A−1 =A
ε →0+

exists and the spectral representation

1
1
1 1
A= dFμ = lim dFμ
0 μ ε →0+ ε μ

holds. If we define Eλ = I − F 1 , 1 ≤ λ < ∞, then

λ

1
∞
1
A=− dE 1 = λ dEλ .
0 μ μ 1

Exercise 27.5. Prove that this {Eλ } is a spectral family which is left-continuous.

The domain D(A) can be characterized as

∞

1

1
D(A) = u ∈ H : λ 2 d(Eλ u, u) < ∞ = u ∈ H : d(Fμ u, u) < ∞ .
1 0 μ2

This proves the theorem for self-adjoint operators that are semibounded from below.
For bounded operators we will only sketch the proof.
Step 1. If A = A∗ and A is bounded, then we can define

pN (A) := a0 I + a1 A + · · · + aN AN , N ∈ N,

where a j ∈ R for j = 0, 1, . . . , N. Then pN (A) is also self-adjoint and bounded

with
pN (A) ≤ sup |pN (t)|.
|t|≤A
290 Part III: Operator Theory and Integral Equations

Step 2. For every continuous real-valued function f on [m, M], where m and M are
as above, we can define f (A) as an approximation by pN (A), i.e., we can prove
that for every ε > 0 there exists pN (A) such that

 f (A) − pN (A) < ε .

Step 3. For every u, v ∈ H let us define the functional L as

L( f ) := ( f (A)u, v).

Then
|L( f )| = |( f (A)u, v)| ≤  f (A) u v ,

that is, L( f ) is a bounded linear functional on C[m, M].

Step 4. (Riesz’s theorem) A continuous positive linear functional L on C[a, b] can
be represented in the form

b
L( f ) = f (x)dν (x),
a

where ν is a measure that satisfies the conditions

(1) L( f ) ≥ 0 for f ≥ 0;
(2) |L( f )| ≤ ν (K)  f K , where K ⊂ [a, b] is compact and

 f K = max | f (x)|.
x∈K

Step 5. It follows from Step 4 that

M
(Au, v) = λ dν (λ ; u, v).
m

Step 6. It is possible to prove that ν (λ ; u, v) is a self-adjoint sesquilinear form,

from which we conclude that there exists a self-adjoint and bounded operator Eλ
such that
ν (λ ; u, v) = (Eλ u, v).

This operator is idempotent, and we may define Eλ ≡ 0 for λ < m and Eλ ≡ I

for λ ≥ M. Thus {Eλ }∞ λ =−∞ is the required spectral family, and the theorem is
proved. See [4] for an alternative proof of this theorem. 
Let A : H → H be a self-adjoint operator in a Hilbert space H. Then by von
Neumann’s spectral theorem we can write

∞
Au = λ dEλ u, u ∈ D(A).
−∞
27 John von Neumann’s Spectral Theorem 291

For every continuous function f we can define

∞

D f := u ∈ H : | f (λ )| d(Eλ u, u) < ∞ .
2
−∞

This set is a linear subspace of H. For every u ∈ D f and v ∈ H let us define the linear
functional
∞

∞

L(v) := f (λ )d(Eλ u, v) = f (λ )dEλ u, v .
−∞ −∞

This functional is continuous because it is bounded. Indeed,


∞ 2
∞
 

|L(v)| ≤ 
2
f (λ )dEλ u v 2
= | f (λ )|2 d(Eλ u, u) v2 = c(u) v2 .
−∞
 −∞

By the Riesz–Fréchet theorem this functional can be expressed in the form of an

inner product, i.e., there exists z ∈ H such that

∞
f (λ )d(Eλ u, v) = (z, v), v ∈ H.
−∞

We set
z := f (A)u, u ∈ Df ,

i.e.,
∞
( f (A)u, v) = f (λ )d(Eλ u, v).
−∞

Remark 27.14. Since in general f is not real-valued, f (A) is not a self-adjoint oper-
ator in general.

Example 27.15. Consider

λ −i
f (λ ) = , λ ∈ R,
λ +i

and a self-adjoint operator A with spectral family Eλ . Define

∞
λ −i
UA := f (A) = dEλ .
−∞ λ +i

The operator UA is called the Cayley transform of A. Since | f (λ )| = 1, we have

D f = D(UA ) = H and
292 Part III: Operator Theory and Integral Equations

∞
β
UA u2 = | f (λ )|2 d(Eλ u, u) = lim d(Eλ u, u)
−∞ α →−∞ α
β →∞
   2 
= lim (Eβ u, u) − (Eα u, u) = lim Eβ u − Eα u2 = u2
α →−∞ α →−∞
β →∞ β →∞

by normalization of {Eλ }. Hence UA is an isometry. There is a one-to-one corre-

spondence between self-adjoint operators and their Cayley transforms. Indeed,

UA = (A − iI)(A + iI)−1

is equivalent to 
I −UA = 2i(A + iI)−1 ,
I +UA = 2A(A + iI)−1 ,

or
A = i(I +UA )(I −UA )−1 .

Example 27.16. Consider

1
f (λ ) = , λ ∈ R, z ∈ C, Im z = 0.
λ −z

Define
∞
1
Rz := (A − zI)−1 = dE .
−∞ λ −z λ

The operator Rz is called the resolvent of A. Since

1 1

λ − z ≤ |Im z|

for all λ ∈ R, we have that Rz is bounded and defined on the whole of H.

Example 27.17. Suppose that K(x, y) ∈ L2 (Ω × Ω ). Define an integral operator on

L2 (Ω ) as

A f (x) = K(x, y) f (y)dy.

Ω

Then

A∗ f (x) = K(y, x) f (y)dy

Ω

and therefore


A∗ A f (x) = K(y, z)K(y, x)dy f (z)dz.
Ω Ω
27 John von Neumann’s Spectral Theorem 293

As we know from Example 26.26, A∗ A is self-adjoint on L2 (Ω ). This fact can also

be checked directly, since

K(y, z)K(y, x)dy = K(y, x)K(y, z)dy.

Ω Ω

Von Neumann’s spectral theorem gives us for this operator and for all s ≥ 0 that

A2
L2 →L2
(A∗ A)s = λ s dEλ ,
0

since A∗ A is positive and bounded by A2L2 →L2 .

Exercise 27.6. Let A = A∗ with spectral family Eλ . Let u ∈ D( f (A)) and v ∈

D(g(A)). Prove that

∞
( f (A)u, g(A)v) = f (λ )g(λ )d(Eλ u, v).
−∞

Exercise 27.7. Let A = A∗ with spectral family Eλ . Let u ∈ D( f (A)). Prove that
f (A)u ∈ D(g(A)) if and only if u ∈ D((g f )(A)) and that

∞
(g f )(A)u = g(λ ) f (λ )dEλ u.
−∞

Remark 27.18. It follows from Exercise 27.7 that

(g f )(A) = ( f g)(A)

on the domain D (( f g)(A)) ∩ D ((g f )(A)).

Chapter 28
Spectra of Self-Adjoint Operators

Definition 28.1. Given a linear operator A on a Hilbert space H with domain D(A),
D(A) = H, the set
 
ρ (A) = z ∈ C : (A − zI)−1 exists as a bounded operator from H to D(A)

is called the resolvent set of A. Its complement

σ (A) = C \ ρ (A)

is called the spectrum of A.

Theorem 28.2.
(1) If A = A then the resolvent set is open and the resolvent operator Rz :=
(A − zI)−1 is an analytic function from ρ (A) to B(H; H), the set of all linear
bounded operators in H. Furthermore, the resolvent identity

Rz − Rξ = (z − ξ )Rz Rξ , z, ξ ∈ ρ (A)

holds and Rz = (Rz )2 .

(2) If A = A∗ then z ∈ ρ (A) if and only if there exists Cz > 0 such that

(A − zI)u ≥ Cz u

for all u ∈ D(A).

Proof. (1) Assume that z0 ∈ ρ (A). Then Rz0 is a bounded linear operator from H to
 −1
D(A) and thus r := Rz0  > 0. Let us define for |z − z0 | < r the operator

Gz0 := (z − z0 )Rz0 .

c Springer International Publishing AG 2017 295

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 28
296 Part III: Operator Theory and Integral Equations
 
Then Gz0 is bounded with Gz0  < 1. Hence it defines the operator

 −1 ∞  j
I − Gz0 = ∑ Gz0 ,
j=0

because this Neumann series converges. But for |z − z0 | < r we have

A − zI = (A − z0 I)(I − Gz0 ),

or
(A − zI)−1 = (I − Gz0 )−1 Rz0 .

Hence Rz exists with D(Rz ) = H and is bounded. It remains to show that R(Rz ) ⊂
D(A). For x ∈ H we know that

y := (A − zI)−1 x ∈ H.

We claim that y ∈ D(A). Indeed,

∞   j+1
y = (A − zI)−1 x = (I − Gz0 )−1 Rz0 x = ∑ (z − z0 ) j Rz0 x
j=0
n   j+1
= lim
n→∞
∑ (z − z0 ) j Rz0 x.
j=0

It follows from this representation that Rz = (A − zI)−1 is an analytic function

from ρ (A) to B(H; H). Next we define
n   j+1
sn x := ∑ (z − z0 ) j Rz0 x.
j=0

It is clear that sn x ∈ D(A) and that limn→∞ sn x = y. Moreover,

lim (A − zI)sn x = x.
n→∞

Writing yn := sn x we conclude from the criterion for closedness that

⎧
⎪ 
⎨yn ∈ D(A), y ∈ D(A),
yn → y, ⇒
⎪
⎩ x = (A − zI)y.
(A − zI)yn → x

Hence y = (A − zI)−1 x ∈ D(A), and therefore ρ (A) is open. The resolvent iden-
tity is proved by a straightforward calculation:
28 Spectra of Self-Adjoint Operators 297

Rz − Rξ = Rz (A − ξ I)Rξ − Rz (A − zI)Rξ = Rz [(A − ξ I) − (A − zI)] Rξ

= (z − ξ )Rz Rξ .

Finally, the limit

Rz − Rξ
lim = lim Rz Rξ = (Rz )2
z→ξ z−ξ z→ξ

exists, and hence Rz = (Rz )2 exists, which proves this part.
(2) Assume that A = A∗ . If z ∈ ρ (A), then by definition Rz maps from H to D(A).
Hence there exists Mz > 0 such that

Rz v ≤ Mz v , v ∈ H.

Since u = Rz (A − zI)u for all u ∈ D(A), we get

u ≤ Mz (A − zI)u , u ∈ D(A).

This is equivalent to

1
(A − zI)u ≥ u , u ∈ D(A).
Mz

Conversely, if there exists Cz > 0 such that

(A − zI)u ≥ Cz u , u ∈ D(A),

then (A − zI)−1 is bounded. Since A is self-adjoint, (A − zI)−1 is defined over all

of H. Indeed, if R(A − zI) = H, then there exists v0 = 0 such that v0 ⊥R(A − zI).
This means that
(v0 , (A − zI)u) = 0, u ∈ D(A),

or
(Au, v0 ) = (zu, v0 ),

or
(u, A∗ v0 ) = (u, zv0 ).

Thus v0 ∈ D(A∗ ) and A∗ v0 = zv0 . Since A = A∗ , it follows that v0 ∈ D(A) and

Av0 = zv0 , or
(A − zI)v0 = 0.

It is easy to check that (A − zI)u2 = (A − zI)u2 for all u ∈ D(A). Therefore,

(A − zI)v0  = (A − zI)v0  ≥ Cz v0  .

298 Part III: Operator Theory and Integral Equations
 
Hence v0 = 0 and D (A − zI)−1 = R(A − zI) = H. This means that z ∈ ρ (A).

Corollary 28.3. If A = A∗ , then σ (A) = 0,
/ σ (A) = σ (A) and σ (A) ⊂ R.
Proof. If z = α + iβ ∈ C with Im z = β = 0, then

(A − zI)x2 = (A − α I)x − iβ x2 = (A − α I)x2 + |β |2 x2 ≥ |β |2 x2 .

This implies (see part (2) of Theorem 28.2) that z ∈ ρ (A), which means that
σ (A) ⊂ R. Since A = A∗ and is therefore closed, the spectrum σ (A) is closed as
the complement of an open set (see part (1) of Theorem 28.2).
It remains to prove that σ (A) = 0.
/ Assume to the contrary that σ (A) = 0. / Then
the resolvent Rz is an entire analytic function. Let us prove that Rz  is uniformly
bounded with respect to z ∈ C. We introduce the functional

Tz (y) := (Rz x, y), x = 1, y ∈ H.

Then Tz (y) is a linear functional on the Hilbert space H. Moreover, since Rz is

bounded for every (fixed) z ∈ C, it follows that

|Tz (y)| ≤ Rz x y ≤ Rz  y = Cz y .

Therefore, Tz (y) is continuous, i.e., {Tz , z ∈ C} is a pointwise bounded family of

continuous linear functionals. By the Banach–Steinhaus theorem (or the uniform
boundedness principle) we conclude that

sup Tz  = c0 < ∞.

z∈C

We therefore have

|Tz (y)| = |(Rz x, y)| ≤ c0 y , x = 1, z ∈ C,

which implies that Rz x ≤ c0 , i.e., Rz  ≤ c0 . By Liouville’s theorem we may

conclude now that Rz is constant with respect to z. But by von Neumann’s spectral
theorem,  ∞
1
Rz = dEλ ,
−∞ λ − z

where {Eλ } is the spectral family of A = A∗ . Due to the estimate

1
Rz  ≤ ,
| Im z|

we may conclude that Rz  → 0 as | Im z| → ∞. Hence Rz ≡ 0. This contradiction

completes the proof. 
28 Spectra of Self-Adjoint Operators 299

Exercise 28.1. Consider A = d

dx defined in L2 (0, 1) with domain

D(A) = {u ∈ W21 (0, 1) : u(1) = 0}.

Show that A = A∗ and σ (A) = 0.

/

Exercise 28.2. [Weyl’s criterion] Let A = A∗ . Prove that λ ∈ σ (A) if and only if
there exists xn ∈ D(A), xn  = 1, such that

lim (A − λ I)xn  = 0.

n→∞

Definition 28.4. Let us assume that A = A. The point spectrum σp (A) of A is the
set of eigenvalues of A, i.e.,

σp (A) = {λ ∈ σ (A) : N(A − λ I) = {0}} .

This means that (A − λ I)−1 does not exist, i.e., there exists a nontrivial u ∈ D(A)
such that Au = λ u. The complement σ (A) \ σp (A) is called the continuous spectrum
σc (A). The discrete spectrum is the set
 
σd (A) = λ ∈ σp (A) : dim N(A − λ I) < ∞ and λ is isolated in σ (A) .

The set σess (A) := σ (A) \ σd (A) is called the essential spectrum of A.

In the framework of this definition, the complex plane can be divided into regions
according to
C = ρ (A) ∪ σ (A),

σ (A) = σp (A) ∪ σc (A),

and
σ (A) = σd (A) ∪ σess (A),

with all the unions disjoint.

Remark 28.5. If A = A∗ , then

(1) λ ∈ σc (A) means that (A − λ I)−1 exists but is not bounded.
(2)

σess (A) = σc (A)

∪ {eigenvalues of infinite multiplicity and their accumulation points}
∪ {accumulation points of σd (A)}.

Exercise 28.3. Let A = A∗ and λ1 , λ2 ∈ σp (A). Prove that if λ1 = λ2 , then

300 Part III: Operator Theory and Integral Equations

N(A − λ1 I)⊥N(A − λ2 I).

Exercise 28.4. Let {e j }∞j=1 be an orthonormal basis in H and let {s j }∞j=1 ⊂ C be

some sequence. Introduce the set

∞
D= x∈H : ∑ |s j | |(x, e j )|
2 2
<∞ .
j=1

Define
∞
Ax = ∑ s j (x, e j )e j , x ∈ D.
j=1

Prove that A = A and that σ (A) = {s j : j = 1, 2, . . .}. Prove also that

∞
1
(A − zI)−1 x = ∑ s j − z (x, e j )e j
j=1

for all z ∈ ρ (A) and x ∈ D.

Exercise 28.5. Prove that the spectrum σ (U) of a unitary operator U lies on the
unit circle in C.

Theorem 28.6. Let A = A∗ and let {Eλ }λ ∈R be its spectral family. Then
(1) μ ∈ σ (A) if and only if Eμ +ε − Eμ −ε = 0 for every ε > 0.
(2) μ ∈ σp (A) if and only if Eμ − Eμ −0 = 0. Here Eμ −0 := limε →0+ Eμ −ε in the
sense of the strong operator topology.

Proof. (1) Suppose that μ ∈ σ (A) but there exists ε > 0 such that Eμ +ε −Eμ −ε = 0.
Then by the spectral theorem we obtain for every x ∈ D(A) that
 ∞ 
(A − μ I)x2 = (λ − μ )2 d(Eλ x, x) ≥ (λ − μ )2 d(Eλ x, x)
−∞ |λ −μ |≥ε
  μ −ε
 ∞ 
≥ ε2 d(Eλ x, x) = ε 2 + d(Eλ x, x)
|λ −μ |≥ε −∞ μ +ε
 
= ε 2 (Eμ −ε x, x) + x2 − (Eμ +ε x, x) = ε 2 x2 .

This inequality means (see part (2) of Theorem 28.2) that μ ∈ / σ (A) but μ ∈
ρ (A). This contradiction proves (1) in one direction. Conversely, if

Pn := Eμ + 1 − Eμ − 1 = 0
n n

for all n ∈ N, then there is a sequence {xn }∞ n=1 such that xn ∈ R(Pn ), i.e.,
xn = Pn xn , i.e., xn ∈ D(A) and xn  = 1. For this sequence it is true that
28 Spectra of Self-Adjoint Operators 301
 ∞
(A − μ I)xn 2 = (λ − μ )2 d(Eλ Pn xn , Pn xn )
−∞

= (λ − μ )2 d(Eλ xn , xn )
|λ −μ |≤1/n

1 ∞ 1 1
≤ d(Eλ xn , xn ) = xn 2 = 2 → 0
n2 −∞ n2 n

as n → ∞. Hence, this sequence satisfies Weyl’s criterion (see Exercise 28.2)

and therefore μ ∈ σ (A).
(2) Suppose μ ∈ R is an eigenvalue of A. Then there exists x0 ∈ D(A), x0 = 0, such
that  ∞
0 = (A − μ I)x0 2 = (λ − μ )2 d(Eλ x0 , x0 ).
−∞

In particular, for all n ∈ N large enough and ε > 0 we have that

 n  n
0= (λ − μ )2 d(Eλ x0 , x0 ) ≥ ε 2 d(Eλ x0 , x0 ) = ε 2 ((En − Eμ +ε )x0 , x0 )
μ +ε μ +ε
 2
= ε 2 (En − Eμ +ε )x0  .

Thus we may conclude that

0 = En x0 − Eμ +ε x0 .

Similarly we can get that

0 = E−n x0 − Eμ −ε x0 .

Letting n → ∞ and ε → 0, we obtain

x0 = Eμ x0 , 0 = Eμ −0 x0 .

Hence
x0 = (Eμ − Eμ −0 )x0

and therefore
Eμ − Eμ −0 = 0.

Conversely, define the projector

P := Eμ − Eμ −0 .

If P = 0, then there exists y ∈ H, y = 0, such that y = Py (e.g., any y ∈ R(P) =

{0} will do). For λ > μ it follows that
302 Part III: Operator Theory and Integral Equations

Eλ y = Eλ Py = Eλ Eμ y − Eλ Eμ −0 y = Py = y.

For λ < μ we have that

Eλ y = Eλ Eμ y − Eλ Eμ −0 y = Eλ y − Eλ y = 0.

Hence
 ∞  ∞
2
(A − μ I)y = (λ − μ ) d(Eλ y, y) =
2
(λ − μ )2 dλ (y, y) = 0.
−∞ μ

Therefore, Ay = μ y and y ∈ D(A), y = 0, i.e., μ is an eigenvalue of A, or

μ ∈ σp (A).


Remark 28.7. The statements of Theorem 28.6 can be reformulated as follows:

(1) μ ∈ σp (A) if and only if Eμ − Eμ −0 = 0.
(2) μ ∈ σc (A) if and only if Eμ − Eμ −0 = 0.

Definition 28.8. Let H and H1 be two Hilbert spaces. A bounded linear operator
K : H → H1 is called compact or completely continuous if it maps bounded sets in
H into precompact sets in H1 , i.e., for every bounded sequence {xn }∞
n=1 ⊂ H the
sequence {Kxn }∞n=1 ⊂ H1 contains a convergent subsequence.

If K : H → H1 is compact, then the following statements hold.

(1) K maps every weakly convergent sequence in H into a norm convergent
sequence in H1 . This condition is also sufficient.
(2) If H = H1 is separable, then every compact operator is a norm limit of a
sequence of operators of finite rank (i.e., operators with finite-dimensional
ranges).
(3) The norm limit of a sequence of compact operators is compact.
Let us prove (2). Let K be a compact operator. Since H is separable, it has an
orthonormal basis {e j }∞j=1 . Consider for n = 1, 2, . . . the projector

n
Pn x := ∑ (x, e j )e j , x ∈ H.
j=1

Then Pn ≤ Pn+1 and (I − Pn )x → 0 as n → ∞. Define

dn := sup K(I − Pn )x ≡ K(I − Pn ) .

x=1

Since R(I − Pn ) ⊃ R(I − Pn+1 ) (see Proposition 27.5), it follows that {dn }∞
n=1 is a
monotonically decreasing sequence of positive numbers. Hence the limit
28 Spectra of Self-Adjoint Operators 303

lim dn := d ≥ 0
n→∞

exists. Let us choose yn ∈ R(I − Pn ), yn  = 1, such that

d
K(I − Pn )yn  = Kyn  ≥ .
2
Then

|(yn , x)| = |((I − Pn )yn , x)| = |(yn , (I − Pn )x)| ≤ yn  (I − Pn )x → 0
w
as n → ∞ for all x ∈ H. This means that yn → 0. The compactness of K implies that
Kyn → 0. Thus d = 0. Therefore,

dn = K − KPn  → 0.

Since Pn is of finite rank, so is KPn , i.e., K is a norm limit of finite-rank operators.

Lemma 28.9. Suppose A = A∗ is compact. Then at least one of the two numbers
± A is an eigenvalue of A.

Proof. Since
A = sup |(Ax, x)|,
x=1

there exists a sequence xn with xn  = 1 such that

A = lim |(Axn , xn )|.

n→∞

In fact, we can assume that limn→∞ (Axn , xn ) exists and equals, say, a. Otherwise, we
would take a subsequence of {xn }. Since A = A∗ , it follows that a is real and A =
|a|. Due to the fact that every bounded set of a Hilbert space is weakly relatively
compact (the unit ball in our case), we can choose a subsequence of {xn }, say {xkn },
w
that converges weakly, i.e., xkn → x. The compactness of A implies that Axkn → y.
Next we observe that

Axkn − axkn 2 = Axkn 2 − 2a(Axkn , xkn ) + a2 ≤ A2 − 2a(Axkn , xkn ) + a2

= 2a2 − 2a(Axkn , xkn ) → 2a2 − 2a2 = 0

as n → ∞. Hence
⎧
⎪ 
⎨Axkn − axkn → 0, xkn → x,
Axkn → y, ⇒
⎪
⎩ w Ax = ax.
xkn → x
304 Part III: Operator Theory and Integral Equations

Since xkn  = 1, we have x = 1 also. Hence x = 0, and a is an eigenvalue of A. 

Remark 28.10. It is not difficult to show that the statement of Lemma 28.9 remains
true if A is just bounded and self-adjoint.

Theorem 28.11 (Riesz–Schauder). Suppose A = A∗ is compact. Then

(1) A has a sequence of real eigenvalues λ j = 0 that can be enumerated in such a
way that
|λ1 | ≥ |λ2 | ≥ · · · ≥ |λ j | ≥ · · · .

(2) If there are infinitely many eigenvalues, then lim j→∞ λ j = 0 and 0 is the only
accumulation point of {λ j }.
(3) The multiplicity of λ j is finite.
(4) If e j is the normalized eigenvector for λ j , then {e j }∞j=1 is an orthonormal system
and
∞ ∞
Ax = ∑ λ j (x, e j )e j = ∑ (Ax, e j )e j , x ∈ H.
j=1 j=1

This means that {e j }∞j=1 is an orthonormal basis of R(A).

(5) σ (A) = {0, λ1 , λ2 , . . . , λ j , . . .}, while 0 is not necessarily an eigenvalue of A.

Proof. Lemma 28.9 gives the existence of an eigenvalue λ1 ∈ R with |λ1 | = A
and a normalized eigenvector e1 . Introduce H1 = e⊥
1 . Then H1 is a closed subspace
of H, and A maps H1 into itself. Indeed,

(Ax, e1 ) = (x, Ae1 ) = (x, λ1 e1 ) = λ1 (x, e1 ) = 0

for every x ∈ H1 . The restriction of the inner product of H to H1 makes H1 a Hilbert

space (since H1 is closed), and the restriction of A to H1 , denoted by A1 = A|H1 , is
again a self-adjoint compact operator that maps in H1 . Clearly, its norm is bounded
by the norm of A, i.e., A1  ≤ A. Applying Lemma 28.9 to A1 on H1 , we get
an eigenvalue λ2 with |λ2 | = A1  and a normalized eigenvector e2 with e2 ⊥e1 . It
is clear that |λ2 | ≤ |λ1 |. Next introduce the closed subspace H2 = (span{e1 , e2 })⊥ .
Again, A leaves H2 invariant, and thus A2 := A1 |H2 = A|H2 is a self-adjoint compact
operator in H2 . Applying Lemma 28.9 to A2 on H2 , we obtain λ3 with |λ3 | = A2 
and a normalized eigenvector e3 with e3 ⊥e2 and e3 ⊥e1 . This process in an infinite-
dimensional Hilbert space leads us to the sequence {λ j }∞j=1 such that |λ j+1 | ≤ |λ j |
and corresponding normalized eigenvectors. Since |λ j | > 0 and the sequence is
monotonically decreasing, there is a limit

lim |λ j | = r.
j→∞

Clearly r ≥ 0. Let us prove that r = 0. If r > 0, then |λ j | ≥ r > 0 for each j = 1, 2, . . .,

or
28 Spectra of Self-Adjoint Operators 305

1 1
≤ < ∞.
|λ j | r

Hence the sequence of vectors

ej
y j :=
λj
w
is bounded, and therefore there is a weakly convergent subsequence y jk → y. The
compactness
 ofA implies
√ the strong convergence of Ay jk ≡ e jk . But for k = m we
 
have e jk − e jm = 2. This contradiction proves (1) and (2).
Exercise 28.6. Prove that if H is an infinite-dimensional Hilbert space, then the
identity operator I is not compact, and the inverse of a compact operator (if it exists)
is unbounded.
Exercise 28.7. Prove part (3) of Theorem 28.11.
Consider now the projector
n
Pn x := ∑ (x, e j )e j , x ∈ H.
j=1

Then I − Pn is a projector onto (span{e1 , . . . , en })⊥ ≡ Hn and hence

A(I − Pn )x ≤ AHn (I − Pn )x ≤ |λn+1 | x → 0

as n → ∞. Since
n n
APn x = ∑ (x, e j )Ae j = ∑ λ j (x, e j )e j
j=1 j=1

and
A(I − Pn )x = Ax − APn x → 0, n → ∞,

we have
∞
Ax = ∑ λ j (x, e j )e j ,
j=1

and part (4) follows. Finally, Exercise 28.4 gives immediately that

σ (A) = {0, λ1 , λ2 , . . . , λ j , . . .}.

This completes the proof. 

Corollary 28.12 (Hilbert–Schmidt theorem). An orthonormal system of eigen-
vectors {e j }∞j=1 of a compact self-adjoint operator A in a Hilbert space H is an
orthonormal basis if and only if N(A) = {0}.
306 Part III: Operator Theory and Integral Equations

Proof. Recall from Exercise 26.7 that

H = N(A∗ ) ⊕ R(A) = N(A) ⊕ R(A).

If N(A) = {0}, then H = R(A). This means that for every x ∈ H and ε > 0 there
exists yε ∈ R(A) such that
x − yε  < ε /2.

But by the Riesz–Schauder theorem,

∞
yε = Axε = ∑ λ j (xε , e j )e j .
j=1

Hence  
 ∞ 
 
x − yε  = x − ∑ λ j (xε , e j )e j  < ε /2.
 j=1


Making use of the Pythagorean theorem, Bessel’s inequality, and Exercise 25.8
yields
   
 n   n 
   
x − ∑ (x, e j )e j  ≤ x − ∑ λ j (xε , e j )e j 
 j=1
  j=1

 
 ∞ ∞ 
 
= x − ∑ λ j (xε , e j )e j + ∑ λ j (xε , e j )e j 
 j=1 j=n+1

 
 ∞ 
 
< ε /2 +  ∑ λ j (xε , e j )e j 
 j=n+1 
 1/2
∞
≤ ε /2 + ∑ |λ j |2 |(xε , e j )|2
j=n+1
 1/2
∞
≤ ε /2 + |λn+1 | ∑ |(xε , e j )| 2
j=n+1

≤ ε /2 + |λn+1 | xε  < ε

for n sufficiently large. This means that {e j }∞j=1 is a basis of H, and moreover, it is
an orthonormal basis.
Conversely, if {e j }∞j=1 is complete in H, then R(A) = H (Riesz–Schauder) and
therefore N(A) = {0}. 
Remark 28.13. The condition N(A) = {0} means that A−1 exists and H must be
separable in this case.
Proposition 28.14 (Riesz). If A is a compact operator on H and μ ∈ C, then the
null space of I − μ A is a finite-dimensional subspace.
28 Spectra of Self-Adjoint Operators 307

Proof. The null space N(I − μ A) is a closed subspace of H, since I − μ A is bounded.

Indeed, for each sequence fn → f and fn − μ A fn = 0 we have that f − μ A f = 0,
since A is continuous.
The operator A is compact on H and therefore also compact from N(I − μ A) onto
N(I − μ A), since N(I − μ A) is closed. Hence, for every f ∈ N(I − μ A) we have

I f = (I − μ A) f + μ A f = μ A f ,

and I is compact on N(I − μ A). Thus N(I − μ A) is finite-dimensional. 

Theorem 28.15 (Riesz’s lemma). If A is a compact operator on H and μ ∈ C, then
R(I − μ A) is closed in H.
Proof. If μ = 0, then R(I − μ A) = H. If μ = 0, then we assume without loss of
generality that μ = 1. Let f ∈ R(I − A), f = 0. Then there exists a sequence {gn } ⊂
H such that
f = lim (I − A)gn .
n→∞

We will prove that f ∈ R(I − A), i.e., there exists g ∈ H such that f = (I − A)g.
Since f = 0, we can assume by the decomposition H = N(I − A) ⊕ N(I − A)⊥ that
gn ∈ N(I − A)⊥ and gn = 0 for all n ∈ N.
Suppose that gn is bounded. Then there is a subsequence {gkn } such that

w
gkn → g.

The compactness of A implies that

Agkn → h = Ag.

Next,
gkn = (I − A)gkn + Agkn → f + h.

Hence g = f + Ag, i.e., f = (I − A)g.

Suppose that gn is not bounded. Then we can assume without loss of generality
that gn  → ∞. Let us introduce a new sequence
gn
un := .
gn 
w
Since un  = 1, there exists a subsequence ukn → u. The compactness of A gives
Aukn → Au. Since (I − A)gn → f , we have

1
(I − A)ukn = (I − A)gkn → 0.
gkn 

This means again that

308 Part III: Operator Theory and Integral Equations

ukn = (I − A)ukn + Aukn → Au

and u = Au, i.e., u ∈ N(I − A). But gn ∈ N(I − A)⊥ . Hence ukn ∈ N(I − A)⊥ and
further u ∈ N(I − A)⊥ , because N(I − A)⊥ is closed. Since ukn  = 1, we have
u = 1. Therefore, u = 0, while

u ∈ N(I − A) ∩ N(I − A)⊥ .

This contradiction shows that unbounded gn cannot occur. 

We are now ready to derive the following fundamental result of Riesz theory.

Theorem 28.16 (Riesz). Let A : H → H be a compact linear operator on a Hilbert

space H. Then for every μ ∈ C the operator I − μ A is injective (i.e., (I − μ A)−1
exists) if and only if it is surjective (i.e., R(I − μ A) = H). Moreover, in this case the
inverse operator (I − μ A)−1 : H → H is bounded.

Proof. If (I − μ A)−1 exists, then (I − μ A∗ )−1 exists too and therefore N(I − μ A∗ ) =
0. Then Riesz’s lemma (Theorem 28.15) and Exercise 26.7 imply H = R(I − μ A),
i.e., I − μ A is surjective.
Conversely, if I − μ A is surjective, then N(I − μ A∗ ) = 0, i.e., I − μ A∗ is injective
and so is I − μ A.
It remains to show that (I − μ A)−1 is bounded on H if I − μ A is injective. Assume
that (I − μ A)−1 is not bounded. Then there exists a sequence fn ∈ H with  fn  = 1
such that  
(I − μ A)−1 fn  ≥ n.

Define
fn (I − μ A)−1 fn
gn := , ϕn := .
(I − μ A)−1 fn  (I − μ A)−1 fn 

Then gn → 0 as n → ∞ and ϕn  = 1. Since A is compact, we can select a subse-

quence ϕkn such that Aϕkn → ϕ as kn → ∞. But

ϕn − μ Aϕn = gn ,

and we observe that ϕkn → μϕ and ϕ ∈ N(I − μ A). Hence ϕ = 0, and this contradicts
ϕn  = 1. 

Theorem 28.17 (Fredholm alternative). Suppose A = A∗ is compact. For given

g ∈ H either the equation
(I − μ A) f = g

has the unique solution (μ −1 ∈

/ σ (A)), in which case f = (I − μ A)−1 g, or else μ −1 ∈
σ (A), and this equation has a solution if and only if g ∈ R(I − μ A), i.e., g⊥N(I −
μ A). In this case, the general solution of the equation is of the form f = f0 + u,
28 Spectra of Self-Adjoint Operators 309

where f0 is a particular solution and u ∈ N(I − μ A) (u is the general solution of

the corresponding homogeneous equation), and the set of all solutions is a finite-
dimensional affine subspace of H.
Proof. Riesz’s lemma (Theorem 28.15) gives

R(I − μ A) = N(I − μ A)⊥ .

If μ −1 ∈
/ σ (A), then (μ )−1 ∈
/ σ (A) also. Thus

R(I − μ A) = N(I − μ A)⊥ = {0}⊥ = H.

Since A = A∗ , this means that (I − μ A)−1 exists, and the unique solution is f =
(I − μ A)−1 g.
If μ −1 ∈ σ (A), then R(I − μ A) is a proper subspace of H, and the equation (I −
μ A) f = g has a solution if and only if g ∈ R(I − μ A). Since the equation is linear,
every solution is of the form

f = f0 + u, u ∈ N(I − μ A),

and the dimension of N(I − μ A) is finite. 

Exercise 28.8. Let A = A∗ be compact and injective. Prove that σp (A) = σd (A) =
σ (A) \ {0} and 0 ∈ σess (A).
Exercise 28.9. Consider the Hilbert space H = l 2 (C) and
x2 xn
A(x1 , x2 , . . . , xn , . . .) = (0, x1 , , . . . , , . . .)
2 n

for (x1 , x2 , . . . , xn , . . .) ∈ l 2 (C). Show that A is compact and has no eigenvalues

(moreover, σ (A) = {0}) and is not self-adjoint.
Exercise 28.10. Consider the Hilbert space H = L2 (R) and

(A f )(t) = t f (t).

Show that the equation A f = f has no nontrivial solutions and that (I − A)−1 does
not exist. This means that the Fredholm alternative does not hold for a noncompact
but self-adjoint operator.
Exercise 28.11. Let H = L2 (Rn ) and let

A f (x) = K(x, y) f (y)dy,
Rn

where K(x, y) ∈ L2 (Rn × Rn ) is such that K(x, y) = K(y, x). Prove that A = A∗ and
that A is compact.
310 Part III: Operator Theory and Integral Equations

Theorem 28.18 (Weyl). If A = A∗ , then λ ∈ σess (A) if and only if there exists an
orthonormal system {xn }∞
n=1 such that

(A − λ I)xn  → 0

as n → ∞.

Proof. We will provide only a partial proof. See [5] for a full proof. Suppose that
λ ∈ σess (A). If λ is an eigenvalue of infinite multiplicity, then there is an infinite
orthonormal system of eigenvectors {xn }∞ n=1 , because dim(Eλ −Eλ −0 )H = ∞ in this
case. Since (A − λ I)xn ≡ 0, it is clear that

(A − λ I)xn → 0.

Next, suppose that λ is an accumulation point of σ (A). This means that λ ∈ σ (A)
and
λ = lim λn ,
n→∞

where λn = λm , n = m, and λn ∈ σ (A). Hence for each n = 1, 2, . . . we have that

Eλn +ε − Eλn −ε = 0

for all ε > 0. Therefore, there exists a sequence rn → 0 such that

Eλn +rn − Eλn −rn = 0.

We can therefore find a normalized vector xn ∈ R(Eλn +rn − Eλn −rn ). Since λn = λm
for n = m, we can find {xn }∞
n=1 as an orthonormal system. By the spectral theorem
we have
 ∞
(A − λ I)xn 2 = (λ − μ )2 d(Eμ xn , xn )
−∞
 ∞
= (λ − μ )2 d(Eμ (Eλn +rn − Eλn −rn )xn , xn )
−∞
 λn +rn
= (λ − μ )2 d(Eμ xn , xn )
λn −rn
 ∞
≤ max (λ − μ )2 d(Eμ xn , xn )
λn −rn ≤μ ≤λn +rn −∞

= max (λ − μ )2 → 0, n → ∞.
λn −rn ≤μ ≤λn +rn

This completes the proof. 

Theorem 28.19 (Weyl). Let A and B be two self-adjoint operators in a Hilbert

space. If there is z ∈ ρ (A) ∩ ρ (B) such that
28 Spectra of Self-Adjoint Operators 311

T := (A − zI)−1 − (B − zI)−1

is a compact operator, then σess (A) = σess (B).

Proof. We show first that σess (A) ⊂ σess (B). Take any λ ∈ σess (A). Then there is
an orthonormal system {xn }∞
n=1 such that

(A − λ I)xn  → 0, n → ∞.

Define the sequence yn as

yn := (A − zI)xn ≡ (A − λ I)xn + (λ − z)xn .

Due to Bessel’s inequality, every orthonormal system in the Hilbert space converges
w
weakly to 0. Hence yn → 0. We also have

|λ − z|
yn  ≥ |λ − z| xn  − (A − λ I)xn  = |λ − z| − (A − λ I)xn  > >0
2
for all n ≥ n0  1. Next we take the identity
 
(B − zI)−1 − (λ − z)−1 yn = −Tyn − (λ − z)−1 (A − λ I)xn .

w
Since T is compact and yn → 0, we deduce that
 
(B − zI)−1 − (λ − z)−1 yn → 0.

Introduce
zn := (B − zI)−1 yn .

Then
zn − (λ − z)−1 yn → 0,

or
yn + (z − λ )zn → 0.

|λ −z| |λ −z|
This fact and yn  > 2 imply that zn  ≥ 3 for all n ≥ n0  1. But

(B − λ I)zn ≡ (B − zI)zn + (z − λ )zn = yn + (z − λ )zn → 0.

Due to zn  ≥ |λ 3−z| > 0, the sequence {zn }∞

n=1 can be chosen as an orthonormal
system. Thus λ ∈ σess (B). This proves that σess (A) ⊂ σess (B). Finally, since −T
is compact too, we can interchange the roles of A and B and obtain the opposite
embedding. 
Chapter 29
Quadratic Forms. Friedrichs Extension.

Definition 29.1. Let D be a linear subspace of a Hilbert space H. A function Q :

D × D → C is called a quadratic form if
(1) Q(α1 x1 + α2 x2 , y) = α1 Q(x1 , y) + α2 Q(x2 , y),
(2) Q(x, β1 y1 + β2 y2 ) = β1 Q(x, y1 ) + β2 Q(x, y2 ),
for all α1 , α2 , β1 , β2 ∈ C and x1 , x2 , x, y1 , y2 , y ∈ D. The space D(Q) := D is called
the domain of Q. Then Q is
(1) densely defined if D(Q) = H.
(2) symmetric if Q(x, y) = Q(y, x).
(3) semibounded from below if there exists λ ∈ R such that Q(x, x) ≥ −λ x2 for
all x ∈ D(Q).
(4) closed (and semibounded) if D(Q) is complete with respect to the norm

xQ := Q(x, x) + (λ + 1) x2 .

(5) bounded (continuous) if there exists M > 0 such that

|Q(x, y)| ≤ M x y

for all x, y ∈ D(Q).

Exercise 29.1. Prove that ·Q is a norm and that

(x, y)Q := Q(x, y) + (λ + 1)(x, y)

is an inner product.

c Springer International Publishing AG 2017 313

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 29
314 Part III: Operator Theory and Integral Equations

Theorem 29.2. Let Q be a densely defined, closed, semibounded, and symmetric

quadratic form in a Hilbert space H such that

Q(x, x) ≥ −λ x2 , x ∈ D(Q).

Then there exists a unique self-adjoint operator A defined by the quadratic form Q
as
Q(x, y) = (Ax, y), x ∈ D(A), y ∈ D(Q)

that is semibounded from below, i.e.,

(Ax, x) ≥ −λ x2 , x ∈ D(A),

and D(A) ⊂ D(Q).

Proof. Let us introduce an inner product on D(Q) by

(x, y)Q := Q(x, y) + (λ + 1)(x, y), x, y ∈ D(Q)

(see Exercise 29.1). Since Q is closed, D(Q) = D(Q) is a closed subspace of H with
respect to the norm ·Q . This means that D(Q) with this inner product defines a
new Hilbert space HQ . It is clear also that

xQ ≥ x

for all x ∈ HQ . Thus, for fixed x ∈ H,

L(y) := (y, x), y ∈ HQ

defines a continuous (bounded) linear functional on the Hilbert space HQ . Applying

the Riesz–Fréchet theorem to HQ , we obtain an element x∗ ∈ HQ (x∗ ∈ D(Q)) such
that
(y, x) ≡ L(y) = (y, x∗ )Q .

It is clear that the map

H
x → x∗ ∈ HQ

defines a linear operator J such that

J : H → HQ , Jx = x∗ .

Hence
(y, x) = (y, Jx)Q , x ∈ H, y ∈ HQ .
29 Quadratic Forms. Friedrichs Extension. 315

Next we prove that J is self-adjoint and that it has an inverse operator J −1 . For all
x, y ∈ H we have

(Jy, x) = (Jy, Jx)Q = (Jx, Jy)Q = (Jx, y) = (y, Jx).

Hence J = J ∗ . It is bounded by the Hellinger–Toeplitz theorem (Exercise 26.2).

Suppose that Jx = 0. Then
(y, x) = (y, Jx)Q = 0

for every y ∈ D(Q). Since D(Q) = H, the last equality implies that x = 0, and there-
fore N(J) = {0} and J −1 exists. Moreover,

H = N(J) ⊕ R(J ∗ ) = R(J)

and R(J) ⊂ HQ . Now we can define a linear operator A on the domain D(A) ≡ R(J)
as
Ax := J −1 x − (λ + 1)x, λ ∈ R.

It is clear that A is densely defined and A = A∗ (J −1 is self-adjoint, since J is). If

now x ∈ D(A) and y ∈ D(Q) ≡ HQ , then

Q(x, y) = (x, y)Q − (λ + 1)(x, y) = (J −1 x, y) − (λ + 1)(x, y) = (Ax, y).

The semiboundedness of A from below follows from that of Q. It remains to prove

that this representation for A is unique. Assume that we have two such representa-
tions, A1 and A2 . Then for every x ∈ D(A1 ) ∩ D(A2 ) and y ∈ D(Q) we have that

Q(x, y) = (A1 x, y) = (A2 x, y).

It follows that
((A1 − A2 )x, y) = 0.

Since D(Q) = H, we must have A1 x = A2 x. This completes the proof. 

Corollary 29.3. Under the same√assumptions as in Theorem 29.2, there exists

√
A + λ I that is self-adjoint on D( A + λ I) ≡ D(Q) = HQ . Moreover,
 
Q(x, y) + λ (x, y) = ( A + λ Ix, A + λ Iy)

for all x, y ∈ D(Q).

Proof. Since A + λ I is self-adjoint and nonnegative, there exists a spectral family

{Eμ }∞
μ =0 such that  ∞
A+λI = μ dEμ .
0
316 Part III: Operator Theory and Integral Equations

We can therefore define the operator

  ∞
√
A + λ I := μ dEμ ,
0

which is also self-adjoint and nonnegative. Then

  ∗
Q(x, y) + λ (x, y) = ((A + λ I)x, y) = ( A + λ Ix, A + λ I y)

√
for all x ∈ D(A) and√y ∈ D(Q). This fact means that x ∈ D( A + λ I) and y ∈
√ ∗
D( A + λ I ). But A + λ I is self-adjoint, and therefore,
  ∗
D( A + λ I) = D( A + λ I ) = D(Q) ≡ HQ .

This completes the proof. 

Theorem 29.4 (Friedrichs extension). Let A be a nonnegative symmetric linear

operator in a Hilbert space H. Then there exists a self-adjoint extension AF of A
that is the smallest among all nonnegative self-adjoint extensions of A in the sense
that its corresponding quadratic form has the smallest domain. This extension AF is
called the Friedrichs extension of A.

Proof Let A be a nonnegative symmetric operator with D(A) = H. Its associated

quadratic form
Q(x, y) := (Ax, y), x, y ∈ D(Q) ≡ D(A),

is densely defined, nonnegative, and symmetric. Let us define a new inner product

(x, y)Q = Q(x, y) + (x, y), x, y ∈ D(Q).

Then D(Q) becomes an inner product space. This inner product space has a com-
pletion HQ with respect to the norm

xQ := Q(x, x) + x2 .

Moreover, the quadratic form Q(x, y) has an extension Q1 (x, y) to this Hilbert space
HQ defined by
Q1 (x, y) = lim Q(xn , yn )
n→∞

HQ HQ
whenever x = limn→∞ xn , y = limn→∞ yn , xn , yn ∈ D(Q) and these limits exist. The
quadratic form Q1 is densely defined, closed, nonnegative, and symmetric. There-
fore, Theorem 29.2, applied to Q1 , gives a unique nonnegative self-adjoint operator
AF such that
29 Quadratic Forms. Friedrichs Extension. 317

Q1 (x, y) = (AF x, y), x ∈ D(AF ) ⊂ HQ , y ∈ D(Q1 ) ≡ HQ .

Since for x, y ∈ D(A) one has

(Ax, y) = Q(x, y) = Q1 (x, y) = (AF x, y),

it follows that AF is a self-adjoint extension of A.

It remains to prove that AF is the smallest nonnegative self-adjoint extension of
A. Suppose that B ≥ 0, B = B∗ , is such that A ⊂ B. The associated quadratic form
QB (x, y) := (Bx, y) is an extension of Q ≡ QA . Hence

QB ⊃ Q = Q1 .

This completes the proof. 

Chapter 30
Elliptic Differential Operators

Let Ω be a domain in Rn , i.e., an open and connected set. We introduce the following
notation:
1 , . . . , xn ) ∈ Ω ;
(1) x = (x
(2) |x| = x12 + · · · + xn2 ;
(3) α = (α1 , . . . , αn ) is a multi-index, i.e., α j ∈ N0 ≡ N ∪ {0}:
(a) |α | = α1 + · · · + αn ,
(b) α ≥ β if α j ≥ β j for all j = 1, 2, . . . , n,
(c) α + β = (α1 + β1 , . . . , αn + βn ),
(d) α − β = (α1 − β1 , . . . , αn − βn ) if α ≥ β ,
(e) xα = x1α1 · · · xnαn with 00 = 1,
(f) α ! = α1 ! · · · αn ! with 0! = 1;
(4) ∂ j = ∂
∂xj and ∂ α = ∂1α1 · · · ∂nαn .

Definition 30.1. An elliptic partial differential operator A(x, ∂ ) of order m on Ω is

an operator of the form
A(x, ∂ ) = ∑ aα (x)∂ α ,
|α |≤m

where aα (x) ∈ C∞ (Ω ), whose principal symbol

a(x, ξ ) = ∑ aα (x)ξ α , ξ ∈ Rn
|α |=m

is invertible for all x ∈ Ω and ξ ∈ Rn \ {0}, that is, a(x, ξ ) = 0 for all x ∈ Ω and
ξ ∈ Rn \ {0}.

Assumption 30.2. We assume that aα (x) are real for |α | = m.

c Springer International Publishing AG 2017 319

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 30
320 Part III: Operator Theory and Integral Equations

Under Assumption 30.2 either a(x, ξ ) > 0 or a(x, ξ ) < 0 for all x ∈ Ω and ξ ∈
Rn \ {0}. Without loss of generality we assume that a(x, ξ ) > 0. Assumption 30.2
implies also that m is even and that for every compact set K ⊂ Ω there exists CK > 0
such that
a(x, ξ ) ≥ CK |ξ |m , x ∈ Ω , ξ ∈ Rn .

Assumption 30.3. We assume that A(x, ∂ ) is formally self-adjoint, i.e.,

A(x, ∂ ) = A (x, ∂ ) := ∑ (−1)|α | ∂ α (aα (x)·).

|α |≤m

Exercise 30.1. Prove that A(x, ∂ ) = A (x, ∂ ) if and only if

aα (x) = ∑ (−1)|β |Cβα ∂ β −α aβ (x),

α ≤β
|β |≤m

where
β!
Cβα = .
α !(β − α )!

Hint: Make use of the generalized Leibniz formula

β
∂ α ( f g) = ∑ Cα ∂ α −β f ∂ β g.
β ≤α

Assumption 30.4. We assume that A(x, ∂ ) has a divergence form

A(x, ∂ ) ≡ ∑ (−1)|α | ∂ α (aαβ (x)∂ β ),

|α |=|β |≤m/2

where aαβ = aβ α and this value is real for all α and β . We assume also the gener-
alized ellipticity condition
 
∑
Ω |α |=|β |=m/2
aαβ (x)∂ α f ∂ β f dx ≥ ν ∑
Ω |α |=m/2
|∂ α f |2 dx, f ∈ C0∞ (Ω ),

where ν > 0 is called the constant of ellipticity.

Remark 30.5. If the coefficients aαβ of A(x, ∂ ) are constants, then this generalized
ellipticity condition reads

∑ aαβ ξ α +β ≥ ν ∑ ξ 2α .
|α |=|β |=m/2 |α |=m/2
30 Elliptic Differential Operators 321

Exercise 30.2. Prove that

∑ ξ 2α
|ξ |m ,
|α |=m/2

i.e.,
c|ξ |m ≤ ∑ ξ 2α ≤ C|ξ |m ,
|α |=m/2

where c and C are some constants.

Example 30.6. Let us consider
n
A(x, ∂ ) = − ∑ ∂ j2 = −Δ , x ∈ Ω ⊂ Rn
j=1

in H = L2 (Ω ) and prove that A ⊂ A∗ with

 
D(A) = C0∞ (Ω ) = f ∈ C∞ (Ω ) : supp f = {x : f (x) = 0} is compact in Ω .

Let u, v ∈ C0∞ (Ω ). Then

n   
(Au, v)L2 = − ∑ ∂ j2 u vdx
j=1 Ω
n  n  
=−∑ ∂ j ((∂ j u) v) dx + ∑ (∂ j u) ∂ j v dx
j=1 Ω j=1 Ω

=− (v∇u, nx )dx + (∇u, ∇v)L2 = (∇u, ∇v)L2 ,
∂Ω

where ∂ Ω is the boundary of Ω and nx is the unit outward normal vector at x ∈ ∂ Ω .

Here we have made use of the divergence theorem. In a similar fashion we obtain
n 
(∇u, ∇v)L2 = − ∑ u∂ j2 vdx = (u, −Δ v)L2 = (u, Av)L2 .
j=1 Ω

Hence A ⊂ A∗ and A is closable.

Example 30.7. Recall from Example 30.6 that

(−Δ u, v)L2 = (∇u, ∇v)L2 , u, v ∈ C0∞ (Ω ).

Hence
(−Δ u, u)L2 = ∇u2L2 ≤ uL2 Δ uL2 , u ∈ C0∞ (Ω ).
322 Part III: Operator Theory and Integral Equations

Therefore,
2 2 2 2
uW 2 = uL2 + ∇uL2 + Δ uL2
2

≤ u2L2 + uL2 Δ uL2 + Δ u2L2
3 3 3
≤ u2L2 + Δ u2L2 ≡ u2A ,
2 2 2

where ·A is a norm that corresponds to the operator A = −Δ as follows:

u2A := u2L2 + −Δ u2L2 .

It is also clear that uA ≤ uW 2 . Combining these inequalities gives

2

2
uW 2 ≤ uA ≤ uW 2
3 2 2

for all u ∈ C0∞ (Ω ). A completion of C0∞ (Ω ) with respect to these norms leads us to
the statement
◦
D(A) = W 22 (Ω ).
◦
Thus A = −Δ on D(A) = W 22 (Ω ). Let us determine D(A∗ ) in this case. By the
definition of D(A∗ ) we have

D((−Δ )∗ ) = v ∈ L2 (Ω ) : there exists v∗ ∈ L2 (Ω ) such that
(−Δ u, v) = (u, v∗ ) for all u ∈ C0∞ (Ω )} .

If we assume that v ∈ W22 (Ω ), then this is equivalent to

(u, (−Δ )∗ v) = (u, v∗ ),

i.e., (−Δ )∗ v = v∗ and D((−Δ )∗ ) = W22 (Ω ). Finally, for Ω ⊂ Rn with Ω = Rn we

obtain that
A ⊂ A ⊂ A∗ ≡ (A)∗

and A = A and A = (A)∗ , that is, the closure of A does not lead us to a self-adjoint
operator.
◦
Remark 30.8. If Ω = Rn , then W 22 (Rn ) ≡ W22 (Rn ) and therefore

A = A∗ = (A)∗ .

Hence the closure of A is self-adjoint in that case, i.e., A is essentially self-adjoint.

30 Elliptic Differential Operators 323

Example 30.9. Consider again A = −Δ on D(A) = C0∞ (Ω ) with Ω = Rn . Since

(−Δ u, u)L2 = ∇u2L2 ≥ 0,

it follows that −Δ is nonnegative with lower bound λ = 0. Therefore,

Q(u, v) := (∇u, ∇v)L2

is a densely defined nonnegative quadratic form with D(Q) ≡ D(A) = C0∞ (Ω ). A

new inner product is defined as

(u, v)Q := (∇u, ∇v)L2 + (u, v)L2

and
u2Q ≡ uW
2
1 (Ω ) .
2

If we apply now the procedure from Theorem 29.4, then we obtain the existence of
Q1 = Q with respect to the norm ·Q , which will also be nonnegative and closed
◦
with D(Q1 ) ≡ W 12 (Ω ). The next step is to obtain the Friedrichs extension AF as

AF = J −1 − I
◦
with D(AF ) ≡ R(J) ⊂ W 12 (Ω ). A more careful examination of Theorem 29.2 leads
us to the fact
◦ ◦
D(AF ) = W 12 (Ω ) ∩ D(A∗ ) = W 12 (Ω ) ∩W22 (Ω ).

Remark 30.10. In general, for a symmetric operator we have

D(AF ) = {u ∈ HQ : Au ∈ H} ,

which is equivalent to

D(AF ) = {u ∈ HQ : u ∈ D(A∗ )} .

Exercise 30.3. Let H = L2 (Ω ) and A(x, D) = −Δ + q(x), where q(x) = q(x) and
q(x) ∈ L∞ (Ω ). Define A, A∗ , and AF .

Exercise 30.4. Let H = L2 (Ω ) and


A(x, ∂ ) = −(∇ + IW(x))2
+ q(x),

where W is an n-dimensional real-valued vector from W∞1 (Ω ) and q is a real-valued

function from L∞ (Ω ). Define A, A∗ , and AF .
324 Part III: Operator Theory and Integral Equations

Consider now a bounded domain Ω ⊂ Rn and an elliptic operator A(x, ∂ ) in Ω of

the form
A(x, ∂ ) = ∑ (−1)|α | ∂ α (aαβ (x)∂ β ),
|α |=|β |≤m/2

where aαβ (x) = aβ α (x) are real. Assume that there exists C0 > 0 such that

m
|aαβ (x)| ≤ C0 , |α |, |β | < ,
2

for all x ∈ Ω . Assume also that A(x, ∂ ) is elliptic, that is,

 
∑
Ω |α |=|β |=m/2
aαβ (x)∂ α f ∂ β f dx ≥ ν
Ω |α |=m/2
∑ |∂ α |2 dx, ν > 0.

Theorem 30.11 (Gårding’s inequality). Suppose that A(x, ∂ ) is as above. Then

for every ε > 0 there is Cε > 0 such that

(A f , f )L2 (Ω ) ≥ (ν − ε )  f 2 m/2 −Cε  f 2L2 (Ω )

W2 (Ω )

for all f ∈ C0∞ (Ω ).

Proof. Let f ∈ C0∞ (Ω ). Then integration by parts yields

(A f , f )L2 (Ω ) = ∑ (−1)|α |
Ω
∂ α (aαβ (x)∂ β f ) f dx
|α |=|β |≤m/2

= ∑ aαβ (x)∂ α f ∂ β f dx
|α |=|β |=m/2 Ω

+ ∑ aαβ (x)∂ α f ∂ β f dx
|α |=|β |<m/2 Ω
 
≥ν ∑ |∂ α f |2 dx −C0 ∑ |∂ α f ||∂ β f |dx
|α |=m/2 Ω |α |=|β |<m/2 Ω
 
≥ν ∑ |∂ α f |2 dx − (C0 + ν ) ∑ |∂ α f |2 dx
|α |≤m/2 Ω |α |<m/2 Ω

= ν  f 2 m/2 − (C0 + ν )  f 2 m/2−1 .

W2 (Ω ) W2 (Ω )

Next we make use of the following lemma.

Lemma 30.12. For all ε > 0 and 0 < δ ≤ m/2 there is Cε (δ ) > 0 such that

(1 + |ξ |2 )m/2−δ ≤ ε (1 + |ξ |2 )m/2 +Cε (δ )

for all ξ ∈ Rn .
30 Elliptic Differential Operators 325

Proof. Let ε > 0 and 0 < δ ≤ m/2. If (1 + |ξ |2 )δ ≥ ε1 , then

(1 + |ξ |2 )−δ ≤ ε .

Hence
(1 + |ξ |2 )m/2−δ ≤ ε (1 + |ξ |2 )m/2 ,

i.e., the claim holds for every positive constant Cε (δ ). For (1 + |ξ |2 )δ < 1
ε we can
obtain
 m/2−δ
1 δ
2 m/2−δ
(1 + |ξ | ) < ≡ Cε (δ ).
ε

This proves the claim. 

Applying this lemma with δ = 1 to the norm of the Sobolev spaces W2k , we conclude
that
 f 2 m/2−1 ≤ ε1  f 2 m/2 +Cε1  f 2L2 (Ω )
W2 (Ω ) W2 (Ω )

for all ε1 > 0. Hence

(A f , f )L2 (Ω ) ≥ ν  f 2 m/2 − (C0 + ν )  f 2 m/2−1

W2 (Ω ) W2 (Ω )

≥ ν  f 2 m/2 − (C0 + ν )ε1  f 2 m/2 − (C0 + ν )Cε1  f 2L2 (Ω )

W2 (Ω ) W2 (Ω )

= (ν − ε )  f 2 m/2 −Cε  f 2L2 (Ω ) .

W2 (Ω )

This proves the theorem. 

Corollary 30.13. There exists a self-adjoint Friedrichs extension AF of A with do-
◦ m/2
main D(AF ) = W 2 (Ω ) ∩W2m (Ω ).
Proof. It follows from Gårding’s inequality that

(A f , f )L2 (Ω ) ≥ −Cε  f 2L2 (Ω ) , f ∈ D(A).

This means that Aμ := A + μ I is positive for μ > Cε , and therefore Theorem 29.4
gives us the existence of
 
Aμ F
≡ (AF )μ = AF + μ I

with domain
  ◦ m/2
D(AF ) = D( Aμ F ) = W 2 (Ω ) ∩ D(A∗ ),

◦ m/2
where W 2 (Ω ) is the domain of the corresponding closed quadratic form (see The-
orem 29.4). If Ω is bounded with smooth boundary ∂ Ω , then it can be proved that
326 Part III: Operator Theory and Integral Equations

D(A∗ ) = W2m (Ω ).

This concludes the proof. 

Gårding’s inequality has two more consequences. Firstly,

 
(AF )μ f  2 ≥ C0  f  2 , C0 > 0,
L L

so that
(AF )−1
μ : L (Ω ) → L (Ω ).
2 2

Secondly,  
(AF )μ f  −m/2 ≥ C0  f  m/2 , C0 > 0,
W2 (Ω ) W2 (Ω )

so that
◦ m/2
(AF )−1
μ : L (Ω ) → W 2 (Ω ).
2

Corollary 30.14. The spectrum σ (AF ) = {λ j }∞j=1 is the sequence of eigenvalues

of finite multiplicity with only one accumulation point at +∞. In short, σ (AF ) =
σd (AF ). The corresponding orthonormal system {ψ j }∞j=1 of eigenfunctions forms
an orthonormal basis and
∞
L2
AF f = ∑ λ j ( f , ψ j )ψ j
j=1

for all f ∈ D(AF ).

Proof. We begin with a lemma.

Lemma 30.15. The embedding

◦ m/2
W 2 (Ω ) → L2 (Ω )

is compact.
◦ m/2
Proof. It is enough to show that for every {ϕk }∞
k=1 ⊂ W 2 (Ω ) with ϕk  m/2 ≤1
W2
there exists {ϕ jk }∞
k=1 that is a Cauchy sequence in L (Ω ). Since Ω is bounded, we
2

have
|ϕk (ξ )| ≤ ϕk L2 |Ω |1/2 ,

i.e., the Fourier transform ϕk (ξ ) (see Chapter 16) is uniformly bounded. Thus there
exists ϕjk (ξ ) that converges pointwise in Rn . Next, using Parseval’s equality and the
definition of the Sobolev spaces H s (Rn ) (see Chapter 20), we have
30 Elliptic Differential Operators 327

  
ϕ j − ϕ j 2 2 = |ϕjk (ξ ) − ϕ
 jm (ξ )| dξ
2
k m L
R
n

= |ϕjk (ξ ) − ϕ
 jm (ξ )| dξ +
2
|ϕjk (ξ ) − ϕ
 jm (ξ )| dξ
2
|ξ |<r |ξ |>r

≤ |ϕjk (ξ ) − ϕ
 jm (ξ )| dξ
2
|ξ |<r

1
+ (1 + |ξ |2 )m/2 |ϕjk (ξ ) − ϕ
 jm (ξ )| dξ
2
(1 + r2 )m/2 Rn
  2
2 −m/2 
= |ϕjk (ξ ) − ϕ
 jm (ξ )| dξ + (1 + r )
2
ϕ jk − ϕ jm W m/2
|ξ |<r 2

=: I1 + I2 .

The first term I1 tends to 0 as k, m → ∞ by the Lebesgue dominated convergence

theorem for every fixed r > 0. The second term converges to 0 as r → ∞ because

ϕ j − ϕ j  m/2 ≤ 2. 
k m W2

Lemma 30.15 gives us that

 −1
Aμ F
: L2 (Ω ) → L2 (Ω )

is a compact operator. Applying the Riesz–Schauder and Hilbert–Schmidt theorems,

we get the following statements:
 −1
(1) σ ( Aμ F ) = {0, μ1 , μ2 , . . .} with μ j ≥ μ j+1 > 0 and μ j → 0 as j → ∞.
(2) μ j is of finite multiplicity.
 −1
(3) Aμ F ψ j = μ j ψ j , where {ψ j }∞j=1 is an orthonormal system.
(4) {ψ j }∞j=1 forms an orthonormal basis in L2 (Ω ).
Since AF ψ j = λ j ψ j with λ j = 1
μj − μ , we conclude that

σ (AF ) = {λ j }∞j=1 , λ j ≤ λ j+1 , λ j → ∞.

Moreover, λ j has finite multiplicity and the ψ j are the corresponding eigenfunctions.
We have also the following representation:

 −1 ∞
Aμ F
f= ∑ μ j ( f , ψ j )ψ j , f ∈ L2 (Ω ).
j=1

Exercise 30.5. Prove that

∞
AF f = ∑ λ j ( f , ψ j )ψ j
j=1
328 Part III: Operator Theory and Integral Equations

for all f ∈ D(AF ).

The corollary is proved. 
In some applications it is quite useful to deal with semigroups of operators. We
consider these semigroups in Hilbert spaces. This approach allows us to characterize
the domains of operators (when they are not bounded); see, e.g., [3].
Let A be a nonnegative self-adjoint operator in a Hilbert space H. By the spectral
theorem we can characterize D(A) as follows: f ∈ D(A) if and only if
 ∞
(1 + λ 2 )d(Eλ f , f ) < ∞
0

and we define a new norm

 f D(A) :=  f H + A f H .

Definition 30.16. Let {G(t)}t>0 be a family of bounded linear operators from H to

H. This family is called an equi-bounded strongly continuous semigroup if
(1) G(t + s) f = G(t)(G(s) f ) for s,t > 0 and f ∈ H,
(2) G(t) f H ≤ M  f H for t > 0 and f ∈ H with M > 0 that does not depend on t
of f ,
(3) limt→0+ G(t) f − f H = 0 for f ∈ H.
Remark 30.17. We can complete this definition by G(0) := I.
Definition 30.18. The infinitesimal generator A of the semigroup {G(t)}t>0 is de-
fined by the formula  
 G(t) − I 
lim 
 − A f =0

t→0 t

with domain D(A) consisting of all f ∈ H such that

G(t) − I
lim f
t→0 t

exists in H.
Remark 30.19. In the sense of the previous definition we write G (0) = A.
Example 30.20. Let H = L2 (Rn ). Let ω (ξ ) be an infinitely differentiable positive
function on Rn \ {0} that is positively homogeneous of order m > 0, i.e., ω (t ξ ) =
|t|m ω (ξ ). Let us define the family {G(t)}t>0 by the formula

G(t) f := F −1 (e−t ω (ξ ) F f ), f ∈ L2 (Rn ).

It is clear that G(t) : L2 (Rn ) → L2 (Rn ). Moreover,

30 Elliptic Differential Operators 329

(1)

G(t + s) f = F −1 (e−(t+s)ω (ξ ) F f )
= F −1 (e−t ω (ξ ) F F −1 (e−sω (ξ ) F f )) = G(t)(G(s) f );

(2)
   
   
G(t) f L2 = F −1 (e−t ω (ξ ) F f ) = e−t ω (ξ ) F f  ≤ F f L2 =  f L2 ;
L2 L2

(3)
   
   
G(t) f − f L2 = F −1 (e−t ω (ξ ) F f − F f ) = (e−t ω (ξ ) − 1)F f  →0
L2 L2

as t → 0 by the Lebesgue dominated convergence theorem. Also by this theorem

we have that
 
G(t) f − f −t ω (ξ ) − 1
−1 e
lim = lim F F f = −F −1 (ω (ξ )F f ) ≡ A f .
t→0 t t→0 t

The domain of A is

D(A) = { f ∈ L2 : ω (ξ )F f L2 < ∞}.

For example, if ω (ξ ) = |ξ |m , then A = −(−Δ )m/2 and D(A) = W2m (Rn ).

Example 30.21. Let A = A∗ ≥ 0. Define
 ∞
G(t) := eItA ≡ eItλ dEλ .
0

Then
(1)
 ∞  ∞
G(t + s) = eI(t+s)λ dEλ = eItλ eIsλ dEλ
0 0
 ∞  ∞
= eItλ dEλ eIsμ dEμ = G(t)G(s);
0 0

(2)  ∞
G(t) f 2 = |eItλ |2 d(Eλ f , f ) =  f 2 ;
0

(3)  ∞
G(t) f − f 2 = |eItλ − 1|2 d(Eλ f , f ) → 0, t → 0,
0
330 Part III: Operator Theory and Integral Equations

and
 ∞ Itλ  ∞
G(t) f − f e −1
= dEλ f → I λ dEλ f ≡ IAf, t → 0,
t 0 t 0

and  
 G(t) − I 
lim 
 f − IAf  = 0.

t→0 t H

These examples reveal a one-to-one correspondence between the infinitesimal gen-

erators of semigroups and self-adjoint operators in Hilbert space.
Chapter 31
Spectral Functions

Let us consider a bounded domain Ω ⊂ Rn and an elliptic differential operator

A(x, ∂ ) in Ω of the form

A(x, ∂ ) = ∑ (−1)|α | ∂ α (aαβ (x)∂ β ),

|α |=|β |≤m/2

where aαβ = aβ α are real, in C∞ (Ω ), and bounded for all α and β . We assume that
 
∑
Ω |α |=|β |=m/2
aαβ (x)∂ α f ∂ β f dx ≥ ν ∑
Ω |α |=m/2
|∂ α f |2 dx, ν > 0.

As was proved above, there exists at least one self-adjoint extension of A with
D(A) = C0∞ (Ω ), namely, the Friedrichs extension AF with

◦ m/2
D(AF ) = W 2 (Ω ) ∩W2m (Ω ).

 of A. Without loss of generality

Let us consider an arbitrary self-adjoint extension A
 
we assume that A ≥ 0. Therefore, A has the spectral representation
 ∞
=
A λ dEλ
0

with domain
    ∞

 = f ∈ L2 (Ω ) :
D A λ d(Eλ f , f ) < ∞ .
2
0

 
 like that for the Friedrichs extension AF .
In general, we have no formula for D A
But we can say that

c Springer International Publishing AG 2017 331

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 31
332 Part III: Operator Theory and Integral Equations

◦  
Wm 
2 (Ω ) ⊂ D A .

Indeed, since aαβ ∈ C∞ (Ω ) and aαβ is bounded, A(x, ∂ ) can be rewritten in the
usual form
A(x, ∂ ) = ∑ aγ (x)∂ γ
|γ |≤m

with bounded coefficients. Hence

A f L2 (Ω ) ≤ c ∑ ∂ γ f L2 (Ω ) ≡ c  f W m (Ω ) .
2
|γ |≤m

This proves the embedding.

But even in this general case, one can obtain more significant results than just
the previous embedding into the domain of the operator. The basis for these results
is the following classical theorem of L. Gårding, which is given here without proof
 is an arbitrary (semibounded
(see, e.g., [14, 15]). In this theorem it is assumed that A
from below) self-adjoint extension of an elliptic differential operator A(x, D) with
smooth and bounded coefficients.
= A
Theorem 31.1 (Gårding). If A  ∗ , then Eλ is an integral operator in L2 (Ω )
such that 
Eλ f (x) = θ (x, y, λ ) f (y)dy,
Ω

where θ (x, y, λ ) is called the spectral function and has the properties
(1) θ (x, y, λ ) = θ (y, x, λ ),
(2) 
θ (x, y, λ ) = θ (x, z, λ )θ (z, y, λ )dz
Ω

and 
θ (x, x, λ ) = |θ (x, z, λ )|2 dz ≥ 0,
Ω

(3)
sup θ (x, ·, λ )L2 (Ω ) ≤ c1 λ k ,
x∈Ω1

where Ω1 = Ω1 ⊂ Ω , k ∈ N with k > n

2m and c1 = c(Ω1 ).

Remark 31.2. It was proved by L. Hörmander that in fact,

θ (x, x, λ ) ≤ c1 λ n/m .
31 Spectral Functions 333
 
Corollary 31.3. Let z ∈ ρ A  . Then (A
 − zI)−1 is an integral operator whose ker-
nel G(x, y, z) is called the Green’s function corresponding to A  and that has the
properties
(1)  ∞
dλ θ (x, y, λ )
G(x, y, z) = ,
0 λ −z

(2) G(x, y, z) = G(y, x, z).

 
Proof. Since z ∈ ρ A  , von Neumann’s spectral theorem gives us

 ∞
 − zI)−1 f =
(A (λ − z)−1 dEλ f .
0

Next, by Theorem 31.1 we get

 ∞ 

 − zI)−1 f =
(A (λ − z)−1 dλ θ (x, y, λ ) f (y)dy
0 Ω
  ∞

= (λ − z)−1 dλ θ (x, y, λ ) f (y)dy = G(x, y, z) f (y)dy,
Ω 0 Ω

where G(x, y, z) is as in (1). Since

 ∞  ∞
dθ (x, y, λ ) dθ (y, x, λ )
G(x, y, z) = = = G(y, x, z),
0 λ −z 0 λ −z

(2) is also proved. 


Exercise 31.1. Prove that θ (x, x, λ ) is a monotonically increasing function with
respect to λ and
(1) |θ (x, y, λ )|2 ≤ θ (x, x, λ )θ (y, y, λ ),
(2) |Eλ f (x)| ≤ θ (x, x, λ )1/2  f L2 (Ω ) .
Exercise 31.2. Prove that
 
|Eλ f (x) − Eμ f (x)| ≤ Eλ f − Eμ f L2 (Ω ) |θ (x, x, λ ) − θ (x, x, μ )|1/2

for all λ > 0 and μ > 0.

Exercise 31.3. Let us assume that n < m. Prove that
 ∞
θ (x, y, λ )dλ
G(x, y, z) =
0 (λ − z)2

and that G(·, y, z) ∈ L2 (Ω ).

334 Part III: Operator Theory and Integral Equations

In the case of the Friedrichs extension for a bounded domain, the spectral func-
tion θ (x, y, λ ) and the Green’s function have a special form. We know from Corol-
lary 7.14 that the spectrum σ (AF ) is the sequence {λ j }∞j=1 of eigenvalues with only
one accumulation point at +∞, and the corresponding orthonormal system {ψ j }∞j=1
forms an orthonormal basis in L2 (Ω ) such that
∞
AF f = ∑ λ j ( f , ψ j )ψ j in L2 .
j=1

This fact implies that


Eλ f = ∑ ( f , ψ j )ψ j = ∑ f (y)ψ j (y)dyψ j (x)
λ j <λ λ j <λ Ω
⎛ ⎞
 
= ⎝ ∑ ψ j (x)ψ j (y)⎠ f (y)dy = θ (x, y, λ ) f (y)dy,
Ω λ j <λ Ω

i.e., the spectral function θ (x, y, λ ) has the following form:

θ (x, y, λ ) = ∑ ψ j (x)ψ j (y).

λ j <λ

Hence (see Corollary 31.3) the Green’s function has the form

∞
ψ j (x)ψ j (y)
G(x, y, z) = ∑ λj −z
in L2 .
j=1

If we assume now that n < m, then we obtain that the Green’s function G(x, y, z)
is uniformly bounded in (x, y) ∈ Ω × Ω . Let us assume for simplicity that z = iz2
and AF ≥ I. Then applying Hörmander’s estimate (see Remark 31.2) for the spectral
function, we obtain
∞ ∞
|ψ j (x)||ψ j (y)| |ψ j (x)||ψ j (y)|
|G(x, y, z)| ≤ ∑  =∑ ∑ 
j=1 λ j2 + z22 k=0 2k ≤λ j <2k+1 λ j2 + z22
⎛ ⎞1 ⎛ ⎞1
2 2
∞
1 ⎝
≤ ∑ 2k
(2 + z 2 )1/2 ∑ |ψ j (x)|2 ⎠ ⎝ ∑ |ψ j (y)|2 ⎠
k=0 2 2k ≤λ <2k+1 2k ≤λ <2k+1
j j
∞
(2k+1 )n/m
≤ ∑ (22k + z2 )1/2 .
k=0 2

Since n < m, this series converges for all z2 .

Chapter 32
The Schrödinger Operator

There are certain physical problems that are connected with the reconstruction of
the quantum-mechanical potential in the Schrödinger operator H = −Δ + q(x). This
operator is defined in Rn . Here and throughout we assume that q is real-valued.
First of all we have to define H as a self-adjoint operator in L2 (Rn ). Our basic
assumption is that the potential q(x) belongs to L p (Rn ) for n2 < p ≤ ∞ and has the
following special behavior at infinity:

|q(x)| ≤ c|x|−μ , |x| > R,

with some μ ≥ 0 and R > 0 sufficiently large. The parameter μ will be specified
later, depending on the situation. We would like to construct the self-adjoint ex-
tension of this operator by Friedrichs’s method, because formally our operator is
defined now only for smooth functions, say for functions from C0∞ (Rn ). In order to
construct such an extension let us consider the Hilbert space H1 defined as follows:

H1 = { f ∈ L2 (Rn ) : ∇ f (x) ∈ L2 (Rn ) and |q(x)|| f (x)|2 dx < ∞}.
Rn

The inner product in H1 is defined by


( f , g)H1 = (∇ f , ∇g)L2 + q(x) f (x)g(x)dx + μ0 ( f , g)L2 ,
Rn

with μ0 > 0 sufficiently large and fixed.

Lemma 32.1. Assume that f ∈ W21 (Rn ) and q ∈ L p (Rn ) for n

2 < p ≤ ∞, n ≥ 2. Then
for every 0 < ε < 1 there exists cε > 0 such that

|(q f , f )L2 | ≤ ε ∇ f 2L2 (Rn ) + cε  f 2L2 (Rn ) .

c Springer International Publishing AG 2017 335

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 32
336 Part III: Operator Theory and Integral Equations

Proof. If p = ∞, then

|(q f , f )L2 | ≤ |q(x)|| f (x)|2 dx ≤ qL∞ (Rn )  f 2L2 (Rn )
Rn
≤ ε ∇ f 2L2 (Rn ) + qL∞ (Rn )  f 2L2 (Rn ) .

If n
2 < p < ∞, then we estimate
 
|(q f , f )L2 | ≤ |q(x)|| f (x)|2 dx + |q(x)|| f (x)|2 dx
|q(x)|<A |q(x)|>A

≤ |q(x)|| f (x)|2 dx + A  f 2L2 (Rn ) .
|q(x)|>A

Let us consider the integral appearing in the last estimate. For n ≥ 3 it follows from
Hölder’s inequality that
   2   n−2
n n 2n n
|q(x)|| f (x)| dx ≤
2
|q(x)| dx 2 | f (x)| n−2 dx
|q(x)|>A |q(x)|>A |q(x)|>A
 2
n
≤ A( 2 −p) n
n 2 2
|q(x)| dx p
c1  f W 1 (Rn )
|q(x)|>A 2

2p 2p
2
≤ c1 A1− n q n
L p (Rn )
 f W 1 (Rn ) .
2

To obtain the last inequality we used the fact that n

2 < p < ∞ and a well known
2n
[1, 3] embedding: W21 (Rn ) ⊂ L n−2 (Rn ), n ≥ 3, with the norm estimate
√
f 2n ≤ c1  f W 1 (Rn ) .
L n−2 (Rn ) 2

Collecting these estimates, we obtain

2p 2p
2 2
|(q f , f )L2 | ≤ c1 A1− n qLnp (Rn )  f W 1 n + A  f L2 (Rn )
2 (R )
2p
 2p

1− 2p 2 1− 2p
= c1 A n qL p (Rn ) ∇ f L2 (Rn ) + A + c1 A
n n qL p (Rn )  f 2L2 (Rn ) .
n

2p
The claim follows now from the last inequality, since A1− n can be chosen suffi-
ciently small for n2 < p < ∞.


Exercise 32.1. Prove Lemma 32.1 for n = 2.

Exercise 32.2. Let us assume that q(x) satisfies the conditions

(1) |q| ≤ c1 |x|−γ1 , |x| < 1, and
(2) |q| ≤ c2 |x|−γ2 , |x| > 1.
Find the conditions on γ1 and γ2 that ensure the statement of Lemma 32.1.
32 The Schrödinger Operator 337

Remark 32.2. Lemma 32.1 holds for every potential q ∈ L p (Rn ) + L∞ (Rn ) for
p > n2 , n ≥ 2.
Using Lemma 32.1, we obtain

 f 2H1 = ∇ f 2L2 (Rn ) + μ0  f 2L2 (Rn ) + (q f , f )L2

≥ ∇ f 2L2 (Rn ) + μ0  f 2L2 (Rn ) − ε ∇ f 2L2 (Rn ) − cε  f 2L2 (Rn )
= (1 − ε ) ∇ f 2L2 (Rn ) + (μ0 − cε )  f 2L2 (Rn ) .

We choose here 0 < ε < 1 and μ0 > cε . On the other hand,

 f 2H1 ≤ (1 + ε ) ∇ f 2L2 (Rn ) + (μ0 + cε )  f 2L2 (Rn ) .

These two inequalities mean that the new Hilbert space H1 is equivalent to the space
W21 (Rn ) up to equivalent norms. Thus we may conclude that for every f ∈ H1 our
operator is well defined by

( f , (H + μ0 ) f )L2 (Rn ) =  f 2H1 .

Moreover, since H + μ0 is positive, we must have

 
 1 2
 f 2H1 = (H + μ0 ) 2 f  2 ,
L (Rn )

and the following statements hold:

1
(1) the domain of (H + μ0 ) 2 is W21 (Rn );
(2) D(H + μ0 ) ≡ D(H) ⊂ W21 (Rn );
(3) D(H) = { f ∈ W21 (Rn ) : H f ∈ L2 (Rn )}.
1 1
Remark 32.3. (H + μ0 ) f = (H + μ0 ) 2 (H + μ0 ) 2 f is equivalent to
1
D(H) = { f ∈ W21 (Rn ) : g := (H + μ0 ) 2 f ∈ W21 (Rn )}.

Remark 32.4. Let us consider this extension procedure from another point of view.
The inequality

( f , (H + μ0 ) f )L2 ≥ (1 − ε ) ∇ f 2L2 (Rn ) + (μ0 − cε )  f 2L2 (Rn )

allows us to conclude that

(1) ( f , (H + μ0 ) f )L2 ≥ c  f 2L2 (Rn ) and
(2) ( f , (H + μ0 ) f )L2 ≥ c  f W
2
1 (Rn )
2

for every f ∈ C0∞ (Rn ). This means that there exists (H + μ0 )−1 that is also defined
for g ∈ C0∞ (Rn ) and satisfies the inequality
338 Part III: Operator Theory and Integral Equations
 
(1) (H + μ0 )−1 gL2 (Rn ) ≤ c1 gL2 (Rn ) or even
 
(2) (H + μ0 )−1 gW 1 (Rn ) ≤ c1 gW −1 (Rn ) , where W2−1 (Rn ) is the dual space of
2 2
W21 (Rn ).

L2 W −1
Since (H + μ0 )−1 is a bounded operator and C0∞ (Rn ) = L2 (Rn ) and C0∞ (Rn ) = 2

−1
W2 (Rn ), we can extend (H + μ0 )−1 as a bounded operator onto L2 (Rn ) in the first
case and onto W2−1 (Rn ) in the second. The extension for the differential operator is
H + μ0 = ((H + μ0 )−1 )−1 and D(H + μ0 ) = R((H + μ0 )−1 ) in both cases. It is also
clear that H + μ0 and (H + μ0 )−1 are self-adjoint operators.
Lemma 32.5. Let us assume that q ∈ L p (Rn ) for 2 ≤ p ≤ ∞ if n = 2, 3 and q ∈
L p (Rn ) for n2 < p ≤ ∞ if n ≥ 4. Then

W22 (Rn ) ⊂ D(H).

Proof. Since H = −Δ + q and D(H) = { f ∈ W21 (Rn ) : H f ∈ L2 (Rn )}, it is enough

to show for the required embedding that for f ∈ W22 (Rn ) it follows that q f ∈ L2 (Rn ).
If p = ∞, then 
|q f |2 dx ≤ q2L∞ (Rn )  f 2L2 (Rn ) < ∞
Rn

for every f ∈ W22 (Rn ), n ≥ 2.

For finite p let us consider first the case n = 2, 3. Since W22 (Rn ) ⊂ C(Rn ) ∩
∞
L (Rn ) (Sobolev embedding), we must have
  
|q f |2 dx = |q f |2 dx + |q f |2 dx
Rn |q|<A |q|>A
 
≤ A2 | f |2 dx +  f 2L∞ (Rn ) |q| p |q|2−p dx
|q|<A |q|>A
2 2
≤ A  f L2 (Rn ) +C  f W 2 (Rn ) A
2 2−p
qLp p (Rn ) < ∞.
2

We will apply the following embeddings:

f ∈ W22 (R4 ) ⊂ L p (R4 ), p < ∞.
2n
f ∈ W22 (Rn ) ⊂ L n−4 (Rn ), if n ≥ 5.
Therefore, on applying Hölder’s inequality we obtain

  
|q f |2 dx = |q f |2 dx + |q f |2 dx
Rn |q|<A |q|>A
  4   n−4
n n 2n n
≤ A2  f 2L2 (Rn ) + |q| 2 dx | f | n−4 dx
|q|>A |q|>A
32 The Schrödinger Operator 339

 4
n 4 n
≤ A2  f 2L2 (Rn ) +CA( 2 −p) n |q| p dx 2
 f W 2 (Rn ) < ∞.
|q|>A 2

if n ≥ 5 and

   2   2
p
p
p
|q f | dx ≤
2
|q| dx
p
| f | dx <∞
R4 R4 R4

if n = 4 for 2 < p < ∞ and p < ∞.


Exercise 32.3. Prove this lemma for q ∈ L p (Rn ) + L∞ (Rn ), n
2 < p ≤ ∞, if n ≥ 4 and
for q ∈ L2 (Rn ) + L∞ (Rn ) if n = 2, 3.
n
Remark 32.6. For n ≥ 5 we may consider q ∈ L 2 (Rn ).
Lemma 32.7. Let us assume that q ∈ Ln (Rn ), n ≥ 3. Then

D(H) = W22 (Rn ).

Proof. The embedding W22 (Rn ) ⊂ D(H) was proved in Lemma 32.5. Let us now
assume that f ∈ D(H), i.e., f ∈ W21 (Rn ) and H f ∈ L2 (Rn ). Note that for
g := H f ∈ L2 we have the following representation:

− f = (−Δ + 1)−1 (q − 1) f − (−Δ + 1)−1 g

= (−Δ + 1)−1 (q f ) − (−Δ + 1)−1 g − (−Δ + 1)−1 f .

It is therefore enough to show that q f ∈ L2 (Rn ). We use the same arguments as

in Lemmas 32.1 and 32.5. So it suffices to show that q f ∈ L2 (Rn ) for every f ∈
2n
W21 (Rn ). From the embedding W21 (Rn ) ⊂ L n−2 (Rn ) for n ≥ 3 we have by Hölder’s
inequality
  
|q(x)|2 | f (x)|2 dx = |q(x)|2 | f (x)|2 dx + |q(x)|2 | f (x)|2 dx
Rn |q|<A |q|>A
  2   n−2
n 2n n
≤A 2
 f 2L2 (Rn ) + |q| dx
n
|f| n−2 dx < ∞.
|q|>A |q|>A

Thus the lemma is proved.


Exercise 32.4. Describe the domain of H for the case n
2 < p < n, n ≥ 3. Hint: Prove
that D(H) ⊂ W22 (Rn ) +Ws2 (Rn ) with some s = s(p).
Let us consider now the Laplacian H0 = −Δ in Rn , n ≥ 1. Since (−Δ f , f )L2 =
∇ f 2L2 (Rn ) ≥ 0 for every f ∈ W21 (Rn ), it follows that H0 is a nonnegative opera-
tor. Moreover, H0 = H0∗ with domain D(H0 ) = W22 (Rn ), and this operator has the
spectral representation
340 Part III: Operator Theory and Integral Equations
 ∞
H0 f = λ dEλ f .
0

It follows that σ (H0 ) ⊂ [0, +∞), but in fact, σ (H0 ) = [0, +∞) and even σ (H0 ) =
σc (H0 ) = σess (H0 ) = [0, +∞). In order to understand this fact it is enough to ob-
serve that for every λ ∈ [0, +∞) the homogeneous equation (H0 − λ )u = 0 has a

solution of the form u(x,k) = ei(k,x) , where (k,k) = λ and k ∈ Rn . These solutions
u(x,k) are called generalized eigenfunctions, but u(x,k) ∈ / L2 (Rn ). These solutions
are bounded and correspond to the continuous spectrum of H0 . Consequently, u(x,k)
are not eigenfunctions, but generalized eigenfunctions. If we consider the solutions
of the equation (H0 − λ )u = 0 for λ < 0, then these solutions will be exponentially
increasing at the infinity. This implies that λ < 0 does not belong to σ (H0 ).
For the spectral representation of H0 we have two forms:
(1) the Neumann spectral representation
 ∞
−Δ f = λ dEλ f , f ∈ W22 (Rn );
0

(2) the scattering theory representation

 
−Δ f = F −1 (|ξ |2 f) = (2π )−n |ξ |2 ei(ξ ,x) dξ e−i(ξ ,y) f (y)dy.
Rn Rn

Exercise 32.5. Determine the connection between these two representations.

There are some important remarks to be made about the resolvent (−Δ − z)−1
for z ∈
/ [0, +∞). A consequence of the spectral theorem is that
 ∞
(−Δ − z)−1 = (λ − z)−1 dEλ , z ∈ C \ [0, +∞),
0

and for such z the operator (−Δ − z)−1 is a bounded operator in L2 (Rn ). Moreover,
/ [0, +∞), the operator (−Δ − z)−1 as an operator-valued function
with respect to z ∈
is a holomorphic function. This fact follows immediately from
 ∞
((−Δ − z)−1 )z = (λ − z)−2 dEλ = (−Δ − z)−2 .
0

The last integral converges as well as the previous one (even better). Now we are in
a position to formulate a theorem about the spectrum of H = −Δ + q.

Theorem 32.8. Assume that q ∈ L p (Rn ), n

2 < p ≤ ∞, n ≥ 2, and q(x) → 0 as
|x| → +∞. Then
(1) σc (H) ⊃ (0, +∞);
(2) σp (H) ⊂ [−c0 , 0] is of finite multiplicity with its only accumulation point at {0}
with c0 such that −Δ + q ≥ −c0 .
32 The Schrödinger Operator 341

In order to prove this theorem we will prove two lemmas.

Lemma 32.9. Assume that the potential q(x) satisfies the assumptions of Theorem
32.8. Assume in addition that q(x) ∈ L2 (Rn ) for n = 2, 3. Then

(−Δ − z)−1 ◦ q : L2 (Rn ) → L2 (Rn )

is a compact operator for z ∈

/ [0, +∞).

Proof. Due to our assumptions on the potential q(x), it can be represented as the
sum q(x) = q1 (x) + q2 (x), where q1 ∈ L p (|x| < R) with the same p and q2 → 0
as |x| → ∞. We may assume (without loss of generality) that q2 is supported in
{x ∈ Rn : |x| > R} and that it is a continuous function. Let us consider first the cases
n = 2, 3. If f ∈ L2 (Rn ), then q1 f ∈ L1 (|x| < R) and (−Δ − z)−1 (q1 f ) ∈ W12 (Rn )
(by the Fourier transform). By the embedding theorem for Sobolev spaces (see, e.g.,
[1, 3]) we have that

2− 2n
(−Δ − z)−1 (q1 f ) ∈ W12 (Rn ) ⊂ W2 (Rn ), n = 2, 3,

with the norm estimate

   
(−Δ − z)−1 (q1 f ) 2 n ≤ (−Δ − z)−1 (q1 f ) 2− n
L (R ) W2 2
 
≤ c (−Δ − z) (q1 f )W 2 ≤ c q1 f L1 (Rn )
−1
1

≤ c q1 L2 (|x|<R)  f L2 (|x|<R) ,

or  
(−Δ − z)−1 ◦ q1  2 ≤ c q1 L2 ,
L (|x|<R)→L2 (Rn )

where c may depend only on z.

In the case n ≥ 4 and q ∈ L p (|x| < R), p > n2 , we may obtain by Hölder’s inequal-
ity that
2n
q1 f ∈ Ls (|x| < R), s > ,
n+4

for f ∈ L2 (Rn ), and therefore, (−Δ −z)−1 (q1 f ) ∈ Ws2 (Rn ). Again by the embedding
theorem for Sobolev spaces we have

2−n( 1s − 21 )
(−Δ − z)−1 (q1 f ) ∈ W2 (Rn )

for some s > 2n

n+4 , with the norm estimate
 
(−Δ − z)−1 ◦ q1  2 ≤ c q1 L p (|x|<R) .
L (|x|<R)→L2 (Rn )
342 Part III: Operator Theory and Integral Equations

In order to prove that (−Δ − z)−1 ◦ q1 is a compact operator we approximate it as

follows:
A := (−Δ − z)−1 ◦ q1 , A j := ϕ j (x)A,

where ϕ j (x) ∈ C0∞ (Rn ), |ϕ j (x)| ≤ C and

 
A − A j  2 2 → 0, j → ∞.
L →L

The reason is that (−Δ − z)−1 ◦ q1 is actually an integral operator with kernel
Kz (x − y) that tends to 0 as |x| → ∞ uniformly with respect to |y| < R (note that
q1 is supported in |y| < R). We therefore can approximate this kernel Kz by the
functions ϕ j ∈ C0∞ (Rn ). But A j is a compact operator for each j = 1, 2, . . ., because
the embedding
W2α (|x| < R) ⊂ L2 (|x| < R)

is compact for positive α . This implies that A is also a compact operator.

Next we consider q2 . Since for f (x) ∈ L2 (Rn ) we know that (−Δ − z)−1 f ∈
W2 (Rn ), we conclude that q2 (−Δ − z)−1 f ∈ L2 (|x| > R). In fact,
2

q2 · : W22 (Rn ) → L2 (|x| > R)

is a compact embedding. In order to establish this fact let us consider again ϕ j (x) ∈
L∞
C0∞ (Rn ), |ϕ j (x)| ≤ c and ϕ j → q2 as j → ∞. We can state this because C0∞ = Ċ. That
is why we required such behavior of q(x) at infinity (q → 0 as |x| → +∞). If we set
A := q2 (−Δ − z)−1 and A j := ϕ j (−Δ − z)−1 , then we obtain
   
A − A j  2 2 ≤ sup |ϕ j − q2 | (−Δ − z)−1  2 2
L →L L →L
x
(32.1)
≤ c sup |ϕ j − q2 | → 0, j → +∞.
x

2
But we know that W2,comp ⊂ Lcomp
2 is a compact embedding. This implies (together
with (32.1)) that A is a compact operator. Since

(−Δ − z)−1 ◦ q2 = (q2 (−Δ − z)−1 )∗ ,

the Lemma is proved.


Lemma 32.10. Let Q be an open and connected set in C. Let A(z) be a compact,
operator-valued, and holomorphic function in Q and in L2 (Rn ). If (I + A(z0 ))−1
exists for some z0 ∈ Q, then (I + A(z))−1 exists in all of Q except for finitely many
points from Q with the only possible accumulation points on ∂ Q.
Proof. We will prove this lemma only for our concrete operator A(z) := (−Δ −
z)−1 q(x) (see [22] for a full proof). Lemma 32.9 shows us that A(z) is a compact
/ [0, +∞). The remarks about Rz = (−Δ − z)−1 show us that A(z) is a
operator for z ∈
32 The Schrödinger Operator 343

holomorphic function in C \ [0, +∞). Also we can prove that (I + (−Δ − z)−1 q)−1
exists for all z ∈ C \ R and for real z < −c0 , where −Δ + q ≥ −c0 . Indeed, if z ∈ C
with Im z = 0, then (I + (−Δ − z)−1 q)u = 0, or (−Δ − z)u = −qu, or (Δ u, u) +
z(u, u) = (qu, u). This implies for z, Im z = 0, that (u, u) = 0 if and only if u = 0. In
the real case z < −c0 , the equality (I + (−Δ − z)−1 q)u = 0 implies

((−Δ + q)u, u) − z(u, u) = 0.

It follows that
(−c0 − z) u2L2 ≤ 0

and thus u = 0. These remarks show us that in C \ [0, +∞) our operator I + (−Δ −
z)−1 q may be noninvertible only on [−c0 , 0).
Let us consider an open and connected set Q in C \ [0, +∞) such that [−c0 , 0) ⊂
Q; see Figure 32.1.
It is easily seen that there exists z0 ∈ Q such that (I + (−Δ − z0 )−1 q)−1 exists
also. It is not difficult to show that there exists δ > 0 such that (I + (−Δ − z)−1 q)−1
exists in Uδ (z0 ). Indeed, let us choose δ > 0 such that

1
A(z) − A(z0 )L2 →L2 < (32.2)
(I + A(z0 ))−1 L2 →L2

for all z such that |z − z0 | < δ . Then

(I + A(z))−1 = (I + A(z0 ))−1 (I + B)−1 ,

where B := (A(z) − A(z0 ))(I + A(z0 ))−1 . But B < 1 due to (32.2), and then

(I + B)−1 = I − B + B2 + · · · + (−1)n Bn + · · ·

exists in the strong topology from L2 to L2 . We may therefore conclude that I + A(z)
may be noninvertible only for finitely many points in Q. This fact follows from
the holomorphicity of A(z) with respect to z by analogy with the theorem about
the zeros of a holomorphic function in complex analysis. Moreover, since A(z) is a
compact operator, it follows by Fredholm’s alternative that Ker(I + A(z)) has finite

Im z

Q
Re z
−c0

Fig. 32.1 The set Q.

344 Part III: Operator Theory and Integral Equations

dimension. We conclude that (I + (−Δ − z)−1 q)−1 does not exist at only a finite
numbers of points (at most) on [−c0 , −ε ] for all ε > 0, and these points are of finite
multiplicity. This completes the proof.


Let us return to the proof of Theorem 32.8.

Proof (Proof of Theorem. 32.8). Let μ be a positive number and μ + c0 > 0 (H ≥

−c0 I). Let us consider for such μ the second resolvent equation

(H + μ )−1 = (H0 + μ )−1 − (H + μ )−1 ◦ q ◦ (H0 + μ )−1 ,

where H0 = −Δ and H = −Δ + q(x). It follows from Lemma 32.9 that q ◦ (H0 +

μ )−1 is a compact operator in L2 (Rn ). This means that (H + μ )−1 is a compact
perturbation of (H0 + μ )−1 . Hence, by Theorem 28.18 above we have

σess ((H + μ )−1 ) = σess ((H0 + μ )−1 ).

But σess ((H0 + μ )−1 ) = [0, μ1 ] = σc ((H0 + μ )−1 ), from which we conclude that

σess (H + μ ) = [μ , +∞].

Outside of this set we have only points of the discrete spectrum with one possible
accumulation point at μ . This statement is a simple corollary of Lemma 32.10.
Moreover, these points of the discrete spectrum are located on [μ − c0 , μ ) and are of
finite multiplicity. Hence the discrete spectrum σd (H) of H belongs to [−c0 , 0) with
only one possible accumulation point at {0}. And (0, +∞) is the continuous part of
σ (H). There is only one problem. Weyl’s theorem states that the operators H and
H0 do not have the same spectrum but the same essential spectrum. Thus on (0, +∞)
there can be eigenvalues of infinite multiplicity (see the definition of σess ). In order
to eliminate such a possibility and to prove that 0 ∈ σc (H) and σd (H) is finite, let
us assume additionally that our potential q(x) has a special behavior at infinity:

|q(x)| ≤ c|x|−μ , |x| → +∞,

where μ > 2. In that case we can prove that on the interval [−c0 , 0) the operator H
has at most finitely many points of the discrete spectrum. And we prove also that
0 ∈ σc (H).
Assume to the contrary that H contains infinitely many points of the discrete
spectrum or that one of them has infinite multiplicity. This means that in D(H) there
exists an infinite-dimensional space of functions {u} that satisfy the equation

(−Δ + q)u = λ u, −c0 ≤ λ ≤ 0.

It follows that
 
(|∇u(x)|2 + q+ (x)|u(x)|2 )dx ≤ q− (x)|u(x)|2 dx,
Rn Rn
32 The Schrödinger Operator 345

where q+ and q− are the positive and negative parts of the potential q(x),
respectively. Let us consider an infinite sequence
 of functions {u(x)} that are
orthogonal with respect to the inner product Rn q− (x)u(x)v(x)dx. This sequence
is uniformly bounded in the metric Rn (|∇u|2 + |q||u|2 )dx, and hence in the metric

Rn (|∇u| + |u| )dx. But for every eigenfunction u(x) of the operator H with eigen-
2 2

value λ ∈ [−c0 , 0] the following inequality holds (see [9]):


|u(x)| ≤ c|λ | |u(y)|dy,
|x−y|≤1

where c does not depend on x. It follows from this inequality that

(1) λ = 0 is not an eigenvalue;
(2) this orthogonal sequence is uniformly bounded in every fixed ball.
Lemma 32.11. Denote by U the set of functions u(x) ∈ D(H) that are uniformly
bounded in every fixed ball in Rn . Then U is a precompact set in the metric

|q||u|2 dx
Rn

if it is a bounded set in the metric


(|∇u|2 + |u|2 )dx.
Rn

Proof. Let {uk (x)}∞

k=1 ⊂ U be an arbitrary sequence that is bounded in the second
metric. Then for u(x) := uk (x) − um (x) we have for r sufficiently large that
  
|u(x)|2
|q(x)||u(x)|2 dx ≤ c μ
dx + |q(x)||u(x)|2 dx
Rn |x|>r |x| |x|≤r,|q(x)|≤A

+ |q(x)||u(x)|2 dx =: I0 + I1 + I2 .
|x|≤r,|q(x)|>A

For n ≥ 3 (for n = 2 the proof needs some changes) and μ > 2 we get
 
I0 ≤ cr2−μ |x|−2 |u(x)|2 dx ≤ cr2−μ |∇u(x)|2 dx, u ∈ W21 (Rn ).
|x|>r Rn

Due to the uniform boundedness of U in every ball, we conclude that


I2 ≤ c |q(x)|dx → 0
|x|≤r,|q(x)|>A

as A → +∞ uniformly on U with fixed r. Since the embedding W21 ⊂ L2 for every

ball is compact, the boundedness of the sequence in the second metric implies the
precompactness in L2 for every ball. We therefore have
346 Part III: Operator Theory and Integral Equations

I1 ≤ A |u(x)|2 dx → 0, m, k → ∞
|x|≤r

with r and A fixed. On passing to the limit, these inequalities for I0 , I1 , and I2 show
that 
|q(x)||u(x)|2 dx → 0, m, k → ∞.
Rn

Thus the lemma is proved.



Let us return to the proof of (1). By Lemma 32.11 we obtain that our sequence
(which is orthogonal with respect to the inner product Rn q− (x)u(x)v(x)dx) is a
Cauchy sequence in the first metric. But this fact contradicts its orthogonality. Thus
(1) is proved.
(2) Let us discuss (briefly) the situation with a positive eigenvalue on the contin-
uous spectrum. If we consider the homogeneous equation

[I + (−Δ − k2 − i0)−1 q] f = 0, k2 > 0,

in the space Ċ(Rn ), then by Green’s formula one can show (see [20] or [21]) that the
n−1
solution f (x) of this equation behaves at infinity as o(|x|− 2 ). We thus conclude
[21] that f (x) ≡ 0 outside some ball in Rn . By the unique continuation principle for
the Schrödinger operator it follows that f ≡ 0 in the whole of Rn .

Let us consider now the spectral representation of the Schrödinger operator H =
−Δ + q(x), with q(x) as in Theorem 32.8 with the behavior O(|x|−μ ), μ > 2, at
infinity (compare with the spectral representation that follows from von Neumann’s
spectral theorem, Theorem 27.13, for the self-adjoint operator −Δ + q in L2 (Rn )).
For all f ∈ D(H), we have
  M
H f (x) = (2π )−n k2 u(x,k)dk f (y)u(y,k)dy + ∑ λ j f j u j (x),
Rn Rn j=1

where u(x,k) are the solutions of the equation Hu = k2 u, u j (x) are the orthonormal
eigenfunctions corresponding to the negative eigenvalues λ j , taking into account the
multiplicity of λ j and f j = ( f , u j )L2 (Rn ) . The functions u(x,k) are called generalized
eigenfunctions. When q ≡ 0, the generalized eigenfunctions have the form u(x,k) =

ei(x,k) . This follows by means of the Fourier transform. Indeed,

(−Δ − k2 )u = 0

if and only if
(|ξ |2 − k2 )
u = 0,
32 The Schrödinger Operator 347

or
u = ∑ cα δ (α ) (ξ −k),
α

since
|ξ |2 = k2

if and only if
ξ −k = 0.

Hence

u(x,k) = ∑ cα F −1 (δ (α ) (ξ −k))(x)
α
 
= ∑ cα ei(x,k) F −1 (δ (α ) (ξ ))(x) = ∑ cα ei(x,k) xα .
α α


But u(x,k) must be bounded, and so u(x,k) = c0 ei(x,k) . We choose c0 = 1. If we have
the Schrödinger operator H = −Δ + q with q = 0, then it is natural to look for the

scattering solutions of Hu = k2 u of the form u(x,k) = ei(x,k) + usc (x,k). Due to this
representation, we have


(−Δ − k2 )(ei(x,k) + usc ) = −qu,

or
(−Δ − k2 )usc = −qu.

In order to find usc , let us recall that from Chapter 22 we know the fundamental
solution of the operator −Δ − k2 . Therefore,


u(x, k) = ei(x,k) − G+
k (|x − y|)q(y)u(y)dy,
Rn

where
  n−2
i |k| 2
(1)
G+
k (|x|) = H n−2 (|k||x|)
4 2π |x| 2

is the fundamental solution for the operator −Δ − k2 . This equation is called the
Lippmann–Schwinger integral equation.
Chapter 33
The Magnetic Schrödinger Operator

As a continuation (and, in some sense, an extension) of the previous chapter, where

the Schrödinger operator was considered, in this chapter we consider the magnetic
Schrödinger operator

 (x))2 u +V (x)u,
Hm u := −(∇ + iW x ∈ Ω ⊂ Rn , n ≥ 2, (33.1)

where Ω is an open set (not necessarily bounded) in Rn , n ≥ 2, with smooth bound-

 (x)
ary. It is assumed that the electric potential V (x) and the magnetic potential W
are real-valued and belong to the following spaces:

V ∈ L p (Ω ) with some 1 < p ≤ ∞ for n = 2 and n/2 ≤ p ≤ ∞ for n ≥ 3,

(33.2)
 ∈ Ls (Ω ) with some 2 < s ≤ ∞ for n = 2 and n ≤ s ≤ ∞ for n ≥ 3.
W

The operator Hm of the form (33.1) is symmetric in the Hilbert space L2 (Ω ) on the
domain C0∞ (Ω ). We want to construct the Friedrichs self-adjoint extension of this
operator and to describe the domain of this extension.
Lemma 33.1. Assume that the conditions (33.2) are satisfied for the coefficients of
Hm . Then for all f ∈ C0∞ (Ω ) the following double inequality holds:

γ1 ∇ f 2L2 (Ω ) −C1  f 2L2 (Ω ) ≤ (Hm f , f )L2 (Ω ) ≤ γ2 ∇ f 2L2 (Ω ) +C2  f 2L2 (Ω ) ,

(33.3)
where 0 < γ1 < 1 < γ2 and C1 ,C2 > 0.

Proof. For all f ∈ C0∞ (Ω ) we have by integration by parts that

 
(Hm f , f )L2 (Ω ) =  ) f |2 dx +
|(∇ + iW V (x)| f |2 dx
Ω Ω
   
= |∇ f |2 dx +  |2 | f |2 dx +
|W V (x)| f |2 dx − 2  f · ∇ f )dx.
Im(W
Ω Ω Ω Ω

c Springer International Publishing AG 2017 349

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 33
350 Part III: Operator Theory and Integral Equations

Therefore, for ε > 0 sufficiently small we obtain the following double inequality:
 2  2
   
(1 − ε ) ∇ f 2L2 (Ω ) − (1/ε − 1) W f 2 − |V |1/2 f  2 ≤ (Hm f , f )L2 (Ω )
L (Ω ) L (Ω )
 2  2
   
≤ (1 + ε ) ∇ f 2L2 (Ω ) + (1/ε + 1) W f  2 + |V |1/2 f  2 .
L (Ω ) L (Ω )

Due to the conditions (33.2) the functions W  and |V |1/2 are from equivalent
spaces (with respect to norm estimates). We shall therefore estimate only the norm
 1/2 2  2
|V | f  2 , and the norm   f
W  can be estimated in the same manner. Let
L (Ω ) L 2 (Ω )
us consider first n ≥ 3 and some p satisfying n/2 ≤ p < ∞. Then for R > 0, using
the Hölder’s inequality we obtain
   
 1/2 2
|V | f  2 ≤ |V || f |2 dx + |V || f |2 dx
L (Ω ) |V (x)|>R |V (x)|≤R
 2/n  (n−2)/n
≤ |V |n/2
dx |f|2n/(n−2)
dx + R  f 2L2 (Ω )
|V (x)|>R |V (x)|>R
2p/n 2 2
≤ C1 R1−2p/n V L p (Ω ∩{x:|V (x)|>R})  f W 1 (Ω ) + R  f L2 (Ω ) .
2

In obtaining the latter inequality we have used the fact that n ≥ 3, n/2 ≤ p < ∞, and
the well known embedding [1, 3] W21 (Ω ) → L2n/(n−2) (Ω ) with the norm estimate

 f L2n/(n−2) (Ω ) ≤ C1  f W 1 (Ω ) .
2

Collecting all these estimates, we get

 
 1/2 2 2p/n 2 2
|V | f  2 ≤ C1 R1−2p/n V L p (Ω ∩{x:|V (x)|>R})  f W 1 (Ω ) + R  f L2 (Ω )
L (Ω ) 2
2
≤ δ (R)  f W + R  f 2L2 (Ω ) ,
2 (Ω )
1

where δ (R) > 0 can be chosen as small as we want if R is sufficiently large. The
same is true (with some evident changes) for p = ∞ and for n = 2. Hence, we have
for arbitrarily small ε > 0 and for arbitrarily small δ (R) > 0 that

[1 − ε − δ (R) − (1/ε − 1)δ (R)] ∇ f 2L2 (Ω )

− [R + δ (R) + (1/ε − 1)(R + δ (R))]  f 2L2 (Ω ) ≤ (Hm f , f )L2 (Ω )
≤ [1 + ε + δ (R) + (1/ε − 1)δ (R)] ∇ f 2L2 (Ω ) + [(2 + 1/ε )(R + δ (R))]  f 2L2 (Ω ) .

Choosing ε > 0 arbitrarily small and R > 0 such that δ (R) = ε 2 , we obtain the
required estimate (33.3). 
33 The Magnetic Schrödinger Operator 351

Exercise 33.1. Prove the previous lemma in the cases p = s = ∞ for n ≥ 2 and
p, s < ∞ for n = 2.

This lemma implies that there exists μ0 > 0 such that Hm + μ0 I is positive and
2
((Hm + μ0 I) f , f )L2 (Ω )  f W 1 (Ω ) .
2

This fact implies that there is a Friedrichs self-adjoint extension of the positive op-
erator Hm + μ0 I, denoted by (Hm + μ0 I)F (see, for example [5]), with the domain
◦
D((Hm + μ0 I)F ) = { f ∈ W 12 (Ω ) : (Hm + μ0 I) f ∈ L2 (Ω )}, (33.4)
◦
where W 12 (Ω ) is the closure of C0∞ (Ω ) with respect to the norm of the Sobolev space
W21 (Ω ). Hence, the Friedrichs extension (Hm )F of Hm can be defined as (Hm )F :=
(Hm + μ0 I)F − μ0 IF with the same domain (33.4).
 ∈ L∞ (Ω ) and ∇· W
Exercise 33.2. Show that if W  ,V ∈ L p (Ω ) with some n ≤ p ≤ ∞
for n ≥ 3 and with some 2 < p ≤ ∞ for n = 2, then
◦
D((Hm )F ) = W 12 (Ω ) ∩W22 (Ω ).

In particular, for Ω = Rn we obtain in this case that

D((Hm )F ) = W22 (Rn ).

Hint. Represent first Hm in the form

 (x)∇u + [|W
Hm u = −Δ − 2iW  |2 +V − i∇ · W
 ]u

and then use the same technique and the same embedding theorems for Sobolev
spaces as in the proof of Lemma 33.1.

Remark 33.2. The Friedrichs self-adjoint extension of the magnetic Schrödinger

operator Hm exists under much “broader” assumptions for the coefficients V and W
 ∈ L2 (Ω ) and V ∈ L1 (Ω ) but
than in Lemma 33.1. Namely, if we just assume that W
V ≥ 0, then since for all f ∈ C0∞ (Ω ) we have
 2
  )f
(Hm f , f )L2 (Ω ) = (∇ + iW  2 + (V f , f )L2 (Ω ) ≥ 0,
L (Ω )

we may conclude that for all μ0 > 0, Hm + μ0 I is positive, and thus the Friedrichs
self-adjoint extension exists (see, for example, [5]). But in this so-called “general”
case we cannot characterize the domain of (Hm )F constructively. We can say only
that
352 Part III: Operator Theory and Integral Equations

 ) : V 1/2 f ∈ L2 (Ω ) and Hm f ∈ L2 (Ω )}.

D((Hm )F ) = { f ∈ D(∇ + iW

But even in this “general” case we may prove the diamagnetic inequality (see [35]).
For all t > 0 we may consider (using von Neumann’s spectral theorem, see The-
orem 27.13) the self-adjoint operators
 ∞
e−t(Hm )F f (x) := e−t λ dEλ f (x),
0
 ∞ (33.5)
(0)
e−t(−Δ )F f (x) := e−t λ dEλ f (x),
0

(0)
where Eλ and Eλ are the spectral families corresponding to the self-adjoint opera-
tors (Hm )F and (−Δ )F , respectively.
Theorem 33.3. Assume that W  ∈ L2 (Ω ), V ∈ L1 (Ω ), V ≥ 0, and that these poten-
tials are real-valued. Then for all f ∈ L2 (Ω ), t > 0, and μ > 0 we have that

|e−t(Hm +μ I)F f (x)| ≤ e−t(−Δ +μ )F | f (x)| (33.6)

holds for almost every x ∈ Ω .

Proof. For brevity we denote ∇ + iW  by D . We have the following two lemmas
W
(which are also of independent interest).
Lemma 33.4. For all f ∈ C0∞ (Ω ) we have

|DW
 f (x)| ≥ |∇| f (x)|| (33.7)

almost everywhere.
Proof. Indeed,

∇| f (x)|2 = ∇( f (x) f (x)) = f ∇ f + f ∇ f = f DW

 f + f DW
 f = 2 Re( f DW
 f ).

This is equivalent to
 f ).
2| f |∇| f | = 2 Re( f DW

The latter equality implies that

|DW
 f || f | ≥ | f ||∇| f ||.

Thus, the lemma is proved. 

Lemma 33.5. For all f ∈ C0∞ (Ω )
and ϕ ≥ 0 sufficiently smooth we have that
   
f
Re DW
 ϕ DW
 f ≥ ∇ϕ ∇| f | (33.8)
|f|

almost everywhere.
33 The Magnetic Schrödinger Operator 353

Proof.Without loss of generality we may assume that f = 0; otherwise, we consider

fε = | f |2 + ε and then take the limit ε → 0+. We have
   
f f  f ϕ = ϕ ∇ f + f ∇ϕ + iW f ϕ
DW
 ϕ =∇ ϕ + iW
|f| |f| |f| |f| |f| |f|

DW f f ∇| f | f
=ϕ − + ∇ϕ .
|f| | f |2 |f|

This equality implies that

   
f DW
 f DW f f ∇| f | f
DW
 ϕ DW
 f =ϕ −  f +
DW ∇ϕ , DW
 f
|f| |f| |f|2 |f|

so
   
f |D f |2 ∇| f | Re( f DW f)
Re DW
 ϕ DW
 f =ϕ W − ϕ 2 Re( f DW
 f)+ ∇ϕ .
|f| |f| |f| |f|

 f ) for f = f 1 + i f 2 , we obtain
Calculating Re( f DW


 f ) = Re[( f 1 − i f 2 )(∇ f 1 + i∇ f 2 + i∇ f 2 − W f 2 )] = f 1 ∇ f 1 + f 2 ∇ f 2 = | f |∇| f |.
Re( f DW

Thus,
   
f |DW
 f|
2
∇| f | | f |∇| f |
Re DW
 ϕ DW
 f = ϕ − ϕ 2 | f |∇| f | + ∇ϕ
|f| |f| |f| |f|
|DW
 f | − |∇| f ||
2 2
=ϕ + ∇ϕ ∇| f | ≥ ∇ϕ ∇| f |
|f|

by Lemma 33.4 and the fact that ϕ ≥ 0. 

To end the proof of Theorem 33.3 we consider μ > 0. Using these two lemmas
we obtain
     
f
∇ϕ ∇| f |dx + μ ϕ | f |dx ≤ Re DW  ϕ D  f dx + μ ϕ | f |dx
Ω Ω Ω |f| W
Ω
 
 
f
≤ DW  ϕ DW  f dx + μ ϕ | f |dx ,
Ω |f| Ω

since Re z + a ≤ |z + a| for real a. Using now integration by parts in both integrals

we have that
354 Part III: Operator Theory and Integral Equations

   
f f

ϕ (−Δ + μ )| f |dx ≤ −DW
2
ϕ + μ f ϕ dx ,
 f
Ω Ω |f| |f|

i.e.,

f
((−Δ + μ )| f |, f )L2 (Ω ) ≤ ϕ (−DW
2
 + μ ) f dx

Ω |f|

≤ ϕ |(−DW
 + μ ) f |dx = (|(−DW
2
 + μ ) f |, ϕ )L2 (Ω ) .
2
Ω

Since (−DW 2 + μ I)−1 exists, by introducing f := (−D2 + μ I)−1 u we can rewrite

 
W
the latter inequality as
−1
((−Δ + μ )| f |, ϕ )L2 (Ω ) = (| f |, (−Δ + μ )ϕ )L2 (Ω ) = (|(−DW
 + μ I) u|, ψ )L2 (Ω )
2

≤ (|u|, ϕ )L2 (Ω ) = (|u|, (−Δ + μ I)−1 ψ )L2 (Ω ) = ((−Δ + μ I)−1 |u|, ψ )L2 (Ω ) ,

where we have used the self-adjointness of all operators and the notation ψ = (−Δ +
μ I)ϕ . Hence, for arbitrary ψ ≥ 0 sufficiently smooth we obtain the inequality
−1 −1
 + μ I) u|, ψ )L2 (Ω ) ≤ ((−Δ + μ ) |u|, ψ )L2 (Ω ) .
(|(−DW
2

Since ψ is an arbitrary function of such type, we may conclude from here that for
every u ∈ L2 (Ω ) we have that
−1 −1
 + μ I) u(x)| ≤ (−Δ + μ I) |u|(x)
|(−DW
2

almost everywhere. Iterating the latter inequality, we obtain for all m ∈ N that
−m
 + μ I)
|(−DW
2
u(x)| ≤ (−Δ + μ I)−m |u|(x).

Hence, for every u ∈ L2 (Ω ) we have


−t(−D2 +μ ) m
m −m
|e W u(x)| ≤ lim (−DW
2
 +μ + ) u(x)
m→∞ t t
m m −m
≤ lim (−Δ + μ + ) |u(x)|
m→∞ t t
m 
m −m
= lim (−Δ + μ + ) |u(x)| = e−t(−Δ +μ ) |u(x)|
m→∞ t t

almost everywhere. Thus (33.6) is proved in the case V = 0. In order to add V ≥ 0

2 + μ I we repeat the above procedure and easily obtain that for all u ∈ L2 (Ω )
to −DW

and ψ ≥ 0 sufficiently smooth, we have
33 The Magnetic Schrödinger Operator 355

−1 −1
 + μ +V ) u|, ψ )L2 (Ω ) ≤ (|u|, (−Δ + μ +V ) ψ )L2 (Ω )
(|(−DW
2

≤ (|u|, (−Δ + μ )−1 ψ )L2 (Ω ) = ((−Δ + μ )−1 |u|, ψ )L2 (Ω ) ,

since
(−Δ + μ +V )−1 ψ ≤ (−Δ + μ )−1 ψ

almost everywhere. This completes the proof of Theorem 33.3. 

There are many applications of the diamagnetic inequality. We will consider
some of them. If A is a nonnegative self-adjoint operator acting in L2 (Ω ), its heat
kernel P(x, y,t) (if it exists) is defined to be a function such that for every t > 0 the
self-adjoint operator e−tA is an integral operator with this kernel (see for comparison
Definition 22.6 and Chapter 45), i.e., for all f ∈ L2 (Ω ),

−tA
e f (x) = P(x, y,t) f (y)dy. (33.9)
Ω

Using this definition, we may conclude that if a heat kernel exists, then for every
μ > 0 the inverse operator (A + μ I)−1 (which exists) is an integral operator with
kernel  ∞
G(x, y, μ ) = e−μ t P(x, y,t) f (y)dt. (33.10)
0

Indeed, by von Neumann’s theorem for A (see Chapter 27), we have that for every
f ∈ L2 (Ω ),  ∞
e−tA f (x) = e−t λ dEλ f ,
0

where {Eλ } is the spectral family corresponding to A.

Since (A + μ I)−1 exists for every μ > 0 and it is self-adjoint, it follows that for
every f ∈ L2 (Ω ) we have
 ∞  ∞  ∞
1
(A + μ I)−1 f (x) = dE f (x) = dEλ f (x) e−t(λ +μ ) dt
0 λ +μ λ 0 0
 ∞  ∞  ∞
= e−t μ dt e−t λ dEλ f (x) = e−t μ P(x, y,t)dt.
0 0 0

There is (at least) one quite general situation in which the heat kernel exists and
has “good” estimates. Let us assume that Ω ⊂ Rn is a bounded domain with smooth
boundary (the smoothness is required for the Sobolev embedding theorem). We con-
sider the magnetic Schrödinger operator Hm in Ω with electric potential V ≥ 0 and
with magnetic potential W satisfying all assumptions of Lemma 33.1. In this case
Hm and H0 = −Δ have Friedrichs self-adjoint extensions, which are denoted by the
same symbols Hm and H0 , respectively. We have the following theorem.
Theorem 33.6. Under the conditions of Lemma 33.1 for μ > 0, the resolvent (Hm +
μ I)−1 is an integral operator with kernel G(x, y, μ ), called the Green’s function
356 Part III: Operator Theory and Integral Equations

corresponding to the Friedrichs extension of Hm . Moreover, the following estimates

are valid:  √
|x − y|2−n e− μ |x−y| , n ≥ 3,
|G(x, y, μ )| ≤ C √ (33.11)
1 + | log( μ |x − y|)|, n = 2.

Proof. Using Lemma 33.1, we conclude that for μ > 0 sufficiently large and for all
◦
f ∈ W 12 (Ω ) we have

2
((Hm + μ I) f , f )L2 (Ω ) ≥ γ  f W 1 (Ω ) , γ > 0.
2

◦
Since the embedding W 12 (Ω ) → L2 (Ω ) is compact (see Lemma 30.15), we have
that (Hm + μ I)−1 is compact. Using now the Riesz–Schauder and Hilbert–Schmidt
theorems (see Theorem 28.10), we conclude that the spectrum σ (Hm ) = {λ j }∞j=1
is discrete and of finite multiplicity with only one accumulation point at infinity.
The corresponding normalized eigenfunctions {ϕ j }∞j=1 form an orthonormal basis
in L2 (Ω ) such that the spectral family for Hm is defined as

Eλ f (x) = ∑ f j ϕ j (x), f j = ( f , ϕ j )L2 (Ω ) .

λ j <λ

Hence
 ∞ ∞
e−tHm f (x) =
0
e−t λ dEλ f (x) = ∑ e−t λ j f j ϕ j (x)
j=1
  
 
∞ ∞
= ∑ e−t λ j Ω
f (y)ϕ j (y)dy ϕ j (x) = ∑ e−t λ j ϕ j (x)ϕ j (y) f (y)dy,
Ω
j=1 j=1

and the heat kernel of Hm will be equal in this case to

∞
P(x, y,t) = ∑ e−t λ j ϕ j (x)ϕ j (y). (33.12)
j=1

It must be mentioned here that all these equalities (and operations) are considered
in the sense of L2 (Ω ). The equality (33.12) implies that (Hm + μ I)−1 is an integral
operator with kernel G(x, y, μ ) defined by
 ∞ ∞  ∞ ∞
ϕ j (x)ϕ j (y)
G(x, y, μ ) = e−μ t P(x, y,t)dt = ∑ ϕ j (x)ϕ j (y) e−t(λ j +μ ) dt =
. ∑
0 j=1 0 j=1 λ j + μ
(33.13)
This function G(x, y, μ ) is called the Green’s function of the Friedrichs extension
of the magnetic Schrödinger operator Hm in the bounded domain. To obtain the
estimates (33.11) we proceed as follows. It is known that the heat kernel P0 (x, y,t)
33 The Magnetic Schrödinger Operator 357

for H0 in the whole of Rn (see, for example, Chapter 22) is equal to

P0 (x, y,t) = (4π t)−n/2 e−|x−y|

2 /(4t)
. (33.14)

At the same time, the heat kernel P0 (x, y,t) of H0 in the bounded domain Ω satisfies
the following boundary value problem:
⎧
⎪  
⎨∂t P0 (x, y,t)
⎪ = Δx P0 (x, y,t), x, y ∈ Ω ,t > 0,

P0 (x, y,t) = 0, x ∈ ∂ Ω , y ∈ Ω , t > 0,
⎪
⎪ ∂Ω
⎩P (x, y, 0) = δ (x − y).
0

If we define P0 (x, y,t) = P0 (x, y,t) + R(x, y,t), then R(x, y,t) has to satisfy
⎧
⎪
⎨∂t R(x, y,t) = Δx R(x, y,t), x, y ∈ Ω , t > 0,
R(x, y,t)|∂ Ω = −P0 (x, y,t), x ∈ ∂ Ω , y ∈ Ω , t > 0
⎪
⎩
R(x, y, 0) = 0, x, y ∈ Ω .

But −P0 (x, y,t) < 0 for x, y ∈ Ω ,t > 0 (see (33.14)). Using then the maximum prin-
ciple for the heat equation (see Theorem 45.7) we obtain that R(x, y,t) ≤ 0 for all
x, y ∈ Ω and t > 0, i.e.,
0 ≤ P0 (x, y,t) ≤ P0 (x, y,t).

The next step is as follows: the diamagnetic inequality (33.6) leads in this case to
 

P(x, y,t) f (y)dy ≤ P0 (x, y,t)| f (y)|dy,
Ω Ω

which holds almost everywhere in x ∈ Ω and for all f ∈ L2 (Ω ). Using the Hardy–
Littlewood maximal function (see, e.g., [18]), we can obtain from the latter inequal-
ity that
|P(x, y,t)| ≤ (4π t)−n/2 e−|x−y| /(4t) , x, y ∈ Ω , t > 0.
2

Using this and (33.10), we get (see Example 22.8)

 ∞
e−μ t (4π t)−n/2 e−|x−y|
2 /(4t)
|G(x, y, μ )| ≤ dt
0
 (2−n)/2
|x − y| √
= (2π )−n/2 √ K(n−2)/2 ( μ |x − y|),
μ

where Kν (z) is the Macdonald function of order ν . Using the asymptotic expansion
for Kν (z) for z → 0 and z → ∞ (see, for example, [23]), we can obtain the following
inequalities (see also the straightforward calculations in Example 22.8):
358 Part III: Operator Theory and Integral Equations
 √
|x − y|2−n e− μ |x−y| , n ≥ 3,
|G(x, y, μ )| ≤ C √
1 + | log( μ |x − y|)|, n = 2,

where x, y ∈ Ω and the constant C > 0 depends only on the dimension n. Thus,
Theorem 33.6 is completely proved. 

One more application of the diamagnetic inequality concerns the estimates of the
normalized eigenfunctions of Hm .

Corollary 33.7. Let ϕ be a normalized eigenfunction of Hm with corresponding

eigenvalue λ > 0. Then
 e n/4
 ϕ  L ∞ (Ω ) ≤ λ n/4 . (33.15)
2π n

Proof. Applying (33.6) to ϕ (x), we obtain (ϕ L2 (Ω ) = 1)


|e−t λ ϕ (x)| = |e−tHm ϕ | ≤ (4π t)−n/2 e−|x−y| /(4t) |ϕ (y)|dy
2

Ω
 1/2
(4π t)−n e−|x−y| /(2t) dy
2
≤ ϕ  L 2 (Ω )
Ω
 1/2
(4π t)−n e−|x−y|
2 /(2t)
≤ dy
Rn
 1/2
−n/2 n/4 −|y|2 /2
= (4π t) t e dy
Rn

= (4π t)−n/2t n/4 (2π )n/4 = π −n/4 2−3n/4t −n/4 ,

i.e., for all t > 0,

|ϕ (x)| ≤ π −n/4 2−3n/4t −n/4 et λ .

Taking the infimum of the right-hand side with respect to t > 0, we

obtain (33.15). 
Chapter 34
Integral Operators with Weak Singularities.
Integral Equations of the First and Second
Kinds.

Let Ω be a bounded domain in Rn . Then


A f (x) = K(x, y) f (y)dy
Ω

is an integral operator in L2 (Ω ) with kernel K.

Definition 34.1. An integral operator A is said to be an operator with weak singu-
larity if its kernel K(x, y) is continuous for all x, y ∈ Ω , x = y, and there are positive
constants M and α ∈ (0, n] such that

|K(x, y)| ≤ M|x − y|α −n , x = y.

Remark 34.2. If K(x, y) is continuous for all x, y ∈ Ω and bounded, then this integral
operator is considered also an operator with weak singularity.

If we have two integral operators A1 and A2 with kernels K1 and K2 , respectively,

then we can consider their composition as follows:
   
(A1 ◦ A2 ) f (x) = K1 (x, y)A2 f (y)dy = K1 (x, y) K2 (y, z) f (z)dz dy
Ω Ω Ω
   
= K1 (x, y)K2 (y, z)dy f (z)dz,
Ω Ω

and analogously
  
(A2 ◦ A1 ) f (x) = K2 (x, y)K1 (y, z)dy f (z)dz,
Ω Ω

assuming that the conditions of Fubini’s theorem are fulfilled.

c Springer International Publishing AG 2017 359

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 34
360 Part III: Operator Theory and Integral Equations

So, we may conclude that the compositions A1 ◦ A2 and A2 ◦ A1 are again integral
operators with kernels

K(x, y) = K1 (x, z)K2 (z, y)dz,
Ω (34.1)
 y) =
K(x, K2 (x, z)K1 (z, y)dz,
Ω

 y), that is, A1 ◦ A2 = A2 ◦ A1 .

respectively. In general, K(x, y) = K(x,
Returning to integral operators with weak singularities, we obtain a very impor-
tant property of them.
Lemma 34.3. If A1 and A2 are integral operators with weak singularities, then
A1 ◦ A2 and A2 ◦ A1 are also integral operators with weak singularities. Moreover, if

|K1 (x, y)| ≤ M1 |x − y|α1 −n and |K2 (x, y)| ≤ M2 |x − y|α2 −n , (34.2)

then there is M > 0 such that

⎧
⎪ α +α −n
⎨|x − y| 1 2 , α1 + α2 < n,
|K(x, y)| ≤ M 1 + | log |x − y||, α1 + α2 = n, (34.3)
⎪
⎩
1, α1 + α2 > n.

 y).
The same estimates hold for the kernel K(x,

Proof. Using (34.1) and (34.2), we obtain


|K(x, y)| ≤ M1 M2 |x − z|α1 −n |z − y|α2 −n dz.
Ω

If α1 + α2 < n, then changing the variable z = y + u|x − y|, we have

x − z = |x − y|(e0 − u), |e0 | = 1,

and 
α1 +α2 −n
|K(x, y)| ≤ M1 M2 |x − y| |u − e0 |α1 −n |u|α2 −n du. (34.4)
Rn

In order to estimate the latter integral we consider three different cases:

|u| ≤ 1/2, 1/2 ≤ |u| ≤ 3/2, |u| ≥ 3/2.

In the first case,

|u − e0 | ≥ |e0 | − |u| ≥ 1 − 1/2 = 1/2,
34 Integral Operators with Weak Singularities 361

and therefore
 
α1 −n α2 −n n−α1
|u − e0 | |u| du ≤ 2 |u|α2 −n du
|u|≤1/2 |u|≤1/2
 1/2 
2n−α1 −α2 n−1
= 2n−α1 rα2 −1 dr dθ = |S |,
0 Sn−1 α2

where |Sn−1 | denotes the area of the unit sphere in Rn .

In the third case,

2 |u|
|u − e0 | ≥ |u| − |e0 | ≥ |u| − 1 ≥ |u| − |u| = ,
3 3
and we have analogously

|u − e0 |α1 −n |u|α2 −n du
|u|≥3/2
 ∞
2n−α1 −α2 3α2 n−1
≤ 3n−α1 |Sn−1 | rα1 +α2 −n−1 dr = |S |.
3/2 n − α1 − α2

In the case 1/2 ≤ |u| ≤ 3/2 we have that |u − e0 | ≤ 5/2, and so


|u − e0 |α1 −n |u|α2 −n du
1/2≤|u|≤3/2
 (34.5)
n−α2 α1 −n 2n−α1 −α2 5α1 n−1
≤2 |u − e0 | du = |S |.
|u−e0 |≤5/2 α1

Combining (34.4)–(34.5), we obtain (34.3) for the case α1 + α2 < n. It can be men-
tioned here that the estimate (34.3) in this case holds also in the case of an arbitrary
(not necessarily bounded) domain Ω .
If now α1 + α2 = n, then the proof of (34.3) will be a little bit different, and it
holds only for a bounded domain Ω . Indeed, for every z ∈ Ω and

|x − y| |x − y|
|x − z| ≤ or |z − y| ≤ ,
2 2
we have in both cases that

|K(x, y)| ≤ M1 M2 2n−α2 |x − y|α2 −n |x − z|α1 −n dz
Ω
 |x−y|/2
≤ M1 M2 2n−α2 |Sn−1 ||x − y|α2 −n rα1 −1 dr
0
M1 M2 n−1 M1 M2 n−1
= |S | or |S |. (34.6)
α1 α2
362 Part III: Operator Theory and Integral Equations

If z ∈ Ω does not belong to these balls with radius |x − y|/2, then we consider two
cases: |z − x| ≥ |z − y| and |z − x| ≤ |z − y|. In both cases we have
  d
dz dr
|K(x, y)| ≤ M1 M2 ≤ M1 M2 |Sn−1 |
Ω \Ω  |x − z|n |x−y|/2 r
(34.7)
2d
= M1 M2 |Sn−1 | log ,
|x − y|

where d = diam Ω . The estimates (34.6) and (34.7) give us (34.3) in the case α1 +
α2 = n.
If finally α1 + α2 > n, then since Ω is bounded, we can analogously obtain (34.3)
in this case. This finishes the proof. 

Remark 34.4. In the case α1 + α2 = n, since for all 0 < t < 1,

| logt| ≤ Cε t −ε , ε > 0,

instead of a logarithmic singularity in (34.3) we may consider a weak singularity for

the kernel K(x, y) as
|K(x, y)| ≤ Mε |x − y|−ε ,

where ε > 0 can be chosen appropriately.

Let A be an integral operator in L2 (Ω ) with weak singularity. Then since 0 <

α ≤ n, we have
 
|x − y|α −n dy ≤ β and |x − y|α −n dx ≤ β ,
Ω Ω

where 
β = sup |x − y|α −n dy < ∞.
x∈Ω Ω

Schur’s test (see Example 26.2) shows that A is bounded in L2 (Ω ) and

AL2 (Ω )→L2 (Ω ) ≤ M β .

We can prove even more.

Theorem 34.5. An integral operator with weak singularity is compact in L2 (Ω ).

Proof. Let us introduce the function

1, 0 ≤ t ≤ σ ,
χσ (t) =
0, t > σ .
34 Integral Operators with Weak Singularities 363

Then for all σ > 0 we may write

K(x, y) = χσ (|x − y|)K(x, y) + (1 − χσ (|x − y|))K(x, y) =: K1 (x, y) + K2 (x, y).

The integral operator with kernel K2 (x, y) is a Hilbert–Schmidt operator for all
σ > 0, since
  
|K2 (x, y)|2 dxdy ≤ M 2 |x − y|2α −2n dxdy
Ω Ω σ ≤|x−y|≤d

is finite. It is therefore compact in L2 (Ω ) (see Exercise 28.10). For the integral

operator A1 with kernel K1 (x, y) we proceed as follows:

A1 f 2L2 (Ω ) = (A1 f , A1 f )L2 (Ω ) = ( f , A∗1 ◦ A1 f )L2 (Ω ) , (34.8)

where A∗1 is the adjoint operator with kernel

K1∗ (x, y) = χσ (|x − y|)K(y, x),

which is also an operator with weak singularity. Using Lemma 34.3, we can estimate
the right-hand side of (34.8) from above as
   
1
|Kσ (x, y)|| f (x)|| f (y)|dxdy ≤ |Kσ (x, y)|| f (x)|2 dxdy
Ω Ω 2 Ω Ω
 
1
+ |Kσ (x, y)|| f (y)|2 dxdy,
2 Ω Ω
(34.9)

where Kσ (x, y) is the kernel of the operator with weak singularity, i.e.,
⎧
⎪ 2α −n , α < n/2,
⎨|x − y|
|Kσ (x, y)| ≤ M |x − y|−ε , α = n/2,
⎪
⎩
1, α > n/2,

where ε > 0 can be chosen as small as we want.

Let us note also that the definition of χσ (t) implies that Kσ (x, y) = 0 for |x − y| >
2σ . Thus (see (34.8) and (34.9)) we have (α < 2n)

A1 f 2L2 (Ω ) ≤ M |x − y|2α −n | f (x)|2 dxdy
|x−y|≤2σ
 
≤M | f (x)|2 |x − y|2α −n dydx
Ω |x−y|≤2σ
(2σ )2α
= M  f 2L2 (Ω ) |Sn−1 | → 0,
2α
364 Part III: Operator Theory and Integral Equations

as σ → 0. This means that

A1 L2 (Ω )→L2 (Ω ) → 0, σ → 0.

The same fact is valid for the cases α ≥ n/2. Thus,

A − A2 L2 (Ω )→L2 (Ω ) ≤ A1 L2 (Ω )→L2 (Ω ) → 0

as σ → 0. But A2 is compact for every σ > 0, and therefore, A is also compact as

the limit of compact operators. This completes the proof. 
We want now to expand the analysis of integral operators with weak singularity
defined on domains in Rn to integral operators with weak singularity defined on
surfaces of dimension n − 1.
Assume that ∂ Ω is the boundary of a bounded domain of class C1 . This means,
roughly speaking, that at every point x ∈ ∂ Ω there is a tangent plane with normal
vector ν (x) that is continuous function on ∂ Ω , and the surface differential dσ (y) in
a neighborhood of each point x ∈ ∂ Ω satisfies the inequality (see [22])

dσ (y) ≤ c0 ρ n−2 dρ dθ ,

where (ρ , θ ) are the polar coordinates in the tangent plane with origin x, and c0 is
independent of x. According to the dimension n − 1 of the surface ∂ Ω , an integral
operator in L2 (∂ Ω ), i.e.,

A f (x) = K(x, y) f (y)dσ (y),
∂Ω

is said to be with weak singularity if its kernel K(x, y) is continuous for all x, y ∈ ∂ Ω ,
x = y, and there are constants M > 0 and α ∈ (0, n − 1] such that

|K(x, y)| ≤ M|x − y|α −(n−1) , x = y.

If K(x, y) is continuous everywhere, we require that K is bounded on ∂ Ω × ∂ Ω . We

can provide now the following theorem.
Theorem 34.6. An integral operator with weak singularity is compact in L2 (∂ Ω ).
Proof. The proof is the same as that for Theorem 34.5. 
For Banach spaces (i.e., complete normed spaces) the same definition of com-
pact operator holds (see Definition 28.8). We will need the compactness of these
integral operators in the Banach space C(Ω ). Let Ω be a compact set in Rn . The
Banach space C(Ω ) is defined as the set of all complex-valued functions ϕ (x) that
are continuous on Ω with norm

ϕ C(Ω ) := max |ϕ (x)| ≡ ϕ L∞ (Ω ) .

Ω
34 Integral Operators with Weak Singularities 365

We will need also the famous Ascoli–Arzelà theorem (for a proof, see [22]).
Theorem 34.7. A set U ⊂ C(Ω ) is relatively compact if and only if
(1) there is a constant M > 0 such that for all ϕ ∈ U we have ϕ L∞ (Ω ) ≤ M (uni-
form boundedness);
(2) for every ε > 0 there is δ > 0 such that

|ϕ (x) − ϕ (y)| < ε

for all x, y ∈ Ω with |x − y| < δ and all ϕ ∈ U (equicontinuity).

Theorem 34.8. An integral operator with continuous kernel is compact on C(Ω ).
Proof. The result follows straightforwardly from the Ascoli–Arzelà theorem. 
Theorem 34.9. An integral operator with weak singularity is compact on C(Ω ).
Proof. The proof follows from Theorem 34.8. Indeed, let us choose a continuous
function h as ⎧
⎪
⎨0, 0 ≤ t ≤ 1/2,
h(t) = 2t − 1, 1/2 ≤ t ≤ 1,
⎪
⎩
1, t ≥ 1,

and the integral operator Ak with kernel Kk (x, y) given by

h(k|x − y|)K(x, y), x = y,

Kk (x, y) =
0, x = y.

The kernel Kk (x, y) is continuous for every k = 1, 2, . . ., and therefore Ak is compact

on C(Ω ). Moreover,
 
 
|Aϕ (x) − Ak ϕ (x)| =  (1 − h(k|x − y|))K(x, y)ϕ (y)dy

Ω

≤ ϕ L∞ (Ω ) |K(x, y)|dy
|x−y|≤1/k

≤ M ϕ L∞ (Ω ) |x − y|α −n dy → 0
|x−y|≤1/k

as k → ∞ uniformly in x ∈ Ω . Thus, A is compact as the norm limit of compact

operators. 
There is a very useful and quite general result for integral operators with weak
singularity for both domains and surfaces in Rn .
Theorem 34.10. An integral operator with weak singularity transforms bounded
functions into continuous functions.
366 Part III: Operator Theory and Integral Equations

Proof. We give the proof for domains in Rn . The proof for surfaces in Rn is the
same. Let x, y ∈ Ω and |x − y| < δ . Then

|A f (x) − A f (y)| ≤ (|K(x, z)| + |K(y, z)|)| f (z)|dz
|x−z|<2δ

+ |K(x, z) − K(y, z)|| f (z)|dz
Ω \{|x−z|<2δ }

≤ M  f L∞ (Ω ) (|x − z|α −n + |y − z|α −n )dz
|x−z|<2δ

+  f L∞ (Ω ) |K(x, z) − K(y, z)|dz =: I1 + I2 .
Ω \{|x−z|<2δ }

Since |z − y| ≤ |x − z| + |x − y|, we have

 3δ
(3δ )α
I1 ≤ 2M  f L∞ (Ω ) |Sn−1 | rα −1 dr = 2M|Sn−1 |  f L∞ (Ω ) →0
0 α

as δ → 0. On the other hand, for |x − y| < δ and |x − z| ≥ 2δ we have that

|y − z| ≥ |x − z| − |x − y| > 2δ − δ = δ .

So the continuity of the kernel K outside of the diagonal implies that

K(x, z) − K(y, z) → 0, δ → 0,

uniformly in z ∈ Ω \ {|x − z| < 2δ }. Since Ω is bounded, we obtain that I2 → 0 as

δ → 0. This completes the proof. 
Exercise 34.1. Prove that if A is as in Theorem 34.10, then f (x) + A f (x) ∈ C(Ω )
for f ∈ L2 (Ω ) implies f ∈ C(Ω ).
We are now in a position to extend the solvability conditions (Fredholm alternative;
see Theorem 28.16) to equations in Hilbert space with compact but not necessarily
self-adjoint operators.
Theorem 34.11 (Fredholm alternative II). Suppose A : H → H is compact. For
all μ ∈ C either the equations

(I − μ A) f = g, (I − μ A∗ ) f  = g ,

have the unique solutions f and f  for any given g and g from H or the correspond-
ing homogeneous equations

(I − μ A) f = 0, (I − μ A∗ ) f  = 0 (34.10)
34 Integral Operators with Weak Singularities 367

have nontrivial solutions such that

dim N(I − μ A) = dim N(I − μ A∗ ) < ∞,

and in this case equations (34.10) have solutions if and only if

g⊥N(I − μ A∗ ) ⇔ g ∈ R(I − μ A),


g ⊥N(I − μ A) ⇔ g ∈ R(I − μ A∗ ),

respectively.

Proof. Riesz’s lemma (see Theorem 28.14) and Exercise 26.7 give

R(I − μ A) = N(I − μ A∗ )⊥ ,
R(I − μ A∗ ) = N(I − μ A)⊥ .

Let us first prove that one always has

dim N(I − μ A) = dim N(I − μ A∗ ).

These two dimensions are finite due to Riesz (see Proposition 28.13). Since every
compact operator is a norm limit of a sequence of operators of finite rank (see
Chapter 28 for details), for every μ ∈ C, μ = 0, we have

I − μ A = −μ A0 + (I − μ A1 ),

where A0 is of finite rank and μ A1  < 1. Then (I − μ A1 )−1 exists and

(I − μ A1 )−1 (I − μ A) = I − μ (I − μ A1 )−1 A0 =: I − A2 ,

where A2 is of finite rank too. Analogously, since (I − μ A∗1 )−1 exists, we must have

(I − μ A∗ )(I − μ A∗1 )−1 = I − μ A∗0 (I − μ A∗1 )−1 =: I − A∗2 ,

where A∗2 is adjoint to A2 and is of finite rank too. These representations allow us to
conclude that

g ∈ N(I − μ A) ⇔ g ∈ N(I − A2 ),
g ∈ N(I − μ A∗2 ) ⇔ (I − μ A∗1 )−1 g ∈ N(I − μ A∗ ).


Thus, it suffices to show that the numbers of independent solutions of the equations

g = A2 g, g = A∗2 g

are equal.
368 Part III: Operator Theory and Integral Equations

Since we know that the ranks of A2 and A∗2 are finite, we may represent the
mappings of the operators I − A2 and I − A∗2 as the mappings of matrices I − M2 and
I − M2∗ with adjoint matrices M2 and M2∗ . But the ranks of the adjoint matrices are
equal, and therefore the numbers of independent solutions of the equations g = A2 g
and g = A∗2 g are equal.
The next step is the following: if R(I − μ A) = H, then N(I − μ A∗ ) = {0}, and
consequently N(I − μ A) = {0} and R(I − μ A∗ ) = H (see Exercise 26.7). This means
that both (I − μ A)−1 and (I − μ A∗ )−1 exist, and the unique solutions of (34.10) are
given by
f = (I − μ A)−1 g, f  = (I − μ A∗ )−1 g .

If N(I − μ A) and N(I − μ A∗ ) are not zero, then R(I − μ A) and R(I − μ A∗ ) are proper
subspaces of H, and equations (34.10) have solutions if and only if

g ∈ R(I − μ A), g ∈ R(I − μ A∗ ).

This is equivalent (see Exercise 26.7) to

g⊥N(I − μ A∗ ), g ⊥N(I − μ A).

This completes the proof. 

We will now demonstrate this Fredholm alternative for integral operators. Let Ω ⊂
Rn be a domain, and let

A f (x) = K(x, y) f (y)dy
Ω

be a compact integral operator in L2 (Ω ). Then its adjoint is defined as


A∗ f (x) = K(y, x) f (y)dy.
Ω

Hence, the Fredholm alternative for these operators reads as follows: either the equa-
tions

f (x) − μ K(x, y) f (y)dy = g(x),
 Ω (34.11)
f  (x) − μ K(y, x) f  (y)dy = g (x),
Ω

are uniquely solvable for all g and g from L2 (Ω ) or the equations


f (x) = μ K(x, y) f (y)dy,
Ω (34.12)
f  (x) = μ K(y, x) f  (y)dy,
Ω
34 Integral Operators with Weak Singularities 369

have the same (finite) number of linearly independent solutions. And in this case,
equations (34.11) are solvable if and only if g and g are orthogonal to every solution
f and f  of the equations (34.12), respectively.

Definition 34.12. Equations (34.11) and (34.12) are called integral equations of the
second and first kinds, respectively.

Exercise 34.2. Consider in L2 (a, b) the integral equation

 b
ϕ (x) − ex−y ϕ (y)dy = f (x), x ∈ [a, b],
a

where f ∈ L2 (a, b). Solve this equation and formulate the Fredholm alternative for
it.

Example 34.13. (Boundary value problems) Consider the second-order ordinary

differential equation

a0 (x)u (x) + a1 (x)u (x) + a2 (x)u(x) = f (x)

on the interval [0, 1] with coefficients f , a2 ∈ L2 (0, 1), a1 ∈ W21 (0, 1) and with
smooth a0 (x) ≥ c0 > 0 subject to the boundary conditions

u(0) = u0 , u(1) = u1 .

Dividing this equation by a0 (x), we may consider the boundary value problem in
the form
u + a1 (x)u + a2 (x)u = f , u(0) = u0 , u(1) = u1 .

Using Green’s function G(x, y) of the form

y(1 − x), 0 ≤ y ≤ x ≤ 1,
G(x, y) =
x(1 − y), 0 ≤ x ≤ y ≤ 1,

we can rewrite this boundary value problem as

 1
u(x) = ϕ0 (x) + G(x, y)(a1 (y)u + a2 (y)u − f (y))dy,
0

where ϕ0 (x) = u0 (1 − x) + u1 x. Integration by parts implies

 1
u(x) = ϕ0 (x) − G(x, y) f (y)dy + G(x, y)a1 (y)u(y)|10
0
 1  1
− ∂y G(x, y)a1 (y) + G(x, y)a1 (y) u(y)dy + G(x, y)a2 (y)u(y)dy.
0 0
370 Part III: Operator Theory and Integral Equations

Since G(x, 1) = G(x, 0) = 0, this equation can be rewritten as

 1
0 (x) −
u(x) = ϕ K(x, y)u(y)dy,
0

where  1
0 (x) = ϕ0 (x) −
ϕ G(x, y) f (y)dy
0

and
K(x, y) = ∂y G(x, y)a1 (y) + G(x, y)a1 (y) − G(x, y)a2 (y).

Exercise 34.3. (1) Prove that K(x, y) is a Hilbert–Schmidt kernel on [0, 1] × [0, 1].
(2) Prove that the boundary value problem and this integral equation of the second
kind are equivalent.
(3) Formulate the solvability condition for the boundary value problem using the
Fredholm alternative for this integral operator.
Chapter 35
Volterra and Singular Integral Equations

In this chapter we consider integral equations of special types on a finite inter-

val [a, b]. We consider the Lebesgue space L∞ (a, b) and the Hölder space Cα (a, b)
(which are not Hilbert spaces but normed spaces) instead of the Hilbert space L2 :
The norms of the spaces L∞ (a, b) and Cα [a, b] are defined as follows:

 f L∞ (a,b) = inf{M : | f (x)| ≤ M a.e. on (a, b)},

| f (x) − f (y)|
 f Cα [a,b] =  f L∞ (a,b) + sup ,
x,y∈[a,b] |x − y|α

where 0 < α ≤ 1.
The fact that f belongs to the Hölder space Cα [a, b] is equivalent to the fact that
f ∈ L∞ (a, b) and there is a constant c0 > 0 such that for all h (sufficiently small),

| f (x + h) − f (x)| ≤ c0 |h|α ,

where x, x + h ∈ [a, b].

Definition 35.1. Integral equations in L∞ (a, b) of the form

 x
f (x) = K(x, y)ϕ (y)dy
a

and  x
ϕ (x) = f (x) + K(x, y)ϕ (y)dy, (35.1)
a

where x ∈ [a, b] and supx,y∈[a,b] |K(x, y)| < ∞, are called Volterra integral equations
of the first and second kinds, respectively.

c Springer International Publishing AG 2017 371

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 35
372 Part III: Operator Theory and Integral Equations

Theorem 35.2. For each f ∈ L∞ (a, b) the Volterra integral equation of the second
kind has a unique solution ϕ ∈ L∞ (a, b) such that

|ϕ (x)| ≤ eM(x−a)  f L∞ (a,b) (35.2)

for all x ∈ [a, b] and

ϕ L∞ (a,b) ≤  f L∞ (a,b) eM(b−a) , (35.3)

where M = supx,y∈[a,b] |K(x, y)|.

Proof. We introduce the iterations of the equation (35.1) by

 x
ϕ j+1 (x) := K(x, y)ϕ j (y)dy, j = 0, 1, 2, . . . ,
a

with ϕ0 = f . Let us prove by induction that

(M(x − a)) j
|ϕ j (x)| ≤  f L∞ (a,b) , j = 0, 1, . . . . (35.4)
j!

Indeed, this estimate clearly holds for j = 0. Assume that (35.4) has been proved
for some j ≥ 0. Then
 x  x
(M(y − a)) j
|ϕ j+1 (x)| ≤ |K(x, y)||ϕ j (y)|dy ≤ M  f L∞ (a,b) dy
a a j!
 x
(y − a) j (x − a) j+1
= M j+1  f L∞ (a,b) dy = M j+1  f L∞ (a,b) .
a j! ( j + 1)!

This proves (35.4).

Let us introduce the function
∞
ϕ (x) := ∑ ϕ j (x). (35.5)
j=0

Then from (35.4) we obtain for all x ∈ [a, b] that

∞
(M(x − a)) j
|ϕ (x)| ≤  f L∞ (a,b) ∑ =  f L∞ (a,b) eM(x−a) .
j=0 j!

Thus, the function ϕ (x) is well defined by the series (35.5), since this series is uni-
formly convergent with respect to x ∈ [a, b].
It remains now to show that this ϕ (x) solves (35.1). Since the series (35.5) con-
verges uniformly, we may integrate it term by term and obtain
35 Volterra and Singular Integral Equations 373
 x ∞  x ∞

a
K(x, y)ϕ (y)dy = ∑ K(x, y)ϕ j (y)dy = ∑ ϕ j+1 (x)
j=0 a j=0
∞
= ∑ ϕ j (x) + ϕ0 (x) − f (x) = ϕ (x) − f (x).
j=1

So (35.1) holds with this ϕ . The estimate (35.3) then follows immediately from
(35.2). Finally, the uniqueness of this solution follows from (35.3) too. 

Corollary 35.3. The homogeneous equation

 x
ϕ (x) = K(x, y)ϕ (y)dy
a

has only the trivial solution in L∞ (a, b).

Proof. The result follows from (35.3).

In general, integral equations of the first kind are more delicate with respect to
solvability than equations of the second kind. However, in some cases, Volterra inte-
gral equations of the first kind can be treated by reducing them to equations of the
second kind. Indeed, consider for x ∈ [a, b],
 x
K(x, y)ϕ (y)dy = f (x), (35.6)
a

and assume that the derivatives ∂∂Kx (x, y) and f  (x) exist and are bounded and that
K(x, x) = 0 for all x ∈ [a, b]. Then, differentiating with respect to x reduces (35.6) to
 x
∂K
ϕ (x)K(x, x) + (x, y)ϕ (y)dy = f  (x),
a ∂x
or
 x ∂K
f  (x) ∂ x (x, y)
ϕ (x) = − ϕ (y)dy. (35.7)
K(x, x) a K(x, x)

Exercise 35.1. Show that (35.6) and (35.7) are equivalent if f (a) = 0.

The second possibility occurs if we assume that

∂K
(x, y)
∂y

exists and is bounded and that K(x, x) = 0 for all x ∈ [a, b]. In this case, setting
 x
ψ (x) := ϕ (y)dy, ψ = ϕ
a
374 Part III: Operator Theory and Integral Equations

and performing integration by parts in (35.6) yields

 x  x
 ∂K
f (x) = K(x, y)ψ (y)dy = K(x, y)ψ (y)|xa − (x, y)ψ (y)dy
a a ∂y
 x
∂K
= K(x, x)ψ (x) − (x, y)ψ (y)dy,
a ∂y
or
 x ∂ K (x, y)
f (x) ∂y
ψ (x) = + ψ (y)dy.
K(x, x) a K(x, x)

There is an interesting generalization of equation (35.1) when the kernel has

weak singularities. More precisely, we consider (35.1) in the space L∞ (a, b) and
assume that the kernel K(x, y) satisfies the estimate

|K(x, y)| ≤ M|x − y|−α , x, y ∈ [a, b], x = y,

with some 0 < α < 1. If we consider again the iterations

 x
ϕ j (x) := K(x, y)ϕ j−1 (y)dy, j = 1, 2, . . . ,
a

with ϕ0 = f , then it can be proved by induction that for all x ∈ [a, b] we have
 j
M(x − a)1−α
|ϕ j (x)| ≤  f L∞ (a,b) , j = 0, 1, . . . .
1−α

Indeed, since this clearly holds for j = 0, assume that it has been proved for some
j ≥ 0. Then
 x
|ϕ j+1 (x)| ≤ |K(x, y)||ϕ j (y)|dy
a
 x
Mj
≤M |x − y|−α ((y − a)1−α ) j  f L∞ (a,b) dy
(1 − α ) j a
 x
M j+1
≤ ((x − a)1−α ) j  f L∞ (a,b) (x − y)−α dy
(1 − α ) j a
M j+1 (x − a)1−α
≤ ((x − a)1−α ) j  f L∞ (a,b)
(1 − α ) j 1−α
  j+1
M(x − a)1−α
=  f L∞ (a,b) .
1−α

If we assume now that

M(b − a)1−α
< 1,
1−α
35 Volterra and Singular Integral Equations 375

then the series

∞
∑ ϕ j (x)
j=0

converges uniformly on the interval [a, b], and the function ϕ defined by
∞
ϕ (x) := ∑ ϕ j (x)
j=0

solves therefore the inhomogeneous integral equation (35.1). Moreover, the follow-
ing estimates hold:

 f L∞ (a,b)
|ϕ (x)| ≤ , x ∈ [a, b],
M(x − a)1−α
1−
1−α

and
 f L∞ (a,b)
ϕ L∞ (a,b) ≤ .
M(b − a)1−α
1−
1−α

Exercise 35.2. Show that the Volterra integral equation of the first kind
 x
ϕ (x) = λ e−(x−y) ϕ (y)dy
a

has, for all λ , only the trivial solution in L∞ (a, b).

Definition 35.4. Let 0 < α < 1, ϕ ∈ Cα [−a, a], and suppose that ϕ is periodic, i.e.,
ϕ (−a) = ϕ (a). In this space an integral equation of the form
 a
ϕ (x + y)dy
ϕ (x) = f (x) + λ p. v. , λ ∈ C, (35.8)
−a y

is understood in the sense that

 a 
ϕ (x + y)dy ϕ (x + y)dy
p. v. = lim (35.9)
−a y ε →0 |y|≥ε ,y∈[−a,a] y

and the function ϕ is extended periodically (with period 2a) to the whole line.

Due to (35.9), we have that  a

dy
p. v. = 0.
−a y
376 Part III: Operator Theory and Integral Equations

Thus
 a  a  a
ϕ (x + y)dy ϕ (x + y) − ϕ (x) ϕ (x + y) − ϕ (x)
p. v. = p. v. dy = dy,
−a y −a y −a y

and the latter integral can be understood in the usual sense for periodic ϕ ∈
Cα [−a, a], since
 a   a  a
 ϕ (x + y) − ϕ (x)  |y|α aα
 dy ≤ c0 dy = 2c0 ξ α −1 dξ = 2c0 . (35.10)
 −a y −a |y| 0 α

Inequality (35.10) shows us that for every ϕ ∈ Cα [−a, a] the integral in (35.8) is
uniformly bounded and also periodic with period 2a. But even more is true.

Proposition 35.5. For every 2a-periodic ϕ ∈ Cα [−a, a] with 0 < α < 1 the integral
in (35.8) defines a 2a-periodic function of x that belongs to the same Hölder space
Cα [−a, a].

Proof. Let us denote by g(x) the integral in (35.8). For |h| > 0 sufficiently small we
have
 a  a
ϕ (x + h + y) − ϕ (x + h) ϕ (x + y) − ϕ (x)
g(x + h) − g(x) = dy − dy
−a y −a y
 
ϕ (x + h + y) − ϕ (x + h) ϕ (x + y) − ϕ (x)
= dy − dy
|y|≤3|h| y |y|≤3|h| y
 
ϕ (x + h + y) − ϕ (x) ϕ (x + y) − ϕ (x)
+ dy − dy
|y|≥3|h| y |y|≥3|h| y
=: I1 + I2 .

For the first integral I1 we have

 
|ϕ (x + y + h) − ϕ (x + h)| |ϕ (x + y) − ϕ (x)|
|I1 | ≤ dy + dy
|y|≤3|h| |y| |y|≤3|h| |y|
 
|y|α |y|α
≤ c0 dy + c0 dy
|y|≤3|h| |y| |y|≤3|h| |y|
 3|h|
(3|h|)α 4c0 3α α
≤ 4c0 ξ α −1 dξ = 4c0 = |h| . (35.11)
0 α α

For the estimation of I2 we first rewrite it as (we change variables in the first integral)
35 Volterra and Singular Integral Equations 377
 
ϕ (z + x) − ϕ (x) ϕ (z + x) − ϕ (x)
I2 = dz − dz
|z−h|≥3|h| z−h |z|≥3|h| z
  
1 1
= (ϕ (z + x) − ϕ (x)) − dz
|z|≥3|h| z−h z

ϕ (z + x) − ϕ (x)
− dz.
{|z−h|≥3|h|}\{|z|≥3|h|} z−h

Then we have
 
|ϕ (z + x) − ϕ (x)||h|dz |ϕ (z + x) − ϕ (x)|
|I2 | ≤ + dz
|z|≥3|h| |z| · |z − h| 2|h|≤|z|≤3|h| |z − h|
z∈[−a,a]
 
|z|α |z|α
≤ c0 |h| dz + c0 dz
a≥|z|≥3|h| |z| · 2|z|/3 2|h|≤|z|≤3|h| |z|/2
 a  3|h|
3c0
= 2· |h| ξ α −2 dξ + 4c0 ξ α −1 dξ
2 3|h| 0
a  
ξ α −1  (3|h|)α (3|h|)α −1 aα −1 4c0 3α α
= 3c0 |h|  + 4c0 = 3c0 |h| − + |h|
α − 1 3|h| α 1−α 1−α α
 
3α c0 α 4c0 3α α α 1 4
< |h| + |h| = c0 3 + |h|α , (35.12)
1−α α 1−α α

since 0 < α < 1. Estimates (35.11)–(35.12) show that this proposition is completely
proved. 

If we denote by  a
ϕ (x + y)dy
Aϕ (x) := p. v. (35.13)
−a y

a periodic linear operator on Cα [−a, a], 0 < α < 1, then Proposition 35.5 gives that
A is bounded in this space. But this operator is not compact there. Nevertheless, the
following holds.

Corollary 35.6. There is λ0 > 0 such that for all λ ∈ C, |λ | < λ0 , and periodic
f ∈ Cα [−a, a], 0 < α < 1, the integral equation (35.8) has a unique solution in
Cα [−a, a], 0 < α < 1.

Proof. Since the operator A from (35.13) is a bounded linear operator in the space
Cα [−a, a], it follows that
ACα →Cα ≤ c0

with some constant c0 > 0. If we choose now λ0 = 1/c0 , then for all λ ∈ C, |λ | < λ0 ,
the operator I − λ A will be invertible in the space Cα [−a, a], since

λ ACα →Cα < 1.

378 Part III: Operator Theory and Integral Equations

This fact implies that the integral equation (35.8) can be solved uniquely in this
space, and the unique solution ϕ can be obtained as

ϕ = (I − λ A)−1 f .

This is equivalent to the fact that (35.8) can be solved by iterations. 

Chapter 36
Approximate Methods

In this chapter we will study approximate solution methods for equations in a Hilbert
space H of the form
Aϕ = f , (I − A)ϕ = f (36.1)

with a bounded or compact operator A. The fundamental concept for solving equa-
tions (36.1) approximately is to replace them by the equations

An ϕn = fn , (I − An )ϕn = fn , (36.2)

respectively. For practical purposes, the approximating equations (36.2) will be cho-
sen so that they can be reduced to a finite-dimensional linear system.
We will begin with some general results that are the basis of our considerations.
Theorem 36.1. Let A : H → H be a bounded linear operator with bounded inverse
A−1 . Assume that the sequence An : H → H of bounded linear operators is norm
convergent to A, i.e.,
An − A → 0, n → ∞.

Then for all n such that  −1 

A (An − A) < 1,

the inverse operators A−1

n exist and
 −1 
 −1  A 
An  ≤ .
1 − A−1 (An − A)

Moreover, the solutions of (36.1) and (36.2) satisfy the error estimate
 −1   
A 
ϕn − ϕ  ≤ −1
(An − A)ϕ  +  fn − f  .
1 − A (An − A)

c Springer International Publishing AG 2017 379

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 36
380 Part III: Operator Theory and Integral Equations

Proof. Since A−1 exists, we may write

A−1 An = I − A−1 (A − An ).

Since  −1 
A (An − A) < 1

for n sufficiently large, for these values of n we have that

 −1
I − A−1 (A − An )

exists by the Neumann series. Thus,

 −1  −1
A−1 An = I − A−1 (A − An ) ,

or  −1
A−1 −1
n A = I − A (A − An ) ,

or  −1 −1
A−1 −1
n = I − A (A − An ) A .

The error estimate follows immediately from the representation

ϕn − ϕ = A−1 −1
n (A − An )ϕ + An ( f n − f ).

This completes the proof. 

Theorem 36.2. Assume that A−1n : H → H exist for all n ≥ n0 and that their norms
are uniformly bounded for such n. Let An − A → 0 as n → ∞. Then the inverse
operator A−1 exists and
 −1 
 −1  An 
A  ≤  
1 − A−1
n (An − A)

 
for all n ≥ n0 with A−1 
n (An − A) < 1.

Exercise 36.1. Prove Theorem 36.2 and obtain the error estimate in this case.

Definition 36.3. A sequence {An }∞ n=1 of compact operators in a Hilbert space H is

said to be collectively compact if for every bounded set U ⊂ H, the image

J = {An ϕ : ϕ ∈ U, n = 1, 2, . . .}

is relatively compact, i.e., if every sequence from J contains a convergent subse-

quence.
36 Approximate Methods 381

Exercise 36.2. Assume that a sequence of compact operators {An }∞

n=1 is collec-
tively compact and converges pointwise to A in H, i.e.,

lim An ϕ = Aϕ , ϕ ∈ H.
n→∞

Prove that the limit operator A is compact.

Exercise 36.3. Under the same assumptions for {An }∞

n=1 as in Exercise 36.2, prove
that
(An − A)A → 0, (An − A)An  → 0,

as n → ∞.

Theorem 36.4. Let A : H → H be a compact operator and let I − A be injective.

Assume that the sequence An : H → H is collectively compact and pointwise con-
vergent to A. Then for all n such that
 
(I − A)−1 (An − A)An  < 1,

the inverse operators (I − An )−1 exist and the solutions of (36.1) and (36.2) satisfy
the error estimate
   
1 + (I − A)−1 An 
ϕn − ϕ  ≤ (A n − A)ϕ  +  f n − f  .
1 − (I − A)−1 (An − A)An 

Proof. By Riesz’s theorem (see Theorem 28.15), the inverse operator (I − A)−1 ex-
ists and is bounded. Due to Exercise 36.3,

(An − A)An  → 0, n → ∞.

Therefore, for n sufficiently large we have

 
(I − A)−1 (An − A)An  < 1.

This fact allows us to conclude (as in Theorem 36.1) that (I − An )−1 exists and
 
 
(I − An )−1  ≤ 1 + (I − A)−1 An 
.
1 − (I − A)−1 (An − A)An 

The error estimate follows from this inequality and the representation
 
ϕn − ϕ = (I − An )−1 (An − A)ϕ + fn − f .

This yields the claim. 

382 Part III: Operator Theory and Integral Equations

Corollary 36.5. Let An be as in Theorem 36.4. Assume that the inverse operators
(I −An )−1 exist and are uniformly bounded for all n ≥ n0 . Then the inverse (I −A)−1
exists if  
(I − An )−1 (An − A)A < 1.

The solutions of (36.1) and (36.2) satisfy the error estimate

   
1 + (I − An )−1 A
ϕn − ϕ  ≤ (An − A)ϕ  +  f n − f  .
1 − (I − An )−1 (An − A)A

Theorem 36.6. Let A : H → H be a bounded linear operator with A < 1. Then
the successive approximations

ϕn+1 := Aϕn + f , n = 0, 1, . . . , (36.3)

converge for each f ∈ H and each ϕ0 ∈ H to the unique solution of (36.1).

Proof. The condition A < 1 implies the existence and boundedness of the inverse
operator (I − A)−1 and the existence of the unique solution of (36.1) as

ϕ = (I − A)−1 f .

It remains only to show that the successive approximations converge to ϕ for all
ϕ0 ∈ H. The definition (36.3) implies

ϕn+1 − ϕn  ≤ A ϕn − ϕn−1  ≤ · · · ≤ An ϕ1 − ϕ0  .

Hence for each p ∈ N we have

   
ϕn+p − ϕn  ≤ ϕn+p − ϕn+p−1  + · · · + ϕn+1 − ϕn 
 
≤ An+p−1 + An+p−2 + · · · + An ϕ1 − ϕ0 
An
≤ ϕ1 − ϕ0  → 0
1 − A

as n → ∞ uniformly in p ∈ N. This means that {ϕn } is a Cauchy sequence in the

Hilbert space H. Therefore, there exists a unique limit

ϕ = lim ϕn .
n→∞

It is clear that this ϕ solves (36.1) uniquely. 

We will return to the integral operators

A f (x) = K(x, y) f (y)dy, (36.4)
Ω
36 Approximate Methods 383

where K(x, y) is assumed to be in L2 (Ω × Ω ). In that case, as we know, A is compact

in L2 (Ω ).
Definition 36.7. A function Kn (x, y) ∈ L2 (Ω × Ω ) is said to be a degenerate kernel
if
n
Kn (x, y) = ∑ a j (x)b j (y),
j=1

with some functions a j , b j ∈ L2 (Ω ).

We consider integral equations of the second kind with a degenerate kernel
Kn (x, y), i.e.,
 n
ϕn (x) −
Ω
∑ a j (x)b j (y)ϕn (y)dy = f (x), (36.5)
j=1

in the form
n
ϕn (x) − ∑ γ j a j (x) = f (x),
j=1

where γ j = (ϕn , b j )L2 (Ω ) . This means that the solution ϕn of (36.5) is necessarily
represented as
n
ϕn (x) = f (x) + ∑ γ j a j (x) (36.6)
j=1

such that the coefficients γ j (which are to be determined) satisfy the linear system

n
γ j − ∑ γk (ak , b j )L2 (Ω ) = ( f , b j )L2 (Ω ) = f j , j = 1, 2, . . . , n. (36.7)
k=1

Hence, the solution ϕn of (36.5) (see also (36.6)) can be obtained whenever we
can solve the linear system (36.7) uniquely with respect to γ j .
Let us consider now integral equations of the second kind with compact self-
adjoint operator (36.4), i.e.,

ϕ (x) − Aϕ (x) = f (x). (36.8)

The main idea is to approximate the kernel K(x, y) from (36.8) by the degenerate
kernel Kn (x, y) from (36.5) such that

K(x, y) − Kn (x, y)L2 (Ω ×Ω ) → 0 (36.9)

as n → ∞ and such that in addition, the inverse operators (I − An )−1 exist and are
uniformly bounded in n.
In that case the system (36.7) is uniquely solvable, and we obtain an approximate
solution ϕn such that
384 Part III: Operator Theory and Integral Equations

ϕ − ϕn L2 (Ω ) → 0, n → ∞.

Indeed, equations (36.5) and (36.8) imply

(ϕ − ϕn ) − An (ϕ − ϕn ) = (A − An )ϕ .

Since (I − An )−1 exist and are uniformly bounded, we have

 
ϕ − ϕn  ≤ (I − An )−1  A − An  ϕ  → 0

as n → ∞ by (36.9). The unique solvability of (36.7) (or the uniqueness of ϕn ) fol-

lows from the existence of the inverse operators (I − An )−1 .
We may justify this choice of the degenerate kernel Kn (x, y) by the following
considerations. Let {e j }∞j=1 be an orthonormal basis in L2 (Ω ). Then K(x, y) ∈
L2 (Ω × Ω ) as a function of x ∈ Ω (with parameter y ∈ Ω ) can be represented by
∞
K(x, y) = ∑ (K(·, y), e j )L2 e j (x).
j=1

Then  
 n 
 
K(x, y) − ∑ (K(·, y), e j )L2 e j  →0
 j=1

L2 (Ω ×Ω )

as n → ∞, and we may consider the degenerate kernel Kn (x, y) in the form

n
Kn (x, y) = ∑ e j (x)b j (y),
j=1

where b j (y) = (K(·, y), e j )L2 . The system (36.7) transforms in this case to

n
γ j − ∑ γk (ek , (e j , K(·, y))L2 )L2 (Ω ) = f j .
k=1

If, for example, e j are the normalized eigenfunctions of the operator A with corre-
sponding eigenvalues λ j , then the latter system can be rewritten as

γ j − λ jγ j = f j, j = 1, 2, . . . , n.

We assume that λ j = 1, so that γ j can be uniquely determined as

fj
γj = ,
1−λj
36 Approximate Methods 385

and therefore ϕn is equal to

n
fj
ϕ (x) = f (x) + ∑ e j (x).
j=1 1 − λj

A different method goes back to Nyström. Let us consider instead of an integral

operator A with kernel K(x, y) a sequence of numerical integration operators
n
(n) (n) (n)
An ϕ (x) = ∑ αj K(x, x j )ϕ (x j ).
j=1

(n) (n)
We assume that the points x j and the weights α j are chosen so that

Aϕ − An ϕ 2L2
 2
  n 
 1 (n) (n) (n) 
= 
Ω Ω
K(x, y)ϕ (y) −
|Ω | ∑ α j K(x, x j )ϕ (x j )dy

dx → 0
j=1

(n) (n)
as n → ∞. The main problem here is to choose the weights α j and the points x j
with this approximation property. The original Nyström method was constructed for
continuous kernels K(x, y).
In Hilbert spaces it is more natural to consider projection methods.
Definition 36.8. Let A : H → H be an injective bounded linear operator. Let Pn :
H → Hn be projection operators such that dim Hn = n. For given f ∈ H, the pro-
jection method generated by Hn and Pn approximates the equation Aϕ = f by the
projection equation
Pn Aϕn = Pn f , ϕn ∈ H. (36.10)

This projection method is said to be convergent if there is n0 ∈ N such that for each
f ∈ H the approximating equation (36.10) has a unique solution ϕn ∈ Hn for all
n ≥ n0 and
ϕn → ϕ , n → ∞,

where ϕ is the unique solution of the equation Aϕ = f .

Theorem 36.9. A projection method converges if and only if there exist n0 ∈ N and
M > 0 such that for all n ≥ n0 the operators

Pn A : H → H

are invertible and the operators (Pn A)−1 Pn A : H → H are uniformly bounded, i.e.,
 
(Pn A)−1 Pn A ≤ M, n ≥ n0 .
386 Part III: Operator Theory and Integral Equations

In case of convergence we have the error estimate

ϕn − ϕ  ≤ (1 + M) inf ψ − ϕ  .
ψ ∈Hn

Proof. If a projection method converges then, by definition, the Pn A are invertible,

and the uniform boundedness follows from the Banach–Steinhaus theorem.
Conversely, under the assumptions of the theorem,

ϕn − ϕ = ((Pn A)−1 Pn A − I)ϕ .

Since for all ψ ∈ Hn we have trivially (Pn A)−1 Pn Aψ = ψ , it follows that

ϕn − ϕ = ((Pn A)−1 Pn A − I)(ϕ − ψ ),

and the error estimate follows. 

Remark 36.10. Projection methods make sense, and we can expect convergence
only if the subspaces Hn possess the denseness property

inf ψ − ϕ  → 0, n → ∞.
ψ ∈Hn

Theorem 36.11. Assume that A : H → H is compact, I − A is injective, and the

projection operators Pn : H → Hn converge pointwise, i.e., Pn ϕ → ϕ , n → ∞ for
each ϕ ∈ H. Then the projection method for I − A converges.

Proof. By Riesz’s theorem (see Theorem 28.15), the operator I − A has a bounded
inverse. Since Pn ϕ → ϕ as n → ∞, we have Pn Aϕ → Aϕ as n → ∞, too. At the same
time, the sequence Pn A is collectively compact, since A is compact and Pn is of finite
rank. Thus, due to Exercise 36.3 we have

(Pn A − A)Pn A → 0, n → ∞. (36.11)

Then the operators (I − Pn A)−1 exist and are uniformly bounded. Indeed, writing

Bn := I + (I − A)−1 Pn A,

we obtain

Bn (I − Pn A) = (I − Pn A) + (I − A)−1 Pn A(I − Pn A)
= I − (I − A)−1 (Pn A − A)Pn A =: I − Sn .

But it is easy to see from (36.11) that

Sn  → 0, n → ∞.
36 Approximate Methods 387

Hence, both I − Pn A and Bn are injective. Since Pn A is compact, we have that (I −

Pn A)−1 is bounded. As a consequence of this fact we have that

(I − Pn A)−1 = (I − Sn )−1 Bn .

The definition of Bn implies

 
Bn  ≤ 1 + (I − A)−1  A .
 
Therefore (I − Pn A)−1  is uniformly bounded in n. The exact equation ϕ − Aϕ = f
and (36.10) with operator I − A lead to

(I − Pn A)(ϕn − ϕ ) = Pn Aϕ − Aϕ + Pn f − f ,

which implies also the error estimate

  
ϕn − ϕ  ≤ (I − Pn A)−1  Pn Aϕ − Aϕ  + Pn f − f  .

The proof is concluded. 

Corollary 36.12. Under the assumptions of Theorem 36.11 and provided addition-
ally that
Pn A − A → 0, n → ∞,

the approximate equation (36.10) with I − A is uniquely solvable for each f ∈ H,

and we have the error estimate

ϕn − ϕ  ≤ M Pn ϕ − ϕ  ,
 
where M is an upper bound for the norm (I − Pn A)−1 .

Proof. The existence of the inverse operators (I − Pn A)−1 and their uniform bound-
edness follow from

I − Pn A = (I − A) − (Pn A − A) = (I − A) I − (I − A)−1 (Pn A − A) ,

−1
(I − Pn A)−1 = I − (I − A)−1 (Pn A − A) (I − A)−1 ,

and  
  (I − A)−1 
(I − Pn A)−1  ≤ .
1 − (I − A)−1  (Pn A − A)

From
(ϕ − ϕn ) − Pn A(ϕ − ϕn ) = ϕ − Pn ϕ
388 Part III: Operator Theory and Integral Equations

and  
ϕn − ϕ  ≤ (I − Pn A)−1  Pn ϕ − ϕ 

we obtain the error estimate. 

Let us return to the projection equation (36.10). It can be rewritten equivalently as

(Aϕn − f , g) = 0, g ∈ Hn . (36.12)

Indeed, if g ∈ Hn , then g = Pn g, Pn∗ = Pn , and hence

0 = (Aϕn − f , g) = (Aϕn − f , Pn g) = (Pn (Aϕn − f ), g),

or
Pn (Aϕn − f ) = 0,

since Hn is considered here to be a Hilbert space. This is the basis for the following
Galerkin projection method.
Assume that {e j }∞j=1 is an orthonormal basis in a Hilbert space H. Considering

Hn = span(e1 , . . . , en ),

we have for the solution ϕn of the projection equation (36.10) the representation
n
ϕn (x) = ∑ γ je j. (36.13)
j=1

The task here is to find (if possible uniquely) the coefficients γ j such that ϕn from
(36.13) solves (36.10). Since (36.12) is equivalent to (36.10), we have from (36.13)
that
(Aϕn , g) = ( f , g), g ∈ Hn ,

or
(Aϕn , ek ) = ( f , ek ) = fk , k = 1, 2, . . . , n,

or
n
∑ γ j (Ae j , ek ) = fk ,
j=1

or
Mγ = f , (36.14)

where γ = (γ1 , . . . , γn ), f = ( f1 , . . . , fn ), and M = {a jk }n×n with a jk = (Ae j , ek ). If

the operator A is invertible, then the matrix M is invertible too, and γ can be obtained
uniquely as
36 Approximate Methods 389

γ = M −1 f .

As a result of this consideration we obtain ϕn (x) uniquely from (36.13). It remains

only to check that this ϕn converges to the solution of the exact equation Aϕ = f . In
order to verify this fact it is enough to apply Theorem 36.9.
We apply now this projection method to the equation (36.8) with compact oper-
ator A.
Theorem 36.13. Let A : H → H be compact and let I − A be injective. Then the
Galerkin projection method converges.

Proof. By Riesz’s theorem, the operator I − A has bounded inverse. Therefore, ϕn

from (36.13) is uniquely defined with γ j that satisfies equation (36.14) with matrix
M = {a jk }n×n , a jk = ((I − A)e j , ek ). Since

∞
Pn ϕ − ϕ 2H = ∑ |(ϕ , e j )|2 → 0, n → ∞,
j=n+1

we may apply Theorem 36.11 and conclude the proof of this theorem. 
 Part IV
Partial Differential Equations
Chapter 37
Introduction

We consider the Euclidean space Rn , n ≥ 1, with elements x = (x1 , . . . , xn ). The

Euclidean length of x is defined by

|x| = x12 + · · · + xn2

and the standard inner product by

(x, y) = x1 y1 + · · · + xn yn .

We use the Cauchy–Bunyakovsky–Schwarz inequality in Rn :

|(x, y)| ≤ |x| · |y|.

Equality here occurs if and only if x = λ y for some λ ∈ R. By BR (x) and SR (x) we
denote the ball and sphere of radius R > 0 with center x:

BR (x) := {y ∈ Rn : |x − y| < R}, SR (x) := {y ∈ Rn : |x − y| = R}.

We say that Ω ⊂ Rn , n ≥ 2, is an open set if for every x ∈ Ω there is R > 0 such that
BR (x) ⊂ Ω . If n = 1, by an open set we mean an open interval (a, b), a < b.
We say that Ω ⊂ Rn , n ≥ 2, is a closed set if Rn \ Ω is open. This is equivalent
to the fact that Ω  ⊂ Ω , where Ω  denotes the set of limit points of Ω , i.e.,

Ω  = {y ∈ Rn : ∃{x(k) }∞ (k)
k=1 ⊂ Ω , |x − y| → 0, k → ∞}.

The closure Ω of the set Ω is defined as Ω = Ω ∪ Ω  . We say that Ω is bounded

if there is R > 0 such that Ω ⊂ BR (0). A closed and bounded set in Rn is called
compact. The boundary ∂ Ω of the set Ω ⊂ Rn is defined as

c Springer International Publishing AG 2017 393

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 37
394 Part IV: Partial Differential Equations

 
∂ Ω = Ω ∩ Rn \ Ω .

An n-tuple α = (α1 , . . . , αn ) of nonnegative integers will be called a multi-index.

We define
(1) |α | = ∑nj=1 α j .
(2) α + β = (α1 + β1 , . . . , αn + βn ) with |α + β | = |α | + |β |.
(3) α ! = α1 ! · · · αn ! with 0! = 1.
(4) α ≥ β if and only if α j ≥ β j for each j = 1, 2, . . . , n. Moreover, α > β if and
only if α ≥ β and there exists j0 such that α j0 > β j0 .
(5) If α ≥ β , then α − β = (α1 − β1 , . . . , αn − βn ) and |α − β | = |α | − |β |.
(6) For x ∈ Rn we define
xα = x1α1 · · · xnαn

with 00 = 1.
We will use the shorthand notation

∂ ∂ |α |
∂j = , ∂ α = ∂1α1 · · · ∂nαn ≡ .
∂xj ∂ x1α1 · · · ∂ xnαn

This part assumes that the reader is familiar also with the following concepts:
(1) The Lebesgue integral in a bounded domain Ω ⊂ Rn and in Rn .
(2) The Banach spaces (L p , 1 ≤ p ≤ ∞, Ck ) and Hilbert spaces (L2 ). If 1 ≤ p < ∞,
then we set
 1/p
L (Ω ) := { f : Ω → C measurable :  f L p (Ω ) :=
p
| f (x)| dx
p
< ∞},
Ω

while

L∞ (Ω ) := { f : Ω → C measurable :  f L∞ (Ω ) := ess sup | f (x)| < ∞}.

x∈Ω

Moreover,

Ck (Ω ) := { f : Ω → C :  f Ck (Ω ) := max ∑
x∈Ω |α |≤k
|∂ α f (x)| < ∞},

where Ω is the closure of Ω . We say that f ∈ C∞ (Ω ) if f ∈ Ck (Ω1 ) for all

k ∈ N and for all bounded subsets Ω1 ⊂ Ω . The space C∞ (Ω ) is not a normed
space. The inner product in L2 (Ω ) is denoted by

( f , g)L2 (Ω ) = f (x)g(x)dx.
Ω
37 Introduction 395

Also in L2 (Ω ), the duality pairing is given by


f , gL2 (Ω ) = f (x)g(x)dx.
Ω


(3) Hölder’s inequality: Let 1 ≤ p ≤ ∞, u ∈ L p and v ∈ L p with

1 1
+ = 1.
p p

Then uv ∈ L1 and
   1   1
p  p
|u(x)v(x)|dx ≤ |u(x)| p dx |v(x)| p dx ,
Ω Ω Ω

where the Hölder conjugate exponent p of p is obtained via

p
p = ,
p−1

with the understanding that p = ∞ if p = 1 and p = 1 if p = ∞.

(4) Lebesgue’s dominated convergence theorem:
Let A ⊂ Rn be measurable and let { fk }∞ k=1 be a sequence of measurable func-
tions converging to f (x) pointwise in A. If there exists a function g ∈ L1 (A)
such that | fk (x)| ≤ g(x) in A, then f ∈ L1 (A) and
 
lim fk (x)dx = f (x)dx.
k→∞ A A

(5) Fubini’s theorem on the interchange of the order of integration:

      
f (x, y)dxdy = dx f (x, y)dy = dy f (x, y)dx ,
X×Y X Y Y X

if f ∈ L1 (X ×Y ).

Exercise 37.1. Prove the generalized Leibniz formula

β
∂ α ( f g) = ∑ Cα ∂ β f ∂ α −β g,
β ≤α

where the generalized binomial coefficients are defined as

β α! α −β
Cα = = Cα .
β !(α − β )!
396 Part IV: Partial Differential Equations

Hypersurfaces

A set S ⊂ Rn is called a hypersurface of class Ck , k = 1, 2, . . . , ∞, if for every x0 ∈ S

there exist an open set V ⊂ Rn containing x0 and a real-valued function ϕ ∈ Ck (V )
such that
∇ϕ ≡ (∂1 ϕ , . . . ∂n ϕ ) = 0 on S ∩ V,

S ∩V = {x ∈ V : ϕ (x) = 0} .

By the implicit function theorem we can solve the equation ϕ (x) = 0 near x0 to
obtain
xn = ψ (x1 , . . . , xn−1 )

for some Ck function ψ . A neighborhood of x0 in S can then be mapped to a piece

of the hyperplane xn = 0 by

x → (x , xn − ψ (x )),

where x = (x1 , . . . , xn−1 ). The vector ∇ϕ is perpendicular to S at x ∈ S ∩ V . The

vector ν (x), which is defined as

∇ϕ
ν (x) := ± ,
|∇ϕ |

is called the normal to S at x. It can be proved that

(∇ψ , −1)
ν (x) = ± .
|∇ψ |2 + 1

If S is the boundary of a domain Ω ⊂ Rn , n ≥ 2, we always choose the orientation

so that ν (x) points out of Ω , and we define the normal derivative of u on S by

∂u ∂u
∂ν u := ν · ∇u ≡ ν1 + · · · + νn .
∂ x1 ∂ xn

Thus ν and ∂ν u are Ck−1 functions.

Example 37.1. Let Sr (y) = {x ∈ Rn : |x − y| = r}. Then

x−y 1 n
∂ ∂
ν (x) =
r
and ∂ν =
r ∑ (x j − y j ) ∂ x j = ∂ r .
j=1
37 Introduction 397

The divergence theorem

Let Ω ⊂ Rn be a bounded domain with C1 boundary S = ∂ Ω and let F be a C1

vector field on Ω . Then
 
∇ · Fdx = F · ν dσ (x).
Ω S

Corollary 37.2 (Integration by parts). Let f and g be C1 functions on Ω . Then

  
∂ j f · gdx = − f · ∂ j gdx + f · gν j dσ (x).
Ω Ω S

Let f and g be locally integrable functions on Rn , i.e., integrable on every bounded

set from Rn . The convolution f ∗ g of f and g is defined by

( f ∗ g)(x) = f (x − y)g(y)dy = (g ∗ f )(x),
Rn

provided that the integral in question exists. The basic theorem on the existence of
convolutions is the following (Young’s inequality for convolution):
Proposition 37.3 (Young’s inequality). Let f ∈ L1 (Rn ) and g ∈ L p (Rn ),
1 ≤ p ≤ ∞. Then f ∗ g ∈ L p (Rn ) and

 f ∗ gL p ≤  f L1 gL p .

Proof. Let p = ∞. Then

 
|( f ∗ g)(x)| ≤ | f (x − y)||g(y)|dy ≤ gL∞ | f (x − y)|dy = gL∞  f L1 .
Rn Rn

Now let 1 ≤ p < ∞. Then it follows from Hölder’s inequality and Fubini’s theorem
that
   p
|( f ∗ g)(x)| dx ≤
p
| f (x − y)||g(y)|dy dx
Rn Rn Rn
   p/p 
≤ | f (x − y)|dy | f (x − y)||g(y)| p dydx
Rn Rn Rn
 
p/p
≤  f L1 | f (x − y)||g(y)| p dydx
Rn Rn
 
p/p
≤  f L1 |g(y)| p dy | f (x − y)|dx
Rn Rn
p/p p/p +1
=  f L1 gLp p  f L1 =  f L1 gLp p .
398 Part IV: Partial Differential Equations

Thus, we have finally

1/p +1/p
 f ∗ gL p ≤  f L1 gL p =  f L1 gL p ,

and the proof is complete. 

Exercise 37.2. Suppose 1 ≤ p, q, r ≤ ∞ and + = + 1. Prove that if f
1
p
1
q
1
r ∈ L p (Rn )
and g ∈ Lq (Rn ), then f ∗ g ∈ Lr (Rn ) and

 f ∗ gr ≤  f  p gq .

In particular,
 f ∗ gL∞ ≤  f L p gL p .

Definition 37.4. Let u ∈ L1 (Rn ) with


u(x)dx = 1.
Rn

Then uε (x) := ε −n u(x/ε ), ε > 0, is called an approximation to the identity.

Proposition 37.5. Let uε (x) be an approximation to the identity. Then for every
function ϕ ∈ L∞ (Rn ) that is continuous at {0} we have

lim uε (x)ϕ (x)dx = ϕ (0).
ε →0+ Rn

Proof. Since uε (x) is an approximation to the identity, we have

 
uε (x)ϕ (x)dx − ϕ (0) = uε (x)(ϕ (x) − ϕ (0))dx,
Rn Rn

and thus




Rn uε (x)ϕ (x)dx − ϕ (0)
≤ |x|≤√ε |uε (x)||ϕ (x) − ϕ (0)|dx

+ √ |uε (x)||ϕ (x) − ϕ (0)|dx
|x|> ε
 
≤ sup
√
|ϕ (x) − ϕ (0)| |uε (x)|dx + 2 ϕ L∞ √ |uε (x)|dx
|x|≤ ε Rn |x|> ε

≤ sup
√
|ϕ (x) − ϕ (0)| · uL1 + 2 ϕ L∞ √ |u(y)|dy → 0
|x|≤ ε |y|>1/ ε

as ε → 0.

37 Introduction 399

Example 37.6. Let u(x) be defined as


sin x1
· · · sin2xn , x ∈ [0, π ]n ,
u(x) = 2
0, x∈/ [0, π ]n .

Then uε (x) is an approximation to the identity and

 επ  επ n
xj
lim (2ε )−n
ε →0 0
···
0
∏ sin ε ϕ (x)dx = ϕ (0).
j=1

Fourier transform

If f ∈ L1 (Rn ), its Fourier transform f or F ( f ) is the bounded function on Rn

defined by 
f(ξ ) = (2π )−n/2 e−ix·ξ f (x)dx.
Rn


Clearly f(ξ ) is well defined for all ξ and f ≤ (2π )−n/2  f 1 .
∞

The Riemann–Lebesgue lemma

If f ∈ L1 (Rn ), then f is continuous and tends to zero at infinity.

Proof. Let us first prove that F f (ξ ) is continuous (even uniformly continuous) in
Rn . Indeed,

|F f (ξ + h) − F f (ξ )| ≤ (2π )−n/2 | f (x)| · |e−i(x,h) − 1|dx
Rn
 
≤ √ | f (x)||x||h|dx + 2 √ | f (x)|dx
|x||h|≤ |h| |x||h|> |h|

≤ |h|  f L1 + 2 √ | f (x)|dx → 0
|x|>1/ |h|

as |h| → 0, since f ∈ L1 (Rn ).

To prove that F f (ξ ) → 0 as |ξ | → 0 we proceed as follows. Since eiπ = −1, we
have
 
−i(x,ξ )
f (x)e−i(x−πξ /|ξ |
−n/2 −n/2 2 ,ξ )
2F f (ξ ) = (2π ) f (x)e dx − (2π ) dx
Rn Rn
 
= (2π )−n/2 f (x)e−i(x,ξ ) dx − (2π )−n/2 f (y + πξ /|ξ |2 )e−i(y,ξ ) dy
Rn Rn

= −(2π )−n/2 ( f (x + πξ /|ξ |2 ) − f (x))e−i(x,ξ ) dx.
Rn
400 Part IV: Partial Differential Equations

Hence

−n/2
2|F f (ξ )| ≤ (2π ) | f (x + πξ /|ξ |2 ) − f (x)|dx
Rn

= (2π )−n/2 f (· + πξ /|ξ |2 ) − f (·) L1 → 0

as |ξ | → ∞, since f ∈ L1 (Rn ). 

∗ g = (2π )n/2 fg.
Exercise 37.3. Prove that if f , g ∈ L1 (Rn ), then f

Exercise 37.4. Suppose f ∈ L1 (Rn ). Prove the following:

(1) If fh (x) = f (x + h), then fh = eih·ξ f.
 
(2) If T : Rn → Rn is linear and invertible, then  f ◦ T = |det T |−1 f (T −1 ) ξ ,
where T  is the adjoint matrix.
f ◦ T = f◦ T .
(3) If T is a rotation, that is, T  = T −1 (and |det T | = 1), then 

Exercise 37.5. Prove that


∂ α f = (−ix)α f, ∂
α f = (iξ )α f.

Exercise 37.6. Prove that if f , g ∈ L1 (Rn ), then

 
f (ξ )
g(ξ )dξ = f(ξ )g(ξ )dξ .
Rn Rn

For f ∈ L1 (Rn ) we define the inverse Fourier transform of f by


F −1 f (x) = (2π )−n/2 eix·ξ f (ξ )dξ .
Rn

It is clear that
F −1 f (x) = F f (−x), F −1 f = F ( f ),

and for f , g ∈ L1 (Rn ),

(F f , g)L2 = ( f , F −1 g)L2 .

The Schwartz space S(Rn ) is defined as

 
∞ α β
S(R ) =
n
f ∈ C (R ) : sup |x ∂ f (x)| < ∞, for all multi-indices α and β .
n
x∈Rn

The Fourier inversion formula

If f ∈ S(Rn ), then (F −1 F ) f = f .
Exercise 37.7. Prove the Fourier inversion formula for f ∈ S(Rn ).
37 Introduction 401

Plancherel’s theorem

The Fourier transform on S extends uniquely to a unitary isomorphism of L2 (Rn )

onto itself, i.e.,

f =  f 2 .
2

This formula is called Parseval’s equality.

The support of a function f : Rn → C, denoted by supp f , is the set

supp f = {x ∈ Rn : f (x) = 0}.

Exercise 37.8. Prove that if f ∈ L1 (Rn ) has compact support, then f extends to an
entire holomorphic function on Cn .

Exercise 37.9. Prove that if f ∈ C0∞ (Rn ), i.e., f ∈ C∞ (Rn ) with compact support, is
supported in {x ∈ Rn : |x| ≤ R}, then for every multi-index α we have

|(iξ )α fˆ(ξ )| ≤ (2π )−n/2 eR|Im ξ | ∂ α f 1 ,

that is, f(ξ ) is decays rapidly |Re ξ | → ∞ when |Im ξ | remains bounded.

Distributions
We say that ϕ j → ϕ in C0∞ (Ω ), Ω ⊂ Rn open, if ϕ j are all supported in a common
compact set K ⊂ Ω and

sup |∂ α ϕ j (x) − ∂ α ϕ (x)| → 0, j→∞

x∈K

for all α . A distribution on Ω is a linear functional u on C0∞ (Ω ) that is continuous,

i.e.,
(1) u : C0∞ (Ω ) → C. The action of u on ϕ ∈ C0∞ (Ω ) is denoted by u, ϕ . The set
of all distributions is denoted by D  (Ω ).
(2) u, c1 ϕ1 + c2 ϕ2  = c1 u, ϕ1  + c2 u, ϕ2 .
(3) If ϕ j → ϕ in C0∞ (Ω ), then u, ϕ j  → u, ϕ  in C as j → ∞. This is equivalent to
the following condition: for all K ⊂ Ω there exist a constant CK and an integer
NK such that for all ϕ ∈ C0∞ (K),

| u, ϕ | ≤ CK ∑ ∂ α ϕ ∞ .
|α |≤NK

Remark 37.7. If u ∈ Lloc1 (Ω ), Ω ⊂ Rn open, then u can be regarded as a distribution

(in that case, a regular distribution) as follows:

402 Part IV: Partial Differential Equations

u, ϕ  := u(x)ϕ (x)dx, ϕ ∈ C0∞ (Ω ).
Ω

The Dirac δ -function

The Dirac δ -function is defined as

δ , ϕ  = ϕ (0), ϕ ∈ C0∞ (Ω ).

It is not a regular distribution.

Example 37.8. Let uε (x) be an approximation to the identity. Then
 
x
uε (ξ ) = (2π )−n/2 ε −n
u( )e−i(x,ξ ) dx = (2π )−n/2 u(y)e−i(y,eξ ) dy = u(εξ ).
Rn ε Rn

In particular,
lim uε (ξ ) = lim u(εξ ) = (2π )−n/2 .
ε →0+ ε →0+

Applying Proposition 37.5, we may conclude that

(1) limε →0+ uε , ϕ  = ϕ (0) i.e. limε →0+ uε = δ in the sense of distributions, and
(2) δ = (2π )−n/2 · 1.
We can extend the operations from functions to distributions as follows:

∂ α u, ϕ  = u, (−1)|α | ∂ α ϕ ,

f u, ϕ  = u, f ϕ , f ∈ C∞ (Ω ),

 ,
u ∗ ψ , ϕ  = u, ϕ ∗ ψ ψ ∈ C0∞ (Ω ),

 (x) = ψ (−x). It is possible to show that u ∗ ψ is actually a C∞ function and

where ψ

∂ α (u ∗ ψ ) = u ∗ ∂ α ψ .

A tempered distribution is a continuous linear functional on S(Rn ). In addition to the

preceding operations for tempered distributions we can define the Fourier transform
by
u, ϕ  = u, ϕ, ϕ ∈ S.


Exercise 37.10. Prove that if u is a tempered distribution and ψ ∈ S, then

u  u.
∗ ψ = (2π )n/2 ψ
37 Introduction 403

Exercise 37.11. Prove that

(1) δ = (2π )−n/2 · 1,  1 = (2π )n/2 δ .
α α
(2) ∂ δ = (iξ ) (2π ) −n/2 .
(3) xα = i|α | ∂ α (
1) = i|α | (2π )n/2 ∂ α δ .
Chapter 38
Local Existence Theory

A partial differential equation of order k ∈ N is an equation of the form

 
F x, (∂ α u)|α |≤k = 0, (38.1)

where F is a function of the variables x ∈ Ω ⊂ Rn , n ≥ 2, and (∂ α u)|α |≤k .

A complex-valued function u(x) on Ω is a classical solution of (38.1) if the
derivatives ∂ α u occurring in F exist on Ω and
 
F x, (∂ α u(x))|α |≤k = 0

pointwise for all x ∈ Ω . The equation (38.1) is said to be linear if it can be written
as
∑ aα (x)∂ α u(x) = f (x) (38.2)
|α |≤k

for some known functions aα and f . In this case we speak of the (linear) differential
operator
L(x, ∂ ) ≡ ∑ aα (x)∂ α
|α |≤k

and write (38.2) simply as Lu = f . If the coefficients aα (x) belong to C∞ (Ω ), we

can apply the operator L to any distribution u ∈ D  (Ω ), and u is called a distribu-
tional solution (or weak solution) of (38.2) if equation (38.2) holds in the sense of
distributions, i.e.,
∑ (−1)|α | u, ∂ α (aα ϕ ) =  f , ϕ ,
|α |≤k

∂u
where ϕ ∈ C0∞ (Ω ). Let us list some examples. Here and throughout we set ut = ∂t ,
∂ 2u
utt = ∂ t2
, and so forth.

c Springer International Publishing AG 2017 405

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 38
406 Part IV: Partial Differential Equations

(1) The eikonal equation

|∇u|2 = c2 ,

where ∇u = (∂1 u, . . . , ∂n u) is the gradient of u.

(2) (a) The heat (or evolution) equation

ut = kΔ u.

(b) The wave equation

utt = c2 Δ u.

(c) The Poisson equation

Δu = f,

where Δ ≡ ∇·∇ = ∂12 +· · ·+ ∂n2 is the Laplacian (or the Laplace operator).
(3) The telegrapher’s equation

utt = c2 Δ u − α ut − m2 u.

(4) The Sine–Gordon equation

utt = c2 Δ u − sin u.

(5) The biharmonic equation

Δ 2 u ≡ Δ (Δ u) = 0.

(6) The Korteweg–de Vries equation

ut + cu · ux + uxxx = 0.

In the linear case, a simple measure of the “strength” of a differential operator is

provided by the notion of characteristics. If L(x, ∂ ) = ∑|α |≤k aα (x)∂ α , then its char-
acteristic form (or principal symbol) at x ∈ Ω is the homogeneous polynomial of
degree k defined by

χL (x, ξ ) = ∑ aα (x)ξ α , ξ ∈ Rn .
|α |=k

A nonzero ξ is said to be characteristic for L at x if χL (x, ξ ) = 0, and the set of all

such ξ is called the characteristic variety of L at x, denoted by charx (L). In other
words,
charx (L) = {ξ
= 0 : χL (x, ξ ) = 0} .

In particular, L is said to be elliptic at x if charx (L) = 0/ and elliptic in Ω if it is

elliptic at every x ∈ Ω .
38 Local Existence Theory 407

Example 38.1. (1) L = ∂1 ∂2 with

 
charx (L) = ξ ∈ R2 : ξ1 = 0 or ξ2 = 0, ξ12 + ξ22 > 0 .

(2) L = 12 (∂1 + i∂2 ) is the Cauchy–Riemann operator on R2 . It is elliptic in R2 .

(3) L = Δ is elliptic in Rn .  
(4) L = ∂1 − ∑nj=2 ∂ j2 , charx (L) = ξ ∈ Rn \ {0} : ξ j = 0, j = 2, 3, . . . , n .
 
(5) L = ∂12 − ∑nj=2 ∂ j2 , charx (L) = ξ ∈ Rn \ {0} : ξ12 = ∑nj=2 ξ j2 .

Let ν (x) be the normal to S at x. A hypersurface S is said to be characteristic for L

at x ∈ S if ν (x) ∈ charx (L), i.e.,

χL (x, ν (x)) = 0,

and S is said to be non-characteristic if it is not characteristic at every point, that is,

if for all x ∈ S,
χL (x, ν (x))
= 0.

It is clear that every S is noncharacteristic for elliptic operators. The lines

S = {x ∈ Rn : x1
= 0, x2 = · · · = xn = 0}

are characteristic for the heat operator, and the cones


S± = {x ∈ Rn : x1 = ± x22 + · · · xn2 }

are characteristic for the wave operator.

Let us consider the first-order linear equation
n
Lu ≡ ∑ a j (x)∂ j u + b(x)u = f (x), (38.3)
j=1

where a j , b, and f are assumed to be C1 functions of x. We assume also that a j ,

b, and f are real-valued. Suppose we wish to find a solution u of (38.3) with given
initial values u = g on the hypersurface S (g is also real-valued). It is clear that
 
charx (L) = ξ
= 0 : A · ξ = 0 ,

where A = (a1 , . . . , an ). This implies that charx (L) ∪ {0} is the hyperplane orthog-
onal to A, and therefore, S is characteristic at x if and only if A is tangent to S at x
(A · ν = 0). Then
408 Part IV: Partial Differential Equations

n n
∑ a j (x)∂ j u(x) = ∑ a j (x)∂ j g(x), x ∈ S,
j=1 j=1

is completely determined as a set of certain directional derivatives of ϕ (see the

definition of S) along S at x, and it may be impossible to make this sum equal to
f (x) − b(x)u(x) (in order to satisfy (38.3)). Indeed, let us assume that u1 and u2
have the same value g on S. This means that u1 − u2 = 0 on S, or (more or less
equivalently)
u1 − u2 = ϕ · γ ,

where ϕ = 0 on S (ϕ defines this surface) and γ

= 0 on S. Next,

(A · ∇)u1 − (A · ∇)u2 = (A · ∇)(ϕγ ) = γ (A · ∇)ϕ + ϕ (A · ∇)γ = 0,

since S is characteristic for L ((A · ∇)ϕ = 0 ⇔ (A · |∇|

∇
)ϕ = 0 ⇔ A · ν = 0). Therefore,
to make the initial value problem well defined we must assume that S is noncharac-
teristic for this problem.
Let us assume that S is noncharacteristic for L and u = g on S. We define the
integral curves for (38.3) as the parametrized curves x(t) that satisfy the system

ẋ = A(x), x = x(t) = (x1 (t), . . . , xn (t)), (38.4)

of ordinary differential equations, where

ẋ = (x1 (t), . . . , xn (t)).

Along one of those curves a solution u of (38.3) must satisfy

du d n
∂u
dt
= (u(x(t))) =
dt ∑ ẋ j ∂ x j = (A · ∇)u = f − bu ≡ f (x(t)) − bu(x(t)),
j=1

or
du
= f − bu. (38.5)
dt
By the existence and uniqueness theorem for ordinary differential equations there
is a unique solution (unique curve) of (38.4) with x(0) = x0 . Along this curve the
solution u(x) of (38.3) must be the solution of (38.5) with u(0) = u(x(0)) = u(x0 ) =
g(x0 ). Moreover, since S is noncharacteristic, x(t) ∈
/ S for t
= 0, at least for small t,
and the curves x(t) fill out a neighborhood of S. Thus we have proved the following
theorem.
Theorem 38.2. Assume that S is a surface of class C1 that is noncharacteristic
for (38.3), and that a j , b, f , and g are real-valued C1 functions. Then for every
sufficiently small neighborhood U of S in Rn there is a unique solution u ∈ C1 of
(38.3) on U that satisfies u = g on S.
38 Local Existence Theory 409

Remark 38.3. The method that was presented above is called the method of char-
acteristics.

The following two examples demonstrate the necessity of noncharacteristic sur-

faces for boundary value problems.
Example 38.4. (1) In R2 , solve x2 ∂1 u + x1 ∂2 u = u with u(x1 , 0) = g(x1 ) on the line
x2 = 0.
Since ν (x) = (0, 1) and since χL (x, ξ ) = x2 ξ1 + x1 ξ2 , we have

χL (x, ν (x)) = x2 · 0 + x1 · 1 = x1
= 0,

so that the lines x1 > 0 and x1 < 0 are noncharacteristic. The system (38.4)–
(38.5) to be solved is

ẋ1 = x2 , ẋ2 = x1 , u̇ = u,

with initial conditions

(x1 , x2 )|t=0 = (x10 , 0), u(0) = g(x10 ),

on S. We obtain

x10 t x10 t
x1 = (e + e−t ), x2 = (e − e−t ), u = g(x10 )et .
2 2
These equations imply

x1 + x2 = x10 et , x1 − x2 = x10 e−t , x12 − x22 = (x10 )2 .

So
x1 + x2
et = ±  , x12 > x22 ,
x12 − x22

and thus 

x1 + x2
u(x1 , x2 ) = ±g ± x12 − x22  ,
x12 − x22

where we have a plus sign for x1 + x2 > 0 and a minus sign for x1 + x2 < 0.
(2) In R2 , solve x1 ∂1 u + x2 ∂2 u = u with u(x1 , 0) = g(x1 ) on the line x2 = 0.
Compared to previous example, in this case the line x2 = 0 is characteristic,
since x1 · 0 + x2 · 1 = 0 on S. The system (38.4)–(38.5) gives in this case that

x1 = x10 et , x2 ≡ 0, u = g(x10 )et .

410 Part IV: Partial Differential Equations

This means that u(x1 , x2 ) is a function of only one variable x1 , and the original
equation transforms to
x1 ∂1 u = u1 ,

which has only the solution u1 (x1 ) = cx1 , where c is a constant. But then we
have a contradiction, since the equality cx1 = g(x1 ) is impossible for an arbi-
trary C1 function g.
Let us consider more examples in which we apply the method of characteristics.

Example 38.5. In R3 , solve x1 ∂1 u + 2x2 ∂2 u + ∂3 u = 3u with u = g(x1 , x2 ) in the

plane x3 = 0.  
Since S = x ∈ R3 : x3 = 0 , we have ν (x) = (0, 0, 1), and since χL (x, ξ ) =
x1 ξ1 + 2x2 ξ2 + ξ3 , we must have

χL (x, ν (x)) = x1 · 0 + 2x2 · 0 + 1 · 1 = 1

= 0,

so that S is noncharacteristic. The system (38.4)–(38.5) to be solved is

ẋ1 = x1 , ẋ2 = 2x2 , ẋ3 = 1, u̇ = 3u,

with initial conditions

(x1 , x2 , x3 )|t=0 = (x10 , x20 , 0), u(0) = g(x10 , x20 ),

on S. We obtain

x1 = x10 et , x2 = x20 e2t , x3 = t, u = g(x10 , x20 )e3t .

These equations imply

x10 = x1 e−t = x1 e−x3 , x20 = x2 e−2t = x2 e−2x3 .

Therefore,
u(x) = u(x1 , x2 , x3 ) = g(x1 e−x3 , x2 e−2x3 )e3x3 .

Example 38.6. In R3 , solve ∂1u + x1 ∂2 u − ∂3 u = u with u(x1 , x2 , 1) = x1 + x2 .

Since S = x ∈ R3 : x3 = 1 , we have ν (x) = (0, 0, 1), and therefore,

χL (x, ν (x)) = 1 · 0 + x1 · 0 − 1 · 1 = −1
= 0,

and S is noncharacteristic. The system (38.4)–(38.5) for this problem becomes

ẋ1 = 1, ẋ2 = x1 , ẋ3 = −1, u̇ = u,

38 Local Existence Theory 411

with
(x1 , x2 , x3 )|t=0 = (x10 , x20 , 1), u(0) = x10 + x20 .

We obtain

t2
x1 = t + x10 , x2 = + tx10 + x20 , x3 = −t + 1, u = (x10 + x20 )et .
2
Then,
t = 1 − x3 , x10 = x1 − t = x1 + x3 − 1,

(1 − x3 )2 1 x2
x20 = x2 − − (1 − x3 )(x1 + x3 − 1) = − x1 + x2 − x3 + x1 x3 + 3 ,
2 2 2
and finally,

x32 1 1−x3
u= + x1 x3 + x2 − e .
2 2

Now let us generalize this technique to quasilinear equations, or to the equations of

the form
n
∑ a j (x, u)∂ j u = b(x, u), (38.6)
j=1

where a j , b, and u are real-valued. If u is a function of x, the normal to the graph of

u in Rn+1 is proportional to (∇u, −1), so (38.6) just says that the vector field

A(x, y) := (a1 , . . . , an , b) ∈ Rn+1

is tangent to the graph y = u(x) at every point. This suggests that we look at the
integral curves of A in Rn+1 given by solving the ordinary differential equations

ẋ j = a j (x, y), j = 1, 2, . . . , n, ẏ = b(x, y).

Suppose we are given initial data u = g on S. If we form the submanifold

S∗ := {(x, g(x)) : x ∈ S}

in Rn+1 , then the graph of the solution should be the hypersurface generated by
the integral curves of A passing through S∗ . Again, we need to assume that S is
noncharacteristic in the sense that the vector

(a1 (x, g(x)), . . . , an (x, g(x))) , x ∈ S,

412 Part IV: Partial Differential Equations

should not be tangent to S at x, or

n
∑ a j (x, g(x))ν j (x)
= 0.
j=1

Suppose u is a solution of (38.6). If we solve

ẋ j = a j (x, u(x)), j = 1, 2, . . . , n,

with x j (0) = x0j , then writing the solution u via integral curves as y(t) = u(x(t)), we
obtain that
n n
ẏ = ∑ ∂ j u · ẋ j = ∑ a j (x, u)∂ j u = b(x, u) = b(x, y).
j=1 j=1

Thus, as in the linear case, u solves (38.6) with given initial data g on S.
Example 38.7. In R2 , solve u∂1 u+ ∂2 u = 1 with u = s/2 on the segment x1 = x2 = s,
where s > 0, s
= 2, is a parameter.
Since ϕ (s) = (s, s), it follows that (x = x1 = s)
 

∂ x1
∂s a1 (s, s, s/2) 1 s/2
det ∂ x2 = det = 1 − s/2
= 0,
∂s a2 (s, s, s/2) 1 1

for s > 0, s
= 2. The system (38.4)–(38.5) for this problem is

ẋ1 = u, ẋ2 = 1, u̇ = 1,

with
x10
(x1 , x2 , u)|t=0 = (x10 , x20 , ) = (s, s, s/2).
2
Then
u = t + s/2, x2 = t + s, ẋ1 = t + s/2,

t2
so that x1 = 2 + st2 + s. This implies

x1 − x2 = t 2 /2 + t(s/2 − 1).

For s and t in terms of x1 and x2 we obtain

s 1 t2 2(x1 − x2 )
= 1+ x1 − x2 − , t= .
2 t 2 x2 − 2
38 Local Existence Theory 413

Hence

2(x1 − x2 ) x1 − x2 t
u= +1+ −
x2 − 2 t 2
2(x1 − x2 ) x2 − 2 x1 − x2
= +1+ −
x2 − 2 2 x2 − 2
x1 − x2 x2 − 2 x1 − x2 x2 2x1 − 4x2 + x22
= +1+ = + = .
x2 − 2 2 x2 − 2 2 2(x2 − 2)

Exercise 38.1. In R2 , solve x12 ∂1 u + x22 ∂2 u = u2 with u ≡ 1 when x2 = 2x1 .

Exercise 38.2. In R2 , solve u∂1 u + x2 ∂2 u = x1 with u(x1 , 1) = 2x1 .

Example 38.8. Consider the Burgers equation

u∂1 u + ∂2 u = 0

 R with u(x1 ,0) = h(x1 ), where h is a known C1 function. It is clear that S :=

in 2

x ∈ R : x2 = 0 is noncharacteristic for this quasilinear equation, since

2

1 h(x1 )
det = 1
= 0,
0 1

and ν (x) = (0, 1). Now we have to solve the ordinary differential equations

ẋ1 = u, ẋ2 = 1, u̇ = 0,

with  
(x1 , x2 , u)|t=0 = x10 , 0, h(x10 ) .

We obtain
x2 = t, u ≡ h(x10 ), x1 = h(x10 )t + x10 ,

so that
x1 − x2 h(x10 ) − x10 = 0.

Let us assume that

−x2 h1 (x10 ) − 1
= 0.

By this condition, the last equation defines an implicit function x10 = g(x1 , x2 ).
Therefore, the solution u of the Burgers equation has the form

u(x1 , x2 ) = h(g(x1 , x2 )).

414 Part IV: Partial Differential Equations

Let us consider two particular cases:

(1) If h(x10 ) = ax10 + b, a
= 0, then

ax1 + b 1
u(x1 , x2 ) = , x2
= − .
ax2 + 1 a

(2) If h(x10 ) = a(x10 )2 + bx10 + c, a

= 0, then
 2
−x2 b − 1 +(x2 b + 1)2 − 4ax2 (cx2 − x1 )
u(x1 , x2 ) = a
2ax2
 
−x2 b − 1 + (x2 b + 1)2 − 4ax2 (cx2 − x1 )
+b + c,
2ax2

with D = (x2 b + 1)2 − 4ax2 (cx2 − x1 ) > 0.

Let us consider again the linear equation (38.2) of order k, i.e.,

∑ aα (x)∂ α u(x) = f (x).

|α |≤k

Let S be a hypersurface of class Ck . If u is a Ck function defined near S, the quantities

u, ∂ν u, . . . , ∂νk−1 u (38.7)

on S are called the Cauchy data of u on S. And the Cauchy problem is to solve
(38.2) with the Cauchy data (38.7). We shall consider Rn , n ≥ 2, to be Rn−1 × R
and denote the coordinates by (x,t), where x = (x1 , . . . , xn−1 ). We can make a change
of coordinates from Rn to Rn−1 × R so that x0 ∈ S is mapped to (0, 0) and a neigh-
borhood of x0 in S is mapped into the hyperplane t = 0. In that case ∂ν = ∂∂t on
S = {(x,t) : t = 0}, and equation (38.2) can be written in the new coordinates as

∑ aα , j (x,t)∂xα ∂t j u = f (x,t) (38.8)

|α |+ j≤k

with the Cauchy data

∂t j u(x, 0) = ϕ j (x), j = 0, 1, . . . , k − 1. (38.9)

Since ν = (0, 0, . . . , 0, 1), the assumption that S is noncharacteristic means that

χL (x, 0, ν (x, 0)) ≡ a0,k (x, 0)

= 0.
38 Local Existence Theory 415

Hence by continuity, a0,k (x,t)

= 0 for small t, and we can solve (38.8) for ∂tk u:
 
 −1
∂tk u(x,t) = a0,k (x,t) f− ∑ aα , j ∂xα ∂t u
j
(38.10)
|α |+ j≤k, j<k

with the Cauchy data (38.9).

Example 38.9. The line t = 0 is noncharacteristic for ∂t2 u = ∂x2 u in R2 . The Cauchy
problem u(x, 0) = g0 (x), ∂t u(x, 0) = g1 (x), has a unique solution in appropriate
classes for g0 and g1 . This can be proved by the method of separation of variables
(see Section 14.2).

Example 38.10. The line t = 0 is characteristic for ∂x ∂t u = 0 in R2 , and we will

therefore have some problems with the solutions. Indeed, if u is a solution of this
equation with Cauchy data u(x, 0) = g0 (x) and ∂t u(x, 0) = g1 (x), then ∂x g1 = 0, that
is, g1 ≡ constant. Thus the Cauchy problem is not solvable in general. On the other
hand, if g1 is constant, then there is no uniqueness, because we can take u(x,t) =
g0 (x) + f (t) with any f (t) such that f (0) = 0 and f  (0) = g1 .

Example 38.11. The line t = 0 is characteristic for ∂x2 u − ∂t u = 0 in R2 . Here if

we are given u(x, 0) = g0 (x), then ∂t u(x, 0) is already completely determined by
∂t u(x, 0) = g0 (x). So, again the Cauchy problem has “bad” behavior.

Let us now formulate and give a sketch of the proof of the famous Cauchy–
Kowalevski theorem for the linear case.
Theorem 38.12. If aα , j (x,t), ϕ0 (x), . . . , ϕk−1 (x) are real-analytic near the origin
in Rn , then there is a neighborhood of the origin on which the Cauchy problem
(38.10)–(38.9) has a unique real-analytic solution.

Proof. The uniqueness of the analytic solution follows from the fact that an analytic
function is completely determined by the values of its derivatives at one point (see
the Taylor formula or the Taylor series). Indeed, for all α and j = 0, 1, . . . , k − 1,

∂xα ∂t j u(x, 0) = ∂xα ϕ j (x).

Therefore,
 
 −1
∂tk u|t=0 = a0,k f (x, 0) − ∑ aα , j (x, 0)∂xα ϕ j (x) ,
|α |+ j≤k, j<k

and moreover,
 
 −1
∂tk u(x,t) = a0,k f (x,t) − ∑ aα , j (x,t)∂xα ∂t j u .
|α |+ j≤k, j<k
416 Part IV: Partial Differential Equations

Then all derivatives of u can be defined from this equation by

 
∂tk+1 u = ∂t ∂tk u .

Next, let us denote by yα , j = ∂xα ∂t j u and by Y = (yα , j ) this vector. Then equation
(38.10) can be rewritten as
 
 −1
y0,k = a0,k f− ∑ aα , j yα , j ,
|α |+ j≤k, j<k

or  
   −1
∂t y0,k−1 = a0,k f− ∑ aα , j ∂x j y(α −j), j ,
|α |+ j≤k, j<k

and therefore, the Cauchy problem (38.10)–(38.9) becomes


∂t Y = ∑n−1
j=1 A j ∂x j Y + B
(38.11)
Y (x, 0) = Φ (x), x ∈ Rn−1 ,

where Y , B, and Φ are analytic vector-valued functions and the A j are analytic
matrix-valued functions. Without loss of generality we can assume that Φ ≡ 0.
( j)
Let Y = (y1 , . . . , yN ), B = (b1 , . . . , bN ), A j = (aml )Nm,l=1 . We seek a solution Y =
(y1 , . . . , yN ) of the form

(m)
ym = ∑ Cα , j xα t j , m = 1, 2, . . . , N.

(m)
The Cauchy data tell us that Cα ,0 = 0 for all α and m, since we assumed Φ ≡ 0. To
(m)
determine Cα , j for j > 0, we substitute ym into (38.11) and get for m = 1, 2, . . . , N,
that
( j)
∂t ym = ∑ aml ∂x j yl + bm (x, y),

or  
(m) ( j) (m)  (m)
∑ Cα , j jxα t j−1 = ∑ ∑ aml
βr
xβ t r ∑ Cα , j α j xα − j t j + ∑ bα j xα t j .
j,l β ,r

(m)
It can be proved that this equation determines uniquely the coefficients Cα , j and
therefore the solution Y = (y1 , . . . , yN ). 

Remark 38.13. Consider the following example in R2 , due to Hadamard, which

sheds light on the Cauchy problem:
√
Δ u = 0, u(x1 , 0) = 0, ∂2 u(x1 , 0) = ke− k
sin(x1 k), k ∈ N.
38 Local Existence Theory 417

This problem is noncharacteristic on R2 , since Δ is elliptic in R2 . We look for

u(x1 , x2 ) = u1 (x1 )u2 (x2 ). Then

u1 u2 + u2 u1 = 0,

which implies that

u1 u
= − 2 = −λ = constant.
u1 u2

Next, the general solutions of

u1 = −λ u1

and
u2 = λ u2

are √ √
u1 = A sin( λ x1 ) + B cos( λ x1 )

and √ √
u2 = C sinh( λ x2 ) + D cosh( λ x2 ),
√
respectively. But u2 (0) = 0, u2 (0) = 1 and u1 (x1 ) = ke− k sin(kx1 ). Thus D = 0,
√ √
B = 0, k = λ , A = ke− k , and C = 1k = √1 . So we finally have
λ

√ 1 √
u(x1 , x2 ) = ke− k
sin(kx1 ) sinh(kx2 ) = e− k sin(kx1 ) sinh(kx2 ).
k
As k → +∞, the Cauchy data √
and their derivatives (for x2 = 0) of all orders tend
uniformly to zero, since e− k decays faster than polynomially. But if x2
= 0 (more
precisely, x2 > 0), then
√
lim e− k
sin(kx1 ) sinh(kx2 ) = ∞,
k→+∞

(k)
if we choose, for example, x2 = 1 and x1 = π /(2k) + 2π . Hence u(x1 , x2 ) is not
bounded. But the solution of the original problem that corresponds to the limiting
case k = ∞ is of course u ≡ 0, since u(x1 , 0) = 0 and ∂2 u(x1 , 0) = 0 in the limiting
case. Hence the solution of the Cauchy problem may not depend continuously on
the Cauchy data. This means by Hadamard that the Cauchy problem for elliptic
operators is “ill-posed,” even when this problem is noncharacteristic.

Remark 38.14. This example of Hadamard’s shows that the solution of the Cauchy
problem may not depend continuously on the Cauchy data. By the terminology of
Hadamard, “the Cauchy problem for the Laplacian is not well posed, but it is ill
418 Part IV: Partial Differential Equations

posed.” Due to Hadamard and Tikhonov, a problem is called well posed if the fol-
lowing conditions are satisfied:
(1) existence;
(2) uniqueness;
(3) stability or continuous dependence on data.
Otherwise, it is called ill posed.

Let us consider one more important example due to H. Lewy. Let L be the first-order
differential operator in R3 ((x, y,t) ∈ R3 ) given by

∂ ∂ ∂
L≡ + i − 2i(x + iy) . (38.12)
∂x ∂y ∂t

Theorem 38.15 (The Hans Lewy example). Let f be a continuous real-valued

function depending only on t. If there is a C1 function u satisfying Lu = f , with the
operator L from (38.12), in some neighborhood of the origin, then f (t) necessarily
is analytic at t = 0.

Remark 38.16. This example shows that the assumption of analyticity of f in

Theorem 38.12 in the linear equation cannot be omitted (it is essential). It appears
necessarily, since Lu = f with L from (38.12) has no C1 solution unless f is analytic.

Proof. Suppose x2 + y2 < R2 , |t| < R, and set z = x + iy = reiθ . Let us denote by
V (t) the function
  2π
V (t) := u(x, y,t)dσ (z) = ir u(r, θ ,t)eiθ dθ ,
|z|=r 0

where u(x, y,t) is the C1 solution of the equation Lu = f with L from (38.12). We
continue to denote u in polar coordinates also by u. By the divergence theorem for
F := (u, iu) we get
 

∂u ∂u
i ∇ · Fdxdy ≡ i +i dxdy = i (u, iu) · ν dσ (z)
|z|<r |z|<r ∂ x ∂y |z|=r
  x  
y
=i u + iu dσ (z) = i ueiθ dσ (z)
|z|=r r r |z|=r
 2π
= ir ueiθ dθ ≡ V (t).
0

But on the other hand, in polar coordinates,


 r  2π

∂u ∂u ∂u ∂u
V (t) ≡ i +i dxdy = i +i (ρ , θ ,t)ρ dρ dθ .
|z|<r ∂x ∂y 0 0 ∂x ∂y
38 Local Existence Theory 419

This implies that

 2π


∂V ∂u ∂u ∂u ∂u dσ (z)
= ir +i (r, θ ,t)dθ = +i (x, y,t)2r
∂r 0 ∂ x ∂ y |z|=r ∂ x ∂ y 2z



∂ u f (t) ∂V dσ (z)
= 2r i + dσ (z) = 2r i + f (t)
|z|=r ∂ t 2z ∂ t |z|=r 2z

∂V
= 2r i + iπ f (t) .
∂t

We therefore have the following equation for V :

1 ∂V ∂V
=i + π f (t) . (38.13)
2r ∂ r ∂t

Let us introduce now a new function U(s,t) = V (s) + π F(t), where s = r2 and
F  = f . The function F exists because f is continuous. It follows from (38.13) that

1 ∂V ∂V ∂U ∂V ∂U ∂U
≡ , = , =i .
2r ∂ r ∂s ∂s ∂s ∂s ∂t
Hence
∂U ∂U
+i = 0. (38.14)
∂t ∂s
Since (38.14) is the Cauchy–Riemann equation, we have that U is a holomorphic
(analytic) function of the variable w = t + is, in the region 0 < s < R2 , |t| < R, and
U is continuous up to s = 0. Next, since U(0,t) = π F(t) (V = 0 when s = 0, i.e.,
r = 0) and f (t) is real-valued, it follows that U(0,t) is also real-valued. Therefore,
by the Schwarz reflection principle (see complex analysis), the formula

U(−s,t) := U(s,t)

gives a holomorphic continuation of U to a full neighborhood of the origin. In par-

ticular, U(0,t) = π F(t) is analytic in t, hence so is f (t) ≡ F  (t). 
Chapter 39
The Laplace Operator

We consider what is perhaps the most important of all partial differential operators,
theLaplace operator (Laplacian) on Rn , defined by
n
Δ= ∑ ∂ j2 ≡ ∇ · ∇.
j=1

We will begin with a quite general fact about partial differential operators.
Definition 39.1. (1) A linear transformation T on Rn is called arotation if T  =
T −1 .
(2) Let h be a fixed vector in Rn . Thetranslation transformation Th f (x) := f (x + h)
is called a.

Theorem 39.2. Suppose that L is a linear partial differential operator on Rn . Then

L commutes with translations and rotations if and only if L is a polynomial in Δ ,
that is, L ≡ ∑mj=0 a j Δ j .

Proof. Let
L(x, ∂ ) ≡ ∑ aα (x)∂ α
|α |≤k

commute with a translation Th . Then

∑ aα (x)∂ α f (x + h) = ∑ aα (x + h)∂ α f (x + h).

|α |≤k |α |≤k

This implies that the aα (x) must be constants (because aα (x) ≡ aα (x + h) for all h),
say aα . Next, since L now has constant coefficients, we have (see Exercise 37.5)

 ξ ) = P(ξ )
Lu( u(ξ ),

c Springer International Publishing AG 2017 421

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 39
422 Part IV: Partial Differential Equations

where the polynomial P(ξ ) is defined by

P(ξ ) = ∑ aα (iξ )α .
|α |≤k

Recall from Exercise 37.4 that if T is a rotation, then

u
◦ T (ξ ) = (
u ◦ T ) (ξ ).

Therefore,
x)(ξ ) = Lu(T
(Lu)(T  ξ ),

or

P(ξ )u(T x)(ξ ) = P(T ξ )
u(T ξ ).

This forces
P(ξ ) = P(T ξ ).

Write ξ = |ξ |θ , where θ ∈ Sn−1 = {x ∈ Rn : |x| = 1} is the direction of ξ . Then

T ξ = |ξ |θ  with some θ  ∈ Sn−1 . But

0 = P(ξ ) − P(T ξ ) = P(|ξ |θ ) − P(|ξ |θ  )

shows that P(ξ ) does not depend on the angle θ of ξ . Therefore, P(ξ ) is radial, that
is,
P(ξ ) = P1 (|ξ |) = ∑ aα |ξ ||α | .
|α |≤k

But since we know that P(ξ ) is a polynomial, |α | must be even:

P(ξ ) = ∑ a j |ξ |2 j .
j

By Exercise 37.5 we have that

Δu(ξ ) = −|ξ |2 u(ξ ).

It follows by induction that


Δ j u(ξ ) = (−1) j |ξ |2 j u
(ξ ), j = 0, 1, . . . .

Taking the inverse Fourier transform, we obtain


u(ξ )) = F −1 ∑ a j |ξ |2 j u(ξ ) = F −1 ∑ aj Δ
Lu = F −1 (P(ξ ) j u(ξ ) =
∑ aj Δ j u.
j j j
39 The Laplace Operator 423

Conversely, let
Lu = ∑ a j Δ j u.
j

It is clear by the chain rule that the Laplacian commutes with translations Th and
rotations T . By induction, the same is true for any power of Δ , and so for L as
well. 

Lemma 39.3. If f (x) = ϕ (r), r = |x|, that is, f is radial, then Δ f = ϕ  (r) +
n−1 
r ϕ (r).

∂r xj
Proof. Since ∂xj = r , it follows that

n n x 
∑ ∂ j (∂ j ϕ (r)) = ∑ ∂ j ϕ  (r)
j
Δf =
j=1 j=1 r
n x  n x2j
∑ ϕ  (r)∂ j +∑ ϕ  (r)
j
= 2
j=1 r j=1 r
 
n 2 n x2j
1 xj
= ∑ − 3
r r
ϕ  (r) + ∑
r2
ϕ  (r)
j=1 j=1
n 1 n
n−1 
= ϕ  (r) − 3
r r ∑ x2j ϕ  (r) + ϕ  (r) = ϕ  (r) + r
ϕ (r).
j=1

This completes the proof. 

Corollary 39.4. If f (x) = ϕ (r), then Δ f = 0 on Rn \ {0} if and only if

a + br2−n , n = 2,
ϕ (r) =
a + b log r, n = 2,

where a and b are arbitrary constants.

Proof. If Δ f = 0, then by Lemma 39.3, we have

n−1 
ϕ  (r) + ϕ (r) = 0.
r

Define ψ (r) := ϕ  (r). Since ψ solves the first-order differential equation

n−1
ψ  (r) + ψ (r) = 0,
r

ψ (r) can be found by the use of an integrating factor. Indeed, multiply by

e(n−1) log r = rn−1 to get
424 Part IV: Partial Differential Equations

rn−1 ψ  (r) + (n − 1)rn−2 ψ (r) = 0,

or

rn−1 ψ (r) = 0.

It follows that
ϕ  (r) = ψ (r) = cr1−n .

Integrate once more to arrive at

⎧
⎪ar + b, n = 1,
cr2−n
+ c1 , n = 2, ⎨
ϕ (r) = 2−n = a log r + b, n = 2,
c log r + c1 , n = 2, ⎪
⎩ 2−n
ar + b, n ≥ 3.

In the opposite direction the result follows from elementary differentiation. 

Definition 39.5. A C2 function u on an open set Ω ⊂ Rn is said to be harmonic on

Ω if Δ u = 0 on Ω .

Exercise 39.1. For u, v ∈ C2 (Ω )∩C1 (Ω ) and for S = ∂ Ω , which is a surfaceGreen’s

identities of class C1 , prove the following:
(1)  
(vΔ u − uΔ v) dx = (v∂ν u − u∂ν v) dσ ;
Ω S

(2)  
(vΔ u + ∇v · ∇u) dx = v∂ν udσ .
Ω S

Exercise 39.2. Prove that if u is harmonic on Ω and u ∈ C1 (Ω ), then


∂ν udσ = 0.
S

Corollary 39.6 (From Green’s identities). Let u ∈ C1 (Ω ) be harmonic on Ω .

(1) if u = 0 on S, then u ≡ 0;
(2) if ∂ν u = 0 on S, then u ≡ constant.

Proof. By taking real and imaginary parts, it suffices to consider real-valued func-
tions. If we let u = v in part (2) of Exercise 39.1, we obtain
 
|∇u|2 dx = u∂ν udσ (x).
Ω S
39 The Laplace Operator 425

In the case (1) we get ∇u ≡ 0, or u ≡ constant. But u ≡ 0 on S implies that u ≡ 0.

In the case (2) we can conclude only that u ≡ constant. 
Theorem 39.7 (The mean value theorem). Suppose u is harmonic on an open set
Ω ⊂ Rn . If x ∈ Ω and r > 0 is small enough that Br (x) ⊂ Ω , then
 
1 1
u(x) = u(y)dσ (y) ≡ u(x + ry)dσ (y),
rn−1 ωn |x−y|=r ωn |y|=1

2π n/2
where ωn = Γ (n/2) is the area of the unit sphere in Rn .

Proof. Let us apply Green’s identity (1) with u and v = |y|2−n if n = 2, and v = log |y|
if n = 2 in the domain

Br (x)\Bε (x) = {y ∈ Rn : ε < |x − y| < r} .

Then for v(y − x) we obtain (n = 2)


0= (vΔ u − uΔ v)dy
Br (x)\Bε (x)
 
= (v∂ν u − u∂ν v)dσ (y) − (v∂ν u − u∂ν v)dσ (y)
|x−y|=r |x−y|=ε
 
= r2−n ∂ν udσ (y) − (2 − n)r1−n udσ (y)
|x−y|=r |x−y|=r
 
− ε 2−n ∂ν udσ (y) + (2 − n)ε 1−n udσ (y). (39.1)
|x−y|=ε |x−y|=ε

In order to get (39.1) we took into account that

x−y x−y d d
∂ν = ν · ∇ = =
r r dr dr
for the sphere. Since u is harmonic, due to Exercise 39.2 we can get from (39.1) that
for all ε > 0, ε < r,
 
ε 1−n udσ (y) = r1−n udσ (y).
|x−y|=ε |x−y|=r

Therefore,
 
lim ε 1−n u(y)dσ (y) = lim u(x + εθ )dθ
ε →0 |x−y|=ε ε →0 |θ |=1

= ωn u(x) = r1−n u(y)dσ (y).
|x−y|=r

This proves the theorem, because the latter steps hold for n = 2 also. 
426 Part IV: Partial Differential Equations

Corollary 39.8. If u and r are as in Theorem 39.7, then

 
n n
u(x) = u(y)dy ≡ u(x + ry)dy, x ∈ Ω. (39.2)
r ωn
n
|x−y|≤r ωn |y|≤1

Proof. Perform integration in polar coordinates and apply Theorem 39.7. 

Remark 39.9. It follows from the latter formula that
ωn
vol {y : |y| ≤ 1} = .
n

Exercise 39.3. Assumethat u is harmonic in Ω . Let χ (x) ∈ C0∞ (B1 (0)) be such
that χ (x) = χ1 (|x|) and Rn χ (x)dx = 1. Define an approximation to the identity by
χε (·) = ε −n χ (ε −1 ·). Prove that

u(x) = χε (x − y)u(y)dy
Bε (x)

for x ∈ Ωε := {x ∈ Ω : Bε (x) ⊂ Ω }.
Corollary 39.10. If u is harmonic on Ω , then u ∈ C∞ (Ω ).
Proof. The statement follows from Exercise 39.3, since the function χε is com-
pactly supported and we may thus differentiate under the integral sign as often as
we please. 
Corollary 39.11. If {uk }∞
k=1 is a sequence of harmonic functions on an open set
Ω ⊂ Rn that converges uniformly on compact subsets of Ω to a limit u, then u is
harmonic on Ω .
Theorem 39.12 (The maximum principle). Suppose Ω ⊂ Rn is open and con-
nected. If u is real-valued and harmonic on Ω with supx∈Ω u(x) = A < ∞, then
either u < A for all x ∈ Ω or u(x) ≡ A in Ω .
Proof. Since u is continuous on Ω , the set {x ∈ Ω : u(x) = A} is closed in Ω . On
the other hand, we may conclude that if u(x) = A at some point x ∈ Ω , then u(y) = A
for all y in a ball about x. Indeed, if y0 ∈ Bσ (x) and u(y0 ) < A, then u(y) < A for all
y from a small neighborhood of y0 . Hence, by Corollary 39.8, for r ≤ σ ,

n
A = u(x) = u(y)dy
rn ωn |x−y|≤r
 
n n
= u(y)dy + n u(y)dy
rn ωn |x−y|≤r,|y0 −y|>ε r ωn |y−y0 |≤ε
   
n n
<A n dy + n dy
r ωn |x−y|≤r,|y0 −y|>ε r ωn |y−y0 |≤ε

n
=A n dy = A,
r ωn |x−y|≤r
39 The Laplace Operator 427

that is, A < A. This contradiction proves our statement. This fact also means that the
set {x ∈ Ω : u(x) = A} is also open. Hence it is either Ω (in this case u ≡ A in Ω )
or the empty set (in this case u(x) < A in Ω ). 
Corollary 39.13. Suppose Ω ⊂ Rn is connected, open, and bounded. If u is real-
valued and harmonic on Ω and continuous on Ω , then the maximum and minimum
of u on Ω are achieved only on ∂ Ω .
Corollary 39.14 (The uniqueness theorem). Suppose Ω is as in Corollary 39.13.
If u1 and u2 are harmonic on Ω and continuous in Ω (possibly complex-valued) and
u1 = u2 on ∂ Ω , then u1 = u2 on Ω .
Proof. The real and imaginary parts of u1 − u2 and u2 − u1 are harmonic on Ω .
Hence they must achieve their maxima on ∂ Ω . These maxima are therefore zero, so
u1 ≡ u2 . 
Theorem 39.15 (Liouville’s theorem). If u is bounded and harmonic on Rn , then
u ≡ constant.
Proof. For all x ∈ Rn and |x| ≤ R, by Corollary 39.8 we have
   
n  
≤ n
|u(x) − u(0)| = u(y)dy − u(y)dy Rn ωn D |u(y)|dy,
Rn ωn  BR (x) BR (0)

where
D = (BR (x)\BR (0)) ∪ (BR (0)\BR (x))

is the symmetric difference of the balls BR (x) and BR (0). Therefore, we obtain
  
n u ∞ n u R+|x|
|u(x) − u(0)| ≤ dy ≤ n ∞ rn−1 dr dθ
R ωn R−|x|≤|y|≤R+|x|
n R ωn R−|x| |θ |=1
 
(R + |x|)n − (R − |x|)n 1
= u ∞ = O u ∞ .
Rn R

Hence the difference |u(x) − u(0)| vanishes as R → ∞, that is, u(x) = u(0). 
Definition 39.16. Afundamental solution for a partial differential operator L is a
distribution K ∈ D  such that
LK = δ .

Remark 39.17. Note that a fundamental solution is not unique. Any two fundamen-
tal solutions differ by a solution of the homogeneous equation Lu = 0.
Exercise 39.4. Show that the characteristic function of the set
 
(x1 , x2 ) ∈ R2 : x1 > 0, x2 > 0

is a fundamental solution for L = ∂1 ∂2 .

428 Part IV: Partial Differential Equations

Exercise 39.5. Prove that the Fourier transform of 1

x1 +ix2 in R2 is equal to − ξ i
.
1 +iξ2

Exercise 39.6. Show that the fundamental solution for the Cauchy–Riemann oper-
ator L = 12 (∂1 + i∂2 ) on R2 is equal to

1 1
.
π x1 + ix2

Since the Laplacian commutes with rotations (Theorem 39.2), it should have a radial
fundamental solution that must be a function of |x| that is harmonic on Rn \ {0}.

Theorem 39.18. Let

|x|2−n
K(x) = (2−n)ωn , n = 2,
(39.3)
2π log |x|, n = 2.
1

Then K is a fundamental solution for Δ .

Proof. For ε > 0 we consider a smoothed-out version Kε of K as

⎧
⎨ (|x|2 +ε 2 ) 2−n
2
Kε (x) = (2−n)ωn , n = 2 (39.4)
⎩ 1
log(|x|2 + ε 2 ), n = 2.
4π

Then Kε → K pointwise (x = 0) as ε → 0+, and Kε and K are dominated by a

fixed locally integrable function for ε ≤ 1 (namely, by |K| for n > 2, | log |x|| + 1
for n = 2, and (|x|2 + 1)1/2 for n = 1). So by Lebesgue’s dominated convergence
theorem, Kε → K in Lloc1 (or in the topology of distributions) as ε → 0+. Hence we

need to show only that Δ Kε → δ as ε → 0 in the sense of distributions, that is,

Δ Kε , ϕ  → ϕ (0), ε →0

for all ϕ ∈ C0∞ (Rn ).

Exercise 39.7. Prove that

Δ Kε (x) = nωn−1 ε 2 (|x|2 + ε 2 )−( 2 +1) ≡ ε −n ψ (ε −1 x)

n

for ψ (y) = nωn−1 (|y|2 + 1)−( 2 +1) .

n

Exercise 39.7 allows us to write

  
Δ Kε , ϕ  = ϕ (x)ε −n ψ (ε −1 x)dx = ϕ (ε z)ψ (z)dz → ϕ (0) ψ (z)dz
Rn Rn Rn
39 The Laplace Operator 429

as ε → 0+. So it remains to show that


ψ (z)dz = 1.
Rn

Using Exercise 39.7, we have

 
n
(|x|2 + 1)−( 2 +1) dx
n
ψ (x)dx =
Rn ωn Rn
 
n ∞ n−1 2
r (r + 1)−( 2 +1) dr
n
= dθ
ωn 0 |θ |=1
 ∞  ∞
n 1
rn−1 (r2 + 1)−( 2 +1) dr =
n n
=n t (n−1)/2 (1 + t)− 2 −1 √ dt
0 2 0 t
 ∞  1 n/2−1
n n n 1 n ds
= t n/2−1 (1 + t)− 2 −1 dt = −1 s 2 +1 2
2 0 2 0 s s
  1
n 1 n
= (1 − s)n/2−1 ds = τ n/2−1 dτ = 1.
2 0 2 0

This means that ε −1 ψ (ε −1 x) is an approximation to the identity and

Δ Kε → δ .

But Kε → K, and so Δ K = δ also. 

Theorem 39.19. Suppose that

(1) f ∈ L1 (Rn ) if n ≥ 3,
(2) R2 | f (y)| (| log |y|| + 1) dy < ∞ if n = 2,
(3) R | f (y)| (1 + |y|) dy < ∞ if n = 1.
Let K be given by (39.3). Then f ∗ K is well defined as a locally integrable function,
and Δ ( f ∗ K) = f in the sense of distributions.

Proof. Let n ≥ 3 and set

1, x ∈ B1 (0),
χ1 (x) =
0, x∈/ B1 (0).

Then χ1 K ∈ L1 (Rn ) and (1 − χ1 )K ∈ L∞ (Rn ). So, for f ∈ L1 (Rn ) we have that

f ∗ (χ1 K) ∈ L1 (Rn ) and f ∗ (1 − χ1 )K ∈ L∞ (Rn ) (see Proposition 37.3). Hence f ∗
1 (Rn ) by addition, and we may calculate
K ∈ Lloc

Δ ( f ∗K), ϕ  =  f ∗ K, Δ ϕ , ϕ ∈ C0∞ (Rn )

  
= ( f ∗ K)(x)Δ ϕ (x)dx = f (y)K(x − y)dyΔ ϕ (x)dx
Rn Rn Rn
430 Part IV: Partial Differential Equations
  
= f (y) K(x − y)Δ ϕ (x)dxdy = f (y)K(x − y), Δ ϕ (x)dy
R Rn Rn
n

= f (y)Δ K(x − y), ϕ (x)dy = f (y)δ (x − y), ϕ (x)dy
R Rn
n

= f (y)ϕ (y)dy =  f , ϕ .
Rn

Hence Δ ( f ∗ K) = f . 

Exercise 39.8. Prove Theorem 39.19 for n = 2.

Exercise 39.9. Prove Theorem 39.19 for n = 1.

Theorem 39.20. Let Ω be a bounded domain in Rn (for n = 1 assume that Ω =

(a, b)) with C1 boundary ∂ Ω = S. If u ∈ C1 (Ω ) is harmonic in Ω , then


u(x) = u(y)∂νy K(x − y) − K(x − y)∂ν u(y) dσ (y), x ∈ Ω, (39.5)
S

where K(x) is the fundamental solution (39.3).

Proof. Let us consider Kε from (39.4). Then since Δ u = 0 in Ω , by Green’s identity

(1) (see Exercise 39.1) we have
 

u(y)Δy Kε (x − y)dy = u(y)∂νy Kε (x − y) − Kε (x − y)∂ν u(y) dσ (y).
Ω S

As ε → 0, the right-hand side of this equation tends to the right-hand side of (39.5)
for each x ∈ Ω , since for x ∈ Ω and y ∈ S there are no singularities in K. On the other
hand, the left-hand side is just (u ∗ Δ Kε ) (x) if we set u ≡ 0 outside Ω . According to
the proof of Theorem 39.18,

(u ∗ Δ Kε ) (x) → u(x), ε → 0,

completing the proof. 

Remark 39.21. If we know that u = f and ∂ν u = g on S, then



u(x) = f (y)∂νy K(x − y) − K(x − y)g(y) dσ (y)
S

is the solution of Δ u = 0 with Cauchy data on S. But this problem is overdetermined,

because we know from Corollary 39.14 that the solution of Δ u = 0 is uniquely
determined by f alone.
39 The Laplace Operator 431

The following theorem concerns the spaces Cα (Ω ) and Ck,α (Ω ) defined by

Cα (Ω ) ≡ C0,α (Ω ) = {u ∈ L∞ (Ω ) : |u(x) − u(y)| ≤ C|x − y|α , x, y ∈ Ω } ,

 
Ck,α (Ω ) ≡ Ck+α (Ω ) = u : ∂ β u ∈ Cα (Ω ), |β | ≤ k ,

for 0 < α < 1 and k ∈ N.

Theorem 39.22 (Regularity in Hölder spaces). Suppose k ≥ 0 is an integer, 0 <
α < 1, and Ω ⊂ Rn is open. If f ∈ Ck+α (Ω ) and u is a distributional solution of
k+2+α
Δ u = f in Ω , then u ∈ Cloc (Ω ).
Proof. Since Δ (∂ β u) = ∂ β Δ u = ∂ β f , we can assume without loss of generality
that k = 0. Given Ω1 ⊂ Ω such that Ω1 ⊂ Ω choose ϕ ∈ C0∞ (Ω ) such that ϕ ≡ 1 on
Ω1 and let g = ϕ f .
Since Δ (g ∗ K) = g (see Theorem 39.19) and therefore Δ (g ∗ K) = f in Ω1 , it
follows that u − (g ∗ K) is harmonic in Ω1 and hence C∞ there. It is therefore enough
to prove that if g is a Cα function with compact support, then g ∗ K ∈ C2+α . To
this end we consider Kε (x) and its derivatives. Straightforward calculations lead to
following formulas (n ≥ 1):

∂
Kε (x) = ωn−1 x j (|x|2 + ε 2 )−n/2 ,
∂xj
(39.6)
∂2 −nxi x j (|x|2 + ε 2 )−n/2−1 , i = j,
Kε (x) = ωn−1 −n/2−1
∂ xi ∂ x j (|x| + ε − nx j )(|x| + ε )
2 2 2 2 2 , i = j.

Exercise 39.10. Prove formulas (39.6).

Since Kε ∈ C∞ , we have g ∗ Kε ∈ C∞ also. Moreover, ∂ j (g ∗ Kε ) = g ∗ ∂ j Kε and
∂i ∂ j (g ∗ Kε ) = g ∗ ∂i ∂ j Kε . The pointwise limits in (39.6) as ε → 0 imply

∂
K(x) = ωn−1 x j |x|−n ,
∂xj
(39.7)
∂2 −nωn−1 xi x j |x|−n−2 , i = j,
K(x) =
∂ xi ∂ x j ωn−1 (|x|2 − nx2j )|x|−n−2 , i = j,

for x = 0. Formulas (39.7) show that ∂ j K(x) is a locally integrable function, and
since g is bounded with compact support, it follows that g ∗ ∂ j K is continuous. Next,
g ∗ ∂ j Kε → g ∗ ∂ j K uniformly as ε → 0+. This is equivalent to ∂ j Kε → ∂ j K in the
topology of distributions (see the definition). Hence ∂ j (g ∗ K) = g ∗ ∂ j K.
This argument does not work for the second derivatives, because ∂i ∂ j K(x) is not
integrable. But there is a different procedure for these terms.
Let i = j. Then ∂i ∂ j Kε (x) and ∂i ∂ j K(x) are odd functions of xi (and x j ); see
(39.6) and (39.7). Due to this fact, their integrals over an annulus 0 < a < |x| < b
vanish. For Kε we can even take a = 0.
432 Part IV: Partial Differential Equations

Exercise 39.11. Prove this fact.

Therefore, for all b > 0 we have
 
g ∗ ∂i ∂ j Kε (x) = g(x − y)∂i ∂ j Kε (y)dy − g(x) ∂i ∂ j Kε (y)dy
Rn |y|<b
 
= (g(x − y) − g(x))∂i ∂ j Kε (y)dy + g(x − y)∂i ∂ j Kε (y)dy.
|y|<b |y|≥b

If we let ε → 0, we obtain

lim g ∗ ∂i ∂ j Kε (x)
ε →0
 
= (g(x − y) − g(x))∂i ∂ j K(y)dy + g(x − y)∂i ∂ j K(y)dy.
|y|<b |y|≥b

This limit exists because

|g(x − y) − g(x)||∂i ∂ j K(y)| ≤ c|y|α |y|−n

(g is Cα ) and because g is compactly supported. Then, since b is arbitrary, we can

let b → +∞ to obtain

∂i ∂ j (g ∗ K)(x) = lim (g(x − y) − g(x))∂i ∂ j K(y)dy
b→∞ |y|<b

+ lim g(x − y)∂i ∂ j K(y)dy
b→∞ |y|≥b

= lim (g(x − y) − g(x))∂i ∂ j K(y)dy. (39.8)
b→∞ |y|<b

A similar result holds for i = j. Indeed,

1
∂ j2 Kε (x) = ε −n ψ (ε −1 x) + K εj (x),
n

where ψ (x) = nωn−1 (|x|2 +1)−n/2−1 and K εj = ωn−1 (|x|2 −nx2j )(|x|2 + ε 2 )−n/2−1 (see
(39.6)). The integral I j of K εj over an annulus a < |y| < b vanishes. Why is that so?
First of all, I j is independent of j by symmetry in the coordinates, that is, I j = Ii for
i = j. So nI j is the integral of ∑nj=1 K εj . But ∑nj=1 K εj = 0. Hence I j = 0 also. We can
therefore apply the same procedure. Since

g ∗ (ε −n ψ (ε −1 x)) → g, ε →0

(because ε −n ψ (ε −1 x) is an approximation to the identity), it follows that

39 The Laplace Operator 433

g(x)
∂ j2 (g ∗ K)(x) = + lim (g(x − y) − g(x))∂ j2 K(y)dy. (39.9)
n b→∞ |y|<b

Since the convergence in (39.8) and (39.9) is uniform, at this point we have shown
that g ∗ K ∈ C2 . But we need to prove more.

Lemma 39.23 (Calderon–Zigmund). Let N be a C1 function on Rn \ {0} that is

homogeneous of degree −n and satisfies

N(y)dy = 0
a<|y|<b

for all 0 < a < b < ∞. Then if g is a Cα function with compact support, 0 < α < 1,
then 
h(x) = lim (g(x − z) − g(x))N(z)dz
b→∞ |z|<b

belongs to Cα .

Proof. Let us write h = h1 + h2 , where


h1 (x) = (g(x − z) − g(x))N(z)dz,
|z|≤3|y|

h2 (x) = lim (g(x − z) − g(x))N(z)dz.
b→∞ 3|y|<|z|<b

We wish to estimate h(x + y) − h(x). Since α > 0, we have


|h1 (x)| ≤ c |z|α |z|−n dz = c |y|α
|z|≤3|y|

and hence
|h1 (x + y) − h1 (x)| ≤ |h1 (x + y)| + |h1 (x)| ≤ 2c |y|α .

On the other hand,


h2 (x + y) − h2 (x) = lim (g(x − z) − g(x))N(z + y)dz
b→∞ 3|y|<|z+y|<b

− lim (g(x − z) − g(x))N(z)dz
b→∞ 3|y|<|z|<b

= lim (g(x − z) − g(x))(N(z + y) − N(z))dz
b→∞ 3|y|<|z|<b

+ lim (g(x − z) − g(x))N(z + y)dz
b→∞ {3|y|<|z+y|<b}\{3|y|<|z|<b}

=: I1 + I2 .
434 Part IV: Partial Differential Equations

It is clear that

{3|y| < |z + y|} \ {3|y| < |z|} ⊂ {2|y| < |z|} \ {3|y| < |z|}
= {2|y| < |z| ≤ 3|y|} .

Therefore,

|I2 | ≤ |g(x − z) − g(x)||N(z + y)|dz
2|y|<|z|≤3|y|
 
≤c |z|α |z + y|−n dz ≤ c |z|α −n dz = c |y|α .
2|y|<|z|≤3|y| 2|y|<|z|≤3|y|

Now we observe that for |z| > 3|y|,

|N(z + y) − N(z)| ≤ |y| sup |∇N(z + ty)|

0≤t≤1

≤ c|y| sup |z + ty|−n−1 ≤ c |y||z|−n−1 ,

0≤t≤1

because ∇N is homogeneous of degree −n − 1, since N is homogeneous of degree

−n. Hence
  ∞
|I1 | ≤ c |z|α |y||z|−n−1 dz = c |y| ρ α −2 dρ = c |y|α .
|z|>3|y| 3|y|

Note that the condition α < 1 is needed here. Collecting the estimates for I1 and I2 ,
we can see that the lemma is proved. 
In order to end the proof of Theorem 39.22 it remains to note that ∂i ∂ j K(x)
satisfies all the conditions of Lemma 39.23. 
Exercise 39.12. Show that a function K1 is a fundamental solution for Δ 2 ≡ Δ (Δ )
on Rn if and only if K1 satisfies the equation

Δ K1 = K,

where K is the fundamental solution for the Laplacian.

Exercise 39.13. Show that the following functions are the fundamental solutions
for Δ 2 on Rn :
(1) n = 4:
log |x|
− ;
4ω4

(2) n = 2:
|x|2 log |x|
;
8π
39 The Laplace Operator 435

(3) n = 2, 4:
|x|4−n
.
2(4 − n)(2 − n)ωn

Exercise 39.14. Show that (4π |x|)−1 e−c|x| is the fundamental solution for −Δ + c2
on R3 for an arbitrary constant c ∈ C.
Chapter 40
The Dirichlet and Neumann Problems

The Dirichlet problem

Given functions f in Ω and g on S = ∂ Ω , find a function u in Ω = Ω ∪ ∂ Ω satisfying

Δ u = f , in Ω
(D)
u = g, on S.

The Neumann problem

Given functions f in Ω and g on S, find a function u in Ω satisfying

Δ u = f , in Ω
(N)
∂ν u = g, on S.

We assume that Ω is bounded with C1 boundary. But we shall not, however, assume
that Ω is connected. The uniqueness theorem (see Corollary 39.14) shows that the
solution of (D) will be unique (if it exists), at least if we require u ∈ C(Ω ). For (N)
uniqueness does not hold: we can add to u(x) any function that is constant on each
connected component of Ω . Moreover, there is an obvious necessary condition for
solvability of (N). If Ω  is a connected component of Ω , then
   
Δ udx = ∂ν udσ (x) = g(x)dσ (x) = f dx,
Ω ∂Ω ∂Ω Ω

that is,  
f (x)dx = g(x)dσ (x).
Ω ∂Ω

It is also clear (by linearity) that (D) can be reduced to the following homoge-
neous problems:

c Springer International Publishing AG 2017 437

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 40
438 Part IV: Partial Differential Equations


Δ v = f , in Ω
(DA )
v = 0, on S

Δ w = 0, in Ω
(DB )
w = g, on S

and u := v + w solves (D). Similar remarks apply to (N), that is,


Δ v = f , in Ω
∂ν v = 0, on S

Δ w = 0, in Ω
∂ν w = g, on S

and u = v + w.

Definition 40.1. The Green’s function for (D) in Ω is the solution G(x, y) of the
boundary value problem

Δx G(x, y) = δ (x − y), x, y ∈ Ω
(40.1)
G(x, y) = 0, x ∈ S, y ∈ Ω .

Analogously, the Green’s function for (N) in Ω is the solution G(x, y) of the bound-
ary value problem

Δx G(x, y) = δ (x − y), x, y ∈ Ω
(40.2)
∂νx G(x, y) = 0, x ∈ S, y ∈ Ω .

This definition allows us to write

G(x, y) = K(x − y) + vy (x), (40.3)

where K is the fundamental solution of Δ in Rn and for all y ∈ Ω , the function vy (x)
satisfies 
Δ vy (x) = 0, in Ω
(40.4)
vy (x) = −K(x − y), on S

in the case of (40.1) and


Δ vy (x) = 0, in Ω
∂νx vy (x) = −∂νx K(x − y), on S
40 The Dirichlet and Neumann Problems 439

in the case of (40.2). Since (40.4) guarantees that vy is real, it follows that so is G
corresponding to (40.1).
Lemma 40.2. The Green’s function (40.1) exists and is unique.
Proof. The uniqueness of G follows again from Corollary 39.14, since K(x − y) in
(40.4) is continuous for all x ∈ S and y ∈ Ω (x = y). The existence will be proved
later. 
Lemma 40.3. For both (40.1) and (40.2) it is true that G(x, y) = G(y, x) for all
x, y ∈ Ω .
Proof. Let G(x, y) and G(x, z) be the Green’s functions for Ω corresponding to
sources located at fixed y and z, y = z, respectively. Let us consider the domain

Ωε = (Ω \ {x : |x − y| < ε }) \ {x : |x − z| < ε } ,

see Figure 40.1.

Fig. 40.1 The domain Ωε .

If x ∈ Ωε , then x = z and x = y, and therefore, Δx G(x, z) = 0 and Δx G(x, y) = 0.

These facts imply

0= (G(x, y)Δx G(x, z) − G(x, z)Δx G(x, y)) dx
Ωε

= (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ (x)
S
− (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ (x)
|x−y|=ε

− (G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ (x).
|x−z|=ε

Hence by (40.1) or (40.2), for arbitrary ε > 0 (sufficiently small),


(G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ (x)
|x−y|=ε

= (G(x, z)∂νx G(x, y) − G(x, y)∂νx G(x, z)) dσ (x).
|x−z|=ε
440 Part IV: Partial Differential Equations

Let n ≥ 3. Due to (40.3) for ε → 0 we have


(G(x, y)∂νx G(x, z) − G(x, z)∂νx G(x, y)) dσ (x)
|x−y|=ε
  
1 (x − y, x − z)
≈ ε 2−n
+ ∂νx vz (x) dσ (x)
ωn |x−y|=ε |x − y||x − z|n

− G(x, z)∂νx G(x, y)dσ (x)
|x−y|=ε

1 2−n n−1 1 (εθ , εθ + y − z)
≈ ε ε dθ − I1 ≈ −I1 ,
ωn ε θ |εθ + y − z|n

where we have defined


I1 = G(x, z)∂νx G(x, y)dσ (x).
|x−y|=ε

The same is true for the integral over |x − z| = ε , that is,


(G(x, z)∂νx G(x, y) − G(x, y)∂νx G(x, z)) dσ (x) ≈ −I2 , ε → 0,
|x−z|=ε

where 
I2 = G(x, y)∂νx G(x, z)dσ (x).
|x−z|=ε

But using the previous techniques we can obtain that


1 1−n n−1
I1 ≈ ε ε G(εθ + y, z)dθ → G(y, z), ε →0
ωn |θ |=1

and 
1 1−n n−1
I2 ≈ ε ε G(εθ + z, y)dθ → G(z, y), ε → 0.
ωn |θ |=1

This means that G(y, z) = G(z, y) for all z = y. This proof holds for n = 2 (and even
for n = 1) with some simple changes. 

Lemma 40.4. In three or more dimensions

K(x − y) < G(x, y) < 0, x, y ∈ Ω , x = y

where G(x, y) is the Green’s function for (D).

Proof. For each fixed y, the function vy (x) := G(x, y) − K(x − y) is harmonic in Ω ;
see (40.4). Moreover, on S = ∂ Ω , vy (x) takes on the positive value
40 The Dirichlet and Neumann Problems 441

|x − y|2−n
−K(x − y) ≡ − .
ωn (2 − n)

By the minimum principle, it follows that vy (x) is strictly positive in Ω . This proves
the first inequality. 

Exercise 40.1. Prove the second inequality in Lemma 40.4.

Exercise 40.2. Show that for n = 2, Lemma 40.4 has the following form:

1 |x − y|
log < G(x, y) < 0, x, y ∈ Ω ,
2π h

where h ≡ maxx,y∈Ω |x − y|.

Exercise 40.3. Obtain the analogue of Lemma 40.4 for n = 1. Hint: show that the
d2
Green’s function for the operator dx 2 on Ω = (0, 1) is


x(y − 1), x < y
G(x, y) =
y(x − 1), x > y.

Remark 40.5. G(x, y) may be extended naturally (because of the symmetry) to Ω ×

Ω by setting G(x, y) = 0 for y ∈ S.

Now we can solve both problems (DA ) and (DB ). Indeed, let us set f = 0 in (DA )
outside Ω and define
 
v(x) := G(x, y) f (y)dy ≡ ( f ∗ K)(x) + (G(x, y) − K(x − y)) f (y)dy.
Ω Ω

Then the Laplacian of the first term is f (see Theorem 39.19), and the second term
is harmonic in x (since vy (x) is harmonic). Also v(x) = 0 on S, because the same is
true for G. Thus, this v(x) solves (DA ).
Consider now (DB ). We assume that g is continuous on S and we wish to find
w that is continuous on Ω . Applying Green’s identity (1) (together with the same
limiting process as in the proof of Lemma 40.3), we obtain

w(x) = (w(y)Δy G(x, y) − G(x, y)Δ w(y)) dy
Ω 
= w(y)∂νy G(x, y)dσ (y) = g(y)∂νy G(x, y)dσ (y).
S S

Let us denote the last integral by (P). Since ∂νy G(x, y) is harmonic in x and continu-
ous in y for x ∈ Ω and y ∈ S, then w(x) is harmonic in Ω . In order to prove that this
w(x) solves (DB ), it remains to prove that w(x) is continuous in Ω and w(x) on S is
g(x). We will prove this general fact later.
442 Part IV: Partial Differential Equations

Definition 40.6. The function ∂νy G(x, y) on Ω × S is called the Poisson kernel for
Ω , and (P) is called the Poisson integral.
Now we are in a position to solve the Dirichlet problem in a half-space. Let
  
Ω = Rn+1
+ = (x , xn+1 ) ∈ R
n+1
: x ∈ Rn , xn+1 > 0 ,

where n ≥ 1 now, and let xn+1 = t. Then

Δn+1 = Δn + ∂t2 , n = 1, 2, . . . .

Denote by K(x,t) a fundamental solution for Δn+1 in Rn+1 , that is,

⎧
⎨ (|x|2 +t 2 ) 1−n
2
K(x,t) = (1−n)ωn+1 , n>1
⎩
4π log(|x| + t ), n = 1.
1 2 2

Let us prove then that the Green’s function for Rn+1

+ is

G(x, y;t, s) = K(x − y,t − s) − K(x − y, −t − s). (40.5)

It is clear (see (40.5)) that G(x, y;t, 0) = G(x, y; 0, s) = 0 and

Δn+1 G = δ (x − y,t − s) − δ (x − y, −t − s) = δ (x − y)δ (t − s),

because for t, s > 0, −t − s < 0, and therefore, δ (−t − s) = 0. Thus G is the Dirichlet
Green’s function for Rn+1+ . From this we immediately have the solution of (DA ) in
R+ as
n+1
  ∞
u(x,t) = G(x, y;t, s) f (y, s)dsdy.
Rn 0

To solve (DB ) we compute the Poisson kernel for this case. Since the outward normal
∂
derivative on ∂ Rn+1
+ is − ∂ t , the Poisson kernel becomes

∂ ∂
− G(x, y;t, s)|s=0 = − (K(x − y,t − s) − K(x − y, −t − s)) |s=0
∂s ∂s
2t
= n+1 . (40.6)
ωn+1 (|x − y|2 + t 2 ) 2

Exercise 40.4. Prove (40.6).

Note that (40.6) holds for all n ≥ 1. According to the formula for (P), the candi-
date for a solution to (DB ) is

2 tg(y)
u(x,t) = dy. (40.7)
ωn+1 Rn (|x − y|2 + t 2 )
n+1
2
40 The Dirichlet and Neumann Problems 443

In other words, if we set

2t
Pt (x) := n+1 , (40.8)
ωn+1 (|x|2 + t 2 ) 2

which is what is usually called the Poisson kernel for Rn+1

+ , the proposed solution
(40.7) is simply equal to
u(x,t) = (g ∗ Pt )(x). (40.9)

Exercise 40.5. Prove that Pt (x) = t −n P1 (t −1 x) and


Pt (y)dy = 1.
Rn

Theorem 40.7. Suppose g ∈ L p (Rn ), 1 ≤ p ≤ ∞. Then u(x,t) from (40.9) is well

defined on Rn+1
+ and is harmonic there. If g is bounded and uniformly continuous,
then u(x,t) is continuous on Rn+1
+ and u(x, 0) = g(x), and

u(·,t) − g(·)∞ → 0

as t → 0+.

Proof. It is clear that for all t > 0, Pt (x) ∈ L1 (Rn ) ∩ L∞ (Rn ); see (40.8). Hence
Pt (x) ∈ Lq (Rn ) for all q ∈ [1, ∞] with respect to x and fixed t > 0. Therefore, the
integral in (40.9) is absolutely convergent, and the same is true if Pt is replaced by
its derivative Δx Pt or ∂t2 Pt (due to Young’s inequality for convolution).
Since G(x, y;t, s) is harmonic for (x,t) = (y, s), it follows that Pt (x) is also har-
monic and
Δx u + ∂t2 u = g ∗ (Δx + ∂t2 )Pt = 0.

It remains to prove that if g is bounded and continuous, then

u(·,t) − g(·)∞ → 0

as t → 0+, and therefore, u(x, 0) = g(x) and u is continuous on Rn+1 + .

We have (see Exercise 40.5)
  
 
g ∗ Pt − g∞ = sup  g(x − y)Pt (y)dy − g(x)Pt (y)dy
n
x∈R Rn Rn

≤ sup |g(x − y) − g(x)||Pt (y)|dy
x∈Rn Rn

= sup |g(x − tz) − g(x)||P1 (z)|dz
x∈Rn Rn
444 Part IV: Partial Differential Equations

= sup |g(x − tz) − g(x)||P1 (z)|dz
x∈Rn |z|<R
 
+ |g(x − tz) − g(x)||P1 (z)|dz
|z|≥R

≤ sup |g(x − tz) − g(x)| + 2 g∞ |P1 (z)|dz < ε
x∈Rn ,|z|<R |z|≥R

for t sufficiently small.

The first term in the sum on the last line can be made less than ε /2, since g is
uniformly continuous on Rn . The second term can be made less than ε /2 for R large
enough, since P1 ∈ L1 (Rn ). Thus, the theorem is proved. 

Remark 40.8. The solution of this problem is not unique: if u(x,t) is a solution,
then so is u(x,t) + ct for all c ∈ C. However, we have the following theorem.

Theorem 40.9. If g ∈ C(Rn ) and limx→∞ g(x) = 0, then u(x,t) := (g ∗ Pt )(x) → 0

as (x,t) → ∞ in Rn+1
+ , and it is the unique solution with this property.

Proof. Assume for the moment that g has compact support, say g = 0 for |x| > R.
Then g ∈ L1 (Rn ) and

g ∗ Pt ∞ ≤ g1 Pt ∞ ≤ ct −n ,

so u(x,t) → 0 as t → ∞ uniformly in x. On the other hand, if 0 < t ≤ T , then

2t
|u(x,t)| ≤ g1 sup |Pt (x − y)| = g1 sup n+1 ≤ cT |x|−n−1 ,
|y|<R |y|<R ωn+1 (|x − y|2 + t 2 ) 2

for |x| > 2R. Hence u(x,t) → 0 as x → ∞ uniformly for t ∈ [0, T ]. This proves that
u(x,t) vanishes at infinity if g(x) has compact support. For general g, choose a se-
quence {gk } of compactly supported functions that converges uniformly (in L∞ (Rn ))
to g and let
uk (x,t) = (gk ∗ Pt )(x).

Then
 
 

uk − uL∞ (Rn+1 ) = sup  (gk − g)(y)Pt (x − y)dy
t,x R n
  
≤ sup gk − gL∞ (Rn ) sup |Pt (x − y)|dy
t x Rn

= gk − gL∞ (Rn ) sup |Pt (y)|dy = gk − gL∞ (Rn ) → 0
t>0 Rn

as k → ∞.
40 The Dirichlet and Neumann Problems 445

Hence u(x,t) vanishes at infinity. Now suppose v is another solution and let
w := v − u. Then w vanishes at infinity and also at t = 0 (see Theorem 40.7). Thus
|w| < ε on the boundary of the cylindrical region {(x,t) : |x| < R, 0 < t < R} for R
sufficiently large, see Figure 40.2.

t =R

R t =0

Fig. 40.2 Geometric illustration of the Poisson integral.

But since w is harmonic, it follows by the maximum principle that |w| < ε in this
region. Letting ε → 0 and R → ∞, we conclude that w ≡ 0. 

Let us consider now the Dirichlet problem in a ball. We use here the following
notation:
B = B1 (0) = {x ∈ Rn : |x| < 1} , ∂ B = S.

Exercise 40.6. Prove that  

x 
|x − y| =  − y|x|

|x|

for x, y ∈ Rn , x = 0, |y| = 1.

Now, assuming first that n > 2, we define

 
x
G(x, y) := K(x − y) − K − y|x|
|x|
  2−n
1 2−n  x
 
= |x − y| −  − y|x| , x = 0. (40.10)
(2 − n)ωn |x|

Exercise 40.6 shows that G(x, y) from (40.10) satisfies G(x, y) = 0, x ∈ B, y ∈ S.

It is also clear that G(x, y) = G(y, x). This is true because
446 Part IV: Partial Differential Equations
 2  2
x   
 − y|x| =  x  − 2(x, y) + |y|2 |x|2 = 1 − 2(x, y) + |y|2 |x|2
 |x|   |x| 
 2  2
y y 
=   − 2(y, x) + |x|2 |y|2 =  − x|y| .
|y| |y|

Next, for x, y ∈ B we have that

 
 x  |x| 1
 
 |x|2  = |x|2 = |x| > 1

and y = x
|x|2
. Hence,
 
x
G(x, y) − K(x − y) ≡ −K − y|x|
|x|

is harmonic in y. But the symmetry of G and K shows also that G(x, y) − K(x − y)
is harmonic in x. Thus, G(x, y) is the Green’s function for B. This also makes clear
how to define G at x = 0 (and at y = 0):

1
G(0, y) = (|y|2−n − 1),
(2 − n)ωn

since  
x 
 − y|x| → 1
 |x| 

as x → 0.
For n = 2 the analogous formulae are
  
1 x  1
G(x, y) = log |x − y| − log  − y|x| ,
 G(0, y) = log |y|.
2π |x| 2π

Now we can compute the Poisson kernel P(x, y) := ∂νy G(x, y), x ∈ B, y ∈ S. Since
∂νy = y · ∇y on S, it follows that
⎛  ⎞
1 ⎝ (y, x − y)
x
|x| − y|x|, y|x| 1 − |x|2
P(x, y) = − −  n ⎠ ≡ , n ≥ 2. (40.11)
ωn |x − y|n x  ωn |x − y|n
 |x| − y|x|

Exercise 40.7. Prove (40.11).

Theorem 40.10. If f ∈ L1 (S), then


u(x) = P(x, y) f (y)dσ (y), x ∈ B,
S
40 The Dirichlet and Neumann Problems 447

is harmonic. If f ∈ C(S), then u extends continuously to B and u = f on S.

Proof. For each x ∈ B (see (40.11)), P(x, y) is a bounded function of y ∈ S, so u(x)

is well defined for f ∈ L1 (S). It is also harmonic in B, because P(x, y) is harmonic
for x = y. Next, we claim that

P(x, y)dσ (y) = 1. (40.12)
S

Since P is harmonic in x, the mean value theorem implies (y ∈ S)


1 = ωn P(0, y) = P(ry , y)dσ (y )
S

for all 0 < r < 1. But

P(ry , y) = P(y, ry ) = P(ry, y )

if y, y ∈ S. The last formula follows from

|ry − y|2 = r2 − 2r(y , y) + 1 = |ry − y |2 .

We therefore conclude that

 
1= P(ry , y)dσ (y ) = P(x, y )dσ (y )
S S

with x = ry. This proves (40.12). We claim also that for all y0 ∈ S and for a neigh-
borhood Bσ (y0 ) ⊂ S,

lim P(ry0 , y)dσ (y) = 0. (40.13)
r→1− S\Bσ (y0 )

Indeed, for y0 , y ∈ S and 0 < r < 1,

|ry0 − y| > r|y0 − y|

and therefore
|ry0 − y|−n < (r|y0 − y|)−n ≤ (rσ )−n

if y ∈ S\Bσ (y0 ), i.e., |y − y0 | ≥ σ . Hence |ry0 − y|−n is bounded uniformly for r →

1− and y ∈ S\Bσ (y0 ). In addition, 1 − |ry0 |2 ≡ 1 − r2 → 0 as r → 1−. This proves
(40.13).
Now suppose f ∈ C(S). Hence f is uniformly continuous, since S is compact.
Hence for every ε > 0 there exists δ > 0 such that

| f (x) − f (y)| < ε , x, y ∈ S, |x − y| < δ .

448 Part IV: Partial Differential Equations

For all x ∈ S and 0 < r < 1, by (40.12),

 
 
|u(rx) − f (x)| =  ( f (y) − f (x))P(rx, y)dσ (y)

S

≤ | f (y) − f (x)||P(rx, y)|dσ (y)
|x−y|<δ

+ | f (y) − f (x)||P(rx, y)|dσ (y)
S\Bδ (x)
 
≤ε |P(rx, y)|dσ (y) + 2  f ∞ |P(rx, y)|dσ (y)
S S\Bδ (x)

≤ ε + 2  f ∞ P(rx, y)dσ (y) → 0,
S\Bδ (x)

as ε → 0 and r → 1− by (40.13). Hence u(rx) → f uniformly as r → 1−. 

Corollary 40.11. (Without proof) If u is as in Theorem 40.10 and f ∈ L p (S), 1 ≤
p ≤ ∞, then
u(r·) − f (·) p → 0

as r → 1−.
Exercise 40.8. Show that the Poisson kernel for the ball BR (x0 ) is

R2 − |x − x0 |2
P(x, y) = , n ≥ 2.
ωn R|x − y|n

Exercise 40.9. (Harnack’s inequality) Suppose u ∈ C(B) is harmonic on B and

u ≥ 0. Show that for |x| = r < 1,

1−r 1+r
u(0) ≤ u(x) ≤ u(0).
(1 + r)n−1 (1 − r)n−1

Theorem 40.12. (The reflection principle) Let Ω ⊂ Rn+1 , n ≥ 1, be open and

satisfy the property that (x, −t) ∈ Ω if (x,t) ∈ Ω . Let Ω+ = {(x,t) ∈ Ω : t > 0}
and Ω0 = {(x,t) ∈ Ω : t = 0}. If u(x,t) is continuous on Ω+ ∪ Ω0 , harmonic in
Ω+ , and u(x, 0) = 0, then we can extend u to be harmonic on Ω by setting
u(x, −t) := −u(x,t).
Proof. See [11, (2.68), p. 110]. 
Definition 40.13. If u is harmonic on Ω \ {x0 }, Ω ⊂ Rn open, then u is said to have
a removable singularity x0 if u can be defined at x0 so as to be harmonic in Ω .
 
Theorem 40.14. Suppose u is harmonic on Ω \ {x0 } and u(x) = o |x − x0 |2−n for
n > 2 and u(x) = o (log |x − x0 |) for n = 2 as x → x0 . Then u has a removable sin-
gularity at x0 .
40 The Dirichlet and Neumann Problems 449

Proof. Without loss of generality we assume that Ω = B := B1 (0) and x0 = 0. Since

u is continuous on ∂ B, by Theorem 40.10 there exists v ∈ C(B) satisfying

Δ v = 0, in B
v = u, on S.

We claim that u = v in B\ {0}, so that we can remove the singularity at {0} by

setting u(0) := v(0). Indeed, given ε > 0 and 0 < δ < 1, consider the function

u(x) − v(x) − ε (|x|2−n − 1), n > 2
gε (x) =
u(x) − v(x) + ε log |x|, n=2

in B\Bδ (0). These functions are real (as we can assume without loss of generality),
harmonic, and continuous for δ ≤ |x| ≤ 1. Moreover, gε (x) = 0 on ∂ B and gε (x) < 0
on ∂ Bδ (0) for all δ sufficiently small. By the maximum principle, gε (x) is negative
in B\ {0}. Letting ε → 0, we see that u − v ≤ 0 in B\ {0}. By the same arguments
we may conclude that also v − u ≤ 0 in B\ {0}. Hence u = v in B\ {0}, and we can
extend u to the whole ball by setting u(0) = v(0). This proves the theorem. 
Chapter 41
Layer Potentials

In this chapter we assume that Ω ⊂ Rn , n ≥ 2 is bounded and open, and that S = ∂ Ω

is a surface of class C2 . We assume also that both Ω and Ω  := Rn \ Ω are connected.

Definition 41.1. Let ν (x) be a normal vector to S at x. Then

∂ν− u(x) := lim ν (x) · ∇u(x + t ν (x)),

t→0−
∂ν+ u(x) := lim ν (x) · ∇u(x + t ν (x)),
t→0+

are called the interior and exterior normal derivatives, respectively, of u.

The interior Dirichlet problem (ID)
Given f ∈ C(S), find u ∈ C2 (Ω ) ∩C(Ω ) such that Δ u = 0 in Ω and u = f on S.
The exterior Dirichlet problem (ED)
Given f ∈ C(S), find u ∈ C2 (Ω  ) ∩C(Ω  ) such that Δ u = 0 in Ω  and at infinity and
u = f on S.
Definition 41.2. A function u is said to be harmonic at infinity if
  
x o(|x|2−n ), n = 2
|x|2−n
u =
|x|2 o (log |x|) , n = 2

as x → 0.
Remark 41.3. This definition implies the following behaviour of u at infinity

o(1), n = 2
u(y) =
o(log |y|), n = 2

as y → ∞.

c Springer International Publishing AG 2017 451

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 41
452 Part IV: Partial Differential Equations

The interior Neumann problem (IN)

Given f ∈ C(S), find u ∈ C2 (Ω ) ∩C(Ω ) such that Δ u = 0 in Ω and ∂ν− u = f exists
on S.
The exterior Neumann problem (EN)
Given f ∈ C(S), find u ∈ C2 (Ω  ) ∩C(Ω  ) such that Δ u = 0 in Ω  and at infinity and
∂ν+ u = f exists on S.
Theorem 41.4. (Uniqueness)
(1) The solutions of (ID) and (ED) are unique.
(2) The solutions of (IN) and (EN) are unique up to a constant on Ω and Ω  ,
respectively. When n > 2 this constant is zero on the unbounded component of
Ω .
Proof. If u solves (ID) with f = 0, then u ≡ 0, because this is just the uniqueness
 If u solves (ED) with f = 0,
theorem for harmonic functions (see Corollary 39.14).
we may assume that {0} ∈ / Ω  . Then u = |x|2−n u |x|x 2 solves (ID) with f = 0 for

the bounded domain Ω  = x : x 2 ∈ Ω  . Hence u ≡ 0, so that u ≡ 0, and part (1)
|x|
is proved.
 
Exercise 41.1. Prove that if u is harmonic, then u = |x|2−n u |x|x 2 , x = 0, is also
harmonic.
Concerning part (2), by Green’s identity we have

|∇u|2 dx = − uΔ udx + u∂ν− udσ (x).

Ω Ω S

Thus ∇u = 0 in Ω , so that u is constant in Ω .

For (EN) let r > 0 be large enough that Ω ⊂ Br (0). Again by Green’s identity we
have

|∇u|2 dx = − uΔ udx + u∂r udσ (x) − u∂ν+ udσ (x)

Br (0)\Ω Br (0)\Ω ∂ Br (0) S

= u∂r udσ (x),

∂ Br (0)

where ∂r u ≡ d
dr u. Since for n > 2 and for large |x| we have
   
u(x) = O |x|2−n , ∂r u(x) = O |x|1−n ,

it follows that

u∂r udσ (x) ≤ cr r
2−n 1−n
dσ (x) = cr3−2n rn−1 = cr2−n → 0
∂ Br (0) ∂ Br (0)
41 Layer Potentials 453

as r → ∞. Hence

|∇u|2 dx = 0.
Ω

This implies that u is constant on Ω  and u = 0 on the unbounded component of Ω  ,

because for large |x|,  
u(x) = O |x|2−n , n > 2.
 
If n = 2 then ∂r u(x) = O r−2 for a function u(x) that is harmonic at infinity.
Exercise 41.2.
 Prove
 that if u is harmonic at infinity, then u is bounded and
∂r u(x) = O r−2 as r → ∞ if n = 2 and ∂r u(x) = O(|x|1−n ), r → ∞, if n > 2.
By Exercise 41.2 we obtain



−2 −1
∂ B (0) u∂r udσ (x) ≤ cr r = cr → 0, r → ∞.
r

Hence ∇u = 0 in Ω  and u is constant in (each component of) Ω  . 

Remark 41.5. If Ω and Ω  are both simply connected, then the solution of (EN) for
n > 2 is unique. This is a consequence of Theorem 41.4 for simply connected Ω  .
We now turn to the problem of finding the solutions (existence problems). Let us
try to solve (ID) by setting

u(x) := f (y)∂νy K(x − y)dσ (y), (41.1)

S

where K is the (known) fundamental solution for Δ .

Remark 41.6. Note that (41.1) involves only the known fundamental solution and
not the Green’s function (which is difficult to find in general) as in the Poisson
integral

w(x) = f (y)∂νy G(x, y)dσ (y). (P)

S

We know that u(x) is harmonic in Ω , because K(x − y) is harmonic for x ∈ Ω ,

y ∈ S. It remains to verify the boundary conditions. Clearly u will not have the
correct boundary values, but in a sense it is not far from correct. We shall prove it
(very soon) that on S,
f
u = + T f ,
2

where T is a compact operator on L2 (S). Thus, what we really want is to take

u(x) = ϕ (y)∂νy K(x − y)dσ (y), x∈

/ S, (41.2)
S
454 Part IV: Partial Differential Equations

where ϕ is the solution of

1
ϕ +Tϕ = f.
2
Similarly, we shall try to solve (IN) (and (EN)) in the form

u(x) = ϕ (y)K(x − y)dσ (y), x∈

/ S. (41.3)
S

Definition 41.7. The functions u(x) from (41.2) and (41.3) are called the double
and single layer potentials with moment (density) ϕ , respectively.

Definition 41.8. Let I(x, y) be continuous on S × S, x = y. We call I a continuous

kernel of order α , 0 ≤ α < n − 1, n ≥ 2, if

|I(x, y)| ≤ c|x − y|−α , 0 < α < n − 1,

and
|I(x, y)| ≤ c1 + c2 |log |x − y|| , α = 0,

where c > 0 and c1 , c2 ≥ 0.

Remark 41.9. Note that a continuous kernel of order 0 is also a continuous kernel
of order α , 0 < α < n − 1.

We denote by I the integral operator

If (x) = I(x, y) f (y)dσ (y), x ∈ S,

S

with kernel I.

Lemma 41.10. If I is a continuous kernel of order α , 0 ≤ α < n − 1, then

(1) I is bounded on L p (S), 1 ≤ p ≤ ∞.
(2) I is compact on L2 (S).

Proof. It is enough to consider 0 < α < n − 1. Let us assume that f ∈ L1 (S). Then
 

 
I f  ≤ |I(x, y)|| f (y)|dσ (y)dσ (x)
L1 (S) S S

≤c | f (y)|dσ (y) |x − y|−α dσ (x)

S S

d
≤ c  f L1 (S) rn−2−α dr = c  f L1 (S) ,
0

where d = diam S = supx,y∈S |x − y|.

If f ∈ L∞ (S), then
41 Layer Potentials 455

 
d
 
I f  ≤ c  f L∞ (S) rn−2−α dr = c  f L∞ (S) .
L∞ (S) 0

For 1 < p < ∞ part (1) follows now by interpolation.

For part (2), let ε > 0 and set

I(x, y), |x − y| > ε ,
Iε (x, y) =
0, |x − y| ≤ ε .

Since Iε is bounded on S × S, it follows that Iε is a Hilbert–Schmidt operator in

L2 (S), so that Iε is compact for each ε > 0.
Exercise 41.3. Prove that a Hilbert–Schmidt operator, i.e., an integral operator
whose kernel I(x, y) satisfies

|I(x, y)|2 dxdy < ∞,

S S

is compact in L2 (S).
On the other hand, due to estimates for convolution,


2 1/2
 
 
I f − Iε f  ≤c | f (y)||x − y| −α
dσ (y) dσ (x)
L2 (S) |x−y|<ε

ε
≤ c  f L2 (S) rn−2−α dr → 0, ε → 0.
0

Thus, I as the limit of Iε is also compact in L2 (S). 

Lemma 41.11.
(1) If I is a continuous kernel of order α , 0 ≤ α < n−1, then Itransforms bounded
functions into continuous functions.
(2) If I is as in part (1), then u + Iu
 ∈ C(S) for u ∈ L2 (S) implies u ∈ C(S).

Proof. Let |x − y| < δ . Then

|If (x) − If (y)| ≤ |I(x, z) − I(y, z)|| f (z)|dσ (z)

S
≤ (|I(x, z)| + |I(y, z)|) | f (z)|dσ (z)
|x−z|<2δ

+ |I(x, z) − I(y, z)|| f (z)|dσ (z)

S\{|x−z|<2δ }

 
≤ c  f ∞ |x − z|−α + |y − z|−α dσ (z)
|x−z|<2δ

+ |I(x, z) − I(y, z)|| f (z)|dσ (z) =: I1 + I2 .

S\{|x−z|<2δ }
456 Part IV: Partial Differential Equations

Since |z − y| ≤ |x − z| + |x − y|, we have

3δ
I1 ≤ c  f ∞ rn−2−α dr → 0, δ → 0.
0

On the other hand, for |x − y| < δ and |x − z| ≥ 2δ we have that

|y − z| ≥ |x − z| − |x − y| > 2δ − δ = δ .

So the continuity of I outside of the diagonal implies that

I(x, z) − I(y, z) → 0, x → y,

uniformly in z ∈ S\ {|x − z| < 2δ }. Hence, I1 and I2 will be small if y is sufficiently

close to x. This proves the first claim.
For the second part, let ε > 0 and let ϕ ∈ C(S × S) be such that 0 ≤ ϕ ≤ 1 and

1, |x − y| < ε /2,
ϕ (x, y) =
0, |x − y| ≥ ε .

 =ϕ
Write Iu Iu + (1
− ϕ )Iu =: I0 u + I1 u. By the Cauchy–Bunyakovsky–Schwarz
inequality we have

1/2
|I1 u(x) − I1 u(y)| ≤ u2 |I1 (x, z) − I1 (y, z)|2 dσ (z) → 0, y → x,
S

since I1 is continuous (see the definition of ϕ ). Now if we set

 − I1 u ≡ u + I0 u,
g := u + Iu

then g is continuous for u ∈ L2 (S) by the conditions of this lemma. Since the operator
norm of I0 can be made less that 1 on L2 (S) and L∞ (S) (we can do this due to the
choice of ε > 0 sufficiently small), then
 −1
u = I + I0 g,

where I is the identity operator. Since g is continuous and the operator norm is less
than 1, we have
∞  j
u = ∑ −I0 g.
j=0

This series converges uniformly, and therefore u is continuous. 

Let us consider now the double layer potential (41.2) with moment ϕ ,
41 Layer Potentials 457

u(x) = ϕ (y)∂νy K(x − y)dσ (y), x ∈ Rn \S.

S

First of all,
(x − y, ν (y))
∂νy K(x − y) = − . (41.4)
ωn |x − y|n

Exercise 41.4. Prove that (41.4) holds for all n ≥ 1.

 (41.4) defines a harmonic function in x ∈ R \S, y ∈ S. More-

It is clearalso that n

over, it is O |x|1−n as x → ∞ (y ∈ S), so that u is also harmonic at infinity.

Exercise 41.5. Prove that (41.4) defines a harmonic function at infinity.

Lemma 41.12. There exists c > 0 such that

|(x − y, ν (y))| ≤ c|x − y|2 , x, y ∈ S.

Proof. It is quite trivial to obtain

|(x − y, ν (y))| ≤ |x − y||ν (y)| = |x − y|.

But the latter inequality allows us to assume that |x − y| ≤ 1. Given y ∈ S, by

a translation and rotation of coordinates we may assume that y = 0 and ν (y) =
(0, 0, . . . , 0, 1). Hence (x − y, ν (y)) transforms to xn , and near y, S is the graph of
the equation xn = ψ (x1 , . . . , xn−1 ), where ψ ∈ C2 (Rn−1 ), ψ (0) = 0, and ∇ψ (0) = 0.
Then
|(x − y, ν (y))| = |xn | ≤ c|(x1 , . . . , xn−1 )|2 ≤ c|x|2 = c|x − y|2

by Taylor’s expansion. 

We denote ∂νy K(x − y) by I(x, y).

Lemma 41.13. I is a continuous kernel of order n − 2, n ≥ 2.

Proof. If x, y ∈ S, then I(x, y) is continuous for x = y; see (41.4). Hence

c|x − y|2
|I(x, y)| ≤ = c |x − y|2−n
ωn |x − y|n

by Lemma 41.12. 

Lemma 41.14. ⎧

⎪
⎨1, x ∈ Ω ,
I(x, y)dσ (y) = 0, x ∈ Ω  , (41.5)
S ⎪
⎩1
2 , x ∈ S.
458 Part IV: Partial Differential Equations

Proof. If x ∈ Ω  , then K(x − y) is harmonic in x ∈

/ S, y ∈ S, and it is also harmonic
in y ∈ Ω , x ∈ Ω  . Hence (see Exercise 39.2)

∂νy K(x − y)dσ (y) = 0,

S

or

I(x, y)dσ (y) = 0, x ∈ Ω .

S

If x ∈ Ω , let δ > 0 be such that Bδ (x) ⊂ Ω . Denote Ωδ = Ω \Bδ (x) and Sδ =

S\ (S ∩ Bδ (x)), see Figure 41.1. Then K(x − y) is harmonic in y in Ωδ , and therefore
by Green’s identity,

0= (1 · Δy K(x − y) − K(x − y)Δ 1) dy

Ω \Bδ (x)

= ∂νy K(x − y)dσ (y) − ∂νy K(x − y)dσ (y)

S |x−y|=δ

δ 1−n
= I(x, y)dσ (y) − dσ (y) = I(x, y)dσ (y) − 1,
S ωn |x−y|=δ S

or

I(x, y)dσ (y) = 1.

S

Now suppose x ∈ S. In this case

I(x, y)dσ (y) = lim I(x, y)dσ (y). (41.6)

S δ →0 Sδ

∂ Bδ+

δ S
Bδ+
Ω x Ω
Bδ−

Sδ Ωδ ∂ Bδ−

Fig. 41.1 Geometric illustration of the boundary near x.

If y ∈ Ωδ , then for x ∈ S we have that x = y. This implies that

41 Layer Potentials 459

0= Δy K(x − y)dy = ∂νy K(x − y)dσ (y) − ∂νy K(x − y)dσ (y).
Ωδ Sδ ∂ B−
δ

Therefore, see (41.4),

lim ∂νy K(x − y)dσ (y) = lim ∂νy K(x − y)dσ (y)
δ →0 Sδ δ →0 ∂ B−
δ

δ 1−n
= lim dσ (y)
δ →0 ωn ∂ B−
δ

δ 1−n  n−1 ωn  1
= lim δ + o(δ n−1 ) = .
δ →0 ωn 2 2

This means that the limit in (41.6) exists and (41.5) is satisfied. 
Lemma 41.15. There exists c > 0 such that

|∂νy K(x − y)|dσ (y) ≤ c, x ∈ Rn .

S

Proof. It follows from Lemma 41.13 that

c
|∂νy K(x − y)|dσ (y) ≤ |x − y|2−n dσ (y) ≤ c1 , x ∈ S.
S ωn S

Next, for x ∈
/ S define dist(x, S) = infy∈S |x − y|.
There are two possibilities now: if dist(x, S) ≥ δ /2, then |x − y| ≥ δ /2 for all
y ∈ S, and therefore

|∂νy K(x − y)|dσ (y) ≤ cδ 1−n dσ (y) = c , (41.7)

S S

where c does not depend on δ > 0 (because δ is fixed).

Suppose now that dist(x, S) < δ /2. If we choose δ > 0 sufficiently small, then
there is a unique x0 ∈ S such that

x = x0 + t ν (x0 ), t ∈ (−δ /2, δ /2).

Set Bδ = {y ∈ S : |x0 − y| < δ }. We estimate the integrals of |I(x, y)| over S\Bδ and
Bδ separately. If y ∈ S\Bδ , then

|x − y| ≥ |x0 − y| − |x − x0 | > δ − δ /2 = δ /2

and
|I(x, y)| ≤ cδ 1−n ,

so that the integral over S\Bδ satisfies (41.7), where again c does not depend on δ .
460 Part IV: Partial Differential Equations

To estimate the integral over Bδ , we note that (see (41.4))

|(x − y, ν (y))| |(x − x0 , ν (y)) + (x0 − y, ν (y))|

|I(x, y)| = =
ωn |x − y|n ωn |x − y|n
|x − x0 | + c|x0 − y|2
≤ . (41.8)
ωn |x − y|n

The latter inequality follows from Lemma 41.12, since x0 , y ∈ S. Moreover, we have
(due to Lemma 41.12)

|x − y|2 = |x − x0 |2 + |x0 − y|2 + 2(x − x0 , x0 − y)

 
x − x0
= |x − x0 |2 + |x0 − y|2 + 2|x − x0 | x0 − y,
|x − x0 |
≥ |x − x0 |2 + |x0 − y|2 − 2|x − x0 ||(x0 − y, ν (x0 ))|
≥ |x − x0 |2 + |x0 − y|2 − 2c|x − x0 ||x0 − y|2
≥ |x − x0 |2 + |x0 − y|2 − |x − x0 ||x0 − y|,

if we choose δ > 0 such that |x0 − y| ≤ 2c 1

, where the constant c > 0 is from Lemma
41.12.  
Since |x − x0 ||x0 − y| ≤ 12 |x − x0 |2 + |x0 − y|2 , we obtain finally

1 
|x − y|2 ≥ |x − x0 |2 + |x0 − y|2
2
and (see (41.4) and (41.8))

|x − x0 | + |x0 − y|2
|I(x, y)| ≤ c
(|x − x0 |2 + |x0 − y|2 )n/2
|x − x0 | c
≤c + .
(|x − x0 |2 + |x0 − y|2 )n/2 |x0 − y|n−2

This implies


δ
δ n−2
|x − x0 | r
|I(x, y)|dσ (y) ≤ c rn−2 dr + c dr
Bδ 0 (|x − x0 | + r )
2 2 n/2 0 rn−2

∞
arn−2
≤ c δ + c dr,
0 (a2 + r2 )n/2

where a := |x − x0 |. For the latter integral we have (t = r/a)

∞
∞
arn−2 t n−2
dr = dt < ∞.
0 (a + r2 )n/2
2 0 (1 + t 2 )n/2
41 Layer Potentials 461

If we combine all estimates, then we may conclude that there is c0 > 0 such that

|∂νy K(x − y)|dσ (y) ≤ c0 , x ∈ Rn ,

S

and this constant does not depend on x. 

Theorem 41.16. Suppose ϕ ∈ C(S) and u is defined by the double layer potential
(41.2) with moment ϕ . Then for all x ∈ S,

ϕ (x)
lim u(x + t ν (x)) = + I(x, y)ϕ (y)dσ (y),
t→0− 2 S

ϕ (x)
lim u(x + t ν (x)) = − + I(x, y)ϕ (y)dσ (y)
t→0+ 2 S

uniformly on S with respect to x.

Proof. If x ∈ S and t < 0, with |t| sufficiently small, then xt := x + t ν (x) ∈ Ω and
u(x + t ν (x)) is well defined by

u(x + t ν (x)) = ϕ (y)I(xt , y)dσ (y) = (ϕ (y) − ϕ (x))I(xt , y)dσ (y) + ϕ (x)
S S

→ ϕ (x) + ϕ (y)I(x, y)dσ (y) − ϕ (x) I(x, y)dσ (y)

S S

= ϕ (x) + ϕ (y)I(x, y)dσ (y) − ϕ (x)/2, t → 0−.

S

If t > 0, the arguments are the same except that

I(xt , y)dσ (y) = 0.

S

Uniform convergence follows from the fact that S is compact and ϕ ∈ C(S). 
Corollary 41.17. For x ∈ S,

ϕ (x) = u− (x) − u+ (x),

where u± = limt→0± u(xt ).

We state without proof that the normal derivative of the double layer potential is
continuous across the boundary in the sense of the following theorem.
Theorem 41.18. Suppose ϕ ∈ C(S) and u is defined by the double layer potential
(41.2) with moment ϕ . Then for all x ∈ S,

lim (ν (x) · ∇u(x + t ν (x)) − ν (x) · ∇u(x − t ν (x))) = 0

t→0+

uniformly on S with respect to x.

462 Part IV: Partial Differential Equations

Let us now consider the single layer potential

u(x) = ϕ (y)K(x − y)dσ (y)

S

with moment ϕ ∈ C(S).

Lemma 41.19. The single layer potential u is continuous on Rn .

Proof. Since u is harmonic in x ∈/ S, we have only to show continuity for x ∈ S.

Given x0 ∈ S and δ > 0, let Bδ = {y ∈ S : |x0 − y| < δ }. Then

|u(x) − u(x0 )| ≤ (|K(x − y)| + |K(x0 − y)|) |ϕ (y)|dσ (y)

Bδ

+ |K(x − y) − K(x0 − y)||ϕ (y)|dσ (y)

S\Bδ

1
≤ cδ (or δ log for n = 2) + ϕ ∞ |K(x − y) − K(x0 − y)|dσ (y) → 0
δ S\Bδ

as x → x0 and δ → 0. 

Exercise 41.6. Prove that


δ, n > 2,
(|K(x − y)| + |K(x0 − y)|) |ϕ (y)|dσ (y) ≤ c ϕ ∞
Bδ δ log δ1 , n = 2.

Definition 41.20. Let us set

(x − y, ν (x))
I ∗ (x, y) := ∂νx K(x − y) ≡ .
ωn |x − y|n

Theorem 41.21. Suppose ϕ ∈ C(S) and u is defined on Rn by the single layer po-
tential (41.3) with moment ϕ . Then for x ∈ S,

ϕ (x)
lim ∂ν u(x + t ν (x)) = − + I ∗ (x, y)ϕ (y)dσ (y),
t→0− 2 S

ϕ (x)
lim ∂ν u(x + t ν (x)) = + I ∗ (x, y)ϕ (y)dσ (y).
t→0+ 2 S

Proof. We consider the double layer potential on Rn \S with moment ϕ ,

v(x) = ϕ (y)∂νy K(x − y)dσ (y),

S

and define the function f on a tubular neighborhood V of S by

41 Layer Potentials 463

v(x) + ∂ν u(x), x ∈ V \S,
f (x) = (41.9)
 ∗
I ϕ (x) + I ϕ (x), x ∈ S,

where u is defined by (41.3).

Here the tubular neighborhood of S is defined as

V = {x + t ν (x) : x ∈ S, |t| < δ } .

We claim that f is continuous on V . It is clearly (see (41.9)) continuous on V \S and

S, so it suffices to show that if x0 ∈ S and x = x0 + t ν (x0 ), then f (x) − f (x0 ) → 0 as
t → 0±. We have

f (x) − f (x0 ) = v(x) + ∂ν u(x) − Iϕ (x0 ) − I∗ ϕ (x0 )

= I(x, y)ϕ (y)dσ (y) + ϕ (y)∂νx K(x − y)dσ (y)

S
S
− I(x0 , y)ϕ (y)dσ (y) − I ∗ (x0 , y)ϕ (y)dσ (y)
S S

= (I(x, y) + I ∗ (x, y) − I(x0 , y) − I ∗ (x0 , y)) ϕ (y)dσ (y).

S

We write this expression as an integral over Bδ = {y ∈ S : |x0 − y| < δ } plus an

integral over S\Bδ . The integral over S\Bδ tends uniformly to 0 as x → x0 , because
|y − x| ≥ δ and |y − x0 | ≥ δ , so that the functions I and I ∗ have no singularities in
this case.
On the other hand, the integral over Bδ can be bounded by

ϕ ∞ (|I(x, y) + I ∗ (x, y)| + |I(x0 , y) + I ∗ (x0 , y)|) dσ (y).

Bδ

Since
(x − y, ν (y))
I(x, y) = −
ωn |x − y|n

and ν (x) = ν (x0 ) for x = x0 + t ν (x0 ) ∈ V , we have

(x − y, ν (x)) (x − y, ν (x0 ))
I ∗ (x, y) = I(y, x) = ≡ . (41.10)
ωn |x − y|n ωn |x − y|n

Hence

(x − y, ν (x0 ) − ν (y)) |x − y||ν (x0 ) − ν (y)|
|I(x, y) + I ∗ (x, y)| = ≤

ωn |x − y|n ωn |x − y|n
|x − y||x0 − y| |x0 − y|
≤c ≤ c = c |x0 − y|2−n ,
ωn |x − y|n |x0 − y|n−1
464 Part IV: Partial Differential Equations

because |x0 − y| ≤ |x0 − x| + |x − y| ≤ 2|x − y|. Here we have also used the fact that
|ν (x0 ) − ν (y)| ≤ c|x0 − y|, since ν is C1 (Figure 41.2).

x0

Fig. 41.2 Geometric illustration of the boundary at the point x0 .

This estimate allows us to obtain that the corresponding integral over Bδ can be
dominated by


δ
c |x0 − y|2−n dσ (y) = c r2−n rn−2 dr = c δ .
|y−x0 |≤δ 0

Thus f := v + ∂ν u extends continuously across S. Therefore, for x ∈ S,

1
Iϕ (x) + I∗ ϕ (x) = v− (x) + ∂ν− u(x) = ϕ (x) + Iϕ (x) + ∂ν− u(x).
2
It follows that
ϕ (x) ∗
∂ν− u(x) = − + I ϕ (x).
2
By similar arguments we obtain

1
Iϕ (x) + I∗ ϕ (x) = v+ (x) + ∂ν+ u(x) = − ϕ (x) + Iϕ (x) + ∂ν+ u(x)
2
and therefore
ϕ (x) ∗
∂ν+ u(x) = + I ϕ (x).
2
This completes the proof. 

Corollary 41.22.
ϕ (x) = ∂ν+ u(x) − ∂ν− u(x),

where u is defined by (41.3).

41 Layer Potentials 465

Lemma 41.23. If f ∈ C(S) and

ϕ ∗ ϕ ∗
+I ϕ = f, − +I ϕ = f,
2 2
then

ϕ dσ = f dσ , f dσ = 0,
S S S

respectively.
Proof. It follows from (41.10) that

1
f (x)dσ (x) = ϕ (x)dσ (x) + ϕ (y)dσ (y) I ∗ (x, y)dσ (x)
S 2 S S S

1 1
= ϕ (x)dσ (x) + ϕ (y)dσ (y) = ϕ (y)dσ (y),
2 S 2 S S

where we have also used Lemma 41.14. 

Lemma 41.24. Let n = 2.
(1) If ϕ ∈ C(S), then the single layer potential u with moment ϕ is harmonic at
infinity if and only if

ϕ (x)dσ (x) = 0,
S

and in that case u → 0 as |x| → ∞.

(2) Let ϕ ∈ C(S) with

ϕ (x)dσ (x) = 0
S

and u as in part (1). If u is constant on Ω , then u ≡ 0.

Proof. Since n = 2, we must have

1
u(x) = log |x − y|ϕ (y)dσ (y)
2π S

1 1
= (log |x − y| − log |x|) ϕ (y)dσ (y) + log |x| ϕ (y)dσ (y).
2π S 2π S

But log |x − y| − log |x| → 0 as x → ∞ uniformly for y ∈ S, and therefore, this term
is harmonic at infinity(we have a removable singularity). Hence u is harmonic at
infinity if and only if S ϕ (x)dσ (x) = 0, and in this case u(x) vanishes at infinity.
This proves part (1).
In part (2), u is harmonic at infinity. If u is constant on Ω , then it solves (ED) with
f ≡ constant on S. But a solution of such a problem must be constant and vanish at
infinity. Therefore, this constant is zero. Thus u ≡ 0. 
Remark 41.25. For n > 2 the single layer potential u is a harmonic function at in-
finity without any additional conditions for the moment ϕ .
466 Part IV: Partial Differential Equations

For solvability of the corresponding integral equations in the space C(S) with
integral operators I and I∗ (see Theorems 41.16 and 41.21) we need the Fredholm
alternative (see in addition Theorems 34.8 and 34.9).

Theorem 41.26 (First Fredholm theorem). The null spaces of 12 I − I and 12 I − I∗
have the same finite dimension
   
1  1 ∗
dim N I − I = dim N I − I < ∞,
2 2

where I denotes the identity operator.

 
Proof. Since I and I∗ are compact operators in C(S), the null spaces N 12 I − I and
     
N 12 I − I∗ are closed subspaces of C(S). If ϕ ∈ N 12 I − I and ψ ∈ N 12 I − I∗ ,
then
2Iϕ = ϕ , 2I∗ ψ = ψ ,

i.e., 2I and 2I∗ are identical on the corresponding null spaces. Since they are com-
pact there, this is possible only when the corresponding null spaces are of finite
dimension. The equality of these dimensions can be checked in the same manner as
in the proof of Theorem 34.11. In this proof, part (2) of Lemma 41.11 must be taken
into account. 

Theorem 41.27 (Second Fredholm theorem). The ranges of the operators 12 I − I

and 12 I − I∗ on C(S) can be described as
   
1  1 ∗
R I − I = { f ∈ C(S) : ( f , ψ )L2 (S) = 0 for any ψ ∈ N I −I }
2 2

and
   
1 1
R I − I∗ = {g ∈ C(S) : (g, ϕ )L2 (S) = 0 for any ϕ ∈ N 
I − I }.
2 2

ϕ
Proof. Let f = 2 − Iϕ for some ϕ ∈ C(S). Then for all ψ ∈ N( 12 I − I∗ ) we have

ϕ 
( f , ψ )L2 (S) = ( − I ϕ , ψ )L2 (S)
2
ϕ ψ
= ( , ψ )L2 (S) − (ϕ , I∗ ψ )L2 (S) = (ϕ , − I∗ ψ )L2 (S) = 0.
2 2

Conversely, assume that f ∈ C(S), f = 0 satisfies ( f , ψ )L2 (S) = 0 for all ψ ∈ N( 12 I −

I∗ ). Assume to the contrary that there is no ϕ ∈ C(S) such that f = 12 ϕ − Iϕ . Then
f can be chosen to be orthogonal to all 12 ϕ − Iϕ , i.e.,
41 Layer Potentials 467

1 1 1
0 = ( f , ϕ − Iϕ )L2 (S) = ( f , ϕ )L2 (S) − (I∗ f , ϕ )L2 (S) = ( f − I∗ f , ϕ )L2 (S) .
2 2 2

This means that f ∈ N( 12 I − I∗ ). But at the same time, f ⊥N( 12 I − I∗ ). Thus f = 0.
This contradiction proves the opposite embedding. For the operator 12 I − I∗ the proof

is the same, since (I∗ )∗ = I. 
Since the ranges of 12 I − I and 12 I − I∗ are closed due to Riesz’s lemma (see Theo-
rem 28.14) and due to part (2) of Lemma 41.11, we obtain the following result.
Theorem 41.28 (Fredholm alternative). Either 12 I − I and 12 I − I∗ are bijective or
 ∗
2 I − I and 2 I − I have nontrivial null spaces with finite dimension
1 1

   
1 1
dim N I − I = dim N I − I∗ < ∞,
2 2

and the ranges of these operators are given by

   
1  1 ∗
R I − I = { f ∈ C(S) : ( f , ψ )L2 (S) = 0 for any ψ ∈ N I −I }
2 2

and
   
1 1
R I − I∗ = {g ∈ C(S) : (g, ϕ )L2 (S) = 0 for any ϕ ∈ N I − I }.
2 2

We will interpret this alternative as follows: either the integral equations

1 1
ϕ − Iϕ = f , ψ − I∗ ψ = g (41.11)
2 2

have unique solutions ϕ and ψ for every given f and g from C(S), or the corre-
sponding homogeneous equations

1 1
ϕ − Iϕ = 0, ψ − I∗ ψ = 0
2 2
have the same number of linearly independent solutions ϕ1 , . . . , ϕm , ψ1 , . . . , ψm , and
in this case equations (41.11) have solutions if and only if f ⊥ϕ j , j = 1, 2, . . . , m,
and g⊥ψ j , j = 1, 2, . . . , m, respectively.
In fact, it is possible to prove a stronger result (which is the analogue of Theo-
rem 28.15) for the complete normed space C(S); see [22] for a proof.
Theorem 41.29 (Riesz). Let A : C(S) → C(S) be a compact linear operator. Then
I − A is injective if and only if it is surjective. If I − A is injective (and therefore also
bijective), then the inverse operator (I − A)−1 : C(S) → C(S) is bounded.
468 Part IV: Partial Differential Equations

Exercise 41.7. Show that

   ⊥    ⊥
1  1  1 1
R I −I = N I −I ∗ , R I − I∗ =N 
I −I ,
2 2 2 2

and then
       
1  1 ∗
1 ∗
1 
C(S) = R I −I ⊕N I −I = R I −I ⊕N I −I .
2 2 2 2

Now we are in a position to prove the main result of this chapter.

Theorem 41.30 (Main theorem). Suppose Ω and Ω  are simply connected. Then
(1) (ID) has a unique solution for every f ∈ C(S).
(2) (ED) has a unique solution for every f ∈ C(S). 
(3) (IN) has a solution for every f ∈ C(S) if and only if S f dσ = 0. The solution
is unique up to a constant. 
(4) (EN) has a unique solution for every f ∈ C(S) if and only if S f dσ = 0.

Proof. We have already proved uniqueness (see Theorem 41.4) and the necessity of
the conditions on f (see Exercise 39.2 and Lemma 41.24). So all that remains is to
establish existence.
For (IN) and (EN) the function f must satisfy the condition

f dσ = 0,
S

or
( f , 1)L2 (S) = 0.

Next, since

I(x, y)dσ (y) = 1/2,

S

 i.e., dim N( 1 I − I)
we may conclude first that 1 ∈ N( 12 I − I),  ≥ 1, and dim N( 1 I −
2 2
∗
I ) ≤ 1 due to the fact that the single layer potential uniquely (up to a constant; see
Theorem 41.4) solves (IN). Thus, due to Theorem 41.28 (Fredholm alternative), we
have    
1  1 ∗
dim N I − I = dim N I − I = 1.
2 2

Using again this alternative, we see that the condition ( f , 1)L2 (S) = 0 is necessary and
sufficient for the solvability of the equation − 12 ϕ + I∗ ϕ = f , which solves (IN). For
(EN) we can solve uniquely the equation 12 ϕ + I∗ ϕ = f if and only if f ⊥N( 12 I + I). 

But since the solution of (ED) is unique, the null space of I + I consists only of the
1
2
41 Layer Potentials 469

trivial solution (we have used here Lemma 41.24 for n = 2). Therefore, the necessary
and sufficient condition for f is automatically satisfied.
Concerning (ID) we consider the integral equation 12 ϕ + Iϕ = f . Let us set
 
1 
ψ := f − ϕ + Iϕ ,
2

so that f = ψ + ( 12 ϕ + Iϕ ). Since f ∈ C(S) and 12 ϕ + Iϕ ∈ R( 12 I + I),

 we can choose
ψ uniquely (using Exercise 41.7) from N( 2 I + I∗ ), i.e.,
1

1
ψ + I∗ ψ = 0.
2
If we consider now the single layer potential with moment ψ ,

v(x) := ψ (y)K(x − y)dσ (y),

S

then it is harmonic in Ω ∪ Ω  and

1
∂ν+ v = ψ + I∗ ψ = 0.
2

Hence ∂ν+ v(x) ≡ 0 for all x ∈ Ω  (due to the uniqueness result), which means that
v ≡ constant in Rn \ Ω , and since v is required to be harmonic at infinity, this constant
is equal to zero (see Theorem 41.4 and Lemma 41.24 for n = 2). The final step is that
a single layer potential v is continuous everywhere in the whole of Rn and v+ ≡ 0 on
Ω  including S. Thus v ≡ 0 in Ω as well. The latter fact can be proved if we consider
the Dirichlet boundary value problem

Δ v = 0, in Ω ,
v = 0, on S.

But Theorem 41.21 leads to ∂ν+ v − ∂ν− v = ψ or ψ = 0. Therefore, we have proved

that (ID) is uniquely solvable for all f ∈ C(S). 

Exercise 41.8. Prove the second part of Theorem 41.30.

Chapter 42
Elliptic Boundary Value Problems

In this chapter we study the equation Lu = f on some bounded domain Ω ⊂ Rn with

Ck -class (with appropriate k ≥ 1) boundary ∂ Ω = S, where u is to satisfy certain
boundary conditions on S. The object of interest is to prove existence, uniqueness,
and regularity theorems. Our approach will be to formulate the problems in terms of
sesquilinear forms and then to apply some Hilbert space theory (see Part III of this
book). Here L will denote a differential operator in the divergence form

L(x, ∂ ) = ∑ (−1)|α | ∂ α (aαβ (x)∂ β ), m = 1, 2, . . . , (42.1)

|α |=|β |≤m

satisfying the coercivity (generalized strong ellipticity) condition with u ∈ C0∞ (Ω ),

 
Re ∑ aαβ (x)∂ β u · ∂ α udx ≥ ν ∑ |∂ α u|2 dx, (42.2)
|α |=|β |=m Ω |α |=m Ω

where ν > 0 is constant. It is easy to see that for the operator (42.1) with constant
coefficients, the condition (42.2) is equivalent to

Re ∑ aαβ ξ α ξ β ≥ ν ∑ |ξ α |2 .
|α |=|β |=m |α |=m

The coefficients of L are assumed to be complex-valued (in general). We introduce

the Dirichlet form D(u, v) associated with this operator L as

D(u, v) = ∑ aαβ (x)∂ β u · ∂ α vdx. (42.3)
|α |=|β |≤m Ω

c Springer International Publishing AG 2017 471

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 42
472 Part IV: Partial Differential Equations

The Dirichlet boundary value problem for the operator L from (42.1) on the domain
Ω can be formulated as follows: given f ∈ L2 (Ω ), find a function u satisfying Lu = f
on Ω in the distributional sense, i.e., u is a distributional solution satisfying the
boundary conditions

u|S = g0 , ∂ν u|S = g1 , . . . , ∂νm−1 uS = gm−1 , (42.4)

where S is a Cm -class surface, ν is an outward normal vector to S, and g j , for j =

0, 1, . . . , m − 1, are from some appropriate space on S.

Remark 42.1 (This example is due to A. Bitsadze). The operator

1
L := (∂x2 + 2i∂x ∂y − ∂y2 )
4

is elliptic on R2 . The general solution of Lu = 0 is

u(x, y) = f (z) + zg(z), z = x + iy,

where f and g are arbitrary holomorphic functions; see [11]. In particular, if we

choose
g(z) = −z f (z),

where f is holomorphic on the unit disk B and continuous on B, then u(x, y) =

(1 − |z|2 ) f (z), and u(x, y) solves the Dirichlet boundary value problem

Lu = 0, in B,
u = 0, on ∂ B.

Hence we have no uniqueness for this problem in B. The reason is that L is elliptic
but not strongly elliptic.

Exercise 42.1. Show that L from Remark 42.1 is elliptic on R2 but not strongly
elliptic in the sense of (42.2).

In order to investigate the solvability of the Dirichlet boundary value problem we

need a regularity theorem for the operator L from (42.1). Here we define

us := uH s (Ω ) .

Theorem 42.2. Suppose that the operator L from (42.1) satisfies the ellipticity con-
dition (42.2) and the coefficients aαβ (x), |α |, |β | ≤ m belong to the Sobolev space
W∞k (Ω ), k = 0, 1, 2, . . .. Then there is a constant C > 0 such that for all u ∈ H0m+k (Ω )
we have
42 Elliptic Boundary Value Problems 473

um+k ≤ C(Luk−m + um+k−1 ). (42.5)

Proof. Since C0∞ (Ω ) is dense in the Sobolev space H0s (Ω ), s ≥ 0, let us consider
u ∈ C0∞ (Ω ). Consider first the case k = 0. Then integration by parts leads to

|(Lu, u)L2 (Ω ) |
 
   
 
= ∑ aαβ (x)∂ β u∂ α udx + ∑ aαβ (x)∂ β u∂ α udx
|α |=|β |=m Ω |α |=|β |≤m−1 Ω

 
≥ Re ∑ aαβ (x)∂ β u∂ α udx −C ∑ |∂ β u∂ α u|dx
|α |=|β |=m Ω |α |=|β |≤m−1 Ω

≥ν ∑ |∂ α u|2 dx −C u2m−1 ,
|α |=m Ω

where the constant C > 0 limits the norms of the coefficients aαβ (x) in L∞ (Ω ).
Using now the Poincaré inequality (Theorem 20.23) and the properties of Sobolev
space, we obtain from the latter inequality that

Lu−m um ≥ ν1 u2m −C um um−1 ,

that is, we have

Lu−m ≥ ν1 um −C um−1 .

Thus (42.5) is proved for k = 0. Consider now the case k = 1. Since the coefficients
 where L
of L belong to W∞1 (Ω ), in that case we have L(∂ j u) = ∂ j Lu + Lu,  is again
an elliptic operator in divergence form with coefficients from L∞ (Ω ). Hence,
       
∂ j u ≤ C(∂ j Lu +    + ∂ j um−1 )
−m  
m
Lu
−m
≤ C(Lu1−m + um + um ).

Here we have used the fact that Lu−m ≤ C um by duality. Using the Poincaré in-
equality (Theorem 20.23), we obtain (42.5) for k = 1. The general case for k follows
by induction. Thus, the theorem is proved. 

Corollary 42.3. Suppose that s ∈ R and s ≥ m. If the coefficients of the operator L

from (42.1)–(42.2) belong to the Sobolev space W∞s−m (Ω ), then there is a constant
C > 0 such that for all u ∈ H0s (Ω ) we have

us ≤ C(Lus−2m + us−1 ). (42.6)

Proof. The result follows by interpolation of Sobolev spaces H0s (Ω ); see [39]. 
474 Part IV: Partial Differential Equations

Theorem 42.4 (Regularity in Sobolev spaces). Suppose that the coefficients of the
operator L from (42.1)–(42.2) belong to C∞ (Ω ). Let u and f be distributions on Ω
s (Ω ) for s ≥ 0, then u ∈ H s+2m (Ω ).
satisfying Lu = f . If f ∈ Hloc loc

Proof. Let ϕ ∈ C0∞ (Ω ). Then the equation ϕ Lu = ϕ f can be rewritten as

L(ϕ u) = ϕ f + L1 u, (42.7)

or
u) = f+ L1 u,
L(

where the operator L1 is of order 2m − 1 with C0∞ (Ω ) coefficients and f ∈ H0s (Ω ).

Our task is to show that u from (42.7) belongs to H0s+2m (Ω ). We use first induction
on s ≥ 0 and 2m ≥ 2. Let us assume that s = 0 and 2m = 2. Then f ∈ L2 (Ω ) with
compact support and u ∈ L2 (Ω ) with compact support (we may assume this without
loss of generality). Applying now Theorem 42.2 (see (42.5)) and using (42.7), we
obtain
 
 

u1 ≤ C(L u0 ) ≤ C( f + L1 u−1 + 
u−1 +  u0 )
−1
 
 
≤ C( f +  u0 + u0 ) < ∞.
0

That is, u ∈ H01 (Ω ). Applying again Theorem 42.2 with this u ∈ H01 (Ω ), we obtain
that
 
 

u2 ≤ C(L u1 ) ≤ C( f + 
u0 + L1 u0 +  u1 + u1 ).
0

Thus, u ∈ H02 (Ω ), and the starting point of induction is checked. Let us assume
now that for every integer s ≥ 1 and 2m ≥ 2 it is true that f ∈ Hloc s (Ω ) implies

u ∈ Hloc (Ω ). Consider now f ∈ Hloc (Ω ). Then f ∈ Hloc (Ω ) as well, and we

s+2m s+1 s

may apply the induction hypothesis, that is, the solution u of Lu = f belongs to
s+2m
Hloc (Ω ). But then we have that (see (42.5) with k = m + s + 1)
 
 

us+2m+1 ≤ C(L u2m+s ) ≤ C( f + L1 us+1 + 
us+1 +  u2m+s )
s+1
 
 
≤ C( f +  us+2m + u2m+s ) < ∞,
s+1

since L1 is of order 2m − 1 and u ∈ Hloc

s+2m
(Ω ). The latter inequality means that this
theorem has been proved for integer s ≥ 0. For arbitrary s ≥ 0 the result follows by
interpolation of Sobolev spaces (see [39]). 
Corollary 42.5. Suppose that the coefficients of the operator L from (42.1)–(42.2)
belong to W∞m (Ω ). Then the distributional solution u of Lu = f with f ∈ L2 (Ω )
2m (Ω ).
belongs to Hloc
42 Elliptic Boundary Value Problems 475

Proof. The result follows from the proof of Theorems 42.2 and 42.4. 
Theorem 42.6 (Gårding’s inequality). Suppose L from (42.1)–(42.2) has L∞ (Ω )
coefficients. Then for all u ∈ H0m (Ω ) we have

Re D(u, u) ≥ c1 u2m − c2 u20 (42.8)

with some positive constants c1 and c2 .

Proof. The proof is much the same as the proofs of previous theorems. Indeed, as
before we can easily obtain that

Re D(u, u) ≥ ν1 u2m −C u2m−1

with positive constants ν1 and C. Since for all u ∈ H0m (Ω ) we have

u2m−1 ≤ ε u2m +Cε u20

with arbitrary ε > 0, we obtain that

Re D(u, u) ≥ (ν1 −Cε ) u2m −CCε u20 .

This completes the proof. 

Exercise 42.2. Show that Theorem 42.6 does not hold for L = Δ 2 on Ω ⊂ Rn for
any u ∈ H 2 (Ω ).
Exercise 42.3. Prove that the range of the operator (−Δ )m + μ I, μ > 0, consid-
ered on H0m (B), B the unit ball in Rn , is complete in H l (B) for all l = 0, 1, 2, . . ..
Hint: Prove the solvability of the equation ((−Δ )m + μ I)u = (|x|2 − 1)P(x) for a
polynomial P(x).
Returning now to the Dirichlet boundary value problem (42.4) for the operator
(42.1)–(42.2), we must search for u in a space of functions for which these bound-
ary conditions make sense. Based on Corollary 42.5, the natural candidate for this
is H 2m (Ω ). The trace formula for Sobolev spaces (see Part II of this book) says that
g j from (42.4) must be from the spaces H 2m− j−1/2 (∂ Ω ), j = 0, 1, . . . , m − 1, respec-
tively. For these given functions g j from (42.4) we may find a function g ∈ H 2m (Ω )
such that 

∂νj g = g j , j = 0, 1, . . . , m − 1. (42.9)
∂Ω

Then, setting w := g − u, we reduce our original problem (42.4) for L to solving


Lw = Lg − f in Ω ,
(42.10)
∂νj w = 0 on S, j = 0, 1, . . . , m − 1.
476 Part IV: Partial Differential Equations

Since (42.9) and (42.10) imply that all derivatives of w of order strictly less than
m vanish on S, we reformulate the Dirichlet boundary value problem (42.10) as
follows: given f ∈ L2 (Ω ), find w ∈ H0m (Ω ) such that

D(w, v) = (Lg − f , v)L2 (Ω ) (42.11)

for all v ∈ H0m (Ω ).

In order to solve the problem (42.11) we consider the properties of sesquilinear
forms. Let H be a Hilbert space.
Definition 42.7. A complex-valued function a(u, v) on H × H is said to be a
sesquilinear form if
(1) a(λ1 u1 + λ2 u2 , v) = λ1 a(u1 , v) + λ2 a(u2 , v),
(2) a(u, μ1 v1 + μ2 v2 ) = μ1 a(u, v1 ) + μ2 a(u, v2 ),
for all u, v, u1 , u2 , v1 , v2 ∈ H and for all λ1 , λ2 , μ1 , μ2 ∈ C.

Definition 42.8. A functional F on H is called a conjugate linear functional if

F(μ1 v1 + μ2 v2 ) = μ1 F(v1 ) + μ2 F(v2 )

for all v1 , v2 ∈ H and for all μ1 , μ2 ∈ C.

Theorem 42.9 (Lax–Milgram). Let a(u, v) be a sesquilinear form on H × H such

that
(1) |a(u, v)| ≤ M u v , u, v ∈ H,
(2) |a(u, u)| ≥ β u2 , u ∈ H,
where M and β are some positive constants. Then for every conjugate linear func-
tional F on H there is a unique u ∈ H such that a(u, v) = F(v). Moreover,

u ≤ c0 FH→C ,

and the constant c0 is independent of F.

Proof. For each fixed u ∈ H the mapping v → a(u, v) is a bounded conjugate linear
functional on H. Hence, the Riesz–Fréchet theorem gives that there is a unique
element w ∈ H such that a(u, v) = (v, w) or a(u, v) = (w, v). Thus we can define an
operator A : H → H mapping u to w as

a(u, v) = (Au, v)

if and only if w = Au. It is clear that A is a bounded linear operator. The linearity
follows from

a(λ1 u1 + λ2 u2 , v) = λ1 a(u1 , v) + λ2 a(u2 , v) = λ1 (w1 , v) + λ2 (w2 , v)

42 Elliptic Boundary Value Problems 477

if and only if A(λ1 u1 + λ2 u2 ) = λ1 Au1 + λ2 Au2 . The boundedness follows from

Au2 = (Au, Au) = a(u, Au) ≤ M u Au ,

which implies Au ≤ M u. Let us note that a(u, Au) is real and positive here.
Next we show that A is one-to-one and that the range of A is equal to H. Indeed,

β u2 ≤ |a(u, u)| = |(Au, u)| ≤ Au u

implies that β u ≤ Au ≤ M u. The first inequality implies that A is one-to-one
and R(A) = R(A). Now we will show that in fact, R(A) = H. Let w ∈ R(A)⊥ . Then

β w2 ≤ |a(w, w)| = |(Aw, w)| = 0,

and therefore w = 0, i.e., R(A) = H = R(A). Next, again due to the Riesz–Fréchet
 ∈ H such that
theorem for F we have that there is a unique w

 v),
F(v) = (w, v ∈ H,

and w  if
 = FH→C . But since R(A) = H, we may find u ∈ H such that Au = w
and only if
 v) = F(v),
a(u, v) = (Au, v) = (w,

which proves the solvability in this theorem. Furthermore,

1 1 1
u ≤  = F .
Au = w
β β β

Finally, we need to show that this element u is unique. If there are two elements u1
and u2 such that
a(u1 , v) = F(v), a(u2 , v) = F(v),

then a(u1 − u2 , v) = 0 for all v ∈ H, and therefore

β u1 − u2 2 ≤ |a(u1 − u2 , u1 − u2 )| = 0.

This completes the proof. 

Remark 42.10. The Lax–Milgram theorem actually says that the operator A that
was constructed there has a bounded inverse such that
 −1  1
A  ≤ ,
H→H β

where β is the same as in condition (2).

478 Part IV: Partial Differential Equations

Assume now that the sesquilinear form a(u, v) satisfies all conditions of the Lax–
Milgram theorem and that the sesquilinear form b(u, v) is only bounded, i.e., there
is a constant M > 0 such that

|b(u, v)| ≤ M u v , u, v ∈ H.

So, we may associate with a(u, v) and b(u, v) two operators A and B, respectively,
such that A has bounded inverse and B is just bounded:

a(u, v) + b(u, v) = (Au, v) + (Bu, v).

In that case the problem of solving

a(u, v) + b(u, v) = F(v), u, v ∈ H,

with a conjugate linear functional F can be reduced to solving

Au + Bu = w

with w from F(v) = (w, v)H .

Theorem 42.11. Let A : H → H be a bijective bounded linear operator with
bounded inverse A−1 and let B : H → H be a compact linear operator. Then A + B is
injective if and only if it is surjective, and in this case (A+B)−1 : H → H is bounded.

Proof. Write
A + B = A(I − (−A−1 )B) =: A(I − K),

where K is compact, since B is compact. Applying Riesz’s theorem (see Theorem

28.15), we obtain that (A + B)−1 = (I − K)−1 A−1 is bounded. 

Corollary 42.12. Let A be as in Theorem 42.11 and let B be bounded (not neces-
sarily compact) with small norm, i.e., B < ε with ε sufficiently small. Then A + B
is bijective and (A + B)−1 : H → H is bounded.

Proof. The operator A + B is bijective. Indeed, let (A + B)u = 0, or u = −A−1 Bu.

Then    
u ≤ A−1  B u < ε A−1  u .

Thus u = 0 is the only possibility, and A + B is injective. The operator A is surjective

with R(A) = H. The same is true for A + B, since

A + B = A(I − (−A−1 )B) = A(I − K)

with K < 1. Thus A + B is injective with bounded inverse, since

42 Elliptic Boundary Value Problems 479

(A + B)−1 = (I − K)−1 A−1 .

The corollary is proved. 

Theorem 42.13. Let D(u, v) be the Dirichlet form (42.3) corresponding to the op-
erator L from (42.1)–(42.2) with L∞ (Ω ) coefficients. Then the following representa-
tion holds:

D(u, v) = (Au, v)H m (Ω ) + (Bu, v)H m (Ω ) , u, v, ∈ H0m (Ω ), (42.12)

where A : H0m (Ω ) → H0m (Ω ) is a linear bounded operator with bounded inverse and
B : H0m (Ω ) → H0m (Ω ) is compact.
Proof. The definition (42.3) of D(u, v) allows us to write

D(u, v) = Dm (u, v) + Dm−1 (u, v),

where Dm (u, v) is a sesquilinear form that satisfies (see (42.2)) the conditions of the
Lax–Milgram theorem (Theorem 42.9) and Dm−1 (u, v) has the form

Dm−1 (u, v) = ∑ aαβ (x)∂ β u∂ α vdx. (42.13)
|α |=|β |≤m−1 Ω

Applying the Lax–Milgram theorem, we obtain that there is a bounded linear oper-
ator A : H0m (Ω ) → H0m (Ω ) with bounded inverse such that

Dm (u, v) = (Au, v)H m (Ω ) . (42.14)

Concerning the sesquilinear form Dm−1 (u, v), we may say that (see (42.13)) since

|Dm−1 (u, v)| ≤ C ∑ |∂ β u| · |∂ α v|dx ≤ C uH m−1 (Ω ) vH m−1 (Ω )
|α |=|β |≤m−1 Ω

≤ C uH m (Ω ) vH m (Ω ) ,

there is a bounded linear operator B : H0m (Ω ) → H0m (Ω ) such that

Dm−1 (u, v) = (Bu, v)H m (Ω ) . (42.15)

We claim that B is compact in H0m (Ω ). To see this, we first note that

Bu2H m (Ω ) = (Bu, Bu)H m (Ω ) = Dm−1 (u, Bu)


= ∑ aαβ (x)∂ β u∂ α (Bu)dx
|α |=|β |≤m−1 Ω

≤ C uH m−1 (Ω ) BuH m−1 (Ω ) ≤ C uH m−1 (Ω ) BuH m (Ω ) ,

480 Part IV: Partial Differential Equations

or
BuH m (Ω ) ≤ C uH m−1 (Ω ) . (42.16)
 
Now let u j ∈ H0m (Ω ) be such that u j H m (Ω ) is bounded. Since H0m (Ω ) is com-
pactly embedded in H0m−1 (Ω ), we have that there is a subsequence u jk that con-
verges strongly in H0m−1 (Ω ), i.e., u jk is a Cauchy sequence in H0m−1 (Ω ). Hence
Bu jk is a Cauchy sequence in H0m (Ω ); see (42.16). This means that Bu jk converges
strongly in H0m (Ω ), and thus B is compact in H0m (Ω ). The theorem now follows
from (42.14) and (42.15). 
Let us now return to the Dirichlet boundary value problem (42.10)–(42.11) for
the elliptic differential operator L from (42.1)–(42.2). Theorem 42.13 allows us to
rewrite (42.11) as

((A + B)w, v)H m (Ω ) = (Lg − f , v)L2 (Ω ) , (42.17)

and the task is to find w ∈ H0m (Ω ) such that (42.17) holds for all v ∈ H0m (Ω ). Since

F : H0m (Ω )
v → (Lg − f , v)L2 (Ω )

is a conjugate linear functional, The Riesz–Fréchet theorem says that there is a

unique f0 ∈ H0m (Ω ) such that

(Lg − f , v)L2 (Ω ) = ( f0 , v)H m (Ω ) .

Due to this fact, (42.17) can be rewritten in operator form as

(A + B)w = f0 . (42.18)

Theorem 42.14 (Unique solvability). The Dirichlet boundary value problem

(42.4) for the operator L from (42.1)–(42.2) has a unique solution if and only if
λ = 0 is not a point of the spectrum of L with homogeneous Dirichlet boundary
conditions. Moreover, this unique solution u can be obtained as

u = g + (A + B)−1 f0 , (42.19)

where g satisfies (42.9) and f0 satisfies (42.18).

Proof. Due to Theorem 42.11, it is enough to show that the operator A + B is in-
jective on H0m (Ω ), i.e., (A + B)u = 0 implies u = 0. This is equivalent to the fact
that
((A + B)u, v) = 0, for all v ∈ H0m (Ω ),

implies u = 0, or D(u, v) = 0 for all v ∈ H0m (Ω ) implies u = 0. But since D(u, v) =

(Lu, v)L2 (Ω ) , the statement of injectivity can be reformulated as
42 Elliptic Boundary Value Problems 481

(Lu, v)L2 (Ω ) = 0, for all v ∈ H0m (Ω ),

or 
Lu = 0,
u ∈ H0m (Ω ) = 0,

or 
Lu = 0, in Ω ,
u = 0, ∂ν u = 0, . . . , ∂ν u = 0, on ∂ Ω .
m−1

Formula (42.19) follows now from (42.18) and (42.9). The theorem is proved. 
Gårding’s inequality (see (42.8)) allows us to get essential information about the
kernel of the operator L and its adjoint. Let us define

W = {u ∈ H0m (Ω ) : (Lu, v)L2 = 0 if and only if D(u, v) = 0 for all v ∈ H0m (Ω )}

and

V = {u ∈ H0m (Ω ) : (L∗ u, v)L2 = 0 if and only if D(v, u) = 0 for all v ∈ H0m (Ω )}.

Theorem 42.15. Suppose L from (42.1)–(42.2) has L∞ (Ω ) coefficients. Then

dimW = dimV < ∞.

Moreover, if f ∈ L2 (Ω ), then there exists u ∈ H0m (Ω ) such that D(u, v) = ( f , v)L2

for all v ∈ H0m (Ω ) if and only if f ⊥V .
Proof. Gårding’s inequality (42.8) implies that

|((L + c2 I)u, u)L2 | ≥ Re[D(u, u) + c2 (u, u)] ≥ c1 u2m , c1 > 0,

for all u ∈ H0m (Ω ). Hence

(L + c2 I)uL2 (Ω ) ≥ c1 um .

The latter inequality means that the operator L + c2 I is invertible and its inverse
(L + c2 I)−1 is compact as an operator in L2 (Ω ), since the embedding H0m (Ω ) →
L2 (Ω ) is compact. Moreover, the range R((L + c2 I)−1 ) is in H0m (Ω ). Now we can
see that u ∈ W if and only if

((L + c2 I)u, v)L2 (Ω ) = c2 (u, v)L2 (Ω ) ,

or
482 Part IV: Partial Differential Equations

(L + c2 I)u = c2 u,

or  
1
I − (L + c2 I)−1 u = 0,
c2

that is, u belongs to the kernel of the operator c12 I − (L + c2 I)−1 . We can say
also that u ∈ V if and only if u belongs to the kernel of the adjoint operator
−1 ∗
c2 I −((L +c2 I) ) . Therefore, the Fredholm alternative (see Theorem 34.11) gives
1

us the statement of this theorem. 

Corollary 42.16. Under the conditions of Theorem 42.15 we have that if in partic-
ular W = V = {0}, then the solution u of the problem D(u, v) = ( f , v)L2 (Ω ) always
exists and is unique for all f ∈ L2 (Ω ).
Proof. The result follows from the Fredholm alternative. 
Corollary 42.17. Suppose that all conditions of Theorem 42.15 are satisfied. As-
sume in addition that L = L∗ , i.e., L is formally self-adjoint (this holds if and only
if aαβ = aβ α for all α , β ). Then there is an orthonormal basis {uk } for L2 (Ω ) con-
sisting of eigenfunctions that satisfy the Dirichlet boundary conditions ∂νj uk = 0 on
S = ∂ Ω for 0 ≤ j ≤ m − 1. The corresponding eigenvalues are real and of finite
multiplicity, and they accumulate only at infinity.
Proof. The proof follows immediately from the Hilbert–Schmidt theorem (Corol-
lary 28.11). 
Example 42.18. Let us consider the Dirichlet boundary value problem in a bounded
domain Ω ⊂ Rn , n ≥ 2 for the Schrödinger operator −Δ + q with complex-valued
potential q ∈ L p (Ω ), n/2 < p ≤ ∞: given f ∈ H 1/2 (∂ Ω ), find u ∈ H 1 (Ω ) such that

(−Δ + q − λ )u = 0, in Ω ,
(42.20)
u = f, on ∂ Ω ,

where λ is real. Then Lemma 32.1 and Theorem 42.14 imply that (42.20) has a
solution if and only if λ = 0 is not a point of the spectrum of this Schrödinger
operator with homogeneous Dirichlet boundary conditions.
Example 42.19. Consider the Dirichlet boundary value problem in a bounded do-
main Ω ⊂ Rn , n ≥ 2, for the Helmholtz operator Δ + k2 n(x) with a complex-valued
function n(x) from L p (Ω ), n/2 < p ≤ ∞: given f ∈ H 1/2 (∂ Ω ), find u ∈ H 1 (Ω ) such
that 
(Δ + k2 n(x))u = 0, in Ω ,
(42.21)
u = f, on ∂ Ω ,

where k is a real or complex number. The values k2 for which there exists a nonzero
function u ∈ H01 (Ω ) satisfying (in the distributional sense)
42 Elliptic Boundary Value Problems 483

Δ u + k2 n(x)u = 0, in Ω ,

are called the Dirichlet eigenvalues of the Helmholtz operator, and the correspond-
ing nonzero solutions are called the eigenfunctions for it. It is clear that k2 = 0 is
not an eigenvalue of this Helmholtz operator. The application of Lemma 32.1 and
Theorem 42.11 lead to the solvability of (42.21). Namely, (42.21) has a unique so-
lution if and only if k2 is not a point of the spectrum (i.e., is not an eigenvalue) of
this Helmholtz operator. In the case of a real-valued function n(x), we may prove
even more.

Theorem 42.20. Assume that n(x) ∈ L p (Ω ), n/2 < p ≤ ∞, is real-valued. Then

there exists an orthonormal basis {uk }∞ k=1 for H0 (Ω ) consisting of eigenfunctions
1

of the Helmholtz operator −Δ − λ n(x). The corresponding eigenvalues {λk }∞ k=1 are
all real and accumulate only at infinity (|λk | → ∞). If in addition n(x) ≥ 0 (n(x) ≡ 0),
then λk > 0 for all k = 1, 2, . . ..

Proof. We may rewrite the eigenvalue problem for (42.21) as (see Lemma 32.1 and
Theorem 42.11)
(A − λ B)u = 0, u ∈ H01 (Ω ), λ = 0, (42.22)

where A is a bounded, self-adjoint, strictly positive linear operator in H01 (Ω ) with

bounded inverse, and B is a compact self-adjoint operator in H01 (Ω ). Next, (42.22)
can be rewritten as
 
1
I − A−1/2 BA−1/2 u = 0, u ∈ H01 (Ω ). (42.23)
λ

This is an eigenvalue problem for the self-adjoint compact operator A−1/2 BA−1/2 .
Using the Riesz–Schauder and Hilbert-Schmidt theorems (see Part III of this book),
we may conclude that there exists a sequence { λ1 }∞ of eigenvalues of A−1/2 BA−1/2
k k=1
such that      
1 1  
  ≥   ≥ ··· ≥  1  ≥ ··· → 0
 λ1   λ2  λ 
k

as k → ∞ with corresponding eigenfunctions {uk }∞

k=1 that form an orthonormal basis
in H01 (Ω ). Hence, the theorem is proved. 

Exercise 42.4. Prove that if {ϕ j }∞j=1 is an orthonormal basis in H0k (Ω ), k = 1, 2, . . .,

then {ϕ j }∞j=1 is a basis (orthogonal) in H0l (Ω ) for every integer 0 ≤ l < k. Show that
the converse is not true.
Chapter 43
The Direct Scattering Problem for the
Helmholtz Equation

In this chapter we will show that the scattering problem for an imperfect conductor
in Rn , n ≥ 2, is well posed. More precisely, we consider a bounded domain Ω ⊂ Rn
(the conductor) containing the origin with connected complement such that ∂ Ω is
in the class C2 . Our aim is to show the existence of a unique solution u ∈ C2 (Rn \
Ω ) ∩C(Rn \ Ω ) of the exterior impedance boundary value problem

Δ u + k2 u = 0, x ∈ Rn \ Ω , (43.1)
ik(x,θ )
u = u0 + usc , u0 = e , θ ∈S , n−1
 
(n−1)/2 ∂ usc
lim r − ikusc = 0, r = |x|,
r→+∞ ∂r
∂u
+ iλ u = 0, x ∈ ∂ Ω ,
∂ν

where the boundary condition on ∂ Ω is assumed in the sense of uniform conver-

gence as x → ∂ Ω , λ (x) ∈ C(∂ Ω ), λ (x) > 0, and ν is the outward unit normal
vector to ∂ Ω .

Theorem 43.1 (Representation formula for an exterior domain). Let usc ∈ C2

(Rn \ Ω ) ∩ C(Rn \ Ω ) be a solution of (43.1) such that ∂∂uνsc exists in the sense of
uniform convergence as x → ∂ Ω . Then for all x ∈ Rn \ Ω we have
  
usc (x) = usc (y)∂νy G+ +
k (|x − y|) − ∂νy usc (y)Gk (|x − y|) dσ (y), (43.2)
∂Ω

where G+
k is defined in Chapter 32.

Proof. Let x ∈ Rn \ Ω be fixed and let ε > 0 be so small that Bε (x) ⊂ Rn \ Ω . Let
BR (0) be a ball of radius R containing both Ω and Bε (x) (Figure 43.1).

c Springer International Publishing AG 2017 485

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 43
486 Part IV: Partial Differential Equations

x ε

Ω
0
R

/ Ω , Bε (x) and BR (0).

Fig. 43.1 Geometric illustration of x ∈

Then from Green’s identity we have that

  
0= usc (y)∂νy G+ +
k (|x − y|) − ∂νy usc (y)Gk (|x − y|) dσ (y).
∂ Ω ∪∂ Bε (x)∪∂ BR (0)

On the spheres ∂ Bε (x) and ∂ BR (0) we have

 
∂  ∂ 
∂ν y = −  , ∂ν y = ,
∂ r r=ε ∂ r r=R

and in addition, on ∂ Bε (x) one has

∂νy |x − y| = νy · ∇y |x − y| = −1.

The well known properties of Hankel functions (see, e.g., [23]) give that
 
(1) (1)
ρ −(n−2)/2 H(n−2)/2 (ρ ) = −ρ −(n−2)/2 Hn/2 (ρ ).
ρ

Taking into account all these facts, we obtain that (n ≥ 2)

 (n−2)/2
i 1 (1)
∂νy G+
k (|x − y|) = kn/2 |x − y|−(n−2)/2 Hn/2 (k|x − y|).
4 2π

(1)
Using now the asymptotic behavior of Hn/2 (k|x − y|) for small arguments (see, e.g.,
[23])
 
(1) −i k|x − y| −n/2
Hn/2 (k|x − y|) ∼ Γ (n/2),
π 2

we obtain that on the sphere ∂ Bε (x), one has

43 The Direct Scattering Problem for the Helmholtz Equation 487

|x − y|1−n
∂νy G+
k (|x − y|) ∼ , n ≥ 2.
ωn

Thus, letting ε → 0+, we see that

  
0∼ usc (y)∂νy G+ +
k (|x − y|) − ∂νy usc (y)Gk (|x − y|) dσ (y)
∂Ω


∂ ∂
+ usc (y) G+ (|x − y|) − u sc (y)G +
(|x − y|) dσ (y)
|y|=R ∂r k ∂r k


ε 1−n ∂
− usc (y) − usc (y)G+ ( ε ) dσ (y),
|x−y|=ε ωn ∂r k

and therefore (using the Sommerfeld radiation condition)

  
usc (x) = usc (y)∂νy G+ +
k (|x − y|) − ∂νy usc (y)Gk (|x − y|) dσ (y)
∂Ω


∂ ∂
+ usc (y) G+ (|x − y|) − u sc (y)G +
(|x − y|) dσ (y)
|y|=R ∂r k ∂r k
  
= usc (y)∂νy G+ +
k (|x − y|) − ∂νy usc (y)Gk (|x − y|) dσ (y)
 ∂Ω 
+ usc (y)o(1/R(n−1)/2 ) − G+ k (|x − y|)o(1/R (n−1)/2
) dσ (y). (43.3)
|y|=R

In order to estimate the latter integral in (43.3) we need the following lemma.

Lemma 43.2. For usc from (43.1) it is true that


lim |usc (y)|2 dσ (y) < ∞.
R→+∞ |y|=R

Proof. The Sommerfeld radiation condition gives

 2
 ∂ usc 
0 = lim  − iku 
sc  dσ (y)

R→+∞ |y|=R ∂ r

   
 ∂ usc 2 ∂ u
= lim   + k |usc (y)| + 2k Im usc
2 2 sc
dσ (y).
R→+∞ |y|=R  ∂ r  ∂r

At the same time, Green’s identity for DR := BR (0) \ Ω implies

  
∂ usc
usc (y) dσ (y) = usc (y)∂νy usc (y)dσ (y) − k2 |usc (y)|2 dy
|y|=R ∂r ∂Ω DR

+ |∇usc (y)|2 dy.
DR
488 Part IV: Partial Differential Equations

This means that for R sufficiently large we have

 
∂ usc
Im usc (y) dσ (y) = Im usc (y)∂νy usc (y)dσ (y).
|y|=R ∂r ∂Ω

We therefore have that


   
 ∂ usc 2
lim   + k |usc (y)| dσ (y) = −2k Im
2 2
usc (y)∂νy usc (y)dσ (y).
R→+∞ |y|=R  ∂ r  ∂Ω

This equality completes the proof of the lemma. 

Taking into account that
 
1
G+
k (|x − y|) = O , |y| = R → ∞,
R(n−1)/2

and Lemma 43.2, we can easily obtain from (43.3), by letting R → ∞, the equality
(43.2). Thus, the theorem is completely proved. 

Corollary 43.3 (Representation formula for an interior domain). Let Ω be a

bounded C2 domain and u ∈ C2 (Ω ) ∩ C1 (Ω ) a solution of (Δ + k2 )u = 0 in Ω .
Then for all x ∈ Ω we have
  
u(x) = ∂νy u(y)G+ +
k (|x − y|) − u(y)∂νy Gk (|x − y|) dσ (y). (43.4)
∂Ω

Corollary 43.4. Under the conditions of Corollary 43.3, u(x) is real-analytic in

x ∈ Ω.

Proof. Since G+
k (|x − y|) for x
= y is real-analytic in x ∈ Ω , the representation for-
mula (43.4) implies that the same is true for u(x). 

Theorem 43.5. Let v ∈ C2 (Rn \ Ω ) ∩ C(Rn \ Ω ) be a solution of the Helmholtz

equation Δ v + k2 v = 0 in Rn \ Ω satisfying the Sommerfeld radiation condition and
the boundary condition ∂ν v + iλ v = 0, λ (x) > 0, on ∂ Ω . Then v ≡ 0.

Proof. Let BR (0) contain Ω in its interior. Then Green’s identity and the boundary
conditions imply that (∂ν = ∂∂r on the spheres)
  
∂ v(y) ∂ v(y)
v(y) − v(y) dσ (y)
|y|=R ∂r ∂r
   
= v(y)∂νy v(y) − v(y)∂νy v(y) dσ (y) = −2i λ (y)|v(y)|2 dσ (y).
∂Ω ∂Ω

But this is equivalent to the equalities

43 The Direct Scattering Problem for the Helmholtz Equation 489
   
∂ v(y)  
2i Im v(y) dσ (y) = 2i Im v(y)∂νy v(y) dσ (y)
|y|=R ∂r ∂Ω

= −2i λ (y)|v(y)|2 dσ (y).
∂Ω

Since λ (y) > 0, these equalities lead to

 
∂ v(y)
Im v(y) dσ (y) = Im v(y)∂νy v(y)dσ (y)
|y|=R ∂r ∂Ω

= λ (y)|v(y)|2 dσ (y) ≥ 0. (43.5)
∂Ω

We have proved in Lemma 43.2 that


   
 ∂ v 2
lim   + k2 |v(y)|2 dσ (y) = −2k Im v(y)∂νy v(y)dσ (y). (43.6)
R→+∞ |y|=R
∂r ∂Ω

But the inequality (43.5) implies that


  
 ∂ v 2
lim   + k |v(y)| dσ (y) = 0
2 2
R→+∞ |y|=R
∂r

and 
Im v(y)∂νy v(y)dσ (y) = 0. (43.7)
∂Ω

Thus, we have from (43.5) and (43.7) (compare with Lemma 43.2)

lim |v(y)|2 dσ (y) = 0
R→+∞ |y|=R

and 
λ (y)|v(y)|2 dσ (y) = 0.
∂Ω

This implies that v ≡ 0 and ∂ν v ≡ 0 on ∂ Ω . These two facts and the representa-
tion formula (43.2) for the scattering solution provide that v(x) ≡ 0 in Rn \ Ω . The
theorem is proved. 
Corollary 43.6. If the solution of the scattering problem (43.1) exists, then it is
unique.
Proof. If two solutions u1 and u2 of (43.1) exist, then their difference v = u1 − u2 =
(1) (2)
usc − usc satisfies the hypothesis of Theorem 43.5. Hence v ≡ 0, i.e., u1 = u2 . 
Theorem 43.7 (Rellich’s lemma). Let u ∈ C2 (Rn \ Ω ) be a solution of the
Helmholtz equation satisfying
490 Part IV: Partial Differential Equations

lim |u(y)|2 dσ (y) = 0.
R→+∞ |y|=R

Then u ≡ 0 in Rn \ Ω .

Proof. Let r  1 and let


U(r) = u(rθ )ϕ (θ )dθ ,
Sn−1

where u is a solution of the Helmholtz equation and ϕ is an arbitrary eigenfunction

of the Laplacian ΔS on the unit sphere, i.e.,

(ΔS + μ 2 )ϕ = 0, μ 2 = m(m + n − 2), m = 0, 1, 2, . . . .,

where
∂2 n−1 ∂ 1
Δ= + + ΔS .
∂ r2 r ∂ r r2

It follows that U(r) satisfies the ordinary differential equation

n−1 
U  (r) + U (r) + (k2 − μ 2 /r2 )U = 0.
r
But the general solution of this equation is given by

(1) (2)
U(r) = K1 r−(n−2)/2 Hν (kr) + K2 r−(n−2)/2 Hν (kr),

(1,2)
where K1 and K2 are arbitrary constants, Hν are Hankel functions of order ν with
(2) (1)
Hν = Hν and ν 2 = μ 2 + ( n−2
2 ) . Since
2 the hypothesis of this theorem implies
that U(r) = o(1/r (n−1)/2 ), we deduce that
K1 and K2 are equal to zero and hence
U(r) ≡ 0 for all r ≥ R0 . The same is true for u(rθ ) due to the completeness of
the eigenfunctions of ΔS on Sn−1 ; see [34]. The claim follows now from the real
analyticity of every solution of the Helmholtz equation. 

Theorem 43.8. Let v ∈ C2 (Rn \ Ω ) ∩ C(Rn \ Ω ) be a solution of the Helmholtz

equation satisfying the Sommerfeld radiation condition at infinity. Let ∂ν (u0 + v)
converge uniformly as x → ∂ Ω and let

Im v(y)∂νy v(y)dσ (y) ≥ 0.
∂Ω

Then v ≡ 0 in Rn \ Ω .

Proof. This follows immediately from (43.6) and Rellich’s lemma

(Theorem 43.7). 
43 The Direct Scattering Problem for the Helmholtz Equation 491

Remark 43.9. All the results of Theorems 43.1–43.8 and their corollaries remain
true if we consider instead of C2 (Rn \ Ω ) ∩ C(Rn \ Ω ) the Sobolev spaces H s .
Namely, we may assume that the problem (43.1) is considered in the space H 2 (Rn \
Ω ) ∩ H 3/2 (Rn \ Ω ).

Here we have considered mostly uniqueness results for these boundary value
problems. The solvability is provided using the results of Chapters 41 and 42 as
follows. As we know, the single layer potential

usc (x) := ϕ (y)G+
k (|x − y|)dσ (y), x ∈ Rn \ ∂ Ω , (43.8)
∂Ω

with continuous density ϕ satisfies the Sommerfeld radiation condition at infinity, is

a solution of the Helmholtz equation in Rn \ ∂ Ω , is continuous in Rn , and satisfies
the discontinuity property (see Theorem 41.21)

∂ν ± usc (x) = ϕ (y)∂νx G+
k (|x − y|)dσ (y) ∓ ϕ (x), x ∈ ∂Ω.
∂Ω

Let us note that these properties of the single layer potential are also valid for ϕ ∈
H −1/2 (∂ Ω ), where the integrals are interpreted in the sense of duality pairing [22,
25]. Thus, (43.8) will solve the scattering problem (43.1), provided that
 
ϕ (x) − 2 ϕ (y)∂νx G+
k (|x − y|)dσ (y) − 2iλ (x) ϕ (y)G+
k (|x − y|)dσ (y)
∂Ω ∂Ω
= 2(∂ν u0 (x) + iλ (x)u0 (x)), x ∈ ∂ Ω , (43.9)

where u0 (x) = eik(x,θ ) . Hence, to establish the existence for the problem (43.1), it
suffices to show the existence of a solution to (43.9) in the normed space C(∂ Ω ). To
this end, we first recall that the integral operators in (43.9) are compact on C(∂ Ω )
(see Theorem 34.9). Hence, by Riesz’s theorem (see Theorem 41.29), it suffices to
show that the homogeneous equation (corresponding to (43.9)) has only the trivial
solution.
Let ϕ be a solution of this homogeneous equation. Then usc from (43.8) will
be a solution of (43.9) with u0 set equal to zero, and hence, by Theorem 43.5 and
Corollary 43.6, we have that this usc (x) is equal to zero for x ∈ Rn \ Ω . By the
continuity of (43.8) across ∂ Ω , usc (x) is a solution of the Helmholtz equation in Ω
as well, and usc (x) = 0 on the boundary ∂ Ω . If we assume now that k2 > 0 is not
a Dirichlet eigenvalue for −Δ in Ω , then usc (x) ≡ 0 in Ω , and by the discontinuity
properties of the single layer potential we have that

0 = ∂ν− usc (x) − ∂ν+ usc (x) = ϕ (x),

i.e., the homogeneous equation under consideration has only the trivial
solution ϕ = 0. Hence by Riesz’s theorem (Theorem 41.29), the corresponding
492 Part IV: Partial Differential Equations

inhomogeneous equation (43.9) has a unique solution that depends continuously

on the right-hand side.
If k2 > 0 is a Dirichlet eigenvalue of −Δ in Ω , then Riesz’s theorem cannot be
applied (it is not valid in this case), and the whole procedure mentioned above does
not work. To obtain an integral equation that is uniquely solvable for all values of
the number k2 > 0 we need to modify the kernel of the representation (43.8). For
the two-dimensional case such a modification is shown in [6, (3.51)]. The same idea
can be considered for higher dimensions with the use of spherical harmonics. So,
we have proved now the following solvability result, valid for all dimensions n ≥ 2.

Theorem 43.10. If k2 > 0 is not a Dirichlet eigenvalue of −Δ in the domain Ω ⊂

Rn , then there exists a unique solution of the scattering problem (43.1) that depends
continuously on u0 (x) = eik(x,θ ) .

It is quite often necessary to consider the scattering problem (43.1) in Sobolev

spaces H s . In that case we define
1
Hloc (Rn \ Ω ) := {u : u ∈ H 1 (BR (0) \ Ω )}

with an arbitrary ball BR (0) of radius R > 0 centered at the origin and containing
Ω . We recall that H −s (∂ Ω ), 0 ≤ s < ∞ is the dual space of H s (∂ Ω ). Then, for
f ∈ H −1/2 (∂ Ω ), a weak solution of

Δ u + k2 u = 0, x ∈ Rn \ Ω , (43.10)
u = u0 + usc , u0 = eik(x,θ ) , θ ∈ Sn−1 ,
 
∂ usc
lim r(n−1)/2 − ikusc = 0, r = |x|,
r→+∞ ∂r
∂ν u(x) + iλ u(x) = 0, x ∈ ∂Ω,

1 (Rn \ Ω ) such that

is defined as a function u ∈ Hloc
  
(∇u · ∇v − k2 uv)dx − i λ uvdσ (x) = − f vdσ (x) (43.11)
R n \Ω ∂Ω ∂Ω

for all v ∈ H 1 (Rn \ Ω ) that are identically equal to zero outside some ball BR (0)
with radius R > 0 sufficiently large. In that case the analogue of Theorem 43.10 for
Sobolev spaces can be proved (see [6, Chapter 8] for details).
Chapter 44
Some Inverse Scattering Problems for the
Schrödinger Operator

The classical inverse scattering problem is to reconstruct the potential q(x) from the
knowledge of the far field data (scattering amplitude, see p. 232) A(k, θ  , θ ), when
k, θ  , and θ are restricted to some given set.
If q ∈ L1 (Rn ), then q(y)u(y, k, θ ) ∈ L1 (Rn ) uniformly with respect to θ ∈ Sn−1
due to

q(y)u(y, k, θ ) = q(y)(eik(θ ,y) + usc (y, k, θ )) = q(y)eik(θ ,y) + |q| 2 · q 1 usc (y, k, θ )
1

and Hölder’s inequality. We may therefore conclude that the scattering amplitude
A(k, θ  , θ ) is well defined and continuous. Also, the following representation holds:


A(k, θ  , θ ) = e−ik(θ ,y) q(y)(eik(θ ,y) + usc )dy
Rn


= e−ik(θ −θ ,y) q(y)dy + R(k, θ  , θ )
Rn
= (2π )n/2 (F q)(k(θ  − θ )) + R(k, θ  , θ ),

where R(k, θ  , θ ) → 0 as k → +∞ uniformly with respect to θ  and θ . This fact

implies that
A(k, θ  , θ ) ≈ (2π )n/2 (F q)(k(θ  − θ )),

or
q(x) ≈ (2π )−n/2 F −1 (A(k, θ  , θ ))(x),

where the inverse Fourier transform must be understood in some special sense.
Let us introduce the cylinders M0 = R × Sn−1 and M = M0 × Sn−1 , and the mea-
sures μθ and μ on M0 and M, respectively, as

c Springer International Publishing AG 2017 493

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 44
494 Part IV: Partial Differential Equations

1
dμθ (k, θ  ) = |k|n−1 dk|θ − θ  |2 dθ  ,
4
1
dμ (k, θ  , θ ) = dθ dμθ (k, θ  ),
|Sn−1 |

2π n/2
where |Sn−1 | = Γ ( 2n ) is the area of the unit sphere Sn−1 , and dθ and dθ  denote the
usual Lebesgue measures on Sn−1 . We shall define the inverse Fourier transform on
M0 and M as

1 
(FM−10 ϕ1 )(x) = e−ik(θ −θ ,x) ϕ1 (k, θ  )dμθ ,
(2π )n/2 M0


1 
(FM−1 ϕ2 )(x) = e−ik(θ −θ ,x) ϕ2 (k, θ  , θ )dμ .
(2π )n/2 M

If we write ξ = k(θ − θ  ), then k and θ  are obtained by

|ξ | ξ
k= , θ  = θ − 2(θ , ξ)ξ, ξ = . (44.1)
2(θ , ξ) |ξ |

Exercise 44.1. Let uθ (k, θ  ) be the coordinate mapping M0 → Rn given as

uθ (k, θ  ) = k(θ − θ  ),

where θ is considered a fixed parameter. Prove that

(1) the formulas (44.1) for k and θ  hold;
(2) the following is true:
 
ϕ ◦ uθ (k, θ  )dμθ (k, θ  ) = ϕ (x)dx
M0 Rn

if ϕ ∈ S is even and
 
ϕ ◦ uθ (k, θ  )dμ (k, θ  , θ ) = ϕ (x)dx
M Rn

if ϕ ∈ S;
(3) in addition,
FM−10 (ϕ ◦ uθ ) = F −1 ϕ

if ϕ ∈ S is even and
FM−1 (ϕ ◦ uθ ) = F −1 ϕ
44 Some Inverse Scattering Problems for the Schrödinger Operator 495

if ϕ ∈ S. Here F −1 is the usual inverse Fourier transform in Rn .

Exercise 44.2. Prove that

(1) A(−k, θ  , θ ) = A(k, θ  , θ );
(2) A(k, θ  , θ ) = A(k, −θ , −θ  ).
n
The approximation q(x) ≈ (2π )− 2 F −1 (A(k, θ  , θ )(x)) for all θ  and θ and for suf-
ficiently large k allows us to introduce the following definitions.

Definition 44.1. The inverse Born approximations qθB (x) and qB (x) of the potential
q(x) are defined by

1 
qθB (x) = (2π )−n/2 (FM−10 A)(x) = e−ik(θ −θ ,x) A(k, θ  , θ )dμθ
(2π )n M0

and

1 
qB (x) = (2π )−n/2 (FM−1 A)(x) = e−ik(θ −θ ,x) A(k, θ  , θ )dμ
(2π )n M

in the sense of distributions.

p
Theorem 44.2 (Uniqueness). Assume that the potential q(x) belongs Lloc (Rn ), n2 <
− μ
p ≤ ∞, n ≥ 3, and has the special behavior |q(x)| ≤ C|x| , μ > 2, |x| → ∞ at
infinity. Then the knowledge of qθB (x) with θ restricted to an (n − 2)-dimensional
hemisphere determines q(x) uniquely.

Proof. It is not difficult to check that if q(x) satisfies the conditions of the present
theorem, then q(x) will satisfy the conditions of Theorem 23.5:

n n+1
q ∈ L p (Rn ), < p≤ ,
2 2
or
n+1 n+1
q ∈ Lσp (Rn ), < p ≤ +∞, σ > 1− .
2 2p

Now we can represent qθB (x) in the form


1 
qθB (x) = e−ik(θ −θ ,x) A(k, θ  , θ )dμθ (k, θ  )
(2π ) M0
n
 
1  
= d μ θ (k, θ 
) e−ik(θ −θ ,x) e−ik(θ ,y) q(y)u(y, k, θ )dy
(2π )n M0 R n
 
1 −ik(θ −θ  ,x−y)
= d μθ e q(y)e−ik(θ ,y) u(y, k, θ )dy,
(2π )n M0 Rn
496 Part IV: Partial Differential Equations

where u(y, k, θ ) is the solution of the Lippmann–Schwinger equation. Setting

v(y, k, θ ) := e−ik(θ ,y) u(y, k, θ )

and making the change of variables ξ = k(θ − θ  ), we obtain

 
 
1 |ξ |
qθB (x) = dξ e−i(ξ ,x−y) q(y)v y, ,θ dy.
(2π )n R n R n
2(θ , ξ)

The usual Fourier transform of qθB (x) is equal to


   
|ξ |
q (ξ ) + (2π )−n/2
θ (ξ ) = q ei(ξ ,y) q(y) v y, ,θ − 1 dy,
2(θ , ξ)
B
Rn

and it implies that

 

|ξ |
|q (ξ )| ≤ (2π )−n/2
θ (ξ ) − q |q(y)| v y, , θ − 1 dy,
B
R n 
2(θ , ξ )

where the function v(y, k, θ ) solves the equation


v(x, k, θ ) = 1 − e−ik(x,θ ) G+
k (|x − y|)e
ik(y,θ )
q(y)v(y, k, θ )dy,
Rn

i.e.,


v = 1−G k (qv),

k = e−ik(x−y,θ ) G+ . For k sufficiently large we obtain that
where G k


−1
v = (I + G k q) (1),

or


v = 1−G q (q), (44.2)



−ik(x−y,θ ) G , and the integral
where G q is an integral operator with kernel Gq = e q

operator Gq with this kernel also satisfies the equation (H − k2 )G q = I. In order to
prove (44.2), we recall that
q = G
G k − Gk qGq ,

and therefore,





G q = Gk − Gk qGq ,
44 Some Inverse Scattering Problems for the Schrödinger Operator 497

or


−1 

G q = (I + Gk q) Gk .

The last equality implies that



−1 

G q (q) = (I + Gk q) Gk (q) = −(v − 1),

because

−1 

(I + G k q) Gk (q) = −(v − 1)

is equivalent to





G k (q) = −(I + Gk q)(v − 1) = −(v − 1) − (Gk q)(v) + (Gk q)(1)




= −v + 1 − 1 + v + G k (q) = Gk (q).

We may therefore apply Theorem 23.5 to obtain


C
v − 1 2p = G q (q) 2p ≤ q 2p ,
p−1
L−σ /2 (Rn )
p−1
L−σ /2 kγ L
p+1
σ /2

where γ , p, and σ are as in that theorem. It remains only to check that the potential
2p
p
q ∈ Lloc (Rn ) with the special behavior at infinity belongs to Lσp+1
/2 (R ). But that is a
n

very simple exercise. Hence, the latter inequality leads to

 γ
|(ξ, θ )|
|q
θ (ξ ) − q
(ξ )| ≤ C q 2 , ξ
= 0
|ξ |
B 2p
p+1
Lσ /2 (Rn )

with the same γ . If q1 and q2 are as q, then

q1 (ξ ) − q
q1 (ξ ) − q2 (ξ )| = |
| θ +q
B
θ −q
B q1 (ξ ) − q
2 (ξ )| ≤ | θ | + |q
B
θ −q
B 2 (ξ )|
 γ  γ
|(ξ, θ )| |(ξ, θ )|
≤ C q1 2p
2
+C q2 2p2
=0
L
p+1
(Rn )
|ξ | L
p+1
(Rn )
|ξ |
σ /2 σ /2

if (ξ, θ ) = 0. Thus, this theorem is proved, because (ξ, θ ) = 0 precisely as θ runs

through an (n − 2)-dimensional hemisphere see [31, 32]. 


Theorem 44.3 (Saito’s formula). Under the same assumptions for q(x) as in
Theorem 44.2,
  
 (2π )n q(y)dy
lim kn−1 e−ik(θ −θ ,x) A(k, θ  , θ )dθ dθ  = ,
k→+∞ Sn−1 Sn−1 π Rn |x − y|n−1
498 Part IV: Partial Differential Equations

where the limit holds in the classical sense for n < p ≤ ∞ and in the sense of distri-
butions for n2 < p ≤ n.

Proof. Let us consider only the case n < p ≤ ∞. The proof for n
2 < p ≤ n requires
some changes.
By definition of the scattering amplitude,
 

I := kn−1 A(k, θ  , θ )e−ik(θ −θ ,x) dθ dθ 
Sn−1 Sn−1
  

= kn−1 q(y)dy eik(θ −θ ,y−x) dθ dθ 
Rn Sn−1 Sn−1
  
 
+ kn−1 q(y)dy e−ik(θ ,y) R(y, k, θ )e−ik(θ −θ ,x) dθ dθ  =: I1 + I2 ,
Rn Sn−1 Sn−1

where R(y, k, θ ) is given by


R(y, k, θ ) = − G+
k (|y − z|)q(z)u(z, k, θ )dz
Rn

and u(z, k, θ ) is the solution of the Lippmann–Schwinger equation. Since

   2
ik(θ −θ  ,y−x)

e 
dθ dθ = eik(θ ,y−x)
dθ
Sn−1 Sn−1 Sn−1
 π 2
4π n−1 ik|y−x| cos ψ
= 2 n−1 e (sin ψ ) dψ
n−2
Γ ( 2 ) 0
J 2n−2 (k|x − y|)
= (2π ) n 2
,
(k|x − y|)n−2

we have that I1 can be represented in the form


q(y)
I1 = (2π )n k J 2n−2 (k|x − y|)dy.
Rn |x − y|n−2 2

We consider two cases: k|x − y| < 1 and k|x − y| > 1. In the first case, using Hölder’s
inequality the integral I1 over {y : k|x − y| < 1} can be estimated by

|q(y)|(k|x − y|)n−2
|I1 | ≤ Ck dy
|x−y|< 1k |x − y|n−2
  1   1
p p
≤ Ck n−1
1
|q(y)| dy
p
1
1 · dy
|x−y|< k |x−y|< k
 1  1
p p
−n n −1
= Ckn−1 k p 1
|q(y)| p
dy = Ck p
1
|q(y)| dy
p
→0
|x−y|< k |x−y|< k
44 Some Inverse Scattering Problems for the Schrödinger Operator 499

as k → +∞, since n < p ≤ ∞. This means that for every fixed x (or even uniformly
with respect to x) I1 approaches zero as k → ∞. Hence, we have only to estimate the
integral I1 over {y : k|x − y| > 1}. The asymptotic behavior of the Bessel function
Jν (·) for large argument implies that

q(y)
I1 = (2π ) k n
|x−y|> 1k |x − y|n−2
  2
2  nπ π  1
× cos k|x − y| − + +O dy
π k|x − y| 4 4 (k|x − y|)3/2

q(y)
= (2π )n k
|x−y|> 1k |x − y|n−2
  
2 cos2 (k|x − y| − n4π + π4 ) 1
× +O dy
π k|x − y| (k|x − y|)2

(2π )n q(y)dy
=
π |x−y|> 1k |x − y|n−1
 
(2π )n q(y) nπ π 
+ cos 2k|x − y| − + dy
π |x−y|> 1k |x − y|n−1 2 2

1 |q(y)|O(1) (1) (2) (3)
+ dy =: I1 + I1 + I1 .
k |x−y|> 1k |x − y|n

It is clear that 
(1) (2π )n q(y)dy
lim I1 =
k→+∞ π Rn |x − y|n−1

and
(2)
lim I = 0.
k→+∞ 1

The latter fact follows from the following arguments. Since q belongs to L p (Rn ) for
p > n and has the special behavior at infinity, we may conclude that the L1 norm of
q(y)
the function |x−y| n−1 is uniformly bounded with respect to x. Hence it follows from
(2)
the Riemann–Lebesgue lemma that I1 approaches zero uniformly with respect to
(3)
x as k → +∞. For I1 we have the estimate

(3) C |q(y)|dy
|I1 | ≤ .
k1−δ Rn |x − y|n−δ
 q(y)dy
If we choose δ such that 1 > δ > np , then Rn |x−y|n−δ will be uniformly bounded
(3)
with respect to x. Therefore, I1 → 0 as k → ∞ uniformly with respect to x. If we
collect all estimates, we obtain that
500 Part IV: Partial Differential Equations

(2π )n q(y)dy
lim I1 = .
k→∞ π Rn |x − y|n−1

Our next task is to prove that I2 → 0 as k → ∞. Since

  
 
I2 = kn−1 q(y)dy e−ik(θ ,y) R(y, k, θ )e−ik(θ −θ ,x) dθ dθ  ,
Rn Sn−1 Sn−1

where 
R(y, k, θ ) = − G+ 
k (|y − z|)q(z)u(z, k, θ )dz = −Gk (qu),
Rn

q (qeik(θ ,z) ). Hence, I2 can be represented as

one can check that R(y, k, θ ) = −G
    
ik(θ  ,x−y)
  ik(θ ,z−x)
I2 = −k n−1
q(y)dy e dθ · Gq q(z) e dθ
Rn Sn−1 Sn−1

 
J n−2 (k|x − y|) J n−2 (k|x − z|)
= −kn−1 (2π )n q(y) 2 q q(z) 2
·G dy
n−2 n−2
Rn (k|x − y|) 2 (k|x − z|) 2

 
J n−2 (k|x − y|) 1 J n−2 (k|x − z|)
= (2π ) k q 1 (y)  |q(z)| 2
n−2 · Kq
n 2 2
n−2 dy,
Rn 2 (|x − y|) 2 (|x − z|) 2

q is an integral operator with kernel

where K

1
Kq (x, y) = −|q(x)| 2 Gq (k, x, y)q 1 (y).
2

q : L2 (Rn ) → L2 (Rn ) with the norm estimate

It follows from Theorem 23.5 that K

q L2 →L2 ≤ C
K ,
kγ
where γ is as in that theorem. We can therefore estimate I2 using Hölder’s inequality
as
 J 2n−2 (k|x − y|)
C
|I2 | ≤ γ k |q(y)| 2 dy.
k Rn |x − y|n−2

By the same arguments as in the proof for I1 we can obtain that

 J 2n−2 (k|x − y|)

k |q(y)| 2
dy < ∞
Rn |x − y|n−2

uniformly with respect to x. This implies that

44 Some Inverse Scattering Problems for the Schrödinger Operator 501

C
|I2 | ≤ →0
kγ
as k → +∞. 


Remark 44.4. This proof holds also for n = 2. In dimension n = 1 there is an analo-
gous result in which we replace the double integral on the left-hand side by the sum
of four values of the integrand at θ = ±1 and θ  = ±1.

Theorem 44.5. Let us assume that n ≥ 2. Under the same assumptions for q1 (x)
and q2 (x) as in Theorem 44.3 let us assume that the corresponding scattering am-
plitudes Aq1 and Aq2 coincide for some sequence k j → ∞ and for all θ  , θ ∈ Sn−1 .
Then q1 (x) = q2 (x) in the sense of L p for n < p ≤ ∞ and in the sense of distributions
for n2 < p ≤ n.

Proof. Saito’s formula shows that we have only to prove that the homogeneous
equation 
q(y)dy
ψ (x) := =0
R n |x − y|n−1

has only the trivial solution q(y) ≡ 0. Let us assume that n < p ≤ ∞. Introduce the
space S0 (Rn ) of all functions from the Schwartz space that vanish in some neigh-
borhood of the origin. Due to the conditions for the potential q(x) we may conclude
(as before) that ψ ∈ L∞ (Rn ), and ψ defines a tempered distribution. Then for every
function ϕ ∈ S0 (Rn ) it follows that

 , ϕ  = Cn |ξ |−1 q(ξ ), ϕ  = Cn 

0 = ψ q(ξ ), |ξ |−1 ϕ .

Since ϕ (ξ ) ∈ S0 (Rn ), we have |ξ |−1 ϕ ∈ S0 (Rn ) also. Hence, for every h ∈ S0 (Rn )
the following equation holds:

q, h = 0.

This means that the support of q(ξ ) is at most at the origin, and therefore q(ξ ) can
be represented as
q(ξ ) = ∑ Cα Dα δ .
|α |≤m

Hence, q(x) is a polynomial. But due to the behavior at infinity we must conclude
that q ≡ 0. This proves Theorem 44.5. 


Let us return now to the Born approximation of q(x). Repeated use of the Lippmann–
Schwinger equation leads to the following representation for the scattering ampli-
tude A(k, θ  , θ ):
502 Part IV: Partial Differential Equations

m 

A(k, θ  , θ ) = ∑  j · (|q| 2 eik(x,θ ) )(y)dy
e−ik(θ ,y) q 1 (y)K
1

j=0 R
n 2


+ m+1 (|q| 21 (u(x, k, θ ))(y)dy,
e−ik(θ ,y) q 1 (y)K
Rn 2

 is an
where u(x, k, θ ) is the solution of the Lippmann–Schwinger equation and K
integral operator with kernel
1
K(x, y) = |q(x)| 2 G+
k (|x − y|)q 1 (y). 2

The equality for A can be reformulated in the sense of integral operators in L2 (Sn−1 )
as
m
 = ∑ Φ0∗ (k) sgn qK
A  j Φ0 (k) + Φ0∗ (k) sgn qK
m+1 Φ (k),
j=0

where Φ0 and Φ (k) are defined by (23.9) and (23.10), and Φ0∗ is the L2 - adjoint of
Φ0 .
Using this equality and the definition of Born’s potential qB (x), we obtain
m    
qB (x) = ∑ FM−1  j Φ0 (k) + F −1 Φ0∗ (k) sgn qK
Φ0∗ (k) sgn qK m+1 Φ (k) ,
M
j=0

where the inverse Fourier transform is applied to the kernels of the corresponding
integral operators. If we rewrite the latter formula as
m
qB (x) = ∑ q j (x) + q
m+1 (x),
j=0

then the term q j has the form

  

|q(z)| 2 e−ik(z,θ ) dz sgn q(z)K j (z, y, k)|q(y)| 2 eik(θ ,y) dy
1 1
q j (x) = FM−1
Rn Rn
 1 ik(θ  ,y)

−1 ∗
= FM Φ0 sgn qK  (|q| 2 e
j
)
  
1 
 j (|q| 21 eik(θ  ,y) ) ,
= e−ik(θ −θ ,x) dμ (k, θ  , θ ) Φ0∗ sgn qK
(2π ) M
n

and a similar formula holds for q

m+1 with obvious changes.
In order to formulate the result about the reconstruction of singularities of the
unknown potential q(x), let us set A(k, θ  , θ ) = 0 for |k| ≤ k0 , where k0 > 0 is arbi-
trarily large.
Theorem 44.6. Assume that the potential q belongs to L2pδ (Rn )∩L1 (Rn ) with (3n−
3)/2 < p ≤ ∞, n ≥ 2, and 2δ > 1 − (n + 1)/(2p). Then for all j ≥ 2 the terms
44 Some Inverse Scattering Problems for the Schrödinger Operator 503

q j (x) and q
j (x) in the Born series belong to the Hölder class Cα (Rn ) for all α ≤
1 − (3n − 3)/(2p).

Proof. For x1 , x2 ∈ Rn we have (see [31])

  
q j (x1 ) − q j (x2 ) = C |k|n−1 dk dθ dθ  (1 − (θ , θ  ))
|k|≥k0 Sn−1 Sn−1 (44.3)
 
 j Φ0 (k, θ , θ  )(e−ik(θ −θ ,x1 ) − e−ik(θ −θ ,x2 ) ).
× Φ0∗ sgn (q)K

For l = 1, 2 let us define el (θ ) = eik(θ ,xl ) ∈ L2 (Sn−1 ) and El (θ ) = el θ ∈ (L2 (Sn−1 ))2 .
Then the latter difference is equal to
 
C |k|n−1 dk (e1 , Φ0∗ sgn (q)K  j Φ0 e1 )L2 (Sn−1 ) − (e2 , Φ0∗ sgn (q)K j Φ0 e2 )L2 (Sn−1 )
|k|≥k0

 j Φ0 E1 )L2 (Sn−1 ) + (E2 , Φ0∗ sgn (q)K
− (E1 , Φ0∗ sgn (q)K  j Φ0 E2 )L2 (Sn−1 ) .
(44.4)
Since E1 − E2 L2 (Sn−1 ) = e1 − e2 L2 (Sn−1 ) and el 2L2 (Sn−1 ) = |Sn−1 |, we obtain
from (44.3)–(44.4) the estimate
  
  j Φ0 
|q j (x1 ) − q j (x2 )| ≤ Cn |k|n−1 e1 − e2 L2 (Sn−1 ) Φ0∗ sgn (q)K  dk.
|k|≥k0

Note that (see [43])


e1 − e2 2L2 (Sn−1 ) = (2 − eik(θ ,x2 −x1 ) − eik(θ ,x1 −x2 ) )dθ
Sn−1
 
J(n−2)/2 (|k||x1 − x2 |)
= 2 |Sn−1 | − (2π )n/2 ,
(|k||x1 − x2 |)(n−2)/2

where Jν is the Bessel function of order ν . By Lemma 23.13 and Theorem 23.5 we
get
 
 ∗  j Φ0  C
Φ0 sgn (q)K  ≤ γ ( j+1)+n−2 ,
|k|

where γ = 1 − (n − 1)/(2p).
If we set r = |x1 − x2 |, we have to estimate the integral

 ∞  1/2
kn−1 J(n−2)/2 (kr)
|Sn−1 | − (2π )n/2 dk.
k0 kγ ( j+1)+n−2 (kr)(n−2)/2

We split this integral into two parts: over 1/r < k < ∞ and over k0 < k < 1/r. By
the asymptotics of the Bessel functions for large argument, the first part can be
estimated from above by
504 Part IV: Partial Differential Equations

 ∞
kn−1
|S n−1
| dk ≤ Cn rγ ( j+1)−2 ,
1/r kγ ( j+1)+n−2

where j is chosen so large that γ ( j + 1) > 2. For the second part of the integral we
use the asymptotics of the Bessel functions for small argument [23], namely,

x(n−2)/2
J(n−2)/2 (x) = (1 + O(x2 )), x → 0.
2(n−2)/2Γ (n/2)

Since |Sn−1 | = 2π n/2 /Γ (n/2), we may estimate the second part from above by
 1/r  1/r
kn dk dk
Cr = Cr ≤ Crmin(1,γ ( j+1)−2) .
k0 kγ ( j+1)+n−2 k0 kγ ( j+1)−2

To finish the proof we use the fact that q j ∈ L∞ (Rn ), which holds since
 ∞
kn−1
|q j (x)| ≤ C ( e L2 (Sn−1 ) + E L2 (Sn−1 ) )dk < ∞
k0 kγ ( j+1)+n−2

for γ ( j + 1) > 2. The latter condition implies, for γ = 1 − (n − 1)/(2p) and j ≥ 2,

that (3n − 3)/2 < p ≤ ∞. For q
j the proof is the same with obvious changes. This
completes the proof. 

The first nonlinear term q1 (x) can be rewritten as

Γ (n/2)
q1 (x) = − G(y − x, z − x)q(y)q(z)dydz, (44.5)
2n+3 π 3n/2 R2n

where G is the tempered distribution

 ∞ 
G(y, z) = + (|y − z|)eik(θ ,z)−ik(θ  ,y) |k|n−1 |θ − θ  |2 dkdθ dθ  .
Gk
−∞ Sn−1 Sn−1

Lemma 44.7. The 2n-dimensional Fourier transform of G equals

 ξ , η ) = −Cn (ξ , η ) ,
G(
|ξ |2 |η |2

where Cn is a positive constant depending only on n.

Proof. The 2n-dimensional Fourier transform of G becomes
 ∞  

 ξ , η) =
G( |k| n−1  2
|θ − θ | dkdθ dθ 
e−i(k(θ −θ )+ξ +η ,y) dy
−∞ Sn−1 Sn−1 Rn

× + (|s|)e−i(η −kθ ,s) ds.
Gk
Rn
44 Some Inverse Scattering Problems for the Schrödinger Operator 505

Using the fact that

+ (·) = lim 1
G ,
k ε →0 | · |2 − k2 − iε

 in the topology of S (R2n ) by G

we may approximate G ε given by
 ∞ 
ε =
G kn−1 |θ − θ  |2 dkdθ dθ 
0 Sn−1 Sn−1

  

e−i(k(θ −θ )+ξ +η ,y) ei(k(θ −θ )−ξ −η ,y)
× + dy.
Rn |η |2 − 2k(θ , η ) − iε |η |2 + 2k(θ , η ) + iε

If we define the variable ζ = k(θ − θ  ) with the Jacobian 12 kn−1 |θ − θ  |2 , then kθ

depends on ζ as
|ζ |2
kθ = ζ − θ .
2(ζ , θ  )

Since the Fourier transform of 1 equals (2π )n δ , it follows that

 
ε (ξ , η ) = 2(2π )n
G
Sn−1 Rn
⎛ ⎞
δ (ζ − ξ − η ) δ (ζ + ξ + η )
×⎝ + ⎠ dζ dθ 
 |ζ |2 (η ,θ  )  |ζ |2 (η ,θ  )
|η | − 2(ζ , θ ) + (ζ ,θ  ) − iε |η | + 2(ζ , θ ) − (ζ ,θ  ) + iε
2 2
  
 1 1
= 2(2π ) n
( f,θ ) + dθ  ,
Sn−1 (g, θ  ) − iε ( f , θ  ) (g, θ  ) + iε ( f , θ  )

where f = ξ + η and g = |ξ |2 (ξ + η ) − |ξ + η |2 ξ . This expression leads us to


ε (ξ , η ) = 4(2π )n p.v.
 ξ , η ) = lim G ( f , θ )  ( f , g) (ξ , η )
G( 
dθ = Cn = −Cn 2 2 ,
ε →0 Sn−1 (g, θ ) |g| 2 |ξ | |η |

where we have used the precise value of the principal value integral; see [30, proof
of Lemma 2.4]. 


Lemma 44.8. Assume that the potential q satisfies all conditions of Theorem 44.6.
Then the first nonlinear term q1 admits the representation
 2
x−y
q1 (x) = Cn q(y)dy

Rn |x − y|n

with some positive constant Cn depending only on n.

506 Part IV: Partial Differential Equations

Proof. The representation (44.5) and Lemma 44.7 imply that

   
−1 (ξ , η ) −1 ξ q(ξ ) η q(η )
q1 (x) = Cn F2n q(ξ )
q(η ) (x, x) = Cn F2n , (x, x),
|ξ |2 |η |2 |ξ |2 |η |2

−1
where F2n denotes the 2n-dimensional inverse Fourier transform. The claim fol-
lows now from F −1 (ξ /|ξ |2 ) = Cn x/|x|n . 


Lemma 44.9. Under the same assumptions on q as in Theorem 44.6 we have that
(1) for 3(n − 1)/2 < p < ∞, q1 belongs to (Wp,2
1
δ −1 (R )) with 1 − (n + 1)/(2p) <
n 2

2δ < n − n/p;
(2) for p = ∞, q1 belongs to the Hölder space C1 (Rn ).

Proof. We introduce the Riesz potential I −1 and Riesz transform R (see [37] and
Chapter 21) as
   
−1 −1 f(ξ ) −1 ξ f(ξ )
I f (x) = F (x), R f (x) = F (x).
|ξ | |ξ |

Note that ∇I −1 = R is bounded in L p (Rn ) for all 1 < p < ∞; see [37]. From [27] we
know that
I −1 : Lσp +1 (Rn ) → Lσp (Rn )

for −n/p < σ < n−1−n/p. This proves (1). Part (2) can be proved like [31, Lemma
2.2]. 


The latter steps lead to the following main result.

Theorem 44.10 (Reconstruction of singularities). Assume that the potential q be-

longs to L2pδ (Rn ) ∩ L1 (Rn ) with p and δ as in Theorem 44.6. Then
(1) for max(3(n − 1)/2, n) < p ≤ ∞ the difference qB − q is a continuous function
in Rn ;
(2) for 3(n − 1)/2 < p ≤ max(3(n − 1)/2, n) the difference qB − q − q1 is a contin-
uous function in Rn .

Proof. The proof of this theorem follows immediately from Lemmas 44.7, 44.8, and
44.9 and the Sobolev embedding theorem. 


The statement of Theorem 44.10 means that all singularities and jumps of the
unknown potential can be recovered by the Born approximation. In particular, if
the potential is the characteristic function of an arbitrary bounded domain, then this
domain can be uniquely determined from the scattering data using a linear method.
Chapter 45
The Heat Operator

We turn our attention now to the heat operator

L = ∂t − Δx , (x,t) ∈ Rn × R.

The heat operator is a prototype of parabolic operators. These are operators of the
form
∂t + ∑ aα (x,t)∂xα ,
|α |≤2m

where the sum satisfies the strong ellipticity condition

(−1)m ∑ aα (x,t)ξ α ≥ ν |ξ |2m ,

|α |=2m

for all (x,t) ∈ Rn × R and ξ ∈ Rn \ {0} with ν > 0 constant.

We begin by considering the initial value problem

∂t u − Δ u = 0, in Rn × (0, ∞),
u(x, 0) = f (x).

This problem is a reasonable problem both physically and mathematically.

Assuming for the moment that f ∈ S, the Schwartz space, and taking the Fourier
transform with respect to x only, we obtain

∂t u(ξ ,t) + |ξ |2 u(ξ ,t) = 0,
(45.1)
u(ξ , 0) = f(ξ ).

c Springer International Publishing AG 2017 507

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 45
508 Part IV: Partial Differential Equations

If we solve the ordinary differential equation (45.1), we obtain

u(ξ ,t) = e−|ξ | t f(ξ ).

2

Thus (at least formally)

   
u(x,t) = F −1 e−|ξ | t f(ξ ) = (2π )−n/2 f ∗ F −1 e−|ξ | t (x,t) = f ∗ Kt (x),
2 2

where   |x|2
Kt (x) = (2π )−n/2 F −1 e−|ξ | t ≡ (4π t)−n/2 e− 4t , t > 0,
2
(45.2)

is called the Gaussian kernel. We define Kt (x) ≡ 0 for t ≤ 0.

Exercise 45.1. Prove (45.2).
Let us first prove that 
Kt (x)dx = 1.
Rn

Indeed, using polar coordinates, we have

   ∞ 
|x|2 r2
Kt (x)dx = (4π t)−n/2 e− 4t dx = (4π t)−n/2 rn−1 e− 4t dr dθ
Rn Rn 0 |θ |=1
 ∞
r2
= ωn (4π t)−n/2 rn−1 e− 4t dr
0
 ∞
−n/2 n−1 1 √ ds
= ωn (4π t) (4st) 2 e−s 4t √
0 2 s

ωn −n/2 ∞ n/2−1 −s
= π s e ds
2 0
ωn −n/2 1 2π n/2 −n/2
= π Γ (n/2) = π Γ (n/2) = 1.
2 2 Γ (n/2)

Theorem 45.1. Suppose that f ∈ L∞ (Rn ) is uniformly continuous. Then u(x,t) :=

( f ∗ Kt )(x) satisfies ∂t u − Δ u = 0 and

u(·,t) − f (·)L∞ (Rn ) → 0

as t → 0+.
Proof. For fixed t > 0,
 
|x−y|2 |x − y|2 n
Δx Kt (x − y) = (4π t)−n/2 e− 4t − ,
4t 2 2t

and for fixed |x − y|

= 0,
45 The Heat Operator 509
 
|x−y|2 |x − y|2 n
∂t Kt (x − y) = (4π t)−n/2 e− 4t − .
4t 2 2t

Therefore, (∂t − Δx )Kt (x − y) = 0.

But we can differentiate (with respect to x and t) under the integral sign, since
this integral will be absolutely convergent for all t > 0. We may therefore conclude
that
∂t u(x,t) − Δx u(x,t) = 0.

It remains only to verify the initial condition. We have


u(x,t) − f (x) = ( f ∗ Kt )(x) − f (x) = f (y)Kt (x − y)dy − f (x)
R n
 
= f (x − z)Kt (z)dz − f (x)Kt (z)dz
R Rn
n

= ( f (x − z) − f (x))Kt (z)dz
R
n
√
= ( f (x − η t) − f (x))K1 (η )dη .
Rn

The assumptions on f imply that

√ 
|u(x,t) − f (x)| ≤ sup | f (x − η t) − f (x)| K1 (η )dη
x∈Rn ,|η |<R Rn

+ 2  f L∞ (Rn ) K1 (η )dη < ε /2 + ε /2
|η |≥R

for small t and for R large enough. So we can see that u(x,t) is continuous (even
uniformly continuous and bounded) for (x,t) ∈ Rn × [0, ∞) and u(x, 0) = f (x). 

Corollary 45.2. u(x,t) ∈ C∞ (Rn × R+ ).

Proof. We can differentiate under the integral sign defining u as often as we please,
because the exponential function increases at infinity faster than any polynomial.
Thus, the heat equation takes arbitrary initial data (bounded and uniformly continu-
ous) and smooths them out. 

Corollary 45.3. Suppose f ∈ L p (Rn ), 1 ≤ p < ∞. Then u(x,t) := ( f ∗ Kt )(x) satis-

fies ∂t u − Δ u = 0 and
u(·,t) − f (·)L p (Rn ) → 0

as t → 0+. And again u(x,t) ∈ C∞ (Rn × R+ ).

Theorem 45.4 (Uniqueness). Suppose u(x,t) ∈ C2 (Rn × R+ ) ∩C(Rn × R+ ) satis-

fies ∂t u − Δ u = 0 for t > 0 and u(x, 0) = 0. If for every ε > 0 there exists cε > 0 such
that
510 Part IV: Partial Differential Equations

|u(x,t)| ≤ cε eε |x| , |∇x u(x,t)| ≤ cε eε |x| ,

2 2
(45.3)

then u ≡ 0.

Proof. For two smooth functions ϕ and ψ , it is true that

n
ϕ (∂t ψ − Δ ψ ) + ψ (∂t ϕ + Δ ϕ ) = ∑ ∂ j (ψ∂ j ϕ − ϕ∂ j ψ ) + ∂t (ϕψ ) = ∇x,t · F,
j=1

where F = (ψ∂1 ϕ − ϕ∂1 ψ , . . . , ψ∂n ϕ − ϕ∂n ψ , ϕψ ). Given x0 ∈ Rn and t0 > 0, let

us take
ψ (x,t) = u(x,t), ϕ (x,t) = Kt0 −t (x − x0 ).

Then
∂t ψ − Δ ψ = 0, t > 0,
∂t ϕ + Δ ϕ = 0, t < t0 .

If we apply the divergence theorem in the region

Ω = {(x,t) ∈ Rn × R+ : |x| < r, 0 < a < t < b < t0 } ,

we obtain
  
0= F · ν dσ = u(x, b)Kt0 −b (x − x0 )dx − u(x, a)Kt0 −a (x − x0 )dx
∂Ω |x|≤r |x|≤r
 b  n
xj
+
a
dt
|x|=r
∑ u(x,t)∂ j Kt0 −t (x − x0 ) − Kt0 −t (x − x0 )∂ j u(x,t)
r
dσ (x).
j=1

Letting r → ∞, the last sum vanishes by assumptions (45.3). We therefore have

 
0= u(x, b)Kt0 −b (x − x0 )dx − u(x, a)Kt0 −a (x − x0 )dx.
Rn Rn

Let us prove that


lim Kt0 −b (x − x0 )u(x, b)dx = u(x0 ,t0 )
b→t0 − Rn

and 
lim Kt0 −a (x − x0 )u(x, a)dx = 0.
a→0+ Rn

Since 
Kt0 −b (x − x0 )dx = 1,
Rn
45 The Heat Operator 511

we have after a change of variables that


Kt0 −b (x − x0 )u(x, b)dx − u(x0 ,t0 )
Rn

= Kt0 −b (x − x0 )[u(x, b) − u(x0 ,t0 )]dx
Rn

= Kτ (z)[u(x0 + z,t0 + τ ) − u(x0 ,t0 )]dz.
Rn

We divide the latter integral into two parts: |z| < δ and |z| > δ . The first part can be
estimated from above by

sup |u(x0 + z,t0 + τ ) − u(x0 ,t0 )| Kτ (z)dz
|z|<δ |z|<δ

≤ sup |u(x0 + z,t0 + τ ) − u(x0 ,t0 )| → 0

|z|<δ

as τ → 0 and δ → 0 due to the continuity of u(x,t) at the point (x0 ,t0 ). The second
part can be estimated from above by (see (45.3))
 
cε
Kτ (z)eε |x0 +z| dz ≤ e−|z| /(4τ )+ε |z| dz
2 2 2
cε
|z|>δ (4πτ ) n/2 |z|>δ

cε −|y|2 /4+ετ |y|2
= τ n/2
√ e dy → 0
(4πτ ) n/2 |y|>δ / τ

as τ → 0. Thus the first limit is justified.

For the second limit we may first rewrite the integral as
 
Kt0 −a (x − x0 )u(x, a)dx = Kt0 −a (z)u(x0 + z, a)dz
Rn Rn
 
= Kt0 −a (z)u(x0 + z, a)dz + Kt0 −a (z)u(x0 + z, a)dz.
|z|≤R |z|>R

The first term in the latter sum can be estimated from above by

sup |u(x0 + z, a)| Kt0 −a (z)dz ≤ sup |u(x0 + z, a)| → 0
|z|≤R |z|≤R |z|≤R

as a → 0+, since u(x,t) is continuous up to the boundary (t = 0) and therefore

uniformly continuous on compact subsets there due to the fact that u(x0 + z, 0) = 0.
The second term can be estimated from above by

cε 0 −a))+ε |x0 +z|
e−|z|
2 /(4(t 2
dz → 0
(4π (t0 − a))n/2 |z|>R

as R → +∞ and a → 0. 
512 Part IV: Partial Differential Equations

Theorem 45.5. The kernel Kt (x) is a fundamental solution for the heat operator.

Proof. Given ε > 0, set


Kt (x), t ≥ ε ,
Kε (x,t) =
0, t < ε.

Clearly Kε (x,t) → Kt (x) as ε → 0 in the sense of distributions. Even more is true,

namely, Kε (x,t) → Kt (x) pointwise as ε → 0 and
  
|Kε (x,t)| dx = Kε (x,t)dx ≤ Kt (x)dx = 1.
Rn Rn Rn

We can therefore apply the dominated convergence theorem and obtain

 
lim Kε (x,t)dx = Kt (x)dx.
ε →0+ Rn Rn

So it remains to show that as ε → 0,

∂t Kε (x,t) − Δx Kε (x,t) → δ (x,t),

or
∂t Kε − Δx Kε , ϕ  → ϕ (0), ϕ ∈ C0∞ (Rn+1 ).

Using integration by parts, we obtain

 ∞ 
∂t Kε − Δx Kε , ϕ  = Kε , −∂t ϕ − Δ ϕ  = dt Kt (x)(−∂t − Δ )ϕ (x,t)dx
ε Rn
  ∞
=− dx Kt (x)∂t ϕ (x,t)dt
ε
 ∞R 
n

− dt Kt (x)Δx ϕ (x,t)dx
ε Rn
 ∞ 
= Kε (x)ϕ (x, ε )dx + dt ∂t Kt (x)ϕ (x,t)dx
ε
 R∞ Rn
n

− dt Δx Kt (x)ϕ (x,t)dx
ε Rn
 ∞ 
= Kε (x)ϕ (x, ε )dx + dt (∂t − Δ )Kt (x)ϕ (x,t)dx
ε
R Rn
n

= Kε (x)ϕ (x, ε )dx → ϕ (0, 0), ε → 0,

Rn

as we know from the proof of Theorem 45.1. 

Theorem 45.6. If f ∈ L1 (Rn+1 ), then

45 The Heat Operator 513
 t 
u(x,t) := ( f ∗ Kt )(x,t) ≡ ds Kt−s (x − y) f (y, s)dy
−∞ Rn

is well defined almost everywhere and is a distributional solution of ∂t u − Δ u = f .

Exercise 45.2. Prove Theorem 45.6.
Let us now consider the heat operator in a bounded domain Ω ⊂ Rn over a time
interval t ∈ [0, T ], 0 < T ≤ ∞. In this case, it is necessary to specify the initial
temperature u(x, 0), x ∈ Ω , and also to prescribe a boundary condition on ∂ Ω ×
[0, T ], see Figure 45.1.

∂Ω
t =T

t =0
Ω

Fig. 45.1 Geometry of the boundary value problem for the Heat equation in Ω and 0 ≤ t ≤ T .

The first basic result concerning such problems is the maximum principle.
Theorem 45.7. Let Ω be a bounded domain in Rn and 0 < T < ∞. Suppose u
is a real-valued continuous function on Ω × [0, T ] that satisfies ∂t u − Δ u = 0 in
Ω × (0, T ). Then u assumes its maximum and minimum either on Ω × {0} or on
∂ Ω × [0, T ].
Proof. Given ε > 0, set v(x,t) := u(x,t) + ε |x|2 . Then ∂t v − Δ v = −2nε . Suppose
0 < T  < T . If the maximum of v in Ω × [0, T  ] occurs at an interior point of Ω ×
(0, T  ), then the first derivatives ∇x,t v vanish there and the second derivative ∂ j2 v
for all j = 1, 2, . . . , n is nonpositive (consider v(x,t) a function of one variable x j ,
j = 1, 2, . . . , n). In particular, ∂t v = 0 and Δ v ≤ 0, which contradicts ∂t v − Δ v =
−2nε < 0 and Δ v = 2nε > 0.
Likewise, if the maximum occurs in Ω × {T  }, then we have ∂t v(x, T  ) ≥ 0 and
Δ v(x, T  ) ≤ 0, which contradicts ∂t v − Δ v < 0. Therefore,

max u ≤ max v ≤ max u + ε max |x|2 .

Ω ×[0,T  ] Ω ×[0,T  ] (Ω ×{0})∪(∂ Ω ×[0,T  ]) Ω
514 Part IV: Partial Differential Equations

It follows that for ε → 0 and T  → T ,

max u ≤ max u ≤ max u.

Ω ×[0,T ] (Ω ×{0})∪(∂ Ω ×[0,T ]) Ω ×[0,T ]

Replacing u by −u, we can obtain the same result for the minimum. 

Corollary 45.8 (Uniqueness). There is at most one continuous function u(x,t) in

Ω ×[0, T ], 0 < T < ∞, that agrees with a given continuous function f (x) in Ω ×{0},
with g(x,t) on ∂ Ω × [0, T ] and satisfies ∂t u − Δ u = 0.

Let us look now more closely at the following problem:

⎧
⎪
⎨∂t u − Δ u = 0, in Ω × (0, ∞),
u(x, 0) = f (x), in Ω , (45.4)
⎪
⎩
u(x,t) = 0, on ∂ Ω × (0, ∞).

This problem can be solved by the method of separation of variables. We begin by

looking for a solution of the form

u(x,t) = F(x)G(t).

Then
∂t u − Δ u = FG − GΔx F = 0

if and only if
G ΔF
= := −λ 2 ,
G F
or
G + λ 2 G = 0, Δ F + λ 2 F = 0,

for some constant λ . The first equation has the general solution

G(t) = ce−λ t ,
2

where c is an arbitrary constant. Without loss of generality we assume that c = 1. It

follows from (45.4) that

Δ F = −λ 2 F, in Ω ,
(45.5)
F = 0, on ∂ Ω ,

because u(x,t) = F(x)G(t) and G(0) = 1.

It remains to solve (45.5), which is an eigenvalue (spectral) problem for the
Laplacian with Dirichlet boundary condition. It is known that the problem (45.5) has
45 The Heat Operator 515

 ∞  ∞
infinitely many solutions Fj (x) j=1 with corresponding λ j2 . The numbers
j=1
−λ j2 are called eigenvalues, and the Fj (x) are called eigenfunctions of the Lapla-
 ∞
cian. It is also known that λ j > 0, j = 1, 2, . . ., λ j2 → ∞, and Fj (x) j=1 can be
  ∞
chosen as a complete orthonormal set in L2 (Ω ) (or Fj (x) j=1 forms an orthonor-
mal basis of L2 (Ω )). This fact allows us to represent f (x) in terms of Fourier series:
∞
f (x) = ∑ f j Fj (x), (45.6)
j=1

where f = ( f , Fj )L2 (Ω ) are called the Fourier coefficients of f with respect to

 ∞ j
Fj j=1 .
If we take now
∞
∑ f j Fj (x)e−λ j t ,
2
u(x,t) = (45.7)
j=1

then we may conclude (at least formally) that

∞ ∞
∂t u = − ∑ f j λ j2 Fj (x)e−λ j t = ∑ f j Δ Fj (x)e−λ j t = Δ u,
2 2

j=1 j=1

that is, u(x,t) from (45.7) satisfies the heat equation and u(x,t) = 0 on ∂ Ω × (0, ∞).
It remains to prove that u(x,t) satisfies the initial condition and to determine for
which functions f (x) the series (45.6) converges and in what sense. This is the main
question in the Fourier method.
It is clear that the series (45.6) and (45.7) (for t ≥ 0) converge in the sense of
L2 (Ω ). It is also clear that if f ∈ C1 (Ω ) vanishes at the boundary, then u will vanish
on ∂ Ω × (0, ∞), and one easily verifies that u is a distributional solution of the heat
equation (t > 0). Hence it is a classical solution, since u(x,t) ∈ C∞ (Ω × (0, ∞)) (see
Corollary 45.3).
Similar considerations apply to the problem
⎧
⎪
⎨∂t u − Δ u = 0, in Ω × (0, ∞),
u(x, 0) = f (x), in Ω
⎪
⎩
∂ν u(x,t) = 0, on ∂ Ω × (0, ∞).

This problem boils down to finding an orthonormal basis of eigenfunctions for the
Laplacian with the Neumann boundary condition. Let us remark that for this prob-
lem, {0} is always an eigenvalue and 1 is an eigenfunction.

Exercise 45.3. Prove that u(x,t) of the form (45.7) is a distributional solution of
the heat equation in Ω × (0, ∞).
516 Part IV: Partial Differential Equations

π
Exercise 45.4. Show that 0 |u(x,t)|2 dx is a decreasing function of t > 0, where
u(x,t) is the solution of

ut − uxx = 0, 0 < x < π ,t > 0,
u(0,t) = u(π ,t) = 0, t > 0.
Chapter 46
The Wave Operator

The wave equation is defined as

∂t2 u(x,t) − Δx u(x,t) = 0, (x,t) ∈ Rn × R. (46.1)

The wave equation is satisfied exactly by the components of the classical electro-
magnetic field in vacuum.
The characteristic variety of (46.1) is
 
charx (L) = (ξ , τ ) ∈ Rn+1 : (ξ , τ ) = 0, τ 2 = |ξ |2 ,

and it is called the light cone. Accordingly, we call

{(ξ , τ ) ∈ charx (L) : τ > 0}

and
{(ξ , τ ) ∈ charx (L) : τ < 0}

the forward and backward light cones, respectively.

The wave operator is a prototype of hyperbolic operators. This means that the
main symbol
∑ aα (x,t)ξ α τ j
|α |+ j=k

has k distinct real roots with respect to τ .

Theorem 46.1. Suppose u(x,t) is a C2 function and that ∂t2 u − Δ u = 0. Suppose

also that u = 0 and ∂ν u = 0 on the ball B = {(x, 0) : |x − x0 | ≤ t0 } in the hyperplane
t = 0. Then u = 0 in the region Ω = {(x,t) : 0 ≤ t ≤ t0 , |x − x0 | ≤ t0 − t}.

c Springer International Publishing AG 2017 517

V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7 46
518 Part IV: Partial Differential Equations

Proof. By considering real and imaginary parts we may assume that u is real. Define
Bt = {x : |x − x0 | < t0 − t}. Let us consider the following integral:
  
1
E(t) = (ut )2 + |∇x u|2 dx,
2 Bt

which represents the energy of the wave in Bt at time t. Next,


 
n
E  (t) = ut utt + ∑ ∂ j u(∂ j u)t dx
Bt j=1
  
1
− (ut )2 + |∇x u|2 dσ (x) =: I1 + I2 .
2 ∂ Bt

Straightforward calculations using the divergence theorem show us that


 
n n
I1 =
Bt
∑ ∂ j [(∂ j u)ut ] − ∑ (∂ j2 u)ut + ut utt dx
j=1 j=1
  n
=
Bt
ut (utt − Δx u)dx +
∂ Bt
∑ (∂ j u)ν j ut dσ (x)
j=1
   
1
≤ |ut | |∇x u|dσ (x) ≤ |ut |2 + |∇x u|2 dσ (x) ≡ −I2 .
∂ Bt 2 ∂ Bt

Hence
dE
≤ −I2 + I2 = 0.
dt

But E(t) ≥ 0 and E(0) = 0 due to the Cauchy data. Therefore, E(t) ≡ 0 if 0 ≤ t ≤ t0
and thus ∇x,t u = 0 in Ω . Since u(x, 0) = 0, it follows that u(x,t) = 0 also in Ω . 

(x0 ,t0 )

x0 t0

Fig. 46.1 Geometric illustration of the backward light cone at (x0 ,t0 ).
46 The Wave Operator 519

Remark 46.2. This theorem shows that the value of u at (x0 ,t0 ) depends only on the
Cauchy data of u in the ball {(x, 0) : |x − x0 | ≤ t0 }, see Figure 46.1.

Conversely, the Cauchy data in a region R in the initial (t = 0) hyperplane influ-

ence only those points inside the forward light cones issuing from points of R.
A similar result holds when the hyperplane t = 0 is replaced by a spacelike hyper-
surface S = {(x,t) : t = ϕ (x)}. A surface S is called spacelike if its normal vector
ν = (ν  , ν0 ) satisfies |ν0 | > |ν  | at every point of S, i.e., if ν lies inside the light
cone. This means that |∇ϕ | < 1.
Let us consider the Cauchy problem for the wave equation:

∂t2 u − Δ u = 0, x ∈ Rn ,t > 0,
(46.2)
u(x, 0) = f (x), ∂t u(x, 0) = g(x).

Definition 46.3. If ϕ is a continuous function on Rn and r > 0, we define the spher-

ical mean Mϕ (x, r) as follows:
 
1 1
Mϕ (x, r) := ϕ (z)dσ (z) = ϕ (x + ry)dσ (y).
rn−1 ωn |x−z|=r ωn |y|=1

Lemma 46.4. If ϕ is a C2 function on Rn , then Mϕ (x, 0) = ϕ (x) and


n−1
Δx Mϕ (x, r) = ∂r2 + ∂r Mϕ (x, r).
r

Proof. It is clear that


1
Mϕ (x, 0) = ϕ (x)dσ (y) = ϕ (x).
ωn |y|=1

For the second part we have by the divergence theorem that

 n 
1 1
∂r Mϕ (x, r) =
ωn ∑
|y|=1 j=1
y j ∂ j ϕ (x + ry)dσ (y) =
ωn |y|≤1
rΔ ϕ (x + ry)dy

1
= Δ ϕ (x + z)dz
rn−1 ωn |z|≤r
 r 
1
= n−1 ρ n−1 dρ Δ ϕ (x + ρ y)dσ (y).
r ωn 0 |y|=1

We therefore have

 rn−1 
∂r rn−1 ∂r Mϕ (x, r) = Δ ϕ (x + ry)dσ (y) ≡ rn−1 Δx Mϕ (x, r).
ωn |y|=1
520 Part IV: Partial Differential Equations

This implies that

(n − 1)rn−2 ∂r Mϕ (x, r) + rn−1 ∂r2 Mϕ (x, r) = rn−1 Δx Mϕ (x, r),

which proves the claim. 

Corollary 46.5. Suppose u(x,t) is a C2 function on Rn+1 and let
 
1 1
Mu (x, r,t) = u(z,t)dσ (z) = u(x + ry,t)dσ (y).
rn−1 ωn |x−z|=r ωn |y|=1

Then u(x,t) satisfies the wave equation if and only if


n−1
∂r +
2
∂r Mu (x, r,t) = ∂t2 Mu (x, r,t). (46.3)
r

Lemma 46.6. If ϕ ∈ Ck+1 (R), k ≥ 1, then

k−1  
∂ k
1
(r2k ϕ  ).
r
∂r2 ∂r r 2k−1
ϕ (r) =
r r

Proof. We employ induction on k. If k = 1, then

k−1  
1
∂r2 ∂r r2k−1 ϕ (r) = ∂r2 (rϕ ) = ∂r (ϕ + rϕ  ) = 2ϕ  + rϕ 
r

and
k

∂r  ∂r
(r ϕ ) =
2k
(r2 ϕ  ) = 2ϕ  + rϕ  .
r r

Assume that
k−1 
1 
∂ k
(r2k ϕ  ).
r
∂r2 ∂r r2k−1 ϕ (r) =
r r

Then

k

k−1
 
1  1 ∂r 2k+1 
∂r2 ∂r r 2k+1
ϕ (r) = ∂r2 ∂r r ϕ
r r r

k−1  
2 1
= ∂r ∂r (2k + 1)r2k−1 ϕ + r2k ϕ 
r

k−1  
k−1  
1 1
= (2k + 1)∂r2 ∂r r2k−1 ϕ + ∂r2 ∂r r2k ϕ 
r r

k  
k
∂r ∂r
= (2k + 1) r2k ϕ  + (r2k (rϕ  ) )
r r
46 The Wave Operator 521

k  
∂r
= (2k + 1)r2k ϕ  + r2k (rϕ  )
r

k  
∂r
= (2k + 1)r2k ϕ  + r2k ϕ  + r2k+1 ϕ 
r

k  
∂r
= (2k + 2)r2k ϕ  + r2k+1 ϕ 
r

k+1  
∂r
= r2k+2 ϕ  .
r

By the principle of induction, the proof is complete. 

Corollary 46.5 gives that if u(x,t) is a solution of the wave equation (46.1) in
Rn × R, then Mu (x, r,t) satisfies (46.3), i.e.,


n−1
∂r2 + ∂r Mu = ∂t2 Mu ,
r

with initial conditions

Mu (x, r, 0) = M f (x, r), ∂t Mu (x, r, 0) = Mg (x, r), (46.4)

since u(x, 0) = f (x) and ∂t u(x, 0) = g(x).

Let us set

 n−3
∂r 2  n−2
u(x, r,t) := r Mu ≡ T Mu ,
r (46.5)
f(x, r) := T M f , g(x, r) := T Mg ,

for n = 2k + 1, k = 1, 2, . . ..

Lemma 46.7. The following is true:

∂r2 u = ∂t2 u,
(46.6)
u|t=0 = f, ∂t u|t=0 = g,

where u, f, and g are defined in (46.5).

Proof. Since n = 2k + 1, we have n−3

2 = k − 1 and n − 2 = 2k − 1. Hence we obtain
from Lemmas 46.4 and 46.6 that

k−1  
k  
∂r ∂r
∂r2 u = ∂r2 T Mu = ∂r2 r 2k−1
Mu = r2k ∂r Mu
r r

k−1  
∂r
= 2kr2k−2 ∂r Mu + r2k−1 ∂r2 Mu
r
522 Part IV: Partial Differential Equations




k−1  
∂r k−1 2k−1 2 n−1 ∂r
= r ∂r Mu + ∂r Mu = r2k−1 ∂t2 Mu
r r r

k−1  
∂r
= ∂t2 r2k−1 Mu = ∂t2 u.
r

Moreover, the initial conditions are satisfied due to (46.4) and (46.5). 

But now, since (46.6) is a one-dimensional problem, we may conclude that

u(x, r,t) from Lemma 46.7 is equal to
  r+t 
1
u(x, r,t) = f(x, r + t) + f(x, r − t) + g(x, s)ds . (46.7)
2 r−t

Lemma 46.8. If n = 2k + 1, k = 1, 2, . . ., then

u(x, r,t)
u(x,t) = Mu (x, 0,t) = lim ,
r→0 (n − 2)!!r

where (n − 2)!! = 1 · 3 · 5 · · · (n − 2) is the solution of (46.2). We have even more,

namely,
1  
u(x,t) = ∂r f|r=t + g(x,t) . (46.8)
(n − 2)!!

Proof. By (46.5) we have

k−1  
k−2  
∂r ∂r
u(x, r,t) = r 2k−1
Mu = (2k − 1)r2k−3 Mu + r2k−2 ∂r Mu
r r
= (2k − 1)(2k − 3) · · · 1 · Mu r + O(r2 ), r → 0,

or
u(x, r,t)
= Mu + O(r), r → 0.
(n − 2)!!r

Hence
u(x, r,t)
Mu (x, 0,t) = lim .
r→0 (n − 2)!!r

But by definition of Mu we have that Mu (x, 0,t) = u(x,t), where u(x,t) is the solution
of (46.2). The initial conditions in (46.2) are satisfied due to (46.5). Next, since
u(x, r,t) satisfies (46.7), we have
46 The Wave Operator 523
 

u(x, r,t) 1 f(x, r + t) + f(x, r − t) 1 r+t
lim = lim + g(x, s)ds
r→0 (n − 2)!!r 2(n − 2)!! r→0 r r r−t
1  
= ∂r f|r=t + ∂r f|r=−t + g(x,t) − g(x, −t) ,
2(n − 2)!!

because f(x,t) and g(x,t) are odd functions of t. We therefore finally obtain

u(x, r,t) 1  
lim = ∂r f|r=t + g(x,t) ,
r→0 (n − 2)!!r (n − 2)!!

and the proof is complete. 

Now we are in a position to prove the main theorem for odd n ≥ 3.
n+3 n+1
Theorem 46.9. Suppose that n ≥ 3 is odd. If f ∈ C 2 (Rn ) and g ∈ C 2 (Rn ), then

 n−3
 
1 ∂t 2
u(x,t) = ∂t t n−2
f (x + ty)dσ (y)
(n − 2)!!ωn t |y|=1

 n−3
  (46.9)
∂t 2
+ t n−2
g(x + ty)dσ (y)
t |y|=1

solves the Cauchy problem (46.2).

Proof. Due to Lemmas 46.7 and 46.8, u(x,t) given by (46.8) is the solution of the
wave equation. It remains only to check that this u satisfies the initial conditions.
But (46.9) gives us for small t that

u(x,t) = M f (x,t) + tMg (x,t) + O(t 2 ),

which implies that

u(x, 0) = M f (x, 0) = f (x), ∂t u(x, 0) = ∂t M f (x, 0) + Mg (x, 0) = g(x).

The last equality follows from the fact that M f (x,t) is even in t, and so its derivative
vanishes at t = 0.
Remark 46.10. If n = 3, then (46.9) becomes

   
1
u(x,t) = ∂t t f (x + ty)dσ (y) + t g(x + ty)dσ (y)
4π |y|=1 |y|=1
 
1
≡ f (x + ty)dσ (y) + t ∇ f (x + ty) · yd σ (y)
4π |y|=1 |y|=1
 
+t g(x + ty)dσ (y) .
|y|=1
524 Part IV: Partial Differential Equations

The solution of (46.2) for even n is readily derived from the solution for odd n
by the “method of descent”. This is just a trivial observation: if u is a solution of the
wave equation in Rn+1 × R that does not depend on xn+1 , then u satisfies the wave
equation in Rn × R. Thus to solve (46.2) in Rn × R with even n, we think of f and g
as functions on Rn+1 that are independent of xn+1 .
n+4 n+2
Theorem 46.11. Suppose that n is even. If f ∈ C 2 (Rn ) and g ∈ C 2 (Rn ), then
the function

 n−2  

2 ∂t 2 f (x + ty)
u(x,t) = ∂t t n−1
 dy
(n − 1)!!ωn+1 t |y|≤1 1 − y2

 n−2  
 (46.10)
∂t 2 g(x + ty)
+ t n−1
 dy
t |y|≤1 1 − y2

solves the Cauchy problem (46.2).

Proof. If n is even, then n + 1 is odd and n + 1 ≥ 3. We can therefore apply (46.9)

in Rn+1 × R to get that

1
u(x,t) =
(n − 1)!!ωn+1

 n−2  

∂t 2
× ∂t t n−1
f (x + ty + tyn+1 )dσ (
y) (46.11)
t y21 +···+y2n +y2n+1 =1

 n−2  

∂t 2
+ t n−1
g(x + ty + tyn+1 )dσ ( y) ,
t y21 +···+y2n +y2n+1 =1

where y = (y, yn+1 ), solves (46.2) in Rn+1 × R (formally). But if we assume now
that f and g do not depend on xn+1 , then u(x,t) does not depend on xn+1 either and
solves (46.2) in Rn × R. It remains only to calculate the integrals in (46.11) under
this assumption. We have
 
f (x + ty + tyn+1 )dσ (
y) = f (x + ty)dσ (
y)
|y|2 +y2n+1 =1 |y|2 +y2n+1 =1

dy
=2 f (x + ty)  ,
|y|≤1 1 − |y|2

because we have the upper and lower hemispheres of the sphere |y|2 + y2n+1 = 1.
Similarly for the second integral in (46.11). This proves the theorem. 

Remark 46.12. If n = 2, then (46.10) becomes

46 The Wave Operator 525
    
1 f (x + ty) g(x + ty)
u(x,t) = ∂t t  dy + t  dy .
2π |y|≤1 1 − y2 |y|≤1 1 − y2

Now we consider the Cauchy problem for the inhomogeneous wave equation

∂t2 u − Δx u = w(x,t),
(46.12)
u(x, 0) = f (x), ∂t u(x, 0) = g(x).

We look for the solution u(x,t) of (46.12) as u = u1 + u2 , where

∂t2 u1 − Δ u1 = 0,
(A)
u1 (x, 0) = f (x), ∂t u1 (x, 0) = g(x),

and
∂t2 u2 − Δ u2 = w,
(B)
u2 (x, 0) = ∂t u2 (x, 0) = 0.

For the problem (B) we will use a method known as Duhamel’s principle.

Theorem 46.13. Suppose w ∈ C[ 2 ]+1 (Rn × R). For s ∈ R let v(x,t; s) be the solu-
n

tion of
∂t2 v(x,t; s) − Δx v(x,t; s) = 0,
v(x, 0; s) = 0, ∂t v(x, 0; s) = w(x, s).

Then  t
u(x,t) := v(x,t − s; s)ds
0

solves (B).

Proof. By definition of u(x,t) it is clear that u(x, 0) = 0. We also have

 t
∂t u(x,t) = v(x, 0;t) + ∂t v(x,t − s; s)ds.
0

This implies that ∂t u(x, 0) = v(x, 0; 0) = 0. Differentiating once more in t, we get

 t
∂t2 u(x,t) = ∂t (v(x, 0;t)) + ∂t v(x, 0;t) + ∂t2 v(x,t − s; s)ds
0
 t
= w(x,t) + Δx v(x,t − s; s)ds
0
 t
= w(x,t) + Δx v(x,t − s; s)ds = w(x,t) + Δx u.
0
526 Part IV: Partial Differential Equations

Thus u solves (B), and the theorem is proved. 

Let us consider again the homogeneous Cauchy problem (46.2). Applying the
Fourier transform with respect to x gives

∂t2 u(ξ ,t) + |ξ |2 u(ξ ,t) = 0,
u(ξ , 0) = f(ξ ), ∂t u(ξ , 0) = g(ξ ).

But this ordinary differential equation with initial conditions can be easily solved to
obtain


sin(|ξ |t)  sin(|ξ |t) sin(|ξ |t)
u(ξ ,t) = f(ξ ) cos(|ξ |t) + g(ξ ) ≡ f (ξ )∂t + g(ξ ) .
|ξ | |ξ | |ξ |

This implies that



sin(|ξ |t) sin(|ξ |t)
u(x,t) = F −1 f(ξ )∂t + F −1 g(ξ )
|ξ | |ξ |





sin(|ξ |t) sin(|ξ |t)
= f ∗ ∂t (2π )−n/2 F −1 + g ∗ (2π )−n/2 F −1
|ξ | |ξ |
= f ∗ ∂t Φ (x,t) + g ∗ Φ (x,t), (46.13)
 
ξ |t)
where Φ (x,t) = (2π )−n/2 F −1 sin(||ξ |
.
The next step is to try to solve the equation

∂t2 F(x,t) − Δx F(x,t) = δ (x)δ (t).

By taking the Fourier transform in x we obtain

 ξ ,t) + |ξ |2 F(
∂t2 F(  ξ ,t) = (2π )−n/2 δ (t).

Therefore, F must be a solution of ∂t2 u + |ξ |2 u = 0 for t = 0, and so

 ξ ,t) = a(ξ ) cos(|ξ |t) + b(ξ ) sin(|ξ |t), t < 0,
F(
c(ξ ) cos(|ξ |t) + d(ξ ) sin(|ξ |t), t > 0.

To obtain the delta function at t = 0 we require that F is continuous at t = 0, but ∂t F

has a jump of size (2π )−n/2 at t = 0. So we have

a(ξ ) = c(ξ ), |ξ |(d(ξ ) − b(ξ )) = (2π )−n/2 .

This gives two equations for the four unknown coefficients a, b, c, d. But it is rea-
sonable to require F(x,t) ≡ 0 for t < 0. Hence, a = b = c = 0 and d = (2π )−n/2 |ξ1 | .
46 The Wave Operator 527

Therefore, ξ |t)
(2π )−n/2 sin(| , t > 0,
 ξ ,t) =
F( |ξ | (46.14)
0, t < 0.

If we compare (46.13) and (46.14), we may conclude that


−n/2 sin(|ξ |t)
F(x,t) = (2π ) Fξ−1 , t > 0,
|ξ |

and
Φ (x,t), t > 0,
Φ+ (x,t) =
0, t < 0,

is the fundamental solution of the wave equation, i.e., F(x,t) with t > 0.
There is one more observation. If we compare (46.9) and (46.10) with (46.13),
then we may conclude that these three formulas are the same. Hence, we may cal-
culate the inverse Fourier transform of

sin(|ξ |t)
(2π )−n/2
|ξ |

in odd and even dimensions respectively with (46.9) and (46.10). In fact, the result
is presented in these two formulas.
In solving the wave equation in the region Ω × (0, ∞), where Ω is a bounded
domain in Rn , it is necessary to specify not only the Cauchy data on Ω ×{0} but also
some conditions on ∂ Ω × (0, ∞) to tell the wave what to do when it hits the bound-
ary. If the boundary conditions on ∂ Ω × (0, ∞) are independent of t, the method of
separation of variables can be used.
Let us (for example) consider the following problem:
⎧
⎪
⎨∂t u − Δx u = 0,
2 in Ω × (0, ∞),
u(x, 0) = f (x), ∂t u(x, 0) = g(x), in Ω , (46.15)
⎪
⎩
u(x,t) = 0, on ∂ Ω × (0, ∞).

We can look for a solution u in the form u(x,t) = F(x)G(t) and get

Δ F(x) + λ 2 F(x) = 0, in Ω ,
(46.16)
F(x) = 0, on ∂ Ω ,

and
G (t) + λ 2 G(t) = 0, 0 < t < ∞. (46.17)
528 Part IV: Partial Differential Equations

The general solution of (46.17) is G(t) = a cos(λ t) +b sin(

∞λ t). Since (46.16) has
 ∞
infinitely many solutions Fj j=1 with corresponding λ j 2 , λ j2 → +∞ (λ j > 0),
 ∞ j=1
and Fj j=1 that can be chosen as an orthonormal basis in L2 (Ω ), the solution
u(x,t) of (46.15) is of the form
∞
u(x,t) = ∑ Fj (x) (a j cos(λ jt) + b j sin(λ jt)) . (46.18)
j=1

At the same time, f (x) and g(x) have the L2 (Ω ) representations

∞ ∞
f (x) = ∑ f j Fj (x), g(x) = ∑ g j Fj (x), (46.19)
j=1 j=1

where f j = ( f , Fj )L2 and g j = (g, Fj )L2 . It follows from (46.15) and (46.18) that

∞ ∞
u(x, 0) = ∑ a j Fj (x), ut (x, 0) = ∑ λ j b j Fj (x). (46.20)
j=1 j=1

Since (46.19) must be satisfied also, we obtain a j = f j and b j = 1

λ j g j. Therefore,
the solution u(x,t) of (46.15) has the form
∞

1
u(x,t) = ∑ j F (x) f j cos(λ j t) +
λj
g j sin(λ j t) .
j=1

 ∞
The series (46.18), (46.19), and (46.20) converge in L2 (Ω ), because Fj j=1 is an
orthonormal basis in L2 (Ω ). It remains only to investigate the convergence of these
series in stronger norms (which depends on f and g, or more precisely, on their
smoothness).
The Neumann problem with ∂ν u(x,t) = 0, x ∈ ∂ Ω , can be considered in a similar
manner.
References

1. Adams R A and Fournier J J F, Sobolev Spaces, Elsevier, Oxford, 2003

2. Agmon S, Spectral properties of Schrödinger operators and scattering theory, Ann. Scuola
Norm. Sup. Pisa Cl. Sci., 4 (1975), 151–218
3. Bergh J and Löfström J, Interpolation Spaces: An Introduction, Springer-Verlag, Berlin Hei-
delberg, 1976
4. Birman M S and Solomjak M Z, Spectral Theory of Self-Adjoint Operators in Hilbert Space,
Kluwer, 1987
5. Blanchard P and Brüning E, Mathematical Methods in Physics, Birkhäuser, 1993
6. Cakoni F and Colton D, A Qualitative Approach To Inverse Scattering Theory, Springer-
Verlag, 2014
7. Colton D and Kress R, Integral Equation Methods in Scattering Theory, Wiley, 1983
8. Colton D, Partial Differential Equations: An Introduction, Dover, 2004
9. Cycon H L, Froese R G, Kirsch W and Simon B, Schrödinger Operators, Academic Press,
New-York, 1989
10. Evans L C, Partial Differential Equations, American Mathematical Society, 1998
11. Folland G B, Introduction to Partial Differential Equations, second edition, Princeton Univer-
sity Press, 1995
12. Folland G B, Fourier Analysis and its Applications, Brooks/Cole Publishing Company, 1992
13. Folland G B, Real Analysis, John Wiley, New York, 1984
14. Gårding L, Singularities in Linear Wave Propagation, Springer, 1987
15. Gårding L, Some Points of Analysis and Their History, American Mathematical Society, 1997
16. Gilbarg D and Trudinger N S, Elliptic Partial Differential Equations of Second Order,
Springer, 2001
17. Gonzalez-Velasco E A, Fourier Analysis and Boundary Value Problems, Academic Press, San
Diego, 1995
18. Grafakos L, Classical and Modern Fourier Analysis, Pearson Education Inc., 2004
19. Hörmander L, The Analysis of Linear Partial Differential Operators I: Distribution Theory
and Fourier Analysis, second edition, Springer-Verlag, Berlin Heidelberg, 2003
20. Ikebe I, Eigenfunction expansions associated with the Schroedinger operators and their appli-
cations to scattering theory, Arch. Rational Mech. Anal., 5 (1960), 1–34
21. Kato T, Growth properties of solutions of the reduced wave equation with a variable coeffi-
cient, Comm. Pure Appl. Math., 12 (1959), 403–425
22. Kress, R, Linear Integral Equations, second edition, Springer-Verlag, New York, 1999
23. Lebedev N N, Special Functions and Their Applications, Dover, New York, 1972
c Springer International Publishing AG 2017 529
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7
530 References

24. Maurin K, Methods of Hilbert Spaces, Polish Scientific Publishers, Warsaw, 1967
25. McLean W, Strongly Elliptic Systems and Boundary Integral Equations, Cambridge Univer-
sity Press, 2000
26. Nikolskii S M, Approximation of Functions of Several Variables and Imbedding Theorems,
Springer-Verlag, Berlin, Heidelberg, New York, 1975
27. Nirenberg L and Walker H, Nullspaces of elliptic partial differential operators in Rn , J. Math.
Anal. Appl., 42 (1973), 271–301.
28. Ouhabaz E M, Analysis of Heat Equations on Domains, Princeton University Press, 2005
29. Pinsky M A, Introduction to Fourier Analysis and Wavelets, Brooks/Cole, 2002
30. Päivärinta L and Serov V, Recovery of singularities of a multidimensional scattering potential,
SIAM J. Math. Anal., 29 (1998), 697–711
31. Päivärinta L and Somersalo E, Inversion of discontinuities for the Schrödinger equation in
three dimensions, SIAM J. Math. Anal., 22 (1991), 480–499
32. Saito Y, Some properties of the scattering amplitude and inverse scattering problem, Osaka,
J. Math.19 (1982), 527–547
33. Saranen J and Vainikko G, Periodic Integral & Pseudodifferential Equations with Numerical
Approximation, Springer, 2001
34. Shubin M A, Pseudodifferential Operators and Spectral Theory, Springer-Verlag, Berlin, 1987
35. Simon B, A Comprehensive Course in Analysis, American Mathematical Society, 2015
36. Stakgold I, Green’s Functions and Boundary Value Problems, second edition, John Wiley,
1997
37. Stein E M, Singular Integrals and Differentiability Properties of Functions, Princeton Univer-
sity Press, Princeton, NJ, 1970
38. Stein E M and Shakarchi R, Fourier Analysis: An Introduction, Princeton University Press,
2003
39. Taylor M E, Partial Differential Equations: Basic Theory, Springer-Verlag, New York, 1996
40. Taylor M E, Partial Differential Equations II: Qualitative Studies of Linear Equations,
Springer-Verlag, New York, 1996
41. Titchmarsh E C, Eigenfunction Expansions Associated with Second-order Differential Equa-
tions, vol. 1, Oxford University Press, Oxford, 1946
42. Triebel H, Theory of Function Spaces, Birkhäuser, Basel, 1983
43. Watson G N, A Treatise on the Theory of Bessel Functions, Cambridge University Press, Cam-
bridge, UK, 1996
44. Wong M W, Discrete Fourier Analysis, Springer Basel AG, 2011
45. Vretblad A, Fourier Analysis and Its Applications, Springer-Verlag, New York, 2003
46. Zigmund A, Trigonometric Series, Vol. II, Cambridge University Press, 1959
Index

Symbols Cayley transform, 291

L2 Hölder condition, 53 characteristic, 406, 407
L p -modulus of continuity, 34 characteristic form, 406
δ -function, 402 characteristic variety, 406
p-integrable function, 4 closable operator, 268
Closed graph theorem, 273
closed operator, 268
A closed subspace, 254
absolute convergence, 17 closure of operator, 268
adjoint operator, 269 collectively compact, 380
approximation to the identity, 398 compact operator, 302
Ascoli–Arzelà theorem, 365 compact set, 393
complete space, 251
completeness relation, 258
B conjugate linear functional, 476
Banach–Steinhaus theorem, 248, 264 conjugate Poisson kernel, 19
band-limited signal, 95 constant of ellipticity, 320
Basic criterion of self-adjointness, 276 continuous kernel, 454
basis, 257 continuous spectrum, 299
Bernstein’s theorem, 54 convergent sequence, 251
Besov space, 46 convolution, 24, 80, 167, 397
Bessel’s inequality, 38, 250 criterion for closedness, 268
biharmonic equation, 406
Bitsadze’s example, 472
bounded operator, 261 D
Burgers equation, 413 d’Alembert’s formula, 118
degenerate kernel, 383
densely defined, 261
C denseness property, 386
Cauchy data, 414 diamagnetic inequality, 352
Cauchy problem, 414 differential operator, 405
Cauchy sequence, 251 dilation, 138
Cauchy–Bunyakovsky–Schwarz inequality, direct Born approximation, 241
251 Dirichlet boundary value problem, 472
Cauchy–Kowalevski theorem, 415 Dirichlet form, 471
Cauchy–Riemann operator, 163, 407, 428 Dirichlet kernel, 59
c Springer International Publishing AG 2017 531
V. Serov, Fourier Series, Fourier Transform and Their Applications
to Mathematical Physics, Applied Mathematical Sciences 197,
DOI 10.1007/978-3-319-65262-7
532 Index

Dirichlet problem, 122 G

Dirichlet–Jordan theorem, 64 Gårding’s inequality, 324, 474
discrete Fourier transform, 78 Galerkin projection method, 388
discrete spectrum, 299 Gaussian kernel, 209, 508
distribution, 401 generalized ellipticity condition, 320
distributional solution, 405 generalized Leibniz formula, 320
divergence theorem, 397 generalized Minkowski inequality, 4
domain, 261 generalized strong ellipticity, 471
double layer potential, 454 gradient, 406
Duhamel’s principle, 525 graph, 267
Green’s function, 333, 355, 438
Green’s identities, 424
E
eigenvalue, 299 H
eigenvalue problem, 101 Hölder condition with exponent α , 8
eikonal equation, 406 Hölder space, 8, 46
elliptic differential operator, 319, 406 Hölder’s inequality, 4
elliptic operator, 208 Hadamard’s example, 416
equi-bounded strongly continuous semi- Hans Lewy example, 418
group, 328 Hardy’s theorem, 69
essential spectrum, 299 harmonic at infinity, 451
essentially self-adjoint operator, 273 harmonic function, 424
even extension, 13 Harnack’s inequality, 448
even function, 5 Hausdorff–Young theorem, 148
evolution equation, 406 heat equation, 99
extension, 268 heat kernel, 209, 355
exterior Dirichlet problem, 451 heat operator, 507
exterior impedance boundary value prob- Hellinger–Toeplitz theorem, 265
lem, 485 Helmholtz operator, 213
exterior Neumann problem, 452 Hilbert space, 251
Hilbert transform, 200
Hilbert–Schmidt norm, 262
homogeneous distribution, 195
F
hyperplane, 396
Fejér kernel, 27
hypersurface, 396
Fejér means, 27
finite-rank operator, 302
formally self-adjoint, 320 I
Fourier coefficient, 14 idempotent, 279
Fourier cosine series, 12 ill-posed problem, 418
Fourier expansion, 257 infinitesimal generator, 328
Fourier inversion formula, 86, 139, 400 injective operator, 308
Fourier series, 12 inner product, 249
Fourier transform, 85, 135, 399 inner product space, 249
Fourier’s method, 100 integrable function, 4
Fourier sine series, 12 integral curves, 408
Fredholm alternative, 467 integral equations of the second and first
Fredholm alternative II, 366 kinds, 369
Friedrichs extension, 316 integral operator, 359
Fubini’s theorem, 4 interior and exterior normal derivatives, 451
function of bounded variation, 6 interior Dirichlet problem, 451
fundamental period, 3 interior Neumann problem, 452
fundamental solution, 207, 427 inverse Born approximation, 495
Index 533

inverse discrete Fourier transform, 78 normal operator, 278

inverse Fourier transform, 85 null space, 261
inverse scattering problem, 493
isometry, 282
O
odd extension, 13
J odd function, 6
John von Neumann’s spectral theorem, 287 open ball, 251
operator with weak singularity, 359
Optical lemma, 229
K orthogonal, 250
kernel, 261, 359 orthogonal complement, 254
Korteweg–de Vries equation, 406 orthogonal functions, 9
orthonormal, 250
orthonormal basis, 257
L
Laplace equation, 19, 121
Laplace operator, 406, 421 P
Lax–Milgram theorem, 476 parallelogram law, 251
Lebesgue space, 253 Parseval’s equality, 23, 38, 79, 140, 258, 401
Leibniz formula, 395 partial sum of the Fourier series, 27
Lemma of Riesz, 307 Peetre’s theorem, 55
length, 250 periodic extension, 3
linear operator, 261 periodic function, 3
linear space, 249 periodization, 93
linear span, 257 piecewise constant function, 6
linear superposition principle, 103 piecewise continuous function, 6
Liouville’s theorem, 427 Plancherel’s theorem, 145, 401
Lippmann–Schwinger equation, 225, 235, Poincaré’s inequality, 188
347 point spectrum, 299
pointwise convergence, 17
Poisson equation, 406
M Poisson identity, 94
magnetic Schrödinger operator, 235, 349 Poisson integral, 442
maximum principle, 426, 513, 515 Poisson kernel, 19, 442
mean square distance, 37 Poisson summation formula, 93
mean value theorem, 425 polarization identity, 252
method of characteristics, 409 positive operator, 281
Minkowski’s inequality, 4 precompact set, 302
modulus of continuity, 8 principal symbol, 208, 319, 406
multi-index, 319, 394 projection method, 385
mutually orthogonal functions, 9 Projection theorem, 254
projector, 279
Pythagorean Theorem, 250
N
neighborhood, 251
Neumann problem, 122 Q
Nikol’skii space, 46 quadratic form, 313
non-characteristic, 407 quasilinear equation, 411
non-negative operator, 281
norm, 251
norm induced by the inner product, 251 R
norm topology, 251 range, 261
normal vector, 396 Reflection principle, 448
534 Index

regular distribution, 401 square-integrable functions, 37

Regularity theorem in Hölder spaces, 431 Stieltjes integral, 7
Regularity theorem in Sobolev spaces, 474 strong ellipticity condition, 507
relatively compact, 380 successive approximations, 382
Rellich’s lemma, 489 summation by parts formula, 20
removable singularity, 448 support, 401
resolvent, 292 surjective operator, 308
resolvent identity, 295 symmetric operator, 273
resolvent set, 295
restriction, 268
Riemann–Lebesgue lemma, 33, 85, 399 T
Riesz kernel, 201 tail sum, 21
Riesz potential, 203 telegrapher’s equation, 406
Riesz’s theorem, 308, 467 tempered distribution, 402
Riesz transform, 201 total variation, 7
Riesz–Fischer theorem, 39 translation, 138, 421
Riesz-Fréchet theorem, 256 triangle inequality, 251
Riesz–Thorin interpolation theorem, 146 trigonometric series, 12
right and left limit, 6 tubular neighborhood, 463
rotation, 421

U
S uniform boundedness principle, 248, 263
scalar product, 249 uniform convergence, 17
scattering amplitude, 228, 232, 240 unitary operator, 282
scattering problem, 224
scattering solutions, 235, 347
Schwartz space, 134, 400 V
self-adjoint operator, 273 vector space, 249
semibounded from below, 313 Volterra integral equations of the first and
separable, 257 second kinds, 371
separation of variables, 100, 514
sequence space, 44, 252
sesquilinear form, 476 W
Sine–Gordon equation, 406 wave equation, 113, 406, 517
single layer potential, 454 wave operator, 517
Sobolev embedding theorem, 185 weak convergence, 268
Sobolev inequality, 222 weak solution, 405
Sobolev space, 8, 46, 86, 177, 188, 254 well-posed problem, 418
Sommerfeld radiation condition, 225 Whittaker–Shannon–Boas theorem, 95
spectral family, 282
spectral function, 332
spectrum, 295 Y
spherical mean, 519 Young’s inequality for convolution, 397